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/bpf-cgroup.h>
8 #include <linux/kernel.h>
9 #include <linux/types.h>
10 #include <linux/slab.h>
11 #include <linux/bpf.h>
12 #include <linux/btf.h>
13 #include <linux/bpf_verifier.h>
14 #include <linux/filter.h>
15 #include <net/netlink.h>
16 #include <linux/file.h>
17 #include <linux/vmalloc.h>
18 #include <linux/stringify.h>
19 #include <linux/bsearch.h>
20 #include <linux/sort.h>
21 #include <linux/perf_event.h>
22 #include <linux/ctype.h>
23 #include <linux/error-injection.h>
24 #include <linux/bpf_lsm.h>
25 #include <linux/btf_ids.h>
26 #include <linux/poison.h>
27 #include <linux/module.h>
31 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
32 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
33 [_id] = & _name ## _verifier_ops,
34 #define BPF_MAP_TYPE(_id, _ops)
35 #define BPF_LINK_TYPE(_id, _name)
36 #include <linux/bpf_types.h>
42 /* bpf_check() is a static code analyzer that walks eBPF program
43 * instruction by instruction and updates register/stack state.
44 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
46 * The first pass is depth-first-search to check that the program is a DAG.
47 * It rejects the following programs:
48 * - larger than BPF_MAXINSNS insns
49 * - if loop is present (detected via back-edge)
50 * - unreachable insns exist (shouldn't be a forest. program = one function)
51 * - out of bounds or malformed jumps
52 * The second pass is all possible path descent from the 1st insn.
53 * Since it's analyzing all paths through the program, the length of the
54 * analysis is limited to 64k insn, which may be hit even if total number of
55 * insn is less then 4K, but there are too many branches that change stack/regs.
56 * Number of 'branches to be analyzed' is limited to 1k
58 * On entry to each instruction, each register has a type, and the instruction
59 * changes the types of the registers depending on instruction semantics.
60 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
63 * All registers are 64-bit.
64 * R0 - return register
65 * R1-R5 argument passing registers
66 * R6-R9 callee saved registers
67 * R10 - frame pointer read-only
69 * At the start of BPF program the register R1 contains a pointer to bpf_context
70 * and has type PTR_TO_CTX.
72 * Verifier tracks arithmetic operations on pointers in case:
73 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
74 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
75 * 1st insn copies R10 (which has FRAME_PTR) type into R1
76 * and 2nd arithmetic instruction is pattern matched to recognize
77 * that it wants to construct a pointer to some element within stack.
78 * So after 2nd insn, the register R1 has type PTR_TO_STACK
79 * (and -20 constant is saved for further stack bounds checking).
80 * Meaning that this reg is a pointer to stack plus known immediate constant.
82 * Most of the time the registers have SCALAR_VALUE type, which
83 * means the register has some value, but it's not a valid pointer.
84 * (like pointer plus pointer becomes SCALAR_VALUE type)
86 * When verifier sees load or store instructions the type of base register
87 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
88 * four pointer types recognized by check_mem_access() function.
90 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
91 * and the range of [ptr, ptr + map's value_size) is accessible.
93 * registers used to pass values to function calls are checked against
94 * function argument constraints.
96 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
97 * It means that the register type passed to this function must be
98 * PTR_TO_STACK and it will be used inside the function as
99 * 'pointer to map element key'
101 * For example the argument constraints for bpf_map_lookup_elem():
102 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
103 * .arg1_type = ARG_CONST_MAP_PTR,
104 * .arg2_type = ARG_PTR_TO_MAP_KEY,
106 * ret_type says that this function returns 'pointer to map elem value or null'
107 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
108 * 2nd argument should be a pointer to stack, which will be used inside
109 * the helper function as a pointer to map element key.
111 * On the kernel side the helper function looks like:
112 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
114 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
115 * void *key = (void *) (unsigned long) r2;
118 * here kernel can access 'key' and 'map' pointers safely, knowing that
119 * [key, key + map->key_size) bytes are valid and were initialized on
120 * the stack of eBPF program.
123 * Corresponding eBPF program may look like:
124 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
125 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
126 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
127 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
128 * here verifier looks at prototype of map_lookup_elem() and sees:
129 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
130 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
132 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
133 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
134 * and were initialized prior to this call.
135 * If it's ok, then verifier allows this BPF_CALL insn and looks at
136 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
137 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
138 * returns either pointer to map value or NULL.
140 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
141 * insn, the register holding that pointer in the true branch changes state to
142 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
143 * branch. See check_cond_jmp_op().
145 * After the call R0 is set to return type of the function and registers R1-R5
146 * are set to NOT_INIT to indicate that they are no longer readable.
148 * The following reference types represent a potential reference to a kernel
149 * resource which, after first being allocated, must be checked and freed by
151 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
153 * When the verifier sees a helper call return a reference type, it allocates a
154 * pointer id for the reference and stores it in the current function state.
155 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
156 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
157 * passes through a NULL-check conditional. For the branch wherein the state is
158 * changed to CONST_IMM, the verifier releases the reference.
160 * For each helper function that allocates a reference, such as
161 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
162 * bpf_sk_release(). When a reference type passes into the release function,
163 * the verifier also releases the reference. If any unchecked or unreleased
164 * reference remains at the end of the program, the verifier rejects it.
167 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
168 struct bpf_verifier_stack_elem {
169 /* verifer state is 'st'
170 * before processing instruction 'insn_idx'
171 * and after processing instruction 'prev_insn_idx'
173 struct bpf_verifier_state st;
176 struct bpf_verifier_stack_elem *next;
177 /* length of verifier log at the time this state was pushed on stack */
181 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
182 #define BPF_COMPLEXITY_LIMIT_STATES 64
184 #define BPF_MAP_KEY_POISON (1ULL << 63)
185 #define BPF_MAP_KEY_SEEN (1ULL << 62)
187 #define BPF_MAP_PTR_UNPRIV 1UL
188 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
189 POISON_POINTER_DELTA))
190 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
192 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx);
193 static int release_reference(struct bpf_verifier_env *env, int ref_obj_id);
194 static void invalidate_non_owning_refs(struct bpf_verifier_env *env);
195 static bool in_rbtree_lock_required_cb(struct bpf_verifier_env *env);
196 static int ref_set_non_owning(struct bpf_verifier_env *env,
197 struct bpf_reg_state *reg);
198 static void specialize_kfunc(struct bpf_verifier_env *env,
199 u32 func_id, u16 offset, unsigned long *addr);
200 static bool is_trusted_reg(const struct bpf_reg_state *reg);
202 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
204 return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON;
207 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
209 return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV;
212 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
213 const struct bpf_map *map, bool unpriv)
215 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
216 unpriv |= bpf_map_ptr_unpriv(aux);
217 aux->map_ptr_state = (unsigned long)map |
218 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
221 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
223 return aux->map_key_state & BPF_MAP_KEY_POISON;
226 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
228 return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
231 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
233 return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
236 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
238 bool poisoned = bpf_map_key_poisoned(aux);
240 aux->map_key_state = state | BPF_MAP_KEY_SEEN |
241 (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
244 static bool bpf_helper_call(const struct bpf_insn *insn)
246 return insn->code == (BPF_JMP | BPF_CALL) &&
250 static bool bpf_pseudo_call(const struct bpf_insn *insn)
252 return insn->code == (BPF_JMP | BPF_CALL) &&
253 insn->src_reg == BPF_PSEUDO_CALL;
256 static bool bpf_pseudo_kfunc_call(const struct bpf_insn *insn)
258 return insn->code == (BPF_JMP | BPF_CALL) &&
259 insn->src_reg == BPF_PSEUDO_KFUNC_CALL;
262 struct bpf_call_arg_meta {
263 struct bpf_map *map_ptr;
280 struct btf_field *kptr_field;
283 struct bpf_kfunc_call_arg_meta {
288 const struct btf_type *func_proto;
289 const char *func_name;
302 /* arg_{btf,btf_id,owning_ref} are used by kfunc-specific handling,
303 * generally to pass info about user-defined local kptr types to later
306 * Record the local kptr type to be drop'd
307 * bpf_refcount_acquire (via KF_ARG_PTR_TO_REFCOUNTED_KPTR arg type)
308 * Record the local kptr type to be refcount_incr'd and use
309 * arg_owning_ref to determine whether refcount_acquire should be
317 struct btf_field *field;
320 struct btf_field *field;
323 enum bpf_dynptr_type type;
326 } initialized_dynptr;
334 struct btf *btf_vmlinux;
336 static DEFINE_MUTEX(bpf_verifier_lock);
338 static const struct bpf_line_info *
339 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
341 const struct bpf_line_info *linfo;
342 const struct bpf_prog *prog;
346 nr_linfo = prog->aux->nr_linfo;
348 if (!nr_linfo || insn_off >= prog->len)
351 linfo = prog->aux->linfo;
352 for (i = 1; i < nr_linfo; i++)
353 if (insn_off < linfo[i].insn_off)
356 return &linfo[i - 1];
359 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
361 struct bpf_verifier_env *env = private_data;
364 if (!bpf_verifier_log_needed(&env->log))
368 bpf_verifier_vlog(&env->log, fmt, args);
372 static const char *ltrim(const char *s)
380 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
382 const char *prefix_fmt, ...)
384 const struct bpf_line_info *linfo;
386 if (!bpf_verifier_log_needed(&env->log))
389 linfo = find_linfo(env, insn_off);
390 if (!linfo || linfo == env->prev_linfo)
396 va_start(args, prefix_fmt);
397 bpf_verifier_vlog(&env->log, prefix_fmt, args);
402 ltrim(btf_name_by_offset(env->prog->aux->btf,
405 env->prev_linfo = linfo;
408 static void verbose_invalid_scalar(struct bpf_verifier_env *env,
409 struct bpf_reg_state *reg,
410 struct tnum *range, const char *ctx,
411 const char *reg_name)
415 verbose(env, "At %s the register %s ", ctx, reg_name);
416 if (!tnum_is_unknown(reg->var_off)) {
417 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
418 verbose(env, "has value %s", tn_buf);
420 verbose(env, "has unknown scalar value");
422 tnum_strn(tn_buf, sizeof(tn_buf), *range);
423 verbose(env, " should have been in %s\n", tn_buf);
426 static bool type_is_pkt_pointer(enum bpf_reg_type type)
428 type = base_type(type);
429 return type == PTR_TO_PACKET ||
430 type == PTR_TO_PACKET_META;
433 static bool type_is_sk_pointer(enum bpf_reg_type type)
435 return type == PTR_TO_SOCKET ||
436 type == PTR_TO_SOCK_COMMON ||
437 type == PTR_TO_TCP_SOCK ||
438 type == PTR_TO_XDP_SOCK;
441 static bool type_may_be_null(u32 type)
443 return type & PTR_MAYBE_NULL;
446 static bool reg_not_null(const struct bpf_reg_state *reg)
448 enum bpf_reg_type type;
451 if (type_may_be_null(type))
454 type = base_type(type);
455 return type == PTR_TO_SOCKET ||
456 type == PTR_TO_TCP_SOCK ||
457 type == PTR_TO_MAP_VALUE ||
458 type == PTR_TO_MAP_KEY ||
459 type == PTR_TO_SOCK_COMMON ||
460 (type == PTR_TO_BTF_ID && is_trusted_reg(reg)) ||
464 static bool type_is_ptr_alloc_obj(u32 type)
466 return base_type(type) == PTR_TO_BTF_ID && type_flag(type) & MEM_ALLOC;
469 static bool type_is_non_owning_ref(u32 type)
471 return type_is_ptr_alloc_obj(type) && type_flag(type) & NON_OWN_REF;
474 static struct btf_record *reg_btf_record(const struct bpf_reg_state *reg)
476 struct btf_record *rec = NULL;
477 struct btf_struct_meta *meta;
479 if (reg->type == PTR_TO_MAP_VALUE) {
480 rec = reg->map_ptr->record;
481 } else if (type_is_ptr_alloc_obj(reg->type)) {
482 meta = btf_find_struct_meta(reg->btf, reg->btf_id);
489 static bool subprog_is_global(const struct bpf_verifier_env *env, int subprog)
491 struct bpf_func_info_aux *aux = env->prog->aux->func_info_aux;
493 return aux && aux[subprog].linkage == BTF_FUNC_GLOBAL;
496 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
498 return btf_record_has_field(reg_btf_record(reg), BPF_SPIN_LOCK);
501 static bool type_is_rdonly_mem(u32 type)
503 return type & MEM_RDONLY;
506 static bool is_acquire_function(enum bpf_func_id func_id,
507 const struct bpf_map *map)
509 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
511 if (func_id == BPF_FUNC_sk_lookup_tcp ||
512 func_id == BPF_FUNC_sk_lookup_udp ||
513 func_id == BPF_FUNC_skc_lookup_tcp ||
514 func_id == BPF_FUNC_ringbuf_reserve ||
515 func_id == BPF_FUNC_kptr_xchg)
518 if (func_id == BPF_FUNC_map_lookup_elem &&
519 (map_type == BPF_MAP_TYPE_SOCKMAP ||
520 map_type == BPF_MAP_TYPE_SOCKHASH))
526 static bool is_ptr_cast_function(enum bpf_func_id func_id)
528 return func_id == BPF_FUNC_tcp_sock ||
529 func_id == BPF_FUNC_sk_fullsock ||
530 func_id == BPF_FUNC_skc_to_tcp_sock ||
531 func_id == BPF_FUNC_skc_to_tcp6_sock ||
532 func_id == BPF_FUNC_skc_to_udp6_sock ||
533 func_id == BPF_FUNC_skc_to_mptcp_sock ||
534 func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
535 func_id == BPF_FUNC_skc_to_tcp_request_sock;
538 static bool is_dynptr_ref_function(enum bpf_func_id func_id)
540 return func_id == BPF_FUNC_dynptr_data;
543 static bool is_callback_calling_kfunc(u32 btf_id);
545 static bool is_callback_calling_function(enum bpf_func_id func_id)
547 return func_id == BPF_FUNC_for_each_map_elem ||
548 func_id == BPF_FUNC_timer_set_callback ||
549 func_id == BPF_FUNC_find_vma ||
550 func_id == BPF_FUNC_loop ||
551 func_id == BPF_FUNC_user_ringbuf_drain;
554 static bool is_async_callback_calling_function(enum bpf_func_id func_id)
556 return func_id == BPF_FUNC_timer_set_callback;
559 static bool is_storage_get_function(enum bpf_func_id func_id)
561 return func_id == BPF_FUNC_sk_storage_get ||
562 func_id == BPF_FUNC_inode_storage_get ||
563 func_id == BPF_FUNC_task_storage_get ||
564 func_id == BPF_FUNC_cgrp_storage_get;
567 static bool helper_multiple_ref_obj_use(enum bpf_func_id func_id,
568 const struct bpf_map *map)
570 int ref_obj_uses = 0;
572 if (is_ptr_cast_function(func_id))
574 if (is_acquire_function(func_id, map))
576 if (is_dynptr_ref_function(func_id))
579 return ref_obj_uses > 1;
582 static bool is_cmpxchg_insn(const struct bpf_insn *insn)
584 return BPF_CLASS(insn->code) == BPF_STX &&
585 BPF_MODE(insn->code) == BPF_ATOMIC &&
586 insn->imm == BPF_CMPXCHG;
589 /* string representation of 'enum bpf_reg_type'
591 * Note that reg_type_str() can not appear more than once in a single verbose()
594 static const char *reg_type_str(struct bpf_verifier_env *env,
595 enum bpf_reg_type type)
597 char postfix[16] = {0}, prefix[64] = {0};
598 static const char * const str[] = {
600 [SCALAR_VALUE] = "scalar",
601 [PTR_TO_CTX] = "ctx",
602 [CONST_PTR_TO_MAP] = "map_ptr",
603 [PTR_TO_MAP_VALUE] = "map_value",
604 [PTR_TO_STACK] = "fp",
605 [PTR_TO_PACKET] = "pkt",
606 [PTR_TO_PACKET_META] = "pkt_meta",
607 [PTR_TO_PACKET_END] = "pkt_end",
608 [PTR_TO_FLOW_KEYS] = "flow_keys",
609 [PTR_TO_SOCKET] = "sock",
610 [PTR_TO_SOCK_COMMON] = "sock_common",
611 [PTR_TO_TCP_SOCK] = "tcp_sock",
612 [PTR_TO_TP_BUFFER] = "tp_buffer",
613 [PTR_TO_XDP_SOCK] = "xdp_sock",
614 [PTR_TO_BTF_ID] = "ptr_",
615 [PTR_TO_MEM] = "mem",
616 [PTR_TO_BUF] = "buf",
617 [PTR_TO_FUNC] = "func",
618 [PTR_TO_MAP_KEY] = "map_key",
619 [CONST_PTR_TO_DYNPTR] = "dynptr_ptr",
622 if (type & PTR_MAYBE_NULL) {
623 if (base_type(type) == PTR_TO_BTF_ID)
624 strncpy(postfix, "or_null_", 16);
626 strncpy(postfix, "_or_null", 16);
629 snprintf(prefix, sizeof(prefix), "%s%s%s%s%s%s%s",
630 type & MEM_RDONLY ? "rdonly_" : "",
631 type & MEM_RINGBUF ? "ringbuf_" : "",
632 type & MEM_USER ? "user_" : "",
633 type & MEM_PERCPU ? "percpu_" : "",
634 type & MEM_RCU ? "rcu_" : "",
635 type & PTR_UNTRUSTED ? "untrusted_" : "",
636 type & PTR_TRUSTED ? "trusted_" : ""
639 snprintf(env->tmp_str_buf, TMP_STR_BUF_LEN, "%s%s%s",
640 prefix, str[base_type(type)], postfix);
641 return env->tmp_str_buf;
644 static char slot_type_char[] = {
645 [STACK_INVALID] = '?',
649 [STACK_DYNPTR] = 'd',
653 static void print_liveness(struct bpf_verifier_env *env,
654 enum bpf_reg_liveness live)
656 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
658 if (live & REG_LIVE_READ)
660 if (live & REG_LIVE_WRITTEN)
662 if (live & REG_LIVE_DONE)
666 static int __get_spi(s32 off)
668 return (-off - 1) / BPF_REG_SIZE;
671 static struct bpf_func_state *func(struct bpf_verifier_env *env,
672 const struct bpf_reg_state *reg)
674 struct bpf_verifier_state *cur = env->cur_state;
676 return cur->frame[reg->frameno];
679 static bool is_spi_bounds_valid(struct bpf_func_state *state, int spi, int nr_slots)
681 int allocated_slots = state->allocated_stack / BPF_REG_SIZE;
683 /* We need to check that slots between [spi - nr_slots + 1, spi] are
684 * within [0, allocated_stack).
686 * Please note that the spi grows downwards. For example, a dynptr
687 * takes the size of two stack slots; the first slot will be at
688 * spi and the second slot will be at spi - 1.
690 return spi - nr_slots + 1 >= 0 && spi < allocated_slots;
693 static int stack_slot_obj_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
694 const char *obj_kind, int nr_slots)
698 if (!tnum_is_const(reg->var_off)) {
699 verbose(env, "%s has to be at a constant offset\n", obj_kind);
703 off = reg->off + reg->var_off.value;
704 if (off % BPF_REG_SIZE) {
705 verbose(env, "cannot pass in %s at an offset=%d\n", obj_kind, off);
709 spi = __get_spi(off);
710 if (spi + 1 < nr_slots) {
711 verbose(env, "cannot pass in %s at an offset=%d\n", obj_kind, off);
715 if (!is_spi_bounds_valid(func(env, reg), spi, nr_slots))
720 static int dynptr_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
722 return stack_slot_obj_get_spi(env, reg, "dynptr", BPF_DYNPTR_NR_SLOTS);
725 static int iter_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg, int nr_slots)
727 return stack_slot_obj_get_spi(env, reg, "iter", nr_slots);
730 static const char *btf_type_name(const struct btf *btf, u32 id)
732 return btf_name_by_offset(btf, btf_type_by_id(btf, id)->name_off);
735 static const char *dynptr_type_str(enum bpf_dynptr_type type)
738 case BPF_DYNPTR_TYPE_LOCAL:
740 case BPF_DYNPTR_TYPE_RINGBUF:
742 case BPF_DYNPTR_TYPE_SKB:
744 case BPF_DYNPTR_TYPE_XDP:
746 case BPF_DYNPTR_TYPE_INVALID:
749 WARN_ONCE(1, "unknown dynptr type %d\n", type);
754 static const char *iter_type_str(const struct btf *btf, u32 btf_id)
756 if (!btf || btf_id == 0)
759 /* we already validated that type is valid and has conforming name */
760 return btf_type_name(btf, btf_id) + sizeof(ITER_PREFIX) - 1;
763 static const char *iter_state_str(enum bpf_iter_state state)
766 case BPF_ITER_STATE_ACTIVE:
768 case BPF_ITER_STATE_DRAINED:
770 case BPF_ITER_STATE_INVALID:
773 WARN_ONCE(1, "unknown iter state %d\n", state);
778 static void mark_reg_scratched(struct bpf_verifier_env *env, u32 regno)
780 env->scratched_regs |= 1U << regno;
783 static void mark_stack_slot_scratched(struct bpf_verifier_env *env, u32 spi)
785 env->scratched_stack_slots |= 1ULL << spi;
788 static bool reg_scratched(const struct bpf_verifier_env *env, u32 regno)
790 return (env->scratched_regs >> regno) & 1;
793 static bool stack_slot_scratched(const struct bpf_verifier_env *env, u64 regno)
795 return (env->scratched_stack_slots >> regno) & 1;
798 static bool verifier_state_scratched(const struct bpf_verifier_env *env)
800 return env->scratched_regs || env->scratched_stack_slots;
803 static void mark_verifier_state_clean(struct bpf_verifier_env *env)
805 env->scratched_regs = 0U;
806 env->scratched_stack_slots = 0ULL;
809 /* Used for printing the entire verifier state. */
810 static void mark_verifier_state_scratched(struct bpf_verifier_env *env)
812 env->scratched_regs = ~0U;
813 env->scratched_stack_slots = ~0ULL;
816 static enum bpf_dynptr_type arg_to_dynptr_type(enum bpf_arg_type arg_type)
818 switch (arg_type & DYNPTR_TYPE_FLAG_MASK) {
819 case DYNPTR_TYPE_LOCAL:
820 return BPF_DYNPTR_TYPE_LOCAL;
821 case DYNPTR_TYPE_RINGBUF:
822 return BPF_DYNPTR_TYPE_RINGBUF;
823 case DYNPTR_TYPE_SKB:
824 return BPF_DYNPTR_TYPE_SKB;
825 case DYNPTR_TYPE_XDP:
826 return BPF_DYNPTR_TYPE_XDP;
828 return BPF_DYNPTR_TYPE_INVALID;
832 static enum bpf_type_flag get_dynptr_type_flag(enum bpf_dynptr_type type)
835 case BPF_DYNPTR_TYPE_LOCAL:
836 return DYNPTR_TYPE_LOCAL;
837 case BPF_DYNPTR_TYPE_RINGBUF:
838 return DYNPTR_TYPE_RINGBUF;
839 case BPF_DYNPTR_TYPE_SKB:
840 return DYNPTR_TYPE_SKB;
841 case BPF_DYNPTR_TYPE_XDP:
842 return DYNPTR_TYPE_XDP;
848 static bool dynptr_type_refcounted(enum bpf_dynptr_type type)
850 return type == BPF_DYNPTR_TYPE_RINGBUF;
853 static void __mark_dynptr_reg(struct bpf_reg_state *reg,
854 enum bpf_dynptr_type type,
855 bool first_slot, int dynptr_id);
857 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
858 struct bpf_reg_state *reg);
860 static void mark_dynptr_stack_regs(struct bpf_verifier_env *env,
861 struct bpf_reg_state *sreg1,
862 struct bpf_reg_state *sreg2,
863 enum bpf_dynptr_type type)
865 int id = ++env->id_gen;
867 __mark_dynptr_reg(sreg1, type, true, id);
868 __mark_dynptr_reg(sreg2, type, false, id);
871 static void mark_dynptr_cb_reg(struct bpf_verifier_env *env,
872 struct bpf_reg_state *reg,
873 enum bpf_dynptr_type type)
875 __mark_dynptr_reg(reg, type, true, ++env->id_gen);
878 static int destroy_if_dynptr_stack_slot(struct bpf_verifier_env *env,
879 struct bpf_func_state *state, int spi);
881 static int mark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
882 enum bpf_arg_type arg_type, int insn_idx, int clone_ref_obj_id)
884 struct bpf_func_state *state = func(env, reg);
885 enum bpf_dynptr_type type;
888 spi = dynptr_get_spi(env, reg);
892 /* We cannot assume both spi and spi - 1 belong to the same dynptr,
893 * hence we need to call destroy_if_dynptr_stack_slot twice for both,
894 * to ensure that for the following example:
897 * So marking spi = 2 should lead to destruction of both d1 and d2. In
898 * case they do belong to same dynptr, second call won't see slot_type
899 * as STACK_DYNPTR and will simply skip destruction.
901 err = destroy_if_dynptr_stack_slot(env, state, spi);
904 err = destroy_if_dynptr_stack_slot(env, state, spi - 1);
908 for (i = 0; i < BPF_REG_SIZE; i++) {
909 state->stack[spi].slot_type[i] = STACK_DYNPTR;
910 state->stack[spi - 1].slot_type[i] = STACK_DYNPTR;
913 type = arg_to_dynptr_type(arg_type);
914 if (type == BPF_DYNPTR_TYPE_INVALID)
917 mark_dynptr_stack_regs(env, &state->stack[spi].spilled_ptr,
918 &state->stack[spi - 1].spilled_ptr, type);
920 if (dynptr_type_refcounted(type)) {
921 /* The id is used to track proper releasing */
924 if (clone_ref_obj_id)
925 id = clone_ref_obj_id;
927 id = acquire_reference_state(env, insn_idx);
932 state->stack[spi].spilled_ptr.ref_obj_id = id;
933 state->stack[spi - 1].spilled_ptr.ref_obj_id = id;
936 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
937 state->stack[spi - 1].spilled_ptr.live |= REG_LIVE_WRITTEN;
942 static void invalidate_dynptr(struct bpf_verifier_env *env, struct bpf_func_state *state, int spi)
946 for (i = 0; i < BPF_REG_SIZE; i++) {
947 state->stack[spi].slot_type[i] = STACK_INVALID;
948 state->stack[spi - 1].slot_type[i] = STACK_INVALID;
951 __mark_reg_not_init(env, &state->stack[spi].spilled_ptr);
952 __mark_reg_not_init(env, &state->stack[spi - 1].spilled_ptr);
954 /* Why do we need to set REG_LIVE_WRITTEN for STACK_INVALID slot?
956 * While we don't allow reading STACK_INVALID, it is still possible to
957 * do <8 byte writes marking some but not all slots as STACK_MISC. Then,
958 * helpers or insns can do partial read of that part without failing,
959 * but check_stack_range_initialized, check_stack_read_var_off, and
960 * check_stack_read_fixed_off will do mark_reg_read for all 8-bytes of
961 * the slot conservatively. Hence we need to prevent those liveness
964 * This was not a problem before because STACK_INVALID is only set by
965 * default (where the default reg state has its reg->parent as NULL), or
966 * in clean_live_states after REG_LIVE_DONE (at which point
967 * mark_reg_read won't walk reg->parent chain), but not randomly during
968 * verifier state exploration (like we did above). Hence, for our case
969 * parentage chain will still be live (i.e. reg->parent may be
970 * non-NULL), while earlier reg->parent was NULL, so we need
971 * REG_LIVE_WRITTEN to screen off read marker propagation when it is
972 * done later on reads or by mark_dynptr_read as well to unnecessary
973 * mark registers in verifier state.
975 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
976 state->stack[spi - 1].spilled_ptr.live |= REG_LIVE_WRITTEN;
979 static int unmark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
981 struct bpf_func_state *state = func(env, reg);
982 int spi, ref_obj_id, i;
984 spi = dynptr_get_spi(env, reg);
988 if (!dynptr_type_refcounted(state->stack[spi].spilled_ptr.dynptr.type)) {
989 invalidate_dynptr(env, state, spi);
993 ref_obj_id = state->stack[spi].spilled_ptr.ref_obj_id;
995 /* If the dynptr has a ref_obj_id, then we need to invalidate
998 * 1) Any dynptrs with a matching ref_obj_id (clones)
999 * 2) Any slices derived from this dynptr.
1002 /* Invalidate any slices associated with this dynptr */
1003 WARN_ON_ONCE(release_reference(env, ref_obj_id));
1005 /* Invalidate any dynptr clones */
1006 for (i = 1; i < state->allocated_stack / BPF_REG_SIZE; i++) {
1007 if (state->stack[i].spilled_ptr.ref_obj_id != ref_obj_id)
1010 /* it should always be the case that if the ref obj id
1011 * matches then the stack slot also belongs to a
1014 if (state->stack[i].slot_type[0] != STACK_DYNPTR) {
1015 verbose(env, "verifier internal error: misconfigured ref_obj_id\n");
1018 if (state->stack[i].spilled_ptr.dynptr.first_slot)
1019 invalidate_dynptr(env, state, i);
1025 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1026 struct bpf_reg_state *reg);
1028 static void mark_reg_invalid(const struct bpf_verifier_env *env, struct bpf_reg_state *reg)
1030 if (!env->allow_ptr_leaks)
1031 __mark_reg_not_init(env, reg);
1033 __mark_reg_unknown(env, reg);
1036 static int destroy_if_dynptr_stack_slot(struct bpf_verifier_env *env,
1037 struct bpf_func_state *state, int spi)
1039 struct bpf_func_state *fstate;
1040 struct bpf_reg_state *dreg;
1043 /* We always ensure that STACK_DYNPTR is never set partially,
1044 * hence just checking for slot_type[0] is enough. This is
1045 * different for STACK_SPILL, where it may be only set for
1046 * 1 byte, so code has to use is_spilled_reg.
1048 if (state->stack[spi].slot_type[0] != STACK_DYNPTR)
1051 /* Reposition spi to first slot */
1052 if (!state->stack[spi].spilled_ptr.dynptr.first_slot)
1055 if (dynptr_type_refcounted(state->stack[spi].spilled_ptr.dynptr.type)) {
1056 verbose(env, "cannot overwrite referenced dynptr\n");
1060 mark_stack_slot_scratched(env, spi);
1061 mark_stack_slot_scratched(env, spi - 1);
1063 /* Writing partially to one dynptr stack slot destroys both. */
1064 for (i = 0; i < BPF_REG_SIZE; i++) {
1065 state->stack[spi].slot_type[i] = STACK_INVALID;
1066 state->stack[spi - 1].slot_type[i] = STACK_INVALID;
1069 dynptr_id = state->stack[spi].spilled_ptr.id;
1070 /* Invalidate any slices associated with this dynptr */
1071 bpf_for_each_reg_in_vstate(env->cur_state, fstate, dreg, ({
1072 /* Dynptr slices are only PTR_TO_MEM_OR_NULL and PTR_TO_MEM */
1073 if (dreg->type != (PTR_TO_MEM | PTR_MAYBE_NULL) && dreg->type != PTR_TO_MEM)
1075 if (dreg->dynptr_id == dynptr_id)
1076 mark_reg_invalid(env, dreg);
1079 /* Do not release reference state, we are destroying dynptr on stack,
1080 * not using some helper to release it. Just reset register.
1082 __mark_reg_not_init(env, &state->stack[spi].spilled_ptr);
1083 __mark_reg_not_init(env, &state->stack[spi - 1].spilled_ptr);
1085 /* Same reason as unmark_stack_slots_dynptr above */
1086 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1087 state->stack[spi - 1].spilled_ptr.live |= REG_LIVE_WRITTEN;
1092 static bool is_dynptr_reg_valid_uninit(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
1096 if (reg->type == CONST_PTR_TO_DYNPTR)
1099 spi = dynptr_get_spi(env, reg);
1101 /* -ERANGE (i.e. spi not falling into allocated stack slots) isn't an
1102 * error because this just means the stack state hasn't been updated yet.
1103 * We will do check_mem_access to check and update stack bounds later.
1105 if (spi < 0 && spi != -ERANGE)
1108 /* We don't need to check if the stack slots are marked by previous
1109 * dynptr initializations because we allow overwriting existing unreferenced
1110 * STACK_DYNPTR slots, see mark_stack_slots_dynptr which calls
1111 * destroy_if_dynptr_stack_slot to ensure dynptr objects at the slots we are
1112 * touching are completely destructed before we reinitialize them for a new
1113 * one. For referenced ones, destroy_if_dynptr_stack_slot returns an error early
1114 * instead of delaying it until the end where the user will get "Unreleased
1120 static bool is_dynptr_reg_valid_init(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
1122 struct bpf_func_state *state = func(env, reg);
1125 /* This already represents first slot of initialized bpf_dynptr.
1127 * CONST_PTR_TO_DYNPTR already has fixed and var_off as 0 due to
1128 * check_func_arg_reg_off's logic, so we don't need to check its
1129 * offset and alignment.
1131 if (reg->type == CONST_PTR_TO_DYNPTR)
1134 spi = dynptr_get_spi(env, reg);
1137 if (!state->stack[spi].spilled_ptr.dynptr.first_slot)
1140 for (i = 0; i < BPF_REG_SIZE; i++) {
1141 if (state->stack[spi].slot_type[i] != STACK_DYNPTR ||
1142 state->stack[spi - 1].slot_type[i] != STACK_DYNPTR)
1149 static bool is_dynptr_type_expected(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
1150 enum bpf_arg_type arg_type)
1152 struct bpf_func_state *state = func(env, reg);
1153 enum bpf_dynptr_type dynptr_type;
1156 /* ARG_PTR_TO_DYNPTR takes any type of dynptr */
1157 if (arg_type == ARG_PTR_TO_DYNPTR)
1160 dynptr_type = arg_to_dynptr_type(arg_type);
1161 if (reg->type == CONST_PTR_TO_DYNPTR) {
1162 return reg->dynptr.type == dynptr_type;
1164 spi = dynptr_get_spi(env, reg);
1167 return state->stack[spi].spilled_ptr.dynptr.type == dynptr_type;
1171 static void __mark_reg_known_zero(struct bpf_reg_state *reg);
1173 static int mark_stack_slots_iter(struct bpf_verifier_env *env,
1174 struct bpf_reg_state *reg, int insn_idx,
1175 struct btf *btf, u32 btf_id, int nr_slots)
1177 struct bpf_func_state *state = func(env, reg);
1180 spi = iter_get_spi(env, reg, nr_slots);
1184 id = acquire_reference_state(env, insn_idx);
1188 for (i = 0; i < nr_slots; i++) {
1189 struct bpf_stack_state *slot = &state->stack[spi - i];
1190 struct bpf_reg_state *st = &slot->spilled_ptr;
1192 __mark_reg_known_zero(st);
1193 st->type = PTR_TO_STACK; /* we don't have dedicated reg type */
1194 st->live |= REG_LIVE_WRITTEN;
1195 st->ref_obj_id = i == 0 ? id : 0;
1197 st->iter.btf_id = btf_id;
1198 st->iter.state = BPF_ITER_STATE_ACTIVE;
1201 for (j = 0; j < BPF_REG_SIZE; j++)
1202 slot->slot_type[j] = STACK_ITER;
1204 mark_stack_slot_scratched(env, spi - i);
1210 static int unmark_stack_slots_iter(struct bpf_verifier_env *env,
1211 struct bpf_reg_state *reg, int nr_slots)
1213 struct bpf_func_state *state = func(env, reg);
1216 spi = iter_get_spi(env, reg, nr_slots);
1220 for (i = 0; i < nr_slots; i++) {
1221 struct bpf_stack_state *slot = &state->stack[spi - i];
1222 struct bpf_reg_state *st = &slot->spilled_ptr;
1225 WARN_ON_ONCE(release_reference(env, st->ref_obj_id));
1227 __mark_reg_not_init(env, st);
1229 /* see unmark_stack_slots_dynptr() for why we need to set REG_LIVE_WRITTEN */
1230 st->live |= REG_LIVE_WRITTEN;
1232 for (j = 0; j < BPF_REG_SIZE; j++)
1233 slot->slot_type[j] = STACK_INVALID;
1235 mark_stack_slot_scratched(env, spi - i);
1241 static bool is_iter_reg_valid_uninit(struct bpf_verifier_env *env,
1242 struct bpf_reg_state *reg, int nr_slots)
1244 struct bpf_func_state *state = func(env, reg);
1247 /* For -ERANGE (i.e. spi not falling into allocated stack slots), we
1248 * will do check_mem_access to check and update stack bounds later, so
1249 * return true for that case.
1251 spi = iter_get_spi(env, reg, nr_slots);
1257 for (i = 0; i < nr_slots; i++) {
1258 struct bpf_stack_state *slot = &state->stack[spi - i];
1260 for (j = 0; j < BPF_REG_SIZE; j++)
1261 if (slot->slot_type[j] == STACK_ITER)
1268 static bool is_iter_reg_valid_init(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
1269 struct btf *btf, u32 btf_id, int nr_slots)
1271 struct bpf_func_state *state = func(env, reg);
1274 spi = iter_get_spi(env, reg, nr_slots);
1278 for (i = 0; i < nr_slots; i++) {
1279 struct bpf_stack_state *slot = &state->stack[spi - i];
1280 struct bpf_reg_state *st = &slot->spilled_ptr;
1282 /* only main (first) slot has ref_obj_id set */
1283 if (i == 0 && !st->ref_obj_id)
1285 if (i != 0 && st->ref_obj_id)
1287 if (st->iter.btf != btf || st->iter.btf_id != btf_id)
1290 for (j = 0; j < BPF_REG_SIZE; j++)
1291 if (slot->slot_type[j] != STACK_ITER)
1298 /* Check if given stack slot is "special":
1299 * - spilled register state (STACK_SPILL);
1300 * - dynptr state (STACK_DYNPTR);
1301 * - iter state (STACK_ITER).
1303 static bool is_stack_slot_special(const struct bpf_stack_state *stack)
1305 enum bpf_stack_slot_type type = stack->slot_type[BPF_REG_SIZE - 1];
1317 WARN_ONCE(1, "unknown stack slot type %d\n", type);
1322 /* The reg state of a pointer or a bounded scalar was saved when
1323 * it was spilled to the stack.
1325 static bool is_spilled_reg(const struct bpf_stack_state *stack)
1327 return stack->slot_type[BPF_REG_SIZE - 1] == STACK_SPILL;
1330 static bool is_spilled_scalar_reg(const struct bpf_stack_state *stack)
1332 return stack->slot_type[BPF_REG_SIZE - 1] == STACK_SPILL &&
1333 stack->spilled_ptr.type == SCALAR_VALUE;
1336 static void scrub_spilled_slot(u8 *stype)
1338 if (*stype != STACK_INVALID)
1339 *stype = STACK_MISC;
1342 static void print_verifier_state(struct bpf_verifier_env *env,
1343 const struct bpf_func_state *state,
1346 const struct bpf_reg_state *reg;
1347 enum bpf_reg_type t;
1351 verbose(env, " frame%d:", state->frameno);
1352 for (i = 0; i < MAX_BPF_REG; i++) {
1353 reg = &state->regs[i];
1357 if (!print_all && !reg_scratched(env, i))
1359 verbose(env, " R%d", i);
1360 print_liveness(env, reg->live);
1362 if (t == SCALAR_VALUE && reg->precise)
1364 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
1365 tnum_is_const(reg->var_off)) {
1366 /* reg->off should be 0 for SCALAR_VALUE */
1367 verbose(env, "%s", t == SCALAR_VALUE ? "" : reg_type_str(env, t));
1368 verbose(env, "%lld", reg->var_off.value + reg->off);
1370 const char *sep = "";
1372 verbose(env, "%s", reg_type_str(env, t));
1373 if (base_type(t) == PTR_TO_BTF_ID)
1374 verbose(env, "%s", btf_type_name(reg->btf, reg->btf_id));
1377 * _a stands for append, was shortened to avoid multiline statements below.
1378 * This macro is used to output a comma separated list of attributes.
1380 #define verbose_a(fmt, ...) ({ verbose(env, "%s" fmt, sep, __VA_ARGS__); sep = ","; })
1383 verbose_a("id=%d", reg->id);
1384 if (reg->ref_obj_id)
1385 verbose_a("ref_obj_id=%d", reg->ref_obj_id);
1386 if (type_is_non_owning_ref(reg->type))
1387 verbose_a("%s", "non_own_ref");
1388 if (t != SCALAR_VALUE)
1389 verbose_a("off=%d", reg->off);
1390 if (type_is_pkt_pointer(t))
1391 verbose_a("r=%d", reg->range);
1392 else if (base_type(t) == CONST_PTR_TO_MAP ||
1393 base_type(t) == PTR_TO_MAP_KEY ||
1394 base_type(t) == PTR_TO_MAP_VALUE)
1395 verbose_a("ks=%d,vs=%d",
1396 reg->map_ptr->key_size,
1397 reg->map_ptr->value_size);
1398 if (tnum_is_const(reg->var_off)) {
1399 /* Typically an immediate SCALAR_VALUE, but
1400 * could be a pointer whose offset is too big
1403 verbose_a("imm=%llx", reg->var_off.value);
1405 if (reg->smin_value != reg->umin_value &&
1406 reg->smin_value != S64_MIN)
1407 verbose_a("smin=%lld", (long long)reg->smin_value);
1408 if (reg->smax_value != reg->umax_value &&
1409 reg->smax_value != S64_MAX)
1410 verbose_a("smax=%lld", (long long)reg->smax_value);
1411 if (reg->umin_value != 0)
1412 verbose_a("umin=%llu", (unsigned long long)reg->umin_value);
1413 if (reg->umax_value != U64_MAX)
1414 verbose_a("umax=%llu", (unsigned long long)reg->umax_value);
1415 if (!tnum_is_unknown(reg->var_off)) {
1418 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1419 verbose_a("var_off=%s", tn_buf);
1421 if (reg->s32_min_value != reg->smin_value &&
1422 reg->s32_min_value != S32_MIN)
1423 verbose_a("s32_min=%d", (int)(reg->s32_min_value));
1424 if (reg->s32_max_value != reg->smax_value &&
1425 reg->s32_max_value != S32_MAX)
1426 verbose_a("s32_max=%d", (int)(reg->s32_max_value));
1427 if (reg->u32_min_value != reg->umin_value &&
1428 reg->u32_min_value != U32_MIN)
1429 verbose_a("u32_min=%d", (int)(reg->u32_min_value));
1430 if (reg->u32_max_value != reg->umax_value &&
1431 reg->u32_max_value != U32_MAX)
1432 verbose_a("u32_max=%d", (int)(reg->u32_max_value));
1439 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
1440 char types_buf[BPF_REG_SIZE + 1];
1444 for (j = 0; j < BPF_REG_SIZE; j++) {
1445 if (state->stack[i].slot_type[j] != STACK_INVALID)
1447 types_buf[j] = slot_type_char[state->stack[i].slot_type[j]];
1449 types_buf[BPF_REG_SIZE] = 0;
1452 if (!print_all && !stack_slot_scratched(env, i))
1454 switch (state->stack[i].slot_type[BPF_REG_SIZE - 1]) {
1456 reg = &state->stack[i].spilled_ptr;
1459 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
1460 print_liveness(env, reg->live);
1461 verbose(env, "=%s", t == SCALAR_VALUE ? "" : reg_type_str(env, t));
1462 if (t == SCALAR_VALUE && reg->precise)
1464 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
1465 verbose(env, "%lld", reg->var_off.value + reg->off);
1468 i += BPF_DYNPTR_NR_SLOTS - 1;
1469 reg = &state->stack[i].spilled_ptr;
1471 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
1472 print_liveness(env, reg->live);
1473 verbose(env, "=dynptr_%s", dynptr_type_str(reg->dynptr.type));
1474 if (reg->ref_obj_id)
1475 verbose(env, "(ref_id=%d)", reg->ref_obj_id);
1478 /* only main slot has ref_obj_id set; skip others */
1479 reg = &state->stack[i].spilled_ptr;
1480 if (!reg->ref_obj_id)
1483 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
1484 print_liveness(env, reg->live);
1485 verbose(env, "=iter_%s(ref_id=%d,state=%s,depth=%u)",
1486 iter_type_str(reg->iter.btf, reg->iter.btf_id),
1487 reg->ref_obj_id, iter_state_str(reg->iter.state),
1493 reg = &state->stack[i].spilled_ptr;
1495 for (j = 0; j < BPF_REG_SIZE; j++)
1496 types_buf[j] = slot_type_char[state->stack[i].slot_type[j]];
1497 types_buf[BPF_REG_SIZE] = 0;
1499 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
1500 print_liveness(env, reg->live);
1501 verbose(env, "=%s", types_buf);
1505 if (state->acquired_refs && state->refs[0].id) {
1506 verbose(env, " refs=%d", state->refs[0].id);
1507 for (i = 1; i < state->acquired_refs; i++)
1508 if (state->refs[i].id)
1509 verbose(env, ",%d", state->refs[i].id);
1511 if (state->in_callback_fn)
1512 verbose(env, " cb");
1513 if (state->in_async_callback_fn)
1514 verbose(env, " async_cb");
1516 mark_verifier_state_clean(env);
1519 static inline u32 vlog_alignment(u32 pos)
1521 return round_up(max(pos + BPF_LOG_MIN_ALIGNMENT / 2, BPF_LOG_ALIGNMENT),
1522 BPF_LOG_MIN_ALIGNMENT) - pos - 1;
1525 static void print_insn_state(struct bpf_verifier_env *env,
1526 const struct bpf_func_state *state)
1528 if (env->prev_log_pos && env->prev_log_pos == env->log.end_pos) {
1529 /* remove new line character */
1530 bpf_vlog_reset(&env->log, env->prev_log_pos - 1);
1531 verbose(env, "%*c;", vlog_alignment(env->prev_insn_print_pos), ' ');
1533 verbose(env, "%d:", env->insn_idx);
1535 print_verifier_state(env, state, false);
1538 /* copy array src of length n * size bytes to dst. dst is reallocated if it's too
1539 * small to hold src. This is different from krealloc since we don't want to preserve
1540 * the contents of dst.
1542 * Leaves dst untouched if src is NULL or length is zero. Returns NULL if memory could
1545 static void *copy_array(void *dst, const void *src, size_t n, size_t size, gfp_t flags)
1551 if (ZERO_OR_NULL_PTR(src))
1554 if (unlikely(check_mul_overflow(n, size, &bytes)))
1557 alloc_bytes = max(ksize(orig), kmalloc_size_roundup(bytes));
1558 dst = krealloc(orig, alloc_bytes, flags);
1564 memcpy(dst, src, bytes);
1566 return dst ? dst : ZERO_SIZE_PTR;
1569 /* resize an array from old_n items to new_n items. the array is reallocated if it's too
1570 * small to hold new_n items. new items are zeroed out if the array grows.
1572 * Contrary to krealloc_array, does not free arr if new_n is zero.
1574 static void *realloc_array(void *arr, size_t old_n, size_t new_n, size_t size)
1579 if (!new_n || old_n == new_n)
1582 alloc_size = kmalloc_size_roundup(size_mul(new_n, size));
1583 new_arr = krealloc(arr, alloc_size, GFP_KERNEL);
1591 memset(arr + old_n * size, 0, (new_n - old_n) * size);
1594 return arr ? arr : ZERO_SIZE_PTR;
1597 static int copy_reference_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
1599 dst->refs = copy_array(dst->refs, src->refs, src->acquired_refs,
1600 sizeof(struct bpf_reference_state), GFP_KERNEL);
1604 dst->acquired_refs = src->acquired_refs;
1608 static int copy_stack_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
1610 size_t n = src->allocated_stack / BPF_REG_SIZE;
1612 dst->stack = copy_array(dst->stack, src->stack, n, sizeof(struct bpf_stack_state),
1617 dst->allocated_stack = src->allocated_stack;
1621 static int resize_reference_state(struct bpf_func_state *state, size_t n)
1623 state->refs = realloc_array(state->refs, state->acquired_refs, n,
1624 sizeof(struct bpf_reference_state));
1628 state->acquired_refs = n;
1632 static int grow_stack_state(struct bpf_func_state *state, int size)
1634 size_t old_n = state->allocated_stack / BPF_REG_SIZE, n = size / BPF_REG_SIZE;
1639 state->stack = realloc_array(state->stack, old_n, n, sizeof(struct bpf_stack_state));
1643 state->allocated_stack = size;
1647 /* Acquire a pointer id from the env and update the state->refs to include
1648 * this new pointer reference.
1649 * On success, returns a valid pointer id to associate with the register
1650 * On failure, returns a negative errno.
1652 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
1654 struct bpf_func_state *state = cur_func(env);
1655 int new_ofs = state->acquired_refs;
1658 err = resize_reference_state(state, state->acquired_refs + 1);
1662 state->refs[new_ofs].id = id;
1663 state->refs[new_ofs].insn_idx = insn_idx;
1664 state->refs[new_ofs].callback_ref = state->in_callback_fn ? state->frameno : 0;
1669 /* release function corresponding to acquire_reference_state(). Idempotent. */
1670 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
1674 last_idx = state->acquired_refs - 1;
1675 for (i = 0; i < state->acquired_refs; i++) {
1676 if (state->refs[i].id == ptr_id) {
1677 /* Cannot release caller references in callbacks */
1678 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno)
1680 if (last_idx && i != last_idx)
1681 memcpy(&state->refs[i], &state->refs[last_idx],
1682 sizeof(*state->refs));
1683 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
1684 state->acquired_refs--;
1691 static void free_func_state(struct bpf_func_state *state)
1696 kfree(state->stack);
1700 static void clear_jmp_history(struct bpf_verifier_state *state)
1702 kfree(state->jmp_history);
1703 state->jmp_history = NULL;
1704 state->jmp_history_cnt = 0;
1707 static void free_verifier_state(struct bpf_verifier_state *state,
1712 for (i = 0; i <= state->curframe; i++) {
1713 free_func_state(state->frame[i]);
1714 state->frame[i] = NULL;
1716 clear_jmp_history(state);
1721 /* copy verifier state from src to dst growing dst stack space
1722 * when necessary to accommodate larger src stack
1724 static int copy_func_state(struct bpf_func_state *dst,
1725 const struct bpf_func_state *src)
1729 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
1730 err = copy_reference_state(dst, src);
1733 return copy_stack_state(dst, src);
1736 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
1737 const struct bpf_verifier_state *src)
1739 struct bpf_func_state *dst;
1742 dst_state->jmp_history = copy_array(dst_state->jmp_history, src->jmp_history,
1743 src->jmp_history_cnt, sizeof(struct bpf_idx_pair),
1745 if (!dst_state->jmp_history)
1747 dst_state->jmp_history_cnt = src->jmp_history_cnt;
1749 /* if dst has more stack frames then src frame, free them */
1750 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
1751 free_func_state(dst_state->frame[i]);
1752 dst_state->frame[i] = NULL;
1754 dst_state->speculative = src->speculative;
1755 dst_state->active_rcu_lock = src->active_rcu_lock;
1756 dst_state->curframe = src->curframe;
1757 dst_state->active_lock.ptr = src->active_lock.ptr;
1758 dst_state->active_lock.id = src->active_lock.id;
1759 dst_state->branches = src->branches;
1760 dst_state->parent = src->parent;
1761 dst_state->first_insn_idx = src->first_insn_idx;
1762 dst_state->last_insn_idx = src->last_insn_idx;
1763 for (i = 0; i <= src->curframe; i++) {
1764 dst = dst_state->frame[i];
1766 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
1769 dst_state->frame[i] = dst;
1771 err = copy_func_state(dst, src->frame[i]);
1778 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
1781 u32 br = --st->branches;
1783 /* WARN_ON(br > 1) technically makes sense here,
1784 * but see comment in push_stack(), hence:
1786 WARN_ONCE((int)br < 0,
1787 "BUG update_branch_counts:branches_to_explore=%d\n",
1795 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
1796 int *insn_idx, bool pop_log)
1798 struct bpf_verifier_state *cur = env->cur_state;
1799 struct bpf_verifier_stack_elem *elem, *head = env->head;
1802 if (env->head == NULL)
1806 err = copy_verifier_state(cur, &head->st);
1811 bpf_vlog_reset(&env->log, head->log_pos);
1813 *insn_idx = head->insn_idx;
1815 *prev_insn_idx = head->prev_insn_idx;
1817 free_verifier_state(&head->st, false);
1824 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
1825 int insn_idx, int prev_insn_idx,
1828 struct bpf_verifier_state *cur = env->cur_state;
1829 struct bpf_verifier_stack_elem *elem;
1832 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1836 elem->insn_idx = insn_idx;
1837 elem->prev_insn_idx = prev_insn_idx;
1838 elem->next = env->head;
1839 elem->log_pos = env->log.end_pos;
1842 err = copy_verifier_state(&elem->st, cur);
1845 elem->st.speculative |= speculative;
1846 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1847 verbose(env, "The sequence of %d jumps is too complex.\n",
1851 if (elem->st.parent) {
1852 ++elem->st.parent->branches;
1853 /* WARN_ON(branches > 2) technically makes sense here,
1855 * 1. speculative states will bump 'branches' for non-branch
1857 * 2. is_state_visited() heuristics may decide not to create
1858 * a new state for a sequence of branches and all such current
1859 * and cloned states will be pointing to a single parent state
1860 * which might have large 'branches' count.
1865 free_verifier_state(env->cur_state, true);
1866 env->cur_state = NULL;
1867 /* pop all elements and return */
1868 while (!pop_stack(env, NULL, NULL, false));
1872 #define CALLER_SAVED_REGS 6
1873 static const int caller_saved[CALLER_SAVED_REGS] = {
1874 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
1877 /* This helper doesn't clear reg->id */
1878 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1880 reg->var_off = tnum_const(imm);
1881 reg->smin_value = (s64)imm;
1882 reg->smax_value = (s64)imm;
1883 reg->umin_value = imm;
1884 reg->umax_value = imm;
1886 reg->s32_min_value = (s32)imm;
1887 reg->s32_max_value = (s32)imm;
1888 reg->u32_min_value = (u32)imm;
1889 reg->u32_max_value = (u32)imm;
1892 /* Mark the unknown part of a register (variable offset or scalar value) as
1893 * known to have the value @imm.
1895 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1897 /* Clear off and union(map_ptr, range) */
1898 memset(((u8 *)reg) + sizeof(reg->type), 0,
1899 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1901 reg->ref_obj_id = 0;
1902 ___mark_reg_known(reg, imm);
1905 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1907 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1908 reg->s32_min_value = (s32)imm;
1909 reg->s32_max_value = (s32)imm;
1910 reg->u32_min_value = (u32)imm;
1911 reg->u32_max_value = (u32)imm;
1914 /* Mark the 'variable offset' part of a register as zero. This should be
1915 * used only on registers holding a pointer type.
1917 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1919 __mark_reg_known(reg, 0);
1922 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1924 __mark_reg_known(reg, 0);
1925 reg->type = SCALAR_VALUE;
1928 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1929 struct bpf_reg_state *regs, u32 regno)
1931 if (WARN_ON(regno >= MAX_BPF_REG)) {
1932 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1933 /* Something bad happened, let's kill all regs */
1934 for (regno = 0; regno < MAX_BPF_REG; regno++)
1935 __mark_reg_not_init(env, regs + regno);
1938 __mark_reg_known_zero(regs + regno);
1941 static void __mark_dynptr_reg(struct bpf_reg_state *reg, enum bpf_dynptr_type type,
1942 bool first_slot, int dynptr_id)
1944 /* reg->type has no meaning for STACK_DYNPTR, but when we set reg for
1945 * callback arguments, it does need to be CONST_PTR_TO_DYNPTR, so simply
1946 * set it unconditionally as it is ignored for STACK_DYNPTR anyway.
1948 __mark_reg_known_zero(reg);
1949 reg->type = CONST_PTR_TO_DYNPTR;
1950 /* Give each dynptr a unique id to uniquely associate slices to it. */
1951 reg->id = dynptr_id;
1952 reg->dynptr.type = type;
1953 reg->dynptr.first_slot = first_slot;
1956 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
1958 if (base_type(reg->type) == PTR_TO_MAP_VALUE) {
1959 const struct bpf_map *map = reg->map_ptr;
1961 if (map->inner_map_meta) {
1962 reg->type = CONST_PTR_TO_MAP;
1963 reg->map_ptr = map->inner_map_meta;
1964 /* transfer reg's id which is unique for every map_lookup_elem
1965 * as UID of the inner map.
1967 if (btf_record_has_field(map->inner_map_meta->record, BPF_TIMER))
1968 reg->map_uid = reg->id;
1969 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
1970 reg->type = PTR_TO_XDP_SOCK;
1971 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
1972 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
1973 reg->type = PTR_TO_SOCKET;
1975 reg->type = PTR_TO_MAP_VALUE;
1980 reg->type &= ~PTR_MAYBE_NULL;
1983 static void mark_reg_graph_node(struct bpf_reg_state *regs, u32 regno,
1984 struct btf_field_graph_root *ds_head)
1986 __mark_reg_known_zero(®s[regno]);
1987 regs[regno].type = PTR_TO_BTF_ID | MEM_ALLOC;
1988 regs[regno].btf = ds_head->btf;
1989 regs[regno].btf_id = ds_head->value_btf_id;
1990 regs[regno].off = ds_head->node_offset;
1993 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1995 return type_is_pkt_pointer(reg->type);
1998 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
2000 return reg_is_pkt_pointer(reg) ||
2001 reg->type == PTR_TO_PACKET_END;
2004 static bool reg_is_dynptr_slice_pkt(const struct bpf_reg_state *reg)
2006 return base_type(reg->type) == PTR_TO_MEM &&
2007 (reg->type & DYNPTR_TYPE_SKB || reg->type & DYNPTR_TYPE_XDP);
2010 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
2011 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
2012 enum bpf_reg_type which)
2014 /* The register can already have a range from prior markings.
2015 * This is fine as long as it hasn't been advanced from its
2018 return reg->type == which &&
2021 tnum_equals_const(reg->var_off, 0);
2024 /* Reset the min/max bounds of a register */
2025 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
2027 reg->smin_value = S64_MIN;
2028 reg->smax_value = S64_MAX;
2029 reg->umin_value = 0;
2030 reg->umax_value = U64_MAX;
2032 reg->s32_min_value = S32_MIN;
2033 reg->s32_max_value = S32_MAX;
2034 reg->u32_min_value = 0;
2035 reg->u32_max_value = U32_MAX;
2038 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
2040 reg->smin_value = S64_MIN;
2041 reg->smax_value = S64_MAX;
2042 reg->umin_value = 0;
2043 reg->umax_value = U64_MAX;
2046 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
2048 reg->s32_min_value = S32_MIN;
2049 reg->s32_max_value = S32_MAX;
2050 reg->u32_min_value = 0;
2051 reg->u32_max_value = U32_MAX;
2054 static void __update_reg32_bounds(struct bpf_reg_state *reg)
2056 struct tnum var32_off = tnum_subreg(reg->var_off);
2058 /* min signed is max(sign bit) | min(other bits) */
2059 reg->s32_min_value = max_t(s32, reg->s32_min_value,
2060 var32_off.value | (var32_off.mask & S32_MIN));
2061 /* max signed is min(sign bit) | max(other bits) */
2062 reg->s32_max_value = min_t(s32, reg->s32_max_value,
2063 var32_off.value | (var32_off.mask & S32_MAX));
2064 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
2065 reg->u32_max_value = min(reg->u32_max_value,
2066 (u32)(var32_off.value | var32_off.mask));
2069 static void __update_reg64_bounds(struct bpf_reg_state *reg)
2071 /* min signed is max(sign bit) | min(other bits) */
2072 reg->smin_value = max_t(s64, reg->smin_value,
2073 reg->var_off.value | (reg->var_off.mask & S64_MIN));
2074 /* max signed is min(sign bit) | max(other bits) */
2075 reg->smax_value = min_t(s64, reg->smax_value,
2076 reg->var_off.value | (reg->var_off.mask & S64_MAX));
2077 reg->umin_value = max(reg->umin_value, reg->var_off.value);
2078 reg->umax_value = min(reg->umax_value,
2079 reg->var_off.value | reg->var_off.mask);
2082 static void __update_reg_bounds(struct bpf_reg_state *reg)
2084 __update_reg32_bounds(reg);
2085 __update_reg64_bounds(reg);
2088 /* Uses signed min/max values to inform unsigned, and vice-versa */
2089 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
2091 /* Learn sign from signed bounds.
2092 * If we cannot cross the sign boundary, then signed and unsigned bounds
2093 * are the same, so combine. This works even in the negative case, e.g.
2094 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
2096 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
2097 reg->s32_min_value = reg->u32_min_value =
2098 max_t(u32, reg->s32_min_value, reg->u32_min_value);
2099 reg->s32_max_value = reg->u32_max_value =
2100 min_t(u32, reg->s32_max_value, reg->u32_max_value);
2103 /* Learn sign from unsigned bounds. Signed bounds cross the sign
2104 * boundary, so we must be careful.
2106 if ((s32)reg->u32_max_value >= 0) {
2107 /* Positive. We can't learn anything from the smin, but smax
2108 * is positive, hence safe.
2110 reg->s32_min_value = reg->u32_min_value;
2111 reg->s32_max_value = reg->u32_max_value =
2112 min_t(u32, reg->s32_max_value, reg->u32_max_value);
2113 } else if ((s32)reg->u32_min_value < 0) {
2114 /* Negative. We can't learn anything from the smax, but smin
2115 * is negative, hence safe.
2117 reg->s32_min_value = reg->u32_min_value =
2118 max_t(u32, reg->s32_min_value, reg->u32_min_value);
2119 reg->s32_max_value = reg->u32_max_value;
2123 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
2125 /* Learn sign from signed bounds.
2126 * If we cannot cross the sign boundary, then signed and unsigned bounds
2127 * are the same, so combine. This works even in the negative case, e.g.
2128 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
2130 if (reg->smin_value >= 0 || reg->smax_value < 0) {
2131 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
2133 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
2137 /* Learn sign from unsigned bounds. Signed bounds cross the sign
2138 * boundary, so we must be careful.
2140 if ((s64)reg->umax_value >= 0) {
2141 /* Positive. We can't learn anything from the smin, but smax
2142 * is positive, hence safe.
2144 reg->smin_value = reg->umin_value;
2145 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
2147 } else if ((s64)reg->umin_value < 0) {
2148 /* Negative. We can't learn anything from the smax, but smin
2149 * is negative, hence safe.
2151 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
2153 reg->smax_value = reg->umax_value;
2157 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
2159 __reg32_deduce_bounds(reg);
2160 __reg64_deduce_bounds(reg);
2163 /* Attempts to improve var_off based on unsigned min/max information */
2164 static void __reg_bound_offset(struct bpf_reg_state *reg)
2166 struct tnum var64_off = tnum_intersect(reg->var_off,
2167 tnum_range(reg->umin_value,
2169 struct tnum var32_off = tnum_intersect(tnum_subreg(var64_off),
2170 tnum_range(reg->u32_min_value,
2171 reg->u32_max_value));
2173 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
2176 static void reg_bounds_sync(struct bpf_reg_state *reg)
2178 /* We might have learned new bounds from the var_off. */
2179 __update_reg_bounds(reg);
2180 /* We might have learned something about the sign bit. */
2181 __reg_deduce_bounds(reg);
2182 /* We might have learned some bits from the bounds. */
2183 __reg_bound_offset(reg);
2184 /* Intersecting with the old var_off might have improved our bounds
2185 * slightly, e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2186 * then new var_off is (0; 0x7f...fc) which improves our umax.
2188 __update_reg_bounds(reg);
2191 static bool __reg32_bound_s64(s32 a)
2193 return a >= 0 && a <= S32_MAX;
2196 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
2198 reg->umin_value = reg->u32_min_value;
2199 reg->umax_value = reg->u32_max_value;
2201 /* Attempt to pull 32-bit signed bounds into 64-bit bounds but must
2202 * be positive otherwise set to worse case bounds and refine later
2205 if (__reg32_bound_s64(reg->s32_min_value) &&
2206 __reg32_bound_s64(reg->s32_max_value)) {
2207 reg->smin_value = reg->s32_min_value;
2208 reg->smax_value = reg->s32_max_value;
2210 reg->smin_value = 0;
2211 reg->smax_value = U32_MAX;
2215 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
2217 /* special case when 64-bit register has upper 32-bit register
2218 * zeroed. Typically happens after zext or <<32, >>32 sequence
2219 * allowing us to use 32-bit bounds directly,
2221 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
2222 __reg_assign_32_into_64(reg);
2224 /* Otherwise the best we can do is push lower 32bit known and
2225 * unknown bits into register (var_off set from jmp logic)
2226 * then learn as much as possible from the 64-bit tnum
2227 * known and unknown bits. The previous smin/smax bounds are
2228 * invalid here because of jmp32 compare so mark them unknown
2229 * so they do not impact tnum bounds calculation.
2231 __mark_reg64_unbounded(reg);
2233 reg_bounds_sync(reg);
2236 static bool __reg64_bound_s32(s64 a)
2238 return a >= S32_MIN && a <= S32_MAX;
2241 static bool __reg64_bound_u32(u64 a)
2243 return a >= U32_MIN && a <= U32_MAX;
2246 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
2248 __mark_reg32_unbounded(reg);
2249 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
2250 reg->s32_min_value = (s32)reg->smin_value;
2251 reg->s32_max_value = (s32)reg->smax_value;
2253 if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
2254 reg->u32_min_value = (u32)reg->umin_value;
2255 reg->u32_max_value = (u32)reg->umax_value;
2257 reg_bounds_sync(reg);
2260 /* Mark a register as having a completely unknown (scalar) value. */
2261 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
2262 struct bpf_reg_state *reg)
2265 * Clear type, off, and union(map_ptr, range) and
2266 * padding between 'type' and union
2268 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
2269 reg->type = SCALAR_VALUE;
2271 reg->ref_obj_id = 0;
2272 reg->var_off = tnum_unknown;
2274 reg->precise = !env->bpf_capable;
2275 __mark_reg_unbounded(reg);
2278 static void mark_reg_unknown(struct bpf_verifier_env *env,
2279 struct bpf_reg_state *regs, u32 regno)
2281 if (WARN_ON(regno >= MAX_BPF_REG)) {
2282 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
2283 /* Something bad happened, let's kill all regs except FP */
2284 for (regno = 0; regno < BPF_REG_FP; regno++)
2285 __mark_reg_not_init(env, regs + regno);
2288 __mark_reg_unknown(env, regs + regno);
2291 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
2292 struct bpf_reg_state *reg)
2294 __mark_reg_unknown(env, reg);
2295 reg->type = NOT_INIT;
2298 static void mark_reg_not_init(struct bpf_verifier_env *env,
2299 struct bpf_reg_state *regs, u32 regno)
2301 if (WARN_ON(regno >= MAX_BPF_REG)) {
2302 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
2303 /* Something bad happened, let's kill all regs except FP */
2304 for (regno = 0; regno < BPF_REG_FP; regno++)
2305 __mark_reg_not_init(env, regs + regno);
2308 __mark_reg_not_init(env, regs + regno);
2311 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
2312 struct bpf_reg_state *regs, u32 regno,
2313 enum bpf_reg_type reg_type,
2314 struct btf *btf, u32 btf_id,
2315 enum bpf_type_flag flag)
2317 if (reg_type == SCALAR_VALUE) {
2318 mark_reg_unknown(env, regs, regno);
2321 mark_reg_known_zero(env, regs, regno);
2322 regs[regno].type = PTR_TO_BTF_ID | flag;
2323 regs[regno].btf = btf;
2324 regs[regno].btf_id = btf_id;
2327 #define DEF_NOT_SUBREG (0)
2328 static void init_reg_state(struct bpf_verifier_env *env,
2329 struct bpf_func_state *state)
2331 struct bpf_reg_state *regs = state->regs;
2334 for (i = 0; i < MAX_BPF_REG; i++) {
2335 mark_reg_not_init(env, regs, i);
2336 regs[i].live = REG_LIVE_NONE;
2337 regs[i].parent = NULL;
2338 regs[i].subreg_def = DEF_NOT_SUBREG;
2342 regs[BPF_REG_FP].type = PTR_TO_STACK;
2343 mark_reg_known_zero(env, regs, BPF_REG_FP);
2344 regs[BPF_REG_FP].frameno = state->frameno;
2347 #define BPF_MAIN_FUNC (-1)
2348 static void init_func_state(struct bpf_verifier_env *env,
2349 struct bpf_func_state *state,
2350 int callsite, int frameno, int subprogno)
2352 state->callsite = callsite;
2353 state->frameno = frameno;
2354 state->subprogno = subprogno;
2355 state->callback_ret_range = tnum_range(0, 0);
2356 init_reg_state(env, state);
2357 mark_verifier_state_scratched(env);
2360 /* Similar to push_stack(), but for async callbacks */
2361 static struct bpf_verifier_state *push_async_cb(struct bpf_verifier_env *env,
2362 int insn_idx, int prev_insn_idx,
2365 struct bpf_verifier_stack_elem *elem;
2366 struct bpf_func_state *frame;
2368 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
2372 elem->insn_idx = insn_idx;
2373 elem->prev_insn_idx = prev_insn_idx;
2374 elem->next = env->head;
2375 elem->log_pos = env->log.end_pos;
2378 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
2380 "The sequence of %d jumps is too complex for async cb.\n",
2384 /* Unlike push_stack() do not copy_verifier_state().
2385 * The caller state doesn't matter.
2386 * This is async callback. It starts in a fresh stack.
2387 * Initialize it similar to do_check_common().
