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++) {
5624 int next_insn, sidx;
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 sidx = find_subprog(env, next_insn);
5636 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5640 if (subprog[sidx].is_async_cb) {
5641 if (subprog[sidx].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 unless called directly */
5646 if (!bpf_pseudo_call(insn + i))
5652 if (subprog[idx].has_tail_call)
5653 tail_call_reachable = true;
5656 if (frame >= MAX_CALL_FRAMES) {
5657 verbose(env, "the call stack of %d frames is too deep !\n",
5663 /* if tail call got detected across bpf2bpf calls then mark each of the
5664 * currently present subprog frames as tail call reachable subprogs;
5665 * this info will be utilized by JIT so that we will be preserving the
5666 * tail call counter throughout bpf2bpf calls combined with tailcalls
5668 if (tail_call_reachable)
5669 for (j = 0; j < frame; j++)
5670 subprog[ret_prog[j]].tail_call_reachable = true;
5671 if (subprog[0].tail_call_reachable)
5672 env->prog->aux->tail_call_reachable = true;
5674 /* end of for() loop means the last insn of the 'subprog'
5675 * was reached. Doesn't matter whether it was JA or EXIT
5679 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
5681 i = ret_insn[frame];
5682 idx = ret_prog[frame];
5686 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
5687 static int get_callee_stack_depth(struct bpf_verifier_env *env,
5688 const struct bpf_insn *insn, int idx)
5690 int start = idx + insn->imm + 1, subprog;
5692 subprog = find_subprog(env, start);
5694 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5698 return env->subprog_info[subprog].stack_depth;
5702 static int __check_buffer_access(struct bpf_verifier_env *env,
5703 const char *buf_info,
5704 const struct bpf_reg_state *reg,
5705 int regno, int off, int size)
5709 "R%d invalid %s buffer access: off=%d, size=%d\n",
5710 regno, buf_info, off, size);
5713 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
5716 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5718 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
5719 regno, off, tn_buf);
5726 static int check_tp_buffer_access(struct bpf_verifier_env *env,
5727 const struct bpf_reg_state *reg,
5728 int regno, int off, int size)
5732 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
5736 if (off + size > env->prog->aux->max_tp_access)
5737 env->prog->aux->max_tp_access = off + size;
5742 static int check_buffer_access(struct bpf_verifier_env *env,
5743 const struct bpf_reg_state *reg,
5744 int regno, int off, int size,
5745 bool zero_size_allowed,
5748 const char *buf_info = type_is_rdonly_mem(reg->type) ? "rdonly" : "rdwr";
5751 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
5755 if (off + size > *max_access)
5756 *max_access = off + size;
5761 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
5762 static void zext_32_to_64(struct bpf_reg_state *reg)
5764 reg->var_off = tnum_subreg(reg->var_off);
5765 __reg_assign_32_into_64(reg);
5768 /* truncate register to smaller size (in bytes)
5769 * must be called with size < BPF_REG_SIZE
5771 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
5775 /* clear high bits in bit representation */
5776 reg->var_off = tnum_cast(reg->var_off, size);
5778 /* fix arithmetic bounds */
5779 mask = ((u64)1 << (size * 8)) - 1;
5780 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
5781 reg->umin_value &= mask;
5782 reg->umax_value &= mask;
5784 reg->umin_value = 0;
5785 reg->umax_value = mask;
5787 reg->smin_value = reg->umin_value;
5788 reg->smax_value = reg->umax_value;
5790 /* If size is smaller than 32bit register the 32bit register
5791 * values are also truncated so we push 64-bit bounds into
5792 * 32-bit bounds. Above were truncated < 32-bits already.
5796 __reg_combine_64_into_32(reg);
5799 static bool bpf_map_is_rdonly(const struct bpf_map *map)
5801 /* A map is considered read-only if the following condition are true:
5803 * 1) BPF program side cannot change any of the map content. The
5804 * BPF_F_RDONLY_PROG flag is throughout the lifetime of a map
5805 * and was set at map creation time.
5806 * 2) The map value(s) have been initialized from user space by a
5807 * loader and then "frozen", such that no new map update/delete
5808 * operations from syscall side are possible for the rest of
5809 * the map's lifetime from that point onwards.
5810 * 3) Any parallel/pending map update/delete operations from syscall
5811 * side have been completed. Only after that point, it's safe to
5812 * assume that map value(s) are immutable.
5814 return (map->map_flags & BPF_F_RDONLY_PROG) &&
5815 READ_ONCE(map->frozen) &&
5816 !bpf_map_write_active(map);
5819 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
5825 err = map->ops->map_direct_value_addr(map, &addr, off);
5828 ptr = (void *)(long)addr + off;
5832 *val = (u64)*(u8 *)ptr;
5835 *val = (u64)*(u16 *)ptr;
5838 *val = (u64)*(u32 *)ptr;
5849 #define BTF_TYPE_SAFE_RCU(__type) __PASTE(__type, __safe_rcu)
5850 #define BTF_TYPE_SAFE_RCU_OR_NULL(__type) __PASTE(__type, __safe_rcu_or_null)
5851 #define BTF_TYPE_SAFE_TRUSTED(__type) __PASTE(__type, __safe_trusted)
5854 * Allow list few fields as RCU trusted or full trusted.
5855 * This logic doesn't allow mix tagging and will be removed once GCC supports
5859 /* RCU trusted: these fields are trusted in RCU CS and never NULL */
5860 BTF_TYPE_SAFE_RCU(struct task_struct) {
5861 const cpumask_t *cpus_ptr;
5862 struct css_set __rcu *cgroups;
5863 struct task_struct __rcu *real_parent;
5864 struct task_struct *group_leader;
5867 BTF_TYPE_SAFE_RCU(struct cgroup) {
5868 /* cgrp->kn is always accessible as documented in kernel/cgroup/cgroup.c */
5869 struct kernfs_node *kn;
5872 BTF_TYPE_SAFE_RCU(struct css_set) {
5873 struct cgroup *dfl_cgrp;
5876 /* RCU trusted: these fields are trusted in RCU CS and can be NULL */
5877 BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct) {
5878 struct file __rcu *exe_file;
5881 /* skb->sk, req->sk are not RCU protected, but we mark them as such
5882 * because bpf prog accessible sockets are SOCK_RCU_FREE.
5884 BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff) {
5888 BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock) {
5892 /* full trusted: these fields are trusted even outside of RCU CS and never NULL */
5893 BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta) {
5894 struct seq_file *seq;
5897 BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task) {
5898 struct bpf_iter_meta *meta;
5899 struct task_struct *task;
5902 BTF_TYPE_SAFE_TRUSTED(struct linux_binprm) {
5906 BTF_TYPE_SAFE_TRUSTED(struct file) {
5907 struct inode *f_inode;
5910 BTF_TYPE_SAFE_TRUSTED(struct dentry) {
5911 /* no negative dentry-s in places where bpf can see it */
5912 struct inode *d_inode;
5915 BTF_TYPE_SAFE_TRUSTED(struct socket) {
5919 static bool type_is_rcu(struct bpf_verifier_env *env,
5920 struct bpf_reg_state *reg,
5921 const char *field_name, u32 btf_id)
5923 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct task_struct));
5924 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct cgroup));
5925 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct css_set));
5927 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_rcu");
5930 static bool type_is_rcu_or_null(struct bpf_verifier_env *env,
5931 struct bpf_reg_state *reg,
5932 const char *field_name, u32 btf_id)
5934 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct));
5935 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff));
5936 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock));
5938 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_rcu_or_null");
5941 static bool type_is_trusted(struct bpf_verifier_env *env,
5942 struct bpf_reg_state *reg,
5943 const char *field_name, u32 btf_id)
5945 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta));
5946 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task));
5947 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct linux_binprm));
5948 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct file));
5949 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct dentry));
5950 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct socket));
5952 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_trusted");
5955 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
5956 struct bpf_reg_state *regs,
5957 int regno, int off, int size,
5958 enum bpf_access_type atype,
5961 struct bpf_reg_state *reg = regs + regno;
5962 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
5963 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
5964 const char *field_name = NULL;
5965 enum bpf_type_flag flag = 0;
5969 if (!env->allow_ptr_leaks) {
5971 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
5975 if (!env->prog->gpl_compatible && btf_is_kernel(reg->btf)) {
5977 "Cannot access kernel 'struct %s' from non-GPL compatible program\n",
5983 "R%d is ptr_%s invalid negative access: off=%d\n",
5987 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
5990 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5992 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
5993 regno, tname, off, tn_buf);
5997 if (reg->type & MEM_USER) {
5999 "R%d is ptr_%s access user memory: off=%d\n",
6004 if (reg->type & MEM_PERCPU) {
6006 "R%d is ptr_%s access percpu memory: off=%d\n",
6011 if (env->ops->btf_struct_access && !type_is_alloc(reg->type) && atype == BPF_WRITE) {
6012 if (!btf_is_kernel(reg->btf)) {
6013 verbose(env, "verifier internal error: reg->btf must be kernel btf\n");
6016 ret = env->ops->btf_struct_access(&env->log, reg, off, size);
6018 /* Writes are permitted with default btf_struct_access for
6019 * program allocated objects (which always have ref_obj_id > 0),
6020 * but not for untrusted PTR_TO_BTF_ID | MEM_ALLOC.
6022 if (atype != BPF_READ && !type_is_ptr_alloc_obj(reg->type)) {
6023 verbose(env, "only read is supported\n");
6027 if (type_is_alloc(reg->type) && !type_is_non_owning_ref(reg->type) &&
6029 verbose(env, "verifier internal error: ref_obj_id for allocated object must be non-zero\n");
6033 ret = btf_struct_access(&env->log, reg, off, size, atype, &btf_id, &flag, &field_name);
6039 if (ret != PTR_TO_BTF_ID) {
6042 } else if (type_flag(reg->type) & PTR_UNTRUSTED) {
6043 /* If this is an untrusted pointer, all pointers formed by walking it
6044 * also inherit the untrusted flag.
6046 flag = PTR_UNTRUSTED;
6048 } else if (is_trusted_reg(reg) || is_rcu_reg(reg)) {
6049 /* By default any pointer obtained from walking a trusted pointer is no
6050 * longer trusted, unless the field being accessed has explicitly been
6051 * marked as inheriting its parent's state of trust (either full or RCU).
6053 * 'cgroups' pointer is untrusted if task->cgroups dereference
6054 * happened in a sleepable program outside of bpf_rcu_read_lock()
6055 * section. In a non-sleepable program it's trusted while in RCU CS (aka MEM_RCU).
6056 * Note bpf_rcu_read_unlock() converts MEM_RCU pointers to PTR_UNTRUSTED.
6058 * A regular RCU-protected pointer with __rcu tag can also be deemed
6059 * trusted if we are in an RCU CS. Such pointer can be NULL.
6061 if (type_is_trusted(env, reg, field_name, btf_id)) {
6062 flag |= PTR_TRUSTED;
6063 } else if (in_rcu_cs(env) && !type_may_be_null(reg->type)) {
6064 if (type_is_rcu(env, reg, field_name, btf_id)) {
6065 /* ignore __rcu tag and mark it MEM_RCU */
6067 } else if (flag & MEM_RCU ||
6068 type_is_rcu_or_null(env, reg, field_name, btf_id)) {
6069 /* __rcu tagged pointers can be NULL */
6070 flag |= MEM_RCU | PTR_MAYBE_NULL;
6071 } else if (flag & (MEM_PERCPU | MEM_USER)) {
6074 /* walking unknown pointers yields old deprecated PTR_TO_BTF_ID */
6075 clear_trusted_flags(&flag);
6079 * If not in RCU CS or MEM_RCU pointer can be NULL then
6080 * aggressively mark as untrusted otherwise such
6081 * pointers will be plain PTR_TO_BTF_ID without flags
6082 * and will be allowed to be passed into helpers for
6085 flag = PTR_UNTRUSTED;
6088 /* Old compat. Deprecated */
6089 clear_trusted_flags(&flag);
6092 if (atype == BPF_READ && value_regno >= 0)
6093 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id, flag);
6098 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
6099 struct bpf_reg_state *regs,
6100 int regno, int off, int size,
6101 enum bpf_access_type atype,
6104 struct bpf_reg_state *reg = regs + regno;
6105 struct bpf_map *map = reg->map_ptr;
6106 struct bpf_reg_state map_reg;
6107 enum bpf_type_flag flag = 0;
6108 const struct btf_type *t;
6114 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
6118 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
6119 verbose(env, "map_ptr access not supported for map type %d\n",
6124 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
6125 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
6127 if (!env->allow_ptr_leaks) {
6129 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
6135 verbose(env, "R%d is %s invalid negative access: off=%d\n",
6140 if (atype != BPF_READ) {
6141 verbose(env, "only read from %s is supported\n", tname);
6145 /* Simulate access to a PTR_TO_BTF_ID */
6146 memset(&map_reg, 0, sizeof(map_reg));
6147 mark_btf_ld_reg(env, &map_reg, 0, PTR_TO_BTF_ID, btf_vmlinux, *map->ops->map_btf_id, 0);
6148 ret = btf_struct_access(&env->log, &map_reg, off, size, atype, &btf_id, &flag, NULL);
6152 if (value_regno >= 0)
6153 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id, flag);
6158 /* Check that the stack access at the given offset is within bounds. The
6159 * maximum valid offset is -1.
6161 * The minimum valid offset is -MAX_BPF_STACK for writes, and
6162 * -state->allocated_stack for reads.
6164 static int check_stack_slot_within_bounds(int off,
6165 struct bpf_func_state *state,
6166 enum bpf_access_type t)
6171 min_valid_off = -MAX_BPF_STACK;
6173 min_valid_off = -state->allocated_stack;
6175 if (off < min_valid_off || off > -1)
6180 /* Check that the stack access at 'regno + off' falls within the maximum stack
6183 * 'off' includes `regno->offset`, but not its dynamic part (if any).
6185 static int check_stack_access_within_bounds(
6186 struct bpf_verifier_env *env,
6187 int regno, int off, int access_size,
6188 enum bpf_access_src src, enum bpf_access_type type)
6190 struct bpf_reg_state *regs = cur_regs(env);
6191 struct bpf_reg_state *reg = regs + regno;
6192 struct bpf_func_state *state = func(env, reg);
6193 int min_off, max_off;
6197 if (src == ACCESS_HELPER)
6198 /* We don't know if helpers are reading or writing (or both). */
6199 err_extra = " indirect access to";
6200 else if (type == BPF_READ)
6201 err_extra = " read from";
6203 err_extra = " write to";
6205 if (tnum_is_const(reg->var_off)) {
6206 min_off = reg->var_off.value + off;
6207 if (access_size > 0)
6208 max_off = min_off + access_size - 1;
6212 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
6213 reg->smin_value <= -BPF_MAX_VAR_OFF) {
6214 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
6218 min_off = reg->smin_value + off;
6219 if (access_size > 0)
6220 max_off = reg->smax_value + off + access_size - 1;
6225 err = check_stack_slot_within_bounds(min_off, state, type);
6227 err = check_stack_slot_within_bounds(max_off, state, type);
6230 if (tnum_is_const(reg->var_off)) {
6231 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
6232 err_extra, regno, off, access_size);
6236 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6237 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
6238 err_extra, regno, tn_buf, access_size);
6244 /* check whether memory at (regno + off) is accessible for t = (read | write)
6245 * if t==write, value_regno is a register which value is stored into memory
6246 * if t==read, value_regno is a register which will receive the value from memory
6247 * if t==write && value_regno==-1, some unknown value is stored into memory
6248 * if t==read && value_regno==-1, don't care what we read from memory
6250 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
6251 int off, int bpf_size, enum bpf_access_type t,
6252 int value_regno, bool strict_alignment_once)
6254 struct bpf_reg_state *regs = cur_regs(env);
6255 struct bpf_reg_state *reg = regs + regno;
6256 struct bpf_func_state *state;
6259 size = bpf_size_to_bytes(bpf_size);
6263 /* alignment checks will add in reg->off themselves */
6264 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
6268 /* for access checks, reg->off is just part of off */
6271 if (reg->type == PTR_TO_MAP_KEY) {
6272 if (t == BPF_WRITE) {
6273 verbose(env, "write to change key R%d not allowed\n", regno);
6277 err = check_mem_region_access(env, regno, off, size,
6278 reg->map_ptr->key_size, false);
6281 if (value_regno >= 0)
6282 mark_reg_unknown(env, regs, value_regno);
6283 } else if (reg->type == PTR_TO_MAP_VALUE) {
6284 struct btf_field *kptr_field = NULL;
6286 if (t == BPF_WRITE && value_regno >= 0 &&
6287 is_pointer_value(env, value_regno)) {
6288 verbose(env, "R%d leaks addr into map\n", value_regno);
6291 err = check_map_access_type(env, regno, off, size, t);
6294 err = check_map_access(env, regno, off, size, false, ACCESS_DIRECT);
6297 if (tnum_is_const(reg->var_off))
6298 kptr_field = btf_record_find(reg->map_ptr->record,
6299 off + reg->var_off.value, BPF_KPTR);
6301 err = check_map_kptr_access(env, regno, value_regno, insn_idx, kptr_field);
6302 } else if (t == BPF_READ && value_regno >= 0) {
6303 struct bpf_map *map = reg->map_ptr;
6305 /* if map is read-only, track its contents as scalars */
6306 if (tnum_is_const(reg->var_off) &&
6307 bpf_map_is_rdonly(map) &&
6308 map->ops->map_direct_value_addr) {
6309 int map_off = off + reg->var_off.value;
6312 err = bpf_map_direct_read(map, map_off, size,
6317 regs[value_regno].type = SCALAR_VALUE;
6318 __mark_reg_known(®s[value_regno], val);
6320 mark_reg_unknown(env, regs, value_regno);
6323 } else if (base_type(reg->type) == PTR_TO_MEM) {
6324 bool rdonly_mem = type_is_rdonly_mem(reg->type);
6326 if (type_may_be_null(reg->type)) {
6327 verbose(env, "R%d invalid mem access '%s'\n", regno,
6328 reg_type_str(env, reg->type));
6332 if (t == BPF_WRITE && rdonly_mem) {
6333 verbose(env, "R%d cannot write into %s\n",
6334 regno, reg_type_str(env, reg->type));
6338 if (t == BPF_WRITE && value_regno >= 0 &&
6339 is_pointer_value(env, value_regno)) {
6340 verbose(env, "R%d leaks addr into mem\n", value_regno);
6344 err = check_mem_region_access(env, regno, off, size,
6345 reg->mem_size, false);
6346 if (!err && value_regno >= 0 && (t == BPF_READ || rdonly_mem))
6347 mark_reg_unknown(env, regs, value_regno);
6348 } else if (reg->type == PTR_TO_CTX) {
6349 enum bpf_reg_type reg_type = SCALAR_VALUE;
6350 struct btf *btf = NULL;
6353 if (t == BPF_WRITE && value_regno >= 0 &&
6354 is_pointer_value(env, value_regno)) {
6355 verbose(env, "R%d leaks addr into ctx\n", value_regno);
6359 err = check_ptr_off_reg(env, reg, regno);
6363 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf,
6366 verbose_linfo(env, insn_idx, "; ");
6367 if (!err && t == BPF_READ && value_regno >= 0) {
6368 /* ctx access returns either a scalar, or a
6369 * PTR_TO_PACKET[_META,_END]. In the latter
6370 * case, we know the offset is zero.
6372 if (reg_type == SCALAR_VALUE) {
6373 mark_reg_unknown(env, regs, value_regno);
6375 mark_reg_known_zero(env, regs,
6377 if (type_may_be_null(reg_type))
6378 regs[value_regno].id = ++env->id_gen;
6379 /* A load of ctx field could have different
6380 * actual load size with the one encoded in the
6381 * insn. When the dst is PTR, it is for sure not
6384 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
6385 if (base_type(reg_type) == PTR_TO_BTF_ID) {
6386 regs[value_regno].btf = btf;
6387 regs[value_regno].btf_id = btf_id;
6390 regs[value_regno].type = reg_type;
6393 } else if (reg->type == PTR_TO_STACK) {
6394 /* Basic bounds checks. */
6395 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
6399 state = func(env, reg);
6400 err = update_stack_depth(env, state, off);
6405 err = check_stack_read(env, regno, off, size,
6408 err = check_stack_write(env, regno, off, size,
6409 value_regno, insn_idx);
6410 } else if (reg_is_pkt_pointer(reg)) {
6411 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
6412 verbose(env, "cannot write into packet\n");
6415 if (t == BPF_WRITE && value_regno >= 0 &&
6416 is_pointer_value(env, value_regno)) {
6417 verbose(env, "R%d leaks addr into packet\n",
6421 err = check_packet_access(env, regno, off, size, false);
6422 if (!err && t == BPF_READ && value_regno >= 0)
6423 mark_reg_unknown(env, regs, value_regno);
6424 } else if (reg->type == PTR_TO_FLOW_KEYS) {
6425 if (t == BPF_WRITE && value_regno >= 0 &&
6426 is_pointer_value(env, value_regno)) {
6427 verbose(env, "R%d leaks addr into flow keys\n",
6432 err = check_flow_keys_access(env, off, size);
6433 if (!err && t == BPF_READ && value_regno >= 0)
6434 mark_reg_unknown(env, regs, value_regno);
6435 } else if (type_is_sk_pointer(reg->type)) {
6436 if (t == BPF_WRITE) {
6437 verbose(env, "R%d cannot write into %s\n",
6438 regno, reg_type_str(env, reg->type));
6441 err = check_sock_access(env, insn_idx, regno, off, size, t);
6442 if (!err && value_regno >= 0)
6443 mark_reg_unknown(env, regs, value_regno);
6444 } else if (reg->type == PTR_TO_TP_BUFFER) {
6445 err = check_tp_buffer_access(env, reg, regno, off, size);
6446 if (!err && t == BPF_READ && value_regno >= 0)
6447 mark_reg_unknown(env, regs, value_regno);
6448 } else if (base_type(reg->type) == PTR_TO_BTF_ID &&
6449 !type_may_be_null(reg->type)) {
6450 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
6452 } else if (reg->type == CONST_PTR_TO_MAP) {
6453 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
6455 } else if (base_type(reg->type) == PTR_TO_BUF) {
6456 bool rdonly_mem = type_is_rdonly_mem(reg->type);
6460 if (t == BPF_WRITE) {
6461 verbose(env, "R%d cannot write into %s\n",
6462 regno, reg_type_str(env, reg->type));
6465 max_access = &env->prog->aux->max_rdonly_access;
6467 max_access = &env->prog->aux->max_rdwr_access;
6470 err = check_buffer_access(env, reg, regno, off, size, false,
6473 if (!err && value_regno >= 0 && (rdonly_mem || t == BPF_READ))
6474 mark_reg_unknown(env, regs, value_regno);
6476 verbose(env, "R%d invalid mem access '%s'\n", regno,
6477 reg_type_str(env, reg->type));
6481 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
6482 regs[value_regno].type == SCALAR_VALUE) {
6483 /* b/h/w load zero-extends, mark upper bits as known 0 */
6484 coerce_reg_to_size(®s[value_regno], size);
6489 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
6494 switch (insn->imm) {
6496 case BPF_ADD | BPF_FETCH:
6498 case BPF_AND | BPF_FETCH:
6500 case BPF_OR | BPF_FETCH:
6502 case BPF_XOR | BPF_FETCH:
6507 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
6511 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
6512 verbose(env, "invalid atomic operand size\n");
6516 /* check src1 operand */
6517 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6521 /* check src2 operand */
6522 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6526 if (insn->imm == BPF_CMPXCHG) {
6527 /* Check comparison of R0 with memory location */
6528 const u32 aux_reg = BPF_REG_0;
6530 err = check_reg_arg(env, aux_reg, SRC_OP);
6534 if (is_pointer_value(env, aux_reg)) {
6535 verbose(env, "R%d leaks addr into mem\n", aux_reg);
6540 if (is_pointer_value(env, insn->src_reg)) {
6541 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
6545 if (is_ctx_reg(env, insn->dst_reg) ||
6546 is_pkt_reg(env, insn->dst_reg) ||
6547 is_flow_key_reg(env, insn->dst_reg) ||
6548 is_sk_reg(env, insn->dst_reg)) {
6549 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
6551 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
6555 if (insn->imm & BPF_FETCH) {
6556 if (insn->imm == BPF_CMPXCHG)
6557 load_reg = BPF_REG_0;
6559 load_reg = insn->src_reg;
6561 /* check and record load of old value */
6562 err = check_reg_arg(env, load_reg, DST_OP);
6566 /* This instruction accesses a memory location but doesn't
6567 * actually load it into a register.
6572 /* Check whether we can read the memory, with second call for fetch
6573 * case to simulate the register fill.
6575 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
6576 BPF_SIZE(insn->code), BPF_READ, -1, true);
6577 if (!err && load_reg >= 0)
6578 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
6579 BPF_SIZE(insn->code), BPF_READ, load_reg,
6584 /* Check whether we can write into the same memory. */
6585 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
6586 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
6593 /* When register 'regno' is used to read the stack (either directly or through
6594 * a helper function) make sure that it's within stack boundary and, depending
6595 * on the access type, that all elements of the stack are initialized.
6597 * 'off' includes 'regno->off', but not its dynamic part (if any).
6599 * All registers that have been spilled on the stack in the slots within the
6600 * read offsets are marked as read.
6602 static int check_stack_range_initialized(
6603 struct bpf_verifier_env *env, int regno, int off,
6604 int access_size, bool zero_size_allowed,
6605 enum bpf_access_src type, struct bpf_call_arg_meta *meta)
6607 struct bpf_reg_state *reg = reg_state(env, regno);
6608 struct bpf_func_state *state = func(env, reg);
6609 int err, min_off, max_off, i, j, slot, spi;
6610 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
6611 enum bpf_access_type bounds_check_type;
6612 /* Some accesses can write anything into the stack, others are
6615 bool clobber = false;
6617 if (access_size == 0 && !zero_size_allowed) {
6618 verbose(env, "invalid zero-sized read\n");
6622 if (type == ACCESS_HELPER) {
6623 /* The bounds checks for writes are more permissive than for
6624 * reads. However, if raw_mode is not set, we'll do extra
6627 bounds_check_type = BPF_WRITE;
6630 bounds_check_type = BPF_READ;
6632 err = check_stack_access_within_bounds(env, regno, off, access_size,
6633 type, bounds_check_type);
6638 if (tnum_is_const(reg->var_off)) {
6639 min_off = max_off = reg->var_off.value + off;
6641 /* Variable offset is prohibited for unprivileged mode for
6642 * simplicity since it requires corresponding support in
6643 * Spectre masking for stack ALU.
6644 * See also retrieve_ptr_limit().
6646 if (!env->bypass_spec_v1) {
6649 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6650 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
6651 regno, err_extra, tn_buf);
6654 /* Only initialized buffer on stack is allowed to be accessed
6655 * with variable offset. With uninitialized buffer it's hard to
6656 * guarantee that whole memory is marked as initialized on
6657 * helper return since specific bounds are unknown what may
6658 * cause uninitialized stack leaking.
6660 if (meta && meta->raw_mode)
6663 min_off = reg->smin_value + off;
6664 max_off = reg->smax_value + off;
6667 if (meta && meta->raw_mode) {
6668 /* Ensure we won't be overwriting dynptrs when simulating byte
6669 * by byte access in check_helper_call using meta.access_size.
6670 * This would be a problem if we have a helper in the future
6673 * helper(uninit_mem, len, dynptr)
6675 * Now, uninint_mem may overlap with dynptr pointer. Hence, it
6676 * may end up writing to dynptr itself when touching memory from
6677 * arg 1. This can be relaxed on a case by case basis for known
6678 * safe cases, but reject due to the possibilitiy of aliasing by
6681 for (i = min_off; i < max_off + access_size; i++) {
6682 int stack_off = -i - 1;
6685 /* raw_mode may write past allocated_stack */
6686 if (state->allocated_stack <= stack_off)
6688 if (state->stack[spi].slot_type[stack_off % BPF_REG_SIZE] == STACK_DYNPTR) {
6689 verbose(env, "potential write to dynptr at off=%d disallowed\n", i);
6693 meta->access_size = access_size;
6694 meta->regno = regno;
6698 for (i = min_off; i < max_off + access_size; i++) {
6702 spi = slot / BPF_REG_SIZE;
6703 if (state->allocated_stack <= slot)
6705 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
6706 if (*stype == STACK_MISC)
6708 if ((*stype == STACK_ZERO) ||
6709 (*stype == STACK_INVALID && env->allow_uninit_stack)) {
6711 /* helper can write anything into the stack */
6712 *stype = STACK_MISC;
6717 if (is_spilled_reg(&state->stack[spi]) &&
6718 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
6719 env->allow_ptr_leaks)) {
6721 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
6722 for (j = 0; j < BPF_REG_SIZE; j++)
6723 scrub_spilled_slot(&state->stack[spi].slot_type[j]);
6729 if (tnum_is_const(reg->var_off)) {
6730 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
6731 err_extra, regno, min_off, i - min_off, access_size);
6735 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6736 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
6737 err_extra, regno, tn_buf, i - min_off, access_size);
6741 /* reading any byte out of 8-byte 'spill_slot' will cause
6742 * the whole slot to be marked as 'read'
6744 mark_reg_read(env, &state->stack[spi].spilled_ptr,
6745 state->stack[spi].spilled_ptr.parent,
6747 /* We do not set REG_LIVE_WRITTEN for stack slot, as we can not
6748 * be sure that whether stack slot is written to or not. Hence,
6749 * we must still conservatively propagate reads upwards even if
6750 * helper may write to the entire memory range.
6753 return update_stack_depth(env, state, min_off);
6756 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
6757 int access_size, bool zero_size_allowed,
6758 struct bpf_call_arg_meta *meta)
6760 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
6763 switch (base_type(reg->type)) {
6765 case PTR_TO_PACKET_META:
6766 return check_packet_access(env, regno, reg->off, access_size,
6768 case PTR_TO_MAP_KEY:
6769 if (meta && meta->raw_mode) {
6770 verbose(env, "R%d cannot write into %s\n", regno,
6771 reg_type_str(env, reg->type));
6774 return check_mem_region_access(env, regno, reg->off, access_size,
6775 reg->map_ptr->key_size, false);
6776 case PTR_TO_MAP_VALUE:
6777 if (check_map_access_type(env, regno, reg->off, access_size,
6778 meta && meta->raw_mode ? BPF_WRITE :
6781 return check_map_access(env, regno, reg->off, access_size,
6782 zero_size_allowed, ACCESS_HELPER);
6784 if (type_is_rdonly_mem(reg->type)) {
6785 if (meta && meta->raw_mode) {
6786 verbose(env, "R%d cannot write into %s\n", regno,
6787 reg_type_str(env, reg->type));
6791 return check_mem_region_access(env, regno, reg->off,
6792 access_size, reg->mem_size,
6795 if (type_is_rdonly_mem(reg->type)) {
6796 if (meta && meta->raw_mode) {
6797 verbose(env, "R%d cannot write into %s\n", regno,
6798 reg_type_str(env, reg->type));
6802 max_access = &env->prog->aux->max_rdonly_access;
6804 max_access = &env->prog->aux->max_rdwr_access;
6806 return check_buffer_access(env, reg, regno, reg->off,
6807 access_size, zero_size_allowed,
6810 return check_stack_range_initialized(
6812 regno, reg->off, access_size,
6813 zero_size_allowed, ACCESS_HELPER, meta);
6815 return check_ptr_to_btf_access(env, regs, regno, reg->off,
6816 access_size, BPF_READ, -1);
6818 /* in case the function doesn't know how to access the context,
6819 * (because we are in a program of type SYSCALL for example), we
6820 * can not statically check its size.
6821 * Dynamically check it now.
6823 if (!env->ops->convert_ctx_access) {
6824 enum bpf_access_type atype = meta && meta->raw_mode ? BPF_WRITE : BPF_READ;
6825 int offset = access_size - 1;
6827 /* Allow zero-byte read from PTR_TO_CTX */
6828 if (access_size == 0)
6829 return zero_size_allowed ? 0 : -EACCES;
6831 return check_mem_access(env, env->insn_idx, regno, offset, BPF_B,
6836 default: /* scalar_value or invalid ptr */
6837 /* Allow zero-byte read from NULL, regardless of pointer type */
6838 if (zero_size_allowed && access_size == 0 &&
6839 register_is_null(reg))
6842 verbose(env, "R%d type=%s ", regno,
6843 reg_type_str(env, reg->type));
6844 verbose(env, "expected=%s\n", reg_type_str(env, PTR_TO_STACK));
6849 static int check_mem_size_reg(struct bpf_verifier_env *env,
6850 struct bpf_reg_state *reg, u32 regno,
6851 bool zero_size_allowed,
6852 struct bpf_call_arg_meta *meta)
6856 /* This is used to refine r0 return value bounds for helpers
6857 * that enforce this value as an upper bound on return values.
6858 * See do_refine_retval_range() for helpers that can refine
6859 * the return value. C type of helper is u32 so we pull register
6860 * bound from umax_value however, if negative verifier errors
6861 * out. Only upper bounds can be learned because retval is an
6862 * int type and negative retvals are allowed.
6864 meta->msize_max_value = reg->umax_value;
6866 /* The register is SCALAR_VALUE; the access check
6867 * happens using its boundaries.
6869 if (!tnum_is_const(reg->var_off))
6870 /* For unprivileged variable accesses, disable raw
6871 * mode so that the program is required to
6872 * initialize all the memory that the helper could
6873 * just partially fill up.
6877 if (reg->smin_value < 0) {
6878 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
6883 if (reg->umin_value == 0) {
6884 err = check_helper_mem_access(env, regno - 1, 0,
6891 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
6892 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
6896 err = check_helper_mem_access(env, regno - 1,
6898 zero_size_allowed, meta);
6900 err = mark_chain_precision(env, regno);
6904 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
6905 u32 regno, u32 mem_size)
6907 bool may_be_null = type_may_be_null(reg->type);
6908 struct bpf_reg_state saved_reg;
6909 struct bpf_call_arg_meta meta;
6912 if (register_is_null(reg))
6915 memset(&meta, 0, sizeof(meta));
6916 /* Assuming that the register contains a value check if the memory
6917 * access is safe. Temporarily save and restore the register's state as
6918 * the conversion shouldn't be visible to a caller.
6922 mark_ptr_not_null_reg(reg);
6925 err = check_helper_mem_access(env, regno, mem_size, true, &meta);
6926 /* Check access for BPF_WRITE */
6927 meta.raw_mode = true;
6928 err = err ?: check_helper_mem_access(env, regno, mem_size, true, &meta);
6936 static int check_kfunc_mem_size_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
6939 struct bpf_reg_state *mem_reg = &cur_regs(env)[regno - 1];
6940 bool may_be_null = type_may_be_null(mem_reg->type);
6941 struct bpf_reg_state saved_reg;
6942 struct bpf_call_arg_meta meta;
6945 WARN_ON_ONCE(regno < BPF_REG_2 || regno > BPF_REG_5);
6947 memset(&meta, 0, sizeof(meta));
6950 saved_reg = *mem_reg;
6951 mark_ptr_not_null_reg(mem_reg);
6954 err = check_mem_size_reg(env, reg, regno, true, &meta);
6955 /* Check access for BPF_WRITE */
6956 meta.raw_mode = true;
6957 err = err ?: check_mem_size_reg(env, reg, regno, true, &meta);
6960 *mem_reg = saved_reg;
6964 /* Implementation details:
6965 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL.
6966 * bpf_obj_new returns PTR_TO_BTF_ID | MEM_ALLOC | PTR_MAYBE_NULL.
6967 * Two bpf_map_lookups (even with the same key) will have different reg->id.
6968 * Two separate bpf_obj_new will also have different reg->id.
6969 * For traditional PTR_TO_MAP_VALUE or PTR_TO_BTF_ID | MEM_ALLOC, the verifier
6970 * clears reg->id after value_or_null->value transition, since the verifier only
6971 * cares about the range of access to valid map value pointer and doesn't care
6972 * about actual address of the map element.
6973 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
6974 * reg->id > 0 after value_or_null->value transition. By doing so
6975 * two bpf_map_lookups will be considered two different pointers that
6976 * point to different bpf_spin_locks. Likewise for pointers to allocated objects
6977 * returned from bpf_obj_new.
6978 * The verifier allows taking only one bpf_spin_lock at a time to avoid
6980 * Since only one bpf_spin_lock is allowed the checks are simpler than
6981 * reg_is_refcounted() logic. The verifier needs to remember only
6982 * one spin_lock instead of array of acquired_refs.
6983 * cur_state->active_lock remembers which map value element or allocated
6984 * object got locked and clears it after bpf_spin_unlock.
6986 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
6989 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
6990 struct bpf_verifier_state *cur = env->cur_state;
6991 bool is_const = tnum_is_const(reg->var_off);
6992 u64 val = reg->var_off.value;
6993 struct bpf_map *map = NULL;
6994 struct btf *btf = NULL;
6995 struct btf_record *rec;
6999 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
7003 if (reg->type == PTR_TO_MAP_VALUE) {
7007 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
7015 rec = reg_btf_record(reg);
7016 if (!btf_record_has_field(rec, BPF_SPIN_LOCK)) {
7017 verbose(env, "%s '%s' has no valid bpf_spin_lock\n", map ? "map" : "local",
7018 map ? map->name : "kptr");
7021 if (rec->spin_lock_off != val + reg->off) {
7022 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock' that is at %d\n",
7023 val + reg->off, rec->spin_lock_off);
7027 if (cur->active_lock.ptr) {
7029 "Locking two bpf_spin_locks are not allowed\n");
7033 cur->active_lock.ptr = map;
7035 cur->active_lock.ptr = btf;
7036 cur->active_lock.id = reg->id;
7045 if (!cur->active_lock.ptr) {
7046 verbose(env, "bpf_spin_unlock without taking a lock\n");
7049 if (cur->active_lock.ptr != ptr ||
7050 cur->active_lock.id != reg->id) {
7051 verbose(env, "bpf_spin_unlock of different lock\n");
7055 invalidate_non_owning_refs(env);
7057 cur->active_lock.ptr = NULL;
7058 cur->active_lock.id = 0;
7063 static int process_timer_func(struct bpf_verifier_env *env, int regno,
7064 struct bpf_call_arg_meta *meta)
7066 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7067 bool is_const = tnum_is_const(reg->var_off);
7068 struct bpf_map *map = reg->map_ptr;
7069 u64 val = reg->var_off.value;
7073 "R%d doesn't have constant offset. bpf_timer has to be at the constant offset\n",
7078 verbose(env, "map '%s' has to have BTF in order to use bpf_timer\n",
7082 if (!btf_record_has_field(map->record, BPF_TIMER)) {
7083 verbose(env, "map '%s' has no valid bpf_timer\n", map->name);
7086 if (map->record->timer_off != val + reg->off) {
7087 verbose(env, "off %lld doesn't point to 'struct bpf_timer' that is at %d\n",
7088 val + reg->off, map->record->timer_off);
7091 if (meta->map_ptr) {
7092 verbose(env, "verifier bug. Two map pointers in a timer helper\n");
7095 meta->map_uid = reg->map_uid;
7096 meta->map_ptr = map;
7100 static int process_kptr_func(struct bpf_verifier_env *env, int regno,
7101 struct bpf_call_arg_meta *meta)
7103 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7104 struct bpf_map *map_ptr = reg->map_ptr;
7105 struct btf_field *kptr_field;
7108 if (!tnum_is_const(reg->var_off)) {
7110 "R%d doesn't have constant offset. kptr has to be at the constant offset\n",
7114 if (!map_ptr->btf) {
7115 verbose(env, "map '%s' has to have BTF in order to use bpf_kptr_xchg\n",
7119 if (!btf_record_has_field(map_ptr->record, BPF_KPTR)) {
7120 verbose(env, "map '%s' has no valid kptr\n", map_ptr->name);
7124 meta->map_ptr = map_ptr;
7125 kptr_off = reg->off + reg->var_off.value;
7126 kptr_field = btf_record_find(map_ptr->record, kptr_off, BPF_KPTR);
7128 verbose(env, "off=%d doesn't point to kptr\n", kptr_off);
7131 if (kptr_field->type != BPF_KPTR_REF) {
7132 verbose(env, "off=%d kptr isn't referenced kptr\n", kptr_off);
7135 meta->kptr_field = kptr_field;
7139 /* There are two register types representing a bpf_dynptr, one is PTR_TO_STACK
7140 * which points to a stack slot, and the other is CONST_PTR_TO_DYNPTR.
7142 * In both cases we deal with the first 8 bytes, but need to mark the next 8
7143 * bytes as STACK_DYNPTR in case of PTR_TO_STACK. In case of
7144 * CONST_PTR_TO_DYNPTR, we are guaranteed to get the beginning of the object.
7146 * Mutability of bpf_dynptr is at two levels, one is at the level of struct
7147 * bpf_dynptr itself, i.e. whether the helper is receiving a pointer to struct
7148 * bpf_dynptr or pointer to const struct bpf_dynptr. In the former case, it can
7149 * mutate the view of the dynptr and also possibly destroy it. In the latter
7150 * case, it cannot mutate the bpf_dynptr itself but it can still mutate the
7151 * memory that dynptr points to.
7153 * The verifier will keep track both levels of mutation (bpf_dynptr's in
7154 * reg->type and the memory's in reg->dynptr.type), but there is no support for
7155 * readonly dynptr view yet, hence only the first case is tracked and checked.
7157 * This is consistent with how C applies the const modifier to a struct object,
7158 * where the pointer itself inside bpf_dynptr becomes const but not what it
7161 * Helpers which do not mutate the bpf_dynptr set MEM_RDONLY in their argument
7162 * type, and declare it as 'const struct bpf_dynptr *' in their prototype.
7164 static int process_dynptr_func(struct bpf_verifier_env *env, int regno, int insn_idx,
7165 enum bpf_arg_type arg_type, int clone_ref_obj_id)
7167 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7170 /* MEM_UNINIT and MEM_RDONLY are exclusive, when applied to an
7171 * ARG_PTR_TO_DYNPTR (or ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_*):
7173 if ((arg_type & (MEM_UNINIT | MEM_RDONLY)) == (MEM_UNINIT | MEM_RDONLY)) {
7174 verbose(env, "verifier internal error: misconfigured dynptr helper type flags\n");
7178 /* MEM_UNINIT - Points to memory that is an appropriate candidate for
7179 * constructing a mutable bpf_dynptr object.
7181 * Currently, this is only possible with PTR_TO_STACK
7182 * pointing to a region of at least 16 bytes which doesn't
7183 * contain an existing bpf_dynptr.
7185 * MEM_RDONLY - Points to a initialized bpf_dynptr that will not be
7186 * mutated or destroyed. However, the memory it points to
7189 * None - Points to a initialized dynptr that can be mutated and
7190 * destroyed, including mutation of the memory it points
7193 if (arg_type & MEM_UNINIT) {
7196 if (!is_dynptr_reg_valid_uninit(env, reg)) {
7197 verbose(env, "Dynptr has to be an uninitialized dynptr\n");
7201 /* we write BPF_DW bits (8 bytes) at a time */
7202 for (i = 0; i < BPF_DYNPTR_SIZE; i += 8) {
7203 err = check_mem_access(env, insn_idx, regno,
7204 i, BPF_DW, BPF_WRITE, -1, false);
7209 err = mark_stack_slots_dynptr(env, reg, arg_type, insn_idx, clone_ref_obj_id);
7210 } else /* MEM_RDONLY and None case from above */ {
7211 /* For the reg->type == PTR_TO_STACK case, bpf_dynptr is never const */
7212 if (reg->type == CONST_PTR_TO_DYNPTR && !(arg_type & MEM_RDONLY)) {
7213 verbose(env, "cannot pass pointer to const bpf_dynptr, the helper mutates it\n");
7217 if (!is_dynptr_reg_valid_init(env, reg)) {
7219 "Expected an initialized dynptr as arg #%d\n",
7224 /* Fold modifiers (in this case, MEM_RDONLY) when checking expected type */
7225 if (!is_dynptr_type_expected(env, reg, arg_type & ~MEM_RDONLY)) {
7227 "Expected a dynptr of type %s as arg #%d\n",
7228 dynptr_type_str(arg_to_dynptr_type(arg_type)), regno);
7232 err = mark_dynptr_read(env, reg);
7237 static u32 iter_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg, int spi)
7239 struct bpf_func_state *state = func(env, reg);
7241 return state->stack[spi].spilled_ptr.ref_obj_id;
7244 static bool is_iter_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7246 return meta->kfunc_flags & (KF_ITER_NEW | KF_ITER_NEXT | KF_ITER_DESTROY);
7249 static bool is_iter_new_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7251 return meta->kfunc_flags & KF_ITER_NEW;
7254 static bool is_iter_next_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7256 return meta->kfunc_flags & KF_ITER_NEXT;
7259 static bool is_iter_destroy_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7261 return meta->kfunc_flags & KF_ITER_DESTROY;
7264 static bool is_kfunc_arg_iter(struct bpf_kfunc_call_arg_meta *meta, int arg)
7266 /* btf_check_iter_kfuncs() guarantees that first argument of any iter
7267 * kfunc is iter state pointer
7269 return arg == 0 && is_iter_kfunc(meta);
7272 static int process_iter_arg(struct bpf_verifier_env *env, int regno, int insn_idx,
7273 struct bpf_kfunc_call_arg_meta *meta)
7275 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7276 const struct btf_type *t;
7277 const struct btf_param *arg;
7278 int spi, err, i, nr_slots;
7281 /* btf_check_iter_kfuncs() ensures we don't need to validate anything here */
7282 arg = &btf_params(meta->func_proto)[0];
7283 t = btf_type_skip_modifiers(meta->btf, arg->type, NULL); /* PTR */
7284 t = btf_type_skip_modifiers(meta->btf, t->type, &btf_id); /* STRUCT */
7285 nr_slots = t->size / BPF_REG_SIZE;
7287 if (is_iter_new_kfunc(meta)) {
7288 /* bpf_iter_<type>_new() expects pointer to uninit iter state */
7289 if (!is_iter_reg_valid_uninit(env, reg, nr_slots)) {
7290 verbose(env, "expected uninitialized iter_%s as arg #%d\n",
7291 iter_type_str(meta->btf, btf_id), regno);
7295 for (i = 0; i < nr_slots * 8; i += BPF_REG_SIZE) {
7296 err = check_mem_access(env, insn_idx, regno,
7297 i, BPF_DW, BPF_WRITE, -1, false);
7302 err = mark_stack_slots_iter(env, reg, insn_idx, meta->btf, btf_id, nr_slots);
7306 /* iter_next() or iter_destroy() expect initialized iter state*/
7307 if (!is_iter_reg_valid_init(env, reg, meta->btf, btf_id, nr_slots)) {
7308 verbose(env, "expected an initialized iter_%s as arg #%d\n",
7309 iter_type_str(meta->btf, btf_id), regno);
7313 spi = iter_get_spi(env, reg, nr_slots);
7317 err = mark_iter_read(env, reg, spi, nr_slots);
7321 /* remember meta->iter info for process_iter_next_call() */
7322 meta->iter.spi = spi;
7323 meta->iter.frameno = reg->frameno;
7324 meta->ref_obj_id = iter_ref_obj_id(env, reg, spi);
7326 if (is_iter_destroy_kfunc(meta)) {
7327 err = unmark_stack_slots_iter(env, reg, nr_slots);
7336 /* process_iter_next_call() is called when verifier gets to iterator's next
7337 * "method" (e.g., bpf_iter_num_next() for numbers iterator) call. We'll refer
7338 * to it as just "iter_next()" in comments below.
7340 * BPF verifier relies on a crucial contract for any iter_next()
7341 * implementation: it should *eventually* return NULL, and once that happens
7342 * it should keep returning NULL. That is, once iterator exhausts elements to
7343 * iterate, it should never reset or spuriously return new elements.
7345 * With the assumption of such contract, process_iter_next_call() simulates
7346 * a fork in the verifier state to validate loop logic correctness and safety
7347 * without having to simulate infinite amount of iterations.
7349 * In current state, we first assume that iter_next() returned NULL and
7350 * iterator state is set to DRAINED (BPF_ITER_STATE_DRAINED). In such
7351 * conditions we should not form an infinite loop and should eventually reach
7354 * Besides that, we also fork current state and enqueue it for later
7355 * verification. In a forked state we keep iterator state as ACTIVE
7356 * (BPF_ITER_STATE_ACTIVE) and assume non-NULL return from iter_next(). We
7357 * also bump iteration depth to prevent erroneous infinite loop detection
7358 * later on (see iter_active_depths_differ() comment for details). In this
7359 * state we assume that we'll eventually loop back to another iter_next()
7360 * calls (it could be in exactly same location or in some other instruction,
7361 * it doesn't matter, we don't make any unnecessary assumptions about this,
7362 * everything revolves around iterator state in a stack slot, not which
7363 * instruction is calling iter_next()). When that happens, we either will come
7364 * to iter_next() with equivalent state and can conclude that next iteration
7365 * will proceed in exactly the same way as we just verified, so it's safe to
7366 * assume that loop converges. If not, we'll go on another iteration
7367 * simulation with a different input state, until all possible starting states
7368 * are validated or we reach maximum number of instructions limit.
7370 * This way, we will either exhaustively discover all possible input states
7371 * that iterator loop can start with and eventually will converge, or we'll
7372 * effectively regress into bounded loop simulation logic and either reach
7373 * maximum number of instructions if loop is not provably convergent, or there
7374 * is some statically known limit on number of iterations (e.g., if there is
7375 * an explicit `if n > 100 then break;` statement somewhere in the loop).
