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>
28 #include <linux/cpumask.h>
33 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
34 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
35 [_id] = & _name ## _verifier_ops,
36 #define BPF_MAP_TYPE(_id, _ops)
37 #define BPF_LINK_TYPE(_id, _name)
38 #include <linux/bpf_types.h>
44 /* bpf_check() is a static code analyzer that walks eBPF program
45 * instruction by instruction and updates register/stack state.
46 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
48 * The first pass is depth-first-search to check that the program is a DAG.
49 * It rejects the following programs:
50 * - larger than BPF_MAXINSNS insns
51 * - if loop is present (detected via back-edge)
52 * - unreachable insns exist (shouldn't be a forest. program = one function)
53 * - out of bounds or malformed jumps
54 * The second pass is all possible path descent from the 1st insn.
55 * Since it's analyzing all paths through the program, the length of the
56 * analysis is limited to 64k insn, which may be hit even if total number of
57 * insn is less then 4K, but there are too many branches that change stack/regs.
58 * Number of 'branches to be analyzed' is limited to 1k
60 * On entry to each instruction, each register has a type, and the instruction
61 * changes the types of the registers depending on instruction semantics.
62 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
65 * All registers are 64-bit.
66 * R0 - return register
67 * R1-R5 argument passing registers
68 * R6-R9 callee saved registers
69 * R10 - frame pointer read-only
71 * At the start of BPF program the register R1 contains a pointer to bpf_context
72 * and has type PTR_TO_CTX.
74 * Verifier tracks arithmetic operations on pointers in case:
75 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
76 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
77 * 1st insn copies R10 (which has FRAME_PTR) type into R1
78 * and 2nd arithmetic instruction is pattern matched to recognize
79 * that it wants to construct a pointer to some element within stack.
80 * So after 2nd insn, the register R1 has type PTR_TO_STACK
81 * (and -20 constant is saved for further stack bounds checking).
82 * Meaning that this reg is a pointer to stack plus known immediate constant.
84 * Most of the time the registers have SCALAR_VALUE type, which
85 * means the register has some value, but it's not a valid pointer.
86 * (like pointer plus pointer becomes SCALAR_VALUE type)
88 * When verifier sees load or store instructions the type of base register
89 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
90 * four pointer types recognized by check_mem_access() function.
92 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
93 * and the range of [ptr, ptr + map's value_size) is accessible.
95 * registers used to pass values to function calls are checked against
96 * function argument constraints.
98 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
99 * It means that the register type passed to this function must be
100 * PTR_TO_STACK and it will be used inside the function as
101 * 'pointer to map element key'
103 * For example the argument constraints for bpf_map_lookup_elem():
104 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
105 * .arg1_type = ARG_CONST_MAP_PTR,
106 * .arg2_type = ARG_PTR_TO_MAP_KEY,
108 * ret_type says that this function returns 'pointer to map elem value or null'
109 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
110 * 2nd argument should be a pointer to stack, which will be used inside
111 * the helper function as a pointer to map element key.
113 * On the kernel side the helper function looks like:
114 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
116 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
117 * void *key = (void *) (unsigned long) r2;
120 * here kernel can access 'key' and 'map' pointers safely, knowing that
121 * [key, key + map->key_size) bytes are valid and were initialized on
122 * the stack of eBPF program.
125 * Corresponding eBPF program may look like:
126 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
127 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
128 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
129 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
130 * here verifier looks at prototype of map_lookup_elem() and sees:
131 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
132 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
134 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
135 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
136 * and were initialized prior to this call.
137 * If it's ok, then verifier allows this BPF_CALL insn and looks at
138 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
139 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
140 * returns either pointer to map value or NULL.
142 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
143 * insn, the register holding that pointer in the true branch changes state to
144 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
145 * branch. See check_cond_jmp_op().
147 * After the call R0 is set to return type of the function and registers R1-R5
148 * are set to NOT_INIT to indicate that they are no longer readable.
150 * The following reference types represent a potential reference to a kernel
151 * resource which, after first being allocated, must be checked and freed by
153 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
155 * When the verifier sees a helper call return a reference type, it allocates a
156 * pointer id for the reference and stores it in the current function state.
157 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
158 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
159 * passes through a NULL-check conditional. For the branch wherein the state is
160 * changed to CONST_IMM, the verifier releases the reference.
162 * For each helper function that allocates a reference, such as
163 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
164 * bpf_sk_release(). When a reference type passes into the release function,
165 * the verifier also releases the reference. If any unchecked or unreleased
166 * reference remains at the end of the program, the verifier rejects it.
169 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
170 struct bpf_verifier_stack_elem {
171 /* verifer state is 'st'
172 * before processing instruction 'insn_idx'
173 * and after processing instruction 'prev_insn_idx'
175 struct bpf_verifier_state st;
178 struct bpf_verifier_stack_elem *next;
179 /* length of verifier log at the time this state was pushed on stack */
183 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
184 #define BPF_COMPLEXITY_LIMIT_STATES 64
186 #define BPF_MAP_KEY_POISON (1ULL << 63)
187 #define BPF_MAP_KEY_SEEN (1ULL << 62)
189 #define BPF_MAP_PTR_UNPRIV 1UL
190 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
191 POISON_POINTER_DELTA))
192 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
194 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx);
195 static int release_reference(struct bpf_verifier_env *env, int ref_obj_id);
196 static void invalidate_non_owning_refs(struct bpf_verifier_env *env);
197 static bool in_rbtree_lock_required_cb(struct bpf_verifier_env *env);
198 static int ref_set_non_owning(struct bpf_verifier_env *env,
199 struct bpf_reg_state *reg);
200 static void specialize_kfunc(struct bpf_verifier_env *env,
201 u32 func_id, u16 offset, unsigned long *addr);
202 static bool is_trusted_reg(const struct bpf_reg_state *reg);
204 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
206 return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON;
209 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
211 return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV;
214 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
215 const struct bpf_map *map, bool unpriv)
217 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
218 unpriv |= bpf_map_ptr_unpriv(aux);
219 aux->map_ptr_state = (unsigned long)map |
220 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
223 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
225 return aux->map_key_state & BPF_MAP_KEY_POISON;
228 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
230 return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
233 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
235 return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
238 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
240 bool poisoned = bpf_map_key_poisoned(aux);
242 aux->map_key_state = state | BPF_MAP_KEY_SEEN |
243 (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
246 static bool bpf_helper_call(const struct bpf_insn *insn)
248 return insn->code == (BPF_JMP | BPF_CALL) &&
252 static bool bpf_pseudo_call(const struct bpf_insn *insn)
254 return insn->code == (BPF_JMP | BPF_CALL) &&
255 insn->src_reg == BPF_PSEUDO_CALL;
258 static bool bpf_pseudo_kfunc_call(const struct bpf_insn *insn)
260 return insn->code == (BPF_JMP | BPF_CALL) &&
261 insn->src_reg == BPF_PSEUDO_KFUNC_CALL;
264 struct bpf_call_arg_meta {
265 struct bpf_map *map_ptr;
282 struct btf_field *kptr_field;
285 struct bpf_kfunc_call_arg_meta {
290 const struct btf_type *func_proto;
291 const char *func_name;
304 /* arg_{btf,btf_id,owning_ref} are used by kfunc-specific handling,
305 * generally to pass info about user-defined local kptr types to later
308 * Record the local kptr type to be drop'd
309 * bpf_refcount_acquire (via KF_ARG_PTR_TO_REFCOUNTED_KPTR arg type)
310 * Record the local kptr type to be refcount_incr'd and use
311 * arg_owning_ref to determine whether refcount_acquire should be
319 struct btf_field *field;
322 struct btf_field *field;
325 enum bpf_dynptr_type type;
328 } initialized_dynptr;
336 struct btf *btf_vmlinux;
338 static DEFINE_MUTEX(bpf_verifier_lock);
340 static const struct bpf_line_info *
341 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
343 const struct bpf_line_info *linfo;
344 const struct bpf_prog *prog;
348 nr_linfo = prog->aux->nr_linfo;
350 if (!nr_linfo || insn_off >= prog->len)
353 linfo = prog->aux->linfo;
354 for (i = 1; i < nr_linfo; i++)
355 if (insn_off < linfo[i].insn_off)
358 return &linfo[i - 1];
361 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
363 struct bpf_verifier_env *env = private_data;
366 if (!bpf_verifier_log_needed(&env->log))
370 bpf_verifier_vlog(&env->log, fmt, args);
374 static const char *ltrim(const char *s)
382 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
384 const char *prefix_fmt, ...)
386 const struct bpf_line_info *linfo;
388 if (!bpf_verifier_log_needed(&env->log))
391 linfo = find_linfo(env, insn_off);
392 if (!linfo || linfo == env->prev_linfo)
398 va_start(args, prefix_fmt);
399 bpf_verifier_vlog(&env->log, prefix_fmt, args);
404 ltrim(btf_name_by_offset(env->prog->aux->btf,
407 env->prev_linfo = linfo;
410 static void verbose_invalid_scalar(struct bpf_verifier_env *env,
411 struct bpf_reg_state *reg,
412 struct tnum *range, const char *ctx,
413 const char *reg_name)
417 verbose(env, "At %s the register %s ", ctx, reg_name);
418 if (!tnum_is_unknown(reg->var_off)) {
419 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
420 verbose(env, "has value %s", tn_buf);
422 verbose(env, "has unknown scalar value");
424 tnum_strn(tn_buf, sizeof(tn_buf), *range);
425 verbose(env, " should have been in %s\n", tn_buf);
428 static bool type_is_pkt_pointer(enum bpf_reg_type type)
430 type = base_type(type);
431 return type == PTR_TO_PACKET ||
432 type == PTR_TO_PACKET_META;
435 static bool type_is_sk_pointer(enum bpf_reg_type type)
437 return type == PTR_TO_SOCKET ||
438 type == PTR_TO_SOCK_COMMON ||
439 type == PTR_TO_TCP_SOCK ||
440 type == PTR_TO_XDP_SOCK;
443 static bool type_may_be_null(u32 type)
445 return type & PTR_MAYBE_NULL;
448 static bool reg_not_null(const struct bpf_reg_state *reg)
450 enum bpf_reg_type type;
453 if (type_may_be_null(type))
456 type = base_type(type);
457 return type == PTR_TO_SOCKET ||
458 type == PTR_TO_TCP_SOCK ||
459 type == PTR_TO_MAP_VALUE ||
460 type == PTR_TO_MAP_KEY ||
461 type == PTR_TO_SOCK_COMMON ||
462 (type == PTR_TO_BTF_ID && is_trusted_reg(reg)) ||
466 static bool type_is_ptr_alloc_obj(u32 type)
468 return base_type(type) == PTR_TO_BTF_ID && type_flag(type) & MEM_ALLOC;
471 static bool type_is_non_owning_ref(u32 type)
473 return type_is_ptr_alloc_obj(type) && type_flag(type) & NON_OWN_REF;
476 static struct btf_record *reg_btf_record(const struct bpf_reg_state *reg)
478 struct btf_record *rec = NULL;
479 struct btf_struct_meta *meta;
481 if (reg->type == PTR_TO_MAP_VALUE) {
482 rec = reg->map_ptr->record;
483 } else if (type_is_ptr_alloc_obj(reg->type)) {
484 meta = btf_find_struct_meta(reg->btf, reg->btf_id);
491 static bool subprog_is_global(const struct bpf_verifier_env *env, int subprog)
493 struct bpf_func_info_aux *aux = env->prog->aux->func_info_aux;
495 return aux && aux[subprog].linkage == BTF_FUNC_GLOBAL;
498 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
500 return btf_record_has_field(reg_btf_record(reg), BPF_SPIN_LOCK);
503 static bool type_is_rdonly_mem(u32 type)
505 return type & MEM_RDONLY;
508 static bool is_acquire_function(enum bpf_func_id func_id,
509 const struct bpf_map *map)
511 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
513 if (func_id == BPF_FUNC_sk_lookup_tcp ||
514 func_id == BPF_FUNC_sk_lookup_udp ||
515 func_id == BPF_FUNC_skc_lookup_tcp ||
516 func_id == BPF_FUNC_ringbuf_reserve ||
517 func_id == BPF_FUNC_kptr_xchg)
520 if (func_id == BPF_FUNC_map_lookup_elem &&
521 (map_type == BPF_MAP_TYPE_SOCKMAP ||
522 map_type == BPF_MAP_TYPE_SOCKHASH))
528 static bool is_ptr_cast_function(enum bpf_func_id func_id)
530 return func_id == BPF_FUNC_tcp_sock ||
531 func_id == BPF_FUNC_sk_fullsock ||
532 func_id == BPF_FUNC_skc_to_tcp_sock ||
533 func_id == BPF_FUNC_skc_to_tcp6_sock ||
534 func_id == BPF_FUNC_skc_to_udp6_sock ||
535 func_id == BPF_FUNC_skc_to_mptcp_sock ||
536 func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
537 func_id == BPF_FUNC_skc_to_tcp_request_sock;
540 static bool is_dynptr_ref_function(enum bpf_func_id func_id)
542 return func_id == BPF_FUNC_dynptr_data;
545 static bool is_callback_calling_kfunc(u32 btf_id);
547 static bool is_callback_calling_function(enum bpf_func_id func_id)
549 return func_id == BPF_FUNC_for_each_map_elem ||
550 func_id == BPF_FUNC_timer_set_callback ||
551 func_id == BPF_FUNC_find_vma ||
552 func_id == BPF_FUNC_loop ||
553 func_id == BPF_FUNC_user_ringbuf_drain;
556 static bool is_async_callback_calling_function(enum bpf_func_id func_id)
558 return func_id == BPF_FUNC_timer_set_callback;
561 static bool is_storage_get_function(enum bpf_func_id func_id)
563 return func_id == BPF_FUNC_sk_storage_get ||
564 func_id == BPF_FUNC_inode_storage_get ||
565 func_id == BPF_FUNC_task_storage_get ||
566 func_id == BPF_FUNC_cgrp_storage_get;
569 static bool helper_multiple_ref_obj_use(enum bpf_func_id func_id,
570 const struct bpf_map *map)
572 int ref_obj_uses = 0;
574 if (is_ptr_cast_function(func_id))
576 if (is_acquire_function(func_id, map))
578 if (is_dynptr_ref_function(func_id))
581 return ref_obj_uses > 1;
584 static bool is_cmpxchg_insn(const struct bpf_insn *insn)
586 return BPF_CLASS(insn->code) == BPF_STX &&
587 BPF_MODE(insn->code) == BPF_ATOMIC &&
588 insn->imm == BPF_CMPXCHG;
591 /* string representation of 'enum bpf_reg_type'
593 * Note that reg_type_str() can not appear more than once in a single verbose()
596 static const char *reg_type_str(struct bpf_verifier_env *env,
597 enum bpf_reg_type type)
599 char postfix[16] = {0}, prefix[64] = {0};
600 static const char * const str[] = {
602 [SCALAR_VALUE] = "scalar",
603 [PTR_TO_CTX] = "ctx",
604 [CONST_PTR_TO_MAP] = "map_ptr",
605 [PTR_TO_MAP_VALUE] = "map_value",
606 [PTR_TO_STACK] = "fp",
607 [PTR_TO_PACKET] = "pkt",
608 [PTR_TO_PACKET_META] = "pkt_meta",
609 [PTR_TO_PACKET_END] = "pkt_end",
610 [PTR_TO_FLOW_KEYS] = "flow_keys",
611 [PTR_TO_SOCKET] = "sock",
612 [PTR_TO_SOCK_COMMON] = "sock_common",
613 [PTR_TO_TCP_SOCK] = "tcp_sock",
614 [PTR_TO_TP_BUFFER] = "tp_buffer",
615 [PTR_TO_XDP_SOCK] = "xdp_sock",
616 [PTR_TO_BTF_ID] = "ptr_",
617 [PTR_TO_MEM] = "mem",
618 [PTR_TO_BUF] = "buf",
619 [PTR_TO_FUNC] = "func",
620 [PTR_TO_MAP_KEY] = "map_key",
621 [CONST_PTR_TO_DYNPTR] = "dynptr_ptr",
624 if (type & PTR_MAYBE_NULL) {
625 if (base_type(type) == PTR_TO_BTF_ID)
626 strncpy(postfix, "or_null_", 16);
628 strncpy(postfix, "_or_null", 16);
631 snprintf(prefix, sizeof(prefix), "%s%s%s%s%s%s%s",
632 type & MEM_RDONLY ? "rdonly_" : "",
633 type & MEM_RINGBUF ? "ringbuf_" : "",
634 type & MEM_USER ? "user_" : "",
635 type & MEM_PERCPU ? "percpu_" : "",
636 type & MEM_RCU ? "rcu_" : "",
637 type & PTR_UNTRUSTED ? "untrusted_" : "",
638 type & PTR_TRUSTED ? "trusted_" : ""
641 snprintf(env->tmp_str_buf, TMP_STR_BUF_LEN, "%s%s%s",
642 prefix, str[base_type(type)], postfix);
643 return env->tmp_str_buf;
646 static char slot_type_char[] = {
647 [STACK_INVALID] = '?',
651 [STACK_DYNPTR] = 'd',
655 static void print_liveness(struct bpf_verifier_env *env,
656 enum bpf_reg_liveness live)
658 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
660 if (live & REG_LIVE_READ)
662 if (live & REG_LIVE_WRITTEN)
664 if (live & REG_LIVE_DONE)
668 static int __get_spi(s32 off)
670 return (-off - 1) / BPF_REG_SIZE;
673 static struct bpf_func_state *func(struct bpf_verifier_env *env,
674 const struct bpf_reg_state *reg)
676 struct bpf_verifier_state *cur = env->cur_state;
678 return cur->frame[reg->frameno];
681 static bool is_spi_bounds_valid(struct bpf_func_state *state, int spi, int nr_slots)
683 int allocated_slots = state->allocated_stack / BPF_REG_SIZE;
685 /* We need to check that slots between [spi - nr_slots + 1, spi] are
686 * within [0, allocated_stack).
688 * Please note that the spi grows downwards. For example, a dynptr
689 * takes the size of two stack slots; the first slot will be at
690 * spi and the second slot will be at spi - 1.
692 return spi - nr_slots + 1 >= 0 && spi < allocated_slots;
695 static int stack_slot_obj_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
696 const char *obj_kind, int nr_slots)
700 if (!tnum_is_const(reg->var_off)) {
701 verbose(env, "%s has to be at a constant offset\n", obj_kind);
705 off = reg->off + reg->var_off.value;
706 if (off % BPF_REG_SIZE) {
707 verbose(env, "cannot pass in %s at an offset=%d\n", obj_kind, off);
711 spi = __get_spi(off);
712 if (spi + 1 < nr_slots) {
713 verbose(env, "cannot pass in %s at an offset=%d\n", obj_kind, off);
717 if (!is_spi_bounds_valid(func(env, reg), spi, nr_slots))
722 static int dynptr_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
724 return stack_slot_obj_get_spi(env, reg, "dynptr", BPF_DYNPTR_NR_SLOTS);
727 static int iter_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg, int nr_slots)
729 return stack_slot_obj_get_spi(env, reg, "iter", nr_slots);
732 static const char *btf_type_name(const struct btf *btf, u32 id)
734 return btf_name_by_offset(btf, btf_type_by_id(btf, id)->name_off);
737 static const char *dynptr_type_str(enum bpf_dynptr_type type)
740 case BPF_DYNPTR_TYPE_LOCAL:
742 case BPF_DYNPTR_TYPE_RINGBUF:
744 case BPF_DYNPTR_TYPE_SKB:
746 case BPF_DYNPTR_TYPE_XDP:
748 case BPF_DYNPTR_TYPE_INVALID:
751 WARN_ONCE(1, "unknown dynptr type %d\n", type);
756 static const char *iter_type_str(const struct btf *btf, u32 btf_id)
758 if (!btf || btf_id == 0)
761 /* we already validated that type is valid and has conforming name */
762 return btf_type_name(btf, btf_id) + sizeof(ITER_PREFIX) - 1;
765 static const char *iter_state_str(enum bpf_iter_state state)
768 case BPF_ITER_STATE_ACTIVE:
770 case BPF_ITER_STATE_DRAINED:
772 case BPF_ITER_STATE_INVALID:
775 WARN_ONCE(1, "unknown iter state %d\n", state);
780 static void mark_reg_scratched(struct bpf_verifier_env *env, u32 regno)
782 env->scratched_regs |= 1U << regno;
785 static void mark_stack_slot_scratched(struct bpf_verifier_env *env, u32 spi)
787 env->scratched_stack_slots |= 1ULL << spi;
790 static bool reg_scratched(const struct bpf_verifier_env *env, u32 regno)
792 return (env->scratched_regs >> regno) & 1;
795 static bool stack_slot_scratched(const struct bpf_verifier_env *env, u64 regno)
797 return (env->scratched_stack_slots >> regno) & 1;
800 static bool verifier_state_scratched(const struct bpf_verifier_env *env)
802 return env->scratched_regs || env->scratched_stack_slots;
805 static void mark_verifier_state_clean(struct bpf_verifier_env *env)
807 env->scratched_regs = 0U;
808 env->scratched_stack_slots = 0ULL;
811 /* Used for printing the entire verifier state. */
812 static void mark_verifier_state_scratched(struct bpf_verifier_env *env)
814 env->scratched_regs = ~0U;
815 env->scratched_stack_slots = ~0ULL;
818 static enum bpf_dynptr_type arg_to_dynptr_type(enum bpf_arg_type arg_type)
820 switch (arg_type & DYNPTR_TYPE_FLAG_MASK) {
821 case DYNPTR_TYPE_LOCAL:
822 return BPF_DYNPTR_TYPE_LOCAL;
823 case DYNPTR_TYPE_RINGBUF:
824 return BPF_DYNPTR_TYPE_RINGBUF;
825 case DYNPTR_TYPE_SKB:
826 return BPF_DYNPTR_TYPE_SKB;
827 case DYNPTR_TYPE_XDP:
828 return BPF_DYNPTR_TYPE_XDP;
830 return BPF_DYNPTR_TYPE_INVALID;
834 static enum bpf_type_flag get_dynptr_type_flag(enum bpf_dynptr_type type)
837 case BPF_DYNPTR_TYPE_LOCAL:
838 return DYNPTR_TYPE_LOCAL;
839 case BPF_DYNPTR_TYPE_RINGBUF:
840 return DYNPTR_TYPE_RINGBUF;
841 case BPF_DYNPTR_TYPE_SKB:
842 return DYNPTR_TYPE_SKB;
843 case BPF_DYNPTR_TYPE_XDP:
844 return DYNPTR_TYPE_XDP;
850 static bool dynptr_type_refcounted(enum bpf_dynptr_type type)
852 return type == BPF_DYNPTR_TYPE_RINGBUF;
855 static void __mark_dynptr_reg(struct bpf_reg_state *reg,
856 enum bpf_dynptr_type type,
857 bool first_slot, int dynptr_id);
859 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
860 struct bpf_reg_state *reg);
862 static void mark_dynptr_stack_regs(struct bpf_verifier_env *env,
863 struct bpf_reg_state *sreg1,
864 struct bpf_reg_state *sreg2,
865 enum bpf_dynptr_type type)
867 int id = ++env->id_gen;
869 __mark_dynptr_reg(sreg1, type, true, id);
870 __mark_dynptr_reg(sreg2, type, false, id);
873 static void mark_dynptr_cb_reg(struct bpf_verifier_env *env,
874 struct bpf_reg_state *reg,
875 enum bpf_dynptr_type type)
877 __mark_dynptr_reg(reg, type, true, ++env->id_gen);
880 static int destroy_if_dynptr_stack_slot(struct bpf_verifier_env *env,
881 struct bpf_func_state *state, int spi);
883 static int mark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
884 enum bpf_arg_type arg_type, int insn_idx, int clone_ref_obj_id)
886 struct bpf_func_state *state = func(env, reg);
887 enum bpf_dynptr_type type;
890 spi = dynptr_get_spi(env, reg);
894 /* We cannot assume both spi and spi - 1 belong to the same dynptr,
895 * hence we need to call destroy_if_dynptr_stack_slot twice for both,
896 * to ensure that for the following example:
899 * So marking spi = 2 should lead to destruction of both d1 and d2. In
900 * case they do belong to same dynptr, second call won't see slot_type
901 * as STACK_DYNPTR and will simply skip destruction.
903 err = destroy_if_dynptr_stack_slot(env, state, spi);
906 err = destroy_if_dynptr_stack_slot(env, state, spi - 1);
910 for (i = 0; i < BPF_REG_SIZE; i++) {
911 state->stack[spi].slot_type[i] = STACK_DYNPTR;
912 state->stack[spi - 1].slot_type[i] = STACK_DYNPTR;
915 type = arg_to_dynptr_type(arg_type);
916 if (type == BPF_DYNPTR_TYPE_INVALID)
919 mark_dynptr_stack_regs(env, &state->stack[spi].spilled_ptr,
920 &state->stack[spi - 1].spilled_ptr, type);
922 if (dynptr_type_refcounted(type)) {
923 /* The id is used to track proper releasing */
926 if (clone_ref_obj_id)
927 id = clone_ref_obj_id;
929 id = acquire_reference_state(env, insn_idx);
934 state->stack[spi].spilled_ptr.ref_obj_id = id;
935 state->stack[spi - 1].spilled_ptr.ref_obj_id = id;
938 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
939 state->stack[spi - 1].spilled_ptr.live |= REG_LIVE_WRITTEN;
944 static void invalidate_dynptr(struct bpf_verifier_env *env, struct bpf_func_state *state, int spi)
948 for (i = 0; i < BPF_REG_SIZE; i++) {
949 state->stack[spi].slot_type[i] = STACK_INVALID;
950 state->stack[spi - 1].slot_type[i] = STACK_INVALID;
953 __mark_reg_not_init(env, &state->stack[spi].spilled_ptr);
954 __mark_reg_not_init(env, &state->stack[spi - 1].spilled_ptr);
956 /* Why do we need to set REG_LIVE_WRITTEN for STACK_INVALID slot?
958 * While we don't allow reading STACK_INVALID, it is still possible to
959 * do <8 byte writes marking some but not all slots as STACK_MISC. Then,
960 * helpers or insns can do partial read of that part without failing,
961 * but check_stack_range_initialized, check_stack_read_var_off, and
962 * check_stack_read_fixed_off will do mark_reg_read for all 8-bytes of
963 * the slot conservatively. Hence we need to prevent those liveness
966 * This was not a problem before because STACK_INVALID is only set by
967 * default (where the default reg state has its reg->parent as NULL), or
968 * in clean_live_states after REG_LIVE_DONE (at which point
969 * mark_reg_read won't walk reg->parent chain), but not randomly during
970 * verifier state exploration (like we did above). Hence, for our case
971 * parentage chain will still be live (i.e. reg->parent may be
972 * non-NULL), while earlier reg->parent was NULL, so we need
973 * REG_LIVE_WRITTEN to screen off read marker propagation when it is
974 * done later on reads or by mark_dynptr_read as well to unnecessary
975 * mark registers in verifier state.
977 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
978 state->stack[spi - 1].spilled_ptr.live |= REG_LIVE_WRITTEN;
981 static int unmark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
983 struct bpf_func_state *state = func(env, reg);
984 int spi, ref_obj_id, i;
986 spi = dynptr_get_spi(env, reg);
990 if (!dynptr_type_refcounted(state->stack[spi].spilled_ptr.dynptr.type)) {
991 invalidate_dynptr(env, state, spi);
995 ref_obj_id = state->stack[spi].spilled_ptr.ref_obj_id;
997 /* If the dynptr has a ref_obj_id, then we need to invalidate
1000 * 1) Any dynptrs with a matching ref_obj_id (clones)
1001 * 2) Any slices derived from this dynptr.
1004 /* Invalidate any slices associated with this dynptr */
1005 WARN_ON_ONCE(release_reference(env, ref_obj_id));
1007 /* Invalidate any dynptr clones */
1008 for (i = 1; i < state->allocated_stack / BPF_REG_SIZE; i++) {
1009 if (state->stack[i].spilled_ptr.ref_obj_id != ref_obj_id)
1012 /* it should always be the case that if the ref obj id
1013 * matches then the stack slot also belongs to a
1016 if (state->stack[i].slot_type[0] != STACK_DYNPTR) {
1017 verbose(env, "verifier internal error: misconfigured ref_obj_id\n");
1020 if (state->stack[i].spilled_ptr.dynptr.first_slot)
1021 invalidate_dynptr(env, state, i);
1027 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1028 struct bpf_reg_state *reg);
1030 static void mark_reg_invalid(const struct bpf_verifier_env *env, struct bpf_reg_state *reg)
1032 if (!env->allow_ptr_leaks)
1033 __mark_reg_not_init(env, reg);
1035 __mark_reg_unknown(env, reg);
1038 static int destroy_if_dynptr_stack_slot(struct bpf_verifier_env *env,
1039 struct bpf_func_state *state, int spi)
1041 struct bpf_func_state *fstate;
1042 struct bpf_reg_state *dreg;
1045 /* We always ensure that STACK_DYNPTR is never set partially,
1046 * hence just checking for slot_type[0] is enough. This is
1047 * different for STACK_SPILL, where it may be only set for
1048 * 1 byte, so code has to use is_spilled_reg.
1050 if (state->stack[spi].slot_type[0] != STACK_DYNPTR)
1053 /* Reposition spi to first slot */
1054 if (!state->stack[spi].spilled_ptr.dynptr.first_slot)
1057 if (dynptr_type_refcounted(state->stack[spi].spilled_ptr.dynptr.type)) {
1058 verbose(env, "cannot overwrite referenced dynptr\n");
1062 mark_stack_slot_scratched(env, spi);
1063 mark_stack_slot_scratched(env, spi - 1);
1065 /* Writing partially to one dynptr stack slot destroys both. */
1066 for (i = 0; i < BPF_REG_SIZE; i++) {
1067 state->stack[spi].slot_type[i] = STACK_INVALID;
1068 state->stack[spi - 1].slot_type[i] = STACK_INVALID;
1071 dynptr_id = state->stack[spi].spilled_ptr.id;
1072 /* Invalidate any slices associated with this dynptr */
1073 bpf_for_each_reg_in_vstate(env->cur_state, fstate, dreg, ({
1074 /* Dynptr slices are only PTR_TO_MEM_OR_NULL and PTR_TO_MEM */
1075 if (dreg->type != (PTR_TO_MEM | PTR_MAYBE_NULL) && dreg->type != PTR_TO_MEM)
1077 if (dreg->dynptr_id == dynptr_id)
1078 mark_reg_invalid(env, dreg);
1081 /* Do not release reference state, we are destroying dynptr on stack,
1082 * not using some helper to release it. Just reset register.
1084 __mark_reg_not_init(env, &state->stack[spi].spilled_ptr);
1085 __mark_reg_not_init(env, &state->stack[spi - 1].spilled_ptr);
1087 /* Same reason as unmark_stack_slots_dynptr above */
1088 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1089 state->stack[spi - 1].spilled_ptr.live |= REG_LIVE_WRITTEN;
1094 static bool is_dynptr_reg_valid_uninit(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
1098 if (reg->type == CONST_PTR_TO_DYNPTR)
1101 spi = dynptr_get_spi(env, reg);
1103 /* -ERANGE (i.e. spi not falling into allocated stack slots) isn't an
1104 * error because this just means the stack state hasn't been updated yet.
1105 * We will do check_mem_access to check and update stack bounds later.
1107 if (spi < 0 && spi != -ERANGE)
1110 /* We don't need to check if the stack slots are marked by previous
1111 * dynptr initializations because we allow overwriting existing unreferenced
1112 * STACK_DYNPTR slots, see mark_stack_slots_dynptr which calls
1113 * destroy_if_dynptr_stack_slot to ensure dynptr objects at the slots we are
1114 * touching are completely destructed before we reinitialize them for a new
1115 * one. For referenced ones, destroy_if_dynptr_stack_slot returns an error early
1116 * instead of delaying it until the end where the user will get "Unreleased
1122 static bool is_dynptr_reg_valid_init(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
1124 struct bpf_func_state *state = func(env, reg);
1127 /* This already represents first slot of initialized bpf_dynptr.
1129 * CONST_PTR_TO_DYNPTR already has fixed and var_off as 0 due to
1130 * check_func_arg_reg_off's logic, so we don't need to check its
1131 * offset and alignment.
1133 if (reg->type == CONST_PTR_TO_DYNPTR)
1136 spi = dynptr_get_spi(env, reg);
1139 if (!state->stack[spi].spilled_ptr.dynptr.first_slot)
1142 for (i = 0; i < BPF_REG_SIZE; i++) {
1143 if (state->stack[spi].slot_type[i] != STACK_DYNPTR ||
1144 state->stack[spi - 1].slot_type[i] != STACK_DYNPTR)
1151 static bool is_dynptr_type_expected(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
1152 enum bpf_arg_type arg_type)
1154 struct bpf_func_state *state = func(env, reg);
1155 enum bpf_dynptr_type dynptr_type;
1158 /* ARG_PTR_TO_DYNPTR takes any type of dynptr */
1159 if (arg_type == ARG_PTR_TO_DYNPTR)
1162 dynptr_type = arg_to_dynptr_type(arg_type);
1163 if (reg->type == CONST_PTR_TO_DYNPTR) {
1164 return reg->dynptr.type == dynptr_type;
1166 spi = dynptr_get_spi(env, reg);
1169 return state->stack[spi].spilled_ptr.dynptr.type == dynptr_type;
1173 static void __mark_reg_known_zero(struct bpf_reg_state *reg);
1175 static int mark_stack_slots_iter(struct bpf_verifier_env *env,
1176 struct bpf_reg_state *reg, int insn_idx,
1177 struct btf *btf, u32 btf_id, int nr_slots)
1179 struct bpf_func_state *state = func(env, reg);
1182 spi = iter_get_spi(env, reg, nr_slots);
1186 id = acquire_reference_state(env, insn_idx);
1190 for (i = 0; i < nr_slots; i++) {
1191 struct bpf_stack_state *slot = &state->stack[spi - i];
1192 struct bpf_reg_state *st = &slot->spilled_ptr;
1194 __mark_reg_known_zero(st);
1195 st->type = PTR_TO_STACK; /* we don't have dedicated reg type */
1196 st->live |= REG_LIVE_WRITTEN;
1197 st->ref_obj_id = i == 0 ? id : 0;
1199 st->iter.btf_id = btf_id;
1200 st->iter.state = BPF_ITER_STATE_ACTIVE;
1203 for (j = 0; j < BPF_REG_SIZE; j++)
1204 slot->slot_type[j] = STACK_ITER;
1206 mark_stack_slot_scratched(env, spi - i);
1212 static int unmark_stack_slots_iter(struct bpf_verifier_env *env,
1213 struct bpf_reg_state *reg, int nr_slots)
1215 struct bpf_func_state *state = func(env, reg);
1218 spi = iter_get_spi(env, reg, nr_slots);
1222 for (i = 0; i < nr_slots; i++) {
1223 struct bpf_stack_state *slot = &state->stack[spi - i];
1224 struct bpf_reg_state *st = &slot->spilled_ptr;
1227 WARN_ON_ONCE(release_reference(env, st->ref_obj_id));
1229 __mark_reg_not_init(env, st);
1231 /* see unmark_stack_slots_dynptr() for why we need to set REG_LIVE_WRITTEN */
1232 st->live |= REG_LIVE_WRITTEN;
1234 for (j = 0; j < BPF_REG_SIZE; j++)
1235 slot->slot_type[j] = STACK_INVALID;
1237 mark_stack_slot_scratched(env, spi - i);
1243 static bool is_iter_reg_valid_uninit(struct bpf_verifier_env *env,
1244 struct bpf_reg_state *reg, int nr_slots)
1246 struct bpf_func_state *state = func(env, reg);
1249 /* For -ERANGE (i.e. spi not falling into allocated stack slots), we
1250 * will do check_mem_access to check and update stack bounds later, so
1251 * return true for that case.
1253 spi = iter_get_spi(env, reg, nr_slots);
1259 for (i = 0; i < nr_slots; i++) {
1260 struct bpf_stack_state *slot = &state->stack[spi - i];
1262 for (j = 0; j < BPF_REG_SIZE; j++)
1263 if (slot->slot_type[j] == STACK_ITER)
1270 static bool is_iter_reg_valid_init(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
1271 struct btf *btf, u32 btf_id, int nr_slots)
1273 struct bpf_func_state *state = func(env, reg);
1276 spi = iter_get_spi(env, reg, nr_slots);
1280 for (i = 0; i < nr_slots; i++) {
1281 struct bpf_stack_state *slot = &state->stack[spi - i];
1282 struct bpf_reg_state *st = &slot->spilled_ptr;
1284 /* only main (first) slot has ref_obj_id set */
1285 if (i == 0 && !st->ref_obj_id)
1287 if (i != 0 && st->ref_obj_id)
1289 if (st->iter.btf != btf || st->iter.btf_id != btf_id)
1292 for (j = 0; j < BPF_REG_SIZE; j++)
1293 if (slot->slot_type[j] != STACK_ITER)
1300 /* Check if given stack slot is "special":
1301 * - spilled register state (STACK_SPILL);
1302 * - dynptr state (STACK_DYNPTR);
1303 * - iter state (STACK_ITER).
1305 static bool is_stack_slot_special(const struct bpf_stack_state *stack)
1307 enum bpf_stack_slot_type type = stack->slot_type[BPF_REG_SIZE - 1];
1319 WARN_ONCE(1, "unknown stack slot type %d\n", type);
1324 /* The reg state of a pointer or a bounded scalar was saved when
1325 * it was spilled to the stack.
1327 static bool is_spilled_reg(const struct bpf_stack_state *stack)
1329 return stack->slot_type[BPF_REG_SIZE - 1] == STACK_SPILL;
1332 static bool is_spilled_scalar_reg(const struct bpf_stack_state *stack)
1334 return stack->slot_type[BPF_REG_SIZE - 1] == STACK_SPILL &&
1335 stack->spilled_ptr.type == SCALAR_VALUE;
1338 static void scrub_spilled_slot(u8 *stype)
1340 if (*stype != STACK_INVALID)
1341 *stype = STACK_MISC;
1344 static void print_verifier_state(struct bpf_verifier_env *env,
1345 const struct bpf_func_state *state,
1348 const struct bpf_reg_state *reg;
1349 enum bpf_reg_type t;
1353 verbose(env, " frame%d:", state->frameno);
1354 for (i = 0; i < MAX_BPF_REG; i++) {
1355 reg = &state->regs[i];
1359 if (!print_all && !reg_scratched(env, i))
1361 verbose(env, " R%d", i);
1362 print_liveness(env, reg->live);
1364 if (t == SCALAR_VALUE && reg->precise)
1366 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
1367 tnum_is_const(reg->var_off)) {
1368 /* reg->off should be 0 for SCALAR_VALUE */
1369 verbose(env, "%s", t == SCALAR_VALUE ? "" : reg_type_str(env, t));
1370 verbose(env, "%lld", reg->var_off.value + reg->off);
1372 const char *sep = "";
1374 verbose(env, "%s", reg_type_str(env, t));
1375 if (base_type(t) == PTR_TO_BTF_ID)
1376 verbose(env, "%s", btf_type_name(reg->btf, reg->btf_id));
1379 * _a stands for append, was shortened to avoid multiline statements below.
1380 * This macro is used to output a comma separated list of attributes.
1382 #define verbose_a(fmt, ...) ({ verbose(env, "%s" fmt, sep, __VA_ARGS__); sep = ","; })
1385 verbose_a("id=%d", reg->id);
1386 if (reg->ref_obj_id)
1387 verbose_a("ref_obj_id=%d", reg->ref_obj_id);
1388 if (type_is_non_owning_ref(reg->type))
1389 verbose_a("%s", "non_own_ref");
1390 if (t != SCALAR_VALUE)
1391 verbose_a("off=%d", reg->off);
1392 if (type_is_pkt_pointer(t))
1393 verbose_a("r=%d", reg->range);
1394 else if (base_type(t) == CONST_PTR_TO_MAP ||
1395 base_type(t) == PTR_TO_MAP_KEY ||
1396 base_type(t) == PTR_TO_MAP_VALUE)
1397 verbose_a("ks=%d,vs=%d",
1398 reg->map_ptr->key_size,
1399 reg->map_ptr->value_size);
1400 if (tnum_is_const(reg->var_off)) {
1401 /* Typically an immediate SCALAR_VALUE, but
1402 * could be a pointer whose offset is too big
1405 verbose_a("imm=%llx", reg->var_off.value);
1407 if (reg->smin_value != reg->umin_value &&
1408 reg->smin_value != S64_MIN)
1409 verbose_a("smin=%lld", (long long)reg->smin_value);
1410 if (reg->smax_value != reg->umax_value &&
1411 reg->smax_value != S64_MAX)
1412 verbose_a("smax=%lld", (long long)reg->smax_value);
1413 if (reg->umin_value != 0)
1414 verbose_a("umin=%llu", (unsigned long long)reg->umin_value);
1415 if (reg->umax_value != U64_MAX)
1416 verbose_a("umax=%llu", (unsigned long long)reg->umax_value);
1417 if (!tnum_is_unknown(reg->var_off)) {
1420 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1421 verbose_a("var_off=%s", tn_buf);
1423 if (reg->s32_min_value != reg->smin_value &&
1424 reg->s32_min_value != S32_MIN)
1425 verbose_a("s32_min=%d", (int)(reg->s32_min_value));
1426 if (reg->s32_max_value != reg->smax_value &&
1427 reg->s32_max_value != S32_MAX)
1428 verbose_a("s32_max=%d", (int)(reg->s32_max_value));
1429 if (reg->u32_min_value != reg->umin_value &&
1430 reg->u32_min_value != U32_MIN)
1431 verbose_a("u32_min=%d", (int)(reg->u32_min_value));
1432 if (reg->u32_max_value != reg->umax_value &&
1433 reg->u32_max_value != U32_MAX)
1434 verbose_a("u32_max=%d", (int)(reg->u32_max_value));
1441 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
1442 char types_buf[BPF_REG_SIZE + 1];
1446 for (j = 0; j < BPF_REG_SIZE; j++) {
1447 if (state->stack[i].slot_type[j] != STACK_INVALID)
1449 types_buf[j] = slot_type_char[state->stack[i].slot_type[j]];
1451 types_buf[BPF_REG_SIZE] = 0;
1454 if (!print_all && !stack_slot_scratched(env, i))
1456 switch (state->stack[i].slot_type[BPF_REG_SIZE - 1]) {
1458 reg = &state->stack[i].spilled_ptr;
1461 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
1462 print_liveness(env, reg->live);
1463 verbose(env, "=%s", t == SCALAR_VALUE ? "" : reg_type_str(env, t));
1464 if (t == SCALAR_VALUE && reg->precise)
1466 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
1467 verbose(env, "%lld", reg->var_off.value + reg->off);
1470 i += BPF_DYNPTR_NR_SLOTS - 1;
1471 reg = &state->stack[i].spilled_ptr;
1473 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
1474 print_liveness(env, reg->live);
1475 verbose(env, "=dynptr_%s", dynptr_type_str(reg->dynptr.type));
1476 if (reg->ref_obj_id)
1477 verbose(env, "(ref_id=%d)", reg->ref_obj_id);
1480 /* only main slot has ref_obj_id set; skip others */
1481 reg = &state->stack[i].spilled_ptr;
1482 if (!reg->ref_obj_id)
1485 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
1486 print_liveness(env, reg->live);
1487 verbose(env, "=iter_%s(ref_id=%d,state=%s,depth=%u)",
1488 iter_type_str(reg->iter.btf, reg->iter.btf_id),
1489 reg->ref_obj_id, iter_state_str(reg->iter.state),
1495 reg = &state->stack[i].spilled_ptr;
1497 for (j = 0; j < BPF_REG_SIZE; j++)
1498 types_buf[j] = slot_type_char[state->stack[i].slot_type[j]];
1499 types_buf[BPF_REG_SIZE] = 0;
1501 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
1502 print_liveness(env, reg->live);
1503 verbose(env, "=%s", types_buf);
1507 if (state->acquired_refs && state->refs[0].id) {
1508 verbose(env, " refs=%d", state->refs[0].id);
1509 for (i = 1; i < state->acquired_refs; i++)
1510 if (state->refs[i].id)
1511 verbose(env, ",%d", state->refs[i].id);
1513 if (state->in_callback_fn)
1514 verbose(env, " cb");
1515 if (state->in_async_callback_fn)
1516 verbose(env, " async_cb");
1519 mark_verifier_state_clean(env);
1522 static inline u32 vlog_alignment(u32 pos)
1524 return round_up(max(pos + BPF_LOG_MIN_ALIGNMENT / 2, BPF_LOG_ALIGNMENT),
1525 BPF_LOG_MIN_ALIGNMENT) - pos - 1;
1528 static void print_insn_state(struct bpf_verifier_env *env,
1529 const struct bpf_func_state *state)
1531 if (env->prev_log_pos && env->prev_log_pos == env->log.end_pos) {
1532 /* remove new line character */
1533 bpf_vlog_reset(&env->log, env->prev_log_pos - 1);
1534 verbose(env, "%*c;", vlog_alignment(env->prev_insn_print_pos), ' ');
1536 verbose(env, "%d:", env->insn_idx);
1538 print_verifier_state(env, state, false);
1541 /* copy array src of length n * size bytes to dst. dst is reallocated if it's too
1542 * small to hold src. This is different from krealloc since we don't want to preserve
1543 * the contents of dst.
1545 * Leaves dst untouched if src is NULL or length is zero. Returns NULL if memory could
1548 static void *copy_array(void *dst, const void *src, size_t n, size_t size, gfp_t flags)
1554 if (ZERO_OR_NULL_PTR(src))
1557 if (unlikely(check_mul_overflow(n, size, &bytes)))
1560 alloc_bytes = max(ksize(orig), kmalloc_size_roundup(bytes));
1561 dst = krealloc(orig, alloc_bytes, flags);
1567 memcpy(dst, src, bytes);
1569 return dst ? dst : ZERO_SIZE_PTR;
1572 /* resize an array from old_n items to new_n items. the array is reallocated if it's too
1573 * small to hold new_n items. new items are zeroed out if the array grows.
1575 * Contrary to krealloc_array, does not free arr if new_n is zero.
1577 static void *realloc_array(void *arr, size_t old_n, size_t new_n, size_t size)
1582 if (!new_n || old_n == new_n)
1585 alloc_size = kmalloc_size_roundup(size_mul(new_n, size));
1586 new_arr = krealloc(arr, alloc_size, GFP_KERNEL);
1594 memset(arr + old_n * size, 0, (new_n - old_n) * size);
1597 return arr ? arr : ZERO_SIZE_PTR;
1600 static int copy_reference_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
1602 dst->refs = copy_array(dst->refs, src->refs, src->acquired_refs,
1603 sizeof(struct bpf_reference_state), GFP_KERNEL);
1607 dst->acquired_refs = src->acquired_refs;
1611 static int copy_stack_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
1613 size_t n = src->allocated_stack / BPF_REG_SIZE;
1615 dst->stack = copy_array(dst->stack, src->stack, n, sizeof(struct bpf_stack_state),
1620 dst->allocated_stack = src->allocated_stack;
1624 static int resize_reference_state(struct bpf_func_state *state, size_t n)
1626 state->refs = realloc_array(state->refs, state->acquired_refs, n,
1627 sizeof(struct bpf_reference_state));
1631 state->acquired_refs = n;
1635 /* Possibly update state->allocated_stack to be at least size bytes. Also
1636 * possibly update the function's high-water mark in its bpf_subprog_info.
1638 static int grow_stack_state(struct bpf_verifier_env *env, struct bpf_func_state *state, int size)
1640 size_t old_n = state->allocated_stack / BPF_REG_SIZE, n = size / BPF_REG_SIZE;
1645 state->stack = realloc_array(state->stack, old_n, n, sizeof(struct bpf_stack_state));
1649 state->allocated_stack = size;
1651 /* update known max for given subprogram */
1652 if (env->subprog_info[state->subprogno].stack_depth < size)
1653 env->subprog_info[state->subprogno].stack_depth = size;
1658 /* Acquire a pointer id from the env and update the state->refs to include
1659 * this new pointer reference.
1660 * On success, returns a valid pointer id to associate with the register
1661 * On failure, returns a negative errno.
1663 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
1665 struct bpf_func_state *state = cur_func(env);
1666 int new_ofs = state->acquired_refs;
1669 err = resize_reference_state(state, state->acquired_refs + 1);
1673 state->refs[new_ofs].id = id;
1674 state->refs[new_ofs].insn_idx = insn_idx;
1675 state->refs[new_ofs].callback_ref = state->in_callback_fn ? state->frameno : 0;
1680 /* release function corresponding to acquire_reference_state(). Idempotent. */
1681 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
1685 last_idx = state->acquired_refs - 1;
1686 for (i = 0; i < state->acquired_refs; i++) {
1687 if (state->refs[i].id == ptr_id) {
1688 /* Cannot release caller references in callbacks */
1689 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno)
1691 if (last_idx && i != last_idx)
1692 memcpy(&state->refs[i], &state->refs[last_idx],
1693 sizeof(*state->refs));
1694 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
1695 state->acquired_refs--;
1702 static void free_func_state(struct bpf_func_state *state)
1707 kfree(state->stack);
1711 static void clear_jmp_history(struct bpf_verifier_state *state)
1713 kfree(state->jmp_history);
1714 state->jmp_history = NULL;
1715 state->jmp_history_cnt = 0;
1718 static void free_verifier_state(struct bpf_verifier_state *state,
1723 for (i = 0; i <= state->curframe; i++) {
1724 free_func_state(state->frame[i]);
1725 state->frame[i] = NULL;
1727 clear_jmp_history(state);
1732 /* copy verifier state from src to dst growing dst stack space
1733 * when necessary to accommodate larger src stack
1735 static int copy_func_state(struct bpf_func_state *dst,
1736 const struct bpf_func_state *src)
1740 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
1741 err = copy_reference_state(dst, src);
1744 return copy_stack_state(dst, src);
1747 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
1748 const struct bpf_verifier_state *src)
1750 struct bpf_func_state *dst;
1753 dst_state->jmp_history = copy_array(dst_state->jmp_history, src->jmp_history,
1754 src->jmp_history_cnt, sizeof(struct bpf_idx_pair),
1756 if (!dst_state->jmp_history)
1758 dst_state->jmp_history_cnt = src->jmp_history_cnt;
1760 /* if dst has more stack frames then src frame, free them */
1761 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
1762 free_func_state(dst_state->frame[i]);
1763 dst_state->frame[i] = NULL;
1765 dst_state->speculative = src->speculative;
1766 dst_state->active_rcu_lock = src->active_rcu_lock;
1767 dst_state->curframe = src->curframe;
1768 dst_state->active_lock.ptr = src->active_lock.ptr;
1769 dst_state->active_lock.id = src->active_lock.id;
1770 dst_state->branches = src->branches;
1771 dst_state->parent = src->parent;
1772 dst_state->first_insn_idx = src->first_insn_idx;
1773 dst_state->last_insn_idx = src->last_insn_idx;
1774 for (i = 0; i <= src->curframe; i++) {
1775 dst = dst_state->frame[i];
1777 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
1780 dst_state->frame[i] = dst;
1782 err = copy_func_state(dst, src->frame[i]);
1789 static u32 state_htab_size(struct bpf_verifier_env *env)
1791 return env->prog->len;
1794 static struct bpf_verifier_state_list **explored_state(struct bpf_verifier_env *env, int idx)
1796 struct bpf_verifier_state *cur = env->cur_state;
1797 struct bpf_func_state *state = cur->frame[cur->curframe];
1799 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
1802 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
1805 u32 br = --st->branches;
1807 /* WARN_ON(br > 1) technically makes sense here,
1808 * but see comment in push_stack(), hence:
1810 WARN_ONCE((int)br < 0,
1811 "BUG update_branch_counts:branches_to_explore=%d\n",
1819 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
1820 int *insn_idx, bool pop_log)
1822 struct bpf_verifier_state *cur = env->cur_state;
1823 struct bpf_verifier_stack_elem *elem, *head = env->head;
1826 if (env->head == NULL)
1830 err = copy_verifier_state(cur, &head->st);
1835 bpf_vlog_reset(&env->log, head->log_pos);
1837 *insn_idx = head->insn_idx;
1839 *prev_insn_idx = head->prev_insn_idx;
1841 free_verifier_state(&head->st, false);
1848 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
1849 int insn_idx, int prev_insn_idx,
1852 struct bpf_verifier_state *cur = env->cur_state;
1853 struct bpf_verifier_stack_elem *elem;
1856 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1860 elem->insn_idx = insn_idx;
1861 elem->prev_insn_idx = prev_insn_idx;
1862 elem->next = env->head;
1863 elem->log_pos = env->log.end_pos;
1866 err = copy_verifier_state(&elem->st, cur);
1869 elem->st.speculative |= speculative;
1870 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1871 verbose(env, "The sequence of %d jumps is too complex.\n",
1875 if (elem->st.parent) {
1876 ++elem->st.parent->branches;
1877 /* WARN_ON(branches > 2) technically makes sense here,
1879 * 1. speculative states will bump 'branches' for non-branch
1881 * 2. is_state_visited() heuristics may decide not to create
1882 * a new state for a sequence of branches and all such current
1883 * and cloned states will be pointing to a single parent state
1884 * which might have large 'branches' count.
1889 free_verifier_state(env->cur_state, true);
1890 env->cur_state = NULL;
1891 /* pop all elements and return */
1892 while (!pop_stack(env, NULL, NULL, false));
1896 #define CALLER_SAVED_REGS 6
1897 static const int caller_saved[CALLER_SAVED_REGS] = {
1898 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
1901 /* This helper doesn't clear reg->id */
1902 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1904 reg->var_off = tnum_const(imm);
1905 reg->smin_value = (s64)imm;
1906 reg->smax_value = (s64)imm;
1907 reg->umin_value = imm;
1908 reg->umax_value = imm;
1910 reg->s32_min_value = (s32)imm;
1911 reg->s32_max_value = (s32)imm;
1912 reg->u32_min_value = (u32)imm;
1913 reg->u32_max_value = (u32)imm;
1916 /* Mark the unknown part of a register (variable offset or scalar value) as
1917 * known to have the value @imm.
1919 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1921 /* Clear off and union(map_ptr, range) */
1922 memset(((u8 *)reg) + sizeof(reg->type), 0,
1923 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1925 reg->ref_obj_id = 0;
1926 ___mark_reg_known(reg, imm);
1929 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1931 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1932 reg->s32_min_value = (s32)imm;
1933 reg->s32_max_value = (s32)imm;
1934 reg->u32_min_value = (u32)imm;
1935 reg->u32_max_value = (u32)imm;
1938 /* Mark the 'variable offset' part of a register as zero. This should be
1939 * used only on registers holding a pointer type.
1941 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1943 __mark_reg_known(reg, 0);
1946 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1948 __mark_reg_known(reg, 0);
1949 reg->type = SCALAR_VALUE;
1952 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1953 struct bpf_reg_state *regs, u32 regno)
1955 if (WARN_ON(regno >= MAX_BPF_REG)) {
1956 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1957 /* Something bad happened, let's kill all regs */
1958 for (regno = 0; regno < MAX_BPF_REG; regno++)
1959 __mark_reg_not_init(env, regs + regno);
1962 __mark_reg_known_zero(regs + regno);
1965 static void __mark_dynptr_reg(struct bpf_reg_state *reg, enum bpf_dynptr_type type,
1966 bool first_slot, int dynptr_id)
1968 /* reg->type has no meaning for STACK_DYNPTR, but when we set reg for
1969 * callback arguments, it does need to be CONST_PTR_TO_DYNPTR, so simply
1970 * set it unconditionally as it is ignored for STACK_DYNPTR anyway.
1972 __mark_reg_known_zero(reg);
1973 reg->type = CONST_PTR_TO_DYNPTR;
1974 /* Give each dynptr a unique id to uniquely associate slices to it. */
1975 reg->id = dynptr_id;
1976 reg->dynptr.type = type;
1977 reg->dynptr.first_slot = first_slot;
1980 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
1982 if (base_type(reg->type) == PTR_TO_MAP_VALUE) {
1983 const struct bpf_map *map = reg->map_ptr;
1985 if (map->inner_map_meta) {
1986 reg->type = CONST_PTR_TO_MAP;
1987 reg->map_ptr = map->inner_map_meta;
1988 /* transfer reg's id which is unique for every map_lookup_elem
1989 * as UID of the inner map.
1991 if (btf_record_has_field(map->inner_map_meta->record, BPF_TIMER))
1992 reg->map_uid = reg->id;
1993 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
1994 reg->type = PTR_TO_XDP_SOCK;
1995 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
1996 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
1997 reg->type = PTR_TO_SOCKET;
1999 reg->type = PTR_TO_MAP_VALUE;
2004 reg->type &= ~PTR_MAYBE_NULL;
2007 static void mark_reg_graph_node(struct bpf_reg_state *regs, u32 regno,
2008 struct btf_field_graph_root *ds_head)
2010 __mark_reg_known_zero(®s[regno]);
2011 regs[regno].type = PTR_TO_BTF_ID | MEM_ALLOC;
2012 regs[regno].btf = ds_head->btf;
2013 regs[regno].btf_id = ds_head->value_btf_id;
2014 regs[regno].off = ds_head->node_offset;
2017 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
2019 return type_is_pkt_pointer(reg->type);
2022 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
2024 return reg_is_pkt_pointer(reg) ||
2025 reg->type == PTR_TO_PACKET_END;
2028 static bool reg_is_dynptr_slice_pkt(const struct bpf_reg_state *reg)
2030 return base_type(reg->type) == PTR_TO_MEM &&
2031 (reg->type & DYNPTR_TYPE_SKB || reg->type & DYNPTR_TYPE_XDP);
2034 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
2035 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
2036 enum bpf_reg_type which)
2038 /* The register can already have a range from prior markings.
2039 * This is fine as long as it hasn't been advanced from its
2042 return reg->type == which &&
2045 tnum_equals_const(reg->var_off, 0);
2048 /* Reset the min/max bounds of a register */
2049 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
2051 reg->smin_value = S64_MIN;
2052 reg->smax_value = S64_MAX;
2053 reg->umin_value = 0;
2054 reg->umax_value = U64_MAX;
2056 reg->s32_min_value = S32_MIN;
2057 reg->s32_max_value = S32_MAX;
2058 reg->u32_min_value = 0;
2059 reg->u32_max_value = U32_MAX;
2062 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
2064 reg->smin_value = S64_MIN;
2065 reg->smax_value = S64_MAX;
2066 reg->umin_value = 0;
2067 reg->umax_value = U64_MAX;
2070 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
2072 reg->s32_min_value = S32_MIN;
2073 reg->s32_max_value = S32_MAX;
2074 reg->u32_min_value = 0;
2075 reg->u32_max_value = U32_MAX;
2078 static void __update_reg32_bounds(struct bpf_reg_state *reg)
2080 struct tnum var32_off = tnum_subreg(reg->var_off);
2082 /* min signed is max(sign bit) | min(other bits) */
2083 reg->s32_min_value = max_t(s32, reg->s32_min_value,
2084 var32_off.value | (var32_off.mask & S32_MIN));
2085 /* max signed is min(sign bit) | max(other bits) */
2086 reg->s32_max_value = min_t(s32, reg->s32_max_value,
2087 var32_off.value | (var32_off.mask & S32_MAX));
2088 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
2089 reg->u32_max_value = min(reg->u32_max_value,
2090 (u32)(var32_off.value | var32_off.mask));
2093 static void __update_reg64_bounds(struct bpf_reg_state *reg)
2095 /* min signed is max(sign bit) | min(other bits) */
2096 reg->smin_value = max_t(s64, reg->smin_value,
2097 reg->var_off.value | (reg->var_off.mask & S64_MIN));
2098 /* max signed is min(sign bit) | max(other bits) */
2099 reg->smax_value = min_t(s64, reg->smax_value,
2100 reg->var_off.value | (reg->var_off.mask & S64_MAX));
2101 reg->umin_value = max(reg->umin_value, reg->var_off.value);
2102 reg->umax_value = min(reg->umax_value,
2103 reg->var_off.value | reg->var_off.mask);
2106 static void __update_reg_bounds(struct bpf_reg_state *reg)
2108 __update_reg32_bounds(reg);
2109 __update_reg64_bounds(reg);
2112 /* Uses signed min/max values to inform unsigned, and vice-versa */
2113 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
2115 /* Learn sign from signed bounds.
2116 * If we cannot cross the sign boundary, then signed and unsigned bounds
2117 * are the same, so combine. This works even in the negative case, e.g.
2118 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
2120 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
2121 reg->s32_min_value = reg->u32_min_value =
2122 max_t(u32, reg->s32_min_value, reg->u32_min_value);
2123 reg->s32_max_value = reg->u32_max_value =
2124 min_t(u32, reg->s32_max_value, reg->u32_max_value);
2127 /* Learn sign from unsigned bounds. Signed bounds cross the sign
2128 * boundary, so we must be careful.
2130 if ((s32)reg->u32_max_value >= 0) {
2131 /* Positive. We can't learn anything from the smin, but smax
2132 * is positive, hence safe.
2134 reg->s32_min_value = reg->u32_min_value;
2135 reg->s32_max_value = reg->u32_max_value =
2136 min_t(u32, reg->s32_max_value, reg->u32_max_value);
2137 } else if ((s32)reg->u32_min_value < 0) {
2138 /* Negative. We can't learn anything from the smax, but smin
2139 * is negative, hence safe.
2141 reg->s32_min_value = reg->u32_min_value =
2142 max_t(u32, reg->s32_min_value, reg->u32_min_value);
2143 reg->s32_max_value = reg->u32_max_value;
2147 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
2149 /* Learn sign from signed bounds.
2150 * If we cannot cross the sign boundary, then signed and unsigned bounds
2151 * are the same, so combine. This works even in the negative case, e.g.
2152 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
2154 if (reg->smin_value >= 0 || reg->smax_value < 0) {
2155 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
2157 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
2161 /* Learn sign from unsigned bounds. Signed bounds cross the sign
2162 * boundary, so we must be careful.
2164 if ((s64)reg->umax_value >= 0) {
2165 /* Positive. We can't learn anything from the smin, but smax
2166 * is positive, hence safe.
2168 reg->smin_value = reg->umin_value;
2169 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
2171 } else if ((s64)reg->umin_value < 0) {
2172 /* Negative. We can't learn anything from the smax, but smin
2173 * is negative, hence safe.
2175 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
2177 reg->smax_value = reg->umax_value;
2181 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
2183 __reg32_deduce_bounds(reg);
2184 __reg64_deduce_bounds(reg);
2187 /* Attempts to improve var_off based on unsigned min/max information */
2188 static void __reg_bound_offset(struct bpf_reg_state *reg)
2190 struct tnum var64_off = tnum_intersect(reg->var_off,
2191 tnum_range(reg->umin_value,
2193 struct tnum var32_off = tnum_intersect(tnum_subreg(var64_off),
2194 tnum_range(reg->u32_min_value,
2195 reg->u32_max_value));
2197 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
2200 static void reg_bounds_sync(struct bpf_reg_state *reg)
2202 /* We might have learned new bounds from the var_off. */
2203 __update_reg_bounds(reg);
2204 /* We might have learned something about the sign bit. */
2205 __reg_deduce_bounds(reg);
2206 /* We might have learned some bits from the bounds. */
2207 __reg_bound_offset(reg);
2208 /* Intersecting with the old var_off might have improved our bounds
2209 * slightly, e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2210 * then new var_off is (0; 0x7f...fc) which improves our umax.
2212 __update_reg_bounds(reg);
2215 static bool __reg32_bound_s64(s32 a)
2217 return a >= 0 && a <= S32_MAX;
2220 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
2222 reg->umin_value = reg->u32_min_value;
2223 reg->umax_value = reg->u32_max_value;
2225 /* Attempt to pull 32-bit signed bounds into 64-bit bounds but must
2226 * be positive otherwise set to worse case bounds and refine later
2229 if (__reg32_bound_s64(reg->s32_min_value) &&
2230 __reg32_bound_s64(reg->s32_max_value)) {
2231 reg->smin_value = reg->s32_min_value;
2232 reg->smax_value = reg->s32_max_value;
2234 reg->smin_value = 0;
2235 reg->smax_value = U32_MAX;
2239 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
2241 /* special case when 64-bit register has upper 32-bit register
2242 * zeroed. Typically happens after zext or <<32, >>32 sequence
2243 * allowing us to use 32-bit bounds directly,
2245 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
2246 __reg_assign_32_into_64(reg);
2248 /* Otherwise the best we can do is push lower 32bit known and
2249 * unknown bits into register (var_off set from jmp logic)
2250 * then learn as much as possible from the 64-bit tnum
2251 * known and unknown bits. The previous smin/smax bounds are
2252 * invalid here because of jmp32 compare so mark them unknown
2253 * so they do not impact tnum bounds calculation.
2255 __mark_reg64_unbounded(reg);
2257 reg_bounds_sync(reg);
2260 static bool __reg64_bound_s32(s64 a)
2262 return a >= S32_MIN && a <= S32_MAX;
2265 static bool __reg64_bound_u32(u64 a)
2267 return a >= U32_MIN && a <= U32_MAX;
2270 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
2272 __mark_reg32_unbounded(reg);
2273 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
2274 reg->s32_min_value = (s32)reg->smin_value;
2275 reg->s32_max_value = (s32)reg->smax_value;
2277 if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
2278 reg->u32_min_value = (u32)reg->umin_value;
2279 reg->u32_max_value = (u32)reg->umax_value;
2281 reg_bounds_sync(reg);
2284 /* Mark a register as having a completely unknown (scalar) value. */
2285 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
2286 struct bpf_reg_state *reg)
2289 * Clear type, off, and union(map_ptr, range) and
2290 * padding between 'type' and union
2292 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
2293 reg->type = SCALAR_VALUE;
2295 reg->ref_obj_id = 0;
2296 reg->var_off = tnum_unknown;
2298 reg->precise = !env->bpf_capable;
2299 __mark_reg_unbounded(reg);
2302 static void mark_reg_unknown(struct bpf_verifier_env *env,
2303 struct bpf_reg_state *regs, u32 regno)
2305 if (WARN_ON(regno >= MAX_BPF_REG)) {
2306 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
2307 /* Something bad happened, let's kill all regs except FP */
2308 for (regno = 0; regno < BPF_REG_FP; regno++)
2309 __mark_reg_not_init(env, regs + regno);
2312 __mark_reg_unknown(env, regs + regno);
2315 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
2316 struct bpf_reg_state *reg)
2318 __mark_reg_unknown(env, reg);
2319 reg->type = NOT_INIT;
2322 static void mark_reg_not_init(struct bpf_verifier_env *env,
2323 struct bpf_reg_state *regs, u32 regno)
2325 if (WARN_ON(regno >= MAX_BPF_REG)) {
2326 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
2327 /* Something bad happened, let's kill all regs except FP */
2328 for (regno = 0; regno < BPF_REG_FP; regno++)
2329 __mark_reg_not_init(env, regs + regno);
2332 __mark_reg_not_init(env, regs + regno);
2335 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
2336 struct bpf_reg_state *regs, u32 regno,
2337 enum bpf_reg_type reg_type,
2338 struct btf *btf, u32 btf_id,
2339 enum bpf_type_flag flag)
2341 if (reg_type == SCALAR_VALUE) {
2342 mark_reg_unknown(env, regs, regno);
2345 mark_reg_known_zero(env, regs, regno);
2346 regs[regno].type = PTR_TO_BTF_ID | flag;
2347 regs[regno].btf = btf;
2348 regs[regno].btf_id = btf_id;
2351 #define DEF_NOT_SUBREG (0)
2352 static void init_reg_state(struct bpf_verifier_env *env,
2353 struct bpf_func_state *state)
2355 struct bpf_reg_state *regs = state->regs;
2358 for (i = 0; i < MAX_BPF_REG; i++) {
2359 mark_reg_not_init(env, regs, i);
2360 regs[i].live = REG_LIVE_NONE;
2361 regs[i].parent = NULL;
2362 regs[i].subreg_def = DEF_NOT_SUBREG;
2366 regs[BPF_REG_FP].type = PTR_TO_STACK;
2367 mark_reg_known_zero(env, regs, BPF_REG_FP);
2368 regs[BPF_REG_FP].frameno = state->frameno;
2371 #define BPF_MAIN_FUNC (-1)
2372 static void init_func_state(struct bpf_verifier_env *env,
2373 struct bpf_func_state *state,
2374 int callsite, int frameno, int subprogno)
2376 state->callsite = callsite;
2377 state->frameno = frameno;
2378 state->subprogno = subprogno;
2379 state->callback_ret_range = tnum_range(0, 0);
2380 init_reg_state(env, state);
2381 mark_verifier_state_scratched(env);
2384 /* Similar to push_stack(), but for async callbacks */
2385 static struct bpf_verifier_state *push_async_cb(struct bpf_verifier_env *env,
2386 int insn_idx, int prev_insn_idx,
2389 struct bpf_verifier_stack_elem *elem;
2390 struct bpf_func_state *frame;
2392 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
2396 elem->insn_idx = insn_idx;
2397 elem->prev_insn_idx = prev_insn_idx;
2398 elem->next = env->head;
2399 elem->log_pos = env->log.end_pos;
2402 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
2404 "The sequence of %d jumps is too complex for async cb.\n",
2408 /* Unlike push_stack() do not copy_verifier_state().
2409 * The caller state doesn't matter.
2410 * This is async callback. It starts in a fresh stack.
2411 * Initialize it similar to do_check_common().
2413 elem->st.branches = 1;
2414 frame = kzalloc(sizeof(*frame), GFP_KERNEL);
2417 init_func_state(env, frame,
2418 BPF_MAIN_FUNC /* callsite */,
2419 0 /* frameno within this callchain */,
2420 subprog /* subprog number within this prog */);
2421 elem->st.frame[0] = frame;
2424 free_verifier_state(env->cur_state, true);
2425 env->cur_state = NULL;
2426 /* pop all elements and return */
2427 while (!pop_stack(env, NULL, NULL, false));
2433 SRC_OP, /* register is used as source operand */
2434 DST_OP, /* register is used as destination operand */
2435 DST_OP_NO_MARK /* same as above, check only, don't mark */
2438 static int cmp_subprogs(const void *a, const void *b)
2440 return ((struct bpf_subprog_info *)a)->start -
2441 ((struct bpf_subprog_info *)b)->start;
2444 static int find_subprog(struct bpf_verifier_env *env, int off)
2446 struct bpf_subprog_info *p;
2448 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
2449 sizeof(env->subprog_info[0]), cmp_subprogs);
2452 return p - env->subprog_info;
2456 static int add_subprog(struct bpf_verifier_env *env, int off)
2458 int insn_cnt = env->prog->len;
2461 if (off >= insn_cnt || off < 0) {
2462 verbose(env, "call to invalid destination\n");
2465 ret = find_subprog(env, off);
2468 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
2469 verbose(env, "too many subprograms\n");
2472 /* determine subprog starts. The end is one before the next starts */
2473 env->subprog_info[env->subprog_cnt++].start = off;
2474 sort(env->subprog_info, env->subprog_cnt,
2475 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
2476 return env->subprog_cnt - 1;
2479 #define MAX_KFUNC_DESCS 256
2480 #define MAX_KFUNC_BTFS 256
2482 struct bpf_kfunc_desc {
2483 struct btf_func_model func_model;
2490 struct bpf_kfunc_btf {
2492 struct module *module;
2496 struct bpf_kfunc_desc_tab {
2497 /* Sorted by func_id (BTF ID) and offset (fd_array offset) during
2498 * verification. JITs do lookups by bpf_insn, where func_id may not be
2499 * available, therefore at the end of verification do_misc_fixups()
2500 * sorts this by imm and offset.