2389 elem->st.branches = 1;
2390 frame = kzalloc(sizeof(*frame), GFP_KERNEL);
2393 init_func_state(env, frame,
2394 BPF_MAIN_FUNC /* callsite */,
2395 0 /* frameno within this callchain */,
2396 subprog /* subprog number within this prog */);
2397 elem->st.frame[0] = frame;
2400 free_verifier_state(env->cur_state, true);
2401 env->cur_state = NULL;
2402 /* pop all elements and return */
2403 while (!pop_stack(env, NULL, NULL, false));
2409 SRC_OP, /* register is used as source operand */
2410 DST_OP, /* register is used as destination operand */
2411 DST_OP_NO_MARK /* same as above, check only, don't mark */
2414 static int cmp_subprogs(const void *a, const void *b)
2416 return ((struct bpf_subprog_info *)a)->start -
2417 ((struct bpf_subprog_info *)b)->start;
2420 static int find_subprog(struct bpf_verifier_env *env, int off)
2422 struct bpf_subprog_info *p;
2424 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
2425 sizeof(env->subprog_info[0]), cmp_subprogs);
2428 return p - env->subprog_info;
2432 static int add_subprog(struct bpf_verifier_env *env, int off)
2434 int insn_cnt = env->prog->len;
2437 if (off >= insn_cnt || off < 0) {
2438 verbose(env, "call to invalid destination\n");
2441 ret = find_subprog(env, off);
2444 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
2445 verbose(env, "too many subprograms\n");
2448 /* determine subprog starts. The end is one before the next starts */
2449 env->subprog_info[env->subprog_cnt++].start = off;
2450 sort(env->subprog_info, env->subprog_cnt,
2451 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
2452 return env->subprog_cnt - 1;
2455 #define MAX_KFUNC_DESCS 256
2456 #define MAX_KFUNC_BTFS 256
2458 struct bpf_kfunc_desc {
2459 struct btf_func_model func_model;
2466 struct bpf_kfunc_btf {
2468 struct module *module;
2472 struct bpf_kfunc_desc_tab {
2473 /* Sorted by func_id (BTF ID) and offset (fd_array offset) during
2474 * verification. JITs do lookups by bpf_insn, where func_id may not be
2475 * available, therefore at the end of verification do_misc_fixups()
2476 * sorts this by imm and offset.
2478 struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
2482 struct bpf_kfunc_btf_tab {
2483 struct bpf_kfunc_btf descs[MAX_KFUNC_BTFS];
2487 static int kfunc_desc_cmp_by_id_off(const void *a, const void *b)
2489 const struct bpf_kfunc_desc *d0 = a;
2490 const struct bpf_kfunc_desc *d1 = b;
2492 /* func_id is not greater than BTF_MAX_TYPE */
2493 return d0->func_id - d1->func_id ?: d0->offset - d1->offset;
2496 static int kfunc_btf_cmp_by_off(const void *a, const void *b)
2498 const struct bpf_kfunc_btf *d0 = a;
2499 const struct bpf_kfunc_btf *d1 = b;
2501 return d0->offset - d1->offset;
2504 static const struct bpf_kfunc_desc *
2505 find_kfunc_desc(const struct bpf_prog *prog, u32 func_id, u16 offset)
2507 struct bpf_kfunc_desc desc = {
2511 struct bpf_kfunc_desc_tab *tab;
2513 tab = prog->aux->kfunc_tab;
2514 return bsearch(&desc, tab->descs, tab->nr_descs,
2515 sizeof(tab->descs[0]), kfunc_desc_cmp_by_id_off);
2518 int bpf_get_kfunc_addr(const struct bpf_prog *prog, u32 func_id,
2519 u16 btf_fd_idx, u8 **func_addr)
2521 const struct bpf_kfunc_desc *desc;
2523 desc = find_kfunc_desc(prog, func_id, btf_fd_idx);
2527 *func_addr = (u8 *)desc->addr;
2531 static struct btf *__find_kfunc_desc_btf(struct bpf_verifier_env *env,
2534 struct bpf_kfunc_btf kf_btf = { .offset = offset };
2535 struct bpf_kfunc_btf_tab *tab;
2536 struct bpf_kfunc_btf *b;
2541 tab = env->prog->aux->kfunc_btf_tab;
2542 b = bsearch(&kf_btf, tab->descs, tab->nr_descs,
2543 sizeof(tab->descs[0]), kfunc_btf_cmp_by_off);
2545 if (tab->nr_descs == MAX_KFUNC_BTFS) {
2546 verbose(env, "too many different module BTFs\n");
2547 return ERR_PTR(-E2BIG);
2550 if (bpfptr_is_null(env->fd_array)) {
2551 verbose(env, "kfunc offset > 0 without fd_array is invalid\n");
2552 return ERR_PTR(-EPROTO);
2555 if (copy_from_bpfptr_offset(&btf_fd, env->fd_array,
2556 offset * sizeof(btf_fd),
2558 return ERR_PTR(-EFAULT);
2560 btf = btf_get_by_fd(btf_fd);
2562 verbose(env, "invalid module BTF fd specified\n");
2566 if (!btf_is_module(btf)) {
2567 verbose(env, "BTF fd for kfunc is not a module BTF\n");
2569 return ERR_PTR(-EINVAL);
2572 mod = btf_try_get_module(btf);
2575 return ERR_PTR(-ENXIO);
2578 b = &tab->descs[tab->nr_descs++];
2583 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2584 kfunc_btf_cmp_by_off, NULL);
2589 void bpf_free_kfunc_btf_tab(struct bpf_kfunc_btf_tab *tab)
2594 while (tab->nr_descs--) {
2595 module_put(tab->descs[tab->nr_descs].module);
2596 btf_put(tab->descs[tab->nr_descs].btf);
2601 static struct btf *find_kfunc_desc_btf(struct bpf_verifier_env *env, s16 offset)
2605 /* In the future, this can be allowed to increase limit
2606 * of fd index into fd_array, interpreted as u16.
2608 verbose(env, "negative offset disallowed for kernel module function call\n");
2609 return ERR_PTR(-EINVAL);
2612 return __find_kfunc_desc_btf(env, offset);
2614 return btf_vmlinux ?: ERR_PTR(-ENOENT);
2617 static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id, s16 offset)
2619 const struct btf_type *func, *func_proto;
2620 struct bpf_kfunc_btf_tab *btf_tab;
2621 struct bpf_kfunc_desc_tab *tab;
2622 struct bpf_prog_aux *prog_aux;
2623 struct bpf_kfunc_desc *desc;
2624 const char *func_name;
2625 struct btf *desc_btf;
2626 unsigned long call_imm;
2630 prog_aux = env->prog->aux;
2631 tab = prog_aux->kfunc_tab;
2632 btf_tab = prog_aux->kfunc_btf_tab;
2635 verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
2639 if (!env->prog->jit_requested) {
2640 verbose(env, "JIT is required for calling kernel function\n");
2644 if (!bpf_jit_supports_kfunc_call()) {
2645 verbose(env, "JIT does not support calling kernel function\n");
2649 if (!env->prog->gpl_compatible) {
2650 verbose(env, "cannot call kernel function from non-GPL compatible program\n");
2654 tab = kzalloc(sizeof(*tab), GFP_KERNEL);
2657 prog_aux->kfunc_tab = tab;
2660 /* func_id == 0 is always invalid, but instead of returning an error, be
2661 * conservative and wait until the code elimination pass before returning
2662 * error, so that invalid calls that get pruned out can be in BPF programs
2663 * loaded from userspace. It is also required that offset be untouched
2666 if (!func_id && !offset)
2669 if (!btf_tab && offset) {
2670 btf_tab = kzalloc(sizeof(*btf_tab), GFP_KERNEL);
2673 prog_aux->kfunc_btf_tab = btf_tab;
2676 desc_btf = find_kfunc_desc_btf(env, offset);
2677 if (IS_ERR(desc_btf)) {
2678 verbose(env, "failed to find BTF for kernel function\n");
2679 return PTR_ERR(desc_btf);
2682 if (find_kfunc_desc(env->prog, func_id, offset))
2685 if (tab->nr_descs == MAX_KFUNC_DESCS) {
2686 verbose(env, "too many different kernel function calls\n");
2690 func = btf_type_by_id(desc_btf, func_id);
2691 if (!func || !btf_type_is_func(func)) {
2692 verbose(env, "kernel btf_id %u is not a function\n",
2696 func_proto = btf_type_by_id(desc_btf, func->type);
2697 if (!func_proto || !btf_type_is_func_proto(func_proto)) {
2698 verbose(env, "kernel function btf_id %u does not have a valid func_proto\n",
2703 func_name = btf_name_by_offset(desc_btf, func->name_off);
2704 addr = kallsyms_lookup_name(func_name);
2706 verbose(env, "cannot find address for kernel function %s\n",
2710 specialize_kfunc(env, func_id, offset, &addr);
2712 if (bpf_jit_supports_far_kfunc_call()) {
2715 call_imm = BPF_CALL_IMM(addr);
2716 /* Check whether the relative offset overflows desc->imm */
2717 if ((unsigned long)(s32)call_imm != call_imm) {
2718 verbose(env, "address of kernel function %s is out of range\n",
2724 if (bpf_dev_bound_kfunc_id(func_id)) {
2725 err = bpf_dev_bound_kfunc_check(&env->log, prog_aux);
2730 desc = &tab->descs[tab->nr_descs++];
2731 desc->func_id = func_id;
2732 desc->imm = call_imm;
2733 desc->offset = offset;
2735 err = btf_distill_func_proto(&env->log, desc_btf,
2736 func_proto, func_name,
2739 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2740 kfunc_desc_cmp_by_id_off, NULL);
2744 static int kfunc_desc_cmp_by_imm_off(const void *a, const void *b)
2746 const struct bpf_kfunc_desc *d0 = a;
2747 const struct bpf_kfunc_desc *d1 = b;
2749 if (d0->imm != d1->imm)
2750 return d0->imm < d1->imm ? -1 : 1;
2751 if (d0->offset != d1->offset)
2752 return d0->offset < d1->offset ? -1 : 1;
2756 static void sort_kfunc_descs_by_imm_off(struct bpf_prog *prog)
2758 struct bpf_kfunc_desc_tab *tab;
2760 tab = prog->aux->kfunc_tab;
2764 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2765 kfunc_desc_cmp_by_imm_off, NULL);
2768 bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
2770 return !!prog->aux->kfunc_tab;
2773 const struct btf_func_model *
2774 bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
2775 const struct bpf_insn *insn)
2777 const struct bpf_kfunc_desc desc = {
2779 .offset = insn->off,
2781 const struct bpf_kfunc_desc *res;
2782 struct bpf_kfunc_desc_tab *tab;
2784 tab = prog->aux->kfunc_tab;
2785 res = bsearch(&desc, tab->descs, tab->nr_descs,
2786 sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm_off);
2788 return res ? &res->func_model : NULL;
2791 static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
2793 struct bpf_subprog_info *subprog = env->subprog_info;
2794 struct bpf_insn *insn = env->prog->insnsi;
2795 int i, ret, insn_cnt = env->prog->len;
2797 /* Add entry function. */
2798 ret = add_subprog(env, 0);
2802 for (i = 0; i < insn_cnt; i++, insn++) {
2803 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
2804 !bpf_pseudo_kfunc_call(insn))
2807 if (!env->bpf_capable) {
2808 verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
2812 if (bpf_pseudo_func(insn) || bpf_pseudo_call(insn))
2813 ret = add_subprog(env, i + insn->imm + 1);
2815 ret = add_kfunc_call(env, insn->imm, insn->off);
2821 /* Add a fake 'exit' subprog which could simplify subprog iteration
2822 * logic. 'subprog_cnt' should not be increased.
2824 subprog[env->subprog_cnt].start = insn_cnt;
2826 if (env->log.level & BPF_LOG_LEVEL2)
2827 for (i = 0; i < env->subprog_cnt; i++)
2828 verbose(env, "func#%d @%d\n", i, subprog[i].start);
2833 static int check_subprogs(struct bpf_verifier_env *env)
2835 int i, subprog_start, subprog_end, off, cur_subprog = 0;
2836 struct bpf_subprog_info *subprog = env->subprog_info;
2837 struct bpf_insn *insn = env->prog->insnsi;
2838 int insn_cnt = env->prog->len;
2840 /* now check that all jumps are within the same subprog */
2841 subprog_start = subprog[cur_subprog].start;
2842 subprog_end = subprog[cur_subprog + 1].start;
2843 for (i = 0; i < insn_cnt; i++) {
2844 u8 code = insn[i].code;
2846 if (code == (BPF_JMP | BPF_CALL) &&
2847 insn[i].src_reg == 0 &&
2848 insn[i].imm == BPF_FUNC_tail_call)
2849 subprog[cur_subprog].has_tail_call = true;
2850 if (BPF_CLASS(code) == BPF_LD &&
2851 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
2852 subprog[cur_subprog].has_ld_abs = true;
2853 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
2855 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
2857 off = i + insn[i].off + 1;
2858 if (off < subprog_start || off >= subprog_end) {
2859 verbose(env, "jump out of range from insn %d to %d\n", i, off);
2863 if (i == subprog_end - 1) {
2864 /* to avoid fall-through from one subprog into another
2865 * the last insn of the subprog should be either exit
2866 * or unconditional jump back
2868 if (code != (BPF_JMP | BPF_EXIT) &&
2869 code != (BPF_JMP | BPF_JA)) {
2870 verbose(env, "last insn is not an exit or jmp\n");
2873 subprog_start = subprog_end;
2875 if (cur_subprog < env->subprog_cnt)
2876 subprog_end = subprog[cur_subprog + 1].start;
2882 /* Parentage chain of this register (or stack slot) should take care of all
2883 * issues like callee-saved registers, stack slot allocation time, etc.
2885 static int mark_reg_read(struct bpf_verifier_env *env,
2886 const struct bpf_reg_state *state,
2887 struct bpf_reg_state *parent, u8 flag)
2889 bool writes = parent == state->parent; /* Observe write marks */
2893 /* if read wasn't screened by an earlier write ... */
2894 if (writes && state->live & REG_LIVE_WRITTEN)
2896 if (parent->live & REG_LIVE_DONE) {
2897 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
2898 reg_type_str(env, parent->type),
2899 parent->var_off.value, parent->off);
2902 /* The first condition is more likely to be true than the
2903 * second, checked it first.
2905 if ((parent->live & REG_LIVE_READ) == flag ||
2906 parent->live & REG_LIVE_READ64)
2907 /* The parentage chain never changes and
2908 * this parent was already marked as LIVE_READ.
2909 * There is no need to keep walking the chain again and
2910 * keep re-marking all parents as LIVE_READ.
2911 * This case happens when the same register is read
2912 * multiple times without writes into it in-between.
2913 * Also, if parent has the stronger REG_LIVE_READ64 set,
2914 * then no need to set the weak REG_LIVE_READ32.
2917 /* ... then we depend on parent's value */
2918 parent->live |= flag;
2919 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
2920 if (flag == REG_LIVE_READ64)
2921 parent->live &= ~REG_LIVE_READ32;
2923 parent = state->parent;
2928 if (env->longest_mark_read_walk < cnt)
2929 env->longest_mark_read_walk = cnt;
2933 static int mark_dynptr_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
2935 struct bpf_func_state *state = func(env, reg);
2938 /* For CONST_PTR_TO_DYNPTR, it must have already been done by
2939 * check_reg_arg in check_helper_call and mark_btf_func_reg_size in
2942 if (reg->type == CONST_PTR_TO_DYNPTR)
2944 spi = dynptr_get_spi(env, reg);
2947 /* Caller ensures dynptr is valid and initialized, which means spi is in
2948 * bounds and spi is the first dynptr slot. Simply mark stack slot as
2951 ret = mark_reg_read(env, &state->stack[spi].spilled_ptr,
2952 state->stack[spi].spilled_ptr.parent, REG_LIVE_READ64);
2955 return mark_reg_read(env, &state->stack[spi - 1].spilled_ptr,
2956 state->stack[spi - 1].spilled_ptr.parent, REG_LIVE_READ64);
2959 static int mark_iter_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
2960 int spi, int nr_slots)
2962 struct bpf_func_state *state = func(env, reg);
2965 for (i = 0; i < nr_slots; i++) {
2966 struct bpf_reg_state *st = &state->stack[spi - i].spilled_ptr;
2968 err = mark_reg_read(env, st, st->parent, REG_LIVE_READ64);
2972 mark_stack_slot_scratched(env, spi - i);
2978 /* This function is supposed to be used by the following 32-bit optimization
2979 * code only. It returns TRUE if the source or destination register operates
2980 * on 64-bit, otherwise return FALSE.
2982 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
2983 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
2988 class = BPF_CLASS(code);
2990 if (class == BPF_JMP) {
2991 /* BPF_EXIT for "main" will reach here. Return TRUE
2996 if (op == BPF_CALL) {
2997 /* BPF to BPF call will reach here because of marking
2998 * caller saved clobber with DST_OP_NO_MARK for which we
2999 * don't care the register def because they are anyway
3000 * marked as NOT_INIT already.
3002 if (insn->src_reg == BPF_PSEUDO_CALL)
3004 /* Helper call will reach here because of arg type
3005 * check, conservatively return TRUE.
3014 if (class == BPF_ALU64 || class == BPF_JMP ||
3015 /* BPF_END always use BPF_ALU class. */
3016 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
3019 if (class == BPF_ALU || class == BPF_JMP32)
3022 if (class == BPF_LDX) {
3024 return BPF_SIZE(code) == BPF_DW;
3025 /* LDX source must be ptr. */
3029 if (class == BPF_STX) {
3030 /* BPF_STX (including atomic variants) has multiple source
3031 * operands, one of which is a ptr. Check whether the caller is
3034 if (t == SRC_OP && reg->type != SCALAR_VALUE)
3036 return BPF_SIZE(code) == BPF_DW;
3039 if (class == BPF_LD) {
3040 u8 mode = BPF_MODE(code);
3043 if (mode == BPF_IMM)
3046 /* Both LD_IND and LD_ABS return 32-bit data. */
3050 /* Implicit ctx ptr. */
3051 if (regno == BPF_REG_6)
3054 /* Explicit source could be any width. */
3058 if (class == BPF_ST)
3059 /* The only source register for BPF_ST is a ptr. */
3062 /* Conservatively return true at default. */
3066 /* Return the regno defined by the insn, or -1. */
3067 static int insn_def_regno(const struct bpf_insn *insn)
3069 switch (BPF_CLASS(insn->code)) {
3075 if (BPF_MODE(insn->code) == BPF_ATOMIC &&
3076 (insn->imm & BPF_FETCH)) {
3077 if (insn->imm == BPF_CMPXCHG)
3080 return insn->src_reg;
3085 return insn->dst_reg;
3089 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
3090 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
3092 int dst_reg = insn_def_regno(insn);
3097 return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
3100 static void mark_insn_zext(struct bpf_verifier_env *env,
3101 struct bpf_reg_state *reg)
3103 s32 def_idx = reg->subreg_def;
3105 if (def_idx == DEF_NOT_SUBREG)
3108 env->insn_aux_data[def_idx - 1].zext_dst = true;
3109 /* The dst will be zero extended, so won't be sub-register anymore. */
3110 reg->subreg_def = DEF_NOT_SUBREG;
3113 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
3114 enum reg_arg_type t)
3116 struct bpf_verifier_state *vstate = env->cur_state;
3117 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3118 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
3119 struct bpf_reg_state *reg, *regs = state->regs;
3122 if (regno >= MAX_BPF_REG) {
3123 verbose(env, "R%d is invalid\n", regno);
3127 mark_reg_scratched(env, regno);
3130 rw64 = is_reg64(env, insn, regno, reg, t);
3132 /* check whether register used as source operand can be read */
3133 if (reg->type == NOT_INIT) {
3134 verbose(env, "R%d !read_ok\n", regno);
3137 /* We don't need to worry about FP liveness because it's read-only */
3138 if (regno == BPF_REG_FP)
3142 mark_insn_zext(env, reg);
3144 return mark_reg_read(env, reg, reg->parent,
3145 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
3147 /* check whether register used as dest operand can be written to */
3148 if (regno == BPF_REG_FP) {
3149 verbose(env, "frame pointer is read only\n");
3152 reg->live |= REG_LIVE_WRITTEN;
3153 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
3155 mark_reg_unknown(env, regs, regno);
3160 static void mark_jmp_point(struct bpf_verifier_env *env, int idx)
3162 env->insn_aux_data[idx].jmp_point = true;
3165 static bool is_jmp_point(struct bpf_verifier_env *env, int insn_idx)
3167 return env->insn_aux_data[insn_idx].jmp_point;
3170 /* for any branch, call, exit record the history of jmps in the given state */
3171 static int push_jmp_history(struct bpf_verifier_env *env,
3172 struct bpf_verifier_state *cur)
3174 u32 cnt = cur->jmp_history_cnt;
3175 struct bpf_idx_pair *p;
3178 if (!is_jmp_point(env, env->insn_idx))
3182 alloc_size = kmalloc_size_roundup(size_mul(cnt, sizeof(*p)));
3183 p = krealloc(cur->jmp_history, alloc_size, GFP_USER);
3186 p[cnt - 1].idx = env->insn_idx;
3187 p[cnt - 1].prev_idx = env->prev_insn_idx;
3188 cur->jmp_history = p;
3189 cur->jmp_history_cnt = cnt;
3193 /* Backtrack one insn at a time. If idx is not at the top of recorded
3194 * history then previous instruction came from straight line execution.
3196 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
3201 if (cnt && st->jmp_history[cnt - 1].idx == i) {
3202 i = st->jmp_history[cnt - 1].prev_idx;
3210 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
3212 const struct btf_type *func;
3213 struct btf *desc_btf;
3215 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
3218 desc_btf = find_kfunc_desc_btf(data, insn->off);
3219 if (IS_ERR(desc_btf))
3222 func = btf_type_by_id(desc_btf, insn->imm);
3223 return btf_name_by_offset(desc_btf, func->name_off);
3226 static inline void bt_init(struct backtrack_state *bt, u32 frame)
3231 static inline void bt_reset(struct backtrack_state *bt)
3233 struct bpf_verifier_env *env = bt->env;
3235 memset(bt, 0, sizeof(*bt));
3239 static inline u32 bt_empty(struct backtrack_state *bt)
3244 for (i = 0; i <= bt->frame; i++)
3245 mask |= bt->reg_masks[i] | bt->stack_masks[i];
3250 static inline int bt_subprog_enter(struct backtrack_state *bt)
3252 if (bt->frame == MAX_CALL_FRAMES - 1) {
3253 verbose(bt->env, "BUG subprog enter from frame %d\n", bt->frame);
3254 WARN_ONCE(1, "verifier backtracking bug");
3261 static inline int bt_subprog_exit(struct backtrack_state *bt)
3263 if (bt->frame == 0) {
3264 verbose(bt->env, "BUG subprog exit from frame 0\n");
3265 WARN_ONCE(1, "verifier backtracking bug");
3272 static inline void bt_set_frame_reg(struct backtrack_state *bt, u32 frame, u32 reg)
3274 bt->reg_masks[frame] |= 1 << reg;
3277 static inline void bt_clear_frame_reg(struct backtrack_state *bt, u32 frame, u32 reg)
3279 bt->reg_masks[frame] &= ~(1 << reg);
3282 static inline void bt_set_reg(struct backtrack_state *bt, u32 reg)
3284 bt_set_frame_reg(bt, bt->frame, reg);
3287 static inline void bt_clear_reg(struct backtrack_state *bt, u32 reg)
3289 bt_clear_frame_reg(bt, bt->frame, reg);
3292 static inline void bt_set_frame_slot(struct backtrack_state *bt, u32 frame, u32 slot)
3294 bt->stack_masks[frame] |= 1ull << slot;
3297 static inline void bt_clear_frame_slot(struct backtrack_state *bt, u32 frame, u32 slot)
3299 bt->stack_masks[frame] &= ~(1ull << slot);
3302 static inline void bt_set_slot(struct backtrack_state *bt, u32 slot)
3304 bt_set_frame_slot(bt, bt->frame, slot);
3307 static inline void bt_clear_slot(struct backtrack_state *bt, u32 slot)
3309 bt_clear_frame_slot(bt, bt->frame, slot);
3312 static inline u32 bt_frame_reg_mask(struct backtrack_state *bt, u32 frame)
3314 return bt->reg_masks[frame];
3317 static inline u32 bt_reg_mask(struct backtrack_state *bt)
3319 return bt->reg_masks[bt->frame];
3322 static inline u64 bt_frame_stack_mask(struct backtrack_state *bt, u32 frame)
3324 return bt->stack_masks[frame];
3327 static inline u64 bt_stack_mask(struct backtrack_state *bt)
3329 return bt->stack_masks[bt->frame];
3332 static inline bool bt_is_reg_set(struct backtrack_state *bt, u32 reg)
3334 return bt->reg_masks[bt->frame] & (1 << reg);
3337 static inline bool bt_is_slot_set(struct backtrack_state *bt, u32 slot)
3339 return bt->stack_masks[bt->frame] & (1ull << slot);
3342 /* format registers bitmask, e.g., "r0,r2,r4" for 0x15 mask */
3343 static void fmt_reg_mask(char *buf, ssize_t buf_sz, u32 reg_mask)
3345 DECLARE_BITMAP(mask, 64);
3351 bitmap_from_u64(mask, reg_mask);
3352 for_each_set_bit(i, mask, 32) {
3353 n = snprintf(buf, buf_sz, "%sr%d", first ? "" : ",", i);
3361 /* format stack slots bitmask, e.g., "-8,-24,-40" for 0x15 mask */
3362 static void fmt_stack_mask(char *buf, ssize_t buf_sz, u64 stack_mask)
3364 DECLARE_BITMAP(mask, 64);
3370 bitmap_from_u64(mask, stack_mask);
3371 for_each_set_bit(i, mask, 64) {
3372 n = snprintf(buf, buf_sz, "%s%d", first ? "" : ",", -(i + 1) * 8);
3381 /* For given verifier state backtrack_insn() is called from the last insn to
3382 * the first insn. Its purpose is to compute a bitmask of registers and
3383 * stack slots that needs precision in the parent verifier state.
3385 * @idx is an index of the instruction we are currently processing;
3386 * @subseq_idx is an index of the subsequent instruction that:
3387 * - *would be* executed next, if jump history is viewed in forward order;
3388 * - *was* processed previously during backtracking.
3390 static int backtrack_insn(struct bpf_verifier_env *env, int idx, int subseq_idx,
3391 struct backtrack_state *bt)
3393 const struct bpf_insn_cbs cbs = {
3394 .cb_call = disasm_kfunc_name,
3395 .cb_print = verbose,
3396 .private_data = env,
3398 struct bpf_insn *insn = env->prog->insnsi + idx;
3399 u8 class = BPF_CLASS(insn->code);
3400 u8 opcode = BPF_OP(insn->code);
3401 u8 mode = BPF_MODE(insn->code);
3402 u32 dreg = insn->dst_reg;
3403 u32 sreg = insn->src_reg;
3406 if (insn->code == 0)
3408 if (env->log.level & BPF_LOG_LEVEL2) {
3409 fmt_reg_mask(env->tmp_str_buf, TMP_STR_BUF_LEN, bt_reg_mask(bt));
3410 verbose(env, "mark_precise: frame%d: regs=%s ",
3411 bt->frame, env->tmp_str_buf);
3412 fmt_stack_mask(env->tmp_str_buf, TMP_STR_BUF_LEN, bt_stack_mask(bt));
3413 verbose(env, "stack=%s before ", env->tmp_str_buf);
3414 verbose(env, "%d: ", idx);
3415 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
3418 if (class == BPF_ALU || class == BPF_ALU64) {
3419 if (!bt_is_reg_set(bt, dreg))
3421 if (opcode == BPF_MOV) {
3422 if (BPF_SRC(insn->code) == BPF_X) {
3424 * dreg needs precision after this insn
3425 * sreg needs precision before this insn
3427 bt_clear_reg(bt, dreg);
3428 bt_set_reg(bt, sreg);
3431 * dreg needs precision after this insn.
3432 * Corresponding register is already marked
3433 * as precise=true in this verifier state.
3434 * No further markings in parent are necessary
3436 bt_clear_reg(bt, dreg);
3439 if (BPF_SRC(insn->code) == BPF_X) {
3441 * both dreg and sreg need precision
3444 bt_set_reg(bt, sreg);
3446 * dreg still needs precision before this insn
3449 } else if (class == BPF_LDX) {
3450 if (!bt_is_reg_set(bt, dreg))
3452 bt_clear_reg(bt, dreg);
3454 /* scalars can only be spilled into stack w/o losing precision.
3455 * Load from any other memory can be zero extended.
3456 * The desire to keep that precision is already indicated
3457 * by 'precise' mark in corresponding register of this state.
3458 * No further tracking necessary.
3460 if (insn->src_reg != BPF_REG_FP)
3463 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
3464 * that [fp - off] slot contains scalar that needs to be
3465 * tracked with precision
3467 spi = (-insn->off - 1) / BPF_REG_SIZE;
3469 verbose(env, "BUG spi %d\n", spi);
3470 WARN_ONCE(1, "verifier backtracking bug");
3473 bt_set_slot(bt, spi);
3474 } else if (class == BPF_STX || class == BPF_ST) {
3475 if (bt_is_reg_set(bt, dreg))
3476 /* stx & st shouldn't be using _scalar_ dst_reg
3477 * to access memory. It means backtracking
3478 * encountered a case of pointer subtraction.
3481 /* scalars can only be spilled into stack */
3482 if (insn->dst_reg != BPF_REG_FP)
3484 spi = (-insn->off - 1) / BPF_REG_SIZE;
3486 verbose(env, "BUG spi %d\n", spi);
3487 WARN_ONCE(1, "verifier backtracking bug");
3490 if (!bt_is_slot_set(bt, spi))
3492 bt_clear_slot(bt, spi);
3493 if (class == BPF_STX)
3494 bt_set_reg(bt, sreg);
3495 } else if (class == BPF_JMP || class == BPF_JMP32) {
3496 if (bpf_pseudo_call(insn)) {
3497 int subprog_insn_idx, subprog;
3499 subprog_insn_idx = idx + insn->imm + 1;
3500 subprog = find_subprog(env, subprog_insn_idx);
3504 if (subprog_is_global(env, subprog)) {
3505 /* check that jump history doesn't have any
3506 * extra instructions from subprog; the next
3507 * instruction after call to global subprog
3508 * should be literally next instruction in
3511 WARN_ONCE(idx + 1 != subseq_idx, "verifier backtracking bug");
3512 /* r1-r5 are invalidated after subprog call,
3513 * so for global func call it shouldn't be set
3516 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3517 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3518 WARN_ONCE(1, "verifier backtracking bug");
3521 /* global subprog always sets R0 */
3522 bt_clear_reg(bt, BPF_REG_0);
3525 /* static subprog call instruction, which
3526 * means that we are exiting current subprog,
3527 * so only r1-r5 could be still requested as
3528 * precise, r0 and r6-r10 or any stack slot in
3529 * the current frame should be zero by now
3531 if (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) {
3532 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3533 WARN_ONCE(1, "verifier backtracking bug");
3536 /* we don't track register spills perfectly,
3537 * so fallback to force-precise instead of failing */
3538 if (bt_stack_mask(bt) != 0)
3540 /* propagate r1-r5 to the caller */
3541 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
3542 if (bt_is_reg_set(bt, i)) {
3543 bt_clear_reg(bt, i);
3544 bt_set_frame_reg(bt, bt->frame - 1, i);
3547 if (bt_subprog_exit(bt))
3551 } else if ((bpf_helper_call(insn) &&
3552 is_callback_calling_function(insn->imm) &&
3553 !is_async_callback_calling_function(insn->imm)) ||
3554 (bpf_pseudo_kfunc_call(insn) && is_callback_calling_kfunc(insn->imm))) {
3555 /* callback-calling helper or kfunc call, which means
3556 * we are exiting from subprog, but unlike the subprog
3557 * call handling above, we shouldn't propagate
3558 * precision of r1-r5 (if any requested), as they are
3559 * not actually arguments passed directly to callback
3562 if (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) {
3563 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3564 WARN_ONCE(1, "verifier backtracking bug");
3567 if (bt_stack_mask(bt) != 0)
3569 /* clear r1-r5 in callback subprog's mask */
3570 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
3571 bt_clear_reg(bt, i);
3572 if (bt_subprog_exit(bt))
3575 } else if (opcode == BPF_CALL) {
3576 /* kfunc with imm==0 is invalid and fixup_kfunc_call will
3577 * catch this error later. Make backtracking conservative
3580 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && insn->imm == 0)
3582 /* regular helper call sets R0 */
3583 bt_clear_reg(bt, BPF_REG_0);
3584 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3585 /* if backtracing was looking for registers R1-R5
3586 * they should have been found already.
3588 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3589 WARN_ONCE(1, "verifier backtracking bug");
3592 } else if (opcode == BPF_EXIT) {
3595 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3596 /* if backtracing was looking for registers R1-R5
3597 * they should have been found already.
3599 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3600 WARN_ONCE(1, "verifier backtracking bug");
3604 /* BPF_EXIT in subprog or callback always returns
3605 * right after the call instruction, so by checking
3606 * whether the instruction at subseq_idx-1 is subprog
3607 * call or not we can distinguish actual exit from
3608 * *subprog* from exit from *callback*. In the former
3609 * case, we need to propagate r0 precision, if
3610 * necessary. In the former we never do that.
3612 r0_precise = subseq_idx - 1 >= 0 &&
3613 bpf_pseudo_call(&env->prog->insnsi[subseq_idx - 1]) &&
3614 bt_is_reg_set(bt, BPF_REG_0);
3616 bt_clear_reg(bt, BPF_REG_0);
3617 if (bt_subprog_enter(bt))
3621 bt_set_reg(bt, BPF_REG_0);
3622 /* r6-r9 and stack slots will stay set in caller frame
3623 * bitmasks until we return back from callee(s)
3626 } else if (BPF_SRC(insn->code) == BPF_X) {
3627 if (!bt_is_reg_set(bt, dreg) && !bt_is_reg_set(bt, sreg))
3630 * Both dreg and sreg need precision before
3631 * this insn. If only sreg was marked precise
3632 * before it would be equally necessary to
3633 * propagate it to dreg.
3635 bt_set_reg(bt, dreg);
3636 bt_set_reg(bt, sreg);
3637 /* else dreg <cond> K
3638 * Only dreg still needs precision before
3639 * this insn, so for the K-based conditional
3640 * there is nothing new to be marked.
3643 } else if (class == BPF_LD) {
3644 if (!bt_is_reg_set(bt, dreg))
3646 bt_clear_reg(bt, dreg);
3647 /* It's ld_imm64 or ld_abs or ld_ind.
3648 * For ld_imm64 no further tracking of precision
3649 * into parent is necessary
3651 if (mode == BPF_IND || mode == BPF_ABS)
3652 /* to be analyzed */
3658 /* the scalar precision tracking algorithm:
3659 * . at the start all registers have precise=false.
3660 * . scalar ranges are tracked as normal through alu and jmp insns.
3661 * . once precise value of the scalar register is used in:
3662 * . ptr + scalar alu
3663 * . if (scalar cond K|scalar)
3664 * . helper_call(.., scalar, ...) where ARG_CONST is expected
3665 * backtrack through the verifier states and mark all registers and
3666 * stack slots with spilled constants that these scalar regisers
3667 * should be precise.
3668 * . during state pruning two registers (or spilled stack slots)
3669 * are equivalent if both are not precise.
3671 * Note the verifier cannot simply walk register parentage chain,
3672 * since many different registers and stack slots could have been
3673 * used to compute single precise scalar.
3675 * The approach of starting with precise=true for all registers and then
3676 * backtrack to mark a register as not precise when the verifier detects
3677 * that program doesn't care about specific value (e.g., when helper
3678 * takes register as ARG_ANYTHING parameter) is not safe.
3680 * It's ok to walk single parentage chain of the verifier states.
3681 * It's possible that this backtracking will go all the way till 1st insn.
3682 * All other branches will be explored for needing precision later.
3684 * The backtracking needs to deal with cases like:
3685 * 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)
3688 * if r5 > 0x79f goto pc+7
3689 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
3692 * call bpf_perf_event_output#25
3693 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
3697 * call foo // uses callee's r6 inside to compute r0
3701 * to track above reg_mask/stack_mask needs to be independent for each frame.
3703 * Also if parent's curframe > frame where backtracking started,
3704 * the verifier need to mark registers in both frames, otherwise callees
3705 * may incorrectly prune callers. This is similar to
3706 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
3708 * For now backtracking falls back into conservative marking.
3710 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
3711 struct bpf_verifier_state *st)
3713 struct bpf_func_state *func;
3714 struct bpf_reg_state *reg;
3717 if (env->log.level & BPF_LOG_LEVEL2) {
3718 verbose(env, "mark_precise: frame%d: falling back to forcing all scalars precise\n",
3722 /* big hammer: mark all scalars precise in this path.
3723 * pop_stack may still get !precise scalars.
3724 * We also skip current state and go straight to first parent state,
3725 * because precision markings in current non-checkpointed state are
3726 * not needed. See why in the comment in __mark_chain_precision below.
3728 for (st = st->parent; st; st = st->parent) {
3729 for (i = 0; i <= st->curframe; i++) {
3730 func = st->frame[i];
3731 for (j = 0; j < BPF_REG_FP; j++) {
3732 reg = &func->regs[j];
3733 if (reg->type != SCALAR_VALUE || reg->precise)
3735 reg->precise = true;
3736 if (env->log.level & BPF_LOG_LEVEL2) {
3737 verbose(env, "force_precise: frame%d: forcing r%d to be precise\n",
3741 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
3742 if (!is_spilled_reg(&func->stack[j]))
3744 reg = &func->stack[j].spilled_ptr;
3745 if (reg->type != SCALAR_VALUE || reg->precise)
3747 reg->precise = true;
3748 if (env->log.level & BPF_LOG_LEVEL2) {
3749 verbose(env, "force_precise: frame%d: forcing fp%d to be precise\n",
3757 static void mark_all_scalars_imprecise(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
3759 struct bpf_func_state *func;
3760 struct bpf_reg_state *reg;
3763 for (i = 0; i <= st->curframe; i++) {
3764 func = st->frame[i];
3765 for (j = 0; j < BPF_REG_FP; j++) {
3766 reg = &func->regs[j];
3767 if (reg->type != SCALAR_VALUE)
3769 reg->precise = false;
3771 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
3772 if (!is_spilled_reg(&func->stack[j]))
3774 reg = &func->stack[j].spilled_ptr;
3775 if (reg->type != SCALAR_VALUE)
3777 reg->precise = false;
3782 static bool idset_contains(struct bpf_idset *s, u32 id)
3786 for (i = 0; i < s->count; ++i)
3787 if (s->ids[i] == id)
3793 static int idset_push(struct bpf_idset *s, u32 id)
3795 if (WARN_ON_ONCE(s->count >= ARRAY_SIZE(s->ids)))
3797 s->ids[s->count++] = id;
3801 static void idset_reset(struct bpf_idset *s)
3806 /* Collect a set of IDs for all registers currently marked as precise in env->bt.
3807 * Mark all registers with these IDs as precise.
3809 static int mark_precise_scalar_ids(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
3811 struct bpf_idset *precise_ids = &env->idset_scratch;
3812 struct backtrack_state *bt = &env->bt;
3813 struct bpf_func_state *func;
3814 struct bpf_reg_state *reg;
3815 DECLARE_BITMAP(mask, 64);
3818 idset_reset(precise_ids);
3820 for (fr = bt->frame; fr >= 0; fr--) {
3821 func = st->frame[fr];
3823 bitmap_from_u64(mask, bt_frame_reg_mask(bt, fr));
3824 for_each_set_bit(i, mask, 32) {
3825 reg = &func->regs[i];
3826 if (!reg->id || reg->type != SCALAR_VALUE)
3828 if (idset_push(precise_ids, reg->id))
3832 bitmap_from_u64(mask, bt_frame_stack_mask(bt, fr));
3833 for_each_set_bit(i, mask, 64) {
3834 if (i >= func->allocated_stack / BPF_REG_SIZE)
3836 if (!is_spilled_scalar_reg(&func->stack[i]))
3838 reg = &func->stack[i].spilled_ptr;
3841 if (idset_push(precise_ids, reg->id))
3846 for (fr = 0; fr <= st->curframe; ++fr) {
3847 func = st->frame[fr];
3849 for (i = BPF_REG_0; i < BPF_REG_10; ++i) {
3850 reg = &func->regs[i];
3853 if (!idset_contains(precise_ids, reg->id))
3855 bt_set_frame_reg(bt, fr, i);
3857 for (i = 0; i < func->allocated_stack / BPF_REG_SIZE; ++i) {
3858 if (!is_spilled_scalar_reg(&func->stack[i]))
3860 reg = &func->stack[i].spilled_ptr;
3863 if (!idset_contains(precise_ids, reg->id))
3865 bt_set_frame_slot(bt, fr, i);
3873 * __mark_chain_precision() backtracks BPF program instruction sequence and
3874 * chain of verifier states making sure that register *regno* (if regno >= 0)
3875 * and/or stack slot *spi* (if spi >= 0) are marked as precisely tracked
3876 * SCALARS, as well as any other registers and slots that contribute to
3877 * a tracked state of given registers/stack slots, depending on specific BPF
3878 * assembly instructions (see backtrack_insns() for exact instruction handling
3879 * logic). This backtracking relies on recorded jmp_history and is able to
3880 * traverse entire chain of parent states. This process ends only when all the
3881 * necessary registers/slots and their transitive dependencies are marked as
3884 * One important and subtle aspect is that precise marks *do not matter* in
3885 * the currently verified state (current state). It is important to understand
3886 * why this is the case.
3888 * First, note that current state is the state that is not yet "checkpointed",
3889 * i.e., it is not yet put into env->explored_states, and it has no children
3890 * states as well. It's ephemeral, and can end up either a) being discarded if
3891 * compatible explored state is found at some point or BPF_EXIT instruction is
3892 * reached or b) checkpointed and put into env->explored_states, branching out
3893 * into one or more children states.
3895 * In the former case, precise markings in current state are completely
3896 * ignored by state comparison code (see regsafe() for details). Only
3897 * checkpointed ("old") state precise markings are important, and if old
3898 * state's register/slot is precise, regsafe() assumes current state's
3899 * register/slot as precise and checks value ranges exactly and precisely. If
3900 * states turn out to be compatible, current state's necessary precise
3901 * markings and any required parent states' precise markings are enforced
3902 * after the fact with propagate_precision() logic, after the fact. But it's
3903 * important to realize that in this case, even after marking current state
3904 * registers/slots as precise, we immediately discard current state. So what
3905 * actually matters is any of the precise markings propagated into current
3906 * state's parent states, which are always checkpointed (due to b) case above).
3907 * As such, for scenario a) it doesn't matter if current state has precise
3908 * markings set or not.
3910 * Now, for the scenario b), checkpointing and forking into child(ren)
3911 * state(s). Note that before current state gets to checkpointing step, any
3912 * processed instruction always assumes precise SCALAR register/slot
3913 * knowledge: if precise value or range is useful to prune jump branch, BPF
3914 * verifier takes this opportunity enthusiastically. Similarly, when
3915 * register's value is used to calculate offset or memory address, exact
3916 * knowledge of SCALAR range is assumed, checked, and enforced. So, similar to
3917 * what we mentioned above about state comparison ignoring precise markings
3918 * during state comparison, BPF verifier ignores and also assumes precise
3919 * markings *at will* during instruction verification process. But as verifier
3920 * assumes precision, it also propagates any precision dependencies across
3921 * parent states, which are not yet finalized, so can be further restricted
3922 * based on new knowledge gained from restrictions enforced by their children
3923 * states. This is so that once those parent states are finalized, i.e., when
3924 * they have no more active children state, state comparison logic in
3925 * is_state_visited() would enforce strict and precise SCALAR ranges, if
3926 * required for correctness.
3928 * To build a bit more intuition, note also that once a state is checkpointed,
3929 * the path we took to get to that state is not important. This is crucial
3930 * property for state pruning. When state is checkpointed and finalized at
3931 * some instruction index, it can be correctly and safely used to "short
3932 * circuit" any *compatible* state that reaches exactly the same instruction
3933 * index. I.e., if we jumped to that instruction from a completely different
3934 * code path than original finalized state was derived from, it doesn't
3935 * matter, current state can be discarded because from that instruction
3936 * forward having a compatible state will ensure we will safely reach the
3937 * exit. States describe preconditions for further exploration, but completely
3938 * forget the history of how we got here.
3940 * This also means that even if we needed precise SCALAR range to get to
3941 * finalized state, but from that point forward *that same* SCALAR register is
3942 * never used in a precise context (i.e., it's precise value is not needed for
3943 * correctness), it's correct and safe to mark such register as "imprecise"
3944 * (i.e., precise marking set to false). This is what we rely on when we do
3945 * not set precise marking in current state. If no child state requires
3946 * precision for any given SCALAR register, it's safe to dictate that it can
3947 * be imprecise. If any child state does require this register to be precise,
3948 * we'll mark it precise later retroactively during precise markings
3949 * propagation from child state to parent states.
3951 * Skipping precise marking setting in current state is a mild version of
3952 * relying on the above observation. But we can utilize this property even
3953 * more aggressively by proactively forgetting any precise marking in the
3954 * current state (which we inherited from the parent state), right before we
3955 * checkpoint it and branch off into new child state. This is done by
3956 * mark_all_scalars_imprecise() to hopefully get more permissive and generic
3957 * finalized states which help in short circuiting more future states.
3959 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno)
3961 struct backtrack_state *bt = &env->bt;
3962 struct bpf_verifier_state *st = env->cur_state;
3963 int first_idx = st->first_insn_idx;
3964 int last_idx = env->insn_idx;
3965 int subseq_idx = -1;
3966 struct bpf_func_state *func;
3967 struct bpf_reg_state *reg;
3968 bool skip_first = true;
3971 if (!env->bpf_capable)
3974 /* set frame number from which we are starting to backtrack */
3975 bt_init(bt, env->cur_state->curframe);
3977 /* Do sanity checks against current state of register and/or stack
3978 * slot, but don't set precise flag in current state, as precision
3979 * tracking in the current state is unnecessary.
3981 func = st->frame[bt->frame];
3983 reg = &func->regs[regno];
3984 if (reg->type != SCALAR_VALUE) {
3985 WARN_ONCE(1, "backtracing misuse");
3988 bt_set_reg(bt, regno);
3995 DECLARE_BITMAP(mask, 64);
3996 u32 history = st->jmp_history_cnt;
3998 if (env->log.level & BPF_LOG_LEVEL2) {
3999 verbose(env, "mark_precise: frame%d: last_idx %d first_idx %d subseq_idx %d \n",
4000 bt->frame, last_idx, first_idx, subseq_idx);
4003 /* If some register with scalar ID is marked as precise,
4004 * make sure that all registers sharing this ID are also precise.
4005 * This is needed to estimate effect of find_equal_scalars().
4006 * Do this at the last instruction of each state,
4007 * bpf_reg_state::id fields are valid for these instructions.
4009 * Allows to track precision in situation like below:
4011 * r2 = unknown value
4015 * r1 = r2 // r1 and r2 now share the same ID
4017 * --- state #1 {r1.id = A, r2.id = A} ---
4019 * if (r2 > 10) goto exit; // find_equal_scalars() assigns range to r1
4021 * --- state #2 {r1.id = A, r2.id = A} ---
4023 * r3 += r1 // need to mark both r1 and r2
4025 if (mark_precise_scalar_ids(env, st))
4029 /* we are at the entry into subprog, which
4030 * is expected for global funcs, but only if
4031 * requested precise registers are R1-R5
4032 * (which are global func's input arguments)
4034 if (st->curframe == 0 &&
4035 st->frame[0]->subprogno > 0 &&
4036 st->frame[0]->callsite == BPF_MAIN_FUNC &&
4037 bt_stack_mask(bt) == 0 &&
4038 (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) == 0) {
4039 bitmap_from_u64(mask, bt_reg_mask(bt));
4040 for_each_set_bit(i, mask, 32) {
4041 reg = &st->frame[0]->regs[i];
4042 if (reg->type != SCALAR_VALUE) {
4043 bt_clear_reg(bt, i);
4046 reg->precise = true;
4051 verbose(env, "BUG backtracking func entry subprog %d reg_mask %x stack_mask %llx\n",
4052 st->frame[0]->subprogno, bt_reg_mask(bt), bt_stack_mask(bt));
4053 WARN_ONCE(1, "verifier backtracking bug");
4057 for (i = last_idx;;) {
4062 err = backtrack_insn(env, i, subseq_idx, bt);
4064 if (err == -ENOTSUPP) {
4065 mark_all_scalars_precise(env, env->cur_state);
4072 /* Found assignment(s) into tracked register in this state.
4073 * Since this state is already marked, just return.
4074 * Nothing to be tracked further in the parent state.
4080 i = get_prev_insn_idx(st, i, &history);
4081 if (i >= env->prog->len) {
4082 /* This can happen if backtracking reached insn 0
4083 * and there are still reg_mask or stack_mask
4085 * It means the backtracking missed the spot where
4086 * particular register was initialized with a constant.
4088 verbose(env, "BUG backtracking idx %d\n", i);
4089 WARN_ONCE(1, "verifier backtracking bug");
4097 for (fr = bt->frame; fr >= 0; fr--) {
4098 func = st->frame[fr];
4099 bitmap_from_u64(mask, bt_frame_reg_mask(bt, fr));
4100 for_each_set_bit(i, mask, 32) {
4101 reg = &func->regs[i];
4102 if (reg->type != SCALAR_VALUE) {
4103 bt_clear_frame_reg(bt, fr, i);
4107 bt_clear_frame_reg(bt, fr, i);
4109 reg->precise = true;
4112 bitmap_from_u64(mask, bt_frame_stack_mask(bt, fr));
4113 for_each_set_bit(i, mask, 64) {
4114 if (i >= func->allocated_stack / BPF_REG_SIZE) {
4115 /* the sequence of instructions:
4117 * 3: (7b) *(u64 *)(r3 -8) = r0
4118 * 4: (79) r4 = *(u64 *)(r10 -8)
4119 * doesn't contain jmps. It's backtracked
4120 * as a single block.
4121 * During backtracking insn 3 is not recognized as
4122 * stack access, so at the end of backtracking
4123 * stack slot fp-8 is still marked in stack_mask.
4124 * However the parent state may not have accessed
4125 * fp-8 and it's "unallocated" stack space.
4126 * In such case fallback to conservative.
4128 mark_all_scalars_precise(env, env->cur_state);
4133 if (!is_spilled_scalar_reg(&func->stack[i])) {
4134 bt_clear_frame_slot(bt, fr, i);
4137 reg = &func->stack[i].spilled_ptr;
4139 bt_clear_frame_slot(bt, fr, i);
4141 reg->precise = true;
4143 if (env->log.level & BPF_LOG_LEVEL2) {
4144 fmt_reg_mask(env->tmp_str_buf, TMP_STR_BUF_LEN,
4145 bt_frame_reg_mask(bt, fr));
4146 verbose(env, "mark_precise: frame%d: parent state regs=%s ",
4147 fr, env->tmp_str_buf);
4148 fmt_stack_mask(env->tmp_str_buf, TMP_STR_BUF_LEN,
4149 bt_frame_stack_mask(bt, fr));
4150 verbose(env, "stack=%s: ", env->tmp_str_buf);
4151 print_verifier_state(env, func, true);
4158 subseq_idx = first_idx;
4159 last_idx = st->last_insn_idx;
4160 first_idx = st->first_insn_idx;
4163 /* if we still have requested precise regs or slots, we missed
4164 * something (e.g., stack access through non-r10 register), so
4165 * fallback to marking all precise
4167 if (!bt_empty(bt)) {
4168 mark_all_scalars_precise(env, env->cur_state);
4175 int mark_chain_precision(struct bpf_verifier_env *env, int regno)
4177 return __mark_chain_precision(env, regno);
4180 /* mark_chain_precision_batch() assumes that env->bt is set in the caller to
4181 * desired reg and stack masks across all relevant frames
4183 static int mark_chain_precision_batch(struct bpf_verifier_env *env)
4185 return __mark_chain_precision(env, -1);
4188 static bool is_spillable_regtype(enum bpf_reg_type type)
4190 switch (base_type(type)) {
4191 case PTR_TO_MAP_VALUE:
4195 case PTR_TO_PACKET_META:
4196 case PTR_TO_PACKET_END:
4197 case PTR_TO_FLOW_KEYS:
4198 case CONST_PTR_TO_MAP:
4200 case PTR_TO_SOCK_COMMON:
4201 case PTR_TO_TCP_SOCK:
4202 case PTR_TO_XDP_SOCK:
4207 case PTR_TO_MAP_KEY:
4214 /* Does this register contain a constant zero? */
4215 static bool register_is_null(struct bpf_reg_state *reg)
4217 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
4220 static bool register_is_const(struct bpf_reg_state *reg)
4222 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
4225 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
4227 return tnum_is_unknown(reg->var_off) &&
4228 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
4229 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
4230 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
4231 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
4234 static bool register_is_bounded(struct bpf_reg_state *reg)
4236 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
4239 static bool __is_pointer_value(bool allow_ptr_leaks,
4240 const struct bpf_reg_state *reg)
4242 if (allow_ptr_leaks)
4245 return reg->type != SCALAR_VALUE;
4248 /* Copy src state preserving dst->parent and dst->live fields */
4249 static void copy_register_state(struct bpf_reg_state *dst, const struct bpf_reg_state *src)
4251 struct bpf_reg_state *parent = dst->parent;
4252 enum bpf_reg_liveness live = dst->live;
4255 dst->parent = parent;
4259 static void save_register_state(struct bpf_func_state *state,
4260 int spi, struct bpf_reg_state *reg,
4265 copy_register_state(&state->stack[spi].spilled_ptr, reg);
4266 if (size == BPF_REG_SIZE)
4267 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
4269 for (i = BPF_REG_SIZE; i > BPF_REG_SIZE - size; i--)
4270 state->stack[spi].slot_type[i - 1] = STACK_SPILL;
4272 /* size < 8 bytes spill */
4274 scrub_spilled_slot(&state->stack[spi].slot_type[i - 1]);
4277 static bool is_bpf_st_mem(struct bpf_insn *insn)
4279 return BPF_CLASS(insn->code) == BPF_ST && BPF_MODE(insn->code) == BPF_MEM;
4282 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
4283 * stack boundary and alignment are checked in check_mem_access()
4285 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
4286 /* stack frame we're writing to */
4287 struct bpf_func_state *state,
4288 int off, int size, int value_regno,
4291 struct bpf_func_state *cur; /* state of the current function */
4292 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
4293 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
4294 struct bpf_reg_state *reg = NULL;
4295 u32 dst_reg = insn->dst_reg;
4297 err = grow_stack_state(state, round_up(slot + 1, BPF_REG_SIZE));
4300 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
4301 * so it's aligned access and [off, off + size) are within stack limits
4303 if (!env->allow_ptr_leaks &&
4304 state->stack[spi].slot_type[0] == STACK_SPILL &&
4305 size != BPF_REG_SIZE) {
4306 verbose(env, "attempt to corrupt spilled pointer on stack\n");
4310 cur = env->cur_state->frame[env->cur_state->curframe];
4311 if (value_regno >= 0)
4312 reg = &cur->regs[value_regno];
4313 if (!env->bypass_spec_v4) {
4314 bool sanitize = reg && is_spillable_regtype(reg->type);
4316 for (i = 0; i < size; i++) {
4317 u8 type = state->stack[spi].slot_type[i];
4319 if (type != STACK_MISC && type != STACK_ZERO) {
4326 env->insn_aux_data[insn_idx].sanitize_stack_spill = true;
4329 err = destroy_if_dynptr_stack_slot(env, state, spi);
4333 mark_stack_slot_scratched(env, spi);
4334 if (reg && !(off % BPF_REG_SIZE) && register_is_bounded(reg) &&
4335 !register_is_null(reg) && env->bpf_capable) {
4336 if (dst_reg != BPF_REG_FP) {
4337 /* The backtracking logic can only recognize explicit
4338 * stack slot address like [fp - 8]. Other spill of
4339 * scalar via different register has to be conservative.
4340 * Backtrack from here and mark all registers as precise
4341 * that contributed into 'reg' being a constant.
4343 err = mark_chain_precision(env, value_regno);
4347 save_register_state(state, spi, reg, size);
4348 /* Break the relation on a narrowing spill. */
4349 if (fls64(reg->umax_value) > BITS_PER_BYTE * size)
4350 state->stack[spi].spilled_ptr.id = 0;
4351 } else if (!reg && !(off % BPF_REG_SIZE) && is_bpf_st_mem(insn) &&
4352 insn->imm != 0 && env->bpf_capable) {
4353 struct bpf_reg_state fake_reg = {};
4355 __mark_reg_known(&fake_reg, (u32)insn->imm);
4356 fake_reg.type = SCALAR_VALUE;
4357 save_register_state(state, spi, &fake_reg, size);
4358 } else if (reg && is_spillable_regtype(reg->type)) {
4359 /* register containing pointer is being spilled into stack */
4360 if (size != BPF_REG_SIZE) {
4361 verbose_linfo(env, insn_idx, "; ");
4362 verbose(env, "invalid size of register spill\n");
4365 if (state != cur && reg->type == PTR_TO_STACK) {
4366 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
4369 save_register_state(state, spi, reg, size);
4371 u8 type = STACK_MISC;
4373 /* regular write of data into stack destroys any spilled ptr */
4374 state->stack[spi].spilled_ptr.type = NOT_INIT;
4375 /* Mark slots as STACK_MISC if they belonged to spilled ptr/dynptr/iter. */
4376 if (is_stack_slot_special(&state->stack[spi]))
4377 for (i = 0; i < BPF_REG_SIZE; i++)
4378 scrub_spilled_slot(&state->stack[spi].slot_type[i]);
4380 /* only mark the slot as written if all 8 bytes were written
4381 * otherwise read propagation may incorrectly stop too soon
4382 * when stack slots are partially written.
4383 * This heuristic means that read propagation will be
4384 * conservative, since it will add reg_live_read marks
4385 * to stack slots all the way to first state when programs
4386 * writes+reads less than 8 bytes
4388 if (size == BPF_REG_SIZE)
4389 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
4391 /* when we zero initialize stack slots mark them as such */
4392 if ((reg && register_is_null(reg)) ||
4393 (!reg && is_bpf_st_mem(insn) && insn->imm == 0)) {
4394 /* backtracking doesn't work for STACK_ZERO yet. */
4395 err = mark_chain_precision(env, value_regno);
4401 /* Mark slots affected by this stack write. */
4402 for (i = 0; i < size; i++)
4403 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
4409 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
4410 * known to contain a variable offset.
4411 * This function checks whether the write is permitted and conservatively
4412 * tracks the effects of the write, considering that each stack slot in the
4413 * dynamic range is potentially written to.
4415 * 'off' includes 'regno->off'.
4416 * 'value_regno' can be -1, meaning that an unknown value is being written to
4419 * Spilled pointers in range are not marked as written because we don't know
4420 * what's going to be actually written. This means that read propagation for
4421 * future reads cannot be terminated by this write.
4423 * For privileged programs, uninitialized stack slots are considered
4424 * initialized by this write (even though we don't know exactly what offsets
4425 * are going to be written to). The idea is that we don't want the verifier to
4426 * reject future reads that access slots written to through variable offsets.
4428 static int check_stack_write_var_off(struct bpf_verifier_env *env,
4429 /* func where register points to */
4430 struct bpf_func_state *state,
4431 int ptr_regno, int off, int size,
4432 int value_regno, int insn_idx)
4434 struct bpf_func_state *cur; /* state of the current function */
4435 int min_off, max_off;
4437 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
4438 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
4439 bool writing_zero = false;
4440 /* set if the fact that we're writing a zero is used to let any
4441 * stack slots remain STACK_ZERO
4443 bool zero_used = false;
4445 cur = env->cur_state->frame[env->cur_state->curframe];
4446 ptr_reg = &cur->regs[ptr_regno];
4447 min_off = ptr_reg->smin_value + off;
4448 max_off = ptr_reg->smax_value + off + size;
4449 if (value_regno >= 0)
4450 value_reg = &cur->regs[value_regno];
4451 if ((value_reg && register_is_null(value_reg)) ||
4452 (!value_reg && is_bpf_st_mem(insn) && insn->imm == 0))
4453 writing_zero = true;
4455 err = grow_stack_state(state, round_up(-min_off, BPF_REG_SIZE));
4459 for (i = min_off; i < max_off; i++) {
4463 err = destroy_if_dynptr_stack_slot(env, state, spi);
4468 /* Variable offset writes destroy any spilled pointers in range. */
4469 for (i = min_off; i < max_off; i++) {
4470 u8 new_type, *stype;
4474 spi = slot / BPF_REG_SIZE;
4475 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
4476 mark_stack_slot_scratched(env, spi);
4478 if (!env->allow_ptr_leaks && *stype != STACK_MISC && *stype != STACK_ZERO) {
4479 /* Reject the write if range we may write to has not
4480 * been initialized beforehand. If we didn't reject
4481 * here, the ptr status would be erased below (even
4482 * though not all slots are actually overwritten),
4483 * possibly opening the door to leaks.
4485 * We do however catch STACK_INVALID case below, and
4486 * only allow reading possibly uninitialized memory
4487 * later for CAP_PERFMON, as the write may not happen to
4490 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
4495 /* Erase all spilled pointers. */
4496 state->stack[spi].spilled_ptr.type = NOT_INIT;
4498 /* Update the slot type. */
4499 new_type = STACK_MISC;
4500 if (writing_zero && *stype == STACK_ZERO) {
4501 new_type = STACK_ZERO;
4504 /* If the slot is STACK_INVALID, we check whether it's OK to
4505 * pretend that it will be initialized by this write. The slot
4506 * might not actually be written to, and so if we mark it as
4507 * initialized future reads might leak uninitialized memory.
4508 * For privileged programs, we will accept such reads to slots
4509 * that may or may not be written because, if we're reject
4510 * them, the error would be too confusing.
4512 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
4513 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
4520 /* backtracking doesn't work for STACK_ZERO yet. */
4521 err = mark_chain_precision(env, value_regno);
4528 /* When register 'dst_regno' is assigned some values from stack[min_off,
4529 * max_off), we set the register's type according to the types of the
4530 * respective stack slots. If all the stack values are known to be zeros, then
4531 * so is the destination reg. Otherwise, the register is considered to be
4532 * SCALAR. This function does not deal with register filling; the caller must
4533 * ensure that all spilled registers in the stack range have been marked as
4536 static void mark_reg_stack_read(struct bpf_verifier_env *env,
4537 /* func where src register points to */
4538 struct bpf_func_state *ptr_state,
4539 int min_off, int max_off, int dst_regno)
4541 struct bpf_verifier_state *vstate = env->cur_state;
4542 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4547 for (i = min_off; i < max_off; i++) {
4549 spi = slot / BPF_REG_SIZE;
4550 mark_stack_slot_scratched(env, spi);
4551 stype = ptr_state->stack[spi].slot_type;
4552 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
4556 if (zeros == max_off - min_off) {
4557 /* any access_size read into register is zero extended,
4558 * so the whole register == const_zero
4560 __mark_reg_const_zero(&state->regs[dst_regno]);
4561 /* backtracking doesn't support STACK_ZERO yet,
4562 * so mark it precise here, so that later
4563 * backtracking can stop here.
4564 * Backtracking may not need this if this register
4565 * doesn't participate in pointer adjustment.
4566 * Forward propagation of precise flag is not
4567 * necessary either. This mark is only to stop
4568 * backtracking. Any register that contributed
4569 * to const 0 was marked precise before spill.
4571 state->regs[dst_regno].precise = true;
4573 /* have read misc data from the stack */
4574 mark_reg_unknown(env, state->regs, dst_regno);
4576 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4579 /* Read the stack at 'off' and put the results into the register indicated by
4580 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
4583 * 'dst_regno' can be -1, meaning that the read value is not going to a
4586 * The access is assumed to be within the current stack bounds.
4588 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
4589 /* func where src register points to */
4590 struct bpf_func_state *reg_state,
4591 int off, int size, int dst_regno)
4593 struct bpf_verifier_state *vstate = env->cur_state;
4594 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4595 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
4596 struct bpf_reg_state *reg;
4599 stype = reg_state->stack[spi].slot_type;
4600 reg = ®_state->stack[spi].spilled_ptr;
4602 mark_stack_slot_scratched(env, spi);
4604 if (is_spilled_reg(®_state->stack[spi])) {
4607 for (i = BPF_REG_SIZE - 1; i > 0 && stype[i - 1] == STACK_SPILL; i--)
4610 if (size != BPF_REG_SIZE || spill_size != BPF_REG_SIZE) {
4611 if (reg->type != SCALAR_VALUE) {
4612 verbose_linfo(env, env->insn_idx, "; ");
4613 verbose(env, "invalid size of register fill\n");
4617 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4621 if (!(off % BPF_REG_SIZE) && size == spill_size) {
4622 /* The earlier check_reg_arg() has decided the
4623 * subreg_def for this insn. Save it first.
4625 s32 subreg_def = state->regs[dst_regno].subreg_def;
4627 copy_register_state(&state->regs[dst_regno], reg);
4628 state->regs[dst_regno].subreg_def = subreg_def;
4630 for (i = 0; i < size; i++) {
4631 type = stype[(slot - i) % BPF_REG_SIZE];
4632 if (type == STACK_SPILL)
4634 if (type == STACK_MISC)
4636 if (type == STACK_INVALID && env->allow_uninit_stack)
4638 verbose(env, "invalid read from stack off %d+%d size %d\n",
4642 mark_reg_unknown(env, state->regs, dst_regno);
4644 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4648 if (dst_regno >= 0) {
4649 /* restore register state from stack */
4650 copy_register_state(&state->regs[dst_regno], reg);
4651 /* mark reg as written since spilled pointer state likely
4652 * has its liveness marks cleared by is_state_visited()
4653 * which resets stack/reg liveness for state transitions
4655 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4656 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
4657 /* If dst_regno==-1, the caller is asking us whether
4658 * it is acceptable to use this value as a SCALAR_VALUE
4660 * We must not allow unprivileged callers to do that
4661 * with spilled pointers.
4663 verbose(env, "leaking pointer from stack off %d\n",
4667 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4669 for (i = 0; i < size; i++) {
4670 type = stype[(slot - i) % BPF_REG_SIZE];
4671 if (type == STACK_MISC)
4673 if (type == STACK_ZERO)
4675 if (type == STACK_INVALID && env->allow_uninit_stack)
4677 verbose(env, "invalid read from stack off %d+%d size %d\n",
4681 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4683 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
4688 enum bpf_access_src {
4689 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
4690 ACCESS_HELPER = 2, /* the access is performed by a helper */
4693 static int check_stack_range_initialized(struct bpf_verifier_env *env,
4694 int regno, int off, int access_size,
4695 bool zero_size_allowed,
4696 enum bpf_access_src type,
4697 struct bpf_call_arg_meta *meta);
4699 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
4701 return cur_regs(env) + regno;
4704 /* Read the stack at 'ptr_regno + off' and put the result into the register
4706 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
4707 * but not its variable offset.
4708 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
4710 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
4711 * filling registers (i.e. reads of spilled register cannot be detected when
4712 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
4713 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
4714 * offset; for a fixed offset check_stack_read_fixed_off should be used
4717 static int check_stack_read_var_off(struct bpf_verifier_env *env,
4718 int ptr_regno, int off, int size, int dst_regno)
4720 /* The state of the source register. */
4721 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
4722 struct bpf_func_state *ptr_state = func(env, reg);
4724 int min_off, max_off;
4726 /* Note that we pass a NULL meta, so raw access will not be permitted.
4728 err = check_stack_range_initialized(env, ptr_regno, off, size,
4729 false, ACCESS_DIRECT, NULL);
4733 min_off = reg->smin_value + off;
4734 max_off = reg->smax_value + off;
4735 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
4739 /* check_stack_read dispatches to check_stack_read_fixed_off or
4740 * check_stack_read_var_off.
4742 * The caller must ensure that the offset falls within the allocated stack
4745 * 'dst_regno' is a register which will receive the value from the stack. It
4746 * can be -1, meaning that the read value is not going to a register.
4748 static int check_stack_read(struct bpf_verifier_env *env,
4749 int ptr_regno, int off, int size,
4752 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
4753 struct bpf_func_state *state = func(env, reg);
4755 /* Some accesses are only permitted with a static offset. */
4756 bool var_off = !tnum_is_const(reg->var_off);
4758 /* The offset is required to be static when reads don't go to a
4759 * register, in order to not leak pointers (see
4760 * check_stack_read_fixed_off).
4762 if (dst_regno < 0 && var_off) {
4765 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4766 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
4770 /* Variable offset is prohibited for unprivileged mode for simplicity
4771 * since it requires corresponding support in Spectre masking for stack
4772 * ALU. See also retrieve_ptr_limit(). The check in
4773 * check_stack_access_for_ptr_arithmetic() called by
4774 * adjust_ptr_min_max_vals() prevents users from creating stack pointers
4775 * with variable offsets, therefore no check is required here. Further,
4776 * just checking it here would be insufficient as speculative stack
4777 * writes could still lead to unsafe speculative behaviour.
4780 off += reg->var_off.value;
4781 err = check_stack_read_fixed_off(env, state, off, size,
4784 /* Variable offset stack reads need more conservative handling
4785 * than fixed offset ones. Note that dst_regno >= 0 on this
4788 err = check_stack_read_var_off(env, ptr_regno, off, size,
4795 /* check_stack_write dispatches to check_stack_write_fixed_off or
4796 * check_stack_write_var_off.
4798 * 'ptr_regno' is the register used as a pointer into the stack.
4799 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
4800 * 'value_regno' is the register whose value we're writing to the stack. It can
4801 * be -1, meaning that we're not writing from a register.