7377 * One very subtle but very important aspect is that we *always* simulate NULL
7378 * condition first (as the current state) before we simulate non-NULL case.
7379 * This has to do with intricacies of scalar precision tracking. By simulating
7380 * "exit condition" of iter_next() returning NULL first, we make sure all the
7381 * relevant precision marks *that will be set **after** we exit iterator loop*
7382 * are propagated backwards to common parent state of NULL and non-NULL
7383 * branches. Thanks to that, state equivalence checks done later in forked
7384 * state, when reaching iter_next() for ACTIVE iterator, can assume that
7385 * precision marks are finalized and won't change. Because simulating another
7386 * ACTIVE iterator iteration won't change them (because given same input
7387 * states we'll end up with exactly same output states which we are currently
7388 * comparing; and verification after the loop already propagated back what
7389 * needs to be **additionally** tracked as precise). It's subtle, grok
7390 * precision tracking for more intuitive understanding.
7392 static int process_iter_next_call(struct bpf_verifier_env *env, int insn_idx,
7393 struct bpf_kfunc_call_arg_meta *meta)
7395 struct bpf_verifier_state *cur_st = env->cur_state, *queued_st;
7396 struct bpf_func_state *cur_fr = cur_st->frame[cur_st->curframe], *queued_fr;
7397 struct bpf_reg_state *cur_iter, *queued_iter;
7398 int iter_frameno = meta->iter.frameno;
7399 int iter_spi = meta->iter.spi;
7401 BTF_TYPE_EMIT(struct bpf_iter);
7403 cur_iter = &env->cur_state->frame[iter_frameno]->stack[iter_spi].spilled_ptr;
7405 if (cur_iter->iter.state != BPF_ITER_STATE_ACTIVE &&
7406 cur_iter->iter.state != BPF_ITER_STATE_DRAINED) {
7407 verbose(env, "verifier internal error: unexpected iterator state %d (%s)\n",
7408 cur_iter->iter.state, iter_state_str(cur_iter->iter.state));
7412 if (cur_iter->iter.state == BPF_ITER_STATE_ACTIVE) {
7413 /* branch out active iter state */
7414 queued_st = push_stack(env, insn_idx + 1, insn_idx, false);
7418 queued_iter = &queued_st->frame[iter_frameno]->stack[iter_spi].spilled_ptr;
7419 queued_iter->iter.state = BPF_ITER_STATE_ACTIVE;
7420 queued_iter->iter.depth++;
7422 queued_fr = queued_st->frame[queued_st->curframe];
7423 mark_ptr_not_null_reg(&queued_fr->regs[BPF_REG_0]);
7426 /* switch to DRAINED state, but keep the depth unchanged */
7427 /* mark current iter state as drained and assume returned NULL */
7428 cur_iter->iter.state = BPF_ITER_STATE_DRAINED;
7429 __mark_reg_const_zero(&cur_fr->regs[BPF_REG_0]);
7434 static bool arg_type_is_mem_size(enum bpf_arg_type type)
7436 return type == ARG_CONST_SIZE ||
7437 type == ARG_CONST_SIZE_OR_ZERO;
7440 static bool arg_type_is_release(enum bpf_arg_type type)
7442 return type & OBJ_RELEASE;
7445 static bool arg_type_is_dynptr(enum bpf_arg_type type)
7447 return base_type(type) == ARG_PTR_TO_DYNPTR;
7450 static int int_ptr_type_to_size(enum bpf_arg_type type)
7452 if (type == ARG_PTR_TO_INT)
7454 else if (type == ARG_PTR_TO_LONG)
7460 static int resolve_map_arg_type(struct bpf_verifier_env *env,
7461 const struct bpf_call_arg_meta *meta,
7462 enum bpf_arg_type *arg_type)
7464 if (!meta->map_ptr) {
7465 /* kernel subsystem misconfigured verifier */
7466 verbose(env, "invalid map_ptr to access map->type\n");
7470 switch (meta->map_ptr->map_type) {
7471 case BPF_MAP_TYPE_SOCKMAP:
7472 case BPF_MAP_TYPE_SOCKHASH:
7473 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
7474 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
7476 verbose(env, "invalid arg_type for sockmap/sockhash\n");
7480 case BPF_MAP_TYPE_BLOOM_FILTER:
7481 if (meta->func_id == BPF_FUNC_map_peek_elem)
7482 *arg_type = ARG_PTR_TO_MAP_VALUE;
7490 struct bpf_reg_types {
7491 const enum bpf_reg_type types[10];
7495 static const struct bpf_reg_types sock_types = {
7505 static const struct bpf_reg_types btf_id_sock_common_types = {
7512 PTR_TO_BTF_ID | PTR_TRUSTED,
7514 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
7518 static const struct bpf_reg_types mem_types = {
7526 PTR_TO_MEM | MEM_RINGBUF,
7528 PTR_TO_BTF_ID | PTR_TRUSTED,
7532 static const struct bpf_reg_types int_ptr_types = {
7542 static const struct bpf_reg_types spin_lock_types = {
7545 PTR_TO_BTF_ID | MEM_ALLOC,
7549 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
7550 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
7551 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
7552 static const struct bpf_reg_types ringbuf_mem_types = { .types = { PTR_TO_MEM | MEM_RINGBUF } };
7553 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
7554 static const struct bpf_reg_types btf_ptr_types = {
7557 PTR_TO_BTF_ID | PTR_TRUSTED,
7558 PTR_TO_BTF_ID | MEM_RCU,
7561 static const struct bpf_reg_types percpu_btf_ptr_types = {
7563 PTR_TO_BTF_ID | MEM_PERCPU,
7564 PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED,
7567 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
7568 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
7569 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
7570 static const struct bpf_reg_types timer_types = { .types = { PTR_TO_MAP_VALUE } };
7571 static const struct bpf_reg_types kptr_types = { .types = { PTR_TO_MAP_VALUE } };
7572 static const struct bpf_reg_types dynptr_types = {
7575 CONST_PTR_TO_DYNPTR,
7579 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
7580 [ARG_PTR_TO_MAP_KEY] = &mem_types,
7581 [ARG_PTR_TO_MAP_VALUE] = &mem_types,
7582 [ARG_CONST_SIZE] = &scalar_types,
7583 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
7584 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
7585 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
7586 [ARG_PTR_TO_CTX] = &context_types,
7587 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
7589 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
7591 [ARG_PTR_TO_SOCKET] = &fullsock_types,
7592 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
7593 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
7594 [ARG_PTR_TO_MEM] = &mem_types,
7595 [ARG_PTR_TO_RINGBUF_MEM] = &ringbuf_mem_types,
7596 [ARG_PTR_TO_INT] = &int_ptr_types,
7597 [ARG_PTR_TO_LONG] = &int_ptr_types,
7598 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
7599 [ARG_PTR_TO_FUNC] = &func_ptr_types,
7600 [ARG_PTR_TO_STACK] = &stack_ptr_types,
7601 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
7602 [ARG_PTR_TO_TIMER] = &timer_types,
7603 [ARG_PTR_TO_KPTR] = &kptr_types,
7604 [ARG_PTR_TO_DYNPTR] = &dynptr_types,
7607 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
7608 enum bpf_arg_type arg_type,
7609 const u32 *arg_btf_id,
7610 struct bpf_call_arg_meta *meta)
7612 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7613 enum bpf_reg_type expected, type = reg->type;
7614 const struct bpf_reg_types *compatible;
7617 compatible = compatible_reg_types[base_type(arg_type)];
7619 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
7623 /* ARG_PTR_TO_MEM + RDONLY is compatible with PTR_TO_MEM and PTR_TO_MEM + RDONLY,
7624 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM and NOT with PTR_TO_MEM + RDONLY
7626 * Same for MAYBE_NULL:
7628 * ARG_PTR_TO_MEM + MAYBE_NULL is compatible with PTR_TO_MEM and PTR_TO_MEM + MAYBE_NULL,
7629 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM but NOT with PTR_TO_MEM + MAYBE_NULL
7631 * ARG_PTR_TO_MEM is compatible with PTR_TO_MEM that is tagged with a dynptr type.
7633 * Therefore we fold these flags depending on the arg_type before comparison.
7635 if (arg_type & MEM_RDONLY)
7636 type &= ~MEM_RDONLY;
7637 if (arg_type & PTR_MAYBE_NULL)
7638 type &= ~PTR_MAYBE_NULL;
7639 if (base_type(arg_type) == ARG_PTR_TO_MEM)
7640 type &= ~DYNPTR_TYPE_FLAG_MASK;
7642 if (meta->func_id == BPF_FUNC_kptr_xchg && type_is_alloc(type))
7645 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
7646 expected = compatible->types[i];
7647 if (expected == NOT_INIT)
7650 if (type == expected)
7654 verbose(env, "R%d type=%s expected=", regno, reg_type_str(env, reg->type));
7655 for (j = 0; j + 1 < i; j++)
7656 verbose(env, "%s, ", reg_type_str(env, compatible->types[j]));
7657 verbose(env, "%s\n", reg_type_str(env, compatible->types[j]));
7661 if (base_type(reg->type) != PTR_TO_BTF_ID)
7664 if (compatible == &mem_types) {
7665 if (!(arg_type & MEM_RDONLY)) {
7667 "%s() may write into memory pointed by R%d type=%s\n",
7668 func_id_name(meta->func_id),
7669 regno, reg_type_str(env, reg->type));
7675 switch ((int)reg->type) {
7677 case PTR_TO_BTF_ID | PTR_TRUSTED:
7678 case PTR_TO_BTF_ID | MEM_RCU:
7679 case PTR_TO_BTF_ID | PTR_MAYBE_NULL:
7680 case PTR_TO_BTF_ID | PTR_MAYBE_NULL | MEM_RCU:
7682 /* For bpf_sk_release, it needs to match against first member
7683 * 'struct sock_common', hence make an exception for it. This
7684 * allows bpf_sk_release to work for multiple socket types.
7686 bool strict_type_match = arg_type_is_release(arg_type) &&
7687 meta->func_id != BPF_FUNC_sk_release;
7689 if (type_may_be_null(reg->type) &&
7690 (!type_may_be_null(arg_type) || arg_type_is_release(arg_type))) {
7691 verbose(env, "Possibly NULL pointer passed to helper arg%d\n", regno);
7696 if (!compatible->btf_id) {
7697 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
7700 arg_btf_id = compatible->btf_id;
7703 if (meta->func_id == BPF_FUNC_kptr_xchg) {
7704 if (map_kptr_match_type(env, meta->kptr_field, reg, regno))
7707 if (arg_btf_id == BPF_PTR_POISON) {
7708 verbose(env, "verifier internal error:");
7709 verbose(env, "R%d has non-overwritten BPF_PTR_POISON type\n",
7714 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
7715 btf_vmlinux, *arg_btf_id,
7716 strict_type_match)) {
7717 verbose(env, "R%d is of type %s but %s is expected\n",
7718 regno, btf_type_name(reg->btf, reg->btf_id),
7719 btf_type_name(btf_vmlinux, *arg_btf_id));
7725 case PTR_TO_BTF_ID | MEM_ALLOC:
7726 if (meta->func_id != BPF_FUNC_spin_lock && meta->func_id != BPF_FUNC_spin_unlock &&
7727 meta->func_id != BPF_FUNC_kptr_xchg) {
7728 verbose(env, "verifier internal error: unimplemented handling of MEM_ALLOC\n");
7731 /* Handled by helper specific checks */
7733 case PTR_TO_BTF_ID | MEM_PERCPU:
7734 case PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED:
7735 /* Handled by helper specific checks */
7738 verbose(env, "verifier internal error: invalid PTR_TO_BTF_ID register for type match\n");
7744 static struct btf_field *
7745 reg_find_field_offset(const struct bpf_reg_state *reg, s32 off, u32 fields)
7747 struct btf_field *field;
7748 struct btf_record *rec;
7750 rec = reg_btf_record(reg);
7754 field = btf_record_find(rec, off, fields);
7761 int check_func_arg_reg_off(struct bpf_verifier_env *env,
7762 const struct bpf_reg_state *reg, int regno,
7763 enum bpf_arg_type arg_type)
7765 u32 type = reg->type;
7767 /* When referenced register is passed to release function, its fixed
7770 * We will check arg_type_is_release reg has ref_obj_id when storing
7771 * meta->release_regno.
7773 if (arg_type_is_release(arg_type)) {
7774 /* ARG_PTR_TO_DYNPTR with OBJ_RELEASE is a bit special, as it
7775 * may not directly point to the object being released, but to
7776 * dynptr pointing to such object, which might be at some offset
7777 * on the stack. In that case, we simply to fallback to the
7780 if (arg_type_is_dynptr(arg_type) && type == PTR_TO_STACK)
7783 if ((type_is_ptr_alloc_obj(type) || type_is_non_owning_ref(type)) && reg->off) {
7784 if (reg_find_field_offset(reg, reg->off, BPF_GRAPH_NODE_OR_ROOT))
7785 return __check_ptr_off_reg(env, reg, regno, true);
7787 verbose(env, "R%d must have zero offset when passed to release func\n",
7789 verbose(env, "No graph node or root found at R%d type:%s off:%d\n", regno,
7790 btf_type_name(reg->btf, reg->btf_id), reg->off);
7794 /* Doing check_ptr_off_reg check for the offset will catch this
7795 * because fixed_off_ok is false, but checking here allows us
7796 * to give the user a better error message.
7799 verbose(env, "R%d must have zero offset when passed to release func or trusted arg to kfunc\n",
7803 return __check_ptr_off_reg(env, reg, regno, false);
7807 /* Pointer types where both fixed and variable offset is explicitly allowed: */
7810 case PTR_TO_PACKET_META:
7811 case PTR_TO_MAP_KEY:
7812 case PTR_TO_MAP_VALUE:
7814 case PTR_TO_MEM | MEM_RDONLY:
7815 case PTR_TO_MEM | MEM_RINGBUF:
7817 case PTR_TO_BUF | MEM_RDONLY:
7820 /* All the rest must be rejected, except PTR_TO_BTF_ID which allows
7824 case PTR_TO_BTF_ID | MEM_ALLOC:
7825 case PTR_TO_BTF_ID | PTR_TRUSTED:
7826 case PTR_TO_BTF_ID | MEM_RCU:
7827 case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF:
7828 /* When referenced PTR_TO_BTF_ID is passed to release function,
7829 * its fixed offset must be 0. In the other cases, fixed offset
7830 * can be non-zero. This was already checked above. So pass
7831 * fixed_off_ok as true to allow fixed offset for all other
7832 * cases. var_off always must be 0 for PTR_TO_BTF_ID, hence we
7833 * still need to do checks instead of returning.
7835 return __check_ptr_off_reg(env, reg, regno, true);
7837 return __check_ptr_off_reg(env, reg, regno, false);
7841 static struct bpf_reg_state *get_dynptr_arg_reg(struct bpf_verifier_env *env,
7842 const struct bpf_func_proto *fn,
7843 struct bpf_reg_state *regs)
7845 struct bpf_reg_state *state = NULL;
7848 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++)
7849 if (arg_type_is_dynptr(fn->arg_type[i])) {
7851 verbose(env, "verifier internal error: multiple dynptr args\n");
7854 state = ®s[BPF_REG_1 + i];
7858 verbose(env, "verifier internal error: no dynptr arg found\n");
7863 static int dynptr_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
7865 struct bpf_func_state *state = func(env, reg);
7868 if (reg->type == CONST_PTR_TO_DYNPTR)
7870 spi = dynptr_get_spi(env, reg);
7873 return state->stack[spi].spilled_ptr.id;
7876 static int dynptr_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
7878 struct bpf_func_state *state = func(env, reg);
7881 if (reg->type == CONST_PTR_TO_DYNPTR)
7882 return reg->ref_obj_id;
7883 spi = dynptr_get_spi(env, reg);
7886 return state->stack[spi].spilled_ptr.ref_obj_id;
7889 static enum bpf_dynptr_type dynptr_get_type(struct bpf_verifier_env *env,
7890 struct bpf_reg_state *reg)
7892 struct bpf_func_state *state = func(env, reg);
7895 if (reg->type == CONST_PTR_TO_DYNPTR)
7896 return reg->dynptr.type;
7898 spi = __get_spi(reg->off);
7900 verbose(env, "verifier internal error: invalid spi when querying dynptr type\n");
7901 return BPF_DYNPTR_TYPE_INVALID;
7904 return state->stack[spi].spilled_ptr.dynptr.type;
7907 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
7908 struct bpf_call_arg_meta *meta,
7909 const struct bpf_func_proto *fn,
7912 u32 regno = BPF_REG_1 + arg;
7913 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7914 enum bpf_arg_type arg_type = fn->arg_type[arg];
7915 enum bpf_reg_type type = reg->type;
7916 u32 *arg_btf_id = NULL;
7919 if (arg_type == ARG_DONTCARE)
7922 err = check_reg_arg(env, regno, SRC_OP);
7926 if (arg_type == ARG_ANYTHING) {
7927 if (is_pointer_value(env, regno)) {
7928 verbose(env, "R%d leaks addr into helper function\n",
7935 if (type_is_pkt_pointer(type) &&
7936 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
7937 verbose(env, "helper access to the packet is not allowed\n");
7941 if (base_type(arg_type) == ARG_PTR_TO_MAP_VALUE) {
7942 err = resolve_map_arg_type(env, meta, &arg_type);
7947 if (register_is_null(reg) && type_may_be_null(arg_type))
7948 /* A NULL register has a SCALAR_VALUE type, so skip
7951 goto skip_type_check;
7953 /* arg_btf_id and arg_size are in a union. */
7954 if (base_type(arg_type) == ARG_PTR_TO_BTF_ID ||
7955 base_type(arg_type) == ARG_PTR_TO_SPIN_LOCK)
7956 arg_btf_id = fn->arg_btf_id[arg];
7958 err = check_reg_type(env, regno, arg_type, arg_btf_id, meta);
7962 err = check_func_arg_reg_off(env, reg, regno, arg_type);
7967 if (arg_type_is_release(arg_type)) {
7968 if (arg_type_is_dynptr(arg_type)) {
7969 struct bpf_func_state *state = func(env, reg);
7972 /* Only dynptr created on stack can be released, thus
7973 * the get_spi and stack state checks for spilled_ptr
7974 * should only be done before process_dynptr_func for
7977 if (reg->type == PTR_TO_STACK) {
7978 spi = dynptr_get_spi(env, reg);
7979 if (spi < 0 || !state->stack[spi].spilled_ptr.ref_obj_id) {
7980 verbose(env, "arg %d is an unacquired reference\n", regno);
7984 verbose(env, "cannot release unowned const bpf_dynptr\n");
7987 } else if (!reg->ref_obj_id && !register_is_null(reg)) {
7988 verbose(env, "R%d must be referenced when passed to release function\n",
7992 if (meta->release_regno) {
7993 verbose(env, "verifier internal error: more than one release argument\n");
7996 meta->release_regno = regno;
7999 if (reg->ref_obj_id) {
8000 if (meta->ref_obj_id) {
8001 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
8002 regno, reg->ref_obj_id,
8006 meta->ref_obj_id = reg->ref_obj_id;
8009 switch (base_type(arg_type)) {
8010 case ARG_CONST_MAP_PTR:
8011 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
8012 if (meta->map_ptr) {
8013 /* Use map_uid (which is unique id of inner map) to reject:
8014 * inner_map1 = bpf_map_lookup_elem(outer_map, key1)
8015 * inner_map2 = bpf_map_lookup_elem(outer_map, key2)
8016 * if (inner_map1 && inner_map2) {
8017 * timer = bpf_map_lookup_elem(inner_map1);
8019 * // mismatch would have been allowed
8020 * bpf_timer_init(timer, inner_map2);
8023 * Comparing map_ptr is enough to distinguish normal and outer maps.
8025 if (meta->map_ptr != reg->map_ptr ||
8026 meta->map_uid != reg->map_uid) {
8028 "timer pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
8029 meta->map_uid, reg->map_uid);
8033 meta->map_ptr = reg->map_ptr;
8034 meta->map_uid = reg->map_uid;
8036 case ARG_PTR_TO_MAP_KEY:
8037 /* bpf_map_xxx(..., map_ptr, ..., key) call:
8038 * check that [key, key + map->key_size) are within
8039 * stack limits and initialized
8041 if (!meta->map_ptr) {
8042 /* in function declaration map_ptr must come before
8043 * map_key, so that it's verified and known before
8044 * we have to check map_key here. Otherwise it means
8045 * that kernel subsystem misconfigured verifier
8047 verbose(env, "invalid map_ptr to access map->key\n");
8050 err = check_helper_mem_access(env, regno,
8051 meta->map_ptr->key_size, false,
8054 case ARG_PTR_TO_MAP_VALUE:
8055 if (type_may_be_null(arg_type) && register_is_null(reg))
8058 /* bpf_map_xxx(..., map_ptr, ..., value) call:
8059 * check [value, value + map->value_size) validity
8061 if (!meta->map_ptr) {
8062 /* kernel subsystem misconfigured verifier */
8063 verbose(env, "invalid map_ptr to access map->value\n");
8066 meta->raw_mode = arg_type & MEM_UNINIT;
8067 err = check_helper_mem_access(env, regno,
8068 meta->map_ptr->value_size, false,
8071 case ARG_PTR_TO_PERCPU_BTF_ID:
8073 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
8076 meta->ret_btf = reg->btf;
8077 meta->ret_btf_id = reg->btf_id;
8079 case ARG_PTR_TO_SPIN_LOCK:
8080 if (in_rbtree_lock_required_cb(env)) {
8081 verbose(env, "can't spin_{lock,unlock} in rbtree cb\n");
8084 if (meta->func_id == BPF_FUNC_spin_lock) {
8085 err = process_spin_lock(env, regno, true);
8088 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
8089 err = process_spin_lock(env, regno, false);
8093 verbose(env, "verifier internal error\n");
8097 case ARG_PTR_TO_TIMER:
8098 err = process_timer_func(env, regno, meta);
8102 case ARG_PTR_TO_FUNC:
8103 meta->subprogno = reg->subprogno;
8105 case ARG_PTR_TO_MEM:
8106 /* The access to this pointer is only checked when we hit the
8107 * next is_mem_size argument below.
8109 meta->raw_mode = arg_type & MEM_UNINIT;
8110 if (arg_type & MEM_FIXED_SIZE) {
8111 err = check_helper_mem_access(env, regno,
8112 fn->arg_size[arg], false,
8116 case ARG_CONST_SIZE:
8117 err = check_mem_size_reg(env, reg, regno, false, meta);
8119 case ARG_CONST_SIZE_OR_ZERO:
8120 err = check_mem_size_reg(env, reg, regno, true, meta);
8122 case ARG_PTR_TO_DYNPTR:
8123 err = process_dynptr_func(env, regno, insn_idx, arg_type, 0);
8127 case ARG_CONST_ALLOC_SIZE_OR_ZERO:
8128 if (!tnum_is_const(reg->var_off)) {
8129 verbose(env, "R%d is not a known constant'\n",
8133 meta->mem_size = reg->var_off.value;
8134 err = mark_chain_precision(env, regno);
8138 case ARG_PTR_TO_INT:
8139 case ARG_PTR_TO_LONG:
8141 int size = int_ptr_type_to_size(arg_type);
8143 err = check_helper_mem_access(env, regno, size, false, meta);
8146 err = check_ptr_alignment(env, reg, 0, size, true);
8149 case ARG_PTR_TO_CONST_STR:
8151 struct bpf_map *map = reg->map_ptr;
8156 if (!bpf_map_is_rdonly(map)) {
8157 verbose(env, "R%d does not point to a readonly map'\n", regno);
8161 if (!tnum_is_const(reg->var_off)) {
8162 verbose(env, "R%d is not a constant address'\n", regno);
8166 if (!map->ops->map_direct_value_addr) {
8167 verbose(env, "no direct value access support for this map type\n");
8171 err = check_map_access(env, regno, reg->off,
8172 map->value_size - reg->off, false,
8177 map_off = reg->off + reg->var_off.value;
8178 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
8180 verbose(env, "direct value access on string failed\n");
8184 str_ptr = (char *)(long)(map_addr);
8185 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
8186 verbose(env, "string is not zero-terminated\n");
8191 case ARG_PTR_TO_KPTR:
8192 err = process_kptr_func(env, regno, meta);
8201 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
8203 enum bpf_attach_type eatype = env->prog->expected_attach_type;
8204 enum bpf_prog_type type = resolve_prog_type(env->prog);
8206 if (func_id != BPF_FUNC_map_update_elem)
8209 /* It's not possible to get access to a locked struct sock in these
8210 * contexts, so updating is safe.
8213 case BPF_PROG_TYPE_TRACING:
8214 if (eatype == BPF_TRACE_ITER)
8217 case BPF_PROG_TYPE_SOCKET_FILTER:
8218 case BPF_PROG_TYPE_SCHED_CLS:
8219 case BPF_PROG_TYPE_SCHED_ACT:
8220 case BPF_PROG_TYPE_XDP:
8221 case BPF_PROG_TYPE_SK_REUSEPORT:
8222 case BPF_PROG_TYPE_FLOW_DISSECTOR:
8223 case BPF_PROG_TYPE_SK_LOOKUP:
8229 verbose(env, "cannot update sockmap in this context\n");
8233 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
8235 return env->prog->jit_requested &&
8236 bpf_jit_supports_subprog_tailcalls();
8239 static int check_map_func_compatibility(struct bpf_verifier_env *env,
8240 struct bpf_map *map, int func_id)
8245 /* We need a two way check, first is from map perspective ... */
8246 switch (map->map_type) {
8247 case BPF_MAP_TYPE_PROG_ARRAY:
8248 if (func_id != BPF_FUNC_tail_call)
8251 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
8252 if (func_id != BPF_FUNC_perf_event_read &&
8253 func_id != BPF_FUNC_perf_event_output &&
8254 func_id != BPF_FUNC_skb_output &&
8255 func_id != BPF_FUNC_perf_event_read_value &&
8256 func_id != BPF_FUNC_xdp_output)
8259 case BPF_MAP_TYPE_RINGBUF:
8260 if (func_id != BPF_FUNC_ringbuf_output &&
8261 func_id != BPF_FUNC_ringbuf_reserve &&
8262 func_id != BPF_FUNC_ringbuf_query &&
8263 func_id != BPF_FUNC_ringbuf_reserve_dynptr &&
8264 func_id != BPF_FUNC_ringbuf_submit_dynptr &&
8265 func_id != BPF_FUNC_ringbuf_discard_dynptr)
8268 case BPF_MAP_TYPE_USER_RINGBUF:
8269 if (func_id != BPF_FUNC_user_ringbuf_drain)
8272 case BPF_MAP_TYPE_STACK_TRACE:
8273 if (func_id != BPF_FUNC_get_stackid)
8276 case BPF_MAP_TYPE_CGROUP_ARRAY:
8277 if (func_id != BPF_FUNC_skb_under_cgroup &&
8278 func_id != BPF_FUNC_current_task_under_cgroup)
8281 case BPF_MAP_TYPE_CGROUP_STORAGE:
8282 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
8283 if (func_id != BPF_FUNC_get_local_storage)
8286 case BPF_MAP_TYPE_DEVMAP:
8287 case BPF_MAP_TYPE_DEVMAP_HASH:
8288 if (func_id != BPF_FUNC_redirect_map &&
8289 func_id != BPF_FUNC_map_lookup_elem)
8292 /* Restrict bpf side of cpumap and xskmap, open when use-cases
8295 case BPF_MAP_TYPE_CPUMAP:
8296 if (func_id != BPF_FUNC_redirect_map)
8299 case BPF_MAP_TYPE_XSKMAP:
8300 if (func_id != BPF_FUNC_redirect_map &&
8301 func_id != BPF_FUNC_map_lookup_elem)
8304 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
8305 case BPF_MAP_TYPE_HASH_OF_MAPS:
8306 if (func_id != BPF_FUNC_map_lookup_elem)
8309 case BPF_MAP_TYPE_SOCKMAP:
8310 if (func_id != BPF_FUNC_sk_redirect_map &&
8311 func_id != BPF_FUNC_sock_map_update &&
8312 func_id != BPF_FUNC_map_delete_elem &&
8313 func_id != BPF_FUNC_msg_redirect_map &&
8314 func_id != BPF_FUNC_sk_select_reuseport &&
8315 func_id != BPF_FUNC_map_lookup_elem &&
8316 !may_update_sockmap(env, func_id))
8319 case BPF_MAP_TYPE_SOCKHASH:
8320 if (func_id != BPF_FUNC_sk_redirect_hash &&
8321 func_id != BPF_FUNC_sock_hash_update &&
8322 func_id != BPF_FUNC_map_delete_elem &&
8323 func_id != BPF_FUNC_msg_redirect_hash &&
8324 func_id != BPF_FUNC_sk_select_reuseport &&
8325 func_id != BPF_FUNC_map_lookup_elem &&
8326 !may_update_sockmap(env, func_id))
8329 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
8330 if (func_id != BPF_FUNC_sk_select_reuseport)
8333 case BPF_MAP_TYPE_QUEUE:
8334 case BPF_MAP_TYPE_STACK:
8335 if (func_id != BPF_FUNC_map_peek_elem &&
8336 func_id != BPF_FUNC_map_pop_elem &&
8337 func_id != BPF_FUNC_map_push_elem)
8340 case BPF_MAP_TYPE_SK_STORAGE:
8341 if (func_id != BPF_FUNC_sk_storage_get &&
8342 func_id != BPF_FUNC_sk_storage_delete &&
8343 func_id != BPF_FUNC_kptr_xchg)
8346 case BPF_MAP_TYPE_INODE_STORAGE:
8347 if (func_id != BPF_FUNC_inode_storage_get &&
8348 func_id != BPF_FUNC_inode_storage_delete &&
8349 func_id != BPF_FUNC_kptr_xchg)
8352 case BPF_MAP_TYPE_TASK_STORAGE:
8353 if (func_id != BPF_FUNC_task_storage_get &&
8354 func_id != BPF_FUNC_task_storage_delete &&
8355 func_id != BPF_FUNC_kptr_xchg)
8358 case BPF_MAP_TYPE_CGRP_STORAGE:
8359 if (func_id != BPF_FUNC_cgrp_storage_get &&
8360 func_id != BPF_FUNC_cgrp_storage_delete &&
8361 func_id != BPF_FUNC_kptr_xchg)
8364 case BPF_MAP_TYPE_BLOOM_FILTER:
8365 if (func_id != BPF_FUNC_map_peek_elem &&
8366 func_id != BPF_FUNC_map_push_elem)
8373 /* ... and second from the function itself. */
8375 case BPF_FUNC_tail_call:
8376 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
8378 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
8379 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
8383 case BPF_FUNC_perf_event_read:
8384 case BPF_FUNC_perf_event_output:
8385 case BPF_FUNC_perf_event_read_value:
8386 case BPF_FUNC_skb_output:
8387 case BPF_FUNC_xdp_output:
8388 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
8391 case BPF_FUNC_ringbuf_output:
8392 case BPF_FUNC_ringbuf_reserve:
8393 case BPF_FUNC_ringbuf_query:
8394 case BPF_FUNC_ringbuf_reserve_dynptr:
8395 case BPF_FUNC_ringbuf_submit_dynptr:
8396 case BPF_FUNC_ringbuf_discard_dynptr:
8397 if (map->map_type != BPF_MAP_TYPE_RINGBUF)
8400 case BPF_FUNC_user_ringbuf_drain:
8401 if (map->map_type != BPF_MAP_TYPE_USER_RINGBUF)
8404 case BPF_FUNC_get_stackid:
8405 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
8408 case BPF_FUNC_current_task_under_cgroup:
8409 case BPF_FUNC_skb_under_cgroup:
8410 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
8413 case BPF_FUNC_redirect_map:
8414 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
8415 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
8416 map->map_type != BPF_MAP_TYPE_CPUMAP &&
8417 map->map_type != BPF_MAP_TYPE_XSKMAP)
8420 case BPF_FUNC_sk_redirect_map:
8421 case BPF_FUNC_msg_redirect_map:
8422 case BPF_FUNC_sock_map_update:
8423 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
8426 case BPF_FUNC_sk_redirect_hash:
8427 case BPF_FUNC_msg_redirect_hash:
8428 case BPF_FUNC_sock_hash_update:
8429 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
8432 case BPF_FUNC_get_local_storage:
8433 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
8434 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
8437 case BPF_FUNC_sk_select_reuseport:
8438 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
8439 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
8440 map->map_type != BPF_MAP_TYPE_SOCKHASH)
8443 case BPF_FUNC_map_pop_elem:
8444 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
8445 map->map_type != BPF_MAP_TYPE_STACK)
8448 case BPF_FUNC_map_peek_elem:
8449 case BPF_FUNC_map_push_elem:
8450 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
8451 map->map_type != BPF_MAP_TYPE_STACK &&
8452 map->map_type != BPF_MAP_TYPE_BLOOM_FILTER)
8455 case BPF_FUNC_map_lookup_percpu_elem:
8456 if (map->map_type != BPF_MAP_TYPE_PERCPU_ARRAY &&
8457 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
8458 map->map_type != BPF_MAP_TYPE_LRU_PERCPU_HASH)
8461 case BPF_FUNC_sk_storage_get:
8462 case BPF_FUNC_sk_storage_delete:
8463 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
8466 case BPF_FUNC_inode_storage_get:
8467 case BPF_FUNC_inode_storage_delete:
8468 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
8471 case BPF_FUNC_task_storage_get:
8472 case BPF_FUNC_task_storage_delete:
8473 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
8476 case BPF_FUNC_cgrp_storage_get:
8477 case BPF_FUNC_cgrp_storage_delete:
8478 if (map->map_type != BPF_MAP_TYPE_CGRP_STORAGE)
8487 verbose(env, "cannot pass map_type %d into func %s#%d\n",
8488 map->map_type, func_id_name(func_id), func_id);
8492 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
8496 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
8498 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
8500 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
8502 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
8504 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
8507 /* We only support one arg being in raw mode at the moment,
8508 * which is sufficient for the helper functions we have
8514 static bool check_args_pair_invalid(const struct bpf_func_proto *fn, int arg)
8516 bool is_fixed = fn->arg_type[arg] & MEM_FIXED_SIZE;
8517 bool has_size = fn->arg_size[arg] != 0;
8518 bool is_next_size = false;
8520 if (arg + 1 < ARRAY_SIZE(fn->arg_type))
8521 is_next_size = arg_type_is_mem_size(fn->arg_type[arg + 1]);
8523 if (base_type(fn->arg_type[arg]) != ARG_PTR_TO_MEM)
8524 return is_next_size;
8526 return has_size == is_next_size || is_next_size == is_fixed;
8529 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
8531 /* bpf_xxx(..., buf, len) call will access 'len'
8532 * bytes from memory 'buf'. Both arg types need
8533 * to be paired, so make sure there's no buggy
8534 * helper function specification.
8536 if (arg_type_is_mem_size(fn->arg1_type) ||
8537 check_args_pair_invalid(fn, 0) ||
8538 check_args_pair_invalid(fn, 1) ||
8539 check_args_pair_invalid(fn, 2) ||
8540 check_args_pair_invalid(fn, 3) ||
8541 check_args_pair_invalid(fn, 4))
8547 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
8551 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
8552 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_BTF_ID)
8553 return !!fn->arg_btf_id[i];
8554 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_SPIN_LOCK)
8555 return fn->arg_btf_id[i] == BPF_PTR_POISON;
8556 if (base_type(fn->arg_type[i]) != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i] &&
8557 /* arg_btf_id and arg_size are in a union. */
8558 (base_type(fn->arg_type[i]) != ARG_PTR_TO_MEM ||
8559 !(fn->arg_type[i] & MEM_FIXED_SIZE)))
8566 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
8568 return check_raw_mode_ok(fn) &&
8569 check_arg_pair_ok(fn) &&
8570 check_btf_id_ok(fn) ? 0 : -EINVAL;
8573 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
8574 * are now invalid, so turn them into unknown SCALAR_VALUE.
8576 * This also applies to dynptr slices belonging to skb and xdp dynptrs,
8577 * since these slices point to packet data.
8579 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
8581 struct bpf_func_state *state;
8582 struct bpf_reg_state *reg;
8584 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
8585 if (reg_is_pkt_pointer_any(reg) || reg_is_dynptr_slice_pkt(reg))
8586 mark_reg_invalid(env, reg);
8592 BEYOND_PKT_END = -2,
8595 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
8597 struct bpf_func_state *state = vstate->frame[vstate->curframe];
8598 struct bpf_reg_state *reg = &state->regs[regn];
8600 if (reg->type != PTR_TO_PACKET)
8601 /* PTR_TO_PACKET_META is not supported yet */
8604 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
8605 * How far beyond pkt_end it goes is unknown.
8606 * if (!range_open) it's the case of pkt >= pkt_end
8607 * if (range_open) it's the case of pkt > pkt_end
8608 * hence this pointer is at least 1 byte bigger than pkt_end
8611 reg->range = BEYOND_PKT_END;
8613 reg->range = AT_PKT_END;
8616 /* The pointer with the specified id has released its reference to kernel
8617 * resources. Identify all copies of the same pointer and clear the reference.
8619 static int release_reference(struct bpf_verifier_env *env,
8622 struct bpf_func_state *state;
8623 struct bpf_reg_state *reg;
8626 err = release_reference_state(cur_func(env), ref_obj_id);
8630 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
8631 if (reg->ref_obj_id == ref_obj_id)
8632 mark_reg_invalid(env, reg);
8638 static void invalidate_non_owning_refs(struct bpf_verifier_env *env)
8640 struct bpf_func_state *unused;
8641 struct bpf_reg_state *reg;
8643 bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
8644 if (type_is_non_owning_ref(reg->type))
8645 mark_reg_invalid(env, reg);
8649 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
8650 struct bpf_reg_state *regs)
8654 /* after the call registers r0 - r5 were scratched */
8655 for (i = 0; i < CALLER_SAVED_REGS; i++) {
8656 mark_reg_not_init(env, regs, caller_saved[i]);
8657 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
8661 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
8662 struct bpf_func_state *caller,
8663 struct bpf_func_state *callee,
8666 static int set_callee_state(struct bpf_verifier_env *env,
8667 struct bpf_func_state *caller,
8668 struct bpf_func_state *callee, int insn_idx);
8670 static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
8671 int *insn_idx, int subprog,
8672 set_callee_state_fn set_callee_state_cb)
8674 struct bpf_verifier_state *state = env->cur_state;
8675 struct bpf_func_state *caller, *callee;
8678 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
8679 verbose(env, "the call stack of %d frames is too deep\n",
8680 state->curframe + 2);
8684 caller = state->frame[state->curframe];
8685 if (state->frame[state->curframe + 1]) {
8686 verbose(env, "verifier bug. Frame %d already allocated\n",
8687 state->curframe + 1);
8691 err = btf_check_subprog_call(env, subprog, caller->regs);
8694 if (subprog_is_global(env, subprog)) {
8696 verbose(env, "Caller passes invalid args into func#%d\n",
8700 if (env->log.level & BPF_LOG_LEVEL)
8702 "Func#%d is global and valid. Skipping.\n",
8704 clear_caller_saved_regs(env, caller->regs);
8706 /* All global functions return a 64-bit SCALAR_VALUE */
8707 mark_reg_unknown(env, caller->regs, BPF_REG_0);
8708 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
8710 /* continue with next insn after call */
8715 /* set_callee_state is used for direct subprog calls, but we are
8716 * interested in validating only BPF helpers that can call subprogs as
8719 if (set_callee_state_cb != set_callee_state) {
8720 if (bpf_pseudo_kfunc_call(insn) &&
8721 !is_callback_calling_kfunc(insn->imm)) {
8722 verbose(env, "verifier bug: kfunc %s#%d not marked as callback-calling\n",
8723 func_id_name(insn->imm), insn->imm);
8725 } else if (!bpf_pseudo_kfunc_call(insn) &&
8726 !is_callback_calling_function(insn->imm)) { /* helper */
8727 verbose(env, "verifier bug: helper %s#%d not marked as callback-calling\n",
8728 func_id_name(insn->imm), insn->imm);
8733 if (insn->code == (BPF_JMP | BPF_CALL) &&
8734 insn->src_reg == 0 &&
8735 insn->imm == BPF_FUNC_timer_set_callback) {
8736 struct bpf_verifier_state *async_cb;
8738 /* there is no real recursion here. timer callbacks are async */
8739 env->subprog_info[subprog].is_async_cb = true;
8740 async_cb = push_async_cb(env, env->subprog_info[subprog].start,
8741 *insn_idx, subprog);
8744 callee = async_cb->frame[0];
8745 callee->async_entry_cnt = caller->async_entry_cnt + 1;
8747 /* Convert bpf_timer_set_callback() args into timer callback args */
8748 err = set_callee_state_cb(env, caller, callee, *insn_idx);
8752 clear_caller_saved_regs(env, caller->regs);
8753 mark_reg_unknown(env, caller->regs, BPF_REG_0);
8754 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
8755 /* continue with next insn after call */
8759 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
8762 state->frame[state->curframe + 1] = callee;
8764 /* callee cannot access r0, r6 - r9 for reading and has to write
8765 * into its own stack before reading from it.
8766 * callee can read/write into caller's stack
8768 init_func_state(env, callee,
8769 /* remember the callsite, it will be used by bpf_exit */
8770 *insn_idx /* callsite */,
8771 state->curframe + 1 /* frameno within this callchain */,
8772 subprog /* subprog number within this prog */);
8774 /* Transfer references to the callee */
8775 err = copy_reference_state(callee, caller);
8779 err = set_callee_state_cb(env, caller, callee, *insn_idx);
8783 clear_caller_saved_regs(env, caller->regs);
8785 /* only increment it after check_reg_arg() finished */
8788 /* and go analyze first insn of the callee */
8789 *insn_idx = env->subprog_info[subprog].start - 1;
8791 if (env->log.level & BPF_LOG_LEVEL) {
8792 verbose(env, "caller:\n");
8793 print_verifier_state(env, caller, true);
8794 verbose(env, "callee:\n");
8795 print_verifier_state(env, callee, true);
8800 free_func_state(callee);
8801 state->frame[state->curframe + 1] = NULL;
8805 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
8806 struct bpf_func_state *caller,
8807 struct bpf_func_state *callee)
8809 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
8810 * void *callback_ctx, u64 flags);
8811 * callback_fn(struct bpf_map *map, void *key, void *value,
8812 * void *callback_ctx);
8814 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
8816 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
8817 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
8818 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
8820 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
8821 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
8822 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
8824 /* pointer to stack or null */
8825 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
8828 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
8832 static int set_callee_state(struct bpf_verifier_env *env,
8833 struct bpf_func_state *caller,
8834 struct bpf_func_state *callee, int insn_idx)
8838 /* copy r1 - r5 args that callee can access. The copy includes parent
8839 * pointers, which connects us up to the liveness chain
8841 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
8842 callee->regs[i] = caller->regs[i];
8846 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
8849 int subprog, target_insn;
8851 target_insn = *insn_idx + insn->imm + 1;
8852 subprog = find_subprog(env, target_insn);
8854 verbose(env, "verifier bug. No program starts at insn %d\n",
8859 return __check_func_call(env, insn, insn_idx, subprog, set_callee_state);
8862 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
8863 struct bpf_func_state *caller,
8864 struct bpf_func_state *callee,
8867 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
8868 struct bpf_map *map;
8871 if (bpf_map_ptr_poisoned(insn_aux)) {
8872 verbose(env, "tail_call abusing map_ptr\n");
8876 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
8877 if (!map->ops->map_set_for_each_callback_args ||
8878 !map->ops->map_for_each_callback) {
8879 verbose(env, "callback function not allowed for map\n");
8883 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
8887 callee->in_callback_fn = true;
8888 callee->callback_ret_range = tnum_range(0, 1);
8892 static int set_loop_callback_state(struct bpf_verifier_env *env,
8893 struct bpf_func_state *caller,
8894 struct bpf_func_state *callee,
8897 /* bpf_loop(u32 nr_loops, void *callback_fn, void *callback_ctx,
8899 * callback_fn(u32 index, void *callback_ctx);
8901 callee->regs[BPF_REG_1].type = SCALAR_VALUE;
8902 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
8905 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
8906 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
8907 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
8909 callee->in_callback_fn = true;
8910 callee->callback_ret_range = tnum_range(0, 1);
8914 static int set_timer_callback_state(struct bpf_verifier_env *env,
8915 struct bpf_func_state *caller,
8916 struct bpf_func_state *callee,
8919 struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
8921 /* bpf_timer_set_callback(struct bpf_timer *timer, void *callback_fn);
8922 * callback_fn(struct bpf_map *map, void *key, void *value);
8924 callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP;
8925 __mark_reg_known_zero(&callee->regs[BPF_REG_1]);
8926 callee->regs[BPF_REG_1].map_ptr = map_ptr;
8928 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
8929 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
8930 callee->regs[BPF_REG_2].map_ptr = map_ptr;
8932 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
8933 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
8934 callee->regs[BPF_REG_3].map_ptr = map_ptr;
8937 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
8938 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
8939 callee->in_async_callback_fn = true;
8940 callee->callback_ret_range = tnum_range(0, 1);
8944 static int set_find_vma_callback_state(struct bpf_verifier_env *env,
8945 struct bpf_func_state *caller,
8946 struct bpf_func_state *callee,
8949 /* bpf_find_vma(struct task_struct *task, u64 addr,
8950 * void *callback_fn, void *callback_ctx, u64 flags)
8951 * (callback_fn)(struct task_struct *task,
8952 * struct vm_area_struct *vma, void *callback_ctx);
8954 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
8956 callee->regs[BPF_REG_2].type = PTR_TO_BTF_ID;
8957 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
8958 callee->regs[BPF_REG_2].btf = btf_vmlinux;
8959 callee->regs[BPF_REG_2].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_VMA],
8961 /* pointer to stack or null */
8962 callee->regs[BPF_REG_3] = caller->regs[BPF_REG_4];
8965 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
8966 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
8967 callee->in_callback_fn = true;
8968 callee->callback_ret_range = tnum_range(0, 1);
8972 static int set_user_ringbuf_callback_state(struct bpf_verifier_env *env,
8973 struct bpf_func_state *caller,
8974 struct bpf_func_state *callee,
8977 /* bpf_user_ringbuf_drain(struct bpf_map *map, void *callback_fn, void
8978 * callback_ctx, u64 flags);
8979 * callback_fn(const struct bpf_dynptr_t* dynptr, void *callback_ctx);
8981 __mark_reg_not_init(env, &callee->regs[BPF_REG_0]);
8982 mark_dynptr_cb_reg(env, &callee->regs[BPF_REG_1], BPF_DYNPTR_TYPE_LOCAL);
8983 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
8986 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
8987 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
8988 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
8990 callee->in_callback_fn = true;
8991 callee->callback_ret_range = tnum_range(0, 1);
8995 static int set_rbtree_add_callback_state(struct bpf_verifier_env *env,
8996 struct bpf_func_state *caller,
8997 struct bpf_func_state *callee,
9000 /* void bpf_rbtree_add_impl(struct bpf_rb_root *root, struct bpf_rb_node *node,
9001 * bool (less)(struct bpf_rb_node *a, const struct bpf_rb_node *b));
9003 * 'struct bpf_rb_node *node' arg to bpf_rbtree_add_impl is the same PTR_TO_BTF_ID w/ offset
9004 * that 'less' callback args will be receiving. However, 'node' arg was release_reference'd
9005 * by this point, so look at 'root'
9007 struct btf_field *field;
9009 field = reg_find_field_offset(&caller->regs[BPF_REG_1], caller->regs[BPF_REG_1].off,
9011 if (!field || !field->graph_root.value_btf_id)
9014 mark_reg_graph_node(callee->regs, BPF_REG_1, &field->graph_root);
9015 ref_set_non_owning(env, &callee->regs[BPF_REG_1]);
9016 mark_reg_graph_node(callee->regs, BPF_REG_2, &field->graph_root);
9017 ref_set_non_owning(env, &callee->regs[BPF_REG_2]);
9019 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
9020 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9021 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9022 callee->in_callback_fn = true;
9023 callee->callback_ret_range = tnum_range(0, 1);
9027 static bool is_rbtree_lock_required_kfunc(u32 btf_id);
9029 /* Are we currently verifying the callback for a rbtree helper that must
9030 * be called with lock held? If so, no need to complain about unreleased
9033 static bool in_rbtree_lock_required_cb(struct bpf_verifier_env *env)
9035 struct bpf_verifier_state *state = env->cur_state;
9036 struct bpf_insn *insn = env->prog->insnsi;
9037 struct bpf_func_state *callee;
9040 if (!state->curframe)
9043 callee = state->frame[state->curframe];
9045 if (!callee->in_callback_fn)
9048 kfunc_btf_id = insn[callee->callsite].imm;
9049 return is_rbtree_lock_required_kfunc(kfunc_btf_id);
9052 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
9054 struct bpf_verifier_state *state = env->cur_state;
9055 struct bpf_func_state *caller, *callee;
9056 struct bpf_reg_state *r0;
9059 callee = state->frame[state->curframe];
9060 r0 = &callee->regs[BPF_REG_0];
9061 if (r0->type == PTR_TO_STACK) {
9062 /* technically it's ok to return caller's stack pointer
9063 * (or caller's caller's pointer) back to the caller,
9064 * since these pointers are valid. Only current stack
9065 * pointer will be invalid as soon as function exits,
9066 * but let's be conservative
9068 verbose(env, "cannot return stack pointer to the caller\n");
9072 caller = state->frame[state->curframe - 1];
9073 if (callee->in_callback_fn) {
9074 /* enforce R0 return value range [0, 1]. */
9075 struct tnum range = callee->callback_ret_range;
9077 if (r0->type != SCALAR_VALUE) {
9078 verbose(env, "R0 not a scalar value\n");
9081 if (!tnum_in(range, r0->var_off)) {
9082 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
9086 /* return to the caller whatever r0 had in the callee */
9087 caller->regs[BPF_REG_0] = *r0;
9090 /* callback_fn frame should have released its own additions to parent's
9091 * reference state at this point, or check_reference_leak would
9092 * complain, hence it must be the same as the caller. There is no need
9095 if (!callee->in_callback_fn) {
9096 /* Transfer references to the caller */
9097 err = copy_reference_state(caller, callee);
9102 *insn_idx = callee->callsite + 1;
9103 if (env->log.level & BPF_LOG_LEVEL) {
9104 verbose(env, "returning from callee:\n");
9105 print_verifier_state(env, callee, true);
9106 verbose(env, "to caller at %d:\n", *insn_idx);
9107 print_verifier_state(env, caller, true);
9109 /* clear everything in the callee */
9110 free_func_state(callee);
9111 state->frame[state->curframe--] = NULL;
9115 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
9117 struct bpf_call_arg_meta *meta)
9119 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
9121 if (ret_type != RET_INTEGER ||
9122 (func_id != BPF_FUNC_get_stack &&
9123 func_id != BPF_FUNC_get_task_stack &&
9124 func_id != BPF_FUNC_probe_read_str &&
9125 func_id != BPF_FUNC_probe_read_kernel_str &&
9126 func_id != BPF_FUNC_probe_read_user_str))
9129 ret_reg->smax_value = meta->msize_max_value;
9130 ret_reg->s32_max_value = meta->msize_max_value;
9131 ret_reg->smin_value = -MAX_ERRNO;
9132 ret_reg->s32_min_value = -MAX_ERRNO;
9133 reg_bounds_sync(ret_reg);
9137 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
9138 int func_id, int insn_idx)
9140 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
9141 struct bpf_map *map = meta->map_ptr;
9143 if (func_id != BPF_FUNC_tail_call &&
9144 func_id != BPF_FUNC_map_lookup_elem &&
9145 func_id != BPF_FUNC_map_update_elem &&
9146 func_id != BPF_FUNC_map_delete_elem &&
9147 func_id != BPF_FUNC_map_push_elem &&
9148 func_id != BPF_FUNC_map_pop_elem &&
9149 func_id != BPF_FUNC_map_peek_elem &&
9150 func_id != BPF_FUNC_for_each_map_elem &&
9151 func_id != BPF_FUNC_redirect_map &&
9152 func_id != BPF_FUNC_map_lookup_percpu_elem)
9156 verbose(env, "kernel subsystem misconfigured verifier\n");
9160 /* In case of read-only, some additional restrictions
9161 * need to be applied in order to prevent altering the
9162 * state of the map from program side.