2502 struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
2506 struct bpf_kfunc_btf_tab {
2507 struct bpf_kfunc_btf descs[MAX_KFUNC_BTFS];
2511 static int kfunc_desc_cmp_by_id_off(const void *a, const void *b)
2513 const struct bpf_kfunc_desc *d0 = a;
2514 const struct bpf_kfunc_desc *d1 = b;
2516 /* func_id is not greater than BTF_MAX_TYPE */
2517 return d0->func_id - d1->func_id ?: d0->offset - d1->offset;
2520 static int kfunc_btf_cmp_by_off(const void *a, const void *b)
2522 const struct bpf_kfunc_btf *d0 = a;
2523 const struct bpf_kfunc_btf *d1 = b;
2525 return d0->offset - d1->offset;
2528 static const struct bpf_kfunc_desc *
2529 find_kfunc_desc(const struct bpf_prog *prog, u32 func_id, u16 offset)
2531 struct bpf_kfunc_desc desc = {
2535 struct bpf_kfunc_desc_tab *tab;
2537 tab = prog->aux->kfunc_tab;
2538 return bsearch(&desc, tab->descs, tab->nr_descs,
2539 sizeof(tab->descs[0]), kfunc_desc_cmp_by_id_off);
2542 int bpf_get_kfunc_addr(const struct bpf_prog *prog, u32 func_id,
2543 u16 btf_fd_idx, u8 **func_addr)
2545 const struct bpf_kfunc_desc *desc;
2547 desc = find_kfunc_desc(prog, func_id, btf_fd_idx);
2551 *func_addr = (u8 *)desc->addr;
2555 static struct btf *__find_kfunc_desc_btf(struct bpf_verifier_env *env,
2558 struct bpf_kfunc_btf kf_btf = { .offset = offset };
2559 struct bpf_kfunc_btf_tab *tab;
2560 struct bpf_kfunc_btf *b;
2565 tab = env->prog->aux->kfunc_btf_tab;
2566 b = bsearch(&kf_btf, tab->descs, tab->nr_descs,
2567 sizeof(tab->descs[0]), kfunc_btf_cmp_by_off);
2569 if (tab->nr_descs == MAX_KFUNC_BTFS) {
2570 verbose(env, "too many different module BTFs\n");
2571 return ERR_PTR(-E2BIG);
2574 if (bpfptr_is_null(env->fd_array)) {
2575 verbose(env, "kfunc offset > 0 without fd_array is invalid\n");
2576 return ERR_PTR(-EPROTO);
2579 if (copy_from_bpfptr_offset(&btf_fd, env->fd_array,
2580 offset * sizeof(btf_fd),
2582 return ERR_PTR(-EFAULT);
2584 btf = btf_get_by_fd(btf_fd);
2586 verbose(env, "invalid module BTF fd specified\n");
2590 if (!btf_is_module(btf)) {
2591 verbose(env, "BTF fd for kfunc is not a module BTF\n");
2593 return ERR_PTR(-EINVAL);
2596 mod = btf_try_get_module(btf);
2599 return ERR_PTR(-ENXIO);
2602 b = &tab->descs[tab->nr_descs++];
2607 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2608 kfunc_btf_cmp_by_off, NULL);
2613 void bpf_free_kfunc_btf_tab(struct bpf_kfunc_btf_tab *tab)
2618 while (tab->nr_descs--) {
2619 module_put(tab->descs[tab->nr_descs].module);
2620 btf_put(tab->descs[tab->nr_descs].btf);
2625 static struct btf *find_kfunc_desc_btf(struct bpf_verifier_env *env, s16 offset)
2629 /* In the future, this can be allowed to increase limit
2630 * of fd index into fd_array, interpreted as u16.
2632 verbose(env, "negative offset disallowed for kernel module function call\n");
2633 return ERR_PTR(-EINVAL);
2636 return __find_kfunc_desc_btf(env, offset);
2638 return btf_vmlinux ?: ERR_PTR(-ENOENT);
2641 static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id, s16 offset)
2643 const struct btf_type *func, *func_proto;
2644 struct bpf_kfunc_btf_tab *btf_tab;
2645 struct bpf_kfunc_desc_tab *tab;
2646 struct bpf_prog_aux *prog_aux;
2647 struct bpf_kfunc_desc *desc;
2648 const char *func_name;
2649 struct btf *desc_btf;
2650 unsigned long call_imm;
2654 prog_aux = env->prog->aux;
2655 tab = prog_aux->kfunc_tab;
2656 btf_tab = prog_aux->kfunc_btf_tab;
2659 verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
2663 if (!env->prog->jit_requested) {
2664 verbose(env, "JIT is required for calling kernel function\n");
2668 if (!bpf_jit_supports_kfunc_call()) {
2669 verbose(env, "JIT does not support calling kernel function\n");
2673 if (!env->prog->gpl_compatible) {
2674 verbose(env, "cannot call kernel function from non-GPL compatible program\n");
2678 tab = kzalloc(sizeof(*tab), GFP_KERNEL);
2681 prog_aux->kfunc_tab = tab;
2684 /* func_id == 0 is always invalid, but instead of returning an error, be
2685 * conservative and wait until the code elimination pass before returning
2686 * error, so that invalid calls that get pruned out can be in BPF programs
2687 * loaded from userspace. It is also required that offset be untouched
2690 if (!func_id && !offset)
2693 if (!btf_tab && offset) {
2694 btf_tab = kzalloc(sizeof(*btf_tab), GFP_KERNEL);
2697 prog_aux->kfunc_btf_tab = btf_tab;
2700 desc_btf = find_kfunc_desc_btf(env, offset);
2701 if (IS_ERR(desc_btf)) {
2702 verbose(env, "failed to find BTF for kernel function\n");
2703 return PTR_ERR(desc_btf);
2706 if (find_kfunc_desc(env->prog, func_id, offset))
2709 if (tab->nr_descs == MAX_KFUNC_DESCS) {
2710 verbose(env, "too many different kernel function calls\n");
2714 func = btf_type_by_id(desc_btf, func_id);
2715 if (!func || !btf_type_is_func(func)) {
2716 verbose(env, "kernel btf_id %u is not a function\n",
2720 func_proto = btf_type_by_id(desc_btf, func->type);
2721 if (!func_proto || !btf_type_is_func_proto(func_proto)) {
2722 verbose(env, "kernel function btf_id %u does not have a valid func_proto\n",
2727 func_name = btf_name_by_offset(desc_btf, func->name_off);
2728 addr = kallsyms_lookup_name(func_name);
2730 verbose(env, "cannot find address for kernel function %s\n",
2734 specialize_kfunc(env, func_id, offset, &addr);
2736 if (bpf_jit_supports_far_kfunc_call()) {
2739 call_imm = BPF_CALL_IMM(addr);
2740 /* Check whether the relative offset overflows desc->imm */
2741 if ((unsigned long)(s32)call_imm != call_imm) {
2742 verbose(env, "address of kernel function %s is out of range\n",
2748 if (bpf_dev_bound_kfunc_id(func_id)) {
2749 err = bpf_dev_bound_kfunc_check(&env->log, prog_aux);
2754 desc = &tab->descs[tab->nr_descs++];
2755 desc->func_id = func_id;
2756 desc->imm = call_imm;
2757 desc->offset = offset;
2759 err = btf_distill_func_proto(&env->log, desc_btf,
2760 func_proto, func_name,
2763 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2764 kfunc_desc_cmp_by_id_off, NULL);
2768 static int kfunc_desc_cmp_by_imm_off(const void *a, const void *b)
2770 const struct bpf_kfunc_desc *d0 = a;
2771 const struct bpf_kfunc_desc *d1 = b;
2773 if (d0->imm != d1->imm)
2774 return d0->imm < d1->imm ? -1 : 1;
2775 if (d0->offset != d1->offset)
2776 return d0->offset < d1->offset ? -1 : 1;
2780 static void sort_kfunc_descs_by_imm_off(struct bpf_prog *prog)
2782 struct bpf_kfunc_desc_tab *tab;
2784 tab = prog->aux->kfunc_tab;
2788 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2789 kfunc_desc_cmp_by_imm_off, NULL);
2792 bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
2794 return !!prog->aux->kfunc_tab;
2797 const struct btf_func_model *
2798 bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
2799 const struct bpf_insn *insn)
2801 const struct bpf_kfunc_desc desc = {
2803 .offset = insn->off,
2805 const struct bpf_kfunc_desc *res;
2806 struct bpf_kfunc_desc_tab *tab;
2808 tab = prog->aux->kfunc_tab;
2809 res = bsearch(&desc, tab->descs, tab->nr_descs,
2810 sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm_off);
2812 return res ? &res->func_model : NULL;
2815 static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
2817 struct bpf_subprog_info *subprog = env->subprog_info;
2818 struct bpf_insn *insn = env->prog->insnsi;
2819 int i, ret, insn_cnt = env->prog->len;
2821 /* Add entry function. */
2822 ret = add_subprog(env, 0);
2826 for (i = 0; i < insn_cnt; i++, insn++) {
2827 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
2828 !bpf_pseudo_kfunc_call(insn))
2831 if (!env->bpf_capable) {
2832 verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
2836 if (bpf_pseudo_func(insn) || bpf_pseudo_call(insn))
2837 ret = add_subprog(env, i + insn->imm + 1);
2839 ret = add_kfunc_call(env, insn->imm, insn->off);
2845 /* Add a fake 'exit' subprog which could simplify subprog iteration
2846 * logic. 'subprog_cnt' should not be increased.
2848 subprog[env->subprog_cnt].start = insn_cnt;
2850 if (env->log.level & BPF_LOG_LEVEL2)
2851 for (i = 0; i < env->subprog_cnt; i++)
2852 verbose(env, "func#%d @%d\n", i, subprog[i].start);
2857 static int check_subprogs(struct bpf_verifier_env *env)
2859 int i, subprog_start, subprog_end, off, cur_subprog = 0;
2860 struct bpf_subprog_info *subprog = env->subprog_info;
2861 struct bpf_insn *insn = env->prog->insnsi;
2862 int insn_cnt = env->prog->len;
2864 /* now check that all jumps are within the same subprog */
2865 subprog_start = subprog[cur_subprog].start;
2866 subprog_end = subprog[cur_subprog + 1].start;
2867 for (i = 0; i < insn_cnt; i++) {
2868 u8 code = insn[i].code;
2870 if (code == (BPF_JMP | BPF_CALL) &&
2871 insn[i].src_reg == 0 &&
2872 insn[i].imm == BPF_FUNC_tail_call)
2873 subprog[cur_subprog].has_tail_call = true;
2874 if (BPF_CLASS(code) == BPF_LD &&
2875 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
2876 subprog[cur_subprog].has_ld_abs = true;
2877 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
2879 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
2881 if (code == (BPF_JMP32 | BPF_JA))
2882 off = i + insn[i].imm + 1;
2884 off = i + insn[i].off + 1;
2885 if (off < subprog_start || off >= subprog_end) {
2886 verbose(env, "jump out of range from insn %d to %d\n", i, off);
2890 if (i == subprog_end - 1) {
2891 /* to avoid fall-through from one subprog into another
2892 * the last insn of the subprog should be either exit
2893 * or unconditional jump back
2895 if (code != (BPF_JMP | BPF_EXIT) &&
2896 code != (BPF_JMP32 | BPF_JA) &&
2897 code != (BPF_JMP | BPF_JA)) {
2898 verbose(env, "last insn is not an exit or jmp\n");
2901 subprog_start = subprog_end;
2903 if (cur_subprog < env->subprog_cnt)
2904 subprog_end = subprog[cur_subprog + 1].start;
2910 /* Parentage chain of this register (or stack slot) should take care of all
2911 * issues like callee-saved registers, stack slot allocation time, etc.
2913 static int mark_reg_read(struct bpf_verifier_env *env,
2914 const struct bpf_reg_state *state,
2915 struct bpf_reg_state *parent, u8 flag)
2917 bool writes = parent == state->parent; /* Observe write marks */
2921 /* if read wasn't screened by an earlier write ... */
2922 if (writes && state->live & REG_LIVE_WRITTEN)
2924 if (parent->live & REG_LIVE_DONE) {
2925 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
2926 reg_type_str(env, parent->type),
2927 parent->var_off.value, parent->off);
2930 /* The first condition is more likely to be true than the
2931 * second, checked it first.
2933 if ((parent->live & REG_LIVE_READ) == flag ||
2934 parent->live & REG_LIVE_READ64)
2935 /* The parentage chain never changes and
2936 * this parent was already marked as LIVE_READ.
2937 * There is no need to keep walking the chain again and
2938 * keep re-marking all parents as LIVE_READ.
2939 * This case happens when the same register is read
2940 * multiple times without writes into it in-between.
2941 * Also, if parent has the stronger REG_LIVE_READ64 set,
2942 * then no need to set the weak REG_LIVE_READ32.
2945 /* ... then we depend on parent's value */
2946 parent->live |= flag;
2947 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
2948 if (flag == REG_LIVE_READ64)
2949 parent->live &= ~REG_LIVE_READ32;
2951 parent = state->parent;
2956 if (env->longest_mark_read_walk < cnt)
2957 env->longest_mark_read_walk = cnt;
2961 static int mark_dynptr_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
2963 struct bpf_func_state *state = func(env, reg);
2966 /* For CONST_PTR_TO_DYNPTR, it must have already been done by
2967 * check_reg_arg in check_helper_call and mark_btf_func_reg_size in
2970 if (reg->type == CONST_PTR_TO_DYNPTR)
2972 spi = dynptr_get_spi(env, reg);
2975 /* Caller ensures dynptr is valid and initialized, which means spi is in
2976 * bounds and spi is the first dynptr slot. Simply mark stack slot as
2979 ret = mark_reg_read(env, &state->stack[spi].spilled_ptr,
2980 state->stack[spi].spilled_ptr.parent, REG_LIVE_READ64);
2983 return mark_reg_read(env, &state->stack[spi - 1].spilled_ptr,
2984 state->stack[spi - 1].spilled_ptr.parent, REG_LIVE_READ64);
2987 static int mark_iter_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
2988 int spi, int nr_slots)
2990 struct bpf_func_state *state = func(env, reg);
2993 for (i = 0; i < nr_slots; i++) {
2994 struct bpf_reg_state *st = &state->stack[spi - i].spilled_ptr;
2996 err = mark_reg_read(env, st, st->parent, REG_LIVE_READ64);
3000 mark_stack_slot_scratched(env, spi - i);
3006 /* This function is supposed to be used by the following 32-bit optimization
3007 * code only. It returns TRUE if the source or destination register operates
3008 * on 64-bit, otherwise return FALSE.
3010 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
3011 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
3016 class = BPF_CLASS(code);
3018 if (class == BPF_JMP) {
3019 /* BPF_EXIT for "main" will reach here. Return TRUE
3024 if (op == BPF_CALL) {
3025 /* BPF to BPF call will reach here because of marking
3026 * caller saved clobber with DST_OP_NO_MARK for which we
3027 * don't care the register def because they are anyway
3028 * marked as NOT_INIT already.
3030 if (insn->src_reg == BPF_PSEUDO_CALL)
3032 /* Helper call will reach here because of arg type
3033 * check, conservatively return TRUE.
3042 if (class == BPF_ALU64 && op == BPF_END && (insn->imm == 16 || insn->imm == 32))
3045 if (class == BPF_ALU64 || class == BPF_JMP ||
3046 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
3049 if (class == BPF_ALU || class == BPF_JMP32)
3052 if (class == BPF_LDX) {
3054 return BPF_SIZE(code) == BPF_DW;
3055 /* LDX source must be ptr. */
3059 if (class == BPF_STX) {
3060 /* BPF_STX (including atomic variants) has multiple source
3061 * operands, one of which is a ptr. Check whether the caller is
3064 if (t == SRC_OP && reg->type != SCALAR_VALUE)
3066 return BPF_SIZE(code) == BPF_DW;
3069 if (class == BPF_LD) {
3070 u8 mode = BPF_MODE(code);
3073 if (mode == BPF_IMM)
3076 /* Both LD_IND and LD_ABS return 32-bit data. */
3080 /* Implicit ctx ptr. */
3081 if (regno == BPF_REG_6)
3084 /* Explicit source could be any width. */
3088 if (class == BPF_ST)
3089 /* The only source register for BPF_ST is a ptr. */
3092 /* Conservatively return true at default. */
3096 /* Return the regno defined by the insn, or -1. */
3097 static int insn_def_regno(const struct bpf_insn *insn)
3099 switch (BPF_CLASS(insn->code)) {
3105 if (BPF_MODE(insn->code) == BPF_ATOMIC &&
3106 (insn->imm & BPF_FETCH)) {
3107 if (insn->imm == BPF_CMPXCHG)
3110 return insn->src_reg;
3115 return insn->dst_reg;
3119 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
3120 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
3122 int dst_reg = insn_def_regno(insn);
3127 return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
3130 static void mark_insn_zext(struct bpf_verifier_env *env,
3131 struct bpf_reg_state *reg)
3133 s32 def_idx = reg->subreg_def;
3135 if (def_idx == DEF_NOT_SUBREG)
3138 env->insn_aux_data[def_idx - 1].zext_dst = true;
3139 /* The dst will be zero extended, so won't be sub-register anymore. */
3140 reg->subreg_def = DEF_NOT_SUBREG;
3143 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
3144 enum reg_arg_type t)
3146 struct bpf_verifier_state *vstate = env->cur_state;
3147 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3148 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
3149 struct bpf_reg_state *reg, *regs = state->regs;
3152 if (regno >= MAX_BPF_REG) {
3153 verbose(env, "R%d is invalid\n", regno);
3157 mark_reg_scratched(env, regno);
3160 rw64 = is_reg64(env, insn, regno, reg, t);
3162 /* check whether register used as source operand can be read */
3163 if (reg->type == NOT_INIT) {
3164 verbose(env, "R%d !read_ok\n", regno);
3167 /* We don't need to worry about FP liveness because it's read-only */
3168 if (regno == BPF_REG_FP)
3172 mark_insn_zext(env, reg);
3174 return mark_reg_read(env, reg, reg->parent,
3175 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
3177 /* check whether register used as dest operand can be written to */
3178 if (regno == BPF_REG_FP) {
3179 verbose(env, "frame pointer is read only\n");
3182 reg->live |= REG_LIVE_WRITTEN;
3183 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
3185 mark_reg_unknown(env, regs, regno);
3190 static void mark_jmp_point(struct bpf_verifier_env *env, int idx)
3192 env->insn_aux_data[idx].jmp_point = true;
3195 static bool is_jmp_point(struct bpf_verifier_env *env, int insn_idx)
3197 return env->insn_aux_data[insn_idx].jmp_point;
3200 /* for any branch, call, exit record the history of jmps in the given state */
3201 static int push_jmp_history(struct bpf_verifier_env *env,
3202 struct bpf_verifier_state *cur)
3204 u32 cnt = cur->jmp_history_cnt;
3205 struct bpf_idx_pair *p;
3208 if (!is_jmp_point(env, env->insn_idx))
3212 alloc_size = kmalloc_size_roundup(size_mul(cnt, sizeof(*p)));
3213 p = krealloc(cur->jmp_history, alloc_size, GFP_USER);
3216 p[cnt - 1].idx = env->insn_idx;
3217 p[cnt - 1].prev_idx = env->prev_insn_idx;
3218 cur->jmp_history = p;
3219 cur->jmp_history_cnt = cnt;
3223 /* Backtrack one insn at a time. If idx is not at the top of recorded
3224 * history then previous instruction came from straight line execution.
3225 * Return -ENOENT if we exhausted all instructions within given state.
3227 * It's legal to have a bit of a looping with the same starting and ending
3228 * insn index within the same state, e.g.: 3->4->5->3, so just because current
3229 * instruction index is the same as state's first_idx doesn't mean we are
3230 * done. If there is still some jump history left, we should keep going. We
3231 * need to take into account that we might have a jump history between given
3232 * state's parent and itself, due to checkpointing. In this case, we'll have
3233 * history entry recording a jump from last instruction of parent state and
3234 * first instruction of given state.
3236 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
3241 if (i == st->first_insn_idx) {
3244 if (cnt == 1 && st->jmp_history[0].idx == i)
3248 if (cnt && st->jmp_history[cnt - 1].idx == i) {
3249 i = st->jmp_history[cnt - 1].prev_idx;
3257 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
3259 const struct btf_type *func;
3260 struct btf *desc_btf;
3262 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
3265 desc_btf = find_kfunc_desc_btf(data, insn->off);
3266 if (IS_ERR(desc_btf))
3269 func = btf_type_by_id(desc_btf, insn->imm);
3270 return btf_name_by_offset(desc_btf, func->name_off);
3273 static inline void bt_init(struct backtrack_state *bt, u32 frame)
3278 static inline void bt_reset(struct backtrack_state *bt)
3280 struct bpf_verifier_env *env = bt->env;
3282 memset(bt, 0, sizeof(*bt));
3286 static inline u32 bt_empty(struct backtrack_state *bt)
3291 for (i = 0; i <= bt->frame; i++)
3292 mask |= bt->reg_masks[i] | bt->stack_masks[i];
3297 static inline int bt_subprog_enter(struct backtrack_state *bt)
3299 if (bt->frame == MAX_CALL_FRAMES - 1) {
3300 verbose(bt->env, "BUG subprog enter from frame %d\n", bt->frame);
3301 WARN_ONCE(1, "verifier backtracking bug");
3308 static inline int bt_subprog_exit(struct backtrack_state *bt)
3310 if (bt->frame == 0) {
3311 verbose(bt->env, "BUG subprog exit from frame 0\n");
3312 WARN_ONCE(1, "verifier backtracking bug");
3319 static inline void bt_set_frame_reg(struct backtrack_state *bt, u32 frame, u32 reg)
3321 bt->reg_masks[frame] |= 1 << reg;
3324 static inline void bt_clear_frame_reg(struct backtrack_state *bt, u32 frame, u32 reg)
3326 bt->reg_masks[frame] &= ~(1 << reg);
3329 static inline void bt_set_reg(struct backtrack_state *bt, u32 reg)
3331 bt_set_frame_reg(bt, bt->frame, reg);
3334 static inline void bt_clear_reg(struct backtrack_state *bt, u32 reg)
3336 bt_clear_frame_reg(bt, bt->frame, reg);
3339 static inline void bt_set_frame_slot(struct backtrack_state *bt, u32 frame, u32 slot)
3341 bt->stack_masks[frame] |= 1ull << slot;
3344 static inline void bt_clear_frame_slot(struct backtrack_state *bt, u32 frame, u32 slot)
3346 bt->stack_masks[frame] &= ~(1ull << slot);
3349 static inline void bt_set_slot(struct backtrack_state *bt, u32 slot)
3351 bt_set_frame_slot(bt, bt->frame, slot);
3354 static inline void bt_clear_slot(struct backtrack_state *bt, u32 slot)
3356 bt_clear_frame_slot(bt, bt->frame, slot);
3359 static inline u32 bt_frame_reg_mask(struct backtrack_state *bt, u32 frame)
3361 return bt->reg_masks[frame];
3364 static inline u32 bt_reg_mask(struct backtrack_state *bt)
3366 return bt->reg_masks[bt->frame];
3369 static inline u64 bt_frame_stack_mask(struct backtrack_state *bt, u32 frame)
3371 return bt->stack_masks[frame];
3374 static inline u64 bt_stack_mask(struct backtrack_state *bt)
3376 return bt->stack_masks[bt->frame];
3379 static inline bool bt_is_reg_set(struct backtrack_state *bt, u32 reg)
3381 return bt->reg_masks[bt->frame] & (1 << reg);
3384 static inline bool bt_is_slot_set(struct backtrack_state *bt, u32 slot)
3386 return bt->stack_masks[bt->frame] & (1ull << slot);
3389 /* format registers bitmask, e.g., "r0,r2,r4" for 0x15 mask */
3390 static void fmt_reg_mask(char *buf, ssize_t buf_sz, u32 reg_mask)
3392 DECLARE_BITMAP(mask, 64);
3398 bitmap_from_u64(mask, reg_mask);
3399 for_each_set_bit(i, mask, 32) {
3400 n = snprintf(buf, buf_sz, "%sr%d", first ? "" : ",", i);
3408 /* format stack slots bitmask, e.g., "-8,-24,-40" for 0x15 mask */
3409 static void fmt_stack_mask(char *buf, ssize_t buf_sz, u64 stack_mask)
3411 DECLARE_BITMAP(mask, 64);
3417 bitmap_from_u64(mask, stack_mask);
3418 for_each_set_bit(i, mask, 64) {
3419 n = snprintf(buf, buf_sz, "%s%d", first ? "" : ",", -(i + 1) * 8);
3428 /* For given verifier state backtrack_insn() is called from the last insn to
3429 * the first insn. Its purpose is to compute a bitmask of registers and
3430 * stack slots that needs precision in the parent verifier state.
3432 * @idx is an index of the instruction we are currently processing;
3433 * @subseq_idx is an index of the subsequent instruction that:
3434 * - *would be* executed next, if jump history is viewed in forward order;
3435 * - *was* processed previously during backtracking.
3437 static int backtrack_insn(struct bpf_verifier_env *env, int idx, int subseq_idx,
3438 struct backtrack_state *bt)
3440 const struct bpf_insn_cbs cbs = {
3441 .cb_call = disasm_kfunc_name,
3442 .cb_print = verbose,
3443 .private_data = env,
3445 struct bpf_insn *insn = env->prog->insnsi + idx;
3446 u8 class = BPF_CLASS(insn->code);
3447 u8 opcode = BPF_OP(insn->code);
3448 u8 mode = BPF_MODE(insn->code);
3449 u32 dreg = insn->dst_reg;
3450 u32 sreg = insn->src_reg;
3453 if (insn->code == 0)
3455 if (env->log.level & BPF_LOG_LEVEL2) {
3456 fmt_reg_mask(env->tmp_str_buf, TMP_STR_BUF_LEN, bt_reg_mask(bt));
3457 verbose(env, "mark_precise: frame%d: regs=%s ",
3458 bt->frame, env->tmp_str_buf);
3459 fmt_stack_mask(env->tmp_str_buf, TMP_STR_BUF_LEN, bt_stack_mask(bt));
3460 verbose(env, "stack=%s before ", env->tmp_str_buf);
3461 verbose(env, "%d: ", idx);
3462 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
3465 if (class == BPF_ALU || class == BPF_ALU64) {
3466 if (!bt_is_reg_set(bt, dreg))
3468 if (opcode == BPF_END || opcode == BPF_NEG) {
3469 /* sreg is reserved and unused
3470 * dreg still need precision before this insn
3473 } else if (opcode == BPF_MOV) {
3474 if (BPF_SRC(insn->code) == BPF_X) {
3475 /* dreg = sreg or dreg = (s8, s16, s32)sreg
3476 * dreg needs precision after this insn
3477 * sreg needs precision before this insn
3479 bt_clear_reg(bt, dreg);
3480 bt_set_reg(bt, sreg);
3483 * dreg needs precision after this insn.
3484 * Corresponding register is already marked
3485 * as precise=true in this verifier state.
3486 * No further markings in parent are necessary
3488 bt_clear_reg(bt, dreg);
3491 if (BPF_SRC(insn->code) == BPF_X) {
3493 * both dreg and sreg need precision
3496 bt_set_reg(bt, sreg);
3498 * dreg still needs precision before this insn
3501 } else if (class == BPF_LDX) {
3502 if (!bt_is_reg_set(bt, dreg))
3504 bt_clear_reg(bt, dreg);
3506 /* scalars can only be spilled into stack w/o losing precision.
3507 * Load from any other memory can be zero extended.
3508 * The desire to keep that precision is already indicated
3509 * by 'precise' mark in corresponding register of this state.
3510 * No further tracking necessary.
3512 if (insn->src_reg != BPF_REG_FP)
3515 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
3516 * that [fp - off] slot contains scalar that needs to be
3517 * tracked with precision
3519 spi = (-insn->off - 1) / BPF_REG_SIZE;
3521 verbose(env, "BUG spi %d\n", spi);
3522 WARN_ONCE(1, "verifier backtracking bug");
3525 bt_set_slot(bt, spi);
3526 } else if (class == BPF_STX || class == BPF_ST) {
3527 if (bt_is_reg_set(bt, dreg))
3528 /* stx & st shouldn't be using _scalar_ dst_reg
3529 * to access memory. It means backtracking
3530 * encountered a case of pointer subtraction.
3533 /* scalars can only be spilled into stack */
3534 if (insn->dst_reg != BPF_REG_FP)
3536 spi = (-insn->off - 1) / BPF_REG_SIZE;
3538 verbose(env, "BUG spi %d\n", spi);
3539 WARN_ONCE(1, "verifier backtracking bug");
3542 if (!bt_is_slot_set(bt, spi))
3544 bt_clear_slot(bt, spi);
3545 if (class == BPF_STX)
3546 bt_set_reg(bt, sreg);
3547 } else if (class == BPF_JMP || class == BPF_JMP32) {
3548 if (bpf_pseudo_call(insn)) {
3549 int subprog_insn_idx, subprog;
3551 subprog_insn_idx = idx + insn->imm + 1;
3552 subprog = find_subprog(env, subprog_insn_idx);
3556 if (subprog_is_global(env, subprog)) {
3557 /* check that jump history doesn't have any
3558 * extra instructions from subprog; the next
3559 * instruction after call to global subprog
3560 * should be literally next instruction in
3563 WARN_ONCE(idx + 1 != subseq_idx, "verifier backtracking bug");
3564 /* r1-r5 are invalidated after subprog call,
3565 * so for global func call it shouldn't be set
3568 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3569 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3570 WARN_ONCE(1, "verifier backtracking bug");
3573 /* global subprog always sets R0 */
3574 bt_clear_reg(bt, BPF_REG_0);
3577 /* static subprog call instruction, which
3578 * means that we are exiting current subprog,
3579 * so only r1-r5 could be still requested as
3580 * precise, r0 and r6-r10 or any stack slot in
3581 * the current frame should be zero by now
3583 if (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) {
3584 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3585 WARN_ONCE(1, "verifier backtracking bug");
3588 /* we don't track register spills perfectly,
3589 * so fallback to force-precise instead of failing */
3590 if (bt_stack_mask(bt) != 0)
3592 /* propagate r1-r5 to the caller */
3593 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
3594 if (bt_is_reg_set(bt, i)) {
3595 bt_clear_reg(bt, i);
3596 bt_set_frame_reg(bt, bt->frame - 1, i);
3599 if (bt_subprog_exit(bt))
3603 } else if ((bpf_helper_call(insn) &&
3604 is_callback_calling_function(insn->imm) &&
3605 !is_async_callback_calling_function(insn->imm)) ||
3606 (bpf_pseudo_kfunc_call(insn) && is_callback_calling_kfunc(insn->imm))) {
3607 /* callback-calling helper or kfunc call, which means
3608 * we are exiting from subprog, but unlike the subprog
3609 * call handling above, we shouldn't propagate
3610 * precision of r1-r5 (if any requested), as they are
3611 * not actually arguments passed directly to callback
3614 if (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) {
3615 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3616 WARN_ONCE(1, "verifier backtracking bug");
3619 if (bt_stack_mask(bt) != 0)
3621 /* clear r1-r5 in callback subprog's mask */
3622 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
3623 bt_clear_reg(bt, i);
3624 if (bt_subprog_exit(bt))
3627 } else if (opcode == BPF_CALL) {
3628 /* kfunc with imm==0 is invalid and fixup_kfunc_call will
3629 * catch this error later. Make backtracking conservative
3632 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && insn->imm == 0)
3634 /* regular helper call sets R0 */
3635 bt_clear_reg(bt, BPF_REG_0);
3636 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3637 /* if backtracing was looking for registers R1-R5
3638 * they should have been found already.
3640 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3641 WARN_ONCE(1, "verifier backtracking bug");
3644 } else if (opcode == BPF_EXIT) {
3647 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3648 /* if backtracing was looking for registers R1-R5
3649 * they should have been found already.
3651 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3652 WARN_ONCE(1, "verifier backtracking bug");
3656 /* BPF_EXIT in subprog or callback always returns
3657 * right after the call instruction, so by checking
3658 * whether the instruction at subseq_idx-1 is subprog
3659 * call or not we can distinguish actual exit from
3660 * *subprog* from exit from *callback*. In the former
3661 * case, we need to propagate r0 precision, if
3662 * necessary. In the former we never do that.
3664 r0_precise = subseq_idx - 1 >= 0 &&
3665 bpf_pseudo_call(&env->prog->insnsi[subseq_idx - 1]) &&
3666 bt_is_reg_set(bt, BPF_REG_0);
3668 bt_clear_reg(bt, BPF_REG_0);
3669 if (bt_subprog_enter(bt))
3673 bt_set_reg(bt, BPF_REG_0);
3674 /* r6-r9 and stack slots will stay set in caller frame
3675 * bitmasks until we return back from callee(s)
3678 } else if (BPF_SRC(insn->code) == BPF_X) {
3679 if (!bt_is_reg_set(bt, dreg) && !bt_is_reg_set(bt, sreg))
3682 * Both dreg and sreg need precision before
3683 * this insn. If only sreg was marked precise
3684 * before it would be equally necessary to
3685 * propagate it to dreg.
3687 bt_set_reg(bt, dreg);
3688 bt_set_reg(bt, sreg);
3689 /* else dreg <cond> K
3690 * Only dreg still needs precision before
3691 * this insn, so for the K-based conditional
3692 * there is nothing new to be marked.
3695 } else if (class == BPF_LD) {
3696 if (!bt_is_reg_set(bt, dreg))
3698 bt_clear_reg(bt, dreg);
3699 /* It's ld_imm64 or ld_abs or ld_ind.
3700 * For ld_imm64 no further tracking of precision
3701 * into parent is necessary
3703 if (mode == BPF_IND || mode == BPF_ABS)
3704 /* to be analyzed */
3710 /* the scalar precision tracking algorithm:
3711 * . at the start all registers have precise=false.
3712 * . scalar ranges are tracked as normal through alu and jmp insns.
3713 * . once precise value of the scalar register is used in:
3714 * . ptr + scalar alu
3715 * . if (scalar cond K|scalar)
3716 * . helper_call(.., scalar, ...) where ARG_CONST is expected
3717 * backtrack through the verifier states and mark all registers and
3718 * stack slots with spilled constants that these scalar regisers
3719 * should be precise.
3720 * . during state pruning two registers (or spilled stack slots)
3721 * are equivalent if both are not precise.
3723 * Note the verifier cannot simply walk register parentage chain,
3724 * since many different registers and stack slots could have been
3725 * used to compute single precise scalar.
3727 * The approach of starting with precise=true for all registers and then
3728 * backtrack to mark a register as not precise when the verifier detects
3729 * that program doesn't care about specific value (e.g., when helper
3730 * takes register as ARG_ANYTHING parameter) is not safe.
3732 * It's ok to walk single parentage chain of the verifier states.
3733 * It's possible that this backtracking will go all the way till 1st insn.
3734 * All other branches will be explored for needing precision later.
3736 * The backtracking needs to deal with cases like:
3737 * 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)
3740 * if r5 > 0x79f goto pc+7
3741 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
3744 * call bpf_perf_event_output#25
3745 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
3749 * call foo // uses callee's r6 inside to compute r0
3753 * to track above reg_mask/stack_mask needs to be independent for each frame.
3755 * Also if parent's curframe > frame where backtracking started,
3756 * the verifier need to mark registers in both frames, otherwise callees
3757 * may incorrectly prune callers. This is similar to
3758 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
3760 * For now backtracking falls back into conservative marking.
3762 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
3763 struct bpf_verifier_state *st)
3765 struct bpf_func_state *func;
3766 struct bpf_reg_state *reg;
3769 if (env->log.level & BPF_LOG_LEVEL2) {
3770 verbose(env, "mark_precise: frame%d: falling back to forcing all scalars precise\n",
3774 /* big hammer: mark all scalars precise in this path.
3775 * pop_stack may still get !precise scalars.
3776 * We also skip current state and go straight to first parent state,
3777 * because precision markings in current non-checkpointed state are
3778 * not needed. See why in the comment in __mark_chain_precision below.
3780 for (st = st->parent; st; st = st->parent) {
3781 for (i = 0; i <= st->curframe; i++) {
3782 func = st->frame[i];
3783 for (j = 0; j < BPF_REG_FP; j++) {
3784 reg = &func->regs[j];
3785 if (reg->type != SCALAR_VALUE || reg->precise)
3787 reg->precise = true;
3788 if (env->log.level & BPF_LOG_LEVEL2) {
3789 verbose(env, "force_precise: frame%d: forcing r%d to be precise\n",
3793 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
3794 if (!is_spilled_reg(&func->stack[j]))
3796 reg = &func->stack[j].spilled_ptr;
3797 if (reg->type != SCALAR_VALUE || reg->precise)
3799 reg->precise = true;
3800 if (env->log.level & BPF_LOG_LEVEL2) {
3801 verbose(env, "force_precise: frame%d: forcing fp%d to be precise\n",
3809 static void mark_all_scalars_imprecise(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
3811 struct bpf_func_state *func;
3812 struct bpf_reg_state *reg;
3815 for (i = 0; i <= st->curframe; i++) {
3816 func = st->frame[i];
3817 for (j = 0; j < BPF_REG_FP; j++) {
3818 reg = &func->regs[j];
3819 if (reg->type != SCALAR_VALUE)
3821 reg->precise = false;
3823 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
3824 if (!is_spilled_reg(&func->stack[j]))
3826 reg = &func->stack[j].spilled_ptr;
3827 if (reg->type != SCALAR_VALUE)
3829 reg->precise = false;
3834 static bool idset_contains(struct bpf_idset *s, u32 id)
3838 for (i = 0; i < s->count; ++i)
3839 if (s->ids[i] == id)
3845 static int idset_push(struct bpf_idset *s, u32 id)
3847 if (WARN_ON_ONCE(s->count >= ARRAY_SIZE(s->ids)))
3849 s->ids[s->count++] = id;
3853 static void idset_reset(struct bpf_idset *s)
3858 /* Collect a set of IDs for all registers currently marked as precise in env->bt.
3859 * Mark all registers with these IDs as precise.
3861 static int mark_precise_scalar_ids(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
3863 struct bpf_idset *precise_ids = &env->idset_scratch;
3864 struct backtrack_state *bt = &env->bt;
3865 struct bpf_func_state *func;
3866 struct bpf_reg_state *reg;
3867 DECLARE_BITMAP(mask, 64);
3870 idset_reset(precise_ids);
3872 for (fr = bt->frame; fr >= 0; fr--) {
3873 func = st->frame[fr];
3875 bitmap_from_u64(mask, bt_frame_reg_mask(bt, fr));
3876 for_each_set_bit(i, mask, 32) {
3877 reg = &func->regs[i];
3878 if (!reg->id || reg->type != SCALAR_VALUE)
3880 if (idset_push(precise_ids, reg->id))
3884 bitmap_from_u64(mask, bt_frame_stack_mask(bt, fr));
3885 for_each_set_bit(i, mask, 64) {
3886 if (i >= func->allocated_stack / BPF_REG_SIZE)
3888 if (!is_spilled_scalar_reg(&func->stack[i]))
3890 reg = &func->stack[i].spilled_ptr;
3893 if (idset_push(precise_ids, reg->id))
3898 for (fr = 0; fr <= st->curframe; ++fr) {
3899 func = st->frame[fr];
3901 for (i = BPF_REG_0; i < BPF_REG_10; ++i) {
3902 reg = &func->regs[i];
3905 if (!idset_contains(precise_ids, reg->id))
3907 bt_set_frame_reg(bt, fr, i);
3909 for (i = 0; i < func->allocated_stack / BPF_REG_SIZE; ++i) {
3910 if (!is_spilled_scalar_reg(&func->stack[i]))
3912 reg = &func->stack[i].spilled_ptr;
3915 if (!idset_contains(precise_ids, reg->id))
3917 bt_set_frame_slot(bt, fr, i);
3925 * __mark_chain_precision() backtracks BPF program instruction sequence and
3926 * chain of verifier states making sure that register *regno* (if regno >= 0)
3927 * and/or stack slot *spi* (if spi >= 0) are marked as precisely tracked
3928 * SCALARS, as well as any other registers and slots that contribute to
3929 * a tracked state of given registers/stack slots, depending on specific BPF
3930 * assembly instructions (see backtrack_insns() for exact instruction handling
3931 * logic). This backtracking relies on recorded jmp_history and is able to
3932 * traverse entire chain of parent states. This process ends only when all the
3933 * necessary registers/slots and their transitive dependencies are marked as
3936 * One important and subtle aspect is that precise marks *do not matter* in
3937 * the currently verified state (current state). It is important to understand
3938 * why this is the case.
3940 * First, note that current state is the state that is not yet "checkpointed",
3941 * i.e., it is not yet put into env->explored_states, and it has no children
3942 * states as well. It's ephemeral, and can end up either a) being discarded if
3943 * compatible explored state is found at some point or BPF_EXIT instruction is
3944 * reached or b) checkpointed and put into env->explored_states, branching out
3945 * into one or more children states.
3947 * In the former case, precise markings in current state are completely
3948 * ignored by state comparison code (see regsafe() for details). Only
3949 * checkpointed ("old") state precise markings are important, and if old
3950 * state's register/slot is precise, regsafe() assumes current state's
3951 * register/slot as precise and checks value ranges exactly and precisely. If
3952 * states turn out to be compatible, current state's necessary precise
3953 * markings and any required parent states' precise markings are enforced
3954 * after the fact with propagate_precision() logic, after the fact. But it's
3955 * important to realize that in this case, even after marking current state
3956 * registers/slots as precise, we immediately discard current state. So what
3957 * actually matters is any of the precise markings propagated into current
3958 * state's parent states, which are always checkpointed (due to b) case above).
3959 * As such, for scenario a) it doesn't matter if current state has precise
3960 * markings set or not.
3962 * Now, for the scenario b), checkpointing and forking into child(ren)
3963 * state(s). Note that before current state gets to checkpointing step, any
3964 * processed instruction always assumes precise SCALAR register/slot
3965 * knowledge: if precise value or range is useful to prune jump branch, BPF
3966 * verifier takes this opportunity enthusiastically. Similarly, when
3967 * register's value is used to calculate offset or memory address, exact
3968 * knowledge of SCALAR range is assumed, checked, and enforced. So, similar to
3969 * what we mentioned above about state comparison ignoring precise markings
3970 * during state comparison, BPF verifier ignores and also assumes precise
3971 * markings *at will* during instruction verification process. But as verifier
3972 * assumes precision, it also propagates any precision dependencies across
3973 * parent states, which are not yet finalized, so can be further restricted
3974 * based on new knowledge gained from restrictions enforced by their children
3975 * states. This is so that once those parent states are finalized, i.e., when
3976 * they have no more active children state, state comparison logic in
3977 * is_state_visited() would enforce strict and precise SCALAR ranges, if
3978 * required for correctness.
3980 * To build a bit more intuition, note also that once a state is checkpointed,
3981 * the path we took to get to that state is not important. This is crucial
3982 * property for state pruning. When state is checkpointed and finalized at
3983 * some instruction index, it can be correctly and safely used to "short
3984 * circuit" any *compatible* state that reaches exactly the same instruction
3985 * index. I.e., if we jumped to that instruction from a completely different
3986 * code path than original finalized state was derived from, it doesn't
3987 * matter, current state can be discarded because from that instruction
3988 * forward having a compatible state will ensure we will safely reach the
3989 * exit. States describe preconditions for further exploration, but completely
3990 * forget the history of how we got here.
3992 * This also means that even if we needed precise SCALAR range to get to
3993 * finalized state, but from that point forward *that same* SCALAR register is
3994 * never used in a precise context (i.e., it's precise value is not needed for
3995 * correctness), it's correct and safe to mark such register as "imprecise"
3996 * (i.e., precise marking set to false). This is what we rely on when we do
3997 * not set precise marking in current state. If no child state requires
3998 * precision for any given SCALAR register, it's safe to dictate that it can
3999 * be imprecise. If any child state does require this register to be precise,
4000 * we'll mark it precise later retroactively during precise markings
4001 * propagation from child state to parent states.
4003 * Skipping precise marking setting in current state is a mild version of
4004 * relying on the above observation. But we can utilize this property even
4005 * more aggressively by proactively forgetting any precise marking in the
4006 * current state (which we inherited from the parent state), right before we
4007 * checkpoint it and branch off into new child state. This is done by
4008 * mark_all_scalars_imprecise() to hopefully get more permissive and generic
4009 * finalized states which help in short circuiting more future states.
4011 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno)
4013 struct backtrack_state *bt = &env->bt;
4014 struct bpf_verifier_state *st = env->cur_state;
4015 int first_idx = st->first_insn_idx;
4016 int last_idx = env->insn_idx;
4017 int subseq_idx = -1;
4018 struct bpf_func_state *func;
4019 struct bpf_reg_state *reg;
4020 bool skip_first = true;
4023 if (!env->bpf_capable)
4026 /* set frame number from which we are starting to backtrack */
4027 bt_init(bt, env->cur_state->curframe);
4029 /* Do sanity checks against current state of register and/or stack
4030 * slot, but don't set precise flag in current state, as precision
4031 * tracking in the current state is unnecessary.
4033 func = st->frame[bt->frame];
4035 reg = &func->regs[regno];
4036 if (reg->type != SCALAR_VALUE) {
4037 WARN_ONCE(1, "backtracing misuse");
4040 bt_set_reg(bt, regno);
4047 DECLARE_BITMAP(mask, 64);
4048 u32 history = st->jmp_history_cnt;
4050 if (env->log.level & BPF_LOG_LEVEL2) {
4051 verbose(env, "mark_precise: frame%d: last_idx %d first_idx %d subseq_idx %d \n",
4052 bt->frame, last_idx, first_idx, subseq_idx);
4055 /* If some register with scalar ID is marked as precise,
4056 * make sure that all registers sharing this ID are also precise.
4057 * This is needed to estimate effect of find_equal_scalars().
4058 * Do this at the last instruction of each state,
4059 * bpf_reg_state::id fields are valid for these instructions.
4061 * Allows to track precision in situation like below:
4063 * r2 = unknown value
4067 * r1 = r2 // r1 and r2 now share the same ID
4069 * --- state #1 {r1.id = A, r2.id = A} ---
4071 * if (r2 > 10) goto exit; // find_equal_scalars() assigns range to r1
4073 * --- state #2 {r1.id = A, r2.id = A} ---
4075 * r3 += r1 // need to mark both r1 and r2
4077 if (mark_precise_scalar_ids(env, st))
4081 /* we are at the entry into subprog, which
4082 * is expected for global funcs, but only if
4083 * requested precise registers are R1-R5
4084 * (which are global func's input arguments)
4086 if (st->curframe == 0 &&
4087 st->frame[0]->subprogno > 0 &&
4088 st->frame[0]->callsite == BPF_MAIN_FUNC &&
4089 bt_stack_mask(bt) == 0 &&
4090 (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) == 0) {
4091 bitmap_from_u64(mask, bt_reg_mask(bt));
4092 for_each_set_bit(i, mask, 32) {
4093 reg = &st->frame[0]->regs[i];
4094 bt_clear_reg(bt, i);
4095 if (reg->type == SCALAR_VALUE)
4096 reg->precise = true;
4101 verbose(env, "BUG backtracking func entry subprog %d reg_mask %x stack_mask %llx\n",
4102 st->frame[0]->subprogno, bt_reg_mask(bt), bt_stack_mask(bt));
4103 WARN_ONCE(1, "verifier backtracking bug");
4107 for (i = last_idx;;) {
4112 err = backtrack_insn(env, i, subseq_idx, bt);
4114 if (err == -ENOTSUPP) {
4115 mark_all_scalars_precise(env, env->cur_state);
4122 /* Found assignment(s) into tracked register in this state.
4123 * Since this state is already marked, just return.
4124 * Nothing to be tracked further in the parent state.
4128 i = get_prev_insn_idx(st, i, &history);
4131 if (i >= env->prog->len) {
4132 /* This can happen if backtracking reached insn 0
4133 * and there are still reg_mask or stack_mask
4135 * It means the backtracking missed the spot where
4136 * particular register was initialized with a constant.
4138 verbose(env, "BUG backtracking idx %d\n", i);
4139 WARN_ONCE(1, "verifier backtracking bug");
4147 for (fr = bt->frame; fr >= 0; fr--) {
4148 func = st->frame[fr];
4149 bitmap_from_u64(mask, bt_frame_reg_mask(bt, fr));
4150 for_each_set_bit(i, mask, 32) {
4151 reg = &func->regs[i];
4152 if (reg->type != SCALAR_VALUE) {
4153 bt_clear_frame_reg(bt, fr, i);
4157 bt_clear_frame_reg(bt, fr, i);
4159 reg->precise = true;
4162 bitmap_from_u64(mask, bt_frame_stack_mask(bt, fr));
4163 for_each_set_bit(i, mask, 64) {
4164 if (i >= func->allocated_stack / BPF_REG_SIZE) {
4165 /* the sequence of instructions:
4167 * 3: (7b) *(u64 *)(r3 -8) = r0
4168 * 4: (79) r4 = *(u64 *)(r10 -8)
4169 * doesn't contain jmps. It's backtracked
4170 * as a single block.
4171 * During backtracking insn 3 is not recognized as
4172 * stack access, so at the end of backtracking
4173 * stack slot fp-8 is still marked in stack_mask.
4174 * However the parent state may not have accessed
4175 * fp-8 and it's "unallocated" stack space.
4176 * In such case fallback to conservative.
4178 mark_all_scalars_precise(env, env->cur_state);
4183 if (!is_spilled_scalar_reg(&func->stack[i])) {
4184 bt_clear_frame_slot(bt, fr, i);
4187 reg = &func->stack[i].spilled_ptr;
4189 bt_clear_frame_slot(bt, fr, i);
4191 reg->precise = true;
4193 if (env->log.level & BPF_LOG_LEVEL2) {
4194 fmt_reg_mask(env->tmp_str_buf, TMP_STR_BUF_LEN,
4195 bt_frame_reg_mask(bt, fr));
4196 verbose(env, "mark_precise: frame%d: parent state regs=%s ",
4197 fr, env->tmp_str_buf);
4198 fmt_stack_mask(env->tmp_str_buf, TMP_STR_BUF_LEN,
4199 bt_frame_stack_mask(bt, fr));
4200 verbose(env, "stack=%s: ", env->tmp_str_buf);
4201 print_verifier_state(env, func, true);
4208 subseq_idx = first_idx;
4209 last_idx = st->last_insn_idx;
4210 first_idx = st->first_insn_idx;
4213 /* if we still have requested precise regs or slots, we missed
4214 * something (e.g., stack access through non-r10 register), so
4215 * fallback to marking all precise
4217 if (!bt_empty(bt)) {
4218 mark_all_scalars_precise(env, env->cur_state);
4225 int mark_chain_precision(struct bpf_verifier_env *env, int regno)
4227 return __mark_chain_precision(env, regno);
4230 /* mark_chain_precision_batch() assumes that env->bt is set in the caller to
4231 * desired reg and stack masks across all relevant frames
4233 static int mark_chain_precision_batch(struct bpf_verifier_env *env)
4235 return __mark_chain_precision(env, -1);
4238 static bool is_spillable_regtype(enum bpf_reg_type type)
4240 switch (base_type(type)) {
4241 case PTR_TO_MAP_VALUE:
4245 case PTR_TO_PACKET_META:
4246 case PTR_TO_PACKET_END:
4247 case PTR_TO_FLOW_KEYS:
4248 case CONST_PTR_TO_MAP:
4250 case PTR_TO_SOCK_COMMON:
4251 case PTR_TO_TCP_SOCK:
4252 case PTR_TO_XDP_SOCK:
4257 case PTR_TO_MAP_KEY:
4264 /* Does this register contain a constant zero? */
4265 static bool register_is_null(struct bpf_reg_state *reg)
4267 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
4270 static bool register_is_const(struct bpf_reg_state *reg)
4272 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
4275 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
4277 return tnum_is_unknown(reg->var_off) &&
4278 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
4279 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
4280 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
4281 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
4284 static bool register_is_bounded(struct bpf_reg_state *reg)
4286 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
4289 static bool __is_pointer_value(bool allow_ptr_leaks,
4290 const struct bpf_reg_state *reg)
4292 if (allow_ptr_leaks)
4295 return reg->type != SCALAR_VALUE;
4298 /* Copy src state preserving dst->parent and dst->live fields */
4299 static void copy_register_state(struct bpf_reg_state *dst, const struct bpf_reg_state *src)
4301 struct bpf_reg_state *parent = dst->parent;
4302 enum bpf_reg_liveness live = dst->live;
4305 dst->parent = parent;
4309 static void save_register_state(struct bpf_func_state *state,
4310 int spi, struct bpf_reg_state *reg,
4315 copy_register_state(&state->stack[spi].spilled_ptr, reg);
4316 if (size == BPF_REG_SIZE)
4317 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
4319 for (i = BPF_REG_SIZE; i > BPF_REG_SIZE - size; i--)
4320 state->stack[spi].slot_type[i - 1] = STACK_SPILL;
4322 /* size < 8 bytes spill */
4324 scrub_spilled_slot(&state->stack[spi].slot_type[i - 1]);
4327 static bool is_bpf_st_mem(struct bpf_insn *insn)
4329 return BPF_CLASS(insn->code) == BPF_ST && BPF_MODE(insn->code) == BPF_MEM;
4332 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
4333 * stack boundary and alignment are checked in check_mem_access()
4335 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
4336 /* stack frame we're writing to */
4337 struct bpf_func_state *state,
4338 int off, int size, int value_regno,
4341 struct bpf_func_state *cur; /* state of the current function */
4342 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
4343 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
4344 struct bpf_reg_state *reg = NULL;
4345 u32 dst_reg = insn->dst_reg;
4347 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
4348 * so it's aligned access and [off, off + size) are within stack limits
4350 if (!env->allow_ptr_leaks &&
4351 is_spilled_reg(&state->stack[spi]) &&
4352 size != BPF_REG_SIZE) {
4353 verbose(env, "attempt to corrupt spilled pointer on stack\n");
4357 cur = env->cur_state->frame[env->cur_state->curframe];
4358 if (value_regno >= 0)
4359 reg = &cur->regs[value_regno];
4360 if (!env->bypass_spec_v4) {
4361 bool sanitize = reg && is_spillable_regtype(reg->type);
4363 for (i = 0; i < size; i++) {
4364 u8 type = state->stack[spi].slot_type[i];
4366 if (type != STACK_MISC && type != STACK_ZERO) {
4373 env->insn_aux_data[insn_idx].sanitize_stack_spill = true;
4376 err = destroy_if_dynptr_stack_slot(env, state, spi);
4380 mark_stack_slot_scratched(env, spi);
4381 if (reg && !(off % BPF_REG_SIZE) && register_is_bounded(reg) &&
4382 !register_is_null(reg) && env->bpf_capable) {
4383 if (dst_reg != BPF_REG_FP) {
4384 /* The backtracking logic can only recognize explicit
4385 * stack slot address like [fp - 8]. Other spill of
4386 * scalar via different register has to be conservative.
4387 * Backtrack from here and mark all registers as precise
4388 * that contributed into 'reg' being a constant.
4390 err = mark_chain_precision(env, value_regno);
4394 save_register_state(state, spi, reg, size);
4395 /* Break the relation on a narrowing spill. */
4396 if (fls64(reg->umax_value) > BITS_PER_BYTE * size)
4397 state->stack[spi].spilled_ptr.id = 0;
4398 } else if (!reg && !(off % BPF_REG_SIZE) && is_bpf_st_mem(insn) &&
4399 insn->imm != 0 && env->bpf_capable) {
4400 struct bpf_reg_state fake_reg = {};
4402 __mark_reg_known(&fake_reg, insn->imm);
4403 fake_reg.type = SCALAR_VALUE;
4404 save_register_state(state, spi, &fake_reg, size);
4405 } else if (reg && is_spillable_regtype(reg->type)) {
4406 /* register containing pointer is being spilled into stack */
4407 if (size != BPF_REG_SIZE) {
4408 verbose_linfo(env, insn_idx, "; ");
4409 verbose(env, "invalid size of register spill\n");
4412 if (state != cur && reg->type == PTR_TO_STACK) {
4413 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
4416 save_register_state(state, spi, reg, size);
4418 u8 type = STACK_MISC;
4420 /* regular write of data into stack destroys any spilled ptr */
4421 state->stack[spi].spilled_ptr.type = NOT_INIT;
4422 /* Mark slots as STACK_MISC if they belonged to spilled ptr/dynptr/iter. */
4423 if (is_stack_slot_special(&state->stack[spi]))
4424 for (i = 0; i < BPF_REG_SIZE; i++)
4425 scrub_spilled_slot(&state->stack[spi].slot_type[i]);
4427 /* only mark the slot as written if all 8 bytes were written
4428 * otherwise read propagation may incorrectly stop too soon
4429 * when stack slots are partially written.
4430 * This heuristic means that read propagation will be
4431 * conservative, since it will add reg_live_read marks
4432 * to stack slots all the way to first state when programs
4433 * writes+reads less than 8 bytes
4435 if (size == BPF_REG_SIZE)
4436 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
4438 /* when we zero initialize stack slots mark them as such */
4439 if ((reg && register_is_null(reg)) ||
4440 (!reg && is_bpf_st_mem(insn) && insn->imm == 0)) {
4441 /* backtracking doesn't work for STACK_ZERO yet. */
4442 err = mark_chain_precision(env, value_regno);
4448 /* Mark slots affected by this stack write. */
4449 for (i = 0; i < size; i++)
4450 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
4456 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
4457 * known to contain a variable offset.
4458 * This function checks whether the write is permitted and conservatively
4459 * tracks the effects of the write, considering that each stack slot in the
4460 * dynamic range is potentially written to.
4462 * 'off' includes 'regno->off'.
4463 * 'value_regno' can be -1, meaning that an unknown value is being written to
4466 * Spilled pointers in range are not marked as written because we don't know
4467 * what's going to be actually written. This means that read propagation for
4468 * future reads cannot be terminated by this write.
4470 * For privileged programs, uninitialized stack slots are considered
4471 * initialized by this write (even though we don't know exactly what offsets
4472 * are going to be written to). The idea is that we don't want the verifier to
4473 * reject future reads that access slots written to through variable offsets.
4475 static int check_stack_write_var_off(struct bpf_verifier_env *env,
4476 /* func where register points to */
4477 struct bpf_func_state *state,
4478 int ptr_regno, int off, int size,
4479 int value_regno, int insn_idx)
4481 struct bpf_func_state *cur; /* state of the current function */
4482 int min_off, max_off;
4484 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
4485 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
4486 bool writing_zero = false;
4487 /* set if the fact that we're writing a zero is used to let any
4488 * stack slots remain STACK_ZERO
4490 bool zero_used = false;
4492 cur = env->cur_state->frame[env->cur_state->curframe];
4493 ptr_reg = &cur->regs[ptr_regno];
4494 min_off = ptr_reg->smin_value + off;
4495 max_off = ptr_reg->smax_value + off + size;
4496 if (value_regno >= 0)
4497 value_reg = &cur->regs[value_regno];
4498 if ((value_reg && register_is_null(value_reg)) ||
4499 (!value_reg && is_bpf_st_mem(insn) && insn->imm == 0))
4500 writing_zero = true;
4502 for (i = min_off; i < max_off; i++) {
4506 err = destroy_if_dynptr_stack_slot(env, state, spi);
4511 /* Variable offset writes destroy any spilled pointers in range. */
4512 for (i = min_off; i < max_off; i++) {
4513 u8 new_type, *stype;
4517 spi = slot / BPF_REG_SIZE;
4518 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
4519 mark_stack_slot_scratched(env, spi);
4521 if (!env->allow_ptr_leaks && *stype != STACK_MISC && *stype != STACK_ZERO) {
4522 /* Reject the write if range we may write to has not
4523 * been initialized beforehand. If we didn't reject
4524 * here, the ptr status would be erased below (even
4525 * though not all slots are actually overwritten),
4526 * possibly opening the door to leaks.
4528 * We do however catch STACK_INVALID case below, and
4529 * only allow reading possibly uninitialized memory
4530 * later for CAP_PERFMON, as the write may not happen to
4533 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
4538 /* Erase all spilled pointers. */
4539 state->stack[spi].spilled_ptr.type = NOT_INIT;
4541 /* Update the slot type. */
4542 new_type = STACK_MISC;
4543 if (writing_zero && *stype == STACK_ZERO) {
4544 new_type = STACK_ZERO;
4547 /* If the slot is STACK_INVALID, we check whether it's OK to
4548 * pretend that it will be initialized by this write. The slot
4549 * might not actually be written to, and so if we mark it as
4550 * initialized future reads might leak uninitialized memory.
4551 * For privileged programs, we will accept such reads to slots
4552 * that may or may not be written because, if we're reject
4553 * them, the error would be too confusing.
4555 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
4556 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
4563 /* backtracking doesn't work for STACK_ZERO yet. */
4564 err = mark_chain_precision(env, value_regno);
4571 /* When register 'dst_regno' is assigned some values from stack[min_off,
4572 * max_off), we set the register's type according to the types of the
4573 * respective stack slots. If all the stack values are known to be zeros, then
4574 * so is the destination reg. Otherwise, the register is considered to be
4575 * SCALAR. This function does not deal with register filling; the caller must
4576 * ensure that all spilled registers in the stack range have been marked as
4579 static void mark_reg_stack_read(struct bpf_verifier_env *env,
4580 /* func where src register points to */
4581 struct bpf_func_state *ptr_state,
4582 int min_off, int max_off, int dst_regno)
4584 struct bpf_verifier_state *vstate = env->cur_state;
4585 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4590 for (i = min_off; i < max_off; i++) {
4592 spi = slot / BPF_REG_SIZE;
4593 mark_stack_slot_scratched(env, spi);
4594 stype = ptr_state->stack[spi].slot_type;
4595 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
4599 if (zeros == max_off - min_off) {
4600 /* any access_size read into register is zero extended,
4601 * so the whole register == const_zero
4603 __mark_reg_const_zero(&state->regs[dst_regno]);
4604 /* backtracking doesn't support STACK_ZERO yet,
4605 * so mark it precise here, so that later
4606 * backtracking can stop here.
4607 * Backtracking may not need this if this register
4608 * doesn't participate in pointer adjustment.
4609 * Forward propagation of precise flag is not
4610 * necessary either. This mark is only to stop
4611 * backtracking. Any register that contributed
4612 * to const 0 was marked precise before spill.
4614 state->regs[dst_regno].precise = true;
4616 /* have read misc data from the stack */
4617 mark_reg_unknown(env, state->regs, dst_regno);
4619 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4622 /* Read the stack at 'off' and put the results into the register indicated by
4623 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
4626 * 'dst_regno' can be -1, meaning that the read value is not going to a
4629 * The access is assumed to be within the current stack bounds.
4631 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
4632 /* func where src register points to */
4633 struct bpf_func_state *reg_state,
4634 int off, int size, int dst_regno)
4636 struct bpf_verifier_state *vstate = env->cur_state;
4637 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4638 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
4639 struct bpf_reg_state *reg;
4642 stype = reg_state->stack[spi].slot_type;
4643 reg = ®_state->stack[spi].spilled_ptr;
4645 mark_stack_slot_scratched(env, spi);
4647 if (is_spilled_reg(®_state->stack[spi])) {
4650 for (i = BPF_REG_SIZE - 1; i > 0 && stype[i - 1] == STACK_SPILL; i--)
4653 if (size != BPF_REG_SIZE || spill_size != BPF_REG_SIZE) {
4654 if (reg->type != SCALAR_VALUE) {
4655 verbose_linfo(env, env->insn_idx, "; ");
4656 verbose(env, "invalid size of register fill\n");
4660 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4664 if (!(off % BPF_REG_SIZE) && size == spill_size) {
4665 /* The earlier check_reg_arg() has decided the
4666 * subreg_def for this insn. Save it first.
4668 s32 subreg_def = state->regs[dst_regno].subreg_def;
4670 copy_register_state(&state->regs[dst_regno], reg);
4671 state->regs[dst_regno].subreg_def = subreg_def;
4673 for (i = 0; i < size; i++) {
4674 type = stype[(slot - i) % BPF_REG_SIZE];
4675 if (type == STACK_SPILL)
4677 if (type == STACK_MISC)
4679 if (type == STACK_INVALID && env->allow_uninit_stack)
4681 verbose(env, "invalid read from stack off %d+%d size %d\n",
4685 mark_reg_unknown(env, state->regs, dst_regno);
4687 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4691 if (dst_regno >= 0) {
4692 /* restore register state from stack */
4693 copy_register_state(&state->regs[dst_regno], reg);
4694 /* mark reg as written since spilled pointer state likely
4695 * has its liveness marks cleared by is_state_visited()
4696 * which resets stack/reg liveness for state transitions
4698 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4699 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
4700 /* If dst_regno==-1, the caller is asking us whether
4701 * it is acceptable to use this value as a SCALAR_VALUE
4703 * We must not allow unprivileged callers to do that
4704 * with spilled pointers.
4706 verbose(env, "leaking pointer from stack off %d\n",
4710 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4712 for (i = 0; i < size; i++) {
4713 type = stype[(slot - i) % BPF_REG_SIZE];
4714 if (type == STACK_MISC)
4716 if (type == STACK_ZERO)
4718 if (type == STACK_INVALID && env->allow_uninit_stack)
4720 verbose(env, "invalid read from stack off %d+%d size %d\n",
4724 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4726 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
4731 enum bpf_access_src {
4732 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
4733 ACCESS_HELPER = 2, /* the access is performed by a helper */
4736 static int check_stack_range_initialized(struct bpf_verifier_env *env,
4737 int regno, int off, int access_size,
4738 bool zero_size_allowed,
4739 enum bpf_access_src type,
4740 struct bpf_call_arg_meta *meta);
4742 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
4744 return cur_regs(env) + regno;
4747 /* Read the stack at 'ptr_regno + off' and put the result into the register
4749 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
4750 * but not its variable offset.
4751 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
4753 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
4754 * filling registers (i.e. reads of spilled register cannot be detected when
4755 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
4756 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
4757 * offset; for a fixed offset check_stack_read_fixed_off should be used
4760 static int check_stack_read_var_off(struct bpf_verifier_env *env,
4761 int ptr_regno, int off, int size, int dst_regno)
4763 /* The state of the source register. */
4764 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
4765 struct bpf_func_state *ptr_state = func(env, reg);
4767 int min_off, max_off;
4769 /* Note that we pass a NULL meta, so raw access will not be permitted.
4771 err = check_stack_range_initialized(env, ptr_regno, off, size,
4772 false, ACCESS_DIRECT, NULL);
4776 min_off = reg->smin_value + off;
4777 max_off = reg->smax_value + off;
4778 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
4782 /* check_stack_read dispatches to check_stack_read_fixed_off or
4783 * check_stack_read_var_off.
4785 * The caller must ensure that the offset falls within the allocated stack
4788 * 'dst_regno' is a register which will receive the value from the stack. It
4789 * can be -1, meaning that the read value is not going to a register.
4791 static int check_stack_read(struct bpf_verifier_env *env,
4792 int ptr_regno, int off, int size,
4795 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
4796 struct bpf_func_state *state = func(env, reg);
4798 /* Some accesses are only permitted with a static offset. */
4799 bool var_off = !tnum_is_const(reg->var_off);
4801 /* The offset is required to be static when reads don't go to a
4802 * register, in order to not leak pointers (see
4803 * check_stack_read_fixed_off).
4805 if (dst_regno < 0 && var_off) {
4808 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4809 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
4813 /* Variable offset is prohibited for unprivileged mode for simplicity
4814 * since it requires corresponding support in Spectre masking for stack
4815 * ALU. See also retrieve_ptr_limit(). The check in
4816 * check_stack_access_for_ptr_arithmetic() called by
4817 * adjust_ptr_min_max_vals() prevents users from creating stack pointers
4818 * with variable offsets, therefore no check is required here. Further,
4819 * just checking it here would be insufficient as speculative stack
4820 * writes could still lead to unsafe speculative behaviour.
4823 off += reg->var_off.value;
4824 err = check_stack_read_fixed_off(env, state, off, size,
4827 /* Variable offset stack reads need more conservative handling
4828 * than fixed offset ones. Note that dst_regno >= 0 on this
4831 err = check_stack_read_var_off(env, ptr_regno, off, size,
4838 /* check_stack_write dispatches to check_stack_write_fixed_off or
4839 * check_stack_write_var_off.
4841 * 'ptr_regno' is the register used as a pointer into the stack.
4842 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
4843 * 'value_regno' is the register whose value we're writing to the stack. It can
4844 * be -1, meaning that we're not writing from a register.
4846 * The caller must ensure that the offset falls within the maximum stack size.
4848 static int check_stack_write(struct bpf_verifier_env *env,
4849 int ptr_regno, int off, int size,
4850 int value_regno, int insn_idx)
4852 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
4853 struct bpf_func_state *state = func(env, reg);
4856 if (tnum_is_const(reg->var_off)) {
4857 off += reg->var_off.value;
4858 err = check_stack_write_fixed_off(env, state, off, size,
4859 value_regno, insn_idx);
4861 /* Variable offset stack reads need more conservative handling
4862 * than fixed offset ones.
4864 err = check_stack_write_var_off(env, state,
4865 ptr_regno, off, size,
4866 value_regno, insn_idx);
4871 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
4872 int off, int size, enum bpf_access_type type)
4874 struct bpf_reg_state *regs = cur_regs(env);
4875 struct bpf_map *map = regs[regno].map_ptr;
4876 u32 cap = bpf_map_flags_to_cap(map);
4878 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
4879 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
4880 map->value_size, off, size);
4884 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
4885 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
4886 map->value_size, off, size);
4893 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
4894 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
4895 int off, int size, u32 mem_size,
4896 bool zero_size_allowed)
4898 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
4899 struct bpf_reg_state *reg;
4901 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
4904 reg = &cur_regs(env)[regno];
4905 switch (reg->type) {
4906 case PTR_TO_MAP_KEY:
4907 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
4908 mem_size, off, size);
4910 case PTR_TO_MAP_VALUE:
4911 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
4912 mem_size, off, size);
4915 case PTR_TO_PACKET_META:
4916 case PTR_TO_PACKET_END:
4917 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
4918 off, size, regno, reg->id, off, mem_size);
4922 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
4923 mem_size, off, size);
4929 /* check read/write into a memory region with possible variable offset */
4930 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
4931 int off, int size, u32 mem_size,
4932 bool zero_size_allowed)
4934 struct bpf_verifier_state *vstate = env->cur_state;
4935 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4936 struct bpf_reg_state *reg = &state->regs[regno];
4939 /* We may have adjusted the register pointing to memory region, so we
4940 * need to try adding each of min_value and max_value to off
4941 * to make sure our theoretical access will be safe.