4803 * The caller must ensure that the offset falls within the maximum stack size.
4805 static int check_stack_write(struct bpf_verifier_env *env,
4806 int ptr_regno, int off, int size,
4807 int value_regno, int insn_idx)
4809 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
4810 struct bpf_func_state *state = func(env, reg);
4813 if (tnum_is_const(reg->var_off)) {
4814 off += reg->var_off.value;
4815 err = check_stack_write_fixed_off(env, state, off, size,
4816 value_regno, insn_idx);
4818 /* Variable offset stack reads need more conservative handling
4819 * than fixed offset ones.
4821 err = check_stack_write_var_off(env, state,
4822 ptr_regno, off, size,
4823 value_regno, insn_idx);
4828 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
4829 int off, int size, enum bpf_access_type type)
4831 struct bpf_reg_state *regs = cur_regs(env);
4832 struct bpf_map *map = regs[regno].map_ptr;
4833 u32 cap = bpf_map_flags_to_cap(map);
4835 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
4836 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
4837 map->value_size, off, size);
4841 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
4842 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
4843 map->value_size, off, size);
4850 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
4851 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
4852 int off, int size, u32 mem_size,
4853 bool zero_size_allowed)
4855 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
4856 struct bpf_reg_state *reg;
4858 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
4861 reg = &cur_regs(env)[regno];
4862 switch (reg->type) {
4863 case PTR_TO_MAP_KEY:
4864 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
4865 mem_size, off, size);
4867 case PTR_TO_MAP_VALUE:
4868 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
4869 mem_size, off, size);
4872 case PTR_TO_PACKET_META:
4873 case PTR_TO_PACKET_END:
4874 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
4875 off, size, regno, reg->id, off, mem_size);
4879 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
4880 mem_size, off, size);
4886 /* check read/write into a memory region with possible variable offset */
4887 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
4888 int off, int size, u32 mem_size,
4889 bool zero_size_allowed)
4891 struct bpf_verifier_state *vstate = env->cur_state;
4892 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4893 struct bpf_reg_state *reg = &state->regs[regno];
4896 /* We may have adjusted the register pointing to memory region, so we
4897 * need to try adding each of min_value and max_value to off
4898 * to make sure our theoretical access will be safe.
4900 * The minimum value is only important with signed
4901 * comparisons where we can't assume the floor of a
4902 * value is 0. If we are using signed variables for our
4903 * index'es we need to make sure that whatever we use
4904 * will have a set floor within our range.
4906 if (reg->smin_value < 0 &&
4907 (reg->smin_value == S64_MIN ||
4908 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
4909 reg->smin_value + off < 0)) {
4910 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
4914 err = __check_mem_access(env, regno, reg->smin_value + off, size,
4915 mem_size, zero_size_allowed);
4917 verbose(env, "R%d min value is outside of the allowed memory range\n",
4922 /* If we haven't set a max value then we need to bail since we can't be
4923 * sure we won't do bad things.
4924 * If reg->umax_value + off could overflow, treat that as unbounded too.
4926 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
4927 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
4931 err = __check_mem_access(env, regno, reg->umax_value + off, size,
4932 mem_size, zero_size_allowed);
4934 verbose(env, "R%d max value is outside of the allowed memory range\n",
4942 static int __check_ptr_off_reg(struct bpf_verifier_env *env,
4943 const struct bpf_reg_state *reg, int regno,
4946 /* Access to this pointer-typed register or passing it to a helper
4947 * is only allowed in its original, unmodified form.
4951 verbose(env, "negative offset %s ptr R%d off=%d disallowed\n",
4952 reg_type_str(env, reg->type), regno, reg->off);
4956 if (!fixed_off_ok && reg->off) {
4957 verbose(env, "dereference of modified %s ptr R%d off=%d disallowed\n",
4958 reg_type_str(env, reg->type), regno, reg->off);
4962 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4965 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4966 verbose(env, "variable %s access var_off=%s disallowed\n",
4967 reg_type_str(env, reg->type), tn_buf);
4974 int check_ptr_off_reg(struct bpf_verifier_env *env,
4975 const struct bpf_reg_state *reg, int regno)
4977 return __check_ptr_off_reg(env, reg, regno, false);
4980 static int map_kptr_match_type(struct bpf_verifier_env *env,
4981 struct btf_field *kptr_field,
4982 struct bpf_reg_state *reg, u32 regno)
4984 const char *targ_name = btf_type_name(kptr_field->kptr.btf, kptr_field->kptr.btf_id);
4985 int perm_flags = PTR_MAYBE_NULL | PTR_TRUSTED | MEM_RCU;
4986 const char *reg_name = "";
4988 /* Only unreferenced case accepts untrusted pointers */
4989 if (kptr_field->type == BPF_KPTR_UNREF)
4990 perm_flags |= PTR_UNTRUSTED;
4992 if (base_type(reg->type) != PTR_TO_BTF_ID || (type_flag(reg->type) & ~perm_flags))
4995 if (!btf_is_kernel(reg->btf)) {
4996 verbose(env, "R%d must point to kernel BTF\n", regno);
4999 /* We need to verify reg->type and reg->btf, before accessing reg->btf */
5000 reg_name = btf_type_name(reg->btf, reg->btf_id);
5002 /* For ref_ptr case, release function check should ensure we get one
5003 * referenced PTR_TO_BTF_ID, and that its fixed offset is 0. For the
5004 * normal store of unreferenced kptr, we must ensure var_off is zero.
5005 * Since ref_ptr cannot be accessed directly by BPF insns, checks for
5006 * reg->off and reg->ref_obj_id are not needed here.
5008 if (__check_ptr_off_reg(env, reg, regno, true))
5011 /* A full type match is needed, as BTF can be vmlinux or module BTF, and
5012 * we also need to take into account the reg->off.
5014 * We want to support cases like:
5022 * v = func(); // PTR_TO_BTF_ID
5023 * val->foo = v; // reg->off is zero, btf and btf_id match type
5024 * val->bar = &v->br; // reg->off is still zero, but we need to retry with
5025 * // first member type of struct after comparison fails
5026 * val->baz = &v->bz; // reg->off is non-zero, so struct needs to be walked
5029 * In the kptr_ref case, check_func_arg_reg_off already ensures reg->off
5030 * is zero. We must also ensure that btf_struct_ids_match does not walk
5031 * the struct to match type against first member of struct, i.e. reject
5032 * second case from above. Hence, when type is BPF_KPTR_REF, we set
5033 * strict mode to true for type match.
5035 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
5036 kptr_field->kptr.btf, kptr_field->kptr.btf_id,
5037 kptr_field->type == BPF_KPTR_REF))
5041 verbose(env, "invalid kptr access, R%d type=%s%s ", regno,
5042 reg_type_str(env, reg->type), reg_name);
5043 verbose(env, "expected=%s%s", reg_type_str(env, PTR_TO_BTF_ID), targ_name);
5044 if (kptr_field->type == BPF_KPTR_UNREF)
5045 verbose(env, " or %s%s\n", reg_type_str(env, PTR_TO_BTF_ID | PTR_UNTRUSTED),
5052 /* The non-sleepable programs and sleepable programs with explicit bpf_rcu_read_lock()
5053 * can dereference RCU protected pointers and result is PTR_TRUSTED.
5055 static bool in_rcu_cs(struct bpf_verifier_env *env)
5057 return env->cur_state->active_rcu_lock || !env->prog->aux->sleepable;
5060 /* Once GCC supports btf_type_tag the following mechanism will be replaced with tag check */
5061 BTF_SET_START(rcu_protected_types)
5062 BTF_ID(struct, prog_test_ref_kfunc)
5063 BTF_ID(struct, cgroup)
5064 BTF_ID(struct, bpf_cpumask)
5065 BTF_ID(struct, task_struct)
5066 BTF_SET_END(rcu_protected_types)
5068 static bool rcu_protected_object(const struct btf *btf, u32 btf_id)
5070 if (!btf_is_kernel(btf))
5072 return btf_id_set_contains(&rcu_protected_types, btf_id);
5075 static bool rcu_safe_kptr(const struct btf_field *field)
5077 const struct btf_field_kptr *kptr = &field->kptr;
5079 return field->type == BPF_KPTR_REF && rcu_protected_object(kptr->btf, kptr->btf_id);
5082 static int check_map_kptr_access(struct bpf_verifier_env *env, u32 regno,
5083 int value_regno, int insn_idx,
5084 struct btf_field *kptr_field)
5086 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
5087 int class = BPF_CLASS(insn->code);
5088 struct bpf_reg_state *val_reg;
5090 /* Things we already checked for in check_map_access and caller:
5091 * - Reject cases where variable offset may touch kptr
5092 * - size of access (must be BPF_DW)
5093 * - tnum_is_const(reg->var_off)
5094 * - kptr_field->offset == off + reg->var_off.value
5096 /* Only BPF_[LDX,STX,ST] | BPF_MEM | BPF_DW is supported */
5097 if (BPF_MODE(insn->code) != BPF_MEM) {
5098 verbose(env, "kptr in map can only be accessed using BPF_MEM instruction mode\n");
5102 /* We only allow loading referenced kptr, since it will be marked as
5103 * untrusted, similar to unreferenced kptr.
5105 if (class != BPF_LDX && kptr_field->type == BPF_KPTR_REF) {
5106 verbose(env, "store to referenced kptr disallowed\n");
5110 if (class == BPF_LDX) {
5111 val_reg = reg_state(env, value_regno);
5112 /* We can simply mark the value_regno receiving the pointer
5113 * value from map as PTR_TO_BTF_ID, with the correct type.
5115 mark_btf_ld_reg(env, cur_regs(env), value_regno, PTR_TO_BTF_ID, kptr_field->kptr.btf,
5116 kptr_field->kptr.btf_id,
5117 rcu_safe_kptr(kptr_field) && in_rcu_cs(env) ?
5118 PTR_MAYBE_NULL | MEM_RCU :
5119 PTR_MAYBE_NULL | PTR_UNTRUSTED);
5120 /* For mark_ptr_or_null_reg */
5121 val_reg->id = ++env->id_gen;
5122 } else if (class == BPF_STX) {
5123 val_reg = reg_state(env, value_regno);
5124 if (!register_is_null(val_reg) &&
5125 map_kptr_match_type(env, kptr_field, val_reg, value_regno))
5127 } else if (class == BPF_ST) {
5129 verbose(env, "BPF_ST imm must be 0 when storing to kptr at off=%u\n",
5130 kptr_field->offset);
5134 verbose(env, "kptr in map can only be accessed using BPF_LDX/BPF_STX/BPF_ST\n");
5140 /* check read/write into a map element with possible variable offset */
5141 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
5142 int off, int size, bool zero_size_allowed,
5143 enum bpf_access_src src)
5145 struct bpf_verifier_state *vstate = env->cur_state;
5146 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5147 struct bpf_reg_state *reg = &state->regs[regno];
5148 struct bpf_map *map = reg->map_ptr;
5149 struct btf_record *rec;
5152 err = check_mem_region_access(env, regno, off, size, map->value_size,
5157 if (IS_ERR_OR_NULL(map->record))
5160 for (i = 0; i < rec->cnt; i++) {
5161 struct btf_field *field = &rec->fields[i];
5162 u32 p = field->offset;
5164 /* If any part of a field can be touched by load/store, reject
5165 * this program. To check that [x1, x2) overlaps with [y1, y2),
5166 * it is sufficient to check x1 < y2 && y1 < x2.
5168 if (reg->smin_value + off < p + btf_field_type_size(field->type) &&
5169 p < reg->umax_value + off + size) {
5170 switch (field->type) {
5171 case BPF_KPTR_UNREF:
5173 if (src != ACCESS_DIRECT) {
5174 verbose(env, "kptr cannot be accessed indirectly by helper\n");
5177 if (!tnum_is_const(reg->var_off)) {
5178 verbose(env, "kptr access cannot have variable offset\n");
5181 if (p != off + reg->var_off.value) {
5182 verbose(env, "kptr access misaligned expected=%u off=%llu\n",
5183 p, off + reg->var_off.value);
5186 if (size != bpf_size_to_bytes(BPF_DW)) {
5187 verbose(env, "kptr access size must be BPF_DW\n");
5192 verbose(env, "%s cannot be accessed directly by load/store\n",
5193 btf_field_type_name(field->type));
5201 #define MAX_PACKET_OFF 0xffff
5203 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
5204 const struct bpf_call_arg_meta *meta,
5205 enum bpf_access_type t)
5207 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
5209 switch (prog_type) {
5210 /* Program types only with direct read access go here! */
5211 case BPF_PROG_TYPE_LWT_IN:
5212 case BPF_PROG_TYPE_LWT_OUT:
5213 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
5214 case BPF_PROG_TYPE_SK_REUSEPORT:
5215 case BPF_PROG_TYPE_FLOW_DISSECTOR:
5216 case BPF_PROG_TYPE_CGROUP_SKB:
5221 /* Program types with direct read + write access go here! */
5222 case BPF_PROG_TYPE_SCHED_CLS:
5223 case BPF_PROG_TYPE_SCHED_ACT:
5224 case BPF_PROG_TYPE_XDP:
5225 case BPF_PROG_TYPE_LWT_XMIT:
5226 case BPF_PROG_TYPE_SK_SKB:
5227 case BPF_PROG_TYPE_SK_MSG:
5229 return meta->pkt_access;
5231 env->seen_direct_write = true;
5234 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
5236 env->seen_direct_write = true;
5245 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
5246 int size, bool zero_size_allowed)
5248 struct bpf_reg_state *regs = cur_regs(env);
5249 struct bpf_reg_state *reg = ®s[regno];
5252 /* We may have added a variable offset to the packet pointer; but any
5253 * reg->range we have comes after that. We are only checking the fixed
5257 /* We don't allow negative numbers, because we aren't tracking enough
5258 * detail to prove they're safe.
5260 if (reg->smin_value < 0) {
5261 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
5266 err = reg->range < 0 ? -EINVAL :
5267 __check_mem_access(env, regno, off, size, reg->range,
5270 verbose(env, "R%d offset is outside of the packet\n", regno);
5274 /* __check_mem_access has made sure "off + size - 1" is within u16.
5275 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
5276 * otherwise find_good_pkt_pointers would have refused to set range info
5277 * that __check_mem_access would have rejected this pkt access.
5278 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
5280 env->prog->aux->max_pkt_offset =
5281 max_t(u32, env->prog->aux->max_pkt_offset,
5282 off + reg->umax_value + size - 1);
5287 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
5288 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
5289 enum bpf_access_type t, enum bpf_reg_type *reg_type,
5290 struct btf **btf, u32 *btf_id)
5292 struct bpf_insn_access_aux info = {
5293 .reg_type = *reg_type,
5297 if (env->ops->is_valid_access &&
5298 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
5299 /* A non zero info.ctx_field_size indicates that this field is a
5300 * candidate for later verifier transformation to load the whole
5301 * field and then apply a mask when accessed with a narrower
5302 * access than actual ctx access size. A zero info.ctx_field_size
5303 * will only allow for whole field access and rejects any other
5304 * type of narrower access.
5306 *reg_type = info.reg_type;
5308 if (base_type(*reg_type) == PTR_TO_BTF_ID) {
5310 *btf_id = info.btf_id;
5312 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
5314 /* remember the offset of last byte accessed in ctx */
5315 if (env->prog->aux->max_ctx_offset < off + size)
5316 env->prog->aux->max_ctx_offset = off + size;
5320 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
5324 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
5327 if (size < 0 || off < 0 ||
5328 (u64)off + size > sizeof(struct bpf_flow_keys)) {
5329 verbose(env, "invalid access to flow keys off=%d size=%d\n",
5336 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
5337 u32 regno, int off, int size,
5338 enum bpf_access_type t)
5340 struct bpf_reg_state *regs = cur_regs(env);
5341 struct bpf_reg_state *reg = ®s[regno];
5342 struct bpf_insn_access_aux info = {};
5345 if (reg->smin_value < 0) {
5346 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
5351 switch (reg->type) {
5352 case PTR_TO_SOCK_COMMON:
5353 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
5356 valid = bpf_sock_is_valid_access(off, size, t, &info);
5358 case PTR_TO_TCP_SOCK:
5359 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
5361 case PTR_TO_XDP_SOCK:
5362 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
5370 env->insn_aux_data[insn_idx].ctx_field_size =
5371 info.ctx_field_size;
5375 verbose(env, "R%d invalid %s access off=%d size=%d\n",
5376 regno, reg_type_str(env, reg->type), off, size);
5381 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
5383 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
5386 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
5388 const struct bpf_reg_state *reg = reg_state(env, regno);
5390 return reg->type == PTR_TO_CTX;
5393 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
5395 const struct bpf_reg_state *reg = reg_state(env, regno);
5397 return type_is_sk_pointer(reg->type);
5400 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
5402 const struct bpf_reg_state *reg = reg_state(env, regno);
5404 return type_is_pkt_pointer(reg->type);
5407 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
5409 const struct bpf_reg_state *reg = reg_state(env, regno);
5411 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
5412 return reg->type == PTR_TO_FLOW_KEYS;
5415 static bool is_trusted_reg(const struct bpf_reg_state *reg)
5417 /* A referenced register is always trusted. */
5418 if (reg->ref_obj_id)
5421 /* If a register is not referenced, it is trusted if it has the
5422 * MEM_ALLOC or PTR_TRUSTED type modifiers, and no others. Some of the
5423 * other type modifiers may be safe, but we elect to take an opt-in
5424 * approach here as some (e.g. PTR_UNTRUSTED and PTR_MAYBE_NULL) are
5427 * Eventually, we should make PTR_TRUSTED the single source of truth
5428 * for whether a register is trusted.
5430 return type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS &&
5431 !bpf_type_has_unsafe_modifiers(reg->type);
5434 static bool is_rcu_reg(const struct bpf_reg_state *reg)
5436 return reg->type & MEM_RCU;
5439 static void clear_trusted_flags(enum bpf_type_flag *flag)
5441 *flag &= ~(BPF_REG_TRUSTED_MODIFIERS | MEM_RCU);
5444 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
5445 const struct bpf_reg_state *reg,
5446 int off, int size, bool strict)
5448 struct tnum reg_off;
5451 /* Byte size accesses are always allowed. */
5452 if (!strict || size == 1)
5455 /* For platforms that do not have a Kconfig enabling
5456 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
5457 * NET_IP_ALIGN is universally set to '2'. And on platforms
5458 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
5459 * to this code only in strict mode where we want to emulate
5460 * the NET_IP_ALIGN==2 checking. Therefore use an
5461 * unconditional IP align value of '2'.
5465 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
5466 if (!tnum_is_aligned(reg_off, size)) {
5469 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5471 "misaligned packet access off %d+%s+%d+%d size %d\n",
5472 ip_align, tn_buf, reg->off, off, size);
5479 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
5480 const struct bpf_reg_state *reg,
5481 const char *pointer_desc,
5482 int off, int size, bool strict)
5484 struct tnum reg_off;
5486 /* Byte size accesses are always allowed. */
5487 if (!strict || size == 1)
5490 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
5491 if (!tnum_is_aligned(reg_off, size)) {
5494 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5495 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
5496 pointer_desc, tn_buf, reg->off, off, size);
5503 static int check_ptr_alignment(struct bpf_verifier_env *env,
5504 const struct bpf_reg_state *reg, int off,
5505 int size, bool strict_alignment_once)
5507 bool strict = env->strict_alignment || strict_alignment_once;
5508 const char *pointer_desc = "";
5510 switch (reg->type) {
5512 case PTR_TO_PACKET_META:
5513 /* Special case, because of NET_IP_ALIGN. Given metadata sits
5514 * right in front, treat it the very same way.
5516 return check_pkt_ptr_alignment(env, reg, off, size, strict);
5517 case PTR_TO_FLOW_KEYS:
5518 pointer_desc = "flow keys ";
5520 case PTR_TO_MAP_KEY:
5521 pointer_desc = "key ";
5523 case PTR_TO_MAP_VALUE:
5524 pointer_desc = "value ";
5527 pointer_desc = "context ";
5530 pointer_desc = "stack ";
5531 /* The stack spill tracking logic in check_stack_write_fixed_off()
5532 * and check_stack_read_fixed_off() relies on stack accesses being
5538 pointer_desc = "sock ";
5540 case PTR_TO_SOCK_COMMON:
5541 pointer_desc = "sock_common ";
5543 case PTR_TO_TCP_SOCK:
5544 pointer_desc = "tcp_sock ";
5546 case PTR_TO_XDP_SOCK:
5547 pointer_desc = "xdp_sock ";
5552 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
5556 static int update_stack_depth(struct bpf_verifier_env *env,
5557 const struct bpf_func_state *func,
5560 u16 stack = env->subprog_info[func->subprogno].stack_depth;
5565 /* update known max for given subprogram */
5566 env->subprog_info[func->subprogno].stack_depth = -off;
5570 /* starting from main bpf function walk all instructions of the function
5571 * and recursively walk all callees that given function can call.
5572 * Ignore jump and exit insns.
5573 * Since recursion is prevented by check_cfg() this algorithm
5574 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
5576 static int check_max_stack_depth(struct bpf_verifier_env *env)
5578 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
5579 struct bpf_subprog_info *subprog = env->subprog_info;
5580 struct bpf_insn *insn = env->prog->insnsi;
5581 bool tail_call_reachable = false;
5582 int ret_insn[MAX_CALL_FRAMES];
5583 int ret_prog[MAX_CALL_FRAMES];
5587 /* protect against potential stack overflow that might happen when
5588 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
5589 * depth for such case down to 256 so that the worst case scenario
5590 * would result in 8k stack size (32 which is tailcall limit * 256 =
5593 * To get the idea what might happen, see an example:
5594 * func1 -> sub rsp, 128
5595 * subfunc1 -> sub rsp, 256
5596 * tailcall1 -> add rsp, 256
5597 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
5598 * subfunc2 -> sub rsp, 64
5599 * subfunc22 -> sub rsp, 128
5600 * tailcall2 -> add rsp, 128
5601 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
5603 * tailcall will unwind the current stack frame but it will not get rid
5604 * of caller's stack as shown on the example above.
5606 if (idx && subprog[idx].has_tail_call && depth >= 256) {
5608 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
5612 /* round up to 32-bytes, since this is granularity
5613 * of interpreter stack size
5615 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
5616 if (depth > MAX_BPF_STACK) {
5617 verbose(env, "combined stack size of %d calls is %d. Too large\n",
5622 subprog_end = subprog[idx + 1].start;
5623 for (; i < subprog_end; i++) {
5626 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
5628 /* remember insn and function to return to */
5629 ret_insn[frame] = i + 1;
5630 ret_prog[frame] = idx;
5632 /* find the callee */
5633 next_insn = i + insn[i].imm + 1;
5634 idx = find_subprog(env, next_insn);
5636 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5640 if (subprog[idx].is_async_cb) {
5641 if (subprog[idx].has_tail_call) {
5642 verbose(env, "verifier bug. subprog has tail_call and async cb\n");
5645 /* async callbacks don't increase bpf prog stack size */
5650 if (subprog[idx].has_tail_call)
5651 tail_call_reachable = true;
5654 if (frame >= MAX_CALL_FRAMES) {
5655 verbose(env, "the call stack of %d frames is too deep !\n",
5661 /* if tail call got detected across bpf2bpf calls then mark each of the
5662 * currently present subprog frames as tail call reachable subprogs;
5663 * this info will be utilized by JIT so that we will be preserving the
5664 * tail call counter throughout bpf2bpf calls combined with tailcalls
5666 if (tail_call_reachable)
5667 for (j = 0; j < frame; j++)
5668 subprog[ret_prog[j]].tail_call_reachable = true;
5669 if (subprog[0].tail_call_reachable)
5670 env->prog->aux->tail_call_reachable = true;
5672 /* end of for() loop means the last insn of the 'subprog'
5673 * was reached. Doesn't matter whether it was JA or EXIT
5677 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
5679 i = ret_insn[frame];
5680 idx = ret_prog[frame];
5684 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
5685 static int get_callee_stack_depth(struct bpf_verifier_env *env,
5686 const struct bpf_insn *insn, int idx)
5688 int start = idx + insn->imm + 1, subprog;
5690 subprog = find_subprog(env, start);
5692 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5696 return env->subprog_info[subprog].stack_depth;
5700 static int __check_buffer_access(struct bpf_verifier_env *env,
5701 const char *buf_info,
5702 const struct bpf_reg_state *reg,
5703 int regno, int off, int size)
5707 "R%d invalid %s buffer access: off=%d, size=%d\n",
5708 regno, buf_info, off, size);
5711 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
5714 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5716 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
5717 regno, off, tn_buf);
5724 static int check_tp_buffer_access(struct bpf_verifier_env *env,
5725 const struct bpf_reg_state *reg,
5726 int regno, int off, int size)
5730 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
5734 if (off + size > env->prog->aux->max_tp_access)
5735 env->prog->aux->max_tp_access = off + size;
5740 static int check_buffer_access(struct bpf_verifier_env *env,
5741 const struct bpf_reg_state *reg,
5742 int regno, int off, int size,
5743 bool zero_size_allowed,
5746 const char *buf_info = type_is_rdonly_mem(reg->type) ? "rdonly" : "rdwr";
5749 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
5753 if (off + size > *max_access)
5754 *max_access = off + size;
5759 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
5760 static void zext_32_to_64(struct bpf_reg_state *reg)
5762 reg->var_off = tnum_subreg(reg->var_off);
5763 __reg_assign_32_into_64(reg);
5766 /* truncate register to smaller size (in bytes)
5767 * must be called with size < BPF_REG_SIZE
5769 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
5773 /* clear high bits in bit representation */
5774 reg->var_off = tnum_cast(reg->var_off, size);
5776 /* fix arithmetic bounds */
5777 mask = ((u64)1 << (size * 8)) - 1;
5778 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
5779 reg->umin_value &= mask;
5780 reg->umax_value &= mask;
5782 reg->umin_value = 0;
5783 reg->umax_value = mask;
5785 reg->smin_value = reg->umin_value;
5786 reg->smax_value = reg->umax_value;
5788 /* If size is smaller than 32bit register the 32bit register
5789 * values are also truncated so we push 64-bit bounds into
5790 * 32-bit bounds. Above were truncated < 32-bits already.
5794 __reg_combine_64_into_32(reg);
5797 static bool bpf_map_is_rdonly(const struct bpf_map *map)
5799 /* A map is considered read-only if the following condition are true:
5801 * 1) BPF program side cannot change any of the map content. The
5802 * BPF_F_RDONLY_PROG flag is throughout the lifetime of a map
5803 * and was set at map creation time.
5804 * 2) The map value(s) have been initialized from user space by a
5805 * loader and then "frozen", such that no new map update/delete
5806 * operations from syscall side are possible for the rest of
5807 * the map's lifetime from that point onwards.
5808 * 3) Any parallel/pending map update/delete operations from syscall
5809 * side have been completed. Only after that point, it's safe to
5810 * assume that map value(s) are immutable.
5812 return (map->map_flags & BPF_F_RDONLY_PROG) &&
5813 READ_ONCE(map->frozen) &&
5814 !bpf_map_write_active(map);
5817 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
5823 err = map->ops->map_direct_value_addr(map, &addr, off);
5826 ptr = (void *)(long)addr + off;
5830 *val = (u64)*(u8 *)ptr;
5833 *val = (u64)*(u16 *)ptr;
5836 *val = (u64)*(u32 *)ptr;
5847 #define BTF_TYPE_SAFE_RCU(__type) __PASTE(__type, __safe_rcu)
5848 #define BTF_TYPE_SAFE_RCU_OR_NULL(__type) __PASTE(__type, __safe_rcu_or_null)
5849 #define BTF_TYPE_SAFE_TRUSTED(__type) __PASTE(__type, __safe_trusted)
5852 * Allow list few fields as RCU trusted or full trusted.
5853 * This logic doesn't allow mix tagging and will be removed once GCC supports
5857 /* RCU trusted: these fields are trusted in RCU CS and never NULL */
5858 BTF_TYPE_SAFE_RCU(struct task_struct) {
5859 const cpumask_t *cpus_ptr;
5860 struct css_set __rcu *cgroups;
5861 struct task_struct __rcu *real_parent;
5862 struct task_struct *group_leader;
5865 BTF_TYPE_SAFE_RCU(struct cgroup) {
5866 /* cgrp->kn is always accessible as documented in kernel/cgroup/cgroup.c */
5867 struct kernfs_node *kn;
5870 BTF_TYPE_SAFE_RCU(struct css_set) {
5871 struct cgroup *dfl_cgrp;
5874 /* RCU trusted: these fields are trusted in RCU CS and can be NULL */
5875 BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct) {
5876 struct file __rcu *exe_file;
5879 /* skb->sk, req->sk are not RCU protected, but we mark them as such
5880 * because bpf prog accessible sockets are SOCK_RCU_FREE.
5882 BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff) {
5886 BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock) {
5890 /* full trusted: these fields are trusted even outside of RCU CS and never NULL */
5891 BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta) {
5892 struct seq_file *seq;
5895 BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task) {
5896 struct bpf_iter_meta *meta;
5897 struct task_struct *task;
5900 BTF_TYPE_SAFE_TRUSTED(struct linux_binprm) {
5904 BTF_TYPE_SAFE_TRUSTED(struct file) {
5905 struct inode *f_inode;
5908 BTF_TYPE_SAFE_TRUSTED(struct dentry) {
5909 /* no negative dentry-s in places where bpf can see it */
5910 struct inode *d_inode;
5913 BTF_TYPE_SAFE_TRUSTED(struct socket) {
5917 static bool type_is_rcu(struct bpf_verifier_env *env,
5918 struct bpf_reg_state *reg,
5919 const char *field_name, u32 btf_id)
5921 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct task_struct));
5922 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct cgroup));
5923 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct css_set));
5925 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_rcu");
5928 static bool type_is_rcu_or_null(struct bpf_verifier_env *env,
5929 struct bpf_reg_state *reg,
5930 const char *field_name, u32 btf_id)
5932 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct));
5933 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff));
5934 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock));
5936 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_rcu_or_null");
5939 static bool type_is_trusted(struct bpf_verifier_env *env,
5940 struct bpf_reg_state *reg,
5941 const char *field_name, u32 btf_id)
5943 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta));
5944 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task));
5945 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct linux_binprm));
5946 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct file));
5947 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct dentry));
5948 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct socket));
5950 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_trusted");
5953 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
5954 struct bpf_reg_state *regs,
5955 int regno, int off, int size,
5956 enum bpf_access_type atype,
5959 struct bpf_reg_state *reg = regs + regno;
5960 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
5961 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
5962 const char *field_name = NULL;
5963 enum bpf_type_flag flag = 0;
5967 if (!env->allow_ptr_leaks) {
5969 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
5973 if (!env->prog->gpl_compatible && btf_is_kernel(reg->btf)) {
5975 "Cannot access kernel 'struct %s' from non-GPL compatible program\n",
5981 "R%d is ptr_%s invalid negative access: off=%d\n",
5985 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
5988 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5990 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
5991 regno, tname, off, tn_buf);
5995 if (reg->type & MEM_USER) {
5997 "R%d is ptr_%s access user memory: off=%d\n",
6002 if (reg->type & MEM_PERCPU) {
6004 "R%d is ptr_%s access percpu memory: off=%d\n",
6009 if (env->ops->btf_struct_access && !type_is_alloc(reg->type) && atype == BPF_WRITE) {
6010 if (!btf_is_kernel(reg->btf)) {
6011 verbose(env, "verifier internal error: reg->btf must be kernel btf\n");
6014 ret = env->ops->btf_struct_access(&env->log, reg, off, size);
6016 /* Writes are permitted with default btf_struct_access for
6017 * program allocated objects (which always have ref_obj_id > 0),
6018 * but not for untrusted PTR_TO_BTF_ID | MEM_ALLOC.
6020 if (atype != BPF_READ && !type_is_ptr_alloc_obj(reg->type)) {
6021 verbose(env, "only read is supported\n");
6025 if (type_is_alloc(reg->type) && !type_is_non_owning_ref(reg->type) &&
6027 verbose(env, "verifier internal error: ref_obj_id for allocated object must be non-zero\n");
6031 ret = btf_struct_access(&env->log, reg, off, size, atype, &btf_id, &flag, &field_name);
6037 if (ret != PTR_TO_BTF_ID) {
6040 } else if (type_flag(reg->type) & PTR_UNTRUSTED) {
6041 /* If this is an untrusted pointer, all pointers formed by walking it
6042 * also inherit the untrusted flag.
6044 flag = PTR_UNTRUSTED;
6046 } else if (is_trusted_reg(reg) || is_rcu_reg(reg)) {
6047 /* By default any pointer obtained from walking a trusted pointer is no
6048 * longer trusted, unless the field being accessed has explicitly been
6049 * marked as inheriting its parent's state of trust (either full or RCU).
6051 * 'cgroups' pointer is untrusted if task->cgroups dereference
6052 * happened in a sleepable program outside of bpf_rcu_read_lock()
6053 * section. In a non-sleepable program it's trusted while in RCU CS (aka MEM_RCU).
6054 * Note bpf_rcu_read_unlock() converts MEM_RCU pointers to PTR_UNTRUSTED.
6056 * A regular RCU-protected pointer with __rcu tag can also be deemed
6057 * trusted if we are in an RCU CS. Such pointer can be NULL.
6059 if (type_is_trusted(env, reg, field_name, btf_id)) {
6060 flag |= PTR_TRUSTED;
6061 } else if (in_rcu_cs(env) && !type_may_be_null(reg->type)) {
6062 if (type_is_rcu(env, reg, field_name, btf_id)) {
6063 /* ignore __rcu tag and mark it MEM_RCU */
6065 } else if (flag & MEM_RCU ||
6066 type_is_rcu_or_null(env, reg, field_name, btf_id)) {
6067 /* __rcu tagged pointers can be NULL */
6068 flag |= MEM_RCU | PTR_MAYBE_NULL;
6069 } else if (flag & (MEM_PERCPU | MEM_USER)) {
6072 /* walking unknown pointers yields old deprecated PTR_TO_BTF_ID */
6073 clear_trusted_flags(&flag);
6077 * If not in RCU CS or MEM_RCU pointer can be NULL then
6078 * aggressively mark as untrusted otherwise such
6079 * pointers will be plain PTR_TO_BTF_ID without flags
6080 * and will be allowed to be passed into helpers for
6083 flag = PTR_UNTRUSTED;
6086 /* Old compat. Deprecated */
6087 clear_trusted_flags(&flag);
6090 if (atype == BPF_READ && value_regno >= 0)
6091 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id, flag);
6096 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
6097 struct bpf_reg_state *regs,
6098 int regno, int off, int size,
6099 enum bpf_access_type atype,
6102 struct bpf_reg_state *reg = regs + regno;
6103 struct bpf_map *map = reg->map_ptr;
6104 struct bpf_reg_state map_reg;
6105 enum bpf_type_flag flag = 0;
6106 const struct btf_type *t;
6112 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
6116 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
6117 verbose(env, "map_ptr access not supported for map type %d\n",
6122 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
6123 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
6125 if (!env->allow_ptr_leaks) {
6127 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
6133 verbose(env, "R%d is %s invalid negative access: off=%d\n",
6138 if (atype != BPF_READ) {
6139 verbose(env, "only read from %s is supported\n", tname);
6143 /* Simulate access to a PTR_TO_BTF_ID */
6144 memset(&map_reg, 0, sizeof(map_reg));
6145 mark_btf_ld_reg(env, &map_reg, 0, PTR_TO_BTF_ID, btf_vmlinux, *map->ops->map_btf_id, 0);
6146 ret = btf_struct_access(&env->log, &map_reg, off, size, atype, &btf_id, &flag, NULL);
6150 if (value_regno >= 0)
6151 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id, flag);
6156 /* Check that the stack access at the given offset is within bounds. The
6157 * maximum valid offset is -1.
6159 * The minimum valid offset is -MAX_BPF_STACK for writes, and
6160 * -state->allocated_stack for reads.
6162 static int check_stack_slot_within_bounds(int off,
6163 struct bpf_func_state *state,
6164 enum bpf_access_type t)
6169 min_valid_off = -MAX_BPF_STACK;
6171 min_valid_off = -state->allocated_stack;
6173 if (off < min_valid_off || off > -1)
6178 /* Check that the stack access at 'regno + off' falls within the maximum stack
6181 * 'off' includes `regno->offset`, but not its dynamic part (if any).
6183 static int check_stack_access_within_bounds(
6184 struct bpf_verifier_env *env,
6185 int regno, int off, int access_size,
6186 enum bpf_access_src src, enum bpf_access_type type)
6188 struct bpf_reg_state *regs = cur_regs(env);
6189 struct bpf_reg_state *reg = regs + regno;
6190 struct bpf_func_state *state = func(env, reg);
6191 int min_off, max_off;
6195 if (src == ACCESS_HELPER)
6196 /* We don't know if helpers are reading or writing (or both). */
6197 err_extra = " indirect access to";
6198 else if (type == BPF_READ)
6199 err_extra = " read from";
6201 err_extra = " write to";
6203 if (tnum_is_const(reg->var_off)) {
6204 min_off = reg->var_off.value + off;
6205 if (access_size > 0)
6206 max_off = min_off + access_size - 1;
6210 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
6211 reg->smin_value <= -BPF_MAX_VAR_OFF) {
6212 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
6216 min_off = reg->smin_value + off;
6217 if (access_size > 0)
6218 max_off = reg->smax_value + off + access_size - 1;
6223 err = check_stack_slot_within_bounds(min_off, state, type);
6225 err = check_stack_slot_within_bounds(max_off, state, type);
6228 if (tnum_is_const(reg->var_off)) {
6229 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
6230 err_extra, regno, off, access_size);
6234 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6235 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
6236 err_extra, regno, tn_buf, access_size);
6242 /* check whether memory at (regno + off) is accessible for t = (read | write)
6243 * if t==write, value_regno is a register which value is stored into memory
6244 * if t==read, value_regno is a register which will receive the value from memory
6245 * if t==write && value_regno==-1, some unknown value is stored into memory
6246 * if t==read && value_regno==-1, don't care what we read from memory
6248 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
6249 int off, int bpf_size, enum bpf_access_type t,
6250 int value_regno, bool strict_alignment_once)
6252 struct bpf_reg_state *regs = cur_regs(env);
6253 struct bpf_reg_state *reg = regs + regno;
6254 struct bpf_func_state *state;
6257 size = bpf_size_to_bytes(bpf_size);
6261 /* alignment checks will add in reg->off themselves */
6262 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
6266 /* for access checks, reg->off is just part of off */
6269 if (reg->type == PTR_TO_MAP_KEY) {
6270 if (t == BPF_WRITE) {
6271 verbose(env, "write to change key R%d not allowed\n", regno);
6275 err = check_mem_region_access(env, regno, off, size,
6276 reg->map_ptr->key_size, false);
6279 if (value_regno >= 0)
6280 mark_reg_unknown(env, regs, value_regno);
6281 } else if (reg->type == PTR_TO_MAP_VALUE) {
6282 struct btf_field *kptr_field = NULL;
6284 if (t == BPF_WRITE && value_regno >= 0 &&
6285 is_pointer_value(env, value_regno)) {
6286 verbose(env, "R%d leaks addr into map\n", value_regno);
6289 err = check_map_access_type(env, regno, off, size, t);
6292 err = check_map_access(env, regno, off, size, false, ACCESS_DIRECT);
6295 if (tnum_is_const(reg->var_off))
6296 kptr_field = btf_record_find(reg->map_ptr->record,
6297 off + reg->var_off.value, BPF_KPTR);
6299 err = check_map_kptr_access(env, regno, value_regno, insn_idx, kptr_field);
6300 } else if (t == BPF_READ && value_regno >= 0) {
6301 struct bpf_map *map = reg->map_ptr;
6303 /* if map is read-only, track its contents as scalars */
6304 if (tnum_is_const(reg->var_off) &&
6305 bpf_map_is_rdonly(map) &&
6306 map->ops->map_direct_value_addr) {
6307 int map_off = off + reg->var_off.value;
6310 err = bpf_map_direct_read(map, map_off, size,
6315 regs[value_regno].type = SCALAR_VALUE;
6316 __mark_reg_known(®s[value_regno], val);
6318 mark_reg_unknown(env, regs, value_regno);
6321 } else if (base_type(reg->type) == PTR_TO_MEM) {
6322 bool rdonly_mem = type_is_rdonly_mem(reg->type);
6324 if (type_may_be_null(reg->type)) {
6325 verbose(env, "R%d invalid mem access '%s'\n", regno,
6326 reg_type_str(env, reg->type));
6330 if (t == BPF_WRITE && rdonly_mem) {
6331 verbose(env, "R%d cannot write into %s\n",
6332 regno, reg_type_str(env, reg->type));
6336 if (t == BPF_WRITE && value_regno >= 0 &&
6337 is_pointer_value(env, value_regno)) {
6338 verbose(env, "R%d leaks addr into mem\n", value_regno);
6342 err = check_mem_region_access(env, regno, off, size,
6343 reg->mem_size, false);
6344 if (!err && value_regno >= 0 && (t == BPF_READ || rdonly_mem))
6345 mark_reg_unknown(env, regs, value_regno);
6346 } else if (reg->type == PTR_TO_CTX) {
6347 enum bpf_reg_type reg_type = SCALAR_VALUE;
6348 struct btf *btf = NULL;
6351 if (t == BPF_WRITE && value_regno >= 0 &&
6352 is_pointer_value(env, value_regno)) {
6353 verbose(env, "R%d leaks addr into ctx\n", value_regno);
6357 err = check_ptr_off_reg(env, reg, regno);
6361 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf,
6364 verbose_linfo(env, insn_idx, "; ");
6365 if (!err && t == BPF_READ && value_regno >= 0) {
6366 /* ctx access returns either a scalar, or a
6367 * PTR_TO_PACKET[_META,_END]. In the latter
6368 * case, we know the offset is zero.
6370 if (reg_type == SCALAR_VALUE) {
6371 mark_reg_unknown(env, regs, value_regno);
6373 mark_reg_known_zero(env, regs,
6375 if (type_may_be_null(reg_type))
6376 regs[value_regno].id = ++env->id_gen;
6377 /* A load of ctx field could have different
6378 * actual load size with the one encoded in the
6379 * insn. When the dst is PTR, it is for sure not
6382 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
6383 if (base_type(reg_type) == PTR_TO_BTF_ID) {
6384 regs[value_regno].btf = btf;
6385 regs[value_regno].btf_id = btf_id;
6388 regs[value_regno].type = reg_type;
6391 } else if (reg->type == PTR_TO_STACK) {
6392 /* Basic bounds checks. */
6393 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
6397 state = func(env, reg);
6398 err = update_stack_depth(env, state, off);
6403 err = check_stack_read(env, regno, off, size,
6406 err = check_stack_write(env, regno, off, size,
6407 value_regno, insn_idx);
6408 } else if (reg_is_pkt_pointer(reg)) {
6409 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
6410 verbose(env, "cannot write into packet\n");
6413 if (t == BPF_WRITE && value_regno >= 0 &&
6414 is_pointer_value(env, value_regno)) {
6415 verbose(env, "R%d leaks addr into packet\n",
6419 err = check_packet_access(env, regno, off, size, false);
6420 if (!err && t == BPF_READ && value_regno >= 0)
6421 mark_reg_unknown(env, regs, value_regno);
6422 } else if (reg->type == PTR_TO_FLOW_KEYS) {
6423 if (t == BPF_WRITE && value_regno >= 0 &&
6424 is_pointer_value(env, value_regno)) {
6425 verbose(env, "R%d leaks addr into flow keys\n",
6430 err = check_flow_keys_access(env, off, size);
6431 if (!err && t == BPF_READ && value_regno >= 0)
6432 mark_reg_unknown(env, regs, value_regno);
6433 } else if (type_is_sk_pointer(reg->type)) {
6434 if (t == BPF_WRITE) {
6435 verbose(env, "R%d cannot write into %s\n",
6436 regno, reg_type_str(env, reg->type));
6439 err = check_sock_access(env, insn_idx, regno, off, size, t);
6440 if (!err && value_regno >= 0)
6441 mark_reg_unknown(env, regs, value_regno);
6442 } else if (reg->type == PTR_TO_TP_BUFFER) {
6443 err = check_tp_buffer_access(env, reg, regno, off, size);
6444 if (!err && t == BPF_READ && value_regno >= 0)
6445 mark_reg_unknown(env, regs, value_regno);
6446 } else if (base_type(reg->type) == PTR_TO_BTF_ID &&
6447 !type_may_be_null(reg->type)) {
6448 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
6450 } else if (reg->type == CONST_PTR_TO_MAP) {
6451 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
6453 } else if (base_type(reg->type) == PTR_TO_BUF) {
6454 bool rdonly_mem = type_is_rdonly_mem(reg->type);
6458 if (t == BPF_WRITE) {
6459 verbose(env, "R%d cannot write into %s\n",
6460 regno, reg_type_str(env, reg->type));
6463 max_access = &env->prog->aux->max_rdonly_access;
6465 max_access = &env->prog->aux->max_rdwr_access;
6468 err = check_buffer_access(env, reg, regno, off, size, false,
6471 if (!err && value_regno >= 0 && (rdonly_mem || t == BPF_READ))
6472 mark_reg_unknown(env, regs, value_regno);
6474 verbose(env, "R%d invalid mem access '%s'\n", regno,
6475 reg_type_str(env, reg->type));
6479 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
6480 regs[value_regno].type == SCALAR_VALUE) {
6481 /* b/h/w load zero-extends, mark upper bits as known 0 */
6482 coerce_reg_to_size(®s[value_regno], size);
6487 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
6492 switch (insn->imm) {
6494 case BPF_ADD | BPF_FETCH:
6496 case BPF_AND | BPF_FETCH:
6498 case BPF_OR | BPF_FETCH:
6500 case BPF_XOR | BPF_FETCH:
6505 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
6509 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
6510 verbose(env, "invalid atomic operand size\n");
6514 /* check src1 operand */
6515 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6519 /* check src2 operand */
6520 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6524 if (insn->imm == BPF_CMPXCHG) {
6525 /* Check comparison of R0 with memory location */
6526 const u32 aux_reg = BPF_REG_0;
6528 err = check_reg_arg(env, aux_reg, SRC_OP);
6532 if (is_pointer_value(env, aux_reg)) {
6533 verbose(env, "R%d leaks addr into mem\n", aux_reg);
6538 if (is_pointer_value(env, insn->src_reg)) {
6539 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
6543 if (is_ctx_reg(env, insn->dst_reg) ||
6544 is_pkt_reg(env, insn->dst_reg) ||
6545 is_flow_key_reg(env, insn->dst_reg) ||
6546 is_sk_reg(env, insn->dst_reg)) {
6547 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
6549 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
6553 if (insn->imm & BPF_FETCH) {
6554 if (insn->imm == BPF_CMPXCHG)
6555 load_reg = BPF_REG_0;
6557 load_reg = insn->src_reg;
6559 /* check and record load of old value */
6560 err = check_reg_arg(env, load_reg, DST_OP);
6564 /* This instruction accesses a memory location but doesn't
6565 * actually load it into a register.
6570 /* Check whether we can read the memory, with second call for fetch
6571 * case to simulate the register fill.
6573 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
6574 BPF_SIZE(insn->code), BPF_READ, -1, true);
6575 if (!err && load_reg >= 0)
6576 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
6577 BPF_SIZE(insn->code), BPF_READ, load_reg,
6582 /* Check whether we can write into the same memory. */
6583 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
6584 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
6591 /* When register 'regno' is used to read the stack (either directly or through
6592 * a helper function) make sure that it's within stack boundary and, depending
6593 * on the access type, that all elements of the stack are initialized.
6595 * 'off' includes 'regno->off', but not its dynamic part (if any).
6597 * All registers that have been spilled on the stack in the slots within the
6598 * read offsets are marked as read.
6600 static int check_stack_range_initialized(
6601 struct bpf_verifier_env *env, int regno, int off,
6602 int access_size, bool zero_size_allowed,
6603 enum bpf_access_src type, struct bpf_call_arg_meta *meta)
6605 struct bpf_reg_state *reg = reg_state(env, regno);
6606 struct bpf_func_state *state = func(env, reg);
6607 int err, min_off, max_off, i, j, slot, spi;
6608 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
6609 enum bpf_access_type bounds_check_type;
6610 /* Some accesses can write anything into the stack, others are
6613 bool clobber = false;
6615 if (access_size == 0 && !zero_size_allowed) {
6616 verbose(env, "invalid zero-sized read\n");
6620 if (type == ACCESS_HELPER) {
6621 /* The bounds checks for writes are more permissive than for
6622 * reads. However, if raw_mode is not set, we'll do extra
6625 bounds_check_type = BPF_WRITE;
6628 bounds_check_type = BPF_READ;
6630 err = check_stack_access_within_bounds(env, regno, off, access_size,
6631 type, bounds_check_type);
6636 if (tnum_is_const(reg->var_off)) {
6637 min_off = max_off = reg->var_off.value + off;
6639 /* Variable offset is prohibited for unprivileged mode for
6640 * simplicity since it requires corresponding support in
6641 * Spectre masking for stack ALU.
6642 * See also retrieve_ptr_limit().
6644 if (!env->bypass_spec_v1) {
6647 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6648 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
6649 regno, err_extra, tn_buf);
6652 /* Only initialized buffer on stack is allowed to be accessed
6653 * with variable offset. With uninitialized buffer it's hard to
6654 * guarantee that whole memory is marked as initialized on
6655 * helper return since specific bounds are unknown what may
6656 * cause uninitialized stack leaking.
6658 if (meta && meta->raw_mode)
6661 min_off = reg->smin_value + off;
6662 max_off = reg->smax_value + off;
6665 if (meta && meta->raw_mode) {
6666 /* Ensure we won't be overwriting dynptrs when simulating byte
6667 * by byte access in check_helper_call using meta.access_size.
6668 * This would be a problem if we have a helper in the future
6671 * helper(uninit_mem, len, dynptr)
6673 * Now, uninint_mem may overlap with dynptr pointer. Hence, it
6674 * may end up writing to dynptr itself when touching memory from
6675 * arg 1. This can be relaxed on a case by case basis for known
6676 * safe cases, but reject due to the possibilitiy of aliasing by
6679 for (i = min_off; i < max_off + access_size; i++) {
6680 int stack_off = -i - 1;
6683 /* raw_mode may write past allocated_stack */
6684 if (state->allocated_stack <= stack_off)
6686 if (state->stack[spi].slot_type[stack_off % BPF_REG_SIZE] == STACK_DYNPTR) {
6687 verbose(env, "potential write to dynptr at off=%d disallowed\n", i);
6691 meta->access_size = access_size;
6692 meta->regno = regno;
6696 for (i = min_off; i < max_off + access_size; i++) {
6700 spi = slot / BPF_REG_SIZE;
6701 if (state->allocated_stack <= slot)
6703 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
6704 if (*stype == STACK_MISC)
6706 if ((*stype == STACK_ZERO) ||
6707 (*stype == STACK_INVALID && env->allow_uninit_stack)) {
6709 /* helper can write anything into the stack */
6710 *stype = STACK_MISC;
6715 if (is_spilled_reg(&state->stack[spi]) &&
6716 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
6717 env->allow_ptr_leaks)) {
6719 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
6720 for (j = 0; j < BPF_REG_SIZE; j++)
6721 scrub_spilled_slot(&state->stack[spi].slot_type[j]);
6727 if (tnum_is_const(reg->var_off)) {
6728 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
6729 err_extra, regno, min_off, i - min_off, access_size);
6733 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6734 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
6735 err_extra, regno, tn_buf, i - min_off, access_size);
6739 /* reading any byte out of 8-byte 'spill_slot' will cause
6740 * the whole slot to be marked as 'read'
6742 mark_reg_read(env, &state->stack[spi].spilled_ptr,
6743 state->stack[spi].spilled_ptr.parent,
6745 /* We do not set REG_LIVE_WRITTEN for stack slot, as we can not
6746 * be sure that whether stack slot is written to or not. Hence,
6747 * we must still conservatively propagate reads upwards even if
6748 * helper may write to the entire memory range.
6751 return update_stack_depth(env, state, min_off);
6754 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
6755 int access_size, bool zero_size_allowed,
6756 struct bpf_call_arg_meta *meta)
6758 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
6761 switch (base_type(reg->type)) {
6763 case PTR_TO_PACKET_META:
6764 return check_packet_access(env, regno, reg->off, access_size,
6766 case PTR_TO_MAP_KEY:
6767 if (meta && meta->raw_mode) {
6768 verbose(env, "R%d cannot write into %s\n", regno,
6769 reg_type_str(env, reg->type));
6772 return check_mem_region_access(env, regno, reg->off, access_size,
6773 reg->map_ptr->key_size, false);
6774 case PTR_TO_MAP_VALUE:
6775 if (check_map_access_type(env, regno, reg->off, access_size,
6776 meta && meta->raw_mode ? BPF_WRITE :
6779 return check_map_access(env, regno, reg->off, access_size,
6780 zero_size_allowed, ACCESS_HELPER);
6782 if (type_is_rdonly_mem(reg->type)) {
6783 if (meta && meta->raw_mode) {
6784 verbose(env, "R%d cannot write into %s\n", regno,
6785 reg_type_str(env, reg->type));
6789 return check_mem_region_access(env, regno, reg->off,
6790 access_size, reg->mem_size,
6793 if (type_is_rdonly_mem(reg->type)) {
6794 if (meta && meta->raw_mode) {
6795 verbose(env, "R%d cannot write into %s\n", regno,
6796 reg_type_str(env, reg->type));
6800 max_access = &env->prog->aux->max_rdonly_access;
6802 max_access = &env->prog->aux->max_rdwr_access;
6804 return check_buffer_access(env, reg, regno, reg->off,
6805 access_size, zero_size_allowed,
6808 return check_stack_range_initialized(
6810 regno, reg->off, access_size,
6811 zero_size_allowed, ACCESS_HELPER, meta);
6813 return check_ptr_to_btf_access(env, regs, regno, reg->off,
6814 access_size, BPF_READ, -1);
6816 /* in case the function doesn't know how to access the context,
6817 * (because we are in a program of type SYSCALL for example), we
6818 * can not statically check its size.
6819 * Dynamically check it now.
6821 if (!env->ops->convert_ctx_access) {
6822 enum bpf_access_type atype = meta && meta->raw_mode ? BPF_WRITE : BPF_READ;
6823 int offset = access_size - 1;
6825 /* Allow zero-byte read from PTR_TO_CTX */
6826 if (access_size == 0)
6827 return zero_size_allowed ? 0 : -EACCES;
6829 return check_mem_access(env, env->insn_idx, regno, offset, BPF_B,
6834 default: /* scalar_value or invalid ptr */
6835 /* Allow zero-byte read from NULL, regardless of pointer type */
6836 if (zero_size_allowed && access_size == 0 &&
6837 register_is_null(reg))
6840 verbose(env, "R%d type=%s ", regno,
6841 reg_type_str(env, reg->type));
6842 verbose(env, "expected=%s\n", reg_type_str(env, PTR_TO_STACK));
6847 static int check_mem_size_reg(struct bpf_verifier_env *env,
6848 struct bpf_reg_state *reg, u32 regno,
6849 bool zero_size_allowed,
6850 struct bpf_call_arg_meta *meta)
6854 /* This is used to refine r0 return value bounds for helpers
6855 * that enforce this value as an upper bound on return values.
6856 * See do_refine_retval_range() for helpers that can refine
6857 * the return value. C type of helper is u32 so we pull register
6858 * bound from umax_value however, if negative verifier errors
6859 * out. Only upper bounds can be learned because retval is an
6860 * int type and negative retvals are allowed.
6862 meta->msize_max_value = reg->umax_value;
6864 /* The register is SCALAR_VALUE; the access check
6865 * happens using its boundaries.
6867 if (!tnum_is_const(reg->var_off))
6868 /* For unprivileged variable accesses, disable raw
6869 * mode so that the program is required to
6870 * initialize all the memory that the helper could
6871 * just partially fill up.
6875 if (reg->smin_value < 0) {
6876 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
6881 if (reg->umin_value == 0) {
6882 err = check_helper_mem_access(env, regno - 1, 0,
6889 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
6890 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
6894 err = check_helper_mem_access(env, regno - 1,
6896 zero_size_allowed, meta);
6898 err = mark_chain_precision(env, regno);
6902 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
6903 u32 regno, u32 mem_size)
6905 bool may_be_null = type_may_be_null(reg->type);
6906 struct bpf_reg_state saved_reg;
6907 struct bpf_call_arg_meta meta;
6910 if (register_is_null(reg))
6913 memset(&meta, 0, sizeof(meta));
6914 /* Assuming that the register contains a value check if the memory
6915 * access is safe. Temporarily save and restore the register's state as
6916 * the conversion shouldn't be visible to a caller.
6920 mark_ptr_not_null_reg(reg);
6923 err = check_helper_mem_access(env, regno, mem_size, true, &meta);
6924 /* Check access for BPF_WRITE */
6925 meta.raw_mode = true;
6926 err = err ?: check_helper_mem_access(env, regno, mem_size, true, &meta);
6934 static int check_kfunc_mem_size_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
6937 struct bpf_reg_state *mem_reg = &cur_regs(env)[regno - 1];
6938 bool may_be_null = type_may_be_null(mem_reg->type);
6939 struct bpf_reg_state saved_reg;
6940 struct bpf_call_arg_meta meta;
6943 WARN_ON_ONCE(regno < BPF_REG_2 || regno > BPF_REG_5);
6945 memset(&meta, 0, sizeof(meta));
6948 saved_reg = *mem_reg;
6949 mark_ptr_not_null_reg(mem_reg);
6952 err = check_mem_size_reg(env, reg, regno, true, &meta);
6953 /* Check access for BPF_WRITE */
6954 meta.raw_mode = true;
6955 err = err ?: check_mem_size_reg(env, reg, regno, true, &meta);
6958 *mem_reg = saved_reg;
6962 /* Implementation details:
6963 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL.
6964 * bpf_obj_new returns PTR_TO_BTF_ID | MEM_ALLOC | PTR_MAYBE_NULL.
6965 * Two bpf_map_lookups (even with the same key) will have different reg->id.
6966 * Two separate bpf_obj_new will also have different reg->id.
6967 * For traditional PTR_TO_MAP_VALUE or PTR_TO_BTF_ID | MEM_ALLOC, the verifier
6968 * clears reg->id after value_or_null->value transition, since the verifier only
6969 * cares about the range of access to valid map value pointer and doesn't care
6970 * about actual address of the map element.
6971 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
6972 * reg->id > 0 after value_or_null->value transition. By doing so
6973 * two bpf_map_lookups will be considered two different pointers that
6974 * point to different bpf_spin_locks. Likewise for pointers to allocated objects
6975 * returned from bpf_obj_new.
6976 * The verifier allows taking only one bpf_spin_lock at a time to avoid
6978 * Since only one bpf_spin_lock is allowed the checks are simpler than
6979 * reg_is_refcounted() logic. The verifier needs to remember only
6980 * one spin_lock instead of array of acquired_refs.
6981 * cur_state->active_lock remembers which map value element or allocated
6982 * object got locked and clears it after bpf_spin_unlock.
6984 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
6987 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
6988 struct bpf_verifier_state *cur = env->cur_state;
6989 bool is_const = tnum_is_const(reg->var_off);
6990 u64 val = reg->var_off.value;
6991 struct bpf_map *map = NULL;
6992 struct btf *btf = NULL;
6993 struct btf_record *rec;
6997 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
7001 if (reg->type == PTR_TO_MAP_VALUE) {
7005 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
7013 rec = reg_btf_record(reg);
7014 if (!btf_record_has_field(rec, BPF_SPIN_LOCK)) {
7015 verbose(env, "%s '%s' has no valid bpf_spin_lock\n", map ? "map" : "local",
7016 map ? map->name : "kptr");
7019 if (rec->spin_lock_off != val + reg->off) {
7020 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock' that is at %d\n",
7021 val + reg->off, rec->spin_lock_off);
7025 if (cur->active_lock.ptr) {
7027 "Locking two bpf_spin_locks are not allowed\n");
7031 cur->active_lock.ptr = map;
7033 cur->active_lock.ptr = btf;
7034 cur->active_lock.id = reg->id;
7043 if (!cur->active_lock.ptr) {
7044 verbose(env, "bpf_spin_unlock without taking a lock\n");
7047 if (cur->active_lock.ptr != ptr ||
7048 cur->active_lock.id != reg->id) {
7049 verbose(env, "bpf_spin_unlock of different lock\n");
7053 invalidate_non_owning_refs(env);
7055 cur->active_lock.ptr = NULL;
7056 cur->active_lock.id = 0;
7061 static int process_timer_func(struct bpf_verifier_env *env, int regno,
7062 struct bpf_call_arg_meta *meta)
7064 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7065 bool is_const = tnum_is_const(reg->var_off);
7066 struct bpf_map *map = reg->map_ptr;
7067 u64 val = reg->var_off.value;
7071 "R%d doesn't have constant offset. bpf_timer has to be at the constant offset\n",
7076 verbose(env, "map '%s' has to have BTF in order to use bpf_timer\n",
7080 if (!btf_record_has_field(map->record, BPF_TIMER)) {
7081 verbose(env, "map '%s' has no valid bpf_timer\n", map->name);
7084 if (map->record->timer_off != val + reg->off) {
7085 verbose(env, "off %lld doesn't point to 'struct bpf_timer' that is at %d\n",
7086 val + reg->off, map->record->timer_off);
7089 if (meta->map_ptr) {
7090 verbose(env, "verifier bug. Two map pointers in a timer helper\n");
7093 meta->map_uid = reg->map_uid;
7094 meta->map_ptr = map;
7098 static int process_kptr_func(struct bpf_verifier_env *env, int regno,
7099 struct bpf_call_arg_meta *meta)
7101 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7102 struct bpf_map *map_ptr = reg->map_ptr;
7103 struct btf_field *kptr_field;
7106 if (!tnum_is_const(reg->var_off)) {
7108 "R%d doesn't have constant offset. kptr has to be at the constant offset\n",
7112 if (!map_ptr->btf) {
7113 verbose(env, "map '%s' has to have BTF in order to use bpf_kptr_xchg\n",
7117 if (!btf_record_has_field(map_ptr->record, BPF_KPTR)) {
7118 verbose(env, "map '%s' has no valid kptr\n", map_ptr->name);
7122 meta->map_ptr = map_ptr;
7123 kptr_off = reg->off + reg->var_off.value;
7124 kptr_field = btf_record_find(map_ptr->record, kptr_off, BPF_KPTR);
7126 verbose(env, "off=%d doesn't point to kptr\n", kptr_off);
7129 if (kptr_field->type != BPF_KPTR_REF) {
7130 verbose(env, "off=%d kptr isn't referenced kptr\n", kptr_off);
7133 meta->kptr_field = kptr_field;
7137 /* There are two register types representing a bpf_dynptr, one is PTR_TO_STACK
7138 * which points to a stack slot, and the other is CONST_PTR_TO_DYNPTR.
7140 * In both cases we deal with the first 8 bytes, but need to mark the next 8
7141 * bytes as STACK_DYNPTR in case of PTR_TO_STACK. In case of
7142 * CONST_PTR_TO_DYNPTR, we are guaranteed to get the beginning of the object.
7144 * Mutability of bpf_dynptr is at two levels, one is at the level of struct
7145 * bpf_dynptr itself, i.e. whether the helper is receiving a pointer to struct
7146 * bpf_dynptr or pointer to const struct bpf_dynptr. In the former case, it can
7147 * mutate the view of the dynptr and also possibly destroy it. In the latter
7148 * case, it cannot mutate the bpf_dynptr itself but it can still mutate the
7149 * memory that dynptr points to.
7151 * The verifier will keep track both levels of mutation (bpf_dynptr's in
7152 * reg->type and the memory's in reg->dynptr.type), but there is no support for
7153 * readonly dynptr view yet, hence only the first case is tracked and checked.
7155 * This is consistent with how C applies the const modifier to a struct object,
7156 * where the pointer itself inside bpf_dynptr becomes const but not what it
7159 * Helpers which do not mutate the bpf_dynptr set MEM_RDONLY in their argument
7160 * type, and declare it as 'const struct bpf_dynptr *' in their prototype.
7162 static int process_dynptr_func(struct bpf_verifier_env *env, int regno, int insn_idx,
7163 enum bpf_arg_type arg_type, int clone_ref_obj_id)
7165 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7168 /* MEM_UNINIT and MEM_RDONLY are exclusive, when applied to an
7169 * ARG_PTR_TO_DYNPTR (or ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_*):
7171 if ((arg_type & (MEM_UNINIT | MEM_RDONLY)) == (MEM_UNINIT | MEM_RDONLY)) {
7172 verbose(env, "verifier internal error: misconfigured dynptr helper type flags\n");
7176 /* MEM_UNINIT - Points to memory that is an appropriate candidate for
7177 * constructing a mutable bpf_dynptr object.
7179 * Currently, this is only possible with PTR_TO_STACK
7180 * pointing to a region of at least 16 bytes which doesn't
7181 * contain an existing bpf_dynptr.
7183 * MEM_RDONLY - Points to a initialized bpf_dynptr that will not be
7184 * mutated or destroyed. However, the memory it points to
7187 * None - Points to a initialized dynptr that can be mutated and
7188 * destroyed, including mutation of the memory it points
7191 if (arg_type & MEM_UNINIT) {
7194 if (!is_dynptr_reg_valid_uninit(env, reg)) {
7195 verbose(env, "Dynptr has to be an uninitialized dynptr\n");
7199 /* we write BPF_DW bits (8 bytes) at a time */
7200 for (i = 0; i < BPF_DYNPTR_SIZE; i += 8) {
7201 err = check_mem_access(env, insn_idx, regno,
7202 i, BPF_DW, BPF_WRITE, -1, false);
7207 err = mark_stack_slots_dynptr(env, reg, arg_type, insn_idx, clone_ref_obj_id);
7208 } else /* MEM_RDONLY and None case from above */ {
7209 /* For the reg->type == PTR_TO_STACK case, bpf_dynptr is never const */
7210 if (reg->type == CONST_PTR_TO_DYNPTR && !(arg_type & MEM_RDONLY)) {
7211 verbose(env, "cannot pass pointer to const bpf_dynptr, the helper mutates it\n");
7215 if (!is_dynptr_reg_valid_init(env, reg)) {
7217 "Expected an initialized dynptr as arg #%d\n",
7222 /* Fold modifiers (in this case, MEM_RDONLY) when checking expected type */
7223 if (!is_dynptr_type_expected(env, reg, arg_type & ~MEM_RDONLY)) {
7225 "Expected a dynptr of type %s as arg #%d\n",
7226 dynptr_type_str(arg_to_dynptr_type(arg_type)), regno);
7230 err = mark_dynptr_read(env, reg);
7235 static u32 iter_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg, int spi)
7237 struct bpf_func_state *state = func(env, reg);
7239 return state->stack[spi].spilled_ptr.ref_obj_id;
7242 static bool is_iter_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7244 return meta->kfunc_flags & (KF_ITER_NEW | KF_ITER_NEXT | KF_ITER_DESTROY);
7247 static bool is_iter_new_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7249 return meta->kfunc_flags & KF_ITER_NEW;
7252 static bool is_iter_next_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7254 return meta->kfunc_flags & KF_ITER_NEXT;
7257 static bool is_iter_destroy_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7259 return meta->kfunc_flags & KF_ITER_DESTROY;
7262 static bool is_kfunc_arg_iter(struct bpf_kfunc_call_arg_meta *meta, int arg)
7264 /* btf_check_iter_kfuncs() guarantees that first argument of any iter
7265 * kfunc is iter state pointer
7267 return arg == 0 && is_iter_kfunc(meta);
7270 static int process_iter_arg(struct bpf_verifier_env *env, int regno, int insn_idx,
7271 struct bpf_kfunc_call_arg_meta *meta)
7273 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7274 const struct btf_type *t;
7275 const struct btf_param *arg;
7276 int spi, err, i, nr_slots;
7279 /* btf_check_iter_kfuncs() ensures we don't need to validate anything here */
7280 arg = &btf_params(meta->func_proto)[0];
7281 t = btf_type_skip_modifiers(meta->btf, arg->type, NULL); /* PTR */
7282 t = btf_type_skip_modifiers(meta->btf, t->type, &btf_id); /* STRUCT */
7283 nr_slots = t->size / BPF_REG_SIZE;
7285 if (is_iter_new_kfunc(meta)) {
7286 /* bpf_iter_<type>_new() expects pointer to uninit iter state */
7287 if (!is_iter_reg_valid_uninit(env, reg, nr_slots)) {
7288 verbose(env, "expected uninitialized iter_%s as arg #%d\n",
7289 iter_type_str(meta->btf, btf_id), regno);
7293 for (i = 0; i < nr_slots * 8; i += BPF_REG_SIZE) {
7294 err = check_mem_access(env, insn_idx, regno,
7295 i, BPF_DW, BPF_WRITE, -1, false);
7300 err = mark_stack_slots_iter(env, reg, insn_idx, meta->btf, btf_id, nr_slots);
7304 /* iter_next() or iter_destroy() expect initialized iter state*/
7305 if (!is_iter_reg_valid_init(env, reg, meta->btf, btf_id, nr_slots)) {
7306 verbose(env, "expected an initialized iter_%s as arg #%d\n",
7307 iter_type_str(meta->btf, btf_id), regno);
7311 spi = iter_get_spi(env, reg, nr_slots);
7315 err = mark_iter_read(env, reg, spi, nr_slots);
7319 /* remember meta->iter info for process_iter_next_call() */
7320 meta->iter.spi = spi;
7321 meta->iter.frameno = reg->frameno;
7322 meta->ref_obj_id = iter_ref_obj_id(env, reg, spi);
7324 if (is_iter_destroy_kfunc(meta)) {
7325 err = unmark_stack_slots_iter(env, reg, nr_slots);
7334 /* process_iter_next_call() is called when verifier gets to iterator's next
7335 * "method" (e.g., bpf_iter_num_next() for numbers iterator) call. We'll refer
7336 * to it as just "iter_next()" in comments below.