9164 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
9165 (func_id == BPF_FUNC_map_delete_elem ||
9166 func_id == BPF_FUNC_map_update_elem ||
9167 func_id == BPF_FUNC_map_push_elem ||
9168 func_id == BPF_FUNC_map_pop_elem)) {
9169 verbose(env, "write into map forbidden\n");
9173 if (!BPF_MAP_PTR(aux->map_ptr_state))
9174 bpf_map_ptr_store(aux, meta->map_ptr,
9175 !meta->map_ptr->bypass_spec_v1);
9176 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
9177 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
9178 !meta->map_ptr->bypass_spec_v1);
9183 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
9184 int func_id, int insn_idx)
9186 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
9187 struct bpf_reg_state *regs = cur_regs(env), *reg;
9188 struct bpf_map *map = meta->map_ptr;
9192 if (func_id != BPF_FUNC_tail_call)
9194 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
9195 verbose(env, "kernel subsystem misconfigured verifier\n");
9199 reg = ®s[BPF_REG_3];
9200 val = reg->var_off.value;
9201 max = map->max_entries;
9203 if (!(register_is_const(reg) && val < max)) {
9204 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
9208 err = mark_chain_precision(env, BPF_REG_3);
9211 if (bpf_map_key_unseen(aux))
9212 bpf_map_key_store(aux, val);
9213 else if (!bpf_map_key_poisoned(aux) &&
9214 bpf_map_key_immediate(aux) != val)
9215 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
9219 static int check_reference_leak(struct bpf_verifier_env *env)
9221 struct bpf_func_state *state = cur_func(env);
9222 bool refs_lingering = false;
9225 if (state->frameno && !state->in_callback_fn)
9228 for (i = 0; i < state->acquired_refs; i++) {
9229 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno)
9231 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
9232 state->refs[i].id, state->refs[i].insn_idx);
9233 refs_lingering = true;
9235 return refs_lingering ? -EINVAL : 0;
9238 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
9239 struct bpf_reg_state *regs)
9241 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
9242 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
9243 struct bpf_map *fmt_map = fmt_reg->map_ptr;
9244 struct bpf_bprintf_data data = {};
9245 int err, fmt_map_off, num_args;
9249 /* data must be an array of u64 */
9250 if (data_len_reg->var_off.value % 8)
9252 num_args = data_len_reg->var_off.value / 8;
9254 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
9255 * and map_direct_value_addr is set.
9257 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
9258 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
9261 verbose(env, "verifier bug\n");
9264 fmt = (char *)(long)fmt_addr + fmt_map_off;
9266 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
9267 * can focus on validating the format specifiers.
9269 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, num_args, &data);
9271 verbose(env, "Invalid format string\n");
9276 static int check_get_func_ip(struct bpf_verifier_env *env)
9278 enum bpf_prog_type type = resolve_prog_type(env->prog);
9279 int func_id = BPF_FUNC_get_func_ip;
9281 if (type == BPF_PROG_TYPE_TRACING) {
9282 if (!bpf_prog_has_trampoline(env->prog)) {
9283 verbose(env, "func %s#%d supported only for fentry/fexit/fmod_ret programs\n",
9284 func_id_name(func_id), func_id);
9288 } else if (type == BPF_PROG_TYPE_KPROBE) {
9292 verbose(env, "func %s#%d not supported for program type %d\n",
9293 func_id_name(func_id), func_id, type);
9297 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
9299 return &env->insn_aux_data[env->insn_idx];
9302 static bool loop_flag_is_zero(struct bpf_verifier_env *env)
9304 struct bpf_reg_state *regs = cur_regs(env);
9305 struct bpf_reg_state *reg = ®s[BPF_REG_4];
9306 bool reg_is_null = register_is_null(reg);
9309 mark_chain_precision(env, BPF_REG_4);
9314 static void update_loop_inline_state(struct bpf_verifier_env *env, u32 subprogno)
9316 struct bpf_loop_inline_state *state = &cur_aux(env)->loop_inline_state;
9318 if (!state->initialized) {
9319 state->initialized = 1;
9320 state->fit_for_inline = loop_flag_is_zero(env);
9321 state->callback_subprogno = subprogno;
9325 if (!state->fit_for_inline)
9328 state->fit_for_inline = (loop_flag_is_zero(env) &&
9329 state->callback_subprogno == subprogno);
9332 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
9335 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
9336 const struct bpf_func_proto *fn = NULL;
9337 enum bpf_return_type ret_type;
9338 enum bpf_type_flag ret_flag;
9339 struct bpf_reg_state *regs;
9340 struct bpf_call_arg_meta meta;
9341 int insn_idx = *insn_idx_p;
9343 int i, err, func_id;
9345 /* find function prototype */
9346 func_id = insn->imm;
9347 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
9348 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
9353 if (env->ops->get_func_proto)
9354 fn = env->ops->get_func_proto(func_id, env->prog);
9356 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
9361 /* eBPF programs must be GPL compatible to use GPL-ed functions */
9362 if (!env->prog->gpl_compatible && fn->gpl_only) {
9363 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
9367 if (fn->allowed && !fn->allowed(env->prog)) {
9368 verbose(env, "helper call is not allowed in probe\n");
9372 if (!env->prog->aux->sleepable && fn->might_sleep) {
9373 verbose(env, "helper call might sleep in a non-sleepable prog\n");
9377 /* With LD_ABS/IND some JITs save/restore skb from r1. */
9378 changes_data = bpf_helper_changes_pkt_data(fn->func);
9379 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
9380 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
9381 func_id_name(func_id), func_id);
9385 memset(&meta, 0, sizeof(meta));
9386 meta.pkt_access = fn->pkt_access;
9388 err = check_func_proto(fn, func_id);
9390 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
9391 func_id_name(func_id), func_id);
9395 if (env->cur_state->active_rcu_lock) {
9396 if (fn->might_sleep) {
9397 verbose(env, "sleepable helper %s#%d in rcu_read_lock region\n",
9398 func_id_name(func_id), func_id);
9402 if (env->prog->aux->sleepable && is_storage_get_function(func_id))
9403 env->insn_aux_data[insn_idx].storage_get_func_atomic = true;
9406 meta.func_id = func_id;
9408 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
9409 err = check_func_arg(env, i, &meta, fn, insn_idx);
9414 err = record_func_map(env, &meta, func_id, insn_idx);
9418 err = record_func_key(env, &meta, func_id, insn_idx);
9422 /* Mark slots with STACK_MISC in case of raw mode, stack offset
9423 * is inferred from register state.
9425 for (i = 0; i < meta.access_size; i++) {
9426 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
9427 BPF_WRITE, -1, false);
9432 regs = cur_regs(env);
9434 if (meta.release_regno) {
9436 /* This can only be set for PTR_TO_STACK, as CONST_PTR_TO_DYNPTR cannot
9437 * be released by any dynptr helper. Hence, unmark_stack_slots_dynptr
9438 * is safe to do directly.
9440 if (arg_type_is_dynptr(fn->arg_type[meta.release_regno - BPF_REG_1])) {
9441 if (regs[meta.release_regno].type == CONST_PTR_TO_DYNPTR) {
9442 verbose(env, "verifier internal error: CONST_PTR_TO_DYNPTR cannot be released\n");
9445 err = unmark_stack_slots_dynptr(env, ®s[meta.release_regno]);
9446 } else if (meta.ref_obj_id) {
9447 err = release_reference(env, meta.ref_obj_id);
9448 } else if (register_is_null(®s[meta.release_regno])) {
9449 /* meta.ref_obj_id can only be 0 if register that is meant to be
9450 * released is NULL, which must be > R0.
9455 verbose(env, "func %s#%d reference has not been acquired before\n",
9456 func_id_name(func_id), func_id);
9462 case BPF_FUNC_tail_call:
9463 err = check_reference_leak(env);
9465 verbose(env, "tail_call would lead to reference leak\n");
9469 case BPF_FUNC_get_local_storage:
9470 /* check that flags argument in get_local_storage(map, flags) is 0,
9471 * this is required because get_local_storage() can't return an error.
9473 if (!register_is_null(®s[BPF_REG_2])) {
9474 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
9478 case BPF_FUNC_for_each_map_elem:
9479 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9480 set_map_elem_callback_state);
9482 case BPF_FUNC_timer_set_callback:
9483 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9484 set_timer_callback_state);
9486 case BPF_FUNC_find_vma:
9487 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9488 set_find_vma_callback_state);
9490 case BPF_FUNC_snprintf:
9491 err = check_bpf_snprintf_call(env, regs);
9494 update_loop_inline_state(env, meta.subprogno);
9495 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9496 set_loop_callback_state);
9498 case BPF_FUNC_dynptr_from_mem:
9499 if (regs[BPF_REG_1].type != PTR_TO_MAP_VALUE) {
9500 verbose(env, "Unsupported reg type %s for bpf_dynptr_from_mem data\n",
9501 reg_type_str(env, regs[BPF_REG_1].type));
9505 case BPF_FUNC_set_retval:
9506 if (prog_type == BPF_PROG_TYPE_LSM &&
9507 env->prog->expected_attach_type == BPF_LSM_CGROUP) {
9508 if (!env->prog->aux->attach_func_proto->type) {
9509 /* Make sure programs that attach to void
9510 * hooks don't try to modify return value.
9512 verbose(env, "BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
9517 case BPF_FUNC_dynptr_data:
9519 struct bpf_reg_state *reg;
9522 reg = get_dynptr_arg_reg(env, fn, regs);
9527 if (meta.dynptr_id) {
9528 verbose(env, "verifier internal error: meta.dynptr_id already set\n");
9531 if (meta.ref_obj_id) {
9532 verbose(env, "verifier internal error: meta.ref_obj_id already set\n");
9536 id = dynptr_id(env, reg);
9538 verbose(env, "verifier internal error: failed to obtain dynptr id\n");
9542 ref_obj_id = dynptr_ref_obj_id(env, reg);
9543 if (ref_obj_id < 0) {
9544 verbose(env, "verifier internal error: failed to obtain dynptr ref_obj_id\n");
9548 meta.dynptr_id = id;
9549 meta.ref_obj_id = ref_obj_id;
9553 case BPF_FUNC_dynptr_write:
9555 enum bpf_dynptr_type dynptr_type;
9556 struct bpf_reg_state *reg;
9558 reg = get_dynptr_arg_reg(env, fn, regs);
9562 dynptr_type = dynptr_get_type(env, reg);
9563 if (dynptr_type == BPF_DYNPTR_TYPE_INVALID)
9566 if (dynptr_type == BPF_DYNPTR_TYPE_SKB)
9567 /* this will trigger clear_all_pkt_pointers(), which will
9568 * invalidate all dynptr slices associated with the skb
9570 changes_data = true;
9574 case BPF_FUNC_user_ringbuf_drain:
9575 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9576 set_user_ringbuf_callback_state);
9583 /* reset caller saved regs */
9584 for (i = 0; i < CALLER_SAVED_REGS; i++) {
9585 mark_reg_not_init(env, regs, caller_saved[i]);
9586 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
9589 /* helper call returns 64-bit value. */
9590 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
9592 /* update return register (already marked as written above) */
9593 ret_type = fn->ret_type;
9594 ret_flag = type_flag(ret_type);
9596 switch (base_type(ret_type)) {
9598 /* sets type to SCALAR_VALUE */
9599 mark_reg_unknown(env, regs, BPF_REG_0);
9602 regs[BPF_REG_0].type = NOT_INIT;
9604 case RET_PTR_TO_MAP_VALUE:
9605 /* There is no offset yet applied, variable or fixed */
9606 mark_reg_known_zero(env, regs, BPF_REG_0);
9607 /* remember map_ptr, so that check_map_access()
9608 * can check 'value_size' boundary of memory access
9609 * to map element returned from bpf_map_lookup_elem()
9611 if (meta.map_ptr == NULL) {
9613 "kernel subsystem misconfigured verifier\n");
9616 regs[BPF_REG_0].map_ptr = meta.map_ptr;
9617 regs[BPF_REG_0].map_uid = meta.map_uid;
9618 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE | ret_flag;
9619 if (!type_may_be_null(ret_type) &&
9620 btf_record_has_field(meta.map_ptr->record, BPF_SPIN_LOCK)) {
9621 regs[BPF_REG_0].id = ++env->id_gen;
9624 case RET_PTR_TO_SOCKET:
9625 mark_reg_known_zero(env, regs, BPF_REG_0);
9626 regs[BPF_REG_0].type = PTR_TO_SOCKET | ret_flag;
9628 case RET_PTR_TO_SOCK_COMMON:
9629 mark_reg_known_zero(env, regs, BPF_REG_0);
9630 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON | ret_flag;
9632 case RET_PTR_TO_TCP_SOCK:
9633 mark_reg_known_zero(env, regs, BPF_REG_0);
9634 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK | ret_flag;
9636 case RET_PTR_TO_MEM:
9637 mark_reg_known_zero(env, regs, BPF_REG_0);
9638 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
9639 regs[BPF_REG_0].mem_size = meta.mem_size;
9641 case RET_PTR_TO_MEM_OR_BTF_ID:
9643 const struct btf_type *t;
9645 mark_reg_known_zero(env, regs, BPF_REG_0);
9646 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
9647 if (!btf_type_is_struct(t)) {
9649 const struct btf_type *ret;
9652 /* resolve the type size of ksym. */
9653 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
9655 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
9656 verbose(env, "unable to resolve the size of type '%s': %ld\n",
9657 tname, PTR_ERR(ret));
9660 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
9661 regs[BPF_REG_0].mem_size = tsize;
9663 /* MEM_RDONLY may be carried from ret_flag, but it
9664 * doesn't apply on PTR_TO_BTF_ID. Fold it, otherwise
9665 * it will confuse the check of PTR_TO_BTF_ID in
9666 * check_mem_access().
9668 ret_flag &= ~MEM_RDONLY;
9670 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
9671 regs[BPF_REG_0].btf = meta.ret_btf;
9672 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
9676 case RET_PTR_TO_BTF_ID:
9678 struct btf *ret_btf;
9681 mark_reg_known_zero(env, regs, BPF_REG_0);
9682 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
9683 if (func_id == BPF_FUNC_kptr_xchg) {
9684 ret_btf = meta.kptr_field->kptr.btf;
9685 ret_btf_id = meta.kptr_field->kptr.btf_id;
9686 if (!btf_is_kernel(ret_btf))
9687 regs[BPF_REG_0].type |= MEM_ALLOC;
9689 if (fn->ret_btf_id == BPF_PTR_POISON) {
9690 verbose(env, "verifier internal error:");
9691 verbose(env, "func %s has non-overwritten BPF_PTR_POISON return type\n",
9692 func_id_name(func_id));
9695 ret_btf = btf_vmlinux;
9696 ret_btf_id = *fn->ret_btf_id;
9698 if (ret_btf_id == 0) {
9699 verbose(env, "invalid return type %u of func %s#%d\n",
9700 base_type(ret_type), func_id_name(func_id),
9704 regs[BPF_REG_0].btf = ret_btf;
9705 regs[BPF_REG_0].btf_id = ret_btf_id;
9709 verbose(env, "unknown return type %u of func %s#%d\n",
9710 base_type(ret_type), func_id_name(func_id), func_id);
9714 if (type_may_be_null(regs[BPF_REG_0].type))
9715 regs[BPF_REG_0].id = ++env->id_gen;
9717 if (helper_multiple_ref_obj_use(func_id, meta.map_ptr)) {
9718 verbose(env, "verifier internal error: func %s#%d sets ref_obj_id more than once\n",
9719 func_id_name(func_id), func_id);
9723 if (is_dynptr_ref_function(func_id))
9724 regs[BPF_REG_0].dynptr_id = meta.dynptr_id;
9726 if (is_ptr_cast_function(func_id) || is_dynptr_ref_function(func_id)) {
9727 /* For release_reference() */
9728 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
9729 } else if (is_acquire_function(func_id, meta.map_ptr)) {
9730 int id = acquire_reference_state(env, insn_idx);
9734 /* For mark_ptr_or_null_reg() */
9735 regs[BPF_REG_0].id = id;
9736 /* For release_reference() */
9737 regs[BPF_REG_0].ref_obj_id = id;
9740 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
9742 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
9746 if ((func_id == BPF_FUNC_get_stack ||
9747 func_id == BPF_FUNC_get_task_stack) &&
9748 !env->prog->has_callchain_buf) {
9749 const char *err_str;
9751 #ifdef CONFIG_PERF_EVENTS
9752 err = get_callchain_buffers(sysctl_perf_event_max_stack);
9753 err_str = "cannot get callchain buffer for func %s#%d\n";
9756 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
9759 verbose(env, err_str, func_id_name(func_id), func_id);
9763 env->prog->has_callchain_buf = true;
9766 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
9767 env->prog->call_get_stack = true;
9769 if (func_id == BPF_FUNC_get_func_ip) {
9770 if (check_get_func_ip(env))
9772 env->prog->call_get_func_ip = true;
9776 clear_all_pkt_pointers(env);
9780 /* mark_btf_func_reg_size() is used when the reg size is determined by
9781 * the BTF func_proto's return value size and argument.
9783 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
9786 struct bpf_reg_state *reg = &cur_regs(env)[regno];
9788 if (regno == BPF_REG_0) {
9789 /* Function return value */
9790 reg->live |= REG_LIVE_WRITTEN;
9791 reg->subreg_def = reg_size == sizeof(u64) ?
9792 DEF_NOT_SUBREG : env->insn_idx + 1;
9794 /* Function argument */
9795 if (reg_size == sizeof(u64)) {
9796 mark_insn_zext(env, reg);
9797 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
9799 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
9804 static bool is_kfunc_acquire(struct bpf_kfunc_call_arg_meta *meta)
9806 return meta->kfunc_flags & KF_ACQUIRE;
9809 static bool is_kfunc_release(struct bpf_kfunc_call_arg_meta *meta)
9811 return meta->kfunc_flags & KF_RELEASE;
9814 static bool is_kfunc_trusted_args(struct bpf_kfunc_call_arg_meta *meta)
9816 return (meta->kfunc_flags & KF_TRUSTED_ARGS) || is_kfunc_release(meta);
9819 static bool is_kfunc_sleepable(struct bpf_kfunc_call_arg_meta *meta)
9821 return meta->kfunc_flags & KF_SLEEPABLE;
9824 static bool is_kfunc_destructive(struct bpf_kfunc_call_arg_meta *meta)
9826 return meta->kfunc_flags & KF_DESTRUCTIVE;
9829 static bool is_kfunc_rcu(struct bpf_kfunc_call_arg_meta *meta)
9831 return meta->kfunc_flags & KF_RCU;
9834 static bool __kfunc_param_match_suffix(const struct btf *btf,
9835 const struct btf_param *arg,
9838 int suffix_len = strlen(suffix), len;
9839 const char *param_name;
9841 /* In the future, this can be ported to use BTF tagging */
9842 param_name = btf_name_by_offset(btf, arg->name_off);
9843 if (str_is_empty(param_name))
9845 len = strlen(param_name);
9846 if (len < suffix_len)
9848 param_name += len - suffix_len;
9849 return !strncmp(param_name, suffix, suffix_len);
9852 static bool is_kfunc_arg_mem_size(const struct btf *btf,
9853 const struct btf_param *arg,
9854 const struct bpf_reg_state *reg)
9856 const struct btf_type *t;
9858 t = btf_type_skip_modifiers(btf, arg->type, NULL);
9859 if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
9862 return __kfunc_param_match_suffix(btf, arg, "__sz");
9865 static bool is_kfunc_arg_const_mem_size(const struct btf *btf,
9866 const struct btf_param *arg,
9867 const struct bpf_reg_state *reg)
9869 const struct btf_type *t;
9871 t = btf_type_skip_modifiers(btf, arg->type, NULL);
9872 if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
9875 return __kfunc_param_match_suffix(btf, arg, "__szk");
9878 static bool is_kfunc_arg_optional(const struct btf *btf, const struct btf_param *arg)
9880 return __kfunc_param_match_suffix(btf, arg, "__opt");
9883 static bool is_kfunc_arg_constant(const struct btf *btf, const struct btf_param *arg)
9885 return __kfunc_param_match_suffix(btf, arg, "__k");
9888 static bool is_kfunc_arg_ignore(const struct btf *btf, const struct btf_param *arg)
9890 return __kfunc_param_match_suffix(btf, arg, "__ign");
9893 static bool is_kfunc_arg_alloc_obj(const struct btf *btf, const struct btf_param *arg)
9895 return __kfunc_param_match_suffix(btf, arg, "__alloc");
9898 static bool is_kfunc_arg_uninit(const struct btf *btf, const struct btf_param *arg)
9900 return __kfunc_param_match_suffix(btf, arg, "__uninit");
9903 static bool is_kfunc_arg_refcounted_kptr(const struct btf *btf, const struct btf_param *arg)
9905 return __kfunc_param_match_suffix(btf, arg, "__refcounted_kptr");
9908 static bool is_kfunc_arg_scalar_with_name(const struct btf *btf,
9909 const struct btf_param *arg,
9912 int len, target_len = strlen(name);
9913 const char *param_name;
9915 param_name = btf_name_by_offset(btf, arg->name_off);
9916 if (str_is_empty(param_name))
9918 len = strlen(param_name);
9919 if (len != target_len)
9921 if (strcmp(param_name, name))
9929 KF_ARG_LIST_HEAD_ID,
9930 KF_ARG_LIST_NODE_ID,
9935 BTF_ID_LIST(kf_arg_btf_ids)
9936 BTF_ID(struct, bpf_dynptr_kern)
9937 BTF_ID(struct, bpf_list_head)
9938 BTF_ID(struct, bpf_list_node)
9939 BTF_ID(struct, bpf_rb_root)
9940 BTF_ID(struct, bpf_rb_node)
9942 static bool __is_kfunc_ptr_arg_type(const struct btf *btf,
9943 const struct btf_param *arg, int type)
9945 const struct btf_type *t;
9948 t = btf_type_skip_modifiers(btf, arg->type, NULL);
9951 if (!btf_type_is_ptr(t))
9953 t = btf_type_skip_modifiers(btf, t->type, &res_id);
9956 return btf_types_are_same(btf, res_id, btf_vmlinux, kf_arg_btf_ids[type]);
9959 static bool is_kfunc_arg_dynptr(const struct btf *btf, const struct btf_param *arg)
9961 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_DYNPTR_ID);
9964 static bool is_kfunc_arg_list_head(const struct btf *btf, const struct btf_param *arg)
9966 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_HEAD_ID);
9969 static bool is_kfunc_arg_list_node(const struct btf *btf, const struct btf_param *arg)
9971 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_NODE_ID);
9974 static bool is_kfunc_arg_rbtree_root(const struct btf *btf, const struct btf_param *arg)
9976 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RB_ROOT_ID);
9979 static bool is_kfunc_arg_rbtree_node(const struct btf *btf, const struct btf_param *arg)
9981 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RB_NODE_ID);
9984 static bool is_kfunc_arg_callback(struct bpf_verifier_env *env, const struct btf *btf,
9985 const struct btf_param *arg)
9987 const struct btf_type *t;
9989 t = btf_type_resolve_func_ptr(btf, arg->type, NULL);
9996 /* Returns true if struct is composed of scalars, 4 levels of nesting allowed */
9997 static bool __btf_type_is_scalar_struct(struct bpf_verifier_env *env,
9998 const struct btf *btf,
9999 const struct btf_type *t, int rec)
10001 const struct btf_type *member_type;
10002 const struct btf_member *member;
10005 if (!btf_type_is_struct(t))
10008 for_each_member(i, t, member) {
10009 const struct btf_array *array;
10011 member_type = btf_type_skip_modifiers(btf, member->type, NULL);
10012 if (btf_type_is_struct(member_type)) {
10014 verbose(env, "max struct nesting depth exceeded\n");
10017 if (!__btf_type_is_scalar_struct(env, btf, member_type, rec + 1))
10021 if (btf_type_is_array(member_type)) {
10022 array = btf_array(member_type);
10023 if (!array->nelems)
10025 member_type = btf_type_skip_modifiers(btf, array->type, NULL);
10026 if (!btf_type_is_scalar(member_type))
10030 if (!btf_type_is_scalar(member_type))
10037 static u32 *reg2btf_ids[__BPF_REG_TYPE_MAX] = {
10039 [PTR_TO_SOCKET] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK],
10040 [PTR_TO_SOCK_COMMON] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
10041 [PTR_TO_TCP_SOCK] = &btf_sock_ids[BTF_SOCK_TYPE_TCP],
10045 enum kfunc_ptr_arg_type {
10047 KF_ARG_PTR_TO_ALLOC_BTF_ID, /* Allocated object */
10048 KF_ARG_PTR_TO_REFCOUNTED_KPTR, /* Refcounted local kptr */
10049 KF_ARG_PTR_TO_DYNPTR,
10050 KF_ARG_PTR_TO_ITER,
10051 KF_ARG_PTR_TO_LIST_HEAD,
10052 KF_ARG_PTR_TO_LIST_NODE,
10053 KF_ARG_PTR_TO_BTF_ID, /* Also covers reg2btf_ids conversions */
10055 KF_ARG_PTR_TO_MEM_SIZE, /* Size derived from next argument, skip it */
10056 KF_ARG_PTR_TO_CALLBACK,
10057 KF_ARG_PTR_TO_RB_ROOT,
10058 KF_ARG_PTR_TO_RB_NODE,
10061 enum special_kfunc_type {
10062 KF_bpf_obj_new_impl,
10063 KF_bpf_obj_drop_impl,
10064 KF_bpf_refcount_acquire_impl,
10065 KF_bpf_list_push_front_impl,
10066 KF_bpf_list_push_back_impl,
10067 KF_bpf_list_pop_front,
10068 KF_bpf_list_pop_back,
10069 KF_bpf_cast_to_kern_ctx,
10070 KF_bpf_rdonly_cast,
10071 KF_bpf_rcu_read_lock,
10072 KF_bpf_rcu_read_unlock,
10073 KF_bpf_rbtree_remove,
10074 KF_bpf_rbtree_add_impl,
10075 KF_bpf_rbtree_first,
10076 KF_bpf_dynptr_from_skb,
10077 KF_bpf_dynptr_from_xdp,
10078 KF_bpf_dynptr_slice,
10079 KF_bpf_dynptr_slice_rdwr,
10080 KF_bpf_dynptr_clone,
10083 BTF_SET_START(special_kfunc_set)
10084 BTF_ID(func, bpf_obj_new_impl)
10085 BTF_ID(func, bpf_obj_drop_impl)
10086 BTF_ID(func, bpf_refcount_acquire_impl)
10087 BTF_ID(func, bpf_list_push_front_impl)
10088 BTF_ID(func, bpf_list_push_back_impl)
10089 BTF_ID(func, bpf_list_pop_front)
10090 BTF_ID(func, bpf_list_pop_back)
10091 BTF_ID(func, bpf_cast_to_kern_ctx)
10092 BTF_ID(func, bpf_rdonly_cast)
10093 BTF_ID(func, bpf_rbtree_remove)
10094 BTF_ID(func, bpf_rbtree_add_impl)
10095 BTF_ID(func, bpf_rbtree_first)
10096 BTF_ID(func, bpf_dynptr_from_skb)
10097 BTF_ID(func, bpf_dynptr_from_xdp)
10098 BTF_ID(func, bpf_dynptr_slice)
10099 BTF_ID(func, bpf_dynptr_slice_rdwr)
10100 BTF_ID(func, bpf_dynptr_clone)
10101 BTF_SET_END(special_kfunc_set)
10103 BTF_ID_LIST(special_kfunc_list)
10104 BTF_ID(func, bpf_obj_new_impl)
10105 BTF_ID(func, bpf_obj_drop_impl)
10106 BTF_ID(func, bpf_refcount_acquire_impl)
10107 BTF_ID(func, bpf_list_push_front_impl)
10108 BTF_ID(func, bpf_list_push_back_impl)
10109 BTF_ID(func, bpf_list_pop_front)
10110 BTF_ID(func, bpf_list_pop_back)
10111 BTF_ID(func, bpf_cast_to_kern_ctx)
10112 BTF_ID(func, bpf_rdonly_cast)
10113 BTF_ID(func, bpf_rcu_read_lock)
10114 BTF_ID(func, bpf_rcu_read_unlock)
10115 BTF_ID(func, bpf_rbtree_remove)
10116 BTF_ID(func, bpf_rbtree_add_impl)
10117 BTF_ID(func, bpf_rbtree_first)
10118 BTF_ID(func, bpf_dynptr_from_skb)
10119 BTF_ID(func, bpf_dynptr_from_xdp)
10120 BTF_ID(func, bpf_dynptr_slice)
10121 BTF_ID(func, bpf_dynptr_slice_rdwr)
10122 BTF_ID(func, bpf_dynptr_clone)
10124 static bool is_kfunc_ret_null(struct bpf_kfunc_call_arg_meta *meta)
10126 if (meta->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl] &&
10127 meta->arg_owning_ref) {
10131 return meta->kfunc_flags & KF_RET_NULL;
10134 static bool is_kfunc_bpf_rcu_read_lock(struct bpf_kfunc_call_arg_meta *meta)
10136 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_lock];
10139 static bool is_kfunc_bpf_rcu_read_unlock(struct bpf_kfunc_call_arg_meta *meta)
10141 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_unlock];
10144 static enum kfunc_ptr_arg_type
10145 get_kfunc_ptr_arg_type(struct bpf_verifier_env *env,
10146 struct bpf_kfunc_call_arg_meta *meta,
10147 const struct btf_type *t, const struct btf_type *ref_t,
10148 const char *ref_tname, const struct btf_param *args,
10149 int argno, int nargs)
10151 u32 regno = argno + 1;
10152 struct bpf_reg_state *regs = cur_regs(env);
10153 struct bpf_reg_state *reg = ®s[regno];
10154 bool arg_mem_size = false;
10156 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx])
10157 return KF_ARG_PTR_TO_CTX;
10159 /* In this function, we verify the kfunc's BTF as per the argument type,
10160 * leaving the rest of the verification with respect to the register
10161 * type to our caller. When a set of conditions hold in the BTF type of
10162 * arguments, we resolve it to a known kfunc_ptr_arg_type.
10164 if (btf_get_prog_ctx_type(&env->log, meta->btf, t, resolve_prog_type(env->prog), argno))
10165 return KF_ARG_PTR_TO_CTX;
10167 if (is_kfunc_arg_alloc_obj(meta->btf, &args[argno]))
10168 return KF_ARG_PTR_TO_ALLOC_BTF_ID;
10170 if (is_kfunc_arg_refcounted_kptr(meta->btf, &args[argno]))
10171 return KF_ARG_PTR_TO_REFCOUNTED_KPTR;
10173 if (is_kfunc_arg_dynptr(meta->btf, &args[argno]))
10174 return KF_ARG_PTR_TO_DYNPTR;
10176 if (is_kfunc_arg_iter(meta, argno))
10177 return KF_ARG_PTR_TO_ITER;
10179 if (is_kfunc_arg_list_head(meta->btf, &args[argno]))
10180 return KF_ARG_PTR_TO_LIST_HEAD;
10182 if (is_kfunc_arg_list_node(meta->btf, &args[argno]))
10183 return KF_ARG_PTR_TO_LIST_NODE;
10185 if (is_kfunc_arg_rbtree_root(meta->btf, &args[argno]))
10186 return KF_ARG_PTR_TO_RB_ROOT;
10188 if (is_kfunc_arg_rbtree_node(meta->btf, &args[argno]))
10189 return KF_ARG_PTR_TO_RB_NODE;
10191 if ((base_type(reg->type) == PTR_TO_BTF_ID || reg2btf_ids[base_type(reg->type)])) {
10192 if (!btf_type_is_struct(ref_t)) {
10193 verbose(env, "kernel function %s args#%d pointer type %s %s is not supported\n",
10194 meta->func_name, argno, btf_type_str(ref_t), ref_tname);
10197 return KF_ARG_PTR_TO_BTF_ID;
10200 if (is_kfunc_arg_callback(env, meta->btf, &args[argno]))
10201 return KF_ARG_PTR_TO_CALLBACK;
10204 if (argno + 1 < nargs &&
10205 (is_kfunc_arg_mem_size(meta->btf, &args[argno + 1], ®s[regno + 1]) ||
10206 is_kfunc_arg_const_mem_size(meta->btf, &args[argno + 1], ®s[regno + 1])))
10207 arg_mem_size = true;
10209 /* This is the catch all argument type of register types supported by
10210 * check_helper_mem_access. However, we only allow when argument type is
10211 * pointer to scalar, or struct composed (recursively) of scalars. When
10212 * arg_mem_size is true, the pointer can be void *.
10214 if (!btf_type_is_scalar(ref_t) && !__btf_type_is_scalar_struct(env, meta->btf, ref_t, 0) &&
10215 (arg_mem_size ? !btf_type_is_void(ref_t) : 1)) {
10216 verbose(env, "arg#%d pointer type %s %s must point to %sscalar, or struct with scalar\n",
10217 argno, btf_type_str(ref_t), ref_tname, arg_mem_size ? "void, " : "");
10220 return arg_mem_size ? KF_ARG_PTR_TO_MEM_SIZE : KF_ARG_PTR_TO_MEM;
10223 static int process_kf_arg_ptr_to_btf_id(struct bpf_verifier_env *env,
10224 struct bpf_reg_state *reg,
10225 const struct btf_type *ref_t,
10226 const char *ref_tname, u32 ref_id,
10227 struct bpf_kfunc_call_arg_meta *meta,
10230 const struct btf_type *reg_ref_t;
10231 bool strict_type_match = false;
10232 const struct btf *reg_btf;
10233 const char *reg_ref_tname;
10236 if (base_type(reg->type) == PTR_TO_BTF_ID) {
10237 reg_btf = reg->btf;
10238 reg_ref_id = reg->btf_id;
10240 reg_btf = btf_vmlinux;
10241 reg_ref_id = *reg2btf_ids[base_type(reg->type)];
10244 /* Enforce strict type matching for calls to kfuncs that are acquiring
10245 * or releasing a reference, or are no-cast aliases. We do _not_
10246 * enforce strict matching for plain KF_TRUSTED_ARGS kfuncs by default,
10247 * as we want to enable BPF programs to pass types that are bitwise
10248 * equivalent without forcing them to explicitly cast with something
10249 * like bpf_cast_to_kern_ctx().
10251 * For example, say we had a type like the following:
10253 * struct bpf_cpumask {
10254 * cpumask_t cpumask;
10255 * refcount_t usage;
10258 * Note that as specified in <linux/cpumask.h>, cpumask_t is typedef'ed
10259 * to a struct cpumask, so it would be safe to pass a struct
10260 * bpf_cpumask * to a kfunc expecting a struct cpumask *.
10262 * The philosophy here is similar to how we allow scalars of different
10263 * types to be passed to kfuncs as long as the size is the same. The
10264 * only difference here is that we're simply allowing
10265 * btf_struct_ids_match() to walk the struct at the 0th offset, and
10268 if (is_kfunc_acquire(meta) ||
10269 (is_kfunc_release(meta) && reg->ref_obj_id) ||
10270 btf_type_ids_nocast_alias(&env->log, reg_btf, reg_ref_id, meta->btf, ref_id))
10271 strict_type_match = true;
10273 WARN_ON_ONCE(is_kfunc_trusted_args(meta) && reg->off);
10275 reg_ref_t = btf_type_skip_modifiers(reg_btf, reg_ref_id, ®_ref_id);
10276 reg_ref_tname = btf_name_by_offset(reg_btf, reg_ref_t->name_off);
10277 if (!btf_struct_ids_match(&env->log, reg_btf, reg_ref_id, reg->off, meta->btf, ref_id, strict_type_match)) {
10278 verbose(env, "kernel function %s args#%d expected pointer to %s %s but R%d has a pointer to %s %s\n",
10279 meta->func_name, argno, btf_type_str(ref_t), ref_tname, argno + 1,
10280 btf_type_str(reg_ref_t), reg_ref_tname);
10286 static int ref_set_non_owning(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
10288 struct bpf_verifier_state *state = env->cur_state;
10290 if (!state->active_lock.ptr) {
10291 verbose(env, "verifier internal error: ref_set_non_owning w/o active lock\n");
10295 if (type_flag(reg->type) & NON_OWN_REF) {
10296 verbose(env, "verifier internal error: NON_OWN_REF already set\n");
10300 reg->type |= NON_OWN_REF;
10304 static int ref_convert_owning_non_owning(struct bpf_verifier_env *env, u32 ref_obj_id)
10306 struct bpf_func_state *state, *unused;
10307 struct bpf_reg_state *reg;
10310 state = cur_func(env);
10313 verbose(env, "verifier internal error: ref_obj_id is zero for "
10314 "owning -> non-owning conversion\n");
10318 for (i = 0; i < state->acquired_refs; i++) {
10319 if (state->refs[i].id != ref_obj_id)
10322 /* Clear ref_obj_id here so release_reference doesn't clobber
10325 bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
10326 if (reg->ref_obj_id == ref_obj_id) {
10327 reg->ref_obj_id = 0;
10328 ref_set_non_owning(env, reg);
10334 verbose(env, "verifier internal error: ref state missing for ref_obj_id\n");
10338 /* Implementation details:
10340 * Each register points to some region of memory, which we define as an
10341 * allocation. Each allocation may embed a bpf_spin_lock which protects any
10342 * special BPF objects (bpf_list_head, bpf_rb_root, etc.) part of the same
10343 * allocation. The lock and the data it protects are colocated in the same
10346 * Hence, everytime a register holds a pointer value pointing to such
10347 * allocation, the verifier preserves a unique reg->id for it.
10349 * The verifier remembers the lock 'ptr' and the lock 'id' whenever
10350 * bpf_spin_lock is called.
10352 * To enable this, lock state in the verifier captures two values:
10353 * active_lock.ptr = Register's type specific pointer
10354 * active_lock.id = A unique ID for each register pointer value
10356 * Currently, PTR_TO_MAP_VALUE and PTR_TO_BTF_ID | MEM_ALLOC are the two
10357 * supported register types.
10359 * The active_lock.ptr in case of map values is the reg->map_ptr, and in case of
10360 * allocated objects is the reg->btf pointer.
10362 * The active_lock.id is non-unique for maps supporting direct_value_addr, as we
10363 * can establish the provenance of the map value statically for each distinct
10364 * lookup into such maps. They always contain a single map value hence unique
10365 * IDs for each pseudo load pessimizes the algorithm and rejects valid programs.
10367 * So, in case of global variables, they use array maps with max_entries = 1,
10368 * hence their active_lock.ptr becomes map_ptr and id = 0 (since they all point
10369 * into the same map value as max_entries is 1, as described above).
10371 * In case of inner map lookups, the inner map pointer has same map_ptr as the
10372 * outer map pointer (in verifier context), but each lookup into an inner map
10373 * assigns a fresh reg->id to the lookup, so while lookups into distinct inner
10374 * maps from the same outer map share the same map_ptr as active_lock.ptr, they
10375 * will get different reg->id assigned to each lookup, hence different
10378 * In case of allocated objects, active_lock.ptr is the reg->btf, and the
10379 * reg->id is a unique ID preserved after the NULL pointer check on the pointer
10380 * returned from bpf_obj_new. Each allocation receives a new reg->id.