4943 * The minimum value is only important with signed
4944 * comparisons where we can't assume the floor of a
4945 * value is 0. If we are using signed variables for our
4946 * index'es we need to make sure that whatever we use
4947 * will have a set floor within our range.
4949 if (reg->smin_value < 0 &&
4950 (reg->smin_value == S64_MIN ||
4951 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
4952 reg->smin_value + off < 0)) {
4953 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
4957 err = __check_mem_access(env, regno, reg->smin_value + off, size,
4958 mem_size, zero_size_allowed);
4960 verbose(env, "R%d min value is outside of the allowed memory range\n",
4965 /* If we haven't set a max value then we need to bail since we can't be
4966 * sure we won't do bad things.
4967 * If reg->umax_value + off could overflow, treat that as unbounded too.
4969 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
4970 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
4974 err = __check_mem_access(env, regno, reg->umax_value + off, size,
4975 mem_size, zero_size_allowed);
4977 verbose(env, "R%d max value is outside of the allowed memory range\n",
4985 static int __check_ptr_off_reg(struct bpf_verifier_env *env,
4986 const struct bpf_reg_state *reg, int regno,
4989 /* Access to this pointer-typed register or passing it to a helper
4990 * is only allowed in its original, unmodified form.
4994 verbose(env, "negative offset %s ptr R%d off=%d disallowed\n",
4995 reg_type_str(env, reg->type), regno, reg->off);
4999 if (!fixed_off_ok && reg->off) {
5000 verbose(env, "dereference of modified %s ptr R%d off=%d disallowed\n",
5001 reg_type_str(env, reg->type), regno, reg->off);
5005 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
5008 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5009 verbose(env, "variable %s access var_off=%s disallowed\n",
5010 reg_type_str(env, reg->type), tn_buf);
5017 int check_ptr_off_reg(struct bpf_verifier_env *env,
5018 const struct bpf_reg_state *reg, int regno)
5020 return __check_ptr_off_reg(env, reg, regno, false);
5023 static int map_kptr_match_type(struct bpf_verifier_env *env,
5024 struct btf_field *kptr_field,
5025 struct bpf_reg_state *reg, u32 regno)
5027 const char *targ_name = btf_type_name(kptr_field->kptr.btf, kptr_field->kptr.btf_id);
5029 const char *reg_name = "";
5031 if (btf_is_kernel(reg->btf)) {
5032 perm_flags = PTR_MAYBE_NULL | PTR_TRUSTED | MEM_RCU;
5034 /* Only unreferenced case accepts untrusted pointers */
5035 if (kptr_field->type == BPF_KPTR_UNREF)
5036 perm_flags |= PTR_UNTRUSTED;
5038 perm_flags = PTR_MAYBE_NULL | MEM_ALLOC;
5041 if (base_type(reg->type) != PTR_TO_BTF_ID || (type_flag(reg->type) & ~perm_flags))
5044 /* We need to verify reg->type and reg->btf, before accessing reg->btf */
5045 reg_name = btf_type_name(reg->btf, reg->btf_id);
5047 /* For ref_ptr case, release function check should ensure we get one
5048 * referenced PTR_TO_BTF_ID, and that its fixed offset is 0. For the
5049 * normal store of unreferenced kptr, we must ensure var_off is zero.
5050 * Since ref_ptr cannot be accessed directly by BPF insns, checks for
5051 * reg->off and reg->ref_obj_id are not needed here.
5053 if (__check_ptr_off_reg(env, reg, regno, true))
5056 /* A full type match is needed, as BTF can be vmlinux, module or prog BTF, and
5057 * we also need to take into account the reg->off.
5059 * We want to support cases like:
5067 * v = func(); // PTR_TO_BTF_ID
5068 * val->foo = v; // reg->off is zero, btf and btf_id match type
5069 * val->bar = &v->br; // reg->off is still zero, but we need to retry with
5070 * // first member type of struct after comparison fails
5071 * val->baz = &v->bz; // reg->off is non-zero, so struct needs to be walked
5074 * In the kptr_ref case, check_func_arg_reg_off already ensures reg->off
5075 * is zero. We must also ensure that btf_struct_ids_match does not walk
5076 * the struct to match type against first member of struct, i.e. reject
5077 * second case from above. Hence, when type is BPF_KPTR_REF, we set
5078 * strict mode to true for type match.
5080 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
5081 kptr_field->kptr.btf, kptr_field->kptr.btf_id,
5082 kptr_field->type == BPF_KPTR_REF))
5086 verbose(env, "invalid kptr access, R%d type=%s%s ", regno,
5087 reg_type_str(env, reg->type), reg_name);
5088 verbose(env, "expected=%s%s", reg_type_str(env, PTR_TO_BTF_ID), targ_name);
5089 if (kptr_field->type == BPF_KPTR_UNREF)
5090 verbose(env, " or %s%s\n", reg_type_str(env, PTR_TO_BTF_ID | PTR_UNTRUSTED),
5097 /* The non-sleepable programs and sleepable programs with explicit bpf_rcu_read_lock()
5098 * can dereference RCU protected pointers and result is PTR_TRUSTED.
5100 static bool in_rcu_cs(struct bpf_verifier_env *env)
5102 return env->cur_state->active_rcu_lock ||
5103 env->cur_state->active_lock.ptr ||
5104 !env->prog->aux->sleepable;
5107 /* Once GCC supports btf_type_tag the following mechanism will be replaced with tag check */
5108 BTF_SET_START(rcu_protected_types)
5109 BTF_ID(struct, prog_test_ref_kfunc)
5110 BTF_ID(struct, cgroup)
5111 BTF_ID(struct, bpf_cpumask)
5112 BTF_ID(struct, task_struct)
5113 BTF_SET_END(rcu_protected_types)
5115 static bool rcu_protected_object(const struct btf *btf, u32 btf_id)
5117 if (!btf_is_kernel(btf))
5119 return btf_id_set_contains(&rcu_protected_types, btf_id);
5122 static bool rcu_safe_kptr(const struct btf_field *field)
5124 const struct btf_field_kptr *kptr = &field->kptr;
5126 return field->type == BPF_KPTR_REF && rcu_protected_object(kptr->btf, kptr->btf_id);
5129 static int check_map_kptr_access(struct bpf_verifier_env *env, u32 regno,
5130 int value_regno, int insn_idx,
5131 struct btf_field *kptr_field)
5133 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
5134 int class = BPF_CLASS(insn->code);
5135 struct bpf_reg_state *val_reg;
5137 /* Things we already checked for in check_map_access and caller:
5138 * - Reject cases where variable offset may touch kptr
5139 * - size of access (must be BPF_DW)
5140 * - tnum_is_const(reg->var_off)
5141 * - kptr_field->offset == off + reg->var_off.value
5143 /* Only BPF_[LDX,STX,ST] | BPF_MEM | BPF_DW is supported */
5144 if (BPF_MODE(insn->code) != BPF_MEM) {
5145 verbose(env, "kptr in map can only be accessed using BPF_MEM instruction mode\n");
5149 /* We only allow loading referenced kptr, since it will be marked as
5150 * untrusted, similar to unreferenced kptr.
5152 if (class != BPF_LDX && kptr_field->type == BPF_KPTR_REF) {
5153 verbose(env, "store to referenced kptr disallowed\n");
5157 if (class == BPF_LDX) {
5158 val_reg = reg_state(env, value_regno);
5159 /* We can simply mark the value_regno receiving the pointer
5160 * value from map as PTR_TO_BTF_ID, with the correct type.
5162 mark_btf_ld_reg(env, cur_regs(env), value_regno, PTR_TO_BTF_ID, kptr_field->kptr.btf,
5163 kptr_field->kptr.btf_id,
5164 rcu_safe_kptr(kptr_field) && in_rcu_cs(env) ?
5165 PTR_MAYBE_NULL | MEM_RCU :
5166 PTR_MAYBE_NULL | PTR_UNTRUSTED);
5167 /* For mark_ptr_or_null_reg */
5168 val_reg->id = ++env->id_gen;
5169 } else if (class == BPF_STX) {
5170 val_reg = reg_state(env, value_regno);
5171 if (!register_is_null(val_reg) &&
5172 map_kptr_match_type(env, kptr_field, val_reg, value_regno))
5174 } else if (class == BPF_ST) {
5176 verbose(env, "BPF_ST imm must be 0 when storing to kptr at off=%u\n",
5177 kptr_field->offset);
5181 verbose(env, "kptr in map can only be accessed using BPF_LDX/BPF_STX/BPF_ST\n");
5187 /* check read/write into a map element with possible variable offset */
5188 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
5189 int off, int size, bool zero_size_allowed,
5190 enum bpf_access_src src)
5192 struct bpf_verifier_state *vstate = env->cur_state;
5193 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5194 struct bpf_reg_state *reg = &state->regs[regno];
5195 struct bpf_map *map = reg->map_ptr;
5196 struct btf_record *rec;
5199 err = check_mem_region_access(env, regno, off, size, map->value_size,
5204 if (IS_ERR_OR_NULL(map->record))
5207 for (i = 0; i < rec->cnt; i++) {
5208 struct btf_field *field = &rec->fields[i];
5209 u32 p = field->offset;
5211 /* If any part of a field can be touched by load/store, reject
5212 * this program. To check that [x1, x2) overlaps with [y1, y2),
5213 * it is sufficient to check x1 < y2 && y1 < x2.
5215 if (reg->smin_value + off < p + btf_field_type_size(field->type) &&
5216 p < reg->umax_value + off + size) {
5217 switch (field->type) {
5218 case BPF_KPTR_UNREF:
5220 if (src != ACCESS_DIRECT) {
5221 verbose(env, "kptr cannot be accessed indirectly by helper\n");
5224 if (!tnum_is_const(reg->var_off)) {
5225 verbose(env, "kptr access cannot have variable offset\n");
5228 if (p != off + reg->var_off.value) {
5229 verbose(env, "kptr access misaligned expected=%u off=%llu\n",
5230 p, off + reg->var_off.value);
5233 if (size != bpf_size_to_bytes(BPF_DW)) {
5234 verbose(env, "kptr access size must be BPF_DW\n");
5239 verbose(env, "%s cannot be accessed directly by load/store\n",
5240 btf_field_type_name(field->type));
5248 #define MAX_PACKET_OFF 0xffff
5250 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
5251 const struct bpf_call_arg_meta *meta,
5252 enum bpf_access_type t)
5254 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
5256 switch (prog_type) {
5257 /* Program types only with direct read access go here! */
5258 case BPF_PROG_TYPE_LWT_IN:
5259 case BPF_PROG_TYPE_LWT_OUT:
5260 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
5261 case BPF_PROG_TYPE_SK_REUSEPORT:
5262 case BPF_PROG_TYPE_FLOW_DISSECTOR:
5263 case BPF_PROG_TYPE_CGROUP_SKB:
5268 /* Program types with direct read + write access go here! */
5269 case BPF_PROG_TYPE_SCHED_CLS:
5270 case BPF_PROG_TYPE_SCHED_ACT:
5271 case BPF_PROG_TYPE_XDP:
5272 case BPF_PROG_TYPE_LWT_XMIT:
5273 case BPF_PROG_TYPE_SK_SKB:
5274 case BPF_PROG_TYPE_SK_MSG:
5276 return meta->pkt_access;
5278 env->seen_direct_write = true;
5281 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
5283 env->seen_direct_write = true;
5292 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
5293 int size, bool zero_size_allowed)
5295 struct bpf_reg_state *regs = cur_regs(env);
5296 struct bpf_reg_state *reg = ®s[regno];
5299 /* We may have added a variable offset to the packet pointer; but any
5300 * reg->range we have comes after that. We are only checking the fixed
5304 /* We don't allow negative numbers, because we aren't tracking enough
5305 * detail to prove they're safe.
5307 if (reg->smin_value < 0) {
5308 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
5313 err = reg->range < 0 ? -EINVAL :
5314 __check_mem_access(env, regno, off, size, reg->range,
5317 verbose(env, "R%d offset is outside of the packet\n", regno);
5321 /* __check_mem_access has made sure "off + size - 1" is within u16.
5322 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
5323 * otherwise find_good_pkt_pointers would have refused to set range info
5324 * that __check_mem_access would have rejected this pkt access.
5325 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
5327 env->prog->aux->max_pkt_offset =
5328 max_t(u32, env->prog->aux->max_pkt_offset,
5329 off + reg->umax_value + size - 1);
5334 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
5335 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
5336 enum bpf_access_type t, enum bpf_reg_type *reg_type,
5337 struct btf **btf, u32 *btf_id)
5339 struct bpf_insn_access_aux info = {
5340 .reg_type = *reg_type,
5344 if (env->ops->is_valid_access &&
5345 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
5346 /* A non zero info.ctx_field_size indicates that this field is a
5347 * candidate for later verifier transformation to load the whole
5348 * field and then apply a mask when accessed with a narrower
5349 * access than actual ctx access size. A zero info.ctx_field_size
5350 * will only allow for whole field access and rejects any other
5351 * type of narrower access.
5353 *reg_type = info.reg_type;
5355 if (base_type(*reg_type) == PTR_TO_BTF_ID) {
5357 *btf_id = info.btf_id;
5359 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
5361 /* remember the offset of last byte accessed in ctx */
5362 if (env->prog->aux->max_ctx_offset < off + size)
5363 env->prog->aux->max_ctx_offset = off + size;
5367 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
5371 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
5374 if (size < 0 || off < 0 ||
5375 (u64)off + size > sizeof(struct bpf_flow_keys)) {
5376 verbose(env, "invalid access to flow keys off=%d size=%d\n",
5383 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
5384 u32 regno, int off, int size,
5385 enum bpf_access_type t)
5387 struct bpf_reg_state *regs = cur_regs(env);
5388 struct bpf_reg_state *reg = ®s[regno];
5389 struct bpf_insn_access_aux info = {};
5392 if (reg->smin_value < 0) {
5393 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
5398 switch (reg->type) {
5399 case PTR_TO_SOCK_COMMON:
5400 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
5403 valid = bpf_sock_is_valid_access(off, size, t, &info);
5405 case PTR_TO_TCP_SOCK:
5406 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
5408 case PTR_TO_XDP_SOCK:
5409 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
5417 env->insn_aux_data[insn_idx].ctx_field_size =
5418 info.ctx_field_size;
5422 verbose(env, "R%d invalid %s access off=%d size=%d\n",
5423 regno, reg_type_str(env, reg->type), off, size);
5428 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
5430 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
5433 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
5435 const struct bpf_reg_state *reg = reg_state(env, regno);
5437 return reg->type == PTR_TO_CTX;
5440 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
5442 const struct bpf_reg_state *reg = reg_state(env, regno);
5444 return type_is_sk_pointer(reg->type);
5447 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
5449 const struct bpf_reg_state *reg = reg_state(env, regno);
5451 return type_is_pkt_pointer(reg->type);
5454 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
5456 const struct bpf_reg_state *reg = reg_state(env, regno);
5458 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
5459 return reg->type == PTR_TO_FLOW_KEYS;
5462 static u32 *reg2btf_ids[__BPF_REG_TYPE_MAX] = {
5464 [PTR_TO_SOCKET] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK],
5465 [PTR_TO_SOCK_COMMON] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
5466 [PTR_TO_TCP_SOCK] = &btf_sock_ids[BTF_SOCK_TYPE_TCP],
5468 [CONST_PTR_TO_MAP] = btf_bpf_map_id,
5471 static bool is_trusted_reg(const struct bpf_reg_state *reg)
5473 /* A referenced register is always trusted. */
5474 if (reg->ref_obj_id)
5477 /* Types listed in the reg2btf_ids are always trusted */
5478 if (reg2btf_ids[base_type(reg->type)])
5481 /* If a register is not referenced, it is trusted if it has the
5482 * MEM_ALLOC or PTR_TRUSTED type modifiers, and no others. Some of the
5483 * other type modifiers may be safe, but we elect to take an opt-in
5484 * approach here as some (e.g. PTR_UNTRUSTED and PTR_MAYBE_NULL) are
5487 * Eventually, we should make PTR_TRUSTED the single source of truth
5488 * for whether a register is trusted.
5490 return type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS &&
5491 !bpf_type_has_unsafe_modifiers(reg->type);
5494 static bool is_rcu_reg(const struct bpf_reg_state *reg)
5496 return reg->type & MEM_RCU;
5499 static void clear_trusted_flags(enum bpf_type_flag *flag)
5501 *flag &= ~(BPF_REG_TRUSTED_MODIFIERS | MEM_RCU);
5504 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
5505 const struct bpf_reg_state *reg,
5506 int off, int size, bool strict)
5508 struct tnum reg_off;
5511 /* Byte size accesses are always allowed. */
5512 if (!strict || size == 1)
5515 /* For platforms that do not have a Kconfig enabling
5516 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
5517 * NET_IP_ALIGN is universally set to '2'. And on platforms
5518 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
5519 * to this code only in strict mode where we want to emulate
5520 * the NET_IP_ALIGN==2 checking. Therefore use an
5521 * unconditional IP align value of '2'.
5525 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
5526 if (!tnum_is_aligned(reg_off, size)) {
5529 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5531 "misaligned packet access off %d+%s+%d+%d size %d\n",
5532 ip_align, tn_buf, reg->off, off, size);
5539 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
5540 const struct bpf_reg_state *reg,
5541 const char *pointer_desc,
5542 int off, int size, bool strict)
5544 struct tnum reg_off;
5546 /* Byte size accesses are always allowed. */
5547 if (!strict || size == 1)
5550 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
5551 if (!tnum_is_aligned(reg_off, size)) {
5554 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5555 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
5556 pointer_desc, tn_buf, reg->off, off, size);
5563 static int check_ptr_alignment(struct bpf_verifier_env *env,
5564 const struct bpf_reg_state *reg, int off,
5565 int size, bool strict_alignment_once)
5567 bool strict = env->strict_alignment || strict_alignment_once;
5568 const char *pointer_desc = "";
5570 switch (reg->type) {
5572 case PTR_TO_PACKET_META:
5573 /* Special case, because of NET_IP_ALIGN. Given metadata sits
5574 * right in front, treat it the very same way.
5576 return check_pkt_ptr_alignment(env, reg, off, size, strict);
5577 case PTR_TO_FLOW_KEYS:
5578 pointer_desc = "flow keys ";
5580 case PTR_TO_MAP_KEY:
5581 pointer_desc = "key ";
5583 case PTR_TO_MAP_VALUE:
5584 pointer_desc = "value ";
5587 pointer_desc = "context ";
5590 pointer_desc = "stack ";
5591 /* The stack spill tracking logic in check_stack_write_fixed_off()
5592 * and check_stack_read_fixed_off() relies on stack accesses being
5598 pointer_desc = "sock ";
5600 case PTR_TO_SOCK_COMMON:
5601 pointer_desc = "sock_common ";
5603 case PTR_TO_TCP_SOCK:
5604 pointer_desc = "tcp_sock ";
5606 case PTR_TO_XDP_SOCK:
5607 pointer_desc = "xdp_sock ";
5612 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
5616 /* starting from main bpf function walk all instructions of the function
5617 * and recursively walk all callees that given function can call.
5618 * Ignore jump and exit insns.
5619 * Since recursion is prevented by check_cfg() this algorithm
5620 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
5622 static int check_max_stack_depth_subprog(struct bpf_verifier_env *env, int idx)
5624 struct bpf_subprog_info *subprog = env->subprog_info;
5625 struct bpf_insn *insn = env->prog->insnsi;
5626 int depth = 0, frame = 0, i, subprog_end;
5627 bool tail_call_reachable = false;
5628 int ret_insn[MAX_CALL_FRAMES];
5629 int ret_prog[MAX_CALL_FRAMES];
5632 i = subprog[idx].start;
5634 /* protect against potential stack overflow that might happen when
5635 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
5636 * depth for such case down to 256 so that the worst case scenario
5637 * would result in 8k stack size (32 which is tailcall limit * 256 =
5640 * To get the idea what might happen, see an example:
5641 * func1 -> sub rsp, 128
5642 * subfunc1 -> sub rsp, 256
5643 * tailcall1 -> add rsp, 256
5644 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
5645 * subfunc2 -> sub rsp, 64
5646 * subfunc22 -> sub rsp, 128
5647 * tailcall2 -> add rsp, 128
5648 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
5650 * tailcall will unwind the current stack frame but it will not get rid
5651 * of caller's stack as shown on the example above.
5653 if (idx && subprog[idx].has_tail_call && depth >= 256) {
5655 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
5659 /* round up to 32-bytes, since this is granularity
5660 * of interpreter stack size
5662 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
5663 if (depth > MAX_BPF_STACK) {
5664 verbose(env, "combined stack size of %d calls is %d. Too large\n",
5669 subprog_end = subprog[idx + 1].start;
5670 for (; i < subprog_end; i++) {
5671 int next_insn, sidx;
5673 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
5675 /* remember insn and function to return to */
5676 ret_insn[frame] = i + 1;
5677 ret_prog[frame] = idx;
5679 /* find the callee */
5680 next_insn = i + insn[i].imm + 1;
5681 sidx = find_subprog(env, next_insn);
5683 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5687 if (subprog[sidx].is_async_cb) {
5688 if (subprog[sidx].has_tail_call) {
5689 verbose(env, "verifier bug. subprog has tail_call and async cb\n");
5692 /* async callbacks don't increase bpf prog stack size unless called directly */
5693 if (!bpf_pseudo_call(insn + i))
5699 if (subprog[idx].has_tail_call)
5700 tail_call_reachable = true;
5703 if (frame >= MAX_CALL_FRAMES) {
5704 verbose(env, "the call stack of %d frames is too deep !\n",
5710 /* if tail call got detected across bpf2bpf calls then mark each of the
5711 * currently present subprog frames as tail call reachable subprogs;
5712 * this info will be utilized by JIT so that we will be preserving the
5713 * tail call counter throughout bpf2bpf calls combined with tailcalls
5715 if (tail_call_reachable)
5716 for (j = 0; j < frame; j++)
5717 subprog[ret_prog[j]].tail_call_reachable = true;
5718 if (subprog[0].tail_call_reachable)
5719 env->prog->aux->tail_call_reachable = true;
5721 /* end of for() loop means the last insn of the 'subprog'
5722 * was reached. Doesn't matter whether it was JA or EXIT
5726 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
5728 i = ret_insn[frame];
5729 idx = ret_prog[frame];
5733 static int check_max_stack_depth(struct bpf_verifier_env *env)
5735 struct bpf_subprog_info *si = env->subprog_info;
5738 for (int i = 0; i < env->subprog_cnt; i++) {
5739 if (!i || si[i].is_async_cb) {
5740 ret = check_max_stack_depth_subprog(env, i);
5749 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
5750 static int get_callee_stack_depth(struct bpf_verifier_env *env,
5751 const struct bpf_insn *insn, int idx)
5753 int start = idx + insn->imm + 1, subprog;
5755 subprog = find_subprog(env, start);
5757 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5761 return env->subprog_info[subprog].stack_depth;
5765 static int __check_buffer_access(struct bpf_verifier_env *env,
5766 const char *buf_info,
5767 const struct bpf_reg_state *reg,
5768 int regno, int off, int size)
5772 "R%d invalid %s buffer access: off=%d, size=%d\n",
5773 regno, buf_info, off, size);
5776 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
5779 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5781 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
5782 regno, off, tn_buf);
5789 static int check_tp_buffer_access(struct bpf_verifier_env *env,
5790 const struct bpf_reg_state *reg,
5791 int regno, int off, int size)
5795 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
5799 if (off + size > env->prog->aux->max_tp_access)
5800 env->prog->aux->max_tp_access = off + size;
5805 static int check_buffer_access(struct bpf_verifier_env *env,
5806 const struct bpf_reg_state *reg,
5807 int regno, int off, int size,
5808 bool zero_size_allowed,
5811 const char *buf_info = type_is_rdonly_mem(reg->type) ? "rdonly" : "rdwr";
5814 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
5818 if (off + size > *max_access)
5819 *max_access = off + size;
5824 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
5825 static void zext_32_to_64(struct bpf_reg_state *reg)
5827 reg->var_off = tnum_subreg(reg->var_off);
5828 __reg_assign_32_into_64(reg);
5831 /* truncate register to smaller size (in bytes)
5832 * must be called with size < BPF_REG_SIZE
5834 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
5838 /* clear high bits in bit representation */
5839 reg->var_off = tnum_cast(reg->var_off, size);
5841 /* fix arithmetic bounds */
5842 mask = ((u64)1 << (size * 8)) - 1;
5843 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
5844 reg->umin_value &= mask;
5845 reg->umax_value &= mask;
5847 reg->umin_value = 0;
5848 reg->umax_value = mask;
5850 reg->smin_value = reg->umin_value;
5851 reg->smax_value = reg->umax_value;
5853 /* If size is smaller than 32bit register the 32bit register
5854 * values are also truncated so we push 64-bit bounds into
5855 * 32-bit bounds. Above were truncated < 32-bits already.
5859 __reg_combine_64_into_32(reg);
5862 static void set_sext64_default_val(struct bpf_reg_state *reg, int size)
5865 reg->smin_value = reg->s32_min_value = S8_MIN;
5866 reg->smax_value = reg->s32_max_value = S8_MAX;
5867 } else if (size == 2) {
5868 reg->smin_value = reg->s32_min_value = S16_MIN;
5869 reg->smax_value = reg->s32_max_value = S16_MAX;
5872 reg->smin_value = reg->s32_min_value = S32_MIN;
5873 reg->smax_value = reg->s32_max_value = S32_MAX;
5875 reg->umin_value = reg->u32_min_value = 0;
5876 reg->umax_value = U64_MAX;
5877 reg->u32_max_value = U32_MAX;
5878 reg->var_off = tnum_unknown;
5881 static void coerce_reg_to_size_sx(struct bpf_reg_state *reg, int size)
5883 s64 init_s64_max, init_s64_min, s64_max, s64_min, u64_cval;
5884 u64 top_smax_value, top_smin_value;
5885 u64 num_bits = size * 8;
5887 if (tnum_is_const(reg->var_off)) {
5888 u64_cval = reg->var_off.value;
5890 reg->var_off = tnum_const((s8)u64_cval);
5892 reg->var_off = tnum_const((s16)u64_cval);
5895 reg->var_off = tnum_const((s32)u64_cval);
5897 u64_cval = reg->var_off.value;
5898 reg->smax_value = reg->smin_value = u64_cval;
5899 reg->umax_value = reg->umin_value = u64_cval;
5900 reg->s32_max_value = reg->s32_min_value = u64_cval;
5901 reg->u32_max_value = reg->u32_min_value = u64_cval;
5905 top_smax_value = ((u64)reg->smax_value >> num_bits) << num_bits;
5906 top_smin_value = ((u64)reg->smin_value >> num_bits) << num_bits;
5908 if (top_smax_value != top_smin_value)
5911 /* find the s64_min and s64_min after sign extension */
5913 init_s64_max = (s8)reg->smax_value;
5914 init_s64_min = (s8)reg->smin_value;
5915 } else if (size == 2) {
5916 init_s64_max = (s16)reg->smax_value;
5917 init_s64_min = (s16)reg->smin_value;
5919 init_s64_max = (s32)reg->smax_value;
5920 init_s64_min = (s32)reg->smin_value;
5923 s64_max = max(init_s64_max, init_s64_min);
5924 s64_min = min(init_s64_max, init_s64_min);
5926 /* both of s64_max/s64_min positive or negative */
5927 if ((s64_max >= 0) == (s64_min >= 0)) {
5928 reg->smin_value = reg->s32_min_value = s64_min;
5929 reg->smax_value = reg->s32_max_value = s64_max;
5930 reg->umin_value = reg->u32_min_value = s64_min;
5931 reg->umax_value = reg->u32_max_value = s64_max;
5932 reg->var_off = tnum_range(s64_min, s64_max);
5937 set_sext64_default_val(reg, size);
5940 static void set_sext32_default_val(struct bpf_reg_state *reg, int size)
5943 reg->s32_min_value = S8_MIN;
5944 reg->s32_max_value = S8_MAX;
5947 reg->s32_min_value = S16_MIN;
5948 reg->s32_max_value = S16_MAX;
5950 reg->u32_min_value = 0;
5951 reg->u32_max_value = U32_MAX;
5954 static void coerce_subreg_to_size_sx(struct bpf_reg_state *reg, int size)
5956 s32 init_s32_max, init_s32_min, s32_max, s32_min, u32_val;
5957 u32 top_smax_value, top_smin_value;
5958 u32 num_bits = size * 8;
5960 if (tnum_is_const(reg->var_off)) {
5961 u32_val = reg->var_off.value;
5963 reg->var_off = tnum_const((s8)u32_val);
5965 reg->var_off = tnum_const((s16)u32_val);
5967 u32_val = reg->var_off.value;
5968 reg->s32_min_value = reg->s32_max_value = u32_val;
5969 reg->u32_min_value = reg->u32_max_value = u32_val;
5973 top_smax_value = ((u32)reg->s32_max_value >> num_bits) << num_bits;
5974 top_smin_value = ((u32)reg->s32_min_value >> num_bits) << num_bits;
5976 if (top_smax_value != top_smin_value)
5979 /* find the s32_min and s32_min after sign extension */
5981 init_s32_max = (s8)reg->s32_max_value;
5982 init_s32_min = (s8)reg->s32_min_value;
5985 init_s32_max = (s16)reg->s32_max_value;
5986 init_s32_min = (s16)reg->s32_min_value;
5988 s32_max = max(init_s32_max, init_s32_min);
5989 s32_min = min(init_s32_max, init_s32_min);
5991 if ((s32_min >= 0) == (s32_max >= 0)) {
5992 reg->s32_min_value = s32_min;
5993 reg->s32_max_value = s32_max;
5994 reg->u32_min_value = (u32)s32_min;
5995 reg->u32_max_value = (u32)s32_max;
6000 set_sext32_default_val(reg, size);
6003 static bool bpf_map_is_rdonly(const struct bpf_map *map)
6005 /* A map is considered read-only if the following condition are true:
6007 * 1) BPF program side cannot change any of the map content. The
6008 * BPF_F_RDONLY_PROG flag is throughout the lifetime of a map
6009 * and was set at map creation time.
6010 * 2) The map value(s) have been initialized from user space by a
6011 * loader and then "frozen", such that no new map update/delete
6012 * operations from syscall side are possible for the rest of
6013 * the map's lifetime from that point onwards.
6014 * 3) Any parallel/pending map update/delete operations from syscall
6015 * side have been completed. Only after that point, it's safe to
6016 * assume that map value(s) are immutable.
6018 return (map->map_flags & BPF_F_RDONLY_PROG) &&
6019 READ_ONCE(map->frozen) &&
6020 !bpf_map_write_active(map);
6023 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val,
6030 err = map->ops->map_direct_value_addr(map, &addr, off);
6033 ptr = (void *)(long)addr + off;
6037 *val = is_ldsx ? (s64)*(s8 *)ptr : (u64)*(u8 *)ptr;
6040 *val = is_ldsx ? (s64)*(s16 *)ptr : (u64)*(u16 *)ptr;
6043 *val = is_ldsx ? (s64)*(s32 *)ptr : (u64)*(u32 *)ptr;
6054 #define BTF_TYPE_SAFE_RCU(__type) __PASTE(__type, __safe_rcu)
6055 #define BTF_TYPE_SAFE_RCU_OR_NULL(__type) __PASTE(__type, __safe_rcu_or_null)
6056 #define BTF_TYPE_SAFE_TRUSTED(__type) __PASTE(__type, __safe_trusted)
6059 * Allow list few fields as RCU trusted or full trusted.
6060 * This logic doesn't allow mix tagging and will be removed once GCC supports
6064 /* RCU trusted: these fields are trusted in RCU CS and never NULL */
6065 BTF_TYPE_SAFE_RCU(struct task_struct) {
6066 const cpumask_t *cpus_ptr;
6067 struct css_set __rcu *cgroups;
6068 struct task_struct __rcu *real_parent;
6069 struct task_struct *group_leader;
6072 BTF_TYPE_SAFE_RCU(struct cgroup) {
6073 /* cgrp->kn is always accessible as documented in kernel/cgroup/cgroup.c */
6074 struct kernfs_node *kn;
6077 BTF_TYPE_SAFE_RCU(struct css_set) {
6078 struct cgroup *dfl_cgrp;
6081 /* RCU trusted: these fields are trusted in RCU CS and can be NULL */
6082 BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct) {
6083 struct file __rcu *exe_file;
6086 /* skb->sk, req->sk are not RCU protected, but we mark them as such
6087 * because bpf prog accessible sockets are SOCK_RCU_FREE.
6089 BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff) {
6093 BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock) {
6097 /* full trusted: these fields are trusted even outside of RCU CS and never NULL */
6098 BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta) {
6099 struct seq_file *seq;
6102 BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task) {
6103 struct bpf_iter_meta *meta;
6104 struct task_struct *task;
6107 BTF_TYPE_SAFE_TRUSTED(struct linux_binprm) {
6111 BTF_TYPE_SAFE_TRUSTED(struct file) {
6112 struct inode *f_inode;
6115 BTF_TYPE_SAFE_TRUSTED(struct dentry) {
6116 /* no negative dentry-s in places where bpf can see it */
6117 struct inode *d_inode;
6120 BTF_TYPE_SAFE_TRUSTED(struct socket) {
6124 static bool type_is_rcu(struct bpf_verifier_env *env,
6125 struct bpf_reg_state *reg,
6126 const char *field_name, u32 btf_id)
6128 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct task_struct));
6129 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct cgroup));
6130 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct css_set));
6132 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_rcu");
6135 static bool type_is_rcu_or_null(struct bpf_verifier_env *env,
6136 struct bpf_reg_state *reg,
6137 const char *field_name, u32 btf_id)
6139 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct));
6140 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff));
6141 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock));
6143 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_rcu_or_null");
6146 static bool type_is_trusted(struct bpf_verifier_env *env,
6147 struct bpf_reg_state *reg,
6148 const char *field_name, u32 btf_id)
6150 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta));
6151 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task));
6152 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct linux_binprm));
6153 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct file));
6154 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct dentry));
6155 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct socket));
6157 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_trusted");
6160 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
6161 struct bpf_reg_state *regs,
6162 int regno, int off, int size,
6163 enum bpf_access_type atype,
6166 struct bpf_reg_state *reg = regs + regno;
6167 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
6168 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
6169 const char *field_name = NULL;
6170 enum bpf_type_flag flag = 0;
6174 if (!env->allow_ptr_leaks) {
6176 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
6180 if (!env->prog->gpl_compatible && btf_is_kernel(reg->btf)) {
6182 "Cannot access kernel 'struct %s' from non-GPL compatible program\n",
6188 "R%d is ptr_%s invalid negative access: off=%d\n",
6192 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
6195 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6197 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
6198 regno, tname, off, tn_buf);
6202 if (reg->type & MEM_USER) {
6204 "R%d is ptr_%s access user memory: off=%d\n",
6209 if (reg->type & MEM_PERCPU) {
6211 "R%d is ptr_%s access percpu memory: off=%d\n",
6216 if (env->ops->btf_struct_access && !type_is_alloc(reg->type) && atype == BPF_WRITE) {
6217 if (!btf_is_kernel(reg->btf)) {
6218 verbose(env, "verifier internal error: reg->btf must be kernel btf\n");
6221 ret = env->ops->btf_struct_access(&env->log, reg, off, size);
6223 /* Writes are permitted with default btf_struct_access for
6224 * program allocated objects (which always have ref_obj_id > 0),
6225 * but not for untrusted PTR_TO_BTF_ID | MEM_ALLOC.
6227 if (atype != BPF_READ && !type_is_ptr_alloc_obj(reg->type)) {
6228 verbose(env, "only read is supported\n");
6232 if (type_is_alloc(reg->type) && !type_is_non_owning_ref(reg->type) &&
6234 verbose(env, "verifier internal error: ref_obj_id for allocated object must be non-zero\n");
6238 ret = btf_struct_access(&env->log, reg, off, size, atype, &btf_id, &flag, &field_name);
6244 if (ret != PTR_TO_BTF_ID) {
6247 } else if (type_flag(reg->type) & PTR_UNTRUSTED) {
6248 /* If this is an untrusted pointer, all pointers formed by walking it
6249 * also inherit the untrusted flag.
6251 flag = PTR_UNTRUSTED;
6253 } else if (is_trusted_reg(reg) || is_rcu_reg(reg)) {
6254 /* By default any pointer obtained from walking a trusted pointer is no
6255 * longer trusted, unless the field being accessed has explicitly been
6256 * marked as inheriting its parent's state of trust (either full or RCU).
6258 * 'cgroups' pointer is untrusted if task->cgroups dereference
6259 * happened in a sleepable program outside of bpf_rcu_read_lock()
6260 * section. In a non-sleepable program it's trusted while in RCU CS (aka MEM_RCU).
6261 * Note bpf_rcu_read_unlock() converts MEM_RCU pointers to PTR_UNTRUSTED.
6263 * A regular RCU-protected pointer with __rcu tag can also be deemed
6264 * trusted if we are in an RCU CS. Such pointer can be NULL.
6266 if (type_is_trusted(env, reg, field_name, btf_id)) {
6267 flag |= PTR_TRUSTED;
6268 } else if (in_rcu_cs(env) && !type_may_be_null(reg->type)) {
6269 if (type_is_rcu(env, reg, field_name, btf_id)) {
6270 /* ignore __rcu tag and mark it MEM_RCU */
6272 } else if (flag & MEM_RCU ||
6273 type_is_rcu_or_null(env, reg, field_name, btf_id)) {
6274 /* __rcu tagged pointers can be NULL */
6275 flag |= MEM_RCU | PTR_MAYBE_NULL;
6277 /* We always trust them */
6278 if (type_is_rcu_or_null(env, reg, field_name, btf_id) &&
6279 flag & PTR_UNTRUSTED)
6280 flag &= ~PTR_UNTRUSTED;
6281 } else if (flag & (MEM_PERCPU | MEM_USER)) {
6284 /* walking unknown pointers yields old deprecated PTR_TO_BTF_ID */
6285 clear_trusted_flags(&flag);
6289 * If not in RCU CS or MEM_RCU pointer can be NULL then
6290 * aggressively mark as untrusted otherwise such
6291 * pointers will be plain PTR_TO_BTF_ID without flags
6292 * and will be allowed to be passed into helpers for
6295 flag = PTR_UNTRUSTED;
6298 /* Old compat. Deprecated */
6299 clear_trusted_flags(&flag);
6302 if (atype == BPF_READ && value_regno >= 0)
6303 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id, flag);
6308 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
6309 struct bpf_reg_state *regs,
6310 int regno, int off, int size,
6311 enum bpf_access_type atype,
6314 struct bpf_reg_state *reg = regs + regno;
6315 struct bpf_map *map = reg->map_ptr;
6316 struct bpf_reg_state map_reg;
6317 enum bpf_type_flag flag = 0;
6318 const struct btf_type *t;
6324 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
6328 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
6329 verbose(env, "map_ptr access not supported for map type %d\n",
6334 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
6335 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
6337 if (!env->allow_ptr_leaks) {
6339 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
6345 verbose(env, "R%d is %s invalid negative access: off=%d\n",
6350 if (atype != BPF_READ) {
6351 verbose(env, "only read from %s is supported\n", tname);
6355 /* Simulate access to a PTR_TO_BTF_ID */
6356 memset(&map_reg, 0, sizeof(map_reg));
6357 mark_btf_ld_reg(env, &map_reg, 0, PTR_TO_BTF_ID, btf_vmlinux, *map->ops->map_btf_id, 0);
6358 ret = btf_struct_access(&env->log, &map_reg, off, size, atype, &btf_id, &flag, NULL);
6362 if (value_regno >= 0)
6363 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id, flag);
6368 /* Check that the stack access at the given offset is within bounds. The
6369 * maximum valid offset is -1.
6371 * The minimum valid offset is -MAX_BPF_STACK for writes, and
6372 * -state->allocated_stack for reads.
6374 static int check_stack_slot_within_bounds(struct bpf_verifier_env *env,
6376 struct bpf_func_state *state,
6377 enum bpf_access_type t)
6381 if (t == BPF_WRITE || env->allow_uninit_stack)
6382 min_valid_off = -MAX_BPF_STACK;
6384 min_valid_off = -state->allocated_stack;
6386 if (off < min_valid_off || off > -1)
6391 /* Check that the stack access at 'regno + off' falls within the maximum stack
6394 * 'off' includes `regno->offset`, but not its dynamic part (if any).
6396 static int check_stack_access_within_bounds(
6397 struct bpf_verifier_env *env,
6398 int regno, int off, int access_size,
6399 enum bpf_access_src src, enum bpf_access_type type)
6401 struct bpf_reg_state *regs = cur_regs(env);
6402 struct bpf_reg_state *reg = regs + regno;
6403 struct bpf_func_state *state = func(env, reg);
6404 s64 min_off, max_off;
6408 if (src == ACCESS_HELPER)
6409 /* We don't know if helpers are reading or writing (or both). */
6410 err_extra = " indirect access to";
6411 else if (type == BPF_READ)
6412 err_extra = " read from";
6414 err_extra = " write to";
6416 if (tnum_is_const(reg->var_off)) {
6417 min_off = (s64)reg->var_off.value + off;
6418 max_off = min_off + access_size;
6420 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
6421 reg->smin_value <= -BPF_MAX_VAR_OFF) {
6422 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
6426 min_off = reg->smin_value + off;
6427 max_off = reg->smax_value + off + access_size;
6430 err = check_stack_slot_within_bounds(env, min_off, state, type);
6431 if (!err && max_off > 0)
6432 err = -EINVAL; /* out of stack access into non-negative offsets */
6435 if (tnum_is_const(reg->var_off)) {
6436 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
6437 err_extra, regno, off, access_size);
6441 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6442 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
6443 err_extra, regno, tn_buf, access_size);
6448 return grow_stack_state(env, state, round_up(-min_off, BPF_REG_SIZE));
6451 /* check whether memory at (regno + off) is accessible for t = (read | write)
6452 * if t==write, value_regno is a register which value is stored into memory
6453 * if t==read, value_regno is a register which will receive the value from memory
6454 * if t==write && value_regno==-1, some unknown value is stored into memory
6455 * if t==read && value_regno==-1, don't care what we read from memory
6457 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
6458 int off, int bpf_size, enum bpf_access_type t,
6459 int value_regno, bool strict_alignment_once, bool is_ldsx)
6461 struct bpf_reg_state *regs = cur_regs(env);
6462 struct bpf_reg_state *reg = regs + regno;
6465 size = bpf_size_to_bytes(bpf_size);
6469 /* alignment checks will add in reg->off themselves */
6470 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
6474 /* for access checks, reg->off is just part of off */
6477 if (reg->type == PTR_TO_MAP_KEY) {
6478 if (t == BPF_WRITE) {
6479 verbose(env, "write to change key R%d not allowed\n", regno);
6483 err = check_mem_region_access(env, regno, off, size,
6484 reg->map_ptr->key_size, false);
6487 if (value_regno >= 0)
6488 mark_reg_unknown(env, regs, value_regno);
6489 } else if (reg->type == PTR_TO_MAP_VALUE) {
6490 struct btf_field *kptr_field = NULL;
6492 if (t == BPF_WRITE && value_regno >= 0 &&
6493 is_pointer_value(env, value_regno)) {
6494 verbose(env, "R%d leaks addr into map\n", value_regno);
6497 err = check_map_access_type(env, regno, off, size, t);
6500 err = check_map_access(env, regno, off, size, false, ACCESS_DIRECT);
6503 if (tnum_is_const(reg->var_off))
6504 kptr_field = btf_record_find(reg->map_ptr->record,
6505 off + reg->var_off.value, BPF_KPTR);
6507 err = check_map_kptr_access(env, regno, value_regno, insn_idx, kptr_field);
6508 } else if (t == BPF_READ && value_regno >= 0) {
6509 struct bpf_map *map = reg->map_ptr;
6511 /* if map is read-only, track its contents as scalars */
6512 if (tnum_is_const(reg->var_off) &&
6513 bpf_map_is_rdonly(map) &&
6514 map->ops->map_direct_value_addr) {
6515 int map_off = off + reg->var_off.value;
6518 err = bpf_map_direct_read(map, map_off, size,
6523 regs[value_regno].type = SCALAR_VALUE;
6524 __mark_reg_known(®s[value_regno], val);
6526 mark_reg_unknown(env, regs, value_regno);
6529 } else if (base_type(reg->type) == PTR_TO_MEM) {
6530 bool rdonly_mem = type_is_rdonly_mem(reg->type);
6532 if (type_may_be_null(reg->type)) {
6533 verbose(env, "R%d invalid mem access '%s'\n", regno,
6534 reg_type_str(env, reg->type));
6538 if (t == BPF_WRITE && rdonly_mem) {
6539 verbose(env, "R%d cannot write into %s\n",
6540 regno, reg_type_str(env, reg->type));
6544 if (t == BPF_WRITE && value_regno >= 0 &&
6545 is_pointer_value(env, value_regno)) {
6546 verbose(env, "R%d leaks addr into mem\n", value_regno);
6550 err = check_mem_region_access(env, regno, off, size,
6551 reg->mem_size, false);
6552 if (!err && value_regno >= 0 && (t == BPF_READ || rdonly_mem))
6553 mark_reg_unknown(env, regs, value_regno);
6554 } else if (reg->type == PTR_TO_CTX) {
6555 enum bpf_reg_type reg_type = SCALAR_VALUE;
6556 struct btf *btf = NULL;
6559 if (t == BPF_WRITE && value_regno >= 0 &&
6560 is_pointer_value(env, value_regno)) {
6561 verbose(env, "R%d leaks addr into ctx\n", value_regno);
6565 err = check_ptr_off_reg(env, reg, regno);
6569 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf,
6572 verbose_linfo(env, insn_idx, "; ");
6573 if (!err && t == BPF_READ && value_regno >= 0) {
6574 /* ctx access returns either a scalar, or a
6575 * PTR_TO_PACKET[_META,_END]. In the latter
6576 * case, we know the offset is zero.
6578 if (reg_type == SCALAR_VALUE) {
6579 mark_reg_unknown(env, regs, value_regno);
6581 mark_reg_known_zero(env, regs,
6583 if (type_may_be_null(reg_type))
6584 regs[value_regno].id = ++env->id_gen;
6585 /* A load of ctx field could have different
6586 * actual load size with the one encoded in the
6587 * insn. When the dst is PTR, it is for sure not
6590 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
6591 if (base_type(reg_type) == PTR_TO_BTF_ID) {
6592 regs[value_regno].btf = btf;
6593 regs[value_regno].btf_id = btf_id;
6596 regs[value_regno].type = reg_type;
6599 } else if (reg->type == PTR_TO_STACK) {
6600 /* Basic bounds checks. */
6601 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
6606 err = check_stack_read(env, regno, off, size,
6609 err = check_stack_write(env, regno, off, size,
6610 value_regno, insn_idx);
6611 } else if (reg_is_pkt_pointer(reg)) {
6612 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
6613 verbose(env, "cannot write into packet\n");
6616 if (t == BPF_WRITE && value_regno >= 0 &&
6617 is_pointer_value(env, value_regno)) {
6618 verbose(env, "R%d leaks addr into packet\n",
6622 err = check_packet_access(env, regno, off, size, false);
6623 if (!err && t == BPF_READ && value_regno >= 0)
6624 mark_reg_unknown(env, regs, value_regno);
6625 } else if (reg->type == PTR_TO_FLOW_KEYS) {
6626 if (t == BPF_WRITE && value_regno >= 0 &&
6627 is_pointer_value(env, value_regno)) {
6628 verbose(env, "R%d leaks addr into flow keys\n",
6633 err = check_flow_keys_access(env, off, size);
6634 if (!err && t == BPF_READ && value_regno >= 0)
6635 mark_reg_unknown(env, regs, value_regno);
6636 } else if (type_is_sk_pointer(reg->type)) {
6637 if (t == BPF_WRITE) {
6638 verbose(env, "R%d cannot write into %s\n",
6639 regno, reg_type_str(env, reg->type));
6642 err = check_sock_access(env, insn_idx, regno, off, size, t);
6643 if (!err && value_regno >= 0)
6644 mark_reg_unknown(env, regs, value_regno);
6645 } else if (reg->type == PTR_TO_TP_BUFFER) {
6646 err = check_tp_buffer_access(env, reg, regno, off, size);
6647 if (!err && t == BPF_READ && value_regno >= 0)
6648 mark_reg_unknown(env, regs, value_regno);
6649 } else if (base_type(reg->type) == PTR_TO_BTF_ID &&
6650 !type_may_be_null(reg->type)) {
6651 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
6653 } else if (reg->type == CONST_PTR_TO_MAP) {
6654 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
6656 } else if (base_type(reg->type) == PTR_TO_BUF) {
6657 bool rdonly_mem = type_is_rdonly_mem(reg->type);
6661 if (t == BPF_WRITE) {
6662 verbose(env, "R%d cannot write into %s\n",
6663 regno, reg_type_str(env, reg->type));
6666 max_access = &env->prog->aux->max_rdonly_access;
6668 max_access = &env->prog->aux->max_rdwr_access;
6671 err = check_buffer_access(env, reg, regno, off, size, false,
6674 if (!err && value_regno >= 0 && (rdonly_mem || t == BPF_READ))
6675 mark_reg_unknown(env, regs, value_regno);
6677 verbose(env, "R%d invalid mem access '%s'\n", regno,
6678 reg_type_str(env, reg->type));
6682 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
6683 regs[value_regno].type == SCALAR_VALUE) {
6685 /* b/h/w load zero-extends, mark upper bits as known 0 */
6686 coerce_reg_to_size(®s[value_regno], size);
6688 coerce_reg_to_size_sx(®s[value_regno], size);
6693 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
6698 switch (insn->imm) {
6700 case BPF_ADD | BPF_FETCH:
6702 case BPF_AND | BPF_FETCH:
6704 case BPF_OR | BPF_FETCH:
6706 case BPF_XOR | BPF_FETCH:
6711 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
6715 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
6716 verbose(env, "invalid atomic operand size\n");
6720 /* check src1 operand */
6721 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6725 /* check src2 operand */
6726 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6730 if (insn->imm == BPF_CMPXCHG) {
6731 /* Check comparison of R0 with memory location */
6732 const u32 aux_reg = BPF_REG_0;
6734 err = check_reg_arg(env, aux_reg, SRC_OP);
6738 if (is_pointer_value(env, aux_reg)) {
6739 verbose(env, "R%d leaks addr into mem\n", aux_reg);
6744 if (is_pointer_value(env, insn->src_reg)) {
6745 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
6749 if (is_ctx_reg(env, insn->dst_reg) ||
6750 is_pkt_reg(env, insn->dst_reg) ||
6751 is_flow_key_reg(env, insn->dst_reg) ||
6752 is_sk_reg(env, insn->dst_reg)) {
6753 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
6755 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
6759 if (insn->imm & BPF_FETCH) {
6760 if (insn->imm == BPF_CMPXCHG)
6761 load_reg = BPF_REG_0;
6763 load_reg = insn->src_reg;
6765 /* check and record load of old value */
6766 err = check_reg_arg(env, load_reg, DST_OP);
6770 /* This instruction accesses a memory location but doesn't
6771 * actually load it into a register.
6776 /* Check whether we can read the memory, with second call for fetch
6777 * case to simulate the register fill.
6779 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
6780 BPF_SIZE(insn->code), BPF_READ, -1, true, false);
6781 if (!err && load_reg >= 0)
6782 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
6783 BPF_SIZE(insn->code), BPF_READ, load_reg,
6788 /* Check whether we can write into the same memory. */
6789 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
6790 BPF_SIZE(insn->code), BPF_WRITE, -1, true, false);
6797 /* When register 'regno' is used to read the stack (either directly or through
6798 * a helper function) make sure that it's within stack boundary and, depending
6799 * on the access type and privileges, that all elements of the stack are
6802 * 'off' includes 'regno->off', but not its dynamic part (if any).
6804 * All registers that have been spilled on the stack in the slots within the
6805 * read offsets are marked as read.
6807 static int check_stack_range_initialized(
6808 struct bpf_verifier_env *env, int regno, int off,
6809 int access_size, bool zero_size_allowed,
6810 enum bpf_access_src type, struct bpf_call_arg_meta *meta)
6812 struct bpf_reg_state *reg = reg_state(env, regno);
6813 struct bpf_func_state *state = func(env, reg);
6814 int err, min_off, max_off, i, j, slot, spi;
6815 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
6816 enum bpf_access_type bounds_check_type;
6817 /* Some accesses can write anything into the stack, others are
6820 bool clobber = false;
6822 if (access_size == 0 && !zero_size_allowed) {
6823 verbose(env, "invalid zero-sized read\n");
6827 if (type == ACCESS_HELPER) {
6828 /* The bounds checks for writes are more permissive than for
6829 * reads. However, if raw_mode is not set, we'll do extra
6832 bounds_check_type = BPF_WRITE;
6835 bounds_check_type = BPF_READ;
6837 err = check_stack_access_within_bounds(env, regno, off, access_size,
6838 type, bounds_check_type);
6843 if (tnum_is_const(reg->var_off)) {
6844 min_off = max_off = reg->var_off.value + off;
6846 /* Variable offset is prohibited for unprivileged mode for
6847 * simplicity since it requires corresponding support in
6848 * Spectre masking for stack ALU.
6849 * See also retrieve_ptr_limit().
6851 if (!env->bypass_spec_v1) {
6854 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6855 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
6856 regno, err_extra, tn_buf);
6859 /* Only initialized buffer on stack is allowed to be accessed
6860 * with variable offset. With uninitialized buffer it's hard to
6861 * guarantee that whole memory is marked as initialized on
6862 * helper return since specific bounds are unknown what may
6863 * cause uninitialized stack leaking.
6865 if (meta && meta->raw_mode)
6868 min_off = reg->smin_value + off;
6869 max_off = reg->smax_value + off;
6872 if (meta && meta->raw_mode) {
6873 /* Ensure we won't be overwriting dynptrs when simulating byte
6874 * by byte access in check_helper_call using meta.access_size.
6875 * This would be a problem if we have a helper in the future
6878 * helper(uninit_mem, len, dynptr)
6880 * Now, uninint_mem may overlap with dynptr pointer. Hence, it
6881 * may end up writing to dynptr itself when touching memory from
6882 * arg 1. This can be relaxed on a case by case basis for known
6883 * safe cases, but reject due to the possibilitiy of aliasing by
6886 for (i = min_off; i < max_off + access_size; i++) {
6887 int stack_off = -i - 1;
6890 /* raw_mode may write past allocated_stack */
6891 if (state->allocated_stack <= stack_off)
6893 if (state->stack[spi].slot_type[stack_off % BPF_REG_SIZE] == STACK_DYNPTR) {
6894 verbose(env, "potential write to dynptr at off=%d disallowed\n", i);
6898 meta->access_size = access_size;
6899 meta->regno = regno;
6903 for (i = min_off; i < max_off + access_size; i++) {
6907 spi = slot / BPF_REG_SIZE;
6908 if (state->allocated_stack <= slot) {
6909 verbose(env, "verifier bug: allocated_stack too small");
6913 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
6914 if (*stype == STACK_MISC)
6916 if ((*stype == STACK_ZERO) ||
6917 (*stype == STACK_INVALID && env->allow_uninit_stack)) {
6919 /* helper can write anything into the stack */
6920 *stype = STACK_MISC;
6925 if (is_spilled_reg(&state->stack[spi]) &&
6926 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
6927 env->allow_ptr_leaks)) {
6929 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
6930 for (j = 0; j < BPF_REG_SIZE; j++)
6931 scrub_spilled_slot(&state->stack[spi].slot_type[j]);
6936 if (tnum_is_const(reg->var_off)) {
6937 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
6938 err_extra, regno, min_off, i - min_off, access_size);
6942 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6943 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
6944 err_extra, regno, tn_buf, i - min_off, access_size);
6948 /* reading any byte out of 8-byte 'spill_slot' will cause
6949 * the whole slot to be marked as 'read'
6951 mark_reg_read(env, &state->stack[spi].spilled_ptr,
6952 state->stack[spi].spilled_ptr.parent,
6954 /* We do not set REG_LIVE_WRITTEN for stack slot, as we can not
6955 * be sure that whether stack slot is written to or not. Hence,
6956 * we must still conservatively propagate reads upwards even if
6957 * helper may write to the entire memory range.
6963 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
6964 int access_size, bool zero_size_allowed,
6965 struct bpf_call_arg_meta *meta)
6967 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
6970 switch (base_type(reg->type)) {
6972 case PTR_TO_PACKET_META:
6973 return check_packet_access(env, regno, reg->off, access_size,
6975 case PTR_TO_MAP_KEY:
6976 if (meta && meta->raw_mode) {
6977 verbose(env, "R%d cannot write into %s\n", regno,
6978 reg_type_str(env, reg->type));
6981 return check_mem_region_access(env, regno, reg->off, access_size,
6982 reg->map_ptr->key_size, false);
6983 case PTR_TO_MAP_VALUE:
6984 if (check_map_access_type(env, regno, reg->off, access_size,
6985 meta && meta->raw_mode ? BPF_WRITE :
6988 return check_map_access(env, regno, reg->off, access_size,
6989 zero_size_allowed, ACCESS_HELPER);
6991 if (type_is_rdonly_mem(reg->type)) {
6992 if (meta && meta->raw_mode) {
6993 verbose(env, "R%d cannot write into %s\n", regno,
6994 reg_type_str(env, reg->type));
6998 return check_mem_region_access(env, regno, reg->off,
6999 access_size, reg->mem_size,
7002 if (type_is_rdonly_mem(reg->type)) {
7003 if (meta && meta->raw_mode) {
7004 verbose(env, "R%d cannot write into %s\n", regno,
7005 reg_type_str(env, reg->type));
7009 max_access = &env->prog->aux->max_rdonly_access;
7011 max_access = &env->prog->aux->max_rdwr_access;
7013 return check_buffer_access(env, reg, regno, reg->off,
7014 access_size, zero_size_allowed,
7017 return check_stack_range_initialized(
7019 regno, reg->off, access_size,
7020 zero_size_allowed, ACCESS_HELPER, meta);
7022 return check_ptr_to_btf_access(env, regs, regno, reg->off,
7023 access_size, BPF_READ, -1);
7025 /* in case the function doesn't know how to access the context,
7026 * (because we are in a program of type SYSCALL for example), we
7027 * can not statically check its size.
7028 * Dynamically check it now.
7030 if (!env->ops->convert_ctx_access) {
7031 enum bpf_access_type atype = meta && meta->raw_mode ? BPF_WRITE : BPF_READ;
7032 int offset = access_size - 1;
7034 /* Allow zero-byte read from PTR_TO_CTX */
7035 if (access_size == 0)
7036 return zero_size_allowed ? 0 : -EACCES;
7038 return check_mem_access(env, env->insn_idx, regno, offset, BPF_B,
7039 atype, -1, false, false);
7043 default: /* scalar_value or invalid ptr */
7044 /* Allow zero-byte read from NULL, regardless of pointer type */
7045 if (zero_size_allowed && access_size == 0 &&
7046 register_is_null(reg))
7049 verbose(env, "R%d type=%s ", regno,
7050 reg_type_str(env, reg->type));
7051 verbose(env, "expected=%s\n", reg_type_str(env, PTR_TO_STACK));
7056 static int check_mem_size_reg(struct bpf_verifier_env *env,
7057 struct bpf_reg_state *reg, u32 regno,
7058 bool zero_size_allowed,
7059 struct bpf_call_arg_meta *meta)
7063 /* This is used to refine r0 return value bounds for helpers
7064 * that enforce this value as an upper bound on return values.
7065 * See do_refine_retval_range() for helpers that can refine
7066 * the return value. C type of helper is u32 so we pull register
7067 * bound from umax_value however, if negative verifier errors
7068 * out. Only upper bounds can be learned because retval is an
7069 * int type and negative retvals are allowed.
7071 meta->msize_max_value = reg->umax_value;
7073 /* The register is SCALAR_VALUE; the access check
7074 * happens using its boundaries.
7076 if (!tnum_is_const(reg->var_off))
7077 /* For unprivileged variable accesses, disable raw
7078 * mode so that the program is required to
7079 * initialize all the memory that the helper could
7080 * just partially fill up.
7084 if (reg->smin_value < 0) {
7085 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
7090 if (reg->umin_value == 0) {
7091 err = check_helper_mem_access(env, regno - 1, 0,
7098 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
7099 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
7103 err = check_helper_mem_access(env, regno - 1,
7105 zero_size_allowed, meta);
7107 err = mark_chain_precision(env, regno);
7111 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
7112 u32 regno, u32 mem_size)
7114 bool may_be_null = type_may_be_null(reg->type);
7115 struct bpf_reg_state saved_reg;
7116 struct bpf_call_arg_meta meta;
7119 if (register_is_null(reg))
7122 memset(&meta, 0, sizeof(meta));
7123 /* Assuming that the register contains a value check if the memory
7124 * access is safe. Temporarily save and restore the register's state as
7125 * the conversion shouldn't be visible to a caller.
7129 mark_ptr_not_null_reg(reg);
7132 err = check_helper_mem_access(env, regno, mem_size, true, &meta);
7133 /* Check access for BPF_WRITE */
7134 meta.raw_mode = true;
7135 err = err ?: check_helper_mem_access(env, regno, mem_size, true, &meta);
7143 static int check_kfunc_mem_size_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
7146 struct bpf_reg_state *mem_reg = &cur_regs(env)[regno - 1];
7147 bool may_be_null = type_may_be_null(mem_reg->type);
7148 struct bpf_reg_state saved_reg;
7149 struct bpf_call_arg_meta meta;
7152 WARN_ON_ONCE(regno < BPF_REG_2 || regno > BPF_REG_5);
7154 memset(&meta, 0, sizeof(meta));
7157 saved_reg = *mem_reg;
7158 mark_ptr_not_null_reg(mem_reg);
7161 err = check_mem_size_reg(env, reg, regno, true, &meta);
7162 /* Check access for BPF_WRITE */
7163 meta.raw_mode = true;
7164 err = err ?: check_mem_size_reg(env, reg, regno, true, &meta);
7167 *mem_reg = saved_reg;
7171 /* Implementation details:
7172 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL.
7173 * bpf_obj_new returns PTR_TO_BTF_ID | MEM_ALLOC | PTR_MAYBE_NULL.
7174 * Two bpf_map_lookups (even with the same key) will have different reg->id.
7175 * Two separate bpf_obj_new will also have different reg->id.
7176 * For traditional PTR_TO_MAP_VALUE or PTR_TO_BTF_ID | MEM_ALLOC, the verifier
7177 * clears reg->id after value_or_null->value transition, since the verifier only
7178 * cares about the range of access to valid map value pointer and doesn't care
7179 * about actual address of the map element.
7180 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
7181 * reg->id > 0 after value_or_null->value transition. By doing so
7182 * two bpf_map_lookups will be considered two different pointers that
7183 * point to different bpf_spin_locks. Likewise for pointers to allocated objects
7184 * returned from bpf_obj_new.
7185 * The verifier allows taking only one bpf_spin_lock at a time to avoid
7187 * Since only one bpf_spin_lock is allowed the checks are simpler than
7188 * reg_is_refcounted() logic. The verifier needs to remember only
7189 * one spin_lock instead of array of acquired_refs.
7190 * cur_state->active_lock remembers which map value element or allocated
7191 * object got locked and clears it after bpf_spin_unlock.
7193 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
7196 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7197 struct bpf_verifier_state *cur = env->cur_state;
7198 bool is_const = tnum_is_const(reg->var_off);
7199 u64 val = reg->var_off.value;
7200 struct bpf_map *map = NULL;
7201 struct btf *btf = NULL;
7202 struct btf_record *rec;
7206 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
7210 if (reg->type == PTR_TO_MAP_VALUE) {
7214 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
7222 rec = reg_btf_record(reg);
7223 if (!btf_record_has_field(rec, BPF_SPIN_LOCK)) {
7224 verbose(env, "%s '%s' has no valid bpf_spin_lock\n", map ? "map" : "local",
7225 map ? map->name : "kptr");
7228 if (rec->spin_lock_off != val + reg->off) {
7229 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock' that is at %d\n",
7230 val + reg->off, rec->spin_lock_off);
7234 if (cur->active_lock.ptr) {
7236 "Locking two bpf_spin_locks are not allowed\n");
7240 cur->active_lock.ptr = map;
7242 cur->active_lock.ptr = btf;
7243 cur->active_lock.id = reg->id;
7252 if (!cur->active_lock.ptr) {
7253 verbose(env, "bpf_spin_unlock without taking a lock\n");
7256 if (cur->active_lock.ptr != ptr ||
7257 cur->active_lock.id != reg->id) {
7258 verbose(env, "bpf_spin_unlock of different lock\n");
7262 invalidate_non_owning_refs(env);
7264 cur->active_lock.ptr = NULL;
7265 cur->active_lock.id = 0;
7270 static int process_timer_func(struct bpf_verifier_env *env, int regno,
7271 struct bpf_call_arg_meta *meta)
7273 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7274 bool is_const = tnum_is_const(reg->var_off);
7275 struct bpf_map *map = reg->map_ptr;
7276 u64 val = reg->var_off.value;
7280 "R%d doesn't have constant offset. bpf_timer has to be at the constant offset\n",
7285 verbose(env, "map '%s' has to have BTF in order to use bpf_timer\n",
7289 if (!btf_record_has_field(map->record, BPF_TIMER)) {
7290 verbose(env, "map '%s' has no valid bpf_timer\n", map->name);
7293 if (map->record->timer_off != val + reg->off) {
7294 verbose(env, "off %lld doesn't point to 'struct bpf_timer' that is at %d\n",
7295 val + reg->off, map->record->timer_off);
7298 if (meta->map_ptr) {
7299 verbose(env, "verifier bug. Two map pointers in a timer helper\n");
7302 meta->map_uid = reg->map_uid;
7303 meta->map_ptr = map;
7307 static int process_kptr_func(struct bpf_verifier_env *env, int regno,
7308 struct bpf_call_arg_meta *meta)
7310 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7311 struct bpf_map *map_ptr = reg->map_ptr;
7312 struct btf_field *kptr_field;
7315 if (!tnum_is_const(reg->var_off)) {
7317 "R%d doesn't have constant offset. kptr has to be at the constant offset\n",
7321 if (!map_ptr->btf) {
7322 verbose(env, "map '%s' has to have BTF in order to use bpf_kptr_xchg\n",
7326 if (!btf_record_has_field(map_ptr->record, BPF_KPTR)) {
7327 verbose(env, "map '%s' has no valid kptr\n", map_ptr->name);
7331 meta->map_ptr = map_ptr;
7332 kptr_off = reg->off + reg->var_off.value;
7333 kptr_field = btf_record_find(map_ptr->record, kptr_off, BPF_KPTR);
7335 verbose(env, "off=%d doesn't point to kptr\n", kptr_off);
7338 if (kptr_field->type != BPF_KPTR_REF) {
7339 verbose(env, "off=%d kptr isn't referenced kptr\n", kptr_off);
7342 meta->kptr_field = kptr_field;
7346 /* There are two register types representing a bpf_dynptr, one is PTR_TO_STACK
7347 * which points to a stack slot, and the other is CONST_PTR_TO_DYNPTR.
7349 * In both cases we deal with the first 8 bytes, but need to mark the next 8
7350 * bytes as STACK_DYNPTR in case of PTR_TO_STACK. In case of
7351 * CONST_PTR_TO_DYNPTR, we are guaranteed to get the beginning of the object.
7353 * Mutability of bpf_dynptr is at two levels, one is at the level of struct
7354 * bpf_dynptr itself, i.e. whether the helper is receiving a pointer to struct
7355 * bpf_dynptr or pointer to const struct bpf_dynptr. In the former case, it can
7356 * mutate the view of the dynptr and also possibly destroy it. In the latter
7357 * case, it cannot mutate the bpf_dynptr itself but it can still mutate the
7358 * memory that dynptr points to.
7360 * The verifier will keep track both levels of mutation (bpf_dynptr's in
7361 * reg->type and the memory's in reg->dynptr.type), but there is no support for
7362 * readonly dynptr view yet, hence only the first case is tracked and checked.
7364 * This is consistent with how C applies the const modifier to a struct object,
7365 * where the pointer itself inside bpf_dynptr becomes const but not what it
7368 * Helpers which do not mutate the bpf_dynptr set MEM_RDONLY in their argument
7369 * type, and declare it as 'const struct bpf_dynptr *' in their prototype.
7371 static int process_dynptr_func(struct bpf_verifier_env *env, int regno, int insn_idx,
7372 enum bpf_arg_type arg_type, int clone_ref_obj_id)
7374 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7377 /* MEM_UNINIT and MEM_RDONLY are exclusive, when applied to an
7378 * ARG_PTR_TO_DYNPTR (or ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_*):
7380 if ((arg_type & (MEM_UNINIT | MEM_RDONLY)) == (MEM_UNINIT | MEM_RDONLY)) {
7381 verbose(env, "verifier internal error: misconfigured dynptr helper type flags\n");
7385 /* MEM_UNINIT - Points to memory that is an appropriate candidate for
7386 * constructing a mutable bpf_dynptr object.
7388 * Currently, this is only possible with PTR_TO_STACK
7389 * pointing to a region of at least 16 bytes which doesn't
7390 * contain an existing bpf_dynptr.