7338 * BPF verifier relies on a crucial contract for any iter_next()
7339 * implementation: it should *eventually* return NULL, and once that happens
7340 * it should keep returning NULL. That is, once iterator exhausts elements to
7341 * iterate, it should never reset or spuriously return new elements.
7343 * With the assumption of such contract, process_iter_next_call() simulates
7344 * a fork in the verifier state to validate loop logic correctness and safety
7345 * without having to simulate infinite amount of iterations.
7347 * In current state, we first assume that iter_next() returned NULL and
7348 * iterator state is set to DRAINED (BPF_ITER_STATE_DRAINED). In such
7349 * conditions we should not form an infinite loop and should eventually reach
7352 * Besides that, we also fork current state and enqueue it for later
7353 * verification. In a forked state we keep iterator state as ACTIVE
7354 * (BPF_ITER_STATE_ACTIVE) and assume non-NULL return from iter_next(). We
7355 * also bump iteration depth to prevent erroneous infinite loop detection
7356 * later on (see iter_active_depths_differ() comment for details). In this
7357 * state we assume that we'll eventually loop back to another iter_next()
7358 * calls (it could be in exactly same location or in some other instruction,
7359 * it doesn't matter, we don't make any unnecessary assumptions about this,
7360 * everything revolves around iterator state in a stack slot, not which
7361 * instruction is calling iter_next()). When that happens, we either will come
7362 * to iter_next() with equivalent state and can conclude that next iteration
7363 * will proceed in exactly the same way as we just verified, so it's safe to
7364 * assume that loop converges. If not, we'll go on another iteration
7365 * simulation with a different input state, until all possible starting states
7366 * are validated or we reach maximum number of instructions limit.
7368 * This way, we will either exhaustively discover all possible input states
7369 * that iterator loop can start with and eventually will converge, or we'll
7370 * effectively regress into bounded loop simulation logic and either reach
7371 * maximum number of instructions if loop is not provably convergent, or there
7372 * is some statically known limit on number of iterations (e.g., if there is
7373 * an explicit `if n > 100 then break;` statement somewhere in the loop).
7375 * One very subtle but very important aspect is that we *always* simulate NULL
7376 * condition first (as the current state) before we simulate non-NULL case.
7377 * This has to do with intricacies of scalar precision tracking. By simulating
7378 * "exit condition" of iter_next() returning NULL first, we make sure all the
7379 * relevant precision marks *that will be set **after** we exit iterator loop*
7380 * are propagated backwards to common parent state of NULL and non-NULL
7381 * branches. Thanks to that, state equivalence checks done later in forked
7382 * state, when reaching iter_next() for ACTIVE iterator, can assume that
7383 * precision marks are finalized and won't change. Because simulating another
7384 * ACTIVE iterator iteration won't change them (because given same input
7385 * states we'll end up with exactly same output states which we are currently
7386 * comparing; and verification after the loop already propagated back what
7387 * needs to be **additionally** tracked as precise). It's subtle, grok
7388 * precision tracking for more intuitive understanding.
7390 static int process_iter_next_call(struct bpf_verifier_env *env, int insn_idx,
7391 struct bpf_kfunc_call_arg_meta *meta)
7393 struct bpf_verifier_state *cur_st = env->cur_state, *queued_st;
7394 struct bpf_func_state *cur_fr = cur_st->frame[cur_st->curframe], *queued_fr;
7395 struct bpf_reg_state *cur_iter, *queued_iter;
7396 int iter_frameno = meta->iter.frameno;
7397 int iter_spi = meta->iter.spi;
7399 BTF_TYPE_EMIT(struct bpf_iter);
7401 cur_iter = &env->cur_state->frame[iter_frameno]->stack[iter_spi].spilled_ptr;
7403 if (cur_iter->iter.state != BPF_ITER_STATE_ACTIVE &&
7404 cur_iter->iter.state != BPF_ITER_STATE_DRAINED) {
7405 verbose(env, "verifier internal error: unexpected iterator state %d (%s)\n",
7406 cur_iter->iter.state, iter_state_str(cur_iter->iter.state));
7410 if (cur_iter->iter.state == BPF_ITER_STATE_ACTIVE) {
7411 /* branch out active iter state */
7412 queued_st = push_stack(env, insn_idx + 1, insn_idx, false);
7416 queued_iter = &queued_st->frame[iter_frameno]->stack[iter_spi].spilled_ptr;
7417 queued_iter->iter.state = BPF_ITER_STATE_ACTIVE;
7418 queued_iter->iter.depth++;
7420 queued_fr = queued_st->frame[queued_st->curframe];
7421 mark_ptr_not_null_reg(&queued_fr->regs[BPF_REG_0]);
7424 /* switch to DRAINED state, but keep the depth unchanged */
7425 /* mark current iter state as drained and assume returned NULL */
7426 cur_iter->iter.state = BPF_ITER_STATE_DRAINED;
7427 __mark_reg_const_zero(&cur_fr->regs[BPF_REG_0]);
7432 static bool arg_type_is_mem_size(enum bpf_arg_type type)
7434 return type == ARG_CONST_SIZE ||
7435 type == ARG_CONST_SIZE_OR_ZERO;
7438 static bool arg_type_is_release(enum bpf_arg_type type)
7440 return type & OBJ_RELEASE;
7443 static bool arg_type_is_dynptr(enum bpf_arg_type type)
7445 return base_type(type) == ARG_PTR_TO_DYNPTR;
7448 static int int_ptr_type_to_size(enum bpf_arg_type type)
7450 if (type == ARG_PTR_TO_INT)
7452 else if (type == ARG_PTR_TO_LONG)
7458 static int resolve_map_arg_type(struct bpf_verifier_env *env,
7459 const struct bpf_call_arg_meta *meta,
7460 enum bpf_arg_type *arg_type)
7462 if (!meta->map_ptr) {
7463 /* kernel subsystem misconfigured verifier */
7464 verbose(env, "invalid map_ptr to access map->type\n");
7468 switch (meta->map_ptr->map_type) {
7469 case BPF_MAP_TYPE_SOCKMAP:
7470 case BPF_MAP_TYPE_SOCKHASH:
7471 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
7472 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
7474 verbose(env, "invalid arg_type for sockmap/sockhash\n");
7478 case BPF_MAP_TYPE_BLOOM_FILTER:
7479 if (meta->func_id == BPF_FUNC_map_peek_elem)
7480 *arg_type = ARG_PTR_TO_MAP_VALUE;
7488 struct bpf_reg_types {
7489 const enum bpf_reg_type types[10];
7493 static const struct bpf_reg_types sock_types = {
7503 static const struct bpf_reg_types btf_id_sock_common_types = {
7510 PTR_TO_BTF_ID | PTR_TRUSTED,
7512 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
7516 static const struct bpf_reg_types mem_types = {
7524 PTR_TO_MEM | MEM_RINGBUF,
7526 PTR_TO_BTF_ID | PTR_TRUSTED,
7530 static const struct bpf_reg_types int_ptr_types = {
7540 static const struct bpf_reg_types spin_lock_types = {
7543 PTR_TO_BTF_ID | MEM_ALLOC,
7547 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
7548 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
7549 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
7550 static const struct bpf_reg_types ringbuf_mem_types = { .types = { PTR_TO_MEM | MEM_RINGBUF } };
7551 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
7552 static const struct bpf_reg_types btf_ptr_types = {
7555 PTR_TO_BTF_ID | PTR_TRUSTED,
7556 PTR_TO_BTF_ID | MEM_RCU,
7559 static const struct bpf_reg_types percpu_btf_ptr_types = {
7561 PTR_TO_BTF_ID | MEM_PERCPU,
7562 PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED,
7565 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
7566 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
7567 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
7568 static const struct bpf_reg_types timer_types = { .types = { PTR_TO_MAP_VALUE } };
7569 static const struct bpf_reg_types kptr_types = { .types = { PTR_TO_MAP_VALUE } };
7570 static const struct bpf_reg_types dynptr_types = {
7573 CONST_PTR_TO_DYNPTR,
7577 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
7578 [ARG_PTR_TO_MAP_KEY] = &mem_types,
7579 [ARG_PTR_TO_MAP_VALUE] = &mem_types,
7580 [ARG_CONST_SIZE] = &scalar_types,
7581 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
7582 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
7583 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
7584 [ARG_PTR_TO_CTX] = &context_types,
7585 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
7587 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
7589 [ARG_PTR_TO_SOCKET] = &fullsock_types,
7590 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
7591 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
7592 [ARG_PTR_TO_MEM] = &mem_types,
7593 [ARG_PTR_TO_RINGBUF_MEM] = &ringbuf_mem_types,
7594 [ARG_PTR_TO_INT] = &int_ptr_types,
7595 [ARG_PTR_TO_LONG] = &int_ptr_types,
7596 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
7597 [ARG_PTR_TO_FUNC] = &func_ptr_types,
7598 [ARG_PTR_TO_STACK] = &stack_ptr_types,
7599 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
7600 [ARG_PTR_TO_TIMER] = &timer_types,
7601 [ARG_PTR_TO_KPTR] = &kptr_types,
7602 [ARG_PTR_TO_DYNPTR] = &dynptr_types,
7605 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
7606 enum bpf_arg_type arg_type,
7607 const u32 *arg_btf_id,
7608 struct bpf_call_arg_meta *meta)
7610 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7611 enum bpf_reg_type expected, type = reg->type;
7612 const struct bpf_reg_types *compatible;
7615 compatible = compatible_reg_types[base_type(arg_type)];
7617 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
7621 /* ARG_PTR_TO_MEM + RDONLY is compatible with PTR_TO_MEM and PTR_TO_MEM + RDONLY,
7622 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM and NOT with PTR_TO_MEM + RDONLY
7624 * Same for MAYBE_NULL:
7626 * ARG_PTR_TO_MEM + MAYBE_NULL is compatible with PTR_TO_MEM and PTR_TO_MEM + MAYBE_NULL,
7627 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM but NOT with PTR_TO_MEM + MAYBE_NULL
7629 * ARG_PTR_TO_MEM is compatible with PTR_TO_MEM that is tagged with a dynptr type.
7631 * Therefore we fold these flags depending on the arg_type before comparison.
7633 if (arg_type & MEM_RDONLY)
7634 type &= ~MEM_RDONLY;
7635 if (arg_type & PTR_MAYBE_NULL)
7636 type &= ~PTR_MAYBE_NULL;
7637 if (base_type(arg_type) == ARG_PTR_TO_MEM)
7638 type &= ~DYNPTR_TYPE_FLAG_MASK;
7640 if (meta->func_id == BPF_FUNC_kptr_xchg && type_is_alloc(type))
7643 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
7644 expected = compatible->types[i];
7645 if (expected == NOT_INIT)
7648 if (type == expected)
7652 verbose(env, "R%d type=%s expected=", regno, reg_type_str(env, reg->type));
7653 for (j = 0; j + 1 < i; j++)
7654 verbose(env, "%s, ", reg_type_str(env, compatible->types[j]));
7655 verbose(env, "%s\n", reg_type_str(env, compatible->types[j]));
7659 if (base_type(reg->type) != PTR_TO_BTF_ID)
7662 if (compatible == &mem_types) {
7663 if (!(arg_type & MEM_RDONLY)) {
7665 "%s() may write into memory pointed by R%d type=%s\n",
7666 func_id_name(meta->func_id),
7667 regno, reg_type_str(env, reg->type));
7673 switch ((int)reg->type) {
7675 case PTR_TO_BTF_ID | PTR_TRUSTED:
7676 case PTR_TO_BTF_ID | MEM_RCU:
7677 case PTR_TO_BTF_ID | PTR_MAYBE_NULL:
7678 case PTR_TO_BTF_ID | PTR_MAYBE_NULL | MEM_RCU:
7680 /* For bpf_sk_release, it needs to match against first member
7681 * 'struct sock_common', hence make an exception for it. This
7682 * allows bpf_sk_release to work for multiple socket types.
7684 bool strict_type_match = arg_type_is_release(arg_type) &&
7685 meta->func_id != BPF_FUNC_sk_release;
7687 if (type_may_be_null(reg->type) &&
7688 (!type_may_be_null(arg_type) || arg_type_is_release(arg_type))) {
7689 verbose(env, "Possibly NULL pointer passed to helper arg%d\n", regno);
7694 if (!compatible->btf_id) {
7695 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
7698 arg_btf_id = compatible->btf_id;
7701 if (meta->func_id == BPF_FUNC_kptr_xchg) {
7702 if (map_kptr_match_type(env, meta->kptr_field, reg, regno))
7705 if (arg_btf_id == BPF_PTR_POISON) {
7706 verbose(env, "verifier internal error:");
7707 verbose(env, "R%d has non-overwritten BPF_PTR_POISON type\n",
7712 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
7713 btf_vmlinux, *arg_btf_id,
7714 strict_type_match)) {
7715 verbose(env, "R%d is of type %s but %s is expected\n",
7716 regno, btf_type_name(reg->btf, reg->btf_id),
7717 btf_type_name(btf_vmlinux, *arg_btf_id));
7723 case PTR_TO_BTF_ID | MEM_ALLOC:
7724 if (meta->func_id != BPF_FUNC_spin_lock && meta->func_id != BPF_FUNC_spin_unlock &&
7725 meta->func_id != BPF_FUNC_kptr_xchg) {
7726 verbose(env, "verifier internal error: unimplemented handling of MEM_ALLOC\n");
7729 /* Handled by helper specific checks */
7731 case PTR_TO_BTF_ID | MEM_PERCPU:
7732 case PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED:
7733 /* Handled by helper specific checks */
7736 verbose(env, "verifier internal error: invalid PTR_TO_BTF_ID register for type match\n");
7742 static struct btf_field *
7743 reg_find_field_offset(const struct bpf_reg_state *reg, s32 off, u32 fields)
7745 struct btf_field *field;
7746 struct btf_record *rec;
7748 rec = reg_btf_record(reg);
7752 field = btf_record_find(rec, off, fields);
7759 int check_func_arg_reg_off(struct bpf_verifier_env *env,
7760 const struct bpf_reg_state *reg, int regno,
7761 enum bpf_arg_type arg_type)
7763 u32 type = reg->type;
7765 /* When referenced register is passed to release function, its fixed
7768 * We will check arg_type_is_release reg has ref_obj_id when storing
7769 * meta->release_regno.
7771 if (arg_type_is_release(arg_type)) {
7772 /* ARG_PTR_TO_DYNPTR with OBJ_RELEASE is a bit special, as it
7773 * may not directly point to the object being released, but to
7774 * dynptr pointing to such object, which might be at some offset
7775 * on the stack. In that case, we simply to fallback to the
7778 if (arg_type_is_dynptr(arg_type) && type == PTR_TO_STACK)
7781 if ((type_is_ptr_alloc_obj(type) || type_is_non_owning_ref(type)) && reg->off) {
7782 if (reg_find_field_offset(reg, reg->off, BPF_GRAPH_NODE_OR_ROOT))
7783 return __check_ptr_off_reg(env, reg, regno, true);
7785 verbose(env, "R%d must have zero offset when passed to release func\n",
7787 verbose(env, "No graph node or root found at R%d type:%s off:%d\n", regno,
7788 btf_type_name(reg->btf, reg->btf_id), reg->off);
7792 /* Doing check_ptr_off_reg check for the offset will catch this
7793 * because fixed_off_ok is false, but checking here allows us
7794 * to give the user a better error message.
7797 verbose(env, "R%d must have zero offset when passed to release func or trusted arg to kfunc\n",
7801 return __check_ptr_off_reg(env, reg, regno, false);
7805 /* Pointer types where both fixed and variable offset is explicitly allowed: */
7808 case PTR_TO_PACKET_META:
7809 case PTR_TO_MAP_KEY:
7810 case PTR_TO_MAP_VALUE:
7812 case PTR_TO_MEM | MEM_RDONLY:
7813 case PTR_TO_MEM | MEM_RINGBUF:
7815 case PTR_TO_BUF | MEM_RDONLY:
7818 /* All the rest must be rejected, except PTR_TO_BTF_ID which allows
7822 case PTR_TO_BTF_ID | MEM_ALLOC:
7823 case PTR_TO_BTF_ID | PTR_TRUSTED:
7824 case PTR_TO_BTF_ID | MEM_RCU:
7825 case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF:
7826 /* When referenced PTR_TO_BTF_ID is passed to release function,
7827 * its fixed offset must be 0. In the other cases, fixed offset
7828 * can be non-zero. This was already checked above. So pass
7829 * fixed_off_ok as true to allow fixed offset for all other
7830 * cases. var_off always must be 0 for PTR_TO_BTF_ID, hence we
7831 * still need to do checks instead of returning.
7833 return __check_ptr_off_reg(env, reg, regno, true);
7835 return __check_ptr_off_reg(env, reg, regno, false);
7839 static struct bpf_reg_state *get_dynptr_arg_reg(struct bpf_verifier_env *env,
7840 const struct bpf_func_proto *fn,
7841 struct bpf_reg_state *regs)
7843 struct bpf_reg_state *state = NULL;
7846 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++)
7847 if (arg_type_is_dynptr(fn->arg_type[i])) {
7849 verbose(env, "verifier internal error: multiple dynptr args\n");
7852 state = ®s[BPF_REG_1 + i];
7856 verbose(env, "verifier internal error: no dynptr arg found\n");
7861 static int dynptr_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
7863 struct bpf_func_state *state = func(env, reg);
7866 if (reg->type == CONST_PTR_TO_DYNPTR)
7868 spi = dynptr_get_spi(env, reg);
7871 return state->stack[spi].spilled_ptr.id;
7874 static int dynptr_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
7876 struct bpf_func_state *state = func(env, reg);
7879 if (reg->type == CONST_PTR_TO_DYNPTR)
7880 return reg->ref_obj_id;
7881 spi = dynptr_get_spi(env, reg);
7884 return state->stack[spi].spilled_ptr.ref_obj_id;
7887 static enum bpf_dynptr_type dynptr_get_type(struct bpf_verifier_env *env,
7888 struct bpf_reg_state *reg)
7890 struct bpf_func_state *state = func(env, reg);
7893 if (reg->type == CONST_PTR_TO_DYNPTR)
7894 return reg->dynptr.type;
7896 spi = __get_spi(reg->off);
7898 verbose(env, "verifier internal error: invalid spi when querying dynptr type\n");
7899 return BPF_DYNPTR_TYPE_INVALID;
7902 return state->stack[spi].spilled_ptr.dynptr.type;
7905 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
7906 struct bpf_call_arg_meta *meta,
7907 const struct bpf_func_proto *fn,
7910 u32 regno = BPF_REG_1 + arg;
7911 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7912 enum bpf_arg_type arg_type = fn->arg_type[arg];
7913 enum bpf_reg_type type = reg->type;
7914 u32 *arg_btf_id = NULL;
7917 if (arg_type == ARG_DONTCARE)
7920 err = check_reg_arg(env, regno, SRC_OP);
7924 if (arg_type == ARG_ANYTHING) {
7925 if (is_pointer_value(env, regno)) {
7926 verbose(env, "R%d leaks addr into helper function\n",
7933 if (type_is_pkt_pointer(type) &&
7934 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
7935 verbose(env, "helper access to the packet is not allowed\n");
7939 if (base_type(arg_type) == ARG_PTR_TO_MAP_VALUE) {
7940 err = resolve_map_arg_type(env, meta, &arg_type);
7945 if (register_is_null(reg) && type_may_be_null(arg_type))
7946 /* A NULL register has a SCALAR_VALUE type, so skip
7949 goto skip_type_check;
7951 /* arg_btf_id and arg_size are in a union. */
7952 if (base_type(arg_type) == ARG_PTR_TO_BTF_ID ||
7953 base_type(arg_type) == ARG_PTR_TO_SPIN_LOCK)
7954 arg_btf_id = fn->arg_btf_id[arg];
7956 err = check_reg_type(env, regno, arg_type, arg_btf_id, meta);
7960 err = check_func_arg_reg_off(env, reg, regno, arg_type);
7965 if (arg_type_is_release(arg_type)) {
7966 if (arg_type_is_dynptr(arg_type)) {
7967 struct bpf_func_state *state = func(env, reg);
7970 /* Only dynptr created on stack can be released, thus
7971 * the get_spi and stack state checks for spilled_ptr
7972 * should only be done before process_dynptr_func for
7975 if (reg->type == PTR_TO_STACK) {
7976 spi = dynptr_get_spi(env, reg);
7977 if (spi < 0 || !state->stack[spi].spilled_ptr.ref_obj_id) {
7978 verbose(env, "arg %d is an unacquired reference\n", regno);
7982 verbose(env, "cannot release unowned const bpf_dynptr\n");
7985 } else if (!reg->ref_obj_id && !register_is_null(reg)) {
7986 verbose(env, "R%d must be referenced when passed to release function\n",
7990 if (meta->release_regno) {
7991 verbose(env, "verifier internal error: more than one release argument\n");
7994 meta->release_regno = regno;
7997 if (reg->ref_obj_id) {
7998 if (meta->ref_obj_id) {
7999 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
8000 regno, reg->ref_obj_id,
8004 meta->ref_obj_id = reg->ref_obj_id;
8007 switch (base_type(arg_type)) {
8008 case ARG_CONST_MAP_PTR:
8009 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
8010 if (meta->map_ptr) {
8011 /* Use map_uid (which is unique id of inner map) to reject:
8012 * inner_map1 = bpf_map_lookup_elem(outer_map, key1)
8013 * inner_map2 = bpf_map_lookup_elem(outer_map, key2)
8014 * if (inner_map1 && inner_map2) {
8015 * timer = bpf_map_lookup_elem(inner_map1);
8017 * // mismatch would have been allowed
8018 * bpf_timer_init(timer, inner_map2);
8021 * Comparing map_ptr is enough to distinguish normal and outer maps.
8023 if (meta->map_ptr != reg->map_ptr ||
8024 meta->map_uid != reg->map_uid) {
8026 "timer pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
8027 meta->map_uid, reg->map_uid);
8031 meta->map_ptr = reg->map_ptr;
8032 meta->map_uid = reg->map_uid;
8034 case ARG_PTR_TO_MAP_KEY:
8035 /* bpf_map_xxx(..., map_ptr, ..., key) call:
8036 * check that [key, key + map->key_size) are within
8037 * stack limits and initialized
8039 if (!meta->map_ptr) {
8040 /* in function declaration map_ptr must come before
8041 * map_key, so that it's verified and known before
8042 * we have to check map_key here. Otherwise it means
8043 * that kernel subsystem misconfigured verifier
8045 verbose(env, "invalid map_ptr to access map->key\n");
8048 err = check_helper_mem_access(env, regno,
8049 meta->map_ptr->key_size, false,
8052 case ARG_PTR_TO_MAP_VALUE:
8053 if (type_may_be_null(arg_type) && register_is_null(reg))
8056 /* bpf_map_xxx(..., map_ptr, ..., value) call:
8057 * check [value, value + map->value_size) validity
8059 if (!meta->map_ptr) {
8060 /* kernel subsystem misconfigured verifier */
8061 verbose(env, "invalid map_ptr to access map->value\n");
8064 meta->raw_mode = arg_type & MEM_UNINIT;
8065 err = check_helper_mem_access(env, regno,
8066 meta->map_ptr->value_size, false,
8069 case ARG_PTR_TO_PERCPU_BTF_ID:
8071 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
8074 meta->ret_btf = reg->btf;
8075 meta->ret_btf_id = reg->btf_id;
8077 case ARG_PTR_TO_SPIN_LOCK:
8078 if (in_rbtree_lock_required_cb(env)) {
8079 verbose(env, "can't spin_{lock,unlock} in rbtree cb\n");
8082 if (meta->func_id == BPF_FUNC_spin_lock) {
8083 err = process_spin_lock(env, regno, true);
8086 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
8087 err = process_spin_lock(env, regno, false);
8091 verbose(env, "verifier internal error\n");
8095 case ARG_PTR_TO_TIMER:
8096 err = process_timer_func(env, regno, meta);
8100 case ARG_PTR_TO_FUNC:
8101 meta->subprogno = reg->subprogno;
8103 case ARG_PTR_TO_MEM:
8104 /* The access to this pointer is only checked when we hit the
8105 * next is_mem_size argument below.
8107 meta->raw_mode = arg_type & MEM_UNINIT;
8108 if (arg_type & MEM_FIXED_SIZE) {
8109 err = check_helper_mem_access(env, regno,
8110 fn->arg_size[arg], false,
8114 case ARG_CONST_SIZE:
8115 err = check_mem_size_reg(env, reg, regno, false, meta);
8117 case ARG_CONST_SIZE_OR_ZERO:
8118 err = check_mem_size_reg(env, reg, regno, true, meta);
8120 case ARG_PTR_TO_DYNPTR:
8121 err = process_dynptr_func(env, regno, insn_idx, arg_type, 0);
8125 case ARG_CONST_ALLOC_SIZE_OR_ZERO:
8126 if (!tnum_is_const(reg->var_off)) {
8127 verbose(env, "R%d is not a known constant'\n",
8131 meta->mem_size = reg->var_off.value;
8132 err = mark_chain_precision(env, regno);
8136 case ARG_PTR_TO_INT:
8137 case ARG_PTR_TO_LONG:
8139 int size = int_ptr_type_to_size(arg_type);
8141 err = check_helper_mem_access(env, regno, size, false, meta);
8144 err = check_ptr_alignment(env, reg, 0, size, true);
8147 case ARG_PTR_TO_CONST_STR:
8149 struct bpf_map *map = reg->map_ptr;
8154 if (!bpf_map_is_rdonly(map)) {
8155 verbose(env, "R%d does not point to a readonly map'\n", regno);
8159 if (!tnum_is_const(reg->var_off)) {
8160 verbose(env, "R%d is not a constant address'\n", regno);
8164 if (!map->ops->map_direct_value_addr) {
8165 verbose(env, "no direct value access support for this map type\n");
8169 err = check_map_access(env, regno, reg->off,
8170 map->value_size - reg->off, false,
8175 map_off = reg->off + reg->var_off.value;
8176 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
8178 verbose(env, "direct value access on string failed\n");
8182 str_ptr = (char *)(long)(map_addr);
8183 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
8184 verbose(env, "string is not zero-terminated\n");
8189 case ARG_PTR_TO_KPTR:
8190 err = process_kptr_func(env, regno, meta);
8199 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
8201 enum bpf_attach_type eatype = env->prog->expected_attach_type;
8202 enum bpf_prog_type type = resolve_prog_type(env->prog);
8204 if (func_id != BPF_FUNC_map_update_elem)
8207 /* It's not possible to get access to a locked struct sock in these
8208 * contexts, so updating is safe.
8211 case BPF_PROG_TYPE_TRACING:
8212 if (eatype == BPF_TRACE_ITER)
8215 case BPF_PROG_TYPE_SOCKET_FILTER:
8216 case BPF_PROG_TYPE_SCHED_CLS:
8217 case BPF_PROG_TYPE_SCHED_ACT:
8218 case BPF_PROG_TYPE_XDP:
8219 case BPF_PROG_TYPE_SK_REUSEPORT:
8220 case BPF_PROG_TYPE_FLOW_DISSECTOR:
8221 case BPF_PROG_TYPE_SK_LOOKUP:
8227 verbose(env, "cannot update sockmap in this context\n");
8231 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
8233 return env->prog->jit_requested &&
8234 bpf_jit_supports_subprog_tailcalls();
8237 static int check_map_func_compatibility(struct bpf_verifier_env *env,
8238 struct bpf_map *map, int func_id)
8243 /* We need a two way check, first is from map perspective ... */
8244 switch (map->map_type) {
8245 case BPF_MAP_TYPE_PROG_ARRAY:
8246 if (func_id != BPF_FUNC_tail_call)
8249 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
8250 if (func_id != BPF_FUNC_perf_event_read &&
8251 func_id != BPF_FUNC_perf_event_output &&
8252 func_id != BPF_FUNC_skb_output &&
8253 func_id != BPF_FUNC_perf_event_read_value &&
8254 func_id != BPF_FUNC_xdp_output)
8257 case BPF_MAP_TYPE_RINGBUF:
8258 if (func_id != BPF_FUNC_ringbuf_output &&
8259 func_id != BPF_FUNC_ringbuf_reserve &&
8260 func_id != BPF_FUNC_ringbuf_query &&
8261 func_id != BPF_FUNC_ringbuf_reserve_dynptr &&
8262 func_id != BPF_FUNC_ringbuf_submit_dynptr &&
8263 func_id != BPF_FUNC_ringbuf_discard_dynptr)
8266 case BPF_MAP_TYPE_USER_RINGBUF:
8267 if (func_id != BPF_FUNC_user_ringbuf_drain)
8270 case BPF_MAP_TYPE_STACK_TRACE:
8271 if (func_id != BPF_FUNC_get_stackid)
8274 case BPF_MAP_TYPE_CGROUP_ARRAY:
8275 if (func_id != BPF_FUNC_skb_under_cgroup &&
8276 func_id != BPF_FUNC_current_task_under_cgroup)
8279 case BPF_MAP_TYPE_CGROUP_STORAGE:
8280 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
8281 if (func_id != BPF_FUNC_get_local_storage)
8284 case BPF_MAP_TYPE_DEVMAP:
8285 case BPF_MAP_TYPE_DEVMAP_HASH:
8286 if (func_id != BPF_FUNC_redirect_map &&
8287 func_id != BPF_FUNC_map_lookup_elem)
8290 /* Restrict bpf side of cpumap and xskmap, open when use-cases
8293 case BPF_MAP_TYPE_CPUMAP:
8294 if (func_id != BPF_FUNC_redirect_map)
8297 case BPF_MAP_TYPE_XSKMAP:
8298 if (func_id != BPF_FUNC_redirect_map &&
8299 func_id != BPF_FUNC_map_lookup_elem)
8302 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
8303 case BPF_MAP_TYPE_HASH_OF_MAPS:
8304 if (func_id != BPF_FUNC_map_lookup_elem)
8307 case BPF_MAP_TYPE_SOCKMAP:
8308 if (func_id != BPF_FUNC_sk_redirect_map &&
8309 func_id != BPF_FUNC_sock_map_update &&
8310 func_id != BPF_FUNC_map_delete_elem &&
8311 func_id != BPF_FUNC_msg_redirect_map &&
8312 func_id != BPF_FUNC_sk_select_reuseport &&
8313 func_id != BPF_FUNC_map_lookup_elem &&
8314 !may_update_sockmap(env, func_id))
8317 case BPF_MAP_TYPE_SOCKHASH:
8318 if (func_id != BPF_FUNC_sk_redirect_hash &&
8319 func_id != BPF_FUNC_sock_hash_update &&
8320 func_id != BPF_FUNC_map_delete_elem &&
8321 func_id != BPF_FUNC_msg_redirect_hash &&
8322 func_id != BPF_FUNC_sk_select_reuseport &&
8323 func_id != BPF_FUNC_map_lookup_elem &&
8324 !may_update_sockmap(env, func_id))
8327 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
8328 if (func_id != BPF_FUNC_sk_select_reuseport)
8331 case BPF_MAP_TYPE_QUEUE:
8332 case BPF_MAP_TYPE_STACK:
8333 if (func_id != BPF_FUNC_map_peek_elem &&
8334 func_id != BPF_FUNC_map_pop_elem &&
8335 func_id != BPF_FUNC_map_push_elem)
8338 case BPF_MAP_TYPE_SK_STORAGE:
8339 if (func_id != BPF_FUNC_sk_storage_get &&
8340 func_id != BPF_FUNC_sk_storage_delete &&
8341 func_id != BPF_FUNC_kptr_xchg)
8344 case BPF_MAP_TYPE_INODE_STORAGE:
8345 if (func_id != BPF_FUNC_inode_storage_get &&
8346 func_id != BPF_FUNC_inode_storage_delete &&
8347 func_id != BPF_FUNC_kptr_xchg)
8350 case BPF_MAP_TYPE_TASK_STORAGE:
8351 if (func_id != BPF_FUNC_task_storage_get &&
8352 func_id != BPF_FUNC_task_storage_delete &&
8353 func_id != BPF_FUNC_kptr_xchg)
8356 case BPF_MAP_TYPE_CGRP_STORAGE:
8357 if (func_id != BPF_FUNC_cgrp_storage_get &&
8358 func_id != BPF_FUNC_cgrp_storage_delete &&
8359 func_id != BPF_FUNC_kptr_xchg)
8362 case BPF_MAP_TYPE_BLOOM_FILTER:
8363 if (func_id != BPF_FUNC_map_peek_elem &&
8364 func_id != BPF_FUNC_map_push_elem)
8371 /* ... and second from the function itself. */
8373 case BPF_FUNC_tail_call:
8374 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
8376 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
8377 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
8381 case BPF_FUNC_perf_event_read:
8382 case BPF_FUNC_perf_event_output:
8383 case BPF_FUNC_perf_event_read_value:
8384 case BPF_FUNC_skb_output:
8385 case BPF_FUNC_xdp_output:
8386 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
8389 case BPF_FUNC_ringbuf_output:
8390 case BPF_FUNC_ringbuf_reserve:
8391 case BPF_FUNC_ringbuf_query:
8392 case BPF_FUNC_ringbuf_reserve_dynptr:
8393 case BPF_FUNC_ringbuf_submit_dynptr:
8394 case BPF_FUNC_ringbuf_discard_dynptr:
8395 if (map->map_type != BPF_MAP_TYPE_RINGBUF)
8398 case BPF_FUNC_user_ringbuf_drain:
8399 if (map->map_type != BPF_MAP_TYPE_USER_RINGBUF)
8402 case BPF_FUNC_get_stackid:
8403 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
8406 case BPF_FUNC_current_task_under_cgroup:
8407 case BPF_FUNC_skb_under_cgroup:
8408 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
8411 case BPF_FUNC_redirect_map:
8412 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
8413 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
8414 map->map_type != BPF_MAP_TYPE_CPUMAP &&
8415 map->map_type != BPF_MAP_TYPE_XSKMAP)
8418 case BPF_FUNC_sk_redirect_map:
8419 case BPF_FUNC_msg_redirect_map:
8420 case BPF_FUNC_sock_map_update:
8421 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
8424 case BPF_FUNC_sk_redirect_hash:
8425 case BPF_FUNC_msg_redirect_hash:
8426 case BPF_FUNC_sock_hash_update:
8427 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
8430 case BPF_FUNC_get_local_storage:
8431 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
8432 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
8435 case BPF_FUNC_sk_select_reuseport:
8436 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
8437 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
8438 map->map_type != BPF_MAP_TYPE_SOCKHASH)
8441 case BPF_FUNC_map_pop_elem:
8442 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
8443 map->map_type != BPF_MAP_TYPE_STACK)
8446 case BPF_FUNC_map_peek_elem:
8447 case BPF_FUNC_map_push_elem:
8448 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
8449 map->map_type != BPF_MAP_TYPE_STACK &&
8450 map->map_type != BPF_MAP_TYPE_BLOOM_FILTER)
8453 case BPF_FUNC_map_lookup_percpu_elem:
8454 if (map->map_type != BPF_MAP_TYPE_PERCPU_ARRAY &&
8455 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
8456 map->map_type != BPF_MAP_TYPE_LRU_PERCPU_HASH)
8459 case BPF_FUNC_sk_storage_get:
8460 case BPF_FUNC_sk_storage_delete:
8461 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
8464 case BPF_FUNC_inode_storage_get:
8465 case BPF_FUNC_inode_storage_delete:
8466 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
8469 case BPF_FUNC_task_storage_get:
8470 case BPF_FUNC_task_storage_delete:
8471 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
8474 case BPF_FUNC_cgrp_storage_get:
8475 case BPF_FUNC_cgrp_storage_delete:
8476 if (map->map_type != BPF_MAP_TYPE_CGRP_STORAGE)
8485 verbose(env, "cannot pass map_type %d into func %s#%d\n",
8486 map->map_type, func_id_name(func_id), func_id);
8490 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
8494 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
8496 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
8498 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
8500 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
8502 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
8505 /* We only support one arg being in raw mode at the moment,
8506 * which is sufficient for the helper functions we have
8512 static bool check_args_pair_invalid(const struct bpf_func_proto *fn, int arg)
8514 bool is_fixed = fn->arg_type[arg] & MEM_FIXED_SIZE;
8515 bool has_size = fn->arg_size[arg] != 0;
8516 bool is_next_size = false;
8518 if (arg + 1 < ARRAY_SIZE(fn->arg_type))
8519 is_next_size = arg_type_is_mem_size(fn->arg_type[arg + 1]);
8521 if (base_type(fn->arg_type[arg]) != ARG_PTR_TO_MEM)
8522 return is_next_size;
8524 return has_size == is_next_size || is_next_size == is_fixed;
8527 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
8529 /* bpf_xxx(..., buf, len) call will access 'len'
8530 * bytes from memory 'buf'. Both arg types need
8531 * to be paired, so make sure there's no buggy
8532 * helper function specification.
8534 if (arg_type_is_mem_size(fn->arg1_type) ||
8535 check_args_pair_invalid(fn, 0) ||
8536 check_args_pair_invalid(fn, 1) ||
8537 check_args_pair_invalid(fn, 2) ||
8538 check_args_pair_invalid(fn, 3) ||
8539 check_args_pair_invalid(fn, 4))
8545 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
8549 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
8550 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_BTF_ID)
8551 return !!fn->arg_btf_id[i];
8552 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_SPIN_LOCK)
8553 return fn->arg_btf_id[i] == BPF_PTR_POISON;
8554 if (base_type(fn->arg_type[i]) != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i] &&
8555 /* arg_btf_id and arg_size are in a union. */
8556 (base_type(fn->arg_type[i]) != ARG_PTR_TO_MEM ||
8557 !(fn->arg_type[i] & MEM_FIXED_SIZE)))
8564 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
8566 return check_raw_mode_ok(fn) &&
8567 check_arg_pair_ok(fn) &&
8568 check_btf_id_ok(fn) ? 0 : -EINVAL;
8571 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
8572 * are now invalid, so turn them into unknown SCALAR_VALUE.
8574 * This also applies to dynptr slices belonging to skb and xdp dynptrs,
8575 * since these slices point to packet data.
8577 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
8579 struct bpf_func_state *state;
8580 struct bpf_reg_state *reg;
8582 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
8583 if (reg_is_pkt_pointer_any(reg) || reg_is_dynptr_slice_pkt(reg))
8584 mark_reg_invalid(env, reg);
8590 BEYOND_PKT_END = -2,
8593 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
8595 struct bpf_func_state *state = vstate->frame[vstate->curframe];
8596 struct bpf_reg_state *reg = &state->regs[regn];
8598 if (reg->type != PTR_TO_PACKET)
8599 /* PTR_TO_PACKET_META is not supported yet */
8602 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
8603 * How far beyond pkt_end it goes is unknown.
8604 * if (!range_open) it's the case of pkt >= pkt_end
8605 * if (range_open) it's the case of pkt > pkt_end
8606 * hence this pointer is at least 1 byte bigger than pkt_end
8609 reg->range = BEYOND_PKT_END;
8611 reg->range = AT_PKT_END;
8614 /* The pointer with the specified id has released its reference to kernel
8615 * resources. Identify all copies of the same pointer and clear the reference.
8617 static int release_reference(struct bpf_verifier_env *env,
8620 struct bpf_func_state *state;
8621 struct bpf_reg_state *reg;
8624 err = release_reference_state(cur_func(env), ref_obj_id);
8628 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
8629 if (reg->ref_obj_id == ref_obj_id)
8630 mark_reg_invalid(env, reg);
8636 static void invalidate_non_owning_refs(struct bpf_verifier_env *env)
8638 struct bpf_func_state *unused;
8639 struct bpf_reg_state *reg;
8641 bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
8642 if (type_is_non_owning_ref(reg->type))
8643 mark_reg_invalid(env, reg);
8647 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
8648 struct bpf_reg_state *regs)
8652 /* after the call registers r0 - r5 were scratched */
8653 for (i = 0; i < CALLER_SAVED_REGS; i++) {
8654 mark_reg_not_init(env, regs, caller_saved[i]);
8655 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
8659 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
8660 struct bpf_func_state *caller,
8661 struct bpf_func_state *callee,
8664 static int set_callee_state(struct bpf_verifier_env *env,
8665 struct bpf_func_state *caller,
8666 struct bpf_func_state *callee, int insn_idx);
8668 static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
8669 int *insn_idx, int subprog,
8670 set_callee_state_fn set_callee_state_cb)
8672 struct bpf_verifier_state *state = env->cur_state;
8673 struct bpf_func_state *caller, *callee;
8676 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
8677 verbose(env, "the call stack of %d frames is too deep\n",
8678 state->curframe + 2);
8682 caller = state->frame[state->curframe];
8683 if (state->frame[state->curframe + 1]) {
8684 verbose(env, "verifier bug. Frame %d already allocated\n",
8685 state->curframe + 1);
8689 err = btf_check_subprog_call(env, subprog, caller->regs);
8692 if (subprog_is_global(env, subprog)) {
8694 verbose(env, "Caller passes invalid args into func#%d\n",
8698 if (env->log.level & BPF_LOG_LEVEL)
8700 "Func#%d is global and valid. Skipping.\n",
8702 clear_caller_saved_regs(env, caller->regs);
8704 /* All global functions return a 64-bit SCALAR_VALUE */
8705 mark_reg_unknown(env, caller->regs, BPF_REG_0);
8706 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
8708 /* continue with next insn after call */
8713 /* set_callee_state is used for direct subprog calls, but we are
8714 * interested in validating only BPF helpers that can call subprogs as
8717 if (set_callee_state_cb != set_callee_state) {
8718 if (bpf_pseudo_kfunc_call(insn) &&
8719 !is_callback_calling_kfunc(insn->imm)) {
8720 verbose(env, "verifier bug: kfunc %s#%d not marked as callback-calling\n",
8721 func_id_name(insn->imm), insn->imm);
8723 } else if (!bpf_pseudo_kfunc_call(insn) &&
8724 !is_callback_calling_function(insn->imm)) { /* helper */
8725 verbose(env, "verifier bug: helper %s#%d not marked as callback-calling\n",
8726 func_id_name(insn->imm), insn->imm);
8731 if (insn->code == (BPF_JMP | BPF_CALL) &&
8732 insn->src_reg == 0 &&
8733 insn->imm == BPF_FUNC_timer_set_callback) {
8734 struct bpf_verifier_state *async_cb;
8736 /* there is no real recursion here. timer callbacks are async */
8737 env->subprog_info[subprog].is_async_cb = true;
8738 async_cb = push_async_cb(env, env->subprog_info[subprog].start,
8739 *insn_idx, subprog);
8742 callee = async_cb->frame[0];
8743 callee->async_entry_cnt = caller->async_entry_cnt + 1;
8745 /* Convert bpf_timer_set_callback() args into timer callback args */
8746 err = set_callee_state_cb(env, caller, callee, *insn_idx);
8750 clear_caller_saved_regs(env, caller->regs);
8751 mark_reg_unknown(env, caller->regs, BPF_REG_0);
8752 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
8753 /* continue with next insn after call */
8757 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
8760 state->frame[state->curframe + 1] = callee;
8762 /* callee cannot access r0, r6 - r9 for reading and has to write
8763 * into its own stack before reading from it.
8764 * callee can read/write into caller's stack
8766 init_func_state(env, callee,
8767 /* remember the callsite, it will be used by bpf_exit */
8768 *insn_idx /* callsite */,
8769 state->curframe + 1 /* frameno within this callchain */,
8770 subprog /* subprog number within this prog */);
8772 /* Transfer references to the callee */
8773 err = copy_reference_state(callee, caller);
8777 err = set_callee_state_cb(env, caller, callee, *insn_idx);
8781 clear_caller_saved_regs(env, caller->regs);
8783 /* only increment it after check_reg_arg() finished */
8786 /* and go analyze first insn of the callee */
8787 *insn_idx = env->subprog_info[subprog].start - 1;
8789 if (env->log.level & BPF_LOG_LEVEL) {
8790 verbose(env, "caller:\n");
8791 print_verifier_state(env, caller, true);
8792 verbose(env, "callee:\n");
8793 print_verifier_state(env, callee, true);
8798 free_func_state(callee);
8799 state->frame[state->curframe + 1] = NULL;
8803 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
8804 struct bpf_func_state *caller,
8805 struct bpf_func_state *callee)
8807 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
8808 * void *callback_ctx, u64 flags);
8809 * callback_fn(struct bpf_map *map, void *key, void *value,
8810 * void *callback_ctx);
8812 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
8814 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
8815 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
8816 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
8818 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
8819 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
8820 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
8822 /* pointer to stack or null */
8823 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
8826 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
8830 static int set_callee_state(struct bpf_verifier_env *env,
8831 struct bpf_func_state *caller,
8832 struct bpf_func_state *callee, int insn_idx)
8836 /* copy r1 - r5 args that callee can access. The copy includes parent
8837 * pointers, which connects us up to the liveness chain
8839 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
8840 callee->regs[i] = caller->regs[i];
8844 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
8847 int subprog, target_insn;
8849 target_insn = *insn_idx + insn->imm + 1;
8850 subprog = find_subprog(env, target_insn);
8852 verbose(env, "verifier bug. No program starts at insn %d\n",
8857 return __check_func_call(env, insn, insn_idx, subprog, set_callee_state);
8860 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
8861 struct bpf_func_state *caller,
8862 struct bpf_func_state *callee,
8865 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
8866 struct bpf_map *map;
8869 if (bpf_map_ptr_poisoned(insn_aux)) {
8870 verbose(env, "tail_call abusing map_ptr\n");
8874 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
8875 if (!map->ops->map_set_for_each_callback_args ||
8876 !map->ops->map_for_each_callback) {
8877 verbose(env, "callback function not allowed for map\n");
8881 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
8885 callee->in_callback_fn = true;
8886 callee->callback_ret_range = tnum_range(0, 1);
8890 static int set_loop_callback_state(struct bpf_verifier_env *env,
8891 struct bpf_func_state *caller,
8892 struct bpf_func_state *callee,
8895 /* bpf_loop(u32 nr_loops, void *callback_fn, void *callback_ctx,
8897 * callback_fn(u32 index, void *callback_ctx);
8899 callee->regs[BPF_REG_1].type = SCALAR_VALUE;
8900 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
8903 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
8904 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
8905 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
8907 callee->in_callback_fn = true;
8908 callee->callback_ret_range = tnum_range(0, 1);
8912 static int set_timer_callback_state(struct bpf_verifier_env *env,
8913 struct bpf_func_state *caller,
8914 struct bpf_func_state *callee,
8917 struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
8919 /* bpf_timer_set_callback(struct bpf_timer *timer, void *callback_fn);
8920 * callback_fn(struct bpf_map *map, void *key, void *value);
8922 callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP;
8923 __mark_reg_known_zero(&callee->regs[BPF_REG_1]);
8924 callee->regs[BPF_REG_1].map_ptr = map_ptr;
8926 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
8927 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
8928 callee->regs[BPF_REG_2].map_ptr = map_ptr;
8930 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
8931 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
8932 callee->regs[BPF_REG_3].map_ptr = map_ptr;
8935 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
8936 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
8937 callee->in_async_callback_fn = true;
8938 callee->callback_ret_range = tnum_range(0, 1);
8942 static int set_find_vma_callback_state(struct bpf_verifier_env *env,
8943 struct bpf_func_state *caller,
8944 struct bpf_func_state *callee,
8947 /* bpf_find_vma(struct task_struct *task, u64 addr,
8948 * void *callback_fn, void *callback_ctx, u64 flags)
8949 * (callback_fn)(struct task_struct *task,
8950 * struct vm_area_struct *vma, void *callback_ctx);
8952 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
8954 callee->regs[BPF_REG_2].type = PTR_TO_BTF_ID;
8955 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
8956 callee->regs[BPF_REG_2].btf = btf_vmlinux;
8957 callee->regs[BPF_REG_2].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_VMA],
8959 /* pointer to stack or null */
8960 callee->regs[BPF_REG_3] = caller->regs[BPF_REG_4];
8963 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
8964 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
8965 callee->in_callback_fn = true;
8966 callee->callback_ret_range = tnum_range(0, 1);
8970 static int set_user_ringbuf_callback_state(struct bpf_verifier_env *env,
8971 struct bpf_func_state *caller,
8972 struct bpf_func_state *callee,
8975 /* bpf_user_ringbuf_drain(struct bpf_map *map, void *callback_fn, void
8976 * callback_ctx, u64 flags);
8977 * callback_fn(const struct bpf_dynptr_t* dynptr, void *callback_ctx);
8979 __mark_reg_not_init(env, &callee->regs[BPF_REG_0]);
8980 mark_dynptr_cb_reg(env, &callee->regs[BPF_REG_1], BPF_DYNPTR_TYPE_LOCAL);
8981 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
8984 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
8985 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
8986 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
8988 callee->in_callback_fn = true;
8989 callee->callback_ret_range = tnum_range(0, 1);
8993 static int set_rbtree_add_callback_state(struct bpf_verifier_env *env,
8994 struct bpf_func_state *caller,
8995 struct bpf_func_state *callee,
8998 /* void bpf_rbtree_add_impl(struct bpf_rb_root *root, struct bpf_rb_node *node,
8999 * bool (less)(struct bpf_rb_node *a, const struct bpf_rb_node *b));
9001 * 'struct bpf_rb_node *node' arg to bpf_rbtree_add_impl is the same PTR_TO_BTF_ID w/ offset
9002 * that 'less' callback args will be receiving. However, 'node' arg was release_reference'd
9003 * by this point, so look at 'root'
9005 struct btf_field *field;
9007 field = reg_find_field_offset(&caller->regs[BPF_REG_1], caller->regs[BPF_REG_1].off,
9009 if (!field || !field->graph_root.value_btf_id)
9012 mark_reg_graph_node(callee->regs, BPF_REG_1, &field->graph_root);
9013 ref_set_non_owning(env, &callee->regs[BPF_REG_1]);
9014 mark_reg_graph_node(callee->regs, BPF_REG_2, &field->graph_root);
9015 ref_set_non_owning(env, &callee->regs[BPF_REG_2]);
9017 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
9018 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9019 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9020 callee->in_callback_fn = true;
9021 callee->callback_ret_range = tnum_range(0, 1);
9025 static bool is_rbtree_lock_required_kfunc(u32 btf_id);
9027 /* Are we currently verifying the callback for a rbtree helper that must
9028 * be called with lock held? If so, no need to complain about unreleased
9031 static bool in_rbtree_lock_required_cb(struct bpf_verifier_env *env)
9033 struct bpf_verifier_state *state = env->cur_state;
9034 struct bpf_insn *insn = env->prog->insnsi;
9035 struct bpf_func_state *callee;
9038 if (!state->curframe)
9041 callee = state->frame[state->curframe];
9043 if (!callee->in_callback_fn)
9046 kfunc_btf_id = insn[callee->callsite].imm;
9047 return is_rbtree_lock_required_kfunc(kfunc_btf_id);
9050 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
9052 struct bpf_verifier_state *state = env->cur_state;
9053 struct bpf_func_state *caller, *callee;
9054 struct bpf_reg_state *r0;
9057 callee = state->frame[state->curframe];
9058 r0 = &callee->regs[BPF_REG_0];
9059 if (r0->type == PTR_TO_STACK) {
9060 /* technically it's ok to return caller's stack pointer
9061 * (or caller's caller's pointer) back to the caller,
9062 * since these pointers are valid. Only current stack
9063 * pointer will be invalid as soon as function exits,
9064 * but let's be conservative
9066 verbose(env, "cannot return stack pointer to the caller\n");
9070 caller = state->frame[state->curframe - 1];
9071 if (callee->in_callback_fn) {
9072 /* enforce R0 return value range [0, 1]. */
9073 struct tnum range = callee->callback_ret_range;
9075 if (r0->type != SCALAR_VALUE) {
9076 verbose(env, "R0 not a scalar value\n");
9079 if (!tnum_in(range, r0->var_off)) {
9080 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
9084 /* return to the caller whatever r0 had in the callee */
9085 caller->regs[BPF_REG_0] = *r0;
9088 /* callback_fn frame should have released its own additions to parent's
9089 * reference state at this point, or check_reference_leak would
9090 * complain, hence it must be the same as the caller. There is no need
9093 if (!callee->in_callback_fn) {
9094 /* Transfer references to the caller */
9095 err = copy_reference_state(caller, callee);
9100 *insn_idx = callee->callsite + 1;
9101 if (env->log.level & BPF_LOG_LEVEL) {
9102 verbose(env, "returning from callee:\n");
9103 print_verifier_state(env, callee, true);
9104 verbose(env, "to caller at %d:\n", *insn_idx);
9105 print_verifier_state(env, caller, true);
9107 /* clear everything in the callee */
9108 free_func_state(callee);
9109 state->frame[state->curframe--] = NULL;
9113 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
9115 struct bpf_call_arg_meta *meta)
9117 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
9119 if (ret_type != RET_INTEGER ||
9120 (func_id != BPF_FUNC_get_stack &&
9121 func_id != BPF_FUNC_get_task_stack &&
9122 func_id != BPF_FUNC_probe_read_str &&
9123 func_id != BPF_FUNC_probe_read_kernel_str &&
9124 func_id != BPF_FUNC_probe_read_user_str))
9127 ret_reg->smax_value = meta->msize_max_value;
9128 ret_reg->s32_max_value = meta->msize_max_value;
9129 ret_reg->smin_value = -MAX_ERRNO;
9130 ret_reg->s32_min_value = -MAX_ERRNO;
9131 reg_bounds_sync(ret_reg);
9135 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
9136 int func_id, int insn_idx)
9138 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
9139 struct bpf_map *map = meta->map_ptr;
9141 if (func_id != BPF_FUNC_tail_call &&
9142 func_id != BPF_FUNC_map_lookup_elem &&
9143 func_id != BPF_FUNC_map_update_elem &&
9144 func_id != BPF_FUNC_map_delete_elem &&
9145 func_id != BPF_FUNC_map_push_elem &&
9146 func_id != BPF_FUNC_map_pop_elem &&
9147 func_id != BPF_FUNC_map_peek_elem &&
9148 func_id != BPF_FUNC_for_each_map_elem &&
9149 func_id != BPF_FUNC_redirect_map &&
9150 func_id != BPF_FUNC_map_lookup_percpu_elem)
9154 verbose(env, "kernel subsystem misconfigured verifier\n");
9158 /* In case of read-only, some additional restrictions
9159 * need to be applied in order to prevent altering the
9160 * state of the map from program side.
9162 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
9163 (func_id == BPF_FUNC_map_delete_elem ||
9164 func_id == BPF_FUNC_map_update_elem ||
9165 func_id == BPF_FUNC_map_push_elem ||
9166 func_id == BPF_FUNC_map_pop_elem)) {
9167 verbose(env, "write into map forbidden\n");
9171 if (!BPF_MAP_PTR(aux->map_ptr_state))
9172 bpf_map_ptr_store(aux, meta->map_ptr,
9173 !meta->map_ptr->bypass_spec_v1);
9174 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
9175 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
9176 !meta->map_ptr->bypass_spec_v1);
9181 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
9182 int func_id, int insn_idx)
9184 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
9185 struct bpf_reg_state *regs = cur_regs(env), *reg;
9186 struct bpf_map *map = meta->map_ptr;
9190 if (func_id != BPF_FUNC_tail_call)
9192 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
9193 verbose(env, "kernel subsystem misconfigured verifier\n");
9197 reg = ®s[BPF_REG_3];
9198 val = reg->var_off.value;
9199 max = map->max_entries;
9201 if (!(register_is_const(reg) && val < max)) {
9202 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
9206 err = mark_chain_precision(env, BPF_REG_3);
9209 if (bpf_map_key_unseen(aux))
9210 bpf_map_key_store(aux, val);
9211 else if (!bpf_map_key_poisoned(aux) &&
9212 bpf_map_key_immediate(aux) != val)
9213 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
9217 static int check_reference_leak(struct bpf_verifier_env *env)
9219 struct bpf_func_state *state = cur_func(env);
9220 bool refs_lingering = false;
9223 if (state->frameno && !state->in_callback_fn)
9226 for (i = 0; i < state->acquired_refs; i++) {
9227 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno)
9229 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
9230 state->refs[i].id, state->refs[i].insn_idx);
9231 refs_lingering = true;
9233 return refs_lingering ? -EINVAL : 0;
9236 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
9237 struct bpf_reg_state *regs)
9239 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
9240 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
9241 struct bpf_map *fmt_map = fmt_reg->map_ptr;
9242 struct bpf_bprintf_data data = {};
9243 int err, fmt_map_off, num_args;
9247 /* data must be an array of u64 */
9248 if (data_len_reg->var_off.value % 8)
9250 num_args = data_len_reg->var_off.value / 8;
9252 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
9253 * and map_direct_value_addr is set.
9255 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
9256 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
9259 verbose(env, "verifier bug\n");
9262 fmt = (char *)(long)fmt_addr + fmt_map_off;
9264 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
9265 * can focus on validating the format specifiers.
9267 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, num_args, &data);
9269 verbose(env, "Invalid format string\n");
9274 static int check_get_func_ip(struct bpf_verifier_env *env)
9276 enum bpf_prog_type type = resolve_prog_type(env->prog);
9277 int func_id = BPF_FUNC_get_func_ip;
9279 if (type == BPF_PROG_TYPE_TRACING) {
9280 if (!bpf_prog_has_trampoline(env->prog)) {
9281 verbose(env, "func %s#%d supported only for fentry/fexit/fmod_ret programs\n",
9282 func_id_name(func_id), func_id);
9286 } else if (type == BPF_PROG_TYPE_KPROBE) {
9290 verbose(env, "func %s#%d not supported for program type %d\n",
9291 func_id_name(func_id), func_id, type);
9295 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
9297 return &env->insn_aux_data[env->insn_idx];
9300 static bool loop_flag_is_zero(struct bpf_verifier_env *env)
9302 struct bpf_reg_state *regs = cur_regs(env);
9303 struct bpf_reg_state *reg = ®s[BPF_REG_4];
9304 bool reg_is_null = register_is_null(reg);
9307 mark_chain_precision(env, BPF_REG_4);
9312 static void update_loop_inline_state(struct bpf_verifier_env *env, u32 subprogno)
9314 struct bpf_loop_inline_state *state = &cur_aux(env)->loop_inline_state;
9316 if (!state->initialized) {
9317 state->initialized = 1;
9318 state->fit_for_inline = loop_flag_is_zero(env);
9319 state->callback_subprogno = subprogno;
9323 if (!state->fit_for_inline)
9326 state->fit_for_inline = (loop_flag_is_zero(env) &&
9327 state->callback_subprogno == subprogno);
9330 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
9333 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
9334 const struct bpf_func_proto *fn = NULL;
9335 enum bpf_return_type ret_type;
9336 enum bpf_type_flag ret_flag;
9337 struct bpf_reg_state *regs;
9338 struct bpf_call_arg_meta meta;
9339 int insn_idx = *insn_idx_p;
9341 int i, err, func_id;
9343 /* find function prototype */
9344 func_id = insn->imm;
9345 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
9346 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
9351 if (env->ops->get_func_proto)
9352 fn = env->ops->get_func_proto(func_id, env->prog);
9354 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
9359 /* eBPF programs must be GPL compatible to use GPL-ed functions */
9360 if (!env->prog->gpl_compatible && fn->gpl_only) {
9361 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
9365 if (fn->allowed && !fn->allowed(env->prog)) {
9366 verbose(env, "helper call is not allowed in probe\n");
9370 if (!env->prog->aux->sleepable && fn->might_sleep) {
9371 verbose(env, "helper call might sleep in a non-sleepable prog\n");
9375 /* With LD_ABS/IND some JITs save/restore skb from r1. */
9376 changes_data = bpf_helper_changes_pkt_data(fn->func);
9377 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
9378 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
9379 func_id_name(func_id), func_id);
9383 memset(&meta, 0, sizeof(meta));
9384 meta.pkt_access = fn->pkt_access;
9386 err = check_func_proto(fn, func_id);
9388 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
9389 func_id_name(func_id), func_id);
9393 if (env->cur_state->active_rcu_lock) {
9394 if (fn->might_sleep) {
9395 verbose(env, "sleepable helper %s#%d in rcu_read_lock region\n",
9396 func_id_name(func_id), func_id);
9400 if (env->prog->aux->sleepable && is_storage_get_function(func_id))
9401 env->insn_aux_data[insn_idx].storage_get_func_atomic = true;
9404 meta.func_id = func_id;
9406 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
9407 err = check_func_arg(env, i, &meta, fn, insn_idx);
9412 err = record_func_map(env, &meta, func_id, insn_idx);
9416 err = record_func_key(env, &meta, func_id, insn_idx);
9420 /* Mark slots with STACK_MISC in case of raw mode, stack offset
9421 * is inferred from register state.
9423 for (i = 0; i < meta.access_size; i++) {
9424 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
9425 BPF_WRITE, -1, false);
9430 regs = cur_regs(env);
9432 if (meta.release_regno) {
9434 /* This can only be set for PTR_TO_STACK, as CONST_PTR_TO_DYNPTR cannot
9435 * be released by any dynptr helper. Hence, unmark_stack_slots_dynptr
9436 * is safe to do directly.
9438 if (arg_type_is_dynptr(fn->arg_type[meta.release_regno - BPF_REG_1])) {
9439 if (regs[meta.release_regno].type == CONST_PTR_TO_DYNPTR) {
9440 verbose(env, "verifier internal error: CONST_PTR_TO_DYNPTR cannot be released\n");
9443 err = unmark_stack_slots_dynptr(env, ®s[meta.release_regno]);
9444 } else if (meta.ref_obj_id) {
9445 err = release_reference(env, meta.ref_obj_id);
9446 } else if (register_is_null(®s[meta.release_regno])) {
9447 /* meta.ref_obj_id can only be 0 if register that is meant to be
9448 * released is NULL, which must be > R0.
9453 verbose(env, "func %s#%d reference has not been acquired before\n",
9454 func_id_name(func_id), func_id);
9460 case BPF_FUNC_tail_call:
9461 err = check_reference_leak(env);
9463 verbose(env, "tail_call would lead to reference leak\n");
9467 case BPF_FUNC_get_local_storage:
9468 /* check that flags argument in get_local_storage(map, flags) is 0,
9469 * this is required because get_local_storage() can't return an error.
9471 if (!register_is_null(®s[BPF_REG_2])) {
9472 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
9476 case BPF_FUNC_for_each_map_elem:
9477 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9478 set_map_elem_callback_state);
9480 case BPF_FUNC_timer_set_callback:
9481 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9482 set_timer_callback_state);
9484 case BPF_FUNC_find_vma:
9485 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9486 set_find_vma_callback_state);
9488 case BPF_FUNC_snprintf:
9489 err = check_bpf_snprintf_call(env, regs);
9492 update_loop_inline_state(env, meta.subprogno);
9493 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9494 set_loop_callback_state);
9496 case BPF_FUNC_dynptr_from_mem:
9497 if (regs[BPF_REG_1].type != PTR_TO_MAP_VALUE) {
9498 verbose(env, "Unsupported reg type %s for bpf_dynptr_from_mem data\n",
9499 reg_type_str(env, regs[BPF_REG_1].type));
9503 case BPF_FUNC_set_retval:
9504 if (prog_type == BPF_PROG_TYPE_LSM &&
9505 env->prog->expected_attach_type == BPF_LSM_CGROUP) {
9506 if (!env->prog->aux->attach_func_proto->type) {
9507 /* Make sure programs that attach to void
9508 * hooks don't try to modify return value.
9510 verbose(env, "BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
9515 case BPF_FUNC_dynptr_data:
9517 struct bpf_reg_state *reg;
9520 reg = get_dynptr_arg_reg(env, fn, regs);
9525 if (meta.dynptr_id) {
9526 verbose(env, "verifier internal error: meta.dynptr_id already set\n");
9529 if (meta.ref_obj_id) {
9530 verbose(env, "verifier internal error: meta.ref_obj_id already set\n");
9534 id = dynptr_id(env, reg);
9536 verbose(env, "verifier internal error: failed to obtain dynptr id\n");
9540 ref_obj_id = dynptr_ref_obj_id(env, reg);
9541 if (ref_obj_id < 0) {
9542 verbose(env, "verifier internal error: failed to obtain dynptr ref_obj_id\n");
9546 meta.dynptr_id = id;
9547 meta.ref_obj_id = ref_obj_id;
9551 case BPF_FUNC_dynptr_write:
9553 enum bpf_dynptr_type dynptr_type;
9554 struct bpf_reg_state *reg;
9556 reg = get_dynptr_arg_reg(env, fn, regs);
9560 dynptr_type = dynptr_get_type(env, reg);
9561 if (dynptr_type == BPF_DYNPTR_TYPE_INVALID)
9564 if (dynptr_type == BPF_DYNPTR_TYPE_SKB)
9565 /* this will trigger clear_all_pkt_pointers(), which will
9566 * invalidate all dynptr slices associated with the skb
9568 changes_data = true;
9572 case BPF_FUNC_user_ringbuf_drain:
9573 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9574 set_user_ringbuf_callback_state);
9581 /* reset caller saved regs */
9582 for (i = 0; i < CALLER_SAVED_REGS; i++) {
9583 mark_reg_not_init(env, regs, caller_saved[i]);
9584 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
9587 /* helper call returns 64-bit value. */
9588 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
9590 /* update return register (already marked as written above) */
9591 ret_type = fn->ret_type;
9592 ret_flag = type_flag(ret_type);
9594 switch (base_type(ret_type)) {
9596 /* sets type to SCALAR_VALUE */
9597 mark_reg_unknown(env, regs, BPF_REG_0);
9600 regs[BPF_REG_0].type = NOT_INIT;
9602 case RET_PTR_TO_MAP_VALUE:
9603 /* There is no offset yet applied, variable or fixed */
9604 mark_reg_known_zero(env, regs, BPF_REG_0);
9605 /* remember map_ptr, so that check_map_access()
9606 * can check 'value_size' boundary of memory access
9607 * to map element returned from bpf_map_lookup_elem()
9609 if (meta.map_ptr == NULL) {
9611 "kernel subsystem misconfigured verifier\n");
9614 regs[BPF_REG_0].map_ptr = meta.map_ptr;
9615 regs[BPF_REG_0].map_uid = meta.map_uid;
9616 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE | ret_flag;
9617 if (!type_may_be_null(ret_type) &&
9618 btf_record_has_field(meta.map_ptr->record, BPF_SPIN_LOCK)) {
9619 regs[BPF_REG_0].id = ++env->id_gen;
9622 case RET_PTR_TO_SOCKET:
9623 mark_reg_known_zero(env, regs, BPF_REG_0);
9624 regs[BPF_REG_0].type = PTR_TO_SOCKET | ret_flag;
9626 case RET_PTR_TO_SOCK_COMMON:
9627 mark_reg_known_zero(env, regs, BPF_REG_0);
9628 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON | ret_flag;
9630 case RET_PTR_TO_TCP_SOCK:
9631 mark_reg_known_zero(env, regs, BPF_REG_0);
9632 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK | ret_flag;
9634 case RET_PTR_TO_MEM:
9635 mark_reg_known_zero(env, regs, BPF_REG_0);
9636 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
9637 regs[BPF_REG_0].mem_size = meta.mem_size;
9639 case RET_PTR_TO_MEM_OR_BTF_ID:
9641 const struct btf_type *t;
9643 mark_reg_known_zero(env, regs, BPF_REG_0);
9644 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
9645 if (!btf_type_is_struct(t)) {
9647 const struct btf_type *ret;
9650 /* resolve the type size of ksym. */
9651 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
9653 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
9654 verbose(env, "unable to resolve the size of type '%s': %ld\n",
9655 tname, PTR_ERR(ret));
9658 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
9659 regs[BPF_REG_0].mem_size = tsize;
9661 /* MEM_RDONLY may be carried from ret_flag, but it
9662 * doesn't apply on PTR_TO_BTF_ID. Fold it, otherwise
9663 * it will confuse the check of PTR_TO_BTF_ID in
9664 * check_mem_access().