10382 static int check_reg_allocation_locked(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
10387 switch ((int)reg->type) {
10388 case PTR_TO_MAP_VALUE:
10389 ptr = reg->map_ptr;
10391 case PTR_TO_BTF_ID | MEM_ALLOC:
10395 verbose(env, "verifier internal error: unknown reg type for lock check\n");
10400 if (!env->cur_state->active_lock.ptr)
10402 if (env->cur_state->active_lock.ptr != ptr ||
10403 env->cur_state->active_lock.id != id) {
10404 verbose(env, "held lock and object are not in the same allocation\n");
10410 static bool is_bpf_list_api_kfunc(u32 btf_id)
10412 return btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
10413 btf_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
10414 btf_id == special_kfunc_list[KF_bpf_list_pop_front] ||
10415 btf_id == special_kfunc_list[KF_bpf_list_pop_back];
10418 static bool is_bpf_rbtree_api_kfunc(u32 btf_id)
10420 return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl] ||
10421 btf_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
10422 btf_id == special_kfunc_list[KF_bpf_rbtree_first];
10425 static bool is_bpf_graph_api_kfunc(u32 btf_id)
10427 return is_bpf_list_api_kfunc(btf_id) || is_bpf_rbtree_api_kfunc(btf_id) ||
10428 btf_id == special_kfunc_list[KF_bpf_refcount_acquire_impl];
10431 static bool is_callback_calling_kfunc(u32 btf_id)
10433 return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl];
10436 static bool is_rbtree_lock_required_kfunc(u32 btf_id)
10438 return is_bpf_rbtree_api_kfunc(btf_id);
10441 static bool check_kfunc_is_graph_root_api(struct bpf_verifier_env *env,
10442 enum btf_field_type head_field_type,
10447 switch (head_field_type) {
10448 case BPF_LIST_HEAD:
10449 ret = is_bpf_list_api_kfunc(kfunc_btf_id);
10452 ret = is_bpf_rbtree_api_kfunc(kfunc_btf_id);
10455 verbose(env, "verifier internal error: unexpected graph root argument type %s\n",
10456 btf_field_type_name(head_field_type));
10461 verbose(env, "verifier internal error: %s head arg for unknown kfunc\n",
10462 btf_field_type_name(head_field_type));
10466 static bool check_kfunc_is_graph_node_api(struct bpf_verifier_env *env,
10467 enum btf_field_type node_field_type,
10472 switch (node_field_type) {
10473 case BPF_LIST_NODE:
10474 ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
10475 kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_back_impl]);
10478 ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
10479 kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl]);
10482 verbose(env, "verifier internal error: unexpected graph node argument type %s\n",
10483 btf_field_type_name(node_field_type));
10488 verbose(env, "verifier internal error: %s node arg for unknown kfunc\n",
10489 btf_field_type_name(node_field_type));
10494 __process_kf_arg_ptr_to_graph_root(struct bpf_verifier_env *env,
10495 struct bpf_reg_state *reg, u32 regno,
10496 struct bpf_kfunc_call_arg_meta *meta,
10497 enum btf_field_type head_field_type,
10498 struct btf_field **head_field)
10500 const char *head_type_name;
10501 struct btf_field *field;
10502 struct btf_record *rec;
10505 if (meta->btf != btf_vmlinux) {
10506 verbose(env, "verifier internal error: unexpected btf mismatch in kfunc call\n");
10510 if (!check_kfunc_is_graph_root_api(env, head_field_type, meta->func_id))
10513 head_type_name = btf_field_type_name(head_field_type);
10514 if (!tnum_is_const(reg->var_off)) {
10516 "R%d doesn't have constant offset. %s has to be at the constant offset\n",
10517 regno, head_type_name);
10521 rec = reg_btf_record(reg);
10522 head_off = reg->off + reg->var_off.value;
10523 field = btf_record_find(rec, head_off, head_field_type);
10525 verbose(env, "%s not found at offset=%u\n", head_type_name, head_off);
10529 /* All functions require bpf_list_head to be protected using a bpf_spin_lock */
10530 if (check_reg_allocation_locked(env, reg)) {
10531 verbose(env, "bpf_spin_lock at off=%d must be held for %s\n",
10532 rec->spin_lock_off, head_type_name);
10537 verbose(env, "verifier internal error: repeating %s arg\n", head_type_name);
10540 *head_field = field;
10544 static int process_kf_arg_ptr_to_list_head(struct bpf_verifier_env *env,
10545 struct bpf_reg_state *reg, u32 regno,
10546 struct bpf_kfunc_call_arg_meta *meta)
10548 return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, BPF_LIST_HEAD,
10549 &meta->arg_list_head.field);
10552 static int process_kf_arg_ptr_to_rbtree_root(struct bpf_verifier_env *env,
10553 struct bpf_reg_state *reg, u32 regno,
10554 struct bpf_kfunc_call_arg_meta *meta)
10556 return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, BPF_RB_ROOT,
10557 &meta->arg_rbtree_root.field);
10561 __process_kf_arg_ptr_to_graph_node(struct bpf_verifier_env *env,
10562 struct bpf_reg_state *reg, u32 regno,
10563 struct bpf_kfunc_call_arg_meta *meta,
10564 enum btf_field_type head_field_type,
10565 enum btf_field_type node_field_type,
10566 struct btf_field **node_field)
10568 const char *node_type_name;
10569 const struct btf_type *et, *t;
10570 struct btf_field *field;
10573 if (meta->btf != btf_vmlinux) {
10574 verbose(env, "verifier internal error: unexpected btf mismatch in kfunc call\n");
10578 if (!check_kfunc_is_graph_node_api(env, node_field_type, meta->func_id))
10581 node_type_name = btf_field_type_name(node_field_type);
10582 if (!tnum_is_const(reg->var_off)) {
10584 "R%d doesn't have constant offset. %s has to be at the constant offset\n",
10585 regno, node_type_name);
10589 node_off = reg->off + reg->var_off.value;
10590 field = reg_find_field_offset(reg, node_off, node_field_type);
10591 if (!field || field->offset != node_off) {
10592 verbose(env, "%s not found at offset=%u\n", node_type_name, node_off);
10596 field = *node_field;
10598 et = btf_type_by_id(field->graph_root.btf, field->graph_root.value_btf_id);
10599 t = btf_type_by_id(reg->btf, reg->btf_id);
10600 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, 0, field->graph_root.btf,
10601 field->graph_root.value_btf_id, true)) {
10602 verbose(env, "operation on %s expects arg#1 %s at offset=%d "
10603 "in struct %s, but arg is at offset=%d in struct %s\n",
10604 btf_field_type_name(head_field_type),
10605 btf_field_type_name(node_field_type),
10606 field->graph_root.node_offset,
10607 btf_name_by_offset(field->graph_root.btf, et->name_off),
10608 node_off, btf_name_by_offset(reg->btf, t->name_off));
10611 meta->arg_btf = reg->btf;
10612 meta->arg_btf_id = reg->btf_id;
10614 if (node_off != field->graph_root.node_offset) {
10615 verbose(env, "arg#1 offset=%d, but expected %s at offset=%d in struct %s\n",
10616 node_off, btf_field_type_name(node_field_type),
10617 field->graph_root.node_offset,
10618 btf_name_by_offset(field->graph_root.btf, et->name_off));
10625 static int process_kf_arg_ptr_to_list_node(struct bpf_verifier_env *env,
10626 struct bpf_reg_state *reg, u32 regno,
10627 struct bpf_kfunc_call_arg_meta *meta)
10629 return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta,
10630 BPF_LIST_HEAD, BPF_LIST_NODE,
10631 &meta->arg_list_head.field);
10634 static int process_kf_arg_ptr_to_rbtree_node(struct bpf_verifier_env *env,
10635 struct bpf_reg_state *reg, u32 regno,
10636 struct bpf_kfunc_call_arg_meta *meta)
10638 return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta,
10639 BPF_RB_ROOT, BPF_RB_NODE,
10640 &meta->arg_rbtree_root.field);
10643 static int check_kfunc_args(struct bpf_verifier_env *env, struct bpf_kfunc_call_arg_meta *meta,
10646 const char *func_name = meta->func_name, *ref_tname;
10647 const struct btf *btf = meta->btf;
10648 const struct btf_param *args;
10649 struct btf_record *rec;
10653 args = (const struct btf_param *)(meta->func_proto + 1);
10654 nargs = btf_type_vlen(meta->func_proto);
10655 if (nargs > MAX_BPF_FUNC_REG_ARGS) {
10656 verbose(env, "Function %s has %d > %d args\n", func_name, nargs,
10657 MAX_BPF_FUNC_REG_ARGS);
10661 /* Check that BTF function arguments match actual types that the
10664 for (i = 0; i < nargs; i++) {
10665 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[i + 1];
10666 const struct btf_type *t, *ref_t, *resolve_ret;
10667 enum bpf_arg_type arg_type = ARG_DONTCARE;
10668 u32 regno = i + 1, ref_id, type_size;
10669 bool is_ret_buf_sz = false;
10672 t = btf_type_skip_modifiers(btf, args[i].type, NULL);
10674 if (is_kfunc_arg_ignore(btf, &args[i]))
10677 if (btf_type_is_scalar(t)) {
10678 if (reg->type != SCALAR_VALUE) {
10679 verbose(env, "R%d is not a scalar\n", regno);
10683 if (is_kfunc_arg_constant(meta->btf, &args[i])) {
10684 if (meta->arg_constant.found) {
10685 verbose(env, "verifier internal error: only one constant argument permitted\n");
10688 if (!tnum_is_const(reg->var_off)) {
10689 verbose(env, "R%d must be a known constant\n", regno);
10692 ret = mark_chain_precision(env, regno);
10695 meta->arg_constant.found = true;
10696 meta->arg_constant.value = reg->var_off.value;
10697 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdonly_buf_size")) {
10698 meta->r0_rdonly = true;
10699 is_ret_buf_sz = true;
10700 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdwr_buf_size")) {
10701 is_ret_buf_sz = true;
10704 if (is_ret_buf_sz) {
10705 if (meta->r0_size) {
10706 verbose(env, "2 or more rdonly/rdwr_buf_size parameters for kfunc");
10710 if (!tnum_is_const(reg->var_off)) {
10711 verbose(env, "R%d is not a const\n", regno);
10715 meta->r0_size = reg->var_off.value;
10716 ret = mark_chain_precision(env, regno);
10723 if (!btf_type_is_ptr(t)) {
10724 verbose(env, "Unrecognized arg#%d type %s\n", i, btf_type_str(t));
10728 if ((is_kfunc_trusted_args(meta) || is_kfunc_rcu(meta)) &&
10729 (register_is_null(reg) || type_may_be_null(reg->type))) {
10730 verbose(env, "Possibly NULL pointer passed to trusted arg%d\n", i);
10734 if (reg->ref_obj_id) {
10735 if (is_kfunc_release(meta) && meta->ref_obj_id) {
10736 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
10737 regno, reg->ref_obj_id,
10741 meta->ref_obj_id = reg->ref_obj_id;
10742 if (is_kfunc_release(meta))
10743 meta->release_regno = regno;
10746 ref_t = btf_type_skip_modifiers(btf, t->type, &ref_id);
10747 ref_tname = btf_name_by_offset(btf, ref_t->name_off);
10749 kf_arg_type = get_kfunc_ptr_arg_type(env, meta, t, ref_t, ref_tname, args, i, nargs);
10750 if (kf_arg_type < 0)
10751 return kf_arg_type;
10753 switch (kf_arg_type) {
10754 case KF_ARG_PTR_TO_ALLOC_BTF_ID:
10755 case KF_ARG_PTR_TO_BTF_ID:
10756 if (!is_kfunc_trusted_args(meta) && !is_kfunc_rcu(meta))
10759 if (!is_trusted_reg(reg)) {
10760 if (!is_kfunc_rcu(meta)) {
10761 verbose(env, "R%d must be referenced or trusted\n", regno);
10764 if (!is_rcu_reg(reg)) {
10765 verbose(env, "R%d must be a rcu pointer\n", regno);
10771 case KF_ARG_PTR_TO_CTX:
10772 /* Trusted arguments have the same offset checks as release arguments */
10773 arg_type |= OBJ_RELEASE;
10775 case KF_ARG_PTR_TO_DYNPTR:
10776 case KF_ARG_PTR_TO_ITER:
10777 case KF_ARG_PTR_TO_LIST_HEAD:
10778 case KF_ARG_PTR_TO_LIST_NODE:
10779 case KF_ARG_PTR_TO_RB_ROOT:
10780 case KF_ARG_PTR_TO_RB_NODE:
10781 case KF_ARG_PTR_TO_MEM:
10782 case KF_ARG_PTR_TO_MEM_SIZE:
10783 case KF_ARG_PTR_TO_CALLBACK:
10784 case KF_ARG_PTR_TO_REFCOUNTED_KPTR:
10785 /* Trusted by default */
10792 if (is_kfunc_release(meta) && reg->ref_obj_id)
10793 arg_type |= OBJ_RELEASE;
10794 ret = check_func_arg_reg_off(env, reg, regno, arg_type);
10798 switch (kf_arg_type) {
10799 case KF_ARG_PTR_TO_CTX:
10800 if (reg->type != PTR_TO_CTX) {
10801 verbose(env, "arg#%d expected pointer to ctx, but got %s\n", i, btf_type_str(t));
10805 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
10806 ret = get_kern_ctx_btf_id(&env->log, resolve_prog_type(env->prog));
10809 meta->ret_btf_id = ret;
10812 case KF_ARG_PTR_TO_ALLOC_BTF_ID:
10813 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
10814 verbose(env, "arg#%d expected pointer to allocated object\n", i);
10817 if (!reg->ref_obj_id) {
10818 verbose(env, "allocated object must be referenced\n");
10821 if (meta->btf == btf_vmlinux &&
10822 meta->func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
10823 meta->arg_btf = reg->btf;
10824 meta->arg_btf_id = reg->btf_id;
10827 case KF_ARG_PTR_TO_DYNPTR:
10829 enum bpf_arg_type dynptr_arg_type = ARG_PTR_TO_DYNPTR;
10830 int clone_ref_obj_id = 0;
10832 if (reg->type != PTR_TO_STACK &&
10833 reg->type != CONST_PTR_TO_DYNPTR) {
10834 verbose(env, "arg#%d expected pointer to stack or dynptr_ptr\n", i);
10838 if (reg->type == CONST_PTR_TO_DYNPTR)
10839 dynptr_arg_type |= MEM_RDONLY;
10841 if (is_kfunc_arg_uninit(btf, &args[i]))
10842 dynptr_arg_type |= MEM_UNINIT;
10844 if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) {
10845 dynptr_arg_type |= DYNPTR_TYPE_SKB;
10846 } else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_xdp]) {
10847 dynptr_arg_type |= DYNPTR_TYPE_XDP;
10848 } else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_clone] &&
10849 (dynptr_arg_type & MEM_UNINIT)) {
10850 enum bpf_dynptr_type parent_type = meta->initialized_dynptr.type;
10852 if (parent_type == BPF_DYNPTR_TYPE_INVALID) {
10853 verbose(env, "verifier internal error: no dynptr type for parent of clone\n");
10857 dynptr_arg_type |= (unsigned int)get_dynptr_type_flag(parent_type);
10858 clone_ref_obj_id = meta->initialized_dynptr.ref_obj_id;
10859 if (dynptr_type_refcounted(parent_type) && !clone_ref_obj_id) {
10860 verbose(env, "verifier internal error: missing ref obj id for parent of clone\n");
10865 ret = process_dynptr_func(env, regno, insn_idx, dynptr_arg_type, clone_ref_obj_id);
10869 if (!(dynptr_arg_type & MEM_UNINIT)) {
10870 int id = dynptr_id(env, reg);
10873 verbose(env, "verifier internal error: failed to obtain dynptr id\n");
10876 meta->initialized_dynptr.id = id;
10877 meta->initialized_dynptr.type = dynptr_get_type(env, reg);
10878 meta->initialized_dynptr.ref_obj_id = dynptr_ref_obj_id(env, reg);
10883 case KF_ARG_PTR_TO_ITER:
10884 ret = process_iter_arg(env, regno, insn_idx, meta);
10888 case KF_ARG_PTR_TO_LIST_HEAD:
10889 if (reg->type != PTR_TO_MAP_VALUE &&
10890 reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
10891 verbose(env, "arg#%d expected pointer to map value or allocated object\n", i);
10894 if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
10895 verbose(env, "allocated object must be referenced\n");
10898 ret = process_kf_arg_ptr_to_list_head(env, reg, regno, meta);
10902 case KF_ARG_PTR_TO_RB_ROOT:
10903 if (reg->type != PTR_TO_MAP_VALUE &&
10904 reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
10905 verbose(env, "arg#%d expected pointer to map value or allocated object\n", i);
10908 if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
10909 verbose(env, "allocated object must be referenced\n");
10912 ret = process_kf_arg_ptr_to_rbtree_root(env, reg, regno, meta);
10916 case KF_ARG_PTR_TO_LIST_NODE:
10917 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
10918 verbose(env, "arg#%d expected pointer to allocated object\n", i);
10921 if (!reg->ref_obj_id) {
10922 verbose(env, "allocated object must be referenced\n");
10925 ret = process_kf_arg_ptr_to_list_node(env, reg, regno, meta);
10929 case KF_ARG_PTR_TO_RB_NODE:
10930 if (meta->func_id == special_kfunc_list[KF_bpf_rbtree_remove]) {
10931 if (!type_is_non_owning_ref(reg->type) || reg->ref_obj_id) {
10932 verbose(env, "rbtree_remove node input must be non-owning ref\n");
10935 if (in_rbtree_lock_required_cb(env)) {
10936 verbose(env, "rbtree_remove not allowed in rbtree cb\n");
10940 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
10941 verbose(env, "arg#%d expected pointer to allocated object\n", i);
10944 if (!reg->ref_obj_id) {
10945 verbose(env, "allocated object must be referenced\n");
10950 ret = process_kf_arg_ptr_to_rbtree_node(env, reg, regno, meta);
10954 case KF_ARG_PTR_TO_BTF_ID:
10955 /* Only base_type is checked, further checks are done here */
10956 if ((base_type(reg->type) != PTR_TO_BTF_ID ||
10957 (bpf_type_has_unsafe_modifiers(reg->type) && !is_rcu_reg(reg))) &&
10958 !reg2btf_ids[base_type(reg->type)]) {
10959 verbose(env, "arg#%d is %s ", i, reg_type_str(env, reg->type));
10960 verbose(env, "expected %s or socket\n",
10961 reg_type_str(env, base_type(reg->type) |
10962 (type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS)));
10965 ret = process_kf_arg_ptr_to_btf_id(env, reg, ref_t, ref_tname, ref_id, meta, i);
10969 case KF_ARG_PTR_TO_MEM:
10970 resolve_ret = btf_resolve_size(btf, ref_t, &type_size);
10971 if (IS_ERR(resolve_ret)) {
10972 verbose(env, "arg#%d reference type('%s %s') size cannot be determined: %ld\n",
10973 i, btf_type_str(ref_t), ref_tname, PTR_ERR(resolve_ret));
10976 ret = check_mem_reg(env, reg, regno, type_size);
10980 case KF_ARG_PTR_TO_MEM_SIZE:
10982 struct bpf_reg_state *buff_reg = ®s[regno];
10983 const struct btf_param *buff_arg = &args[i];
10984 struct bpf_reg_state *size_reg = ®s[regno + 1];
10985 const struct btf_param *size_arg = &args[i + 1];
10987 if (!register_is_null(buff_reg) || !is_kfunc_arg_optional(meta->btf, buff_arg)) {
10988 ret = check_kfunc_mem_size_reg(env, size_reg, regno + 1);
10990 verbose(env, "arg#%d arg#%d memory, len pair leads to invalid memory access\n", i, i + 1);
10995 if (is_kfunc_arg_const_mem_size(meta->btf, size_arg, size_reg)) {
10996 if (meta->arg_constant.found) {
10997 verbose(env, "verifier internal error: only one constant argument permitted\n");
11000 if (!tnum_is_const(size_reg->var_off)) {
11001 verbose(env, "R%d must be a known constant\n", regno + 1);
11004 meta->arg_constant.found = true;
11005 meta->arg_constant.value = size_reg->var_off.value;
11008 /* Skip next '__sz' or '__szk' argument */
11012 case KF_ARG_PTR_TO_CALLBACK:
11013 meta->subprogno = reg->subprogno;
11015 case KF_ARG_PTR_TO_REFCOUNTED_KPTR:
11016 if (!type_is_ptr_alloc_obj(reg->type)) {
11017 verbose(env, "arg#%d is neither owning or non-owning ref\n", i);
11020 if (!type_is_non_owning_ref(reg->type))
11021 meta->arg_owning_ref = true;
11023 rec = reg_btf_record(reg);
11025 verbose(env, "verifier internal error: Couldn't find btf_record\n");
11029 if (rec->refcount_off < 0) {
11030 verbose(env, "arg#%d doesn't point to a type with bpf_refcount field\n", i);
11033 if (rec->refcount_off >= 0) {
11034 verbose(env, "bpf_refcount_acquire calls are disabled for now\n");
11037 meta->arg_btf = reg->btf;
11038 meta->arg_btf_id = reg->btf_id;
11043 if (is_kfunc_release(meta) && !meta->release_regno) {
11044 verbose(env, "release kernel function %s expects refcounted PTR_TO_BTF_ID\n",
11052 static int fetch_kfunc_meta(struct bpf_verifier_env *env,
11053 struct bpf_insn *insn,
11054 struct bpf_kfunc_call_arg_meta *meta,
11055 const char **kfunc_name)
11057 const struct btf_type *func, *func_proto;
11058 u32 func_id, *kfunc_flags;
11059 const char *func_name;
11060 struct btf *desc_btf;
11063 *kfunc_name = NULL;
11068 desc_btf = find_kfunc_desc_btf(env, insn->off);
11069 if (IS_ERR(desc_btf))
11070 return PTR_ERR(desc_btf);
11072 func_id = insn->imm;
11073 func = btf_type_by_id(desc_btf, func_id);
11074 func_name = btf_name_by_offset(desc_btf, func->name_off);
11076 *kfunc_name = func_name;
11077 func_proto = btf_type_by_id(desc_btf, func->type);
11079 kfunc_flags = btf_kfunc_id_set_contains(desc_btf, func_id, env->prog);
11080 if (!kfunc_flags) {
11084 memset(meta, 0, sizeof(*meta));
11085 meta->btf = desc_btf;
11086 meta->func_id = func_id;
11087 meta->kfunc_flags = *kfunc_flags;
11088 meta->func_proto = func_proto;
11089 meta->func_name = func_name;
11094 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
11097 const struct btf_type *t, *ptr_type;
11098 u32 i, nargs, ptr_type_id, release_ref_obj_id;
11099 struct bpf_reg_state *regs = cur_regs(env);
11100 const char *func_name, *ptr_type_name;
11101 bool sleepable, rcu_lock, rcu_unlock;
11102 struct bpf_kfunc_call_arg_meta meta;
11103 struct bpf_insn_aux_data *insn_aux;
11104 int err, insn_idx = *insn_idx_p;
11105 const struct btf_param *args;
11106 const struct btf_type *ret_t;
11107 struct btf *desc_btf;
11109 /* skip for now, but return error when we find this in fixup_kfunc_call */
11113 err = fetch_kfunc_meta(env, insn, &meta, &func_name);
11114 if (err == -EACCES && func_name)
11115 verbose(env, "calling kernel function %s is not allowed\n", func_name);
11118 desc_btf = meta.btf;
11119 insn_aux = &env->insn_aux_data[insn_idx];
11121 insn_aux->is_iter_next = is_iter_next_kfunc(&meta);
11123 if (is_kfunc_destructive(&meta) && !capable(CAP_SYS_BOOT)) {
11124 verbose(env, "destructive kfunc calls require CAP_SYS_BOOT capability\n");
11128 sleepable = is_kfunc_sleepable(&meta);
11129 if (sleepable && !env->prog->aux->sleepable) {
11130 verbose(env, "program must be sleepable to call sleepable kfunc %s\n", func_name);
11134 rcu_lock = is_kfunc_bpf_rcu_read_lock(&meta);
11135 rcu_unlock = is_kfunc_bpf_rcu_read_unlock(&meta);
11137 if (env->cur_state->active_rcu_lock) {
11138 struct bpf_func_state *state;
11139 struct bpf_reg_state *reg;
11142 verbose(env, "nested rcu read lock (kernel function %s)\n", func_name);
11144 } else if (rcu_unlock) {
11145 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
11146 if (reg->type & MEM_RCU) {
11147 reg->type &= ~(MEM_RCU | PTR_MAYBE_NULL);
11148 reg->type |= PTR_UNTRUSTED;
11151 env->cur_state->active_rcu_lock = false;
11152 } else if (sleepable) {
11153 verbose(env, "kernel func %s is sleepable within rcu_read_lock region\n", func_name);
11156 } else if (rcu_lock) {
11157 env->cur_state->active_rcu_lock = true;
11158 } else if (rcu_unlock) {
11159 verbose(env, "unmatched rcu read unlock (kernel function %s)\n", func_name);
11163 /* Check the arguments */
11164 err = check_kfunc_args(env, &meta, insn_idx);
11167 /* In case of release function, we get register number of refcounted
11168 * PTR_TO_BTF_ID in bpf_kfunc_arg_meta, do the release now.
11170 if (meta.release_regno) {
11171 err = release_reference(env, regs[meta.release_regno].ref_obj_id);
11173 verbose(env, "kfunc %s#%d reference has not been acquired before\n",
11174 func_name, meta.func_id);
11179 if (meta.func_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
11180 meta.func_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
11181 meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
11182 release_ref_obj_id = regs[BPF_REG_2].ref_obj_id;
11183 insn_aux->insert_off = regs[BPF_REG_2].off;
11184 insn_aux->kptr_struct_meta = btf_find_struct_meta(meta.arg_btf, meta.arg_btf_id);
11185 err = ref_convert_owning_non_owning(env, release_ref_obj_id);
11187 verbose(env, "kfunc %s#%d conversion of owning ref to non-owning failed\n",
11188 func_name, meta.func_id);
11192 err = release_reference(env, release_ref_obj_id);
11194 verbose(env, "kfunc %s#%d reference has not been acquired before\n",
11195 func_name, meta.func_id);
11200 if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
11201 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
11202 set_rbtree_add_callback_state);
11204 verbose(env, "kfunc %s#%d failed callback verification\n",
11205 func_name, meta.func_id);
11210 for (i = 0; i < CALLER_SAVED_REGS; i++)
11211 mark_reg_not_init(env, regs, caller_saved[i]);
11213 /* Check return type */
11214 t = btf_type_skip_modifiers(desc_btf, meta.func_proto->type, NULL);
11216 if (is_kfunc_acquire(&meta) && !btf_type_is_struct_ptr(meta.btf, t)) {
11217 /* Only exception is bpf_obj_new_impl */
11218 if (meta.btf != btf_vmlinux ||
11219 (meta.func_id != special_kfunc_list[KF_bpf_obj_new_impl] &&
11220 meta.func_id != special_kfunc_list[KF_bpf_refcount_acquire_impl])) {
11221 verbose(env, "acquire kernel function does not return PTR_TO_BTF_ID\n");
11226 if (btf_type_is_scalar(t)) {
11227 mark_reg_unknown(env, regs, BPF_REG_0);
11228 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
11229 } else if (btf_type_is_ptr(t)) {
11230 ptr_type = btf_type_skip_modifiers(desc_btf, t->type, &ptr_type_id);
11232 if (meta.btf == btf_vmlinux && btf_id_set_contains(&special_kfunc_set, meta.func_id)) {
11233 if (meta.func_id == special_kfunc_list[KF_bpf_obj_new_impl]) {
11234 struct btf *ret_btf;
11237 if (unlikely(!bpf_global_ma_set))
11240 if (((u64)(u32)meta.arg_constant.value) != meta.arg_constant.value) {
11241 verbose(env, "local type ID argument must be in range [0, U32_MAX]\n");
11245 ret_btf = env->prog->aux->btf;
11246 ret_btf_id = meta.arg_constant.value;
11248 /* This may be NULL due to user not supplying a BTF */
11250 verbose(env, "bpf_obj_new requires prog BTF\n");
11254 ret_t = btf_type_by_id(ret_btf, ret_btf_id);
11255 if (!ret_t || !__btf_type_is_struct(ret_t)) {
11256 verbose(env, "bpf_obj_new type ID argument must be of a struct\n");
11260 mark_reg_known_zero(env, regs, BPF_REG_0);
11261 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
11262 regs[BPF_REG_0].btf = ret_btf;
11263 regs[BPF_REG_0].btf_id = ret_btf_id;
11265 insn_aux->obj_new_size = ret_t->size;
11266 insn_aux->kptr_struct_meta =
11267 btf_find_struct_meta(ret_btf, ret_btf_id);
11268 } else if (meta.func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) {
11269 mark_reg_known_zero(env, regs, BPF_REG_0);
11270 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
11271 regs[BPF_REG_0].btf = meta.arg_btf;
11272 regs[BPF_REG_0].btf_id = meta.arg_btf_id;
11274 insn_aux->kptr_struct_meta =
11275 btf_find_struct_meta(meta.arg_btf,
11277 } else if (meta.func_id == special_kfunc_list[KF_bpf_list_pop_front] ||
11278 meta.func_id == special_kfunc_list[KF_bpf_list_pop_back]) {
11279 struct btf_field *field = meta.arg_list_head.field;
11281 mark_reg_graph_node(regs, BPF_REG_0, &field->graph_root);
11282 } else if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
11283 meta.func_id == special_kfunc_list[KF_bpf_rbtree_first]) {
11284 struct btf_field *field = meta.arg_rbtree_root.field;
11286 mark_reg_graph_node(regs, BPF_REG_0, &field->graph_root);
11287 } else if (meta.func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
11288 mark_reg_known_zero(env, regs, BPF_REG_0);
11289 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_TRUSTED;
11290 regs[BPF_REG_0].btf = desc_btf;
11291 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
11292 } else if (meta.func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
11293 ret_t = btf_type_by_id(desc_btf, meta.arg_constant.value);
11294 if (!ret_t || !btf_type_is_struct(ret_t)) {
11296 "kfunc bpf_rdonly_cast type ID argument must be of a struct\n");
11300 mark_reg_known_zero(env, regs, BPF_REG_0);
11301 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
11302 regs[BPF_REG_0].btf = desc_btf;
11303 regs[BPF_REG_0].btf_id = meta.arg_constant.value;
11304 } else if (meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice] ||
11305 meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice_rdwr]) {
11306 enum bpf_type_flag type_flag = get_dynptr_type_flag(meta.initialized_dynptr.type);
11308 mark_reg_known_zero(env, regs, BPF_REG_0);
11310 if (!meta.arg_constant.found) {
11311 verbose(env, "verifier internal error: bpf_dynptr_slice(_rdwr) no constant size\n");
11315 regs[BPF_REG_0].mem_size = meta.arg_constant.value;
11317 /* PTR_MAYBE_NULL will be added when is_kfunc_ret_null is checked */
11318 regs[BPF_REG_0].type = PTR_TO_MEM | type_flag;
11320 if (meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice]) {
11321 regs[BPF_REG_0].type |= MEM_RDONLY;
11323 /* this will set env->seen_direct_write to true */
11324 if (!may_access_direct_pkt_data(env, NULL, BPF_WRITE)) {
11325 verbose(env, "the prog does not allow writes to packet data\n");
11330 if (!meta.initialized_dynptr.id) {
11331 verbose(env, "verifier internal error: no dynptr id\n");
11334 regs[BPF_REG_0].dynptr_id = meta.initialized_dynptr.id;
11336 /* we don't need to set BPF_REG_0's ref obj id
11337 * because packet slices are not refcounted (see
11338 * dynptr_type_refcounted)
11341 verbose(env, "kernel function %s unhandled dynamic return type\n",
11345 } else if (!__btf_type_is_struct(ptr_type)) {
11346 if (!meta.r0_size) {
11349 if (!IS_ERR(btf_resolve_size(desc_btf, ptr_type, &sz))) {
11351 meta.r0_rdonly = true;
11354 if (!meta.r0_size) {
11355 ptr_type_name = btf_name_by_offset(desc_btf,
11356 ptr_type->name_off);
11358 "kernel function %s returns pointer type %s %s is not supported\n",
11360 btf_type_str(ptr_type),
11365 mark_reg_known_zero(env, regs, BPF_REG_0);
11366 regs[BPF_REG_0].type = PTR_TO_MEM;
11367 regs[BPF_REG_0].mem_size = meta.r0_size;
11369 if (meta.r0_rdonly)
11370 regs[BPF_REG_0].type |= MEM_RDONLY;
11372 /* Ensures we don't access the memory after a release_reference() */
11373 if (meta.ref_obj_id)
11374 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
11376 mark_reg_known_zero(env, regs, BPF_REG_0);
11377 regs[BPF_REG_0].btf = desc_btf;
11378 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
11379 regs[BPF_REG_0].btf_id = ptr_type_id;
11382 if (is_kfunc_ret_null(&meta)) {
11383 regs[BPF_REG_0].type |= PTR_MAYBE_NULL;
11384 /* For mark_ptr_or_null_reg, see 93c230e3f5bd6 */
11385 regs[BPF_REG_0].id = ++env->id_gen;
11387 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
11388 if (is_kfunc_acquire(&meta)) {
11389 int id = acquire_reference_state(env, insn_idx);
11393 if (is_kfunc_ret_null(&meta))
11394 regs[BPF_REG_0].id = id;
11395 regs[BPF_REG_0].ref_obj_id = id;
11396 } else if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_first]) {
11397 ref_set_non_owning(env, ®s[BPF_REG_0]);
11400 if (reg_may_point_to_spin_lock(®s[BPF_REG_0]) && !regs[BPF_REG_0].id)
11401 regs[BPF_REG_0].id = ++env->id_gen;
11402 } else if (btf_type_is_void(t)) {
11403 if (meta.btf == btf_vmlinux && btf_id_set_contains(&special_kfunc_set, meta.func_id)) {
11404 if (meta.func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
11405 insn_aux->kptr_struct_meta =
11406 btf_find_struct_meta(meta.arg_btf,
11412 nargs = btf_type_vlen(meta.func_proto);
11413 args = (const struct btf_param *)(meta.func_proto + 1);
11414 for (i = 0; i < nargs; i++) {
11417 t = btf_type_skip_modifiers(desc_btf, args[i].type, NULL);
11418 if (btf_type_is_ptr(t))
11419 mark_btf_func_reg_size(env, regno, sizeof(void *));
11421 /* scalar. ensured by btf_check_kfunc_arg_match() */
11422 mark_btf_func_reg_size(env, regno, t->size);
11425 if (is_iter_next_kfunc(&meta)) {
11426 err = process_iter_next_call(env, insn_idx, &meta);
11434 static bool signed_add_overflows(s64 a, s64 b)
11436 /* Do the add in u64, where overflow is well-defined */
11437 s64 res = (s64)((u64)a + (u64)b);
11444 static bool signed_add32_overflows(s32 a, s32 b)
11446 /* Do the add in u32, where overflow is well-defined */
11447 s32 res = (s32)((u32)a + (u32)b);
11454 static bool signed_sub_overflows(s64 a, s64 b)
11456 /* Do the sub in u64, where overflow is well-defined */
11457 s64 res = (s64)((u64)a - (u64)b);
11464 static bool signed_sub32_overflows(s32 a, s32 b)
11466 /* Do the sub in u32, where overflow is well-defined */
11467 s32 res = (s32)((u32)a - (u32)b);
11474 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
11475 const struct bpf_reg_state *reg,
11476 enum bpf_reg_type type)
11478 bool known = tnum_is_const(reg->var_off);
11479 s64 val = reg->var_off.value;
11480 s64 smin = reg->smin_value;
11482 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
11483 verbose(env, "math between %s pointer and %lld is not allowed\n",
11484 reg_type_str(env, type), val);
11488 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
11489 verbose(env, "%s pointer offset %d is not allowed\n",
11490 reg_type_str(env, type), reg->off);
11494 if (smin == S64_MIN) {
11495 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
11496 reg_type_str(env, type));
11500 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
11501 verbose(env, "value %lld makes %s pointer be out of bounds\n",
11502 smin, reg_type_str(env, type));
11510 REASON_BOUNDS = -1,
11517 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
11518 u32 *alu_limit, bool mask_to_left)
11520 u32 max = 0, ptr_limit = 0;
11522 switch (ptr_reg->type) {
11524 /* Offset 0 is out-of-bounds, but acceptable start for the
11525 * left direction, see BPF_REG_FP. Also, unknown scalar
11526 * offset where we would need to deal with min/max bounds is
11527 * currently prohibited for unprivileged.
11529 max = MAX_BPF_STACK + mask_to_left;
11530 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
11532 case PTR_TO_MAP_VALUE:
11533 max = ptr_reg->map_ptr->value_size;
11534 ptr_limit = (mask_to_left ?
11535 ptr_reg->smin_value :
11536 ptr_reg->umax_value) + ptr_reg->off;
11539 return REASON_TYPE;
11542 if (ptr_limit >= max)
11543 return REASON_LIMIT;
11544 *alu_limit = ptr_limit;
11548 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
11549 const struct bpf_insn *insn)
11551 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
11554 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
11555 u32 alu_state, u32 alu_limit)
11557 /* If we arrived here from different branches with different
11558 * state or limits to sanitize, then this won't work.
11560 if (aux->alu_state &&
11561 (aux->alu_state != alu_state ||
11562 aux->alu_limit != alu_limit))
11563 return REASON_PATHS;
11565 /* Corresponding fixup done in do_misc_fixups(). */
11566 aux->alu_state = alu_state;
11567 aux->alu_limit = alu_limit;
11571 static int sanitize_val_alu(struct bpf_verifier_env *env,
11572 struct bpf_insn *insn)
11574 struct bpf_insn_aux_data *aux = cur_aux(env);
11576 if (can_skip_alu_sanitation(env, insn))
11579 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
11582 static bool sanitize_needed(u8 opcode)
11584 return opcode == BPF_ADD || opcode == BPF_SUB;
11587 struct bpf_sanitize_info {
11588 struct bpf_insn_aux_data aux;
11592 static struct bpf_verifier_state *
11593 sanitize_speculative_path(struct bpf_verifier_env *env,
11594 const struct bpf_insn *insn,
11595 u32 next_idx, u32 curr_idx)
11597 struct bpf_verifier_state *branch;
11598 struct bpf_reg_state *regs;
11600 branch = push_stack(env, next_idx, curr_idx, true);
11601 if (branch && insn) {
11602 regs = branch->frame[branch->curframe]->regs;
11603 if (BPF_SRC(insn->code) == BPF_K) {
11604 mark_reg_unknown(env, regs, insn->dst_reg);
11605 } else if (BPF_SRC(insn->code) == BPF_X) {
11606 mark_reg_unknown(env, regs, insn->dst_reg);
11607 mark_reg_unknown(env, regs, insn->src_reg);
11613 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
11614 struct bpf_insn *insn,
11615 const struct bpf_reg_state *ptr_reg,
11616 const struct bpf_reg_state *off_reg,
11617 struct bpf_reg_state *dst_reg,
11618 struct bpf_sanitize_info *info,
11619 const bool commit_window)
11621 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
11622 struct bpf_verifier_state *vstate = env->cur_state;
11623 bool off_is_imm = tnum_is_const(off_reg->var_off);
11624 bool off_is_neg = off_reg->smin_value < 0;
11625 bool ptr_is_dst_reg = ptr_reg == dst_reg;
11626 u8 opcode = BPF_OP(insn->code);
11627 u32 alu_state, alu_limit;
11628 struct bpf_reg_state tmp;
11632 if (can_skip_alu_sanitation(env, insn))
11635 /* We already marked aux for masking from non-speculative
11636 * paths, thus we got here in the first place. We only care
11637 * to explore bad access from here.
11639 if (vstate->speculative)
11642 if (!commit_window) {
11643 if (!tnum_is_const(off_reg->var_off) &&
11644 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
11645 return REASON_BOUNDS;
11647 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
11648 (opcode == BPF_SUB && !off_is_neg);
11651 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
11655 if (commit_window) {
11656 /* In commit phase we narrow the masking window based on
11657 * the observed pointer move after the simulated operation.
11659 alu_state = info->aux.alu_state;
11660 alu_limit = abs(info->aux.alu_limit - alu_limit);
11662 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
11663 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
11664 alu_state |= ptr_is_dst_reg ?
11665 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
11667 /* Limit pruning on unknown scalars to enable deep search for
11668 * potential masking differences from other program paths.
11671 env->explore_alu_limits = true;
11674 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
11678 /* If we're in commit phase, we're done here given we already
11679 * pushed the truncated dst_reg into the speculative verification
11682 * Also, when register is a known constant, we rewrite register-based
11683 * operation to immediate-based, and thus do not need masking (and as
11684 * a consequence, do not need to simulate the zero-truncation either).
11686 if (commit_window || off_is_imm)
11689 /* Simulate and find potential out-of-bounds access under
11690 * speculative execution from truncation as a result of
11691 * masking when off was not within expected range. If off
11692 * sits in dst, then we temporarily need to move ptr there
11693 * to simulate dst (== 0) +/-= ptr. Needed, for example,
11694 * for cases where we use K-based arithmetic in one direction
11695 * and truncated reg-based in the other in order to explore
11698 if (!ptr_is_dst_reg) {
11700 copy_register_state(dst_reg, ptr_reg);
11702 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
11704 if (!ptr_is_dst_reg && ret)
11706 return !ret ? REASON_STACK : 0;
11709 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
11711 struct bpf_verifier_state *vstate = env->cur_state;
11713 /* If we simulate paths under speculation, we don't update the
11714 * insn as 'seen' such that when we verify unreachable paths in
11715 * the non-speculative domain, sanitize_dead_code() can still
11716 * rewrite/sanitize them.
11718 if (!vstate->speculative)
11719 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
11722 static int sanitize_err(struct bpf_verifier_env *env,
11723 const struct bpf_insn *insn, int reason,
11724 const struct bpf_reg_state *off_reg,
11725 const struct bpf_reg_state *dst_reg)
11727 static const char *err = "pointer arithmetic with it prohibited for !root";
11728 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
11729 u32 dst = insn->dst_reg, src = insn->src_reg;
11732 case REASON_BOUNDS:
11733 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
11734 off_reg == dst_reg ? dst : src, err);
11737 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
11738 off_reg == dst_reg ? src : dst, err);
11741 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
11745 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
11749 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
11753 verbose(env, "verifier internal error: unknown reason (%d)\n",
11761 /* check that stack access falls within stack limits and that 'reg' doesn't
11762 * have a variable offset.
11764 * Variable offset is prohibited for unprivileged mode for simplicity since it
11765 * requires corresponding support in Spectre masking for stack ALU. See also
11766 * retrieve_ptr_limit().
11769 * 'off' includes 'reg->off'.
11771 static int check_stack_access_for_ptr_arithmetic(
11772 struct bpf_verifier_env *env,
11774 const struct bpf_reg_state *reg,
11777 if (!tnum_is_const(reg->var_off)) {
11780 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
11781 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
11782 regno, tn_buf, off);
11786 if (off >= 0 || off < -MAX_BPF_STACK) {
11787 verbose(env, "R%d stack pointer arithmetic goes out of range, "
11788 "prohibited for !root; off=%d\n", regno, off);
11795 static int sanitize_check_bounds(struct bpf_verifier_env *env,
11796 const struct bpf_insn *insn,
11797 const struct bpf_reg_state *dst_reg)
11799 u32 dst = insn->dst_reg;
11801 /* For unprivileged we require that resulting offset must be in bounds
11802 * in order to be able to sanitize access later on.
11804 if (env->bypass_spec_v1)
11807 switch (dst_reg->type) {
11809 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
11810 dst_reg->off + dst_reg->var_off.value))
11813 case PTR_TO_MAP_VALUE:
11814 if (check_map_access(env, dst, dst_reg->off, 1, false, ACCESS_HELPER)) {
11815 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
11816 "prohibited for !root\n", dst);
11827 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
11828 * Caller should also handle BPF_MOV case separately.
11829 * If we return -EACCES, caller may want to try again treating pointer as a
11830 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
11832 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
11833 struct bpf_insn *insn,
11834 const struct bpf_reg_state *ptr_reg,
11835 const struct bpf_reg_state *off_reg)
11837 struct bpf_verifier_state *vstate = env->cur_state;
11838 struct bpf_func_state *state = vstate->frame[vstate->curframe];
11839 struct bpf_reg_state *regs = state->regs, *dst_reg;
11840 bool known = tnum_is_const(off_reg->var_off);
11841 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
11842 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
11843 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
11844 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
11845 struct bpf_sanitize_info info = {};
11846 u8 opcode = BPF_OP(insn->code);
11847 u32 dst = insn->dst_reg;
11850 dst_reg = ®s[dst];
11852 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
11853 smin_val > smax_val || umin_val > umax_val) {
11854 /* Taint dst register if offset had invalid bounds derived from
11855 * e.g. dead branches.
11857 __mark_reg_unknown(env, dst_reg);
11861 if (BPF_CLASS(insn->code) != BPF_ALU64) {
11862 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
11863 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
11864 __mark_reg_unknown(env, dst_reg);
11869 "R%d 32-bit pointer arithmetic prohibited\n",
11874 if (ptr_reg->type & PTR_MAYBE_NULL) {
11875 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
11876 dst, reg_type_str(env, ptr_reg->type));
11880 switch (base_type(ptr_reg->type)) {
11881 case CONST_PTR_TO_MAP:
11882 /* smin_val represents the known value */
11883 if (known && smin_val == 0 && opcode == BPF_ADD)
11886 case PTR_TO_PACKET_END:
11887 case PTR_TO_SOCKET:
11888 case PTR_TO_SOCK_COMMON:
11889 case PTR_TO_TCP_SOCK:
11890 case PTR_TO_XDP_SOCK:
11891 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
11892 dst, reg_type_str(env, ptr_reg->type));
11898 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
11899 * The id may be overwritten later if we create a new variable offset.
11901 dst_reg->type = ptr_reg->type;
11902 dst_reg->id = ptr_reg->id;
11904 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
11905 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
11908 /* pointer types do not carry 32-bit bounds at the moment. */
11909 __mark_reg32_unbounded(dst_reg);
11911 if (sanitize_needed(opcode)) {
11912 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
11915 return sanitize_err(env, insn, ret, off_reg, dst_reg);
11920 /* We can take a fixed offset as long as it doesn't overflow
11921 * the s32 'off' field
11923 if (known && (ptr_reg->off + smin_val ==
11924 (s64)(s32)(ptr_reg->off + smin_val))) {
11925 /* pointer += K. Accumulate it into fixed offset */
11926 dst_reg->smin_value = smin_ptr;
11927 dst_reg->smax_value = smax_ptr;
11928 dst_reg->umin_value = umin_ptr;
11929 dst_reg->umax_value = umax_ptr;
11930 dst_reg->var_off = ptr_reg->var_off;
11931 dst_reg->off = ptr_reg->off + smin_val;
11932 dst_reg->raw = ptr_reg->raw;
11935 /* A new variable offset is created. Note that off_reg->off
11936 * == 0, since it's a scalar.
11937 * dst_reg gets the pointer type and since some positive
11938 * integer value was added to the pointer, give it a new 'id'
11939 * if it's a PTR_TO_PACKET.
11940 * this creates a new 'base' pointer, off_reg (variable) gets
11941 * added into the variable offset, and we copy the fixed offset
11944 if (signed_add_overflows(smin_ptr, smin_val) ||
11945 signed_add_overflows(smax_ptr, smax_val)) {
11946 dst_reg->smin_value = S64_MIN;
11947 dst_reg->smax_value = S64_MAX;
11949 dst_reg->smin_value = smin_ptr + smin_val;
11950 dst_reg->smax_value = smax_ptr + smax_val;
11952 if (umin_ptr + umin_val < umin_ptr ||
11953 umax_ptr + umax_val < umax_ptr) {
11954 dst_reg->umin_value = 0;
11955 dst_reg->umax_value = U64_MAX;
11957 dst_reg->umin_value = umin_ptr + umin_val;
11958 dst_reg->umax_value = umax_ptr + umax_val;
11960 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
11961 dst_reg->off = ptr_reg->off;
11962 dst_reg->raw = ptr_reg->raw;
11963 if (reg_is_pkt_pointer(ptr_reg)) {
11964 dst_reg->id = ++env->id_gen;
11965 /* something was added to pkt_ptr, set range to zero */
11966 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
11970 if (dst_reg == off_reg) {
11971 /* scalar -= pointer. Creates an unknown scalar */
11972 verbose(env, "R%d tried to subtract pointer from scalar\n",
11976 /* We don't allow subtraction from FP, because (according to
11977 * test_verifier.c test "invalid fp arithmetic", JITs might not
11978 * be able to deal with it.
11980 if (ptr_reg->type == PTR_TO_STACK) {
11981 verbose(env, "R%d subtraction from stack pointer prohibited\n",
11985 if (known && (ptr_reg->off - smin_val ==
11986 (s64)(s32)(ptr_reg->off - smin_val))) {
11987 /* pointer -= K. Subtract it from fixed offset */
11988 dst_reg->smin_value = smin_ptr;
11989 dst_reg->smax_value = smax_ptr;
11990 dst_reg->umin_value = umin_ptr;
11991 dst_reg->umax_value = umax_ptr;
11992 dst_reg->var_off = ptr_reg->var_off;
11993 dst_reg->id = ptr_reg->id;
11994 dst_reg->off = ptr_reg->off - smin_val;
11995 dst_reg->raw = ptr_reg->raw;
11998 /* A new variable offset is created. If the subtrahend is known
11999 * nonnegative, then any reg->range we had before is still good.
12001 if (signed_sub_overflows(smin_ptr, smax_val) ||
12002 signed_sub_overflows(smax_ptr, smin_val)) {
12003 /* Overflow possible, we know nothing */
12004 dst_reg->smin_value = S64_MIN;
12005 dst_reg->smax_value = S64_MAX;
12007 dst_reg->smin_value = smin_ptr - smax_val;
12008 dst_reg->smax_value = smax_ptr - smin_val;
12010 if (umin_ptr < umax_val) {
12011 /* Overflow possible, we know nothing */
12012 dst_reg->umin_value = 0;
12013 dst_reg->umax_value = U64_MAX;
12015 /* Cannot overflow (as long as bounds are consistent) */
12016 dst_reg->umin_value = umin_ptr - umax_val;
12017 dst_reg->umax_value = umax_ptr - umin_val;
12019 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
12020 dst_reg->off = ptr_reg->off;
12021 dst_reg->raw = ptr_reg->raw;
12022 if (reg_is_pkt_pointer(ptr_reg)) {
12023 dst_reg->id = ++env->id_gen;
12024 /* something was added to pkt_ptr, set range to zero */
12026 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
12032 /* bitwise ops on pointers are troublesome, prohibit. */
12033 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
12034 dst, bpf_alu_string[opcode >> 4]);
12037 /* other operators (e.g. MUL,LSH) produce non-pointer results */
12038 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
12039 dst, bpf_alu_string[opcode >> 4]);
12043 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
12045 reg_bounds_sync(dst_reg);
12046 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
12048 if (sanitize_needed(opcode)) {
12049 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
12052 return sanitize_err(env, insn, ret, off_reg, dst_reg);
12058 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
12059 struct bpf_reg_state *src_reg)
12061 s32 smin_val = src_reg->s32_min_value;
12062 s32 smax_val = src_reg->s32_max_value;
12063 u32 umin_val = src_reg->u32_min_value;
12064 u32 umax_val = src_reg->u32_max_value;
12066 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
12067 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
12068 dst_reg->s32_min_value = S32_MIN;
12069 dst_reg->s32_max_value = S32_MAX;
12071 dst_reg->s32_min_value += smin_val;
12072 dst_reg->s32_max_value += smax_val;
12074 if (dst_reg->u32_min_value + umin_val < umin_val ||
12075 dst_reg->u32_max_value + umax_val < umax_val) {
12076 dst_reg->u32_min_value = 0;
12077 dst_reg->u32_max_value = U32_MAX;
12079 dst_reg->u32_min_value += umin_val;
12080 dst_reg->u32_max_value += umax_val;
12084 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
12085 struct bpf_reg_state *src_reg)
12087 s64 smin_val = src_reg->smin_value;
12088 s64 smax_val = src_reg->smax_value;
12089 u64 umin_val = src_reg->umin_value;
12090 u64 umax_val = src_reg->umax_value;
12092 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
12093 signed_add_overflows(dst_reg->smax_value, smax_val)) {
12094 dst_reg->smin_value = S64_MIN;
12095 dst_reg->smax_value = S64_MAX;
12097 dst_reg->smin_value += smin_val;
12098 dst_reg->smax_value += smax_val;
12100 if (dst_reg->umin_value + umin_val < umin_val ||
12101 dst_reg->umax_value + umax_val < umax_val) {
12102 dst_reg->umin_value = 0;
12103 dst_reg->umax_value = U64_MAX;
12105 dst_reg->umin_value += umin_val;
12106 dst_reg->umax_value += umax_val;
12110 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
12111 struct bpf_reg_state *src_reg)
12113 s32 smin_val = src_reg->s32_min_value;
12114 s32 smax_val = src_reg->s32_max_value;
12115 u32 umin_val = src_reg->u32_min_value;
12116 u32 umax_val = src_reg->u32_max_value;
12118 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
12119 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
12120 /* Overflow possible, we know nothing */
12121 dst_reg->s32_min_value = S32_MIN;
12122 dst_reg->s32_max_value = S32_MAX;
12124 dst_reg->s32_min_value -= smax_val;
12125 dst_reg->s32_max_value -= smin_val;
12127 if (dst_reg->u32_min_value < umax_val) {
12128 /* Overflow possible, we know nothing */
12129 dst_reg->u32_min_value = 0;
12130 dst_reg->u32_max_value = U32_MAX;
12132 /* Cannot overflow (as long as bounds are consistent) */
12133 dst_reg->u32_min_value -= umax_val;
12134 dst_reg->u32_max_value -= umin_val;
12138 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
12139 struct bpf_reg_state *src_reg)
12141 s64 smin_val = src_reg->smin_value;
12142 s64 smax_val = src_reg->smax_value;
12143 u64 umin_val = src_reg->umin_value;
12144 u64 umax_val = src_reg->umax_value;
12146 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
12147 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
12148 /* Overflow possible, we know nothing */
12149 dst_reg->smin_value = S64_MIN;
12150 dst_reg->smax_value = S64_MAX;
12152 dst_reg->smin_value -= smax_val;
12153 dst_reg->smax_value -= smin_val;
12155 if (dst_reg->umin_value < umax_val) {
12156 /* Overflow possible, we know nothing */
12157 dst_reg->umin_value = 0;
12158 dst_reg->umax_value = U64_MAX;
12160 /* Cannot overflow (as long as bounds are consistent) */
12161 dst_reg->umin_value -= umax_val;
12162 dst_reg->umax_value -= umin_val;
12166 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
12167 struct bpf_reg_state *src_reg)
12169 s32 smin_val = src_reg->s32_min_value;
12170 u32 umin_val = src_reg->u32_min_value;
12171 u32 umax_val = src_reg->u32_max_value;
12173 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
12174 /* Ain't nobody got time to multiply that sign */
12175 __mark_reg32_unbounded(dst_reg);
12178 /* Both values are positive, so we can work with unsigned and
12179 * copy the result to signed (unless it exceeds S32_MAX).
12181 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
12182 /* Potential overflow, we know nothing */
12183 __mark_reg32_unbounded(dst_reg);
12186 dst_reg->u32_min_value *= umin_val;
12187 dst_reg->u32_max_value *= umax_val;
12188 if (dst_reg->u32_max_value > S32_MAX) {
12189 /* Overflow possible, we know nothing */
12190 dst_reg->s32_min_value = S32_MIN;
12191 dst_reg->s32_max_value = S32_MAX;
12193 dst_reg->s32_min_value = dst_reg->u32_min_value;
12194 dst_reg->s32_max_value = dst_reg->u32_max_value;
12198 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
12199 struct bpf_reg_state *src_reg)
12201 s64 smin_val = src_reg->smin_value;
12202 u64 umin_val = src_reg->umin_value;
12203 u64 umax_val = src_reg->umax_value;
12205 if (smin_val < 0 || dst_reg->smin_value < 0) {
12206 /* Ain't nobody got time to multiply that sign */
12207 __mark_reg64_unbounded(dst_reg);
12210 /* Both values are positive, so we can work with unsigned and
12211 * copy the result to signed (unless it exceeds S64_MAX).
