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");
1518 mark_verifier_state_clean(env);
1521 static inline u32 vlog_alignment(u32 pos)
1523 return round_up(max(pos + BPF_LOG_MIN_ALIGNMENT / 2, BPF_LOG_ALIGNMENT),
1524 BPF_LOG_MIN_ALIGNMENT) - pos - 1;
1527 static void print_insn_state(struct bpf_verifier_env *env,
1528 const struct bpf_func_state *state)
1530 if (env->prev_log_pos && env->prev_log_pos == env->log.end_pos) {
1531 /* remove new line character */
1532 bpf_vlog_reset(&env->log, env->prev_log_pos - 1);
1533 verbose(env, "%*c;", vlog_alignment(env->prev_insn_print_pos), ' ');
1535 verbose(env, "%d:", env->insn_idx);
1537 print_verifier_state(env, state, false);
1540 /* copy array src of length n * size bytes to dst. dst is reallocated if it's too
1541 * small to hold src. This is different from krealloc since we don't want to preserve
1542 * the contents of dst.
1544 * Leaves dst untouched if src is NULL or length is zero. Returns NULL if memory could
1547 static void *copy_array(void *dst, const void *src, size_t n, size_t size, gfp_t flags)
1553 if (ZERO_OR_NULL_PTR(src))
1556 if (unlikely(check_mul_overflow(n, size, &bytes)))
1559 alloc_bytes = max(ksize(orig), kmalloc_size_roundup(bytes));
1560 dst = krealloc(orig, alloc_bytes, flags);
1566 memcpy(dst, src, bytes);
1568 return dst ? dst : ZERO_SIZE_PTR;
1571 /* resize an array from old_n items to new_n items. the array is reallocated if it's too
1572 * small to hold new_n items. new items are zeroed out if the array grows.
1574 * Contrary to krealloc_array, does not free arr if new_n is zero.
1576 static void *realloc_array(void *arr, size_t old_n, size_t new_n, size_t size)
1581 if (!new_n || old_n == new_n)
1584 alloc_size = kmalloc_size_roundup(size_mul(new_n, size));
1585 new_arr = krealloc(arr, alloc_size, GFP_KERNEL);
1593 memset(arr + old_n * size, 0, (new_n - old_n) * size);
1596 return arr ? arr : ZERO_SIZE_PTR;
1599 static int copy_reference_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
1601 dst->refs = copy_array(dst->refs, src->refs, src->acquired_refs,
1602 sizeof(struct bpf_reference_state), GFP_KERNEL);
1606 dst->acquired_refs = src->acquired_refs;
1610 static int copy_stack_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
1612 size_t n = src->allocated_stack / BPF_REG_SIZE;
1614 dst->stack = copy_array(dst->stack, src->stack, n, sizeof(struct bpf_stack_state),
1619 dst->allocated_stack = src->allocated_stack;
1623 static int resize_reference_state(struct bpf_func_state *state, size_t n)
1625 state->refs = realloc_array(state->refs, state->acquired_refs, n,
1626 sizeof(struct bpf_reference_state));
1630 state->acquired_refs = n;
1634 static int grow_stack_state(struct bpf_func_state *state, int size)
1636 size_t old_n = state->allocated_stack / BPF_REG_SIZE, n = size / BPF_REG_SIZE;
1641 state->stack = realloc_array(state->stack, old_n, n, sizeof(struct bpf_stack_state));
1645 state->allocated_stack = size;
1649 /* Acquire a pointer id from the env and update the state->refs to include
1650 * this new pointer reference.
1651 * On success, returns a valid pointer id to associate with the register
1652 * On failure, returns a negative errno.