9666 ret_flag &= ~MEM_RDONLY;
9668 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
9669 regs[BPF_REG_0].btf = meta.ret_btf;
9670 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
9674 case RET_PTR_TO_BTF_ID:
9676 struct btf *ret_btf;
9679 mark_reg_known_zero(env, regs, BPF_REG_0);
9680 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
9681 if (func_id == BPF_FUNC_kptr_xchg) {
9682 ret_btf = meta.kptr_field->kptr.btf;
9683 ret_btf_id = meta.kptr_field->kptr.btf_id;
9684 if (!btf_is_kernel(ret_btf))
9685 regs[BPF_REG_0].type |= MEM_ALLOC;
9687 if (fn->ret_btf_id == BPF_PTR_POISON) {
9688 verbose(env, "verifier internal error:");
9689 verbose(env, "func %s has non-overwritten BPF_PTR_POISON return type\n",
9690 func_id_name(func_id));
9693 ret_btf = btf_vmlinux;
9694 ret_btf_id = *fn->ret_btf_id;
9696 if (ret_btf_id == 0) {
9697 verbose(env, "invalid return type %u of func %s#%d\n",
9698 base_type(ret_type), func_id_name(func_id),
9702 regs[BPF_REG_0].btf = ret_btf;
9703 regs[BPF_REG_0].btf_id = ret_btf_id;
9707 verbose(env, "unknown return type %u of func %s#%d\n",
9708 base_type(ret_type), func_id_name(func_id), func_id);
9712 if (type_may_be_null(regs[BPF_REG_0].type))
9713 regs[BPF_REG_0].id = ++env->id_gen;
9715 if (helper_multiple_ref_obj_use(func_id, meta.map_ptr)) {
9716 verbose(env, "verifier internal error: func %s#%d sets ref_obj_id more than once\n",
9717 func_id_name(func_id), func_id);
9721 if (is_dynptr_ref_function(func_id))
9722 regs[BPF_REG_0].dynptr_id = meta.dynptr_id;
9724 if (is_ptr_cast_function(func_id) || is_dynptr_ref_function(func_id)) {
9725 /* For release_reference() */
9726 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
9727 } else if (is_acquire_function(func_id, meta.map_ptr)) {
9728 int id = acquire_reference_state(env, insn_idx);
9732 /* For mark_ptr_or_null_reg() */
9733 regs[BPF_REG_0].id = id;
9734 /* For release_reference() */
9735 regs[BPF_REG_0].ref_obj_id = id;
9738 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
9740 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
9744 if ((func_id == BPF_FUNC_get_stack ||
9745 func_id == BPF_FUNC_get_task_stack) &&
9746 !env->prog->has_callchain_buf) {
9747 const char *err_str;
9749 #ifdef CONFIG_PERF_EVENTS
9750 err = get_callchain_buffers(sysctl_perf_event_max_stack);
9751 err_str = "cannot get callchain buffer for func %s#%d\n";
9754 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
9757 verbose(env, err_str, func_id_name(func_id), func_id);
9761 env->prog->has_callchain_buf = true;
9764 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
9765 env->prog->call_get_stack = true;
9767 if (func_id == BPF_FUNC_get_func_ip) {
9768 if (check_get_func_ip(env))
9770 env->prog->call_get_func_ip = true;
9774 clear_all_pkt_pointers(env);
9778 /* mark_btf_func_reg_size() is used when the reg size is determined by
9779 * the BTF func_proto's return value size and argument.
9781 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
9784 struct bpf_reg_state *reg = &cur_regs(env)[regno];
9786 if (regno == BPF_REG_0) {
9787 /* Function return value */
9788 reg->live |= REG_LIVE_WRITTEN;
9789 reg->subreg_def = reg_size == sizeof(u64) ?
9790 DEF_NOT_SUBREG : env->insn_idx + 1;
9792 /* Function argument */
9793 if (reg_size == sizeof(u64)) {
9794 mark_insn_zext(env, reg);
9795 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
9797 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
9802 static bool is_kfunc_acquire(struct bpf_kfunc_call_arg_meta *meta)
9804 return meta->kfunc_flags & KF_ACQUIRE;
9807 static bool is_kfunc_release(struct bpf_kfunc_call_arg_meta *meta)
9809 return meta->kfunc_flags & KF_RELEASE;
9812 static bool is_kfunc_trusted_args(struct bpf_kfunc_call_arg_meta *meta)
9814 return (meta->kfunc_flags & KF_TRUSTED_ARGS) || is_kfunc_release(meta);
9817 static bool is_kfunc_sleepable(struct bpf_kfunc_call_arg_meta *meta)
9819 return meta->kfunc_flags & KF_SLEEPABLE;
9822 static bool is_kfunc_destructive(struct bpf_kfunc_call_arg_meta *meta)
9824 return meta->kfunc_flags & KF_DESTRUCTIVE;
9827 static bool is_kfunc_rcu(struct bpf_kfunc_call_arg_meta *meta)
9829 return meta->kfunc_flags & KF_RCU;
9832 static bool __kfunc_param_match_suffix(const struct btf *btf,
9833 const struct btf_param *arg,
9836 int suffix_len = strlen(suffix), len;
9837 const char *param_name;
9839 /* In the future, this can be ported to use BTF tagging */
9840 param_name = btf_name_by_offset(btf, arg->name_off);
9841 if (str_is_empty(param_name))
9843 len = strlen(param_name);
9844 if (len < suffix_len)
9846 param_name += len - suffix_len;
9847 return !strncmp(param_name, suffix, suffix_len);
9850 static bool is_kfunc_arg_mem_size(const struct btf *btf,
9851 const struct btf_param *arg,
9852 const struct bpf_reg_state *reg)
9854 const struct btf_type *t;
9856 t = btf_type_skip_modifiers(btf, arg->type, NULL);
9857 if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
9860 return __kfunc_param_match_suffix(btf, arg, "__sz");
9863 static bool is_kfunc_arg_const_mem_size(const struct btf *btf,
9864 const struct btf_param *arg,
9865 const struct bpf_reg_state *reg)
9867 const struct btf_type *t;
9869 t = btf_type_skip_modifiers(btf, arg->type, NULL);
9870 if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
9873 return __kfunc_param_match_suffix(btf, arg, "__szk");
9876 static bool is_kfunc_arg_optional(const struct btf *btf, const struct btf_param *arg)
9878 return __kfunc_param_match_suffix(btf, arg, "__opt");
9881 static bool is_kfunc_arg_constant(const struct btf *btf, const struct btf_param *arg)
9883 return __kfunc_param_match_suffix(btf, arg, "__k");
9886 static bool is_kfunc_arg_ignore(const struct btf *btf, const struct btf_param *arg)
9888 return __kfunc_param_match_suffix(btf, arg, "__ign");
9891 static bool is_kfunc_arg_alloc_obj(const struct btf *btf, const struct btf_param *arg)
9893 return __kfunc_param_match_suffix(btf, arg, "__alloc");
9896 static bool is_kfunc_arg_uninit(const struct btf *btf, const struct btf_param *arg)
9898 return __kfunc_param_match_suffix(btf, arg, "__uninit");
9901 static bool is_kfunc_arg_refcounted_kptr(const struct btf *btf, const struct btf_param *arg)
9903 return __kfunc_param_match_suffix(btf, arg, "__refcounted_kptr");
9906 static bool is_kfunc_arg_scalar_with_name(const struct btf *btf,
9907 const struct btf_param *arg,
9910 int len, target_len = strlen(name);
9911 const char *param_name;
9913 param_name = btf_name_by_offset(btf, arg->name_off);
9914 if (str_is_empty(param_name))
9916 len = strlen(param_name);
9917 if (len != target_len)
9919 if (strcmp(param_name, name))
9927 KF_ARG_LIST_HEAD_ID,
9928 KF_ARG_LIST_NODE_ID,
9933 BTF_ID_LIST(kf_arg_btf_ids)
9934 BTF_ID(struct, bpf_dynptr_kern)
9935 BTF_ID(struct, bpf_list_head)
9936 BTF_ID(struct, bpf_list_node)
9937 BTF_ID(struct, bpf_rb_root)
9938 BTF_ID(struct, bpf_rb_node)
9940 static bool __is_kfunc_ptr_arg_type(const struct btf *btf,
9941 const struct btf_param *arg, int type)
9943 const struct btf_type *t;
9946 t = btf_type_skip_modifiers(btf, arg->type, NULL);
9949 if (!btf_type_is_ptr(t))
9951 t = btf_type_skip_modifiers(btf, t->type, &res_id);
9954 return btf_types_are_same(btf, res_id, btf_vmlinux, kf_arg_btf_ids[type]);
9957 static bool is_kfunc_arg_dynptr(const struct btf *btf, const struct btf_param *arg)
9959 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_DYNPTR_ID);
9962 static bool is_kfunc_arg_list_head(const struct btf *btf, const struct btf_param *arg)
9964 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_HEAD_ID);
9967 static bool is_kfunc_arg_list_node(const struct btf *btf, const struct btf_param *arg)
9969 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_NODE_ID);
9972 static bool is_kfunc_arg_rbtree_root(const struct btf *btf, const struct btf_param *arg)
9974 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RB_ROOT_ID);
9977 static bool is_kfunc_arg_rbtree_node(const struct btf *btf, const struct btf_param *arg)
9979 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RB_NODE_ID);
9982 static bool is_kfunc_arg_callback(struct bpf_verifier_env *env, const struct btf *btf,
9983 const struct btf_param *arg)
9985 const struct btf_type *t;
9987 t = btf_type_resolve_func_ptr(btf, arg->type, NULL);
9994 /* Returns true if struct is composed of scalars, 4 levels of nesting allowed */
9995 static bool __btf_type_is_scalar_struct(struct bpf_verifier_env *env,
9996 const struct btf *btf,
9997 const struct btf_type *t, int rec)
9999 const struct btf_type *member_type;
10000 const struct btf_member *member;
10003 if (!btf_type_is_struct(t))
10006 for_each_member(i, t, member) {
10007 const struct btf_array *array;
10009 member_type = btf_type_skip_modifiers(btf, member->type, NULL);
10010 if (btf_type_is_struct(member_type)) {
10012 verbose(env, "max struct nesting depth exceeded\n");
10015 if (!__btf_type_is_scalar_struct(env, btf, member_type, rec + 1))
10019 if (btf_type_is_array(member_type)) {
10020 array = btf_array(member_type);
10021 if (!array->nelems)
10023 member_type = btf_type_skip_modifiers(btf, array->type, NULL);
10024 if (!btf_type_is_scalar(member_type))
10028 if (!btf_type_is_scalar(member_type))
10035 static u32 *reg2btf_ids[__BPF_REG_TYPE_MAX] = {
10037 [PTR_TO_SOCKET] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK],
10038 [PTR_TO_SOCK_COMMON] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
10039 [PTR_TO_TCP_SOCK] = &btf_sock_ids[BTF_SOCK_TYPE_TCP],
10043 enum kfunc_ptr_arg_type {
10045 KF_ARG_PTR_TO_ALLOC_BTF_ID, /* Allocated object */
10046 KF_ARG_PTR_TO_REFCOUNTED_KPTR, /* Refcounted local kptr */
10047 KF_ARG_PTR_TO_DYNPTR,
10048 KF_ARG_PTR_TO_ITER,
10049 KF_ARG_PTR_TO_LIST_HEAD,
10050 KF_ARG_PTR_TO_LIST_NODE,
10051 KF_ARG_PTR_TO_BTF_ID, /* Also covers reg2btf_ids conversions */
10053 KF_ARG_PTR_TO_MEM_SIZE, /* Size derived from next argument, skip it */
10054 KF_ARG_PTR_TO_CALLBACK,
10055 KF_ARG_PTR_TO_RB_ROOT,
10056 KF_ARG_PTR_TO_RB_NODE,
10059 enum special_kfunc_type {
10060 KF_bpf_obj_new_impl,
10061 KF_bpf_obj_drop_impl,
10062 KF_bpf_refcount_acquire_impl,
10063 KF_bpf_list_push_front_impl,
10064 KF_bpf_list_push_back_impl,
10065 KF_bpf_list_pop_front,
10066 KF_bpf_list_pop_back,
10067 KF_bpf_cast_to_kern_ctx,
10068 KF_bpf_rdonly_cast,
10069 KF_bpf_rcu_read_lock,
10070 KF_bpf_rcu_read_unlock,
10071 KF_bpf_rbtree_remove,
10072 KF_bpf_rbtree_add_impl,
10073 KF_bpf_rbtree_first,
10074 KF_bpf_dynptr_from_skb,
10075 KF_bpf_dynptr_from_xdp,
10076 KF_bpf_dynptr_slice,
10077 KF_bpf_dynptr_slice_rdwr,
10078 KF_bpf_dynptr_clone,
10081 BTF_SET_START(special_kfunc_set)
10082 BTF_ID(func, bpf_obj_new_impl)
10083 BTF_ID(func, bpf_obj_drop_impl)
10084 BTF_ID(func, bpf_refcount_acquire_impl)
10085 BTF_ID(func, bpf_list_push_front_impl)
10086 BTF_ID(func, bpf_list_push_back_impl)
10087 BTF_ID(func, bpf_list_pop_front)
10088 BTF_ID(func, bpf_list_pop_back)
10089 BTF_ID(func, bpf_cast_to_kern_ctx)
10090 BTF_ID(func, bpf_rdonly_cast)
10091 BTF_ID(func, bpf_rbtree_remove)
10092 BTF_ID(func, bpf_rbtree_add_impl)
10093 BTF_ID(func, bpf_rbtree_first)
10094 BTF_ID(func, bpf_dynptr_from_skb)
10095 BTF_ID(func, bpf_dynptr_from_xdp)
10096 BTF_ID(func, bpf_dynptr_slice)
10097 BTF_ID(func, bpf_dynptr_slice_rdwr)
10098 BTF_ID(func, bpf_dynptr_clone)
10099 BTF_SET_END(special_kfunc_set)
10101 BTF_ID_LIST(special_kfunc_list)
10102 BTF_ID(func, bpf_obj_new_impl)
10103 BTF_ID(func, bpf_obj_drop_impl)
10104 BTF_ID(func, bpf_refcount_acquire_impl)
10105 BTF_ID(func, bpf_list_push_front_impl)
10106 BTF_ID(func, bpf_list_push_back_impl)
10107 BTF_ID(func, bpf_list_pop_front)
10108 BTF_ID(func, bpf_list_pop_back)
10109 BTF_ID(func, bpf_cast_to_kern_ctx)
10110 BTF_ID(func, bpf_rdonly_cast)
10111 BTF_ID(func, bpf_rcu_read_lock)
10112 BTF_ID(func, bpf_rcu_read_unlock)
10113 BTF_ID(func, bpf_rbtree_remove)
10114 BTF_ID(func, bpf_rbtree_add_impl)
10115 BTF_ID(func, bpf_rbtree_first)
10116 BTF_ID(func, bpf_dynptr_from_skb)
10117 BTF_ID(func, bpf_dynptr_from_xdp)
10118 BTF_ID(func, bpf_dynptr_slice)
10119 BTF_ID(func, bpf_dynptr_slice_rdwr)
10120 BTF_ID(func, bpf_dynptr_clone)
10122 static bool is_kfunc_ret_null(struct bpf_kfunc_call_arg_meta *meta)
10124 if (meta->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl] &&
10125 meta->arg_owning_ref) {
10129 return meta->kfunc_flags & KF_RET_NULL;
10132 static bool is_kfunc_bpf_rcu_read_lock(struct bpf_kfunc_call_arg_meta *meta)
10134 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_lock];
10137 static bool is_kfunc_bpf_rcu_read_unlock(struct bpf_kfunc_call_arg_meta *meta)
10139 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_unlock];
10142 static enum kfunc_ptr_arg_type
10143 get_kfunc_ptr_arg_type(struct bpf_verifier_env *env,
10144 struct bpf_kfunc_call_arg_meta *meta,
10145 const struct btf_type *t, const struct btf_type *ref_t,
10146 const char *ref_tname, const struct btf_param *args,
10147 int argno, int nargs)
10149 u32 regno = argno + 1;
10150 struct bpf_reg_state *regs = cur_regs(env);
10151 struct bpf_reg_state *reg = ®s[regno];
10152 bool arg_mem_size = false;
10154 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx])
10155 return KF_ARG_PTR_TO_CTX;
10157 /* In this function, we verify the kfunc's BTF as per the argument type,
10158 * leaving the rest of the verification with respect to the register
10159 * type to our caller. When a set of conditions hold in the BTF type of
10160 * arguments, we resolve it to a known kfunc_ptr_arg_type.
10162 if (btf_get_prog_ctx_type(&env->log, meta->btf, t, resolve_prog_type(env->prog), argno))
10163 return KF_ARG_PTR_TO_CTX;
10165 if (is_kfunc_arg_alloc_obj(meta->btf, &args[argno]))
10166 return KF_ARG_PTR_TO_ALLOC_BTF_ID;
10168 if (is_kfunc_arg_refcounted_kptr(meta->btf, &args[argno]))
10169 return KF_ARG_PTR_TO_REFCOUNTED_KPTR;
10171 if (is_kfunc_arg_dynptr(meta->btf, &args[argno]))
10172 return KF_ARG_PTR_TO_DYNPTR;
10174 if (is_kfunc_arg_iter(meta, argno))
10175 return KF_ARG_PTR_TO_ITER;
10177 if (is_kfunc_arg_list_head(meta->btf, &args[argno]))
10178 return KF_ARG_PTR_TO_LIST_HEAD;
10180 if (is_kfunc_arg_list_node(meta->btf, &args[argno]))
10181 return KF_ARG_PTR_TO_LIST_NODE;
10183 if (is_kfunc_arg_rbtree_root(meta->btf, &args[argno]))
10184 return KF_ARG_PTR_TO_RB_ROOT;
10186 if (is_kfunc_arg_rbtree_node(meta->btf, &args[argno]))
10187 return KF_ARG_PTR_TO_RB_NODE;
10189 if ((base_type(reg->type) == PTR_TO_BTF_ID || reg2btf_ids[base_type(reg->type)])) {
10190 if (!btf_type_is_struct(ref_t)) {
10191 verbose(env, "kernel function %s args#%d pointer type %s %s is not supported\n",
10192 meta->func_name, argno, btf_type_str(ref_t), ref_tname);
10195 return KF_ARG_PTR_TO_BTF_ID;
10198 if (is_kfunc_arg_callback(env, meta->btf, &args[argno]))
10199 return KF_ARG_PTR_TO_CALLBACK;
10202 if (argno + 1 < nargs &&
10203 (is_kfunc_arg_mem_size(meta->btf, &args[argno + 1], ®s[regno + 1]) ||
10204 is_kfunc_arg_const_mem_size(meta->btf, &args[argno + 1], ®s[regno + 1])))
10205 arg_mem_size = true;
10207 /* This is the catch all argument type of register types supported by
10208 * check_helper_mem_access. However, we only allow when argument type is
10209 * pointer to scalar, or struct composed (recursively) of scalars. When
10210 * arg_mem_size is true, the pointer can be void *.
10212 if (!btf_type_is_scalar(ref_t) && !__btf_type_is_scalar_struct(env, meta->btf, ref_t, 0) &&
10213 (arg_mem_size ? !btf_type_is_void(ref_t) : 1)) {
10214 verbose(env, "arg#%d pointer type %s %s must point to %sscalar, or struct with scalar\n",
10215 argno, btf_type_str(ref_t), ref_tname, arg_mem_size ? "void, " : "");
10218 return arg_mem_size ? KF_ARG_PTR_TO_MEM_SIZE : KF_ARG_PTR_TO_MEM;
10221 static int process_kf_arg_ptr_to_btf_id(struct bpf_verifier_env *env,
10222 struct bpf_reg_state *reg,
10223 const struct btf_type *ref_t,
10224 const char *ref_tname, u32 ref_id,
10225 struct bpf_kfunc_call_arg_meta *meta,
10228 const struct btf_type *reg_ref_t;
10229 bool strict_type_match = false;
10230 const struct btf *reg_btf;
10231 const char *reg_ref_tname;
10234 if (base_type(reg->type) == PTR_TO_BTF_ID) {
10235 reg_btf = reg->btf;
10236 reg_ref_id = reg->btf_id;
10238 reg_btf = btf_vmlinux;
10239 reg_ref_id = *reg2btf_ids[base_type(reg->type)];
10242 /* Enforce strict type matching for calls to kfuncs that are acquiring
10243 * or releasing a reference, or are no-cast aliases. We do _not_
10244 * enforce strict matching for plain KF_TRUSTED_ARGS kfuncs by default,
10245 * as we want to enable BPF programs to pass types that are bitwise
10246 * equivalent without forcing them to explicitly cast with something
10247 * like bpf_cast_to_kern_ctx().
10249 * For example, say we had a type like the following:
10251 * struct bpf_cpumask {
10252 * cpumask_t cpumask;
10253 * refcount_t usage;
10256 * Note that as specified in <linux/cpumask.h>, cpumask_t is typedef'ed
10257 * to a struct cpumask, so it would be safe to pass a struct
10258 * bpf_cpumask * to a kfunc expecting a struct cpumask *.
10260 * The philosophy here is similar to how we allow scalars of different
10261 * types to be passed to kfuncs as long as the size is the same. The
10262 * only difference here is that we're simply allowing
10263 * btf_struct_ids_match() to walk the struct at the 0th offset, and
10266 if (is_kfunc_acquire(meta) ||
10267 (is_kfunc_release(meta) && reg->ref_obj_id) ||
10268 btf_type_ids_nocast_alias(&env->log, reg_btf, reg_ref_id, meta->btf, ref_id))
10269 strict_type_match = true;
10271 WARN_ON_ONCE(is_kfunc_trusted_args(meta) && reg->off);
10273 reg_ref_t = btf_type_skip_modifiers(reg_btf, reg_ref_id, ®_ref_id);
10274 reg_ref_tname = btf_name_by_offset(reg_btf, reg_ref_t->name_off);
10275 if (!btf_struct_ids_match(&env->log, reg_btf, reg_ref_id, reg->off, meta->btf, ref_id, strict_type_match)) {
10276 verbose(env, "kernel function %s args#%d expected pointer to %s %s but R%d has a pointer to %s %s\n",
10277 meta->func_name, argno, btf_type_str(ref_t), ref_tname, argno + 1,
10278 btf_type_str(reg_ref_t), reg_ref_tname);
10284 static int ref_set_non_owning(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
10286 struct bpf_verifier_state *state = env->cur_state;
10288 if (!state->active_lock.ptr) {
10289 verbose(env, "verifier internal error: ref_set_non_owning w/o active lock\n");
10293 if (type_flag(reg->type) & NON_OWN_REF) {
10294 verbose(env, "verifier internal error: NON_OWN_REF already set\n");
10298 reg->type |= NON_OWN_REF;
10302 static int ref_convert_owning_non_owning(struct bpf_verifier_env *env, u32 ref_obj_id)
10304 struct bpf_func_state *state, *unused;
10305 struct bpf_reg_state *reg;
10308 state = cur_func(env);
10311 verbose(env, "verifier internal error: ref_obj_id is zero for "
10312 "owning -> non-owning conversion\n");
10316 for (i = 0; i < state->acquired_refs; i++) {
10317 if (state->refs[i].id != ref_obj_id)
10320 /* Clear ref_obj_id here so release_reference doesn't clobber
10323 bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
10324 if (reg->ref_obj_id == ref_obj_id) {
10325 reg->ref_obj_id = 0;
10326 ref_set_non_owning(env, reg);
10332 verbose(env, "verifier internal error: ref state missing for ref_obj_id\n");
10336 /* Implementation details:
10338 * Each register points to some region of memory, which we define as an
10339 * allocation. Each allocation may embed a bpf_spin_lock which protects any
10340 * special BPF objects (bpf_list_head, bpf_rb_root, etc.) part of the same
10341 * allocation. The lock and the data it protects are colocated in the same
10344 * Hence, everytime a register holds a pointer value pointing to such
10345 * allocation, the verifier preserves a unique reg->id for it.
10347 * The verifier remembers the lock 'ptr' and the lock 'id' whenever
10348 * bpf_spin_lock is called.
10350 * To enable this, lock state in the verifier captures two values:
10351 * active_lock.ptr = Register's type specific pointer
10352 * active_lock.id = A unique ID for each register pointer value
10354 * Currently, PTR_TO_MAP_VALUE and PTR_TO_BTF_ID | MEM_ALLOC are the two
10355 * supported register types.
10357 * The active_lock.ptr in case of map values is the reg->map_ptr, and in case of
10358 * allocated objects is the reg->btf pointer.
10360 * The active_lock.id is non-unique for maps supporting direct_value_addr, as we
10361 * can establish the provenance of the map value statically for each distinct
10362 * lookup into such maps. They always contain a single map value hence unique
10363 * IDs for each pseudo load pessimizes the algorithm and rejects valid programs.
10365 * So, in case of global variables, they use array maps with max_entries = 1,
10366 * hence their active_lock.ptr becomes map_ptr and id = 0 (since they all point
10367 * into the same map value as max_entries is 1, as described above).
10369 * In case of inner map lookups, the inner map pointer has same map_ptr as the
10370 * outer map pointer (in verifier context), but each lookup into an inner map
10371 * assigns a fresh reg->id to the lookup, so while lookups into distinct inner
10372 * maps from the same outer map share the same map_ptr as active_lock.ptr, they
10373 * will get different reg->id assigned to each lookup, hence different
10376 * In case of allocated objects, active_lock.ptr is the reg->btf, and the
10377 * reg->id is a unique ID preserved after the NULL pointer check on the pointer
10378 * returned from bpf_obj_new. Each allocation receives a new reg->id.
10380 static int check_reg_allocation_locked(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
10385 switch ((int)reg->type) {
10386 case PTR_TO_MAP_VALUE:
10387 ptr = reg->map_ptr;
10389 case PTR_TO_BTF_ID | MEM_ALLOC:
10393 verbose(env, "verifier internal error: unknown reg type for lock check\n");
10398 if (!env->cur_state->active_lock.ptr)
10400 if (env->cur_state->active_lock.ptr != ptr ||
10401 env->cur_state->active_lock.id != id) {
10402 verbose(env, "held lock and object are not in the same allocation\n");
10408 static bool is_bpf_list_api_kfunc(u32 btf_id)
10410 return btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
10411 btf_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
10412 btf_id == special_kfunc_list[KF_bpf_list_pop_front] ||
10413 btf_id == special_kfunc_list[KF_bpf_list_pop_back];
10416 static bool is_bpf_rbtree_api_kfunc(u32 btf_id)
10418 return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl] ||
10419 btf_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
10420 btf_id == special_kfunc_list[KF_bpf_rbtree_first];
10423 static bool is_bpf_graph_api_kfunc(u32 btf_id)
10425 return is_bpf_list_api_kfunc(btf_id) || is_bpf_rbtree_api_kfunc(btf_id) ||
10426 btf_id == special_kfunc_list[KF_bpf_refcount_acquire_impl];
10429 static bool is_callback_calling_kfunc(u32 btf_id)
10431 return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl];
10434 static bool is_rbtree_lock_required_kfunc(u32 btf_id)
10436 return is_bpf_rbtree_api_kfunc(btf_id);
10439 static bool check_kfunc_is_graph_root_api(struct bpf_verifier_env *env,
10440 enum btf_field_type head_field_type,
10445 switch (head_field_type) {
10446 case BPF_LIST_HEAD:
10447 ret = is_bpf_list_api_kfunc(kfunc_btf_id);
10450 ret = is_bpf_rbtree_api_kfunc(kfunc_btf_id);
10453 verbose(env, "verifier internal error: unexpected graph root argument type %s\n",
10454 btf_field_type_name(head_field_type));
10459 verbose(env, "verifier internal error: %s head arg for unknown kfunc\n",
10460 btf_field_type_name(head_field_type));
10464 static bool check_kfunc_is_graph_node_api(struct bpf_verifier_env *env,
10465 enum btf_field_type node_field_type,
10470 switch (node_field_type) {
10471 case BPF_LIST_NODE:
10472 ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
10473 kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_back_impl]);
10476 ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
10477 kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl]);
10480 verbose(env, "verifier internal error: unexpected graph node argument type %s\n",
10481 btf_field_type_name(node_field_type));
10486 verbose(env, "verifier internal error: %s node arg for unknown kfunc\n",
10487 btf_field_type_name(node_field_type));
10492 __process_kf_arg_ptr_to_graph_root(struct bpf_verifier_env *env,
10493 struct bpf_reg_state *reg, u32 regno,
10494 struct bpf_kfunc_call_arg_meta *meta,
10495 enum btf_field_type head_field_type,
10496 struct btf_field **head_field)
10498 const char *head_type_name;
10499 struct btf_field *field;
10500 struct btf_record *rec;
10503 if (meta->btf != btf_vmlinux) {
10504 verbose(env, "verifier internal error: unexpected btf mismatch in kfunc call\n");
10508 if (!check_kfunc_is_graph_root_api(env, head_field_type, meta->func_id))
10511 head_type_name = btf_field_type_name(head_field_type);
10512 if (!tnum_is_const(reg->var_off)) {
10514 "R%d doesn't have constant offset. %s has to be at the constant offset\n",
10515 regno, head_type_name);
10519 rec = reg_btf_record(reg);
10520 head_off = reg->off + reg->var_off.value;
10521 field = btf_record_find(rec, head_off, head_field_type);
10523 verbose(env, "%s not found at offset=%u\n", head_type_name, head_off);
10527 /* All functions require bpf_list_head to be protected using a bpf_spin_lock */
10528 if (check_reg_allocation_locked(env, reg)) {
10529 verbose(env, "bpf_spin_lock at off=%d must be held for %s\n",
10530 rec->spin_lock_off, head_type_name);
10535 verbose(env, "verifier internal error: repeating %s arg\n", head_type_name);
10538 *head_field = field;
10542 static int process_kf_arg_ptr_to_list_head(struct bpf_verifier_env *env,
10543 struct bpf_reg_state *reg, u32 regno,
10544 struct bpf_kfunc_call_arg_meta *meta)
10546 return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, BPF_LIST_HEAD,
10547 &meta->arg_list_head.field);
10550 static int process_kf_arg_ptr_to_rbtree_root(struct bpf_verifier_env *env,
10551 struct bpf_reg_state *reg, u32 regno,
10552 struct bpf_kfunc_call_arg_meta *meta)
10554 return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, BPF_RB_ROOT,
10555 &meta->arg_rbtree_root.field);
10559 __process_kf_arg_ptr_to_graph_node(struct bpf_verifier_env *env,
10560 struct bpf_reg_state *reg, u32 regno,
10561 struct bpf_kfunc_call_arg_meta *meta,
10562 enum btf_field_type head_field_type,
10563 enum btf_field_type node_field_type,
10564 struct btf_field **node_field)
10566 const char *node_type_name;
10567 const struct btf_type *et, *t;
10568 struct btf_field *field;
10571 if (meta->btf != btf_vmlinux) {
10572 verbose(env, "verifier internal error: unexpected btf mismatch in kfunc call\n");
10576 if (!check_kfunc_is_graph_node_api(env, node_field_type, meta->func_id))
10579 node_type_name = btf_field_type_name(node_field_type);
10580 if (!tnum_is_const(reg->var_off)) {
10582 "R%d doesn't have constant offset. %s has to be at the constant offset\n",
10583 regno, node_type_name);
10587 node_off = reg->off + reg->var_off.value;
10588 field = reg_find_field_offset(reg, node_off, node_field_type);
10589 if (!field || field->offset != node_off) {
10590 verbose(env, "%s not found at offset=%u\n", node_type_name, node_off);
10594 field = *node_field;
10596 et = btf_type_by_id(field->graph_root.btf, field->graph_root.value_btf_id);
10597 t = btf_type_by_id(reg->btf, reg->btf_id);
10598 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, 0, field->graph_root.btf,
10599 field->graph_root.value_btf_id, true)) {
10600 verbose(env, "operation on %s expects arg#1 %s at offset=%d "
10601 "in struct %s, but arg is at offset=%d in struct %s\n",
10602 btf_field_type_name(head_field_type),
10603 btf_field_type_name(node_field_type),
10604 field->graph_root.node_offset,
10605 btf_name_by_offset(field->graph_root.btf, et->name_off),
10606 node_off, btf_name_by_offset(reg->btf, t->name_off));
10609 meta->arg_btf = reg->btf;
10610 meta->arg_btf_id = reg->btf_id;
10612 if (node_off != field->graph_root.node_offset) {
10613 verbose(env, "arg#1 offset=%d, but expected %s at offset=%d in struct %s\n",
10614 node_off, btf_field_type_name(node_field_type),
10615 field->graph_root.node_offset,
10616 btf_name_by_offset(field->graph_root.btf, et->name_off));
10623 static int process_kf_arg_ptr_to_list_node(struct bpf_verifier_env *env,
10624 struct bpf_reg_state *reg, u32 regno,
10625 struct bpf_kfunc_call_arg_meta *meta)
10627 return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta,
10628 BPF_LIST_HEAD, BPF_LIST_NODE,
10629 &meta->arg_list_head.field);
10632 static int process_kf_arg_ptr_to_rbtree_node(struct bpf_verifier_env *env,
10633 struct bpf_reg_state *reg, u32 regno,
10634 struct bpf_kfunc_call_arg_meta *meta)
10636 return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta,
10637 BPF_RB_ROOT, BPF_RB_NODE,
10638 &meta->arg_rbtree_root.field);
10641 static int check_kfunc_args(struct bpf_verifier_env *env, struct bpf_kfunc_call_arg_meta *meta,
10644 const char *func_name = meta->func_name, *ref_tname;
10645 const struct btf *btf = meta->btf;
10646 const struct btf_param *args;
10647 struct btf_record *rec;
10651 args = (const struct btf_param *)(meta->func_proto + 1);
10652 nargs = btf_type_vlen(meta->func_proto);
10653 if (nargs > MAX_BPF_FUNC_REG_ARGS) {
10654 verbose(env, "Function %s has %d > %d args\n", func_name, nargs,
10655 MAX_BPF_FUNC_REG_ARGS);
10659 /* Check that BTF function arguments match actual types that the
10662 for (i = 0; i < nargs; i++) {
10663 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[i + 1];
10664 const struct btf_type *t, *ref_t, *resolve_ret;
10665 enum bpf_arg_type arg_type = ARG_DONTCARE;
10666 u32 regno = i + 1, ref_id, type_size;
10667 bool is_ret_buf_sz = false;
10670 t = btf_type_skip_modifiers(btf, args[i].type, NULL);
10672 if (is_kfunc_arg_ignore(btf, &args[i]))
10675 if (btf_type_is_scalar(t)) {
10676 if (reg->type != SCALAR_VALUE) {
10677 verbose(env, "R%d is not a scalar\n", regno);
10681 if (is_kfunc_arg_constant(meta->btf, &args[i])) {
10682 if (meta->arg_constant.found) {
10683 verbose(env, "verifier internal error: only one constant argument permitted\n");
10686 if (!tnum_is_const(reg->var_off)) {
10687 verbose(env, "R%d must be a known constant\n", regno);
10690 ret = mark_chain_precision(env, regno);
10693 meta->arg_constant.found = true;
10694 meta->arg_constant.value = reg->var_off.value;
10695 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdonly_buf_size")) {
10696 meta->r0_rdonly = true;
10697 is_ret_buf_sz = true;
10698 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdwr_buf_size")) {
10699 is_ret_buf_sz = true;
10702 if (is_ret_buf_sz) {
10703 if (meta->r0_size) {
10704 verbose(env, "2 or more rdonly/rdwr_buf_size parameters for kfunc");
10708 if (!tnum_is_const(reg->var_off)) {
10709 verbose(env, "R%d is not a const\n", regno);
10713 meta->r0_size = reg->var_off.value;
10714 ret = mark_chain_precision(env, regno);
10721 if (!btf_type_is_ptr(t)) {
10722 verbose(env, "Unrecognized arg#%d type %s\n", i, btf_type_str(t));
10726 if ((is_kfunc_trusted_args(meta) || is_kfunc_rcu(meta)) &&
10727 (register_is_null(reg) || type_may_be_null(reg->type))) {
10728 verbose(env, "Possibly NULL pointer passed to trusted arg%d\n", i);
10732 if (reg->ref_obj_id) {
10733 if (is_kfunc_release(meta) && meta->ref_obj_id) {
10734 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
10735 regno, reg->ref_obj_id,
10739 meta->ref_obj_id = reg->ref_obj_id;
10740 if (is_kfunc_release(meta))
10741 meta->release_regno = regno;
10744 ref_t = btf_type_skip_modifiers(btf, t->type, &ref_id);
10745 ref_tname = btf_name_by_offset(btf, ref_t->name_off);
10747 kf_arg_type = get_kfunc_ptr_arg_type(env, meta, t, ref_t, ref_tname, args, i, nargs);
10748 if (kf_arg_type < 0)
10749 return kf_arg_type;
10751 switch (kf_arg_type) {
10752 case KF_ARG_PTR_TO_ALLOC_BTF_ID:
10753 case KF_ARG_PTR_TO_BTF_ID:
10754 if (!is_kfunc_trusted_args(meta) && !is_kfunc_rcu(meta))
10757 if (!is_trusted_reg(reg)) {
10758 if (!is_kfunc_rcu(meta)) {
10759 verbose(env, "R%d must be referenced or trusted\n", regno);
10762 if (!is_rcu_reg(reg)) {
10763 verbose(env, "R%d must be a rcu pointer\n", regno);
10769 case KF_ARG_PTR_TO_CTX:
10770 /* Trusted arguments have the same offset checks as release arguments */
10771 arg_type |= OBJ_RELEASE;
10773 case KF_ARG_PTR_TO_DYNPTR:
10774 case KF_ARG_PTR_TO_ITER:
10775 case KF_ARG_PTR_TO_LIST_HEAD:
10776 case KF_ARG_PTR_TO_LIST_NODE:
10777 case KF_ARG_PTR_TO_RB_ROOT:
10778 case KF_ARG_PTR_TO_RB_NODE:
10779 case KF_ARG_PTR_TO_MEM:
10780 case KF_ARG_PTR_TO_MEM_SIZE:
10781 case KF_ARG_PTR_TO_CALLBACK:
10782 case KF_ARG_PTR_TO_REFCOUNTED_KPTR:
10783 /* Trusted by default */
10790 if (is_kfunc_release(meta) && reg->ref_obj_id)
10791 arg_type |= OBJ_RELEASE;
10792 ret = check_func_arg_reg_off(env, reg, regno, arg_type);
10796 switch (kf_arg_type) {
10797 case KF_ARG_PTR_TO_CTX:
10798 if (reg->type != PTR_TO_CTX) {
10799 verbose(env, "arg#%d expected pointer to ctx, but got %s\n", i, btf_type_str(t));
10803 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
10804 ret = get_kern_ctx_btf_id(&env->log, resolve_prog_type(env->prog));
10807 meta->ret_btf_id = ret;
10810 case KF_ARG_PTR_TO_ALLOC_BTF_ID:
10811 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
10812 verbose(env, "arg#%d expected pointer to allocated object\n", i);
10815 if (!reg->ref_obj_id) {
10816 verbose(env, "allocated object must be referenced\n");
10819 if (meta->btf == btf_vmlinux &&
10820 meta->func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
10821 meta->arg_btf = reg->btf;
10822 meta->arg_btf_id = reg->btf_id;
10825 case KF_ARG_PTR_TO_DYNPTR:
10827 enum bpf_arg_type dynptr_arg_type = ARG_PTR_TO_DYNPTR;
10828 int clone_ref_obj_id = 0;
10830 if (reg->type != PTR_TO_STACK &&
10831 reg->type != CONST_PTR_TO_DYNPTR) {
10832 verbose(env, "arg#%d expected pointer to stack or dynptr_ptr\n", i);
10836 if (reg->type == CONST_PTR_TO_DYNPTR)
10837 dynptr_arg_type |= MEM_RDONLY;
10839 if (is_kfunc_arg_uninit(btf, &args[i]))
10840 dynptr_arg_type |= MEM_UNINIT;
10842 if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) {
10843 dynptr_arg_type |= DYNPTR_TYPE_SKB;
10844 } else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_xdp]) {
10845 dynptr_arg_type |= DYNPTR_TYPE_XDP;
10846 } else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_clone] &&
10847 (dynptr_arg_type & MEM_UNINIT)) {
10848 enum bpf_dynptr_type parent_type = meta->initialized_dynptr.type;
10850 if (parent_type == BPF_DYNPTR_TYPE_INVALID) {
10851 verbose(env, "verifier internal error: no dynptr type for parent of clone\n");
10855 dynptr_arg_type |= (unsigned int)get_dynptr_type_flag(parent_type);
10856 clone_ref_obj_id = meta->initialized_dynptr.ref_obj_id;
10857 if (dynptr_type_refcounted(parent_type) && !clone_ref_obj_id) {
10858 verbose(env, "verifier internal error: missing ref obj id for parent of clone\n");
10863 ret = process_dynptr_func(env, regno, insn_idx, dynptr_arg_type, clone_ref_obj_id);
10867 if (!(dynptr_arg_type & MEM_UNINIT)) {
10868 int id = dynptr_id(env, reg);
10871 verbose(env, "verifier internal error: failed to obtain dynptr id\n");
10874 meta->initialized_dynptr.id = id;
10875 meta->initialized_dynptr.type = dynptr_get_type(env, reg);
10876 meta->initialized_dynptr.ref_obj_id = dynptr_ref_obj_id(env, reg);
10881 case KF_ARG_PTR_TO_ITER:
10882 ret = process_iter_arg(env, regno, insn_idx, meta);
10886 case KF_ARG_PTR_TO_LIST_HEAD:
10887 if (reg->type != PTR_TO_MAP_VALUE &&
10888 reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
10889 verbose(env, "arg#%d expected pointer to map value or allocated object\n", i);
10892 if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
10893 verbose(env, "allocated object must be referenced\n");
10896 ret = process_kf_arg_ptr_to_list_head(env, reg, regno, meta);
10900 case KF_ARG_PTR_TO_RB_ROOT:
10901 if (reg->type != PTR_TO_MAP_VALUE &&
10902 reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
10903 verbose(env, "arg#%d expected pointer to map value or allocated object\n", i);
10906 if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
10907 verbose(env, "allocated object must be referenced\n");
10910 ret = process_kf_arg_ptr_to_rbtree_root(env, reg, regno, meta);
10914 case KF_ARG_PTR_TO_LIST_NODE:
10915 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
10916 verbose(env, "arg#%d expected pointer to allocated object\n", i);
10919 if (!reg->ref_obj_id) {
10920 verbose(env, "allocated object must be referenced\n");
10923 ret = process_kf_arg_ptr_to_list_node(env, reg, regno, meta);
10927 case KF_ARG_PTR_TO_RB_NODE:
10928 if (meta->func_id == special_kfunc_list[KF_bpf_rbtree_remove]) {
10929 if (!type_is_non_owning_ref(reg->type) || reg->ref_obj_id) {
10930 verbose(env, "rbtree_remove node input must be non-owning ref\n");
10933 if (in_rbtree_lock_required_cb(env)) {
10934 verbose(env, "rbtree_remove not allowed in rbtree cb\n");
10938 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
10939 verbose(env, "arg#%d expected pointer to allocated object\n", i);
10942 if (!reg->ref_obj_id) {
10943 verbose(env, "allocated object must be referenced\n");
10948 ret = process_kf_arg_ptr_to_rbtree_node(env, reg, regno, meta);
10952 case KF_ARG_PTR_TO_BTF_ID:
10953 /* Only base_type is checked, further checks are done here */
10954 if ((base_type(reg->type) != PTR_TO_BTF_ID ||
10955 (bpf_type_has_unsafe_modifiers(reg->type) && !is_rcu_reg(reg))) &&
10956 !reg2btf_ids[base_type(reg->type)]) {
10957 verbose(env, "arg#%d is %s ", i, reg_type_str(env, reg->type));
10958 verbose(env, "expected %s or socket\n",
10959 reg_type_str(env, base_type(reg->type) |
10960 (type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS)));
10963 ret = process_kf_arg_ptr_to_btf_id(env, reg, ref_t, ref_tname, ref_id, meta, i);
10967 case KF_ARG_PTR_TO_MEM:
10968 resolve_ret = btf_resolve_size(btf, ref_t, &type_size);
10969 if (IS_ERR(resolve_ret)) {
10970 verbose(env, "arg#%d reference type('%s %s') size cannot be determined: %ld\n",
10971 i, btf_type_str(ref_t), ref_tname, PTR_ERR(resolve_ret));
10974 ret = check_mem_reg(env, reg, regno, type_size);
10978 case KF_ARG_PTR_TO_MEM_SIZE:
10980 struct bpf_reg_state *buff_reg = ®s[regno];
10981 const struct btf_param *buff_arg = &args[i];
10982 struct bpf_reg_state *size_reg = ®s[regno + 1];
10983 const struct btf_param *size_arg = &args[i + 1];
10985 if (!register_is_null(buff_reg) || !is_kfunc_arg_optional(meta->btf, buff_arg)) {
10986 ret = check_kfunc_mem_size_reg(env, size_reg, regno + 1);
10988 verbose(env, "arg#%d arg#%d memory, len pair leads to invalid memory access\n", i, i + 1);
10993 if (is_kfunc_arg_const_mem_size(meta->btf, size_arg, size_reg)) {
10994 if (meta->arg_constant.found) {
10995 verbose(env, "verifier internal error: only one constant argument permitted\n");
10998 if (!tnum_is_const(size_reg->var_off)) {
10999 verbose(env, "R%d must be a known constant\n", regno + 1);
11002 meta->arg_constant.found = true;
11003 meta->arg_constant.value = size_reg->var_off.value;
11006 /* Skip next '__sz' or '__szk' argument */
11010 case KF_ARG_PTR_TO_CALLBACK:
11011 meta->subprogno = reg->subprogno;
11013 case KF_ARG_PTR_TO_REFCOUNTED_KPTR:
11014 if (!type_is_ptr_alloc_obj(reg->type)) {
11015 verbose(env, "arg#%d is neither owning or non-owning ref\n", i);
11018 if (!type_is_non_owning_ref(reg->type))
11019 meta->arg_owning_ref = true;
11021 rec = reg_btf_record(reg);
11023 verbose(env, "verifier internal error: Couldn't find btf_record\n");
11027 if (rec->refcount_off < 0) {
11028 verbose(env, "arg#%d doesn't point to a type with bpf_refcount field\n", i);
11031 if (rec->refcount_off >= 0) {
11032 verbose(env, "bpf_refcount_acquire calls are disabled for now\n");
11035 meta->arg_btf = reg->btf;
11036 meta->arg_btf_id = reg->btf_id;
11041 if (is_kfunc_release(meta) && !meta->release_regno) {
11042 verbose(env, "release kernel function %s expects refcounted PTR_TO_BTF_ID\n",
11050 static int fetch_kfunc_meta(struct bpf_verifier_env *env,
11051 struct bpf_insn *insn,
11052 struct bpf_kfunc_call_arg_meta *meta,
11053 const char **kfunc_name)
11055 const struct btf_type *func, *func_proto;
11056 u32 func_id, *kfunc_flags;
11057 const char *func_name;
11058 struct btf *desc_btf;
11061 *kfunc_name = NULL;
11066 desc_btf = find_kfunc_desc_btf(env, insn->off);
11067 if (IS_ERR(desc_btf))
11068 return PTR_ERR(desc_btf);
11070 func_id = insn->imm;
11071 func = btf_type_by_id(desc_btf, func_id);
11072 func_name = btf_name_by_offset(desc_btf, func->name_off);
11074 *kfunc_name = func_name;
11075 func_proto = btf_type_by_id(desc_btf, func->type);
11077 kfunc_flags = btf_kfunc_id_set_contains(desc_btf, func_id, env->prog);
11078 if (!kfunc_flags) {
11082 memset(meta, 0, sizeof(*meta));
11083 meta->btf = desc_btf;
11084 meta->func_id = func_id;
11085 meta->kfunc_flags = *kfunc_flags;
11086 meta->func_proto = func_proto;
11087 meta->func_name = func_name;
11092 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
11095 const struct btf_type *t, *ptr_type;
11096 u32 i, nargs, ptr_type_id, release_ref_obj_id;
11097 struct bpf_reg_state *regs = cur_regs(env);
11098 const char *func_name, *ptr_type_name;
11099 bool sleepable, rcu_lock, rcu_unlock;
11100 struct bpf_kfunc_call_arg_meta meta;
11101 struct bpf_insn_aux_data *insn_aux;
11102 int err, insn_idx = *insn_idx_p;
11103 const struct btf_param *args;
11104 const struct btf_type *ret_t;
11105 struct btf *desc_btf;
11107 /* skip for now, but return error when we find this in fixup_kfunc_call */
11111 err = fetch_kfunc_meta(env, insn, &meta, &func_name);
11112 if (err == -EACCES && func_name)
11113 verbose(env, "calling kernel function %s is not allowed\n", func_name);
11116 desc_btf = meta.btf;
11117 insn_aux = &env->insn_aux_data[insn_idx];
11119 insn_aux->is_iter_next = is_iter_next_kfunc(&meta);
11121 if (is_kfunc_destructive(&meta) && !capable(CAP_SYS_BOOT)) {
11122 verbose(env, "destructive kfunc calls require CAP_SYS_BOOT capability\n");
11126 sleepable = is_kfunc_sleepable(&meta);
11127 if (sleepable && !env->prog->aux->sleepable) {
11128 verbose(env, "program must be sleepable to call sleepable kfunc %s\n", func_name);
11132 rcu_lock = is_kfunc_bpf_rcu_read_lock(&meta);
11133 rcu_unlock = is_kfunc_bpf_rcu_read_unlock(&meta);
11135 if (env->cur_state->active_rcu_lock) {
11136 struct bpf_func_state *state;
11137 struct bpf_reg_state *reg;
11140 verbose(env, "nested rcu read lock (kernel function %s)\n", func_name);
11142 } else if (rcu_unlock) {
11143 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
11144 if (reg->type & MEM_RCU) {
11145 reg->type &= ~(MEM_RCU | PTR_MAYBE_NULL);
11146 reg->type |= PTR_UNTRUSTED;
11149 env->cur_state->active_rcu_lock = false;
11150 } else if (sleepable) {
11151 verbose(env, "kernel func %s is sleepable within rcu_read_lock region\n", func_name);
11154 } else if (rcu_lock) {
11155 env->cur_state->active_rcu_lock = true;
11156 } else if (rcu_unlock) {
11157 verbose(env, "unmatched rcu read unlock (kernel function %s)\n", func_name);
11161 /* Check the arguments */
11162 err = check_kfunc_args(env, &meta, insn_idx);
11165 /* In case of release function, we get register number of refcounted
11166 * PTR_TO_BTF_ID in bpf_kfunc_arg_meta, do the release now.
11168 if (meta.release_regno) {
11169 err = release_reference(env, regs[meta.release_regno].ref_obj_id);
11171 verbose(env, "kfunc %s#%d reference has not been acquired before\n",
11172 func_name, meta.func_id);
11177 if (meta.func_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
11178 meta.func_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
11179 meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
11180 release_ref_obj_id = regs[BPF_REG_2].ref_obj_id;
11181 insn_aux->insert_off = regs[BPF_REG_2].off;
11182 insn_aux->kptr_struct_meta = btf_find_struct_meta(meta.arg_btf, meta.arg_btf_id);
11183 err = ref_convert_owning_non_owning(env, release_ref_obj_id);
11185 verbose(env, "kfunc %s#%d conversion of owning ref to non-owning failed\n",
11186 func_name, meta.func_id);
11190 err = release_reference(env, release_ref_obj_id);
11192 verbose(env, "kfunc %s#%d reference has not been acquired before\n",
11193 func_name, meta.func_id);
11198 if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
11199 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
11200 set_rbtree_add_callback_state);
11202 verbose(env, "kfunc %s#%d failed callback verification\n",
11203 func_name, meta.func_id);
11208 for (i = 0; i < CALLER_SAVED_REGS; i++)
11209 mark_reg_not_init(env, regs, caller_saved[i]);
11211 /* Check return type */
11212 t = btf_type_skip_modifiers(desc_btf, meta.func_proto->type, NULL);
11214 if (is_kfunc_acquire(&meta) && !btf_type_is_struct_ptr(meta.btf, t)) {
11215 /* Only exception is bpf_obj_new_impl */
11216 if (meta.btf != btf_vmlinux ||
11217 (meta.func_id != special_kfunc_list[KF_bpf_obj_new_impl] &&
11218 meta.func_id != special_kfunc_list[KF_bpf_refcount_acquire_impl])) {
11219 verbose(env, "acquire kernel function does not return PTR_TO_BTF_ID\n");
11224 if (btf_type_is_scalar(t)) {
11225 mark_reg_unknown(env, regs, BPF_REG_0);
11226 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
11227 } else if (btf_type_is_ptr(t)) {
11228 ptr_type = btf_type_skip_modifiers(desc_btf, t->type, &ptr_type_id);
11230 if (meta.btf == btf_vmlinux && btf_id_set_contains(&special_kfunc_set, meta.func_id)) {
11231 if (meta.func_id == special_kfunc_list[KF_bpf_obj_new_impl]) {
11232 struct btf *ret_btf;
11235 if (unlikely(!bpf_global_ma_set))
11238 if (((u64)(u32)meta.arg_constant.value) != meta.arg_constant.value) {
11239 verbose(env, "local type ID argument must be in range [0, U32_MAX]\n");
11243 ret_btf = env->prog->aux->btf;
11244 ret_btf_id = meta.arg_constant.value;
11246 /* This may be NULL due to user not supplying a BTF */
11248 verbose(env, "bpf_obj_new requires prog BTF\n");
11252 ret_t = btf_type_by_id(ret_btf, ret_btf_id);
11253 if (!ret_t || !__btf_type_is_struct(ret_t)) {
11254 verbose(env, "bpf_obj_new type ID argument must be of a struct\n");
11258 mark_reg_known_zero(env, regs, BPF_REG_0);
11259 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
11260 regs[BPF_REG_0].btf = ret_btf;
11261 regs[BPF_REG_0].btf_id = ret_btf_id;
11263 insn_aux->obj_new_size = ret_t->size;
11264 insn_aux->kptr_struct_meta =
11265 btf_find_struct_meta(ret_btf, ret_btf_id);
11266 } else if (meta.func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) {
11267 mark_reg_known_zero(env, regs, BPF_REG_0);
11268 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
11269 regs[BPF_REG_0].btf = meta.arg_btf;
11270 regs[BPF_REG_0].btf_id = meta.arg_btf_id;
11272 insn_aux->kptr_struct_meta =
11273 btf_find_struct_meta(meta.arg_btf,
11275 } else if (meta.func_id == special_kfunc_list[KF_bpf_list_pop_front] ||
11276 meta.func_id == special_kfunc_list[KF_bpf_list_pop_back]) {
11277 struct btf_field *field = meta.arg_list_head.field;
11279 mark_reg_graph_node(regs, BPF_REG_0, &field->graph_root);
11280 } else if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
11281 meta.func_id == special_kfunc_list[KF_bpf_rbtree_first]) {
11282 struct btf_field *field = meta.arg_rbtree_root.field;
11284 mark_reg_graph_node(regs, BPF_REG_0, &field->graph_root);
11285 } else if (meta.func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
11286 mark_reg_known_zero(env, regs, BPF_REG_0);
11287 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_TRUSTED;
11288 regs[BPF_REG_0].btf = desc_btf;
11289 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
11290 } else if (meta.func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
11291 ret_t = btf_type_by_id(desc_btf, meta.arg_constant.value);
11292 if (!ret_t || !btf_type_is_struct(ret_t)) {
11294 "kfunc bpf_rdonly_cast type ID argument must be of a struct\n");
11298 mark_reg_known_zero(env, regs, BPF_REG_0);
11299 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
11300 regs[BPF_REG_0].btf = desc_btf;
11301 regs[BPF_REG_0].btf_id = meta.arg_constant.value;
11302 } else if (meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice] ||
11303 meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice_rdwr]) {
11304 enum bpf_type_flag type_flag = get_dynptr_type_flag(meta.initialized_dynptr.type);
11306 mark_reg_known_zero(env, regs, BPF_REG_0);
11308 if (!meta.arg_constant.found) {
11309 verbose(env, "verifier internal error: bpf_dynptr_slice(_rdwr) no constant size\n");
11313 regs[BPF_REG_0].mem_size = meta.arg_constant.value;
11315 /* PTR_MAYBE_NULL will be added when is_kfunc_ret_null is checked */
11316 regs[BPF_REG_0].type = PTR_TO_MEM | type_flag;
11318 if (meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice]) {
11319 regs[BPF_REG_0].type |= MEM_RDONLY;
11321 /* this will set env->seen_direct_write to true */
11322 if (!may_access_direct_pkt_data(env, NULL, BPF_WRITE)) {
11323 verbose(env, "the prog does not allow writes to packet data\n");
11328 if (!meta.initialized_dynptr.id) {
11329 verbose(env, "verifier internal error: no dynptr id\n");
11332 regs[BPF_REG_0].dynptr_id = meta.initialized_dynptr.id;
11334 /* we don't need to set BPF_REG_0's ref obj id
11335 * because packet slices are not refcounted (see
11336 * dynptr_type_refcounted)
11339 verbose(env, "kernel function %s unhandled dynamic return type\n",
11343 } else if (!__btf_type_is_struct(ptr_type)) {
11344 if (!meta.r0_size) {
11347 if (!IS_ERR(btf_resolve_size(desc_btf, ptr_type, &sz))) {
11349 meta.r0_rdonly = true;
11352 if (!meta.r0_size) {
11353 ptr_type_name = btf_name_by_offset(desc_btf,
11354 ptr_type->name_off);
11356 "kernel function %s returns pointer type %s %s is not supported\n",
11358 btf_type_str(ptr_type),
11363 mark_reg_known_zero(env, regs, BPF_REG_0);
11364 regs[BPF_REG_0].type = PTR_TO_MEM;
11365 regs[BPF_REG_0].mem_size = meta.r0_size;
11367 if (meta.r0_rdonly)
11368 regs[BPF_REG_0].type |= MEM_RDONLY;
11370 /* Ensures we don't access the memory after a release_reference() */
11371 if (meta.ref_obj_id)
11372 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
11374 mark_reg_known_zero(env, regs, BPF_REG_0);
11375 regs[BPF_REG_0].btf = desc_btf;
11376 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
11377 regs[BPF_REG_0].btf_id = ptr_type_id;
11380 if (is_kfunc_ret_null(&meta)) {
11381 regs[BPF_REG_0].type |= PTR_MAYBE_NULL;
11382 /* For mark_ptr_or_null_reg, see 93c230e3f5bd6 */
11383 regs[BPF_REG_0].id = ++env->id_gen;
11385 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
11386 if (is_kfunc_acquire(&meta)) {
11387 int id = acquire_reference_state(env, insn_idx);
11391 if (is_kfunc_ret_null(&meta))
11392 regs[BPF_REG_0].id = id;
11393 regs[BPF_REG_0].ref_obj_id = id;
11394 } else if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_first]) {
11395 ref_set_non_owning(env, ®s[BPF_REG_0]);
11398 if (reg_may_point_to_spin_lock(®s[BPF_REG_0]) && !regs[BPF_REG_0].id)
11399 regs[BPF_REG_0].id = ++env->id_gen;
11400 } else if (btf_type_is_void(t)) {
11401 if (meta.btf == btf_vmlinux && btf_id_set_contains(&special_kfunc_set, meta.func_id)) {
11402 if (meta.func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
11403 insn_aux->kptr_struct_meta =
11404 btf_find_struct_meta(meta.arg_btf,
11410 nargs = btf_type_vlen(meta.func_proto);
11411 args = (const struct btf_param *)(meta.func_proto + 1);
11412 for (i = 0; i < nargs; i++) {
11415 t = btf_type_skip_modifiers(desc_btf, args[i].type, NULL);
11416 if (btf_type_is_ptr(t))
11417 mark_btf_func_reg_size(env, regno, sizeof(void *));
11419 /* scalar. ensured by btf_check_kfunc_arg_match() */
11420 mark_btf_func_reg_size(env, regno, t->size);
11423 if (is_iter_next_kfunc(&meta)) {
11424 err = process_iter_next_call(env, insn_idx, &meta);
11432 static bool signed_add_overflows(s64 a, s64 b)
11434 /* Do the add in u64, where overflow is well-defined */
11435 s64 res = (s64)((u64)a + (u64)b);
11442 static bool signed_add32_overflows(s32 a, s32 b)
11444 /* Do the add in u32, where overflow is well-defined */
11445 s32 res = (s32)((u32)a + (u32)b);
11452 static bool signed_sub_overflows(s64 a, s64 b)
11454 /* Do the sub in u64, where overflow is well-defined */
11455 s64 res = (s64)((u64)a - (u64)b);
11462 static bool signed_sub32_overflows(s32 a, s32 b)
11464 /* Do the sub in u32, where overflow is well-defined */
11465 s32 res = (s32)((u32)a - (u32)b);
11472 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
11473 const struct bpf_reg_state *reg,
11474 enum bpf_reg_type type)
11476 bool known = tnum_is_const(reg->var_off);
11477 s64 val = reg->var_off.value;
11478 s64 smin = reg->smin_value;
11480 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
11481 verbose(env, "math between %s pointer and %lld is not allowed\n",
11482 reg_type_str(env, type), val);
11486 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
11487 verbose(env, "%s pointer offset %d is not allowed\n",
11488 reg_type_str(env, type), reg->off);
11492 if (smin == S64_MIN) {
11493 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
11494 reg_type_str(env, type));
11498 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
11499 verbose(env, "value %lld makes %s pointer be out of bounds\n",
11500 smin, reg_type_str(env, type));
11508 REASON_BOUNDS = -1,
11515 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
11516 u32 *alu_limit, bool mask_to_left)
11518 u32 max = 0, ptr_limit = 0;
11520 switch (ptr_reg->type) {
11522 /* Offset 0 is out-of-bounds, but acceptable start for the
11523 * left direction, see BPF_REG_FP. Also, unknown scalar
11524 * offset where we would need to deal with min/max bounds is
11525 * currently prohibited for unprivileged.
11527 max = MAX_BPF_STACK + mask_to_left;
11528 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
11530 case PTR_TO_MAP_VALUE:
11531 max = ptr_reg->map_ptr->value_size;
11532 ptr_limit = (mask_to_left ?
11533 ptr_reg->smin_value :
11534 ptr_reg->umax_value) + ptr_reg->off;
11537 return REASON_TYPE;
11540 if (ptr_limit >= max)
11541 return REASON_LIMIT;
11542 *alu_limit = ptr_limit;
11546 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
11547 const struct bpf_insn *insn)
11549 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
11552 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
11553 u32 alu_state, u32 alu_limit)
11555 /* If we arrived here from different branches with different
11556 * state or limits to sanitize, then this won't work.
11558 if (aux->alu_state &&
11559 (aux->alu_state != alu_state ||
11560 aux->alu_limit != alu_limit))
11561 return REASON_PATHS;
11563 /* Corresponding fixup done in do_misc_fixups(). */
11564 aux->alu_state = alu_state;
11565 aux->alu_limit = alu_limit;
11569 static int sanitize_val_alu(struct bpf_verifier_env *env,
11570 struct bpf_insn *insn)
11572 struct bpf_insn_aux_data *aux = cur_aux(env);
11574 if (can_skip_alu_sanitation(env, insn))
11577 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
11580 static bool sanitize_needed(u8 opcode)
11582 return opcode == BPF_ADD || opcode == BPF_SUB;
11585 struct bpf_sanitize_info {
11586 struct bpf_insn_aux_data aux;
11590 static struct bpf_verifier_state *
11591 sanitize_speculative_path(struct bpf_verifier_env *env,
11592 const struct bpf_insn *insn,
11593 u32 next_idx, u32 curr_idx)
11595 struct bpf_verifier_state *branch;
11596 struct bpf_reg_state *regs;
11598 branch = push_stack(env, next_idx, curr_idx, true);
11599 if (branch && insn) {
11600 regs = branch->frame[branch->curframe]->regs;
11601 if (BPF_SRC(insn->code) == BPF_K) {
11602 mark_reg_unknown(env, regs, insn->dst_reg);
11603 } else if (BPF_SRC(insn->code) == BPF_X) {
11604 mark_reg_unknown(env, regs, insn->dst_reg);
11605 mark_reg_unknown(env, regs, insn->src_reg);
11611 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
11612 struct bpf_insn *insn,
11613 const struct bpf_reg_state *ptr_reg,
11614 const struct bpf_reg_state *off_reg,
11615 struct bpf_reg_state *dst_reg,
11616 struct bpf_sanitize_info *info,
11617 const bool commit_window)
11619 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
11620 struct bpf_verifier_state *vstate = env->cur_state;
11621 bool off_is_imm = tnum_is_const(off_reg->var_off);
11622 bool off_is_neg = off_reg->smin_value < 0;
11623 bool ptr_is_dst_reg = ptr_reg == dst_reg;
11624 u8 opcode = BPF_OP(insn->code);
11625 u32 alu_state, alu_limit;
11626 struct bpf_reg_state tmp;
11630 if (can_skip_alu_sanitation(env, insn))
11633 /* We already marked aux for masking from non-speculative
11634 * paths, thus we got here in the first place. We only care
11635 * to explore bad access from here.
11637 if (vstate->speculative)
11640 if (!commit_window) {
11641 if (!tnum_is_const(off_reg->var_off) &&
11642 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
11643 return REASON_BOUNDS;
11645 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
11646 (opcode == BPF_SUB && !off_is_neg);
11649 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
11653 if (commit_window) {
11654 /* In commit phase we narrow the masking window based on
11655 * the observed pointer move after the simulated operation.
11657 alu_state = info->aux.alu_state;
11658 alu_limit = abs(info->aux.alu_limit - alu_limit);
11660 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
11661 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
11662 alu_state |= ptr_is_dst_reg ?
11663 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
11665 /* Limit pruning on unknown scalars to enable deep search for
11666 * potential masking differences from other program paths.
11669 env->explore_alu_limits = true;
11672 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
11676 /* If we're in commit phase, we're done here given we already
11677 * pushed the truncated dst_reg into the speculative verification
11680 * Also, when register is a known constant, we rewrite register-based
11681 * operation to immediate-based, and thus do not need masking (and as
11682 * a consequence, do not need to simulate the zero-truncation either).
11684 if (commit_window || off_is_imm)
11687 /* Simulate and find potential out-of-bounds access under
11688 * speculative execution from truncation as a result of
11689 * masking when off was not within expected range. If off
11690 * sits in dst, then we temporarily need to move ptr there
11691 * to simulate dst (== 0) +/-= ptr. Needed, for example,
11692 * for cases where we use K-based arithmetic in one direction
11693 * and truncated reg-based in the other in order to explore
11696 if (!ptr_is_dst_reg) {
11698 copy_register_state(dst_reg, ptr_reg);
11700 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
11702 if (!ptr_is_dst_reg && ret)
11704 return !ret ? REASON_STACK : 0;
11707 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
11709 struct bpf_verifier_state *vstate = env->cur_state;
11711 /* If we simulate paths under speculation, we don't update the
11712 * insn as 'seen' such that when we verify unreachable paths in
11713 * the non-speculative domain, sanitize_dead_code() can still
11714 * rewrite/sanitize them.
11716 if (!vstate->speculative)
11717 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
11720 static int sanitize_err(struct bpf_verifier_env *env,
11721 const struct bpf_insn *insn, int reason,
11722 const struct bpf_reg_state *off_reg,
11723 const struct bpf_reg_state *dst_reg)
11725 static const char *err = "pointer arithmetic with it prohibited for !root";
11726 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
11727 u32 dst = insn->dst_reg, src = insn->src_reg;
11730 case REASON_BOUNDS:
11731 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
11732 off_reg == dst_reg ? dst : src, err);
11735 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
11736 off_reg == dst_reg ? src : dst, err);
11739 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
11743 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
11747 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
11751 verbose(env, "verifier internal error: unknown reason (%d)\n",
11759 /* check that stack access falls within stack limits and that 'reg' doesn't
11760 * have a variable offset.
11762 * Variable offset is prohibited for unprivileged mode for simplicity since it
11763 * requires corresponding support in Spectre masking for stack ALU. See also
11764 * retrieve_ptr_limit().
11767 * 'off' includes 'reg->off'.
11769 static int check_stack_access_for_ptr_arithmetic(
11770 struct bpf_verifier_env *env,
11772 const struct bpf_reg_state *reg,
11775 if (!tnum_is_const(reg->var_off)) {
11778 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
11779 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
11780 regno, tn_buf, off);
11784 if (off >= 0 || off < -MAX_BPF_STACK) {
11785 verbose(env, "R%d stack pointer arithmetic goes out of range, "
11786 "prohibited for !root; off=%d\n", regno, off);
11793 static int sanitize_check_bounds(struct bpf_verifier_env *env,
11794 const struct bpf_insn *insn,
11795 const struct bpf_reg_state *dst_reg)
11797 u32 dst = insn->dst_reg;
11799 /* For unprivileged we require that resulting offset must be in bounds
11800 * in order to be able to sanitize access later on.
11802 if (env->bypass_spec_v1)
11805 switch (dst_reg->type) {
11807 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
11808 dst_reg->off + dst_reg->var_off.value))
11811 case PTR_TO_MAP_VALUE:
11812 if (check_map_access(env, dst, dst_reg->off, 1, false, ACCESS_HELPER)) {
11813 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
11814 "prohibited for !root\n", dst);
11825 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
11826 * Caller should also handle BPF_MOV case separately.
11827 * If we return -EACCES, caller may want to try again treating pointer as a
11828 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
11830 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
11831 struct bpf_insn *insn,
11832 const struct bpf_reg_state *ptr_reg,
11833 const struct bpf_reg_state *off_reg)
11835 struct bpf_verifier_state *vstate = env->cur_state;
11836 struct bpf_func_state *state = vstate->frame[vstate->curframe];
11837 struct bpf_reg_state *regs = state->regs, *dst_reg;
11838 bool known = tnum_is_const(off_reg->var_off);
11839 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
11840 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
11841 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
11842 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
11843 struct bpf_sanitize_info info = {};
11844 u8 opcode = BPF_OP(insn->code);
11845 u32 dst = insn->dst_reg;
11848 dst_reg = ®s[dst];
11850 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
11851 smin_val > smax_val || umin_val > umax_val) {
11852 /* Taint dst register if offset had invalid bounds derived from
11853 * e.g. dead branches.
11855 __mark_reg_unknown(env, dst_reg);
11859 if (BPF_CLASS(insn->code) != BPF_ALU64) {
11860 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
11861 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
11862 __mark_reg_unknown(env, dst_reg);
11867 "R%d 32-bit pointer arithmetic prohibited\n",
11872 if (ptr_reg->type & PTR_MAYBE_NULL) {
11873 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
11874 dst, reg_type_str(env, ptr_reg->type));
11878 switch (base_type(ptr_reg->type)) {
11879 case CONST_PTR_TO_MAP:
11880 /* smin_val represents the known value */
11881 if (known && smin_val == 0 && opcode == BPF_ADD)
11884 case PTR_TO_PACKET_END:
11885 case PTR_TO_SOCKET:
11886 case PTR_TO_SOCK_COMMON:
11887 case PTR_TO_TCP_SOCK:
11888 case PTR_TO_XDP_SOCK:
11889 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
11890 dst, reg_type_str(env, ptr_reg->type));
11896 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
11897 * The id may be overwritten later if we create a new variable offset.