12213 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
12214 /* Potential overflow, we know nothing */
12215 __mark_reg64_unbounded(dst_reg);
12218 dst_reg->umin_value *= umin_val;
12219 dst_reg->umax_value *= umax_val;
12220 if (dst_reg->umax_value > S64_MAX) {
12221 /* Overflow possible, we know nothing */
12222 dst_reg->smin_value = S64_MIN;
12223 dst_reg->smax_value = S64_MAX;
12225 dst_reg->smin_value = dst_reg->umin_value;
12226 dst_reg->smax_value = dst_reg->umax_value;
12230 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
12231 struct bpf_reg_state *src_reg)
12233 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12234 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12235 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12236 s32 smin_val = src_reg->s32_min_value;
12237 u32 umax_val = src_reg->u32_max_value;
12239 if (src_known && dst_known) {
12240 __mark_reg32_known(dst_reg, var32_off.value);
12244 /* We get our minimum from the var_off, since that's inherently
12245 * bitwise. Our maximum is the minimum of the operands' maxima.
12247 dst_reg->u32_min_value = var32_off.value;
12248 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
12249 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
12250 /* Lose signed bounds when ANDing negative numbers,
12251 * ain't nobody got time for that.
12253 dst_reg->s32_min_value = S32_MIN;
12254 dst_reg->s32_max_value = S32_MAX;
12256 /* ANDing two positives gives a positive, so safe to
12257 * cast result into s64.
12259 dst_reg->s32_min_value = dst_reg->u32_min_value;
12260 dst_reg->s32_max_value = dst_reg->u32_max_value;
12264 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
12265 struct bpf_reg_state *src_reg)
12267 bool src_known = tnum_is_const(src_reg->var_off);
12268 bool dst_known = tnum_is_const(dst_reg->var_off);
12269 s64 smin_val = src_reg->smin_value;
12270 u64 umax_val = src_reg->umax_value;
12272 if (src_known && dst_known) {
12273 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12277 /* We get our minimum from the var_off, since that's inherently
12278 * bitwise. Our maximum is the minimum of the operands' maxima.
12280 dst_reg->umin_value = dst_reg->var_off.value;
12281 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
12282 if (dst_reg->smin_value < 0 || smin_val < 0) {
12283 /* Lose signed bounds when ANDing negative numbers,
12284 * ain't nobody got time for that.
12286 dst_reg->smin_value = S64_MIN;
12287 dst_reg->smax_value = S64_MAX;
12289 /* ANDing two positives gives a positive, so safe to
12290 * cast result into s64.
12292 dst_reg->smin_value = dst_reg->umin_value;
12293 dst_reg->smax_value = dst_reg->umax_value;
12295 /* We may learn something more from the var_off */
12296 __update_reg_bounds(dst_reg);
12299 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
12300 struct bpf_reg_state *src_reg)
12302 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12303 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12304 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12305 s32 smin_val = src_reg->s32_min_value;
12306 u32 umin_val = src_reg->u32_min_value;
12308 if (src_known && dst_known) {
12309 __mark_reg32_known(dst_reg, var32_off.value);
12313 /* We get our maximum from the var_off, and our minimum is the
12314 * maximum of the operands' minima
12316 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
12317 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
12318 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
12319 /* Lose signed bounds when ORing negative numbers,
12320 * ain't nobody got time for that.
12322 dst_reg->s32_min_value = S32_MIN;
12323 dst_reg->s32_max_value = S32_MAX;
12325 /* ORing two positives gives a positive, so safe to
12326 * cast result into s64.
12328 dst_reg->s32_min_value = dst_reg->u32_min_value;
12329 dst_reg->s32_max_value = dst_reg->u32_max_value;
12333 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
12334 struct bpf_reg_state *src_reg)
12336 bool src_known = tnum_is_const(src_reg->var_off);
12337 bool dst_known = tnum_is_const(dst_reg->var_off);
12338 s64 smin_val = src_reg->smin_value;
12339 u64 umin_val = src_reg->umin_value;
12341 if (src_known && dst_known) {
12342 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12346 /* We get our maximum from the var_off, and our minimum is the
12347 * maximum of the operands' minima
12349 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
12350 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
12351 if (dst_reg->smin_value < 0 || smin_val < 0) {
12352 /* Lose signed bounds when ORing negative numbers,
12353 * ain't nobody got time for that.
12355 dst_reg->smin_value = S64_MIN;
12356 dst_reg->smax_value = S64_MAX;
12358 /* ORing two positives gives a positive, so safe to
12359 * cast result into s64.
12361 dst_reg->smin_value = dst_reg->umin_value;
12362 dst_reg->smax_value = dst_reg->umax_value;
12364 /* We may learn something more from the var_off */
12365 __update_reg_bounds(dst_reg);
12368 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
12369 struct bpf_reg_state *src_reg)
12371 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12372 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12373 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12374 s32 smin_val = src_reg->s32_min_value;
12376 if (src_known && dst_known) {
12377 __mark_reg32_known(dst_reg, var32_off.value);
12381 /* We get both minimum and maximum from the var32_off. */
12382 dst_reg->u32_min_value = var32_off.value;
12383 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
12385 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
12386 /* XORing two positive sign numbers gives a positive,
12387 * so safe to cast u32 result into s32.
12389 dst_reg->s32_min_value = dst_reg->u32_min_value;
12390 dst_reg->s32_max_value = dst_reg->u32_max_value;
12392 dst_reg->s32_min_value = S32_MIN;
12393 dst_reg->s32_max_value = S32_MAX;
12397 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
12398 struct bpf_reg_state *src_reg)
12400 bool src_known = tnum_is_const(src_reg->var_off);
12401 bool dst_known = tnum_is_const(dst_reg->var_off);
12402 s64 smin_val = src_reg->smin_value;
12404 if (src_known && dst_known) {
12405 /* dst_reg->var_off.value has been updated earlier */
12406 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12410 /* We get both minimum and maximum from the var_off. */
12411 dst_reg->umin_value = dst_reg->var_off.value;
12412 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
12414 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
12415 /* XORing two positive sign numbers gives a positive,
12416 * so safe to cast u64 result into s64.
12418 dst_reg->smin_value = dst_reg->umin_value;
12419 dst_reg->smax_value = dst_reg->umax_value;
12421 dst_reg->smin_value = S64_MIN;
12422 dst_reg->smax_value = S64_MAX;
12425 __update_reg_bounds(dst_reg);
12428 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
12429 u64 umin_val, u64 umax_val)
12431 /* We lose all sign bit information (except what we can pick
12434 dst_reg->s32_min_value = S32_MIN;
12435 dst_reg->s32_max_value = S32_MAX;
12436 /* If we might shift our top bit out, then we know nothing */
12437 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
12438 dst_reg->u32_min_value = 0;
12439 dst_reg->u32_max_value = U32_MAX;
12441 dst_reg->u32_min_value <<= umin_val;
12442 dst_reg->u32_max_value <<= umax_val;
12446 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
12447 struct bpf_reg_state *src_reg)
12449 u32 umax_val = src_reg->u32_max_value;
12450 u32 umin_val = src_reg->u32_min_value;
12451 /* u32 alu operation will zext upper bits */
12452 struct tnum subreg = tnum_subreg(dst_reg->var_off);
12454 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
12455 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
12456 /* Not required but being careful mark reg64 bounds as unknown so
12457 * that we are forced to pick them up from tnum and zext later and
12458 * if some path skips this step we are still safe.
12460 __mark_reg64_unbounded(dst_reg);
12461 __update_reg32_bounds(dst_reg);
12464 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
12465 u64 umin_val, u64 umax_val)
12467 /* Special case <<32 because it is a common compiler pattern to sign
12468 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
12469 * positive we know this shift will also be positive so we can track
12470 * bounds correctly. Otherwise we lose all sign bit information except
12471 * what we can pick up from var_off. Perhaps we can generalize this
12472 * later to shifts of any length.
12474 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
12475 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
12477 dst_reg->smax_value = S64_MAX;
12479 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
12480 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
12482 dst_reg->smin_value = S64_MIN;
12484 /* If we might shift our top bit out, then we know nothing */
12485 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
12486 dst_reg->umin_value = 0;
12487 dst_reg->umax_value = U64_MAX;
12489 dst_reg->umin_value <<= umin_val;
12490 dst_reg->umax_value <<= umax_val;
12494 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
12495 struct bpf_reg_state *src_reg)
12497 u64 umax_val = src_reg->umax_value;
12498 u64 umin_val = src_reg->umin_value;
12500 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
12501 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
12502 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
12504 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
12505 /* We may learn something more from the var_off */
12506 __update_reg_bounds(dst_reg);
12509 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
12510 struct bpf_reg_state *src_reg)
12512 struct tnum subreg = tnum_subreg(dst_reg->var_off);
12513 u32 umax_val = src_reg->u32_max_value;
12514 u32 umin_val = src_reg->u32_min_value;
12516 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
12517 * be negative, then either:
12518 * 1) src_reg might be zero, so the sign bit of the result is
12519 * unknown, so we lose our signed bounds
12520 * 2) it's known negative, thus the unsigned bounds capture the
12522 * 3) the signed bounds cross zero, so they tell us nothing
12524 * If the value in dst_reg is known nonnegative, then again the
12525 * unsigned bounds capture the signed bounds.
12526 * Thus, in all cases it suffices to blow away our signed bounds
12527 * and rely on inferring new ones from the unsigned bounds and
12528 * var_off of the result.
12530 dst_reg->s32_min_value = S32_MIN;
12531 dst_reg->s32_max_value = S32_MAX;
12533 dst_reg->var_off = tnum_rshift(subreg, umin_val);
12534 dst_reg->u32_min_value >>= umax_val;
12535 dst_reg->u32_max_value >>= umin_val;
12537 __mark_reg64_unbounded(dst_reg);
12538 __update_reg32_bounds(dst_reg);
12541 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
12542 struct bpf_reg_state *src_reg)
12544 u64 umax_val = src_reg->umax_value;
12545 u64 umin_val = src_reg->umin_value;
12547 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
12548 * be negative, then either:
12549 * 1) src_reg might be zero, so the sign bit of the result is
12550 * unknown, so we lose our signed bounds
12551 * 2) it's known negative, thus the unsigned bounds capture the
12553 * 3) the signed bounds cross zero, so they tell us nothing
12555 * If the value in dst_reg is known nonnegative, then again the
12556 * unsigned bounds capture the signed bounds.
12557 * Thus, in all cases it suffices to blow away our signed bounds
12558 * and rely on inferring new ones from the unsigned bounds and
12559 * var_off of the result.
12561 dst_reg->smin_value = S64_MIN;
12562 dst_reg->smax_value = S64_MAX;
12563 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
12564 dst_reg->umin_value >>= umax_val;
12565 dst_reg->umax_value >>= umin_val;
12567 /* Its not easy to operate on alu32 bounds here because it depends
12568 * on bits being shifted in. Take easy way out and mark unbounded
12569 * so we can recalculate later from tnum.
12571 __mark_reg32_unbounded(dst_reg);
12572 __update_reg_bounds(dst_reg);
12575 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
12576 struct bpf_reg_state *src_reg)
12578 u64 umin_val = src_reg->u32_min_value;
12580 /* Upon reaching here, src_known is true and
12581 * umax_val is equal to umin_val.
12583 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
12584 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
12586 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
12588 /* blow away the dst_reg umin_value/umax_value and rely on
12589 * dst_reg var_off to refine the result.
12591 dst_reg->u32_min_value = 0;
12592 dst_reg->u32_max_value = U32_MAX;
12594 __mark_reg64_unbounded(dst_reg);
12595 __update_reg32_bounds(dst_reg);
12598 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
12599 struct bpf_reg_state *src_reg)
12601 u64 umin_val = src_reg->umin_value;
12603 /* Upon reaching here, src_known is true and umax_val is equal
12606 dst_reg->smin_value >>= umin_val;
12607 dst_reg->smax_value >>= umin_val;
12609 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
12611 /* blow away the dst_reg umin_value/umax_value and rely on
12612 * dst_reg var_off to refine the result.
12614 dst_reg->umin_value = 0;
12615 dst_reg->umax_value = U64_MAX;
12617 /* Its not easy to operate on alu32 bounds here because it depends
12618 * on bits being shifted in from upper 32-bits. Take easy way out
12619 * and mark unbounded so we can recalculate later from tnum.
12621 __mark_reg32_unbounded(dst_reg);
12622 __update_reg_bounds(dst_reg);
12625 /* WARNING: This function does calculations on 64-bit values, but the actual
12626 * execution may occur on 32-bit values. Therefore, things like bitshifts
12627 * need extra checks in the 32-bit case.
12629 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
12630 struct bpf_insn *insn,
12631 struct bpf_reg_state *dst_reg,
12632 struct bpf_reg_state src_reg)
12634 struct bpf_reg_state *regs = cur_regs(env);
12635 u8 opcode = BPF_OP(insn->code);
12637 s64 smin_val, smax_val;
12638 u64 umin_val, umax_val;
12639 s32 s32_min_val, s32_max_val;
12640 u32 u32_min_val, u32_max_val;
12641 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
12642 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
12645 smin_val = src_reg.smin_value;
12646 smax_val = src_reg.smax_value;
12647 umin_val = src_reg.umin_value;
12648 umax_val = src_reg.umax_value;
12650 s32_min_val = src_reg.s32_min_value;
12651 s32_max_val = src_reg.s32_max_value;
12652 u32_min_val = src_reg.u32_min_value;
12653 u32_max_val = src_reg.u32_max_value;
12656 src_known = tnum_subreg_is_const(src_reg.var_off);
12658 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
12659 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
12660 /* Taint dst register if offset had invalid bounds
12661 * derived from e.g. dead branches.
12663 __mark_reg_unknown(env, dst_reg);
12667 src_known = tnum_is_const(src_reg.var_off);
12669 (smin_val != smax_val || umin_val != umax_val)) ||
12670 smin_val > smax_val || umin_val > umax_val) {
12671 /* Taint dst register if offset had invalid bounds
12672 * derived from e.g. dead branches.
12674 __mark_reg_unknown(env, dst_reg);
12680 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
12681 __mark_reg_unknown(env, dst_reg);
12685 if (sanitize_needed(opcode)) {
12686 ret = sanitize_val_alu(env, insn);
12688 return sanitize_err(env, insn, ret, NULL, NULL);
12691 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
12692 * There are two classes of instructions: The first class we track both
12693 * alu32 and alu64 sign/unsigned bounds independently this provides the
12694 * greatest amount of precision when alu operations are mixed with jmp32
12695 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
12696 * and BPF_OR. This is possible because these ops have fairly easy to
12697 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
12698 * See alu32 verifier tests for examples. The second class of
12699 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
12700 * with regards to tracking sign/unsigned bounds because the bits may
12701 * cross subreg boundaries in the alu64 case. When this happens we mark
12702 * the reg unbounded in the subreg bound space and use the resulting
12703 * tnum to calculate an approximation of the sign/unsigned bounds.
12707 scalar32_min_max_add(dst_reg, &src_reg);
12708 scalar_min_max_add(dst_reg, &src_reg);
12709 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
12712 scalar32_min_max_sub(dst_reg, &src_reg);
12713 scalar_min_max_sub(dst_reg, &src_reg);
12714 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
12717 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
12718 scalar32_min_max_mul(dst_reg, &src_reg);
12719 scalar_min_max_mul(dst_reg, &src_reg);
12722 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
12723 scalar32_min_max_and(dst_reg, &src_reg);
12724 scalar_min_max_and(dst_reg, &src_reg);
12727 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
12728 scalar32_min_max_or(dst_reg, &src_reg);
12729 scalar_min_max_or(dst_reg, &src_reg);
12732 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
12733 scalar32_min_max_xor(dst_reg, &src_reg);
12734 scalar_min_max_xor(dst_reg, &src_reg);
12737 if (umax_val >= insn_bitness) {
12738 /* Shifts greater than 31 or 63 are undefined.
12739 * This includes shifts by a negative number.
12741 mark_reg_unknown(env, regs, insn->dst_reg);
12745 scalar32_min_max_lsh(dst_reg, &src_reg);
12747 scalar_min_max_lsh(dst_reg, &src_reg);
12750 if (umax_val >= insn_bitness) {
12751 /* Shifts greater than 31 or 63 are undefined.
12752 * This includes shifts by a negative number.
12754 mark_reg_unknown(env, regs, insn->dst_reg);
12758 scalar32_min_max_rsh(dst_reg, &src_reg);
12760 scalar_min_max_rsh(dst_reg, &src_reg);
12763 if (umax_val >= insn_bitness) {
12764 /* Shifts greater than 31 or 63 are undefined.
12765 * This includes shifts by a negative number.
12767 mark_reg_unknown(env, regs, insn->dst_reg);
12771 scalar32_min_max_arsh(dst_reg, &src_reg);
12773 scalar_min_max_arsh(dst_reg, &src_reg);
12776 mark_reg_unknown(env, regs, insn->dst_reg);
12780 /* ALU32 ops are zero extended into 64bit register */
12782 zext_32_to_64(dst_reg);
12783 reg_bounds_sync(dst_reg);
12787 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
12790 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
12791 struct bpf_insn *insn)
12793 struct bpf_verifier_state *vstate = env->cur_state;
12794 struct bpf_func_state *state = vstate->frame[vstate->curframe];
12795 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
12796 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
12797 u8 opcode = BPF_OP(insn->code);
12800 dst_reg = ®s[insn->dst_reg];
12802 if (dst_reg->type != SCALAR_VALUE)
12805 /* Make sure ID is cleared otherwise dst_reg min/max could be
12806 * incorrectly propagated into other registers by find_equal_scalars()
12809 if (BPF_SRC(insn->code) == BPF_X) {
12810 src_reg = ®s[insn->src_reg];
12811 if (src_reg->type != SCALAR_VALUE) {
12812 if (dst_reg->type != SCALAR_VALUE) {
12813 /* Combining two pointers by any ALU op yields
12814 * an arbitrary scalar. Disallow all math except
12815 * pointer subtraction
12817 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
12818 mark_reg_unknown(env, regs, insn->dst_reg);
12821 verbose(env, "R%d pointer %s pointer prohibited\n",
12823 bpf_alu_string[opcode >> 4]);
12826 /* scalar += pointer
12827 * This is legal, but we have to reverse our
12828 * src/dest handling in computing the range
12830 err = mark_chain_precision(env, insn->dst_reg);
12833 return adjust_ptr_min_max_vals(env, insn,
12836 } else if (ptr_reg) {
12837 /* pointer += scalar */
12838 err = mark_chain_precision(env, insn->src_reg);
12841 return adjust_ptr_min_max_vals(env, insn,
12843 } else if (dst_reg->precise) {
12844 /* if dst_reg is precise, src_reg should be precise as well */
12845 err = mark_chain_precision(env, insn->src_reg);
12850 /* Pretend the src is a reg with a known value, since we only
12851 * need to be able to read from this state.
12853 off_reg.type = SCALAR_VALUE;
12854 __mark_reg_known(&off_reg, insn->imm);
12855 src_reg = &off_reg;
12856 if (ptr_reg) /* pointer += K */
12857 return adjust_ptr_min_max_vals(env, insn,
12861 /* Got here implies adding two SCALAR_VALUEs */
12862 if (WARN_ON_ONCE(ptr_reg)) {
12863 print_verifier_state(env, state, true);
12864 verbose(env, "verifier internal error: unexpected ptr_reg\n");
12867 if (WARN_ON(!src_reg)) {
12868 print_verifier_state(env, state, true);
12869 verbose(env, "verifier internal error: no src_reg\n");
12872 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
12875 /* check validity of 32-bit and 64-bit arithmetic operations */
12876 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
12878 struct bpf_reg_state *regs = cur_regs(env);
12879 u8 opcode = BPF_OP(insn->code);
12882 if (opcode == BPF_END || opcode == BPF_NEG) {
12883 if (opcode == BPF_NEG) {
12884 if (BPF_SRC(insn->code) != BPF_K ||
12885 insn->src_reg != BPF_REG_0 ||
12886 insn->off != 0 || insn->imm != 0) {
12887 verbose(env, "BPF_NEG uses reserved fields\n");
12891 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
12892 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
12893 BPF_CLASS(insn->code) == BPF_ALU64) {
12894 verbose(env, "BPF_END uses reserved fields\n");
12899 /* check src operand */
12900 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
12904 if (is_pointer_value(env, insn->dst_reg)) {
12905 verbose(env, "R%d pointer arithmetic prohibited\n",
12910 /* check dest operand */
12911 err = check_reg_arg(env, insn->dst_reg, DST_OP);
12915 } else if (opcode == BPF_MOV) {
12917 if (BPF_SRC(insn->code) == BPF_X) {
12918 if (insn->imm != 0 || insn->off != 0) {
12919 verbose(env, "BPF_MOV uses reserved fields\n");
12923 /* check src operand */
12924 err = check_reg_arg(env, insn->src_reg, SRC_OP);
12928 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
12929 verbose(env, "BPF_MOV uses reserved fields\n");
12934 /* check dest operand, mark as required later */
12935 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
12939 if (BPF_SRC(insn->code) == BPF_X) {
12940 struct bpf_reg_state *src_reg = regs + insn->src_reg;
12941 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
12942 bool need_id = src_reg->type == SCALAR_VALUE && !src_reg->id &&
12943 !tnum_is_const(src_reg->var_off);
12945 if (BPF_CLASS(insn->code) == BPF_ALU64) {
12947 * copy register state to dest reg
12950 /* Assign src and dst registers the same ID
12951 * that will be used by find_equal_scalars()
12952 * to propagate min/max range.
12954 src_reg->id = ++env->id_gen;
12955 copy_register_state(dst_reg, src_reg);
12956 dst_reg->live |= REG_LIVE_WRITTEN;
12957 dst_reg->subreg_def = DEF_NOT_SUBREG;
12959 /* R1 = (u32) R2 */
12960 if (is_pointer_value(env, insn->src_reg)) {
12962 "R%d partial copy of pointer\n",
12965 } else if (src_reg->type == SCALAR_VALUE) {
12966 bool is_src_reg_u32 = src_reg->umax_value <= U32_MAX;
12968 if (is_src_reg_u32 && need_id)
12969 src_reg->id = ++env->id_gen;
12970 copy_register_state(dst_reg, src_reg);
12971 /* Make sure ID is cleared if src_reg is not in u32 range otherwise
12972 * dst_reg min/max could be incorrectly
12973 * propagated into src_reg by find_equal_scalars()
12975 if (!is_src_reg_u32)
12977 dst_reg->live |= REG_LIVE_WRITTEN;
12978 dst_reg->subreg_def = env->insn_idx + 1;
12980 mark_reg_unknown(env, regs,
12983 zext_32_to_64(dst_reg);
12984 reg_bounds_sync(dst_reg);
12988 * remember the value we stored into this reg
12990 /* clear any state __mark_reg_known doesn't set */
12991 mark_reg_unknown(env, regs, insn->dst_reg);
12992 regs[insn->dst_reg].type = SCALAR_VALUE;
12993 if (BPF_CLASS(insn->code) == BPF_ALU64) {
12994 __mark_reg_known(regs + insn->dst_reg,
12997 __mark_reg_known(regs + insn->dst_reg,
13002 } else if (opcode > BPF_END) {
13003 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
13006 } else { /* all other ALU ops: and, sub, xor, add, ... */
13008 if (BPF_SRC(insn->code) == BPF_X) {
13009 if (insn->imm != 0 || insn->off != 0) {
13010 verbose(env, "BPF_ALU uses reserved fields\n");
13013 /* check src1 operand */
13014 err = check_reg_arg(env, insn->src_reg, SRC_OP);
13018 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
13019 verbose(env, "BPF_ALU uses reserved fields\n");
13024 /* check src2 operand */
13025 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
13029 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
13030 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
13031 verbose(env, "div by zero\n");
13035 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
13036 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
13037 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
13039 if (insn->imm < 0 || insn->imm >= size) {
13040 verbose(env, "invalid shift %d\n", insn->imm);
13045 /* check dest operand */
13046 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
13050 return adjust_reg_min_max_vals(env, insn);
13056 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
13057 struct bpf_reg_state *dst_reg,
13058 enum bpf_reg_type type,
13059 bool range_right_open)
13061 struct bpf_func_state *state;
13062 struct bpf_reg_state *reg;
13065 if (dst_reg->off < 0 ||
13066 (dst_reg->off == 0 && range_right_open))
13067 /* This doesn't give us any range */
13070 if (dst_reg->umax_value > MAX_PACKET_OFF ||
13071 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
13072 /* Risk of overflow. For instance, ptr + (1<<63) may be less
13073 * than pkt_end, but that's because it's also less than pkt.
13077 new_range = dst_reg->off;
13078 if (range_right_open)
13081 /* Examples for register markings:
13083 * pkt_data in dst register:
13087 * if (r2 > pkt_end) goto <handle exception>
13092 * if (r2 < pkt_end) goto <access okay>
13093 * <handle exception>
13096 * r2 == dst_reg, pkt_end == src_reg
13097 * r2=pkt(id=n,off=8,r=0)
13098 * r3=pkt(id=n,off=0,r=0)
13100 * pkt_data in src register:
13104 * if (pkt_end >= r2) goto <access okay>
13105 * <handle exception>
13109 * if (pkt_end <= r2) goto <handle exception>
13113 * pkt_end == dst_reg, r2 == src_reg
13114 * r2=pkt(id=n,off=8,r=0)
13115 * r3=pkt(id=n,off=0,r=0)
13117 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
13118 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
13119 * and [r3, r3 + 8-1) respectively is safe to access depending on
13123 /* If our ids match, then we must have the same max_value. And we
13124 * don't care about the other reg's fixed offset, since if it's too big
13125 * the range won't allow anything.
13126 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
13128 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
13129 if (reg->type == type && reg->id == dst_reg->id)
13130 /* keep the maximum range already checked */
13131 reg->range = max(reg->range, new_range);
13135 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
13137 struct tnum subreg = tnum_subreg(reg->var_off);
13138 s32 sval = (s32)val;
13142 if (tnum_is_const(subreg))
13143 return !!tnum_equals_const(subreg, val);
13144 else if (val < reg->u32_min_value || val > reg->u32_max_value)
13148 if (tnum_is_const(subreg))
13149 return !tnum_equals_const(subreg, val);
13150 else if (val < reg->u32_min_value || val > reg->u32_max_value)
13154 if ((~subreg.mask & subreg.value) & val)
13156 if (!((subreg.mask | subreg.value) & val))
13160 if (reg->u32_min_value > val)
13162 else if (reg->u32_max_value <= val)
13166 if (reg->s32_min_value > sval)
13168 else if (reg->s32_max_value <= sval)
13172 if (reg->u32_max_value < val)
13174 else if (reg->u32_min_value >= val)
13178 if (reg->s32_max_value < sval)
13180 else if (reg->s32_min_value >= sval)
13184 if (reg->u32_min_value >= val)
13186 else if (reg->u32_max_value < val)
13190 if (reg->s32_min_value >= sval)
13192 else if (reg->s32_max_value < sval)
13196 if (reg->u32_max_value <= val)
13198 else if (reg->u32_min_value > val)
13202 if (reg->s32_max_value <= sval)
13204 else if (reg->s32_min_value > sval)
13213 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
13215 s64 sval = (s64)val;
13219 if (tnum_is_const(reg->var_off))
13220 return !!tnum_equals_const(reg->var_off, val);
13221 else if (val < reg->umin_value || val > reg->umax_value)
13225 if (tnum_is_const(reg->var_off))
13226 return !tnum_equals_const(reg->var_off, val);
13227 else if (val < reg->umin_value || val > reg->umax_value)
13231 if ((~reg->var_off.mask & reg->var_off.value) & val)
13233 if (!((reg->var_off.mask | reg->var_off.value) & val))
13237 if (reg->umin_value > val)
13239 else if (reg->umax_value <= val)
13243 if (reg->smin_value > sval)
13245 else if (reg->smax_value <= sval)
13249 if (reg->umax_value < val)
13251 else if (reg->umin_value >= val)
13255 if (reg->smax_value < sval)
13257 else if (reg->smin_value >= sval)
13261 if (reg->umin_value >= val)
13263 else if (reg->umax_value < val)
13267 if (reg->smin_value >= sval)
13269 else if (reg->smax_value < sval)
13273 if (reg->umax_value <= val)
13275 else if (reg->umin_value > val)
13279 if (reg->smax_value <= sval)
13281 else if (reg->smin_value > sval)
13289 /* compute branch direction of the expression "if (reg opcode val) goto target;"
13291 * 1 - branch will be taken and "goto target" will be executed
13292 * 0 - branch will not be taken and fall-through to next insn
13293 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
13296 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
13299 if (__is_pointer_value(false, reg)) {
13300 if (!reg_not_null(reg))
13303 /* If pointer is valid tests against zero will fail so we can
13304 * use this to direct branch taken.
13320 return is_branch32_taken(reg, val, opcode);
13321 return is_branch64_taken(reg, val, opcode);
13324 static int flip_opcode(u32 opcode)
13326 /* How can we transform "a <op> b" into "b <op> a"? */
13327 static const u8 opcode_flip[16] = {
13328 /* these stay the same */
13329 [BPF_JEQ >> 4] = BPF_JEQ,
13330 [BPF_JNE >> 4] = BPF_JNE,
13331 [BPF_JSET >> 4] = BPF_JSET,
13332 /* these swap "lesser" and "greater" (L and G in the opcodes) */
13333 [BPF_JGE >> 4] = BPF_JLE,
13334 [BPF_JGT >> 4] = BPF_JLT,
13335 [BPF_JLE >> 4] = BPF_JGE,
13336 [BPF_JLT >> 4] = BPF_JGT,
13337 [BPF_JSGE >> 4] = BPF_JSLE,
13338 [BPF_JSGT >> 4] = BPF_JSLT,
13339 [BPF_JSLE >> 4] = BPF_JSGE,
13340 [BPF_JSLT >> 4] = BPF_JSGT
13342 return opcode_flip[opcode >> 4];
13345 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
13346 struct bpf_reg_state *src_reg,
13349 struct bpf_reg_state *pkt;
13351 if (src_reg->type == PTR_TO_PACKET_END) {
13353 } else if (dst_reg->type == PTR_TO_PACKET_END) {
13355 opcode = flip_opcode(opcode);
13360 if (pkt->range >= 0)
13365 /* pkt <= pkt_end */
13368 /* pkt > pkt_end */
13369 if (pkt->range == BEYOND_PKT_END)
13370 /* pkt has at last one extra byte beyond pkt_end */
13371 return opcode == BPF_JGT;
13374 /* pkt < pkt_end */
13377 /* pkt >= pkt_end */
13378 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
13379 return opcode == BPF_JGE;
13385 /* Adjusts the register min/max values in the case that the dst_reg is the
13386 * variable register that we are working on, and src_reg is a constant or we're
13387 * simply doing a BPF_K check.
13388 * In JEQ/JNE cases we also adjust the var_off values.
13390 static void reg_set_min_max(struct bpf_reg_state *true_reg,
13391 struct bpf_reg_state *false_reg,
13392 u64 val, u32 val32,
13393 u8 opcode, bool is_jmp32)
13395 struct tnum false_32off = tnum_subreg(false_reg->var_off);
13396 struct tnum false_64off = false_reg->var_off;
13397 struct tnum true_32off = tnum_subreg(true_reg->var_off);
13398 struct tnum true_64off = true_reg->var_off;
13399 s64 sval = (s64)val;
13400 s32 sval32 = (s32)val32;
13402 /* If the dst_reg is a pointer, we can't learn anything about its
13403 * variable offset from the compare (unless src_reg were a pointer into
13404 * the same object, but we don't bother with that.
13405 * Since false_reg and true_reg have the same type by construction, we
13406 * only need to check one of them for pointerness.
13408 if (__is_pointer_value(false, false_reg))
13412 /* JEQ/JNE comparison doesn't change the register equivalence.
13415 * if (r1 == 42) goto label;
13417 * label: // here both r1 and r2 are known to be 42.
13419 * Hence when marking register as known preserve it's ID.
13423 __mark_reg32_known(true_reg, val32);
13424 true_32off = tnum_subreg(true_reg->var_off);
13426 ___mark_reg_known(true_reg, val);
13427 true_64off = true_reg->var_off;
13432 __mark_reg32_known(false_reg, val32);
13433 false_32off = tnum_subreg(false_reg->var_off);
13435 ___mark_reg_known(false_reg, val);
13436 false_64off = false_reg->var_off;
13441 false_32off = tnum_and(false_32off, tnum_const(~val32));
13442 if (is_power_of_2(val32))
13443 true_32off = tnum_or(true_32off,
13444 tnum_const(val32));
13446 false_64off = tnum_and(false_64off, tnum_const(~val));
13447 if (is_power_of_2(val))
13448 true_64off = tnum_or(true_64off,
13456 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
13457 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
13459 false_reg->u32_max_value = min(false_reg->u32_max_value,
13461 true_reg->u32_min_value = max(true_reg->u32_min_value,
13464 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
13465 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
13467 false_reg->umax_value = min(false_reg->umax_value, false_umax);
13468 true_reg->umin_value = max(true_reg->umin_value, true_umin);
13476 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
13477 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
13479 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
13480 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
13482 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
13483 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
13485 false_reg->smax_value = min(false_reg->smax_value, false_smax);
13486 true_reg->smin_value = max(true_reg->smin_value, true_smin);
13494 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
13495 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
13497 false_reg->u32_min_value = max(false_reg->u32_min_value,
13499 true_reg->u32_max_value = min(true_reg->u32_max_value,
13502 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
13503 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
13505 false_reg->umin_value = max(false_reg->umin_value, false_umin);
13506 true_reg->umax_value = min(true_reg->umax_value, true_umax);
13514 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
13515 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
13517 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
13518 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
13520 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
13521 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
13523 false_reg->smin_value = max(false_reg->smin_value, false_smin);
13524 true_reg->smax_value = min(true_reg->smax_value, true_smax);
13533 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
13534 tnum_subreg(false_32off));
13535 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
13536 tnum_subreg(true_32off));
13537 __reg_combine_32_into_64(false_reg);
13538 __reg_combine_32_into_64(true_reg);
13540 false_reg->var_off = false_64off;
13541 true_reg->var_off = true_64off;
13542 __reg_combine_64_into_32(false_reg);
13543 __reg_combine_64_into_32(true_reg);
13547 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
13548 * the variable reg.
13550 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
13551 struct bpf_reg_state *false_reg,
13552 u64 val, u32 val32,
13553 u8 opcode, bool is_jmp32)
13555 opcode = flip_opcode(opcode);
13556 /* This uses zero as "not present in table"; luckily the zero opcode,
13557 * BPF_JA, can't get here.
13560 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
13563 /* Regs are known to be equal, so intersect their min/max/var_off */
13564 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
13565 struct bpf_reg_state *dst_reg)
13567 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
13568 dst_reg->umin_value);
13569 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
13570 dst_reg->umax_value);
13571 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
13572 dst_reg->smin_value);
13573 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
13574 dst_reg->smax_value);
13575 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
13577 reg_bounds_sync(src_reg);
13578 reg_bounds_sync(dst_reg);
13581 static void reg_combine_min_max(struct bpf_reg_state *true_src,
13582 struct bpf_reg_state *true_dst,
13583 struct bpf_reg_state *false_src,
13584 struct bpf_reg_state *false_dst,
13589 __reg_combine_min_max(true_src, true_dst);
13592 __reg_combine_min_max(false_src, false_dst);
13597 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
13598 struct bpf_reg_state *reg, u32 id,
13601 if (type_may_be_null(reg->type) && reg->id == id &&
13602 (is_rcu_reg(reg) || !WARN_ON_ONCE(!reg->id))) {
13603 /* Old offset (both fixed and variable parts) should have been
13604 * known-zero, because we don't allow pointer arithmetic on
13605 * pointers that might be NULL. If we see this happening, don't
13606 * convert the register.
13608 * But in some cases, some helpers that return local kptrs
13609 * advance offset for the returned pointer. In those cases, it
13610 * is fine to expect to see reg->off.
13612 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value || !tnum_equals_const(reg->var_off, 0)))
13614 if (!(type_is_ptr_alloc_obj(reg->type) || type_is_non_owning_ref(reg->type)) &&
13615 WARN_ON_ONCE(reg->off))
13619 reg->type = SCALAR_VALUE;
13620 /* We don't need id and ref_obj_id from this point
13621 * onwards anymore, thus we should better reset it,
13622 * so that state pruning has chances to take effect.
13625 reg->ref_obj_id = 0;
13630 mark_ptr_not_null_reg(reg);
13632 if (!reg_may_point_to_spin_lock(reg)) {
13633 /* For not-NULL ptr, reg->ref_obj_id will be reset
13634 * in release_reference().
13636 * reg->id is still used by spin_lock ptr. Other
13637 * than spin_lock ptr type, reg->id can be reset.
13644 /* The logic is similar to find_good_pkt_pointers(), both could eventually
13645 * be folded together at some point.
13647 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
13650 struct bpf_func_state *state = vstate->frame[vstate->curframe];
13651 struct bpf_reg_state *regs = state->regs, *reg;
13652 u32 ref_obj_id = regs[regno].ref_obj_id;
13653 u32 id = regs[regno].id;
13655 if (ref_obj_id && ref_obj_id == id && is_null)
13656 /* regs[regno] is in the " == NULL" branch.
13657 * No one could have freed the reference state before
13658 * doing the NULL check.
13660 WARN_ON_ONCE(release_reference_state(state, id));
13662 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
13663 mark_ptr_or_null_reg(state, reg, id, is_null);
13667 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
13668 struct bpf_reg_state *dst_reg,
13669 struct bpf_reg_state *src_reg,
13670 struct bpf_verifier_state *this_branch,
13671 struct bpf_verifier_state *other_branch)
13673 if (BPF_SRC(insn->code) != BPF_X)
13676 /* Pointers are always 64-bit. */
13677 if (BPF_CLASS(insn->code) == BPF_JMP32)
13680 switch (BPF_OP(insn->code)) {
13682 if ((dst_reg->type == PTR_TO_PACKET &&
13683 src_reg->type == PTR_TO_PACKET_END) ||
13684 (dst_reg->type == PTR_TO_PACKET_META &&
13685 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13686 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
13687 find_good_pkt_pointers(this_branch, dst_reg,
13688 dst_reg->type, false);
13689 mark_pkt_end(other_branch, insn->dst_reg, true);
13690 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
13691 src_reg->type == PTR_TO_PACKET) ||
13692 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
13693 src_reg->type == PTR_TO_PACKET_META)) {
13694 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
13695 find_good_pkt_pointers(other_branch, src_reg,
13696 src_reg->type, true);
13697 mark_pkt_end(this_branch, insn->src_reg, false);
13703 if ((dst_reg->type == PTR_TO_PACKET &&
13704 src_reg->type == PTR_TO_PACKET_END) ||
13705 (dst_reg->type == PTR_TO_PACKET_META &&
13706 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13707 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
13708 find_good_pkt_pointers(other_branch, dst_reg,
13709 dst_reg->type, true);
13710 mark_pkt_end(this_branch, insn->dst_reg, false);
13711 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
13712 src_reg->type == PTR_TO_PACKET) ||
13713 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
13714 src_reg->type == PTR_TO_PACKET_META)) {
13715 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
13716 find_good_pkt_pointers(this_branch, src_reg,
13717 src_reg->type, false);
13718 mark_pkt_end(other_branch, insn->src_reg, true);
13724 if ((dst_reg->type == PTR_TO_PACKET &&
13725 src_reg->type == PTR_TO_PACKET_END) ||
13726 (dst_reg->type == PTR_TO_PACKET_META &&
13727 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13728 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
13729 find_good_pkt_pointers(this_branch, dst_reg,
13730 dst_reg->type, true);
13731 mark_pkt_end(other_branch, insn->dst_reg, false);
13732 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
13733 src_reg->type == PTR_TO_PACKET) ||
13734 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
13735 src_reg->type == PTR_TO_PACKET_META)) {
13736 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
13737 find_good_pkt_pointers(other_branch, src_reg,
13738 src_reg->type, false);
13739 mark_pkt_end(this_branch, insn->src_reg, true);
13745 if ((dst_reg->type == PTR_TO_PACKET &&
13746 src_reg->type == PTR_TO_PACKET_END) ||
13747 (dst_reg->type == PTR_TO_PACKET_META &&
13748 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13749 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
13750 find_good_pkt_pointers(other_branch, dst_reg,
13751 dst_reg->type, false);
13752 mark_pkt_end(this_branch, insn->dst_reg, true);
13753 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
13754 src_reg->type == PTR_TO_PACKET) ||
13755 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
13756 src_reg->type == PTR_TO_PACKET_META)) {
13757 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
13758 find_good_pkt_pointers(this_branch, src_reg,
13759 src_reg->type, true);
13760 mark_pkt_end(other_branch, insn->src_reg, false);
13772 static void find_equal_scalars(struct bpf_verifier_state *vstate,
13773 struct bpf_reg_state *known_reg)
13775 struct bpf_func_state *state;
13776 struct bpf_reg_state *reg;
13778 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
13779 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
13780 copy_register_state(reg, known_reg);
13784 static int check_cond_jmp_op(struct bpf_verifier_env *env,
13785 struct bpf_insn *insn, int *insn_idx)
13787 struct bpf_verifier_state *this_branch = env->cur_state;
13788 struct bpf_verifier_state *other_branch;
13789 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
13790 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
13791 struct bpf_reg_state *eq_branch_regs;
13792 u8 opcode = BPF_OP(insn->code);
13797 /* Only conditional jumps are expected to reach here. */
13798 if (opcode == BPF_JA || opcode > BPF_JSLE) {
13799 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
13803 if (BPF_SRC(insn->code) == BPF_X) {
13804 if (insn->imm != 0) {
13805 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
13809 /* check src1 operand */
13810 err = check_reg_arg(env, insn->src_reg, SRC_OP);
13814 if (is_pointer_value(env, insn->src_reg)) {
13815 verbose(env, "R%d pointer comparison prohibited\n",
13819 src_reg = ®s[insn->src_reg];
13821 if (insn->src_reg != BPF_REG_0) {
13822 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
13827 /* check src2 operand */
13828 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
13832 dst_reg = ®s[insn->dst_reg];
13833 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
13835 if (BPF_SRC(insn->code) == BPF_K) {
13836 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
13837 } else if (src_reg->type == SCALAR_VALUE &&
13838 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
13839 pred = is_branch_taken(dst_reg,
13840 tnum_subreg(src_reg->var_off).value,
13843 } else if (src_reg->type == SCALAR_VALUE &&
13844 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
13845 pred = is_branch_taken(dst_reg,
13846 src_reg->var_off.value,
13849 } else if (dst_reg->type == SCALAR_VALUE &&
13850 is_jmp32 && tnum_is_const(tnum_subreg(dst_reg->var_off))) {
13851 pred = is_branch_taken(src_reg,
13852 tnum_subreg(dst_reg->var_off).value,
13853 flip_opcode(opcode),
13855 } else if (dst_reg->type == SCALAR_VALUE &&
13856 !is_jmp32 && tnum_is_const(dst_reg->var_off)) {
13857 pred = is_branch_taken(src_reg,
13858 dst_reg->var_off.value,
13859 flip_opcode(opcode),
13861 } else if (reg_is_pkt_pointer_any(dst_reg) &&
13862 reg_is_pkt_pointer_any(src_reg) &&
13864 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
13868 /* If we get here with a dst_reg pointer type it is because
13869 * above is_branch_taken() special cased the 0 comparison.
13871 if (!__is_pointer_value(false, dst_reg))
13872 err = mark_chain_precision(env, insn->dst_reg);
13873 if (BPF_SRC(insn->code) == BPF_X && !err &&
13874 !__is_pointer_value(false, src_reg))
13875 err = mark_chain_precision(env, insn->src_reg);
13881 /* Only follow the goto, ignore fall-through. If needed, push
13882 * the fall-through branch for simulation under speculative
13885 if (!env->bypass_spec_v1 &&
13886 !sanitize_speculative_path(env, insn, *insn_idx + 1,
13889 *insn_idx += insn->off;
13891 } else if (pred == 0) {
13892 /* Only follow the fall-through branch, since that's where the
13893 * program will go. If needed, push the goto branch for
13894 * simulation under speculative execution.
13896 if (!env->bypass_spec_v1 &&
13897 !sanitize_speculative_path(env, insn,
13898 *insn_idx + insn->off + 1,
13904 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
13908 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
13910 /* detect if we are comparing against a constant value so we can adjust
13911 * our min/max values for our dst register.
13912 * this is only legit if both are scalars (or pointers to the same
13913 * object, I suppose, see the PTR_MAYBE_NULL related if block below),
13914 * because otherwise the different base pointers mean the offsets aren't
13917 if (BPF_SRC(insn->code) == BPF_X) {
13918 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
13920 if (dst_reg->type == SCALAR_VALUE &&
13921 src_reg->type == SCALAR_VALUE) {
13922 if (tnum_is_const(src_reg->var_off) ||
13924 tnum_is_const(tnum_subreg(src_reg->var_off))))
13925 reg_set_min_max(&other_branch_regs[insn->dst_reg],
13927 src_reg->var_off.value,
13928 tnum_subreg(src_reg->var_off).value,
13930 else if (tnum_is_const(dst_reg->var_off) ||
13932 tnum_is_const(tnum_subreg(dst_reg->var_off))))
13933 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
13935 dst_reg->var_off.value,
13936 tnum_subreg(dst_reg->var_off).value,
13938 else if (!is_jmp32 &&
13939 (opcode == BPF_JEQ || opcode == BPF_JNE))
13940 /* Comparing for equality, we can combine knowledge */
13941 reg_combine_min_max(&other_branch_regs[insn->src_reg],
13942 &other_branch_regs[insn->dst_reg],
13943 src_reg, dst_reg, opcode);
13945 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
13946 find_equal_scalars(this_branch, src_reg);
13947 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
13951 } else if (dst_reg->type == SCALAR_VALUE) {
13952 reg_set_min_max(&other_branch_regs[insn->dst_reg],
13953 dst_reg, insn->imm, (u32)insn->imm,
13957 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
13958 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
13959 find_equal_scalars(this_branch, dst_reg);
13960 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
13963 /* if one pointer register is compared to another pointer
13964 * register check if PTR_MAYBE_NULL could be lifted.
13965 * E.g. register A - maybe null
13966 * register B - not null
13967 * for JNE A, B, ... - A is not null in the false branch;
13968 * for JEQ A, B, ... - A is not null in the true branch.
13970 * Since PTR_TO_BTF_ID points to a kernel struct that does
13971 * not need to be null checked by the BPF program, i.e.,
13972 * could be null even without PTR_MAYBE_NULL marking, so
13973 * only propagate nullness when neither reg is that type.
13975 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_X &&
13976 __is_pointer_value(false, src_reg) && __is_pointer_value(false, dst_reg) &&
13977 type_may_be_null(src_reg->type) != type_may_be_null(dst_reg->type) &&
13978 base_type(src_reg->type) != PTR_TO_BTF_ID &&
13979 base_type(dst_reg->type) != PTR_TO_BTF_ID) {
13980 eq_branch_regs = NULL;
13983 eq_branch_regs = other_branch_regs;
13986 eq_branch_regs = regs;
13992 if (eq_branch_regs) {
13993 if (type_may_be_null(src_reg->type))
13994 mark_ptr_not_null_reg(&eq_branch_regs[insn->src_reg]);
13996 mark_ptr_not_null_reg(&eq_branch_regs[insn->dst_reg]);
14000 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
14001 * NOTE: these optimizations below are related with pointer comparison
14002 * which will never be JMP32.
14004 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
14005 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
14006 type_may_be_null(dst_reg->type)) {
14007 /* Mark all identical registers in each branch as either
14008 * safe or unknown depending R == 0 or R != 0 conditional.
14010 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
14011 opcode == BPF_JNE);
14012 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
14013 opcode == BPF_JEQ);
14014 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
14015 this_branch, other_branch) &&
14016 is_pointer_value(env, insn->dst_reg)) {
14017 verbose(env, "R%d pointer comparison prohibited\n",
14021 if (env->log.level & BPF_LOG_LEVEL)
14022 print_insn_state(env, this_branch->frame[this_branch->curframe]);
14026 /* verify BPF_LD_IMM64 instruction */
14027 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
14029 struct bpf_insn_aux_data *aux = cur_aux(env);
14030 struct bpf_reg_state *regs = cur_regs(env);
14031 struct bpf_reg_state *dst_reg;
14032 struct bpf_map *map;
14035 if (BPF_SIZE(insn->code) != BPF_DW) {
14036 verbose(env, "invalid BPF_LD_IMM insn\n");
14039 if (insn->off != 0) {
14040 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
14044 err = check_reg_arg(env, insn->dst_reg, DST_OP);
14048 dst_reg = ®s[insn->dst_reg];
14049 if (insn->src_reg == 0) {
14050 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
14052 dst_reg->type = SCALAR_VALUE;
14053 __mark_reg_known(®s[insn->dst_reg], imm);
14057 /* All special src_reg cases are listed below. From this point onwards
14058 * we either succeed and assign a corresponding dst_reg->type after
14059 * zeroing the offset, or fail and reject the program.