1654 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
1656 struct bpf_func_state *state = cur_func(env);
1657 int new_ofs = state->acquired_refs;
1660 err = resize_reference_state(state, state->acquired_refs + 1);
1664 state->refs[new_ofs].id = id;
1665 state->refs[new_ofs].insn_idx = insn_idx;
1666 state->refs[new_ofs].callback_ref = state->in_callback_fn ? state->frameno : 0;
1671 /* release function corresponding to acquire_reference_state(). Idempotent. */
1672 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
1676 last_idx = state->acquired_refs - 1;
1677 for (i = 0; i < state->acquired_refs; i++) {
1678 if (state->refs[i].id == ptr_id) {
1679 /* Cannot release caller references in callbacks */
1680 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno)
1682 if (last_idx && i != last_idx)
1683 memcpy(&state->refs[i], &state->refs[last_idx],
1684 sizeof(*state->refs));
1685 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
1686 state->acquired_refs--;
1693 static void free_func_state(struct bpf_func_state *state)
1698 kfree(state->stack);
1702 static void clear_jmp_history(struct bpf_verifier_state *state)
1704 kfree(state->jmp_history);
1705 state->jmp_history = NULL;
1706 state->jmp_history_cnt = 0;
1709 static void free_verifier_state(struct bpf_verifier_state *state,
1714 for (i = 0; i <= state->curframe; i++) {
1715 free_func_state(state->frame[i]);
1716 state->frame[i] = NULL;
1718 clear_jmp_history(state);
1723 /* copy verifier state from src to dst growing dst stack space
1724 * when necessary to accommodate larger src stack
1726 static int copy_func_state(struct bpf_func_state *dst,
1727 const struct bpf_func_state *src)
1731 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
1732 err = copy_reference_state(dst, src);
1735 return copy_stack_state(dst, src);
1738 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
1739 const struct bpf_verifier_state *src)
1741 struct bpf_func_state *dst;
1744 dst_state->jmp_history = copy_array(dst_state->jmp_history, src->jmp_history,
1745 src->jmp_history_cnt, sizeof(struct bpf_idx_pair),
1747 if (!dst_state->jmp_history)
1749 dst_state->jmp_history_cnt = src->jmp_history_cnt;
1751 /* if dst has more stack frames then src frame, free them */
1752 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
1753 free_func_state(dst_state->frame[i]);
1754 dst_state->frame[i] = NULL;
1756 dst_state->speculative = src->speculative;
1757 dst_state->active_rcu_lock = src->active_rcu_lock;
1758 dst_state->curframe = src->curframe;
1759 dst_state->active_lock.ptr = src->active_lock.ptr;
1760 dst_state->active_lock.id = src->active_lock.id;
1761 dst_state->branches = src->branches;
1762 dst_state->parent = src->parent;
1763 dst_state->first_insn_idx = src->first_insn_idx;
1764 dst_state->last_insn_idx = src->last_insn_idx;
1765 for (i = 0; i <= src->curframe; i++) {
1766 dst = dst_state->frame[i];
1768 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
1771 dst_state->frame[i] = dst;
1773 err = copy_func_state(dst, src->frame[i]);
1780 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
1783 u32 br = --st->branches;
1785 /* WARN_ON(br > 1) technically makes sense here,
1786 * but see comment in push_stack(), hence:
1788 WARN_ONCE((int)br < 0,
1789 "BUG update_branch_counts:branches_to_explore=%d\n",
1797 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
1798 int *insn_idx, bool pop_log)
1800 struct bpf_verifier_state *cur = env->cur_state;
1801 struct bpf_verifier_stack_elem *elem, *head = env->head;
1804 if (env->head == NULL)
1808 err = copy_verifier_state(cur, &head->st);
1813 bpf_vlog_reset(&env->log, head->log_pos);
1815 *insn_idx = head->insn_idx;
1817 *prev_insn_idx = head->prev_insn_idx;
1819 free_verifier_state(&head->st, false);
1826 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
1827 int insn_idx, int prev_insn_idx,
1830 struct bpf_verifier_state *cur = env->cur_state;
1831 struct bpf_verifier_stack_elem *elem;
1834 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1838 elem->insn_idx = insn_idx;
1839 elem->prev_insn_idx = prev_insn_idx;
1840 elem->next = env->head;
1841 elem->log_pos = env->log.end_pos;
1844 err = copy_verifier_state(&elem->st, cur);
1847 elem->st.speculative |= speculative;
1848 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1849 verbose(env, "The sequence of %d jumps is too complex.\n",
1853 if (elem->st.parent) {
1854 ++elem->st.parent->branches;
1855 /* WARN_ON(branches > 2) technically makes sense here,
1857 * 1. speculative states will bump 'branches' for non-branch
1859 * 2. is_state_visited() heuristics may decide not to create
1860 * a new state for a sequence of branches and all such current
1861 * and cloned states will be pointing to a single parent state
1862 * which might have large 'branches' count.