11899 dst_reg->type = ptr_reg->type;
11900 dst_reg->id = ptr_reg->id;
11902 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
11903 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
11906 /* pointer types do not carry 32-bit bounds at the moment. */
11907 __mark_reg32_unbounded(dst_reg);
11909 if (sanitize_needed(opcode)) {
11910 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
11913 return sanitize_err(env, insn, ret, off_reg, dst_reg);
11918 /* We can take a fixed offset as long as it doesn't overflow
11919 * the s32 'off' field
11921 if (known && (ptr_reg->off + smin_val ==
11922 (s64)(s32)(ptr_reg->off + smin_val))) {
11923 /* pointer += K. Accumulate it into fixed offset */
11924 dst_reg->smin_value = smin_ptr;
11925 dst_reg->smax_value = smax_ptr;
11926 dst_reg->umin_value = umin_ptr;
11927 dst_reg->umax_value = umax_ptr;
11928 dst_reg->var_off = ptr_reg->var_off;
11929 dst_reg->off = ptr_reg->off + smin_val;
11930 dst_reg->raw = ptr_reg->raw;
11933 /* A new variable offset is created. Note that off_reg->off
11934 * == 0, since it's a scalar.
11935 * dst_reg gets the pointer type and since some positive
11936 * integer value was added to the pointer, give it a new 'id'
11937 * if it's a PTR_TO_PACKET.
11938 * this creates a new 'base' pointer, off_reg (variable) gets
11939 * added into the variable offset, and we copy the fixed offset
11942 if (signed_add_overflows(smin_ptr, smin_val) ||
11943 signed_add_overflows(smax_ptr, smax_val)) {
11944 dst_reg->smin_value = S64_MIN;
11945 dst_reg->smax_value = S64_MAX;
11947 dst_reg->smin_value = smin_ptr + smin_val;
11948 dst_reg->smax_value = smax_ptr + smax_val;
11950 if (umin_ptr + umin_val < umin_ptr ||
11951 umax_ptr + umax_val < umax_ptr) {
11952 dst_reg->umin_value = 0;
11953 dst_reg->umax_value = U64_MAX;
11955 dst_reg->umin_value = umin_ptr + umin_val;
11956 dst_reg->umax_value = umax_ptr + umax_val;
11958 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
11959 dst_reg->off = ptr_reg->off;
11960 dst_reg->raw = ptr_reg->raw;
11961 if (reg_is_pkt_pointer(ptr_reg)) {
11962 dst_reg->id = ++env->id_gen;
11963 /* something was added to pkt_ptr, set range to zero */
11964 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
11968 if (dst_reg == off_reg) {
11969 /* scalar -= pointer. Creates an unknown scalar */
11970 verbose(env, "R%d tried to subtract pointer from scalar\n",
11974 /* We don't allow subtraction from FP, because (according to
11975 * test_verifier.c test "invalid fp arithmetic", JITs might not
11976 * be able to deal with it.
11978 if (ptr_reg->type == PTR_TO_STACK) {
11979 verbose(env, "R%d subtraction from stack pointer prohibited\n",
11983 if (known && (ptr_reg->off - smin_val ==
11984 (s64)(s32)(ptr_reg->off - smin_val))) {
11985 /* pointer -= K. Subtract it from fixed offset */
11986 dst_reg->smin_value = smin_ptr;
11987 dst_reg->smax_value = smax_ptr;
11988 dst_reg->umin_value = umin_ptr;
11989 dst_reg->umax_value = umax_ptr;
11990 dst_reg->var_off = ptr_reg->var_off;
11991 dst_reg->id = ptr_reg->id;
11992 dst_reg->off = ptr_reg->off - smin_val;
11993 dst_reg->raw = ptr_reg->raw;
11996 /* A new variable offset is created. If the subtrahend is known
11997 * nonnegative, then any reg->range we had before is still good.
11999 if (signed_sub_overflows(smin_ptr, smax_val) ||
12000 signed_sub_overflows(smax_ptr, smin_val)) {
12001 /* Overflow possible, we know nothing */
12002 dst_reg->smin_value = S64_MIN;
12003 dst_reg->smax_value = S64_MAX;
12005 dst_reg->smin_value = smin_ptr - smax_val;
12006 dst_reg->smax_value = smax_ptr - smin_val;
12008 if (umin_ptr < umax_val) {
12009 /* Overflow possible, we know nothing */
12010 dst_reg->umin_value = 0;
12011 dst_reg->umax_value = U64_MAX;
12013 /* Cannot overflow (as long as bounds are consistent) */
12014 dst_reg->umin_value = umin_ptr - umax_val;
12015 dst_reg->umax_value = umax_ptr - umin_val;
12017 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
12018 dst_reg->off = ptr_reg->off;
12019 dst_reg->raw = ptr_reg->raw;
12020 if (reg_is_pkt_pointer(ptr_reg)) {
12021 dst_reg->id = ++env->id_gen;
12022 /* something was added to pkt_ptr, set range to zero */
12024 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
12030 /* bitwise ops on pointers are troublesome, prohibit. */
12031 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
12032 dst, bpf_alu_string[opcode >> 4]);
12035 /* other operators (e.g. MUL,LSH) produce non-pointer results */
12036 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
12037 dst, bpf_alu_string[opcode >> 4]);
12041 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
12043 reg_bounds_sync(dst_reg);
12044 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
12046 if (sanitize_needed(opcode)) {
12047 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
12050 return sanitize_err(env, insn, ret, off_reg, dst_reg);
12056 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
12057 struct bpf_reg_state *src_reg)
12059 s32 smin_val = src_reg->s32_min_value;
12060 s32 smax_val = src_reg->s32_max_value;
12061 u32 umin_val = src_reg->u32_min_value;
12062 u32 umax_val = src_reg->u32_max_value;
12064 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
12065 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
12066 dst_reg->s32_min_value = S32_MIN;
12067 dst_reg->s32_max_value = S32_MAX;
12069 dst_reg->s32_min_value += smin_val;
12070 dst_reg->s32_max_value += smax_val;
12072 if (dst_reg->u32_min_value + umin_val < umin_val ||
12073 dst_reg->u32_max_value + umax_val < umax_val) {
12074 dst_reg->u32_min_value = 0;
12075 dst_reg->u32_max_value = U32_MAX;
12077 dst_reg->u32_min_value += umin_val;
12078 dst_reg->u32_max_value += umax_val;
12082 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
12083 struct bpf_reg_state *src_reg)
12085 s64 smin_val = src_reg->smin_value;
12086 s64 smax_val = src_reg->smax_value;
12087 u64 umin_val = src_reg->umin_value;
12088 u64 umax_val = src_reg->umax_value;
12090 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
12091 signed_add_overflows(dst_reg->smax_value, smax_val)) {
12092 dst_reg->smin_value = S64_MIN;
12093 dst_reg->smax_value = S64_MAX;
12095 dst_reg->smin_value += smin_val;
12096 dst_reg->smax_value += smax_val;
12098 if (dst_reg->umin_value + umin_val < umin_val ||
12099 dst_reg->umax_value + umax_val < umax_val) {
12100 dst_reg->umin_value = 0;
12101 dst_reg->umax_value = U64_MAX;
12103 dst_reg->umin_value += umin_val;
12104 dst_reg->umax_value += umax_val;
12108 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
12109 struct bpf_reg_state *src_reg)
12111 s32 smin_val = src_reg->s32_min_value;
12112 s32 smax_val = src_reg->s32_max_value;
12113 u32 umin_val = src_reg->u32_min_value;
12114 u32 umax_val = src_reg->u32_max_value;
12116 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
12117 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
12118 /* Overflow possible, we know nothing */
12119 dst_reg->s32_min_value = S32_MIN;
12120 dst_reg->s32_max_value = S32_MAX;
12122 dst_reg->s32_min_value -= smax_val;
12123 dst_reg->s32_max_value -= smin_val;
12125 if (dst_reg->u32_min_value < umax_val) {
12126 /* Overflow possible, we know nothing */
12127 dst_reg->u32_min_value = 0;
12128 dst_reg->u32_max_value = U32_MAX;
12130 /* Cannot overflow (as long as bounds are consistent) */
12131 dst_reg->u32_min_value -= umax_val;
12132 dst_reg->u32_max_value -= umin_val;
12136 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
12137 struct bpf_reg_state *src_reg)
12139 s64 smin_val = src_reg->smin_value;
12140 s64 smax_val = src_reg->smax_value;
12141 u64 umin_val = src_reg->umin_value;
12142 u64 umax_val = src_reg->umax_value;
12144 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
12145 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
12146 /* Overflow possible, we know nothing */
12147 dst_reg->smin_value = S64_MIN;
12148 dst_reg->smax_value = S64_MAX;
12150 dst_reg->smin_value -= smax_val;
12151 dst_reg->smax_value -= smin_val;
12153 if (dst_reg->umin_value < umax_val) {
12154 /* Overflow possible, we know nothing */
12155 dst_reg->umin_value = 0;
12156 dst_reg->umax_value = U64_MAX;
12158 /* Cannot overflow (as long as bounds are consistent) */
12159 dst_reg->umin_value -= umax_val;
12160 dst_reg->umax_value -= umin_val;
12164 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
12165 struct bpf_reg_state *src_reg)
12167 s32 smin_val = src_reg->s32_min_value;
12168 u32 umin_val = src_reg->u32_min_value;
12169 u32 umax_val = src_reg->u32_max_value;
12171 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
12172 /* Ain't nobody got time to multiply that sign */
12173 __mark_reg32_unbounded(dst_reg);
12176 /* Both values are positive, so we can work with unsigned and
12177 * copy the result to signed (unless it exceeds S32_MAX).
12179 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
12180 /* Potential overflow, we know nothing */
12181 __mark_reg32_unbounded(dst_reg);
12184 dst_reg->u32_min_value *= umin_val;
12185 dst_reg->u32_max_value *= umax_val;
12186 if (dst_reg->u32_max_value > S32_MAX) {
12187 /* Overflow possible, we know nothing */
12188 dst_reg->s32_min_value = S32_MIN;
12189 dst_reg->s32_max_value = S32_MAX;
12191 dst_reg->s32_min_value = dst_reg->u32_min_value;
12192 dst_reg->s32_max_value = dst_reg->u32_max_value;
12196 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
12197 struct bpf_reg_state *src_reg)
12199 s64 smin_val = src_reg->smin_value;
12200 u64 umin_val = src_reg->umin_value;
12201 u64 umax_val = src_reg->umax_value;
12203 if (smin_val < 0 || dst_reg->smin_value < 0) {
12204 /* Ain't nobody got time to multiply that sign */
12205 __mark_reg64_unbounded(dst_reg);
12208 /* Both values are positive, so we can work with unsigned and
12209 * copy the result to signed (unless it exceeds S64_MAX).
12211 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
12212 /* Potential overflow, we know nothing */
12213 __mark_reg64_unbounded(dst_reg);
12216 dst_reg->umin_value *= umin_val;
12217 dst_reg->umax_value *= umax_val;
12218 if (dst_reg->umax_value > S64_MAX) {
12219 /* Overflow possible, we know nothing */
12220 dst_reg->smin_value = S64_MIN;
12221 dst_reg->smax_value = S64_MAX;
12223 dst_reg->smin_value = dst_reg->umin_value;
12224 dst_reg->smax_value = dst_reg->umax_value;
12228 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
12229 struct bpf_reg_state *src_reg)
12231 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12232 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12233 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12234 s32 smin_val = src_reg->s32_min_value;
12235 u32 umax_val = src_reg->u32_max_value;
12237 if (src_known && dst_known) {
12238 __mark_reg32_known(dst_reg, var32_off.value);
12242 /* We get our minimum from the var_off, since that's inherently
12243 * bitwise. Our maximum is the minimum of the operands' maxima.
12245 dst_reg->u32_min_value = var32_off.value;
12246 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
12247 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
12248 /* Lose signed bounds when ANDing negative numbers,
12249 * ain't nobody got time for that.
12251 dst_reg->s32_min_value = S32_MIN;
12252 dst_reg->s32_max_value = S32_MAX;
12254 /* ANDing two positives gives a positive, so safe to
12255 * cast result into s64.
12257 dst_reg->s32_min_value = dst_reg->u32_min_value;
12258 dst_reg->s32_max_value = dst_reg->u32_max_value;
12262 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
12263 struct bpf_reg_state *src_reg)
12265 bool src_known = tnum_is_const(src_reg->var_off);
12266 bool dst_known = tnum_is_const(dst_reg->var_off);
12267 s64 smin_val = src_reg->smin_value;
12268 u64 umax_val = src_reg->umax_value;
12270 if (src_known && dst_known) {
12271 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12275 /* We get our minimum from the var_off, since that's inherently
12276 * bitwise. Our maximum is the minimum of the operands' maxima.
12278 dst_reg->umin_value = dst_reg->var_off.value;
12279 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
12280 if (dst_reg->smin_value < 0 || smin_val < 0) {
12281 /* Lose signed bounds when ANDing negative numbers,
12282 * ain't nobody got time for that.
12284 dst_reg->smin_value = S64_MIN;
12285 dst_reg->smax_value = S64_MAX;
12287 /* ANDing two positives gives a positive, so safe to
12288 * cast result into s64.
12290 dst_reg->smin_value = dst_reg->umin_value;
12291 dst_reg->smax_value = dst_reg->umax_value;
12293 /* We may learn something more from the var_off */
12294 __update_reg_bounds(dst_reg);
12297 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
12298 struct bpf_reg_state *src_reg)
12300 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12301 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12302 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12303 s32 smin_val = src_reg->s32_min_value;
12304 u32 umin_val = src_reg->u32_min_value;
12306 if (src_known && dst_known) {
12307 __mark_reg32_known(dst_reg, var32_off.value);
12311 /* We get our maximum from the var_off, and our minimum is the
12312 * maximum of the operands' minima
12314 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
12315 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
12316 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
12317 /* Lose signed bounds when ORing negative numbers,
12318 * ain't nobody got time for that.
12320 dst_reg->s32_min_value = S32_MIN;
12321 dst_reg->s32_max_value = S32_MAX;
12323 /* ORing two positives gives a positive, so safe to
12324 * cast result into s64.
12326 dst_reg->s32_min_value = dst_reg->u32_min_value;
12327 dst_reg->s32_max_value = dst_reg->u32_max_value;
12331 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
12332 struct bpf_reg_state *src_reg)
12334 bool src_known = tnum_is_const(src_reg->var_off);
12335 bool dst_known = tnum_is_const(dst_reg->var_off);
12336 s64 smin_val = src_reg->smin_value;
12337 u64 umin_val = src_reg->umin_value;
12339 if (src_known && dst_known) {
12340 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12344 /* We get our maximum from the var_off, and our minimum is the
12345 * maximum of the operands' minima
12347 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
12348 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
12349 if (dst_reg->smin_value < 0 || smin_val < 0) {
12350 /* Lose signed bounds when ORing negative numbers,
12351 * ain't nobody got time for that.
12353 dst_reg->smin_value = S64_MIN;
12354 dst_reg->smax_value = S64_MAX;
12356 /* ORing two positives gives a positive, so safe to
12357 * cast result into s64.
12359 dst_reg->smin_value = dst_reg->umin_value;
12360 dst_reg->smax_value = dst_reg->umax_value;
12362 /* We may learn something more from the var_off */
12363 __update_reg_bounds(dst_reg);
12366 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
12367 struct bpf_reg_state *src_reg)
12369 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12370 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12371 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12372 s32 smin_val = src_reg->s32_min_value;
12374 if (src_known && dst_known) {
12375 __mark_reg32_known(dst_reg, var32_off.value);
12379 /* We get both minimum and maximum from the var32_off. */
12380 dst_reg->u32_min_value = var32_off.value;
12381 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
12383 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
12384 /* XORing two positive sign numbers gives a positive,
12385 * so safe to cast u32 result into s32.
12387 dst_reg->s32_min_value = dst_reg->u32_min_value;
12388 dst_reg->s32_max_value = dst_reg->u32_max_value;
12390 dst_reg->s32_min_value = S32_MIN;
12391 dst_reg->s32_max_value = S32_MAX;
12395 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
12396 struct bpf_reg_state *src_reg)
12398 bool src_known = tnum_is_const(src_reg->var_off);
12399 bool dst_known = tnum_is_const(dst_reg->var_off);
12400 s64 smin_val = src_reg->smin_value;
12402 if (src_known && dst_known) {
12403 /* dst_reg->var_off.value has been updated earlier */
12404 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12408 /* We get both minimum and maximum from the var_off. */
12409 dst_reg->umin_value = dst_reg->var_off.value;
12410 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
12412 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
12413 /* XORing two positive sign numbers gives a positive,
12414 * so safe to cast u64 result into s64.
12416 dst_reg->smin_value = dst_reg->umin_value;
12417 dst_reg->smax_value = dst_reg->umax_value;
12419 dst_reg->smin_value = S64_MIN;
12420 dst_reg->smax_value = S64_MAX;
12423 __update_reg_bounds(dst_reg);
12426 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
12427 u64 umin_val, u64 umax_val)
12429 /* We lose all sign bit information (except what we can pick
12432 dst_reg->s32_min_value = S32_MIN;
12433 dst_reg->s32_max_value = S32_MAX;
12434 /* If we might shift our top bit out, then we know nothing */
12435 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
12436 dst_reg->u32_min_value = 0;
12437 dst_reg->u32_max_value = U32_MAX;
12439 dst_reg->u32_min_value <<= umin_val;
12440 dst_reg->u32_max_value <<= umax_val;
12444 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
12445 struct bpf_reg_state *src_reg)
12447 u32 umax_val = src_reg->u32_max_value;
12448 u32 umin_val = src_reg->u32_min_value;
12449 /* u32 alu operation will zext upper bits */
12450 struct tnum subreg = tnum_subreg(dst_reg->var_off);
12452 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
12453 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
12454 /* Not required but being careful mark reg64 bounds as unknown so
12455 * that we are forced to pick them up from tnum and zext later and
12456 * if some path skips this step we are still safe.
12458 __mark_reg64_unbounded(dst_reg);
12459 __update_reg32_bounds(dst_reg);
12462 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
12463 u64 umin_val, u64 umax_val)
12465 /* Special case <<32 because it is a common compiler pattern to sign
12466 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
12467 * positive we know this shift will also be positive so we can track
12468 * bounds correctly. Otherwise we lose all sign bit information except
12469 * what we can pick up from var_off. Perhaps we can generalize this
12470 * later to shifts of any length.
12472 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
12473 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
12475 dst_reg->smax_value = S64_MAX;
12477 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
12478 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
12480 dst_reg->smin_value = S64_MIN;
12482 /* If we might shift our top bit out, then we know nothing */
12483 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
12484 dst_reg->umin_value = 0;
12485 dst_reg->umax_value = U64_MAX;
12487 dst_reg->umin_value <<= umin_val;
12488 dst_reg->umax_value <<= umax_val;
12492 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
12493 struct bpf_reg_state *src_reg)
12495 u64 umax_val = src_reg->umax_value;
12496 u64 umin_val = src_reg->umin_value;
12498 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
12499 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
12500 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
12502 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
12503 /* We may learn something more from the var_off */
12504 __update_reg_bounds(dst_reg);
12507 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
12508 struct bpf_reg_state *src_reg)
12510 struct tnum subreg = tnum_subreg(dst_reg->var_off);
12511 u32 umax_val = src_reg->u32_max_value;
12512 u32 umin_val = src_reg->u32_min_value;
12514 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
12515 * be negative, then either:
12516 * 1) src_reg might be zero, so the sign bit of the result is
12517 * unknown, so we lose our signed bounds
12518 * 2) it's known negative, thus the unsigned bounds capture the
12520 * 3) the signed bounds cross zero, so they tell us nothing
12522 * If the value in dst_reg is known nonnegative, then again the
12523 * unsigned bounds capture the signed bounds.
12524 * Thus, in all cases it suffices to blow away our signed bounds
12525 * and rely on inferring new ones from the unsigned bounds and
12526 * var_off of the result.
12528 dst_reg->s32_min_value = S32_MIN;
12529 dst_reg->s32_max_value = S32_MAX;
12531 dst_reg->var_off = tnum_rshift(subreg, umin_val);
12532 dst_reg->u32_min_value >>= umax_val;
12533 dst_reg->u32_max_value >>= umin_val;
12535 __mark_reg64_unbounded(dst_reg);
12536 __update_reg32_bounds(dst_reg);
12539 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
12540 struct bpf_reg_state *src_reg)
12542 u64 umax_val = src_reg->umax_value;
12543 u64 umin_val = src_reg->umin_value;
12545 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
12546 * be negative, then either:
12547 * 1) src_reg might be zero, so the sign bit of the result is
12548 * unknown, so we lose our signed bounds
12549 * 2) it's known negative, thus the unsigned bounds capture the
12551 * 3) the signed bounds cross zero, so they tell us nothing
12553 * If the value in dst_reg is known nonnegative, then again the
12554 * unsigned bounds capture the signed bounds.
12555 * Thus, in all cases it suffices to blow away our signed bounds
12556 * and rely on inferring new ones from the unsigned bounds and
12557 * var_off of the result.
12559 dst_reg->smin_value = S64_MIN;
12560 dst_reg->smax_value = S64_MAX;
12561 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
12562 dst_reg->umin_value >>= umax_val;
12563 dst_reg->umax_value >>= umin_val;
12565 /* Its not easy to operate on alu32 bounds here because it depends
12566 * on bits being shifted in. Take easy way out and mark unbounded
12567 * so we can recalculate later from tnum.
12569 __mark_reg32_unbounded(dst_reg);
12570 __update_reg_bounds(dst_reg);
12573 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
12574 struct bpf_reg_state *src_reg)
12576 u64 umin_val = src_reg->u32_min_value;
12578 /* Upon reaching here, src_known is true and
12579 * umax_val is equal to umin_val.
12581 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
12582 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
12584 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
12586 /* blow away the dst_reg umin_value/umax_value and rely on
12587 * dst_reg var_off to refine the result.
12589 dst_reg->u32_min_value = 0;
12590 dst_reg->u32_max_value = U32_MAX;
12592 __mark_reg64_unbounded(dst_reg);
12593 __update_reg32_bounds(dst_reg);
12596 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
12597 struct bpf_reg_state *src_reg)
12599 u64 umin_val = src_reg->umin_value;
12601 /* Upon reaching here, src_known is true and umax_val is equal
12604 dst_reg->smin_value >>= umin_val;
12605 dst_reg->smax_value >>= umin_val;
12607 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
12609 /* blow away the dst_reg umin_value/umax_value and rely on
12610 * dst_reg var_off to refine the result.
12612 dst_reg->umin_value = 0;
12613 dst_reg->umax_value = U64_MAX;
12615 /* Its not easy to operate on alu32 bounds here because it depends
12616 * on bits being shifted in from upper 32-bits. Take easy way out
12617 * and mark unbounded so we can recalculate later from tnum.
12619 __mark_reg32_unbounded(dst_reg);
12620 __update_reg_bounds(dst_reg);
12623 /* WARNING: This function does calculations on 64-bit values, but the actual
12624 * execution may occur on 32-bit values. Therefore, things like bitshifts
12625 * need extra checks in the 32-bit case.
12627 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
12628 struct bpf_insn *insn,
12629 struct bpf_reg_state *dst_reg,
12630 struct bpf_reg_state src_reg)
12632 struct bpf_reg_state *regs = cur_regs(env);
12633 u8 opcode = BPF_OP(insn->code);
12635 s64 smin_val, smax_val;
12636 u64 umin_val, umax_val;
12637 s32 s32_min_val, s32_max_val;
12638 u32 u32_min_val, u32_max_val;
12639 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
12640 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
12643 smin_val = src_reg.smin_value;
12644 smax_val = src_reg.smax_value;
12645 umin_val = src_reg.umin_value;
12646 umax_val = src_reg.umax_value;
12648 s32_min_val = src_reg.s32_min_value;
12649 s32_max_val = src_reg.s32_max_value;
12650 u32_min_val = src_reg.u32_min_value;
12651 u32_max_val = src_reg.u32_max_value;
12654 src_known = tnum_subreg_is_const(src_reg.var_off);
12656 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
12657 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
12658 /* Taint dst register if offset had invalid bounds
12659 * derived from e.g. dead branches.
12661 __mark_reg_unknown(env, dst_reg);
12665 src_known = tnum_is_const(src_reg.var_off);
12667 (smin_val != smax_val || umin_val != umax_val)) ||
12668 smin_val > smax_val || umin_val > umax_val) {
12669 /* Taint dst register if offset had invalid bounds
12670 * derived from e.g. dead branches.
12672 __mark_reg_unknown(env, dst_reg);
12678 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
12679 __mark_reg_unknown(env, dst_reg);
12683 if (sanitize_needed(opcode)) {
12684 ret = sanitize_val_alu(env, insn);
12686 return sanitize_err(env, insn, ret, NULL, NULL);
12689 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
12690 * There are two classes of instructions: The first class we track both
12691 * alu32 and alu64 sign/unsigned bounds independently this provides the
12692 * greatest amount of precision when alu operations are mixed with jmp32
12693 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
12694 * and BPF_OR. This is possible because these ops have fairly easy to
12695 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
12696 * See alu32 verifier tests for examples. The second class of
12697 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
12698 * with regards to tracking sign/unsigned bounds because the bits may
12699 * cross subreg boundaries in the alu64 case. When this happens we mark
12700 * the reg unbounded in the subreg bound space and use the resulting
12701 * tnum to calculate an approximation of the sign/unsigned bounds.
12705 scalar32_min_max_add(dst_reg, &src_reg);
12706 scalar_min_max_add(dst_reg, &src_reg);
12707 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
12710 scalar32_min_max_sub(dst_reg, &src_reg);
12711 scalar_min_max_sub(dst_reg, &src_reg);
12712 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
12715 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
12716 scalar32_min_max_mul(dst_reg, &src_reg);
12717 scalar_min_max_mul(dst_reg, &src_reg);
12720 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
12721 scalar32_min_max_and(dst_reg, &src_reg);
12722 scalar_min_max_and(dst_reg, &src_reg);
12725 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
12726 scalar32_min_max_or(dst_reg, &src_reg);
12727 scalar_min_max_or(dst_reg, &src_reg);
12730 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
12731 scalar32_min_max_xor(dst_reg, &src_reg);
12732 scalar_min_max_xor(dst_reg, &src_reg);
12735 if (umax_val >= insn_bitness) {
12736 /* Shifts greater than 31 or 63 are undefined.
12737 * This includes shifts by a negative number.
12739 mark_reg_unknown(env, regs, insn->dst_reg);
12743 scalar32_min_max_lsh(dst_reg, &src_reg);
12745 scalar_min_max_lsh(dst_reg, &src_reg);
12748 if (umax_val >= insn_bitness) {
12749 /* Shifts greater than 31 or 63 are undefined.
12750 * This includes shifts by a negative number.
12752 mark_reg_unknown(env, regs, insn->dst_reg);
12756 scalar32_min_max_rsh(dst_reg, &src_reg);
12758 scalar_min_max_rsh(dst_reg, &src_reg);
12761 if (umax_val >= insn_bitness) {
12762 /* Shifts greater than 31 or 63 are undefined.
12763 * This includes shifts by a negative number.
12765 mark_reg_unknown(env, regs, insn->dst_reg);
12769 scalar32_min_max_arsh(dst_reg, &src_reg);
12771 scalar_min_max_arsh(dst_reg, &src_reg);
12774 mark_reg_unknown(env, regs, insn->dst_reg);
12778 /* ALU32 ops are zero extended into 64bit register */
12780 zext_32_to_64(dst_reg);
12781 reg_bounds_sync(dst_reg);
12785 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
12788 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
12789 struct bpf_insn *insn)
12791 struct bpf_verifier_state *vstate = env->cur_state;
12792 struct bpf_func_state *state = vstate->frame[vstate->curframe];
12793 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
12794 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
12795 u8 opcode = BPF_OP(insn->code);
12798 dst_reg = ®s[insn->dst_reg];
12800 if (dst_reg->type != SCALAR_VALUE)
12803 /* Make sure ID is cleared otherwise dst_reg min/max could be
12804 * incorrectly propagated into other registers by find_equal_scalars()
12807 if (BPF_SRC(insn->code) == BPF_X) {
12808 src_reg = ®s[insn->src_reg];
12809 if (src_reg->type != SCALAR_VALUE) {
12810 if (dst_reg->type != SCALAR_VALUE) {
12811 /* Combining two pointers by any ALU op yields
12812 * an arbitrary scalar. Disallow all math except
12813 * pointer subtraction
12815 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
12816 mark_reg_unknown(env, regs, insn->dst_reg);
12819 verbose(env, "R%d pointer %s pointer prohibited\n",
12821 bpf_alu_string[opcode >> 4]);
12824 /* scalar += pointer
12825 * This is legal, but we have to reverse our
12826 * src/dest handling in computing the range
12828 err = mark_chain_precision(env, insn->dst_reg);
12831 return adjust_ptr_min_max_vals(env, insn,
12834 } else if (ptr_reg) {
12835 /* pointer += scalar */
12836 err = mark_chain_precision(env, insn->src_reg);
12839 return adjust_ptr_min_max_vals(env, insn,
12841 } else if (dst_reg->precise) {
12842 /* if dst_reg is precise, src_reg should be precise as well */
12843 err = mark_chain_precision(env, insn->src_reg);
12848 /* Pretend the src is a reg with a known value, since we only
12849 * need to be able to read from this state.
12851 off_reg.type = SCALAR_VALUE;
12852 __mark_reg_known(&off_reg, insn->imm);
12853 src_reg = &off_reg;
12854 if (ptr_reg) /* pointer += K */
12855 return adjust_ptr_min_max_vals(env, insn,
12859 /* Got here implies adding two SCALAR_VALUEs */
12860 if (WARN_ON_ONCE(ptr_reg)) {
12861 print_verifier_state(env, state, true);
12862 verbose(env, "verifier internal error: unexpected ptr_reg\n");
12865 if (WARN_ON(!src_reg)) {
12866 print_verifier_state(env, state, true);
12867 verbose(env, "verifier internal error: no src_reg\n");
12870 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
12873 /* check validity of 32-bit and 64-bit arithmetic operations */
12874 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
12876 struct bpf_reg_state *regs = cur_regs(env);
12877 u8 opcode = BPF_OP(insn->code);
12880 if (opcode == BPF_END || opcode == BPF_NEG) {
12881 if (opcode == BPF_NEG) {
12882 if (BPF_SRC(insn->code) != BPF_K ||
12883 insn->src_reg != BPF_REG_0 ||
12884 insn->off != 0 || insn->imm != 0) {
12885 verbose(env, "BPF_NEG uses reserved fields\n");
12889 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
12890 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
12891 BPF_CLASS(insn->code) == BPF_ALU64) {
12892 verbose(env, "BPF_END uses reserved fields\n");
12897 /* check src operand */
12898 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
12902 if (is_pointer_value(env, insn->dst_reg)) {
12903 verbose(env, "R%d pointer arithmetic prohibited\n",
12908 /* check dest operand */
12909 err = check_reg_arg(env, insn->dst_reg, DST_OP);
12913 } else if (opcode == BPF_MOV) {
12915 if (BPF_SRC(insn->code) == BPF_X) {
12916 if (insn->imm != 0 || insn->off != 0) {
12917 verbose(env, "BPF_MOV uses reserved fields\n");
12921 /* check src operand */
12922 err = check_reg_arg(env, insn->src_reg, SRC_OP);
12926 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
12927 verbose(env, "BPF_MOV uses reserved fields\n");
12932 /* check dest operand, mark as required later */
12933 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
12937 if (BPF_SRC(insn->code) == BPF_X) {
12938 struct bpf_reg_state *src_reg = regs + insn->src_reg;
12939 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
12940 bool need_id = src_reg->type == SCALAR_VALUE && !src_reg->id &&
12941 !tnum_is_const(src_reg->var_off);
12943 if (BPF_CLASS(insn->code) == BPF_ALU64) {
12945 * copy register state to dest reg
12948 /* Assign src and dst registers the same ID
12949 * that will be used by find_equal_scalars()
12950 * to propagate min/max range.
12952 src_reg->id = ++env->id_gen;
12953 copy_register_state(dst_reg, src_reg);
12954 dst_reg->live |= REG_LIVE_WRITTEN;
12955 dst_reg->subreg_def = DEF_NOT_SUBREG;
12957 /* R1 = (u32) R2 */
12958 if (is_pointer_value(env, insn->src_reg)) {
12960 "R%d partial copy of pointer\n",
12963 } else if (src_reg->type == SCALAR_VALUE) {
12964 bool is_src_reg_u32 = src_reg->umax_value <= U32_MAX;
12966 if (is_src_reg_u32 && need_id)
12967 src_reg->id = ++env->id_gen;
12968 copy_register_state(dst_reg, src_reg);
12969 /* Make sure ID is cleared if src_reg is not in u32 range otherwise
12970 * dst_reg min/max could be incorrectly
12971 * propagated into src_reg by find_equal_scalars()
12973 if (!is_src_reg_u32)
12975 dst_reg->live |= REG_LIVE_WRITTEN;
12976 dst_reg->subreg_def = env->insn_idx + 1;
12978 mark_reg_unknown(env, regs,
12981 zext_32_to_64(dst_reg);
12982 reg_bounds_sync(dst_reg);
12986 * remember the value we stored into this reg
12988 /* clear any state __mark_reg_known doesn't set */
12989 mark_reg_unknown(env, regs, insn->dst_reg);
12990 regs[insn->dst_reg].type = SCALAR_VALUE;
12991 if (BPF_CLASS(insn->code) == BPF_ALU64) {
12992 __mark_reg_known(regs + insn->dst_reg,
12995 __mark_reg_known(regs + insn->dst_reg,
13000 } else if (opcode > BPF_END) {
13001 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
13004 } else { /* all other ALU ops: and, sub, xor, add, ... */
13006 if (BPF_SRC(insn->code) == BPF_X) {
13007 if (insn->imm != 0 || insn->off != 0) {
13008 verbose(env, "BPF_ALU uses reserved fields\n");
13011 /* check src1 operand */
13012 err = check_reg_arg(env, insn->src_reg, SRC_OP);
13016 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
13017 verbose(env, "BPF_ALU uses reserved fields\n");
13022 /* check src2 operand */
13023 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
13027 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
13028 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
13029 verbose(env, "div by zero\n");
13033 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
13034 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
13035 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
13037 if (insn->imm < 0 || insn->imm >= size) {
13038 verbose(env, "invalid shift %d\n", insn->imm);
13043 /* check dest operand */
13044 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
13048 return adjust_reg_min_max_vals(env, insn);
13054 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
13055 struct bpf_reg_state *dst_reg,
13056 enum bpf_reg_type type,
13057 bool range_right_open)
13059 struct bpf_func_state *state;
13060 struct bpf_reg_state *reg;
13063 if (dst_reg->off < 0 ||
13064 (dst_reg->off == 0 && range_right_open))
13065 /* This doesn't give us any range */
13068 if (dst_reg->umax_value > MAX_PACKET_OFF ||
13069 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
13070 /* Risk of overflow. For instance, ptr + (1<<63) may be less
13071 * than pkt_end, but that's because it's also less than pkt.
13075 new_range = dst_reg->off;
13076 if (range_right_open)
13079 /* Examples for register markings:
13081 * pkt_data in dst register:
13085 * if (r2 > pkt_end) goto <handle exception>
13090 * if (r2 < pkt_end) goto <access okay>
13091 * <handle exception>
13094 * r2 == dst_reg, pkt_end == src_reg
13095 * r2=pkt(id=n,off=8,r=0)
13096 * r3=pkt(id=n,off=0,r=0)
13098 * pkt_data in src register:
13102 * if (pkt_end >= r2) goto <access okay>
13103 * <handle exception>
13107 * if (pkt_end <= r2) goto <handle exception>
13111 * pkt_end == dst_reg, r2 == src_reg
13112 * r2=pkt(id=n,off=8,r=0)
13113 * r3=pkt(id=n,off=0,r=0)
13115 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
13116 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
13117 * and [r3, r3 + 8-1) respectively is safe to access depending on
13121 /* If our ids match, then we must have the same max_value. And we
13122 * don't care about the other reg's fixed offset, since if it's too big
13123 * the range won't allow anything.
13124 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
13126 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
13127 if (reg->type == type && reg->id == dst_reg->id)
13128 /* keep the maximum range already checked */
13129 reg->range = max(reg->range, new_range);
13133 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
13135 struct tnum subreg = tnum_subreg(reg->var_off);
13136 s32 sval = (s32)val;
13140 if (tnum_is_const(subreg))
13141 return !!tnum_equals_const(subreg, val);
13142 else if (val < reg->u32_min_value || val > reg->u32_max_value)
13146 if (tnum_is_const(subreg))
13147 return !tnum_equals_const(subreg, val);
13148 else if (val < reg->u32_min_value || val > reg->u32_max_value)
13152 if ((~subreg.mask & subreg.value) & val)
13154 if (!((subreg.mask | subreg.value) & val))
13158 if (reg->u32_min_value > val)
13160 else if (reg->u32_max_value <= val)
13164 if (reg->s32_min_value > sval)
13166 else if (reg->s32_max_value <= sval)
13170 if (reg->u32_max_value < val)
13172 else if (reg->u32_min_value >= val)
13176 if (reg->s32_max_value < sval)
13178 else if (reg->s32_min_value >= sval)
13182 if (reg->u32_min_value >= val)
13184 else if (reg->u32_max_value < val)
13188 if (reg->s32_min_value >= sval)
13190 else if (reg->s32_max_value < sval)
13194 if (reg->u32_max_value <= val)
13196 else if (reg->u32_min_value > val)
13200 if (reg->s32_max_value <= sval)
13202 else if (reg->s32_min_value > sval)
13211 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
13213 s64 sval = (s64)val;
13217 if (tnum_is_const(reg->var_off))
13218 return !!tnum_equals_const(reg->var_off, val);
13219 else if (val < reg->umin_value || val > reg->umax_value)
13223 if (tnum_is_const(reg->var_off))
13224 return !tnum_equals_const(reg->var_off, val);
13225 else if (val < reg->umin_value || val > reg->umax_value)
13229 if ((~reg->var_off.mask & reg->var_off.value) & val)
13231 if (!((reg->var_off.mask | reg->var_off.value) & val))
13235 if (reg->umin_value > val)
13237 else if (reg->umax_value <= val)
13241 if (reg->smin_value > sval)
13243 else if (reg->smax_value <= sval)
13247 if (reg->umax_value < val)
13249 else if (reg->umin_value >= val)
13253 if (reg->smax_value < sval)
13255 else if (reg->smin_value >= sval)
13259 if (reg->umin_value >= val)
13261 else if (reg->umax_value < val)
13265 if (reg->smin_value >= sval)
13267 else if (reg->smax_value < sval)
13271 if (reg->umax_value <= val)
13273 else if (reg->umin_value > val)
13277 if (reg->smax_value <= sval)
13279 else if (reg->smin_value > sval)
13287 /* compute branch direction of the expression "if (reg opcode val) goto target;"
13289 * 1 - branch will be taken and "goto target" will be executed
13290 * 0 - branch will not be taken and fall-through to next insn
13291 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
13294 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
13297 if (__is_pointer_value(false, reg)) {
13298 if (!reg_not_null(reg))
13301 /* If pointer is valid tests against zero will fail so we can
13302 * use this to direct branch taken.
13318 return is_branch32_taken(reg, val, opcode);
13319 return is_branch64_taken(reg, val, opcode);
13322 static int flip_opcode(u32 opcode)
13324 /* How can we transform "a <op> b" into "b <op> a"? */
13325 static const u8 opcode_flip[16] = {
13326 /* these stay the same */
13327 [BPF_JEQ >> 4] = BPF_JEQ,
13328 [BPF_JNE >> 4] = BPF_JNE,
13329 [BPF_JSET >> 4] = BPF_JSET,
13330 /* these swap "lesser" and "greater" (L and G in the opcodes) */
13331 [BPF_JGE >> 4] = BPF_JLE,
13332 [BPF_JGT >> 4] = BPF_JLT,
13333 [BPF_JLE >> 4] = BPF_JGE,
13334 [BPF_JLT >> 4] = BPF_JGT,
13335 [BPF_JSGE >> 4] = BPF_JSLE,
13336 [BPF_JSGT >> 4] = BPF_JSLT,
13337 [BPF_JSLE >> 4] = BPF_JSGE,
13338 [BPF_JSLT >> 4] = BPF_JSGT
13340 return opcode_flip[opcode >> 4];
13343 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
13344 struct bpf_reg_state *src_reg,
13347 struct bpf_reg_state *pkt;
13349 if (src_reg->type == PTR_TO_PACKET_END) {
13351 } else if (dst_reg->type == PTR_TO_PACKET_END) {
13353 opcode = flip_opcode(opcode);
13358 if (pkt->range >= 0)
13363 /* pkt <= pkt_end */
13366 /* pkt > pkt_end */
13367 if (pkt->range == BEYOND_PKT_END)
13368 /* pkt has at last one extra byte beyond pkt_end */
13369 return opcode == BPF_JGT;
13372 /* pkt < pkt_end */
13375 /* pkt >= pkt_end */
13376 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
13377 return opcode == BPF_JGE;
13383 /* Adjusts the register min/max values in the case that the dst_reg is the
13384 * variable register that we are working on, and src_reg is a constant or we're
13385 * simply doing a BPF_K check.
13386 * In JEQ/JNE cases we also adjust the var_off values.
13388 static void reg_set_min_max(struct bpf_reg_state *true_reg,
13389 struct bpf_reg_state *false_reg,
13390 u64 val, u32 val32,
13391 u8 opcode, bool is_jmp32)
13393 struct tnum false_32off = tnum_subreg(false_reg->var_off);
13394 struct tnum false_64off = false_reg->var_off;
13395 struct tnum true_32off = tnum_subreg(true_reg->var_off);
13396 struct tnum true_64off = true_reg->var_off;
13397 s64 sval = (s64)val;
13398 s32 sval32 = (s32)val32;
13400 /* If the dst_reg is a pointer, we can't learn anything about its
13401 * variable offset from the compare (unless src_reg were a pointer into
13402 * the same object, but we don't bother with that.
13403 * Since false_reg and true_reg have the same type by construction, we
13404 * only need to check one of them for pointerness.
13406 if (__is_pointer_value(false, false_reg))
13410 /* JEQ/JNE comparison doesn't change the register equivalence.
13413 * if (r1 == 42) goto label;
13415 * label: // here both r1 and r2 are known to be 42.
13417 * Hence when marking register as known preserve it's ID.
13421 __mark_reg32_known(true_reg, val32);
13422 true_32off = tnum_subreg(true_reg->var_off);
13424 ___mark_reg_known(true_reg, val);
13425 true_64off = true_reg->var_off;
13430 __mark_reg32_known(false_reg, val32);
13431 false_32off = tnum_subreg(false_reg->var_off);
13433 ___mark_reg_known(false_reg, val);
13434 false_64off = false_reg->var_off;
13439 false_32off = tnum_and(false_32off, tnum_const(~val32));
13440 if (is_power_of_2(val32))
13441 true_32off = tnum_or(true_32off,
13442 tnum_const(val32));
13444 false_64off = tnum_and(false_64off, tnum_const(~val));
13445 if (is_power_of_2(val))
13446 true_64off = tnum_or(true_64off,
13454 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
13455 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
13457 false_reg->u32_max_value = min(false_reg->u32_max_value,
13459 true_reg->u32_min_value = max(true_reg->u32_min_value,
13462 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
13463 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
13465 false_reg->umax_value = min(false_reg->umax_value, false_umax);
13466 true_reg->umin_value = max(true_reg->umin_value, true_umin);
13474 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
13475 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
13477 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
13478 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
13480 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
13481 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
13483 false_reg->smax_value = min(false_reg->smax_value, false_smax);
13484 true_reg->smin_value = max(true_reg->smin_value, true_smin);
13492 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
13493 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
13495 false_reg->u32_min_value = max(false_reg->u32_min_value,
13497 true_reg->u32_max_value = min(true_reg->u32_max_value,
13500 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
13501 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
13503 false_reg->umin_value = max(false_reg->umin_value, false_umin);
13504 true_reg->umax_value = min(true_reg->umax_value, true_umax);
13512 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
13513 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
13515 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
13516 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
13518 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
13519 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
13521 false_reg->smin_value = max(false_reg->smin_value, false_smin);
13522 true_reg->smax_value = min(true_reg->smax_value, true_smax);
13531 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
13532 tnum_subreg(false_32off));
13533 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
13534 tnum_subreg(true_32off));
13535 __reg_combine_32_into_64(false_reg);
13536 __reg_combine_32_into_64(true_reg);
13538 false_reg->var_off = false_64off;
13539 true_reg->var_off = true_64off;
13540 __reg_combine_64_into_32(false_reg);
13541 __reg_combine_64_into_32(true_reg);
13545 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
13546 * the variable reg.
13548 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
13549 struct bpf_reg_state *false_reg,
13550 u64 val, u32 val32,
13551 u8 opcode, bool is_jmp32)
13553 opcode = flip_opcode(opcode);
13554 /* This uses zero as "not present in table"; luckily the zero opcode,
13555 * BPF_JA, can't get here.
13558 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
13561 /* Regs are known to be equal, so intersect their min/max/var_off */
13562 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
13563 struct bpf_reg_state *dst_reg)
13565 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
13566 dst_reg->umin_value);
13567 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
13568 dst_reg->umax_value);
13569 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
13570 dst_reg->smin_value);
13571 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
13572 dst_reg->smax_value);
13573 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
13575 reg_bounds_sync(src_reg);
13576 reg_bounds_sync(dst_reg);
13579 static void reg_combine_min_max(struct bpf_reg_state *true_src,
13580 struct bpf_reg_state *true_dst,
13581 struct bpf_reg_state *false_src,
13582 struct bpf_reg_state *false_dst,
13587 __reg_combine_min_max(true_src, true_dst);
13590 __reg_combine_min_max(false_src, false_dst);
13595 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
13596 struct bpf_reg_state *reg, u32 id,
13599 if (type_may_be_null(reg->type) && reg->id == id &&
13600 (is_rcu_reg(reg) || !WARN_ON_ONCE(!reg->id))) {
13601 /* Old offset (both fixed and variable parts) should have been
13602 * known-zero, because we don't allow pointer arithmetic on
13603 * pointers that might be NULL. If we see this happening, don't
13604 * convert the register.
13606 * But in some cases, some helpers that return local kptrs
13607 * advance offset for the returned pointer. In those cases, it
13608 * is fine to expect to see reg->off.
13610 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value || !tnum_equals_const(reg->var_off, 0)))
13612 if (!(type_is_ptr_alloc_obj(reg->type) || type_is_non_owning_ref(reg->type)) &&
13613 WARN_ON_ONCE(reg->off))
13617 reg->type = SCALAR_VALUE;
13618 /* We don't need id and ref_obj_id from this point
13619 * onwards anymore, thus we should better reset it,
13620 * so that state pruning has chances to take effect.
13623 reg->ref_obj_id = 0;
13628 mark_ptr_not_null_reg(reg);
13630 if (!reg_may_point_to_spin_lock(reg)) {
13631 /* For not-NULL ptr, reg->ref_obj_id will be reset
13632 * in release_reference().
13634 * reg->id is still used by spin_lock ptr. Other
13635 * than spin_lock ptr type, reg->id can be reset.
13642 /* The logic is similar to find_good_pkt_pointers(), both could eventually
13643 * be folded together at some point.
13645 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
13648 struct bpf_func_state *state = vstate->frame[vstate->curframe];
13649 struct bpf_reg_state *regs = state->regs, *reg;
13650 u32 ref_obj_id = regs[regno].ref_obj_id;
13651 u32 id = regs[regno].id;
13653 if (ref_obj_id && ref_obj_id == id && is_null)
13654 /* regs[regno] is in the " == NULL" branch.
13655 * No one could have freed the reference state before
13656 * doing the NULL check.
13658 WARN_ON_ONCE(release_reference_state(state, id));
13660 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
13661 mark_ptr_or_null_reg(state, reg, id, is_null);
13665 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
13666 struct bpf_reg_state *dst_reg,
13667 struct bpf_reg_state *src_reg,
13668 struct bpf_verifier_state *this_branch,
13669 struct bpf_verifier_state *other_branch)
13671 if (BPF_SRC(insn->code) != BPF_X)
13674 /* Pointers are always 64-bit. */
13675 if (BPF_CLASS(insn->code) == BPF_JMP32)
13678 switch (BPF_OP(insn->code)) {
13680 if ((dst_reg->type == PTR_TO_PACKET &&
13681 src_reg->type == PTR_TO_PACKET_END) ||
13682 (dst_reg->type == PTR_TO_PACKET_META &&
13683 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13684 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
13685 find_good_pkt_pointers(this_branch, dst_reg,
13686 dst_reg->type, false);
13687 mark_pkt_end(other_branch, insn->dst_reg, true);
13688 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
13689 src_reg->type == PTR_TO_PACKET) ||
13690 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
13691 src_reg->type == PTR_TO_PACKET_META)) {
13692 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
13693 find_good_pkt_pointers(other_branch, src_reg,
13694 src_reg->type, true);
13695 mark_pkt_end(this_branch, insn->src_reg, false);
13701 if ((dst_reg->type == PTR_TO_PACKET &&
13702 src_reg->type == PTR_TO_PACKET_END) ||
13703 (dst_reg->type == PTR_TO_PACKET_META &&
13704 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13705 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
13706 find_good_pkt_pointers(other_branch, dst_reg,
13707 dst_reg->type, true);
13708 mark_pkt_end(this_branch, insn->dst_reg, false);
13709 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
13710 src_reg->type == PTR_TO_PACKET) ||
13711 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
13712 src_reg->type == PTR_TO_PACKET_META)) {
13713 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
13714 find_good_pkt_pointers(this_branch, src_reg,
13715 src_reg->type, false);
13716 mark_pkt_end(other_branch, insn->src_reg, true);
13722 if ((dst_reg->type == PTR_TO_PACKET &&
13723 src_reg->type == PTR_TO_PACKET_END) ||
13724 (dst_reg->type == PTR_TO_PACKET_META &&
13725 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13726 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
13727 find_good_pkt_pointers(this_branch, dst_reg,
13728 dst_reg->type, true);
13729 mark_pkt_end(other_branch, insn->dst_reg, false);
13730 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
13731 src_reg->type == PTR_TO_PACKET) ||
13732 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
13733 src_reg->type == PTR_TO_PACKET_META)) {
13734 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
13735 find_good_pkt_pointers(other_branch, src_reg,
13736 src_reg->type, false);
13737 mark_pkt_end(this_branch, insn->src_reg, true);
13743 if ((dst_reg->type == PTR_TO_PACKET &&
13744 src_reg->type == PTR_TO_PACKET_END) ||
13745 (dst_reg->type == PTR_TO_PACKET_META &&
13746 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13747 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
13748 find_good_pkt_pointers(other_branch, dst_reg,
13749 dst_reg->type, false);
13750 mark_pkt_end(this_branch, insn->dst_reg, true);
13751 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
13752 src_reg->type == PTR_TO_PACKET) ||
13753 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
13754 src_reg->type == PTR_TO_PACKET_META)) {
13755 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
13756 find_good_pkt_pointers(this_branch, src_reg,
13757 src_reg->type, true);
13758 mark_pkt_end(other_branch, insn->src_reg, false);
13770 static void find_equal_scalars(struct bpf_verifier_state *vstate,
13771 struct bpf_reg_state *known_reg)
13773 struct bpf_func_state *state;
13774 struct bpf_reg_state *reg;
13776 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
13777 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
13778 copy_register_state(reg, known_reg);
13782 static int check_cond_jmp_op(struct bpf_verifier_env *env,
13783 struct bpf_insn *insn, int *insn_idx)
13785 struct bpf_verifier_state *this_branch = env->cur_state;
13786 struct bpf_verifier_state *other_branch;
13787 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
13788 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
13789 struct bpf_reg_state *eq_branch_regs;
13790 u8 opcode = BPF_OP(insn->code);
13795 /* Only conditional jumps are expected to reach here. */
13796 if (opcode == BPF_JA || opcode > BPF_JSLE) {
13797 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
13801 if (BPF_SRC(insn->code) == BPF_X) {
13802 if (insn->imm != 0) {
13803 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
13807 /* check src1 operand */
13808 err = check_reg_arg(env, insn->src_reg, SRC_OP);
13812 if (is_pointer_value(env, insn->src_reg)) {
13813 verbose(env, "R%d pointer comparison prohibited\n",
13817 src_reg = ®s[insn->src_reg];
13819 if (insn->src_reg != BPF_REG_0) {
13820 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
13825 /* check src2 operand */
13826 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
13830 dst_reg = ®s[insn->dst_reg];
13831 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
13833 if (BPF_SRC(insn->code) == BPF_K) {
13834 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
13835 } else if (src_reg->type == SCALAR_VALUE &&
13836 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
13837 pred = is_branch_taken(dst_reg,
13838 tnum_subreg(src_reg->var_off).value,
13841 } else if (src_reg->type == SCALAR_VALUE &&
13842 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
13843 pred = is_branch_taken(dst_reg,
13844 src_reg->var_off.value,
13847 } else if (dst_reg->type == SCALAR_VALUE &&
13848 is_jmp32 && tnum_is_const(tnum_subreg(dst_reg->var_off))) {
13849 pred = is_branch_taken(src_reg,
13850 tnum_subreg(dst_reg->var_off).value,
13851 flip_opcode(opcode),
13853 } else if (dst_reg->type == SCALAR_VALUE &&
13854 !is_jmp32 && tnum_is_const(dst_reg->var_off)) {
13855 pred = is_branch_taken(src_reg,
13856 dst_reg->var_off.value,
13857 flip_opcode(opcode),
13859 } else if (reg_is_pkt_pointer_any(dst_reg) &&
13860 reg_is_pkt_pointer_any(src_reg) &&
13862 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
13866 /* If we get here with a dst_reg pointer type it is because
13867 * above is_branch_taken() special cased the 0 comparison.
13869 if (!__is_pointer_value(false, dst_reg))
13870 err = mark_chain_precision(env, insn->dst_reg);
13871 if (BPF_SRC(insn->code) == BPF_X && !err &&
13872 !__is_pointer_value(false, src_reg))
13873 err = mark_chain_precision(env, insn->src_reg);
13879 /* Only follow the goto, ignore fall-through. If needed, push
13880 * the fall-through branch for simulation under speculative
13883 if (!env->bypass_spec_v1 &&
13884 !sanitize_speculative_path(env, insn, *insn_idx + 1,
13887 *insn_idx += insn->off;
13889 } else if (pred == 0) {
13890 /* Only follow the fall-through branch, since that's where the
13891 * program will go. If needed, push the goto branch for
13892 * simulation under speculative execution.
13894 if (!env->bypass_spec_v1 &&
13895 !sanitize_speculative_path(env, insn,
13896 *insn_idx + insn->off + 1,
13902 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
13906 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
13908 /* detect if we are comparing against a constant value so we can adjust
13909 * our min/max values for our dst register.
13910 * this is only legit if both are scalars (or pointers to the same
13911 * object, I suppose, see the PTR_MAYBE_NULL related if block below),
13912 * because otherwise the different base pointers mean the offsets aren't
13915 if (BPF_SRC(insn->code) == BPF_X) {
13916 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
13918 if (dst_reg->type == SCALAR_VALUE &&
13919 src_reg->type == SCALAR_VALUE) {
13920 if (tnum_is_const(src_reg->var_off) ||
13922 tnum_is_const(tnum_subreg(src_reg->var_off))))
13923 reg_set_min_max(&other_branch_regs[insn->dst_reg],
13925 src_reg->var_off.value,
13926 tnum_subreg(src_reg->var_off).value,
13928 else if (tnum_is_const(dst_reg->var_off) ||
13930 tnum_is_const(tnum_subreg(dst_reg->var_off))))
13931 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
13933 dst_reg->var_off.value,
13934 tnum_subreg(dst_reg->var_off).value,
13936 else if (!is_jmp32 &&
13937 (opcode == BPF_JEQ || opcode == BPF_JNE))
13938 /* Comparing for equality, we can combine knowledge */
13939 reg_combine_min_max(&other_branch_regs[insn->src_reg],
13940 &other_branch_regs[insn->dst_reg],
13941 src_reg, dst_reg, opcode);
13943 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
13944 find_equal_scalars(this_branch, src_reg);
13945 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
13949 } else if (dst_reg->type == SCALAR_VALUE) {
13950 reg_set_min_max(&other_branch_regs[insn->dst_reg],
13951 dst_reg, insn->imm, (u32)insn->imm,
13955 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
13956 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
13957 find_equal_scalars(this_branch, dst_reg);
13958 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
13961 /* if one pointer register is compared to another pointer
13962 * register check if PTR_MAYBE_NULL could be lifted.
13963 * E.g. register A - maybe null
13964 * register B - not null
13965 * for JNE A, B, ... - A is not null in the false branch;
13966 * for JEQ A, B, ... - A is not null in the true branch.
13968 * Since PTR_TO_BTF_ID points to a kernel struct that does
13969 * not need to be null checked by the BPF program, i.e.,
13970 * could be null even without PTR_MAYBE_NULL marking, so
13971 * only propagate nullness when neither reg is that type.
13973 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_X &&
13974 __is_pointer_value(false, src_reg) && __is_pointer_value(false, dst_reg) &&
13975 type_may_be_null(src_reg->type) != type_may_be_null(dst_reg->type) &&
13976 base_type(src_reg->type) != PTR_TO_BTF_ID &&
13977 base_type(dst_reg->type) != PTR_TO_BTF_ID) {
13978 eq_branch_regs = NULL;
13981 eq_branch_regs = other_branch_regs;
13984 eq_branch_regs = regs;
13990 if (eq_branch_regs) {
13991 if (type_may_be_null(src_reg->type))
13992 mark_ptr_not_null_reg(&eq_branch_regs[insn->src_reg]);
13994 mark_ptr_not_null_reg(&eq_branch_regs[insn->dst_reg]);
13998 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
13999 * NOTE: these optimizations below are related with pointer comparison
14000 * which will never be JMP32.
14002 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
14003 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
14004 type_may_be_null(dst_reg->type)) {
14005 /* Mark all identical registers in each branch as either
14006 * safe or unknown depending R == 0 or R != 0 conditional.
14008 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
14009 opcode == BPF_JNE);
14010 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
14011 opcode == BPF_JEQ);
14012 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
14013 this_branch, other_branch) &&
14014 is_pointer_value(env, insn->dst_reg)) {
14015 verbose(env, "R%d pointer comparison prohibited\n",
14019 if (env->log.level & BPF_LOG_LEVEL)
14020 print_insn_state(env, this_branch->frame[this_branch->curframe]);
14024 /* verify BPF_LD_IMM64 instruction */
14025 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
14027 struct bpf_insn_aux_data *aux = cur_aux(env);
14028 struct bpf_reg_state *regs = cur_regs(env);
14029 struct bpf_reg_state *dst_reg;
14030 struct bpf_map *map;
14033 if (BPF_SIZE(insn->code) != BPF_DW) {
14034 verbose(env, "invalid BPF_LD_IMM insn\n");
14037 if (insn->off != 0) {
14038 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
14042 err = check_reg_arg(env, insn->dst_reg, DST_OP);
14046 dst_reg = ®s[insn->dst_reg];
14047 if (insn->src_reg == 0) {
14048 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
14050 dst_reg->type = SCALAR_VALUE;
14051 __mark_reg_known(®s[insn->dst_reg], imm);
14055 /* All special src_reg cases are listed below. From this point onwards
14056 * we either succeed and assign a corresponding dst_reg->type after
14057 * zeroing the offset, or fail and reject the program.
14059 mark_reg_known_zero(env, regs, insn->dst_reg);
14061 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
14062 dst_reg->type = aux->btf_var.reg_type;
14063 switch (base_type(dst_reg->type)) {
14065 dst_reg->mem_size = aux->btf_var.mem_size;
14067 case PTR_TO_BTF_ID:
14068 dst_reg->btf = aux->btf_var.btf;
14069 dst_reg->btf_id = aux->btf_var.btf_id;
14072 verbose(env, "bpf verifier is misconfigured\n");
14078 if (insn->src_reg == BPF_PSEUDO_FUNC) {
14079 struct bpf_prog_aux *aux = env->prog->aux;
14080 u32 subprogno = find_subprog(env,
14081 env->insn_idx + insn->imm + 1);
14083 if (!aux->func_info) {
14084 verbose(env, "missing btf func_info\n");
14087 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
14088 verbose(env, "callback function not static\n");
14092 dst_reg->type = PTR_TO_FUNC;
14093 dst_reg->subprogno = subprogno;
14097 map = env->used_maps[aux->map_index];
14098 dst_reg->map_ptr = map;
14100 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
14101 insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
14102 dst_reg->type = PTR_TO_MAP_VALUE;
14103 dst_reg->off = aux->map_off;
14104 WARN_ON_ONCE(map->max_entries != 1);
14105 /* We want reg->id to be same (0) as map_value is not distinct */
14106 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
14107 insn->src_reg == BPF_PSEUDO_MAP_IDX) {
14108 dst_reg->type = CONST_PTR_TO_MAP;
14110 verbose(env, "bpf verifier is misconfigured\n");
14117 static bool may_access_skb(enum bpf_prog_type type)
14120 case BPF_PROG_TYPE_SOCKET_FILTER:
14121 case BPF_PROG_TYPE_SCHED_CLS:
14122 case BPF_PROG_TYPE_SCHED_ACT:
14129 /* verify safety of LD_ABS|LD_IND instructions:
14130 * - they can only appear in the programs where ctx == skb
14131 * - since they are wrappers of function calls, they scratch R1-R5 registers,
14132 * preserve R6-R9, and store return value into R0
14135 * ctx == skb == R6 == CTX
14138 * SRC == any register
14139 * IMM == 32-bit immediate
14142 * R0 - 8/16/32-bit skb data converted to cpu endianness
14144 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
14146 struct bpf_reg_state *regs = cur_regs(env);
14147 static const int ctx_reg = BPF_REG_6;
14148 u8 mode = BPF_MODE(insn->code);
14151 if (!may_access_skb(resolve_prog_type(env->prog))) {
14152 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
14156 if (!env->ops->gen_ld_abs) {
14157 verbose(env, "bpf verifier is misconfigured\n");
14161 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
14162 BPF_SIZE(insn->code) == BPF_DW ||
14163 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
14164 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
14168 /* check whether implicit source operand (register R6) is readable */
14169 err = check_reg_arg(env, ctx_reg, SRC_OP);
14173 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
14174 * gen_ld_abs() may terminate the program at runtime, leading to
14177 err = check_reference_leak(env);
14179 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
14183 if (env->cur_state->active_lock.ptr) {
14184 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
14188 if (env->cur_state->active_rcu_lock) {
14189 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_rcu_read_lock-ed region\n");
14193 if (regs[ctx_reg].type != PTR_TO_CTX) {
14195 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
14199 if (mode == BPF_IND) {
14200 /* check explicit source operand */
14201 err = check_reg_arg(env, insn->src_reg, SRC_OP);
14206 err = check_ptr_off_reg(env, ®s[ctx_reg], ctx_reg);
14210 /* reset caller saved regs to unreadable */
14211 for (i = 0; i < CALLER_SAVED_REGS; i++) {
14212 mark_reg_not_init(env, regs, caller_saved[i]);
14213 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
14216 /* mark destination R0 register as readable, since it contains
14217 * the value fetched from the packet.
14218 * Already marked as written above.
14220 mark_reg_unknown(env, regs, BPF_REG_0);
14221 /* ld_abs load up to 32-bit skb data. */
14222 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
14226 static int check_return_code(struct bpf_verifier_env *env)
14228 struct tnum enforce_attach_type_range = tnum_unknown;
14229 const struct bpf_prog *prog = env->prog;
14230 struct bpf_reg_state *reg;
14231 struct tnum range = tnum_range(0, 1);
14232 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
14234 struct bpf_func_state *frame = env->cur_state->frame[0];
14235 const bool is_subprog = frame->subprogno;
14237 /* LSM and struct_ops func-ptr's return type could be "void" */
14239 switch (prog_type) {
14240 case BPF_PROG_TYPE_LSM:
14241 if (prog->expected_attach_type == BPF_LSM_CGROUP)
14242 /* See below, can be 0 or 0-1 depending on hook. */
14245 case BPF_PROG_TYPE_STRUCT_OPS:
14246 if (!prog->aux->attach_func_proto->type)
14254 /* eBPF calling convention is such that R0 is used
14255 * to return the value from eBPF program.
14256 * Make sure that it's readable at this time
14257 * of bpf_exit, which means that program wrote
14258 * something into it earlier
14260 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
14264 if (is_pointer_value(env, BPF_REG_0)) {
14265 verbose(env, "R0 leaks addr as return value\n");
14269 reg = cur_regs(env) + BPF_REG_0;
14271 if (frame->in_async_callback_fn) {
14272 /* enforce return zero from async callbacks like timer */
14273 if (reg->type != SCALAR_VALUE) {
14274 verbose(env, "In async callback the register R0 is not a known value (%s)\n",
14275 reg_type_str(env, reg->type));
14279 if (!tnum_in(tnum_const(0), reg->var_off)) {
14280 verbose_invalid_scalar(env, reg, &range, "async callback", "R0");
14287 if (reg->type != SCALAR_VALUE) {
14288 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
14289 reg_type_str(env, reg->type));
14295 switch (prog_type) {
14296 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
14297 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
14298 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
14299 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
14300 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
14301 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
14302 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
14303 range = tnum_range(1, 1);
14304 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
14305 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
14306 range = tnum_range(0, 3);
14308 case BPF_PROG_TYPE_CGROUP_SKB:
14309 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
14310 range = tnum_range(0, 3);
14311 enforce_attach_type_range = tnum_range(2, 3);
14314 case BPF_PROG_TYPE_CGROUP_SOCK:
14315 case BPF_PROG_TYPE_SOCK_OPS:
14316 case BPF_PROG_TYPE_CGROUP_DEVICE:
14317 case BPF_PROG_TYPE_CGROUP_SYSCTL:
14318 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
14320 case BPF_PROG_TYPE_RAW_TRACEPOINT:
14321 if (!env->prog->aux->attach_btf_id)
14323 range = tnum_const(0);
14325 case BPF_PROG_TYPE_TRACING:
14326 switch (env->prog->expected_attach_type) {
14327 case BPF_TRACE_FENTRY:
14328 case BPF_TRACE_FEXIT:
14329 range = tnum_const(0);
14331 case BPF_TRACE_RAW_TP:
14332 case BPF_MODIFY_RETURN:
14334 case BPF_TRACE_ITER:
14340 case BPF_PROG_TYPE_SK_LOOKUP:
14341 range = tnum_range(SK_DROP, SK_PASS);
14344 case BPF_PROG_TYPE_LSM:
14345 if (env->prog->expected_attach_type != BPF_LSM_CGROUP) {
14346 /* Regular BPF_PROG_TYPE_LSM programs can return
14351 if (!env->prog->aux->attach_func_proto->type) {
14352 /* Make sure programs that attach to void
14353 * hooks don't try to modify return value.
14355 range = tnum_range(1, 1);
14359 case BPF_PROG_TYPE_NETFILTER:
14360 range = tnum_range(NF_DROP, NF_ACCEPT);
14362 case BPF_PROG_TYPE_EXT:
14363 /* freplace program can return anything as its return value
14364 * depends on the to-be-replaced kernel func or bpf program.