14061 mark_reg_known_zero(env, regs, insn->dst_reg);
14063 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
14064 dst_reg->type = aux->btf_var.reg_type;
14065 switch (base_type(dst_reg->type)) {
14067 dst_reg->mem_size = aux->btf_var.mem_size;
14069 case PTR_TO_BTF_ID:
14070 dst_reg->btf = aux->btf_var.btf;
14071 dst_reg->btf_id = aux->btf_var.btf_id;
14074 verbose(env, "bpf verifier is misconfigured\n");
14080 if (insn->src_reg == BPF_PSEUDO_FUNC) {
14081 struct bpf_prog_aux *aux = env->prog->aux;
14082 u32 subprogno = find_subprog(env,
14083 env->insn_idx + insn->imm + 1);
14085 if (!aux->func_info) {
14086 verbose(env, "missing btf func_info\n");
14089 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
14090 verbose(env, "callback function not static\n");
14094 dst_reg->type = PTR_TO_FUNC;
14095 dst_reg->subprogno = subprogno;
14099 map = env->used_maps[aux->map_index];
14100 dst_reg->map_ptr = map;
14102 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
14103 insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
14104 dst_reg->type = PTR_TO_MAP_VALUE;
14105 dst_reg->off = aux->map_off;
14106 WARN_ON_ONCE(map->max_entries != 1);
14107 /* We want reg->id to be same (0) as map_value is not distinct */
14108 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
14109 insn->src_reg == BPF_PSEUDO_MAP_IDX) {
14110 dst_reg->type = CONST_PTR_TO_MAP;
14112 verbose(env, "bpf verifier is misconfigured\n");
14119 static bool may_access_skb(enum bpf_prog_type type)
14122 case BPF_PROG_TYPE_SOCKET_FILTER:
14123 case BPF_PROG_TYPE_SCHED_CLS:
14124 case BPF_PROG_TYPE_SCHED_ACT:
14131 /* verify safety of LD_ABS|LD_IND instructions:
14132 * - they can only appear in the programs where ctx == skb
14133 * - since they are wrappers of function calls, they scratch R1-R5 registers,
14134 * preserve R6-R9, and store return value into R0
14137 * ctx == skb == R6 == CTX
14140 * SRC == any register
14141 * IMM == 32-bit immediate
14144 * R0 - 8/16/32-bit skb data converted to cpu endianness
14146 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
14148 struct bpf_reg_state *regs = cur_regs(env);
14149 static const int ctx_reg = BPF_REG_6;
14150 u8 mode = BPF_MODE(insn->code);
14153 if (!may_access_skb(resolve_prog_type(env->prog))) {
14154 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
14158 if (!env->ops->gen_ld_abs) {
14159 verbose(env, "bpf verifier is misconfigured\n");
14163 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
14164 BPF_SIZE(insn->code) == BPF_DW ||
14165 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
14166 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
14170 /* check whether implicit source operand (register R6) is readable */
14171 err = check_reg_arg(env, ctx_reg, SRC_OP);
14175 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
14176 * gen_ld_abs() may terminate the program at runtime, leading to
14179 err = check_reference_leak(env);
14181 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
14185 if (env->cur_state->active_lock.ptr) {
14186 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
14190 if (env->cur_state->active_rcu_lock) {
14191 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_rcu_read_lock-ed region\n");
14195 if (regs[ctx_reg].type != PTR_TO_CTX) {
14197 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
14201 if (mode == BPF_IND) {
14202 /* check explicit source operand */
14203 err = check_reg_arg(env, insn->src_reg, SRC_OP);
14208 err = check_ptr_off_reg(env, ®s[ctx_reg], ctx_reg);
14212 /* reset caller saved regs to unreadable */
14213 for (i = 0; i < CALLER_SAVED_REGS; i++) {
14214 mark_reg_not_init(env, regs, caller_saved[i]);
14215 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
14218 /* mark destination R0 register as readable, since it contains
14219 * the value fetched from the packet.
14220 * Already marked as written above.
14222 mark_reg_unknown(env, regs, BPF_REG_0);
14223 /* ld_abs load up to 32-bit skb data. */
14224 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
14228 static int check_return_code(struct bpf_verifier_env *env)
14230 struct tnum enforce_attach_type_range = tnum_unknown;
14231 const struct bpf_prog *prog = env->prog;
14232 struct bpf_reg_state *reg;
14233 struct tnum range = tnum_range(0, 1);
14234 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
14236 struct bpf_func_state *frame = env->cur_state->frame[0];
14237 const bool is_subprog = frame->subprogno;
14239 /* LSM and struct_ops func-ptr's return type could be "void" */
14241 switch (prog_type) {
14242 case BPF_PROG_TYPE_LSM:
14243 if (prog->expected_attach_type == BPF_LSM_CGROUP)
14244 /* See below, can be 0 or 0-1 depending on hook. */
14247 case BPF_PROG_TYPE_STRUCT_OPS:
14248 if (!prog->aux->attach_func_proto->type)
14256 /* eBPF calling convention is such that R0 is used
14257 * to return the value from eBPF program.
14258 * Make sure that it's readable at this time
14259 * of bpf_exit, which means that program wrote
14260 * something into it earlier
14262 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
14266 if (is_pointer_value(env, BPF_REG_0)) {
14267 verbose(env, "R0 leaks addr as return value\n");
14271 reg = cur_regs(env) + BPF_REG_0;
14273 if (frame->in_async_callback_fn) {
14274 /* enforce return zero from async callbacks like timer */
14275 if (reg->type != SCALAR_VALUE) {
14276 verbose(env, "In async callback the register R0 is not a known value (%s)\n",
14277 reg_type_str(env, reg->type));
14281 if (!tnum_in(tnum_const(0), reg->var_off)) {
14282 verbose_invalid_scalar(env, reg, &range, "async callback", "R0");
14289 if (reg->type != SCALAR_VALUE) {
14290 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
14291 reg_type_str(env, reg->type));
14297 switch (prog_type) {
14298 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
14299 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
14300 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
14301 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
14302 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
14303 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
14304 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
14305 range = tnum_range(1, 1);
14306 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
14307 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
14308 range = tnum_range(0, 3);
14310 case BPF_PROG_TYPE_CGROUP_SKB:
14311 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
14312 range = tnum_range(0, 3);
14313 enforce_attach_type_range = tnum_range(2, 3);
14316 case BPF_PROG_TYPE_CGROUP_SOCK:
14317 case BPF_PROG_TYPE_SOCK_OPS:
14318 case BPF_PROG_TYPE_CGROUP_DEVICE:
14319 case BPF_PROG_TYPE_CGROUP_SYSCTL:
14320 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
14322 case BPF_PROG_TYPE_RAW_TRACEPOINT:
14323 if (!env->prog->aux->attach_btf_id)
14325 range = tnum_const(0);
14327 case BPF_PROG_TYPE_TRACING:
14328 switch (env->prog->expected_attach_type) {
14329 case BPF_TRACE_FENTRY:
14330 case BPF_TRACE_FEXIT:
14331 range = tnum_const(0);
14333 case BPF_TRACE_RAW_TP:
14334 case BPF_MODIFY_RETURN:
14336 case BPF_TRACE_ITER:
14342 case BPF_PROG_TYPE_SK_LOOKUP:
14343 range = tnum_range(SK_DROP, SK_PASS);
14346 case BPF_PROG_TYPE_LSM:
14347 if (env->prog->expected_attach_type != BPF_LSM_CGROUP) {
14348 /* Regular BPF_PROG_TYPE_LSM programs can return
14353 if (!env->prog->aux->attach_func_proto->type) {
14354 /* Make sure programs that attach to void
14355 * hooks don't try to modify return value.
14357 range = tnum_range(1, 1);
14361 case BPF_PROG_TYPE_NETFILTER:
14362 range = tnum_range(NF_DROP, NF_ACCEPT);
14364 case BPF_PROG_TYPE_EXT:
14365 /* freplace program can return anything as its return value
14366 * depends on the to-be-replaced kernel func or bpf program.
14372 if (reg->type != SCALAR_VALUE) {
14373 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
14374 reg_type_str(env, reg->type));
14378 if (!tnum_in(range, reg->var_off)) {
14379 verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
14380 if (prog->expected_attach_type == BPF_LSM_CGROUP &&
14381 prog_type == BPF_PROG_TYPE_LSM &&
14382 !prog->aux->attach_func_proto->type)
14383 verbose(env, "Note, BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
14387 if (!tnum_is_unknown(enforce_attach_type_range) &&
14388 tnum_in(enforce_attach_type_range, reg->var_off))
14389 env->prog->enforce_expected_attach_type = 1;
14393 /* non-recursive DFS pseudo code
14394 * 1 procedure DFS-iterative(G,v):
14395 * 2 label v as discovered
14396 * 3 let S be a stack
14398 * 5 while S is not empty
14400 * 7 if t is what we're looking for:
14402 * 9 for all edges e in G.adjacentEdges(t) do
14403 * 10 if edge e is already labelled
14404 * 11 continue with the next edge
14405 * 12 w <- G.adjacentVertex(t,e)
14406 * 13 if vertex w is not discovered and not explored
14407 * 14 label e as tree-edge
14408 * 15 label w as discovered
14411 * 18 else if vertex w is discovered
14412 * 19 label e as back-edge
14414 * 21 // vertex w is explored
14415 * 22 label e as forward- or cross-edge
14416 * 23 label t as explored
14420 * 0x10 - discovered
14421 * 0x11 - discovered and fall-through edge labelled
14422 * 0x12 - discovered and fall-through and branch edges labelled
14433 static u32 state_htab_size(struct bpf_verifier_env *env)
14435 return env->prog->len;
14438 static struct bpf_verifier_state_list **explored_state(
14439 struct bpf_verifier_env *env,
14442 struct bpf_verifier_state *cur = env->cur_state;
14443 struct bpf_func_state *state = cur->frame[cur->curframe];
14445 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
14448 static void mark_prune_point(struct bpf_verifier_env *env, int idx)
14450 env->insn_aux_data[idx].prune_point = true;
14453 static bool is_prune_point(struct bpf_verifier_env *env, int insn_idx)
14455 return env->insn_aux_data[insn_idx].prune_point;
14458 static void mark_force_checkpoint(struct bpf_verifier_env *env, int idx)
14460 env->insn_aux_data[idx].force_checkpoint = true;
14463 static bool is_force_checkpoint(struct bpf_verifier_env *env, int insn_idx)
14465 return env->insn_aux_data[insn_idx].force_checkpoint;
14470 DONE_EXPLORING = 0,
14471 KEEP_EXPLORING = 1,
14474 /* t, w, e - match pseudo-code above:
14475 * t - index of current instruction
14476 * w - next instruction
14479 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
14482 int *insn_stack = env->cfg.insn_stack;
14483 int *insn_state = env->cfg.insn_state;
14485 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
14486 return DONE_EXPLORING;
14488 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
14489 return DONE_EXPLORING;
14491 if (w < 0 || w >= env->prog->len) {
14492 verbose_linfo(env, t, "%d: ", t);
14493 verbose(env, "jump out of range from insn %d to %d\n", t, w);
14498 /* mark branch target for state pruning */
14499 mark_prune_point(env, w);
14500 mark_jmp_point(env, w);
14503 if (insn_state[w] == 0) {
14505 insn_state[t] = DISCOVERED | e;
14506 insn_state[w] = DISCOVERED;
14507 if (env->cfg.cur_stack >= env->prog->len)
14509 insn_stack[env->cfg.cur_stack++] = w;
14510 return KEEP_EXPLORING;
14511 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
14512 if (loop_ok && env->bpf_capable)
14513 return DONE_EXPLORING;
14514 verbose_linfo(env, t, "%d: ", t);
14515 verbose_linfo(env, w, "%d: ", w);
14516 verbose(env, "back-edge from insn %d to %d\n", t, w);
14518 } else if (insn_state[w] == EXPLORED) {
14519 /* forward- or cross-edge */
14520 insn_state[t] = DISCOVERED | e;
14522 verbose(env, "insn state internal bug\n");
14525 return DONE_EXPLORING;
14528 static int visit_func_call_insn(int t, struct bpf_insn *insns,
14529 struct bpf_verifier_env *env,
14534 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
14538 mark_prune_point(env, t + 1);
14539 /* when we exit from subprog, we need to record non-linear history */
14540 mark_jmp_point(env, t + 1);
14542 if (visit_callee) {
14543 mark_prune_point(env, t);
14544 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env,
14545 /* It's ok to allow recursion from CFG point of
14546 * view. __check_func_call() will do the actual
14549 bpf_pseudo_func(insns + t));
14554 /* Visits the instruction at index t and returns one of the following:
14555 * < 0 - an error occurred
14556 * DONE_EXPLORING - the instruction was fully explored
14557 * KEEP_EXPLORING - there is still work to be done before it is fully explored
14559 static int visit_insn(int t, struct bpf_verifier_env *env)
14561 struct bpf_insn *insns = env->prog->insnsi, *insn = &insns[t];
14564 if (bpf_pseudo_func(insn))
14565 return visit_func_call_insn(t, insns, env, true);
14567 /* All non-branch instructions have a single fall-through edge. */
14568 if (BPF_CLASS(insn->code) != BPF_JMP &&
14569 BPF_CLASS(insn->code) != BPF_JMP32)
14570 return push_insn(t, t + 1, FALLTHROUGH, env, false);
14572 switch (BPF_OP(insn->code)) {
14574 return DONE_EXPLORING;
14577 if (insn->src_reg == 0 && insn->imm == BPF_FUNC_timer_set_callback)
14578 /* Mark this call insn as a prune point to trigger
14579 * is_state_visited() check before call itself is
14580 * processed by __check_func_call(). Otherwise new
14581 * async state will be pushed for further exploration.
14583 mark_prune_point(env, t);
14584 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
14585 struct bpf_kfunc_call_arg_meta meta;
14587 ret = fetch_kfunc_meta(env, insn, &meta, NULL);
14588 if (ret == 0 && is_iter_next_kfunc(&meta)) {
14589 mark_prune_point(env, t);
14590 /* Checking and saving state checkpoints at iter_next() call
14591 * is crucial for fast convergence of open-coded iterator loop
14592 * logic, so we need to force it. If we don't do that,
14593 * is_state_visited() might skip saving a checkpoint, causing
14594 * unnecessarily long sequence of not checkpointed
14595 * instructions and jumps, leading to exhaustion of jump
14596 * history buffer, and potentially other undesired outcomes.
14597 * It is expected that with correct open-coded iterators
14598 * convergence will happen quickly, so we don't run a risk of
14599 * exhausting memory.
14601 mark_force_checkpoint(env, t);
14604 return visit_func_call_insn(t, insns, env, insn->src_reg == BPF_PSEUDO_CALL);
14607 if (BPF_SRC(insn->code) != BPF_K)
14610 /* unconditional jump with single edge */
14611 ret = push_insn(t, t + insn->off + 1, FALLTHROUGH, env,
14616 mark_prune_point(env, t + insn->off + 1);
14617 mark_jmp_point(env, t + insn->off + 1);
14622 /* conditional jump with two edges */
14623 mark_prune_point(env, t);
14625 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
14629 return push_insn(t, t + insn->off + 1, BRANCH, env, true);
14633 /* non-recursive depth-first-search to detect loops in BPF program
14634 * loop == back-edge in directed graph
14636 static int check_cfg(struct bpf_verifier_env *env)
14638 int insn_cnt = env->prog->len;
14639 int *insn_stack, *insn_state;
14643 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
14647 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
14649 kvfree(insn_state);
14653 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
14654 insn_stack[0] = 0; /* 0 is the first instruction */
14655 env->cfg.cur_stack = 1;
14657 while (env->cfg.cur_stack > 0) {
14658 int t = insn_stack[env->cfg.cur_stack - 1];
14660 ret = visit_insn(t, env);
14662 case DONE_EXPLORING:
14663 insn_state[t] = EXPLORED;
14664 env->cfg.cur_stack--;
14666 case KEEP_EXPLORING:
14670 verbose(env, "visit_insn internal bug\n");
14677 if (env->cfg.cur_stack < 0) {
14678 verbose(env, "pop stack internal bug\n");
14683 for (i = 0; i < insn_cnt; i++) {
14684 if (insn_state[i] != EXPLORED) {
14685 verbose(env, "unreachable insn %d\n", i);
14690 ret = 0; /* cfg looks good */
14693 kvfree(insn_state);
14694 kvfree(insn_stack);
14695 env->cfg.insn_state = env->cfg.insn_stack = NULL;
14699 static int check_abnormal_return(struct bpf_verifier_env *env)
14703 for (i = 1; i < env->subprog_cnt; i++) {
14704 if (env->subprog_info[i].has_ld_abs) {
14705 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
14708 if (env->subprog_info[i].has_tail_call) {
14709 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
14716 /* The minimum supported BTF func info size */
14717 #define MIN_BPF_FUNCINFO_SIZE 8
14718 #define MAX_FUNCINFO_REC_SIZE 252
14720 static int check_btf_func(struct bpf_verifier_env *env,
14721 const union bpf_attr *attr,
14724 const struct btf_type *type, *func_proto, *ret_type;
14725 u32 i, nfuncs, urec_size, min_size;
14726 u32 krec_size = sizeof(struct bpf_func_info);
14727 struct bpf_func_info *krecord;
14728 struct bpf_func_info_aux *info_aux = NULL;
14729 struct bpf_prog *prog;
14730 const struct btf *btf;
14732 u32 prev_offset = 0;
14733 bool scalar_return;
14736 nfuncs = attr->func_info_cnt;
14738 if (check_abnormal_return(env))
14743 if (nfuncs != env->subprog_cnt) {
14744 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
14748 urec_size = attr->func_info_rec_size;
14749 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
14750 urec_size > MAX_FUNCINFO_REC_SIZE ||
14751 urec_size % sizeof(u32)) {
14752 verbose(env, "invalid func info rec size %u\n", urec_size);
14757 btf = prog->aux->btf;
14759 urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
14760 min_size = min_t(u32, krec_size, urec_size);
14762 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
14765 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
14769 for (i = 0; i < nfuncs; i++) {
14770 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
14772 if (ret == -E2BIG) {
14773 verbose(env, "nonzero tailing record in func info");
14774 /* set the size kernel expects so loader can zero
14775 * out the rest of the record.
14777 if (copy_to_bpfptr_offset(uattr,
14778 offsetof(union bpf_attr, func_info_rec_size),
14779 &min_size, sizeof(min_size)))
14785 if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
14790 /* check insn_off */
14793 if (krecord[i].insn_off) {
14795 "nonzero insn_off %u for the first func info record",
14796 krecord[i].insn_off);
14799 } else if (krecord[i].insn_off <= prev_offset) {
14801 "same or smaller insn offset (%u) than previous func info record (%u)",
14802 krecord[i].insn_off, prev_offset);
14806 if (env->subprog_info[i].start != krecord[i].insn_off) {
14807 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
14811 /* check type_id */
14812 type = btf_type_by_id(btf, krecord[i].type_id);
14813 if (!type || !btf_type_is_func(type)) {
14814 verbose(env, "invalid type id %d in func info",
14815 krecord[i].type_id);
14818 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
14820 func_proto = btf_type_by_id(btf, type->type);
14821 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
14822 /* btf_func_check() already verified it during BTF load */
14824 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
14826 btf_type_is_small_int(ret_type) || btf_is_any_enum(ret_type);
14827 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
14828 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
14831 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
14832 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
14836 prev_offset = krecord[i].insn_off;
14837 bpfptr_add(&urecord, urec_size);
14840 prog->aux->func_info = krecord;
14841 prog->aux->func_info_cnt = nfuncs;
14842 prog->aux->func_info_aux = info_aux;
14851 static void adjust_btf_func(struct bpf_verifier_env *env)
14853 struct bpf_prog_aux *aux = env->prog->aux;
14856 if (!aux->func_info)
14859 for (i = 0; i < env->subprog_cnt; i++)
14860 aux->func_info[i].insn_off = env->subprog_info[i].start;
14863 #define MIN_BPF_LINEINFO_SIZE offsetofend(struct bpf_line_info, line_col)
14864 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
14866 static int check_btf_line(struct bpf_verifier_env *env,
14867 const union bpf_attr *attr,
14870 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
14871 struct bpf_subprog_info *sub;
14872 struct bpf_line_info *linfo;
14873 struct bpf_prog *prog;
14874 const struct btf *btf;
14878 nr_linfo = attr->line_info_cnt;
14881 if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info))
14884 rec_size = attr->line_info_rec_size;
14885 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
14886 rec_size > MAX_LINEINFO_REC_SIZE ||
14887 rec_size & (sizeof(u32) - 1))
14890 /* Need to zero it in case the userspace may
14891 * pass in a smaller bpf_line_info object.
14893 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
14894 GFP_KERNEL | __GFP_NOWARN);
14899 btf = prog->aux->btf;
14902 sub = env->subprog_info;
14903 ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
14904 expected_size = sizeof(struct bpf_line_info);
14905 ncopy = min_t(u32, expected_size, rec_size);
14906 for (i = 0; i < nr_linfo; i++) {
14907 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
14909 if (err == -E2BIG) {
14910 verbose(env, "nonzero tailing record in line_info");
14911 if (copy_to_bpfptr_offset(uattr,
14912 offsetof(union bpf_attr, line_info_rec_size),
14913 &expected_size, sizeof(expected_size)))
14919 if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
14925 * Check insn_off to ensure
14926 * 1) strictly increasing AND
14927 * 2) bounded by prog->len
14929 * The linfo[0].insn_off == 0 check logically falls into
14930 * the later "missing bpf_line_info for func..." case
14931 * because the first linfo[0].insn_off must be the
14932 * first sub also and the first sub must have
14933 * subprog_info[0].start == 0.
14935 if ((i && linfo[i].insn_off <= prev_offset) ||
14936 linfo[i].insn_off >= prog->len) {
14937 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
14938 i, linfo[i].insn_off, prev_offset,
14944 if (!prog->insnsi[linfo[i].insn_off].code) {
14946 "Invalid insn code at line_info[%u].insn_off\n",
14952 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
14953 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
14954 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
14959 if (s != env->subprog_cnt) {
14960 if (linfo[i].insn_off == sub[s].start) {
14961 sub[s].linfo_idx = i;
14963 } else if (sub[s].start < linfo[i].insn_off) {
14964 verbose(env, "missing bpf_line_info for func#%u\n", s);
14970 prev_offset = linfo[i].insn_off;
14971 bpfptr_add(&ulinfo, rec_size);
14974 if (s != env->subprog_cnt) {
14975 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
14976 env->subprog_cnt - s, s);
14981 prog->aux->linfo = linfo;
14982 prog->aux->nr_linfo = nr_linfo;
14991 #define MIN_CORE_RELO_SIZE sizeof(struct bpf_core_relo)
14992 #define MAX_CORE_RELO_SIZE MAX_FUNCINFO_REC_SIZE
14994 static int check_core_relo(struct bpf_verifier_env *env,
14995 const union bpf_attr *attr,
14998 u32 i, nr_core_relo, ncopy, expected_size, rec_size;
14999 struct bpf_core_relo core_relo = {};
15000 struct bpf_prog *prog = env->prog;
15001 const struct btf *btf = prog->aux->btf;
15002 struct bpf_core_ctx ctx = {
15006 bpfptr_t u_core_relo;
15009 nr_core_relo = attr->core_relo_cnt;
15012 if (nr_core_relo > INT_MAX / sizeof(struct bpf_core_relo))
15015 rec_size = attr->core_relo_rec_size;
15016 if (rec_size < MIN_CORE_RELO_SIZE ||
15017 rec_size > MAX_CORE_RELO_SIZE ||
15018 rec_size % sizeof(u32))
15021 u_core_relo = make_bpfptr(attr->core_relos, uattr.is_kernel);
15022 expected_size = sizeof(struct bpf_core_relo);
15023 ncopy = min_t(u32, expected_size, rec_size);
15025 /* Unlike func_info and line_info, copy and apply each CO-RE
15026 * relocation record one at a time.
15028 for (i = 0; i < nr_core_relo; i++) {
15029 /* future proofing when sizeof(bpf_core_relo) changes */
15030 err = bpf_check_uarg_tail_zero(u_core_relo, expected_size, rec_size);
15032 if (err == -E2BIG) {
15033 verbose(env, "nonzero tailing record in core_relo");
15034 if (copy_to_bpfptr_offset(uattr,
15035 offsetof(union bpf_attr, core_relo_rec_size),
15036 &expected_size, sizeof(expected_size)))
15042 if (copy_from_bpfptr(&core_relo, u_core_relo, ncopy)) {
15047 if (core_relo.insn_off % 8 || core_relo.insn_off / 8 >= prog->len) {
15048 verbose(env, "Invalid core_relo[%u].insn_off:%u prog->len:%u\n",
15049 i, core_relo.insn_off, prog->len);
15054 err = bpf_core_apply(&ctx, &core_relo, i,
15055 &prog->insnsi[core_relo.insn_off / 8]);
15058 bpfptr_add(&u_core_relo, rec_size);
15063 static int check_btf_info(struct bpf_verifier_env *env,
15064 const union bpf_attr *attr,
15070 if (!attr->func_info_cnt && !attr->line_info_cnt) {
15071 if (check_abnormal_return(env))
15076 btf = btf_get_by_fd(attr->prog_btf_fd);
15078 return PTR_ERR(btf);
15079 if (btf_is_kernel(btf)) {
15083 env->prog->aux->btf = btf;
15085 err = check_btf_func(env, attr, uattr);
15089 err = check_btf_line(env, attr, uattr);
15093 err = check_core_relo(env, attr, uattr);
15100 /* check %cur's range satisfies %old's */
15101 static bool range_within(struct bpf_reg_state *old,
15102 struct bpf_reg_state *cur)
15104 return old->umin_value <= cur->umin_value &&
15105 old->umax_value >= cur->umax_value &&
15106 old->smin_value <= cur->smin_value &&
15107 old->smax_value >= cur->smax_value &&
15108 old->u32_min_value <= cur->u32_min_value &&
15109 old->u32_max_value >= cur->u32_max_value &&
15110 old->s32_min_value <= cur->s32_min_value &&
15111 old->s32_max_value >= cur->s32_max_value;
15114 /* If in the old state two registers had the same id, then they need to have
15115 * the same id in the new state as well. But that id could be different from
15116 * the old state, so we need to track the mapping from old to new ids.
15117 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
15118 * regs with old id 5 must also have new id 9 for the new state to be safe. But
15119 * regs with a different old id could still have new id 9, we don't care about
15121 * So we look through our idmap to see if this old id has been seen before. If
15122 * so, we require the new id to match; otherwise, we add the id pair to the map.
15124 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap)
15126 struct bpf_id_pair *map = idmap->map;
15129 /* either both IDs should be set or both should be zero */
15130 if (!!old_id != !!cur_id)
15133 if (old_id == 0) /* cur_id == 0 as well */
15136 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
15138 /* Reached an empty slot; haven't seen this id before */
15139 map[i].old = old_id;
15140 map[i].cur = cur_id;
15143 if (map[i].old == old_id)
15144 return map[i].cur == cur_id;
15145 if (map[i].cur == cur_id)
15148 /* We ran out of idmap slots, which should be impossible */
15153 /* Similar to check_ids(), but allocate a unique temporary ID
15154 * for 'old_id' or 'cur_id' of zero.
15155 * This makes pairs like '0 vs unique ID', 'unique ID vs 0' valid.
15157 static bool check_scalar_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap)
15159 old_id = old_id ? old_id : ++idmap->tmp_id_gen;
15160 cur_id = cur_id ? cur_id : ++idmap->tmp_id_gen;
15162 return check_ids(old_id, cur_id, idmap);
15165 static void clean_func_state(struct bpf_verifier_env *env,
15166 struct bpf_func_state *st)
15168 enum bpf_reg_liveness live;
15171 for (i = 0; i < BPF_REG_FP; i++) {
15172 live = st->regs[i].live;
15173 /* liveness must not touch this register anymore */
15174 st->regs[i].live |= REG_LIVE_DONE;
15175 if (!(live & REG_LIVE_READ))
15176 /* since the register is unused, clear its state
15177 * to make further comparison simpler
15179 __mark_reg_not_init(env, &st->regs[i]);
15182 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
15183 live = st->stack[i].spilled_ptr.live;
15184 /* liveness must not touch this stack slot anymore */
15185 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
15186 if (!(live & REG_LIVE_READ)) {
15187 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
15188 for (j = 0; j < BPF_REG_SIZE; j++)
15189 st->stack[i].slot_type[j] = STACK_INVALID;
15194 static void clean_verifier_state(struct bpf_verifier_env *env,
15195 struct bpf_verifier_state *st)
15199 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
15200 /* all regs in this state in all frames were already marked */
15203 for (i = 0; i <= st->curframe; i++)
15204 clean_func_state(env, st->frame[i]);
15207 /* the parentage chains form a tree.
15208 * the verifier states are added to state lists at given insn and
15209 * pushed into state stack for future exploration.
15210 * when the verifier reaches bpf_exit insn some of the verifer states
15211 * stored in the state lists have their final liveness state already,
15212 * but a lot of states will get revised from liveness point of view when
15213 * the verifier explores other branches.
15216 * 2: if r1 == 100 goto pc+1
15219 * when the verifier reaches exit insn the register r0 in the state list of
15220 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
15221 * of insn 2 and goes exploring further. At the insn 4 it will walk the
15222 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
15224 * Since the verifier pushes the branch states as it sees them while exploring
15225 * the program the condition of walking the branch instruction for the second
15226 * time means that all states below this branch were already explored and
15227 * their final liveness marks are already propagated.
15228 * Hence when the verifier completes the search of state list in is_state_visited()
15229 * we can call this clean_live_states() function to mark all liveness states
15230 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
15231 * will not be used.
15232 * This function also clears the registers and stack for states that !READ
15233 * to simplify state merging.
15235 * Important note here that walking the same branch instruction in the callee
15236 * doesn't meant that the states are DONE. The verifier has to compare
15239 static void clean_live_states(struct bpf_verifier_env *env, int insn,
15240 struct bpf_verifier_state *cur)
15242 struct bpf_verifier_state_list *sl;
15245 sl = *explored_state(env, insn);
15247 if (sl->state.branches)
15249 if (sl->state.insn_idx != insn ||
15250 sl->state.curframe != cur->curframe)
15252 for (i = 0; i <= cur->curframe; i++)
15253 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
15255 clean_verifier_state(env, &sl->state);
15261 static bool regs_exact(const struct bpf_reg_state *rold,
15262 const struct bpf_reg_state *rcur,
15263 struct bpf_idmap *idmap)
15265 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
15266 check_ids(rold->id, rcur->id, idmap) &&
15267 check_ids(rold->ref_obj_id, rcur->ref_obj_id, idmap);
15270 /* Returns true if (rold safe implies rcur safe) */
15271 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
15272 struct bpf_reg_state *rcur, struct bpf_idmap *idmap)
15274 if (!(rold->live & REG_LIVE_READ))
15275 /* explored state didn't use this */
15277 if (rold->type == NOT_INIT)
15278 /* explored state can't have used this */
15280 if (rcur->type == NOT_INIT)
15283 /* Enforce that register types have to match exactly, including their
15284 * modifiers (like PTR_MAYBE_NULL, MEM_RDONLY, etc), as a general
15287 * One can make a point that using a pointer register as unbounded
15288 * SCALAR would be technically acceptable, but this could lead to
15289 * pointer leaks because scalars are allowed to leak while pointers
15290 * are not. We could make this safe in special cases if root is
15291 * calling us, but it's probably not worth the hassle.
15293 * Also, register types that are *not* MAYBE_NULL could technically be
15294 * safe to use as their MAYBE_NULL variants (e.g., PTR_TO_MAP_VALUE
15295 * is safe to be used as PTR_TO_MAP_VALUE_OR_NULL, provided both point
15296 * to the same map).
15297 * However, if the old MAYBE_NULL register then got NULL checked,
15298 * doing so could have affected others with the same id, and we can't
15299 * check for that because we lost the id when we converted to
15300 * a non-MAYBE_NULL variant.
15301 * So, as a general rule we don't allow mixing MAYBE_NULL and
15302 * non-MAYBE_NULL registers as well.
15304 if (rold->type != rcur->type)
15307 switch (base_type(rold->type)) {
15309 if (env->explore_alu_limits) {
15310 /* explore_alu_limits disables tnum_in() and range_within()
15311 * logic and requires everything to be strict
15313 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
15314 check_scalar_ids(rold->id, rcur->id, idmap);
15316 if (!rold->precise)
15318 /* Why check_ids() for scalar registers?
15320 * Consider the following BPF code:
15321 * 1: r6 = ... unbound scalar, ID=a ...
15322 * 2: r7 = ... unbound scalar, ID=b ...
15323 * 3: if (r6 > r7) goto +1
15325 * 5: if (r6 > X) goto ...
15326 * 6: ... memory operation using r7 ...
15328 * First verification path is [1-6]:
15329 * - at (4) same bpf_reg_state::id (b) would be assigned to r6 and r7;
15330 * - at (5) r6 would be marked <= X, find_equal_scalars() would also mark
15331 * r7 <= X, because r6 and r7 share same id.
15332 * Next verification path is [1-4, 6].
15334 * Instruction (6) would be reached in two states:
15335 * I. r6{.id=b}, r7{.id=b} via path 1-6;
15336 * II. r6{.id=a}, r7{.id=b} via path 1-4, 6.
15338 * Use check_ids() to distinguish these states.
15340 * Also verify that new value satisfies old value range knowledge.
15342 return range_within(rold, rcur) &&
15343 tnum_in(rold->var_off, rcur->var_off) &&
15344 check_scalar_ids(rold->id, rcur->id, idmap);
15345 case PTR_TO_MAP_KEY:
15346 case PTR_TO_MAP_VALUE:
15349 case PTR_TO_TP_BUFFER:
15350 /* If the new min/max/var_off satisfy the old ones and
15351 * everything else matches, we are OK.
15353 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, var_off)) == 0 &&
15354 range_within(rold, rcur) &&
15355 tnum_in(rold->var_off, rcur->var_off) &&
15356 check_ids(rold->id, rcur->id, idmap) &&
15357 check_ids(rold->ref_obj_id, rcur->ref_obj_id, idmap);
15358 case PTR_TO_PACKET_META:
15359 case PTR_TO_PACKET:
15360 /* We must have at least as much range as the old ptr
15361 * did, so that any accesses which were safe before are
15362 * still safe. This is true even if old range < old off,
15363 * since someone could have accessed through (ptr - k), or
15364 * even done ptr -= k in a register, to get a safe access.
15366 if (rold->range > rcur->range)
15368 /* If the offsets don't match, we can't trust our alignment;
15369 * nor can we be sure that we won't fall out of range.
15371 if (rold->off != rcur->off)
15373 /* id relations must be preserved */
15374 if (!check_ids(rold->id, rcur->id, idmap))
15376 /* new val must satisfy old val knowledge */
15377 return range_within(rold, rcur) &&
15378 tnum_in(rold->var_off, rcur->var_off);
15380 /* two stack pointers are equal only if they're pointing to
15381 * the same stack frame, since fp-8 in foo != fp-8 in bar
15383 return regs_exact(rold, rcur, idmap) && rold->frameno == rcur->frameno;
15385 return regs_exact(rold, rcur, idmap);
15389 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
15390 struct bpf_func_state *cur, struct bpf_idmap *idmap)
15394 /* walk slots of the explored stack and ignore any additional
15395 * slots in the current stack, since explored(safe) state
15398 for (i = 0; i < old->allocated_stack; i++) {
15399 struct bpf_reg_state *old_reg, *cur_reg;
15401 spi = i / BPF_REG_SIZE;
15403 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
15404 i += BPF_REG_SIZE - 1;
15405 /* explored state didn't use this */
15409 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
15412 if (env->allow_uninit_stack &&
15413 old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC)
15416 /* explored stack has more populated slots than current stack
15417 * and these slots were used
15419 if (i >= cur->allocated_stack)
15422 /* if old state was safe with misc data in the stack
15423 * it will be safe with zero-initialized stack.
15424 * The opposite is not true
15426 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
15427 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
15429 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
15430 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
15431 /* Ex: old explored (safe) state has STACK_SPILL in
15432 * this stack slot, but current has STACK_MISC ->
15433 * this verifier states are not equivalent,
15434 * return false to continue verification of this path
15437 if (i % BPF_REG_SIZE != BPF_REG_SIZE - 1)
15439 /* Both old and cur are having same slot_type */
15440 switch (old->stack[spi].slot_type[BPF_REG_SIZE - 1]) {
15442 /* when explored and current stack slot are both storing
15443 * spilled registers, check that stored pointers types
15444 * are the same as well.
15445 * Ex: explored safe path could have stored
15446 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
15447 * but current path has stored:
15448 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
15449 * such verifier states are not equivalent.
15450 * return false to continue verification of this path
15452 if (!regsafe(env, &old->stack[spi].spilled_ptr,
15453 &cur->stack[spi].spilled_ptr, idmap))
15457 old_reg = &old->stack[spi].spilled_ptr;
15458 cur_reg = &cur->stack[spi].spilled_ptr;
15459 if (old_reg->dynptr.type != cur_reg->dynptr.type ||
15460 old_reg->dynptr.first_slot != cur_reg->dynptr.first_slot ||
15461 !check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap))
15465 old_reg = &old->stack[spi].spilled_ptr;
15466 cur_reg = &cur->stack[spi].spilled_ptr;
15467 /* iter.depth is not compared between states as it
15468 * doesn't matter for correctness and would otherwise
15469 * prevent convergence; we maintain it only to prevent
15470 * infinite loop check triggering, see
15471 * iter_active_depths_differ()
15473 if (old_reg->iter.btf != cur_reg->iter.btf ||
15474 old_reg->iter.btf_id != cur_reg->iter.btf_id ||
15475 old_reg->iter.state != cur_reg->iter.state ||
15476 /* ignore {old_reg,cur_reg}->iter.depth, see above */
15477 !check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap))
15482 case STACK_INVALID:
15484 /* Ensure that new unhandled slot types return false by default */
15492 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur,
15493 struct bpf_idmap *idmap)
15497 if (old->acquired_refs != cur->acquired_refs)
15500 for (i = 0; i < old->acquired_refs; i++) {
15501 if (!check_ids(old->refs[i].id, cur->refs[i].id, idmap))
15508 /* compare two verifier states
15510 * all states stored in state_list are known to be valid, since
15511 * verifier reached 'bpf_exit' instruction through them
15513 * this function is called when verifier exploring different branches of
15514 * execution popped from the state stack. If it sees an old state that has
15515 * more strict register state and more strict stack state then this execution
15516 * branch doesn't need to be explored further, since verifier already
15517 * concluded that more strict state leads to valid finish.
15519 * Therefore two states are equivalent if register state is more conservative
15520 * and explored stack state is more conservative than the current one.
15523 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
15524 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
15526 * In other words if current stack state (one being explored) has more
15527 * valid slots than old one that already passed validation, it means
15528 * the verifier can stop exploring and conclude that current state is valid too
15530 * Similarly with registers. If explored state has register type as invalid
15531 * whereas register type in current state is meaningful, it means that
15532 * the current state will reach 'bpf_exit' instruction safely
15534 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
15535 struct bpf_func_state *cur)
15539 for (i = 0; i < MAX_BPF_REG; i++)
15540 if (!regsafe(env, &old->regs[i], &cur->regs[i],
15541 &env->idmap_scratch))
15544 if (!stacksafe(env, old, cur, &env->idmap_scratch))
15547 if (!refsafe(old, cur, &env->idmap_scratch))
15553 static bool states_equal(struct bpf_verifier_env *env,
15554 struct bpf_verifier_state *old,
15555 struct bpf_verifier_state *cur)
15559 if (old->curframe != cur->curframe)
15562 env->idmap_scratch.tmp_id_gen = env->id_gen;
15563 memset(&env->idmap_scratch.map, 0, sizeof(env->idmap_scratch.map));
15565 /* Verification state from speculative execution simulation
15566 * must never prune a non-speculative execution one.
15568 if (old->speculative && !cur->speculative)
15571 if (old->active_lock.ptr != cur->active_lock.ptr)
15574 /* Old and cur active_lock's have to be either both present
15577 if (!!old->active_lock.id != !!cur->active_lock.id)
15580 if (old->active_lock.id &&
15581 !check_ids(old->active_lock.id, cur->active_lock.id, &env->idmap_scratch))
15584 if (old->active_rcu_lock != cur->active_rcu_lock)
15587 /* for states to be equal callsites have to be the same
15588 * and all frame states need to be equivalent
15590 for (i = 0; i <= old->curframe; i++) {
15591 if (old->frame[i]->callsite != cur->frame[i]->callsite)
15593 if (!func_states_equal(env, old->frame[i], cur->frame[i]))
15599 /* Return 0 if no propagation happened. Return negative error code if error
15600 * happened. Otherwise, return the propagated bit.
15602 static int propagate_liveness_reg(struct bpf_verifier_env *env,
15603 struct bpf_reg_state *reg,
15604 struct bpf_reg_state *parent_reg)
15606 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
15607 u8 flag = reg->live & REG_LIVE_READ;
15610 /* When comes here, read flags of PARENT_REG or REG could be any of
15611 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
15612 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
15614 if (parent_flag == REG_LIVE_READ64 ||
15615 /* Or if there is no read flag from REG. */
15617 /* Or if the read flag from REG is the same as PARENT_REG. */
15618 parent_flag == flag)
15621 err = mark_reg_read(env, reg, parent_reg, flag);
15628 /* A write screens off any subsequent reads; but write marks come from the
15629 * straight-line code between a state and its parent. When we arrive at an
15630 * equivalent state (jump target or such) we didn't arrive by the straight-line
15631 * code, so read marks in the state must propagate to the parent regardless
15632 * of the state's write marks. That's what 'parent == state->parent' comparison
15633 * in mark_reg_read() is for.
15635 static int propagate_liveness(struct bpf_verifier_env *env,
15636 const struct bpf_verifier_state *vstate,
15637 struct bpf_verifier_state *vparent)
15639 struct bpf_reg_state *state_reg, *parent_reg;
15640 struct bpf_func_state *state, *parent;
15641 int i, frame, err = 0;
15643 if (vparent->curframe != vstate->curframe) {
15644 WARN(1, "propagate_live: parent frame %d current frame %d\n",
15645 vparent->curframe, vstate->curframe);
15648 /* Propagate read liveness of registers... */
15649 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
15650 for (frame = 0; frame <= vstate->curframe; frame++) {
15651 parent = vparent->frame[frame];
15652 state = vstate->frame[frame];
15653 parent_reg = parent->regs;
15654 state_reg = state->regs;
15655 /* We don't need to worry about FP liveness, it's read-only */
15656 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
15657 err = propagate_liveness_reg(env, &state_reg[i],
15661 if (err == REG_LIVE_READ64)
15662 mark_insn_zext(env, &parent_reg[i]);
15665 /* Propagate stack slots. */
15666 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
15667 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
15668 parent_reg = &parent->stack[i].spilled_ptr;
15669 state_reg = &state->stack[i].spilled_ptr;
15670 err = propagate_liveness_reg(env, state_reg,
15679 /* find precise scalars in the previous equivalent state and
15680 * propagate them into the current state
15682 static int propagate_precision(struct bpf_verifier_env *env,
15683 const struct bpf_verifier_state *old)
15685 struct bpf_reg_state *state_reg;
15686 struct bpf_func_state *state;
15687 int i, err = 0, fr;
15690 for (fr = old->curframe; fr >= 0; fr--) {
15691 state = old->frame[fr];
15692 state_reg = state->regs;
15694 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
15695 if (state_reg->type != SCALAR_VALUE ||
15696 !state_reg->precise ||
15697 !(state_reg->live & REG_LIVE_READ))
15699 if (env->log.level & BPF_LOG_LEVEL2) {
15701 verbose(env, "frame %d: propagating r%d", fr, i);
15703 verbose(env, ",r%d", i);
15705 bt_set_frame_reg(&env->bt, fr, i);
15709 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
15710 if (!is_spilled_reg(&state->stack[i]))
15712 state_reg = &state->stack[i].spilled_ptr;
15713 if (state_reg->type != SCALAR_VALUE ||
15714 !state_reg->precise ||
15715 !(state_reg->live & REG_LIVE_READ))
15717 if (env->log.level & BPF_LOG_LEVEL2) {
15719 verbose(env, "frame %d: propagating fp%d",
15720 fr, (-i - 1) * BPF_REG_SIZE);
15722 verbose(env, ",fp%d", (-i - 1) * BPF_REG_SIZE);
15724 bt_set_frame_slot(&env->bt, fr, i);
15728 verbose(env, "\n");
15731 err = mark_chain_precision_batch(env);
15738 static bool states_maybe_looping(struct bpf_verifier_state *old,
15739 struct bpf_verifier_state *cur)
15741 struct bpf_func_state *fold, *fcur;
15742 int i, fr = cur->curframe;
15744 if (old->curframe != fr)
15747 fold = old->frame[fr];
15748 fcur = cur->frame[fr];
15749 for (i = 0; i < MAX_BPF_REG; i++)
15750 if (memcmp(&fold->regs[i], &fcur->regs[i],
15751 offsetof(struct bpf_reg_state, parent)))
15756 static bool is_iter_next_insn(struct bpf_verifier_env *env, int insn_idx)
15758 return env->insn_aux_data[insn_idx].is_iter_next;
15761 /* is_state_visited() handles iter_next() (see process_iter_next_call() for
15762 * terminology) calls specially: as opposed to bounded BPF loops, it *expects*
15763 * states to match, which otherwise would look like an infinite loop. So while
15764 * iter_next() calls are taken care of, we still need to be careful and
15765 * prevent erroneous and too eager declaration of "ininite loop", when
15766 * iterators are involved.
15768 * Here's a situation in pseudo-BPF assembly form:
15770 * 0: again: ; set up iter_next() call args
15771 * 1: r1 = &it ; <CHECKPOINT HERE>
15772 * 2: call bpf_iter_num_next ; this is iter_next() call
15773 * 3: if r0 == 0 goto done
15774 * 4: ... something useful here ...
15775 * 5: goto again ; another iteration
15778 * 8: call bpf_iter_num_destroy ; clean up iter state
15781 * This is a typical loop. Let's assume that we have a prune point at 1:,
15782 * before we get to `call bpf_iter_num_next` (e.g., because of that `goto
15783 * again`, assuming other heuristics don't get in a way).
15785 * When we first time come to 1:, let's say we have some state X. We proceed
15786 * to 2:, fork states, enqueue ACTIVE, validate NULL case successfully, exit.
15787 * Now we come back to validate that forked ACTIVE state. We proceed through
15788 * 3-5, come to goto, jump to 1:. Let's assume our state didn't change, so we
15789 * are converging. But the problem is that we don't know that yet, as this
15790 * convergence has to happen at iter_next() call site only. So if nothing is
15791 * done, at 1: verifier will use bounded loop logic and declare infinite
15792 * looping (and would be *technically* correct, if not for iterator's
15793 * "eventual sticky NULL" contract, see process_iter_next_call()). But we
15794 * don't want that. So what we do in process_iter_next_call() when we go on
15795 * another ACTIVE iteration, we bump slot->iter.depth, to mark that it's
15796 * a different iteration. So when we suspect an infinite loop, we additionally
15797 * check if any of the *ACTIVE* iterator states depths differ. If yes, we
15798 * pretend we are not looping and wait for next iter_next() call.
15800 * This only applies to ACTIVE state. In DRAINED state we don't expect to
15801 * loop, because that would actually mean infinite loop, as DRAINED state is
15802 * "sticky", and so we'll keep returning into the same instruction with the
15803 * same state (at least in one of possible code paths).
15805 * This approach allows to keep infinite loop heuristic even in the face of
15806 * active iterator. E.g., C snippet below is and will be detected as
15807 * inifintely looping:
15809 * struct bpf_iter_num it;
15812 * bpf_iter_num_new(&it, 0, 10);
15813 * while ((p = bpf_iter_num_next(&t))) {
15815 * while (x--) {} // <<-- infinite loop here
15819 static bool iter_active_depths_differ(struct bpf_verifier_state *old, struct bpf_verifier_state *cur)
15821 struct bpf_reg_state *slot, *cur_slot;
15822 struct bpf_func_state *state;
15825 for (fr = old->curframe; fr >= 0; fr--) {
15826 state = old->frame[fr];
15827 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
15828 if (state->stack[i].slot_type[0] != STACK_ITER)
15831 slot = &state->stack[i].spilled_ptr;
15832 if (slot->iter.state != BPF_ITER_STATE_ACTIVE)
15835 cur_slot = &cur->frame[fr]->stack[i].spilled_ptr;
15836 if (cur_slot->iter.depth != slot->iter.depth)
15843 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
15845 struct bpf_verifier_state_list *new_sl;
15846 struct bpf_verifier_state_list *sl, **pprev;
15847 struct bpf_verifier_state *cur = env->cur_state, *new;
15848 int i, j, err, states_cnt = 0;
15849 bool force_new_state = env->test_state_freq || is_force_checkpoint(env, insn_idx);
15850 bool add_new_state = force_new_state;
15852 /* bpf progs typically have pruning point every 4 instructions
15853 * http://vger.kernel.org/bpfconf2019.html#session-1
15854 * Do not add new state for future pruning if the verifier hasn't seen
15855 * at least 2 jumps and at least 8 instructions.
15856 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
15857 * In tests that amounts to up to 50% reduction into total verifier
15858 * memory consumption and 20% verifier time speedup.
15860 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
15861 env->insn_processed - env->prev_insn_processed >= 8)
15862 add_new_state = true;
15864 pprev = explored_state(env, insn_idx);
15867 clean_live_states(env, insn_idx, cur);
15871 if (sl->state.insn_idx != insn_idx)
15874 if (sl->state.branches) {
15875 struct bpf_func_state *frame = sl->state.frame[sl->state.curframe];
15877 if (frame->in_async_callback_fn &&
15878 frame->async_entry_cnt != cur->frame[cur->curframe]->async_entry_cnt) {
15879 /* Different async_entry_cnt means that the verifier is
15880 * processing another entry into async callback.
15881 * Seeing the same state is not an indication of infinite
15882 * loop or infinite recursion.
15883 * But finding the same state doesn't mean that it's safe
15884 * to stop processing the current state. The previous state
15885 * hasn't yet reached bpf_exit, since state.branches > 0.
15886 * Checking in_async_callback_fn alone is not enough either.