1867 free_verifier_state(env->cur_state, true);
1868 env->cur_state = NULL;
1869 /* pop all elements and return */
1870 while (!pop_stack(env, NULL, NULL, false));
1874 #define CALLER_SAVED_REGS 6
1875 static const int caller_saved[CALLER_SAVED_REGS] = {
1876 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
1879 /* This helper doesn't clear reg->id */
1880 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1882 reg->var_off = tnum_const(imm);
1883 reg->smin_value = (s64)imm;
1884 reg->smax_value = (s64)imm;
1885 reg->umin_value = imm;
1886 reg->umax_value = imm;
1888 reg->s32_min_value = (s32)imm;
1889 reg->s32_max_value = (s32)imm;
1890 reg->u32_min_value = (u32)imm;
1891 reg->u32_max_value = (u32)imm;
1894 /* Mark the unknown part of a register (variable offset or scalar value) as
1895 * known to have the value @imm.
1897 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1899 /* Clear off and union(map_ptr, range) */
1900 memset(((u8 *)reg) + sizeof(reg->type), 0,
1901 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1903 reg->ref_obj_id = 0;
1904 ___mark_reg_known(reg, imm);
1907 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1909 reg->var_off = tnum_const_subreg(reg->var_off, 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 'variable offset' part of a register as zero. This should be
1917 * used only on registers holding a pointer type.
1919 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1921 __mark_reg_known(reg, 0);
1924 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1926 __mark_reg_known(reg, 0);
1927 reg->type = SCALAR_VALUE;
1930 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1931 struct bpf_reg_state *regs, u32 regno)
1933 if (WARN_ON(regno >= MAX_BPF_REG)) {
1934 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1935 /* Something bad happened, let's kill all regs */
1936 for (regno = 0; regno < MAX_BPF_REG; regno++)
1937 __mark_reg_not_init(env, regs + regno);
1940 __mark_reg_known_zero(regs + regno);
1943 static void __mark_dynptr_reg(struct bpf_reg_state *reg, enum bpf_dynptr_type type,
1944 bool first_slot, int dynptr_id)
1946 /* reg->type has no meaning for STACK_DYNPTR, but when we set reg for
1947 * callback arguments, it does need to be CONST_PTR_TO_DYNPTR, so simply
1948 * set it unconditionally as it is ignored for STACK_DYNPTR anyway.
1950 __mark_reg_known_zero(reg);
1951 reg->type = CONST_PTR_TO_DYNPTR;
1952 /* Give each dynptr a unique id to uniquely associate slices to it. */
1953 reg->id = dynptr_id;
1954 reg->dynptr.type = type;
1955 reg->dynptr.first_slot = first_slot;
1958 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
1960 if (base_type(reg->type) == PTR_TO_MAP_VALUE) {
1961 const struct bpf_map *map = reg->map_ptr;
1963 if (map->inner_map_meta) {
1964 reg->type = CONST_PTR_TO_MAP;
1965 reg->map_ptr = map->inner_map_meta;
1966 /* transfer reg's id which is unique for every map_lookup_elem
1967 * as UID of the inner map.
1969 if (btf_record_has_field(map->inner_map_meta->record, BPF_TIMER))
1970 reg->map_uid = reg->id;
1971 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
1972 reg->type = PTR_TO_XDP_SOCK;
1973 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
1974 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
1975 reg->type = PTR_TO_SOCKET;
1977 reg->type = PTR_TO_MAP_VALUE;
1982 reg->type &= ~PTR_MAYBE_NULL;
1985 static void mark_reg_graph_node(struct bpf_reg_state *regs, u32 regno,
1986 struct btf_field_graph_root *ds_head)
1988 __mark_reg_known_zero(®s[regno]);
1989 regs[regno].type = PTR_TO_BTF_ID | MEM_ALLOC;
1990 regs[regno].btf = ds_head->btf;
1991 regs[regno].btf_id = ds_head->value_btf_id;
1992 regs[regno].off = ds_head->node_offset;
1995 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1997 return type_is_pkt_pointer(reg->type);
2000 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
2002 return reg_is_pkt_pointer(reg) ||
2003 reg->type == PTR_TO_PACKET_END;
2006 static bool reg_is_dynptr_slice_pkt(const struct bpf_reg_state *reg)
2008 return base_type(reg->type) == PTR_TO_MEM &&
2009 (reg->type & DYNPTR_TYPE_SKB || reg->type & DYNPTR_TYPE_XDP);
2012 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
2013 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
2014 enum bpf_reg_type which)
2016 /* The register can already have a range from prior markings.