14370 if (reg->type != SCALAR_VALUE) {
14371 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
14372 reg_type_str(env, reg->type));
14376 if (!tnum_in(range, reg->var_off)) {
14377 verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
14378 if (prog->expected_attach_type == BPF_LSM_CGROUP &&
14379 prog_type == BPF_PROG_TYPE_LSM &&
14380 !prog->aux->attach_func_proto->type)
14381 verbose(env, "Note, BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
14385 if (!tnum_is_unknown(enforce_attach_type_range) &&
14386 tnum_in(enforce_attach_type_range, reg->var_off))
14387 env->prog->enforce_expected_attach_type = 1;
14391 /* non-recursive DFS pseudo code
14392 * 1 procedure DFS-iterative(G,v):
14393 * 2 label v as discovered
14394 * 3 let S be a stack
14396 * 5 while S is not empty
14398 * 7 if t is what we're looking for:
14400 * 9 for all edges e in G.adjacentEdges(t) do
14401 * 10 if edge e is already labelled
14402 * 11 continue with the next edge
14403 * 12 w <- G.adjacentVertex(t,e)
14404 * 13 if vertex w is not discovered and not explored
14405 * 14 label e as tree-edge
14406 * 15 label w as discovered
14409 * 18 else if vertex w is discovered
14410 * 19 label e as back-edge
14412 * 21 // vertex w is explored
14413 * 22 label e as forward- or cross-edge
14414 * 23 label t as explored
14418 * 0x10 - discovered
14419 * 0x11 - discovered and fall-through edge labelled
14420 * 0x12 - discovered and fall-through and branch edges labelled
14431 static u32 state_htab_size(struct bpf_verifier_env *env)
14433 return env->prog->len;
14436 static struct bpf_verifier_state_list **explored_state(
14437 struct bpf_verifier_env *env,
14440 struct bpf_verifier_state *cur = env->cur_state;
14441 struct bpf_func_state *state = cur->frame[cur->curframe];
14443 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
14446 static void mark_prune_point(struct bpf_verifier_env *env, int idx)
14448 env->insn_aux_data[idx].prune_point = true;
14451 static bool is_prune_point(struct bpf_verifier_env *env, int insn_idx)
14453 return env->insn_aux_data[insn_idx].prune_point;
14456 static void mark_force_checkpoint(struct bpf_verifier_env *env, int idx)
14458 env->insn_aux_data[idx].force_checkpoint = true;
14461 static bool is_force_checkpoint(struct bpf_verifier_env *env, int insn_idx)
14463 return env->insn_aux_data[insn_idx].force_checkpoint;
14468 DONE_EXPLORING = 0,
14469 KEEP_EXPLORING = 1,
14472 /* t, w, e - match pseudo-code above:
14473 * t - index of current instruction
14474 * w - next instruction
14477 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
14480 int *insn_stack = env->cfg.insn_stack;
14481 int *insn_state = env->cfg.insn_state;
14483 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
14484 return DONE_EXPLORING;
14486 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
14487 return DONE_EXPLORING;
14489 if (w < 0 || w >= env->prog->len) {
14490 verbose_linfo(env, t, "%d: ", t);
14491 verbose(env, "jump out of range from insn %d to %d\n", t, w);
14496 /* mark branch target for state pruning */
14497 mark_prune_point(env, w);
14498 mark_jmp_point(env, w);
14501 if (insn_state[w] == 0) {
14503 insn_state[t] = DISCOVERED | e;
14504 insn_state[w] = DISCOVERED;
14505 if (env->cfg.cur_stack >= env->prog->len)
14507 insn_stack[env->cfg.cur_stack++] = w;
14508 return KEEP_EXPLORING;
14509 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
14510 if (loop_ok && env->bpf_capable)
14511 return DONE_EXPLORING;
14512 verbose_linfo(env, t, "%d: ", t);
14513 verbose_linfo(env, w, "%d: ", w);
14514 verbose(env, "back-edge from insn %d to %d\n", t, w);
14516 } else if (insn_state[w] == EXPLORED) {
14517 /* forward- or cross-edge */
14518 insn_state[t] = DISCOVERED | e;
14520 verbose(env, "insn state internal bug\n");
14523 return DONE_EXPLORING;
14526 static int visit_func_call_insn(int t, struct bpf_insn *insns,
14527 struct bpf_verifier_env *env,
14532 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
14536 mark_prune_point(env, t + 1);
14537 /* when we exit from subprog, we need to record non-linear history */
14538 mark_jmp_point(env, t + 1);
14540 if (visit_callee) {
14541 mark_prune_point(env, t);
14542 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env,
14543 /* It's ok to allow recursion from CFG point of
14544 * view. __check_func_call() will do the actual
14547 bpf_pseudo_func(insns + t));
14552 /* Visits the instruction at index t and returns one of the following:
14553 * < 0 - an error occurred
14554 * DONE_EXPLORING - the instruction was fully explored
14555 * KEEP_EXPLORING - there is still work to be done before it is fully explored
14557 static int visit_insn(int t, struct bpf_verifier_env *env)
14559 struct bpf_insn *insns = env->prog->insnsi, *insn = &insns[t];
14562 if (bpf_pseudo_func(insn))
14563 return visit_func_call_insn(t, insns, env, true);
14565 /* All non-branch instructions have a single fall-through edge. */
14566 if (BPF_CLASS(insn->code) != BPF_JMP &&
14567 BPF_CLASS(insn->code) != BPF_JMP32)
14568 return push_insn(t, t + 1, FALLTHROUGH, env, false);
14570 switch (BPF_OP(insn->code)) {
14572 return DONE_EXPLORING;
14575 if (insn->src_reg == 0 && insn->imm == BPF_FUNC_timer_set_callback)
14576 /* Mark this call insn as a prune point to trigger
14577 * is_state_visited() check before call itself is
14578 * processed by __check_func_call(). Otherwise new
14579 * async state will be pushed for further exploration.
14581 mark_prune_point(env, t);
14582 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
14583 struct bpf_kfunc_call_arg_meta meta;
14585 ret = fetch_kfunc_meta(env, insn, &meta, NULL);
14586 if (ret == 0 && is_iter_next_kfunc(&meta)) {
14587 mark_prune_point(env, t);
14588 /* Checking and saving state checkpoints at iter_next() call
14589 * is crucial for fast convergence of open-coded iterator loop
14590 * logic, so we need to force it. If we don't do that,
14591 * is_state_visited() might skip saving a checkpoint, causing
14592 * unnecessarily long sequence of not checkpointed
14593 * instructions and jumps, leading to exhaustion of jump
14594 * history buffer, and potentially other undesired outcomes.
14595 * It is expected that with correct open-coded iterators
14596 * convergence will happen quickly, so we don't run a risk of
14597 * exhausting memory.
14599 mark_force_checkpoint(env, t);
14602 return visit_func_call_insn(t, insns, env, insn->src_reg == BPF_PSEUDO_CALL);
14605 if (BPF_SRC(insn->code) != BPF_K)
14608 /* unconditional jump with single edge */
14609 ret = push_insn(t, t + insn->off + 1, FALLTHROUGH, env,
14614 mark_prune_point(env, t + insn->off + 1);
14615 mark_jmp_point(env, t + insn->off + 1);
14620 /* conditional jump with two edges */
14621 mark_prune_point(env, t);
14623 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
14627 return push_insn(t, t + insn->off + 1, BRANCH, env, true);
14631 /* non-recursive depth-first-search to detect loops in BPF program
14632 * loop == back-edge in directed graph
14634 static int check_cfg(struct bpf_verifier_env *env)
14636 int insn_cnt = env->prog->len;
14637 int *insn_stack, *insn_state;
14641 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
14645 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
14647 kvfree(insn_state);
14651 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
14652 insn_stack[0] = 0; /* 0 is the first instruction */
14653 env->cfg.cur_stack = 1;
14655 while (env->cfg.cur_stack > 0) {
14656 int t = insn_stack[env->cfg.cur_stack - 1];
14658 ret = visit_insn(t, env);
14660 case DONE_EXPLORING:
14661 insn_state[t] = EXPLORED;
14662 env->cfg.cur_stack--;
14664 case KEEP_EXPLORING:
14668 verbose(env, "visit_insn internal bug\n");
14675 if (env->cfg.cur_stack < 0) {
14676 verbose(env, "pop stack internal bug\n");
14681 for (i = 0; i < insn_cnt; i++) {
14682 if (insn_state[i] != EXPLORED) {
14683 verbose(env, "unreachable insn %d\n", i);
14688 ret = 0; /* cfg looks good */
14691 kvfree(insn_state);
14692 kvfree(insn_stack);
14693 env->cfg.insn_state = env->cfg.insn_stack = NULL;
14697 static int check_abnormal_return(struct bpf_verifier_env *env)
14701 for (i = 1; i < env->subprog_cnt; i++) {
14702 if (env->subprog_info[i].has_ld_abs) {
14703 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
14706 if (env->subprog_info[i].has_tail_call) {
14707 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
14714 /* The minimum supported BTF func info size */
14715 #define MIN_BPF_FUNCINFO_SIZE 8
14716 #define MAX_FUNCINFO_REC_SIZE 252
14718 static int check_btf_func(struct bpf_verifier_env *env,
14719 const union bpf_attr *attr,
14722 const struct btf_type *type, *func_proto, *ret_type;
14723 u32 i, nfuncs, urec_size, min_size;
14724 u32 krec_size = sizeof(struct bpf_func_info);
14725 struct bpf_func_info *krecord;
14726 struct bpf_func_info_aux *info_aux = NULL;
14727 struct bpf_prog *prog;
14728 const struct btf *btf;
14730 u32 prev_offset = 0;
14731 bool scalar_return;
14734 nfuncs = attr->func_info_cnt;
14736 if (check_abnormal_return(env))
14741 if (nfuncs != env->subprog_cnt) {
14742 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
14746 urec_size = attr->func_info_rec_size;
14747 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
14748 urec_size > MAX_FUNCINFO_REC_SIZE ||
14749 urec_size % sizeof(u32)) {
14750 verbose(env, "invalid func info rec size %u\n", urec_size);
14755 btf = prog->aux->btf;
14757 urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
14758 min_size = min_t(u32, krec_size, urec_size);
14760 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
14763 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
14767 for (i = 0; i < nfuncs; i++) {
14768 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
14770 if (ret == -E2BIG) {
14771 verbose(env, "nonzero tailing record in func info");
14772 /* set the size kernel expects so loader can zero
14773 * out the rest of the record.
14775 if (copy_to_bpfptr_offset(uattr,
14776 offsetof(union bpf_attr, func_info_rec_size),
14777 &min_size, sizeof(min_size)))
14783 if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
14788 /* check insn_off */
14791 if (krecord[i].insn_off) {
14793 "nonzero insn_off %u for the first func info record",
14794 krecord[i].insn_off);
14797 } else if (krecord[i].insn_off <= prev_offset) {
14799 "same or smaller insn offset (%u) than previous func info record (%u)",
14800 krecord[i].insn_off, prev_offset);
14804 if (env->subprog_info[i].start != krecord[i].insn_off) {
14805 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
14809 /* check type_id */
14810 type = btf_type_by_id(btf, krecord[i].type_id);
14811 if (!type || !btf_type_is_func(type)) {
14812 verbose(env, "invalid type id %d in func info",
14813 krecord[i].type_id);
14816 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
14818 func_proto = btf_type_by_id(btf, type->type);
14819 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
14820 /* btf_func_check() already verified it during BTF load */
14822 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
14824 btf_type_is_small_int(ret_type) || btf_is_any_enum(ret_type);
14825 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
14826 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
14829 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
14830 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
14834 prev_offset = krecord[i].insn_off;
14835 bpfptr_add(&urecord, urec_size);
14838 prog->aux->func_info = krecord;
14839 prog->aux->func_info_cnt = nfuncs;
14840 prog->aux->func_info_aux = info_aux;
14849 static void adjust_btf_func(struct bpf_verifier_env *env)
14851 struct bpf_prog_aux *aux = env->prog->aux;
14854 if (!aux->func_info)
14857 for (i = 0; i < env->subprog_cnt; i++)
14858 aux->func_info[i].insn_off = env->subprog_info[i].start;
14861 #define MIN_BPF_LINEINFO_SIZE offsetofend(struct bpf_line_info, line_col)
14862 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
14864 static int check_btf_line(struct bpf_verifier_env *env,
14865 const union bpf_attr *attr,
14868 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
14869 struct bpf_subprog_info *sub;
14870 struct bpf_line_info *linfo;
14871 struct bpf_prog *prog;
14872 const struct btf *btf;
14876 nr_linfo = attr->line_info_cnt;
14879 if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info))
14882 rec_size = attr->line_info_rec_size;
14883 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
14884 rec_size > MAX_LINEINFO_REC_SIZE ||
14885 rec_size & (sizeof(u32) - 1))
14888 /* Need to zero it in case the userspace may
14889 * pass in a smaller bpf_line_info object.
14891 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
14892 GFP_KERNEL | __GFP_NOWARN);
14897 btf = prog->aux->btf;
14900 sub = env->subprog_info;
14901 ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
14902 expected_size = sizeof(struct bpf_line_info);
14903 ncopy = min_t(u32, expected_size, rec_size);
14904 for (i = 0; i < nr_linfo; i++) {
14905 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
14907 if (err == -E2BIG) {
14908 verbose(env, "nonzero tailing record in line_info");
14909 if (copy_to_bpfptr_offset(uattr,
14910 offsetof(union bpf_attr, line_info_rec_size),
14911 &expected_size, sizeof(expected_size)))
14917 if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
14923 * Check insn_off to ensure
14924 * 1) strictly increasing AND
14925 * 2) bounded by prog->len
14927 * The linfo[0].insn_off == 0 check logically falls into
14928 * the later "missing bpf_line_info for func..." case
14929 * because the first linfo[0].insn_off must be the
14930 * first sub also and the first sub must have
14931 * subprog_info[0].start == 0.
14933 if ((i && linfo[i].insn_off <= prev_offset) ||
14934 linfo[i].insn_off >= prog->len) {
14935 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
14936 i, linfo[i].insn_off, prev_offset,
14942 if (!prog->insnsi[linfo[i].insn_off].code) {
14944 "Invalid insn code at line_info[%u].insn_off\n",
14950 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
14951 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
14952 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
14957 if (s != env->subprog_cnt) {
14958 if (linfo[i].insn_off == sub[s].start) {
14959 sub[s].linfo_idx = i;
14961 } else if (sub[s].start < linfo[i].insn_off) {
14962 verbose(env, "missing bpf_line_info for func#%u\n", s);
14968 prev_offset = linfo[i].insn_off;
14969 bpfptr_add(&ulinfo, rec_size);
14972 if (s != env->subprog_cnt) {
14973 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
14974 env->subprog_cnt - s, s);
14979 prog->aux->linfo = linfo;
14980 prog->aux->nr_linfo = nr_linfo;
14989 #define MIN_CORE_RELO_SIZE sizeof(struct bpf_core_relo)
14990 #define MAX_CORE_RELO_SIZE MAX_FUNCINFO_REC_SIZE
14992 static int check_core_relo(struct bpf_verifier_env *env,
14993 const union bpf_attr *attr,
14996 u32 i, nr_core_relo, ncopy, expected_size, rec_size;
14997 struct bpf_core_relo core_relo = {};
14998 struct bpf_prog *prog = env->prog;
14999 const struct btf *btf = prog->aux->btf;
15000 struct bpf_core_ctx ctx = {
15004 bpfptr_t u_core_relo;
15007 nr_core_relo = attr->core_relo_cnt;
15010 if (nr_core_relo > INT_MAX / sizeof(struct bpf_core_relo))
15013 rec_size = attr->core_relo_rec_size;
15014 if (rec_size < MIN_CORE_RELO_SIZE ||
15015 rec_size > MAX_CORE_RELO_SIZE ||
15016 rec_size % sizeof(u32))
15019 u_core_relo = make_bpfptr(attr->core_relos, uattr.is_kernel);
15020 expected_size = sizeof(struct bpf_core_relo);
15021 ncopy = min_t(u32, expected_size, rec_size);
15023 /* Unlike func_info and line_info, copy and apply each CO-RE
15024 * relocation record one at a time.
15026 for (i = 0; i < nr_core_relo; i++) {
15027 /* future proofing when sizeof(bpf_core_relo) changes */
15028 err = bpf_check_uarg_tail_zero(u_core_relo, expected_size, rec_size);
15030 if (err == -E2BIG) {
15031 verbose(env, "nonzero tailing record in core_relo");
15032 if (copy_to_bpfptr_offset(uattr,
15033 offsetof(union bpf_attr, core_relo_rec_size),
15034 &expected_size, sizeof(expected_size)))
15040 if (copy_from_bpfptr(&core_relo, u_core_relo, ncopy)) {
15045 if (core_relo.insn_off % 8 || core_relo.insn_off / 8 >= prog->len) {
15046 verbose(env, "Invalid core_relo[%u].insn_off:%u prog->len:%u\n",
15047 i, core_relo.insn_off, prog->len);
15052 err = bpf_core_apply(&ctx, &core_relo, i,
15053 &prog->insnsi[core_relo.insn_off / 8]);
15056 bpfptr_add(&u_core_relo, rec_size);
15061 static int check_btf_info(struct bpf_verifier_env *env,
15062 const union bpf_attr *attr,
15068 if (!attr->func_info_cnt && !attr->line_info_cnt) {
15069 if (check_abnormal_return(env))
15074 btf = btf_get_by_fd(attr->prog_btf_fd);
15076 return PTR_ERR(btf);
15077 if (btf_is_kernel(btf)) {
15081 env->prog->aux->btf = btf;
15083 err = check_btf_func(env, attr, uattr);
15087 err = check_btf_line(env, attr, uattr);
15091 err = check_core_relo(env, attr, uattr);
15098 /* check %cur's range satisfies %old's */
15099 static bool range_within(struct bpf_reg_state *old,
15100 struct bpf_reg_state *cur)
15102 return old->umin_value <= cur->umin_value &&
15103 old->umax_value >= cur->umax_value &&
15104 old->smin_value <= cur->smin_value &&
15105 old->smax_value >= cur->smax_value &&
15106 old->u32_min_value <= cur->u32_min_value &&
15107 old->u32_max_value >= cur->u32_max_value &&
15108 old->s32_min_value <= cur->s32_min_value &&
15109 old->s32_max_value >= cur->s32_max_value;
15112 /* If in the old state two registers had the same id, then they need to have
15113 * the same id in the new state as well. But that id could be different from
15114 * the old state, so we need to track the mapping from old to new ids.
15115 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
15116 * regs with old id 5 must also have new id 9 for the new state to be safe. But
15117 * regs with a different old id could still have new id 9, we don't care about
15119 * So we look through our idmap to see if this old id has been seen before. If
15120 * so, we require the new id to match; otherwise, we add the id pair to the map.
15122 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap)
15124 struct bpf_id_pair *map = idmap->map;
15127 /* either both IDs should be set or both should be zero */
15128 if (!!old_id != !!cur_id)
15131 if (old_id == 0) /* cur_id == 0 as well */
15134 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
15136 /* Reached an empty slot; haven't seen this id before */
15137 map[i].old = old_id;
15138 map[i].cur = cur_id;
15141 if (map[i].old == old_id)
15142 return map[i].cur == cur_id;
15143 if (map[i].cur == cur_id)
15146 /* We ran out of idmap slots, which should be impossible */
15151 /* Similar to check_ids(), but allocate a unique temporary ID
15152 * for 'old_id' or 'cur_id' of zero.
15153 * This makes pairs like '0 vs unique ID', 'unique ID vs 0' valid.
15155 static bool check_scalar_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap)
15157 old_id = old_id ? old_id : ++idmap->tmp_id_gen;
15158 cur_id = cur_id ? cur_id : ++idmap->tmp_id_gen;
15160 return check_ids(old_id, cur_id, idmap);
15163 static void clean_func_state(struct bpf_verifier_env *env,
15164 struct bpf_func_state *st)
15166 enum bpf_reg_liveness live;
15169 for (i = 0; i < BPF_REG_FP; i++) {
15170 live = st->regs[i].live;
15171 /* liveness must not touch this register anymore */
15172 st->regs[i].live |= REG_LIVE_DONE;
15173 if (!(live & REG_LIVE_READ))
15174 /* since the register is unused, clear its state
15175 * to make further comparison simpler
15177 __mark_reg_not_init(env, &st->regs[i]);
15180 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
15181 live = st->stack[i].spilled_ptr.live;
15182 /* liveness must not touch this stack slot anymore */
15183 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
15184 if (!(live & REG_LIVE_READ)) {
15185 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
15186 for (j = 0; j < BPF_REG_SIZE; j++)
15187 st->stack[i].slot_type[j] = STACK_INVALID;
15192 static void clean_verifier_state(struct bpf_verifier_env *env,
15193 struct bpf_verifier_state *st)
15197 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
15198 /* all regs in this state in all frames were already marked */
15201 for (i = 0; i <= st->curframe; i++)
15202 clean_func_state(env, st->frame[i]);
15205 /* the parentage chains form a tree.
15206 * the verifier states are added to state lists at given insn and
15207 * pushed into state stack for future exploration.
15208 * when the verifier reaches bpf_exit insn some of the verifer states
15209 * stored in the state lists have their final liveness state already,
15210 * but a lot of states will get revised from liveness point of view when
15211 * the verifier explores other branches.
15214 * 2: if r1 == 100 goto pc+1
15217 * when the verifier reaches exit insn the register r0 in the state list of
15218 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
15219 * of insn 2 and goes exploring further. At the insn 4 it will walk the
15220 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
15222 * Since the verifier pushes the branch states as it sees them while exploring
15223 * the program the condition of walking the branch instruction for the second
15224 * time means that all states below this branch were already explored and
15225 * their final liveness marks are already propagated.
15226 * Hence when the verifier completes the search of state list in is_state_visited()
15227 * we can call this clean_live_states() function to mark all liveness states
15228 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
15229 * will not be used.
15230 * This function also clears the registers and stack for states that !READ
15231 * to simplify state merging.
15233 * Important note here that walking the same branch instruction in the callee
15234 * doesn't meant that the states are DONE. The verifier has to compare
15237 static void clean_live_states(struct bpf_verifier_env *env, int insn,
15238 struct bpf_verifier_state *cur)
15240 struct bpf_verifier_state_list *sl;
15243 sl = *explored_state(env, insn);
15245 if (sl->state.branches)
15247 if (sl->state.insn_idx != insn ||
15248 sl->state.curframe != cur->curframe)
15250 for (i = 0; i <= cur->curframe; i++)
15251 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
15253 clean_verifier_state(env, &sl->state);
15259 static bool regs_exact(const struct bpf_reg_state *rold,
15260 const struct bpf_reg_state *rcur,
15261 struct bpf_idmap *idmap)
15263 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
15264 check_ids(rold->id, rcur->id, idmap) &&
15265 check_ids(rold->ref_obj_id, rcur->ref_obj_id, idmap);
15268 /* Returns true if (rold safe implies rcur safe) */
15269 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
15270 struct bpf_reg_state *rcur, struct bpf_idmap *idmap)
15272 if (!(rold->live & REG_LIVE_READ))
15273 /* explored state didn't use this */
15275 if (rold->type == NOT_INIT)
15276 /* explored state can't have used this */
15278 if (rcur->type == NOT_INIT)
15281 /* Enforce that register types have to match exactly, including their
15282 * modifiers (like PTR_MAYBE_NULL, MEM_RDONLY, etc), as a general
15285 * One can make a point that using a pointer register as unbounded
15286 * SCALAR would be technically acceptable, but this could lead to
15287 * pointer leaks because scalars are allowed to leak while pointers
15288 * are not. We could make this safe in special cases if root is
15289 * calling us, but it's probably not worth the hassle.
15291 * Also, register types that are *not* MAYBE_NULL could technically be
15292 * safe to use as their MAYBE_NULL variants (e.g., PTR_TO_MAP_VALUE
15293 * is safe to be used as PTR_TO_MAP_VALUE_OR_NULL, provided both point
15294 * to the same map).
15295 * However, if the old MAYBE_NULL register then got NULL checked,
15296 * doing so could have affected others with the same id, and we can't
15297 * check for that because we lost the id when we converted to
15298 * a non-MAYBE_NULL variant.
15299 * So, as a general rule we don't allow mixing MAYBE_NULL and
15300 * non-MAYBE_NULL registers as well.
15302 if (rold->type != rcur->type)
15305 switch (base_type(rold->type)) {
15307 if (env->explore_alu_limits) {
15308 /* explore_alu_limits disables tnum_in() and range_within()
15309 * logic and requires everything to be strict
15311 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
15312 check_scalar_ids(rold->id, rcur->id, idmap);
15314 if (!rold->precise)
15316 /* Why check_ids() for scalar registers?
15318 * Consider the following BPF code:
15319 * 1: r6 = ... unbound scalar, ID=a ...
15320 * 2: r7 = ... unbound scalar, ID=b ...
15321 * 3: if (r6 > r7) goto +1
15323 * 5: if (r6 > X) goto ...
15324 * 6: ... memory operation using r7 ...
15326 * First verification path is [1-6]:
15327 * - at (4) same bpf_reg_state::id (b) would be assigned to r6 and r7;
15328 * - at (5) r6 would be marked <= X, find_equal_scalars() would also mark
15329 * r7 <= X, because r6 and r7 share same id.
15330 * Next verification path is [1-4, 6].
15332 * Instruction (6) would be reached in two states:
15333 * I. r6{.id=b}, r7{.id=b} via path 1-6;
15334 * II. r6{.id=a}, r7{.id=b} via path 1-4, 6.
15336 * Use check_ids() to distinguish these states.
15338 * Also verify that new value satisfies old value range knowledge.
15340 return range_within(rold, rcur) &&
15341 tnum_in(rold->var_off, rcur->var_off) &&
15342 check_scalar_ids(rold->id, rcur->id, idmap);
15343 case PTR_TO_MAP_KEY:
15344 case PTR_TO_MAP_VALUE:
15347 case PTR_TO_TP_BUFFER:
15348 /* If the new min/max/var_off satisfy the old ones and
15349 * everything else matches, we are OK.
15351 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, var_off)) == 0 &&
15352 range_within(rold, rcur) &&
15353 tnum_in(rold->var_off, rcur->var_off) &&
15354 check_ids(rold->id, rcur->id, idmap) &&
15355 check_ids(rold->ref_obj_id, rcur->ref_obj_id, idmap);
15356 case PTR_TO_PACKET_META:
15357 case PTR_TO_PACKET:
15358 /* We must have at least as much range as the old ptr
15359 * did, so that any accesses which were safe before are
15360 * still safe. This is true even if old range < old off,
15361 * since someone could have accessed through (ptr - k), or
15362 * even done ptr -= k in a register, to get a safe access.
15364 if (rold->range > rcur->range)
15366 /* If the offsets don't match, we can't trust our alignment;
15367 * nor can we be sure that we won't fall out of range.
15369 if (rold->off != rcur->off)
15371 /* id relations must be preserved */
15372 if (!check_ids(rold->id, rcur->id, idmap))
15374 /* new val must satisfy old val knowledge */
15375 return range_within(rold, rcur) &&
15376 tnum_in(rold->var_off, rcur->var_off);
15378 /* two stack pointers are equal only if they're pointing to
15379 * the same stack frame, since fp-8 in foo != fp-8 in bar
15381 return regs_exact(rold, rcur, idmap) && rold->frameno == rcur->frameno;
15383 return regs_exact(rold, rcur, idmap);
15387 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
15388 struct bpf_func_state *cur, struct bpf_idmap *idmap)
15392 /* walk slots of the explored stack and ignore any additional
15393 * slots in the current stack, since explored(safe) state
15396 for (i = 0; i < old->allocated_stack; i++) {
15397 struct bpf_reg_state *old_reg, *cur_reg;
15399 spi = i / BPF_REG_SIZE;
15401 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
15402 i += BPF_REG_SIZE - 1;
15403 /* explored state didn't use this */
15407 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
15410 if (env->allow_uninit_stack &&
15411 old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC)
15414 /* explored stack has more populated slots than current stack
15415 * and these slots were used
15417 if (i >= cur->allocated_stack)
15420 /* if old state was safe with misc data in the stack
15421 * it will be safe with zero-initialized stack.
15422 * The opposite is not true
15424 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
15425 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
15427 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
15428 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
15429 /* Ex: old explored (safe) state has STACK_SPILL in
15430 * this stack slot, but current has STACK_MISC ->
15431 * this verifier states are not equivalent,
15432 * return false to continue verification of this path
15435 if (i % BPF_REG_SIZE != BPF_REG_SIZE - 1)
15437 /* Both old and cur are having same slot_type */
15438 switch (old->stack[spi].slot_type[BPF_REG_SIZE - 1]) {
15440 /* when explored and current stack slot are both storing
15441 * spilled registers, check that stored pointers types
15442 * are the same as well.
15443 * Ex: explored safe path could have stored
15444 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
15445 * but current path has stored:
15446 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
15447 * such verifier states are not equivalent.
15448 * return false to continue verification of this path
15450 if (!regsafe(env, &old->stack[spi].spilled_ptr,
15451 &cur->stack[spi].spilled_ptr, idmap))
15455 old_reg = &old->stack[spi].spilled_ptr;
15456 cur_reg = &cur->stack[spi].spilled_ptr;
15457 if (old_reg->dynptr.type != cur_reg->dynptr.type ||
15458 old_reg->dynptr.first_slot != cur_reg->dynptr.first_slot ||
15459 !check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap))
15463 old_reg = &old->stack[spi].spilled_ptr;
15464 cur_reg = &cur->stack[spi].spilled_ptr;
15465 /* iter.depth is not compared between states as it
15466 * doesn't matter for correctness and would otherwise
15467 * prevent convergence; we maintain it only to prevent
15468 * infinite loop check triggering, see
15469 * iter_active_depths_differ()
15471 if (old_reg->iter.btf != cur_reg->iter.btf ||
15472 old_reg->iter.btf_id != cur_reg->iter.btf_id ||
15473 old_reg->iter.state != cur_reg->iter.state ||
15474 /* ignore {old_reg,cur_reg}->iter.depth, see above */
15475 !check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap))
15480 case STACK_INVALID:
15482 /* Ensure that new unhandled slot types return false by default */
15490 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur,
15491 struct bpf_idmap *idmap)
15495 if (old->acquired_refs != cur->acquired_refs)
15498 for (i = 0; i < old->acquired_refs; i++) {
15499 if (!check_ids(old->refs[i].id, cur->refs[i].id, idmap))
15506 /* compare two verifier states
15508 * all states stored in state_list are known to be valid, since
15509 * verifier reached 'bpf_exit' instruction through them
15511 * this function is called when verifier exploring different branches of
15512 * execution popped from the state stack. If it sees an old state that has
15513 * more strict register state and more strict stack state then this execution
15514 * branch doesn't need to be explored further, since verifier already
15515 * concluded that more strict state leads to valid finish.
15517 * Therefore two states are equivalent if register state is more conservative
15518 * and explored stack state is more conservative than the current one.
15521 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
15522 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
15524 * In other words if current stack state (one being explored) has more
15525 * valid slots than old one that already passed validation, it means
15526 * the verifier can stop exploring and conclude that current state is valid too
15528 * Similarly with registers. If explored state has register type as invalid
15529 * whereas register type in current state is meaningful, it means that
15530 * the current state will reach 'bpf_exit' instruction safely
15532 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
15533 struct bpf_func_state *cur)
15537 for (i = 0; i < MAX_BPF_REG; i++)
15538 if (!regsafe(env, &old->regs[i], &cur->regs[i],
15539 &env->idmap_scratch))
15542 if (!stacksafe(env, old, cur, &env->idmap_scratch))
15545 if (!refsafe(old, cur, &env->idmap_scratch))
15551 static bool states_equal(struct bpf_verifier_env *env,
15552 struct bpf_verifier_state *old,
15553 struct bpf_verifier_state *cur)
15557 if (old->curframe != cur->curframe)
15560 env->idmap_scratch.tmp_id_gen = env->id_gen;
15561 memset(&env->idmap_scratch.map, 0, sizeof(env->idmap_scratch.map));
15563 /* Verification state from speculative execution simulation
15564 * must never prune a non-speculative execution one.
15566 if (old->speculative && !cur->speculative)
15569 if (old->active_lock.ptr != cur->active_lock.ptr)
15572 /* Old and cur active_lock's have to be either both present
15575 if (!!old->active_lock.id != !!cur->active_lock.id)
15578 if (old->active_lock.id &&
15579 !check_ids(old->active_lock.id, cur->active_lock.id, &env->idmap_scratch))
15582 if (old->active_rcu_lock != cur->active_rcu_lock)
15585 /* for states to be equal callsites have to be the same
15586 * and all frame states need to be equivalent
15588 for (i = 0; i <= old->curframe; i++) {
15589 if (old->frame[i]->callsite != cur->frame[i]->callsite)
15591 if (!func_states_equal(env, old->frame[i], cur->frame[i]))
15597 /* Return 0 if no propagation happened. Return negative error code if error
15598 * happened. Otherwise, return the propagated bit.
15600 static int propagate_liveness_reg(struct bpf_verifier_env *env,
15601 struct bpf_reg_state *reg,
15602 struct bpf_reg_state *parent_reg)
15604 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
15605 u8 flag = reg->live & REG_LIVE_READ;
15608 /* When comes here, read flags of PARENT_REG or REG could be any of
15609 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
15610 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
15612 if (parent_flag == REG_LIVE_READ64 ||
15613 /* Or if there is no read flag from REG. */
15615 /* Or if the read flag from REG is the same as PARENT_REG. */
15616 parent_flag == flag)
15619 err = mark_reg_read(env, reg, parent_reg, flag);
15626 /* A write screens off any subsequent reads; but write marks come from the
15627 * straight-line code between a state and its parent. When we arrive at an
15628 * equivalent state (jump target or such) we didn't arrive by the straight-line
15629 * code, so read marks in the state must propagate to the parent regardless
15630 * of the state's write marks. That's what 'parent == state->parent' comparison
15631 * in mark_reg_read() is for.
15633 static int propagate_liveness(struct bpf_verifier_env *env,
15634 const struct bpf_verifier_state *vstate,
15635 struct bpf_verifier_state *vparent)
15637 struct bpf_reg_state *state_reg, *parent_reg;
15638 struct bpf_func_state *state, *parent;
15639 int i, frame, err = 0;
15641 if (vparent->curframe != vstate->curframe) {
15642 WARN(1, "propagate_live: parent frame %d current frame %d\n",
15643 vparent->curframe, vstate->curframe);
15646 /* Propagate read liveness of registers... */
15647 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
15648 for (frame = 0; frame <= vstate->curframe; frame++) {
15649 parent = vparent->frame[frame];
15650 state = vstate->frame[frame];
15651 parent_reg = parent->regs;
15652 state_reg = state->regs;
15653 /* We don't need to worry about FP liveness, it's read-only */
15654 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
15655 err = propagate_liveness_reg(env, &state_reg[i],
15659 if (err == REG_LIVE_READ64)
15660 mark_insn_zext(env, &parent_reg[i]);
15663 /* Propagate stack slots. */
15664 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
15665 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
15666 parent_reg = &parent->stack[i].spilled_ptr;
15667 state_reg = &state->stack[i].spilled_ptr;
15668 err = propagate_liveness_reg(env, state_reg,
15677 /* find precise scalars in the previous equivalent state and
15678 * propagate them into the current state
15680 static int propagate_precision(struct bpf_verifier_env *env,
15681 const struct bpf_verifier_state *old)
15683 struct bpf_reg_state *state_reg;
15684 struct bpf_func_state *state;
15685 int i, err = 0, fr;
15688 for (fr = old->curframe; fr >= 0; fr--) {
15689 state = old->frame[fr];
15690 state_reg = state->regs;
15692 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
15693 if (state_reg->type != SCALAR_VALUE ||
15694 !state_reg->precise ||
15695 !(state_reg->live & REG_LIVE_READ))
15697 if (env->log.level & BPF_LOG_LEVEL2) {
15699 verbose(env, "frame %d: propagating r%d", fr, i);
15701 verbose(env, ",r%d", i);
15703 bt_set_frame_reg(&env->bt, fr, i);
15707 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
15708 if (!is_spilled_reg(&state->stack[i]))
15710 state_reg = &state->stack[i].spilled_ptr;
15711 if (state_reg->type != SCALAR_VALUE ||
15712 !state_reg->precise ||
15713 !(state_reg->live & REG_LIVE_READ))
15715 if (env->log.level & BPF_LOG_LEVEL2) {
15717 verbose(env, "frame %d: propagating fp%d",
15718 fr, (-i - 1) * BPF_REG_SIZE);
15720 verbose(env, ",fp%d", (-i - 1) * BPF_REG_SIZE);
15722 bt_set_frame_slot(&env->bt, fr, i);
15726 verbose(env, "\n");
15729 err = mark_chain_precision_batch(env);
15736 static bool states_maybe_looping(struct bpf_verifier_state *old,
15737 struct bpf_verifier_state *cur)
15739 struct bpf_func_state *fold, *fcur;
15740 int i, fr = cur->curframe;
15742 if (old->curframe != fr)
15745 fold = old->frame[fr];
15746 fcur = cur->frame[fr];
15747 for (i = 0; i < MAX_BPF_REG; i++)
15748 if (memcmp(&fold->regs[i], &fcur->regs[i],
15749 offsetof(struct bpf_reg_state, parent)))
15754 static bool is_iter_next_insn(struct bpf_verifier_env *env, int insn_idx)
15756 return env->insn_aux_data[insn_idx].is_iter_next;
15759 /* is_state_visited() handles iter_next() (see process_iter_next_call() for
15760 * terminology) calls specially: as opposed to bounded BPF loops, it *expects*
15761 * states to match, which otherwise would look like an infinite loop. So while
15762 * iter_next() calls are taken care of, we still need to be careful and
15763 * prevent erroneous and too eager declaration of "ininite loop", when
15764 * iterators are involved.
15766 * Here's a situation in pseudo-BPF assembly form:
15768 * 0: again: ; set up iter_next() call args
15769 * 1: r1 = &it ; <CHECKPOINT HERE>
15770 * 2: call bpf_iter_num_next ; this is iter_next() call
15771 * 3: if r0 == 0 goto done
15772 * 4: ... something useful here ...
15773 * 5: goto again ; another iteration
15776 * 8: call bpf_iter_num_destroy ; clean up iter state
15779 * This is a typical loop. Let's assume that we have a prune point at 1:,
15780 * before we get to `call bpf_iter_num_next` (e.g., because of that `goto
15781 * again`, assuming other heuristics don't get in a way).
15783 * When we first time come to 1:, let's say we have some state X. We proceed
15784 * to 2:, fork states, enqueue ACTIVE, validate NULL case successfully, exit.
15785 * Now we come back to validate that forked ACTIVE state. We proceed through
15786 * 3-5, come to goto, jump to 1:. Let's assume our state didn't change, so we
15787 * are converging. But the problem is that we don't know that yet, as this
15788 * convergence has to happen at iter_next() call site only. So if nothing is
15789 * done, at 1: verifier will use bounded loop logic and declare infinite
15790 * looping (and would be *technically* correct, if not for iterator's
15791 * "eventual sticky NULL" contract, see process_iter_next_call()). But we
15792 * don't want that. So what we do in process_iter_next_call() when we go on
15793 * another ACTIVE iteration, we bump slot->iter.depth, to mark that it's
15794 * a different iteration. So when we suspect an infinite loop, we additionally
15795 * check if any of the *ACTIVE* iterator states depths differ. If yes, we
15796 * pretend we are not looping and wait for next iter_next() call.
15798 * This only applies to ACTIVE state. In DRAINED state we don't expect to
15799 * loop, because that would actually mean infinite loop, as DRAINED state is
15800 * "sticky", and so we'll keep returning into the same instruction with the
15801 * same state (at least in one of possible code paths).
15803 * This approach allows to keep infinite loop heuristic even in the face of
15804 * active iterator. E.g., C snippet below is and will be detected as
15805 * inifintely looping:
15807 * struct bpf_iter_num it;
15810 * bpf_iter_num_new(&it, 0, 10);
15811 * while ((p = bpf_iter_num_next(&t))) {
15813 * while (x--) {} // <<-- infinite loop here
15817 static bool iter_active_depths_differ(struct bpf_verifier_state *old, struct bpf_verifier_state *cur)
15819 struct bpf_reg_state *slot, *cur_slot;
15820 struct bpf_func_state *state;
15823 for (fr = old->curframe; fr >= 0; fr--) {
15824 state = old->frame[fr];
15825 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
15826 if (state->stack[i].slot_type[0] != STACK_ITER)
15829 slot = &state->stack[i].spilled_ptr;
15830 if (slot->iter.state != BPF_ITER_STATE_ACTIVE)
15833 cur_slot = &cur->frame[fr]->stack[i].spilled_ptr;
15834 if (cur_slot->iter.depth != slot->iter.depth)
15841 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
15843 struct bpf_verifier_state_list *new_sl;
15844 struct bpf_verifier_state_list *sl, **pprev;
15845 struct bpf_verifier_state *cur = env->cur_state, *new;
15846 int i, j, err, states_cnt = 0;
15847 bool force_new_state = env->test_state_freq || is_force_checkpoint(env, insn_idx);
15848 bool add_new_state = force_new_state;
15850 /* bpf progs typically have pruning point every 4 instructions
15851 * http://vger.kernel.org/bpfconf2019.html#session-1
15852 * Do not add new state for future pruning if the verifier hasn't seen
15853 * at least 2 jumps and at least 8 instructions.
15854 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
15855 * In tests that amounts to up to 50% reduction into total verifier
15856 * memory consumption and 20% verifier time speedup.
15858 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
15859 env->insn_processed - env->prev_insn_processed >= 8)
15860 add_new_state = true;
15862 pprev = explored_state(env, insn_idx);
15865 clean_live_states(env, insn_idx, cur);
15869 if (sl->state.insn_idx != insn_idx)
15872 if (sl->state.branches) {
15873 struct bpf_func_state *frame = sl->state.frame[sl->state.curframe];
15875 if (frame->in_async_callback_fn &&
15876 frame->async_entry_cnt != cur->frame[cur->curframe]->async_entry_cnt) {
15877 /* Different async_entry_cnt means that the verifier is
15878 * processing another entry into async callback.
15879 * Seeing the same state is not an indication of infinite
15880 * loop or infinite recursion.
15881 * But finding the same state doesn't mean that it's safe
15882 * to stop processing the current state. The previous state
15883 * hasn't yet reached bpf_exit, since state.branches > 0.
15884 * Checking in_async_callback_fn alone is not enough either.
15885 * Since the verifier still needs to catch infinite loops
15886 * inside async callbacks.
15888 goto skip_inf_loop_check;
15890 /* BPF open-coded iterators loop detection is special.
15891 * states_maybe_looping() logic is too simplistic in detecting
15892 * states that *might* be equivalent, because it doesn't know
15893 * about ID remapping, so don't even perform it.
15894 * See process_iter_next_call() and iter_active_depths_differ()
15895 * for overview of the logic. When current and one of parent
15896 * states are detected as equivalent, it's a good thing: we prove
15897 * convergence and can stop simulating further iterations.
15898 * It's safe to assume that iterator loop will finish, taking into
15899 * account iter_next() contract of eventually returning
15900 * sticky NULL result.
15902 if (is_iter_next_insn(env, insn_idx)) {
15903 if (states_equal(env, &sl->state, cur)) {
15904 struct bpf_func_state *cur_frame;
15905 struct bpf_reg_state *iter_state, *iter_reg;
15908 cur_frame = cur->frame[cur->curframe];
15909 /* btf_check_iter_kfuncs() enforces that
15910 * iter state pointer is always the first arg
15912 iter_reg = &cur_frame->regs[BPF_REG_1];
15913 /* current state is valid due to states_equal(),
15914 * so we can assume valid iter and reg state,
15915 * no need for extra (re-)validations
15917 spi = __get_spi(iter_reg->off + iter_reg->var_off.value);
15918 iter_state = &func(env, iter_reg)->stack[spi].spilled_ptr;
15919 if (iter_state->iter.state == BPF_ITER_STATE_ACTIVE)
15922 goto skip_inf_loop_check;
15924 /* attempt to detect infinite loop to avoid unnecessary doomed work */
15925 if (states_maybe_looping(&sl->state, cur) &&
15926 states_equal(env, &sl->state, cur) &&
15927 !iter_active_depths_differ(&sl->state, cur)) {
15928 verbose_linfo(env, insn_idx, "; ");
15929 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
15932 /* if the verifier is processing a loop, avoid adding new state
15933 * too often, since different loop iterations have distinct
15934 * states and may not help future pruning.
15935 * This threshold shouldn't be too low to make sure that
15936 * a loop with large bound will be rejected quickly.
15937 * The most abusive loop will be:
15939 * if r1 < 1000000 goto pc-2
15940 * 1M insn_procssed limit / 100 == 10k peak states.
15941 * This threshold shouldn't be too high either, since states
15942 * at the end of the loop are likely to be useful in pruning.
15944 skip_inf_loop_check:
15945 if (!force_new_state &&
15946 env->jmps_processed - env->prev_jmps_processed < 20 &&
15947 env->insn_processed - env->prev_insn_processed < 100)
15948 add_new_state = false;
15951 if (states_equal(env, &sl->state, cur)) {
15954 /* reached equivalent register/stack state,
15955 * prune the search.
15956 * Registers read by the continuation are read by us.
15957 * If we have any write marks in env->cur_state, they
15958 * will prevent corresponding reads in the continuation
15959 * from reaching our parent (an explored_state). Our
15960 * own state will get the read marks recorded, but
15961 * they'll be immediately forgotten as we're pruning
15962 * this state and will pop a new one.
15964 err = propagate_liveness(env, &sl->state, cur);
15966 /* if previous state reached the exit with precision and
15967 * current state is equivalent to it (except precsion marks)
15968 * the precision needs to be propagated back in
15969 * the current state.
15971 err = err ? : push_jmp_history(env, cur);
15972 err = err ? : propagate_precision(env, &sl->state);
15978 /* when new state is not going to be added do not increase miss count.
15979 * Otherwise several loop iterations will remove the state
15980 * recorded earlier. The goal of these heuristics is to have
15981 * states from some iterations of the loop (some in the beginning
15982 * and some at the end) to help pruning.
15986 /* heuristic to determine whether this state is beneficial
15987 * to keep checking from state equivalence point of view.
15988 * Higher numbers increase max_states_per_insn and verification time,
15989 * but do not meaningfully decrease insn_processed.
15991 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
15992 /* the state is unlikely to be useful. Remove it to
15993 * speed up verification
15996 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
15997 u32 br = sl->state.branches;
16000 "BUG live_done but branches_to_explore %d\n",
16002 free_verifier_state(&sl->state, false);
16004 env->peak_states--;
16006 /* cannot free this state, since parentage chain may
16007 * walk it later. Add it for free_list instead to
16008 * be freed at the end of verification
16010 sl->next = env->free_list;
16011 env->free_list = sl;
16021 if (env->max_states_per_insn < states_cnt)
16022 env->max_states_per_insn = states_cnt;
16024 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
16027 if (!add_new_state)
16030 /* There were no equivalent states, remember the current one.
16031 * Technically the current state is not proven to be safe yet,
16032 * but it will either reach outer most bpf_exit (which means it's safe)
16033 * or it will be rejected. When there are no loops the verifier won't be
16034 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
16035 * again on the way to bpf_exit.
16036 * When looping the sl->state.branches will be > 0 and this state
16037 * will not be considered for equivalence until branches == 0.
16039 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
16042 env->total_states++;
16043 env->peak_states++;
16044 env->prev_jmps_processed = env->jmps_processed;
16045 env->prev_insn_processed = env->insn_processed;
16047 /* forget precise markings we inherited, see __mark_chain_precision */
16048 if (env->bpf_capable)
16049 mark_all_scalars_imprecise(env, cur);
16051 /* add new state to the head of linked list */
16052 new = &new_sl->state;
16053 err = copy_verifier_state(new, cur);
16055 free_verifier_state(new, false);
16059 new->insn_idx = insn_idx;
16060 WARN_ONCE(new->branches != 1,
16061 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
16064 cur->first_insn_idx = insn_idx;
16065 clear_jmp_history(cur);
16066 new_sl->next = *explored_state(env, insn_idx);
16067 *explored_state(env, insn_idx) = new_sl;
16068 /* connect new state to parentage chain. Current frame needs all
16069 * registers connected. Only r6 - r9 of the callers are alive (pushed
16070 * to the stack implicitly by JITs) so in callers' frames connect just
16071 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
16072 * the state of the call instruction (with WRITTEN set), and r0 comes
16073 * from callee with its full parentage chain, anyway.
16075 /* clear write marks in current state: the writes we did are not writes
16076 * our child did, so they don't screen off its reads from us.
16077 * (There are no read marks in current state, because reads always mark
16078 * their parent and current state never has children yet. Only
16079 * explored_states can get read marks.)
16081 for (j = 0; j <= cur->curframe; j++) {
16082 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
16083 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
16084 for (i = 0; i < BPF_REG_FP; i++)
16085 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
16088 /* all stack frames are accessible from callee, clear them all */
16089 for (j = 0; j <= cur->curframe; j++) {
16090 struct bpf_func_state *frame = cur->frame[j];
16091 struct bpf_func_state *newframe = new->frame[j];
16093 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
16094 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
16095 frame->stack[i].spilled_ptr.parent =
16096 &newframe->stack[i].spilled_ptr;
16102 /* Return true if it's OK to have the same insn return a different type. */
16103 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
16105 switch (base_type(type)) {
16107 case PTR_TO_SOCKET:
16108 case PTR_TO_SOCK_COMMON:
16109 case PTR_TO_TCP_SOCK:
16110 case PTR_TO_XDP_SOCK:
16111 case PTR_TO_BTF_ID:
16118 /* If an instruction was previously used with particular pointer types, then we
16119 * need to be careful to avoid cases such as the below, where it may be ok
16120 * for one branch accessing the pointer, but not ok for the other branch:
16125 * R1 = some_other_valid_ptr;
16128 * R2 = *(u32 *)(R1 + 0);
16130 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
16132 return src != prev && (!reg_type_mismatch_ok(src) ||
16133 !reg_type_mismatch_ok(prev));
16136 static int save_aux_ptr_type(struct bpf_verifier_env *env, enum bpf_reg_type type,
16137 bool allow_trust_missmatch)
16139 enum bpf_reg_type *prev_type = &env->insn_aux_data[env->insn_idx].ptr_type;
16141 if (*prev_type == NOT_INIT) {
16142 /* Saw a valid insn
16143 * dst_reg = *(u32 *)(src_reg + off)
16144 * save type to validate intersecting paths
16147 } else if (reg_type_mismatch(type, *prev_type)) {
16148 /* Abuser program is trying to use the same insn
16149 * dst_reg = *(u32*) (src_reg + off)
16150 * with different pointer types:
16151 * src_reg == ctx in one branch and
16152 * src_reg == stack|map in some other branch.
16155 if (allow_trust_missmatch &&
16156 base_type(type) == PTR_TO_BTF_ID &&
16157 base_type(*prev_type) == PTR_TO_BTF_ID) {
16159 * Have to support a use case when one path through
16160 * the program yields TRUSTED pointer while another
16161 * is UNTRUSTED. Fallback to UNTRUSTED to generate
16164 *prev_type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
16166 verbose(env, "same insn cannot be used with different pointers\n");
16174 static int do_check(struct bpf_verifier_env *env)
16176 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
16177 struct bpf_verifier_state *state = env->cur_state;
16178 struct bpf_insn *insns = env->prog->insnsi;
16179 struct bpf_reg_state *regs;
16180 int insn_cnt = env->prog->len;
16181 bool do_print_state = false;
16182 int prev_insn_idx = -1;
16185 struct bpf_insn *insn;
16189 env->prev_insn_idx = prev_insn_idx;
16190 if (env->insn_idx >= insn_cnt) {
16191 verbose(env, "invalid insn idx %d insn_cnt %d\n",
16192 env->insn_idx, insn_cnt);
16196 insn = &insns[env->insn_idx];
16197 class = BPF_CLASS(insn->code);
16199 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
16201 "BPF program is too large. Processed %d insn\n",
16202 env->insn_processed);
16206 state->last_insn_idx = env->prev_insn_idx;
16208 if (is_prune_point(env, env->insn_idx)) {
16209 err = is_state_visited(env, env->insn_idx);
16213 /* found equivalent state, can prune the search */
16214 if (env->log.level & BPF_LOG_LEVEL) {
16215 if (do_print_state)
16216 verbose(env, "\nfrom %d to %d%s: safe\n",
16217 env->prev_insn_idx, env->insn_idx,
16218 env->cur_state->speculative ?
16219 " (speculative execution)" : "");
16221 verbose(env, "%d: safe\n", env->insn_idx);
16223 goto process_bpf_exit;
16227 if (is_jmp_point(env, env->insn_idx)) {
16228 err = push_jmp_history(env, state);
16233 if (signal_pending(current))
16236 if (need_resched())
16239 if (env->log.level & BPF_LOG_LEVEL2 && do_print_state) {
16240 verbose(env, "\nfrom %d to %d%s:",
16241 env->prev_insn_idx, env->insn_idx,
16242 env->cur_state->speculative ?
16243 " (speculative execution)" : "");
16244 print_verifier_state(env, state->frame[state->curframe], true);
16245 do_print_state = false;
16248 if (env->log.level & BPF_LOG_LEVEL) {
16249 const struct bpf_insn_cbs cbs = {
16250 .cb_call = disasm_kfunc_name,
16251 .cb_print = verbose,
16252 .private_data = env,
16255 if (verifier_state_scratched(env))
16256 print_insn_state(env, state->frame[state->curframe]);
16258 verbose_linfo(env, env->insn_idx, "; ");
16259 env->prev_log_pos = env->log.end_pos;
16260 verbose(env, "%d: ", env->insn_idx);
16261 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
16262 env->prev_insn_print_pos = env->log.end_pos - env->prev_log_pos;
16263 env->prev_log_pos = env->log.end_pos;
16266 if (bpf_prog_is_offloaded(env->prog->aux)) {
16267 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
16268 env->prev_insn_idx);
16273 regs = cur_regs(env);
16274 sanitize_mark_insn_seen(env);
16275 prev_insn_idx = env->insn_idx;
16277 if (class == BPF_ALU || class == BPF_ALU64) {
16278 err = check_alu_op(env, insn);
16282 } else if (class == BPF_LDX) {
16283 enum bpf_reg_type src_reg_type;
16285 /* check for reserved fields is already done */
16287 /* check src operand */
16288 err = check_reg_arg(env, insn->src_reg, SRC_OP);
16292 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
16296 src_reg_type = regs[insn->src_reg].type;
16298 /* check that memory (src_reg + off) is readable,
16299 * the state of dst_reg will be updated by this func
16301 err = check_mem_access(env, env->insn_idx, insn->src_reg,
16302 insn->off, BPF_SIZE(insn->code),
16303 BPF_READ, insn->dst_reg, false);
16307 err = save_aux_ptr_type(env, src_reg_type, true);
16310 } else if (class == BPF_STX) {
16311 enum bpf_reg_type dst_reg_type;
16313 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
16314 err = check_atomic(env, env->insn_idx, insn);
16321 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
16322 verbose(env, "BPF_STX uses reserved fields\n");
16326 /* check src1 operand */
16327 err = check_reg_arg(env, insn->src_reg, SRC_OP);
16330 /* check src2 operand */
16331 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
16335 dst_reg_type = regs[insn->dst_reg].type;
16337 /* check that memory (dst_reg + off) is writeable */
16338 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
16339 insn->off, BPF_SIZE(insn->code),
16340 BPF_WRITE, insn->src_reg, false);
16344 err = save_aux_ptr_type(env, dst_reg_type, false);
16347 } else if (class == BPF_ST) {
16348 enum bpf_reg_type dst_reg_type;
16350 if (BPF_MODE(insn->code) != BPF_MEM ||
16351 insn->src_reg != BPF_REG_0) {
16352 verbose(env, "BPF_ST uses reserved fields\n");
16355 /* check src operand */
16356 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
16360 dst_reg_type = regs[insn->dst_reg].type;
16362 /* check that memory (dst_reg + off) is writeable */
16363 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
16364 insn->off, BPF_SIZE(insn->code),
16365 BPF_WRITE, -1, false);
16369 err = save_aux_ptr_type(env, dst_reg_type, false);
16372 } else if (class == BPF_JMP || class == BPF_JMP32) {
16373 u8 opcode = BPF_OP(insn->code);
16375 env->jmps_processed++;
16376 if (opcode == BPF_CALL) {
16377 if (BPF_SRC(insn->code) != BPF_K ||
16378 (insn->src_reg != BPF_PSEUDO_KFUNC_CALL
16379 && insn->off != 0) ||
16380 (insn->src_reg != BPF_REG_0 &&
16381 insn->src_reg != BPF_PSEUDO_CALL &&
16382 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
16383 insn->dst_reg != BPF_REG_0 ||
16384 class == BPF_JMP32) {
16385 verbose(env, "BPF_CALL uses reserved fields\n");
16389 if (env->cur_state->active_lock.ptr) {
16390 if ((insn->src_reg == BPF_REG_0 && insn->imm != BPF_FUNC_spin_unlock) ||
16391 (insn->src_reg == BPF_PSEUDO_CALL) ||
16392 (insn->src_reg == BPF_PSEUDO_KFUNC_CALL &&
16393 (insn->off != 0 || !is_bpf_graph_api_kfunc(insn->imm)))) {
16394 verbose(env, "function calls are not allowed while holding a lock\n");
16398 if (insn->src_reg == BPF_PSEUDO_CALL)
16399 err = check_func_call(env, insn, &env->insn_idx);
16400 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
16401 err = check_kfunc_call(env, insn, &env->insn_idx);
16403 err = check_helper_call(env, insn, &env->insn_idx);
16407 mark_reg_scratched(env, BPF_REG_0);
16408 } else if (opcode == BPF_JA) {
16409 if (BPF_SRC(insn->code) != BPF_K ||
16411 insn->src_reg != BPF_REG_0 ||
16412 insn->dst_reg != BPF_REG_0 ||
16413 class == BPF_JMP32) {
16414 verbose(env, "BPF_JA uses reserved fields\n");
16418 env->insn_idx += insn->off + 1;
16421 } else if (opcode == BPF_EXIT) {
16422 if (BPF_SRC(insn->code) != BPF_K ||
16424 insn->src_reg != BPF_REG_0 ||
16425 insn->dst_reg != BPF_REG_0 ||
16426 class == BPF_JMP32) {
16427 verbose(env, "BPF_EXIT uses reserved fields\n");
16431 if (env->cur_state->active_lock.ptr &&
16432 !in_rbtree_lock_required_cb(env)) {
16433 verbose(env, "bpf_spin_unlock is missing\n");
16437 if (env->cur_state->active_rcu_lock) {
16438 verbose(env, "bpf_rcu_read_unlock is missing\n");
16442 /* We must do check_reference_leak here before
16443 * prepare_func_exit to handle the case when
16444 * state->curframe > 0, it may be a callback
16445 * function, for which reference_state must
16446 * match caller reference state when it exits.
16448 err = check_reference_leak(env);
16452 if (state->curframe) {
16453 /* exit from nested function */
16454 err = prepare_func_exit(env, &env->insn_idx);
16457 do_print_state = true;
16461 err = check_return_code(env);
16465 mark_verifier_state_scratched(env);
16466 update_branch_counts(env, env->cur_state);
16467 err = pop_stack(env, &prev_insn_idx,
16468 &env->insn_idx, pop_log);
16470 if (err != -ENOENT)
16474 do_print_state = true;
16478 err = check_cond_jmp_op(env, insn, &env->insn_idx);
16482 } else if (class == BPF_LD) {
16483 u8 mode = BPF_MODE(insn->code);
16485 if (mode == BPF_ABS || mode == BPF_IND) {
16486 err = check_ld_abs(env, insn);
16490 } else if (mode == BPF_IMM) {
16491 err = check_ld_imm(env, insn);
16496 sanitize_mark_insn_seen(env);
16498 verbose(env, "invalid BPF_LD mode\n");
16502 verbose(env, "unknown insn class %d\n", class);
16512 static int find_btf_percpu_datasec(struct btf *btf)
16514 const struct btf_type *t;
16519 * Both vmlinux and module each have their own ".data..percpu"
16520 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
16521 * types to look at only module's own BTF types.
16523 n = btf_nr_types(btf);
16524 if (btf_is_module(btf))
16525 i = btf_nr_types(btf_vmlinux);
16529 for(; i < n; i++) {
16530 t = btf_type_by_id(btf, i);
16531 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
16534 tname = btf_name_by_offset(btf, t->name_off);
16535 if (!strcmp(tname, ".data..percpu"))
16542 /* replace pseudo btf_id with kernel symbol address */
16543 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
16544 struct bpf_insn *insn,
16545 struct bpf_insn_aux_data *aux)
16547 const struct btf_var_secinfo *vsi;
16548 const struct btf_type *datasec;
16549 struct btf_mod_pair *btf_mod;
16550 const struct btf_type *t;
16551 const char *sym_name;
16552 bool percpu = false;
16553 u32 type, id = insn->imm;
16557 int i, btf_fd, err;
16559 btf_fd = insn[1].imm;
16561 btf = btf_get_by_fd(btf_fd);
16563 verbose(env, "invalid module BTF object FD specified.\n");
16567 if (!btf_vmlinux) {
16568 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
16575 t = btf_type_by_id(btf, id);
16577 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
16582 if (!btf_type_is_var(t) && !btf_type_is_func(t)) {
16583 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR or KIND_FUNC\n", id);
16588 sym_name = btf_name_by_offset(btf, t->name_off);
16589 addr = kallsyms_lookup_name(sym_name);
16591 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
16596 insn[0].imm = (u32)addr;
16597 insn[1].imm = addr >> 32;
16599 if (btf_type_is_func(t)) {
16600 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
16601 aux->btf_var.mem_size = 0;
16605 datasec_id = find_btf_percpu_datasec(btf);
16606 if (datasec_id > 0) {
16607 datasec = btf_type_by_id(btf, datasec_id);
16608 for_each_vsi(i, datasec, vsi) {
16609 if (vsi->type == id) {
16617 t = btf_type_skip_modifiers(btf, type, NULL);
16619 aux->btf_var.reg_type = PTR_TO_BTF_ID | MEM_PERCPU;
16620 aux->btf_var.btf = btf;
16621 aux->btf_var.btf_id = type;
16622 } else if (!btf_type_is_struct(t)) {
16623 const struct btf_type *ret;
16627 /* resolve the type size of ksym. */
16628 ret = btf_resolve_size(btf, t, &tsize);
16630 tname = btf_name_by_offset(btf, t->name_off);
16631 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
16632 tname, PTR_ERR(ret));
16636 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
16637 aux->btf_var.mem_size = tsize;
16639 aux->btf_var.reg_type = PTR_TO_BTF_ID;
16640 aux->btf_var.btf = btf;
16641 aux->btf_var.btf_id = type;
16644 /* check whether we recorded this BTF (and maybe module) already */
16645 for (i = 0; i < env->used_btf_cnt; i++) {
16646 if (env->used_btfs[i].btf == btf) {
16652 if (env->used_btf_cnt >= MAX_USED_BTFS) {
16657 btf_mod = &env->used_btfs[env->used_btf_cnt];
16658 btf_mod->btf = btf;
16659 btf_mod->module = NULL;
16661 /* if we reference variables from kernel module, bump its refcount */
16662 if (btf_is_module(btf)) {
16663 btf_mod->module = btf_try_get_module(btf);
16664 if (!btf_mod->module) {
16670 env->used_btf_cnt++;
16678 static bool is_tracing_prog_type(enum bpf_prog_type type)
16681 case BPF_PROG_TYPE_KPROBE:
16682 case BPF_PROG_TYPE_TRACEPOINT:
16683 case BPF_PROG_TYPE_PERF_EVENT:
16684 case BPF_PROG_TYPE_RAW_TRACEPOINT:
16685 case BPF_PROG_TYPE_RAW_TRACEPOINT_WRITABLE:
16692 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
16693 struct bpf_map *map,
16694 struct bpf_prog *prog)
16697 enum bpf_prog_type prog_type = resolve_prog_type(prog);
16699 if (btf_record_has_field(map->record, BPF_LIST_HEAD) ||
16700 btf_record_has_field(map->record, BPF_RB_ROOT)) {
16701 if (is_tracing_prog_type(prog_type)) {
16702 verbose(env, "tracing progs cannot use bpf_{list_head,rb_root} yet\n");
16707 if (btf_record_has_field(map->record, BPF_SPIN_LOCK)) {
16708 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
16709 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
16713 if (is_tracing_prog_type(prog_type)) {
16714 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
16718 if (prog->aux->sleepable) {
16719 verbose(env, "sleepable progs cannot use bpf_spin_lock yet\n");
16724 if (btf_record_has_field(map->record, BPF_TIMER)) {
16725 if (is_tracing_prog_type(prog_type)) {
16726 verbose(env, "tracing progs cannot use bpf_timer yet\n");
16731 if ((bpf_prog_is_offloaded(prog->aux) || bpf_map_is_offloaded(map)) &&
16732 !bpf_offload_prog_map_match(prog, map)) {
16733 verbose(env, "offload device mismatch between prog and map\n");
16737 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
16738 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
16742 if (prog->aux->sleepable)
16743 switch (map->map_type) {
16744 case BPF_MAP_TYPE_HASH:
16745 case BPF_MAP_TYPE_LRU_HASH:
16746 case BPF_MAP_TYPE_ARRAY:
16747 case BPF_MAP_TYPE_PERCPU_HASH:
16748 case BPF_MAP_TYPE_PERCPU_ARRAY:
16749 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
16750 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
16751 case BPF_MAP_TYPE_HASH_OF_MAPS:
16752 case BPF_MAP_TYPE_RINGBUF:
16753 case BPF_MAP_TYPE_USER_RINGBUF:
16754 case BPF_MAP_TYPE_INODE_STORAGE:
16755 case BPF_MAP_TYPE_SK_STORAGE:
16756 case BPF_MAP_TYPE_TASK_STORAGE:
16757 case BPF_MAP_TYPE_CGRP_STORAGE:
16761 "Sleepable programs can only use array, hash, ringbuf and local storage maps\n");
16768 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
16770 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
16771 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
16774 /* find and rewrite pseudo imm in ld_imm64 instructions:
16776 * 1. if it accesses map FD, replace it with actual map pointer.
16777 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
16779 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
16781 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
16783 struct bpf_insn *insn = env->prog->insnsi;
16784 int insn_cnt = env->prog->len;
16787 err = bpf_prog_calc_tag(env->prog);
16791 for (i = 0; i < insn_cnt; i++, insn++) {
16792 if (BPF_CLASS(insn->code) == BPF_LDX &&
16793 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
16794 verbose(env, "BPF_LDX uses reserved fields\n");
16798 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
16799 struct bpf_insn_aux_data *aux;
16800 struct bpf_map *map;
16805 if (i == insn_cnt - 1 || insn[1].code != 0 ||
16806 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
16807 insn[1].off != 0) {
16808 verbose(env, "invalid bpf_ld_imm64 insn\n");
16812 if (insn[0].src_reg == 0)
16813 /* valid generic load 64-bit imm */
16816 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
16817 aux = &env->insn_aux_data[i];
16818 err = check_pseudo_btf_id(env, insn, aux);
16824 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
16825 aux = &env->insn_aux_data[i];
16826 aux->ptr_type = PTR_TO_FUNC;
16830 /* In final convert_pseudo_ld_imm64() step, this is
16831 * converted into regular 64-bit imm load insn.
16833 switch (insn[0].src_reg) {
16834 case BPF_PSEUDO_MAP_VALUE:
16835 case BPF_PSEUDO_MAP_IDX_VALUE:
16837 case BPF_PSEUDO_MAP_FD:
16838 case BPF_PSEUDO_MAP_IDX:
16839 if (insn[1].imm == 0)
16843 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
16847 switch (insn[0].src_reg) {
16848 case BPF_PSEUDO_MAP_IDX_VALUE:
16849 case BPF_PSEUDO_MAP_IDX:
16850 if (bpfptr_is_null(env->fd_array)) {
16851 verbose(env, "fd_idx without fd_array is invalid\n");
16854 if (copy_from_bpfptr_offset(&fd, env->fd_array,
16855 insn[0].imm * sizeof(fd),
16865 map = __bpf_map_get(f);
16867 verbose(env, "fd %d is not pointing to valid bpf_map\n",
16869 return PTR_ERR(map);
16872 err = check_map_prog_compatibility(env, map, env->prog);
16878 aux = &env->insn_aux_data[i];
16879 if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
16880 insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
16881 addr = (unsigned long)map;
16883 u32 off = insn[1].imm;
16885 if (off >= BPF_MAX_VAR_OFF) {
16886 verbose(env, "direct value offset of %u is not allowed\n", off);
16891 if (!map->ops->map_direct_value_addr) {
16892 verbose(env, "no direct value access support for this map type\n");
16897 err = map->ops->map_direct_value_addr(map, &addr, off);
16899 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
16900 map->value_size, off);
16905 aux->map_off = off;
16909 insn[0].imm = (u32)addr;
16910 insn[1].imm = addr >> 32;
16912 /* check whether we recorded this map already */
16913 for (j = 0; j < env->used_map_cnt; j++) {
16914 if (env->used_maps[j] == map) {
16915 aux->map_index = j;
16921 if (env->used_map_cnt >= MAX_USED_MAPS) {
16926 /* hold the map. If the program is rejected by verifier,
16927 * the map will be released by release_maps() or it
16928 * will be used by the valid program until it's unloaded
16929 * and all maps are released in free_used_maps()
16933 aux->map_index = env->used_map_cnt;
16934 env->used_maps[env->used_map_cnt++] = map;
16936 if (bpf_map_is_cgroup_storage(map) &&
16937 bpf_cgroup_storage_assign(env->prog->aux, map)) {
16938 verbose(env, "only one cgroup storage of each type is allowed\n");
16950 /* Basic sanity check before we invest more work here. */
16951 if (!bpf_opcode_in_insntable(insn->code)) {
16952 verbose(env, "unknown opcode %02x\n", insn->code);
16957 /* now all pseudo BPF_LD_IMM64 instructions load valid
16958 * 'struct bpf_map *' into a register instead of user map_fd.
16959 * These pointers will be used later by verifier to validate map access.
16964 /* drop refcnt of maps used by the rejected program */
16965 static void release_maps(struct bpf_verifier_env *env)
16967 __bpf_free_used_maps(env->prog->aux, env->used_maps,
16968 env->used_map_cnt);
16971 /* drop refcnt of maps used by the rejected program */
16972 static void release_btfs(struct bpf_verifier_env *env)
16974 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
16975 env->used_btf_cnt);
16978 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
16979 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
16981 struct bpf_insn *insn = env->prog->insnsi;
16982 int insn_cnt = env->prog->len;
16985 for (i = 0; i < insn_cnt; i++, insn++) {
16986 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
16988 if (insn->src_reg == BPF_PSEUDO_FUNC)
16994 /* single env->prog->insni[off] instruction was replaced with the range
16995 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
16996 * [0, off) and [off, end) to new locations, so the patched range stays zero
16998 static void adjust_insn_aux_data(struct bpf_verifier_env *env,
16999 struct bpf_insn_aux_data *new_data,
17000 struct bpf_prog *new_prog, u32 off, u32 cnt)
17002 struct bpf_insn_aux_data *old_data = env->insn_aux_data;
17003 struct bpf_insn *insn = new_prog->insnsi;
17004 u32 old_seen = old_data[off].seen;
17008 /* aux info at OFF always needs adjustment, no matter fast path
17009 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
17010 * original insn at old prog.