7392 * MEM_RDONLY - Points to a initialized bpf_dynptr that will not be
7393 * mutated or destroyed. However, the memory it points to
7396 * None - Points to a initialized dynptr that can be mutated and
7397 * destroyed, including mutation of the memory it points
7400 if (arg_type & MEM_UNINIT) {
7403 if (!is_dynptr_reg_valid_uninit(env, reg)) {
7404 verbose(env, "Dynptr has to be an uninitialized dynptr\n");
7408 /* we write BPF_DW bits (8 bytes) at a time */
7409 for (i = 0; i < BPF_DYNPTR_SIZE; i += 8) {
7410 err = check_mem_access(env, insn_idx, regno,
7411 i, BPF_DW, BPF_WRITE, -1, false, false);
7416 err = mark_stack_slots_dynptr(env, reg, arg_type, insn_idx, clone_ref_obj_id);
7417 } else /* MEM_RDONLY and None case from above */ {
7418 /* For the reg->type == PTR_TO_STACK case, bpf_dynptr is never const */
7419 if (reg->type == CONST_PTR_TO_DYNPTR && !(arg_type & MEM_RDONLY)) {
7420 verbose(env, "cannot pass pointer to const bpf_dynptr, the helper mutates it\n");
7424 if (!is_dynptr_reg_valid_init(env, reg)) {
7426 "Expected an initialized dynptr as arg #%d\n",
7431 /* Fold modifiers (in this case, MEM_RDONLY) when checking expected type */
7432 if (!is_dynptr_type_expected(env, reg, arg_type & ~MEM_RDONLY)) {
7434 "Expected a dynptr of type %s as arg #%d\n",
7435 dynptr_type_str(arg_to_dynptr_type(arg_type)), regno);
7439 err = mark_dynptr_read(env, reg);
7444 static u32 iter_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg, int spi)
7446 struct bpf_func_state *state = func(env, reg);
7448 return state->stack[spi].spilled_ptr.ref_obj_id;
7451 static bool is_iter_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7453 return meta->kfunc_flags & (KF_ITER_NEW | KF_ITER_NEXT | KF_ITER_DESTROY);
7456 static bool is_iter_new_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7458 return meta->kfunc_flags & KF_ITER_NEW;
7461 static bool is_iter_next_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7463 return meta->kfunc_flags & KF_ITER_NEXT;
7466 static bool is_iter_destroy_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7468 return meta->kfunc_flags & KF_ITER_DESTROY;
7471 static bool is_kfunc_arg_iter(struct bpf_kfunc_call_arg_meta *meta, int arg)
7473 /* btf_check_iter_kfuncs() guarantees that first argument of any iter
7474 * kfunc is iter state pointer
7476 return arg == 0 && is_iter_kfunc(meta);
7479 static int process_iter_arg(struct bpf_verifier_env *env, int regno, int insn_idx,
7480 struct bpf_kfunc_call_arg_meta *meta)
7482 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7483 const struct btf_type *t;
7484 const struct btf_param *arg;
7485 int spi, err, i, nr_slots;
7488 /* btf_check_iter_kfuncs() ensures we don't need to validate anything here */
7489 arg = &btf_params(meta->func_proto)[0];
7490 t = btf_type_skip_modifiers(meta->btf, arg->type, NULL); /* PTR */
7491 t = btf_type_skip_modifiers(meta->btf, t->type, &btf_id); /* STRUCT */
7492 nr_slots = t->size / BPF_REG_SIZE;
7494 if (is_iter_new_kfunc(meta)) {
7495 /* bpf_iter_<type>_new() expects pointer to uninit iter state */
7496 if (!is_iter_reg_valid_uninit(env, reg, nr_slots)) {
7497 verbose(env, "expected uninitialized iter_%s as arg #%d\n",
7498 iter_type_str(meta->btf, btf_id), regno);
7502 for (i = 0; i < nr_slots * 8; i += BPF_REG_SIZE) {
7503 err = check_mem_access(env, insn_idx, regno,
7504 i, BPF_DW, BPF_WRITE, -1, false, false);
7509 err = mark_stack_slots_iter(env, reg, insn_idx, meta->btf, btf_id, nr_slots);
7513 /* iter_next() or iter_destroy() expect initialized iter state*/
7514 if (!is_iter_reg_valid_init(env, reg, meta->btf, btf_id, nr_slots)) {
7515 verbose(env, "expected an initialized iter_%s as arg #%d\n",
7516 iter_type_str(meta->btf, btf_id), regno);
7520 spi = iter_get_spi(env, reg, nr_slots);
7524 err = mark_iter_read(env, reg, spi, nr_slots);
7528 /* remember meta->iter info for process_iter_next_call() */
7529 meta->iter.spi = spi;
7530 meta->iter.frameno = reg->frameno;
7531 meta->ref_obj_id = iter_ref_obj_id(env, reg, spi);
7533 if (is_iter_destroy_kfunc(meta)) {
7534 err = unmark_stack_slots_iter(env, reg, nr_slots);
7543 /* process_iter_next_call() is called when verifier gets to iterator's next
7544 * "method" (e.g., bpf_iter_num_next() for numbers iterator) call. We'll refer
7545 * to it as just "iter_next()" in comments below.
7547 * BPF verifier relies on a crucial contract for any iter_next()
7548 * implementation: it should *eventually* return NULL, and once that happens
7549 * it should keep returning NULL. That is, once iterator exhausts elements to
7550 * iterate, it should never reset or spuriously return new elements.
7552 * With the assumption of such contract, process_iter_next_call() simulates
7553 * a fork in the verifier state to validate loop logic correctness and safety
7554 * without having to simulate infinite amount of iterations.
7556 * In current state, we first assume that iter_next() returned NULL and
7557 * iterator state is set to DRAINED (BPF_ITER_STATE_DRAINED). In such
7558 * conditions we should not form an infinite loop and should eventually reach
7561 * Besides that, we also fork current state and enqueue it for later
7562 * verification. In a forked state we keep iterator state as ACTIVE
7563 * (BPF_ITER_STATE_ACTIVE) and assume non-NULL return from iter_next(). We
7564 * also bump iteration depth to prevent erroneous infinite loop detection
7565 * later on (see iter_active_depths_differ() comment for details). In this
7566 * state we assume that we'll eventually loop back to another iter_next()
7567 * calls (it could be in exactly same location or in some other instruction,
7568 * it doesn't matter, we don't make any unnecessary assumptions about this,
7569 * everything revolves around iterator state in a stack slot, not which
7570 * instruction is calling iter_next()). When that happens, we either will come
7571 * to iter_next() with equivalent state and can conclude that next iteration
7572 * will proceed in exactly the same way as we just verified, so it's safe to
7573 * assume that loop converges. If not, we'll go on another iteration
7574 * simulation with a different input state, until all possible starting states
7575 * are validated or we reach maximum number of instructions limit.
7577 * This way, we will either exhaustively discover all possible input states
7578 * that iterator loop can start with and eventually will converge, or we'll
7579 * effectively regress into bounded loop simulation logic and either reach
7580 * maximum number of instructions if loop is not provably convergent, or there
7581 * is some statically known limit on number of iterations (e.g., if there is
7582 * an explicit `if n > 100 then break;` statement somewhere in the loop).
7584 * One very subtle but very important aspect is that we *always* simulate NULL
7585 * condition first (as the current state) before we simulate non-NULL case.
7586 * This has to do with intricacies of scalar precision tracking. By simulating
7587 * "exit condition" of iter_next() returning NULL first, we make sure all the
7588 * relevant precision marks *that will be set **after** we exit iterator loop*
7589 * are propagated backwards to common parent state of NULL and non-NULL
7590 * branches. Thanks to that, state equivalence checks done later in forked
7591 * state, when reaching iter_next() for ACTIVE iterator, can assume that
7592 * precision marks are finalized and won't change. Because simulating another
7593 * ACTIVE iterator iteration won't change them (because given same input
7594 * states we'll end up with exactly same output states which we are currently
7595 * comparing; and verification after the loop already propagated back what
7596 * needs to be **additionally** tracked as precise). It's subtle, grok
7597 * precision tracking for more intuitive understanding.
7599 static int process_iter_next_call(struct bpf_verifier_env *env, int insn_idx,
7600 struct bpf_kfunc_call_arg_meta *meta)
7602 struct bpf_verifier_state *cur_st = env->cur_state, *queued_st;
7603 struct bpf_func_state *cur_fr = cur_st->frame[cur_st->curframe], *queued_fr;
7604 struct bpf_reg_state *cur_iter, *queued_iter;
7605 int iter_frameno = meta->iter.frameno;
7606 int iter_spi = meta->iter.spi;
7608 BTF_TYPE_EMIT(struct bpf_iter);
7610 cur_iter = &env->cur_state->frame[iter_frameno]->stack[iter_spi].spilled_ptr;
7612 if (cur_iter->iter.state != BPF_ITER_STATE_ACTIVE &&
7613 cur_iter->iter.state != BPF_ITER_STATE_DRAINED) {
7614 verbose(env, "verifier internal error: unexpected iterator state %d (%s)\n",
7615 cur_iter->iter.state, iter_state_str(cur_iter->iter.state));
7619 if (cur_iter->iter.state == BPF_ITER_STATE_ACTIVE) {
7620 /* branch out active iter state */
7621 queued_st = push_stack(env, insn_idx + 1, insn_idx, false);
7625 queued_iter = &queued_st->frame[iter_frameno]->stack[iter_spi].spilled_ptr;
7626 queued_iter->iter.state = BPF_ITER_STATE_ACTIVE;
7627 queued_iter->iter.depth++;
7629 queued_fr = queued_st->frame[queued_st->curframe];
7630 mark_ptr_not_null_reg(&queued_fr->regs[BPF_REG_0]);
7633 /* switch to DRAINED state, but keep the depth unchanged */
7634 /* mark current iter state as drained and assume returned NULL */
7635 cur_iter->iter.state = BPF_ITER_STATE_DRAINED;
7636 __mark_reg_const_zero(&cur_fr->regs[BPF_REG_0]);
7641 static bool arg_type_is_mem_size(enum bpf_arg_type type)
7643 return type == ARG_CONST_SIZE ||
7644 type == ARG_CONST_SIZE_OR_ZERO;
7647 static bool arg_type_is_release(enum bpf_arg_type type)
7649 return type & OBJ_RELEASE;
7652 static bool arg_type_is_dynptr(enum bpf_arg_type type)
7654 return base_type(type) == ARG_PTR_TO_DYNPTR;
7657 static int int_ptr_type_to_size(enum bpf_arg_type type)
7659 if (type == ARG_PTR_TO_INT)
7661 else if (type == ARG_PTR_TO_LONG)
7667 static int resolve_map_arg_type(struct bpf_verifier_env *env,
7668 const struct bpf_call_arg_meta *meta,
7669 enum bpf_arg_type *arg_type)
7671 if (!meta->map_ptr) {
7672 /* kernel subsystem misconfigured verifier */
7673 verbose(env, "invalid map_ptr to access map->type\n");
7677 switch (meta->map_ptr->map_type) {
7678 case BPF_MAP_TYPE_SOCKMAP:
7679 case BPF_MAP_TYPE_SOCKHASH:
7680 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
7681 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
7683 verbose(env, "invalid arg_type for sockmap/sockhash\n");
7687 case BPF_MAP_TYPE_BLOOM_FILTER:
7688 if (meta->func_id == BPF_FUNC_map_peek_elem)
7689 *arg_type = ARG_PTR_TO_MAP_VALUE;
7697 struct bpf_reg_types {
7698 const enum bpf_reg_type types[10];
7702 static const struct bpf_reg_types sock_types = {
7712 static const struct bpf_reg_types btf_id_sock_common_types = {
7719 PTR_TO_BTF_ID | PTR_TRUSTED,
7721 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
7725 static const struct bpf_reg_types mem_types = {
7733 PTR_TO_MEM | MEM_RINGBUF,
7735 PTR_TO_BTF_ID | PTR_TRUSTED,
7739 static const struct bpf_reg_types int_ptr_types = {
7749 static const struct bpf_reg_types spin_lock_types = {
7752 PTR_TO_BTF_ID | MEM_ALLOC,
7756 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
7757 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
7758 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
7759 static const struct bpf_reg_types ringbuf_mem_types = { .types = { PTR_TO_MEM | MEM_RINGBUF } };
7760 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
7761 static const struct bpf_reg_types btf_ptr_types = {
7764 PTR_TO_BTF_ID | PTR_TRUSTED,
7765 PTR_TO_BTF_ID | MEM_RCU,
7768 static const struct bpf_reg_types percpu_btf_ptr_types = {
7770 PTR_TO_BTF_ID | MEM_PERCPU,
7771 PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED,
7774 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
7775 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
7776 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
7777 static const struct bpf_reg_types timer_types = { .types = { PTR_TO_MAP_VALUE } };
7778 static const struct bpf_reg_types kptr_types = { .types = { PTR_TO_MAP_VALUE } };
7779 static const struct bpf_reg_types dynptr_types = {
7782 CONST_PTR_TO_DYNPTR,
7786 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
7787 [ARG_PTR_TO_MAP_KEY] = &mem_types,
7788 [ARG_PTR_TO_MAP_VALUE] = &mem_types,
7789 [ARG_CONST_SIZE] = &scalar_types,
7790 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
7791 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
7792 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
7793 [ARG_PTR_TO_CTX] = &context_types,
7794 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
7796 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
7798 [ARG_PTR_TO_SOCKET] = &fullsock_types,
7799 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
7800 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
7801 [ARG_PTR_TO_MEM] = &mem_types,
7802 [ARG_PTR_TO_RINGBUF_MEM] = &ringbuf_mem_types,
7803 [ARG_PTR_TO_INT] = &int_ptr_types,
7804 [ARG_PTR_TO_LONG] = &int_ptr_types,
7805 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
7806 [ARG_PTR_TO_FUNC] = &func_ptr_types,
7807 [ARG_PTR_TO_STACK] = &stack_ptr_types,
7808 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
7809 [ARG_PTR_TO_TIMER] = &timer_types,
7810 [ARG_PTR_TO_KPTR] = &kptr_types,
7811 [ARG_PTR_TO_DYNPTR] = &dynptr_types,
7814 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
7815 enum bpf_arg_type arg_type,
7816 const u32 *arg_btf_id,
7817 struct bpf_call_arg_meta *meta)
7819 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7820 enum bpf_reg_type expected, type = reg->type;
7821 const struct bpf_reg_types *compatible;
7824 compatible = compatible_reg_types[base_type(arg_type)];
7826 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
7830 /* ARG_PTR_TO_MEM + RDONLY is compatible with PTR_TO_MEM and PTR_TO_MEM + RDONLY,
7831 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM and NOT with PTR_TO_MEM + RDONLY
7833 * Same for MAYBE_NULL:
7835 * ARG_PTR_TO_MEM + MAYBE_NULL is compatible with PTR_TO_MEM and PTR_TO_MEM + MAYBE_NULL,
7836 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM but NOT with PTR_TO_MEM + MAYBE_NULL
7838 * ARG_PTR_TO_MEM is compatible with PTR_TO_MEM that is tagged with a dynptr type.
7840 * Therefore we fold these flags depending on the arg_type before comparison.
7842 if (arg_type & MEM_RDONLY)
7843 type &= ~MEM_RDONLY;
7844 if (arg_type & PTR_MAYBE_NULL)
7845 type &= ~PTR_MAYBE_NULL;
7846 if (base_type(arg_type) == ARG_PTR_TO_MEM)
7847 type &= ~DYNPTR_TYPE_FLAG_MASK;
7849 if (meta->func_id == BPF_FUNC_kptr_xchg && type_is_alloc(type))
7852 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
7853 expected = compatible->types[i];
7854 if (expected == NOT_INIT)
7857 if (type == expected)
7861 verbose(env, "R%d type=%s expected=", regno, reg_type_str(env, reg->type));
7862 for (j = 0; j + 1 < i; j++)
7863 verbose(env, "%s, ", reg_type_str(env, compatible->types[j]));
7864 verbose(env, "%s\n", reg_type_str(env, compatible->types[j]));
7868 if (base_type(reg->type) != PTR_TO_BTF_ID)
7871 if (compatible == &mem_types) {
7872 if (!(arg_type & MEM_RDONLY)) {
7874 "%s() may write into memory pointed by R%d type=%s\n",
7875 func_id_name(meta->func_id),
7876 regno, reg_type_str(env, reg->type));
7882 switch ((int)reg->type) {
7884 case PTR_TO_BTF_ID | PTR_TRUSTED:
7885 case PTR_TO_BTF_ID | MEM_RCU:
7886 case PTR_TO_BTF_ID | PTR_MAYBE_NULL:
7887 case PTR_TO_BTF_ID | PTR_MAYBE_NULL | MEM_RCU:
7889 /* For bpf_sk_release, it needs to match against first member
7890 * 'struct sock_common', hence make an exception for it. This
7891 * allows bpf_sk_release to work for multiple socket types.
7893 bool strict_type_match = arg_type_is_release(arg_type) &&
7894 meta->func_id != BPF_FUNC_sk_release;
7896 if (type_may_be_null(reg->type) &&
7897 (!type_may_be_null(arg_type) || arg_type_is_release(arg_type))) {
7898 verbose(env, "Possibly NULL pointer passed to helper arg%d\n", regno);
7903 if (!compatible->btf_id) {
7904 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
7907 arg_btf_id = compatible->btf_id;
7910 if (meta->func_id == BPF_FUNC_kptr_xchg) {
7911 if (map_kptr_match_type(env, meta->kptr_field, reg, regno))
7914 if (arg_btf_id == BPF_PTR_POISON) {
7915 verbose(env, "verifier internal error:");
7916 verbose(env, "R%d has non-overwritten BPF_PTR_POISON type\n",
7921 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
7922 btf_vmlinux, *arg_btf_id,
7923 strict_type_match)) {
7924 verbose(env, "R%d is of type %s but %s is expected\n",
7925 regno, btf_type_name(reg->btf, reg->btf_id),
7926 btf_type_name(btf_vmlinux, *arg_btf_id));
7932 case PTR_TO_BTF_ID | MEM_ALLOC:
7933 if (meta->func_id != BPF_FUNC_spin_lock && meta->func_id != BPF_FUNC_spin_unlock &&
7934 meta->func_id != BPF_FUNC_kptr_xchg) {
7935 verbose(env, "verifier internal error: unimplemented handling of MEM_ALLOC\n");
7938 if (meta->func_id == BPF_FUNC_kptr_xchg) {
7939 if (map_kptr_match_type(env, meta->kptr_field, reg, regno))
7943 case PTR_TO_BTF_ID | MEM_PERCPU:
7944 case PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED:
7945 /* Handled by helper specific checks */
7948 verbose(env, "verifier internal error: invalid PTR_TO_BTF_ID register for type match\n");
7954 static struct btf_field *
7955 reg_find_field_offset(const struct bpf_reg_state *reg, s32 off, u32 fields)
7957 struct btf_field *field;
7958 struct btf_record *rec;
7960 rec = reg_btf_record(reg);
7964 field = btf_record_find(rec, off, fields);
7971 int check_func_arg_reg_off(struct bpf_verifier_env *env,
7972 const struct bpf_reg_state *reg, int regno,
7973 enum bpf_arg_type arg_type)
7975 u32 type = reg->type;
7977 /* When referenced register is passed to release function, its fixed
7980 * We will check arg_type_is_release reg has ref_obj_id when storing
7981 * meta->release_regno.
7983 if (arg_type_is_release(arg_type)) {
7984 /* ARG_PTR_TO_DYNPTR with OBJ_RELEASE is a bit special, as it
7985 * may not directly point to the object being released, but to
7986 * dynptr pointing to such object, which might be at some offset
7987 * on the stack. In that case, we simply to fallback to the
7990 if (arg_type_is_dynptr(arg_type) && type == PTR_TO_STACK)
7993 /* Doing check_ptr_off_reg check for the offset will catch this
7994 * because fixed_off_ok is false, but checking here allows us
7995 * to give the user a better error message.
7998 verbose(env, "R%d must have zero offset when passed to release func or trusted arg to kfunc\n",
8002 return __check_ptr_off_reg(env, reg, regno, false);
8006 /* Pointer types where both fixed and variable offset is explicitly allowed: */
8009 case PTR_TO_PACKET_META:
8010 case PTR_TO_MAP_KEY:
8011 case PTR_TO_MAP_VALUE:
8013 case PTR_TO_MEM | MEM_RDONLY:
8014 case PTR_TO_MEM | MEM_RINGBUF:
8016 case PTR_TO_BUF | MEM_RDONLY:
8019 /* All the rest must be rejected, except PTR_TO_BTF_ID which allows
8023 case PTR_TO_BTF_ID | MEM_ALLOC:
8024 case PTR_TO_BTF_ID | PTR_TRUSTED:
8025 case PTR_TO_BTF_ID | MEM_RCU:
8026 case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF:
8027 case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF | MEM_RCU:
8028 /* When referenced PTR_TO_BTF_ID is passed to release function,
8029 * its fixed offset must be 0. In the other cases, fixed offset
8030 * can be non-zero. This was already checked above. So pass
8031 * fixed_off_ok as true to allow fixed offset for all other
8032 * cases. var_off always must be 0 for PTR_TO_BTF_ID, hence we
8033 * still need to do checks instead of returning.
8035 return __check_ptr_off_reg(env, reg, regno, true);
8037 return __check_ptr_off_reg(env, reg, regno, false);
8041 static struct bpf_reg_state *get_dynptr_arg_reg(struct bpf_verifier_env *env,
8042 const struct bpf_func_proto *fn,
8043 struct bpf_reg_state *regs)
8045 struct bpf_reg_state *state = NULL;
8048 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++)
8049 if (arg_type_is_dynptr(fn->arg_type[i])) {
8051 verbose(env, "verifier internal error: multiple dynptr args\n");
8054 state = ®s[BPF_REG_1 + i];
8058 verbose(env, "verifier internal error: no dynptr arg found\n");
8063 static int dynptr_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
8065 struct bpf_func_state *state = func(env, reg);
8068 if (reg->type == CONST_PTR_TO_DYNPTR)
8070 spi = dynptr_get_spi(env, reg);
8073 return state->stack[spi].spilled_ptr.id;
8076 static int dynptr_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
8078 struct bpf_func_state *state = func(env, reg);
8081 if (reg->type == CONST_PTR_TO_DYNPTR)
8082 return reg->ref_obj_id;
8083 spi = dynptr_get_spi(env, reg);
8086 return state->stack[spi].spilled_ptr.ref_obj_id;
8089 static enum bpf_dynptr_type dynptr_get_type(struct bpf_verifier_env *env,
8090 struct bpf_reg_state *reg)
8092 struct bpf_func_state *state = func(env, reg);
8095 if (reg->type == CONST_PTR_TO_DYNPTR)
8096 return reg->dynptr.type;
8098 spi = __get_spi(reg->off);
8100 verbose(env, "verifier internal error: invalid spi when querying dynptr type\n");
8101 return BPF_DYNPTR_TYPE_INVALID;
8104 return state->stack[spi].spilled_ptr.dynptr.type;
8107 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
8108 struct bpf_call_arg_meta *meta,
8109 const struct bpf_func_proto *fn,
8112 u32 regno = BPF_REG_1 + arg;
8113 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
8114 enum bpf_arg_type arg_type = fn->arg_type[arg];
8115 enum bpf_reg_type type = reg->type;
8116 u32 *arg_btf_id = NULL;
8119 if (arg_type == ARG_DONTCARE)
8122 err = check_reg_arg(env, regno, SRC_OP);
8126 if (arg_type == ARG_ANYTHING) {
8127 if (is_pointer_value(env, regno)) {
8128 verbose(env, "R%d leaks addr into helper function\n",
8135 if (type_is_pkt_pointer(type) &&
8136 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
8137 verbose(env, "helper access to the packet is not allowed\n");
8141 if (base_type(arg_type) == ARG_PTR_TO_MAP_VALUE) {
8142 err = resolve_map_arg_type(env, meta, &arg_type);
8147 if (register_is_null(reg) && type_may_be_null(arg_type))
8148 /* A NULL register has a SCALAR_VALUE type, so skip
8151 goto skip_type_check;
8153 /* arg_btf_id and arg_size are in a union. */
8154 if (base_type(arg_type) == ARG_PTR_TO_BTF_ID ||
8155 base_type(arg_type) == ARG_PTR_TO_SPIN_LOCK)
8156 arg_btf_id = fn->arg_btf_id[arg];
8158 err = check_reg_type(env, regno, arg_type, arg_btf_id, meta);
8162 err = check_func_arg_reg_off(env, reg, regno, arg_type);
8167 if (arg_type_is_release(arg_type)) {
8168 if (arg_type_is_dynptr(arg_type)) {
8169 struct bpf_func_state *state = func(env, reg);
8172 /* Only dynptr created on stack can be released, thus
8173 * the get_spi and stack state checks for spilled_ptr
8174 * should only be done before process_dynptr_func for
8177 if (reg->type == PTR_TO_STACK) {
8178 spi = dynptr_get_spi(env, reg);
8179 if (spi < 0 || !state->stack[spi].spilled_ptr.ref_obj_id) {
8180 verbose(env, "arg %d is an unacquired reference\n", regno);
8184 verbose(env, "cannot release unowned const bpf_dynptr\n");
8187 } else if (!reg->ref_obj_id && !register_is_null(reg)) {
8188 verbose(env, "R%d must be referenced when passed to release function\n",
8192 if (meta->release_regno) {
8193 verbose(env, "verifier internal error: more than one release argument\n");
8196 meta->release_regno = regno;
8199 if (reg->ref_obj_id) {
8200 if (meta->ref_obj_id) {
8201 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
8202 regno, reg->ref_obj_id,
8206 meta->ref_obj_id = reg->ref_obj_id;
8209 switch (base_type(arg_type)) {
8210 case ARG_CONST_MAP_PTR:
8211 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
8212 if (meta->map_ptr) {
8213 /* Use map_uid (which is unique id of inner map) to reject:
8214 * inner_map1 = bpf_map_lookup_elem(outer_map, key1)
8215 * inner_map2 = bpf_map_lookup_elem(outer_map, key2)
8216 * if (inner_map1 && inner_map2) {
8217 * timer = bpf_map_lookup_elem(inner_map1);
8219 * // mismatch would have been allowed
8220 * bpf_timer_init(timer, inner_map2);
8223 * Comparing map_ptr is enough to distinguish normal and outer maps.
8225 if (meta->map_ptr != reg->map_ptr ||
8226 meta->map_uid != reg->map_uid) {
8228 "timer pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
8229 meta->map_uid, reg->map_uid);
8233 meta->map_ptr = reg->map_ptr;
8234 meta->map_uid = reg->map_uid;
8236 case ARG_PTR_TO_MAP_KEY:
8237 /* bpf_map_xxx(..., map_ptr, ..., key) call:
8238 * check that [key, key + map->key_size) are within
8239 * stack limits and initialized
8241 if (!meta->map_ptr) {
8242 /* in function declaration map_ptr must come before
8243 * map_key, so that it's verified and known before
8244 * we have to check map_key here. Otherwise it means
8245 * that kernel subsystem misconfigured verifier
8247 verbose(env, "invalid map_ptr to access map->key\n");
8250 err = check_helper_mem_access(env, regno,
8251 meta->map_ptr->key_size, false,
8254 case ARG_PTR_TO_MAP_VALUE:
8255 if (type_may_be_null(arg_type) && register_is_null(reg))
8258 /* bpf_map_xxx(..., map_ptr, ..., value) call:
8259 * check [value, value + map->value_size) validity
8261 if (!meta->map_ptr) {
8262 /* kernel subsystem misconfigured verifier */
8263 verbose(env, "invalid map_ptr to access map->value\n");
8266 meta->raw_mode = arg_type & MEM_UNINIT;
8267 err = check_helper_mem_access(env, regno,
8268 meta->map_ptr->value_size, false,
8271 case ARG_PTR_TO_PERCPU_BTF_ID:
8273 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
8276 meta->ret_btf = reg->btf;
8277 meta->ret_btf_id = reg->btf_id;
8279 case ARG_PTR_TO_SPIN_LOCK:
8280 if (in_rbtree_lock_required_cb(env)) {
8281 verbose(env, "can't spin_{lock,unlock} in rbtree cb\n");
8284 if (meta->func_id == BPF_FUNC_spin_lock) {
8285 err = process_spin_lock(env, regno, true);
8288 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
8289 err = process_spin_lock(env, regno, false);
8293 verbose(env, "verifier internal error\n");
8297 case ARG_PTR_TO_TIMER:
8298 err = process_timer_func(env, regno, meta);
8302 case ARG_PTR_TO_FUNC:
8303 meta->subprogno = reg->subprogno;
8305 case ARG_PTR_TO_MEM:
8306 /* The access to this pointer is only checked when we hit the
8307 * next is_mem_size argument below.
8309 meta->raw_mode = arg_type & MEM_UNINIT;
8310 if (arg_type & MEM_FIXED_SIZE) {
8311 err = check_helper_mem_access(env, regno,
8312 fn->arg_size[arg], false,
8316 case ARG_CONST_SIZE:
8317 err = check_mem_size_reg(env, reg, regno, false, meta);
8319 case ARG_CONST_SIZE_OR_ZERO:
8320 err = check_mem_size_reg(env, reg, regno, true, meta);
8322 case ARG_PTR_TO_DYNPTR:
8323 err = process_dynptr_func(env, regno, insn_idx, arg_type, 0);
8327 case ARG_CONST_ALLOC_SIZE_OR_ZERO:
8328 if (!tnum_is_const(reg->var_off)) {
8329 verbose(env, "R%d is not a known constant'\n",
8333 meta->mem_size = reg->var_off.value;
8334 err = mark_chain_precision(env, regno);
8338 case ARG_PTR_TO_INT:
8339 case ARG_PTR_TO_LONG:
8341 int size = int_ptr_type_to_size(arg_type);
8343 err = check_helper_mem_access(env, regno, size, false, meta);
8346 err = check_ptr_alignment(env, reg, 0, size, true);
8349 case ARG_PTR_TO_CONST_STR:
8351 struct bpf_map *map = reg->map_ptr;
8356 if (!bpf_map_is_rdonly(map)) {
8357 verbose(env, "R%d does not point to a readonly map'\n", regno);
8361 if (!tnum_is_const(reg->var_off)) {
8362 verbose(env, "R%d is not a constant address'\n", regno);
8366 if (!map->ops->map_direct_value_addr) {
8367 verbose(env, "no direct value access support for this map type\n");
8371 err = check_map_access(env, regno, reg->off,
8372 map->value_size - reg->off, false,
8377 map_off = reg->off + reg->var_off.value;
8378 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
8380 verbose(env, "direct value access on string failed\n");
8384 str_ptr = (char *)(long)(map_addr);
8385 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
8386 verbose(env, "string is not zero-terminated\n");
8391 case ARG_PTR_TO_KPTR:
8392 err = process_kptr_func(env, regno, meta);
8401 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
8403 enum bpf_attach_type eatype = env->prog->expected_attach_type;
8404 enum bpf_prog_type type = resolve_prog_type(env->prog);
8406 if (func_id != BPF_FUNC_map_update_elem)
8409 /* It's not possible to get access to a locked struct sock in these
8410 * contexts, so updating is safe.
8413 case BPF_PROG_TYPE_TRACING:
8414 if (eatype == BPF_TRACE_ITER)
8417 case BPF_PROG_TYPE_SOCKET_FILTER:
8418 case BPF_PROG_TYPE_SCHED_CLS:
8419 case BPF_PROG_TYPE_SCHED_ACT:
8420 case BPF_PROG_TYPE_XDP:
8421 case BPF_PROG_TYPE_SK_REUSEPORT:
8422 case BPF_PROG_TYPE_FLOW_DISSECTOR:
8423 case BPF_PROG_TYPE_SK_LOOKUP:
8429 verbose(env, "cannot update sockmap in this context\n");
8433 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
8435 return env->prog->jit_requested &&
8436 bpf_jit_supports_subprog_tailcalls();
8439 static int check_map_func_compatibility(struct bpf_verifier_env *env,
8440 struct bpf_map *map, int func_id)
8445 /* We need a two way check, first is from map perspective ... */
8446 switch (map->map_type) {
8447 case BPF_MAP_TYPE_PROG_ARRAY:
8448 if (func_id != BPF_FUNC_tail_call)
8451 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
8452 if (func_id != BPF_FUNC_perf_event_read &&
8453 func_id != BPF_FUNC_perf_event_output &&
8454 func_id != BPF_FUNC_skb_output &&
8455 func_id != BPF_FUNC_perf_event_read_value &&
8456 func_id != BPF_FUNC_xdp_output)
8459 case BPF_MAP_TYPE_RINGBUF:
8460 if (func_id != BPF_FUNC_ringbuf_output &&
8461 func_id != BPF_FUNC_ringbuf_reserve &&
8462 func_id != BPF_FUNC_ringbuf_query &&
8463 func_id != BPF_FUNC_ringbuf_reserve_dynptr &&
8464 func_id != BPF_FUNC_ringbuf_submit_dynptr &&
8465 func_id != BPF_FUNC_ringbuf_discard_dynptr)
8468 case BPF_MAP_TYPE_USER_RINGBUF:
8469 if (func_id != BPF_FUNC_user_ringbuf_drain)
8472 case BPF_MAP_TYPE_STACK_TRACE:
8473 if (func_id != BPF_FUNC_get_stackid)
8476 case BPF_MAP_TYPE_CGROUP_ARRAY:
8477 if (func_id != BPF_FUNC_skb_under_cgroup &&
8478 func_id != BPF_FUNC_current_task_under_cgroup)
8481 case BPF_MAP_TYPE_CGROUP_STORAGE:
8482 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
8483 if (func_id != BPF_FUNC_get_local_storage)
8486 case BPF_MAP_TYPE_DEVMAP:
8487 case BPF_MAP_TYPE_DEVMAP_HASH:
8488 if (func_id != BPF_FUNC_redirect_map &&
8489 func_id != BPF_FUNC_map_lookup_elem)
8492 /* Restrict bpf side of cpumap and xskmap, open when use-cases
8495 case BPF_MAP_TYPE_CPUMAP:
8496 if (func_id != BPF_FUNC_redirect_map)
8499 case BPF_MAP_TYPE_XSKMAP:
8500 if (func_id != BPF_FUNC_redirect_map &&
8501 func_id != BPF_FUNC_map_lookup_elem)
8504 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
8505 case BPF_MAP_TYPE_HASH_OF_MAPS:
8506 if (func_id != BPF_FUNC_map_lookup_elem)
8509 case BPF_MAP_TYPE_SOCKMAP:
8510 if (func_id != BPF_FUNC_sk_redirect_map &&
8511 func_id != BPF_FUNC_sock_map_update &&
8512 func_id != BPF_FUNC_map_delete_elem &&
8513 func_id != BPF_FUNC_msg_redirect_map &&
8514 func_id != BPF_FUNC_sk_select_reuseport &&
8515 func_id != BPF_FUNC_map_lookup_elem &&
8516 !may_update_sockmap(env, func_id))
8519 case BPF_MAP_TYPE_SOCKHASH:
8520 if (func_id != BPF_FUNC_sk_redirect_hash &&
8521 func_id != BPF_FUNC_sock_hash_update &&
8522 func_id != BPF_FUNC_map_delete_elem &&
8523 func_id != BPF_FUNC_msg_redirect_hash &&
8524 func_id != BPF_FUNC_sk_select_reuseport &&
8525 func_id != BPF_FUNC_map_lookup_elem &&
8526 !may_update_sockmap(env, func_id))
8529 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
8530 if (func_id != BPF_FUNC_sk_select_reuseport)
8533 case BPF_MAP_TYPE_QUEUE:
8534 case BPF_MAP_TYPE_STACK:
8535 if (func_id != BPF_FUNC_map_peek_elem &&
8536 func_id != BPF_FUNC_map_pop_elem &&
8537 func_id != BPF_FUNC_map_push_elem)
8540 case BPF_MAP_TYPE_SK_STORAGE:
8541 if (func_id != BPF_FUNC_sk_storage_get &&
8542 func_id != BPF_FUNC_sk_storage_delete &&
8543 func_id != BPF_FUNC_kptr_xchg)
8546 case BPF_MAP_TYPE_INODE_STORAGE:
8547 if (func_id != BPF_FUNC_inode_storage_get &&
8548 func_id != BPF_FUNC_inode_storage_delete &&
8549 func_id != BPF_FUNC_kptr_xchg)
8552 case BPF_MAP_TYPE_TASK_STORAGE:
8553 if (func_id != BPF_FUNC_task_storage_get &&
8554 func_id != BPF_FUNC_task_storage_delete &&
8555 func_id != BPF_FUNC_kptr_xchg)
8558 case BPF_MAP_TYPE_CGRP_STORAGE:
8559 if (func_id != BPF_FUNC_cgrp_storage_get &&
8560 func_id != BPF_FUNC_cgrp_storage_delete &&
8561 func_id != BPF_FUNC_kptr_xchg)
8564 case BPF_MAP_TYPE_BLOOM_FILTER:
8565 if (func_id != BPF_FUNC_map_peek_elem &&
8566 func_id != BPF_FUNC_map_push_elem)
8573 /* ... and second from the function itself. */
8575 case BPF_FUNC_tail_call:
8576 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
8578 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
8579 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
8583 case BPF_FUNC_perf_event_read:
8584 case BPF_FUNC_perf_event_output:
8585 case BPF_FUNC_perf_event_read_value:
8586 case BPF_FUNC_skb_output:
8587 case BPF_FUNC_xdp_output:
8588 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
8591 case BPF_FUNC_ringbuf_output:
8592 case BPF_FUNC_ringbuf_reserve:
8593 case BPF_FUNC_ringbuf_query:
8594 case BPF_FUNC_ringbuf_reserve_dynptr:
8595 case BPF_FUNC_ringbuf_submit_dynptr:
8596 case BPF_FUNC_ringbuf_discard_dynptr:
8597 if (map->map_type != BPF_MAP_TYPE_RINGBUF)
8600 case BPF_FUNC_user_ringbuf_drain:
8601 if (map->map_type != BPF_MAP_TYPE_USER_RINGBUF)
8604 case BPF_FUNC_get_stackid:
8605 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
8608 case BPF_FUNC_current_task_under_cgroup:
8609 case BPF_FUNC_skb_under_cgroup:
8610 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
8613 case BPF_FUNC_redirect_map:
8614 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
8615 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
8616 map->map_type != BPF_MAP_TYPE_CPUMAP &&
8617 map->map_type != BPF_MAP_TYPE_XSKMAP)
8620 case BPF_FUNC_sk_redirect_map:
8621 case BPF_FUNC_msg_redirect_map:
8622 case BPF_FUNC_sock_map_update:
8623 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
8626 case BPF_FUNC_sk_redirect_hash:
8627 case BPF_FUNC_msg_redirect_hash:
8628 case BPF_FUNC_sock_hash_update:
8629 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
8632 case BPF_FUNC_get_local_storage:
8633 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
8634 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
8637 case BPF_FUNC_sk_select_reuseport:
8638 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
8639 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
8640 map->map_type != BPF_MAP_TYPE_SOCKHASH)
8643 case BPF_FUNC_map_pop_elem:
8644 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
8645 map->map_type != BPF_MAP_TYPE_STACK)
8648 case BPF_FUNC_map_peek_elem:
8649 case BPF_FUNC_map_push_elem:
8650 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
8651 map->map_type != BPF_MAP_TYPE_STACK &&
8652 map->map_type != BPF_MAP_TYPE_BLOOM_FILTER)
8655 case BPF_FUNC_map_lookup_percpu_elem:
8656 if (map->map_type != BPF_MAP_TYPE_PERCPU_ARRAY &&
8657 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
8658 map->map_type != BPF_MAP_TYPE_LRU_PERCPU_HASH)
8661 case BPF_FUNC_sk_storage_get:
8662 case BPF_FUNC_sk_storage_delete:
8663 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
8666 case BPF_FUNC_inode_storage_get:
8667 case BPF_FUNC_inode_storage_delete:
8668 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
8671 case BPF_FUNC_task_storage_get:
8672 case BPF_FUNC_task_storage_delete:
8673 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
8676 case BPF_FUNC_cgrp_storage_get:
8677 case BPF_FUNC_cgrp_storage_delete:
8678 if (map->map_type != BPF_MAP_TYPE_CGRP_STORAGE)
8687 verbose(env, "cannot pass map_type %d into func %s#%d\n",
8688 map->map_type, func_id_name(func_id), func_id);
8692 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
8696 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
8698 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
8700 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
8702 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
8704 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
8707 /* We only support one arg being in raw mode at the moment,
8708 * which is sufficient for the helper functions we have
8714 static bool check_args_pair_invalid(const struct bpf_func_proto *fn, int arg)
8716 bool is_fixed = fn->arg_type[arg] & MEM_FIXED_SIZE;
8717 bool has_size = fn->arg_size[arg] != 0;
8718 bool is_next_size = false;
8720 if (arg + 1 < ARRAY_SIZE(fn->arg_type))
8721 is_next_size = arg_type_is_mem_size(fn->arg_type[arg + 1]);
8723 if (base_type(fn->arg_type[arg]) != ARG_PTR_TO_MEM)
8724 return is_next_size;
8726 return has_size == is_next_size || is_next_size == is_fixed;
8729 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
8731 /* bpf_xxx(..., buf, len) call will access 'len'
8732 * bytes from memory 'buf'. Both arg types need
8733 * to be paired, so make sure there's no buggy
8734 * helper function specification.
8736 if (arg_type_is_mem_size(fn->arg1_type) ||
8737 check_args_pair_invalid(fn, 0) ||
8738 check_args_pair_invalid(fn, 1) ||
8739 check_args_pair_invalid(fn, 2) ||
8740 check_args_pair_invalid(fn, 3) ||
8741 check_args_pair_invalid(fn, 4))
8747 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
8751 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
8752 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_BTF_ID)
8753 return !!fn->arg_btf_id[i];
8754 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_SPIN_LOCK)
8755 return fn->arg_btf_id[i] == BPF_PTR_POISON;
8756 if (base_type(fn->arg_type[i]) != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i] &&
8757 /* arg_btf_id and arg_size are in a union. */
8758 (base_type(fn->arg_type[i]) != ARG_PTR_TO_MEM ||
8759 !(fn->arg_type[i] & MEM_FIXED_SIZE)))
8766 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
8768 return check_raw_mode_ok(fn) &&
8769 check_arg_pair_ok(fn) &&
8770 check_btf_id_ok(fn) ? 0 : -EINVAL;
8773 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
8774 * are now invalid, so turn them into unknown SCALAR_VALUE.
8776 * This also applies to dynptr slices belonging to skb and xdp dynptrs,
8777 * since these slices point to packet data.
8779 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
8781 struct bpf_func_state *state;
8782 struct bpf_reg_state *reg;
8784 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
8785 if (reg_is_pkt_pointer_any(reg) || reg_is_dynptr_slice_pkt(reg))
8786 mark_reg_invalid(env, reg);
8792 BEYOND_PKT_END = -2,
8795 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
8797 struct bpf_func_state *state = vstate->frame[vstate->curframe];
8798 struct bpf_reg_state *reg = &state->regs[regn];
8800 if (reg->type != PTR_TO_PACKET)
8801 /* PTR_TO_PACKET_META is not supported yet */
8804 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
8805 * How far beyond pkt_end it goes is unknown.
8806 * if (!range_open) it's the case of pkt >= pkt_end
8807 * if (range_open) it's the case of pkt > pkt_end
8808 * hence this pointer is at least 1 byte bigger than pkt_end
8811 reg->range = BEYOND_PKT_END;
8813 reg->range = AT_PKT_END;
8816 /* The pointer with the specified id has released its reference to kernel
8817 * resources. Identify all copies of the same pointer and clear the reference.
8819 static int release_reference(struct bpf_verifier_env *env,
8822 struct bpf_func_state *state;
8823 struct bpf_reg_state *reg;
8826 err = release_reference_state(cur_func(env), ref_obj_id);
8830 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
8831 if (reg->ref_obj_id == ref_obj_id)
8832 mark_reg_invalid(env, reg);
8838 static void invalidate_non_owning_refs(struct bpf_verifier_env *env)
8840 struct bpf_func_state *unused;
8841 struct bpf_reg_state *reg;
8843 bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
8844 if (type_is_non_owning_ref(reg->type))
8845 mark_reg_invalid(env, reg);
8849 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
8850 struct bpf_reg_state *regs)
8854 /* after the call registers r0 - r5 were scratched */
8855 for (i = 0; i < CALLER_SAVED_REGS; i++) {
8856 mark_reg_not_init(env, regs, caller_saved[i]);
8857 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
8861 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
8862 struct bpf_func_state *caller,
8863 struct bpf_func_state *callee,
8866 static int set_callee_state(struct bpf_verifier_env *env,
8867 struct bpf_func_state *caller,
8868 struct bpf_func_state *callee, int insn_idx);
8870 static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
8871 int *insn_idx, int subprog,
8872 set_callee_state_fn set_callee_state_cb)
8874 struct bpf_verifier_state *state = env->cur_state;
8875 struct bpf_func_state *caller, *callee;
8878 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
8879 verbose(env, "the call stack of %d frames is too deep\n",
8880 state->curframe + 2);
8884 caller = state->frame[state->curframe];
8885 if (state->frame[state->curframe + 1]) {
8886 verbose(env, "verifier bug. Frame %d already allocated\n",
8887 state->curframe + 1);
8891 err = btf_check_subprog_call(env, subprog, caller->regs);
8894 if (subprog_is_global(env, subprog)) {
8896 verbose(env, "Caller passes invalid args into func#%d\n",
8900 if (env->log.level & BPF_LOG_LEVEL)
8902 "Func#%d is global and valid. Skipping.\n",
8904 clear_caller_saved_regs(env, caller->regs);
8906 /* All global functions return a 64-bit SCALAR_VALUE */
8907 mark_reg_unknown(env, caller->regs, BPF_REG_0);
8908 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
8910 /* continue with next insn after call */
8915 /* set_callee_state is used for direct subprog calls, but we are
8916 * interested in validating only BPF helpers that can call subprogs as
8919 if (set_callee_state_cb != set_callee_state) {
8920 if (bpf_pseudo_kfunc_call(insn) &&
8921 !is_callback_calling_kfunc(insn->imm)) {
8922 verbose(env, "verifier bug: kfunc %s#%d not marked as callback-calling\n",
8923 func_id_name(insn->imm), insn->imm);
8925 } else if (!bpf_pseudo_kfunc_call(insn) &&
8926 !is_callback_calling_function(insn->imm)) { /* helper */
8927 verbose(env, "verifier bug: helper %s#%d not marked as callback-calling\n",
8928 func_id_name(insn->imm), insn->imm);
8933 if (insn->code == (BPF_JMP | BPF_CALL) &&
8934 insn->src_reg == 0 &&
8935 insn->imm == BPF_FUNC_timer_set_callback) {
8936 struct bpf_verifier_state *async_cb;
8938 /* there is no real recursion here. timer callbacks are async */
8939 env->subprog_info[subprog].is_async_cb = true;
8940 async_cb = push_async_cb(env, env->subprog_info[subprog].start,
8941 *insn_idx, subprog);
8944 callee = async_cb->frame[0];
8945 callee->async_entry_cnt = caller->async_entry_cnt + 1;
8947 /* Convert bpf_timer_set_callback() args into timer callback args */
8948 err = set_callee_state_cb(env, caller, callee, *insn_idx);
8952 clear_caller_saved_regs(env, caller->regs);
8953 mark_reg_unknown(env, caller->regs, BPF_REG_0);
8954 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
8955 /* continue with next insn after call */
8959 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
8962 state->frame[state->curframe + 1] = callee;
8964 /* callee cannot access r0, r6 - r9 for reading and has to write
8965 * into its own stack before reading from it.
8966 * callee can read/write into caller's stack
8968 init_func_state(env, callee,
8969 /* remember the callsite, it will be used by bpf_exit */
8970 *insn_idx /* callsite */,
8971 state->curframe + 1 /* frameno within this callchain */,
8972 subprog /* subprog number within this prog */);
8974 /* Transfer references to the callee */
8975 err = copy_reference_state(callee, caller);
8979 err = set_callee_state_cb(env, caller, callee, *insn_idx);
8983 clear_caller_saved_regs(env, caller->regs);
8985 /* only increment it after check_reg_arg() finished */
8988 /* and go analyze first insn of the callee */
8989 *insn_idx = env->subprog_info[subprog].start - 1;
8991 if (env->log.level & BPF_LOG_LEVEL) {
8992 verbose(env, "caller:\n");
8993 print_verifier_state(env, caller, true);
8994 verbose(env, "callee:\n");
8995 print_verifier_state(env, callee, true);
9000 free_func_state(callee);
9001 state->frame[state->curframe + 1] = NULL;
9005 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
9006 struct bpf_func_state *caller,
9007 struct bpf_func_state *callee)
9009 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
9010 * void *callback_ctx, u64 flags);
9011 * callback_fn(struct bpf_map *map, void *key, void *value,
9012 * void *callback_ctx);
9014 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
9016 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
9017 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
9018 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
9020 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
9021 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
9022 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
9024 /* pointer to stack or null */
9025 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
9028 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9032 static int set_callee_state(struct bpf_verifier_env *env,
9033 struct bpf_func_state *caller,
9034 struct bpf_func_state *callee, int insn_idx)
9038 /* copy r1 - r5 args that callee can access. The copy includes parent
9039 * pointers, which connects us up to the liveness chain
9041 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
9042 callee->regs[i] = caller->regs[i];
9046 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
9049 int subprog, target_insn;
9051 target_insn = *insn_idx + insn->imm + 1;
9052 subprog = find_subprog(env, target_insn);
9054 verbose(env, "verifier bug. No program starts at insn %d\n",
9059 return __check_func_call(env, insn, insn_idx, subprog, set_callee_state);
9062 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
9063 struct bpf_func_state *caller,
9064 struct bpf_func_state *callee,
9067 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
9068 struct bpf_map *map;
9071 if (bpf_map_ptr_poisoned(insn_aux)) {
9072 verbose(env, "tail_call abusing map_ptr\n");
9076 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
9077 if (!map->ops->map_set_for_each_callback_args ||
9078 !map->ops->map_for_each_callback) {
9079 verbose(env, "callback function not allowed for map\n");
9083 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
9087 callee->in_callback_fn = true;
9088 callee->callback_ret_range = tnum_range(0, 1);
9092 static int set_loop_callback_state(struct bpf_verifier_env *env,
9093 struct bpf_func_state *caller,
9094 struct bpf_func_state *callee,
9097 /* bpf_loop(u32 nr_loops, void *callback_fn, void *callback_ctx,
9099 * callback_fn(u32 index, void *callback_ctx);
9101 callee->regs[BPF_REG_1].type = SCALAR_VALUE;
9102 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
9105 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
9106 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9107 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9109 callee->in_callback_fn = true;
9110 callee->callback_ret_range = tnum_range(0, 1);
9114 static int set_timer_callback_state(struct bpf_verifier_env *env,
9115 struct bpf_func_state *caller,
9116 struct bpf_func_state *callee,
9119 struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
9121 /* bpf_timer_set_callback(struct bpf_timer *timer, void *callback_fn);
9122 * callback_fn(struct bpf_map *map, void *key, void *value);
9124 callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP;
9125 __mark_reg_known_zero(&callee->regs[BPF_REG_1]);
9126 callee->regs[BPF_REG_1].map_ptr = map_ptr;
9128 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
9129 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
9130 callee->regs[BPF_REG_2].map_ptr = map_ptr;
9132 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
9133 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
9134 callee->regs[BPF_REG_3].map_ptr = map_ptr;
9137 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9138 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9139 callee->in_async_callback_fn = true;
9140 callee->callback_ret_range = tnum_range(0, 1);
9144 static int set_find_vma_callback_state(struct bpf_verifier_env *env,
9145 struct bpf_func_state *caller,
9146 struct bpf_func_state *callee,
9149 /* bpf_find_vma(struct task_struct *task, u64 addr,
9150 * void *callback_fn, void *callback_ctx, u64 flags)
9151 * (callback_fn)(struct task_struct *task,
9152 * struct vm_area_struct *vma, void *callback_ctx);
9154 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
9156 callee->regs[BPF_REG_2].type = PTR_TO_BTF_ID;
9157 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
9158 callee->regs[BPF_REG_2].btf = btf_vmlinux;
9159 callee->regs[BPF_REG_2].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_VMA],
9161 /* pointer to stack or null */
9162 callee->regs[BPF_REG_3] = caller->regs[BPF_REG_4];
9165 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9166 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9167 callee->in_callback_fn = true;
9168 callee->callback_ret_range = tnum_range(0, 1);
9172 static int set_user_ringbuf_callback_state(struct bpf_verifier_env *env,
9173 struct bpf_func_state *caller,
9174 struct bpf_func_state *callee,
9177 /* bpf_user_ringbuf_drain(struct bpf_map *map, void *callback_fn, void
9178 * callback_ctx, u64 flags);
9179 * callback_fn(const struct bpf_dynptr_t* dynptr, void *callback_ctx);
9181 __mark_reg_not_init(env, &callee->regs[BPF_REG_0]);
9182 mark_dynptr_cb_reg(env, &callee->regs[BPF_REG_1], BPF_DYNPTR_TYPE_LOCAL);
9183 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
9186 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
9187 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9188 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9190 callee->in_callback_fn = true;
9191 callee->callback_ret_range = tnum_range(0, 1);
9195 static int set_rbtree_add_callback_state(struct bpf_verifier_env *env,
9196 struct bpf_func_state *caller,
9197 struct bpf_func_state *callee,
9200 /* void bpf_rbtree_add_impl(struct bpf_rb_root *root, struct bpf_rb_node *node,
9201 * bool (less)(struct bpf_rb_node *a, const struct bpf_rb_node *b));
9203 * 'struct bpf_rb_node *node' arg to bpf_rbtree_add_impl is the same PTR_TO_BTF_ID w/ offset
9204 * that 'less' callback args will be receiving. However, 'node' arg was release_reference'd
9205 * by this point, so look at 'root'
9207 struct btf_field *field;
9209 field = reg_find_field_offset(&caller->regs[BPF_REG_1], caller->regs[BPF_REG_1].off,
9211 if (!field || !field->graph_root.value_btf_id)
9214 mark_reg_graph_node(callee->regs, BPF_REG_1, &field->graph_root);
9215 ref_set_non_owning(env, &callee->regs[BPF_REG_1]);
9216 mark_reg_graph_node(callee->regs, BPF_REG_2, &field->graph_root);
9217 ref_set_non_owning(env, &callee->regs[BPF_REG_2]);
9219 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
9220 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9221 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9222 callee->in_callback_fn = true;
9223 callee->callback_ret_range = tnum_range(0, 1);
9227 static bool is_rbtree_lock_required_kfunc(u32 btf_id);
9229 /* Are we currently verifying the callback for a rbtree helper that must
9230 * be called with lock held? If so, no need to complain about unreleased
9233 static bool in_rbtree_lock_required_cb(struct bpf_verifier_env *env)
9235 struct bpf_verifier_state *state = env->cur_state;
9236 struct bpf_insn *insn = env->prog->insnsi;
9237 struct bpf_func_state *callee;
9240 if (!state->curframe)
9243 callee = state->frame[state->curframe];
9245 if (!callee->in_callback_fn)
9248 kfunc_btf_id = insn[callee->callsite].imm;
9249 return is_rbtree_lock_required_kfunc(kfunc_btf_id);
9252 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
9254 struct bpf_verifier_state *state = env->cur_state;
9255 struct bpf_func_state *caller, *callee;
9256 struct bpf_reg_state *r0;
9259 callee = state->frame[state->curframe];
9260 r0 = &callee->regs[BPF_REG_0];
9261 if (r0->type == PTR_TO_STACK) {
9262 /* technically it's ok to return caller's stack pointer
9263 * (or caller's caller's pointer) back to the caller,
9264 * since these pointers are valid. Only current stack
9265 * pointer will be invalid as soon as function exits,
9266 * but let's be conservative
9268 verbose(env, "cannot return stack pointer to the caller\n");
9272 caller = state->frame[state->curframe - 1];
9273 if (callee->in_callback_fn) {
9274 /* enforce R0 return value range [0, 1]. */
9275 struct tnum range = callee->callback_ret_range;
9277 if (r0->type != SCALAR_VALUE) {
9278 verbose(env, "R0 not a scalar value\n");
9282 /* we are going to rely on register's precise value */
9283 err = mark_reg_read(env, r0, r0->parent, REG_LIVE_READ64);
9284 err = err ?: mark_chain_precision(env, BPF_REG_0);
9288 if (!tnum_in(range, r0->var_off)) {
9289 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
9293 /* return to the caller whatever r0 had in the callee */
9294 caller->regs[BPF_REG_0] = *r0;
9297 /* callback_fn frame should have released its own additions to parent's
9298 * reference state at this point, or check_reference_leak would
9299 * complain, hence it must be the same as the caller. There is no need
9302 if (!callee->in_callback_fn) {
9303 /* Transfer references to the caller */
9304 err = copy_reference_state(caller, callee);
9309 *insn_idx = callee->callsite + 1;
9310 if (env->log.level & BPF_LOG_LEVEL) {
9311 verbose(env, "returning from callee:\n");
9312 print_verifier_state(env, callee, true);
9313 verbose(env, "to caller at %d:\n", *insn_idx);
9314 print_verifier_state(env, caller, true);
9316 /* clear everything in the callee */
9317 free_func_state(callee);
9318 state->frame[state->curframe--] = NULL;
9322 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
9324 struct bpf_call_arg_meta *meta)
9326 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
9328 if (ret_type != RET_INTEGER)
9332 case BPF_FUNC_get_stack:
9333 case BPF_FUNC_get_task_stack:
9334 case BPF_FUNC_probe_read_str:
9335 case BPF_FUNC_probe_read_kernel_str:
9336 case BPF_FUNC_probe_read_user_str:
9337 ret_reg->smax_value = meta->msize_max_value;
9338 ret_reg->s32_max_value = meta->msize_max_value;
9339 ret_reg->smin_value = -MAX_ERRNO;
9340 ret_reg->s32_min_value = -MAX_ERRNO;
9341 reg_bounds_sync(ret_reg);
9343 case BPF_FUNC_get_smp_processor_id:
9344 ret_reg->umax_value = nr_cpu_ids - 1;
9345 ret_reg->u32_max_value = nr_cpu_ids - 1;
9346 ret_reg->smax_value = nr_cpu_ids - 1;
9347 ret_reg->s32_max_value = nr_cpu_ids - 1;
9348 ret_reg->umin_value = 0;
9349 ret_reg->u32_min_value = 0;
9350 ret_reg->smin_value = 0;
9351 ret_reg->s32_min_value = 0;
9352 reg_bounds_sync(ret_reg);
9358 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
9359 int func_id, int insn_idx)
9361 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
9362 struct bpf_map *map = meta->map_ptr;
9364 if (func_id != BPF_FUNC_tail_call &&
9365 func_id != BPF_FUNC_map_lookup_elem &&
9366 func_id != BPF_FUNC_map_update_elem &&
9367 func_id != BPF_FUNC_map_delete_elem &&
9368 func_id != BPF_FUNC_map_push_elem &&
9369 func_id != BPF_FUNC_map_pop_elem &&
9370 func_id != BPF_FUNC_map_peek_elem &&
9371 func_id != BPF_FUNC_for_each_map_elem &&
9372 func_id != BPF_FUNC_redirect_map &&
9373 func_id != BPF_FUNC_map_lookup_percpu_elem)
9377 verbose(env, "kernel subsystem misconfigured verifier\n");
9381 /* In case of read-only, some additional restrictions
9382 * need to be applied in order to prevent altering the
9383 * state of the map from program side.
9385 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
9386 (func_id == BPF_FUNC_map_delete_elem ||
9387 func_id == BPF_FUNC_map_update_elem ||
9388 func_id == BPF_FUNC_map_push_elem ||
9389 func_id == BPF_FUNC_map_pop_elem)) {
9390 verbose(env, "write into map forbidden\n");
9394 if (!BPF_MAP_PTR(aux->map_ptr_state))
9395 bpf_map_ptr_store(aux, meta->map_ptr,
9396 !meta->map_ptr->bypass_spec_v1);
9397 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
9398 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
9399 !meta->map_ptr->bypass_spec_v1);
9404 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
9405 int func_id, int insn_idx)
9407 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
9408 struct bpf_reg_state *regs = cur_regs(env), *reg;
9409 struct bpf_map *map = meta->map_ptr;
9413 if (func_id != BPF_FUNC_tail_call)
9415 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
9416 verbose(env, "kernel subsystem misconfigured verifier\n");
9420 reg = ®s[BPF_REG_3];
9421 val = reg->var_off.value;
9422 max = map->max_entries;
9424 if (!(register_is_const(reg) && val < max)) {
9425 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
9429 err = mark_chain_precision(env, BPF_REG_3);
9432 if (bpf_map_key_unseen(aux))
9433 bpf_map_key_store(aux, val);
9434 else if (!bpf_map_key_poisoned(aux) &&
9435 bpf_map_key_immediate(aux) != val)
9436 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
9440 static int check_reference_leak(struct bpf_verifier_env *env)
9442 struct bpf_func_state *state = cur_func(env);
9443 bool refs_lingering = false;
9446 if (state->frameno && !state->in_callback_fn)
9449 for (i = 0; i < state->acquired_refs; i++) {
9450 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno)
9452 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
9453 state->refs[i].id, state->refs[i].insn_idx);
9454 refs_lingering = true;
9456 return refs_lingering ? -EINVAL : 0;
9459 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
9460 struct bpf_reg_state *regs)
9462 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
9463 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
9464 struct bpf_map *fmt_map = fmt_reg->map_ptr;
9465 struct bpf_bprintf_data data = {};
9466 int err, fmt_map_off, num_args;
9470 /* data must be an array of u64 */
9471 if (data_len_reg->var_off.value % 8)
9473 num_args = data_len_reg->var_off.value / 8;
9475 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
9476 * and map_direct_value_addr is set.
9478 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
9479 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
9482 verbose(env, "verifier bug\n");
9485 fmt = (char *)(long)fmt_addr + fmt_map_off;
9487 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
9488 * can focus on validating the format specifiers.
9490 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, num_args, &data);
9492 verbose(env, "Invalid format string\n");
9497 static int check_get_func_ip(struct bpf_verifier_env *env)
9499 enum bpf_prog_type type = resolve_prog_type(env->prog);
9500 int func_id = BPF_FUNC_get_func_ip;
9502 if (type == BPF_PROG_TYPE_TRACING) {
9503 if (!bpf_prog_has_trampoline(env->prog)) {
9504 verbose(env, "func %s#%d supported only for fentry/fexit/fmod_ret programs\n",
9505 func_id_name(func_id), func_id);
9509 } else if (type == BPF_PROG_TYPE_KPROBE) {
9513 verbose(env, "func %s#%d not supported for program type %d\n",
9514 func_id_name(func_id), func_id, type);
9518 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
9520 return &env->insn_aux_data[env->insn_idx];
9523 static bool loop_flag_is_zero(struct bpf_verifier_env *env)
9525 struct bpf_reg_state *regs = cur_regs(env);
9526 struct bpf_reg_state *reg = ®s[BPF_REG_4];
9527 bool reg_is_null = register_is_null(reg);
9530 mark_chain_precision(env, BPF_REG_4);
9535 static void update_loop_inline_state(struct bpf_verifier_env *env, u32 subprogno)
9537 struct bpf_loop_inline_state *state = &cur_aux(env)->loop_inline_state;
9539 if (!state->initialized) {
9540 state->initialized = 1;
9541 state->fit_for_inline = loop_flag_is_zero(env);
9542 state->callback_subprogno = subprogno;
9546 if (!state->fit_for_inline)
9549 state->fit_for_inline = (loop_flag_is_zero(env) &&
9550 state->callback_subprogno == subprogno);
9553 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
9556 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
9557 const struct bpf_func_proto *fn = NULL;
9558 enum bpf_return_type ret_type;
9559 enum bpf_type_flag ret_flag;
9560 struct bpf_reg_state *regs;
9561 struct bpf_call_arg_meta meta;
9562 int insn_idx = *insn_idx_p;
9564 int i, err, func_id;
9566 /* find function prototype */
9567 func_id = insn->imm;
9568 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
9569 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
9574 if (env->ops->get_func_proto)
9575 fn = env->ops->get_func_proto(func_id, env->prog);
9577 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
9582 /* eBPF programs must be GPL compatible to use GPL-ed functions */
9583 if (!env->prog->gpl_compatible && fn->gpl_only) {
9584 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
9588 if (fn->allowed && !fn->allowed(env->prog)) {
9589 verbose(env, "helper call is not allowed in probe\n");
9593 if (!env->prog->aux->sleepable && fn->might_sleep) {
9594 verbose(env, "helper call might sleep in a non-sleepable prog\n");
9598 /* With LD_ABS/IND some JITs save/restore skb from r1. */
9599 changes_data = bpf_helper_changes_pkt_data(fn->func);
9600 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
9601 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
9602 func_id_name(func_id), func_id);
9606 memset(&meta, 0, sizeof(meta));
9607 meta.pkt_access = fn->pkt_access;
9609 err = check_func_proto(fn, func_id);
9611 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
9612 func_id_name(func_id), func_id);
9616 if (env->cur_state->active_rcu_lock) {
9617 if (fn->might_sleep) {
9618 verbose(env, "sleepable helper %s#%d in rcu_read_lock region\n",
9619 func_id_name(func_id), func_id);
9623 if (env->prog->aux->sleepable && is_storage_get_function(func_id))
9624 env->insn_aux_data[insn_idx].storage_get_func_atomic = true;
9627 meta.func_id = func_id;
9629 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
9630 err = check_func_arg(env, i, &meta, fn, insn_idx);
9635 err = record_func_map(env, &meta, func_id, insn_idx);
9639 err = record_func_key(env, &meta, func_id, insn_idx);
9643 /* Mark slots with STACK_MISC in case of raw mode, stack offset
9644 * is inferred from register state.
9646 for (i = 0; i < meta.access_size; i++) {
9647 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
9648 BPF_WRITE, -1, false, false);
9653 regs = cur_regs(env);
9655 if (meta.release_regno) {
9657 /* This can only be set for PTR_TO_STACK, as CONST_PTR_TO_DYNPTR cannot
9658 * be released by any dynptr helper. Hence, unmark_stack_slots_dynptr
9659 * is safe to do directly.
9661 if (arg_type_is_dynptr(fn->arg_type[meta.release_regno - BPF_REG_1])) {
9662 if (regs[meta.release_regno].type == CONST_PTR_TO_DYNPTR) {
9663 verbose(env, "verifier internal error: CONST_PTR_TO_DYNPTR cannot be released\n");
9666 err = unmark_stack_slots_dynptr(env, ®s[meta.release_regno]);
9667 } else if (meta.ref_obj_id) {
9668 err = release_reference(env, meta.ref_obj_id);
9669 } else if (register_is_null(®s[meta.release_regno])) {
9670 /* meta.ref_obj_id can only be 0 if register that is meant to be
9671 * released is NULL, which must be > R0.
9676 verbose(env, "func %s#%d reference has not been acquired before\n",
9677 func_id_name(func_id), func_id);
9683 case BPF_FUNC_tail_call:
9684 err = check_reference_leak(env);
9686 verbose(env, "tail_call would lead to reference leak\n");
9690 case BPF_FUNC_get_local_storage:
9691 /* check that flags argument in get_local_storage(map, flags) is 0,
9692 * this is required because get_local_storage() can't return an error.
9694 if (!register_is_null(®s[BPF_REG_2])) {
9695 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
9699 case BPF_FUNC_for_each_map_elem:
9700 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9701 set_map_elem_callback_state);
9703 case BPF_FUNC_timer_set_callback:
9704 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9705 set_timer_callback_state);
9707 case BPF_FUNC_find_vma:
9708 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9709 set_find_vma_callback_state);
9711 case BPF_FUNC_snprintf:
9712 err = check_bpf_snprintf_call(env, regs);
9715 update_loop_inline_state(env, meta.subprogno);
9716 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9717 set_loop_callback_state);
9719 case BPF_FUNC_dynptr_from_mem:
9720 if (regs[BPF_REG_1].type != PTR_TO_MAP_VALUE) {
9721 verbose(env, "Unsupported reg type %s for bpf_dynptr_from_mem data\n",
9722 reg_type_str(env, regs[BPF_REG_1].type));
9726 case BPF_FUNC_set_retval:
9727 if (prog_type == BPF_PROG_TYPE_LSM &&
9728 env->prog->expected_attach_type == BPF_LSM_CGROUP) {
9729 if (!env->prog->aux->attach_func_proto->type) {
9730 /* Make sure programs that attach to void
9731 * hooks don't try to modify return value.
9733 verbose(env, "BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
9738 case BPF_FUNC_dynptr_data:
9740 struct bpf_reg_state *reg;
9743 reg = get_dynptr_arg_reg(env, fn, regs);
9748 if (meta.dynptr_id) {
9749 verbose(env, "verifier internal error: meta.dynptr_id already set\n");
9752 if (meta.ref_obj_id) {
9753 verbose(env, "verifier internal error: meta.ref_obj_id already set\n");
9757 id = dynptr_id(env, reg);
9759 verbose(env, "verifier internal error: failed to obtain dynptr id\n");
9763 ref_obj_id = dynptr_ref_obj_id(env, reg);
9764 if (ref_obj_id < 0) {
9765 verbose(env, "verifier internal error: failed to obtain dynptr ref_obj_id\n");
9769 meta.dynptr_id = id;
9770 meta.ref_obj_id = ref_obj_id;
9774 case BPF_FUNC_dynptr_write:
9776 enum bpf_dynptr_type dynptr_type;
9777 struct bpf_reg_state *reg;
9779 reg = get_dynptr_arg_reg(env, fn, regs);
9783 dynptr_type = dynptr_get_type(env, reg);
9784 if (dynptr_type == BPF_DYNPTR_TYPE_INVALID)
9787 if (dynptr_type == BPF_DYNPTR_TYPE_SKB)
9788 /* this will trigger clear_all_pkt_pointers(), which will
9789 * invalidate all dynptr slices associated with the skb
9791 changes_data = true;
9795 case BPF_FUNC_user_ringbuf_drain:
9796 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9797 set_user_ringbuf_callback_state);
9804 /* reset caller saved regs */
9805 for (i = 0; i < CALLER_SAVED_REGS; i++) {
9806 mark_reg_not_init(env, regs, caller_saved[i]);
9807 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
9810 /* helper call returns 64-bit value. */
9811 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
9813 /* update return register (already marked as written above) */
9814 ret_type = fn->ret_type;
9815 ret_flag = type_flag(ret_type);
9817 switch (base_type(ret_type)) {
9819 /* sets type to SCALAR_VALUE */
9820 mark_reg_unknown(env, regs, BPF_REG_0);
9823 regs[BPF_REG_0].type = NOT_INIT;
9825 case RET_PTR_TO_MAP_VALUE:
9826 /* There is no offset yet applied, variable or fixed */
9827 mark_reg_known_zero(env, regs, BPF_REG_0);
9828 /* remember map_ptr, so that check_map_access()
9829 * can check 'value_size' boundary of memory access
9830 * to map element returned from bpf_map_lookup_elem()
9832 if (meta.map_ptr == NULL) {
9834 "kernel subsystem misconfigured verifier\n");
9837 regs[BPF_REG_0].map_ptr = meta.map_ptr;
9838 regs[BPF_REG_0].map_uid = meta.map_uid;
9839 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE | ret_flag;
9840 if (!type_may_be_null(ret_type) &&
9841 btf_record_has_field(meta.map_ptr->record, BPF_SPIN_LOCK)) {
9842 regs[BPF_REG_0].id = ++env->id_gen;
9845 case RET_PTR_TO_SOCKET:
9846 mark_reg_known_zero(env, regs, BPF_REG_0);
9847 regs[BPF_REG_0].type = PTR_TO_SOCKET | ret_flag;
9849 case RET_PTR_TO_SOCK_COMMON:
9850 mark_reg_known_zero(env, regs, BPF_REG_0);
9851 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON | ret_flag;
9853 case RET_PTR_TO_TCP_SOCK:
9854 mark_reg_known_zero(env, regs, BPF_REG_0);
9855 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK | ret_flag;
9857 case RET_PTR_TO_MEM:
9858 mark_reg_known_zero(env, regs, BPF_REG_0);
9859 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
9860 regs[BPF_REG_0].mem_size = meta.mem_size;
9862 case RET_PTR_TO_MEM_OR_BTF_ID:
9864 const struct btf_type *t;
9866 mark_reg_known_zero(env, regs, BPF_REG_0);
9867 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
9868 if (!btf_type_is_struct(t)) {
9870 const struct btf_type *ret;
9873 /* resolve the type size of ksym. */
9874 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
9876 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
9877 verbose(env, "unable to resolve the size of type '%s': %ld\n",
9878 tname, PTR_ERR(ret));
9881 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
9882 regs[BPF_REG_0].mem_size = tsize;
9884 /* MEM_RDONLY may be carried from ret_flag, but it
9885 * doesn't apply on PTR_TO_BTF_ID. Fold it, otherwise
9886 * it will confuse the check of PTR_TO_BTF_ID in
9887 * check_mem_access().