2017 * This is fine as long as it hasn't been advanced from its
2020 return reg->type == which &&
2023 tnum_equals_const(reg->var_off, 0);
2026 /* Reset the min/max bounds of a register */
2027 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
2029 reg->smin_value = S64_MIN;
2030 reg->smax_value = S64_MAX;
2031 reg->umin_value = 0;
2032 reg->umax_value = U64_MAX;
2034 reg->s32_min_value = S32_MIN;
2035 reg->s32_max_value = S32_MAX;
2036 reg->u32_min_value = 0;
2037 reg->u32_max_value = U32_MAX;
2040 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
2042 reg->smin_value = S64_MIN;
2043 reg->smax_value = S64_MAX;
2044 reg->umin_value = 0;
2045 reg->umax_value = U64_MAX;
2048 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
2050 reg->s32_min_value = S32_MIN;
2051 reg->s32_max_value = S32_MAX;
2052 reg->u32_min_value = 0;
2053 reg->u32_max_value = U32_MAX;
2056 static void __update_reg32_bounds(struct bpf_reg_state *reg)
2058 struct tnum var32_off = tnum_subreg(reg->var_off);
2060 /* min signed is max(sign bit) | min(other bits) */
2061 reg->s32_min_value = max_t(s32, reg->s32_min_value,
2062 var32_off.value | (var32_off.mask & S32_MIN));
2063 /* max signed is min(sign bit) | max(other bits) */
2064 reg->s32_max_value = min_t(s32, reg->s32_max_value,
2065 var32_off.value | (var32_off.mask & S32_MAX));
2066 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
2067 reg->u32_max_value = min(reg->u32_max_value,
2068 (u32)(var32_off.value | var32_off.mask));
2071 static void __update_reg64_bounds(struct bpf_reg_state *reg)
2073 /* min signed is max(sign bit) | min(other bits) */
2074 reg->smin_value = max_t(s64, reg->smin_value,
2075 reg->var_off.value | (reg->var_off.mask & S64_MIN));
2076 /* max signed is min(sign bit) | max(other bits) */
2077 reg->smax_value = min_t(s64, reg->smax_value,
2078 reg->var_off.value | (reg->var_off.mask & S64_MAX));
2079 reg->umin_value = max(reg->umin_value, reg->var_off.value);
2080 reg->umax_value = min(reg->umax_value,
2081 reg->var_off.value | reg->var_off.mask);
2084 static void __update_reg_bounds(struct bpf_reg_state *reg)
2086 __update_reg32_bounds(reg);
2087 __update_reg64_bounds(reg);
2090 /* Uses signed min/max values to inform unsigned, and vice-versa */
2091 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
2093 /* Learn sign from signed bounds.
2094 * If we cannot cross the sign boundary, then signed and unsigned bounds
2095 * are the same, so combine. This works even in the negative case, e.g.
2096 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
2098 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
2099 reg->s32_min_value = reg->u32_min_value =
2100 max_t(u32, reg->s32_min_value, reg->u32_min_value);
2101 reg->s32_max_value = reg->u32_max_value =
2102 min_t(u32, reg->s32_max_value, reg->u32_max_value);
2105 /* Learn sign from unsigned bounds. Signed bounds cross the sign
2106 * boundary, so we must be careful.
2108 if ((s32)reg->u32_max_value >= 0) {
2109 /* Positive. We can't learn anything from the smin, but smax
2110 * is positive, hence safe.