17012 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
17016 prog_len = new_prog->len;
17018 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
17019 memcpy(new_data + off + cnt - 1, old_data + off,
17020 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
17021 for (i = off; i < off + cnt - 1; i++) {
17022 /* Expand insni[off]'s seen count to the patched range. */
17023 new_data[i].seen = old_seen;
17024 new_data[i].zext_dst = insn_has_def32(env, insn + i);
17026 env->insn_aux_data = new_data;
17030 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
17036 /* NOTE: fake 'exit' subprog should be updated as well. */
17037 for (i = 0; i <= env->subprog_cnt; i++) {
17038 if (env->subprog_info[i].start <= off)
17040 env->subprog_info[i].start += len - 1;
17044 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
17046 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
17047 int i, sz = prog->aux->size_poke_tab;
17048 struct bpf_jit_poke_descriptor *desc;
17050 for (i = 0; i < sz; i++) {
17052 if (desc->insn_idx <= off)
17054 desc->insn_idx += len - 1;
17058 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
17059 const struct bpf_insn *patch, u32 len)
17061 struct bpf_prog *new_prog;
17062 struct bpf_insn_aux_data *new_data = NULL;
17065 new_data = vzalloc(array_size(env->prog->len + len - 1,
17066 sizeof(struct bpf_insn_aux_data)));
17071 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
17072 if (IS_ERR(new_prog)) {
17073 if (PTR_ERR(new_prog) == -ERANGE)
17075 "insn %d cannot be patched due to 16-bit range\n",
17076 env->insn_aux_data[off].orig_idx);
17080 adjust_insn_aux_data(env, new_data, new_prog, off, len);
17081 adjust_subprog_starts(env, off, len);
17082 adjust_poke_descs(new_prog, off, len);
17086 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
17091 /* find first prog starting at or after off (first to remove) */
17092 for (i = 0; i < env->subprog_cnt; i++)
17093 if (env->subprog_info[i].start >= off)
17095 /* find first prog starting at or after off + cnt (first to stay) */
17096 for (j = i; j < env->subprog_cnt; j++)
17097 if (env->subprog_info[j].start >= off + cnt)
17099 /* if j doesn't start exactly at off + cnt, we are just removing
17100 * the front of previous prog
17102 if (env->subprog_info[j].start != off + cnt)
17106 struct bpf_prog_aux *aux = env->prog->aux;
17109 /* move fake 'exit' subprog as well */
17110 move = env->subprog_cnt + 1 - j;
17112 memmove(env->subprog_info + i,
17113 env->subprog_info + j,
17114 sizeof(*env->subprog_info) * move);
17115 env->subprog_cnt -= j - i;
17117 /* remove func_info */
17118 if (aux->func_info) {
17119 move = aux->func_info_cnt - j;
17121 memmove(aux->func_info + i,
17122 aux->func_info + j,
17123 sizeof(*aux->func_info) * move);
17124 aux->func_info_cnt -= j - i;
17125 /* func_info->insn_off is set after all code rewrites,
17126 * in adjust_btf_func() - no need to adjust
17130 /* convert i from "first prog to remove" to "first to adjust" */
17131 if (env->subprog_info[i].start == off)
17135 /* update fake 'exit' subprog as well */
17136 for (; i <= env->subprog_cnt; i++)
17137 env->subprog_info[i].start -= cnt;
17142 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
17145 struct bpf_prog *prog = env->prog;
17146 u32 i, l_off, l_cnt, nr_linfo;
17147 struct bpf_line_info *linfo;
17149 nr_linfo = prog->aux->nr_linfo;
17153 linfo = prog->aux->linfo;
17155 /* find first line info to remove, count lines to be removed */
17156 for (i = 0; i < nr_linfo; i++)
17157 if (linfo[i].insn_off >= off)
17162 for (; i < nr_linfo; i++)
17163 if (linfo[i].insn_off < off + cnt)
17168 /* First live insn doesn't match first live linfo, it needs to "inherit"
17169 * last removed linfo. prog is already modified, so prog->len == off
17170 * means no live instructions after (tail of the program was removed).
17172 if (prog->len != off && l_cnt &&
17173 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
17175 linfo[--i].insn_off = off + cnt;
17178 /* remove the line info which refer to the removed instructions */
17180 memmove(linfo + l_off, linfo + i,
17181 sizeof(*linfo) * (nr_linfo - i));
17183 prog->aux->nr_linfo -= l_cnt;
17184 nr_linfo = prog->aux->nr_linfo;
17187 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
17188 for (i = l_off; i < nr_linfo; i++)
17189 linfo[i].insn_off -= cnt;
17191 /* fix up all subprogs (incl. 'exit') which start >= off */
17192 for (i = 0; i <= env->subprog_cnt; i++)
17193 if (env->subprog_info[i].linfo_idx > l_off) {
17194 /* program may have started in the removed region but
17195 * may not be fully removed
17197 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
17198 env->subprog_info[i].linfo_idx -= l_cnt;
17200 env->subprog_info[i].linfo_idx = l_off;
17206 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
17208 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17209 unsigned int orig_prog_len = env->prog->len;
17212 if (bpf_prog_is_offloaded(env->prog->aux))
17213 bpf_prog_offload_remove_insns(env, off, cnt);
17215 err = bpf_remove_insns(env->prog, off, cnt);
17219 err = adjust_subprog_starts_after_remove(env, off, cnt);
17223 err = bpf_adj_linfo_after_remove(env, off, cnt);
17227 memmove(aux_data + off, aux_data + off + cnt,
17228 sizeof(*aux_data) * (orig_prog_len - off - cnt));
17233 /* The verifier does more data flow analysis than llvm and will not
17234 * explore branches that are dead at run time. Malicious programs can
17235 * have dead code too. Therefore replace all dead at-run-time code
17238 * Just nops are not optimal, e.g. if they would sit at the end of the
17239 * program and through another bug we would manage to jump there, then
17240 * we'd execute beyond program memory otherwise. Returning exception
17241 * code also wouldn't work since we can have subprogs where the dead
17242 * code could be located.
17244 static void sanitize_dead_code(struct bpf_verifier_env *env)
17246 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17247 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
17248 struct bpf_insn *insn = env->prog->insnsi;
17249 const int insn_cnt = env->prog->len;
17252 for (i = 0; i < insn_cnt; i++) {
17253 if (aux_data[i].seen)
17255 memcpy(insn + i, &trap, sizeof(trap));
17256 aux_data[i].zext_dst = false;
17260 static bool insn_is_cond_jump(u8 code)
17264 if (BPF_CLASS(code) == BPF_JMP32)
17267 if (BPF_CLASS(code) != BPF_JMP)
17271 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
17274 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
17276 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17277 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
17278 struct bpf_insn *insn = env->prog->insnsi;
17279 const int insn_cnt = env->prog->len;
17282 for (i = 0; i < insn_cnt; i++, insn++) {
17283 if (!insn_is_cond_jump(insn->code))
17286 if (!aux_data[i + 1].seen)
17287 ja.off = insn->off;
17288 else if (!aux_data[i + 1 + insn->off].seen)
17293 if (bpf_prog_is_offloaded(env->prog->aux))
17294 bpf_prog_offload_replace_insn(env, i, &ja);
17296 memcpy(insn, &ja, sizeof(ja));
17300 static int opt_remove_dead_code(struct bpf_verifier_env *env)
17302 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17303 int insn_cnt = env->prog->len;
17306 for (i = 0; i < insn_cnt; i++) {
17310 while (i + j < insn_cnt && !aux_data[i + j].seen)
17315 err = verifier_remove_insns(env, i, j);
17318 insn_cnt = env->prog->len;
17324 static int opt_remove_nops(struct bpf_verifier_env *env)
17326 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
17327 struct bpf_insn *insn = env->prog->insnsi;
17328 int insn_cnt = env->prog->len;
17331 for (i = 0; i < insn_cnt; i++) {
17332 if (memcmp(&insn[i], &ja, sizeof(ja)))
17335 err = verifier_remove_insns(env, i, 1);
17345 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
17346 const union bpf_attr *attr)
17348 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
17349 struct bpf_insn_aux_data *aux = env->insn_aux_data;
17350 int i, patch_len, delta = 0, len = env->prog->len;
17351 struct bpf_insn *insns = env->prog->insnsi;
17352 struct bpf_prog *new_prog;
17355 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
17356 zext_patch[1] = BPF_ZEXT_REG(0);
17357 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
17358 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
17359 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
17360 for (i = 0; i < len; i++) {
17361 int adj_idx = i + delta;
17362 struct bpf_insn insn;
17365 insn = insns[adj_idx];
17366 load_reg = insn_def_regno(&insn);
17367 if (!aux[adj_idx].zext_dst) {
17375 class = BPF_CLASS(code);
17376 if (load_reg == -1)
17379 /* NOTE: arg "reg" (the fourth one) is only used for
17380 * BPF_STX + SRC_OP, so it is safe to pass NULL
17383 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
17384 if (class == BPF_LD &&
17385 BPF_MODE(code) == BPF_IMM)
17390 /* ctx load could be transformed into wider load. */
17391 if (class == BPF_LDX &&
17392 aux[adj_idx].ptr_type == PTR_TO_CTX)
17395 imm_rnd = get_random_u32();
17396 rnd_hi32_patch[0] = insn;
17397 rnd_hi32_patch[1].imm = imm_rnd;
17398 rnd_hi32_patch[3].dst_reg = load_reg;
17399 patch = rnd_hi32_patch;
17401 goto apply_patch_buffer;
17404 /* Add in an zero-extend instruction if a) the JIT has requested
17405 * it or b) it's a CMPXCHG.
17407 * The latter is because: BPF_CMPXCHG always loads a value into
17408 * R0, therefore always zero-extends. However some archs'
17409 * equivalent instruction only does this load when the
17410 * comparison is successful. This detail of CMPXCHG is
17411 * orthogonal to the general zero-extension behaviour of the
17412 * CPU, so it's treated independently of bpf_jit_needs_zext.
17414 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
17417 /* Zero-extension is done by the caller. */
17418 if (bpf_pseudo_kfunc_call(&insn))
17421 if (WARN_ON(load_reg == -1)) {
17422 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
17426 zext_patch[0] = insn;
17427 zext_patch[1].dst_reg = load_reg;
17428 zext_patch[1].src_reg = load_reg;
17429 patch = zext_patch;
17431 apply_patch_buffer:
17432 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
17435 env->prog = new_prog;
17436 insns = new_prog->insnsi;
17437 aux = env->insn_aux_data;
17438 delta += patch_len - 1;
17444 /* convert load instructions that access fields of a context type into a
17445 * sequence of instructions that access fields of the underlying structure:
17446 * struct __sk_buff -> struct sk_buff
17447 * struct bpf_sock_ops -> struct sock
17449 static int convert_ctx_accesses(struct bpf_verifier_env *env)
17451 const struct bpf_verifier_ops *ops = env->ops;
17452 int i, cnt, size, ctx_field_size, delta = 0;
17453 const int insn_cnt = env->prog->len;
17454 struct bpf_insn insn_buf[16], *insn;
17455 u32 target_size, size_default, off;
17456 struct bpf_prog *new_prog;
17457 enum bpf_access_type type;
17458 bool is_narrower_load;
17460 if (ops->gen_prologue || env->seen_direct_write) {
17461 if (!ops->gen_prologue) {
17462 verbose(env, "bpf verifier is misconfigured\n");
17465 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
17467 if (cnt >= ARRAY_SIZE(insn_buf)) {
17468 verbose(env, "bpf verifier is misconfigured\n");
17471 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
17475 env->prog = new_prog;
17480 if (bpf_prog_is_offloaded(env->prog->aux))
17483 insn = env->prog->insnsi + delta;
17485 for (i = 0; i < insn_cnt; i++, insn++) {
17486 bpf_convert_ctx_access_t convert_ctx_access;
17488 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
17489 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
17490 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
17491 insn->code == (BPF_LDX | BPF_MEM | BPF_DW)) {
17493 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
17494 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
17495 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
17496 insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
17497 insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
17498 insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
17499 insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
17500 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
17506 if (type == BPF_WRITE &&
17507 env->insn_aux_data[i + delta].sanitize_stack_spill) {
17508 struct bpf_insn patch[] = {
17513 cnt = ARRAY_SIZE(patch);
17514 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
17519 env->prog = new_prog;
17520 insn = new_prog->insnsi + i + delta;
17524 switch ((int)env->insn_aux_data[i + delta].ptr_type) {
17526 if (!ops->convert_ctx_access)
17528 convert_ctx_access = ops->convert_ctx_access;
17530 case PTR_TO_SOCKET:
17531 case PTR_TO_SOCK_COMMON:
17532 convert_ctx_access = bpf_sock_convert_ctx_access;
17534 case PTR_TO_TCP_SOCK:
17535 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
17537 case PTR_TO_XDP_SOCK:
17538 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
17540 case PTR_TO_BTF_ID:
17541 case PTR_TO_BTF_ID | PTR_UNTRUSTED:
17542 /* PTR_TO_BTF_ID | MEM_ALLOC always has a valid lifetime, unlike
17543 * PTR_TO_BTF_ID, and an active ref_obj_id, but the same cannot
17544 * be said once it is marked PTR_UNTRUSTED, hence we must handle
17545 * any faults for loads into such types. BPF_WRITE is disallowed
17548 case PTR_TO_BTF_ID | MEM_ALLOC | PTR_UNTRUSTED:
17549 if (type == BPF_READ) {
17550 insn->code = BPF_LDX | BPF_PROBE_MEM |
17551 BPF_SIZE((insn)->code);
17552 env->prog->aux->num_exentries++;
17559 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
17560 size = BPF_LDST_BYTES(insn);
17562 /* If the read access is a narrower load of the field,
17563 * convert to a 4/8-byte load, to minimum program type specific
17564 * convert_ctx_access changes. If conversion is successful,
17565 * we will apply proper mask to the result.
17567 is_narrower_load = size < ctx_field_size;
17568 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
17570 if (is_narrower_load) {
17573 if (type == BPF_WRITE) {
17574 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
17579 if (ctx_field_size == 4)
17581 else if (ctx_field_size == 8)
17582 size_code = BPF_DW;
17584 insn->off = off & ~(size_default - 1);
17585 insn->code = BPF_LDX | BPF_MEM | size_code;
17589 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
17591 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
17592 (ctx_field_size && !target_size)) {
17593 verbose(env, "bpf verifier is misconfigured\n");
17597 if (is_narrower_load && size < target_size) {
17598 u8 shift = bpf_ctx_narrow_access_offset(
17599 off, size, size_default) * 8;
17600 if (shift && cnt + 1 >= ARRAY_SIZE(insn_buf)) {
17601 verbose(env, "bpf verifier narrow ctx load misconfigured\n");
17604 if (ctx_field_size <= 4) {
17606 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
17609 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
17610 (1 << size * 8) - 1);
17613 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
17616 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
17617 (1ULL << size * 8) - 1);
17621 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
17627 /* keep walking new program and skip insns we just inserted */
17628 env->prog = new_prog;
17629 insn = new_prog->insnsi + i + delta;
17635 static int jit_subprogs(struct bpf_verifier_env *env)
17637 struct bpf_prog *prog = env->prog, **func, *tmp;
17638 int i, j, subprog_start, subprog_end = 0, len, subprog;
17639 struct bpf_map *map_ptr;
17640 struct bpf_insn *insn;
17641 void *old_bpf_func;
17642 int err, num_exentries;
17644 if (env->subprog_cnt <= 1)
17647 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
17648 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn))
17651 /* Upon error here we cannot fall back to interpreter but
17652 * need a hard reject of the program. Thus -EFAULT is
17653 * propagated in any case.
17655 subprog = find_subprog(env, i + insn->imm + 1);
17657 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
17658 i + insn->imm + 1);
17661 /* temporarily remember subprog id inside insn instead of
17662 * aux_data, since next loop will split up all insns into funcs
17664 insn->off = subprog;
17665 /* remember original imm in case JIT fails and fallback
17666 * to interpreter will be needed
17668 env->insn_aux_data[i].call_imm = insn->imm;
17669 /* point imm to __bpf_call_base+1 from JITs point of view */
17671 if (bpf_pseudo_func(insn))
17672 /* jit (e.g. x86_64) may emit fewer instructions
17673 * if it learns a u32 imm is the same as a u64 imm.
17674 * Force a non zero here.
17679 err = bpf_prog_alloc_jited_linfo(prog);
17681 goto out_undo_insn;
17684 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
17686 goto out_undo_insn;
17688 for (i = 0; i < env->subprog_cnt; i++) {
17689 subprog_start = subprog_end;
17690 subprog_end = env->subprog_info[i + 1].start;
17692 len = subprog_end - subprog_start;
17693 /* bpf_prog_run() doesn't call subprogs directly,
17694 * hence main prog stats include the runtime of subprogs.
17695 * subprogs don't have IDs and not reachable via prog_get_next_id
17696 * func[i]->stats will never be accessed and stays NULL
17698 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
17701 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
17702 len * sizeof(struct bpf_insn));
17703 func[i]->type = prog->type;
17704 func[i]->len = len;
17705 if (bpf_prog_calc_tag(func[i]))
17707 func[i]->is_func = 1;
17708 func[i]->aux->func_idx = i;
17709 /* Below members will be freed only at prog->aux */
17710 func[i]->aux->btf = prog->aux->btf;
17711 func[i]->aux->func_info = prog->aux->func_info;
17712 func[i]->aux->func_info_cnt = prog->aux->func_info_cnt;
17713 func[i]->aux->poke_tab = prog->aux->poke_tab;
17714 func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
17716 for (j = 0; j < prog->aux->size_poke_tab; j++) {
17717 struct bpf_jit_poke_descriptor *poke;
17719 poke = &prog->aux->poke_tab[j];
17720 if (poke->insn_idx < subprog_end &&
17721 poke->insn_idx >= subprog_start)
17722 poke->aux = func[i]->aux;
17725 func[i]->aux->name[0] = 'F';
17726 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
17727 func[i]->jit_requested = 1;
17728 func[i]->blinding_requested = prog->blinding_requested;
17729 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
17730 func[i]->aux->kfunc_btf_tab = prog->aux->kfunc_btf_tab;
17731 func[i]->aux->linfo = prog->aux->linfo;
17732 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
17733 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
17734 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
17736 insn = func[i]->insnsi;
17737 for (j = 0; j < func[i]->len; j++, insn++) {
17738 if (BPF_CLASS(insn->code) == BPF_LDX &&
17739 BPF_MODE(insn->code) == BPF_PROBE_MEM)
17742 func[i]->aux->num_exentries = num_exentries;
17743 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
17744 func[i] = bpf_int_jit_compile(func[i]);
17745 if (!func[i]->jited) {
17752 /* at this point all bpf functions were successfully JITed
17753 * now populate all bpf_calls with correct addresses and
17754 * run last pass of JIT
17756 for (i = 0; i < env->subprog_cnt; i++) {
17757 insn = func[i]->insnsi;
17758 for (j = 0; j < func[i]->len; j++, insn++) {
17759 if (bpf_pseudo_func(insn)) {
17760 subprog = insn->off;
17761 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
17762 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
17765 if (!bpf_pseudo_call(insn))
17767 subprog = insn->off;
17768 insn->imm = BPF_CALL_IMM(func[subprog]->bpf_func);
17771 /* we use the aux data to keep a list of the start addresses
17772 * of the JITed images for each function in the program
17774 * for some architectures, such as powerpc64, the imm field
17775 * might not be large enough to hold the offset of the start
17776 * address of the callee's JITed image from __bpf_call_base
17778 * in such cases, we can lookup the start address of a callee
17779 * by using its subprog id, available from the off field of
17780 * the call instruction, as an index for this list
17782 func[i]->aux->func = func;
17783 func[i]->aux->func_cnt = env->subprog_cnt;
17785 for (i = 0; i < env->subprog_cnt; i++) {
17786 old_bpf_func = func[i]->bpf_func;
17787 tmp = bpf_int_jit_compile(func[i]);
17788 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
17789 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
17796 /* finally lock prog and jit images for all functions and
17797 * populate kallsysm. Begin at the first subprogram, since
17798 * bpf_prog_load will add the kallsyms for the main program.
17800 for (i = 1; i < env->subprog_cnt; i++) {
17801 bpf_prog_lock_ro(func[i]);
17802 bpf_prog_kallsyms_add(func[i]);
17805 /* Last step: make now unused interpreter insns from main
17806 * prog consistent for later dump requests, so they can
17807 * later look the same as if they were interpreted only.
17809 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
17810 if (bpf_pseudo_func(insn)) {
17811 insn[0].imm = env->insn_aux_data[i].call_imm;
17812 insn[1].imm = insn->off;
17816 if (!bpf_pseudo_call(insn))
17818 insn->off = env->insn_aux_data[i].call_imm;
17819 subprog = find_subprog(env, i + insn->off + 1);
17820 insn->imm = subprog;
17824 prog->bpf_func = func[0]->bpf_func;
17825 prog->jited_len = func[0]->jited_len;
17826 prog->aux->extable = func[0]->aux->extable;
17827 prog->aux->num_exentries = func[0]->aux->num_exentries;
17828 prog->aux->func = func;
17829 prog->aux->func_cnt = env->subprog_cnt;
17830 bpf_prog_jit_attempt_done(prog);
17833 /* We failed JIT'ing, so at this point we need to unregister poke
17834 * descriptors from subprogs, so that kernel is not attempting to
17835 * patch it anymore as we're freeing the subprog JIT memory.
17837 for (i = 0; i < prog->aux->size_poke_tab; i++) {
17838 map_ptr = prog->aux->poke_tab[i].tail_call.map;
17839 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
17841 /* At this point we're guaranteed that poke descriptors are not
17842 * live anymore. We can just unlink its descriptor table as it's
17843 * released with the main prog.
17845 for (i = 0; i < env->subprog_cnt; i++) {
17848 func[i]->aux->poke_tab = NULL;
17849 bpf_jit_free(func[i]);
17853 /* cleanup main prog to be interpreted */
17854 prog->jit_requested = 0;
17855 prog->blinding_requested = 0;
17856 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
17857 if (!bpf_pseudo_call(insn))
17860 insn->imm = env->insn_aux_data[i].call_imm;
17862 bpf_prog_jit_attempt_done(prog);
17866 static int fixup_call_args(struct bpf_verifier_env *env)
17868 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
17869 struct bpf_prog *prog = env->prog;
17870 struct bpf_insn *insn = prog->insnsi;
17871 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
17876 if (env->prog->jit_requested &&
17877 !bpf_prog_is_offloaded(env->prog->aux)) {
17878 err = jit_subprogs(env);
17881 if (err == -EFAULT)
17884 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
17885 if (has_kfunc_call) {
17886 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
17889 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
17890 /* When JIT fails the progs with bpf2bpf calls and tail_calls
17891 * have to be rejected, since interpreter doesn't support them yet.
17893 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
17896 for (i = 0; i < prog->len; i++, insn++) {
17897 if (bpf_pseudo_func(insn)) {
17898 /* When JIT fails the progs with callback calls
17899 * have to be rejected, since interpreter doesn't support them yet.
17901 verbose(env, "callbacks are not allowed in non-JITed programs\n");
17905 if (!bpf_pseudo_call(insn))
17907 depth = get_callee_stack_depth(env, insn, i);
17910 bpf_patch_call_args(insn, depth);
17917 /* replace a generic kfunc with a specialized version if necessary */
17918 static void specialize_kfunc(struct bpf_verifier_env *env,
17919 u32 func_id, u16 offset, unsigned long *addr)
17921 struct bpf_prog *prog = env->prog;
17922 bool seen_direct_write;
17926 if (bpf_dev_bound_kfunc_id(func_id)) {
17927 xdp_kfunc = bpf_dev_bound_resolve_kfunc(prog, func_id);
17929 *addr = (unsigned long)xdp_kfunc;
17932 /* fallback to default kfunc when not supported by netdev */
17938 if (func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) {
17939 seen_direct_write = env->seen_direct_write;
17940 is_rdonly = !may_access_direct_pkt_data(env, NULL, BPF_WRITE);
17943 *addr = (unsigned long)bpf_dynptr_from_skb_rdonly;
17945 /* restore env->seen_direct_write to its original value, since
17946 * may_access_direct_pkt_data mutates it
17948 env->seen_direct_write = seen_direct_write;
17952 static void __fixup_collection_insert_kfunc(struct bpf_insn_aux_data *insn_aux,
17953 u16 struct_meta_reg,
17954 u16 node_offset_reg,
17955 struct bpf_insn *insn,
17956 struct bpf_insn *insn_buf,
17959 struct btf_struct_meta *kptr_struct_meta = insn_aux->kptr_struct_meta;
17960 struct bpf_insn addr[2] = { BPF_LD_IMM64(struct_meta_reg, (long)kptr_struct_meta) };
17962 insn_buf[0] = addr[0];
17963 insn_buf[1] = addr[1];
17964 insn_buf[2] = BPF_MOV64_IMM(node_offset_reg, insn_aux->insert_off);
17965 insn_buf[3] = *insn;
17969 static int fixup_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
17970 struct bpf_insn *insn_buf, int insn_idx, int *cnt)
17972 const struct bpf_kfunc_desc *desc;
17975 verbose(env, "invalid kernel function call not eliminated in verifier pass\n");
17981 /* insn->imm has the btf func_id. Replace it with an offset relative to
17982 * __bpf_call_base, unless the JIT needs to call functions that are
17983 * further than 32 bits away (bpf_jit_supports_far_kfunc_call()).
17985 desc = find_kfunc_desc(env->prog, insn->imm, insn->off);
17987 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
17992 if (!bpf_jit_supports_far_kfunc_call())
17993 insn->imm = BPF_CALL_IMM(desc->addr);
17996 if (desc->func_id == special_kfunc_list[KF_bpf_obj_new_impl]) {
17997 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
17998 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
17999 u64 obj_new_size = env->insn_aux_data[insn_idx].obj_new_size;
18001 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_1, obj_new_size);
18002 insn_buf[1] = addr[0];
18003 insn_buf[2] = addr[1];
18004 insn_buf[3] = *insn;
18006 } else if (desc->func_id == special_kfunc_list[KF_bpf_obj_drop_impl] ||
18007 desc->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) {
18008 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18009 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
18011 insn_buf[0] = addr[0];
18012 insn_buf[1] = addr[1];
18013 insn_buf[2] = *insn;
18015 } else if (desc->func_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
18016 desc->func_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
18017 desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
18018 int struct_meta_reg = BPF_REG_3;
18019 int node_offset_reg = BPF_REG_4;
18021 /* rbtree_add has extra 'less' arg, so args-to-fixup are in diff regs */
18022 if (desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
18023 struct_meta_reg = BPF_REG_4;
18024 node_offset_reg = BPF_REG_5;
18027 __fixup_collection_insert_kfunc(&env->insn_aux_data[insn_idx], struct_meta_reg,
18028 node_offset_reg, insn, insn_buf, cnt);
18029 } else if (desc->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx] ||
18030 desc->func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
18031 insn_buf[0] = BPF_MOV64_REG(BPF_REG_0, BPF_REG_1);
18037 /* Do various post-verification rewrites in a single program pass.
18038 * These rewrites simplify JIT and interpreter implementations.
18040 static int do_misc_fixups(struct bpf_verifier_env *env)
18042 struct bpf_prog *prog = env->prog;
18043 enum bpf_attach_type eatype = prog->expected_attach_type;
18044 enum bpf_prog_type prog_type = resolve_prog_type(prog);
18045 struct bpf_insn *insn = prog->insnsi;
18046 const struct bpf_func_proto *fn;
18047 const int insn_cnt = prog->len;
18048 const struct bpf_map_ops *ops;
18049 struct bpf_insn_aux_data *aux;
18050 struct bpf_insn insn_buf[16];
18051 struct bpf_prog *new_prog;
18052 struct bpf_map *map_ptr;
18053 int i, ret, cnt, delta = 0;
18055 for (i = 0; i < insn_cnt; i++, insn++) {
18056 /* Make divide-by-zero exceptions impossible. */
18057 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
18058 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
18059 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
18060 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
18061 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
18062 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
18063 struct bpf_insn *patchlet;
18064 struct bpf_insn chk_and_div[] = {
18065 /* [R,W]x div 0 -> 0 */
18066 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
18067 BPF_JNE | BPF_K, insn->src_reg,
18069 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
18070 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
18073 struct bpf_insn chk_and_mod[] = {
18074 /* [R,W]x mod 0 -> [R,W]x */
18075 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
18076 BPF_JEQ | BPF_K, insn->src_reg,
18077 0, 1 + (is64 ? 0 : 1), 0),
18079 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
18080 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
18083 patchlet = isdiv ? chk_and_div : chk_and_mod;
18084 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
18085 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
18087 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
18092 env->prog = prog = new_prog;
18093 insn = new_prog->insnsi + i + delta;
18097 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
18098 if (BPF_CLASS(insn->code) == BPF_LD &&
18099 (BPF_MODE(insn->code) == BPF_ABS ||
18100 BPF_MODE(insn->code) == BPF_IND)) {
18101 cnt = env->ops->gen_ld_abs(insn, insn_buf);
18102 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
18103 verbose(env, "bpf verifier is misconfigured\n");
18107 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18112 env->prog = prog = new_prog;
18113 insn = new_prog->insnsi + i + delta;
18117 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
18118 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
18119 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
18120 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
18121 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
18122 struct bpf_insn *patch = &insn_buf[0];
18123 bool issrc, isneg, isimm;
18126 aux = &env->insn_aux_data[i + delta];
18127 if (!aux->alu_state ||
18128 aux->alu_state == BPF_ALU_NON_POINTER)
18131 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
18132 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
18133 BPF_ALU_SANITIZE_SRC;
18134 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
18136 off_reg = issrc ? insn->src_reg : insn->dst_reg;
18138 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
18141 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
18142 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
18143 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
18144 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
18145 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
18146 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
18147 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
18150 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
18151 insn->src_reg = BPF_REG_AX;
18153 insn->code = insn->code == code_add ?
18154 code_sub : code_add;
18156 if (issrc && isneg && !isimm)
18157 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
18158 cnt = patch - insn_buf;
18160 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18165 env->prog = prog = new_prog;
18166 insn = new_prog->insnsi + i + delta;
18170 if (insn->code != (BPF_JMP | BPF_CALL))
18172 if (insn->src_reg == BPF_PSEUDO_CALL)
18174 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
18175 ret = fixup_kfunc_call(env, insn, insn_buf, i + delta, &cnt);
18181 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18186 env->prog = prog = new_prog;
18187 insn = new_prog->insnsi + i + delta;
18191 if (insn->imm == BPF_FUNC_get_route_realm)
18192 prog->dst_needed = 1;
18193 if (insn->imm == BPF_FUNC_get_prandom_u32)
18194 bpf_user_rnd_init_once();
18195 if (insn->imm == BPF_FUNC_override_return)
18196 prog->kprobe_override = 1;
18197 if (insn->imm == BPF_FUNC_tail_call) {
18198 /* If we tail call into other programs, we
18199 * cannot make any assumptions since they can
18200 * be replaced dynamically during runtime in
18201 * the program array.
18203 prog->cb_access = 1;
18204 if (!allow_tail_call_in_subprogs(env))
18205 prog->aux->stack_depth = MAX_BPF_STACK;
18206 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
18208 /* mark bpf_tail_call as different opcode to avoid
18209 * conditional branch in the interpreter for every normal
18210 * call and to prevent accidental JITing by JIT compiler
18211 * that doesn't support bpf_tail_call yet
18214 insn->code = BPF_JMP | BPF_TAIL_CALL;
18216 aux = &env->insn_aux_data[i + delta];
18217 if (env->bpf_capable && !prog->blinding_requested &&
18218 prog->jit_requested &&
18219 !bpf_map_key_poisoned(aux) &&
18220 !bpf_map_ptr_poisoned(aux) &&
18221 !bpf_map_ptr_unpriv(aux)) {
18222 struct bpf_jit_poke_descriptor desc = {
18223 .reason = BPF_POKE_REASON_TAIL_CALL,
18224 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
18225 .tail_call.key = bpf_map_key_immediate(aux),
18226 .insn_idx = i + delta,
18229 ret = bpf_jit_add_poke_descriptor(prog, &desc);
18231 verbose(env, "adding tail call poke descriptor failed\n");
18235 insn->imm = ret + 1;
18239 if (!bpf_map_ptr_unpriv(aux))
18242 /* instead of changing every JIT dealing with tail_call
18243 * emit two extra insns:
18244 * if (index >= max_entries) goto out;
18245 * index &= array->index_mask;
18246 * to avoid out-of-bounds cpu speculation
18248 if (bpf_map_ptr_poisoned(aux)) {
18249 verbose(env, "tail_call abusing map_ptr\n");
18253 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
18254 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
18255 map_ptr->max_entries, 2);
18256 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
18257 container_of(map_ptr,
18260 insn_buf[2] = *insn;
18262 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18267 env->prog = prog = new_prog;
18268 insn = new_prog->insnsi + i + delta;
18272 if (insn->imm == BPF_FUNC_timer_set_callback) {
18273 /* The verifier will process callback_fn as many times as necessary
18274 * with different maps and the register states prepared by
18275 * set_timer_callback_state will be accurate.
18277 * The following use case is valid:
18278 * map1 is shared by prog1, prog2, prog3.
18279 * prog1 calls bpf_timer_init for some map1 elements
18280 * prog2 calls bpf_timer_set_callback for some map1 elements.
18281 * Those that were not bpf_timer_init-ed will return -EINVAL.
18282 * prog3 calls bpf_timer_start for some map1 elements.
18283 * Those that were not both bpf_timer_init-ed and
18284 * bpf_timer_set_callback-ed will return -EINVAL.
18286 struct bpf_insn ld_addrs[2] = {
18287 BPF_LD_IMM64(BPF_REG_3, (long)prog->aux),
18290 insn_buf[0] = ld_addrs[0];
18291 insn_buf[1] = ld_addrs[1];
18292 insn_buf[2] = *insn;
18295 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18300 env->prog = prog = new_prog;
18301 insn = new_prog->insnsi + i + delta;
18302 goto patch_call_imm;
18305 if (is_storage_get_function(insn->imm)) {
18306 if (!env->prog->aux->sleepable ||
18307 env->insn_aux_data[i + delta].storage_get_func_atomic)
18308 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_ATOMIC);
18310 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_KERNEL);
18311 insn_buf[1] = *insn;
18314 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18319 env->prog = prog = new_prog;
18320 insn = new_prog->insnsi + i + delta;
18321 goto patch_call_imm;
18324 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
18325 * and other inlining handlers are currently limited to 64 bit
18328 if (prog->jit_requested && BITS_PER_LONG == 64 &&
18329 (insn->imm == BPF_FUNC_map_lookup_elem ||
18330 insn->imm == BPF_FUNC_map_update_elem ||
18331 insn->imm == BPF_FUNC_map_delete_elem ||
18332 insn->imm == BPF_FUNC_map_push_elem ||
18333 insn->imm == BPF_FUNC_map_pop_elem ||
18334 insn->imm == BPF_FUNC_map_peek_elem ||
18335 insn->imm == BPF_FUNC_redirect_map ||
18336 insn->imm == BPF_FUNC_for_each_map_elem ||
18337 insn->imm == BPF_FUNC_map_lookup_percpu_elem)) {
18338 aux = &env->insn_aux_data[i + delta];
18339 if (bpf_map_ptr_poisoned(aux))
18340 goto patch_call_imm;
18342 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
18343 ops = map_ptr->ops;
18344 if (insn->imm == BPF_FUNC_map_lookup_elem &&
18345 ops->map_gen_lookup) {
18346 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
18347 if (cnt == -EOPNOTSUPP)
18348 goto patch_map_ops_generic;
18349 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
18350 verbose(env, "bpf verifier is misconfigured\n");
18354 new_prog = bpf_patch_insn_data(env, i + delta,
18360 env->prog = prog = new_prog;
18361 insn = new_prog->insnsi + i + delta;
18365 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
18366 (void *(*)(struct bpf_map *map, void *key))NULL));
18367 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
18368 (long (*)(struct bpf_map *map, void *key))NULL));
18369 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
18370 (long (*)(struct bpf_map *map, void *key, void *value,
18372 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
18373 (long (*)(struct bpf_map *map, void *value,
18375 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
18376 (long (*)(struct bpf_map *map, void *value))NULL));
18377 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
18378 (long (*)(struct bpf_map *map, void *value))NULL));
18379 BUILD_BUG_ON(!__same_type(ops->map_redirect,
18380 (long (*)(struct bpf_map *map, u64 index, u64 flags))NULL));
18381 BUILD_BUG_ON(!__same_type(ops->map_for_each_callback,
18382 (long (*)(struct bpf_map *map,
18383 bpf_callback_t callback_fn,
18384 void *callback_ctx,
18386 BUILD_BUG_ON(!__same_type(ops->map_lookup_percpu_elem,
18387 (void *(*)(struct bpf_map *map, void *key, u32 cpu))NULL));
18389 patch_map_ops_generic:
18390 switch (insn->imm) {
18391 case BPF_FUNC_map_lookup_elem:
18392 insn->imm = BPF_CALL_IMM(ops->map_lookup_elem);
18394 case BPF_FUNC_map_update_elem:
18395 insn->imm = BPF_CALL_IMM(ops->map_update_elem);
18397 case BPF_FUNC_map_delete_elem:
18398 insn->imm = BPF_CALL_IMM(ops->map_delete_elem);
18400 case BPF_FUNC_map_push_elem:
18401 insn->imm = BPF_CALL_IMM(ops->map_push_elem);
18403 case BPF_FUNC_map_pop_elem:
18404 insn->imm = BPF_CALL_IMM(ops->map_pop_elem);
18406 case BPF_FUNC_map_peek_elem:
18407 insn->imm = BPF_CALL_IMM(ops->map_peek_elem);
18409 case BPF_FUNC_redirect_map:
18410 insn->imm = BPF_CALL_IMM(ops->map_redirect);
18412 case BPF_FUNC_for_each_map_elem:
18413 insn->imm = BPF_CALL_IMM(ops->map_for_each_callback);
18415 case BPF_FUNC_map_lookup_percpu_elem:
18416 insn->imm = BPF_CALL_IMM(ops->map_lookup_percpu_elem);
18420 goto patch_call_imm;
18423 /* Implement bpf_jiffies64 inline. */
18424 if (prog->jit_requested && BITS_PER_LONG == 64 &&
18425 insn->imm == BPF_FUNC_jiffies64) {
18426 struct bpf_insn ld_jiffies_addr[2] = {
18427 BPF_LD_IMM64(BPF_REG_0,
18428 (unsigned long)&jiffies),
18431 insn_buf[0] = ld_jiffies_addr[0];
18432 insn_buf[1] = ld_jiffies_addr[1];
18433 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
18437 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
18443 env->prog = prog = new_prog;
18444 insn = new_prog->insnsi + i + delta;
18448 /* Implement bpf_get_func_arg inline. */
18449 if (prog_type == BPF_PROG_TYPE_TRACING &&
18450 insn->imm == BPF_FUNC_get_func_arg) {
18451 /* Load nr_args from ctx - 8 */
18452 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
18453 insn_buf[1] = BPF_JMP32_REG(BPF_JGE, BPF_REG_2, BPF_REG_0, 6);
18454 insn_buf[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_2, 3);
18455 insn_buf[3] = BPF_ALU64_REG(BPF_ADD, BPF_REG_2, BPF_REG_1);
18456 insn_buf[4] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_2, 0);
18457 insn_buf[5] = BPF_STX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
18458 insn_buf[6] = BPF_MOV64_IMM(BPF_REG_0, 0);
18459 insn_buf[7] = BPF_JMP_A(1);
18460 insn_buf[8] = BPF_MOV64_IMM(BPF_REG_0, -EINVAL);
18463 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18468 env->prog = prog = new_prog;
18469 insn = new_prog->insnsi + i + delta;
18473 /* Implement bpf_get_func_ret inline. */
18474 if (prog_type == BPF_PROG_TYPE_TRACING &&
18475 insn->imm == BPF_FUNC_get_func_ret) {
18476 if (eatype == BPF_TRACE_FEXIT ||
18477 eatype == BPF_MODIFY_RETURN) {
18478 /* Load nr_args from ctx - 8 */
18479 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
18480 insn_buf[1] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_0, 3);
18481 insn_buf[2] = BPF_ALU64_REG(BPF_ADD, BPF_REG_0, BPF_REG_1);
18482 insn_buf[3] = BPF_LDX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
18483 insn_buf[4] = BPF_STX_MEM(BPF_DW, BPF_REG_2, BPF_REG_3, 0);
18484 insn_buf[5] = BPF_MOV64_IMM(BPF_REG_0, 0);
18487 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, -EOPNOTSUPP);
18491 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18496 env->prog = prog = new_prog;
18497 insn = new_prog->insnsi + i + delta;
18501 /* Implement get_func_arg_cnt inline. */
18502 if (prog_type == BPF_PROG_TYPE_TRACING &&
18503 insn->imm == BPF_FUNC_get_func_arg_cnt) {
18504 /* Load nr_args from ctx - 8 */
18505 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
18507 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
18511 env->prog = prog = new_prog;
18512 insn = new_prog->insnsi + i + delta;
18516 /* Implement bpf_get_func_ip inline. */
18517 if (prog_type == BPF_PROG_TYPE_TRACING &&
18518 insn->imm == BPF_FUNC_get_func_ip) {
18519 /* Load IP address from ctx - 16 */
18520 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -16);
18522 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
18526 env->prog = prog = new_prog;
18527 insn = new_prog->insnsi + i + delta;
18532 fn = env->ops->get_func_proto(insn->imm, env->prog);
18533 /* all functions that have prototype and verifier allowed
18534 * programs to call them, must be real in-kernel functions
18538 "kernel subsystem misconfigured func %s#%d\n",
18539 func_id_name(insn->imm), insn->imm);
18542 insn->imm = fn->func - __bpf_call_base;
18545 /* Since poke tab is now finalized, publish aux to tracker. */
18546 for (i = 0; i < prog->aux->size_poke_tab; i++) {
18547 map_ptr = prog->aux->poke_tab[i].tail_call.map;
18548 if (!map_ptr->ops->map_poke_track ||
18549 !map_ptr->ops->map_poke_untrack ||
18550 !map_ptr->ops->map_poke_run) {
18551 verbose(env, "bpf verifier is misconfigured\n");
18555 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
18557 verbose(env, "tracking tail call prog failed\n");
18562 sort_kfunc_descs_by_imm_off(env->prog);
18567 static struct bpf_prog *inline_bpf_loop(struct bpf_verifier_env *env,
18570 u32 callback_subprogno,
18573 s32 r6_offset = stack_base + 0 * BPF_REG_SIZE;
18574 s32 r7_offset = stack_base + 1 * BPF_REG_SIZE;
18575 s32 r8_offset = stack_base + 2 * BPF_REG_SIZE;
18576 int reg_loop_max = BPF_REG_6;
18577 int reg_loop_cnt = BPF_REG_7;
18578 int reg_loop_ctx = BPF_REG_8;
18580 struct bpf_prog *new_prog;
18581 u32 callback_start;
18582 u32 call_insn_offset;
18583 s32 callback_offset;
18585 /* This represents an inlined version of bpf_iter.c:bpf_loop,
18586 * be careful to modify this code in sync.
18588 struct bpf_insn insn_buf[] = {
18589 /* Return error and jump to the end of the patch if
18590 * expected number of iterations is too big.
18592 BPF_JMP_IMM(BPF_JLE, BPF_REG_1, BPF_MAX_LOOPS, 2),
18593 BPF_MOV32_IMM(BPF_REG_0, -E2BIG),
18594 BPF_JMP_IMM(BPF_JA, 0, 0, 16),
18595 /* spill R6, R7, R8 to use these as loop vars */
18596 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_6, r6_offset),
18597 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_7, r7_offset),
18598 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_8, r8_offset),
18599 /* initialize loop vars */
18600 BPF_MOV64_REG(reg_loop_max, BPF_REG_1),
18601 BPF_MOV32_IMM(reg_loop_cnt, 0),
18602 BPF_MOV64_REG(reg_loop_ctx, BPF_REG_3),
18604 * if reg_loop_cnt >= reg_loop_max skip the loop body
18606 BPF_JMP_REG(BPF_JGE, reg_loop_cnt, reg_loop_max, 5),
18608 * correct callback offset would be set after patching
18610 BPF_MOV64_REG(BPF_REG_1, reg_loop_cnt),
18611 BPF_MOV64_REG(BPF_REG_2, reg_loop_ctx),
18613 /* increment loop counter */
18614 BPF_ALU64_IMM(BPF_ADD, reg_loop_cnt, 1),
18615 /* jump to loop header if callback returned 0 */
18616 BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, -6),
18617 /* return value of bpf_loop,
18618 * set R0 to the number of iterations
18620 BPF_MOV64_REG(BPF_REG_0, reg_loop_cnt),
18621 /* restore original values of R6, R7, R8 */
18622 BPF_LDX_MEM(BPF_DW, BPF_REG_6, BPF_REG_10, r6_offset),
18623 BPF_LDX_MEM(BPF_DW, BPF_REG_7, BPF_REG_10, r7_offset),
18624 BPF_LDX_MEM(BPF_DW, BPF_REG_8, BPF_REG_10, r8_offset),
18627 *cnt = ARRAY_SIZE(insn_buf);
18628 new_prog = bpf_patch_insn_data(env, position, insn_buf, *cnt);
18632 /* callback start is known only after patching */
18633 callback_start = env->subprog_info[callback_subprogno].start;
18634 /* Note: insn_buf[12] is an offset of BPF_CALL_REL instruction */
18635 call_insn_offset = position + 12;
18636 callback_offset = callback_start - call_insn_offset - 1;
18637 new_prog->insnsi[call_insn_offset].imm = callback_offset;
18642 static bool is_bpf_loop_call(struct bpf_insn *insn)
18644 return insn->code == (BPF_JMP | BPF_CALL) &&
18645 insn->src_reg == 0 &&
18646 insn->imm == BPF_FUNC_loop;
18649 /* For all sub-programs in the program (including main) check
18650 * insn_aux_data to see if there are bpf_loop calls that require
18651 * inlining. If such calls are found the calls are replaced with a
18652 * sequence of instructions produced by `inline_bpf_loop` function and
18653 * subprog stack_depth is increased by the size of 3 registers.
18654 * This stack space is used to spill values of the R6, R7, R8. These
18655 * registers are used to store the loop bound, counter and context
18658 static int optimize_bpf_loop(struct bpf_verifier_env *env)
18660 struct bpf_subprog_info *subprogs = env->subprog_info;
18661 int i, cur_subprog = 0, cnt, delta = 0;
18662 struct bpf_insn *insn = env->prog->insnsi;
18663 int insn_cnt = env->prog->len;
18664 u16 stack_depth = subprogs[cur_subprog].stack_depth;
18665 u16 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
18666 u16 stack_depth_extra = 0;
18668 for (i = 0; i < insn_cnt; i++, insn++) {
18669 struct bpf_loop_inline_state *inline_state =
18670 &env->insn_aux_data[i + delta].loop_inline_state;
18672 if (is_bpf_loop_call(insn) && inline_state->fit_for_inline) {
18673 struct bpf_prog *new_prog;
18675 stack_depth_extra = BPF_REG_SIZE * 3 + stack_depth_roundup;
18676 new_prog = inline_bpf_loop(env,
18678 -(stack_depth + stack_depth_extra),
18679 inline_state->callback_subprogno,
18685 env->prog = new_prog;
18686 insn = new_prog->insnsi + i + delta;
18689 if (subprogs[cur_subprog + 1].start == i + delta + 1) {
18690 subprogs[cur_subprog].stack_depth += stack_depth_extra;
18692 stack_depth = subprogs[cur_subprog].stack_depth;
18693 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
18694 stack_depth_extra = 0;
18698 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
18703 static void free_states(struct bpf_verifier_env *env)
18705 struct bpf_verifier_state_list *sl, *sln;
18708 sl = env->free_list;
18711 free_verifier_state(&sl->state, false);
18715 env->free_list = NULL;
18717 if (!env->explored_states)
18720 for (i = 0; i < state_htab_size(env); i++) {
18721 sl = env->explored_states[i];
18725 free_verifier_state(&sl->state, false);
18729 env->explored_states[i] = NULL;
18733 static int do_check_common(struct bpf_verifier_env *env, int subprog)
18735 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
18736 struct bpf_verifier_state *state;
18737 struct bpf_reg_state *regs;
18740 env->prev_linfo = NULL;
18743 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
18746 state->curframe = 0;
18747 state->speculative = false;
18748 state->branches = 1;
18749 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
18750 if (!state->frame[0]) {
18754 env->cur_state = state;
18755 init_func_state(env, state->frame[0],
18756 BPF_MAIN_FUNC /* callsite */,
18759 state->first_insn_idx = env->subprog_info[subprog].start;
18760 state->last_insn_idx = -1;
18762 regs = state->frame[state->curframe]->regs;
18763 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
18764 ret = btf_prepare_func_args(env, subprog, regs);
18767 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
18768 if (regs[i].type == PTR_TO_CTX)
18769 mark_reg_known_zero(env, regs, i);
18770 else if (regs[i].type == SCALAR_VALUE)
18771 mark_reg_unknown(env, regs, i);
18772 else if (base_type(regs[i].type) == PTR_TO_MEM) {
18773 const u32 mem_size = regs[i].mem_size;
18775 mark_reg_known_zero(env, regs, i);
18776 regs[i].mem_size = mem_size;
18777 regs[i].id = ++env->id_gen;
18781 /* 1st arg to a function */
18782 regs[BPF_REG_1].type = PTR_TO_CTX;
18783 mark_reg_known_zero(env, regs, BPF_REG_1);
18784 ret = btf_check_subprog_arg_match(env, subprog, regs);
18785 if (ret == -EFAULT)
18786 /* unlikely verifier bug. abort.
18787 * ret == 0 and ret < 0 are sadly acceptable for
18788 * main() function due to backward compatibility.
18789 * Like socket filter program may be written as:
18790 * int bpf_prog(struct pt_regs *ctx)
18791 * and never dereference that ctx in the program.
18792 * 'struct pt_regs' is a type mismatch for socket
18793 * filter that should be using 'struct __sk_buff'.
18798 ret = do_check(env);
18800 /* check for NULL is necessary, since cur_state can be freed inside
18801 * do_check() under memory pressure.
18803 if (env->cur_state) {
18804 free_verifier_state(env->cur_state, true);
18805 env->cur_state = NULL;
18807 while (!pop_stack(env, NULL, NULL, false));
18808 if (!ret && pop_log)
18809 bpf_vlog_reset(&env->log, 0);
18814 /* Verify all global functions in a BPF program one by one based on their BTF.
18815 * All global functions must pass verification. Otherwise the whole program is rejected.
18826 * foo() will be verified first for R1=any_scalar_value. During verification it
18827 * will be assumed that bar() already verified successfully and call to bar()
18828 * from foo() will be checked for type match only. Later bar() will be verified
18829 * independently to check that it's safe for R1=any_scalar_value.
18831 static int do_check_subprogs(struct bpf_verifier_env *env)
18833 struct bpf_prog_aux *aux = env->prog->aux;
18836 if (!aux->func_info)
18839 for (i = 1; i < env->subprog_cnt; i++) {
18840 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
18842 env->insn_idx = env->subprog_info[i].start;
18843 WARN_ON_ONCE(env->insn_idx == 0);
18844 ret = do_check_common(env, i);
18847 } else if (env->log.level & BPF_LOG_LEVEL) {
18849 "Func#%d is safe for any args that match its prototype\n",
18856 static int do_check_main(struct bpf_verifier_env *env)
18861 ret = do_check_common(env, 0);
18863 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
18868 static void print_verification_stats(struct bpf_verifier_env *env)
18872 if (env->log.level & BPF_LOG_STATS) {
18873 verbose(env, "verification time %lld usec\n",
18874 div_u64(env->verification_time, 1000));
18875 verbose(env, "stack depth ");
18876 for (i = 0; i < env->subprog_cnt; i++) {
18877 u32 depth = env->subprog_info[i].stack_depth;
18879 verbose(env, "%d", depth);
18880 if (i + 1 < env->subprog_cnt)
18883 verbose(env, "\n");
18885 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
18886 "total_states %d peak_states %d mark_read %d\n",
18887 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
18888 env->max_states_per_insn, env->total_states,
18889 env->peak_states, env->longest_mark_read_walk);
18892 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
18894 const struct btf_type *t, *func_proto;
18895 const struct bpf_struct_ops *st_ops;
18896 const struct btf_member *member;
18897 struct bpf_prog *prog = env->prog;
18898 u32 btf_id, member_idx;
18901 if (!prog->gpl_compatible) {
18902 verbose(env, "struct ops programs must have a GPL compatible license\n");
18906 btf_id = prog->aux->attach_btf_id;
18907 st_ops = bpf_struct_ops_find(btf_id);
18909 verbose(env, "attach_btf_id %u is not a supported struct\n",
18915 member_idx = prog->expected_attach_type;
18916 if (member_idx >= btf_type_vlen(t)) {
18917 verbose(env, "attach to invalid member idx %u of struct %s\n",
18918 member_idx, st_ops->name);
18922 member = &btf_type_member(t)[member_idx];
18923 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
18924 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
18927 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
18928 mname, member_idx, st_ops->name);
18932 if (st_ops->check_member) {
18933 int err = st_ops->check_member(t, member, prog);
18936 verbose(env, "attach to unsupported member %s of struct %s\n",
18937 mname, st_ops->name);
18942 prog->aux->attach_func_proto = func_proto;
18943 prog->aux->attach_func_name = mname;
18944 env->ops = st_ops->verifier_ops;
18948 #define SECURITY_PREFIX "security_"
18950 static int check_attach_modify_return(unsigned long addr, const char *func_name)
18952 if (within_error_injection_list(addr) ||
18953 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
18959 /* list of non-sleepable functions that are otherwise on
18960 * ALLOW_ERROR_INJECTION list
18962 BTF_SET_START(btf_non_sleepable_error_inject)
18963 /* Three functions below can be called from sleepable and non-sleepable context.
18964 * Assume non-sleepable from bpf safety point of view.
18966 BTF_ID(func, __filemap_add_folio)
18967 BTF_ID(func, should_fail_alloc_page)
18968 BTF_ID(func, should_failslab)
18969 BTF_SET_END(btf_non_sleepable_error_inject)
18971 static int check_non_sleepable_error_inject(u32 btf_id)
18973 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
18976 int bpf_check_attach_target(struct bpf_verifier_log *log,
18977 const struct bpf_prog *prog,
18978 const struct bpf_prog *tgt_prog,
18980 struct bpf_attach_target_info *tgt_info)
18982 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
18983 const char prefix[] = "btf_trace_";
18984 int ret = 0, subprog = -1, i;
18985 const struct btf_type *t;
18986 bool conservative = true;
18990 struct module *mod = NULL;
18993 bpf_log(log, "Tracing programs must provide btf_id\n");
18996 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
18999 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
19002 t = btf_type_by_id(btf, btf_id);
19004 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
19007 tname = btf_name_by_offset(btf, t->name_off);
19009 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
19013 struct bpf_prog_aux *aux = tgt_prog->aux;
19015 if (bpf_prog_is_dev_bound(prog->aux) &&
19016 !bpf_prog_dev_bound_match(prog, tgt_prog)) {
19017 bpf_log(log, "Target program bound device mismatch");
19021 for (i = 0; i < aux->func_info_cnt; i++)
19022 if (aux->func_info[i].type_id == btf_id) {
19026 if (subprog == -1) {
19027 bpf_log(log, "Subprog %s doesn't exist\n", tname);
19030 conservative = aux->func_info_aux[subprog].unreliable;
19031 if (prog_extension) {
19032 if (conservative) {
19034 "Cannot replace static functions\n");
19037 if (!prog->jit_requested) {
19039 "Extension programs should be JITed\n");
19043 if (!tgt_prog->jited) {
19044 bpf_log(log, "Can attach to only JITed progs\n");
19047 if (tgt_prog->type == prog->type) {
19048 /* Cannot fentry/fexit another fentry/fexit program.
19049 * Cannot attach program extension to another extension.
19050 * It's ok to attach fentry/fexit to extension program.
19052 bpf_log(log, "Cannot recursively attach\n");
19055 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
19057 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
19058 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
19059 /* Program extensions can extend all program types
19060 * except fentry/fexit. The reason is the following.
19061 * The fentry/fexit programs are used for performance
19062 * analysis, stats and can be attached to any program
19063 * type except themselves. When extension program is
19064 * replacing XDP function it is necessary to allow
19065 * performance analysis of all functions. Both original
19066 * XDP program and its program extension. Hence
19067 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
19068 * allowed. If extending of fentry/fexit was allowed it
19069 * would be possible to create long call chain
19070 * fentry->extension->fentry->extension beyond
19071 * reasonable stack size. Hence extending fentry is not
19074 bpf_log(log, "Cannot extend fentry/fexit\n");
19078 if (prog_extension) {
19079 bpf_log(log, "Cannot replace kernel functions\n");
19084 switch (prog->expected_attach_type) {
19085 case BPF_TRACE_RAW_TP:
19088 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
19091 if (!btf_type_is_typedef(t)) {
19092 bpf_log(log, "attach_btf_id %u is not a typedef\n",
19096 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
19097 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
19101 tname += sizeof(prefix) - 1;
19102 t = btf_type_by_id(btf, t->type);
19103 if (!btf_type_is_ptr(t))
19104 /* should never happen in valid vmlinux build */
19106 t = btf_type_by_id(btf, t->type);
19107 if (!btf_type_is_func_proto(t))
19108 /* should never happen in valid vmlinux build */
19112 case BPF_TRACE_ITER:
19113 if (!btf_type_is_func(t)) {
19114 bpf_log(log, "attach_btf_id %u is not a function\n",
19118 t = btf_type_by_id(btf, t->type);
19119 if (!btf_type_is_func_proto(t))
19121 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
19126 if (!prog_extension)
19129 case BPF_MODIFY_RETURN:
19131 case BPF_LSM_CGROUP:
19132 case BPF_TRACE_FENTRY:
19133 case BPF_TRACE_FEXIT:
19134 if (!btf_type_is_func(t)) {
19135 bpf_log(log, "attach_btf_id %u is not a function\n",
19139 if (prog_extension &&
19140 btf_check_type_match(log, prog, btf, t))
19142 t = btf_type_by_id(btf, t->type);
19143 if (!btf_type_is_func_proto(t))
19146 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
19147 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
19148 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
19151 if (tgt_prog && conservative)
19154 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
19160 addr = (long) tgt_prog->bpf_func;
19162 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
19164 if (btf_is_module(btf)) {
19165 mod = btf_try_get_module(btf);
19167 addr = find_kallsyms_symbol_value(mod, tname);
19171 addr = kallsyms_lookup_name(tname);
19176 "The address of function %s cannot be found\n",
19182 if (prog->aux->sleepable) {
19184 switch (prog->type) {
19185 case BPF_PROG_TYPE_TRACING:
19187 /* fentry/fexit/fmod_ret progs can be sleepable if they are
19188 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
19190 if (!check_non_sleepable_error_inject(btf_id) &&
19191 within_error_injection_list(addr))
19193 /* fentry/fexit/fmod_ret progs can also be sleepable if they are
19194 * in the fmodret id set with the KF_SLEEPABLE flag.
19197 u32 *flags = btf_kfunc_is_modify_return(btf, btf_id,
19200 if (flags && (*flags & KF_SLEEPABLE))
19204 case BPF_PROG_TYPE_LSM:
19205 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
19206 * Only some of them are sleepable.
19208 if (bpf_lsm_is_sleepable_hook(btf_id))
19216 bpf_log(log, "%s is not sleepable\n", tname);
19219 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
19222 bpf_log(log, "can't modify return codes of BPF programs\n");
19226 if (btf_kfunc_is_modify_return(btf, btf_id, prog) ||
19227 !check_attach_modify_return(addr, tname))
19231 bpf_log(log, "%s() is not modifiable\n", tname);
19238 tgt_info->tgt_addr = addr;
19239 tgt_info->tgt_name = tname;
19240 tgt_info->tgt_type = t;
19241 tgt_info->tgt_mod = mod;
19245 BTF_SET_START(btf_id_deny)
19248 BTF_ID(func, migrate_disable)
19249 BTF_ID(func, migrate_enable)
19251 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
19252 BTF_ID(func, rcu_read_unlock_strict)
19254 #if defined(CONFIG_DEBUG_PREEMPT) || defined(CONFIG_TRACE_PREEMPT_TOGGLE)
19255 BTF_ID(func, preempt_count_add)
19256 BTF_ID(func, preempt_count_sub)
19258 #ifdef CONFIG_PREEMPT_RCU
19259 BTF_ID(func, __rcu_read_lock)
19260 BTF_ID(func, __rcu_read_unlock)
19262 BTF_SET_END(btf_id_deny)
19264 static bool can_be_sleepable(struct bpf_prog *prog)
19266 if (prog->type == BPF_PROG_TYPE_TRACING) {
19267 switch (prog->expected_attach_type) {
19268 case BPF_TRACE_FENTRY:
19269 case BPF_TRACE_FEXIT:
19270 case BPF_MODIFY_RETURN:
19271 case BPF_TRACE_ITER:
19277 return prog->type == BPF_PROG_TYPE_LSM ||
19278 prog->type == BPF_PROG_TYPE_KPROBE /* only for uprobes */ ||
19279 prog->type == BPF_PROG_TYPE_STRUCT_OPS;
19282 static int check_attach_btf_id(struct bpf_verifier_env *env)
19284 struct bpf_prog *prog = env->prog;
19285 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
19286 struct bpf_attach_target_info tgt_info = {};
19287 u32 btf_id = prog->aux->attach_btf_id;
19288 struct bpf_trampoline *tr;
19292 if (prog->type == BPF_PROG_TYPE_SYSCALL) {
19293 if (prog->aux->sleepable)
19294 /* attach_btf_id checked to be zero already */
19296 verbose(env, "Syscall programs can only be sleepable\n");
19300 if (prog->aux->sleepable && !can_be_sleepable(prog)) {
19301 verbose(env, "Only fentry/fexit/fmod_ret, lsm, iter, uprobe, and struct_ops programs can be sleepable\n");
19305 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
19306 return check_struct_ops_btf_id(env);
19308 if (prog->type != BPF_PROG_TYPE_TRACING &&
19309 prog->type != BPF_PROG_TYPE_LSM &&
19310 prog->type != BPF_PROG_TYPE_EXT)
19313 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
19317 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
19318 /* to make freplace equivalent to their targets, they need to
19319 * inherit env->ops and expected_attach_type for the rest of the
19322 env->ops = bpf_verifier_ops[tgt_prog->type];
19323 prog->expected_attach_type = tgt_prog->expected_attach_type;
19326 /* store info about the attachment target that will be used later */
19327 prog->aux->attach_func_proto = tgt_info.tgt_type;
19328 prog->aux->attach_func_name = tgt_info.tgt_name;
19329 prog->aux->mod = tgt_info.tgt_mod;
19332 prog->aux->saved_dst_prog_type = tgt_prog->type;
19333 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
19336 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
19337 prog->aux->attach_btf_trace = true;
19339 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
19340 if (!bpf_iter_prog_supported(prog))
19345 if (prog->type == BPF_PROG_TYPE_LSM) {
19346 ret = bpf_lsm_verify_prog(&env->log, prog);
19349 } else if (prog->type == BPF_PROG_TYPE_TRACING &&
19350 btf_id_set_contains(&btf_id_deny, btf_id)) {
19354 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
19355 tr = bpf_trampoline_get(key, &tgt_info);
19359 prog->aux->dst_trampoline = tr;
19363 struct btf *bpf_get_btf_vmlinux(void)
19365 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
19366 mutex_lock(&bpf_verifier_lock);
19368 btf_vmlinux = btf_parse_vmlinux();
19369 mutex_unlock(&bpf_verifier_lock);
19371 return btf_vmlinux;
19374 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr, __u32 uattr_size)
19376 u64 start_time = ktime_get_ns();
19377 struct bpf_verifier_env *env;
19378 int i, len, ret = -EINVAL, err;
19382 /* no program is valid */
19383 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
19386 /* 'struct bpf_verifier_env' can be global, but since it's not small,
19387 * allocate/free it every time bpf_check() is called
19389 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
19395 len = (*prog)->len;
19396 env->insn_aux_data =
19397 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
19399 if (!env->insn_aux_data)
19401 for (i = 0; i < len; i++)
19402 env->insn_aux_data[i].orig_idx = i;
19404 env->ops = bpf_verifier_ops[env->prog->type];
19405 env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
19406 is_priv = bpf_capable();
19408 bpf_get_btf_vmlinux();
19410 /* grab the mutex to protect few globals used by verifier */
19412 mutex_lock(&bpf_verifier_lock);
19414 /* user could have requested verbose verifier output
19415 * and supplied buffer to store the verification trace
19417 ret = bpf_vlog_init(&env->log, attr->log_level,
19418 (char __user *) (unsigned long) attr->log_buf,
19423 mark_verifier_state_clean(env);
19425 if (IS_ERR(btf_vmlinux)) {
19426 /* Either gcc or pahole or kernel are broken. */
19427 verbose(env, "in-kernel BTF is malformed\n");
19428 ret = PTR_ERR(btf_vmlinux);
19429 goto skip_full_check;
19432 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
19433 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
19434 env->strict_alignment = true;
19435 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
19436 env->strict_alignment = false;
19438 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
19439 env->allow_uninit_stack = bpf_allow_uninit_stack();
19440 env->bypass_spec_v1 = bpf_bypass_spec_v1();
19441 env->bypass_spec_v4 = bpf_bypass_spec_v4();
19442 env->bpf_capable = bpf_capable();
19445 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
19447 env->explored_states = kvcalloc(state_htab_size(env),
19448 sizeof(struct bpf_verifier_state_list *),
19451 if (!env->explored_states)
19452 goto skip_full_check;
19454 ret = add_subprog_and_kfunc(env);
19456 goto skip_full_check;
19458 ret = check_subprogs(env);
19460 goto skip_full_check;
19462 ret = check_btf_info(env, attr, uattr);
19464 goto skip_full_check;
19466 ret = check_attach_btf_id(env);
19468 goto skip_full_check;
19470 ret = resolve_pseudo_ldimm64(env);
19472 goto skip_full_check;
19474 if (bpf_prog_is_offloaded(env->prog->aux)) {
19475 ret = bpf_prog_offload_verifier_prep(env->prog);
19477 goto skip_full_check;
19480 ret = check_cfg(env);
19482 goto skip_full_check;
19484 ret = do_check_subprogs(env);
19485 ret = ret ?: do_check_main(env);
19487 if (ret == 0 && bpf_prog_is_offloaded(env->prog->aux))
19488 ret = bpf_prog_offload_finalize(env);
19491 kvfree(env->explored_states);
19494 ret = check_max_stack_depth(env);
19496 /* instruction rewrites happen after this point */
19498 ret = optimize_bpf_loop(env);
19502 opt_hard_wire_dead_code_branches(env);
19504 ret = opt_remove_dead_code(env);
19506 ret = opt_remove_nops(env);
19509 sanitize_dead_code(env);
19513 /* program is valid, convert *(u32*)(ctx + off) accesses */
19514 ret = convert_ctx_accesses(env);
19517 ret = do_misc_fixups(env);
19519 /* do 32-bit optimization after insn patching has done so those patched
19520 * insns could be handled correctly.
19522 if (ret == 0 && !bpf_prog_is_offloaded(env->prog->aux)) {
19523 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
19524 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
19529 ret = fixup_call_args(env);
19531 env->verification_time = ktime_get_ns() - start_time;
19532 print_verification_stats(env);
19533 env->prog->aux->verified_insns = env->insn_processed;
19535 /* preserve original error even if log finalization is successful */
19536 err = bpf_vlog_finalize(&env->log, &log_true_size);
19540 if (uattr_size >= offsetofend(union bpf_attr, log_true_size) &&
19541 copy_to_bpfptr_offset(uattr, offsetof(union bpf_attr, log_true_size),
19542 &log_true_size, sizeof(log_true_size))) {
19544 goto err_release_maps;
19548 goto err_release_maps;
19550 if (env->used_map_cnt) {
19551 /* if program passed verifier, update used_maps in bpf_prog_info */
19552 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
19553 sizeof(env->used_maps[0]),
19556 if (!env->prog->aux->used_maps) {
19558 goto err_release_maps;
19561 memcpy(env->prog->aux->used_maps, env->used_maps,
19562 sizeof(env->used_maps[0]) * env->used_map_cnt);
19563 env->prog->aux->used_map_cnt = env->used_map_cnt;
19565 if (env->used_btf_cnt) {
19566 /* if program passed verifier, update used_btfs in bpf_prog_aux */
19567 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
19568 sizeof(env->used_btfs[0]),
19570 if (!env->prog->aux->used_btfs) {
19572 goto err_release_maps;
19575 memcpy(env->prog->aux->used_btfs, env->used_btfs,
19576 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
19577 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
19579 if (env->used_map_cnt || env->used_btf_cnt) {
19580 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
19581 * bpf_ld_imm64 instructions
19583 convert_pseudo_ld_imm64(env);
19586 adjust_btf_func(env);
19589 if (!env->prog->aux->used_maps)
19590 /* if we didn't copy map pointers into bpf_prog_info, release
19591 * them now. Otherwise free_used_maps() will release them.
19594 if (!env->prog->aux->used_btfs)
19597 /* extension progs temporarily inherit the attach_type of their targets
19598 for verification purposes, so set it back to zero before returning
19600 if (env->prog->type == BPF_PROG_TYPE_EXT)
19601 env->prog->expected_attach_type = 0;
19606 mutex_unlock(&bpf_verifier_lock);
19607 vfree(env->insn_aux_data);