15887 * Since the verifier still needs to catch infinite loops
15888 * inside async callbacks.
15890 goto skip_inf_loop_check;
15892 /* BPF open-coded iterators loop detection is special.
15893 * states_maybe_looping() logic is too simplistic in detecting
15894 * states that *might* be equivalent, because it doesn't know
15895 * about ID remapping, so don't even perform it.
15896 * See process_iter_next_call() and iter_active_depths_differ()
15897 * for overview of the logic. When current and one of parent
15898 * states are detected as equivalent, it's a good thing: we prove
15899 * convergence and can stop simulating further iterations.
15900 * It's safe to assume that iterator loop will finish, taking into
15901 * account iter_next() contract of eventually returning
15902 * sticky NULL result.
15904 if (is_iter_next_insn(env, insn_idx)) {
15905 if (states_equal(env, &sl->state, cur)) {
15906 struct bpf_func_state *cur_frame;
15907 struct bpf_reg_state *iter_state, *iter_reg;
15910 cur_frame = cur->frame[cur->curframe];
15911 /* btf_check_iter_kfuncs() enforces that
15912 * iter state pointer is always the first arg
15914 iter_reg = &cur_frame->regs[BPF_REG_1];
15915 /* current state is valid due to states_equal(),
15916 * so we can assume valid iter and reg state,
15917 * no need for extra (re-)validations
15919 spi = __get_spi(iter_reg->off + iter_reg->var_off.value);
15920 iter_state = &func(env, iter_reg)->stack[spi].spilled_ptr;
15921 if (iter_state->iter.state == BPF_ITER_STATE_ACTIVE)
15924 goto skip_inf_loop_check;
15926 /* attempt to detect infinite loop to avoid unnecessary doomed work */
15927 if (states_maybe_looping(&sl->state, cur) &&
15928 states_equal(env, &sl->state, cur) &&
15929 !iter_active_depths_differ(&sl->state, cur)) {
15930 verbose_linfo(env, insn_idx, "; ");
15931 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
15934 /* if the verifier is processing a loop, avoid adding new state
15935 * too often, since different loop iterations have distinct
15936 * states and may not help future pruning.
15937 * This threshold shouldn't be too low to make sure that
15938 * a loop with large bound will be rejected quickly.
15939 * The most abusive loop will be:
15941 * if r1 < 1000000 goto pc-2
15942 * 1M insn_procssed limit / 100 == 10k peak states.
15943 * This threshold shouldn't be too high either, since states
15944 * at the end of the loop are likely to be useful in pruning.
15946 skip_inf_loop_check:
15947 if (!force_new_state &&
15948 env->jmps_processed - env->prev_jmps_processed < 20 &&
15949 env->insn_processed - env->prev_insn_processed < 100)
15950 add_new_state = false;
15953 if (states_equal(env, &sl->state, cur)) {
15956 /* reached equivalent register/stack state,
15957 * prune the search.
15958 * Registers read by the continuation are read by us.
15959 * If we have any write marks in env->cur_state, they
15960 * will prevent corresponding reads in the continuation
15961 * from reaching our parent (an explored_state). Our
15962 * own state will get the read marks recorded, but
15963 * they'll be immediately forgotten as we're pruning
15964 * this state and will pop a new one.
15966 err = propagate_liveness(env, &sl->state, cur);
15968 /* if previous state reached the exit with precision and
15969 * current state is equivalent to it (except precsion marks)
15970 * the precision needs to be propagated back in
15971 * the current state.
15973 err = err ? : push_jmp_history(env, cur);
15974 err = err ? : propagate_precision(env, &sl->state);
15980 /* when new state is not going to be added do not increase miss count.
15981 * Otherwise several loop iterations will remove the state
15982 * recorded earlier. The goal of these heuristics is to have
15983 * states from some iterations of the loop (some in the beginning
15984 * and some at the end) to help pruning.
15988 /* heuristic to determine whether this state is beneficial
15989 * to keep checking from state equivalence point of view.
15990 * Higher numbers increase max_states_per_insn and verification time,
15991 * but do not meaningfully decrease insn_processed.
15993 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
15994 /* the state is unlikely to be useful. Remove it to
15995 * speed up verification
15998 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
15999 u32 br = sl->state.branches;
16002 "BUG live_done but branches_to_explore %d\n",
16004 free_verifier_state(&sl->state, false);
16006 env->peak_states--;
16008 /* cannot free this state, since parentage chain may
16009 * walk it later. Add it for free_list instead to
16010 * be freed at the end of verification
16012 sl->next = env->free_list;
16013 env->free_list = sl;
16023 if (env->max_states_per_insn < states_cnt)
16024 env->max_states_per_insn = states_cnt;
16026 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
16029 if (!add_new_state)
16032 /* There were no equivalent states, remember the current one.
16033 * Technically the current state is not proven to be safe yet,
16034 * but it will either reach outer most bpf_exit (which means it's safe)
16035 * or it will be rejected. When there are no loops the verifier won't be
16036 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
16037 * again on the way to bpf_exit.
16038 * When looping the sl->state.branches will be > 0 and this state
16039 * will not be considered for equivalence until branches == 0.
16041 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
16044 env->total_states++;
16045 env->peak_states++;
16046 env->prev_jmps_processed = env->jmps_processed;
16047 env->prev_insn_processed = env->insn_processed;
16049 /* forget precise markings we inherited, see __mark_chain_precision */
16050 if (env->bpf_capable)
16051 mark_all_scalars_imprecise(env, cur);
16053 /* add new state to the head of linked list */
16054 new = &new_sl->state;
16055 err = copy_verifier_state(new, cur);
16057 free_verifier_state(new, false);
16061 new->insn_idx = insn_idx;
16062 WARN_ONCE(new->branches != 1,
16063 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
16066 cur->first_insn_idx = insn_idx;
16067 clear_jmp_history(cur);
16068 new_sl->next = *explored_state(env, insn_idx);
16069 *explored_state(env, insn_idx) = new_sl;
16070 /* connect new state to parentage chain. Current frame needs all
16071 * registers connected. Only r6 - r9 of the callers are alive (pushed
16072 * to the stack implicitly by JITs) so in callers' frames connect just
16073 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
16074 * the state of the call instruction (with WRITTEN set), and r0 comes
16075 * from callee with its full parentage chain, anyway.
16077 /* clear write marks in current state: the writes we did are not writes
16078 * our child did, so they don't screen off its reads from us.
16079 * (There are no read marks in current state, because reads always mark
16080 * their parent and current state never has children yet. Only
16081 * explored_states can get read marks.)
16083 for (j = 0; j <= cur->curframe; j++) {
16084 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
16085 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
16086 for (i = 0; i < BPF_REG_FP; i++)
16087 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
16090 /* all stack frames are accessible from callee, clear them all */
16091 for (j = 0; j <= cur->curframe; j++) {
16092 struct bpf_func_state *frame = cur->frame[j];
16093 struct bpf_func_state *newframe = new->frame[j];
16095 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
16096 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
16097 frame->stack[i].spilled_ptr.parent =
16098 &newframe->stack[i].spilled_ptr;
16104 /* Return true if it's OK to have the same insn return a different type. */
16105 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
16107 switch (base_type(type)) {
16109 case PTR_TO_SOCKET:
16110 case PTR_TO_SOCK_COMMON:
16111 case PTR_TO_TCP_SOCK:
16112 case PTR_TO_XDP_SOCK:
16113 case PTR_TO_BTF_ID:
16120 /* If an instruction was previously used with particular pointer types, then we
16121 * need to be careful to avoid cases such as the below, where it may be ok
16122 * for one branch accessing the pointer, but not ok for the other branch:
16127 * R1 = some_other_valid_ptr;
16130 * R2 = *(u32 *)(R1 + 0);
16132 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
16134 return src != prev && (!reg_type_mismatch_ok(src) ||
16135 !reg_type_mismatch_ok(prev));
16138 static int save_aux_ptr_type(struct bpf_verifier_env *env, enum bpf_reg_type type,
16139 bool allow_trust_missmatch)
16141 enum bpf_reg_type *prev_type = &env->insn_aux_data[env->insn_idx].ptr_type;
16143 if (*prev_type == NOT_INIT) {
16144 /* Saw a valid insn
16145 * dst_reg = *(u32 *)(src_reg + off)
16146 * save type to validate intersecting paths
16149 } else if (reg_type_mismatch(type, *prev_type)) {
16150 /* Abuser program is trying to use the same insn
16151 * dst_reg = *(u32*) (src_reg + off)
16152 * with different pointer types:
16153 * src_reg == ctx in one branch and
16154 * src_reg == stack|map in some other branch.
16157 if (allow_trust_missmatch &&
16158 base_type(type) == PTR_TO_BTF_ID &&
16159 base_type(*prev_type) == PTR_TO_BTF_ID) {
16161 * Have to support a use case when one path through
16162 * the program yields TRUSTED pointer while another
16163 * is UNTRUSTED. Fallback to UNTRUSTED to generate
16166 *prev_type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
16168 verbose(env, "same insn cannot be used with different pointers\n");
16176 static int do_check(struct bpf_verifier_env *env)
16178 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
16179 struct bpf_verifier_state *state = env->cur_state;
16180 struct bpf_insn *insns = env->prog->insnsi;
16181 struct bpf_reg_state *regs;
16182 int insn_cnt = env->prog->len;
16183 bool do_print_state = false;
16184 int prev_insn_idx = -1;
16187 struct bpf_insn *insn;
16191 env->prev_insn_idx = prev_insn_idx;
16192 if (env->insn_idx >= insn_cnt) {
16193 verbose(env, "invalid insn idx %d insn_cnt %d\n",
16194 env->insn_idx, insn_cnt);
16198 insn = &insns[env->insn_idx];
16199 class = BPF_CLASS(insn->code);
16201 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
16203 "BPF program is too large. Processed %d insn\n",
16204 env->insn_processed);
16208 state->last_insn_idx = env->prev_insn_idx;
16210 if (is_prune_point(env, env->insn_idx)) {
16211 err = is_state_visited(env, env->insn_idx);
16215 /* found equivalent state, can prune the search */
16216 if (env->log.level & BPF_LOG_LEVEL) {
16217 if (do_print_state)
16218 verbose(env, "\nfrom %d to %d%s: safe\n",
16219 env->prev_insn_idx, env->insn_idx,
16220 env->cur_state->speculative ?
16221 " (speculative execution)" : "");
16223 verbose(env, "%d: safe\n", env->insn_idx);
16225 goto process_bpf_exit;
16229 if (is_jmp_point(env, env->insn_idx)) {
16230 err = push_jmp_history(env, state);
16235 if (signal_pending(current))
16238 if (need_resched())
16241 if (env->log.level & BPF_LOG_LEVEL2 && do_print_state) {
16242 verbose(env, "\nfrom %d to %d%s:",
16243 env->prev_insn_idx, env->insn_idx,
16244 env->cur_state->speculative ?
16245 " (speculative execution)" : "");
16246 print_verifier_state(env, state->frame[state->curframe], true);
16247 do_print_state = false;
16250 if (env->log.level & BPF_LOG_LEVEL) {
16251 const struct bpf_insn_cbs cbs = {
16252 .cb_call = disasm_kfunc_name,
16253 .cb_print = verbose,
16254 .private_data = env,
16257 if (verifier_state_scratched(env))
16258 print_insn_state(env, state->frame[state->curframe]);
16260 verbose_linfo(env, env->insn_idx, "; ");
16261 env->prev_log_pos = env->log.end_pos;
16262 verbose(env, "%d: ", env->insn_idx);
16263 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
16264 env->prev_insn_print_pos = env->log.end_pos - env->prev_log_pos;
16265 env->prev_log_pos = env->log.end_pos;
16268 if (bpf_prog_is_offloaded(env->prog->aux)) {
16269 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
16270 env->prev_insn_idx);
16275 regs = cur_regs(env);
16276 sanitize_mark_insn_seen(env);
16277 prev_insn_idx = env->insn_idx;
16279 if (class == BPF_ALU || class == BPF_ALU64) {
16280 err = check_alu_op(env, insn);
16284 } else if (class == BPF_LDX) {
16285 enum bpf_reg_type src_reg_type;
16287 /* check for reserved fields is already done */
16289 /* check src operand */
16290 err = check_reg_arg(env, insn->src_reg, SRC_OP);
16294 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
16298 src_reg_type = regs[insn->src_reg].type;
16300 /* check that memory (src_reg + off) is readable,
16301 * the state of dst_reg will be updated by this func
16303 err = check_mem_access(env, env->insn_idx, insn->src_reg,
16304 insn->off, BPF_SIZE(insn->code),
16305 BPF_READ, insn->dst_reg, false);
16309 err = save_aux_ptr_type(env, src_reg_type, true);
16312 } else if (class == BPF_STX) {
16313 enum bpf_reg_type dst_reg_type;
16315 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
16316 err = check_atomic(env, env->insn_idx, insn);
16323 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
16324 verbose(env, "BPF_STX uses reserved fields\n");
16328 /* check src1 operand */
16329 err = check_reg_arg(env, insn->src_reg, SRC_OP);
16332 /* check src2 operand */
16333 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
16337 dst_reg_type = regs[insn->dst_reg].type;
16339 /* check that memory (dst_reg + off) is writeable */
16340 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
16341 insn->off, BPF_SIZE(insn->code),
16342 BPF_WRITE, insn->src_reg, false);
16346 err = save_aux_ptr_type(env, dst_reg_type, false);
16349 } else if (class == BPF_ST) {
16350 enum bpf_reg_type dst_reg_type;
16352 if (BPF_MODE(insn->code) != BPF_MEM ||
16353 insn->src_reg != BPF_REG_0) {
16354 verbose(env, "BPF_ST uses reserved fields\n");
16357 /* check src operand */
16358 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
16362 dst_reg_type = regs[insn->dst_reg].type;
16364 /* check that memory (dst_reg + off) is writeable */
16365 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
16366 insn->off, BPF_SIZE(insn->code),
16367 BPF_WRITE, -1, false);
16371 err = save_aux_ptr_type(env, dst_reg_type, false);
16374 } else if (class == BPF_JMP || class == BPF_JMP32) {
16375 u8 opcode = BPF_OP(insn->code);
16377 env->jmps_processed++;
16378 if (opcode == BPF_CALL) {
16379 if (BPF_SRC(insn->code) != BPF_K ||
16380 (insn->src_reg != BPF_PSEUDO_KFUNC_CALL
16381 && insn->off != 0) ||
16382 (insn->src_reg != BPF_REG_0 &&
16383 insn->src_reg != BPF_PSEUDO_CALL &&
16384 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
16385 insn->dst_reg != BPF_REG_0 ||
16386 class == BPF_JMP32) {
16387 verbose(env, "BPF_CALL uses reserved fields\n");
16391 if (env->cur_state->active_lock.ptr) {
16392 if ((insn->src_reg == BPF_REG_0 && insn->imm != BPF_FUNC_spin_unlock) ||
16393 (insn->src_reg == BPF_PSEUDO_CALL) ||
16394 (insn->src_reg == BPF_PSEUDO_KFUNC_CALL &&
16395 (insn->off != 0 || !is_bpf_graph_api_kfunc(insn->imm)))) {
16396 verbose(env, "function calls are not allowed while holding a lock\n");
16400 if (insn->src_reg == BPF_PSEUDO_CALL)
16401 err = check_func_call(env, insn, &env->insn_idx);
16402 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
16403 err = check_kfunc_call(env, insn, &env->insn_idx);
16405 err = check_helper_call(env, insn, &env->insn_idx);
16409 mark_reg_scratched(env, BPF_REG_0);
16410 } else if (opcode == BPF_JA) {
16411 if (BPF_SRC(insn->code) != BPF_K ||
16413 insn->src_reg != BPF_REG_0 ||
16414 insn->dst_reg != BPF_REG_0 ||
16415 class == BPF_JMP32) {
16416 verbose(env, "BPF_JA uses reserved fields\n");
16420 env->insn_idx += insn->off + 1;
16423 } else if (opcode == BPF_EXIT) {
16424 if (BPF_SRC(insn->code) != BPF_K ||
16426 insn->src_reg != BPF_REG_0 ||
16427 insn->dst_reg != BPF_REG_0 ||
16428 class == BPF_JMP32) {
16429 verbose(env, "BPF_EXIT uses reserved fields\n");
16433 if (env->cur_state->active_lock.ptr &&
16434 !in_rbtree_lock_required_cb(env)) {
16435 verbose(env, "bpf_spin_unlock is missing\n");
16439 if (env->cur_state->active_rcu_lock) {
16440 verbose(env, "bpf_rcu_read_unlock is missing\n");
16444 /* We must do check_reference_leak here before
16445 * prepare_func_exit to handle the case when
16446 * state->curframe > 0, it may be a callback
16447 * function, for which reference_state must
16448 * match caller reference state when it exits.
16450 err = check_reference_leak(env);
16454 if (state->curframe) {
16455 /* exit from nested function */
16456 err = prepare_func_exit(env, &env->insn_idx);
16459 do_print_state = true;
16463 err = check_return_code(env);
16467 mark_verifier_state_scratched(env);
16468 update_branch_counts(env, env->cur_state);
16469 err = pop_stack(env, &prev_insn_idx,
16470 &env->insn_idx, pop_log);
16472 if (err != -ENOENT)
16476 do_print_state = true;
16480 err = check_cond_jmp_op(env, insn, &env->insn_idx);
16484 } else if (class == BPF_LD) {
16485 u8 mode = BPF_MODE(insn->code);
16487 if (mode == BPF_ABS || mode == BPF_IND) {
16488 err = check_ld_abs(env, insn);
16492 } else if (mode == BPF_IMM) {
16493 err = check_ld_imm(env, insn);
16498 sanitize_mark_insn_seen(env);
16500 verbose(env, "invalid BPF_LD mode\n");
16504 verbose(env, "unknown insn class %d\n", class);
16514 static int find_btf_percpu_datasec(struct btf *btf)
16516 const struct btf_type *t;
16521 * Both vmlinux and module each have their own ".data..percpu"
16522 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
16523 * types to look at only module's own BTF types.
16525 n = btf_nr_types(btf);
16526 if (btf_is_module(btf))
16527 i = btf_nr_types(btf_vmlinux);
16531 for(; i < n; i++) {
16532 t = btf_type_by_id(btf, i);
16533 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
16536 tname = btf_name_by_offset(btf, t->name_off);
16537 if (!strcmp(tname, ".data..percpu"))
16544 /* replace pseudo btf_id with kernel symbol address */
16545 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
16546 struct bpf_insn *insn,
16547 struct bpf_insn_aux_data *aux)
16549 const struct btf_var_secinfo *vsi;
16550 const struct btf_type *datasec;
16551 struct btf_mod_pair *btf_mod;
16552 const struct btf_type *t;
16553 const char *sym_name;
16554 bool percpu = false;
16555 u32 type, id = insn->imm;
16559 int i, btf_fd, err;
16561 btf_fd = insn[1].imm;
16563 btf = btf_get_by_fd(btf_fd);
16565 verbose(env, "invalid module BTF object FD specified.\n");
16569 if (!btf_vmlinux) {
16570 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
16577 t = btf_type_by_id(btf, id);
16579 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
16584 if (!btf_type_is_var(t) && !btf_type_is_func(t)) {
16585 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR or KIND_FUNC\n", id);
16590 sym_name = btf_name_by_offset(btf, t->name_off);
16591 addr = kallsyms_lookup_name(sym_name);
16593 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
16598 insn[0].imm = (u32)addr;
16599 insn[1].imm = addr >> 32;
16601 if (btf_type_is_func(t)) {
16602 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
16603 aux->btf_var.mem_size = 0;
16607 datasec_id = find_btf_percpu_datasec(btf);
16608 if (datasec_id > 0) {
16609 datasec = btf_type_by_id(btf, datasec_id);
16610 for_each_vsi(i, datasec, vsi) {
16611 if (vsi->type == id) {
16619 t = btf_type_skip_modifiers(btf, type, NULL);
16621 aux->btf_var.reg_type = PTR_TO_BTF_ID | MEM_PERCPU;
16622 aux->btf_var.btf = btf;
16623 aux->btf_var.btf_id = type;
16624 } else if (!btf_type_is_struct(t)) {
16625 const struct btf_type *ret;
16629 /* resolve the type size of ksym. */
16630 ret = btf_resolve_size(btf, t, &tsize);
16632 tname = btf_name_by_offset(btf, t->name_off);
16633 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
16634 tname, PTR_ERR(ret));
16638 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
16639 aux->btf_var.mem_size = tsize;
16641 aux->btf_var.reg_type = PTR_TO_BTF_ID;
16642 aux->btf_var.btf = btf;
16643 aux->btf_var.btf_id = type;
16646 /* check whether we recorded this BTF (and maybe module) already */
16647 for (i = 0; i < env->used_btf_cnt; i++) {
16648 if (env->used_btfs[i].btf == btf) {
16654 if (env->used_btf_cnt >= MAX_USED_BTFS) {
16659 btf_mod = &env->used_btfs[env->used_btf_cnt];
16660 btf_mod->btf = btf;
16661 btf_mod->module = NULL;
16663 /* if we reference variables from kernel module, bump its refcount */
16664 if (btf_is_module(btf)) {
16665 btf_mod->module = btf_try_get_module(btf);
16666 if (!btf_mod->module) {
16672 env->used_btf_cnt++;
16680 static bool is_tracing_prog_type(enum bpf_prog_type type)
16683 case BPF_PROG_TYPE_KPROBE:
16684 case BPF_PROG_TYPE_TRACEPOINT:
16685 case BPF_PROG_TYPE_PERF_EVENT:
16686 case BPF_PROG_TYPE_RAW_TRACEPOINT:
16687 case BPF_PROG_TYPE_RAW_TRACEPOINT_WRITABLE:
16694 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
16695 struct bpf_map *map,
16696 struct bpf_prog *prog)
16699 enum bpf_prog_type prog_type = resolve_prog_type(prog);
16701 if (btf_record_has_field(map->record, BPF_LIST_HEAD) ||
16702 btf_record_has_field(map->record, BPF_RB_ROOT)) {
16703 if (is_tracing_prog_type(prog_type)) {
16704 verbose(env, "tracing progs cannot use bpf_{list_head,rb_root} yet\n");
16709 if (btf_record_has_field(map->record, BPF_SPIN_LOCK)) {
16710 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
16711 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
16715 if (is_tracing_prog_type(prog_type)) {
16716 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
16720 if (prog->aux->sleepable) {
16721 verbose(env, "sleepable progs cannot use bpf_spin_lock yet\n");
16726 if (btf_record_has_field(map->record, BPF_TIMER)) {
16727 if (is_tracing_prog_type(prog_type)) {
16728 verbose(env, "tracing progs cannot use bpf_timer yet\n");
16733 if ((bpf_prog_is_offloaded(prog->aux) || bpf_map_is_offloaded(map)) &&
16734 !bpf_offload_prog_map_match(prog, map)) {
16735 verbose(env, "offload device mismatch between prog and map\n");
16739 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
16740 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
16744 if (prog->aux->sleepable)
16745 switch (map->map_type) {
16746 case BPF_MAP_TYPE_HASH:
16747 case BPF_MAP_TYPE_LRU_HASH:
16748 case BPF_MAP_TYPE_ARRAY:
16749 case BPF_MAP_TYPE_PERCPU_HASH:
16750 case BPF_MAP_TYPE_PERCPU_ARRAY:
16751 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
16752 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
16753 case BPF_MAP_TYPE_HASH_OF_MAPS:
16754 case BPF_MAP_TYPE_RINGBUF:
16755 case BPF_MAP_TYPE_USER_RINGBUF:
16756 case BPF_MAP_TYPE_INODE_STORAGE:
16757 case BPF_MAP_TYPE_SK_STORAGE:
16758 case BPF_MAP_TYPE_TASK_STORAGE:
16759 case BPF_MAP_TYPE_CGRP_STORAGE:
16763 "Sleepable programs can only use array, hash, ringbuf and local storage maps\n");
16770 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
16772 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
16773 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
16776 /* find and rewrite pseudo imm in ld_imm64 instructions:
16778 * 1. if it accesses map FD, replace it with actual map pointer.
16779 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
16781 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
16783 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
16785 struct bpf_insn *insn = env->prog->insnsi;
16786 int insn_cnt = env->prog->len;
16789 err = bpf_prog_calc_tag(env->prog);
16793 for (i = 0; i < insn_cnt; i++, insn++) {
16794 if (BPF_CLASS(insn->code) == BPF_LDX &&
16795 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
16796 verbose(env, "BPF_LDX uses reserved fields\n");
16800 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
16801 struct bpf_insn_aux_data *aux;
16802 struct bpf_map *map;
16807 if (i == insn_cnt - 1 || insn[1].code != 0 ||
16808 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
16809 insn[1].off != 0) {
16810 verbose(env, "invalid bpf_ld_imm64 insn\n");
16814 if (insn[0].src_reg == 0)
16815 /* valid generic load 64-bit imm */
16818 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
16819 aux = &env->insn_aux_data[i];
16820 err = check_pseudo_btf_id(env, insn, aux);
16826 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
16827 aux = &env->insn_aux_data[i];
16828 aux->ptr_type = PTR_TO_FUNC;
16832 /* In final convert_pseudo_ld_imm64() step, this is
16833 * converted into regular 64-bit imm load insn.
16835 switch (insn[0].src_reg) {
16836 case BPF_PSEUDO_MAP_VALUE:
16837 case BPF_PSEUDO_MAP_IDX_VALUE:
16839 case BPF_PSEUDO_MAP_FD:
16840 case BPF_PSEUDO_MAP_IDX:
16841 if (insn[1].imm == 0)
16845 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
16849 switch (insn[0].src_reg) {
16850 case BPF_PSEUDO_MAP_IDX_VALUE:
16851 case BPF_PSEUDO_MAP_IDX:
16852 if (bpfptr_is_null(env->fd_array)) {
16853 verbose(env, "fd_idx without fd_array is invalid\n");
16856 if (copy_from_bpfptr_offset(&fd, env->fd_array,
16857 insn[0].imm * sizeof(fd),
16867 map = __bpf_map_get(f);
16869 verbose(env, "fd %d is not pointing to valid bpf_map\n",
16871 return PTR_ERR(map);
16874 err = check_map_prog_compatibility(env, map, env->prog);
16880 aux = &env->insn_aux_data[i];
16881 if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
16882 insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
16883 addr = (unsigned long)map;
16885 u32 off = insn[1].imm;
16887 if (off >= BPF_MAX_VAR_OFF) {
16888 verbose(env, "direct value offset of %u is not allowed\n", off);
16893 if (!map->ops->map_direct_value_addr) {
16894 verbose(env, "no direct value access support for this map type\n");
16899 err = map->ops->map_direct_value_addr(map, &addr, off);
16901 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
16902 map->value_size, off);
16907 aux->map_off = off;
16911 insn[0].imm = (u32)addr;
16912 insn[1].imm = addr >> 32;
16914 /* check whether we recorded this map already */
16915 for (j = 0; j < env->used_map_cnt; j++) {
16916 if (env->used_maps[j] == map) {
16917 aux->map_index = j;
16923 if (env->used_map_cnt >= MAX_USED_MAPS) {
16928 /* hold the map. If the program is rejected by verifier,
16929 * the map will be released by release_maps() or it
16930 * will be used by the valid program until it's unloaded
16931 * and all maps are released in free_used_maps()
16935 aux->map_index = env->used_map_cnt;
16936 env->used_maps[env->used_map_cnt++] = map;
16938 if (bpf_map_is_cgroup_storage(map) &&
16939 bpf_cgroup_storage_assign(env->prog->aux, map)) {
16940 verbose(env, "only one cgroup storage of each type is allowed\n");
16952 /* Basic sanity check before we invest more work here. */
16953 if (!bpf_opcode_in_insntable(insn->code)) {
16954 verbose(env, "unknown opcode %02x\n", insn->code);
16959 /* now all pseudo BPF_LD_IMM64 instructions load valid
16960 * 'struct bpf_map *' into a register instead of user map_fd.
16961 * These pointers will be used later by verifier to validate map access.
16966 /* drop refcnt of maps used by the rejected program */
16967 static void release_maps(struct bpf_verifier_env *env)
16969 __bpf_free_used_maps(env->prog->aux, env->used_maps,
16970 env->used_map_cnt);
16973 /* drop refcnt of maps used by the rejected program */
16974 static void release_btfs(struct bpf_verifier_env *env)
16976 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
16977 env->used_btf_cnt);
16980 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
16981 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
16983 struct bpf_insn *insn = env->prog->insnsi;
16984 int insn_cnt = env->prog->len;
16987 for (i = 0; i < insn_cnt; i++, insn++) {
16988 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
16990 if (insn->src_reg == BPF_PSEUDO_FUNC)
16996 /* single env->prog->insni[off] instruction was replaced with the range
16997 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
16998 * [0, off) and [off, end) to new locations, so the patched range stays zero
17000 static void adjust_insn_aux_data(struct bpf_verifier_env *env,
17001 struct bpf_insn_aux_data *new_data,
17002 struct bpf_prog *new_prog, u32 off, u32 cnt)
17004 struct bpf_insn_aux_data *old_data = env->insn_aux_data;
17005 struct bpf_insn *insn = new_prog->insnsi;
17006 u32 old_seen = old_data[off].seen;
17010 /* aux info at OFF always needs adjustment, no matter fast path
17011 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
17012 * original insn at old prog.
17014 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
17018 prog_len = new_prog->len;
17020 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
17021 memcpy(new_data + off + cnt - 1, old_data + off,
17022 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
17023 for (i = off; i < off + cnt - 1; i++) {
17024 /* Expand insni[off]'s seen count to the patched range. */
17025 new_data[i].seen = old_seen;
17026 new_data[i].zext_dst = insn_has_def32(env, insn + i);
17028 env->insn_aux_data = new_data;
17032 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
17038 /* NOTE: fake 'exit' subprog should be updated as well. */
17039 for (i = 0; i <= env->subprog_cnt; i++) {
17040 if (env->subprog_info[i].start <= off)
17042 env->subprog_info[i].start += len - 1;
17046 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
17048 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
17049 int i, sz = prog->aux->size_poke_tab;
17050 struct bpf_jit_poke_descriptor *desc;
17052 for (i = 0; i < sz; i++) {
17054 if (desc->insn_idx <= off)
17056 desc->insn_idx += len - 1;
17060 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
17061 const struct bpf_insn *patch, u32 len)
17063 struct bpf_prog *new_prog;
17064 struct bpf_insn_aux_data *new_data = NULL;
17067 new_data = vzalloc(array_size(env->prog->len + len - 1,
17068 sizeof(struct bpf_insn_aux_data)));
17073 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
17074 if (IS_ERR(new_prog)) {
17075 if (PTR_ERR(new_prog) == -ERANGE)
17077 "insn %d cannot be patched due to 16-bit range\n",
17078 env->insn_aux_data[off].orig_idx);
17082 adjust_insn_aux_data(env, new_data, new_prog, off, len);
17083 adjust_subprog_starts(env, off, len);
17084 adjust_poke_descs(new_prog, off, len);
17088 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
17093 /* find first prog starting at or after off (first to remove) */
17094 for (i = 0; i < env->subprog_cnt; i++)
17095 if (env->subprog_info[i].start >= off)
17097 /* find first prog starting at or after off + cnt (first to stay) */
17098 for (j = i; j < env->subprog_cnt; j++)
17099 if (env->subprog_info[j].start >= off + cnt)
17101 /* if j doesn't start exactly at off + cnt, we are just removing
17102 * the front of previous prog
17104 if (env->subprog_info[j].start != off + cnt)
17108 struct bpf_prog_aux *aux = env->prog->aux;
17111 /* move fake 'exit' subprog as well */
17112 move = env->subprog_cnt + 1 - j;
17114 memmove(env->subprog_info + i,
17115 env->subprog_info + j,
17116 sizeof(*env->subprog_info) * move);
17117 env->subprog_cnt -= j - i;
17119 /* remove func_info */
17120 if (aux->func_info) {
17121 move = aux->func_info_cnt - j;
17123 memmove(aux->func_info + i,
17124 aux->func_info + j,
17125 sizeof(*aux->func_info) * move);
17126 aux->func_info_cnt -= j - i;
17127 /* func_info->insn_off is set after all code rewrites,
17128 * in adjust_btf_func() - no need to adjust
17132 /* convert i from "first prog to remove" to "first to adjust" */
17133 if (env->subprog_info[i].start == off)
17137 /* update fake 'exit' subprog as well */
17138 for (; i <= env->subprog_cnt; i++)
17139 env->subprog_info[i].start -= cnt;
17144 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
17147 struct bpf_prog *prog = env->prog;
17148 u32 i, l_off, l_cnt, nr_linfo;
17149 struct bpf_line_info *linfo;
17151 nr_linfo = prog->aux->nr_linfo;
17155 linfo = prog->aux->linfo;
17157 /* find first line info to remove, count lines to be removed */
17158 for (i = 0; i < nr_linfo; i++)
17159 if (linfo[i].insn_off >= off)
17164 for (; i < nr_linfo; i++)
17165 if (linfo[i].insn_off < off + cnt)
17170 /* First live insn doesn't match first live linfo, it needs to "inherit"
17171 * last removed linfo. prog is already modified, so prog->len == off
17172 * means no live instructions after (tail of the program was removed).
17174 if (prog->len != off && l_cnt &&
17175 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
17177 linfo[--i].insn_off = off + cnt;
17180 /* remove the line info which refer to the removed instructions */
17182 memmove(linfo + l_off, linfo + i,
17183 sizeof(*linfo) * (nr_linfo - i));
17185 prog->aux->nr_linfo -= l_cnt;
17186 nr_linfo = prog->aux->nr_linfo;
17189 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
17190 for (i = l_off; i < nr_linfo; i++)
17191 linfo[i].insn_off -= cnt;
17193 /* fix up all subprogs (incl. 'exit') which start >= off */
17194 for (i = 0; i <= env->subprog_cnt; i++)
17195 if (env->subprog_info[i].linfo_idx > l_off) {
17196 /* program may have started in the removed region but
17197 * may not be fully removed
17199 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
17200 env->subprog_info[i].linfo_idx -= l_cnt;
17202 env->subprog_info[i].linfo_idx = l_off;
17208 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
17210 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17211 unsigned int orig_prog_len = env->prog->len;
17214 if (bpf_prog_is_offloaded(env->prog->aux))
17215 bpf_prog_offload_remove_insns(env, off, cnt);
17217 err = bpf_remove_insns(env->prog, off, cnt);
17221 err = adjust_subprog_starts_after_remove(env, off, cnt);
17225 err = bpf_adj_linfo_after_remove(env, off, cnt);
17229 memmove(aux_data + off, aux_data + off + cnt,
17230 sizeof(*aux_data) * (orig_prog_len - off - cnt));
17235 /* The verifier does more data flow analysis than llvm and will not
17236 * explore branches that are dead at run time. Malicious programs can
17237 * have dead code too. Therefore replace all dead at-run-time code
17240 * Just nops are not optimal, e.g. if they would sit at the end of the
17241 * program and through another bug we would manage to jump there, then
17242 * we'd execute beyond program memory otherwise. Returning exception
17243 * code also wouldn't work since we can have subprogs where the dead
17244 * code could be located.
17246 static void sanitize_dead_code(struct bpf_verifier_env *env)
17248 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17249 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
17250 struct bpf_insn *insn = env->prog->insnsi;
17251 const int insn_cnt = env->prog->len;
17254 for (i = 0; i < insn_cnt; i++) {
17255 if (aux_data[i].seen)
17257 memcpy(insn + i, &trap, sizeof(trap));
17258 aux_data[i].zext_dst = false;
17262 static bool insn_is_cond_jump(u8 code)
17266 if (BPF_CLASS(code) == BPF_JMP32)
17269 if (BPF_CLASS(code) != BPF_JMP)
17273 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
17276 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
17278 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17279 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
17280 struct bpf_insn *insn = env->prog->insnsi;
17281 const int insn_cnt = env->prog->len;
17284 for (i = 0; i < insn_cnt; i++, insn++) {
17285 if (!insn_is_cond_jump(insn->code))
17288 if (!aux_data[i + 1].seen)
17289 ja.off = insn->off;
17290 else if (!aux_data[i + 1 + insn->off].seen)
17295 if (bpf_prog_is_offloaded(env->prog->aux))
17296 bpf_prog_offload_replace_insn(env, i, &ja);
17298 memcpy(insn, &ja, sizeof(ja));
17302 static int opt_remove_dead_code(struct bpf_verifier_env *env)
17304 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17305 int insn_cnt = env->prog->len;
17308 for (i = 0; i < insn_cnt; i++) {
17312 while (i + j < insn_cnt && !aux_data[i + j].seen)
17317 err = verifier_remove_insns(env, i, j);
17320 insn_cnt = env->prog->len;
17326 static int opt_remove_nops(struct bpf_verifier_env *env)
17328 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
17329 struct bpf_insn *insn = env->prog->insnsi;
17330 int insn_cnt = env->prog->len;
17333 for (i = 0; i < insn_cnt; i++) {
17334 if (memcmp(&insn[i], &ja, sizeof(ja)))
17337 err = verifier_remove_insns(env, i, 1);
17347 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
17348 const union bpf_attr *attr)
17350 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
17351 struct bpf_insn_aux_data *aux = env->insn_aux_data;
17352 int i, patch_len, delta = 0, len = env->prog->len;
17353 struct bpf_insn *insns = env->prog->insnsi;
17354 struct bpf_prog *new_prog;
17357 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
17358 zext_patch[1] = BPF_ZEXT_REG(0);
17359 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
17360 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
17361 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
17362 for (i = 0; i < len; i++) {
17363 int adj_idx = i + delta;
17364 struct bpf_insn insn;
17367 insn = insns[adj_idx];
17368 load_reg = insn_def_regno(&insn);
17369 if (!aux[adj_idx].zext_dst) {
17377 class = BPF_CLASS(code);
17378 if (load_reg == -1)
17381 /* NOTE: arg "reg" (the fourth one) is only used for
17382 * BPF_STX + SRC_OP, so it is safe to pass NULL
17385 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
17386 if (class == BPF_LD &&
17387 BPF_MODE(code) == BPF_IMM)
17392 /* ctx load could be transformed into wider load. */
17393 if (class == BPF_LDX &&
17394 aux[adj_idx].ptr_type == PTR_TO_CTX)
17397 imm_rnd = get_random_u32();
17398 rnd_hi32_patch[0] = insn;
17399 rnd_hi32_patch[1].imm = imm_rnd;
17400 rnd_hi32_patch[3].dst_reg = load_reg;
17401 patch = rnd_hi32_patch;
17403 goto apply_patch_buffer;
17406 /* Add in an zero-extend instruction if a) the JIT has requested
17407 * it or b) it's a CMPXCHG.
17409 * The latter is because: BPF_CMPXCHG always loads a value into
17410 * R0, therefore always zero-extends. However some archs'
17411 * equivalent instruction only does this load when the
17412 * comparison is successful. This detail of CMPXCHG is
17413 * orthogonal to the general zero-extension behaviour of the
17414 * CPU, so it's treated independently of bpf_jit_needs_zext.
17416 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
17419 /* Zero-extension is done by the caller. */
17420 if (bpf_pseudo_kfunc_call(&insn))
17423 if (WARN_ON(load_reg == -1)) {
17424 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
17428 zext_patch[0] = insn;
17429 zext_patch[1].dst_reg = load_reg;
17430 zext_patch[1].src_reg = load_reg;
17431 patch = zext_patch;
17433 apply_patch_buffer:
17434 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
17437 env->prog = new_prog;
17438 insns = new_prog->insnsi;
17439 aux = env->insn_aux_data;
17440 delta += patch_len - 1;
17446 /* convert load instructions that access fields of a context type into a
17447 * sequence of instructions that access fields of the underlying structure:
17448 * struct __sk_buff -> struct sk_buff
17449 * struct bpf_sock_ops -> struct sock
17451 static int convert_ctx_accesses(struct bpf_verifier_env *env)
17453 const struct bpf_verifier_ops *ops = env->ops;
17454 int i, cnt, size, ctx_field_size, delta = 0;
17455 const int insn_cnt = env->prog->len;
17456 struct bpf_insn insn_buf[16], *insn;
17457 u32 target_size, size_default, off;
17458 struct bpf_prog *new_prog;
17459 enum bpf_access_type type;
17460 bool is_narrower_load;
17462 if (ops->gen_prologue || env->seen_direct_write) {
17463 if (!ops->gen_prologue) {
17464 verbose(env, "bpf verifier is misconfigured\n");
17467 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
17469 if (cnt >= ARRAY_SIZE(insn_buf)) {
17470 verbose(env, "bpf verifier is misconfigured\n");
17473 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
17477 env->prog = new_prog;
17482 if (bpf_prog_is_offloaded(env->prog->aux))
17485 insn = env->prog->insnsi + delta;
17487 for (i = 0; i < insn_cnt; i++, insn++) {
17488 bpf_convert_ctx_access_t convert_ctx_access;
17490 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
17491 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
17492 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
17493 insn->code == (BPF_LDX | BPF_MEM | BPF_DW)) {
17495 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
17496 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
17497 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
17498 insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
17499 insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
17500 insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
17501 insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
17502 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
17508 if (type == BPF_WRITE &&
17509 env->insn_aux_data[i + delta].sanitize_stack_spill) {
17510 struct bpf_insn patch[] = {
17515 cnt = ARRAY_SIZE(patch);
17516 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
17521 env->prog = new_prog;
17522 insn = new_prog->insnsi + i + delta;
17526 switch ((int)env->insn_aux_data[i + delta].ptr_type) {
17528 if (!ops->convert_ctx_access)
17530 convert_ctx_access = ops->convert_ctx_access;
17532 case PTR_TO_SOCKET:
17533 case PTR_TO_SOCK_COMMON:
17534 convert_ctx_access = bpf_sock_convert_ctx_access;
17536 case PTR_TO_TCP_SOCK:
17537 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
17539 case PTR_TO_XDP_SOCK:
17540 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
17542 case PTR_TO_BTF_ID:
17543 case PTR_TO_BTF_ID | PTR_UNTRUSTED:
17544 /* PTR_TO_BTF_ID | MEM_ALLOC always has a valid lifetime, unlike
17545 * PTR_TO_BTF_ID, and an active ref_obj_id, but the same cannot
17546 * be said once it is marked PTR_UNTRUSTED, hence we must handle
17547 * any faults for loads into such types. BPF_WRITE is disallowed
17550 case PTR_TO_BTF_ID | MEM_ALLOC | PTR_UNTRUSTED:
17551 if (type == BPF_READ) {
17552 insn->code = BPF_LDX | BPF_PROBE_MEM |
17553 BPF_SIZE((insn)->code);
17554 env->prog->aux->num_exentries++;
17561 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
17562 size = BPF_LDST_BYTES(insn);
17564 /* If the read access is a narrower load of the field,
17565 * convert to a 4/8-byte load, to minimum program type specific
17566 * convert_ctx_access changes. If conversion is successful,
17567 * we will apply proper mask to the result.
17569 is_narrower_load = size < ctx_field_size;
17570 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
17572 if (is_narrower_load) {
17575 if (type == BPF_WRITE) {
17576 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
17581 if (ctx_field_size == 4)
17583 else if (ctx_field_size == 8)
17584 size_code = BPF_DW;
17586 insn->off = off & ~(size_default - 1);
17587 insn->code = BPF_LDX | BPF_MEM | size_code;
17591 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
17593 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
17594 (ctx_field_size && !target_size)) {
17595 verbose(env, "bpf verifier is misconfigured\n");
17599 if (is_narrower_load && size < target_size) {
17600 u8 shift = bpf_ctx_narrow_access_offset(
17601 off, size, size_default) * 8;
17602 if (shift && cnt + 1 >= ARRAY_SIZE(insn_buf)) {
17603 verbose(env, "bpf verifier narrow ctx load misconfigured\n");
17606 if (ctx_field_size <= 4) {
17608 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
17611 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
17612 (1 << size * 8) - 1);
17615 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
17618 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
17619 (1ULL << size * 8) - 1);
17623 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
17629 /* keep walking new program and skip insns we just inserted */
17630 env->prog = new_prog;
17631 insn = new_prog->insnsi + i + delta;
17637 static int jit_subprogs(struct bpf_verifier_env *env)
17639 struct bpf_prog *prog = env->prog, **func, *tmp;
17640 int i, j, subprog_start, subprog_end = 0, len, subprog;
17641 struct bpf_map *map_ptr;
17642 struct bpf_insn *insn;
17643 void *old_bpf_func;
17644 int err, num_exentries;
17646 if (env->subprog_cnt <= 1)
17649 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
17650 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn))
17653 /* Upon error here we cannot fall back to interpreter but
17654 * need a hard reject of the program. Thus -EFAULT is
17655 * propagated in any case.
17657 subprog = find_subprog(env, i + insn->imm + 1);
17659 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
17660 i + insn->imm + 1);
17663 /* temporarily remember subprog id inside insn instead of
17664 * aux_data, since next loop will split up all insns into funcs
17666 insn->off = subprog;
17667 /* remember original imm in case JIT fails and fallback
17668 * to interpreter will be needed
17670 env->insn_aux_data[i].call_imm = insn->imm;
17671 /* point imm to __bpf_call_base+1 from JITs point of view */
17673 if (bpf_pseudo_func(insn))
17674 /* jit (e.g. x86_64) may emit fewer instructions
17675 * if it learns a u32 imm is the same as a u64 imm.
17676 * Force a non zero here.
17681 err = bpf_prog_alloc_jited_linfo(prog);
17683 goto out_undo_insn;
17686 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
17688 goto out_undo_insn;
17690 for (i = 0; i < env->subprog_cnt; i++) {
17691 subprog_start = subprog_end;
17692 subprog_end = env->subprog_info[i + 1].start;
17694 len = subprog_end - subprog_start;
17695 /* bpf_prog_run() doesn't call subprogs directly,
17696 * hence main prog stats include the runtime of subprogs.
17697 * subprogs don't have IDs and not reachable via prog_get_next_id
17698 * func[i]->stats will never be accessed and stays NULL
17700 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
17703 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
17704 len * sizeof(struct bpf_insn));
17705 func[i]->type = prog->type;
17706 func[i]->len = len;
17707 if (bpf_prog_calc_tag(func[i]))
17709 func[i]->is_func = 1;
17710 func[i]->aux->func_idx = i;
17711 /* Below members will be freed only at prog->aux */
17712 func[i]->aux->btf = prog->aux->btf;
17713 func[i]->aux->func_info = prog->aux->func_info;
17714 func[i]->aux->func_info_cnt = prog->aux->func_info_cnt;
17715 func[i]->aux->poke_tab = prog->aux->poke_tab;
17716 func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
17718 for (j = 0; j < prog->aux->size_poke_tab; j++) {
17719 struct bpf_jit_poke_descriptor *poke;
17721 poke = &prog->aux->poke_tab[j];
17722 if (poke->insn_idx < subprog_end &&
17723 poke->insn_idx >= subprog_start)
17724 poke->aux = func[i]->aux;
17727 func[i]->aux->name[0] = 'F';
17728 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
17729 func[i]->jit_requested = 1;
17730 func[i]->blinding_requested = prog->blinding_requested;
17731 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
17732 func[i]->aux->kfunc_btf_tab = prog->aux->kfunc_btf_tab;
17733 func[i]->aux->linfo = prog->aux->linfo;
17734 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
17735 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
17736 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
17738 insn = func[i]->insnsi;
17739 for (j = 0; j < func[i]->len; j++, insn++) {
17740 if (BPF_CLASS(insn->code) == BPF_LDX &&
17741 BPF_MODE(insn->code) == BPF_PROBE_MEM)
17744 func[i]->aux->num_exentries = num_exentries;
17745 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
17746 func[i] = bpf_int_jit_compile(func[i]);
17747 if (!func[i]->jited) {
17754 /* at this point all bpf functions were successfully JITed
17755 * now populate all bpf_calls with correct addresses and
17756 * run last pass of JIT
17758 for (i = 0; i < env->subprog_cnt; i++) {
17759 insn = func[i]->insnsi;
17760 for (j = 0; j < func[i]->len; j++, insn++) {
17761 if (bpf_pseudo_func(insn)) {
17762 subprog = insn->off;
17763 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
17764 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
17767 if (!bpf_pseudo_call(insn))
17769 subprog = insn->off;
17770 insn->imm = BPF_CALL_IMM(func[subprog]->bpf_func);
17773 /* we use the aux data to keep a list of the start addresses
17774 * of the JITed images for each function in the program
17776 * for some architectures, such as powerpc64, the imm field
17777 * might not be large enough to hold the offset of the start
17778 * address of the callee's JITed image from __bpf_call_base
17780 * in such cases, we can lookup the start address of a callee
17781 * by using its subprog id, available from the off field of
17782 * the call instruction, as an index for this list
17784 func[i]->aux->func = func;
17785 func[i]->aux->func_cnt = env->subprog_cnt;
17787 for (i = 0; i < env->subprog_cnt; i++) {
17788 old_bpf_func = func[i]->bpf_func;
17789 tmp = bpf_int_jit_compile(func[i]);
17790 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
17791 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
17798 /* finally lock prog and jit images for all functions and
17799 * populate kallsysm. Begin at the first subprogram, since
17800 * bpf_prog_load will add the kallsyms for the main program.
17802 for (i = 1; i < env->subprog_cnt; i++) {
17803 bpf_prog_lock_ro(func[i]);
17804 bpf_prog_kallsyms_add(func[i]);
17807 /* Last step: make now unused interpreter insns from main
17808 * prog consistent for later dump requests, so they can
17809 * later look the same as if they were interpreted only.