2112 reg->s32_min_value = reg->u32_min_value;
2113 reg->s32_max_value = reg->u32_max_value =
2114 min_t(u32, reg->s32_max_value, reg->u32_max_value);
2115 } else if ((s32)reg->u32_min_value < 0) {
2116 /* Negative. We can't learn anything from the smax, but smin
2117 * is negative, hence safe.
2119 reg->s32_min_value = reg->u32_min_value =
2120 max_t(u32, reg->s32_min_value, reg->u32_min_value);
2121 reg->s32_max_value = reg->u32_max_value;
2125 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
2127 /* Learn sign from signed bounds.
2128 * If we cannot cross the sign boundary, then signed and unsigned bounds
2129 * are the same, so combine. This works even in the negative case, e.g.
2130 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
2132 if (reg->smin_value >= 0 || reg->smax_value < 0) {
2133 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
2135 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
2139 /* Learn sign from unsigned bounds. Signed bounds cross the sign
2140 * boundary, so we must be careful.
2142 if ((s64)reg->umax_value >= 0) {
2143 /* Positive. We can't learn anything from the smin, but smax
2144 * is positive, hence safe.
2146 reg->smin_value = reg->umin_value;
2147 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
2149 } else if ((s64)reg->umin_value < 0) {
2150 /* Negative. We can't learn anything from the smax, but smin
2151 * is negative, hence safe.
2153 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
2155 reg->smax_value = reg->umax_value;
2159 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
2161 __reg32_deduce_bounds(reg);
2162 __reg64_deduce_bounds(reg);
2165 /* Attempts to improve var_off based on unsigned min/max information */
2166 static void __reg_bound_offset(struct bpf_reg_state *reg)
2168 struct tnum var64_off = tnum_intersect(reg->var_off,
2169 tnum_range(reg->umin_value,
2171 struct tnum var32_off = tnum_intersect(tnum_subreg(var64_off),
2172 tnum_range(reg->u32_min_value,
2173 reg->u32_max_value));
2175 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
2178 static void reg_bounds_sync(struct bpf_reg_state *reg)
2180 /* We might have learned new bounds from the var_off. */
2181 __update_reg_bounds(reg);
2182 /* We might have learned something about the sign bit. */
2183 __reg_deduce_bounds(reg);
2184 /* We might have learned some bits from the bounds. */
2185 __reg_bound_offset(reg);
2186 /* Intersecting with the old var_off might have improved our bounds
2187 * slightly, e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2188 * then new var_off is (0; 0x7f...fc) which improves our umax.
2190 __update_reg_bounds(reg);
2193 static bool __reg32_bound_s64(s32 a)
2195 return a >= 0 && a <= S32_MAX;
2198 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
2200 reg->umin_value = reg->u32_min_value;
2201 reg->umax_value = reg->u32_max_value;
2203 /* Attempt to pull 32-bit signed bounds into 64-bit bounds but must
2204 * be positive otherwise set to worse case bounds and refine later
2207 if (__reg32_bound_s64(reg->s32_min_value) &&
2208 __reg32_bound_s64(reg->s32_max_value)) {
2209 reg->smin_value = reg->s32_min_value;
2210 reg->smax_value = reg->s32_max_value;
2212 reg->smin_value = 0;
2213 reg->smax_value = U32_MAX;
2217 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
2219 /* special case when 64-bit register has upper 32-bit register
2220 * zeroed. Typically happens after zext or <<32, >>32 sequence
2221 * allowing us to use 32-bit bounds directly,
2223 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
2224 __reg_assign_32_into_64(reg);
2226 /* Otherwise the best we can do is push lower 32bit known and
2227 * unknown bits into register (var_off set from jmp logic)
2228 * then learn as much as possible from the 64-bit tnum
2229 * known and unknown bits. The previous smin/smax bounds are
2230 * invalid here because of jmp32 compare so mark them unknown
2231 * so they do not impact tnum bounds calculation.