9889 ret_flag &= ~MEM_RDONLY;
9891 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
9892 regs[BPF_REG_0].btf = meta.ret_btf;
9893 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
9897 case RET_PTR_TO_BTF_ID:
9899 struct btf *ret_btf;
9902 mark_reg_known_zero(env, regs, BPF_REG_0);
9903 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
9904 if (func_id == BPF_FUNC_kptr_xchg) {
9905 ret_btf = meta.kptr_field->kptr.btf;
9906 ret_btf_id = meta.kptr_field->kptr.btf_id;
9907 if (!btf_is_kernel(ret_btf))
9908 regs[BPF_REG_0].type |= MEM_ALLOC;
9910 if (fn->ret_btf_id == BPF_PTR_POISON) {
9911 verbose(env, "verifier internal error:");
9912 verbose(env, "func %s has non-overwritten BPF_PTR_POISON return type\n",
9913 func_id_name(func_id));
9916 ret_btf = btf_vmlinux;
9917 ret_btf_id = *fn->ret_btf_id;
9919 if (ret_btf_id == 0) {
9920 verbose(env, "invalid return type %u of func %s#%d\n",
9921 base_type(ret_type), func_id_name(func_id),
9925 regs[BPF_REG_0].btf = ret_btf;
9926 regs[BPF_REG_0].btf_id = ret_btf_id;
9930 verbose(env, "unknown return type %u of func %s#%d\n",
9931 base_type(ret_type), func_id_name(func_id), func_id);
9935 if (type_may_be_null(regs[BPF_REG_0].type))
9936 regs[BPF_REG_0].id = ++env->id_gen;
9938 if (helper_multiple_ref_obj_use(func_id, meta.map_ptr)) {
9939 verbose(env, "verifier internal error: func %s#%d sets ref_obj_id more than once\n",
9940 func_id_name(func_id), func_id);
9944 if (is_dynptr_ref_function(func_id))
9945 regs[BPF_REG_0].dynptr_id = meta.dynptr_id;
9947 if (is_ptr_cast_function(func_id) || is_dynptr_ref_function(func_id)) {
9948 /* For release_reference() */
9949 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
9950 } else if (is_acquire_function(func_id, meta.map_ptr)) {
9951 int id = acquire_reference_state(env, insn_idx);
9955 /* For mark_ptr_or_null_reg() */
9956 regs[BPF_REG_0].id = id;
9957 /* For release_reference() */
9958 regs[BPF_REG_0].ref_obj_id = id;
9961 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
9963 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
9967 if ((func_id == BPF_FUNC_get_stack ||
9968 func_id == BPF_FUNC_get_task_stack) &&
9969 !env->prog->has_callchain_buf) {
9970 const char *err_str;
9972 #ifdef CONFIG_PERF_EVENTS
9973 err = get_callchain_buffers(sysctl_perf_event_max_stack);
9974 err_str = "cannot get callchain buffer for func %s#%d\n";
9977 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
9980 verbose(env, err_str, func_id_name(func_id), func_id);
9984 env->prog->has_callchain_buf = true;
9987 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
9988 env->prog->call_get_stack = true;
9990 if (func_id == BPF_FUNC_get_func_ip) {
9991 if (check_get_func_ip(env))
9993 env->prog->call_get_func_ip = true;
9997 clear_all_pkt_pointers(env);
10001 /* mark_btf_func_reg_size() is used when the reg size is determined by
10002 * the BTF func_proto's return value size and argument.
10004 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
10007 struct bpf_reg_state *reg = &cur_regs(env)[regno];
10009 if (regno == BPF_REG_0) {
10010 /* Function return value */
10011 reg->live |= REG_LIVE_WRITTEN;
10012 reg->subreg_def = reg_size == sizeof(u64) ?
10013 DEF_NOT_SUBREG : env->insn_idx + 1;
10015 /* Function argument */
10016 if (reg_size == sizeof(u64)) {
10017 mark_insn_zext(env, reg);
10018 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
10020 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
10025 static bool is_kfunc_acquire(struct bpf_kfunc_call_arg_meta *meta)
10027 return meta->kfunc_flags & KF_ACQUIRE;
10030 static bool is_kfunc_release(struct bpf_kfunc_call_arg_meta *meta)
10032 return meta->kfunc_flags & KF_RELEASE;
10035 static bool is_kfunc_trusted_args(struct bpf_kfunc_call_arg_meta *meta)
10037 return (meta->kfunc_flags & KF_TRUSTED_ARGS) || is_kfunc_release(meta);
10040 static bool is_kfunc_sleepable(struct bpf_kfunc_call_arg_meta *meta)
10042 return meta->kfunc_flags & KF_SLEEPABLE;
10045 static bool is_kfunc_destructive(struct bpf_kfunc_call_arg_meta *meta)
10047 return meta->kfunc_flags & KF_DESTRUCTIVE;
10050 static bool is_kfunc_rcu(struct bpf_kfunc_call_arg_meta *meta)
10052 return meta->kfunc_flags & KF_RCU;
10055 static bool __kfunc_param_match_suffix(const struct btf *btf,
10056 const struct btf_param *arg,
10057 const char *suffix)
10059 int suffix_len = strlen(suffix), len;
10060 const char *param_name;
10062 /* In the future, this can be ported to use BTF tagging */
10063 param_name = btf_name_by_offset(btf, arg->name_off);
10064 if (str_is_empty(param_name))
10066 len = strlen(param_name);
10067 if (len < suffix_len)
10069 param_name += len - suffix_len;
10070 return !strncmp(param_name, suffix, suffix_len);
10073 static bool is_kfunc_arg_mem_size(const struct btf *btf,
10074 const struct btf_param *arg,
10075 const struct bpf_reg_state *reg)
10077 const struct btf_type *t;
10079 t = btf_type_skip_modifiers(btf, arg->type, NULL);
10080 if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
10083 return __kfunc_param_match_suffix(btf, arg, "__sz");
10086 static bool is_kfunc_arg_const_mem_size(const struct btf *btf,
10087 const struct btf_param *arg,
10088 const struct bpf_reg_state *reg)
10090 const struct btf_type *t;
10092 t = btf_type_skip_modifiers(btf, arg->type, NULL);
10093 if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
10096 return __kfunc_param_match_suffix(btf, arg, "__szk");
10099 static bool is_kfunc_arg_optional(const struct btf *btf, const struct btf_param *arg)
10101 return __kfunc_param_match_suffix(btf, arg, "__opt");
10104 static bool is_kfunc_arg_constant(const struct btf *btf, const struct btf_param *arg)
10106 return __kfunc_param_match_suffix(btf, arg, "__k");
10109 static bool is_kfunc_arg_ignore(const struct btf *btf, const struct btf_param *arg)
10111 return __kfunc_param_match_suffix(btf, arg, "__ign");
10114 static bool is_kfunc_arg_alloc_obj(const struct btf *btf, const struct btf_param *arg)
10116 return __kfunc_param_match_suffix(btf, arg, "__alloc");
10119 static bool is_kfunc_arg_uninit(const struct btf *btf, const struct btf_param *arg)
10121 return __kfunc_param_match_suffix(btf, arg, "__uninit");
10124 static bool is_kfunc_arg_refcounted_kptr(const struct btf *btf, const struct btf_param *arg)
10126 return __kfunc_param_match_suffix(btf, arg, "__refcounted_kptr");
10129 static bool is_kfunc_arg_scalar_with_name(const struct btf *btf,
10130 const struct btf_param *arg,
10133 int len, target_len = strlen(name);
10134 const char *param_name;
10136 param_name = btf_name_by_offset(btf, arg->name_off);
10137 if (str_is_empty(param_name))
10139 len = strlen(param_name);
10140 if (len != target_len)
10142 if (strcmp(param_name, name))
10150 KF_ARG_LIST_HEAD_ID,
10151 KF_ARG_LIST_NODE_ID,
10156 BTF_ID_LIST(kf_arg_btf_ids)
10157 BTF_ID(struct, bpf_dynptr_kern)
10158 BTF_ID(struct, bpf_list_head)
10159 BTF_ID(struct, bpf_list_node)
10160 BTF_ID(struct, bpf_rb_root)
10161 BTF_ID(struct, bpf_rb_node)
10163 static bool __is_kfunc_ptr_arg_type(const struct btf *btf,
10164 const struct btf_param *arg, int type)
10166 const struct btf_type *t;
10169 t = btf_type_skip_modifiers(btf, arg->type, NULL);
10172 if (!btf_type_is_ptr(t))
10174 t = btf_type_skip_modifiers(btf, t->type, &res_id);
10177 return btf_types_are_same(btf, res_id, btf_vmlinux, kf_arg_btf_ids[type]);
10180 static bool is_kfunc_arg_dynptr(const struct btf *btf, const struct btf_param *arg)
10182 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_DYNPTR_ID);
10185 static bool is_kfunc_arg_list_head(const struct btf *btf, const struct btf_param *arg)
10187 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_HEAD_ID);
10190 static bool is_kfunc_arg_list_node(const struct btf *btf, const struct btf_param *arg)
10192 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_NODE_ID);
10195 static bool is_kfunc_arg_rbtree_root(const struct btf *btf, const struct btf_param *arg)
10197 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RB_ROOT_ID);
10200 static bool is_kfunc_arg_rbtree_node(const struct btf *btf, const struct btf_param *arg)
10202 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RB_NODE_ID);
10205 static bool is_kfunc_arg_callback(struct bpf_verifier_env *env, const struct btf *btf,
10206 const struct btf_param *arg)
10208 const struct btf_type *t;
10210 t = btf_type_resolve_func_ptr(btf, arg->type, NULL);
10217 /* Returns true if struct is composed of scalars, 4 levels of nesting allowed */
10218 static bool __btf_type_is_scalar_struct(struct bpf_verifier_env *env,
10219 const struct btf *btf,
10220 const struct btf_type *t, int rec)
10222 const struct btf_type *member_type;
10223 const struct btf_member *member;
10226 if (!btf_type_is_struct(t))
10229 for_each_member(i, t, member) {
10230 const struct btf_array *array;
10232 member_type = btf_type_skip_modifiers(btf, member->type, NULL);
10233 if (btf_type_is_struct(member_type)) {
10235 verbose(env, "max struct nesting depth exceeded\n");
10238 if (!__btf_type_is_scalar_struct(env, btf, member_type, rec + 1))
10242 if (btf_type_is_array(member_type)) {
10243 array = btf_array(member_type);
10244 if (!array->nelems)
10246 member_type = btf_type_skip_modifiers(btf, array->type, NULL);
10247 if (!btf_type_is_scalar(member_type))
10251 if (!btf_type_is_scalar(member_type))
10257 enum kfunc_ptr_arg_type {
10259 KF_ARG_PTR_TO_ALLOC_BTF_ID, /* Allocated object */
10260 KF_ARG_PTR_TO_REFCOUNTED_KPTR, /* Refcounted local kptr */
10261 KF_ARG_PTR_TO_DYNPTR,
10262 KF_ARG_PTR_TO_ITER,
10263 KF_ARG_PTR_TO_LIST_HEAD,
10264 KF_ARG_PTR_TO_LIST_NODE,
10265 KF_ARG_PTR_TO_BTF_ID, /* Also covers reg2btf_ids conversions */
10267 KF_ARG_PTR_TO_MEM_SIZE, /* Size derived from next argument, skip it */
10268 KF_ARG_PTR_TO_CALLBACK,
10269 KF_ARG_PTR_TO_RB_ROOT,
10270 KF_ARG_PTR_TO_RB_NODE,
10273 enum special_kfunc_type {
10274 KF_bpf_obj_new_impl,
10275 KF_bpf_obj_drop_impl,
10276 KF_bpf_refcount_acquire_impl,
10277 KF_bpf_list_push_front_impl,
10278 KF_bpf_list_push_back_impl,
10279 KF_bpf_list_pop_front,
10280 KF_bpf_list_pop_back,
10281 KF_bpf_cast_to_kern_ctx,
10282 KF_bpf_rdonly_cast,
10283 KF_bpf_rcu_read_lock,
10284 KF_bpf_rcu_read_unlock,
10285 KF_bpf_rbtree_remove,
10286 KF_bpf_rbtree_add_impl,
10287 KF_bpf_rbtree_first,
10288 KF_bpf_dynptr_from_skb,
10289 KF_bpf_dynptr_from_xdp,
10290 KF_bpf_dynptr_slice,
10291 KF_bpf_dynptr_slice_rdwr,
10292 KF_bpf_dynptr_clone,
10295 BTF_SET_START(special_kfunc_set)
10296 BTF_ID(func, bpf_obj_new_impl)
10297 BTF_ID(func, bpf_obj_drop_impl)
10298 BTF_ID(func, bpf_refcount_acquire_impl)
10299 BTF_ID(func, bpf_list_push_front_impl)
10300 BTF_ID(func, bpf_list_push_back_impl)
10301 BTF_ID(func, bpf_list_pop_front)
10302 BTF_ID(func, bpf_list_pop_back)
10303 BTF_ID(func, bpf_cast_to_kern_ctx)
10304 BTF_ID(func, bpf_rdonly_cast)
10305 BTF_ID(func, bpf_rbtree_remove)
10306 BTF_ID(func, bpf_rbtree_add_impl)
10307 BTF_ID(func, bpf_rbtree_first)
10308 BTF_ID(func, bpf_dynptr_from_skb)
10309 BTF_ID(func, bpf_dynptr_from_xdp)
10310 BTF_ID(func, bpf_dynptr_slice)
10311 BTF_ID(func, bpf_dynptr_slice_rdwr)
10312 BTF_ID(func, bpf_dynptr_clone)
10313 BTF_SET_END(special_kfunc_set)
10315 BTF_ID_LIST(special_kfunc_list)
10316 BTF_ID(func, bpf_obj_new_impl)
10317 BTF_ID(func, bpf_obj_drop_impl)
10318 BTF_ID(func, bpf_refcount_acquire_impl)
10319 BTF_ID(func, bpf_list_push_front_impl)
10320 BTF_ID(func, bpf_list_push_back_impl)
10321 BTF_ID(func, bpf_list_pop_front)
10322 BTF_ID(func, bpf_list_pop_back)
10323 BTF_ID(func, bpf_cast_to_kern_ctx)
10324 BTF_ID(func, bpf_rdonly_cast)
10325 BTF_ID(func, bpf_rcu_read_lock)
10326 BTF_ID(func, bpf_rcu_read_unlock)
10327 BTF_ID(func, bpf_rbtree_remove)
10328 BTF_ID(func, bpf_rbtree_add_impl)
10329 BTF_ID(func, bpf_rbtree_first)
10330 BTF_ID(func, bpf_dynptr_from_skb)
10331 BTF_ID(func, bpf_dynptr_from_xdp)
10332 BTF_ID(func, bpf_dynptr_slice)
10333 BTF_ID(func, bpf_dynptr_slice_rdwr)
10334 BTF_ID(func, bpf_dynptr_clone)
10336 static bool is_kfunc_ret_null(struct bpf_kfunc_call_arg_meta *meta)
10338 if (meta->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl] &&
10339 meta->arg_owning_ref) {
10343 return meta->kfunc_flags & KF_RET_NULL;
10346 static bool is_kfunc_bpf_rcu_read_lock(struct bpf_kfunc_call_arg_meta *meta)
10348 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_lock];
10351 static bool is_kfunc_bpf_rcu_read_unlock(struct bpf_kfunc_call_arg_meta *meta)
10353 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_unlock];
10356 static enum kfunc_ptr_arg_type
10357 get_kfunc_ptr_arg_type(struct bpf_verifier_env *env,
10358 struct bpf_kfunc_call_arg_meta *meta,
10359 const struct btf_type *t, const struct btf_type *ref_t,
10360 const char *ref_tname, const struct btf_param *args,
10361 int argno, int nargs)
10363 u32 regno = argno + 1;
10364 struct bpf_reg_state *regs = cur_regs(env);
10365 struct bpf_reg_state *reg = ®s[regno];
10366 bool arg_mem_size = false;
10368 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx])
10369 return KF_ARG_PTR_TO_CTX;
10371 /* In this function, we verify the kfunc's BTF as per the argument type,
10372 * leaving the rest of the verification with respect to the register
10373 * type to our caller. When a set of conditions hold in the BTF type of
10374 * arguments, we resolve it to a known kfunc_ptr_arg_type.
10376 if (btf_get_prog_ctx_type(&env->log, meta->btf, t, resolve_prog_type(env->prog), argno))
10377 return KF_ARG_PTR_TO_CTX;
10379 if (is_kfunc_arg_alloc_obj(meta->btf, &args[argno]))
10380 return KF_ARG_PTR_TO_ALLOC_BTF_ID;
10382 if (is_kfunc_arg_refcounted_kptr(meta->btf, &args[argno]))
10383 return KF_ARG_PTR_TO_REFCOUNTED_KPTR;
10385 if (is_kfunc_arg_dynptr(meta->btf, &args[argno]))
10386 return KF_ARG_PTR_TO_DYNPTR;
10388 if (is_kfunc_arg_iter(meta, argno))
10389 return KF_ARG_PTR_TO_ITER;
10391 if (is_kfunc_arg_list_head(meta->btf, &args[argno]))
10392 return KF_ARG_PTR_TO_LIST_HEAD;
10394 if (is_kfunc_arg_list_node(meta->btf, &args[argno]))
10395 return KF_ARG_PTR_TO_LIST_NODE;
10397 if (is_kfunc_arg_rbtree_root(meta->btf, &args[argno]))
10398 return KF_ARG_PTR_TO_RB_ROOT;
10400 if (is_kfunc_arg_rbtree_node(meta->btf, &args[argno]))
10401 return KF_ARG_PTR_TO_RB_NODE;
10403 if ((base_type(reg->type) == PTR_TO_BTF_ID || reg2btf_ids[base_type(reg->type)])) {
10404 if (!btf_type_is_struct(ref_t)) {
10405 verbose(env, "kernel function %s args#%d pointer type %s %s is not supported\n",
10406 meta->func_name, argno, btf_type_str(ref_t), ref_tname);
10409 return KF_ARG_PTR_TO_BTF_ID;
10412 if (is_kfunc_arg_callback(env, meta->btf, &args[argno]))
10413 return KF_ARG_PTR_TO_CALLBACK;
10416 if (argno + 1 < nargs &&
10417 (is_kfunc_arg_mem_size(meta->btf, &args[argno + 1], ®s[regno + 1]) ||
10418 is_kfunc_arg_const_mem_size(meta->btf, &args[argno + 1], ®s[regno + 1])))
10419 arg_mem_size = true;
10421 /* This is the catch all argument type of register types supported by
10422 * check_helper_mem_access. However, we only allow when argument type is
10423 * pointer to scalar, or struct composed (recursively) of scalars. When
10424 * arg_mem_size is true, the pointer can be void *.
10426 if (!btf_type_is_scalar(ref_t) && !__btf_type_is_scalar_struct(env, meta->btf, ref_t, 0) &&
10427 (arg_mem_size ? !btf_type_is_void(ref_t) : 1)) {
10428 verbose(env, "arg#%d pointer type %s %s must point to %sscalar, or struct with scalar\n",
10429 argno, btf_type_str(ref_t), ref_tname, arg_mem_size ? "void, " : "");
10432 return arg_mem_size ? KF_ARG_PTR_TO_MEM_SIZE : KF_ARG_PTR_TO_MEM;
10435 static int process_kf_arg_ptr_to_btf_id(struct bpf_verifier_env *env,
10436 struct bpf_reg_state *reg,
10437 const struct btf_type *ref_t,
10438 const char *ref_tname, u32 ref_id,
10439 struct bpf_kfunc_call_arg_meta *meta,
10442 const struct btf_type *reg_ref_t;
10443 bool strict_type_match = false;
10444 const struct btf *reg_btf;
10445 const char *reg_ref_tname;
10448 if (base_type(reg->type) == PTR_TO_BTF_ID) {
10449 reg_btf = reg->btf;
10450 reg_ref_id = reg->btf_id;
10452 reg_btf = btf_vmlinux;
10453 reg_ref_id = *reg2btf_ids[base_type(reg->type)];
10456 /* Enforce strict type matching for calls to kfuncs that are acquiring
10457 * or releasing a reference, or are no-cast aliases. We do _not_
10458 * enforce strict matching for plain KF_TRUSTED_ARGS kfuncs by default,
10459 * as we want to enable BPF programs to pass types that are bitwise
10460 * equivalent without forcing them to explicitly cast with something
10461 * like bpf_cast_to_kern_ctx().
10463 * For example, say we had a type like the following:
10465 * struct bpf_cpumask {
10466 * cpumask_t cpumask;
10467 * refcount_t usage;
10470 * Note that as specified in <linux/cpumask.h>, cpumask_t is typedef'ed
10471 * to a struct cpumask, so it would be safe to pass a struct
10472 * bpf_cpumask * to a kfunc expecting a struct cpumask *.
10474 * The philosophy here is similar to how we allow scalars of different
10475 * types to be passed to kfuncs as long as the size is the same. The
10476 * only difference here is that we're simply allowing
10477 * btf_struct_ids_match() to walk the struct at the 0th offset, and
10480 if (is_kfunc_acquire(meta) ||
10481 (is_kfunc_release(meta) && reg->ref_obj_id) ||
10482 btf_type_ids_nocast_alias(&env->log, reg_btf, reg_ref_id, meta->btf, ref_id))
10483 strict_type_match = true;
10485 WARN_ON_ONCE(is_kfunc_trusted_args(meta) && reg->off);
10487 reg_ref_t = btf_type_skip_modifiers(reg_btf, reg_ref_id, ®_ref_id);
10488 reg_ref_tname = btf_name_by_offset(reg_btf, reg_ref_t->name_off);
10489 if (!btf_struct_ids_match(&env->log, reg_btf, reg_ref_id, reg->off, meta->btf, ref_id, strict_type_match)) {
10490 verbose(env, "kernel function %s args#%d expected pointer to %s %s but R%d has a pointer to %s %s\n",
10491 meta->func_name, argno, btf_type_str(ref_t), ref_tname, argno + 1,
10492 btf_type_str(reg_ref_t), reg_ref_tname);
10498 static int ref_set_non_owning(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
10500 struct bpf_verifier_state *state = env->cur_state;
10501 struct btf_record *rec = reg_btf_record(reg);
10503 if (!state->active_lock.ptr) {
10504 verbose(env, "verifier internal error: ref_set_non_owning w/o active lock\n");
10508 if (type_flag(reg->type) & NON_OWN_REF) {
10509 verbose(env, "verifier internal error: NON_OWN_REF already set\n");
10513 reg->type |= NON_OWN_REF;
10514 if (rec->refcount_off >= 0)
10515 reg->type |= MEM_RCU;
10520 static int ref_convert_owning_non_owning(struct bpf_verifier_env *env, u32 ref_obj_id)
10522 struct bpf_func_state *state, *unused;
10523 struct bpf_reg_state *reg;
10526 state = cur_func(env);
10529 verbose(env, "verifier internal error: ref_obj_id is zero for "
10530 "owning -> non-owning conversion\n");
10534 for (i = 0; i < state->acquired_refs; i++) {
10535 if (state->refs[i].id != ref_obj_id)
10538 /* Clear ref_obj_id here so release_reference doesn't clobber
10541 bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
10542 if (reg->ref_obj_id == ref_obj_id) {
10543 reg->ref_obj_id = 0;
10544 ref_set_non_owning(env, reg);
10550 verbose(env, "verifier internal error: ref state missing for ref_obj_id\n");
10554 /* Implementation details:
10556 * Each register points to some region of memory, which we define as an
10557 * allocation. Each allocation may embed a bpf_spin_lock which protects any
10558 * special BPF objects (bpf_list_head, bpf_rb_root, etc.) part of the same
10559 * allocation. The lock and the data it protects are colocated in the same
10562 * Hence, everytime a register holds a pointer value pointing to such
10563 * allocation, the verifier preserves a unique reg->id for it.
10565 * The verifier remembers the lock 'ptr' and the lock 'id' whenever
10566 * bpf_spin_lock is called.
10568 * To enable this, lock state in the verifier captures two values:
10569 * active_lock.ptr = Register's type specific pointer
10570 * active_lock.id = A unique ID for each register pointer value
10572 * Currently, PTR_TO_MAP_VALUE and PTR_TO_BTF_ID | MEM_ALLOC are the two
10573 * supported register types.
10575 * The active_lock.ptr in case of map values is the reg->map_ptr, and in case of
10576 * allocated objects is the reg->btf pointer.
10578 * The active_lock.id is non-unique for maps supporting direct_value_addr, as we
10579 * can establish the provenance of the map value statically for each distinct
10580 * lookup into such maps. They always contain a single map value hence unique
10581 * IDs for each pseudo load pessimizes the algorithm and rejects valid programs.
10583 * So, in case of global variables, they use array maps with max_entries = 1,
10584 * hence their active_lock.ptr becomes map_ptr and id = 0 (since they all point
10585 * into the same map value as max_entries is 1, as described above).
10587 * In case of inner map lookups, the inner map pointer has same map_ptr as the
10588 * outer map pointer (in verifier context), but each lookup into an inner map
10589 * assigns a fresh reg->id to the lookup, so while lookups into distinct inner
10590 * maps from the same outer map share the same map_ptr as active_lock.ptr, they
10591 * will get different reg->id assigned to each lookup, hence different
10594 * In case of allocated objects, active_lock.ptr is the reg->btf, and the
10595 * reg->id is a unique ID preserved after the NULL pointer check on the pointer
10596 * returned from bpf_obj_new. Each allocation receives a new reg->id.
10598 static int check_reg_allocation_locked(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
10603 switch ((int)reg->type) {
10604 case PTR_TO_MAP_VALUE:
10605 ptr = reg->map_ptr;
10607 case PTR_TO_BTF_ID | MEM_ALLOC:
10611 verbose(env, "verifier internal error: unknown reg type for lock check\n");
10616 if (!env->cur_state->active_lock.ptr)
10618 if (env->cur_state->active_lock.ptr != ptr ||
10619 env->cur_state->active_lock.id != id) {
10620 verbose(env, "held lock and object are not in the same allocation\n");
10626 static bool is_bpf_list_api_kfunc(u32 btf_id)
10628 return btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
10629 btf_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
10630 btf_id == special_kfunc_list[KF_bpf_list_pop_front] ||
10631 btf_id == special_kfunc_list[KF_bpf_list_pop_back];
10634 static bool is_bpf_rbtree_api_kfunc(u32 btf_id)
10636 return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl] ||
10637 btf_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
10638 btf_id == special_kfunc_list[KF_bpf_rbtree_first];
10641 static bool is_bpf_graph_api_kfunc(u32 btf_id)
10643 return is_bpf_list_api_kfunc(btf_id) || is_bpf_rbtree_api_kfunc(btf_id) ||
10644 btf_id == special_kfunc_list[KF_bpf_refcount_acquire_impl];
10647 static bool is_callback_calling_kfunc(u32 btf_id)
10649 return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl];
10652 static bool is_rbtree_lock_required_kfunc(u32 btf_id)
10654 return is_bpf_rbtree_api_kfunc(btf_id);
10657 static bool check_kfunc_is_graph_root_api(struct bpf_verifier_env *env,
10658 enum btf_field_type head_field_type,
10663 switch (head_field_type) {
10664 case BPF_LIST_HEAD:
10665 ret = is_bpf_list_api_kfunc(kfunc_btf_id);
10668 ret = is_bpf_rbtree_api_kfunc(kfunc_btf_id);
10671 verbose(env, "verifier internal error: unexpected graph root argument type %s\n",
10672 btf_field_type_name(head_field_type));
10677 verbose(env, "verifier internal error: %s head arg for unknown kfunc\n",
10678 btf_field_type_name(head_field_type));
10682 static bool check_kfunc_is_graph_node_api(struct bpf_verifier_env *env,
10683 enum btf_field_type node_field_type,
10688 switch (node_field_type) {
10689 case BPF_LIST_NODE:
10690 ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
10691 kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_back_impl]);
10694 ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
10695 kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl]);
10698 verbose(env, "verifier internal error: unexpected graph node argument type %s\n",
10699 btf_field_type_name(node_field_type));
10704 verbose(env, "verifier internal error: %s node arg for unknown kfunc\n",
10705 btf_field_type_name(node_field_type));
10710 __process_kf_arg_ptr_to_graph_root(struct bpf_verifier_env *env,
10711 struct bpf_reg_state *reg, u32 regno,
10712 struct bpf_kfunc_call_arg_meta *meta,
10713 enum btf_field_type head_field_type,
10714 struct btf_field **head_field)
10716 const char *head_type_name;
10717 struct btf_field *field;
10718 struct btf_record *rec;
10721 if (meta->btf != btf_vmlinux) {
10722 verbose(env, "verifier internal error: unexpected btf mismatch in kfunc call\n");
10726 if (!check_kfunc_is_graph_root_api(env, head_field_type, meta->func_id))
10729 head_type_name = btf_field_type_name(head_field_type);
10730 if (!tnum_is_const(reg->var_off)) {
10732 "R%d doesn't have constant offset. %s has to be at the constant offset\n",
10733 regno, head_type_name);
10737 rec = reg_btf_record(reg);
10738 head_off = reg->off + reg->var_off.value;
10739 field = btf_record_find(rec, head_off, head_field_type);
10741 verbose(env, "%s not found at offset=%u\n", head_type_name, head_off);
10745 /* All functions require bpf_list_head to be protected using a bpf_spin_lock */
10746 if (check_reg_allocation_locked(env, reg)) {
10747 verbose(env, "bpf_spin_lock at off=%d must be held for %s\n",
10748 rec->spin_lock_off, head_type_name);
10753 verbose(env, "verifier internal error: repeating %s arg\n", head_type_name);
10756 *head_field = field;
10760 static int process_kf_arg_ptr_to_list_head(struct bpf_verifier_env *env,
10761 struct bpf_reg_state *reg, u32 regno,
10762 struct bpf_kfunc_call_arg_meta *meta)
10764 return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, BPF_LIST_HEAD,
10765 &meta->arg_list_head.field);
10768 static int process_kf_arg_ptr_to_rbtree_root(struct bpf_verifier_env *env,
10769 struct bpf_reg_state *reg, u32 regno,
10770 struct bpf_kfunc_call_arg_meta *meta)
10772 return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, BPF_RB_ROOT,
10773 &meta->arg_rbtree_root.field);
10777 __process_kf_arg_ptr_to_graph_node(struct bpf_verifier_env *env,
10778 struct bpf_reg_state *reg, u32 regno,
10779 struct bpf_kfunc_call_arg_meta *meta,
10780 enum btf_field_type head_field_type,
10781 enum btf_field_type node_field_type,
10782 struct btf_field **node_field)
10784 const char *node_type_name;
10785 const struct btf_type *et, *t;
10786 struct btf_field *field;
10789 if (meta->btf != btf_vmlinux) {
10790 verbose(env, "verifier internal error: unexpected btf mismatch in kfunc call\n");
10794 if (!check_kfunc_is_graph_node_api(env, node_field_type, meta->func_id))
10797 node_type_name = btf_field_type_name(node_field_type);
10798 if (!tnum_is_const(reg->var_off)) {
10800 "R%d doesn't have constant offset. %s has to be at the constant offset\n",
10801 regno, node_type_name);
10805 node_off = reg->off + reg->var_off.value;
10806 field = reg_find_field_offset(reg, node_off, node_field_type);
10807 if (!field || field->offset != node_off) {
10808 verbose(env, "%s not found at offset=%u\n", node_type_name, node_off);
10812 field = *node_field;
10814 et = btf_type_by_id(field->graph_root.btf, field->graph_root.value_btf_id);
10815 t = btf_type_by_id(reg->btf, reg->btf_id);
10816 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, 0, field->graph_root.btf,
10817 field->graph_root.value_btf_id, true)) {
10818 verbose(env, "operation on %s expects arg#1 %s at offset=%d "
10819 "in struct %s, but arg is at offset=%d in struct %s\n",
10820 btf_field_type_name(head_field_type),
10821 btf_field_type_name(node_field_type),
10822 field->graph_root.node_offset,
10823 btf_name_by_offset(field->graph_root.btf, et->name_off),
10824 node_off, btf_name_by_offset(reg->btf, t->name_off));
10827 meta->arg_btf = reg->btf;
10828 meta->arg_btf_id = reg->btf_id;
10830 if (node_off != field->graph_root.node_offset) {
10831 verbose(env, "arg#1 offset=%d, but expected %s at offset=%d in struct %s\n",
10832 node_off, btf_field_type_name(node_field_type),
10833 field->graph_root.node_offset,
10834 btf_name_by_offset(field->graph_root.btf, et->name_off));
10841 static int process_kf_arg_ptr_to_list_node(struct bpf_verifier_env *env,
10842 struct bpf_reg_state *reg, u32 regno,
10843 struct bpf_kfunc_call_arg_meta *meta)
10845 return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta,
10846 BPF_LIST_HEAD, BPF_LIST_NODE,
10847 &meta->arg_list_head.field);
10850 static int process_kf_arg_ptr_to_rbtree_node(struct bpf_verifier_env *env,
10851 struct bpf_reg_state *reg, u32 regno,
10852 struct bpf_kfunc_call_arg_meta *meta)
10854 return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta,
10855 BPF_RB_ROOT, BPF_RB_NODE,
10856 &meta->arg_rbtree_root.field);
10859 static int check_kfunc_args(struct bpf_verifier_env *env, struct bpf_kfunc_call_arg_meta *meta,
10862 const char *func_name = meta->func_name, *ref_tname;
10863 const struct btf *btf = meta->btf;
10864 const struct btf_param *args;
10865 struct btf_record *rec;
10869 args = (const struct btf_param *)(meta->func_proto + 1);
10870 nargs = btf_type_vlen(meta->func_proto);
10871 if (nargs > MAX_BPF_FUNC_REG_ARGS) {
10872 verbose(env, "Function %s has %d > %d args\n", func_name, nargs,
10873 MAX_BPF_FUNC_REG_ARGS);
10877 /* Check that BTF function arguments match actual types that the
10880 for (i = 0; i < nargs; i++) {
10881 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[i + 1];
10882 const struct btf_type *t, *ref_t, *resolve_ret;
10883 enum bpf_arg_type arg_type = ARG_DONTCARE;
10884 u32 regno = i + 1, ref_id, type_size;
10885 bool is_ret_buf_sz = false;
10888 t = btf_type_skip_modifiers(btf, args[i].type, NULL);
10890 if (is_kfunc_arg_ignore(btf, &args[i]))
10893 if (btf_type_is_scalar(t)) {
10894 if (reg->type != SCALAR_VALUE) {
10895 verbose(env, "R%d is not a scalar\n", regno);
10899 if (is_kfunc_arg_constant(meta->btf, &args[i])) {
10900 if (meta->arg_constant.found) {
10901 verbose(env, "verifier internal error: only one constant argument permitted\n");
10904 if (!tnum_is_const(reg->var_off)) {
10905 verbose(env, "R%d must be a known constant\n", regno);
10908 ret = mark_chain_precision(env, regno);
10911 meta->arg_constant.found = true;
10912 meta->arg_constant.value = reg->var_off.value;
10913 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdonly_buf_size")) {
10914 meta->r0_rdonly = true;
10915 is_ret_buf_sz = true;
10916 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdwr_buf_size")) {
10917 is_ret_buf_sz = true;
10920 if (is_ret_buf_sz) {
10921 if (meta->r0_size) {
10922 verbose(env, "2 or more rdonly/rdwr_buf_size parameters for kfunc");
10926 if (!tnum_is_const(reg->var_off)) {
10927 verbose(env, "R%d is not a const\n", regno);
10931 meta->r0_size = reg->var_off.value;
10932 ret = mark_chain_precision(env, regno);
10939 if (!btf_type_is_ptr(t)) {
10940 verbose(env, "Unrecognized arg#%d type %s\n", i, btf_type_str(t));
10944 if ((is_kfunc_trusted_args(meta) || is_kfunc_rcu(meta)) &&
10945 (register_is_null(reg) || type_may_be_null(reg->type))) {
10946 verbose(env, "Possibly NULL pointer passed to trusted arg%d\n", i);
10950 if (reg->ref_obj_id) {
10951 if (is_kfunc_release(meta) && meta->ref_obj_id) {
10952 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
10953 regno, reg->ref_obj_id,
10957 meta->ref_obj_id = reg->ref_obj_id;
10958 if (is_kfunc_release(meta))
10959 meta->release_regno = regno;
10962 ref_t = btf_type_skip_modifiers(btf, t->type, &ref_id);
10963 ref_tname = btf_name_by_offset(btf, ref_t->name_off);
10965 kf_arg_type = get_kfunc_ptr_arg_type(env, meta, t, ref_t, ref_tname, args, i, nargs);
10966 if (kf_arg_type < 0)
10967 return kf_arg_type;
10969 switch (kf_arg_type) {
10970 case KF_ARG_PTR_TO_ALLOC_BTF_ID:
10971 case KF_ARG_PTR_TO_BTF_ID:
10972 if (!is_kfunc_trusted_args(meta) && !is_kfunc_rcu(meta))
10975 if (!is_trusted_reg(reg)) {
10976 if (!is_kfunc_rcu(meta)) {
10977 verbose(env, "R%d must be referenced or trusted\n", regno);
10980 if (!is_rcu_reg(reg)) {
10981 verbose(env, "R%d must be a rcu pointer\n", regno);
10987 case KF_ARG_PTR_TO_CTX:
10988 /* Trusted arguments have the same offset checks as release arguments */
10989 arg_type |= OBJ_RELEASE;
10991 case KF_ARG_PTR_TO_DYNPTR:
10992 case KF_ARG_PTR_TO_ITER:
10993 case KF_ARG_PTR_TO_LIST_HEAD:
10994 case KF_ARG_PTR_TO_LIST_NODE:
10995 case KF_ARG_PTR_TO_RB_ROOT:
10996 case KF_ARG_PTR_TO_RB_NODE:
10997 case KF_ARG_PTR_TO_MEM:
10998 case KF_ARG_PTR_TO_MEM_SIZE:
10999 case KF_ARG_PTR_TO_CALLBACK:
11000 case KF_ARG_PTR_TO_REFCOUNTED_KPTR:
11001 /* Trusted by default */
11008 if (is_kfunc_release(meta) && reg->ref_obj_id)
11009 arg_type |= OBJ_RELEASE;
11010 ret = check_func_arg_reg_off(env, reg, regno, arg_type);
11014 switch (kf_arg_type) {
11015 case KF_ARG_PTR_TO_CTX:
11016 if (reg->type != PTR_TO_CTX) {
11017 verbose(env, "arg#%d expected pointer to ctx, but got %s\n", i, btf_type_str(t));
11021 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
11022 ret = get_kern_ctx_btf_id(&env->log, resolve_prog_type(env->prog));
11025 meta->ret_btf_id = ret;
11028 case KF_ARG_PTR_TO_ALLOC_BTF_ID:
11029 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11030 verbose(env, "arg#%d expected pointer to allocated object\n", i);
11033 if (!reg->ref_obj_id) {
11034 verbose(env, "allocated object must be referenced\n");
11037 if (meta->btf == btf_vmlinux &&
11038 meta->func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
11039 meta->arg_btf = reg->btf;
11040 meta->arg_btf_id = reg->btf_id;
11043 case KF_ARG_PTR_TO_DYNPTR:
11045 enum bpf_arg_type dynptr_arg_type = ARG_PTR_TO_DYNPTR;
11046 int clone_ref_obj_id = 0;
11048 if (reg->type != PTR_TO_STACK &&
11049 reg->type != CONST_PTR_TO_DYNPTR) {
11050 verbose(env, "arg#%d expected pointer to stack or dynptr_ptr\n", i);
11054 if (reg->type == CONST_PTR_TO_DYNPTR)
11055 dynptr_arg_type |= MEM_RDONLY;
11057 if (is_kfunc_arg_uninit(btf, &args[i]))
11058 dynptr_arg_type |= MEM_UNINIT;
11060 if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) {
11061 dynptr_arg_type |= DYNPTR_TYPE_SKB;
11062 } else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_xdp]) {
11063 dynptr_arg_type |= DYNPTR_TYPE_XDP;
11064 } else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_clone] &&
11065 (dynptr_arg_type & MEM_UNINIT)) {
11066 enum bpf_dynptr_type parent_type = meta->initialized_dynptr.type;
11068 if (parent_type == BPF_DYNPTR_TYPE_INVALID) {
11069 verbose(env, "verifier internal error: no dynptr type for parent of clone\n");
11073 dynptr_arg_type |= (unsigned int)get_dynptr_type_flag(parent_type);
11074 clone_ref_obj_id = meta->initialized_dynptr.ref_obj_id;
11075 if (dynptr_type_refcounted(parent_type) && !clone_ref_obj_id) {
11076 verbose(env, "verifier internal error: missing ref obj id for parent of clone\n");
11081 ret = process_dynptr_func(env, regno, insn_idx, dynptr_arg_type, clone_ref_obj_id);
11085 if (!(dynptr_arg_type & MEM_UNINIT)) {
11086 int id = dynptr_id(env, reg);
11089 verbose(env, "verifier internal error: failed to obtain dynptr id\n");
11092 meta->initialized_dynptr.id = id;
11093 meta->initialized_dynptr.type = dynptr_get_type(env, reg);
11094 meta->initialized_dynptr.ref_obj_id = dynptr_ref_obj_id(env, reg);
11099 case KF_ARG_PTR_TO_ITER:
11100 ret = process_iter_arg(env, regno, insn_idx, meta);
11104 case KF_ARG_PTR_TO_LIST_HEAD:
11105 if (reg->type != PTR_TO_MAP_VALUE &&
11106 reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11107 verbose(env, "arg#%d expected pointer to map value or allocated object\n", i);
11110 if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
11111 verbose(env, "allocated object must be referenced\n");
11114 ret = process_kf_arg_ptr_to_list_head(env, reg, regno, meta);
11118 case KF_ARG_PTR_TO_RB_ROOT:
11119 if (reg->type != PTR_TO_MAP_VALUE &&
11120 reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11121 verbose(env, "arg#%d expected pointer to map value or allocated object\n", i);
11124 if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
11125 verbose(env, "allocated object must be referenced\n");
11128 ret = process_kf_arg_ptr_to_rbtree_root(env, reg, regno, meta);
11132 case KF_ARG_PTR_TO_LIST_NODE:
11133 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11134 verbose(env, "arg#%d expected pointer to allocated object\n", i);
11137 if (!reg->ref_obj_id) {
11138 verbose(env, "allocated object must be referenced\n");
11141 ret = process_kf_arg_ptr_to_list_node(env, reg, regno, meta);
11145 case KF_ARG_PTR_TO_RB_NODE:
11146 if (meta->func_id == special_kfunc_list[KF_bpf_rbtree_remove]) {
11147 if (!type_is_non_owning_ref(reg->type) || reg->ref_obj_id) {
11148 verbose(env, "rbtree_remove node input must be non-owning ref\n");
11151 if (in_rbtree_lock_required_cb(env)) {
11152 verbose(env, "rbtree_remove not allowed in rbtree cb\n");
11156 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11157 verbose(env, "arg#%d expected pointer to allocated object\n", i);
11160 if (!reg->ref_obj_id) {
11161 verbose(env, "allocated object must be referenced\n");
11166 ret = process_kf_arg_ptr_to_rbtree_node(env, reg, regno, meta);
11170 case KF_ARG_PTR_TO_BTF_ID:
11171 /* Only base_type is checked, further checks are done here */
11172 if ((base_type(reg->type) != PTR_TO_BTF_ID ||
11173 (bpf_type_has_unsafe_modifiers(reg->type) && !is_rcu_reg(reg))) &&
11174 !reg2btf_ids[base_type(reg->type)]) {
11175 verbose(env, "arg#%d is %s ", i, reg_type_str(env, reg->type));
11176 verbose(env, "expected %s or socket\n",
11177 reg_type_str(env, base_type(reg->type) |
11178 (type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS)));
11181 ret = process_kf_arg_ptr_to_btf_id(env, reg, ref_t, ref_tname, ref_id, meta, i);
11185 case KF_ARG_PTR_TO_MEM:
11186 resolve_ret = btf_resolve_size(btf, ref_t, &type_size);
11187 if (IS_ERR(resolve_ret)) {
11188 verbose(env, "arg#%d reference type('%s %s') size cannot be determined: %ld\n",
11189 i, btf_type_str(ref_t), ref_tname, PTR_ERR(resolve_ret));
11192 ret = check_mem_reg(env, reg, regno, type_size);
11196 case KF_ARG_PTR_TO_MEM_SIZE:
11198 struct bpf_reg_state *buff_reg = ®s[regno];
11199 const struct btf_param *buff_arg = &args[i];
11200 struct bpf_reg_state *size_reg = ®s[regno + 1];
11201 const struct btf_param *size_arg = &args[i + 1];
11203 if (!register_is_null(buff_reg) || !is_kfunc_arg_optional(meta->btf, buff_arg)) {
11204 ret = check_kfunc_mem_size_reg(env, size_reg, regno + 1);
11206 verbose(env, "arg#%d arg#%d memory, len pair leads to invalid memory access\n", i, i + 1);
11211 if (is_kfunc_arg_const_mem_size(meta->btf, size_arg, size_reg)) {
11212 if (meta->arg_constant.found) {
11213 verbose(env, "verifier internal error: only one constant argument permitted\n");
11216 if (!tnum_is_const(size_reg->var_off)) {
11217 verbose(env, "R%d must be a known constant\n", regno + 1);
11220 meta->arg_constant.found = true;
11221 meta->arg_constant.value = size_reg->var_off.value;
11224 /* Skip next '__sz' or '__szk' argument */
11228 case KF_ARG_PTR_TO_CALLBACK:
11229 if (reg->type != PTR_TO_FUNC) {
11230 verbose(env, "arg%d expected pointer to func\n", i);
11233 meta->subprogno = reg->subprogno;
11235 case KF_ARG_PTR_TO_REFCOUNTED_KPTR:
11236 if (!type_is_ptr_alloc_obj(reg->type)) {
11237 verbose(env, "arg#%d is neither owning or non-owning ref\n", i);
11240 if (!type_is_non_owning_ref(reg->type))
11241 meta->arg_owning_ref = true;
11243 rec = reg_btf_record(reg);
11245 verbose(env, "verifier internal error: Couldn't find btf_record\n");
11249 if (rec->refcount_off < 0) {
11250 verbose(env, "arg#%d doesn't point to a type with bpf_refcount field\n", i);
11254 meta->arg_btf = reg->btf;
11255 meta->arg_btf_id = reg->btf_id;
11260 if (is_kfunc_release(meta) && !meta->release_regno) {
11261 verbose(env, "release kernel function %s expects refcounted PTR_TO_BTF_ID\n",
11269 static int fetch_kfunc_meta(struct bpf_verifier_env *env,
11270 struct bpf_insn *insn,
11271 struct bpf_kfunc_call_arg_meta *meta,
11272 const char **kfunc_name)
11274 const struct btf_type *func, *func_proto;
11275 u32 func_id, *kfunc_flags;
11276 const char *func_name;
11277 struct btf *desc_btf;
11280 *kfunc_name = NULL;
11285 desc_btf = find_kfunc_desc_btf(env, insn->off);
11286 if (IS_ERR(desc_btf))
11287 return PTR_ERR(desc_btf);
11289 func_id = insn->imm;
11290 func = btf_type_by_id(desc_btf, func_id);
11291 func_name = btf_name_by_offset(desc_btf, func->name_off);
11293 *kfunc_name = func_name;
11294 func_proto = btf_type_by_id(desc_btf, func->type);
11296 kfunc_flags = btf_kfunc_id_set_contains(desc_btf, func_id, env->prog);
11297 if (!kfunc_flags) {
11301 memset(meta, 0, sizeof(*meta));
11302 meta->btf = desc_btf;
11303 meta->func_id = func_id;
11304 meta->kfunc_flags = *kfunc_flags;
11305 meta->func_proto = func_proto;
11306 meta->func_name = func_name;
11311 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
11314 const struct btf_type *t, *ptr_type;
11315 u32 i, nargs, ptr_type_id, release_ref_obj_id;
11316 struct bpf_reg_state *regs = cur_regs(env);
11317 const char *func_name, *ptr_type_name;
11318 bool sleepable, rcu_lock, rcu_unlock;
11319 struct bpf_kfunc_call_arg_meta meta;
11320 struct bpf_insn_aux_data *insn_aux;
11321 int err, insn_idx = *insn_idx_p;
11322 const struct btf_param *args;
11323 const struct btf_type *ret_t;
11324 struct btf *desc_btf;
11326 /* skip for now, but return error when we find this in fixup_kfunc_call */
11330 err = fetch_kfunc_meta(env, insn, &meta, &func_name);
11331 if (err == -EACCES && func_name)
11332 verbose(env, "calling kernel function %s is not allowed\n", func_name);
11335 desc_btf = meta.btf;
11336 insn_aux = &env->insn_aux_data[insn_idx];
11338 insn_aux->is_iter_next = is_iter_next_kfunc(&meta);
11340 if (is_kfunc_destructive(&meta) && !capable(CAP_SYS_BOOT)) {
11341 verbose(env, "destructive kfunc calls require CAP_SYS_BOOT capability\n");
11345 sleepable = is_kfunc_sleepable(&meta);
11346 if (sleepable && !env->prog->aux->sleepable) {
11347 verbose(env, "program must be sleepable to call sleepable kfunc %s\n", func_name);
11351 rcu_lock = is_kfunc_bpf_rcu_read_lock(&meta);
11352 rcu_unlock = is_kfunc_bpf_rcu_read_unlock(&meta);
11354 if (env->cur_state->active_rcu_lock) {
11355 struct bpf_func_state *state;
11356 struct bpf_reg_state *reg;
11358 if (in_rbtree_lock_required_cb(env) && (rcu_lock || rcu_unlock)) {
11359 verbose(env, "Calling bpf_rcu_read_{lock,unlock} in unnecessary rbtree callback\n");
11364 verbose(env, "nested rcu read lock (kernel function %s)\n", func_name);
11366 } else if (rcu_unlock) {
11367 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
11368 if (reg->type & MEM_RCU) {
11369 reg->type &= ~(MEM_RCU | PTR_MAYBE_NULL);
11370 reg->type |= PTR_UNTRUSTED;
11373 env->cur_state->active_rcu_lock = false;
11374 } else if (sleepable) {
11375 verbose(env, "kernel func %s is sleepable within rcu_read_lock region\n", func_name);
11378 } else if (rcu_lock) {
11379 env->cur_state->active_rcu_lock = true;
11380 } else if (rcu_unlock) {
11381 verbose(env, "unmatched rcu read unlock (kernel function %s)\n", func_name);
11385 /* Check the arguments */
11386 err = check_kfunc_args(env, &meta, insn_idx);
11389 /* In case of release function, we get register number of refcounted
11390 * PTR_TO_BTF_ID in bpf_kfunc_arg_meta, do the release now.
11392 if (meta.release_regno) {
11393 err = release_reference(env, regs[meta.release_regno].ref_obj_id);
11395 verbose(env, "kfunc %s#%d reference has not been acquired before\n",
11396 func_name, meta.func_id);
11401 if (meta.func_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
11402 meta.func_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
11403 meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
11404 release_ref_obj_id = regs[BPF_REG_2].ref_obj_id;
11405 insn_aux->insert_off = regs[BPF_REG_2].off;
11406 insn_aux->kptr_struct_meta = btf_find_struct_meta(meta.arg_btf, meta.arg_btf_id);
11407 err = ref_convert_owning_non_owning(env, release_ref_obj_id);
11409 verbose(env, "kfunc %s#%d conversion of owning ref to non-owning failed\n",
11410 func_name, meta.func_id);
11414 err = release_reference(env, release_ref_obj_id);
11416 verbose(env, "kfunc %s#%d reference has not been acquired before\n",
11417 func_name, meta.func_id);
11422 if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
11423 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
11424 set_rbtree_add_callback_state);
11426 verbose(env, "kfunc %s#%d failed callback verification\n",
11427 func_name, meta.func_id);
11432 for (i = 0; i < CALLER_SAVED_REGS; i++)
11433 mark_reg_not_init(env, regs, caller_saved[i]);
11435 /* Check return type */
11436 t = btf_type_skip_modifiers(desc_btf, meta.func_proto->type, NULL);
11438 if (is_kfunc_acquire(&meta) && !btf_type_is_struct_ptr(meta.btf, t)) {
11439 /* Only exception is bpf_obj_new_impl */
11440 if (meta.btf != btf_vmlinux ||
11441 (meta.func_id != special_kfunc_list[KF_bpf_obj_new_impl] &&
11442 meta.func_id != special_kfunc_list[KF_bpf_refcount_acquire_impl])) {
11443 verbose(env, "acquire kernel function does not return PTR_TO_BTF_ID\n");
11448 if (btf_type_is_scalar(t)) {
11449 mark_reg_unknown(env, regs, BPF_REG_0);
11450 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
11451 } else if (btf_type_is_ptr(t)) {
11452 ptr_type = btf_type_skip_modifiers(desc_btf, t->type, &ptr_type_id);
11454 if (meta.btf == btf_vmlinux && btf_id_set_contains(&special_kfunc_set, meta.func_id)) {
11455 if (meta.func_id == special_kfunc_list[KF_bpf_obj_new_impl]) {
11456 struct btf *ret_btf;
11459 if (unlikely(!bpf_global_ma_set))
11462 if (((u64)(u32)meta.arg_constant.value) != meta.arg_constant.value) {
11463 verbose(env, "local type ID argument must be in range [0, U32_MAX]\n");
11467 ret_btf = env->prog->aux->btf;
11468 ret_btf_id = meta.arg_constant.value;
11470 /* This may be NULL due to user not supplying a BTF */
11472 verbose(env, "bpf_obj_new requires prog BTF\n");
11476 ret_t = btf_type_by_id(ret_btf, ret_btf_id);
11477 if (!ret_t || !__btf_type_is_struct(ret_t)) {
11478 verbose(env, "bpf_obj_new type ID argument must be of a struct\n");
11482 mark_reg_known_zero(env, regs, BPF_REG_0);
11483 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
11484 regs[BPF_REG_0].btf = ret_btf;
11485 regs[BPF_REG_0].btf_id = ret_btf_id;
11487 insn_aux->obj_new_size = ret_t->size;
11488 insn_aux->kptr_struct_meta =
11489 btf_find_struct_meta(ret_btf, ret_btf_id);
11490 } else if (meta.func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) {
11491 mark_reg_known_zero(env, regs, BPF_REG_0);
11492 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
11493 regs[BPF_REG_0].btf = meta.arg_btf;
11494 regs[BPF_REG_0].btf_id = meta.arg_btf_id;
11496 insn_aux->kptr_struct_meta =
11497 btf_find_struct_meta(meta.arg_btf,
11499 } else if (meta.func_id == special_kfunc_list[KF_bpf_list_pop_front] ||
11500 meta.func_id == special_kfunc_list[KF_bpf_list_pop_back]) {
11501 struct btf_field *field = meta.arg_list_head.field;
11503 mark_reg_graph_node(regs, BPF_REG_0, &field->graph_root);
11504 } else if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
11505 meta.func_id == special_kfunc_list[KF_bpf_rbtree_first]) {
11506 struct btf_field *field = meta.arg_rbtree_root.field;
11508 mark_reg_graph_node(regs, BPF_REG_0, &field->graph_root);
11509 } else if (meta.func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
11510 mark_reg_known_zero(env, regs, BPF_REG_0);
11511 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_TRUSTED;
11512 regs[BPF_REG_0].btf = desc_btf;
11513 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
11514 } else if (meta.func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
11515 ret_t = btf_type_by_id(desc_btf, meta.arg_constant.value);
11516 if (!ret_t || !btf_type_is_struct(ret_t)) {
11518 "kfunc bpf_rdonly_cast type ID argument must be of a struct\n");
11522 mark_reg_known_zero(env, regs, BPF_REG_0);
11523 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
11524 regs[BPF_REG_0].btf = desc_btf;
11525 regs[BPF_REG_0].btf_id = meta.arg_constant.value;
11526 } else if (meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice] ||
11527 meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice_rdwr]) {
11528 enum bpf_type_flag type_flag = get_dynptr_type_flag(meta.initialized_dynptr.type);
11530 mark_reg_known_zero(env, regs, BPF_REG_0);
11532 if (!meta.arg_constant.found) {
11533 verbose(env, "verifier internal error: bpf_dynptr_slice(_rdwr) no constant size\n");
11537 regs[BPF_REG_0].mem_size = meta.arg_constant.value;
11539 /* PTR_MAYBE_NULL will be added when is_kfunc_ret_null is checked */
11540 regs[BPF_REG_0].type = PTR_TO_MEM | type_flag;
11542 if (meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice]) {
11543 regs[BPF_REG_0].type |= MEM_RDONLY;
11545 /* this will set env->seen_direct_write to true */
11546 if (!may_access_direct_pkt_data(env, NULL, BPF_WRITE)) {
11547 verbose(env, "the prog does not allow writes to packet data\n");
11552 if (!meta.initialized_dynptr.id) {
11553 verbose(env, "verifier internal error: no dynptr id\n");
11556 regs[BPF_REG_0].dynptr_id = meta.initialized_dynptr.id;
11558 /* we don't need to set BPF_REG_0's ref obj id
11559 * because packet slices are not refcounted (see
11560 * dynptr_type_refcounted)
11563 verbose(env, "kernel function %s unhandled dynamic return type\n",
11567 } else if (!__btf_type_is_struct(ptr_type)) {
11568 if (!meta.r0_size) {
11571 if (!IS_ERR(btf_resolve_size(desc_btf, ptr_type, &sz))) {
11573 meta.r0_rdonly = true;
11576 if (!meta.r0_size) {
11577 ptr_type_name = btf_name_by_offset(desc_btf,
11578 ptr_type->name_off);
11580 "kernel function %s returns pointer type %s %s is not supported\n",
11582 btf_type_str(ptr_type),
11587 mark_reg_known_zero(env, regs, BPF_REG_0);
11588 regs[BPF_REG_0].type = PTR_TO_MEM;
11589 regs[BPF_REG_0].mem_size = meta.r0_size;
11591 if (meta.r0_rdonly)
11592 regs[BPF_REG_0].type |= MEM_RDONLY;
11594 /* Ensures we don't access the memory after a release_reference() */
11595 if (meta.ref_obj_id)
11596 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
11598 mark_reg_known_zero(env, regs, BPF_REG_0);
11599 regs[BPF_REG_0].btf = desc_btf;
11600 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
11601 regs[BPF_REG_0].btf_id = ptr_type_id;
11604 if (is_kfunc_ret_null(&meta)) {
11605 regs[BPF_REG_0].type |= PTR_MAYBE_NULL;
11606 /* For mark_ptr_or_null_reg, see 93c230e3f5bd6 */
11607 regs[BPF_REG_0].id = ++env->id_gen;
11609 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
11610 if (is_kfunc_acquire(&meta)) {
11611 int id = acquire_reference_state(env, insn_idx);
11615 if (is_kfunc_ret_null(&meta))
11616 regs[BPF_REG_0].id = id;
11617 regs[BPF_REG_0].ref_obj_id = id;
11618 } else if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_first]) {
11619 ref_set_non_owning(env, ®s[BPF_REG_0]);
11622 if (reg_may_point_to_spin_lock(®s[BPF_REG_0]) && !regs[BPF_REG_0].id)
11623 regs[BPF_REG_0].id = ++env->id_gen;
11624 } else if (btf_type_is_void(t)) {
11625 if (meta.btf == btf_vmlinux && btf_id_set_contains(&special_kfunc_set, meta.func_id)) {
11626 if (meta.func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
11627 insn_aux->kptr_struct_meta =
11628 btf_find_struct_meta(meta.arg_btf,
11634 nargs = btf_type_vlen(meta.func_proto);
11635 args = (const struct btf_param *)(meta.func_proto + 1);
11636 for (i = 0; i < nargs; i++) {
11639 t = btf_type_skip_modifiers(desc_btf, args[i].type, NULL);
11640 if (btf_type_is_ptr(t))
11641 mark_btf_func_reg_size(env, regno, sizeof(void *));
11643 /* scalar. ensured by btf_check_kfunc_arg_match() */
11644 mark_btf_func_reg_size(env, regno, t->size);
11647 if (is_iter_next_kfunc(&meta)) {
11648 err = process_iter_next_call(env, insn_idx, &meta);
11656 static bool signed_add_overflows(s64 a, s64 b)
11658 /* Do the add in u64, where overflow is well-defined */
11659 s64 res = (s64)((u64)a + (u64)b);
11666 static bool signed_add32_overflows(s32 a, s32 b)
11668 /* Do the add in u32, where overflow is well-defined */
11669 s32 res = (s32)((u32)a + (u32)b);
11676 static bool signed_sub_overflows(s64 a, s64 b)
11678 /* Do the sub in u64, where overflow is well-defined */
11679 s64 res = (s64)((u64)a - (u64)b);
11686 static bool signed_sub32_overflows(s32 a, s32 b)
11688 /* Do the sub in u32, where overflow is well-defined */
11689 s32 res = (s32)((u32)a - (u32)b);
11696 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
11697 const struct bpf_reg_state *reg,
11698 enum bpf_reg_type type)
11700 bool known = tnum_is_const(reg->var_off);
11701 s64 val = reg->var_off.value;
11702 s64 smin = reg->smin_value;
11704 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
11705 verbose(env, "math between %s pointer and %lld is not allowed\n",
11706 reg_type_str(env, type), val);
11710 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
11711 verbose(env, "%s pointer offset %d is not allowed\n",
11712 reg_type_str(env, type), reg->off);
11716 if (smin == S64_MIN) {
11717 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
11718 reg_type_str(env, type));
11722 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
11723 verbose(env, "value %lld makes %s pointer be out of bounds\n",
11724 smin, reg_type_str(env, type));
11732 REASON_BOUNDS = -1,
11739 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
11740 u32 *alu_limit, bool mask_to_left)
11742 u32 max = 0, ptr_limit = 0;
11744 switch (ptr_reg->type) {
11746 /* Offset 0 is out-of-bounds, but acceptable start for the
11747 * left direction, see BPF_REG_FP. Also, unknown scalar
11748 * offset where we would need to deal with min/max bounds is
11749 * currently prohibited for unprivileged.
11751 max = MAX_BPF_STACK + mask_to_left;
11752 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
11754 case PTR_TO_MAP_VALUE:
11755 max = ptr_reg->map_ptr->value_size;
11756 ptr_limit = (mask_to_left ?
11757 ptr_reg->smin_value :
11758 ptr_reg->umax_value) + ptr_reg->off;
11761 return REASON_TYPE;
11764 if (ptr_limit >= max)
11765 return REASON_LIMIT;
11766 *alu_limit = ptr_limit;
11770 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
11771 const struct bpf_insn *insn)
11773 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
11776 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
11777 u32 alu_state, u32 alu_limit)
11779 /* If we arrived here from different branches with different
11780 * state or limits to sanitize, then this won't work.
11782 if (aux->alu_state &&
11783 (aux->alu_state != alu_state ||
11784 aux->alu_limit != alu_limit))
11785 return REASON_PATHS;
11787 /* Corresponding fixup done in do_misc_fixups(). */
11788 aux->alu_state = alu_state;
11789 aux->alu_limit = alu_limit;
11793 static int sanitize_val_alu(struct bpf_verifier_env *env,
11794 struct bpf_insn *insn)
11796 struct bpf_insn_aux_data *aux = cur_aux(env);
11798 if (can_skip_alu_sanitation(env, insn))
11801 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
11804 static bool sanitize_needed(u8 opcode)
11806 return opcode == BPF_ADD || opcode == BPF_SUB;
11809 struct bpf_sanitize_info {
11810 struct bpf_insn_aux_data aux;
11814 static struct bpf_verifier_state *
11815 sanitize_speculative_path(struct bpf_verifier_env *env,
11816 const struct bpf_insn *insn,
11817 u32 next_idx, u32 curr_idx)
11819 struct bpf_verifier_state *branch;
11820 struct bpf_reg_state *regs;
11822 branch = push_stack(env, next_idx, curr_idx, true);
11823 if (branch && insn) {
11824 regs = branch->frame[branch->curframe]->regs;
11825 if (BPF_SRC(insn->code) == BPF_K) {
11826 mark_reg_unknown(env, regs, insn->dst_reg);
11827 } else if (BPF_SRC(insn->code) == BPF_X) {
11828 mark_reg_unknown(env, regs, insn->dst_reg);
11829 mark_reg_unknown(env, regs, insn->src_reg);
11835 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
11836 struct bpf_insn *insn,
11837 const struct bpf_reg_state *ptr_reg,
11838 const struct bpf_reg_state *off_reg,
11839 struct bpf_reg_state *dst_reg,
11840 struct bpf_sanitize_info *info,
11841 const bool commit_window)
11843 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
11844 struct bpf_verifier_state *vstate = env->cur_state;
11845 bool off_is_imm = tnum_is_const(off_reg->var_off);
11846 bool off_is_neg = off_reg->smin_value < 0;
11847 bool ptr_is_dst_reg = ptr_reg == dst_reg;
11848 u8 opcode = BPF_OP(insn->code);
11849 u32 alu_state, alu_limit;
11850 struct bpf_reg_state tmp;
11854 if (can_skip_alu_sanitation(env, insn))
11857 /* We already marked aux for masking from non-speculative
11858 * paths, thus we got here in the first place. We only care
11859 * to explore bad access from here.
11861 if (vstate->speculative)
11864 if (!commit_window) {
11865 if (!tnum_is_const(off_reg->var_off) &&
11866 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
11867 return REASON_BOUNDS;
11869 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
11870 (opcode == BPF_SUB && !off_is_neg);
11873 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
11877 if (commit_window) {
11878 /* In commit phase we narrow the masking window based on
11879 * the observed pointer move after the simulated operation.
11881 alu_state = info->aux.alu_state;
11882 alu_limit = abs(info->aux.alu_limit - alu_limit);
11884 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
11885 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
11886 alu_state |= ptr_is_dst_reg ?
11887 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
11889 /* Limit pruning on unknown scalars to enable deep search for
11890 * potential masking differences from other program paths.
11893 env->explore_alu_limits = true;
11896 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
11900 /* If we're in commit phase, we're done here given we already
11901 * pushed the truncated dst_reg into the speculative verification
11904 * Also, when register is a known constant, we rewrite register-based
11905 * operation to immediate-based, and thus do not need masking (and as
11906 * a consequence, do not need to simulate the zero-truncation either).
11908 if (commit_window || off_is_imm)
11911 /* Simulate and find potential out-of-bounds access under
11912 * speculative execution from truncation as a result of
11913 * masking when off was not within expected range. If off
11914 * sits in dst, then we temporarily need to move ptr there
11915 * to simulate dst (== 0) +/-= ptr. Needed, for example,
11916 * for cases where we use K-based arithmetic in one direction
11917 * and truncated reg-based in the other in order to explore
11920 if (!ptr_is_dst_reg) {
11922 copy_register_state(dst_reg, ptr_reg);
11924 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
11926 if (!ptr_is_dst_reg && ret)
11928 return !ret ? REASON_STACK : 0;
11931 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
11933 struct bpf_verifier_state *vstate = env->cur_state;
11935 /* If we simulate paths under speculation, we don't update the
11936 * insn as 'seen' such that when we verify unreachable paths in
11937 * the non-speculative domain, sanitize_dead_code() can still
11938 * rewrite/sanitize them.
11940 if (!vstate->speculative)
11941 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
11944 static int sanitize_err(struct bpf_verifier_env *env,
11945 const struct bpf_insn *insn, int reason,
11946 const struct bpf_reg_state *off_reg,
11947 const struct bpf_reg_state *dst_reg)
11949 static const char *err = "pointer arithmetic with it prohibited for !root";
11950 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
11951 u32 dst = insn->dst_reg, src = insn->src_reg;
11954 case REASON_BOUNDS:
11955 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
11956 off_reg == dst_reg ? dst : src, err);
11959 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
11960 off_reg == dst_reg ? src : dst, err);
11963 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
11967 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
11971 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
11975 verbose(env, "verifier internal error: unknown reason (%d)\n",
11983 /* check that stack access falls within stack limits and that 'reg' doesn't
11984 * have a variable offset.
11986 * Variable offset is prohibited for unprivileged mode for simplicity since it
11987 * requires corresponding support in Spectre masking for stack ALU. See also
11988 * retrieve_ptr_limit().
11991 * 'off' includes 'reg->off'.
11993 static int check_stack_access_for_ptr_arithmetic(
11994 struct bpf_verifier_env *env,
11996 const struct bpf_reg_state *reg,
11999 if (!tnum_is_const(reg->var_off)) {
12002 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
12003 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
12004 regno, tn_buf, off);
12008 if (off >= 0 || off < -MAX_BPF_STACK) {
12009 verbose(env, "R%d stack pointer arithmetic goes out of range, "
12010 "prohibited for !root; off=%d\n", regno, off);
12017 static int sanitize_check_bounds(struct bpf_verifier_env *env,
12018 const struct bpf_insn *insn,
12019 const struct bpf_reg_state *dst_reg)
12021 u32 dst = insn->dst_reg;
12023 /* For unprivileged we require that resulting offset must be in bounds
12024 * in order to be able to sanitize access later on.
12026 if (env->bypass_spec_v1)
12029 switch (dst_reg->type) {
12031 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
12032 dst_reg->off + dst_reg->var_off.value))
12035 case PTR_TO_MAP_VALUE:
12036 if (check_map_access(env, dst, dst_reg->off, 1, false, ACCESS_HELPER)) {
12037 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
12038 "prohibited for !root\n", dst);
12049 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
12050 * Caller should also handle BPF_MOV case separately.
12051 * If we return -EACCES, caller may want to try again treating pointer as a
12052 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
12054 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
12055 struct bpf_insn *insn,
12056 const struct bpf_reg_state *ptr_reg,
12057 const struct bpf_reg_state *off_reg)
12059 struct bpf_verifier_state *vstate = env->cur_state;
12060 struct bpf_func_state *state = vstate->frame[vstate->curframe];
12061 struct bpf_reg_state *regs = state->regs, *dst_reg;
12062 bool known = tnum_is_const(off_reg->var_off);
12063 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
12064 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
12065 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
12066 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
12067 struct bpf_sanitize_info info = {};
12068 u8 opcode = BPF_OP(insn->code);
12069 u32 dst = insn->dst_reg;
12072 dst_reg = ®s[dst];
12074 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
12075 smin_val > smax_val || umin_val > umax_val) {
12076 /* Taint dst register if offset had invalid bounds derived from
12077 * e.g. dead branches.