17811 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
17812 if (bpf_pseudo_func(insn)) {
17813 insn[0].imm = env->insn_aux_data[i].call_imm;
17814 insn[1].imm = insn->off;
17818 if (!bpf_pseudo_call(insn))
17820 insn->off = env->insn_aux_data[i].call_imm;
17821 subprog = find_subprog(env, i + insn->off + 1);
17822 insn->imm = subprog;
17826 prog->bpf_func = func[0]->bpf_func;
17827 prog->jited_len = func[0]->jited_len;
17828 prog->aux->extable = func[0]->aux->extable;
17829 prog->aux->num_exentries = func[0]->aux->num_exentries;
17830 prog->aux->func = func;
17831 prog->aux->func_cnt = env->subprog_cnt;
17832 bpf_prog_jit_attempt_done(prog);
17835 /* We failed JIT'ing, so at this point we need to unregister poke
17836 * descriptors from subprogs, so that kernel is not attempting to
17837 * patch it anymore as we're freeing the subprog JIT memory.
17839 for (i = 0; i < prog->aux->size_poke_tab; i++) {
17840 map_ptr = prog->aux->poke_tab[i].tail_call.map;
17841 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
17843 /* At this point we're guaranteed that poke descriptors are not
17844 * live anymore. We can just unlink its descriptor table as it's
17845 * released with the main prog.
17847 for (i = 0; i < env->subprog_cnt; i++) {
17850 func[i]->aux->poke_tab = NULL;
17851 bpf_jit_free(func[i]);
17855 /* cleanup main prog to be interpreted */
17856 prog->jit_requested = 0;
17857 prog->blinding_requested = 0;
17858 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
17859 if (!bpf_pseudo_call(insn))
17862 insn->imm = env->insn_aux_data[i].call_imm;
17864 bpf_prog_jit_attempt_done(prog);
17868 static int fixup_call_args(struct bpf_verifier_env *env)
17870 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
17871 struct bpf_prog *prog = env->prog;
17872 struct bpf_insn *insn = prog->insnsi;
17873 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
17878 if (env->prog->jit_requested &&
17879 !bpf_prog_is_offloaded(env->prog->aux)) {
17880 err = jit_subprogs(env);
17883 if (err == -EFAULT)
17886 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
17887 if (has_kfunc_call) {
17888 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
17891 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
17892 /* When JIT fails the progs with bpf2bpf calls and tail_calls
17893 * have to be rejected, since interpreter doesn't support them yet.
17895 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
17898 for (i = 0; i < prog->len; i++, insn++) {
17899 if (bpf_pseudo_func(insn)) {
17900 /* When JIT fails the progs with callback calls
17901 * have to be rejected, since interpreter doesn't support them yet.
17903 verbose(env, "callbacks are not allowed in non-JITed programs\n");
17907 if (!bpf_pseudo_call(insn))
17909 depth = get_callee_stack_depth(env, insn, i);
17912 bpf_patch_call_args(insn, depth);
17919 /* replace a generic kfunc with a specialized version if necessary */
17920 static void specialize_kfunc(struct bpf_verifier_env *env,
17921 u32 func_id, u16 offset, unsigned long *addr)
17923 struct bpf_prog *prog = env->prog;
17924 bool seen_direct_write;
17928 if (bpf_dev_bound_kfunc_id(func_id)) {
17929 xdp_kfunc = bpf_dev_bound_resolve_kfunc(prog, func_id);
17931 *addr = (unsigned long)xdp_kfunc;
17934 /* fallback to default kfunc when not supported by netdev */
17940 if (func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) {
17941 seen_direct_write = env->seen_direct_write;
17942 is_rdonly = !may_access_direct_pkt_data(env, NULL, BPF_WRITE);
17945 *addr = (unsigned long)bpf_dynptr_from_skb_rdonly;
17947 /* restore env->seen_direct_write to its original value, since
17948 * may_access_direct_pkt_data mutates it
17950 env->seen_direct_write = seen_direct_write;
17954 static void __fixup_collection_insert_kfunc(struct bpf_insn_aux_data *insn_aux,
17955 u16 struct_meta_reg,
17956 u16 node_offset_reg,
17957 struct bpf_insn *insn,
17958 struct bpf_insn *insn_buf,
17961 struct btf_struct_meta *kptr_struct_meta = insn_aux->kptr_struct_meta;
17962 struct bpf_insn addr[2] = { BPF_LD_IMM64(struct_meta_reg, (long)kptr_struct_meta) };
17964 insn_buf[0] = addr[0];
17965 insn_buf[1] = addr[1];
17966 insn_buf[2] = BPF_MOV64_IMM(node_offset_reg, insn_aux->insert_off);
17967 insn_buf[3] = *insn;
17971 static int fixup_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
17972 struct bpf_insn *insn_buf, int insn_idx, int *cnt)
17974 const struct bpf_kfunc_desc *desc;
17977 verbose(env, "invalid kernel function call not eliminated in verifier pass\n");
17983 /* insn->imm has the btf func_id. Replace it with an offset relative to
17984 * __bpf_call_base, unless the JIT needs to call functions that are
17985 * further than 32 bits away (bpf_jit_supports_far_kfunc_call()).
17987 desc = find_kfunc_desc(env->prog, insn->imm, insn->off);
17989 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
17994 if (!bpf_jit_supports_far_kfunc_call())
17995 insn->imm = BPF_CALL_IMM(desc->addr);
17998 if (desc->func_id == special_kfunc_list[KF_bpf_obj_new_impl]) {
17999 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18000 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
18001 u64 obj_new_size = env->insn_aux_data[insn_idx].obj_new_size;
18003 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_1, obj_new_size);
18004 insn_buf[1] = addr[0];
18005 insn_buf[2] = addr[1];
18006 insn_buf[3] = *insn;
18008 } else if (desc->func_id == special_kfunc_list[KF_bpf_obj_drop_impl] ||
18009 desc->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) {
18010 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18011 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
18013 insn_buf[0] = addr[0];
18014 insn_buf[1] = addr[1];
18015 insn_buf[2] = *insn;
18017 } else if (desc->func_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
18018 desc->func_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
18019 desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
18020 int struct_meta_reg = BPF_REG_3;
18021 int node_offset_reg = BPF_REG_4;
18023 /* rbtree_add has extra 'less' arg, so args-to-fixup are in diff regs */
18024 if (desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
18025 struct_meta_reg = BPF_REG_4;
18026 node_offset_reg = BPF_REG_5;
18029 __fixup_collection_insert_kfunc(&env->insn_aux_data[insn_idx], struct_meta_reg,
18030 node_offset_reg, insn, insn_buf, cnt);
18031 } else if (desc->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx] ||
18032 desc->func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
18033 insn_buf[0] = BPF_MOV64_REG(BPF_REG_0, BPF_REG_1);
18039 /* Do various post-verification rewrites in a single program pass.
18040 * These rewrites simplify JIT and interpreter implementations.
18042 static int do_misc_fixups(struct bpf_verifier_env *env)
18044 struct bpf_prog *prog = env->prog;
18045 enum bpf_attach_type eatype = prog->expected_attach_type;
18046 enum bpf_prog_type prog_type = resolve_prog_type(prog);
18047 struct bpf_insn *insn = prog->insnsi;
18048 const struct bpf_func_proto *fn;
18049 const int insn_cnt = prog->len;
18050 const struct bpf_map_ops *ops;
18051 struct bpf_insn_aux_data *aux;
18052 struct bpf_insn insn_buf[16];
18053 struct bpf_prog *new_prog;
18054 struct bpf_map *map_ptr;
18055 int i, ret, cnt, delta = 0;
18057 for (i = 0; i < insn_cnt; i++, insn++) {
18058 /* Make divide-by-zero exceptions impossible. */
18059 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
18060 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
18061 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
18062 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
18063 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
18064 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
18065 struct bpf_insn *patchlet;
18066 struct bpf_insn chk_and_div[] = {
18067 /* [R,W]x div 0 -> 0 */
18068 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
18069 BPF_JNE | BPF_K, insn->src_reg,
18071 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
18072 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
18075 struct bpf_insn chk_and_mod[] = {
18076 /* [R,W]x mod 0 -> [R,W]x */
18077 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
18078 BPF_JEQ | BPF_K, insn->src_reg,
18079 0, 1 + (is64 ? 0 : 1), 0),
18081 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
18082 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
18085 patchlet = isdiv ? chk_and_div : chk_and_mod;
18086 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
18087 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
18089 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
18094 env->prog = prog = new_prog;
18095 insn = new_prog->insnsi + i + delta;
18099 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
18100 if (BPF_CLASS(insn->code) == BPF_LD &&
18101 (BPF_MODE(insn->code) == BPF_ABS ||
18102 BPF_MODE(insn->code) == BPF_IND)) {
18103 cnt = env->ops->gen_ld_abs(insn, insn_buf);
18104 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
18105 verbose(env, "bpf verifier is misconfigured\n");
18109 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18114 env->prog = prog = new_prog;
18115 insn = new_prog->insnsi + i + delta;
18119 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
18120 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
18121 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
18122 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
18123 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
18124 struct bpf_insn *patch = &insn_buf[0];
18125 bool issrc, isneg, isimm;
18128 aux = &env->insn_aux_data[i + delta];
18129 if (!aux->alu_state ||
18130 aux->alu_state == BPF_ALU_NON_POINTER)
18133 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
18134 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
18135 BPF_ALU_SANITIZE_SRC;
18136 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
18138 off_reg = issrc ? insn->src_reg : insn->dst_reg;
18140 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
18143 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
18144 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
18145 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
18146 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
18147 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
18148 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
18149 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
18152 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
18153 insn->src_reg = BPF_REG_AX;
18155 insn->code = insn->code == code_add ?
18156 code_sub : code_add;
18158 if (issrc && isneg && !isimm)
18159 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
18160 cnt = patch - insn_buf;
18162 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18167 env->prog = prog = new_prog;
18168 insn = new_prog->insnsi + i + delta;
18172 if (insn->code != (BPF_JMP | BPF_CALL))
18174 if (insn->src_reg == BPF_PSEUDO_CALL)
18176 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
18177 ret = fixup_kfunc_call(env, insn, insn_buf, i + delta, &cnt);
18183 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18188 env->prog = prog = new_prog;
18189 insn = new_prog->insnsi + i + delta;
18193 if (insn->imm == BPF_FUNC_get_route_realm)
18194 prog->dst_needed = 1;
18195 if (insn->imm == BPF_FUNC_get_prandom_u32)
18196 bpf_user_rnd_init_once();
18197 if (insn->imm == BPF_FUNC_override_return)
18198 prog->kprobe_override = 1;
18199 if (insn->imm == BPF_FUNC_tail_call) {
18200 /* If we tail call into other programs, we
18201 * cannot make any assumptions since they can
18202 * be replaced dynamically during runtime in
18203 * the program array.
18205 prog->cb_access = 1;
18206 if (!allow_tail_call_in_subprogs(env))
18207 prog->aux->stack_depth = MAX_BPF_STACK;
18208 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
18210 /* mark bpf_tail_call as different opcode to avoid
18211 * conditional branch in the interpreter for every normal
18212 * call and to prevent accidental JITing by JIT compiler
18213 * that doesn't support bpf_tail_call yet
18216 insn->code = BPF_JMP | BPF_TAIL_CALL;
18218 aux = &env->insn_aux_data[i + delta];
18219 if (env->bpf_capable && !prog->blinding_requested &&
18220 prog->jit_requested &&
18221 !bpf_map_key_poisoned(aux) &&
18222 !bpf_map_ptr_poisoned(aux) &&
18223 !bpf_map_ptr_unpriv(aux)) {
18224 struct bpf_jit_poke_descriptor desc = {
18225 .reason = BPF_POKE_REASON_TAIL_CALL,
18226 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
18227 .tail_call.key = bpf_map_key_immediate(aux),
18228 .insn_idx = i + delta,
18231 ret = bpf_jit_add_poke_descriptor(prog, &desc);
18233 verbose(env, "adding tail call poke descriptor failed\n");
18237 insn->imm = ret + 1;
18241 if (!bpf_map_ptr_unpriv(aux))
18244 /* instead of changing every JIT dealing with tail_call
18245 * emit two extra insns:
18246 * if (index >= max_entries) goto out;
18247 * index &= array->index_mask;
18248 * to avoid out-of-bounds cpu speculation
18250 if (bpf_map_ptr_poisoned(aux)) {
18251 verbose(env, "tail_call abusing map_ptr\n");
18255 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
18256 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
18257 map_ptr->max_entries, 2);
18258 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
18259 container_of(map_ptr,
18262 insn_buf[2] = *insn;
18264 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18269 env->prog = prog = new_prog;
18270 insn = new_prog->insnsi + i + delta;
18274 if (insn->imm == BPF_FUNC_timer_set_callback) {
18275 /* The verifier will process callback_fn as many times as necessary
18276 * with different maps and the register states prepared by
18277 * set_timer_callback_state will be accurate.
18279 * The following use case is valid:
18280 * map1 is shared by prog1, prog2, prog3.
18281 * prog1 calls bpf_timer_init for some map1 elements
18282 * prog2 calls bpf_timer_set_callback for some map1 elements.
18283 * Those that were not bpf_timer_init-ed will return -EINVAL.
18284 * prog3 calls bpf_timer_start for some map1 elements.
18285 * Those that were not both bpf_timer_init-ed and
18286 * bpf_timer_set_callback-ed will return -EINVAL.
18288 struct bpf_insn ld_addrs[2] = {
18289 BPF_LD_IMM64(BPF_REG_3, (long)prog->aux),
18292 insn_buf[0] = ld_addrs[0];
18293 insn_buf[1] = ld_addrs[1];
18294 insn_buf[2] = *insn;
18297 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18302 env->prog = prog = new_prog;
18303 insn = new_prog->insnsi + i + delta;
18304 goto patch_call_imm;
18307 if (is_storage_get_function(insn->imm)) {
18308 if (!env->prog->aux->sleepable ||
18309 env->insn_aux_data[i + delta].storage_get_func_atomic)
18310 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_ATOMIC);
18312 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_KERNEL);
18313 insn_buf[1] = *insn;
18316 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18321 env->prog = prog = new_prog;
18322 insn = new_prog->insnsi + i + delta;
18323 goto patch_call_imm;
18326 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
18327 * and other inlining handlers are currently limited to 64 bit
18330 if (prog->jit_requested && BITS_PER_LONG == 64 &&
18331 (insn->imm == BPF_FUNC_map_lookup_elem ||
18332 insn->imm == BPF_FUNC_map_update_elem ||
18333 insn->imm == BPF_FUNC_map_delete_elem ||
18334 insn->imm == BPF_FUNC_map_push_elem ||
18335 insn->imm == BPF_FUNC_map_pop_elem ||
18336 insn->imm == BPF_FUNC_map_peek_elem ||
18337 insn->imm == BPF_FUNC_redirect_map ||
18338 insn->imm == BPF_FUNC_for_each_map_elem ||
18339 insn->imm == BPF_FUNC_map_lookup_percpu_elem)) {
18340 aux = &env->insn_aux_data[i + delta];
18341 if (bpf_map_ptr_poisoned(aux))
18342 goto patch_call_imm;
18344 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
18345 ops = map_ptr->ops;
18346 if (insn->imm == BPF_FUNC_map_lookup_elem &&
18347 ops->map_gen_lookup) {
18348 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
18349 if (cnt == -EOPNOTSUPP)
18350 goto patch_map_ops_generic;
18351 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
18352 verbose(env, "bpf verifier is misconfigured\n");
18356 new_prog = bpf_patch_insn_data(env, i + delta,
18362 env->prog = prog = new_prog;
18363 insn = new_prog->insnsi + i + delta;
18367 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
18368 (void *(*)(struct bpf_map *map, void *key))NULL));
18369 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
18370 (long (*)(struct bpf_map *map, void *key))NULL));
18371 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
18372 (long (*)(struct bpf_map *map, void *key, void *value,
18374 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
18375 (long (*)(struct bpf_map *map, void *value,
18377 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
18378 (long (*)(struct bpf_map *map, void *value))NULL));
18379 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
18380 (long (*)(struct bpf_map *map, void *value))NULL));
18381 BUILD_BUG_ON(!__same_type(ops->map_redirect,
18382 (long (*)(struct bpf_map *map, u64 index, u64 flags))NULL));
18383 BUILD_BUG_ON(!__same_type(ops->map_for_each_callback,
18384 (long (*)(struct bpf_map *map,
18385 bpf_callback_t callback_fn,
18386 void *callback_ctx,
18388 BUILD_BUG_ON(!__same_type(ops->map_lookup_percpu_elem,
18389 (void *(*)(struct bpf_map *map, void *key, u32 cpu))NULL));
18391 patch_map_ops_generic:
18392 switch (insn->imm) {
18393 case BPF_FUNC_map_lookup_elem:
18394 insn->imm = BPF_CALL_IMM(ops->map_lookup_elem);
18396 case BPF_FUNC_map_update_elem:
18397 insn->imm = BPF_CALL_IMM(ops->map_update_elem);
18399 case BPF_FUNC_map_delete_elem:
18400 insn->imm = BPF_CALL_IMM(ops->map_delete_elem);
18402 case BPF_FUNC_map_push_elem:
18403 insn->imm = BPF_CALL_IMM(ops->map_push_elem);
18405 case BPF_FUNC_map_pop_elem:
18406 insn->imm = BPF_CALL_IMM(ops->map_pop_elem);
18408 case BPF_FUNC_map_peek_elem:
18409 insn->imm = BPF_CALL_IMM(ops->map_peek_elem);
18411 case BPF_FUNC_redirect_map:
18412 insn->imm = BPF_CALL_IMM(ops->map_redirect);
18414 case BPF_FUNC_for_each_map_elem:
18415 insn->imm = BPF_CALL_IMM(ops->map_for_each_callback);
18417 case BPF_FUNC_map_lookup_percpu_elem:
18418 insn->imm = BPF_CALL_IMM(ops->map_lookup_percpu_elem);
18422 goto patch_call_imm;
18425 /* Implement bpf_jiffies64 inline. */
18426 if (prog->jit_requested && BITS_PER_LONG == 64 &&
18427 insn->imm == BPF_FUNC_jiffies64) {
18428 struct bpf_insn ld_jiffies_addr[2] = {
18429 BPF_LD_IMM64(BPF_REG_0,
18430 (unsigned long)&jiffies),
18433 insn_buf[0] = ld_jiffies_addr[0];
18434 insn_buf[1] = ld_jiffies_addr[1];
18435 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
18439 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
18445 env->prog = prog = new_prog;
18446 insn = new_prog->insnsi + i + delta;
18450 /* Implement bpf_get_func_arg inline. */
18451 if (prog_type == BPF_PROG_TYPE_TRACING &&
18452 insn->imm == BPF_FUNC_get_func_arg) {
18453 /* Load nr_args from ctx - 8 */
18454 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
18455 insn_buf[1] = BPF_JMP32_REG(BPF_JGE, BPF_REG_2, BPF_REG_0, 6);
18456 insn_buf[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_2, 3);
18457 insn_buf[3] = BPF_ALU64_REG(BPF_ADD, BPF_REG_2, BPF_REG_1);
18458 insn_buf[4] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_2, 0);
18459 insn_buf[5] = BPF_STX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
18460 insn_buf[6] = BPF_MOV64_IMM(BPF_REG_0, 0);
18461 insn_buf[7] = BPF_JMP_A(1);
18462 insn_buf[8] = BPF_MOV64_IMM(BPF_REG_0, -EINVAL);
18465 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18470 env->prog = prog = new_prog;
18471 insn = new_prog->insnsi + i + delta;
18475 /* Implement bpf_get_func_ret inline. */
18476 if (prog_type == BPF_PROG_TYPE_TRACING &&
18477 insn->imm == BPF_FUNC_get_func_ret) {
18478 if (eatype == BPF_TRACE_FEXIT ||
18479 eatype == BPF_MODIFY_RETURN) {
18480 /* Load nr_args from ctx - 8 */
18481 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
18482 insn_buf[1] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_0, 3);
18483 insn_buf[2] = BPF_ALU64_REG(BPF_ADD, BPF_REG_0, BPF_REG_1);
18484 insn_buf[3] = BPF_LDX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
18485 insn_buf[4] = BPF_STX_MEM(BPF_DW, BPF_REG_2, BPF_REG_3, 0);
18486 insn_buf[5] = BPF_MOV64_IMM(BPF_REG_0, 0);
18489 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, -EOPNOTSUPP);
18493 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18498 env->prog = prog = new_prog;
18499 insn = new_prog->insnsi + i + delta;
18503 /* Implement get_func_arg_cnt inline. */
18504 if (prog_type == BPF_PROG_TYPE_TRACING &&
18505 insn->imm == BPF_FUNC_get_func_arg_cnt) {
18506 /* Load nr_args from ctx - 8 */
18507 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
18509 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
18513 env->prog = prog = new_prog;
18514 insn = new_prog->insnsi + i + delta;
18518 /* Implement bpf_get_func_ip inline. */
18519 if (prog_type == BPF_PROG_TYPE_TRACING &&
18520 insn->imm == BPF_FUNC_get_func_ip) {
18521 /* Load IP address from ctx - 16 */
18522 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -16);
18524 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
18528 env->prog = prog = new_prog;
18529 insn = new_prog->insnsi + i + delta;
18534 fn = env->ops->get_func_proto(insn->imm, env->prog);
18535 /* all functions that have prototype and verifier allowed
18536 * programs to call them, must be real in-kernel functions
18540 "kernel subsystem misconfigured func %s#%d\n",
18541 func_id_name(insn->imm), insn->imm);
18544 insn->imm = fn->func - __bpf_call_base;
18547 /* Since poke tab is now finalized, publish aux to tracker. */
18548 for (i = 0; i < prog->aux->size_poke_tab; i++) {
18549 map_ptr = prog->aux->poke_tab[i].tail_call.map;
18550 if (!map_ptr->ops->map_poke_track ||
18551 !map_ptr->ops->map_poke_untrack ||
18552 !map_ptr->ops->map_poke_run) {
18553 verbose(env, "bpf verifier is misconfigured\n");
18557 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
18559 verbose(env, "tracking tail call prog failed\n");
18564 sort_kfunc_descs_by_imm_off(env->prog);
18569 static struct bpf_prog *inline_bpf_loop(struct bpf_verifier_env *env,
18572 u32 callback_subprogno,
18575 s32 r6_offset = stack_base + 0 * BPF_REG_SIZE;
18576 s32 r7_offset = stack_base + 1 * BPF_REG_SIZE;
18577 s32 r8_offset = stack_base + 2 * BPF_REG_SIZE;
18578 int reg_loop_max = BPF_REG_6;
18579 int reg_loop_cnt = BPF_REG_7;
18580 int reg_loop_ctx = BPF_REG_8;
18582 struct bpf_prog *new_prog;
18583 u32 callback_start;
18584 u32 call_insn_offset;
18585 s32 callback_offset;
18587 /* This represents an inlined version of bpf_iter.c:bpf_loop,
18588 * be careful to modify this code in sync.
18590 struct bpf_insn insn_buf[] = {
18591 /* Return error and jump to the end of the patch if
18592 * expected number of iterations is too big.
18594 BPF_JMP_IMM(BPF_JLE, BPF_REG_1, BPF_MAX_LOOPS, 2),
18595 BPF_MOV32_IMM(BPF_REG_0, -E2BIG),
18596 BPF_JMP_IMM(BPF_JA, 0, 0, 16),
18597 /* spill R6, R7, R8 to use these as loop vars */
18598 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_6, r6_offset),
18599 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_7, r7_offset),
18600 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_8, r8_offset),
18601 /* initialize loop vars */
18602 BPF_MOV64_REG(reg_loop_max, BPF_REG_1),
18603 BPF_MOV32_IMM(reg_loop_cnt, 0),
18604 BPF_MOV64_REG(reg_loop_ctx, BPF_REG_3),
18606 * if reg_loop_cnt >= reg_loop_max skip the loop body
18608 BPF_JMP_REG(BPF_JGE, reg_loop_cnt, reg_loop_max, 5),
18610 * correct callback offset would be set after patching
18612 BPF_MOV64_REG(BPF_REG_1, reg_loop_cnt),
18613 BPF_MOV64_REG(BPF_REG_2, reg_loop_ctx),
18615 /* increment loop counter */
18616 BPF_ALU64_IMM(BPF_ADD, reg_loop_cnt, 1),
18617 /* jump to loop header if callback returned 0 */
18618 BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, -6),
18619 /* return value of bpf_loop,
18620 * set R0 to the number of iterations
18622 BPF_MOV64_REG(BPF_REG_0, reg_loop_cnt),
18623 /* restore original values of R6, R7, R8 */
18624 BPF_LDX_MEM(BPF_DW, BPF_REG_6, BPF_REG_10, r6_offset),
18625 BPF_LDX_MEM(BPF_DW, BPF_REG_7, BPF_REG_10, r7_offset),
18626 BPF_LDX_MEM(BPF_DW, BPF_REG_8, BPF_REG_10, r8_offset),
18629 *cnt = ARRAY_SIZE(insn_buf);
18630 new_prog = bpf_patch_insn_data(env, position, insn_buf, *cnt);
18634 /* callback start is known only after patching */
18635 callback_start = env->subprog_info[callback_subprogno].start;
18636 /* Note: insn_buf[12] is an offset of BPF_CALL_REL instruction */
18637 call_insn_offset = position + 12;
18638 callback_offset = callback_start - call_insn_offset - 1;
18639 new_prog->insnsi[call_insn_offset].imm = callback_offset;
18644 static bool is_bpf_loop_call(struct bpf_insn *insn)
18646 return insn->code == (BPF_JMP | BPF_CALL) &&
18647 insn->src_reg == 0 &&
18648 insn->imm == BPF_FUNC_loop;
18651 /* For all sub-programs in the program (including main) check
18652 * insn_aux_data to see if there are bpf_loop calls that require
18653 * inlining. If such calls are found the calls are replaced with a
18654 * sequence of instructions produced by `inline_bpf_loop` function and
18655 * subprog stack_depth is increased by the size of 3 registers.
18656 * This stack space is used to spill values of the R6, R7, R8. These
18657 * registers are used to store the loop bound, counter and context
18660 static int optimize_bpf_loop(struct bpf_verifier_env *env)
18662 struct bpf_subprog_info *subprogs = env->subprog_info;
18663 int i, cur_subprog = 0, cnt, delta = 0;
18664 struct bpf_insn *insn = env->prog->insnsi;
18665 int insn_cnt = env->prog->len;
18666 u16 stack_depth = subprogs[cur_subprog].stack_depth;
18667 u16 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
18668 u16 stack_depth_extra = 0;
18670 for (i = 0; i < insn_cnt; i++, insn++) {
18671 struct bpf_loop_inline_state *inline_state =
18672 &env->insn_aux_data[i + delta].loop_inline_state;
18674 if (is_bpf_loop_call(insn) && inline_state->fit_for_inline) {
18675 struct bpf_prog *new_prog;
18677 stack_depth_extra = BPF_REG_SIZE * 3 + stack_depth_roundup;
18678 new_prog = inline_bpf_loop(env,
18680 -(stack_depth + stack_depth_extra),
18681 inline_state->callback_subprogno,
18687 env->prog = new_prog;
18688 insn = new_prog->insnsi + i + delta;
18691 if (subprogs[cur_subprog + 1].start == i + delta + 1) {
18692 subprogs[cur_subprog].stack_depth += stack_depth_extra;
18694 stack_depth = subprogs[cur_subprog].stack_depth;
18695 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
18696 stack_depth_extra = 0;
18700 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
18705 static void free_states(struct bpf_verifier_env *env)
18707 struct bpf_verifier_state_list *sl, *sln;
18710 sl = env->free_list;
18713 free_verifier_state(&sl->state, false);
18717 env->free_list = NULL;
18719 if (!env->explored_states)
18722 for (i = 0; i < state_htab_size(env); i++) {
18723 sl = env->explored_states[i];
18727 free_verifier_state(&sl->state, false);
18731 env->explored_states[i] = NULL;
18735 static int do_check_common(struct bpf_verifier_env *env, int subprog)
18737 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
18738 struct bpf_verifier_state *state;
18739 struct bpf_reg_state *regs;
18742 env->prev_linfo = NULL;
18745 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
18748 state->curframe = 0;
18749 state->speculative = false;
18750 state->branches = 1;
18751 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
18752 if (!state->frame[0]) {
18756 env->cur_state = state;
18757 init_func_state(env, state->frame[0],
18758 BPF_MAIN_FUNC /* callsite */,
18761 state->first_insn_idx = env->subprog_info[subprog].start;
18762 state->last_insn_idx = -1;
18764 regs = state->frame[state->curframe]->regs;
18765 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
18766 ret = btf_prepare_func_args(env, subprog, regs);
18769 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
18770 if (regs[i].type == PTR_TO_CTX)
18771 mark_reg_known_zero(env, regs, i);
18772 else if (regs[i].type == SCALAR_VALUE)
18773 mark_reg_unknown(env, regs, i);
18774 else if (base_type(regs[i].type) == PTR_TO_MEM) {
18775 const u32 mem_size = regs[i].mem_size;
18777 mark_reg_known_zero(env, regs, i);
18778 regs[i].mem_size = mem_size;
18779 regs[i].id = ++env->id_gen;
18783 /* 1st arg to a function */
18784 regs[BPF_REG_1].type = PTR_TO_CTX;
18785 mark_reg_known_zero(env, regs, BPF_REG_1);
18786 ret = btf_check_subprog_arg_match(env, subprog, regs);
18787 if (ret == -EFAULT)
18788 /* unlikely verifier bug. abort.
18789 * ret == 0 and ret < 0 are sadly acceptable for
18790 * main() function due to backward compatibility.
18791 * Like socket filter program may be written as:
18792 * int bpf_prog(struct pt_regs *ctx)
18793 * and never dereference that ctx in the program.
18794 * 'struct pt_regs' is a type mismatch for socket
18795 * filter that should be using 'struct __sk_buff'.
18800 ret = do_check(env);
18802 /* check for NULL is necessary, since cur_state can be freed inside
18803 * do_check() under memory pressure.
18805 if (env->cur_state) {
18806 free_verifier_state(env->cur_state, true);
18807 env->cur_state = NULL;
18809 while (!pop_stack(env, NULL, NULL, false));
18810 if (!ret && pop_log)
18811 bpf_vlog_reset(&env->log, 0);
18816 /* Verify all global functions in a BPF program one by one based on their BTF.
18817 * All global functions must pass verification. Otherwise the whole program is rejected.
18828 * foo() will be verified first for R1=any_scalar_value. During verification it
18829 * will be assumed that bar() already verified successfully and call to bar()
18830 * from foo() will be checked for type match only. Later bar() will be verified
18831 * independently to check that it's safe for R1=any_scalar_value.
18833 static int do_check_subprogs(struct bpf_verifier_env *env)
18835 struct bpf_prog_aux *aux = env->prog->aux;
18838 if (!aux->func_info)
18841 for (i = 1; i < env->subprog_cnt; i++) {
18842 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
18844 env->insn_idx = env->subprog_info[i].start;
18845 WARN_ON_ONCE(env->insn_idx == 0);
18846 ret = do_check_common(env, i);
18849 } else if (env->log.level & BPF_LOG_LEVEL) {
18851 "Func#%d is safe for any args that match its prototype\n",
18858 static int do_check_main(struct bpf_verifier_env *env)
18863 ret = do_check_common(env, 0);
18865 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
18870 static void print_verification_stats(struct bpf_verifier_env *env)
18874 if (env->log.level & BPF_LOG_STATS) {
18875 verbose(env, "verification time %lld usec\n",
18876 div_u64(env->verification_time, 1000));
18877 verbose(env, "stack depth ");
18878 for (i = 0; i < env->subprog_cnt; i++) {
18879 u32 depth = env->subprog_info[i].stack_depth;
18881 verbose(env, "%d", depth);
18882 if (i + 1 < env->subprog_cnt)
18885 verbose(env, "\n");
18887 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
18888 "total_states %d peak_states %d mark_read %d\n",
18889 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
18890 env->max_states_per_insn, env->total_states,
18891 env->peak_states, env->longest_mark_read_walk);
18894 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
18896 const struct btf_type *t, *func_proto;
18897 const struct bpf_struct_ops *st_ops;
18898 const struct btf_member *member;
18899 struct bpf_prog *prog = env->prog;
18900 u32 btf_id, member_idx;
18903 if (!prog->gpl_compatible) {
18904 verbose(env, "struct ops programs must have a GPL compatible license\n");
18908 btf_id = prog->aux->attach_btf_id;
18909 st_ops = bpf_struct_ops_find(btf_id);
18911 verbose(env, "attach_btf_id %u is not a supported struct\n",
18917 member_idx = prog->expected_attach_type;
18918 if (member_idx >= btf_type_vlen(t)) {
18919 verbose(env, "attach to invalid member idx %u of struct %s\n",
18920 member_idx, st_ops->name);
18924 member = &btf_type_member(t)[member_idx];
18925 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
18926 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
18929 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
18930 mname, member_idx, st_ops->name);
18934 if (st_ops->check_member) {
18935 int err = st_ops->check_member(t, member, prog);
18938 verbose(env, "attach to unsupported member %s of struct %s\n",
18939 mname, st_ops->name);
18944 prog->aux->attach_func_proto = func_proto;
18945 prog->aux->attach_func_name = mname;
18946 env->ops = st_ops->verifier_ops;
18950 #define SECURITY_PREFIX "security_"
18952 static int check_attach_modify_return(unsigned long addr, const char *func_name)
18954 if (within_error_injection_list(addr) ||
18955 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
18961 /* list of non-sleepable functions that are otherwise on
18962 * ALLOW_ERROR_INJECTION list
18964 BTF_SET_START(btf_non_sleepable_error_inject)
18965 /* Three functions below can be called from sleepable and non-sleepable context.
18966 * Assume non-sleepable from bpf safety point of view.
18968 BTF_ID(func, __filemap_add_folio)
18969 BTF_ID(func, should_fail_alloc_page)
18970 BTF_ID(func, should_failslab)
18971 BTF_SET_END(btf_non_sleepable_error_inject)
18973 static int check_non_sleepable_error_inject(u32 btf_id)
18975 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
18978 int bpf_check_attach_target(struct bpf_verifier_log *log,
18979 const struct bpf_prog *prog,
18980 const struct bpf_prog *tgt_prog,
18982 struct bpf_attach_target_info *tgt_info)
18984 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
18985 const char prefix[] = "btf_trace_";
18986 int ret = 0, subprog = -1, i;
18987 const struct btf_type *t;
18988 bool conservative = true;
18992 struct module *mod = NULL;
18995 bpf_log(log, "Tracing programs must provide btf_id\n");
18998 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
19001 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
19004 t = btf_type_by_id(btf, btf_id);
19006 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
19009 tname = btf_name_by_offset(btf, t->name_off);
19011 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
19015 struct bpf_prog_aux *aux = tgt_prog->aux;
19017 if (bpf_prog_is_dev_bound(prog->aux) &&
19018 !bpf_prog_dev_bound_match(prog, tgt_prog)) {
19019 bpf_log(log, "Target program bound device mismatch");
19023 for (i = 0; i < aux->func_info_cnt; i++)
19024 if (aux->func_info[i].type_id == btf_id) {
19028 if (subprog == -1) {
19029 bpf_log(log, "Subprog %s doesn't exist\n", tname);
19032 conservative = aux->func_info_aux[subprog].unreliable;
19033 if (prog_extension) {
19034 if (conservative) {
19036 "Cannot replace static functions\n");
19039 if (!prog->jit_requested) {
19041 "Extension programs should be JITed\n");
19045 if (!tgt_prog->jited) {
19046 bpf_log(log, "Can attach to only JITed progs\n");
19049 if (tgt_prog->type == prog->type) {
19050 /* Cannot fentry/fexit another fentry/fexit program.
19051 * Cannot attach program extension to another extension.
19052 * It's ok to attach fentry/fexit to extension program.
19054 bpf_log(log, "Cannot recursively attach\n");
19057 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
19059 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
19060 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
19061 /* Program extensions can extend all program types
19062 * except fentry/fexit. The reason is the following.
19063 * The fentry/fexit programs are used for performance
19064 * analysis, stats and can be attached to any program
19065 * type except themselves. When extension program is
19066 * replacing XDP function it is necessary to allow
19067 * performance analysis of all functions. Both original
19068 * XDP program and its program extension. Hence
19069 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
19070 * allowed. If extending of fentry/fexit was allowed it
19071 * would be possible to create long call chain
19072 * fentry->extension->fentry->extension beyond
19073 * reasonable stack size. Hence extending fentry is not
19076 bpf_log(log, "Cannot extend fentry/fexit\n");
19080 if (prog_extension) {
19081 bpf_log(log, "Cannot replace kernel functions\n");
19086 switch (prog->expected_attach_type) {
19087 case BPF_TRACE_RAW_TP:
19090 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
19093 if (!btf_type_is_typedef(t)) {
19094 bpf_log(log, "attach_btf_id %u is not a typedef\n",
19098 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
19099 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
19103 tname += sizeof(prefix) - 1;
19104 t = btf_type_by_id(btf, t->type);
19105 if (!btf_type_is_ptr(t))
19106 /* should never happen in valid vmlinux build */
19108 t = btf_type_by_id(btf, t->type);
19109 if (!btf_type_is_func_proto(t))
19110 /* should never happen in valid vmlinux build */
19114 case BPF_TRACE_ITER:
19115 if (!btf_type_is_func(t)) {
19116 bpf_log(log, "attach_btf_id %u is not a function\n",
19120 t = btf_type_by_id(btf, t->type);
19121 if (!btf_type_is_func_proto(t))
19123 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
19128 if (!prog_extension)
19131 case BPF_MODIFY_RETURN:
19133 case BPF_LSM_CGROUP:
19134 case BPF_TRACE_FENTRY:
19135 case BPF_TRACE_FEXIT:
19136 if (!btf_type_is_func(t)) {
19137 bpf_log(log, "attach_btf_id %u is not a function\n",
19141 if (prog_extension &&
19142 btf_check_type_match(log, prog, btf, t))
19144 t = btf_type_by_id(btf, t->type);
19145 if (!btf_type_is_func_proto(t))
19148 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
19149 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
19150 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
19153 if (tgt_prog && conservative)
19156 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
19162 addr = (long) tgt_prog->bpf_func;
19164 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
19166 if (btf_is_module(btf)) {
19167 mod = btf_try_get_module(btf);
19169 addr = find_kallsyms_symbol_value(mod, tname);
19173 addr = kallsyms_lookup_name(tname);
19178 "The address of function %s cannot be found\n",
19184 if (prog->aux->sleepable) {
19186 switch (prog->type) {
19187 case BPF_PROG_TYPE_TRACING:
19189 /* fentry/fexit/fmod_ret progs can be sleepable if they are
19190 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
19192 if (!check_non_sleepable_error_inject(btf_id) &&
19193 within_error_injection_list(addr))
19195 /* fentry/fexit/fmod_ret progs can also be sleepable if they are
19196 * in the fmodret id set with the KF_SLEEPABLE flag.
19199 u32 *flags = btf_kfunc_is_modify_return(btf, btf_id,
19202 if (flags && (*flags & KF_SLEEPABLE))
19206 case BPF_PROG_TYPE_LSM:
19207 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
19208 * Only some of them are sleepable.
19210 if (bpf_lsm_is_sleepable_hook(btf_id))
19218 bpf_log(log, "%s is not sleepable\n", tname);
19221 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
19224 bpf_log(log, "can't modify return codes of BPF programs\n");
19228 if (btf_kfunc_is_modify_return(btf, btf_id, prog) ||
19229 !check_attach_modify_return(addr, tname))
19233 bpf_log(log, "%s() is not modifiable\n", tname);
19240 tgt_info->tgt_addr = addr;
19241 tgt_info->tgt_name = tname;
19242 tgt_info->tgt_type = t;
19243 tgt_info->tgt_mod = mod;
19247 BTF_SET_START(btf_id_deny)
19250 BTF_ID(func, migrate_disable)
19251 BTF_ID(func, migrate_enable)
19253 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
19254 BTF_ID(func, rcu_read_unlock_strict)
19256 #if defined(CONFIG_DEBUG_PREEMPT) || defined(CONFIG_TRACE_PREEMPT_TOGGLE)
19257 BTF_ID(func, preempt_count_add)
19258 BTF_ID(func, preempt_count_sub)
19260 #ifdef CONFIG_PREEMPT_RCU
19261 BTF_ID(func, __rcu_read_lock)
19262 BTF_ID(func, __rcu_read_unlock)
19264 BTF_SET_END(btf_id_deny)
19266 static bool can_be_sleepable(struct bpf_prog *prog)
19268 if (prog->type == BPF_PROG_TYPE_TRACING) {
19269 switch (prog->expected_attach_type) {
19270 case BPF_TRACE_FENTRY:
19271 case BPF_TRACE_FEXIT:
19272 case BPF_MODIFY_RETURN:
19273 case BPF_TRACE_ITER:
19279 return prog->type == BPF_PROG_TYPE_LSM ||
19280 prog->type == BPF_PROG_TYPE_KPROBE /* only for uprobes */ ||
19281 prog->type == BPF_PROG_TYPE_STRUCT_OPS;
19284 static int check_attach_btf_id(struct bpf_verifier_env *env)
19286 struct bpf_prog *prog = env->prog;
19287 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
19288 struct bpf_attach_target_info tgt_info = {};
19289 u32 btf_id = prog->aux->attach_btf_id;
19290 struct bpf_trampoline *tr;
19294 if (prog->type == BPF_PROG_TYPE_SYSCALL) {
19295 if (prog->aux->sleepable)
19296 /* attach_btf_id checked to be zero already */
19298 verbose(env, "Syscall programs can only be sleepable\n");
19302 if (prog->aux->sleepable && !can_be_sleepable(prog)) {
19303 verbose(env, "Only fentry/fexit/fmod_ret, lsm, iter, uprobe, and struct_ops programs can be sleepable\n");
19307 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
19308 return check_struct_ops_btf_id(env);
19310 if (prog->type != BPF_PROG_TYPE_TRACING &&
19311 prog->type != BPF_PROG_TYPE_LSM &&
19312 prog->type != BPF_PROG_TYPE_EXT)
19315 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
19319 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
19320 /* to make freplace equivalent to their targets, they need to
19321 * inherit env->ops and expected_attach_type for the rest of the
19324 env->ops = bpf_verifier_ops[tgt_prog->type];
19325 prog->expected_attach_type = tgt_prog->expected_attach_type;
19328 /* store info about the attachment target that will be used later */
19329 prog->aux->attach_func_proto = tgt_info.tgt_type;
19330 prog->aux->attach_func_name = tgt_info.tgt_name;
19331 prog->aux->mod = tgt_info.tgt_mod;
19334 prog->aux->saved_dst_prog_type = tgt_prog->type;
19335 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
19338 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
19339 prog->aux->attach_btf_trace = true;
19341 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
19342 if (!bpf_iter_prog_supported(prog))
19347 if (prog->type == BPF_PROG_TYPE_LSM) {
19348 ret = bpf_lsm_verify_prog(&env->log, prog);
19351 } else if (prog->type == BPF_PROG_TYPE_TRACING &&
19352 btf_id_set_contains(&btf_id_deny, btf_id)) {
19356 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
19357 tr = bpf_trampoline_get(key, &tgt_info);
19361 prog->aux->dst_trampoline = tr;
19365 struct btf *bpf_get_btf_vmlinux(void)
19367 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
19368 mutex_lock(&bpf_verifier_lock);
19370 btf_vmlinux = btf_parse_vmlinux();
19371 mutex_unlock(&bpf_verifier_lock);
19373 return btf_vmlinux;
19376 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr, __u32 uattr_size)
19378 u64 start_time = ktime_get_ns();
19379 struct bpf_verifier_env *env;
19380 int i, len, ret = -EINVAL, err;
19384 /* no program is valid */
19385 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
19388 /* 'struct bpf_verifier_env' can be global, but since it's not small,
19389 * allocate/free it every time bpf_check() is called
19391 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
19397 len = (*prog)->len;
19398 env->insn_aux_data =
19399 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
19401 if (!env->insn_aux_data)
19403 for (i = 0; i < len; i++)
19404 env->insn_aux_data[i].orig_idx = i;
19406 env->ops = bpf_verifier_ops[env->prog->type];
19407 env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
19408 is_priv = bpf_capable();
19410 bpf_get_btf_vmlinux();
19412 /* grab the mutex to protect few globals used by verifier */
19414 mutex_lock(&bpf_verifier_lock);
19416 /* user could have requested verbose verifier output
19417 * and supplied buffer to store the verification trace
19419 ret = bpf_vlog_init(&env->log, attr->log_level,
19420 (char __user *) (unsigned long) attr->log_buf,
19425 mark_verifier_state_clean(env);
19427 if (IS_ERR(btf_vmlinux)) {
19428 /* Either gcc or pahole or kernel are broken. */
19429 verbose(env, "in-kernel BTF is malformed\n");
19430 ret = PTR_ERR(btf_vmlinux);
19431 goto skip_full_check;
19434 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
19435 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
19436 env->strict_alignment = true;
19437 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
19438 env->strict_alignment = false;
19440 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
19441 env->allow_uninit_stack = bpf_allow_uninit_stack();
19442 env->bypass_spec_v1 = bpf_bypass_spec_v1();
19443 env->bypass_spec_v4 = bpf_bypass_spec_v4();
19444 env->bpf_capable = bpf_capable();
19447 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
19449 env->explored_states = kvcalloc(state_htab_size(env),
19450 sizeof(struct bpf_verifier_state_list *),
19453 if (!env->explored_states)
19454 goto skip_full_check;
19456 ret = add_subprog_and_kfunc(env);
19458 goto skip_full_check;
19460 ret = check_subprogs(env);
19462 goto skip_full_check;
19464 ret = check_btf_info(env, attr, uattr);
19466 goto skip_full_check;
19468 ret = check_attach_btf_id(env);
19470 goto skip_full_check;
19472 ret = resolve_pseudo_ldimm64(env);
19474 goto skip_full_check;
19476 if (bpf_prog_is_offloaded(env->prog->aux)) {
19477 ret = bpf_prog_offload_verifier_prep(env->prog);
19479 goto skip_full_check;
19482 ret = check_cfg(env);
19484 goto skip_full_check;
19486 ret = do_check_subprogs(env);
19487 ret = ret ?: do_check_main(env);
19489 if (ret == 0 && bpf_prog_is_offloaded(env->prog->aux))
19490 ret = bpf_prog_offload_finalize(env);
19493 kvfree(env->explored_states);
19496 ret = check_max_stack_depth(env);
19498 /* instruction rewrites happen after this point */
19500 ret = optimize_bpf_loop(env);
19504 opt_hard_wire_dead_code_branches(env);
19506 ret = opt_remove_dead_code(env);
19508 ret = opt_remove_nops(env);
19511 sanitize_dead_code(env);
19515 /* program is valid, convert *(u32*)(ctx + off) accesses */
19516 ret = convert_ctx_accesses(env);
19519 ret = do_misc_fixups(env);
19521 /* do 32-bit optimization after insn patching has done so those patched
19522 * insns could be handled correctly.
19524 if (ret == 0 && !bpf_prog_is_offloaded(env->prog->aux)) {
19525 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
19526 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
19531 ret = fixup_call_args(env);
19533 env->verification_time = ktime_get_ns() - start_time;
19534 print_verification_stats(env);
19535 env->prog->aux->verified_insns = env->insn_processed;
19537 /* preserve original error even if log finalization is successful */
19538 err = bpf_vlog_finalize(&env->log, &log_true_size);
19542 if (uattr_size >= offsetofend(union bpf_attr, log_true_size) &&
19543 copy_to_bpfptr_offset(uattr, offsetof(union bpf_attr, log_true_size),
19544 &log_true_size, sizeof(log_true_size))) {
19546 goto err_release_maps;
19550 goto err_release_maps;
19552 if (env->used_map_cnt) {
19553 /* if program passed verifier, update used_maps in bpf_prog_info */
19554 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
19555 sizeof(env->used_maps[0]),
19558 if (!env->prog->aux->used_maps) {
19560 goto err_release_maps;
19563 memcpy(env->prog->aux->used_maps, env->used_maps,
19564 sizeof(env->used_maps[0]) * env->used_map_cnt);
19565 env->prog->aux->used_map_cnt = env->used_map_cnt;
19567 if (env->used_btf_cnt) {
19568 /* if program passed verifier, update used_btfs in bpf_prog_aux */
19569 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
19570 sizeof(env->used_btfs[0]),
19572 if (!env->prog->aux->used_btfs) {
19574 goto err_release_maps;
19577 memcpy(env->prog->aux->used_btfs, env->used_btfs,
19578 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
19579 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
19581 if (env->used_map_cnt || env->used_btf_cnt) {
19582 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
19583 * bpf_ld_imm64 instructions
19585 convert_pseudo_ld_imm64(env);
19588 adjust_btf_func(env);
19591 if (!env->prog->aux->used_maps)
19592 /* if we didn't copy map pointers into bpf_prog_info, release
19593 * them now. Otherwise free_used_maps() will release them.
19596 if (!env->prog->aux->used_btfs)
19599 /* extension progs temporarily inherit the attach_type of their targets
19600 for verification purposes, so set it back to zero before returning
19602 if (env->prog->type == BPF_PROG_TYPE_EXT)
19603 env->prog->expected_attach_type = 0;
19608 mutex_unlock(&bpf_verifier_lock);
19609 vfree(env->insn_aux_data);