2233 __mark_reg64_unbounded(reg);
2235 reg_bounds_sync(reg);
2238 static bool __reg64_bound_s32(s64 a)
2240 return a >= S32_MIN && a <= S32_MAX;
2243 static bool __reg64_bound_u32(u64 a)
2245 return a >= U32_MIN && a <= U32_MAX;
2248 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
2250 __mark_reg32_unbounded(reg);
2251 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
2252 reg->s32_min_value = (s32)reg->smin_value;
2253 reg->s32_max_value = (s32)reg->smax_value;
2255 if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
2256 reg->u32_min_value = (u32)reg->umin_value;
2257 reg->u32_max_value = (u32)reg->umax_value;
2259 reg_bounds_sync(reg);
2262 /* Mark a register as having a completely unknown (scalar) value. */
2263 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
2264 struct bpf_reg_state *reg)
2267 * Clear type, off, and union(map_ptr, range) and
2268 * padding between 'type' and union
2270 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
2271 reg->type = SCALAR_VALUE;
2273 reg->ref_obj_id = 0;
2274 reg->var_off = tnum_unknown;
2276 reg->precise = !env->bpf_capable;
2277 __mark_reg_unbounded(reg);
2280 static void mark_reg_unknown(struct bpf_verifier_env *env,
2281 struct bpf_reg_state *regs, u32 regno)
2283 if (WARN_ON(regno >= MAX_BPF_REG)) {
2284 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
2285 /* Something bad happened, let's kill all regs except FP */
2286 for (regno = 0; regno < BPF_REG_FP; regno++)
2287 __mark_reg_not_init(env, regs + regno);
2290 __mark_reg_unknown(env, regs + regno);
2293 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
2294 struct bpf_reg_state *reg)
2296 __mark_reg_unknown(env, reg);
2297 reg->type = NOT_INIT;
2300 static void mark_reg_not_init(struct bpf_verifier_env *env,
2301 struct bpf_reg_state *regs, u32 regno)
2303 if (WARN_ON(regno >= MAX_BPF_REG)) {
2304 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
2305 /* Something bad happened, let's kill all regs except FP */
2306 for (regno = 0; regno < BPF_REG_FP; regno++)
2307 __mark_reg_not_init(env, regs + regno);
2310 __mark_reg_not_init(env, regs + regno);
2313 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
2314 struct bpf_reg_state *regs, u32 regno,
2315 enum bpf_reg_type reg_type,
2316 struct btf *btf, u32 btf_id,
2317 enum bpf_type_flag flag)
2319 if (reg_type == SCALAR_VALUE) {
2320 mark_reg_unknown(env, regs, regno);
2323 mark_reg_known_zero(env, regs, regno);
2324 regs[regno].type = PTR_TO_BTF_ID | flag;
2325 regs[regno].btf = btf;
2326 regs[regno].btf_id = btf_id;
2329 #define DEF_NOT_SUBREG (0)
2330 static void init_reg_state(struct bpf_verifier_env *env,
2331 struct bpf_func_state *state)
2333 struct bpf_reg_state *regs = state->regs;
2336 for (i = 0; i < MAX_BPF_REG; i++) {
2337 mark_reg_not_init(env, regs, i);
2338 regs[i].live = REG_LIVE_NONE;
2339 regs[i].parent = NULL;
2340 regs[i].subreg_def = DEF_NOT_SUBREG;
2344 regs[BPF_REG_FP].type = PTR_TO_STACK;
2345 mark_reg_known_zero(env, regs, BPF_REG_FP);
2346 regs[BPF_REG_FP].frameno = state->frameno;
2349 #define BPF_MAIN_FUNC (-1)
2350 static void init_func_state(struct bpf_verifier_env *env,
2351 struct bpf_func_state *state,
2352 int callsite, int frameno, int subprogno)
2354 state->callsite = callsite;
2355 state->frameno = frameno;
2356 state->subprogno = subprogno;
2357 state->callback_ret_range = tnum_range(0, 0);
2358 init_reg_state(env, state);
2359 mark_verifier_state_scratched(env);
2362 /* Similar to push_stack(), but for async callbacks */
2363 static struct bpf_verifier_state *push_async_cb(struct bpf_verifier_env *env,
2364 int insn_idx, int prev_insn_idx,
2367 struct bpf_verifier_stack_elem *elem;
2368 struct bpf_func_state *frame;
2370 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
2374 elem->insn_idx = insn_idx;
2375 elem->prev_insn_idx = prev_insn_idx;
2376 elem->next = env->head;
2377 elem->log_pos = env->log.end_pos;
2380 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
2382 "The sequence of %d jumps is too complex for async cb.\n",
2386 /* Unlike push_stack() do not copy_verifier_state().