12079 __mark_reg_unknown(env, dst_reg);
12083 if (BPF_CLASS(insn->code) != BPF_ALU64) {
12084 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
12085 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
12086 __mark_reg_unknown(env, dst_reg);
12091 "R%d 32-bit pointer arithmetic prohibited\n",
12096 if (ptr_reg->type & PTR_MAYBE_NULL) {
12097 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
12098 dst, reg_type_str(env, ptr_reg->type));
12102 switch (base_type(ptr_reg->type)) {
12103 case PTR_TO_FLOW_KEYS:
12107 case CONST_PTR_TO_MAP:
12108 /* smin_val represents the known value */
12109 if (known && smin_val == 0 && opcode == BPF_ADD)
12112 case PTR_TO_PACKET_END:
12113 case PTR_TO_SOCKET:
12114 case PTR_TO_SOCK_COMMON:
12115 case PTR_TO_TCP_SOCK:
12116 case PTR_TO_XDP_SOCK:
12117 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
12118 dst, reg_type_str(env, ptr_reg->type));
12124 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
12125 * The id may be overwritten later if we create a new variable offset.
12127 dst_reg->type = ptr_reg->type;
12128 dst_reg->id = ptr_reg->id;
12130 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
12131 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
12134 /* pointer types do not carry 32-bit bounds at the moment. */
12135 __mark_reg32_unbounded(dst_reg);
12137 if (sanitize_needed(opcode)) {
12138 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
12141 return sanitize_err(env, insn, ret, off_reg, dst_reg);
12146 /* We can take a fixed offset as long as it doesn't overflow
12147 * the s32 'off' field
12149 if (known && (ptr_reg->off + smin_val ==
12150 (s64)(s32)(ptr_reg->off + smin_val))) {
12151 /* pointer += K. Accumulate it into fixed offset */
12152 dst_reg->smin_value = smin_ptr;
12153 dst_reg->smax_value = smax_ptr;
12154 dst_reg->umin_value = umin_ptr;
12155 dst_reg->umax_value = umax_ptr;
12156 dst_reg->var_off = ptr_reg->var_off;
12157 dst_reg->off = ptr_reg->off + smin_val;
12158 dst_reg->raw = ptr_reg->raw;
12161 /* A new variable offset is created. Note that off_reg->off
12162 * == 0, since it's a scalar.
12163 * dst_reg gets the pointer type and since some positive
12164 * integer value was added to the pointer, give it a new 'id'
12165 * if it's a PTR_TO_PACKET.
12166 * this creates a new 'base' pointer, off_reg (variable) gets
12167 * added into the variable offset, and we copy the fixed offset
12170 if (signed_add_overflows(smin_ptr, smin_val) ||
12171 signed_add_overflows(smax_ptr, smax_val)) {
12172 dst_reg->smin_value = S64_MIN;
12173 dst_reg->smax_value = S64_MAX;
12175 dst_reg->smin_value = smin_ptr + smin_val;
12176 dst_reg->smax_value = smax_ptr + smax_val;
12178 if (umin_ptr + umin_val < umin_ptr ||
12179 umax_ptr + umax_val < umax_ptr) {
12180 dst_reg->umin_value = 0;
12181 dst_reg->umax_value = U64_MAX;
12183 dst_reg->umin_value = umin_ptr + umin_val;
12184 dst_reg->umax_value = umax_ptr + umax_val;
12186 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
12187 dst_reg->off = ptr_reg->off;
12188 dst_reg->raw = ptr_reg->raw;
12189 if (reg_is_pkt_pointer(ptr_reg)) {
12190 dst_reg->id = ++env->id_gen;
12191 /* something was added to pkt_ptr, set range to zero */
12192 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
12196 if (dst_reg == off_reg) {
12197 /* scalar -= pointer. Creates an unknown scalar */
12198 verbose(env, "R%d tried to subtract pointer from scalar\n",
12202 /* We don't allow subtraction from FP, because (according to
12203 * test_verifier.c test "invalid fp arithmetic", JITs might not
12204 * be able to deal with it.
12206 if (ptr_reg->type == PTR_TO_STACK) {
12207 verbose(env, "R%d subtraction from stack pointer prohibited\n",
12211 if (known && (ptr_reg->off - smin_val ==
12212 (s64)(s32)(ptr_reg->off - smin_val))) {
12213 /* pointer -= K. Subtract it from fixed offset */
12214 dst_reg->smin_value = smin_ptr;
12215 dst_reg->smax_value = smax_ptr;
12216 dst_reg->umin_value = umin_ptr;
12217 dst_reg->umax_value = umax_ptr;
12218 dst_reg->var_off = ptr_reg->var_off;
12219 dst_reg->id = ptr_reg->id;
12220 dst_reg->off = ptr_reg->off - smin_val;
12221 dst_reg->raw = ptr_reg->raw;
12224 /* A new variable offset is created. If the subtrahend is known
12225 * nonnegative, then any reg->range we had before is still good.
12227 if (signed_sub_overflows(smin_ptr, smax_val) ||
12228 signed_sub_overflows(smax_ptr, smin_val)) {
12229 /* Overflow possible, we know nothing */
12230 dst_reg->smin_value = S64_MIN;
12231 dst_reg->smax_value = S64_MAX;
12233 dst_reg->smin_value = smin_ptr - smax_val;
12234 dst_reg->smax_value = smax_ptr - smin_val;
12236 if (umin_ptr < umax_val) {
12237 /* Overflow possible, we know nothing */
12238 dst_reg->umin_value = 0;
12239 dst_reg->umax_value = U64_MAX;
12241 /* Cannot overflow (as long as bounds are consistent) */
12242 dst_reg->umin_value = umin_ptr - umax_val;
12243 dst_reg->umax_value = umax_ptr - umin_val;
12245 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
12246 dst_reg->off = ptr_reg->off;
12247 dst_reg->raw = ptr_reg->raw;
12248 if (reg_is_pkt_pointer(ptr_reg)) {
12249 dst_reg->id = ++env->id_gen;
12250 /* something was added to pkt_ptr, set range to zero */
12252 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
12258 /* bitwise ops on pointers are troublesome, prohibit. */
12259 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
12260 dst, bpf_alu_string[opcode >> 4]);
12263 /* other operators (e.g. MUL,LSH) produce non-pointer results */
12264 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
12265 dst, bpf_alu_string[opcode >> 4]);
12269 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
12271 reg_bounds_sync(dst_reg);
12272 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
12274 if (sanitize_needed(opcode)) {
12275 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
12278 return sanitize_err(env, insn, ret, off_reg, dst_reg);
12284 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
12285 struct bpf_reg_state *src_reg)
12287 s32 smin_val = src_reg->s32_min_value;
12288 s32 smax_val = src_reg->s32_max_value;
12289 u32 umin_val = src_reg->u32_min_value;
12290 u32 umax_val = src_reg->u32_max_value;
12292 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
12293 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
12294 dst_reg->s32_min_value = S32_MIN;
12295 dst_reg->s32_max_value = S32_MAX;
12297 dst_reg->s32_min_value += smin_val;
12298 dst_reg->s32_max_value += smax_val;
12300 if (dst_reg->u32_min_value + umin_val < umin_val ||
12301 dst_reg->u32_max_value + umax_val < umax_val) {
12302 dst_reg->u32_min_value = 0;
12303 dst_reg->u32_max_value = U32_MAX;
12305 dst_reg->u32_min_value += umin_val;
12306 dst_reg->u32_max_value += umax_val;
12310 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
12311 struct bpf_reg_state *src_reg)
12313 s64 smin_val = src_reg->smin_value;
12314 s64 smax_val = src_reg->smax_value;
12315 u64 umin_val = src_reg->umin_value;
12316 u64 umax_val = src_reg->umax_value;
12318 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
12319 signed_add_overflows(dst_reg->smax_value, smax_val)) {
12320 dst_reg->smin_value = S64_MIN;
12321 dst_reg->smax_value = S64_MAX;
12323 dst_reg->smin_value += smin_val;
12324 dst_reg->smax_value += smax_val;
12326 if (dst_reg->umin_value + umin_val < umin_val ||
12327 dst_reg->umax_value + umax_val < umax_val) {
12328 dst_reg->umin_value = 0;
12329 dst_reg->umax_value = U64_MAX;
12331 dst_reg->umin_value += umin_val;
12332 dst_reg->umax_value += umax_val;
12336 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
12337 struct bpf_reg_state *src_reg)
12339 s32 smin_val = src_reg->s32_min_value;
12340 s32 smax_val = src_reg->s32_max_value;
12341 u32 umin_val = src_reg->u32_min_value;
12342 u32 umax_val = src_reg->u32_max_value;
12344 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
12345 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
12346 /* Overflow possible, we know nothing */
12347 dst_reg->s32_min_value = S32_MIN;
12348 dst_reg->s32_max_value = S32_MAX;
12350 dst_reg->s32_min_value -= smax_val;
12351 dst_reg->s32_max_value -= smin_val;
12353 if (dst_reg->u32_min_value < umax_val) {
12354 /* Overflow possible, we know nothing */
12355 dst_reg->u32_min_value = 0;
12356 dst_reg->u32_max_value = U32_MAX;
12358 /* Cannot overflow (as long as bounds are consistent) */
12359 dst_reg->u32_min_value -= umax_val;
12360 dst_reg->u32_max_value -= umin_val;
12364 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
12365 struct bpf_reg_state *src_reg)
12367 s64 smin_val = src_reg->smin_value;
12368 s64 smax_val = src_reg->smax_value;
12369 u64 umin_val = src_reg->umin_value;
12370 u64 umax_val = src_reg->umax_value;
12372 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
12373 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
12374 /* Overflow possible, we know nothing */
12375 dst_reg->smin_value = S64_MIN;
12376 dst_reg->smax_value = S64_MAX;
12378 dst_reg->smin_value -= smax_val;
12379 dst_reg->smax_value -= smin_val;
12381 if (dst_reg->umin_value < umax_val) {
12382 /* Overflow possible, we know nothing */
12383 dst_reg->umin_value = 0;
12384 dst_reg->umax_value = U64_MAX;
12386 /* Cannot overflow (as long as bounds are consistent) */
12387 dst_reg->umin_value -= umax_val;
12388 dst_reg->umax_value -= umin_val;
12392 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
12393 struct bpf_reg_state *src_reg)
12395 s32 smin_val = src_reg->s32_min_value;
12396 u32 umin_val = src_reg->u32_min_value;
12397 u32 umax_val = src_reg->u32_max_value;
12399 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
12400 /* Ain't nobody got time to multiply that sign */
12401 __mark_reg32_unbounded(dst_reg);
12404 /* Both values are positive, so we can work with unsigned and
12405 * copy the result to signed (unless it exceeds S32_MAX).
12407 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
12408 /* Potential overflow, we know nothing */
12409 __mark_reg32_unbounded(dst_reg);
12412 dst_reg->u32_min_value *= umin_val;
12413 dst_reg->u32_max_value *= umax_val;
12414 if (dst_reg->u32_max_value > S32_MAX) {
12415 /* Overflow possible, we know nothing */
12416 dst_reg->s32_min_value = S32_MIN;
12417 dst_reg->s32_max_value = S32_MAX;
12419 dst_reg->s32_min_value = dst_reg->u32_min_value;
12420 dst_reg->s32_max_value = dst_reg->u32_max_value;
12424 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
12425 struct bpf_reg_state *src_reg)
12427 s64 smin_val = src_reg->smin_value;
12428 u64 umin_val = src_reg->umin_value;
12429 u64 umax_val = src_reg->umax_value;
12431 if (smin_val < 0 || dst_reg->smin_value < 0) {
12432 /* Ain't nobody got time to multiply that sign */
12433 __mark_reg64_unbounded(dst_reg);
12436 /* Both values are positive, so we can work with unsigned and
12437 * copy the result to signed (unless it exceeds S64_MAX).
12439 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
12440 /* Potential overflow, we know nothing */
12441 __mark_reg64_unbounded(dst_reg);
12444 dst_reg->umin_value *= umin_val;
12445 dst_reg->umax_value *= umax_val;
12446 if (dst_reg->umax_value > S64_MAX) {
12447 /* Overflow possible, we know nothing */
12448 dst_reg->smin_value = S64_MIN;
12449 dst_reg->smax_value = S64_MAX;
12451 dst_reg->smin_value = dst_reg->umin_value;
12452 dst_reg->smax_value = dst_reg->umax_value;
12456 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
12457 struct bpf_reg_state *src_reg)
12459 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12460 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12461 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12462 s32 smin_val = src_reg->s32_min_value;
12463 u32 umax_val = src_reg->u32_max_value;
12465 if (src_known && dst_known) {
12466 __mark_reg32_known(dst_reg, var32_off.value);
12470 /* We get our minimum from the var_off, since that's inherently
12471 * bitwise. Our maximum is the minimum of the operands' maxima.
12473 dst_reg->u32_min_value = var32_off.value;
12474 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
12475 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
12476 /* Lose signed bounds when ANDing negative numbers,
12477 * ain't nobody got time for that.
12479 dst_reg->s32_min_value = S32_MIN;
12480 dst_reg->s32_max_value = S32_MAX;
12482 /* ANDing two positives gives a positive, so safe to
12483 * cast result into s64.
12485 dst_reg->s32_min_value = dst_reg->u32_min_value;
12486 dst_reg->s32_max_value = dst_reg->u32_max_value;
12490 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
12491 struct bpf_reg_state *src_reg)
12493 bool src_known = tnum_is_const(src_reg->var_off);
12494 bool dst_known = tnum_is_const(dst_reg->var_off);
12495 s64 smin_val = src_reg->smin_value;
12496 u64 umax_val = src_reg->umax_value;
12498 if (src_known && dst_known) {
12499 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12503 /* We get our minimum from the var_off, since that's inherently
12504 * bitwise. Our maximum is the minimum of the operands' maxima.
12506 dst_reg->umin_value = dst_reg->var_off.value;
12507 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
12508 if (dst_reg->smin_value < 0 || smin_val < 0) {
12509 /* Lose signed bounds when ANDing negative numbers,
12510 * ain't nobody got time for that.
12512 dst_reg->smin_value = S64_MIN;
12513 dst_reg->smax_value = S64_MAX;
12515 /* ANDing two positives gives a positive, so safe to
12516 * cast result into s64.
12518 dst_reg->smin_value = dst_reg->umin_value;
12519 dst_reg->smax_value = dst_reg->umax_value;
12521 /* We may learn something more from the var_off */
12522 __update_reg_bounds(dst_reg);
12525 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
12526 struct bpf_reg_state *src_reg)
12528 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12529 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12530 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12531 s32 smin_val = src_reg->s32_min_value;
12532 u32 umin_val = src_reg->u32_min_value;
12534 if (src_known && dst_known) {
12535 __mark_reg32_known(dst_reg, var32_off.value);
12539 /* We get our maximum from the var_off, and our minimum is the
12540 * maximum of the operands' minima
12542 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
12543 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
12544 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
12545 /* Lose signed bounds when ORing negative numbers,
12546 * ain't nobody got time for that.
12548 dst_reg->s32_min_value = S32_MIN;
12549 dst_reg->s32_max_value = S32_MAX;
12551 /* ORing two positives gives a positive, so safe to
12552 * cast result into s64.
12554 dst_reg->s32_min_value = dst_reg->u32_min_value;
12555 dst_reg->s32_max_value = dst_reg->u32_max_value;
12559 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
12560 struct bpf_reg_state *src_reg)
12562 bool src_known = tnum_is_const(src_reg->var_off);
12563 bool dst_known = tnum_is_const(dst_reg->var_off);
12564 s64 smin_val = src_reg->smin_value;
12565 u64 umin_val = src_reg->umin_value;
12567 if (src_known && dst_known) {
12568 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12572 /* We get our maximum from the var_off, and our minimum is the
12573 * maximum of the operands' minima
12575 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
12576 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
12577 if (dst_reg->smin_value < 0 || smin_val < 0) {
12578 /* Lose signed bounds when ORing negative numbers,
12579 * ain't nobody got time for that.
12581 dst_reg->smin_value = S64_MIN;
12582 dst_reg->smax_value = S64_MAX;
12584 /* ORing two positives gives a positive, so safe to
12585 * cast result into s64.
12587 dst_reg->smin_value = dst_reg->umin_value;
12588 dst_reg->smax_value = dst_reg->umax_value;
12590 /* We may learn something more from the var_off */
12591 __update_reg_bounds(dst_reg);
12594 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
12595 struct bpf_reg_state *src_reg)
12597 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12598 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12599 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12600 s32 smin_val = src_reg->s32_min_value;
12602 if (src_known && dst_known) {
12603 __mark_reg32_known(dst_reg, var32_off.value);
12607 /* We get both minimum and maximum from the var32_off. */
12608 dst_reg->u32_min_value = var32_off.value;
12609 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
12611 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
12612 /* XORing two positive sign numbers gives a positive,
12613 * so safe to cast u32 result into s32.
12615 dst_reg->s32_min_value = dst_reg->u32_min_value;
12616 dst_reg->s32_max_value = dst_reg->u32_max_value;
12618 dst_reg->s32_min_value = S32_MIN;
12619 dst_reg->s32_max_value = S32_MAX;
12623 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
12624 struct bpf_reg_state *src_reg)
12626 bool src_known = tnum_is_const(src_reg->var_off);
12627 bool dst_known = tnum_is_const(dst_reg->var_off);
12628 s64 smin_val = src_reg->smin_value;
12630 if (src_known && dst_known) {
12631 /* dst_reg->var_off.value has been updated earlier */
12632 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12636 /* We get both minimum and maximum from the var_off. */
12637 dst_reg->umin_value = dst_reg->var_off.value;
12638 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
12640 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
12641 /* XORing two positive sign numbers gives a positive,
12642 * so safe to cast u64 result into s64.
12644 dst_reg->smin_value = dst_reg->umin_value;
12645 dst_reg->smax_value = dst_reg->umax_value;
12647 dst_reg->smin_value = S64_MIN;
12648 dst_reg->smax_value = S64_MAX;
12651 __update_reg_bounds(dst_reg);
12654 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
12655 u64 umin_val, u64 umax_val)
12657 /* We lose all sign bit information (except what we can pick
12660 dst_reg->s32_min_value = S32_MIN;
12661 dst_reg->s32_max_value = S32_MAX;
12662 /* If we might shift our top bit out, then we know nothing */
12663 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
12664 dst_reg->u32_min_value = 0;
12665 dst_reg->u32_max_value = U32_MAX;
12667 dst_reg->u32_min_value <<= umin_val;
12668 dst_reg->u32_max_value <<= umax_val;
12672 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
12673 struct bpf_reg_state *src_reg)
12675 u32 umax_val = src_reg->u32_max_value;
12676 u32 umin_val = src_reg->u32_min_value;
12677 /* u32 alu operation will zext upper bits */
12678 struct tnum subreg = tnum_subreg(dst_reg->var_off);
12680 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
12681 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
12682 /* Not required but being careful mark reg64 bounds as unknown so
12683 * that we are forced to pick them up from tnum and zext later and
12684 * if some path skips this step we are still safe.
12686 __mark_reg64_unbounded(dst_reg);
12687 __update_reg32_bounds(dst_reg);
12690 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
12691 u64 umin_val, u64 umax_val)
12693 /* Special case <<32 because it is a common compiler pattern to sign
12694 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
12695 * positive we know this shift will also be positive so we can track
12696 * bounds correctly. Otherwise we lose all sign bit information except
12697 * what we can pick up from var_off. Perhaps we can generalize this
12698 * later to shifts of any length.
12700 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
12701 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
12703 dst_reg->smax_value = S64_MAX;
12705 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
12706 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
12708 dst_reg->smin_value = S64_MIN;
12710 /* If we might shift our top bit out, then we know nothing */
12711 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
12712 dst_reg->umin_value = 0;
12713 dst_reg->umax_value = U64_MAX;
12715 dst_reg->umin_value <<= umin_val;
12716 dst_reg->umax_value <<= umax_val;
12720 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
12721 struct bpf_reg_state *src_reg)
12723 u64 umax_val = src_reg->umax_value;
12724 u64 umin_val = src_reg->umin_value;
12726 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
12727 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
12728 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
12730 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
12731 /* We may learn something more from the var_off */
12732 __update_reg_bounds(dst_reg);
12735 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
12736 struct bpf_reg_state *src_reg)
12738 struct tnum subreg = tnum_subreg(dst_reg->var_off);
12739 u32 umax_val = src_reg->u32_max_value;
12740 u32 umin_val = src_reg->u32_min_value;
12742 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
12743 * be negative, then either:
12744 * 1) src_reg might be zero, so the sign bit of the result is
12745 * unknown, so we lose our signed bounds
12746 * 2) it's known negative, thus the unsigned bounds capture the
12748 * 3) the signed bounds cross zero, so they tell us nothing
12750 * If the value in dst_reg is known nonnegative, then again the
12751 * unsigned bounds capture the signed bounds.
12752 * Thus, in all cases it suffices to blow away our signed bounds
12753 * and rely on inferring new ones from the unsigned bounds and
12754 * var_off of the result.
12756 dst_reg->s32_min_value = S32_MIN;
12757 dst_reg->s32_max_value = S32_MAX;
12759 dst_reg->var_off = tnum_rshift(subreg, umin_val);
12760 dst_reg->u32_min_value >>= umax_val;
12761 dst_reg->u32_max_value >>= umin_val;
12763 __mark_reg64_unbounded(dst_reg);
12764 __update_reg32_bounds(dst_reg);
12767 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
12768 struct bpf_reg_state *src_reg)
12770 u64 umax_val = src_reg->umax_value;
12771 u64 umin_val = src_reg->umin_value;
12773 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
12774 * be negative, then either:
12775 * 1) src_reg might be zero, so the sign bit of the result is
12776 * unknown, so we lose our signed bounds
12777 * 2) it's known negative, thus the unsigned bounds capture the
12779 * 3) the signed bounds cross zero, so they tell us nothing
12781 * If the value in dst_reg is known nonnegative, then again the
12782 * unsigned bounds capture the signed bounds.
12783 * Thus, in all cases it suffices to blow away our signed bounds
12784 * and rely on inferring new ones from the unsigned bounds and
12785 * var_off of the result.
12787 dst_reg->smin_value = S64_MIN;
12788 dst_reg->smax_value = S64_MAX;
12789 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
12790 dst_reg->umin_value >>= umax_val;
12791 dst_reg->umax_value >>= umin_val;
12793 /* Its not easy to operate on alu32 bounds here because it depends
12794 * on bits being shifted in. Take easy way out and mark unbounded
12795 * so we can recalculate later from tnum.
12797 __mark_reg32_unbounded(dst_reg);
12798 __update_reg_bounds(dst_reg);
12801 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
12802 struct bpf_reg_state *src_reg)
12804 u64 umin_val = src_reg->u32_min_value;
12806 /* Upon reaching here, src_known is true and
12807 * umax_val is equal to umin_val.
12809 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
12810 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
12812 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
12814 /* blow away the dst_reg umin_value/umax_value and rely on
12815 * dst_reg var_off to refine the result.
12817 dst_reg->u32_min_value = 0;
12818 dst_reg->u32_max_value = U32_MAX;
12820 __mark_reg64_unbounded(dst_reg);
12821 __update_reg32_bounds(dst_reg);
12824 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
12825 struct bpf_reg_state *src_reg)
12827 u64 umin_val = src_reg->umin_value;
12829 /* Upon reaching here, src_known is true and umax_val is equal
12832 dst_reg->smin_value >>= umin_val;
12833 dst_reg->smax_value >>= umin_val;
12835 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
12837 /* blow away the dst_reg umin_value/umax_value and rely on
12838 * dst_reg var_off to refine the result.
12840 dst_reg->umin_value = 0;
12841 dst_reg->umax_value = U64_MAX;
12843 /* Its not easy to operate on alu32 bounds here because it depends
12844 * on bits being shifted in from upper 32-bits. Take easy way out
12845 * and mark unbounded so we can recalculate later from tnum.
12847 __mark_reg32_unbounded(dst_reg);
12848 __update_reg_bounds(dst_reg);
12851 /* WARNING: This function does calculations on 64-bit values, but the actual
12852 * execution may occur on 32-bit values. Therefore, things like bitshifts
12853 * need extra checks in the 32-bit case.
12855 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
12856 struct bpf_insn *insn,
12857 struct bpf_reg_state *dst_reg,
12858 struct bpf_reg_state src_reg)
12860 struct bpf_reg_state *regs = cur_regs(env);
12861 u8 opcode = BPF_OP(insn->code);
12863 s64 smin_val, smax_val;
12864 u64 umin_val, umax_val;
12865 s32 s32_min_val, s32_max_val;
12866 u32 u32_min_val, u32_max_val;
12867 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
12868 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
12871 smin_val = src_reg.smin_value;
12872 smax_val = src_reg.smax_value;
12873 umin_val = src_reg.umin_value;
12874 umax_val = src_reg.umax_value;
12876 s32_min_val = src_reg.s32_min_value;
12877 s32_max_val = src_reg.s32_max_value;
12878 u32_min_val = src_reg.u32_min_value;
12879 u32_max_val = src_reg.u32_max_value;
12882 src_known = tnum_subreg_is_const(src_reg.var_off);
12884 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
12885 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
12886 /* Taint dst register if offset had invalid bounds
12887 * derived from e.g. dead branches.
12889 __mark_reg_unknown(env, dst_reg);
12893 src_known = tnum_is_const(src_reg.var_off);
12895 (smin_val != smax_val || umin_val != umax_val)) ||
12896 smin_val > smax_val || umin_val > umax_val) {
12897 /* Taint dst register if offset had invalid bounds
12898 * derived from e.g. dead branches.
12900 __mark_reg_unknown(env, dst_reg);
12906 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
12907 __mark_reg_unknown(env, dst_reg);
12911 if (sanitize_needed(opcode)) {
12912 ret = sanitize_val_alu(env, insn);
12914 return sanitize_err(env, insn, ret, NULL, NULL);
12917 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
12918 * There are two classes of instructions: The first class we track both
12919 * alu32 and alu64 sign/unsigned bounds independently this provides the
12920 * greatest amount of precision when alu operations are mixed with jmp32
12921 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
12922 * and BPF_OR. This is possible because these ops have fairly easy to
12923 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
12924 * See alu32 verifier tests for examples. The second class of
12925 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
12926 * with regards to tracking sign/unsigned bounds because the bits may
12927 * cross subreg boundaries in the alu64 case. When this happens we mark
12928 * the reg unbounded in the subreg bound space and use the resulting
12929 * tnum to calculate an approximation of the sign/unsigned bounds.
12933 scalar32_min_max_add(dst_reg, &src_reg);
12934 scalar_min_max_add(dst_reg, &src_reg);
12935 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
12938 scalar32_min_max_sub(dst_reg, &src_reg);
12939 scalar_min_max_sub(dst_reg, &src_reg);
12940 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
12943 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
12944 scalar32_min_max_mul(dst_reg, &src_reg);
12945 scalar_min_max_mul(dst_reg, &src_reg);
12948 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
12949 scalar32_min_max_and(dst_reg, &src_reg);
12950 scalar_min_max_and(dst_reg, &src_reg);
12953 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
12954 scalar32_min_max_or(dst_reg, &src_reg);
12955 scalar_min_max_or(dst_reg, &src_reg);
12958 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
12959 scalar32_min_max_xor(dst_reg, &src_reg);
12960 scalar_min_max_xor(dst_reg, &src_reg);
12963 if (umax_val >= insn_bitness) {
12964 /* Shifts greater than 31 or 63 are undefined.
12965 * This includes shifts by a negative number.
12967 mark_reg_unknown(env, regs, insn->dst_reg);
12971 scalar32_min_max_lsh(dst_reg, &src_reg);
12973 scalar_min_max_lsh(dst_reg, &src_reg);
12976 if (umax_val >= insn_bitness) {
12977 /* Shifts greater than 31 or 63 are undefined.
12978 * This includes shifts by a negative number.
12980 mark_reg_unknown(env, regs, insn->dst_reg);
12984 scalar32_min_max_rsh(dst_reg, &src_reg);
12986 scalar_min_max_rsh(dst_reg, &src_reg);
12989 if (umax_val >= insn_bitness) {
12990 /* Shifts greater than 31 or 63 are undefined.
12991 * This includes shifts by a negative number.
12993 mark_reg_unknown(env, regs, insn->dst_reg);
12997 scalar32_min_max_arsh(dst_reg, &src_reg);
12999 scalar_min_max_arsh(dst_reg, &src_reg);
13002 mark_reg_unknown(env, regs, insn->dst_reg);
13006 /* ALU32 ops are zero extended into 64bit register */
13008 zext_32_to_64(dst_reg);
13009 reg_bounds_sync(dst_reg);
13013 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
13016 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
13017 struct bpf_insn *insn)
13019 struct bpf_verifier_state *vstate = env->cur_state;
13020 struct bpf_func_state *state = vstate->frame[vstate->curframe];
13021 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
13022 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
13023 u8 opcode = BPF_OP(insn->code);
13026 dst_reg = ®s[insn->dst_reg];
13028 if (dst_reg->type != SCALAR_VALUE)
13031 /* Make sure ID is cleared otherwise dst_reg min/max could be
13032 * incorrectly propagated into other registers by find_equal_scalars()
13035 if (BPF_SRC(insn->code) == BPF_X) {
13036 src_reg = ®s[insn->src_reg];
13037 if (src_reg->type != SCALAR_VALUE) {
13038 if (dst_reg->type != SCALAR_VALUE) {
13039 /* Combining two pointers by any ALU op yields
13040 * an arbitrary scalar. Disallow all math except
13041 * pointer subtraction
13043 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
13044 mark_reg_unknown(env, regs, insn->dst_reg);
13047 verbose(env, "R%d pointer %s pointer prohibited\n",
13049 bpf_alu_string[opcode >> 4]);
13052 /* scalar += pointer
13053 * This is legal, but we have to reverse our
13054 * src/dest handling in computing the range
13056 err = mark_chain_precision(env, insn->dst_reg);
13059 return adjust_ptr_min_max_vals(env, insn,
13062 } else if (ptr_reg) {
13063 /* pointer += scalar */
13064 err = mark_chain_precision(env, insn->src_reg);
13067 return adjust_ptr_min_max_vals(env, insn,
13069 } else if (dst_reg->precise) {
13070 /* if dst_reg is precise, src_reg should be precise as well */
13071 err = mark_chain_precision(env, insn->src_reg);
13076 /* Pretend the src is a reg with a known value, since we only
13077 * need to be able to read from this state.
13079 off_reg.type = SCALAR_VALUE;
13080 __mark_reg_known(&off_reg, insn->imm);
13081 src_reg = &off_reg;
13082 if (ptr_reg) /* pointer += K */
13083 return adjust_ptr_min_max_vals(env, insn,
13087 /* Got here implies adding two SCALAR_VALUEs */
13088 if (WARN_ON_ONCE(ptr_reg)) {
13089 print_verifier_state(env, state, true);
13090 verbose(env, "verifier internal error: unexpected ptr_reg\n");
13093 if (WARN_ON(!src_reg)) {
13094 print_verifier_state(env, state, true);
13095 verbose(env, "verifier internal error: no src_reg\n");
13098 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
13101 /* check validity of 32-bit and 64-bit arithmetic operations */
13102 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
13104 struct bpf_reg_state *regs = cur_regs(env);
13105 u8 opcode = BPF_OP(insn->code);
13108 if (opcode == BPF_END || opcode == BPF_NEG) {
13109 if (opcode == BPF_NEG) {
13110 if (BPF_SRC(insn->code) != BPF_K ||
13111 insn->src_reg != BPF_REG_0 ||
13112 insn->off != 0 || insn->imm != 0) {
13113 verbose(env, "BPF_NEG uses reserved fields\n");
13117 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
13118 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
13119 (BPF_CLASS(insn->code) == BPF_ALU64 &&
13120 BPF_SRC(insn->code) != BPF_TO_LE)) {
13121 verbose(env, "BPF_END uses reserved fields\n");
13126 /* check src operand */
13127 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
13131 if (is_pointer_value(env, insn->dst_reg)) {
13132 verbose(env, "R%d pointer arithmetic prohibited\n",
13137 /* check dest operand */
13138 err = check_reg_arg(env, insn->dst_reg, DST_OP);
13142 } else if (opcode == BPF_MOV) {
13144 if (BPF_SRC(insn->code) == BPF_X) {
13145 if (insn->imm != 0) {
13146 verbose(env, "BPF_MOV uses reserved fields\n");
13150 if (BPF_CLASS(insn->code) == BPF_ALU) {
13151 if (insn->off != 0 && insn->off != 8 && insn->off != 16) {
13152 verbose(env, "BPF_MOV uses reserved fields\n");
13156 if (insn->off != 0 && insn->off != 8 && insn->off != 16 &&
13158 verbose(env, "BPF_MOV uses reserved fields\n");
13163 /* check src operand */
13164 err = check_reg_arg(env, insn->src_reg, SRC_OP);
13168 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
13169 verbose(env, "BPF_MOV uses reserved fields\n");
13174 /* check dest operand, mark as required later */
13175 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
13179 if (BPF_SRC(insn->code) == BPF_X) {
13180 struct bpf_reg_state *src_reg = regs + insn->src_reg;
13181 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
13182 bool need_id = src_reg->type == SCALAR_VALUE && !src_reg->id &&
13183 !tnum_is_const(src_reg->var_off);
13185 if (BPF_CLASS(insn->code) == BPF_ALU64) {
13186 if (insn->off == 0) {
13188 * copy register state to dest reg
13191 /* Assign src and dst registers the same ID
13192 * that will be used by find_equal_scalars()
13193 * to propagate min/max range.
13195 src_reg->id = ++env->id_gen;
13196 copy_register_state(dst_reg, src_reg);
13197 dst_reg->live |= REG_LIVE_WRITTEN;
13198 dst_reg->subreg_def = DEF_NOT_SUBREG;
13200 /* case: R1 = (s8, s16 s32)R2 */
13201 if (is_pointer_value(env, insn->src_reg)) {
13203 "R%d sign-extension part of pointer\n",
13206 } else if (src_reg->type == SCALAR_VALUE) {
13209 no_sext = src_reg->umax_value < (1ULL << (insn->off - 1));
13210 if (no_sext && need_id)
13211 src_reg->id = ++env->id_gen;
13212 copy_register_state(dst_reg, src_reg);
13215 coerce_reg_to_size_sx(dst_reg, insn->off >> 3);
13216 dst_reg->live |= REG_LIVE_WRITTEN;
13217 dst_reg->subreg_def = DEF_NOT_SUBREG;
13219 mark_reg_unknown(env, regs, insn->dst_reg);
13223 /* R1 = (u32) R2 */
13224 if (is_pointer_value(env, insn->src_reg)) {
13226 "R%d partial copy of pointer\n",
13229 } else if (src_reg->type == SCALAR_VALUE) {
13230 if (insn->off == 0) {
13231 bool is_src_reg_u32 = src_reg->umax_value <= U32_MAX;
13233 if (is_src_reg_u32 && need_id)
13234 src_reg->id = ++env->id_gen;
13235 copy_register_state(dst_reg, src_reg);
13236 /* Make sure ID is cleared if src_reg is not in u32
13237 * range otherwise dst_reg min/max could be incorrectly
13238 * propagated into src_reg by find_equal_scalars()
13240 if (!is_src_reg_u32)
13242 dst_reg->live |= REG_LIVE_WRITTEN;
13243 dst_reg->subreg_def = env->insn_idx + 1;
13245 /* case: W1 = (s8, s16)W2 */
13246 bool no_sext = src_reg->umax_value < (1ULL << (insn->off - 1));
13248 if (no_sext && need_id)
13249 src_reg->id = ++env->id_gen;
13250 copy_register_state(dst_reg, src_reg);
13253 dst_reg->live |= REG_LIVE_WRITTEN;
13254 dst_reg->subreg_def = env->insn_idx + 1;
13255 coerce_subreg_to_size_sx(dst_reg, insn->off >> 3);
13258 mark_reg_unknown(env, regs,
13261 zext_32_to_64(dst_reg);
13262 reg_bounds_sync(dst_reg);
13266 * remember the value we stored into this reg
13268 /* clear any state __mark_reg_known doesn't set */
13269 mark_reg_unknown(env, regs, insn->dst_reg);
13270 regs[insn->dst_reg].type = SCALAR_VALUE;
13271 if (BPF_CLASS(insn->code) == BPF_ALU64) {
13272 __mark_reg_known(regs + insn->dst_reg,
13275 __mark_reg_known(regs + insn->dst_reg,
13280 } else if (opcode > BPF_END) {
13281 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
13284 } else { /* all other ALU ops: and, sub, xor, add, ... */
13286 if (BPF_SRC(insn->code) == BPF_X) {
13287 if (insn->imm != 0 || insn->off > 1 ||
13288 (insn->off == 1 && opcode != BPF_MOD && opcode != BPF_DIV)) {
13289 verbose(env, "BPF_ALU uses reserved fields\n");
13292 /* check src1 operand */
13293 err = check_reg_arg(env, insn->src_reg, SRC_OP);
13297 if (insn->src_reg != BPF_REG_0 || insn->off > 1 ||
13298 (insn->off == 1 && opcode != BPF_MOD && opcode != BPF_DIV)) {
13299 verbose(env, "BPF_ALU uses reserved fields\n");
13304 /* check src2 operand */
13305 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
13309 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
13310 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
13311 verbose(env, "div by zero\n");
13315 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
13316 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
13317 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
13319 if (insn->imm < 0 || insn->imm >= size) {
13320 verbose(env, "invalid shift %d\n", insn->imm);
13325 /* check dest operand */
13326 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
13330 return adjust_reg_min_max_vals(env, insn);
13336 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
13337 struct bpf_reg_state *dst_reg,
13338 enum bpf_reg_type type,
13339 bool range_right_open)
13341 struct bpf_func_state *state;
13342 struct bpf_reg_state *reg;
13345 if (dst_reg->off < 0 ||
13346 (dst_reg->off == 0 && range_right_open))
13347 /* This doesn't give us any range */
13350 if (dst_reg->umax_value > MAX_PACKET_OFF ||
13351 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
13352 /* Risk of overflow. For instance, ptr + (1<<63) may be less
13353 * than pkt_end, but that's because it's also less than pkt.
13357 new_range = dst_reg->off;
13358 if (range_right_open)
13361 /* Examples for register markings:
13363 * pkt_data in dst register:
13367 * if (r2 > pkt_end) goto <handle exception>
13372 * if (r2 < pkt_end) goto <access okay>
13373 * <handle exception>
13376 * r2 == dst_reg, pkt_end == src_reg
13377 * r2=pkt(id=n,off=8,r=0)
13378 * r3=pkt(id=n,off=0,r=0)
13380 * pkt_data in src register:
13384 * if (pkt_end >= r2) goto <access okay>
13385 * <handle exception>
13389 * if (pkt_end <= r2) goto <handle exception>
13393 * pkt_end == dst_reg, r2 == src_reg
13394 * r2=pkt(id=n,off=8,r=0)
13395 * r3=pkt(id=n,off=0,r=0)
13397 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
13398 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
13399 * and [r3, r3 + 8-1) respectively is safe to access depending on
13403 /* If our ids match, then we must have the same max_value. And we
13404 * don't care about the other reg's fixed offset, since if it's too big
13405 * the range won't allow anything.
13406 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
13408 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
13409 if (reg->type == type && reg->id == dst_reg->id)
13410 /* keep the maximum range already checked */
13411 reg->range = max(reg->range, new_range);
13415 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
13417 struct tnum subreg = tnum_subreg(reg->var_off);
13418 s32 sval = (s32)val;
13422 if (tnum_is_const(subreg))
13423 return !!tnum_equals_const(subreg, val);
13424 else if (val < reg->u32_min_value || val > reg->u32_max_value)
13428 if (tnum_is_const(subreg))
13429 return !tnum_equals_const(subreg, val);
13430 else if (val < reg->u32_min_value || val > reg->u32_max_value)
13434 if ((~subreg.mask & subreg.value) & val)
13436 if (!((subreg.mask | subreg.value) & val))
13440 if (reg->u32_min_value > val)
13442 else if (reg->u32_max_value <= val)
13446 if (reg->s32_min_value > sval)
13448 else if (reg->s32_max_value <= sval)
13452 if (reg->u32_max_value < val)
13454 else if (reg->u32_min_value >= val)
13458 if (reg->s32_max_value < sval)
13460 else if (reg->s32_min_value >= sval)
13464 if (reg->u32_min_value >= val)
13466 else if (reg->u32_max_value < val)
13470 if (reg->s32_min_value >= sval)
13472 else if (reg->s32_max_value < sval)
13476 if (reg->u32_max_value <= val)
13478 else if (reg->u32_min_value > val)
13482 if (reg->s32_max_value <= sval)
13484 else if (reg->s32_min_value > sval)
13493 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
13495 s64 sval = (s64)val;
13499 if (tnum_is_const(reg->var_off))
13500 return !!tnum_equals_const(reg->var_off, val);
13501 else if (val < reg->umin_value || val > reg->umax_value)
13505 if (tnum_is_const(reg->var_off))
13506 return !tnum_equals_const(reg->var_off, val);
13507 else if (val < reg->umin_value || val > reg->umax_value)
13511 if ((~reg->var_off.mask & reg->var_off.value) & val)
13513 if (!((reg->var_off.mask | reg->var_off.value) & val))
13517 if (reg->umin_value > val)
13519 else if (reg->umax_value <= val)
13523 if (reg->smin_value > sval)
13525 else if (reg->smax_value <= sval)
13529 if (reg->umax_value < val)
13531 else if (reg->umin_value >= val)
13535 if (reg->smax_value < sval)
13537 else if (reg->smin_value >= sval)
13541 if (reg->umin_value >= val)
13543 else if (reg->umax_value < val)
13547 if (reg->smin_value >= sval)
13549 else if (reg->smax_value < sval)
13553 if (reg->umax_value <= val)
13555 else if (reg->umin_value > val)
13559 if (reg->smax_value <= sval)
13561 else if (reg->smin_value > sval)
13569 /* compute branch direction of the expression "if (reg opcode val) goto target;"
13571 * 1 - branch will be taken and "goto target" will be executed
13572 * 0 - branch will not be taken and fall-through to next insn
13573 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
13576 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
13579 if (__is_pointer_value(false, reg)) {
13580 if (!reg_not_null(reg))
13583 /* If pointer is valid tests against zero will fail so we can
13584 * use this to direct branch taken.
13600 return is_branch32_taken(reg, val, opcode);
13601 return is_branch64_taken(reg, val, opcode);
13604 static int flip_opcode(u32 opcode)
13606 /* How can we transform "a <op> b" into "b <op> a"? */
13607 static const u8 opcode_flip[16] = {
13608 /* these stay the same */
13609 [BPF_JEQ >> 4] = BPF_JEQ,
13610 [BPF_JNE >> 4] = BPF_JNE,
13611 [BPF_JSET >> 4] = BPF_JSET,
13612 /* these swap "lesser" and "greater" (L and G in the opcodes) */
13613 [BPF_JGE >> 4] = BPF_JLE,
13614 [BPF_JGT >> 4] = BPF_JLT,
13615 [BPF_JLE >> 4] = BPF_JGE,
13616 [BPF_JLT >> 4] = BPF_JGT,
13617 [BPF_JSGE >> 4] = BPF_JSLE,
13618 [BPF_JSGT >> 4] = BPF_JSLT,
13619 [BPF_JSLE >> 4] = BPF_JSGE,
13620 [BPF_JSLT >> 4] = BPF_JSGT
13622 return opcode_flip[opcode >> 4];
13625 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
13626 struct bpf_reg_state *src_reg,
13629 struct bpf_reg_state *pkt;
13631 if (src_reg->type == PTR_TO_PACKET_END) {
13633 } else if (dst_reg->type == PTR_TO_PACKET_END) {
13635 opcode = flip_opcode(opcode);
13640 if (pkt->range >= 0)
13645 /* pkt <= pkt_end */
13648 /* pkt > pkt_end */
13649 if (pkt->range == BEYOND_PKT_END)
13650 /* pkt has at last one extra byte beyond pkt_end */
13651 return opcode == BPF_JGT;
13654 /* pkt < pkt_end */
13657 /* pkt >= pkt_end */
13658 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
13659 return opcode == BPF_JGE;
13665 /* Adjusts the register min/max values in the case that the dst_reg is the
13666 * variable register that we are working on, and src_reg is a constant or we're
13667 * simply doing a BPF_K check.
13668 * In JEQ/JNE cases we also adjust the var_off values.
13670 static void reg_set_min_max(struct bpf_reg_state *true_reg,
13671 struct bpf_reg_state *false_reg,
13672 u64 val, u32 val32,
13673 u8 opcode, bool is_jmp32)
13675 struct tnum false_32off = tnum_subreg(false_reg->var_off);
13676 struct tnum false_64off = false_reg->var_off;
13677 struct tnum true_32off = tnum_subreg(true_reg->var_off);
13678 struct tnum true_64off = true_reg->var_off;
13679 s64 sval = (s64)val;
13680 s32 sval32 = (s32)val32;
13682 /* If the dst_reg is a pointer, we can't learn anything about its
13683 * variable offset from the compare (unless src_reg were a pointer into
13684 * the same object, but we don't bother with that.
13685 * Since false_reg and true_reg have the same type by construction, we
13686 * only need to check one of them for pointerness.
13688 if (__is_pointer_value(false, false_reg))
13692 /* JEQ/JNE comparison doesn't change the register equivalence.
13695 * if (r1 == 42) goto label;
13697 * label: // here both r1 and r2 are known to be 42.
13699 * Hence when marking register as known preserve it's ID.
13703 __mark_reg32_known(true_reg, val32);
13704 true_32off = tnum_subreg(true_reg->var_off);
13706 ___mark_reg_known(true_reg, val);
13707 true_64off = true_reg->var_off;
13712 __mark_reg32_known(false_reg, val32);
13713 false_32off = tnum_subreg(false_reg->var_off);
13715 ___mark_reg_known(false_reg, val);
13716 false_64off = false_reg->var_off;
13721 false_32off = tnum_and(false_32off, tnum_const(~val32));
13722 if (is_power_of_2(val32))
13723 true_32off = tnum_or(true_32off,
13724 tnum_const(val32));
13726 false_64off = tnum_and(false_64off, tnum_const(~val));
13727 if (is_power_of_2(val))
13728 true_64off = tnum_or(true_64off,
13736 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
13737 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
13739 false_reg->u32_max_value = min(false_reg->u32_max_value,
13741 true_reg->u32_min_value = max(true_reg->u32_min_value,
13744 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
13745 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
13747 false_reg->umax_value = min(false_reg->umax_value, false_umax);
13748 true_reg->umin_value = max(true_reg->umin_value, true_umin);
13756 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
13757 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
13759 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
13760 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
13762 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
13763 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
13765 false_reg->smax_value = min(false_reg->smax_value, false_smax);
13766 true_reg->smin_value = max(true_reg->smin_value, true_smin);
13774 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
13775 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
13777 false_reg->u32_min_value = max(false_reg->u32_min_value,
13779 true_reg->u32_max_value = min(true_reg->u32_max_value,
13782 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
13783 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
13785 false_reg->umin_value = max(false_reg->umin_value, false_umin);
13786 true_reg->umax_value = min(true_reg->umax_value, true_umax);
13794 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
13795 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
13797 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
13798 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
13800 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
13801 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
13803 false_reg->smin_value = max(false_reg->smin_value, false_smin);
13804 true_reg->smax_value = min(true_reg->smax_value, true_smax);
13813 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
13814 tnum_subreg(false_32off));
13815 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
13816 tnum_subreg(true_32off));
13817 __reg_combine_32_into_64(false_reg);
13818 __reg_combine_32_into_64(true_reg);
13820 false_reg->var_off = false_64off;
13821 true_reg->var_off = true_64off;
13822 __reg_combine_64_into_32(false_reg);
13823 __reg_combine_64_into_32(true_reg);
13827 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
13828 * the variable reg.
13830 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
13831 struct bpf_reg_state *false_reg,
13832 u64 val, u32 val32,
13833 u8 opcode, bool is_jmp32)
13835 opcode = flip_opcode(opcode);
13836 /* This uses zero as "not present in table"; luckily the zero opcode,
13837 * BPF_JA, can't get here.
13840 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
13843 /* Regs are known to be equal, so intersect their min/max/var_off */
13844 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
13845 struct bpf_reg_state *dst_reg)
13847 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
13848 dst_reg->umin_value);
13849 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
13850 dst_reg->umax_value);
13851 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
13852 dst_reg->smin_value);
13853 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
13854 dst_reg->smax_value);
13855 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
13857 reg_bounds_sync(src_reg);
13858 reg_bounds_sync(dst_reg);
13861 static void reg_combine_min_max(struct bpf_reg_state *true_src,
13862 struct bpf_reg_state *true_dst,
13863 struct bpf_reg_state *false_src,
13864 struct bpf_reg_state *false_dst,
13869 __reg_combine_min_max(true_src, true_dst);
13872 __reg_combine_min_max(false_src, false_dst);
13877 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
13878 struct bpf_reg_state *reg, u32 id,
13881 if (type_may_be_null(reg->type) && reg->id == id &&
13882 (is_rcu_reg(reg) || !WARN_ON_ONCE(!reg->id))) {
13883 /* Old offset (both fixed and variable parts) should have been
13884 * known-zero, because we don't allow pointer arithmetic on
13885 * pointers that might be NULL. If we see this happening, don't
13886 * convert the register.
13888 * But in some cases, some helpers that return local kptrs
13889 * advance offset for the returned pointer. In those cases, it
13890 * is fine to expect to see reg->off.
13892 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value || !tnum_equals_const(reg->var_off, 0)))
13894 if (!(type_is_ptr_alloc_obj(reg->type) || type_is_non_owning_ref(reg->type)) &&
13895 WARN_ON_ONCE(reg->off))
13899 reg->type = SCALAR_VALUE;
13900 /* We don't need id and ref_obj_id from this point
13901 * onwards anymore, thus we should better reset it,
13902 * so that state pruning has chances to take effect.
13905 reg->ref_obj_id = 0;
13910 mark_ptr_not_null_reg(reg);
13912 if (!reg_may_point_to_spin_lock(reg)) {
13913 /* For not-NULL ptr, reg->ref_obj_id will be reset
13914 * in release_reference().
13916 * reg->id is still used by spin_lock ptr. Other
13917 * than spin_lock ptr type, reg->id can be reset.
13924 /* The logic is similar to find_good_pkt_pointers(), both could eventually
13925 * be folded together at some point.
13927 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
13930 struct bpf_func_state *state = vstate->frame[vstate->curframe];
13931 struct bpf_reg_state *regs = state->regs, *reg;
13932 u32 ref_obj_id = regs[regno].ref_obj_id;
13933 u32 id = regs[regno].id;
13935 if (ref_obj_id && ref_obj_id == id && is_null)
13936 /* regs[regno] is in the " == NULL" branch.
13937 * No one could have freed the reference state before
13938 * doing the NULL check.
13940 WARN_ON_ONCE(release_reference_state(state, id));
13942 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
13943 mark_ptr_or_null_reg(state, reg, id, is_null);
13947 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
13948 struct bpf_reg_state *dst_reg,
13949 struct bpf_reg_state *src_reg,
13950 struct bpf_verifier_state *this_branch,
13951 struct bpf_verifier_state *other_branch)
13953 if (BPF_SRC(insn->code) != BPF_X)
13956 /* Pointers are always 64-bit. */
13957 if (BPF_CLASS(insn->code) == BPF_JMP32)
13960 switch (BPF_OP(insn->code)) {
13962 if ((dst_reg->type == PTR_TO_PACKET &&
13963 src_reg->type == PTR_TO_PACKET_END) ||
13964 (dst_reg->type == PTR_TO_PACKET_META &&
13965 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13966 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
13967 find_good_pkt_pointers(this_branch, dst_reg,
13968 dst_reg->type, false);
13969 mark_pkt_end(other_branch, insn->dst_reg, true);
13970 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
13971 src_reg->type == PTR_TO_PACKET) ||
13972 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
13973 src_reg->type == PTR_TO_PACKET_META)) {
13974 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
13975 find_good_pkt_pointers(other_branch, src_reg,
13976 src_reg->type, true);
13977 mark_pkt_end(this_branch, insn->src_reg, false);
13983 if ((dst_reg->type == PTR_TO_PACKET &&
13984 src_reg->type == PTR_TO_PACKET_END) ||
13985 (dst_reg->type == PTR_TO_PACKET_META &&
13986 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13987 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
13988 find_good_pkt_pointers(other_branch, dst_reg,
13989 dst_reg->type, true);
13990 mark_pkt_end(this_branch, insn->dst_reg, false);
13991 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
13992 src_reg->type == PTR_TO_PACKET) ||
13993 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
13994 src_reg->type == PTR_TO_PACKET_META)) {
13995 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
13996 find_good_pkt_pointers(this_branch, src_reg,
13997 src_reg->type, false);
13998 mark_pkt_end(other_branch, insn->src_reg, true);
14004 if ((dst_reg->type == PTR_TO_PACKET &&
14005 src_reg->type == PTR_TO_PACKET_END) ||
14006 (dst_reg->type == PTR_TO_PACKET_META &&
14007 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
14008 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
14009 find_good_pkt_pointers(this_branch, dst_reg,
14010 dst_reg->type, true);
14011 mark_pkt_end(other_branch, insn->dst_reg, false);
14012 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
14013 src_reg->type == PTR_TO_PACKET) ||
14014 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
14015 src_reg->type == PTR_TO_PACKET_META)) {
14016 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
14017 find_good_pkt_pointers(other_branch, src_reg,
14018 src_reg->type, false);
14019 mark_pkt_end(this_branch, insn->src_reg, true);
14025 if ((dst_reg->type == PTR_TO_PACKET &&
14026 src_reg->type == PTR_TO_PACKET_END) ||
14027 (dst_reg->type == PTR_TO_PACKET_META &&
14028 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
14029 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
14030 find_good_pkt_pointers(other_branch, dst_reg,
14031 dst_reg->type, false);
14032 mark_pkt_end(this_branch, insn->dst_reg, true);
14033 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
14034 src_reg->type == PTR_TO_PACKET) ||
14035 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
14036 src_reg->type == PTR_TO_PACKET_META)) {
14037 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
14038 find_good_pkt_pointers(this_branch, src_reg,
14039 src_reg->type, true);
14040 mark_pkt_end(other_branch, insn->src_reg, false);
14052 static void find_equal_scalars(struct bpf_verifier_state *vstate,
14053 struct bpf_reg_state *known_reg)
14055 struct bpf_func_state *state;
14056 struct bpf_reg_state *reg;
14058 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
14059 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
14060 copy_register_state(reg, known_reg);
14064 static int check_cond_jmp_op(struct bpf_verifier_env *env,
14065 struct bpf_insn *insn, int *insn_idx)
14067 struct bpf_verifier_state *this_branch = env->cur_state;
14068 struct bpf_verifier_state *other_branch;
14069 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
14070 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
14071 struct bpf_reg_state *eq_branch_regs;
14072 u8 opcode = BPF_OP(insn->code);
14077 /* Only conditional jumps are expected to reach here. */
14078 if (opcode == BPF_JA || opcode > BPF_JSLE) {
14079 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
14083 /* check src2 operand */
14084 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
14088 dst_reg = ®s[insn->dst_reg];
14089 if (BPF_SRC(insn->code) == BPF_X) {
14090 if (insn->imm != 0) {
14091 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
14095 /* check src1 operand */
14096 err = check_reg_arg(env, insn->src_reg, SRC_OP);
14100 src_reg = ®s[insn->src_reg];
14101 if (!(reg_is_pkt_pointer_any(dst_reg) && reg_is_pkt_pointer_any(src_reg)) &&
14102 is_pointer_value(env, insn->src_reg)) {
14103 verbose(env, "R%d pointer comparison prohibited\n",
14108 if (insn->src_reg != BPF_REG_0) {
14109 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
14114 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
14116 if (BPF_SRC(insn->code) == BPF_K) {
14117 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
14118 } else if (src_reg->type == SCALAR_VALUE &&
14119 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
14120 pred = is_branch_taken(dst_reg,
14121 tnum_subreg(src_reg->var_off).value,
14124 } else if (src_reg->type == SCALAR_VALUE &&
14125 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
14126 pred = is_branch_taken(dst_reg,
14127 src_reg->var_off.value,
14130 } else if (dst_reg->type == SCALAR_VALUE &&
14131 is_jmp32 && tnum_is_const(tnum_subreg(dst_reg->var_off))) {
14132 pred = is_branch_taken(src_reg,
14133 tnum_subreg(dst_reg->var_off).value,
14134 flip_opcode(opcode),
14136 } else if (dst_reg->type == SCALAR_VALUE &&
14137 !is_jmp32 && tnum_is_const(dst_reg->var_off)) {
14138 pred = is_branch_taken(src_reg,
14139 dst_reg->var_off.value,
14140 flip_opcode(opcode),
14142 } else if (reg_is_pkt_pointer_any(dst_reg) &&
14143 reg_is_pkt_pointer_any(src_reg) &&
14145 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
14149 /* If we get here with a dst_reg pointer type it is because
14150 * above is_branch_taken() special cased the 0 comparison.
14152 if (!__is_pointer_value(false, dst_reg))
14153 err = mark_chain_precision(env, insn->dst_reg);
14154 if (BPF_SRC(insn->code) == BPF_X && !err &&
14155 !__is_pointer_value(false, src_reg))
14156 err = mark_chain_precision(env, insn->src_reg);
14162 /* Only follow the goto, ignore fall-through. If needed, push
14163 * the fall-through branch for simulation under speculative
14166 if (!env->bypass_spec_v1 &&
14167 !sanitize_speculative_path(env, insn, *insn_idx + 1,
14170 if (env->log.level & BPF_LOG_LEVEL)
14171 print_insn_state(env, this_branch->frame[this_branch->curframe]);
14172 *insn_idx += insn->off;
14174 } else if (pred == 0) {
14175 /* Only follow the fall-through branch, since that's where the
14176 * program will go. If needed, push the goto branch for
14177 * simulation under speculative execution.
14179 if (!env->bypass_spec_v1 &&
14180 !sanitize_speculative_path(env, insn,
14181 *insn_idx + insn->off + 1,
14184 if (env->log.level & BPF_LOG_LEVEL)
14185 print_insn_state(env, this_branch->frame[this_branch->curframe]);
14189 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
14193 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
14195 /* detect if we are comparing against a constant value so we can adjust
14196 * our min/max values for our dst register.
14197 * this is only legit if both are scalars (or pointers to the same
14198 * object, I suppose, see the PTR_MAYBE_NULL related if block below),
14199 * because otherwise the different base pointers mean the offsets aren't
14202 if (BPF_SRC(insn->code) == BPF_X) {
14203 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
14205 if (dst_reg->type == SCALAR_VALUE &&
14206 src_reg->type == SCALAR_VALUE) {
14207 if (tnum_is_const(src_reg->var_off) ||
14209 tnum_is_const(tnum_subreg(src_reg->var_off))))
14210 reg_set_min_max(&other_branch_regs[insn->dst_reg],
14212 src_reg->var_off.value,
14213 tnum_subreg(src_reg->var_off).value,
14215 else if (tnum_is_const(dst_reg->var_off) ||
14217 tnum_is_const(tnum_subreg(dst_reg->var_off))))
14218 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
14220 dst_reg->var_off.value,
14221 tnum_subreg(dst_reg->var_off).value,
14223 else if (!is_jmp32 &&
14224 (opcode == BPF_JEQ || opcode == BPF_JNE))
14225 /* Comparing for equality, we can combine knowledge */
14226 reg_combine_min_max(&other_branch_regs[insn->src_reg],
14227 &other_branch_regs[insn->dst_reg],
14228 src_reg, dst_reg, opcode);
14230 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
14231 find_equal_scalars(this_branch, src_reg);
14232 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
14236 } else if (dst_reg->type == SCALAR_VALUE) {
14237 reg_set_min_max(&other_branch_regs[insn->dst_reg],
14238 dst_reg, insn->imm, (u32)insn->imm,
14242 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
14243 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
14244 find_equal_scalars(this_branch, dst_reg);
14245 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
14248 /* if one pointer register is compared to another pointer
14249 * register check if PTR_MAYBE_NULL could be lifted.
14250 * E.g. register A - maybe null
14251 * register B - not null
14252 * for JNE A, B, ... - A is not null in the false branch;
14253 * for JEQ A, B, ... - A is not null in the true branch.
14255 * Since PTR_TO_BTF_ID points to a kernel struct that does
14256 * not need to be null checked by the BPF program, i.e.,
14257 * could be null even without PTR_MAYBE_NULL marking, so
14258 * only propagate nullness when neither reg is that type.
14260 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_X &&
14261 __is_pointer_value(false, src_reg) && __is_pointer_value(false, dst_reg) &&
14262 type_may_be_null(src_reg->type) != type_may_be_null(dst_reg->type) &&
14263 base_type(src_reg->type) != PTR_TO_BTF_ID &&
14264 base_type(dst_reg->type) != PTR_TO_BTF_ID) {
14265 eq_branch_regs = NULL;
14268 eq_branch_regs = other_branch_regs;
14271 eq_branch_regs = regs;
14277 if (eq_branch_regs) {
14278 if (type_may_be_null(src_reg->type))
14279 mark_ptr_not_null_reg(&eq_branch_regs[insn->src_reg]);
14281 mark_ptr_not_null_reg(&eq_branch_regs[insn->dst_reg]);
14285 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
14286 * NOTE: these optimizations below are related with pointer comparison
14287 * which will never be JMP32.
14289 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
14290 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
14291 type_may_be_null(dst_reg->type)) {
14292 /* Mark all identical registers in each branch as either
14293 * safe or unknown depending R == 0 or R != 0 conditional.
14295 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
14296 opcode == BPF_JNE);
14297 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
14298 opcode == BPF_JEQ);
14299 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
14300 this_branch, other_branch) &&
14301 is_pointer_value(env, insn->dst_reg)) {
14302 verbose(env, "R%d pointer comparison prohibited\n",
14306 if (env->log.level & BPF_LOG_LEVEL)
14307 print_insn_state(env, this_branch->frame[this_branch->curframe]);
14311 /* verify BPF_LD_IMM64 instruction */
14312 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
14314 struct bpf_insn_aux_data *aux = cur_aux(env);
14315 struct bpf_reg_state *regs = cur_regs(env);
14316 struct bpf_reg_state *dst_reg;
14317 struct bpf_map *map;
14320 if (BPF_SIZE(insn->code) != BPF_DW) {
14321 verbose(env, "invalid BPF_LD_IMM insn\n");
14324 if (insn->off != 0) {
14325 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
14329 err = check_reg_arg(env, insn->dst_reg, DST_OP);
14333 dst_reg = ®s[insn->dst_reg];
14334 if (insn->src_reg == 0) {
14335 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
14337 dst_reg->type = SCALAR_VALUE;
14338 __mark_reg_known(®s[insn->dst_reg], imm);
14342 /* All special src_reg cases are listed below. From this point onwards
14343 * we either succeed and assign a corresponding dst_reg->type after
14344 * zeroing the offset, or fail and reject the program.
14346 mark_reg_known_zero(env, regs, insn->dst_reg);
14348 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
14349 dst_reg->type = aux->btf_var.reg_type;
14350 switch (base_type(dst_reg->type)) {
14352 dst_reg->mem_size = aux->btf_var.mem_size;
14354 case PTR_TO_BTF_ID:
14355 dst_reg->btf = aux->btf_var.btf;
14356 dst_reg->btf_id = aux->btf_var.btf_id;
14359 verbose(env, "bpf verifier is misconfigured\n");
14365 if (insn->src_reg == BPF_PSEUDO_FUNC) {
14366 struct bpf_prog_aux *aux = env->prog->aux;
14367 u32 subprogno = find_subprog(env,
14368 env->insn_idx + insn->imm + 1);
14370 if (!aux->func_info) {
14371 verbose(env, "missing btf func_info\n");
14374 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
14375 verbose(env, "callback function not static\n");
14379 dst_reg->type = PTR_TO_FUNC;
14380 dst_reg->subprogno = subprogno;
14384 map = env->used_maps[aux->map_index];
14385 dst_reg->map_ptr = map;
14387 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
14388 insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
14389 dst_reg->type = PTR_TO_MAP_VALUE;
14390 dst_reg->off = aux->map_off;
14391 WARN_ON_ONCE(map->max_entries != 1);
14392 /* We want reg->id to be same (0) as map_value is not distinct */
14393 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
14394 insn->src_reg == BPF_PSEUDO_MAP_IDX) {
14395 dst_reg->type = CONST_PTR_TO_MAP;
14397 verbose(env, "bpf verifier is misconfigured\n");
14404 static bool may_access_skb(enum bpf_prog_type type)
14407 case BPF_PROG_TYPE_SOCKET_FILTER:
14408 case BPF_PROG_TYPE_SCHED_CLS:
14409 case BPF_PROG_TYPE_SCHED_ACT:
14416 /* verify safety of LD_ABS|LD_IND instructions:
14417 * - they can only appear in the programs where ctx == skb
14418 * - since they are wrappers of function calls, they scratch R1-R5 registers,
14419 * preserve R6-R9, and store return value into R0
14422 * ctx == skb == R6 == CTX
14425 * SRC == any register
14426 * IMM == 32-bit immediate
14429 * R0 - 8/16/32-bit skb data converted to cpu endianness
14431 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
14433 struct bpf_reg_state *regs = cur_regs(env);
14434 static const int ctx_reg = BPF_REG_6;
14435 u8 mode = BPF_MODE(insn->code);
14438 if (!may_access_skb(resolve_prog_type(env->prog))) {
14439 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
14443 if (!env->ops->gen_ld_abs) {
14444 verbose(env, "bpf verifier is misconfigured\n");
14448 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
14449 BPF_SIZE(insn->code) == BPF_DW ||
14450 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
14451 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
14455 /* check whether implicit source operand (register R6) is readable */
14456 err = check_reg_arg(env, ctx_reg, SRC_OP);
14460 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
14461 * gen_ld_abs() may terminate the program at runtime, leading to
14464 err = check_reference_leak(env);
14466 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
14470 if (env->cur_state->active_lock.ptr) {
14471 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
14475 if (env->cur_state->active_rcu_lock) {
14476 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_rcu_read_lock-ed region\n");
14480 if (regs[ctx_reg].type != PTR_TO_CTX) {
14482 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
14486 if (mode == BPF_IND) {
14487 /* check explicit source operand */
14488 err = check_reg_arg(env, insn->src_reg, SRC_OP);
14493 err = check_ptr_off_reg(env, ®s[ctx_reg], ctx_reg);
14497 /* reset caller saved regs to unreadable */
14498 for (i = 0; i < CALLER_SAVED_REGS; i++) {
14499 mark_reg_not_init(env, regs, caller_saved[i]);
14500 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
14503 /* mark destination R0 register as readable, since it contains
14504 * the value fetched from the packet.
14505 * Already marked as written above.
14507 mark_reg_unknown(env, regs, BPF_REG_0);
14508 /* ld_abs load up to 32-bit skb data. */
14509 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
14513 static int check_return_code(struct bpf_verifier_env *env)
14515 struct tnum enforce_attach_type_range = tnum_unknown;
14516 const struct bpf_prog *prog = env->prog;
14517 struct bpf_reg_state *reg;
14518 struct tnum range = tnum_range(0, 1), const_0 = tnum_const(0);
14519 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
14521 struct bpf_func_state *frame = env->cur_state->frame[0];
14522 const bool is_subprog = frame->subprogno;
14524 /* LSM and struct_ops func-ptr's return type could be "void" */
14526 switch (prog_type) {
14527 case BPF_PROG_TYPE_LSM:
14528 if (prog->expected_attach_type == BPF_LSM_CGROUP)
14529 /* See below, can be 0 or 0-1 depending on hook. */
14532 case BPF_PROG_TYPE_STRUCT_OPS:
14533 if (!prog->aux->attach_func_proto->type)
14541 /* eBPF calling convention is such that R0 is used
14542 * to return the value from eBPF program.
14543 * Make sure that it's readable at this time
14544 * of bpf_exit, which means that program wrote
14545 * something into it earlier
14547 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
14551 if (is_pointer_value(env, BPF_REG_0)) {
14552 verbose(env, "R0 leaks addr as return value\n");
14556 reg = cur_regs(env) + BPF_REG_0;
14558 if (frame->in_async_callback_fn) {
14559 /* enforce return zero from async callbacks like timer */
14560 if (reg->type != SCALAR_VALUE) {
14561 verbose(env, "In async callback the register R0 is not a known value (%s)\n",
14562 reg_type_str(env, reg->type));
14566 if (!tnum_in(const_0, reg->var_off)) {
14567 verbose_invalid_scalar(env, reg, &const_0, "async callback", "R0");
14574 if (reg->type != SCALAR_VALUE) {
14575 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
14576 reg_type_str(env, reg->type));
14582 switch (prog_type) {
14583 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
14584 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
14585 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
14586 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
14587 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
14588 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
14589 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
14590 range = tnum_range(1, 1);
14591 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
14592 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
14593 range = tnum_range(0, 3);
14595 case BPF_PROG_TYPE_CGROUP_SKB:
14596 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
14597 range = tnum_range(0, 3);
14598 enforce_attach_type_range = tnum_range(2, 3);
14601 case BPF_PROG_TYPE_CGROUP_SOCK:
14602 case BPF_PROG_TYPE_SOCK_OPS:
14603 case BPF_PROG_TYPE_CGROUP_DEVICE:
14604 case BPF_PROG_TYPE_CGROUP_SYSCTL:
14605 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
14607 case BPF_PROG_TYPE_RAW_TRACEPOINT:
14608 if (!env->prog->aux->attach_btf_id)
14610 range = tnum_const(0);
14612 case BPF_PROG_TYPE_TRACING:
14613 switch (env->prog->expected_attach_type) {
14614 case BPF_TRACE_FENTRY:
14615 case BPF_TRACE_FEXIT:
14616 range = tnum_const(0);
14618 case BPF_TRACE_RAW_TP:
14619 case BPF_MODIFY_RETURN:
14621 case BPF_TRACE_ITER:
14627 case BPF_PROG_TYPE_SK_LOOKUP:
14628 range = tnum_range(SK_DROP, SK_PASS);
14631 case BPF_PROG_TYPE_LSM:
14632 if (env->prog->expected_attach_type != BPF_LSM_CGROUP) {
14633 /* Regular BPF_PROG_TYPE_LSM programs can return
14638 if (!env->prog->aux->attach_func_proto->type) {
14639 /* Make sure programs that attach to void
14640 * hooks don't try to modify return value.
14642 range = tnum_range(1, 1);
14646 case BPF_PROG_TYPE_NETFILTER:
14647 range = tnum_range(NF_DROP, NF_ACCEPT);
14649 case BPF_PROG_TYPE_EXT:
14650 /* freplace program can return anything as its return value
14651 * depends on the to-be-replaced kernel func or bpf program.
14657 if (reg->type != SCALAR_VALUE) {
14658 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
14659 reg_type_str(env, reg->type));
14663 if (!tnum_in(range, reg->var_off)) {
14664 verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
14665 if (prog->expected_attach_type == BPF_LSM_CGROUP &&
14666 prog_type == BPF_PROG_TYPE_LSM &&
14667 !prog->aux->attach_func_proto->type)
14668 verbose(env, "Note, BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
14672 if (!tnum_is_unknown(enforce_attach_type_range) &&
14673 tnum_in(enforce_attach_type_range, reg->var_off))
14674 env->prog->enforce_expected_attach_type = 1;
14678 /* non-recursive DFS pseudo code
14679 * 1 procedure DFS-iterative(G,v):
14680 * 2 label v as discovered
14681 * 3 let S be a stack
14683 * 5 while S is not empty
14685 * 7 if t is what we're looking for:
14687 * 9 for all edges e in G.adjacentEdges(t) do
14688 * 10 if edge e is already labelled
14689 * 11 continue with the next edge
14690 * 12 w <- G.adjacentVertex(t,e)
14691 * 13 if vertex w is not discovered and not explored
14692 * 14 label e as tree-edge
14693 * 15 label w as discovered
14696 * 18 else if vertex w is discovered
14697 * 19 label e as back-edge
14699 * 21 // vertex w is explored
14700 * 22 label e as forward- or cross-edge
14701 * 23 label t as explored
14705 * 0x10 - discovered
14706 * 0x11 - discovered and fall-through edge labelled
14707 * 0x12 - discovered and fall-through and branch edges labelled
14718 static void mark_prune_point(struct bpf_verifier_env *env, int idx)
14720 env->insn_aux_data[idx].prune_point = true;
14723 static bool is_prune_point(struct bpf_verifier_env *env, int insn_idx)
14725 return env->insn_aux_data[insn_idx].prune_point;
14728 static void mark_force_checkpoint(struct bpf_verifier_env *env, int idx)
14730 env->insn_aux_data[idx].force_checkpoint = true;
14733 static bool is_force_checkpoint(struct bpf_verifier_env *env, int insn_idx)
14735 return env->insn_aux_data[insn_idx].force_checkpoint;
14740 DONE_EXPLORING = 0,
14741 KEEP_EXPLORING = 1,
14744 /* t, w, e - match pseudo-code above:
14745 * t - index of current instruction
14746 * w - next instruction
14749 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
14751 int *insn_stack = env->cfg.insn_stack;
14752 int *insn_state = env->cfg.insn_state;
14754 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
14755 return DONE_EXPLORING;
14757 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
14758 return DONE_EXPLORING;
14760 if (w < 0 || w >= env->prog->len) {
14761 verbose_linfo(env, t, "%d: ", t);
14762 verbose(env, "jump out of range from insn %d to %d\n", t, w);
14767 /* mark branch target for state pruning */
14768 mark_prune_point(env, w);
14769 mark_jmp_point(env, w);
14772 if (insn_state[w] == 0) {
14774 insn_state[t] = DISCOVERED | e;
14775 insn_state[w] = DISCOVERED;
14776 if (env->cfg.cur_stack >= env->prog->len)
14778 insn_stack[env->cfg.cur_stack++] = w;
14779 return KEEP_EXPLORING;
14780 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
14781 if (env->bpf_capable)
14782 return DONE_EXPLORING;
14783 verbose_linfo(env, t, "%d: ", t);
14784 verbose_linfo(env, w, "%d: ", w);
14785 verbose(env, "back-edge from insn %d to %d\n", t, w);
14787 } else if (insn_state[w] == EXPLORED) {
14788 /* forward- or cross-edge */
14789 insn_state[t] = DISCOVERED | e;
14791 verbose(env, "insn state internal bug\n");
14794 return DONE_EXPLORING;
14797 static int visit_func_call_insn(int t, struct bpf_insn *insns,
14798 struct bpf_verifier_env *env,
14803 insn_sz = bpf_is_ldimm64(&insns[t]) ? 2 : 1;
14804 ret = push_insn(t, t + insn_sz, FALLTHROUGH, env);
14808 mark_prune_point(env, t + insn_sz);
14809 /* when we exit from subprog, we need to record non-linear history */
14810 mark_jmp_point(env, t + insn_sz);
14812 if (visit_callee) {
14813 mark_prune_point(env, t);
14814 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env);
14819 /* Visits the instruction at index t and returns one of the following:
14820 * < 0 - an error occurred
14821 * DONE_EXPLORING - the instruction was fully explored
14822 * KEEP_EXPLORING - there is still work to be done before it is fully explored
14824 static int visit_insn(int t, struct bpf_verifier_env *env)
14826 struct bpf_insn *insns = env->prog->insnsi, *insn = &insns[t];
14827 int ret, off, insn_sz;
14829 if (bpf_pseudo_func(insn))
14830 return visit_func_call_insn(t, insns, env, true);
14832 /* All non-branch instructions have a single fall-through edge. */
14833 if (BPF_CLASS(insn->code) != BPF_JMP &&
14834 BPF_CLASS(insn->code) != BPF_JMP32) {
14835 insn_sz = bpf_is_ldimm64(insn) ? 2 : 1;
14836 return push_insn(t, t + insn_sz, FALLTHROUGH, env);
14839 switch (BPF_OP(insn->code)) {
14841 return DONE_EXPLORING;
14844 if (insn->src_reg == 0 && insn->imm == BPF_FUNC_timer_set_callback)
14845 /* Mark this call insn as a prune point to trigger
14846 * is_state_visited() check before call itself is
14847 * processed by __check_func_call(). Otherwise new
14848 * async state will be pushed for further exploration.
14850 mark_prune_point(env, t);
14851 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
14852 struct bpf_kfunc_call_arg_meta meta;
14854 ret = fetch_kfunc_meta(env, insn, &meta, NULL);
14855 if (ret == 0 && is_iter_next_kfunc(&meta)) {
14856 mark_prune_point(env, t);
14857 /* Checking and saving state checkpoints at iter_next() call
14858 * is crucial for fast convergence of open-coded iterator loop
14859 * logic, so we need to force it. If we don't do that,
14860 * is_state_visited() might skip saving a checkpoint, causing
14861 * unnecessarily long sequence of not checkpointed
14862 * instructions and jumps, leading to exhaustion of jump
14863 * history buffer, and potentially other undesired outcomes.
14864 * It is expected that with correct open-coded iterators
14865 * convergence will happen quickly, so we don't run a risk of
14866 * exhausting memory.
14868 mark_force_checkpoint(env, t);
14871 return visit_func_call_insn(t, insns, env, insn->src_reg == BPF_PSEUDO_CALL);
14874 if (BPF_SRC(insn->code) != BPF_K)
14877 if (BPF_CLASS(insn->code) == BPF_JMP)
14882 /* unconditional jump with single edge */
14883 ret = push_insn(t, t + off + 1, FALLTHROUGH, env);
14887 mark_prune_point(env, t + off + 1);
14888 mark_jmp_point(env, t + off + 1);
14893 /* conditional jump with two edges */
14894 mark_prune_point(env, t);
14896 ret = push_insn(t, t + 1, FALLTHROUGH, env);
14900 return push_insn(t, t + insn->off + 1, BRANCH, env);
14904 /* non-recursive depth-first-search to detect loops in BPF program
14905 * loop == back-edge in directed graph
14907 static int check_cfg(struct bpf_verifier_env *env)
14909 int insn_cnt = env->prog->len;
14910 int *insn_stack, *insn_state;
14914 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
14918 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
14920 kvfree(insn_state);
14924 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
14925 insn_stack[0] = 0; /* 0 is the first instruction */
14926 env->cfg.cur_stack = 1;
14928 while (env->cfg.cur_stack > 0) {
14929 int t = insn_stack[env->cfg.cur_stack - 1];
14931 ret = visit_insn(t, env);
14933 case DONE_EXPLORING:
14934 insn_state[t] = EXPLORED;
14935 env->cfg.cur_stack--;
14937 case KEEP_EXPLORING:
14941 verbose(env, "visit_insn internal bug\n");
14948 if (env->cfg.cur_stack < 0) {
14949 verbose(env, "pop stack internal bug\n");
14954 for (i = 0; i < insn_cnt; i++) {
14955 struct bpf_insn *insn = &env->prog->insnsi[i];
14957 if (insn_state[i] != EXPLORED) {
14958 verbose(env, "unreachable insn %d\n", i);
14962 if (bpf_is_ldimm64(insn)) {
14963 if (insn_state[i + 1] != 0) {
14964 verbose(env, "jump into the middle of ldimm64 insn %d\n", i);
14968 i++; /* skip second half of ldimm64 */
14971 ret = 0; /* cfg looks good */
14974 kvfree(insn_state);
14975 kvfree(insn_stack);
14976 env->cfg.insn_state = env->cfg.insn_stack = NULL;
14980 static int check_abnormal_return(struct bpf_verifier_env *env)
14984 for (i = 1; i < env->subprog_cnt; i++) {
14985 if (env->subprog_info[i].has_ld_abs) {
14986 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
14989 if (env->subprog_info[i].has_tail_call) {
14990 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
14997 /* The minimum supported BTF func info size */
14998 #define MIN_BPF_FUNCINFO_SIZE 8
14999 #define MAX_FUNCINFO_REC_SIZE 252
15001 static int check_btf_func(struct bpf_verifier_env *env,
15002 const union bpf_attr *attr,
15005 const struct btf_type *type, *func_proto, *ret_type;
15006 u32 i, nfuncs, urec_size, min_size;
15007 u32 krec_size = sizeof(struct bpf_func_info);
15008 struct bpf_func_info *krecord;
15009 struct bpf_func_info_aux *info_aux = NULL;
15010 struct bpf_prog *prog;
15011 const struct btf *btf;
15013 u32 prev_offset = 0;
15014 bool scalar_return;
15017 nfuncs = attr->func_info_cnt;
15019 if (check_abnormal_return(env))
15024 if (nfuncs != env->subprog_cnt) {
15025 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
15029 urec_size = attr->func_info_rec_size;
15030 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
15031 urec_size > MAX_FUNCINFO_REC_SIZE ||
15032 urec_size % sizeof(u32)) {
15033 verbose(env, "invalid func info rec size %u\n", urec_size);
15038 btf = prog->aux->btf;
15040 urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
15041 min_size = min_t(u32, krec_size, urec_size);
15043 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
15046 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
15050 for (i = 0; i < nfuncs; i++) {
15051 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
15053 if (ret == -E2BIG) {
15054 verbose(env, "nonzero tailing record in func info");
15055 /* set the size kernel expects so loader can zero
15056 * out the rest of the record.
15058 if (copy_to_bpfptr_offset(uattr,
15059 offsetof(union bpf_attr, func_info_rec_size),
15060 &min_size, sizeof(min_size)))
15066 if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
15071 /* check insn_off */
15074 if (krecord[i].insn_off) {
15076 "nonzero insn_off %u for the first func info record",
15077 krecord[i].insn_off);
15080 } else if (krecord[i].insn_off <= prev_offset) {
15082 "same or smaller insn offset (%u) than previous func info record (%u)",
15083 krecord[i].insn_off, prev_offset);
15087 if (env->subprog_info[i].start != krecord[i].insn_off) {
15088 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
15092 /* check type_id */
15093 type = btf_type_by_id(btf, krecord[i].type_id);
15094 if (!type || !btf_type_is_func(type)) {
15095 verbose(env, "invalid type id %d in func info",
15096 krecord[i].type_id);
15099 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
15101 func_proto = btf_type_by_id(btf, type->type);
15102 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
15103 /* btf_func_check() already verified it during BTF load */
15105 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
15107 btf_type_is_small_int(ret_type) || btf_is_any_enum(ret_type);
15108 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
15109 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
15112 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
15113 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
15117 prev_offset = krecord[i].insn_off;
15118 bpfptr_add(&urecord, urec_size);
15121 prog->aux->func_info = krecord;
15122 prog->aux->func_info_cnt = nfuncs;
15123 prog->aux->func_info_aux = info_aux;
15132 static void adjust_btf_func(struct bpf_verifier_env *env)
15134 struct bpf_prog_aux *aux = env->prog->aux;
15137 if (!aux->func_info)
15140 for (i = 0; i < env->subprog_cnt; i++)
15141 aux->func_info[i].insn_off = env->subprog_info[i].start;
15144 #define MIN_BPF_LINEINFO_SIZE offsetofend(struct bpf_line_info, line_col)
15145 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
15147 static int check_btf_line(struct bpf_verifier_env *env,
15148 const union bpf_attr *attr,
15151 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
15152 struct bpf_subprog_info *sub;
15153 struct bpf_line_info *linfo;
15154 struct bpf_prog *prog;
15155 const struct btf *btf;
15159 nr_linfo = attr->line_info_cnt;
15162 if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info))
15165 rec_size = attr->line_info_rec_size;
15166 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
15167 rec_size > MAX_LINEINFO_REC_SIZE ||
15168 rec_size & (sizeof(u32) - 1))
15171 /* Need to zero it in case the userspace may
15172 * pass in a smaller bpf_line_info object.
15174 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
15175 GFP_KERNEL | __GFP_NOWARN);
15180 btf = prog->aux->btf;
15183 sub = env->subprog_info;
15184 ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
15185 expected_size = sizeof(struct bpf_line_info);
15186 ncopy = min_t(u32, expected_size, rec_size);
15187 for (i = 0; i < nr_linfo; i++) {
15188 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
15190 if (err == -E2BIG) {
15191 verbose(env, "nonzero tailing record in line_info");
15192 if (copy_to_bpfptr_offset(uattr,
15193 offsetof(union bpf_attr, line_info_rec_size),
15194 &expected_size, sizeof(expected_size)))
15200 if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
15206 * Check insn_off to ensure
15207 * 1) strictly increasing AND
15208 * 2) bounded by prog->len
15210 * The linfo[0].insn_off == 0 check logically falls into
15211 * the later "missing bpf_line_info for func..." case
15212 * because the first linfo[0].insn_off must be the
15213 * first sub also and the first sub must have
15214 * subprog_info[0].start == 0.
15216 if ((i && linfo[i].insn_off <= prev_offset) ||
15217 linfo[i].insn_off >= prog->len) {
15218 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
15219 i, linfo[i].insn_off, prev_offset,
15225 if (!prog->insnsi[linfo[i].insn_off].code) {
15227 "Invalid insn code at line_info[%u].insn_off\n",
15233 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
15234 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
15235 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
15240 if (s != env->subprog_cnt) {
15241 if (linfo[i].insn_off == sub[s].start) {
15242 sub[s].linfo_idx = i;
15244 } else if (sub[s].start < linfo[i].insn_off) {
15245 verbose(env, "missing bpf_line_info for func#%u\n", s);
15251 prev_offset = linfo[i].insn_off;
15252 bpfptr_add(&ulinfo, rec_size);
15255 if (s != env->subprog_cnt) {
15256 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
15257 env->subprog_cnt - s, s);
15262 prog->aux->linfo = linfo;
15263 prog->aux->nr_linfo = nr_linfo;
15272 #define MIN_CORE_RELO_SIZE sizeof(struct bpf_core_relo)
15273 #define MAX_CORE_RELO_SIZE MAX_FUNCINFO_REC_SIZE
15275 static int check_core_relo(struct bpf_verifier_env *env,
15276 const union bpf_attr *attr,
15279 u32 i, nr_core_relo, ncopy, expected_size, rec_size;
15280 struct bpf_core_relo core_relo = {};
15281 struct bpf_prog *prog = env->prog;
15282 const struct btf *btf = prog->aux->btf;
15283 struct bpf_core_ctx ctx = {
15287 bpfptr_t u_core_relo;
15290 nr_core_relo = attr->core_relo_cnt;
15293 if (nr_core_relo > INT_MAX / sizeof(struct bpf_core_relo))
15296 rec_size = attr->core_relo_rec_size;
15297 if (rec_size < MIN_CORE_RELO_SIZE ||
15298 rec_size > MAX_CORE_RELO_SIZE ||
15299 rec_size % sizeof(u32))
15302 u_core_relo = make_bpfptr(attr->core_relos, uattr.is_kernel);
15303 expected_size = sizeof(struct bpf_core_relo);
15304 ncopy = min_t(u32, expected_size, rec_size);
15306 /* Unlike func_info and line_info, copy and apply each CO-RE
15307 * relocation record one at a time.
15309 for (i = 0; i < nr_core_relo; i++) {
15310 /* future proofing when sizeof(bpf_core_relo) changes */
15311 err = bpf_check_uarg_tail_zero(u_core_relo, expected_size, rec_size);
15313 if (err == -E2BIG) {
15314 verbose(env, "nonzero tailing record in core_relo");
15315 if (copy_to_bpfptr_offset(uattr,
15316 offsetof(union bpf_attr, core_relo_rec_size),
15317 &expected_size, sizeof(expected_size)))
15323 if (copy_from_bpfptr(&core_relo, u_core_relo, ncopy)) {
15328 if (core_relo.insn_off % 8 || core_relo.insn_off / 8 >= prog->len) {
15329 verbose(env, "Invalid core_relo[%u].insn_off:%u prog->len:%u\n",
15330 i, core_relo.insn_off, prog->len);
15335 err = bpf_core_apply(&ctx, &core_relo, i,
15336 &prog->insnsi[core_relo.insn_off / 8]);
15339 bpfptr_add(&u_core_relo, rec_size);
15344 static int check_btf_info(struct bpf_verifier_env *env,
15345 const union bpf_attr *attr,
15351 if (!attr->func_info_cnt && !attr->line_info_cnt) {
15352 if (check_abnormal_return(env))
15357 btf = btf_get_by_fd(attr->prog_btf_fd);
15359 return PTR_ERR(btf);
15360 if (btf_is_kernel(btf)) {
15364 env->prog->aux->btf = btf;
15366 err = check_btf_func(env, attr, uattr);
15370 err = check_btf_line(env, attr, uattr);
15374 err = check_core_relo(env, attr, uattr);
15381 /* check %cur's range satisfies %old's */
15382 static bool range_within(struct bpf_reg_state *old,
15383 struct bpf_reg_state *cur)
15385 return old->umin_value <= cur->umin_value &&
15386 old->umax_value >= cur->umax_value &&
15387 old->smin_value <= cur->smin_value &&
15388 old->smax_value >= cur->smax_value &&
15389 old->u32_min_value <= cur->u32_min_value &&
15390 old->u32_max_value >= cur->u32_max_value &&
15391 old->s32_min_value <= cur->s32_min_value &&
15392 old->s32_max_value >= cur->s32_max_value;
15395 /* If in the old state two registers had the same id, then they need to have
15396 * the same id in the new state as well. But that id could be different from
15397 * the old state, so we need to track the mapping from old to new ids.
15398 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
15399 * regs with old id 5 must also have new id 9 for the new state to be safe. But
15400 * regs with a different old id could still have new id 9, we don't care about
15402 * So we look through our idmap to see if this old id has been seen before. If
15403 * so, we require the new id to match; otherwise, we add the id pair to the map.
15405 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap)
15407 struct bpf_id_pair *map = idmap->map;
15410 /* either both IDs should be set or both should be zero */
15411 if (!!old_id != !!cur_id)
15414 if (old_id == 0) /* cur_id == 0 as well */
15417 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
15419 /* Reached an empty slot; haven't seen this id before */
15420 map[i].old = old_id;
15421 map[i].cur = cur_id;
15424 if (map[i].old == old_id)
15425 return map[i].cur == cur_id;
15426 if (map[i].cur == cur_id)
15429 /* We ran out of idmap slots, which should be impossible */
15434 /* Similar to check_ids(), but allocate a unique temporary ID
15435 * for 'old_id' or 'cur_id' of zero.
15436 * This makes pairs like '0 vs unique ID', 'unique ID vs 0' valid.
15438 static bool check_scalar_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap)
15440 old_id = old_id ? old_id : ++idmap->tmp_id_gen;
15441 cur_id = cur_id ? cur_id : ++idmap->tmp_id_gen;
15443 return check_ids(old_id, cur_id, idmap);
15446 static void clean_func_state(struct bpf_verifier_env *env,
15447 struct bpf_func_state *st)
15449 enum bpf_reg_liveness live;
15452 for (i = 0; i < BPF_REG_FP; i++) {
15453 live = st->regs[i].live;
15454 /* liveness must not touch this register anymore */
15455 st->regs[i].live |= REG_LIVE_DONE;
15456 if (!(live & REG_LIVE_READ))
15457 /* since the register is unused, clear its state
15458 * to make further comparison simpler
15460 __mark_reg_not_init(env, &st->regs[i]);
15463 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
15464 live = st->stack[i].spilled_ptr.live;
15465 /* liveness must not touch this stack slot anymore */
15466 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
15467 if (!(live & REG_LIVE_READ)) {
15468 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
15469 for (j = 0; j < BPF_REG_SIZE; j++)
15470 st->stack[i].slot_type[j] = STACK_INVALID;
15475 static void clean_verifier_state(struct bpf_verifier_env *env,
15476 struct bpf_verifier_state *st)
15480 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
15481 /* all regs in this state in all frames were already marked */
15484 for (i = 0; i <= st->curframe; i++)
15485 clean_func_state(env, st->frame[i]);
15488 /* the parentage chains form a tree.
15489 * the verifier states are added to state lists at given insn and
15490 * pushed into state stack for future exploration.
15491 * when the verifier reaches bpf_exit insn some of the verifer states
15492 * stored in the state lists have their final liveness state already,
15493 * but a lot of states will get revised from liveness point of view when
15494 * the verifier explores other branches.
15497 * 2: if r1 == 100 goto pc+1
15500 * when the verifier reaches exit insn the register r0 in the state list of
15501 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
15502 * of insn 2 and goes exploring further. At the insn 4 it will walk the
15503 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
15505 * Since the verifier pushes the branch states as it sees them while exploring
15506 * the program the condition of walking the branch instruction for the second
15507 * time means that all states below this branch were already explored and
15508 * their final liveness marks are already propagated.
15509 * Hence when the verifier completes the search of state list in is_state_visited()
15510 * we can call this clean_live_states() function to mark all liveness states
15511 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
15512 * will not be used.
15513 * This function also clears the registers and stack for states that !READ
15514 * to simplify state merging.
15516 * Important note here that walking the same branch instruction in the callee
15517 * doesn't meant that the states are DONE. The verifier has to compare
15520 static void clean_live_states(struct bpf_verifier_env *env, int insn,
15521 struct bpf_verifier_state *cur)
15523 struct bpf_verifier_state_list *sl;
15526 sl = *explored_state(env, insn);
15528 if (sl->state.branches)
15530 if (sl->state.insn_idx != insn ||
15531 sl->state.curframe != cur->curframe)
15533 for (i = 0; i <= cur->curframe; i++)
15534 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
15536 clean_verifier_state(env, &sl->state);
15542 static bool regs_exact(const struct bpf_reg_state *rold,
15543 const struct bpf_reg_state *rcur,
15544 struct bpf_idmap *idmap)
15546 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
15547 check_ids(rold->id, rcur->id, idmap) &&
15548 check_ids(rold->ref_obj_id, rcur->ref_obj_id, idmap);
15551 /* Returns true if (rold safe implies rcur safe) */
15552 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
15553 struct bpf_reg_state *rcur, struct bpf_idmap *idmap)
15555 if (!(rold->live & REG_LIVE_READ))
15556 /* explored state didn't use this */
15558 if (rold->type == NOT_INIT)
15559 /* explored state can't have used this */
15561 if (rcur->type == NOT_INIT)
15564 /* Enforce that register types have to match exactly, including their
15565 * modifiers (like PTR_MAYBE_NULL, MEM_RDONLY, etc), as a general
15568 * One can make a point that using a pointer register as unbounded
15569 * SCALAR would be technically acceptable, but this could lead to
15570 * pointer leaks because scalars are allowed to leak while pointers
15571 * are not. We could make this safe in special cases if root is
15572 * calling us, but it's probably not worth the hassle.
15574 * Also, register types that are *not* MAYBE_NULL could technically be
15575 * safe to use as their MAYBE_NULL variants (e.g., PTR_TO_MAP_VALUE
15576 * is safe to be used as PTR_TO_MAP_VALUE_OR_NULL, provided both point
15577 * to the same map).
15578 * However, if the old MAYBE_NULL register then got NULL checked,
15579 * doing so could have affected others with the same id, and we can't
15580 * check for that because we lost the id when we converted to
15581 * a non-MAYBE_NULL variant.
15582 * So, as a general rule we don't allow mixing MAYBE_NULL and
15583 * non-MAYBE_NULL registers as well.
15585 if (rold->type != rcur->type)
15588 switch (base_type(rold->type)) {
15590 if (env->explore_alu_limits) {
15591 /* explore_alu_limits disables tnum_in() and range_within()
15592 * logic and requires everything to be strict
15594 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
15595 check_scalar_ids(rold->id, rcur->id, idmap);
15597 if (!rold->precise)
15599 /* Why check_ids() for scalar registers?
15601 * Consider the following BPF code:
15602 * 1: r6 = ... unbound scalar, ID=a ...
15603 * 2: r7 = ... unbound scalar, ID=b ...
15604 * 3: if (r6 > r7) goto +1
15606 * 5: if (r6 > X) goto ...
15607 * 6: ... memory operation using r7 ...
15609 * First verification path is [1-6]:
15610 * - at (4) same bpf_reg_state::id (b) would be assigned to r6 and r7;
15611 * - at (5) r6 would be marked <= X, find_equal_scalars() would also mark
15612 * r7 <= X, because r6 and r7 share same id.
15613 * Next verification path is [1-4, 6].
15615 * Instruction (6) would be reached in two states:
15616 * I. r6{.id=b}, r7{.id=b} via path 1-6;
15617 * II. r6{.id=a}, r7{.id=b} via path 1-4, 6.
15619 * Use check_ids() to distinguish these states.
15621 * Also verify that new value satisfies old value range knowledge.
15623 return range_within(rold, rcur) &&
15624 tnum_in(rold->var_off, rcur->var_off) &&
15625 check_scalar_ids(rold->id, rcur->id, idmap);
15626 case PTR_TO_MAP_KEY:
15627 case PTR_TO_MAP_VALUE:
15630 case PTR_TO_TP_BUFFER:
15631 /* If the new min/max/var_off satisfy the old ones and
15632 * everything else matches, we are OK.
15634 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, var_off)) == 0 &&
15635 range_within(rold, rcur) &&
15636 tnum_in(rold->var_off, rcur->var_off) &&
15637 check_ids(rold->id, rcur->id, idmap) &&
15638 check_ids(rold->ref_obj_id, rcur->ref_obj_id, idmap);
15639 case PTR_TO_PACKET_META:
15640 case PTR_TO_PACKET:
15641 /* We must have at least as much range as the old ptr
15642 * did, so that any accesses which were safe before are
15643 * still safe. This is true even if old range < old off,
15644 * since someone could have accessed through (ptr - k), or
15645 * even done ptr -= k in a register, to get a safe access.
15647 if (rold->range > rcur->range)
15649 /* If the offsets don't match, we can't trust our alignment;
15650 * nor can we be sure that we won't fall out of range.
15652 if (rold->off != rcur->off)
15654 /* id relations must be preserved */
15655 if (!check_ids(rold->id, rcur->id, idmap))
15657 /* new val must satisfy old val knowledge */
15658 return range_within(rold, rcur) &&
15659 tnum_in(rold->var_off, rcur->var_off);
15661 /* two stack pointers are equal only if they're pointing to
15662 * the same stack frame, since fp-8 in foo != fp-8 in bar
15664 return regs_exact(rold, rcur, idmap) && rold->frameno == rcur->frameno;
15666 return regs_exact(rold, rcur, idmap);
15670 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
15671 struct bpf_func_state *cur, struct bpf_idmap *idmap)
15675 /* walk slots of the explored stack and ignore any additional
15676 * slots in the current stack, since explored(safe) state
15679 for (i = 0; i < old->allocated_stack; i++) {
15680 struct bpf_reg_state *old_reg, *cur_reg;
15682 spi = i / BPF_REG_SIZE;
15684 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
15685 i += BPF_REG_SIZE - 1;
15686 /* explored state didn't use this */
15690 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
15693 if (env->allow_uninit_stack &&
15694 old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC)
15697 /* explored stack has more populated slots than current stack
15698 * and these slots were used
15700 if (i >= cur->allocated_stack)
15703 /* if old state was safe with misc data in the stack
15704 * it will be safe with zero-initialized stack.
15705 * The opposite is not true
15707 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
15708 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
15710 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
15711 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
15712 /* Ex: old explored (safe) state has STACK_SPILL in
15713 * this stack slot, but current has STACK_MISC ->
15714 * this verifier states are not equivalent,
15715 * return false to continue verification of this path
15718 if (i % BPF_REG_SIZE != BPF_REG_SIZE - 1)
15720 /* Both old and cur are having same slot_type */
15721 switch (old->stack[spi].slot_type[BPF_REG_SIZE - 1]) {
15723 /* when explored and current stack slot are both storing
15724 * spilled registers, check that stored pointers types
15725 * are the same as well.
15726 * Ex: explored safe path could have stored
15727 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
15728 * but current path has stored:
15729 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
15730 * such verifier states are not equivalent.
15731 * return false to continue verification of this path
15733 if (!regsafe(env, &old->stack[spi].spilled_ptr,
15734 &cur->stack[spi].spilled_ptr, idmap))
15738 old_reg = &old->stack[spi].spilled_ptr;
15739 cur_reg = &cur->stack[spi].spilled_ptr;
15740 if (old_reg->dynptr.type != cur_reg->dynptr.type ||
15741 old_reg->dynptr.first_slot != cur_reg->dynptr.first_slot ||
15742 !check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap))
15746 old_reg = &old->stack[spi].spilled_ptr;
15747 cur_reg = &cur->stack[spi].spilled_ptr;
15748 /* iter.depth is not compared between states as it
15749 * doesn't matter for correctness and would otherwise
15750 * prevent convergence; we maintain it only to prevent
15751 * infinite loop check triggering, see
15752 * iter_active_depths_differ()
15754 if (old_reg->iter.btf != cur_reg->iter.btf ||
15755 old_reg->iter.btf_id != cur_reg->iter.btf_id ||
15756 old_reg->iter.state != cur_reg->iter.state ||
15757 /* ignore {old_reg,cur_reg}->iter.depth, see above */
15758 !check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap))
15763 case STACK_INVALID:
15765 /* Ensure that new unhandled slot types return false by default */
15773 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur,
15774 struct bpf_idmap *idmap)
15778 if (old->acquired_refs != cur->acquired_refs)
15781 for (i = 0; i < old->acquired_refs; i++) {
15782 if (!check_ids(old->refs[i].id, cur->refs[i].id, idmap))
15789 /* compare two verifier states
15791 * all states stored in state_list are known to be valid, since
15792 * verifier reached 'bpf_exit' instruction through them
15794 * this function is called when verifier exploring different branches of
15795 * execution popped from the state stack. If it sees an old state that has
15796 * more strict register state and more strict stack state then this execution
15797 * branch doesn't need to be explored further, since verifier already
15798 * concluded that more strict state leads to valid finish.
15800 * Therefore two states are equivalent if register state is more conservative
15801 * and explored stack state is more conservative than the current one.
15804 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
15805 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
15807 * In other words if current stack state (one being explored) has more
15808 * valid slots than old one that already passed validation, it means
15809 * the verifier can stop exploring and conclude that current state is valid too
15811 * Similarly with registers. If explored state has register type as invalid
15812 * whereas register type in current state is meaningful, it means that
15813 * the current state will reach 'bpf_exit' instruction safely
15815 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
15816 struct bpf_func_state *cur)
15820 for (i = 0; i < MAX_BPF_REG; i++)
15821 if (!regsafe(env, &old->regs[i], &cur->regs[i],
15822 &env->idmap_scratch))
15825 if (!stacksafe(env, old, cur, &env->idmap_scratch))
15828 if (!refsafe(old, cur, &env->idmap_scratch))
15834 static bool states_equal(struct bpf_verifier_env *env,
15835 struct bpf_verifier_state *old,
15836 struct bpf_verifier_state *cur)
15840 if (old->curframe != cur->curframe)
15843 env->idmap_scratch.tmp_id_gen = env->id_gen;
15844 memset(&env->idmap_scratch.map, 0, sizeof(env->idmap_scratch.map));
15846 /* Verification state from speculative execution simulation
15847 * must never prune a non-speculative execution one.
15849 if (old->speculative && !cur->speculative)
15852 if (old->active_lock.ptr != cur->active_lock.ptr)
15855 /* Old and cur active_lock's have to be either both present
15858 if (!!old->active_lock.id != !!cur->active_lock.id)
15861 if (old->active_lock.id &&
15862 !check_ids(old->active_lock.id, cur->active_lock.id, &env->idmap_scratch))
15865 if (old->active_rcu_lock != cur->active_rcu_lock)
15868 /* for states to be equal callsites have to be the same
15869 * and all frame states need to be equivalent
15871 for (i = 0; i <= old->curframe; i++) {
15872 if (old->frame[i]->callsite != cur->frame[i]->callsite)
15874 if (!func_states_equal(env, old->frame[i], cur->frame[i]))
15880 /* Return 0 if no propagation happened. Return negative error code if error
15881 * happened. Otherwise, return the propagated bit.
15883 static int propagate_liveness_reg(struct bpf_verifier_env *env,
15884 struct bpf_reg_state *reg,
15885 struct bpf_reg_state *parent_reg)
15887 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
15888 u8 flag = reg->live & REG_LIVE_READ;
15891 /* When comes here, read flags of PARENT_REG or REG could be any of
15892 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
15893 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
15895 if (parent_flag == REG_LIVE_READ64 ||
15896 /* Or if there is no read flag from REG. */
15898 /* Or if the read flag from REG is the same as PARENT_REG. */
15899 parent_flag == flag)
15902 err = mark_reg_read(env, reg, parent_reg, flag);
15909 /* A write screens off any subsequent reads; but write marks come from the
15910 * straight-line code between a state and its parent. When we arrive at an
15911 * equivalent state (jump target or such) we didn't arrive by the straight-line
15912 * code, so read marks in the state must propagate to the parent regardless
15913 * of the state's write marks. That's what 'parent == state->parent' comparison
15914 * in mark_reg_read() is for.
15916 static int propagate_liveness(struct bpf_verifier_env *env,
15917 const struct bpf_verifier_state *vstate,
15918 struct bpf_verifier_state *vparent)
15920 struct bpf_reg_state *state_reg, *parent_reg;
15921 struct bpf_func_state *state, *parent;
15922 int i, frame, err = 0;
15924 if (vparent->curframe != vstate->curframe) {
15925 WARN(1, "propagate_live: parent frame %d current frame %d\n",
15926 vparent->curframe, vstate->curframe);
15929 /* Propagate read liveness of registers... */
15930 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
15931 for (frame = 0; frame <= vstate->curframe; frame++) {
15932 parent = vparent->frame[frame];
15933 state = vstate->frame[frame];
15934 parent_reg = parent->regs;
15935 state_reg = state->regs;
15936 /* We don't need to worry about FP liveness, it's read-only */
15937 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
15938 err = propagate_liveness_reg(env, &state_reg[i],
15942 if (err == REG_LIVE_READ64)
15943 mark_insn_zext(env, &parent_reg[i]);
15946 /* Propagate stack slots. */
15947 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
15948 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
15949 parent_reg = &parent->stack[i].spilled_ptr;
15950 state_reg = &state->stack[i].spilled_ptr;
15951 err = propagate_liveness_reg(env, state_reg,
15960 /* find precise scalars in the previous equivalent state and
15961 * propagate them into the current state
15963 static int propagate_precision(struct bpf_verifier_env *env,
15964 const struct bpf_verifier_state *old)
15966 struct bpf_reg_state *state_reg;
15967 struct bpf_func_state *state;
15968 int i, err = 0, fr;
15971 for (fr = old->curframe; fr >= 0; fr--) {
15972 state = old->frame[fr];
15973 state_reg = state->regs;
15975 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
15976 if (state_reg->type != SCALAR_VALUE ||
15977 !state_reg->precise ||
15978 !(state_reg->live & REG_LIVE_READ))
15980 if (env->log.level & BPF_LOG_LEVEL2) {
15982 verbose(env, "frame %d: propagating r%d", fr, i);
15984 verbose(env, ",r%d", i);
15986 bt_set_frame_reg(&env->bt, fr, i);
15990 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
15991 if (!is_spilled_reg(&state->stack[i]))
15993 state_reg = &state->stack[i].spilled_ptr;
15994 if (state_reg->type != SCALAR_VALUE ||
15995 !state_reg->precise ||
15996 !(state_reg->live & REG_LIVE_READ))
15998 if (env->log.level & BPF_LOG_LEVEL2) {
16000 verbose(env, "frame %d: propagating fp%d",
16001 fr, (-i - 1) * BPF_REG_SIZE);
16003 verbose(env, ",fp%d", (-i - 1) * BPF_REG_SIZE);
16005 bt_set_frame_slot(&env->bt, fr, i);
16009 verbose(env, "\n");
16012 err = mark_chain_precision_batch(env);
16019 static bool states_maybe_looping(struct bpf_verifier_state *old,
16020 struct bpf_verifier_state *cur)
16022 struct bpf_func_state *fold, *fcur;
16023 int i, fr = cur->curframe;
16025 if (old->curframe != fr)
16028 fold = old->frame[fr];
16029 fcur = cur->frame[fr];
16030 for (i = 0; i < MAX_BPF_REG; i++)
16031 if (memcmp(&fold->regs[i], &fcur->regs[i],
16032 offsetof(struct bpf_reg_state, parent)))
16037 static bool is_iter_next_insn(struct bpf_verifier_env *env, int insn_idx)
16039 return env->insn_aux_data[insn_idx].is_iter_next;
16042 /* is_state_visited() handles iter_next() (see process_iter_next_call() for
16043 * terminology) calls specially: as opposed to bounded BPF loops, it *expects*
16044 * states to match, which otherwise would look like an infinite loop. So while
16045 * iter_next() calls are taken care of, we still need to be careful and
16046 * prevent erroneous and too eager declaration of "ininite loop", when
16047 * iterators are involved.
16049 * Here's a situation in pseudo-BPF assembly form:
16051 * 0: again: ; set up iter_next() call args
16052 * 1: r1 = &it ; <CHECKPOINT HERE>
16053 * 2: call bpf_iter_num_next ; this is iter_next() call
16054 * 3: if r0 == 0 goto done
16055 * 4: ... something useful here ...
16056 * 5: goto again ; another iteration
16059 * 8: call bpf_iter_num_destroy ; clean up iter state
16062 * This is a typical loop. Let's assume that we have a prune point at 1:,
16063 * before we get to `call bpf_iter_num_next` (e.g., because of that `goto
16064 * again`, assuming other heuristics don't get in a way).
16066 * When we first time come to 1:, let's say we have some state X. We proceed
16067 * to 2:, fork states, enqueue ACTIVE, validate NULL case successfully, exit.
16068 * Now we come back to validate that forked ACTIVE state. We proceed through
16069 * 3-5, come to goto, jump to 1:. Let's assume our state didn't change, so we
16070 * are converging. But the problem is that we don't know that yet, as this
16071 * convergence has to happen at iter_next() call site only. So if nothing is
16072 * done, at 1: verifier will use bounded loop logic and declare infinite
16073 * looping (and would be *technically* correct, if not for iterator's
16074 * "eventual sticky NULL" contract, see process_iter_next_call()). But we
16075 * don't want that. So what we do in process_iter_next_call() when we go on
16076 * another ACTIVE iteration, we bump slot->iter.depth, to mark that it's
16077 * a different iteration. So when we suspect an infinite loop, we additionally
16078 * check if any of the *ACTIVE* iterator states depths differ. If yes, we
16079 * pretend we are not looping and wait for next iter_next() call.
16081 * This only applies to ACTIVE state. In DRAINED state we don't expect to
16082 * loop, because that would actually mean infinite loop, as DRAINED state is
16083 * "sticky", and so we'll keep returning into the same instruction with the
16084 * same state (at least in one of possible code paths).
16086 * This approach allows to keep infinite loop heuristic even in the face of
16087 * active iterator. E.g., C snippet below is and will be detected as
16088 * inifintely looping:
16090 * struct bpf_iter_num it;
16093 * bpf_iter_num_new(&it, 0, 10);
16094 * while ((p = bpf_iter_num_next(&t))) {
16096 * while (x--) {} // <<-- infinite loop here
16100 static bool iter_active_depths_differ(struct bpf_verifier_state *old, struct bpf_verifier_state *cur)
16102 struct bpf_reg_state *slot, *cur_slot;
16103 struct bpf_func_state *state;
16106 for (fr = old->curframe; fr >= 0; fr--) {
16107 state = old->frame[fr];
16108 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
16109 if (state->stack[i].slot_type[0] != STACK_ITER)
16112 slot = &state->stack[i].spilled_ptr;
16113 if (slot->iter.state != BPF_ITER_STATE_ACTIVE)
16116 cur_slot = &cur->frame[fr]->stack[i].spilled_ptr;
16117 if (cur_slot->iter.depth != slot->iter.depth)
16124 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
16126 struct bpf_verifier_state_list *new_sl;
16127 struct bpf_verifier_state_list *sl, **pprev;
16128 struct bpf_verifier_state *cur = env->cur_state, *new;
16129 int i, j, err, states_cnt = 0;
16130 bool force_new_state = env->test_state_freq || is_force_checkpoint(env, insn_idx);
16131 bool add_new_state = force_new_state;
16133 /* bpf progs typically have pruning point every 4 instructions
16134 * http://vger.kernel.org/bpfconf2019.html#session-1
16135 * Do not add new state for future pruning if the verifier hasn't seen
16136 * at least 2 jumps and at least 8 instructions.
16137 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
16138 * In tests that amounts to up to 50% reduction into total verifier
16139 * memory consumption and 20% verifier time speedup.
16141 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
16142 env->insn_processed - env->prev_insn_processed >= 8)
16143 add_new_state = true;
16145 pprev = explored_state(env, insn_idx);
16148 clean_live_states(env, insn_idx, cur);
16152 if (sl->state.insn_idx != insn_idx)
16155 if (sl->state.branches) {
16156 struct bpf_func_state *frame = sl->state.frame[sl->state.curframe];
16158 if (frame->in_async_callback_fn &&
16159 frame->async_entry_cnt != cur->frame[cur->curframe]->async_entry_cnt) {
16160 /* Different async_entry_cnt means that the verifier is
16161 * processing another entry into async callback.
16162 * Seeing the same state is not an indication of infinite
16163 * loop or infinite recursion.
16164 * But finding the same state doesn't mean that it's safe
16165 * to stop processing the current state. The previous state
16166 * hasn't yet reached bpf_exit, since state.branches > 0.
16167 * Checking in_async_callback_fn alone is not enough either.
16168 * Since the verifier still needs to catch infinite loops
16169 * inside async callbacks.
16171 goto skip_inf_loop_check;
16173 /* BPF open-coded iterators loop detection is special.
16174 * states_maybe_looping() logic is too simplistic in detecting
16175 * states that *might* be equivalent, because it doesn't know
16176 * about ID remapping, so don't even perform it.
16177 * See process_iter_next_call() and iter_active_depths_differ()
16178 * for overview of the logic. When current and one of parent
16179 * states are detected as equivalent, it's a good thing: we prove
16180 * convergence and can stop simulating further iterations.
16181 * It's safe to assume that iterator loop will finish, taking into
16182 * account iter_next() contract of eventually returning
16183 * sticky NULL result.
16185 if (is_iter_next_insn(env, insn_idx)) {
16186 if (states_equal(env, &sl->state, cur)) {
16187 struct bpf_func_state *cur_frame;
16188 struct bpf_reg_state *iter_state, *iter_reg;
16191 cur_frame = cur->frame[cur->curframe];
16192 /* btf_check_iter_kfuncs() enforces that
16193 * iter state pointer is always the first arg
16195 iter_reg = &cur_frame->regs[BPF_REG_1];
16196 /* current state is valid due to states_equal(),
16197 * so we can assume valid iter and reg state,
16198 * no need for extra (re-)validations
16200 spi = __get_spi(iter_reg->off + iter_reg->var_off.value);
16201 iter_state = &func(env, iter_reg)->stack[spi].spilled_ptr;
16202 if (iter_state->iter.state == BPF_ITER_STATE_ACTIVE)
16205 goto skip_inf_loop_check;
16207 /* attempt to detect infinite loop to avoid unnecessary doomed work */
16208 if (states_maybe_looping(&sl->state, cur) &&
16209 states_equal(env, &sl->state, cur) &&
16210 !iter_active_depths_differ(&sl->state, cur)) {
16211 verbose_linfo(env, insn_idx, "; ");
16212 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
16215 /* if the verifier is processing a loop, avoid adding new state
16216 * too often, since different loop iterations have distinct
16217 * states and may not help future pruning.
16218 * This threshold shouldn't be too low to make sure that
16219 * a loop with large bound will be rejected quickly.
16220 * The most abusive loop will be:
16222 * if r1 < 1000000 goto pc-2
16223 * 1M insn_procssed limit / 100 == 10k peak states.
16224 * This threshold shouldn't be too high either, since states
16225 * at the end of the loop are likely to be useful in pruning.
16227 skip_inf_loop_check:
16228 if (!force_new_state &&
16229 env->jmps_processed - env->prev_jmps_processed < 20 &&
16230 env->insn_processed - env->prev_insn_processed < 100)
16231 add_new_state = false;
16234 if (states_equal(env, &sl->state, cur)) {
16237 /* reached equivalent register/stack state,
16238 * prune the search.
16239 * Registers read by the continuation are read by us.
16240 * If we have any write marks in env->cur_state, they
16241 * will prevent corresponding reads in the continuation
16242 * from reaching our parent (an explored_state). Our
16243 * own state will get the read marks recorded, but
16244 * they'll be immediately forgotten as we're pruning
16245 * this state and will pop a new one.
16247 err = propagate_liveness(env, &sl->state, cur);
16249 /* if previous state reached the exit with precision and
16250 * current state is equivalent to it (except precsion marks)
16251 * the precision needs to be propagated back in
16252 * the current state.
16254 err = err ? : push_jmp_history(env, cur);
16255 err = err ? : propagate_precision(env, &sl->state);
16261 /* when new state is not going to be added do not increase miss count.
16262 * Otherwise several loop iterations will remove the state
16263 * recorded earlier. The goal of these heuristics is to have
16264 * states from some iterations of the loop (some in the beginning
16265 * and some at the end) to help pruning.
16269 /* heuristic to determine whether this state is beneficial
16270 * to keep checking from state equivalence point of view.
16271 * Higher numbers increase max_states_per_insn and verification time,
16272 * but do not meaningfully decrease insn_processed.
16274 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
16275 /* the state is unlikely to be useful. Remove it to
16276 * speed up verification
16279 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
16280 u32 br = sl->state.branches;
16283 "BUG live_done but branches_to_explore %d\n",
16285 free_verifier_state(&sl->state, false);
16287 env->peak_states--;
16289 /* cannot free this state, since parentage chain may
16290 * walk it later. Add it for free_list instead to
16291 * be freed at the end of verification
16293 sl->next = env->free_list;
16294 env->free_list = sl;
16304 if (env->max_states_per_insn < states_cnt)
16305 env->max_states_per_insn = states_cnt;
16307 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
16310 if (!add_new_state)
16313 /* There were no equivalent states, remember the current one.
16314 * Technically the current state is not proven to be safe yet,
16315 * but it will either reach outer most bpf_exit (which means it's safe)
16316 * or it will be rejected. When there are no loops the verifier won't be
16317 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
16318 * again on the way to bpf_exit.
16319 * When looping the sl->state.branches will be > 0 and this state
16320 * will not be considered for equivalence until branches == 0.
16322 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
16325 env->total_states++;
16326 env->peak_states++;
16327 env->prev_jmps_processed = env->jmps_processed;
16328 env->prev_insn_processed = env->insn_processed;
16330 /* forget precise markings we inherited, see __mark_chain_precision */
16331 if (env->bpf_capable)
16332 mark_all_scalars_imprecise(env, cur);
16334 /* add new state to the head of linked list */
16335 new = &new_sl->state;
16336 err = copy_verifier_state(new, cur);
16338 free_verifier_state(new, false);
16342 new->insn_idx = insn_idx;
16343 WARN_ONCE(new->branches != 1,
16344 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
16347 cur->first_insn_idx = insn_idx;
16348 clear_jmp_history(cur);
16349 new_sl->next = *explored_state(env, insn_idx);
16350 *explored_state(env, insn_idx) = new_sl;
16351 /* connect new state to parentage chain. Current frame needs all
16352 * registers connected. Only r6 - r9 of the callers are alive (pushed
16353 * to the stack implicitly by JITs) so in callers' frames connect just
16354 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
16355 * the state of the call instruction (with WRITTEN set), and r0 comes
16356 * from callee with its full parentage chain, anyway.
16358 /* clear write marks in current state: the writes we did are not writes
16359 * our child did, so they don't screen off its reads from us.
16360 * (There are no read marks in current state, because reads always mark
16361 * their parent and current state never has children yet. Only
16362 * explored_states can get read marks.)
16364 for (j = 0; j <= cur->curframe; j++) {
16365 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
16366 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
16367 for (i = 0; i < BPF_REG_FP; i++)
16368 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
16371 /* all stack frames are accessible from callee, clear them all */
16372 for (j = 0; j <= cur->curframe; j++) {
16373 struct bpf_func_state *frame = cur->frame[j];
16374 struct bpf_func_state *newframe = new->frame[j];
16376 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
16377 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
16378 frame->stack[i].spilled_ptr.parent =
16379 &newframe->stack[i].spilled_ptr;
16385 /* Return true if it's OK to have the same insn return a different type. */
16386 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
16388 switch (base_type(type)) {
16390 case PTR_TO_SOCKET:
16391 case PTR_TO_SOCK_COMMON:
16392 case PTR_TO_TCP_SOCK:
16393 case PTR_TO_XDP_SOCK:
16394 case PTR_TO_BTF_ID:
16401 /* If an instruction was previously used with particular pointer types, then we
16402 * need to be careful to avoid cases such as the below, where it may be ok
16403 * for one branch accessing the pointer, but not ok for the other branch:
16408 * R1 = some_other_valid_ptr;
16411 * R2 = *(u32 *)(R1 + 0);
16413 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
16415 return src != prev && (!reg_type_mismatch_ok(src) ||
16416 !reg_type_mismatch_ok(prev));
16419 static int save_aux_ptr_type(struct bpf_verifier_env *env, enum bpf_reg_type type,
16420 bool allow_trust_missmatch)
16422 enum bpf_reg_type *prev_type = &env->insn_aux_data[env->insn_idx].ptr_type;
16424 if (*prev_type == NOT_INIT) {
16425 /* Saw a valid insn
16426 * dst_reg = *(u32 *)(src_reg + off)
16427 * save type to validate intersecting paths
16430 } else if (reg_type_mismatch(type, *prev_type)) {
16431 /* Abuser program is trying to use the same insn
16432 * dst_reg = *(u32*) (src_reg + off)
16433 * with different pointer types:
16434 * src_reg == ctx in one branch and
16435 * src_reg == stack|map in some other branch.
16438 if (allow_trust_missmatch &&
16439 base_type(type) == PTR_TO_BTF_ID &&
16440 base_type(*prev_type) == PTR_TO_BTF_ID) {
16442 * Have to support a use case when one path through
16443 * the program yields TRUSTED pointer while another
16444 * is UNTRUSTED. Fallback to UNTRUSTED to generate
16445 * BPF_PROBE_MEM/BPF_PROBE_MEMSX.
16447 *prev_type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
16449 verbose(env, "same insn cannot be used with different pointers\n");
16457 static int do_check(struct bpf_verifier_env *env)
16459 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
16460 struct bpf_verifier_state *state = env->cur_state;
16461 struct bpf_insn *insns = env->prog->insnsi;
16462 struct bpf_reg_state *regs;
16463 int insn_cnt = env->prog->len;
16464 bool do_print_state = false;
16465 int prev_insn_idx = -1;
16468 struct bpf_insn *insn;
16472 env->prev_insn_idx = prev_insn_idx;
16473 if (env->insn_idx >= insn_cnt) {
16474 verbose(env, "invalid insn idx %d insn_cnt %d\n",
16475 env->insn_idx, insn_cnt);
16479 insn = &insns[env->insn_idx];
16480 class = BPF_CLASS(insn->code);
16482 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
16484 "BPF program is too large. Processed %d insn\n",
16485 env->insn_processed);
16489 state->last_insn_idx = env->prev_insn_idx;
16491 if (is_prune_point(env, env->insn_idx)) {
16492 err = is_state_visited(env, env->insn_idx);
16496 /* found equivalent state, can prune the search */
16497 if (env->log.level & BPF_LOG_LEVEL) {
16498 if (do_print_state)
16499 verbose(env, "\nfrom %d to %d%s: safe\n",
16500 env->prev_insn_idx, env->insn_idx,
16501 env->cur_state->speculative ?
16502 " (speculative execution)" : "");
16504 verbose(env, "%d: safe\n", env->insn_idx);
16506 goto process_bpf_exit;
16510 if (is_jmp_point(env, env->insn_idx)) {
16511 err = push_jmp_history(env, state);
16516 if (signal_pending(current))
16519 if (need_resched())
16522 if (env->log.level & BPF_LOG_LEVEL2 && do_print_state) {
16523 verbose(env, "\nfrom %d to %d%s:",
16524 env->prev_insn_idx, env->insn_idx,
16525 env->cur_state->speculative ?
16526 " (speculative execution)" : "");
16527 print_verifier_state(env, state->frame[state->curframe], true);
16528 do_print_state = false;
16531 if (env->log.level & BPF_LOG_LEVEL) {
16532 const struct bpf_insn_cbs cbs = {
16533 .cb_call = disasm_kfunc_name,
16534 .cb_print = verbose,
16535 .private_data = env,
16538 if (verifier_state_scratched(env))
16539 print_insn_state(env, state->frame[state->curframe]);
16541 verbose_linfo(env, env->insn_idx, "; ");
16542 env->prev_log_pos = env->log.end_pos;
16543 verbose(env, "%d: ", env->insn_idx);
16544 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
16545 env->prev_insn_print_pos = env->log.end_pos - env->prev_log_pos;
16546 env->prev_log_pos = env->log.end_pos;
16549 if (bpf_prog_is_offloaded(env->prog->aux)) {
16550 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
16551 env->prev_insn_idx);
16556 regs = cur_regs(env);
16557 sanitize_mark_insn_seen(env);
16558 prev_insn_idx = env->insn_idx;
16560 if (class == BPF_ALU || class == BPF_ALU64) {
16561 err = check_alu_op(env, insn);
16565 } else if (class == BPF_LDX) {
16566 enum bpf_reg_type src_reg_type;
16568 /* check for reserved fields is already done */
16570 /* check src operand */
16571 err = check_reg_arg(env, insn->src_reg, SRC_OP);
16575 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
16579 src_reg_type = regs[insn->src_reg].type;
16581 /* check that memory (src_reg + off) is readable,
16582 * the state of dst_reg will be updated by this func
16584 err = check_mem_access(env, env->insn_idx, insn->src_reg,
16585 insn->off, BPF_SIZE(insn->code),
16586 BPF_READ, insn->dst_reg, false,
16587 BPF_MODE(insn->code) == BPF_MEMSX);
16591 err = save_aux_ptr_type(env, src_reg_type, true);
16594 } else if (class == BPF_STX) {
16595 enum bpf_reg_type dst_reg_type;
16597 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
16598 err = check_atomic(env, env->insn_idx, insn);
16605 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
16606 verbose(env, "BPF_STX uses reserved fields\n");
16610 /* check src1 operand */
16611 err = check_reg_arg(env, insn->src_reg, SRC_OP);
16614 /* check src2 operand */
16615 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
16619 dst_reg_type = regs[insn->dst_reg].type;
16621 /* check that memory (dst_reg + off) is writeable */
16622 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
16623 insn->off, BPF_SIZE(insn->code),
16624 BPF_WRITE, insn->src_reg, false, false);
16628 err = save_aux_ptr_type(env, dst_reg_type, false);
16631 } else if (class == BPF_ST) {
16632 enum bpf_reg_type dst_reg_type;
16634 if (BPF_MODE(insn->code) != BPF_MEM ||
16635 insn->src_reg != BPF_REG_0) {
16636 verbose(env, "BPF_ST uses reserved fields\n");
16639 /* check src operand */
16640 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
16644 dst_reg_type = regs[insn->dst_reg].type;
16646 /* check that memory (dst_reg + off) is writeable */
16647 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
16648 insn->off, BPF_SIZE(insn->code),
16649 BPF_WRITE, -1, false, false);
16653 err = save_aux_ptr_type(env, dst_reg_type, false);
16656 } else if (class == BPF_JMP || class == BPF_JMP32) {
16657 u8 opcode = BPF_OP(insn->code);
16659 env->jmps_processed++;
16660 if (opcode == BPF_CALL) {
16661 if (BPF_SRC(insn->code) != BPF_K ||
16662 (insn->src_reg != BPF_PSEUDO_KFUNC_CALL
16663 && insn->off != 0) ||
16664 (insn->src_reg != BPF_REG_0 &&
16665 insn->src_reg != BPF_PSEUDO_CALL &&
16666 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
16667 insn->dst_reg != BPF_REG_0 ||
16668 class == BPF_JMP32) {
16669 verbose(env, "BPF_CALL uses reserved fields\n");
16673 if (env->cur_state->active_lock.ptr) {
16674 if ((insn->src_reg == BPF_REG_0 && insn->imm != BPF_FUNC_spin_unlock) ||
16675 (insn->src_reg == BPF_PSEUDO_CALL) ||
16676 (insn->src_reg == BPF_PSEUDO_KFUNC_CALL &&
16677 (insn->off != 0 || !is_bpf_graph_api_kfunc(insn->imm)))) {
16678 verbose(env, "function calls are not allowed while holding a lock\n");
16682 if (insn->src_reg == BPF_PSEUDO_CALL)
16683 err = check_func_call(env, insn, &env->insn_idx);
16684 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
16685 err = check_kfunc_call(env, insn, &env->insn_idx);
16687 err = check_helper_call(env, insn, &env->insn_idx);
16691 mark_reg_scratched(env, BPF_REG_0);
16692 } else if (opcode == BPF_JA) {
16693 if (BPF_SRC(insn->code) != BPF_K ||
16694 insn->src_reg != BPF_REG_0 ||
16695 insn->dst_reg != BPF_REG_0 ||
16696 (class == BPF_JMP && insn->imm != 0) ||
16697 (class == BPF_JMP32 && insn->off != 0)) {
16698 verbose(env, "BPF_JA uses reserved fields\n");
16702 if (class == BPF_JMP)
16703 env->insn_idx += insn->off + 1;
16705 env->insn_idx += insn->imm + 1;
16708 } else if (opcode == BPF_EXIT) {
16709 if (BPF_SRC(insn->code) != BPF_K ||
16711 insn->src_reg != BPF_REG_0 ||
16712 insn->dst_reg != BPF_REG_0 ||
16713 class == BPF_JMP32) {
16714 verbose(env, "BPF_EXIT uses reserved fields\n");
16718 if (env->cur_state->active_lock.ptr &&
16719 !in_rbtree_lock_required_cb(env)) {
16720 verbose(env, "bpf_spin_unlock is missing\n");
16724 if (env->cur_state->active_rcu_lock &&
16725 !in_rbtree_lock_required_cb(env)) {
16726 verbose(env, "bpf_rcu_read_unlock is missing\n");
16730 /* We must do check_reference_leak here before
16731 * prepare_func_exit to handle the case when
16732 * state->curframe > 0, it may be a callback
16733 * function, for which reference_state must
16734 * match caller reference state when it exits.
16736 err = check_reference_leak(env);
16740 if (state->curframe) {
16741 /* exit from nested function */
16742 err = prepare_func_exit(env, &env->insn_idx);
16745 do_print_state = true;
16749 err = check_return_code(env);
16753 mark_verifier_state_scratched(env);
16754 update_branch_counts(env, env->cur_state);
16755 err = pop_stack(env, &prev_insn_idx,
16756 &env->insn_idx, pop_log);
16758 if (err != -ENOENT)
16762 do_print_state = true;
16766 err = check_cond_jmp_op(env, insn, &env->insn_idx);
16770 } else if (class == BPF_LD) {
16771 u8 mode = BPF_MODE(insn->code);
16773 if (mode == BPF_ABS || mode == BPF_IND) {
16774 err = check_ld_abs(env, insn);
16778 } else if (mode == BPF_IMM) {
16779 err = check_ld_imm(env, insn);
16784 sanitize_mark_insn_seen(env);
16786 verbose(env, "invalid BPF_LD mode\n");
16790 verbose(env, "unknown insn class %d\n", class);
16800 static int find_btf_percpu_datasec(struct btf *btf)
16802 const struct btf_type *t;
16807 * Both vmlinux and module each have their own ".data..percpu"
16808 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
16809 * types to look at only module's own BTF types.
16811 n = btf_nr_types(btf);
16812 if (btf_is_module(btf))
16813 i = btf_nr_types(btf_vmlinux);
16817 for(; i < n; i++) {
16818 t = btf_type_by_id(btf, i);
16819 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
16822 tname = btf_name_by_offset(btf, t->name_off);
16823 if (!strcmp(tname, ".data..percpu"))
16830 /* replace pseudo btf_id with kernel symbol address */
16831 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
16832 struct bpf_insn *insn,
16833 struct bpf_insn_aux_data *aux)
16835 const struct btf_var_secinfo *vsi;
16836 const struct btf_type *datasec;
16837 struct btf_mod_pair *btf_mod;
16838 const struct btf_type *t;
16839 const char *sym_name;
16840 bool percpu = false;
16841 u32 type, id = insn->imm;
16845 int i, btf_fd, err;
16847 btf_fd = insn[1].imm;
16849 btf = btf_get_by_fd(btf_fd);
16851 verbose(env, "invalid module BTF object FD specified.\n");
16855 if (!btf_vmlinux) {
16856 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
16863 t = btf_type_by_id(btf, id);
16865 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
16870 if (!btf_type_is_var(t) && !btf_type_is_func(t)) {
16871 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR or KIND_FUNC\n", id);
16876 sym_name = btf_name_by_offset(btf, t->name_off);
16877 addr = kallsyms_lookup_name(sym_name);
16879 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
16884 insn[0].imm = (u32)addr;
16885 insn[1].imm = addr >> 32;
16887 if (btf_type_is_func(t)) {
16888 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
16889 aux->btf_var.mem_size = 0;
16893 datasec_id = find_btf_percpu_datasec(btf);
16894 if (datasec_id > 0) {
16895 datasec = btf_type_by_id(btf, datasec_id);
16896 for_each_vsi(i, datasec, vsi) {
16897 if (vsi->type == id) {
16905 t = btf_type_skip_modifiers(btf, type, NULL);
16907 aux->btf_var.reg_type = PTR_TO_BTF_ID | MEM_PERCPU;
16908 aux->btf_var.btf = btf;
16909 aux->btf_var.btf_id = type;
16910 } else if (!btf_type_is_struct(t)) {
16911 const struct btf_type *ret;
16915 /* resolve the type size of ksym. */
16916 ret = btf_resolve_size(btf, t, &tsize);
16918 tname = btf_name_by_offset(btf, t->name_off);
16919 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
16920 tname, PTR_ERR(ret));
16924 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
16925 aux->btf_var.mem_size = tsize;
16927 aux->btf_var.reg_type = PTR_TO_BTF_ID;
16928 aux->btf_var.btf = btf;
16929 aux->btf_var.btf_id = type;
16932 /* check whether we recorded this BTF (and maybe module) already */
16933 for (i = 0; i < env->used_btf_cnt; i++) {
16934 if (env->used_btfs[i].btf == btf) {
16940 if (env->used_btf_cnt >= MAX_USED_BTFS) {
16945 btf_mod = &env->used_btfs[env->used_btf_cnt];
16946 btf_mod->btf = btf;
16947 btf_mod->module = NULL;
16949 /* if we reference variables from kernel module, bump its refcount */
16950 if (btf_is_module(btf)) {
16951 btf_mod->module = btf_try_get_module(btf);
16952 if (!btf_mod->module) {
16958 env->used_btf_cnt++;
16966 static bool is_tracing_prog_type(enum bpf_prog_type type)
16969 case BPF_PROG_TYPE_KPROBE:
16970 case BPF_PROG_TYPE_TRACEPOINT:
16971 case BPF_PROG_TYPE_PERF_EVENT:
16972 case BPF_PROG_TYPE_RAW_TRACEPOINT:
16973 case BPF_PROG_TYPE_RAW_TRACEPOINT_WRITABLE:
16980 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
16981 struct bpf_map *map,
16982 struct bpf_prog *prog)
16985 enum bpf_prog_type prog_type = resolve_prog_type(prog);
16987 if (btf_record_has_field(map->record, BPF_LIST_HEAD) ||
16988 btf_record_has_field(map->record, BPF_RB_ROOT)) {
16989 if (is_tracing_prog_type(prog_type)) {
16990 verbose(env, "tracing progs cannot use bpf_{list_head,rb_root} yet\n");
16995 if (btf_record_has_field(map->record, BPF_SPIN_LOCK)) {
16996 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
16997 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
17001 if (is_tracing_prog_type(prog_type)) {
17002 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
17007 if (btf_record_has_field(map->record, BPF_TIMER)) {
17008 if (is_tracing_prog_type(prog_type)) {
17009 verbose(env, "tracing progs cannot use bpf_timer yet\n");
17014 if ((bpf_prog_is_offloaded(prog->aux) || bpf_map_is_offloaded(map)) &&
17015 !bpf_offload_prog_map_match(prog, map)) {
17016 verbose(env, "offload device mismatch between prog and map\n");
17020 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
17021 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
17025 if (prog->aux->sleepable)
17026 switch (map->map_type) {
17027 case BPF_MAP_TYPE_HASH:
17028 case BPF_MAP_TYPE_LRU_HASH:
17029 case BPF_MAP_TYPE_ARRAY:
17030 case BPF_MAP_TYPE_PERCPU_HASH:
17031 case BPF_MAP_TYPE_PERCPU_ARRAY:
17032 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
17033 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
17034 case BPF_MAP_TYPE_HASH_OF_MAPS:
17035 case BPF_MAP_TYPE_RINGBUF:
17036 case BPF_MAP_TYPE_USER_RINGBUF:
17037 case BPF_MAP_TYPE_INODE_STORAGE:
17038 case BPF_MAP_TYPE_SK_STORAGE:
17039 case BPF_MAP_TYPE_TASK_STORAGE:
17040 case BPF_MAP_TYPE_CGRP_STORAGE:
17044 "Sleepable programs can only use array, hash, ringbuf and local storage maps\n");
17051 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
17053 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
17054 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
17057 /* find and rewrite pseudo imm in ld_imm64 instructions:
17059 * 1. if it accesses map FD, replace it with actual map pointer.
17060 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
17062 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
17064 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
17066 struct bpf_insn *insn = env->prog->insnsi;
17067 int insn_cnt = env->prog->len;
17070 err = bpf_prog_calc_tag(env->prog);
17074 for (i = 0; i < insn_cnt; i++, insn++) {
17075 if (BPF_CLASS(insn->code) == BPF_LDX &&
17076 ((BPF_MODE(insn->code) != BPF_MEM && BPF_MODE(insn->code) != BPF_MEMSX) ||
17078 verbose(env, "BPF_LDX uses reserved fields\n");
17082 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
17083 struct bpf_insn_aux_data *aux;
17084 struct bpf_map *map;
17089 if (i == insn_cnt - 1 || insn[1].code != 0 ||
17090 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
17091 insn[1].off != 0) {
17092 verbose(env, "invalid bpf_ld_imm64 insn\n");
17096 if (insn[0].src_reg == 0)
17097 /* valid generic load 64-bit imm */
17100 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
17101 aux = &env->insn_aux_data[i];
17102 err = check_pseudo_btf_id(env, insn, aux);
17108 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
17109 aux = &env->insn_aux_data[i];
17110 aux->ptr_type = PTR_TO_FUNC;
17114 /* In final convert_pseudo_ld_imm64() step, this is
17115 * converted into regular 64-bit imm load insn.
17117 switch (insn[0].src_reg) {
17118 case BPF_PSEUDO_MAP_VALUE:
17119 case BPF_PSEUDO_MAP_IDX_VALUE:
17121 case BPF_PSEUDO_MAP_FD:
17122 case BPF_PSEUDO_MAP_IDX:
17123 if (insn[1].imm == 0)
17127 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
17131 switch (insn[0].src_reg) {
17132 case BPF_PSEUDO_MAP_IDX_VALUE:
17133 case BPF_PSEUDO_MAP_IDX:
17134 if (bpfptr_is_null(env->fd_array)) {
17135 verbose(env, "fd_idx without fd_array is invalid\n");
17138 if (copy_from_bpfptr_offset(&fd, env->fd_array,
17139 insn[0].imm * sizeof(fd),
17149 map = __bpf_map_get(f);
17151 verbose(env, "fd %d is not pointing to valid bpf_map\n",
17153 return PTR_ERR(map);
17156 err = check_map_prog_compatibility(env, map, env->prog);
17162 aux = &env->insn_aux_data[i];
17163 if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
17164 insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
17165 addr = (unsigned long)map;
17167 u32 off = insn[1].imm;
17169 if (off >= BPF_MAX_VAR_OFF) {
17170 verbose(env, "direct value offset of %u is not allowed\n", off);
17175 if (!map->ops->map_direct_value_addr) {
17176 verbose(env, "no direct value access support for this map type\n");
17181 err = map->ops->map_direct_value_addr(map, &addr, off);
17183 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
17184 map->value_size, off);
17189 aux->map_off = off;
17193 insn[0].imm = (u32)addr;
17194 insn[1].imm = addr >> 32;
17196 /* check whether we recorded this map already */
17197 for (j = 0; j < env->used_map_cnt; j++) {
17198 if (env->used_maps[j] == map) {
17199 aux->map_index = j;
17205 if (env->used_map_cnt >= MAX_USED_MAPS) {
17210 /* hold the map. If the program is rejected by verifier,
17211 * the map will be released by release_maps() or it
17212 * will be used by the valid program until it's unloaded
17213 * and all maps are released in free_used_maps()
17217 aux->map_index = env->used_map_cnt;
17218 env->used_maps[env->used_map_cnt++] = map;
17220 if (bpf_map_is_cgroup_storage(map) &&
17221 bpf_cgroup_storage_assign(env->prog->aux, map)) {
17222 verbose(env, "only one cgroup storage of each type is allowed\n");
17234 /* Basic sanity check before we invest more work here. */
17235 if (!bpf_opcode_in_insntable(insn->code)) {
17236 verbose(env, "unknown opcode %02x\n", insn->code);
17241 /* now all pseudo BPF_LD_IMM64 instructions load valid
17242 * 'struct bpf_map *' into a register instead of user map_fd.
17243 * These pointers will be used later by verifier to validate map access.
17248 /* drop refcnt of maps used by the rejected program */
17249 static void release_maps(struct bpf_verifier_env *env)
17251 __bpf_free_used_maps(env->prog->aux, env->used_maps,
17252 env->used_map_cnt);
17255 /* drop refcnt of maps used by the rejected program */
17256 static void release_btfs(struct bpf_verifier_env *env)
17258 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
17259 env->used_btf_cnt);
17262 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
17263 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
17265 struct bpf_insn *insn = env->prog->insnsi;
17266 int insn_cnt = env->prog->len;
17269 for (i = 0; i < insn_cnt; i++, insn++) {
17270 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
17272 if (insn->src_reg == BPF_PSEUDO_FUNC)
17278 /* single env->prog->insni[off] instruction was replaced with the range
17279 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
17280 * [0, off) and [off, end) to new locations, so the patched range stays zero
17282 static void adjust_insn_aux_data(struct bpf_verifier_env *env,
17283 struct bpf_insn_aux_data *new_data,
17284 struct bpf_prog *new_prog, u32 off, u32 cnt)
17286 struct bpf_insn_aux_data *old_data = env->insn_aux_data;
17287 struct bpf_insn *insn = new_prog->insnsi;
17288 u32 old_seen = old_data[off].seen;
17292 /* aux info at OFF always needs adjustment, no matter fast path
17293 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
17294 * original insn at old prog.
17296 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
17300 prog_len = new_prog->len;
17302 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
17303 memcpy(new_data + off + cnt - 1, old_data + off,
17304 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
17305 for (i = off; i < off + cnt - 1; i++) {
17306 /* Expand insni[off]'s seen count to the patched range. */
17307 new_data[i].seen = old_seen;
17308 new_data[i].zext_dst = insn_has_def32(env, insn + i);
17310 env->insn_aux_data = new_data;
17314 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
17320 /* NOTE: fake 'exit' subprog should be updated as well. */
17321 for (i = 0; i <= env->subprog_cnt; i++) {
17322 if (env->subprog_info[i].start <= off)
17324 env->subprog_info[i].start += len - 1;
17328 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
17330 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
17331 int i, sz = prog->aux->size_poke_tab;
17332 struct bpf_jit_poke_descriptor *desc;
17334 for (i = 0; i < sz; i++) {
17336 if (desc->insn_idx <= off)
17338 desc->insn_idx += len - 1;
17342 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
17343 const struct bpf_insn *patch, u32 len)
17345 struct bpf_prog *new_prog;
17346 struct bpf_insn_aux_data *new_data = NULL;
17349 new_data = vzalloc(array_size(env->prog->len + len - 1,
17350 sizeof(struct bpf_insn_aux_data)));
17355 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
17356 if (IS_ERR(new_prog)) {
17357 if (PTR_ERR(new_prog) == -ERANGE)
17359 "insn %d cannot be patched due to 16-bit range\n",
17360 env->insn_aux_data[off].orig_idx);
17364 adjust_insn_aux_data(env, new_data, new_prog, off, len);
17365 adjust_subprog_starts(env, off, len);
17366 adjust_poke_descs(new_prog, off, len);
17370 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
17375 /* find first prog starting at or after off (first to remove) */
17376 for (i = 0; i < env->subprog_cnt; i++)
17377 if (env->subprog_info[i].start >= off)
17379 /* find first prog starting at or after off + cnt (first to stay) */
17380 for (j = i; j < env->subprog_cnt; j++)
17381 if (env->subprog_info[j].start >= off + cnt)
17383 /* if j doesn't start exactly at off + cnt, we are just removing
17384 * the front of previous prog
17386 if (env->subprog_info[j].start != off + cnt)
17390 struct bpf_prog_aux *aux = env->prog->aux;
17393 /* move fake 'exit' subprog as well */
17394 move = env->subprog_cnt + 1 - j;
17396 memmove(env->subprog_info + i,
17397 env->subprog_info + j,
17398 sizeof(*env->subprog_info) * move);
17399 env->subprog_cnt -= j - i;
17401 /* remove func_info */
17402 if (aux->func_info) {
17403 move = aux->func_info_cnt - j;
17405 memmove(aux->func_info + i,
17406 aux->func_info + j,
17407 sizeof(*aux->func_info) * move);
17408 aux->func_info_cnt -= j - i;
17409 /* func_info->insn_off is set after all code rewrites,
17410 * in adjust_btf_func() - no need to adjust
17414 /* convert i from "first prog to remove" to "first to adjust" */
17415 if (env->subprog_info[i].start == off)
17419 /* update fake 'exit' subprog as well */
17420 for (; i <= env->subprog_cnt; i++)
17421 env->subprog_info[i].start -= cnt;
17426 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
17429 struct bpf_prog *prog = env->prog;
17430 u32 i, l_off, l_cnt, nr_linfo;
17431 struct bpf_line_info *linfo;
17433 nr_linfo = prog->aux->nr_linfo;
17437 linfo = prog->aux->linfo;
17439 /* find first line info to remove, count lines to be removed */
17440 for (i = 0; i < nr_linfo; i++)
17441 if (linfo[i].insn_off >= off)
17446 for (; i < nr_linfo; i++)
17447 if (linfo[i].insn_off < off + cnt)
17452 /* First live insn doesn't match first live linfo, it needs to "inherit"
17453 * last removed linfo. prog is already modified, so prog->len == off
17454 * means no live instructions after (tail of the program was removed).
17456 if (prog->len != off && l_cnt &&
17457 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
17459 linfo[--i].insn_off = off + cnt;
17462 /* remove the line info which refer to the removed instructions */
17464 memmove(linfo + l_off, linfo + i,
17465 sizeof(*linfo) * (nr_linfo - i));
17467 prog->aux->nr_linfo -= l_cnt;
17468 nr_linfo = prog->aux->nr_linfo;
17471 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
17472 for (i = l_off; i < nr_linfo; i++)
17473 linfo[i].insn_off -= cnt;
17475 /* fix up all subprogs (incl. 'exit') which start >= off */
17476 for (i = 0; i <= env->subprog_cnt; i++)
17477 if (env->subprog_info[i].linfo_idx > l_off) {
17478 /* program may have started in the removed region but
17479 * may not be fully removed
17481 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
17482 env->subprog_info[i].linfo_idx -= l_cnt;
17484 env->subprog_info[i].linfo_idx = l_off;
17490 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
17492 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17493 unsigned int orig_prog_len = env->prog->len;
17496 if (bpf_prog_is_offloaded(env->prog->aux))
17497 bpf_prog_offload_remove_insns(env, off, cnt);
17499 err = bpf_remove_insns(env->prog, off, cnt);
17503 err = adjust_subprog_starts_after_remove(env, off, cnt);
17507 err = bpf_adj_linfo_after_remove(env, off, cnt);
17511 memmove(aux_data + off, aux_data + off + cnt,
17512 sizeof(*aux_data) * (orig_prog_len - off - cnt));
17517 /* The verifier does more data flow analysis than llvm and will not
17518 * explore branches that are dead at run time. Malicious programs can
17519 * have dead code too. Therefore replace all dead at-run-time code
17522 * Just nops are not optimal, e.g. if they would sit at the end of the
17523 * program and through another bug we would manage to jump there, then
17524 * we'd execute beyond program memory otherwise. Returning exception
17525 * code also wouldn't work since we can have subprogs where the dead
17526 * code could be located.
17528 static void sanitize_dead_code(struct bpf_verifier_env *env)
17530 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17531 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
17532 struct bpf_insn *insn = env->prog->insnsi;
17533 const int insn_cnt = env->prog->len;
17536 for (i = 0; i < insn_cnt; i++) {
17537 if (aux_data[i].seen)
17539 memcpy(insn + i, &trap, sizeof(trap));
17540 aux_data[i].zext_dst = false;
17544 static bool insn_is_cond_jump(u8 code)
17549 if (BPF_CLASS(code) == BPF_JMP32)
17550 return op != BPF_JA;
17552 if (BPF_CLASS(code) != BPF_JMP)
17555 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
17558 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
17560 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17561 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
17562 struct bpf_insn *insn = env->prog->insnsi;
17563 const int insn_cnt = env->prog->len;
17566 for (i = 0; i < insn_cnt; i++, insn++) {
17567 if (!insn_is_cond_jump(insn->code))
17570 if (!aux_data[i + 1].seen)
17571 ja.off = insn->off;
17572 else if (!aux_data[i + 1 + insn->off].seen)
17577 if (bpf_prog_is_offloaded(env->prog->aux))
17578 bpf_prog_offload_replace_insn(env, i, &ja);
17580 memcpy(insn, &ja, sizeof(ja));
17584 static int opt_remove_dead_code(struct bpf_verifier_env *env)
17586 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17587 int insn_cnt = env->prog->len;
17590 for (i = 0; i < insn_cnt; i++) {
17594 while (i + j < insn_cnt && !aux_data[i + j].seen)
17599 err = verifier_remove_insns(env, i, j);
17602 insn_cnt = env->prog->len;
17608 static int opt_remove_nops(struct bpf_verifier_env *env)
17610 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
17611 struct bpf_insn *insn = env->prog->insnsi;
17612 int insn_cnt = env->prog->len;
17615 for (i = 0; i < insn_cnt; i++) {
17616 if (memcmp(&insn[i], &ja, sizeof(ja)))
17619 err = verifier_remove_insns(env, i, 1);
17629 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
17630 const union bpf_attr *attr)
17632 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
17633 struct bpf_insn_aux_data *aux = env->insn_aux_data;
17634 int i, patch_len, delta = 0, len = env->prog->len;
17635 struct bpf_insn *insns = env->prog->insnsi;
17636 struct bpf_prog *new_prog;
17639 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
17640 zext_patch[1] = BPF_ZEXT_REG(0);
17641 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
17642 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
17643 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
17644 for (i = 0; i < len; i++) {
17645 int adj_idx = i + delta;
17646 struct bpf_insn insn;
17649 insn = insns[adj_idx];
17650 load_reg = insn_def_regno(&insn);
17651 if (!aux[adj_idx].zext_dst) {
17659 class = BPF_CLASS(code);
17660 if (load_reg == -1)
17663 /* NOTE: arg "reg" (the fourth one) is only used for
17664 * BPF_STX + SRC_OP, so it is safe to pass NULL
17667 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
17668 if (class == BPF_LD &&
17669 BPF_MODE(code) == BPF_IMM)
17674 /* ctx load could be transformed into wider load. */
17675 if (class == BPF_LDX &&
17676 aux[adj_idx].ptr_type == PTR_TO_CTX)
17679 imm_rnd = get_random_u32();
17680 rnd_hi32_patch[0] = insn;
17681 rnd_hi32_patch[1].imm = imm_rnd;
17682 rnd_hi32_patch[3].dst_reg = load_reg;
17683 patch = rnd_hi32_patch;
17685 goto apply_patch_buffer;
17688 /* Add in an zero-extend instruction if a) the JIT has requested
17689 * it or b) it's a CMPXCHG.
17691 * The latter is because: BPF_CMPXCHG always loads a value into
17692 * R0, therefore always zero-extends. However some archs'
17693 * equivalent instruction only does this load when the
17694 * comparison is successful. This detail of CMPXCHG is
17695 * orthogonal to the general zero-extension behaviour of the
17696 * CPU, so it's treated independently of bpf_jit_needs_zext.
17698 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
17701 /* Zero-extension is done by the caller. */
17702 if (bpf_pseudo_kfunc_call(&insn))
17705 if (WARN_ON(load_reg == -1)) {
17706 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
17710 zext_patch[0] = insn;
17711 zext_patch[1].dst_reg = load_reg;
17712 zext_patch[1].src_reg = load_reg;
17713 patch = zext_patch;
17715 apply_patch_buffer:
17716 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
17719 env->prog = new_prog;
17720 insns = new_prog->insnsi;
17721 aux = env->insn_aux_data;
17722 delta += patch_len - 1;
17728 /* convert load instructions that access fields of a context type into a
17729 * sequence of instructions that access fields of the underlying structure:
17730 * struct __sk_buff -> struct sk_buff
17731 * struct bpf_sock_ops -> struct sock
17733 static int convert_ctx_accesses(struct bpf_verifier_env *env)
17735 const struct bpf_verifier_ops *ops = env->ops;
17736 int i, cnt, size, ctx_field_size, delta = 0;
17737 const int insn_cnt = env->prog->len;
17738 struct bpf_insn insn_buf[16], *insn;
17739 u32 target_size, size_default, off;
17740 struct bpf_prog *new_prog;
17741 enum bpf_access_type type;
17742 bool is_narrower_load;
17744 if (ops->gen_prologue || env->seen_direct_write) {
17745 if (!ops->gen_prologue) {
17746 verbose(env, "bpf verifier is misconfigured\n");
17749 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
17751 if (cnt >= ARRAY_SIZE(insn_buf)) {
17752 verbose(env, "bpf verifier is misconfigured\n");
17755 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
17759 env->prog = new_prog;
17764 if (bpf_prog_is_offloaded(env->prog->aux))
17767 insn = env->prog->insnsi + delta;
17769 for (i = 0; i < insn_cnt; i++, insn++) {
17770 bpf_convert_ctx_access_t convert_ctx_access;
17773 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
17774 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
17775 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
17776 insn->code == (BPF_LDX | BPF_MEM | BPF_DW) ||
17777 insn->code == (BPF_LDX | BPF_MEMSX | BPF_B) ||
17778 insn->code == (BPF_LDX | BPF_MEMSX | BPF_H) ||
17779 insn->code == (BPF_LDX | BPF_MEMSX | BPF_W)) {
17781 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
17782 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
17783 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
17784 insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
17785 insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
17786 insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
17787 insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
17788 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
17794 if (type == BPF_WRITE &&
17795 env->insn_aux_data[i + delta].sanitize_stack_spill) {
17796 struct bpf_insn patch[] = {
17801 cnt = ARRAY_SIZE(patch);
17802 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
17807 env->prog = new_prog;
17808 insn = new_prog->insnsi + i + delta;
17812 switch ((int)env->insn_aux_data[i + delta].ptr_type) {
17814 if (!ops->convert_ctx_access)
17816 convert_ctx_access = ops->convert_ctx_access;
17818 case PTR_TO_SOCKET:
17819 case PTR_TO_SOCK_COMMON:
17820 convert_ctx_access = bpf_sock_convert_ctx_access;
17822 case PTR_TO_TCP_SOCK:
17823 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
17825 case PTR_TO_XDP_SOCK:
17826 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
17828 case PTR_TO_BTF_ID:
17829 case PTR_TO_BTF_ID | PTR_UNTRUSTED:
17830 /* PTR_TO_BTF_ID | MEM_ALLOC always has a valid lifetime, unlike
17831 * PTR_TO_BTF_ID, and an active ref_obj_id, but the same cannot
17832 * be said once it is marked PTR_UNTRUSTED, hence we must handle
17833 * any faults for loads into such types. BPF_WRITE is disallowed
17836 case PTR_TO_BTF_ID | MEM_ALLOC | PTR_UNTRUSTED:
17837 if (type == BPF_READ) {
17838 if (BPF_MODE(insn->code) == BPF_MEM)
17839 insn->code = BPF_LDX | BPF_PROBE_MEM |
17840 BPF_SIZE((insn)->code);
17842 insn->code = BPF_LDX | BPF_PROBE_MEMSX |
17843 BPF_SIZE((insn)->code);
17844 env->prog->aux->num_exentries++;
17851 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
17852 size = BPF_LDST_BYTES(insn);
17853 mode = BPF_MODE(insn->code);
17855 /* If the read access is a narrower load of the field,
17856 * convert to a 4/8-byte load, to minimum program type specific
17857 * convert_ctx_access changes. If conversion is successful,
17858 * we will apply proper mask to the result.
17860 is_narrower_load = size < ctx_field_size;
17861 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
17863 if (is_narrower_load) {
17866 if (type == BPF_WRITE) {
17867 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
17872 if (ctx_field_size == 4)
17874 else if (ctx_field_size == 8)
17875 size_code = BPF_DW;
17877 insn->off = off & ~(size_default - 1);
17878 insn->code = BPF_LDX | BPF_MEM | size_code;
17882 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
17884 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
17885 (ctx_field_size && !target_size)) {
17886 verbose(env, "bpf verifier is misconfigured\n");
17890 if (is_narrower_load && size < target_size) {
17891 u8 shift = bpf_ctx_narrow_access_offset(
17892 off, size, size_default) * 8;
17893 if (shift && cnt + 1 >= ARRAY_SIZE(insn_buf)) {
17894 verbose(env, "bpf verifier narrow ctx load misconfigured\n");
17897 if (ctx_field_size <= 4) {
17899 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
17902 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
17903 (1 << size * 8) - 1);
17906 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
17909 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
17910 (1ULL << size * 8) - 1);
17913 if (mode == BPF_MEMSX)
17914 insn_buf[cnt++] = BPF_RAW_INSN(BPF_ALU64 | BPF_MOV | BPF_X,
17915 insn->dst_reg, insn->dst_reg,
17918 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
17924 /* keep walking new program and skip insns we just inserted */
17925 env->prog = new_prog;
17926 insn = new_prog->insnsi + i + delta;
17932 static int jit_subprogs(struct bpf_verifier_env *env)
17934 struct bpf_prog *prog = env->prog, **func, *tmp;
17935 int i, j, subprog_start, subprog_end = 0, len, subprog;
17936 struct bpf_map *map_ptr;
17937 struct bpf_insn *insn;
17938 void *old_bpf_func;
17939 int err, num_exentries;
17941 if (env->subprog_cnt <= 1)
17944 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
17945 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn))
17948 /* Upon error here we cannot fall back to interpreter but
17949 * need a hard reject of the program. Thus -EFAULT is
17950 * propagated in any case.
17952 subprog = find_subprog(env, i + insn->imm + 1);
17954 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
17955 i + insn->imm + 1);
17958 /* temporarily remember subprog id inside insn instead of
17959 * aux_data, since next loop will split up all insns into funcs
17961 insn->off = subprog;
17962 /* remember original imm in case JIT fails and fallback
17963 * to interpreter will be needed
17965 env->insn_aux_data[i].call_imm = insn->imm;
17966 /* point imm to __bpf_call_base+1 from JITs point of view */
17968 if (bpf_pseudo_func(insn))
17969 /* jit (e.g. x86_64) may emit fewer instructions
17970 * if it learns a u32 imm is the same as a u64 imm.
17971 * Force a non zero here.
17976 err = bpf_prog_alloc_jited_linfo(prog);
17978 goto out_undo_insn;
17981 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
17983 goto out_undo_insn;
17985 for (i = 0; i < env->subprog_cnt; i++) {
17986 subprog_start = subprog_end;
17987 subprog_end = env->subprog_info[i + 1].start;
17989 len = subprog_end - subprog_start;
17990 /* bpf_prog_run() doesn't call subprogs directly,
17991 * hence main prog stats include the runtime of subprogs.
17992 * subprogs don't have IDs and not reachable via prog_get_next_id
17993 * func[i]->stats will never be accessed and stays NULL
17995 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
17998 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
17999 len * sizeof(struct bpf_insn));
18000 func[i]->type = prog->type;
18001 func[i]->len = len;
18002 if (bpf_prog_calc_tag(func[i]))
18004 func[i]->is_func = 1;
18005 func[i]->aux->func_idx = i;
18006 /* Below members will be freed only at prog->aux */
18007 func[i]->aux->btf = prog->aux->btf;
18008 func[i]->aux->func_info = prog->aux->func_info;
18009 func[i]->aux->func_info_cnt = prog->aux->func_info_cnt;
18010 func[i]->aux->poke_tab = prog->aux->poke_tab;
18011 func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
18013 for (j = 0; j < prog->aux->size_poke_tab; j++) {
18014 struct bpf_jit_poke_descriptor *poke;
18016 poke = &prog->aux->poke_tab[j];
18017 if (poke->insn_idx < subprog_end &&
18018 poke->insn_idx >= subprog_start)
18019 poke->aux = func[i]->aux;
18022 func[i]->aux->name[0] = 'F';
18023 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
18024 func[i]->jit_requested = 1;
18025 func[i]->blinding_requested = prog->blinding_requested;
18026 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
18027 func[i]->aux->kfunc_btf_tab = prog->aux->kfunc_btf_tab;
18028 func[i]->aux->linfo = prog->aux->linfo;
18029 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
18030 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
18031 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
18033 insn = func[i]->insnsi;
18034 for (j = 0; j < func[i]->len; j++, insn++) {
18035 if (BPF_CLASS(insn->code) == BPF_LDX &&
18036 (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
18037 BPF_MODE(insn->code) == BPF_PROBE_MEMSX))
18040 func[i]->aux->num_exentries = num_exentries;
18041 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
18042 func[i] = bpf_int_jit_compile(func[i]);
18043 if (!func[i]->jited) {
18050 /* at this point all bpf functions were successfully JITed
18051 * now populate all bpf_calls with correct addresses and
18052 * run last pass of JIT
18054 for (i = 0; i < env->subprog_cnt; i++) {
18055 insn = func[i]->insnsi;
18056 for (j = 0; j < func[i]->len; j++, insn++) {
18057 if (bpf_pseudo_func(insn)) {
18058 subprog = insn->off;
18059 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
18060 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
18063 if (!bpf_pseudo_call(insn))
18065 subprog = insn->off;
18066 insn->imm = BPF_CALL_IMM(func[subprog]->bpf_func);
18069 /* we use the aux data to keep a list of the start addresses
18070 * of the JITed images for each function in the program
18072 * for some architectures, such as powerpc64, the imm field
18073 * might not be large enough to hold the offset of the start
18074 * address of the callee's JITed image from __bpf_call_base
18076 * in such cases, we can lookup the start address of a callee
18077 * by using its subprog id, available from the off field of
18078 * the call instruction, as an index for this list
18080 func[i]->aux->func = func;
18081 func[i]->aux->func_cnt = env->subprog_cnt;
18083 for (i = 0; i < env->subprog_cnt; i++) {
18084 old_bpf_func = func[i]->bpf_func;
18085 tmp = bpf_int_jit_compile(func[i]);
18086 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
18087 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
18094 /* finally lock prog and jit images for all functions and
18095 * populate kallsysm. Begin at the first subprogram, since
18096 * bpf_prog_load will add the kallsyms for the main program.
18098 for (i = 1; i < env->subprog_cnt; i++) {
18099 bpf_prog_lock_ro(func[i]);
18100 bpf_prog_kallsyms_add(func[i]);
18103 /* Last step: make now unused interpreter insns from main
18104 * prog consistent for later dump requests, so they can
18105 * later look the same as if they were interpreted only.
18107 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
18108 if (bpf_pseudo_func(insn)) {
18109 insn[0].imm = env->insn_aux_data[i].call_imm;
18110 insn[1].imm = insn->off;
18114 if (!bpf_pseudo_call(insn))
18116 insn->off = env->insn_aux_data[i].call_imm;
18117 subprog = find_subprog(env, i + insn->off + 1);
18118 insn->imm = subprog;
18122 prog->bpf_func = func[0]->bpf_func;
18123 prog->jited_len = func[0]->jited_len;
18124 prog->aux->extable = func[0]->aux->extable;
18125 prog->aux->num_exentries = func[0]->aux->num_exentries;
18126 prog->aux->func = func;
18127 prog->aux->func_cnt = env->subprog_cnt;
18128 bpf_prog_jit_attempt_done(prog);
18131 /* We failed JIT'ing, so at this point we need to unregister poke
18132 * descriptors from subprogs, so that kernel is not attempting to
18133 * patch it anymore as we're freeing the subprog JIT memory.
18135 for (i = 0; i < prog->aux->size_poke_tab; i++) {
18136 map_ptr = prog->aux->poke_tab[i].tail_call.map;
18137 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
18139 /* At this point we're guaranteed that poke descriptors are not
18140 * live anymore. We can just unlink its descriptor table as it's
18141 * released with the main prog.
18143 for (i = 0; i < env->subprog_cnt; i++) {
18146 func[i]->aux->poke_tab = NULL;
18147 bpf_jit_free(func[i]);
18151 /* cleanup main prog to be interpreted */
18152 prog->jit_requested = 0;
18153 prog->blinding_requested = 0;
18154 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
18155 if (!bpf_pseudo_call(insn))
18158 insn->imm = env->insn_aux_data[i].call_imm;
18160 bpf_prog_jit_attempt_done(prog);
18164 static int fixup_call_args(struct bpf_verifier_env *env)
18166 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
18167 struct bpf_prog *prog = env->prog;
18168 struct bpf_insn *insn = prog->insnsi;
18169 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
18174 if (env->prog->jit_requested &&
18175 !bpf_prog_is_offloaded(env->prog->aux)) {
18176 err = jit_subprogs(env);
18179 if (err == -EFAULT)
18182 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
18183 if (has_kfunc_call) {
18184 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
18187 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
18188 /* When JIT fails the progs with bpf2bpf calls and tail_calls
18189 * have to be rejected, since interpreter doesn't support them yet.
18191 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
18194 for (i = 0; i < prog->len; i++, insn++) {
18195 if (bpf_pseudo_func(insn)) {
18196 /* When JIT fails the progs with callback calls
18197 * have to be rejected, since interpreter doesn't support them yet.
18199 verbose(env, "callbacks are not allowed in non-JITed programs\n");
18203 if (!bpf_pseudo_call(insn))
18205 depth = get_callee_stack_depth(env, insn, i);
18208 bpf_patch_call_args(insn, depth);
18215 /* replace a generic kfunc with a specialized version if necessary */
18216 static void specialize_kfunc(struct bpf_verifier_env *env,
18217 u32 func_id, u16 offset, unsigned long *addr)
18219 struct bpf_prog *prog = env->prog;
18220 bool seen_direct_write;
18224 if (bpf_dev_bound_kfunc_id(func_id)) {
18225 xdp_kfunc = bpf_dev_bound_resolve_kfunc(prog, func_id);
18227 *addr = (unsigned long)xdp_kfunc;
18230 /* fallback to default kfunc when not supported by netdev */
18236 if (func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) {
18237 seen_direct_write = env->seen_direct_write;
18238 is_rdonly = !may_access_direct_pkt_data(env, NULL, BPF_WRITE);
18241 *addr = (unsigned long)bpf_dynptr_from_skb_rdonly;
18243 /* restore env->seen_direct_write to its original value, since
18244 * may_access_direct_pkt_data mutates it
18246 env->seen_direct_write = seen_direct_write;
18250 static void __fixup_collection_insert_kfunc(struct bpf_insn_aux_data *insn_aux,
18251 u16 struct_meta_reg,
18252 u16 node_offset_reg,
18253 struct bpf_insn *insn,
18254 struct bpf_insn *insn_buf,
18257 struct btf_struct_meta *kptr_struct_meta = insn_aux->kptr_struct_meta;
18258 struct bpf_insn addr[2] = { BPF_LD_IMM64(struct_meta_reg, (long)kptr_struct_meta) };
18260 insn_buf[0] = addr[0];
18261 insn_buf[1] = addr[1];
18262 insn_buf[2] = BPF_MOV64_IMM(node_offset_reg, insn_aux->insert_off);
18263 insn_buf[3] = *insn;
18267 static int fixup_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
18268 struct bpf_insn *insn_buf, int insn_idx, int *cnt)
18270 const struct bpf_kfunc_desc *desc;
18273 verbose(env, "invalid kernel function call not eliminated in verifier pass\n");
18279 /* insn->imm has the btf func_id. Replace it with an offset relative to
18280 * __bpf_call_base, unless the JIT needs to call functions that are
18281 * further than 32 bits away (bpf_jit_supports_far_kfunc_call()).
18283 desc = find_kfunc_desc(env->prog, insn->imm, insn->off);
18285 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
18290 if (!bpf_jit_supports_far_kfunc_call())
18291 insn->imm = BPF_CALL_IMM(desc->addr);
18294 if (desc->func_id == special_kfunc_list[KF_bpf_obj_new_impl]) {
18295 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18296 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
18297 u64 obj_new_size = env->insn_aux_data[insn_idx].obj_new_size;
18299 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_1, obj_new_size);
18300 insn_buf[1] = addr[0];
18301 insn_buf[2] = addr[1];
18302 insn_buf[3] = *insn;
18304 } else if (desc->func_id == special_kfunc_list[KF_bpf_obj_drop_impl] ||
18305 desc->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) {
18306 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18307 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
18309 if (desc->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl] &&
18310 !kptr_struct_meta) {
18311 verbose(env, "verifier internal error: kptr_struct_meta expected at insn_idx %d\n",
18316 insn_buf[0] = addr[0];
18317 insn_buf[1] = addr[1];
18318 insn_buf[2] = *insn;
18320 } else if (desc->func_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
18321 desc->func_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
18322 desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
18323 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18324 int struct_meta_reg = BPF_REG_3;
18325 int node_offset_reg = BPF_REG_4;
18327 /* rbtree_add has extra 'less' arg, so args-to-fixup are in diff regs */
18328 if (desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
18329 struct_meta_reg = BPF_REG_4;
18330 node_offset_reg = BPF_REG_5;
18333 if (!kptr_struct_meta) {
18334 verbose(env, "verifier internal error: kptr_struct_meta expected at insn_idx %d\n",
18339 __fixup_collection_insert_kfunc(&env->insn_aux_data[insn_idx], struct_meta_reg,
18340 node_offset_reg, insn, insn_buf, cnt);
18341 } else if (desc->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx] ||
18342 desc->func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
18343 insn_buf[0] = BPF_MOV64_REG(BPF_REG_0, BPF_REG_1);
18349 /* Do various post-verification rewrites in a single program pass.
18350 * These rewrites simplify JIT and interpreter implementations.
18352 static int do_misc_fixups(struct bpf_verifier_env *env)
18354 struct bpf_prog *prog = env->prog;
18355 enum bpf_attach_type eatype = prog->expected_attach_type;
18356 enum bpf_prog_type prog_type = resolve_prog_type(prog);
18357 struct bpf_insn *insn = prog->insnsi;
18358 const struct bpf_func_proto *fn;
18359 const int insn_cnt = prog->len;
18360 const struct bpf_map_ops *ops;
18361 struct bpf_insn_aux_data *aux;
18362 struct bpf_insn insn_buf[16];
18363 struct bpf_prog *new_prog;
18364 struct bpf_map *map_ptr;
18365 int i, ret, cnt, delta = 0;
18367 for (i = 0; i < insn_cnt; i++, insn++) {
18368 /* Make divide-by-zero exceptions impossible. */
18369 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
18370 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
18371 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
18372 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
18373 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
18374 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
18375 struct bpf_insn *patchlet;
18376 struct bpf_insn chk_and_div[] = {
18377 /* [R,W]x div 0 -> 0 */
18378 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
18379 BPF_JNE | BPF_K, insn->src_reg,
18381 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
18382 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
18385 struct bpf_insn chk_and_mod[] = {
18386 /* [R,W]x mod 0 -> [R,W]x */
18387 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
18388 BPF_JEQ | BPF_K, insn->src_reg,
18389 0, 1 + (is64 ? 0 : 1), 0),
18391 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
18392 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
18395 patchlet = isdiv ? chk_and_div : chk_and_mod;
18396 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
18397 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
18399 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
18404 env->prog = prog = new_prog;
18405 insn = new_prog->insnsi + i + delta;
18409 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
18410 if (BPF_CLASS(insn->code) == BPF_LD &&
18411 (BPF_MODE(insn->code) == BPF_ABS ||
18412 BPF_MODE(insn->code) == BPF_IND)) {
18413 cnt = env->ops->gen_ld_abs(insn, insn_buf);
18414 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
18415 verbose(env, "bpf verifier is misconfigured\n");
18419 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18424 env->prog = prog = new_prog;
18425 insn = new_prog->insnsi + i + delta;
18429 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
18430 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
18431 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
18432 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
18433 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
18434 struct bpf_insn *patch = &insn_buf[0];
18435 bool issrc, isneg, isimm;
18438 aux = &env->insn_aux_data[i + delta];
18439 if (!aux->alu_state ||
18440 aux->alu_state == BPF_ALU_NON_POINTER)
18443 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
18444 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
18445 BPF_ALU_SANITIZE_SRC;
18446 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
18448 off_reg = issrc ? insn->src_reg : insn->dst_reg;
18450 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
18453 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
18454 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
18455 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
18456 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
18457 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
18458 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
18459 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
18462 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
18463 insn->src_reg = BPF_REG_AX;
18465 insn->code = insn->code == code_add ?
18466 code_sub : code_add;
18468 if (issrc && isneg && !isimm)
18469 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
18470 cnt = patch - insn_buf;
18472 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18477 env->prog = prog = new_prog;
18478 insn = new_prog->insnsi + i + delta;
18482 if (insn->code != (BPF_JMP | BPF_CALL))
18484 if (insn->src_reg == BPF_PSEUDO_CALL)
18486 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
18487 ret = fixup_kfunc_call(env, insn, insn_buf, i + delta, &cnt);
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 if (insn->imm == BPF_FUNC_get_route_realm)
18504 prog->dst_needed = 1;
18505 if (insn->imm == BPF_FUNC_get_prandom_u32)
18506 bpf_user_rnd_init_once();
18507 if (insn->imm == BPF_FUNC_override_return)
18508 prog->kprobe_override = 1;
18509 if (insn->imm == BPF_FUNC_tail_call) {
18510 /* If we tail call into other programs, we
18511 * cannot make any assumptions since they can
18512 * be replaced dynamically during runtime in
18513 * the program array.
18515 prog->cb_access = 1;
18516 if (!allow_tail_call_in_subprogs(env))
18517 prog->aux->stack_depth = MAX_BPF_STACK;
18518 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
18520 /* mark bpf_tail_call as different opcode to avoid
18521 * conditional branch in the interpreter for every normal
18522 * call and to prevent accidental JITing by JIT compiler
18523 * that doesn't support bpf_tail_call yet
18526 insn->code = BPF_JMP | BPF_TAIL_CALL;
18528 aux = &env->insn_aux_data[i + delta];
18529 if (env->bpf_capable && !prog->blinding_requested &&
18530 prog->jit_requested &&
18531 !bpf_map_key_poisoned(aux) &&
18532 !bpf_map_ptr_poisoned(aux) &&
18533 !bpf_map_ptr_unpriv(aux)) {
18534 struct bpf_jit_poke_descriptor desc = {
18535 .reason = BPF_POKE_REASON_TAIL_CALL,
18536 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
18537 .tail_call.key = bpf_map_key_immediate(aux),
18538 .insn_idx = i + delta,
18541 ret = bpf_jit_add_poke_descriptor(prog, &desc);
18543 verbose(env, "adding tail call poke descriptor failed\n");
18547 insn->imm = ret + 1;
18551 if (!bpf_map_ptr_unpriv(aux))
18554 /* instead of changing every JIT dealing with tail_call
18555 * emit two extra insns:
18556 * if (index >= max_entries) goto out;
18557 * index &= array->index_mask;
18558 * to avoid out-of-bounds cpu speculation
18560 if (bpf_map_ptr_poisoned(aux)) {
18561 verbose(env, "tail_call abusing map_ptr\n");
18565 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
18566 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
18567 map_ptr->max_entries, 2);
18568 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
18569 container_of(map_ptr,
18572 insn_buf[2] = *insn;
18574 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18579 env->prog = prog = new_prog;
18580 insn = new_prog->insnsi + i + delta;
18584 if (insn->imm == BPF_FUNC_timer_set_callback) {
18585 /* The verifier will process callback_fn as many times as necessary
18586 * with different maps and the register states prepared by
18587 * set_timer_callback_state will be accurate.
18589 * The following use case is valid:
18590 * map1 is shared by prog1, prog2, prog3.
18591 * prog1 calls bpf_timer_init for some map1 elements
18592 * prog2 calls bpf_timer_set_callback for some map1 elements.
18593 * Those that were not bpf_timer_init-ed will return -EINVAL.
18594 * prog3 calls bpf_timer_start for some map1 elements.
18595 * Those that were not both bpf_timer_init-ed and
18596 * bpf_timer_set_callback-ed will return -EINVAL.
18598 struct bpf_insn ld_addrs[2] = {
18599 BPF_LD_IMM64(BPF_REG_3, (long)prog->aux),
18602 insn_buf[0] = ld_addrs[0];
18603 insn_buf[1] = ld_addrs[1];
18604 insn_buf[2] = *insn;
18607 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18612 env->prog = prog = new_prog;
18613 insn = new_prog->insnsi + i + delta;
18614 goto patch_call_imm;
18617 if (is_storage_get_function(insn->imm)) {
18618 if (!env->prog->aux->sleepable ||
18619 env->insn_aux_data[i + delta].storage_get_func_atomic)
18620 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_ATOMIC);
18622 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_KERNEL);
18623 insn_buf[1] = *insn;
18626 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18631 env->prog = prog = new_prog;
18632 insn = new_prog->insnsi + i + delta;
18633 goto patch_call_imm;
18636 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
18637 * and other inlining handlers are currently limited to 64 bit
18640 if (prog->jit_requested && BITS_PER_LONG == 64 &&
18641 (insn->imm == BPF_FUNC_map_lookup_elem ||
18642 insn->imm == BPF_FUNC_map_update_elem ||
18643 insn->imm == BPF_FUNC_map_delete_elem ||
18644 insn->imm == BPF_FUNC_map_push_elem ||
18645 insn->imm == BPF_FUNC_map_pop_elem ||
18646 insn->imm == BPF_FUNC_map_peek_elem ||
18647 insn->imm == BPF_FUNC_redirect_map ||
18648 insn->imm == BPF_FUNC_for_each_map_elem ||
18649 insn->imm == BPF_FUNC_map_lookup_percpu_elem)) {
18650 aux = &env->insn_aux_data[i + delta];
18651 if (bpf_map_ptr_poisoned(aux))
18652 goto patch_call_imm;
18654 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
18655 ops = map_ptr->ops;
18656 if (insn->imm == BPF_FUNC_map_lookup_elem &&
18657 ops->map_gen_lookup) {
18658 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
18659 if (cnt == -EOPNOTSUPP)
18660 goto patch_map_ops_generic;
18661 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
18662 verbose(env, "bpf verifier is misconfigured\n");
18666 new_prog = bpf_patch_insn_data(env, i + delta,
18672 env->prog = prog = new_prog;
18673 insn = new_prog->insnsi + i + delta;
18677 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
18678 (void *(*)(struct bpf_map *map, void *key))NULL));
18679 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
18680 (long (*)(struct bpf_map *map, void *key))NULL));
18681 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
18682 (long (*)(struct bpf_map *map, void *key, void *value,
18684 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
18685 (long (*)(struct bpf_map *map, void *value,
18687 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
18688 (long (*)(struct bpf_map *map, void *value))NULL));
18689 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
18690 (long (*)(struct bpf_map *map, void *value))NULL));
18691 BUILD_BUG_ON(!__same_type(ops->map_redirect,
18692 (long (*)(struct bpf_map *map, u64 index, u64 flags))NULL));
18693 BUILD_BUG_ON(!__same_type(ops->map_for_each_callback,
18694 (long (*)(struct bpf_map *map,
18695 bpf_callback_t callback_fn,
18696 void *callback_ctx,
18698 BUILD_BUG_ON(!__same_type(ops->map_lookup_percpu_elem,
18699 (void *(*)(struct bpf_map *map, void *key, u32 cpu))NULL));
18701 patch_map_ops_generic:
18702 switch (insn->imm) {
18703 case BPF_FUNC_map_lookup_elem:
18704 insn->imm = BPF_CALL_IMM(ops->map_lookup_elem);
18706 case BPF_FUNC_map_update_elem:
18707 insn->imm = BPF_CALL_IMM(ops->map_update_elem);
18709 case BPF_FUNC_map_delete_elem:
18710 insn->imm = BPF_CALL_IMM(ops->map_delete_elem);
18712 case BPF_FUNC_map_push_elem:
18713 insn->imm = BPF_CALL_IMM(ops->map_push_elem);
18715 case BPF_FUNC_map_pop_elem:
18716 insn->imm = BPF_CALL_IMM(ops->map_pop_elem);
18718 case BPF_FUNC_map_peek_elem:
18719 insn->imm = BPF_CALL_IMM(ops->map_peek_elem);
18721 case BPF_FUNC_redirect_map:
18722 insn->imm = BPF_CALL_IMM(ops->map_redirect);
18724 case BPF_FUNC_for_each_map_elem:
18725 insn->imm = BPF_CALL_IMM(ops->map_for_each_callback);
18727 case BPF_FUNC_map_lookup_percpu_elem:
18728 insn->imm = BPF_CALL_IMM(ops->map_lookup_percpu_elem);
18732 goto patch_call_imm;
18735 /* Implement bpf_jiffies64 inline. */
18736 if (prog->jit_requested && BITS_PER_LONG == 64 &&
18737 insn->imm == BPF_FUNC_jiffies64) {
18738 struct bpf_insn ld_jiffies_addr[2] = {
18739 BPF_LD_IMM64(BPF_REG_0,
18740 (unsigned long)&jiffies),
18743 insn_buf[0] = ld_jiffies_addr[0];
18744 insn_buf[1] = ld_jiffies_addr[1];
18745 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
18749 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
18755 env->prog = prog = new_prog;
18756 insn = new_prog->insnsi + i + delta;
18760 /* Implement bpf_get_func_arg inline. */
18761 if (prog_type == BPF_PROG_TYPE_TRACING &&
18762 insn->imm == BPF_FUNC_get_func_arg) {
18763 /* Load nr_args from ctx - 8 */
18764 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
18765 insn_buf[1] = BPF_JMP32_REG(BPF_JGE, BPF_REG_2, BPF_REG_0, 6);
18766 insn_buf[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_2, 3);
18767 insn_buf[3] = BPF_ALU64_REG(BPF_ADD, BPF_REG_2, BPF_REG_1);
18768 insn_buf[4] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_2, 0);
18769 insn_buf[5] = BPF_STX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
18770 insn_buf[6] = BPF_MOV64_IMM(BPF_REG_0, 0);
18771 insn_buf[7] = BPF_JMP_A(1);
18772 insn_buf[8] = BPF_MOV64_IMM(BPF_REG_0, -EINVAL);
18775 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18780 env->prog = prog = new_prog;
18781 insn = new_prog->insnsi + i + delta;
18785 /* Implement bpf_get_func_ret inline. */
18786 if (prog_type == BPF_PROG_TYPE_TRACING &&
18787 insn->imm == BPF_FUNC_get_func_ret) {
18788 if (eatype == BPF_TRACE_FEXIT ||
18789 eatype == BPF_MODIFY_RETURN) {
18790 /* Load nr_args from ctx - 8 */
18791 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
18792 insn_buf[1] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_0, 3);
18793 insn_buf[2] = BPF_ALU64_REG(BPF_ADD, BPF_REG_0, BPF_REG_1);
18794 insn_buf[3] = BPF_LDX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
18795 insn_buf[4] = BPF_STX_MEM(BPF_DW, BPF_REG_2, BPF_REG_3, 0);
18796 insn_buf[5] = BPF_MOV64_IMM(BPF_REG_0, 0);
18799 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, -EOPNOTSUPP);
18803 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18808 env->prog = prog = new_prog;
18809 insn = new_prog->insnsi + i + delta;
18813 /* Implement get_func_arg_cnt inline. */
18814 if (prog_type == BPF_PROG_TYPE_TRACING &&
18815 insn->imm == BPF_FUNC_get_func_arg_cnt) {
18816 /* Load nr_args from ctx - 8 */
18817 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
18819 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
18823 env->prog = prog = new_prog;
18824 insn = new_prog->insnsi + i + delta;
18828 /* Implement bpf_get_func_ip inline. */
18829 if (prog_type == BPF_PROG_TYPE_TRACING &&
18830 insn->imm == BPF_FUNC_get_func_ip) {
18831 /* Load IP address from ctx - 16 */
18832 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -16);
18834 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
18838 env->prog = prog = new_prog;
18839 insn = new_prog->insnsi + i + delta;
18844 fn = env->ops->get_func_proto(insn->imm, env->prog);
18845 /* all functions that have prototype and verifier allowed
18846 * programs to call them, must be real in-kernel functions
18850 "kernel subsystem misconfigured func %s#%d\n",
18851 func_id_name(insn->imm), insn->imm);
18854 insn->imm = fn->func - __bpf_call_base;
18857 /* Since poke tab is now finalized, publish aux to tracker. */
18858 for (i = 0; i < prog->aux->size_poke_tab; i++) {
18859 map_ptr = prog->aux->poke_tab[i].tail_call.map;
18860 if (!map_ptr->ops->map_poke_track ||
18861 !map_ptr->ops->map_poke_untrack ||
18862 !map_ptr->ops->map_poke_run) {
18863 verbose(env, "bpf verifier is misconfigured\n");
18867 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
18869 verbose(env, "tracking tail call prog failed\n");
18874 sort_kfunc_descs_by_imm_off(env->prog);
18879 static struct bpf_prog *inline_bpf_loop(struct bpf_verifier_env *env,
18882 u32 callback_subprogno,
18885 s32 r6_offset = stack_base + 0 * BPF_REG_SIZE;
18886 s32 r7_offset = stack_base + 1 * BPF_REG_SIZE;
18887 s32 r8_offset = stack_base + 2 * BPF_REG_SIZE;
18888 int reg_loop_max = BPF_REG_6;
18889 int reg_loop_cnt = BPF_REG_7;
18890 int reg_loop_ctx = BPF_REG_8;
18892 struct bpf_prog *new_prog;
18893 u32 callback_start;
18894 u32 call_insn_offset;
18895 s32 callback_offset;
18897 /* This represents an inlined version of bpf_iter.c:bpf_loop,
18898 * be careful to modify this code in sync.
18900 struct bpf_insn insn_buf[] = {
18901 /* Return error and jump to the end of the patch if
18902 * expected number of iterations is too big.
18904 BPF_JMP_IMM(BPF_JLE, BPF_REG_1, BPF_MAX_LOOPS, 2),
18905 BPF_MOV32_IMM(BPF_REG_0, -E2BIG),
18906 BPF_JMP_IMM(BPF_JA, 0, 0, 16),
18907 /* spill R6, R7, R8 to use these as loop vars */
18908 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_6, r6_offset),
18909 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_7, r7_offset),
18910 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_8, r8_offset),
18911 /* initialize loop vars */
18912 BPF_MOV64_REG(reg_loop_max, BPF_REG_1),
18913 BPF_MOV32_IMM(reg_loop_cnt, 0),
18914 BPF_MOV64_REG(reg_loop_ctx, BPF_REG_3),
18916 * if reg_loop_cnt >= reg_loop_max skip the loop body
18918 BPF_JMP_REG(BPF_JGE, reg_loop_cnt, reg_loop_max, 5),
18920 * correct callback offset would be set after patching
18922 BPF_MOV64_REG(BPF_REG_1, reg_loop_cnt),
18923 BPF_MOV64_REG(BPF_REG_2, reg_loop_ctx),
18925 /* increment loop counter */
18926 BPF_ALU64_IMM(BPF_ADD, reg_loop_cnt, 1),
18927 /* jump to loop header if callback returned 0 */
18928 BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, -6),
18929 /* return value of bpf_loop,
18930 * set R0 to the number of iterations
18932 BPF_MOV64_REG(BPF_REG_0, reg_loop_cnt),
18933 /* restore original values of R6, R7, R8 */
18934 BPF_LDX_MEM(BPF_DW, BPF_REG_6, BPF_REG_10, r6_offset),
18935 BPF_LDX_MEM(BPF_DW, BPF_REG_7, BPF_REG_10, r7_offset),
18936 BPF_LDX_MEM(BPF_DW, BPF_REG_8, BPF_REG_10, r8_offset),
18939 *cnt = ARRAY_SIZE(insn_buf);
18940 new_prog = bpf_patch_insn_data(env, position, insn_buf, *cnt);
18944 /* callback start is known only after patching */
18945 callback_start = env->subprog_info[callback_subprogno].start;
18946 /* Note: insn_buf[12] is an offset of BPF_CALL_REL instruction */
18947 call_insn_offset = position + 12;
18948 callback_offset = callback_start - call_insn_offset - 1;
18949 new_prog->insnsi[call_insn_offset].imm = callback_offset;
18954 static bool is_bpf_loop_call(struct bpf_insn *insn)
18956 return insn->code == (BPF_JMP | BPF_CALL) &&
18957 insn->src_reg == 0 &&
18958 insn->imm == BPF_FUNC_loop;
18961 /* For all sub-programs in the program (including main) check
18962 * insn_aux_data to see if there are bpf_loop calls that require
18963 * inlining. If such calls are found the calls are replaced with a
18964 * sequence of instructions produced by `inline_bpf_loop` function and
18965 * subprog stack_depth is increased by the size of 3 registers.
18966 * This stack space is used to spill values of the R6, R7, R8. These
18967 * registers are used to store the loop bound, counter and context
18970 static int optimize_bpf_loop(struct bpf_verifier_env *env)
18972 struct bpf_subprog_info *subprogs = env->subprog_info;
18973 int i, cur_subprog = 0, cnt, delta = 0;
18974 struct bpf_insn *insn = env->prog->insnsi;
18975 int insn_cnt = env->prog->len;
18976 u16 stack_depth = subprogs[cur_subprog].stack_depth;
18977 u16 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
18978 u16 stack_depth_extra = 0;
18980 for (i = 0; i < insn_cnt; i++, insn++) {
18981 struct bpf_loop_inline_state *inline_state =
18982 &env->insn_aux_data[i + delta].loop_inline_state;
18984 if (is_bpf_loop_call(insn) && inline_state->fit_for_inline) {
18985 struct bpf_prog *new_prog;
18987 stack_depth_extra = BPF_REG_SIZE * 3 + stack_depth_roundup;
18988 new_prog = inline_bpf_loop(env,
18990 -(stack_depth + stack_depth_extra),
18991 inline_state->callback_subprogno,
18997 env->prog = new_prog;
18998 insn = new_prog->insnsi + i + delta;
19001 if (subprogs[cur_subprog + 1].start == i + delta + 1) {
19002 subprogs[cur_subprog].stack_depth += stack_depth_extra;
19004 stack_depth = subprogs[cur_subprog].stack_depth;
19005 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
19006 stack_depth_extra = 0;
19010 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
19015 static void free_states(struct bpf_verifier_env *env)
19017 struct bpf_verifier_state_list *sl, *sln;
19020 sl = env->free_list;
19023 free_verifier_state(&sl->state, false);
19027 env->free_list = NULL;
19029 if (!env->explored_states)
19032 for (i = 0; i < state_htab_size(env); i++) {
19033 sl = env->explored_states[i];
19037 free_verifier_state(&sl->state, false);
19041 env->explored_states[i] = NULL;
19045 static int do_check_common(struct bpf_verifier_env *env, int subprog)
19047 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
19048 struct bpf_verifier_state *state;
19049 struct bpf_reg_state *regs;
19052 env->prev_linfo = NULL;
19055 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
19058 state->curframe = 0;
19059 state->speculative = false;
19060 state->branches = 1;
19061 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
19062 if (!state->frame[0]) {
19066 env->cur_state = state;
19067 init_func_state(env, state->frame[0],
19068 BPF_MAIN_FUNC /* callsite */,
19071 state->first_insn_idx = env->subprog_info[subprog].start;
19072 state->last_insn_idx = -1;
19074 regs = state->frame[state->curframe]->regs;
19075 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
19076 ret = btf_prepare_func_args(env, subprog, regs);
19079 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
19080 if (regs[i].type == PTR_TO_CTX)
19081 mark_reg_known_zero(env, regs, i);
19082 else if (regs[i].type == SCALAR_VALUE)
19083 mark_reg_unknown(env, regs, i);
19084 else if (base_type(regs[i].type) == PTR_TO_MEM) {
19085 const u32 mem_size = regs[i].mem_size;
19087 mark_reg_known_zero(env, regs, i);
19088 regs[i].mem_size = mem_size;
19089 regs[i].id = ++env->id_gen;
19093 /* 1st arg to a function */
19094 regs[BPF_REG_1].type = PTR_TO_CTX;
19095 mark_reg_known_zero(env, regs, BPF_REG_1);
19096 ret = btf_check_subprog_arg_match(env, subprog, regs);
19097 if (ret == -EFAULT)
19098 /* unlikely verifier bug. abort.
19099 * ret == 0 and ret < 0 are sadly acceptable for
19100 * main() function due to backward compatibility.
19101 * Like socket filter program may be written as:
19102 * int bpf_prog(struct pt_regs *ctx)
19103 * and never dereference that ctx in the program.
19104 * 'struct pt_regs' is a type mismatch for socket
19105 * filter that should be using 'struct __sk_buff'.
19110 ret = do_check(env);
19112 /* check for NULL is necessary, since cur_state can be freed inside
19113 * do_check() under memory pressure.
19115 if (env->cur_state) {
19116 free_verifier_state(env->cur_state, true);
19117 env->cur_state = NULL;
19119 while (!pop_stack(env, NULL, NULL, false));
19120 if (!ret && pop_log)
19121 bpf_vlog_reset(&env->log, 0);
19126 /* Verify all global functions in a BPF program one by one based on their BTF.
19127 * All global functions must pass verification. Otherwise the whole program is rejected.
19138 * foo() will be verified first for R1=any_scalar_value. During verification it
19139 * will be assumed that bar() already verified successfully and call to bar()
19140 * from foo() will be checked for type match only. Later bar() will be verified
19141 * independently to check that it's safe for R1=any_scalar_value.
19143 static int do_check_subprogs(struct bpf_verifier_env *env)
19145 struct bpf_prog_aux *aux = env->prog->aux;
19148 if (!aux->func_info)
19151 for (i = 1; i < env->subprog_cnt; i++) {
19152 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
19154 env->insn_idx = env->subprog_info[i].start;
19155 WARN_ON_ONCE(env->insn_idx == 0);
19156 ret = do_check_common(env, i);
19159 } else if (env->log.level & BPF_LOG_LEVEL) {
19161 "Func#%d is safe for any args that match its prototype\n",
19168 static int do_check_main(struct bpf_verifier_env *env)
19173 ret = do_check_common(env, 0);
19175 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
19180 static void print_verification_stats(struct bpf_verifier_env *env)
19184 if (env->log.level & BPF_LOG_STATS) {
19185 verbose(env, "verification time %lld usec\n",
19186 div_u64(env->verification_time, 1000));
19187 verbose(env, "stack depth ");
19188 for (i = 0; i < env->subprog_cnt; i++) {
19189 u32 depth = env->subprog_info[i].stack_depth;
19191 verbose(env, "%d", depth);
19192 if (i + 1 < env->subprog_cnt)
19195 verbose(env, "\n");
19197 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
19198 "total_states %d peak_states %d mark_read %d\n",
19199 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
19200 env->max_states_per_insn, env->total_states,
19201 env->peak_states, env->longest_mark_read_walk);
19204 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
19206 const struct btf_type *t, *func_proto;
19207 const struct bpf_struct_ops *st_ops;
19208 const struct btf_member *member;
19209 struct bpf_prog *prog = env->prog;
19210 u32 btf_id, member_idx;
19213 if (!prog->gpl_compatible) {
19214 verbose(env, "struct ops programs must have a GPL compatible license\n");
19218 btf_id = prog->aux->attach_btf_id;
19219 st_ops = bpf_struct_ops_find(btf_id);
19221 verbose(env, "attach_btf_id %u is not a supported struct\n",
19227 member_idx = prog->expected_attach_type;
19228 if (member_idx >= btf_type_vlen(t)) {
19229 verbose(env, "attach to invalid member idx %u of struct %s\n",
19230 member_idx, st_ops->name);
19234 member = &btf_type_member(t)[member_idx];
19235 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
19236 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
19239 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
19240 mname, member_idx, st_ops->name);
19244 if (st_ops->check_member) {
19245 int err = st_ops->check_member(t, member, prog);
19248 verbose(env, "attach to unsupported member %s of struct %s\n",
19249 mname, st_ops->name);
19254 prog->aux->attach_func_proto = func_proto;
19255 prog->aux->attach_func_name = mname;
19256 env->ops = st_ops->verifier_ops;
19260 #define SECURITY_PREFIX "security_"
19262 static int check_attach_modify_return(unsigned long addr, const char *func_name)
19264 if (within_error_injection_list(addr) ||
19265 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
19271 /* list of non-sleepable functions that are otherwise on
19272 * ALLOW_ERROR_INJECTION list
19274 BTF_SET_START(btf_non_sleepable_error_inject)
19275 /* Three functions below can be called from sleepable and non-sleepable context.
19276 * Assume non-sleepable from bpf safety point of view.
19278 BTF_ID(func, __filemap_add_folio)
19279 BTF_ID(func, should_fail_alloc_page)
19280 BTF_ID(func, should_failslab)
19281 BTF_SET_END(btf_non_sleepable_error_inject)
19283 static int check_non_sleepable_error_inject(u32 btf_id)
19285 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
19288 int bpf_check_attach_target(struct bpf_verifier_log *log,
19289 const struct bpf_prog *prog,
19290 const struct bpf_prog *tgt_prog,
19292 struct bpf_attach_target_info *tgt_info)
19294 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
19295 const char prefix[] = "btf_trace_";
19296 int ret = 0, subprog = -1, i;
19297 const struct btf_type *t;
19298 bool conservative = true;
19302 struct module *mod = NULL;
19305 bpf_log(log, "Tracing programs must provide btf_id\n");
19308 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
19311 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
19314 t = btf_type_by_id(btf, btf_id);
19316 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
19319 tname = btf_name_by_offset(btf, t->name_off);
19321 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
19325 struct bpf_prog_aux *aux = tgt_prog->aux;
19327 if (bpf_prog_is_dev_bound(prog->aux) &&
19328 !bpf_prog_dev_bound_match(prog, tgt_prog)) {
19329 bpf_log(log, "Target program bound device mismatch");
19333 for (i = 0; i < aux->func_info_cnt; i++)
19334 if (aux->func_info[i].type_id == btf_id) {
19338 if (subprog == -1) {
19339 bpf_log(log, "Subprog %s doesn't exist\n", tname);
19342 conservative = aux->func_info_aux[subprog].unreliable;
19343 if (prog_extension) {
19344 if (conservative) {
19346 "Cannot replace static functions\n");
19349 if (!prog->jit_requested) {
19351 "Extension programs should be JITed\n");
19355 if (!tgt_prog->jited) {
19356 bpf_log(log, "Can attach to only JITed progs\n");
19359 if (tgt_prog->type == prog->type) {
19360 /* Cannot fentry/fexit another fentry/fexit program.
19361 * Cannot attach program extension to another extension.
19362 * It's ok to attach fentry/fexit to extension program.
19364 bpf_log(log, "Cannot recursively attach\n");
19367 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
19369 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
19370 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
19371 /* Program extensions can extend all program types
19372 * except fentry/fexit. The reason is the following.
19373 * The fentry/fexit programs are used for performance
19374 * analysis, stats and can be attached to any program
19375 * type except themselves. When extension program is
19376 * replacing XDP function it is necessary to allow
19377 * performance analysis of all functions. Both original
19378 * XDP program and its program extension. Hence
19379 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
19380 * allowed. If extending of fentry/fexit was allowed it
19381 * would be possible to create long call chain
19382 * fentry->extension->fentry->extension beyond
19383 * reasonable stack size. Hence extending fentry is not
19386 bpf_log(log, "Cannot extend fentry/fexit\n");
19390 if (prog_extension) {
19391 bpf_log(log, "Cannot replace kernel functions\n");
19396 switch (prog->expected_attach_type) {
19397 case BPF_TRACE_RAW_TP:
19400 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
19403 if (!btf_type_is_typedef(t)) {
19404 bpf_log(log, "attach_btf_id %u is not a typedef\n",
19408 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
19409 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
19413 tname += sizeof(prefix) - 1;
19414 t = btf_type_by_id(btf, t->type);
19415 if (!btf_type_is_ptr(t))
19416 /* should never happen in valid vmlinux build */
19418 t = btf_type_by_id(btf, t->type);
19419 if (!btf_type_is_func_proto(t))
19420 /* should never happen in valid vmlinux build */
19424 case BPF_TRACE_ITER:
19425 if (!btf_type_is_func(t)) {
19426 bpf_log(log, "attach_btf_id %u is not a function\n",
19430 t = btf_type_by_id(btf, t->type);
19431 if (!btf_type_is_func_proto(t))
19433 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
19438 if (!prog_extension)
19441 case BPF_MODIFY_RETURN:
19443 case BPF_LSM_CGROUP:
19444 case BPF_TRACE_FENTRY:
19445 case BPF_TRACE_FEXIT:
19446 if (!btf_type_is_func(t)) {
19447 bpf_log(log, "attach_btf_id %u is not a function\n",
19451 if (prog_extension &&
19452 btf_check_type_match(log, prog, btf, t))
19454 t = btf_type_by_id(btf, t->type);
19455 if (!btf_type_is_func_proto(t))
19458 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
19459 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
19460 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
19463 if (tgt_prog && conservative)
19466 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
19472 addr = (long) tgt_prog->bpf_func;
19474 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
19476 if (btf_is_module(btf)) {
19477 mod = btf_try_get_module(btf);
19479 addr = find_kallsyms_symbol_value(mod, tname);
19483 addr = kallsyms_lookup_name(tname);
19488 "The address of function %s cannot be found\n",
19494 if (prog->aux->sleepable) {
19496 switch (prog->type) {
19497 case BPF_PROG_TYPE_TRACING:
19499 /* fentry/fexit/fmod_ret progs can be sleepable if they are
19500 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
19502 if (!check_non_sleepable_error_inject(btf_id) &&
19503 within_error_injection_list(addr))
19505 /* fentry/fexit/fmod_ret progs can also be sleepable if they are
19506 * in the fmodret id set with the KF_SLEEPABLE flag.
19509 u32 *flags = btf_kfunc_is_modify_return(btf, btf_id,
19512 if (flags && (*flags & KF_SLEEPABLE))
19516 case BPF_PROG_TYPE_LSM:
19517 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
19518 * Only some of them are sleepable.
19520 if (bpf_lsm_is_sleepable_hook(btf_id))
19528 bpf_log(log, "%s is not sleepable\n", tname);
19531 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
19534 bpf_log(log, "can't modify return codes of BPF programs\n");
19538 if (btf_kfunc_is_modify_return(btf, btf_id, prog) ||
19539 !check_attach_modify_return(addr, tname))
19543 bpf_log(log, "%s() is not modifiable\n", tname);
19550 tgt_info->tgt_addr = addr;
19551 tgt_info->tgt_name = tname;
19552 tgt_info->tgt_type = t;
19553 tgt_info->tgt_mod = mod;
19557 BTF_SET_START(btf_id_deny)
19560 BTF_ID(func, migrate_disable)
19561 BTF_ID(func, migrate_enable)
19563 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
19564 BTF_ID(func, rcu_read_unlock_strict)
19566 #if defined(CONFIG_DEBUG_PREEMPT) || defined(CONFIG_TRACE_PREEMPT_TOGGLE)
19567 BTF_ID(func, preempt_count_add)
19568 BTF_ID(func, preempt_count_sub)
19570 #ifdef CONFIG_PREEMPT_RCU
19571 BTF_ID(func, __rcu_read_lock)
19572 BTF_ID(func, __rcu_read_unlock)
19574 BTF_SET_END(btf_id_deny)
19576 static bool can_be_sleepable(struct bpf_prog *prog)
19578 if (prog->type == BPF_PROG_TYPE_TRACING) {
19579 switch (prog->expected_attach_type) {
19580 case BPF_TRACE_FENTRY:
19581 case BPF_TRACE_FEXIT:
19582 case BPF_MODIFY_RETURN:
19583 case BPF_TRACE_ITER:
19589 return prog->type == BPF_PROG_TYPE_LSM ||
19590 prog->type == BPF_PROG_TYPE_KPROBE /* only for uprobes */ ||
19591 prog->type == BPF_PROG_TYPE_STRUCT_OPS;
19594 static int check_attach_btf_id(struct bpf_verifier_env *env)
19596 struct bpf_prog *prog = env->prog;
19597 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
19598 struct bpf_attach_target_info tgt_info = {};
19599 u32 btf_id = prog->aux->attach_btf_id;
19600 struct bpf_trampoline *tr;
19604 if (prog->type == BPF_PROG_TYPE_SYSCALL) {
19605 if (prog->aux->sleepable)
19606 /* attach_btf_id checked to be zero already */
19608 verbose(env, "Syscall programs can only be sleepable\n");
19612 if (prog->aux->sleepable && !can_be_sleepable(prog)) {
19613 verbose(env, "Only fentry/fexit/fmod_ret, lsm, iter, uprobe, and struct_ops programs can be sleepable\n");
19617 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
19618 return check_struct_ops_btf_id(env);
19620 if (prog->type != BPF_PROG_TYPE_TRACING &&
19621 prog->type != BPF_PROG_TYPE_LSM &&
19622 prog->type != BPF_PROG_TYPE_EXT)
19625 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
19629 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
19630 /* to make freplace equivalent to their targets, they need to
19631 * inherit env->ops and expected_attach_type for the rest of the
19634 env->ops = bpf_verifier_ops[tgt_prog->type];
19635 prog->expected_attach_type = tgt_prog->expected_attach_type;
19638 /* store info about the attachment target that will be used later */
19639 prog->aux->attach_func_proto = tgt_info.tgt_type;
19640 prog->aux->attach_func_name = tgt_info.tgt_name;
19641 prog->aux->mod = tgt_info.tgt_mod;
19644 prog->aux->saved_dst_prog_type = tgt_prog->type;
19645 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
19648 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
19649 prog->aux->attach_btf_trace = true;
19651 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
19652 if (!bpf_iter_prog_supported(prog))
19657 if (prog->type == BPF_PROG_TYPE_LSM) {
19658 ret = bpf_lsm_verify_prog(&env->log, prog);
19661 } else if (prog->type == BPF_PROG_TYPE_TRACING &&
19662 btf_id_set_contains(&btf_id_deny, btf_id)) {
19666 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
19667 tr = bpf_trampoline_get(key, &tgt_info);
19671 if (tgt_prog && tgt_prog->aux->tail_call_reachable)
19672 tr->flags = BPF_TRAMP_F_TAIL_CALL_CTX;
19674 prog->aux->dst_trampoline = tr;
19678 struct btf *bpf_get_btf_vmlinux(void)
19680 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
19681 mutex_lock(&bpf_verifier_lock);
19683 btf_vmlinux = btf_parse_vmlinux();
19684 mutex_unlock(&bpf_verifier_lock);
19686 return btf_vmlinux;
19689 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr, __u32 uattr_size)
19691 u64 start_time = ktime_get_ns();
19692 struct bpf_verifier_env *env;
19693 int i, len, ret = -EINVAL, err;
19697 /* no program is valid */
19698 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
19701 /* 'struct bpf_verifier_env' can be global, but since it's not small,
19702 * allocate/free it every time bpf_check() is called
19704 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
19710 len = (*prog)->len;
19711 env->insn_aux_data =
19712 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
19714 if (!env->insn_aux_data)
19716 for (i = 0; i < len; i++)
19717 env->insn_aux_data[i].orig_idx = i;
19719 env->ops = bpf_verifier_ops[env->prog->type];
19720 env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
19721 is_priv = bpf_capable();
19723 bpf_get_btf_vmlinux();
19725 /* grab the mutex to protect few globals used by verifier */
19727 mutex_lock(&bpf_verifier_lock);
19729 /* user could have requested verbose verifier output
19730 * and supplied buffer to store the verification trace
19732 ret = bpf_vlog_init(&env->log, attr->log_level,
19733 (char __user *) (unsigned long) attr->log_buf,
19738 mark_verifier_state_clean(env);
19740 if (IS_ERR(btf_vmlinux)) {
19741 /* Either gcc or pahole or kernel are broken. */
19742 verbose(env, "in-kernel BTF is malformed\n");
19743 ret = PTR_ERR(btf_vmlinux);
19744 goto skip_full_check;
19747 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
19748 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
19749 env->strict_alignment = true;
19750 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
19751 env->strict_alignment = false;
19753 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
19754 env->allow_uninit_stack = bpf_allow_uninit_stack();
19755 env->bypass_spec_v1 = bpf_bypass_spec_v1();
19756 env->bypass_spec_v4 = bpf_bypass_spec_v4();
19757 env->bpf_capable = bpf_capable();
19760 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
19762 env->explored_states = kvcalloc(state_htab_size(env),
19763 sizeof(struct bpf_verifier_state_list *),
19766 if (!env->explored_states)
19767 goto skip_full_check;
19769 ret = add_subprog_and_kfunc(env);
19771 goto skip_full_check;
19773 ret = check_subprogs(env);
19775 goto skip_full_check;
19777 ret = check_btf_info(env, attr, uattr);
19779 goto skip_full_check;
19781 ret = check_attach_btf_id(env);
19783 goto skip_full_check;
19785 ret = resolve_pseudo_ldimm64(env);
19787 goto skip_full_check;
19789 if (bpf_prog_is_offloaded(env->prog->aux)) {
19790 ret = bpf_prog_offload_verifier_prep(env->prog);
19792 goto skip_full_check;
19795 ret = check_cfg(env);
19797 goto skip_full_check;
19799 ret = do_check_subprogs(env);
19800 ret = ret ?: do_check_main(env);
19802 if (ret == 0 && bpf_prog_is_offloaded(env->prog->aux))
19803 ret = bpf_prog_offload_finalize(env);
19806 kvfree(env->explored_states);
19809 ret = check_max_stack_depth(env);
19811 /* instruction rewrites happen after this point */
19813 ret = optimize_bpf_loop(env);
19817 opt_hard_wire_dead_code_branches(env);
19819 ret = opt_remove_dead_code(env);
19821 ret = opt_remove_nops(env);
19824 sanitize_dead_code(env);
19828 /* program is valid, convert *(u32*)(ctx + off) accesses */
19829 ret = convert_ctx_accesses(env);
19832 ret = do_misc_fixups(env);
19834 /* do 32-bit optimization after insn patching has done so those patched
19835 * insns could be handled correctly.
19837 if (ret == 0 && !bpf_prog_is_offloaded(env->prog->aux)) {
19838 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
19839 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
19844 ret = fixup_call_args(env);
19846 env->verification_time = ktime_get_ns() - start_time;
19847 print_verification_stats(env);
19848 env->prog->aux->verified_insns = env->insn_processed;
19850 /* preserve original error even if log finalization is successful */
19851 err = bpf_vlog_finalize(&env->log, &log_true_size);
19855 if (uattr_size >= offsetofend(union bpf_attr, log_true_size) &&
19856 copy_to_bpfptr_offset(uattr, offsetof(union bpf_attr, log_true_size),
19857 &log_true_size, sizeof(log_true_size))) {
19859 goto err_release_maps;
19863 goto err_release_maps;
19865 if (env->used_map_cnt) {
19866 /* if program passed verifier, update used_maps in bpf_prog_info */
19867 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
19868 sizeof(env->used_maps[0]),
19871 if (!env->prog->aux->used_maps) {
19873 goto err_release_maps;
19876 memcpy(env->prog->aux->used_maps, env->used_maps,
19877 sizeof(env->used_maps[0]) * env->used_map_cnt);
19878 env->prog->aux->used_map_cnt = env->used_map_cnt;
19880 if (env->used_btf_cnt) {
19881 /* if program passed verifier, update used_btfs in bpf_prog_aux */
19882 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
19883 sizeof(env->used_btfs[0]),
19885 if (!env->prog->aux->used_btfs) {
19887 goto err_release_maps;
19890 memcpy(env->prog->aux->used_btfs, env->used_btfs,
19891 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
19892 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
19894 if (env->used_map_cnt || env->used_btf_cnt) {
19895 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
19896 * bpf_ld_imm64 instructions
19898 convert_pseudo_ld_imm64(env);
19901 adjust_btf_func(env);
19904 if (!env->prog->aux->used_maps)
19905 /* if we didn't copy map pointers into bpf_prog_info, release
19906 * them now. Otherwise free_used_maps() will release them.
19909 if (!env->prog->aux->used_btfs)
19912 /* extension progs temporarily inherit the attach_type of their targets
19913 for verification purposes, so set it back to zero before returning
19915 if (env->prog->type == BPF_PROG_TYPE_EXT)
19916 env->prog->expected_attach_type = 0;
19921 mutex_unlock(&bpf_verifier_lock);
19922 vfree(env->insn_aux_data);