2387 * The caller state doesn't matter.
2388 * This is async callback. It starts in a fresh stack.
2389 * Initialize it similar to do_check_common().
2391 elem->st.branches = 1;
2392 frame = kzalloc(sizeof(*frame), GFP_KERNEL);
2395 init_func_state(env, frame,
2396 BPF_MAIN_FUNC /* callsite */,
2397 0 /* frameno within this callchain */,
2398 subprog /* subprog number within this prog */);
2399 elem->st.frame[0] = frame;
2402 free_verifier_state(env->cur_state, true);
2403 env->cur_state = NULL;
2404 /* pop all elements and return */
2405 while (!pop_stack(env, NULL, NULL, false));
2411 SRC_OP, /* register is used as source operand */
2412 DST_OP, /* register is used as destination operand */
2413 DST_OP_NO_MARK /* same as above, check only, don't mark */
2416 static int cmp_subprogs(const void *a, const void *b)
2418 return ((struct bpf_subprog_info *)a)->start -
2419 ((struct bpf_subprog_info *)b)->start;
2422 static int find_subprog(struct bpf_verifier_env *env, int off)
2424 struct bpf_subprog_info *p;
2426 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
2427 sizeof(env->subprog_info[0]), cmp_subprogs);
2430 return p - env->subprog_info;
2434 static int add_subprog(struct bpf_verifier_env *env, int off)
2436 int insn_cnt = env->prog->len;
2439 if (off >= insn_cnt || off < 0) {
2440 verbose(env, "call to invalid destination\n");
2443 ret = find_subprog(env, off);
2446 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
2447 verbose(env, "too many subprograms\n");
2450 /* determine subprog starts. The end is one before the next starts */
2451 env->subprog_info[env->subprog_cnt++].start = off;
2452 sort(env->subprog_info, env->subprog_cnt,
2453 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
2454 return env->subprog_cnt - 1;
2457 #define MAX_KFUNC_DESCS 256
2458 #define MAX_KFUNC_BTFS 256
2460 struct bpf_kfunc_desc {
2461 struct btf_func_model func_model;
2468 struct bpf_kfunc_btf {
2470 struct module *module;
2474 struct bpf_kfunc_desc_tab {
2475 /* Sorted by func_id (BTF ID) and offset (fd_array offset) during
2476 * verification. JITs do lookups by bpf_insn, where func_id may not be
2477 * available, therefore at the end of verification do_misc_fixups()
2478 * sorts this by imm and offset.
2480 struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
2484 struct bpf_kfunc_btf_tab {
2485 struct bpf_kfunc_btf descs[MAX_KFUNC_BTFS];
2489 static int kfunc_desc_cmp_by_id_off(const void *a, const void *b)
2491 const struct bpf_kfunc_desc *d0 = a;
2492 const struct bpf_kfunc_desc *d1 = b;
2494 /* func_id is not greater than BTF_MAX_TYPE */
2495 return d0->func_id - d1->func_id ?: d0->offset - d1->offset;
2498 static int kfunc_btf_cmp_by_off(const void *a, const void *b)
2500 const struct bpf_kfunc_btf *d0 = a;
2501 const struct bpf_kfunc_btf *d1 = b;
2503 return d0->offset - d1->offset;
2506 static const struct bpf_kfunc_desc *
2507 find_kfunc_desc(const struct bpf_prog *prog, u32 func_id, u16 offset)
2509 struct bpf_kfunc_desc desc = {
2513 struct bpf_kfunc_desc_tab *tab;
2515 tab = prog->aux->kfunc_tab;
2516 return bsearch(&desc, tab->descs, tab->nr_descs,
2517 sizeof(tab->descs[0]), kfunc_desc_cmp_by_id_off);
2520 int bpf_get_kfunc_addr(const struct bpf_prog *prog, u32 func_id,
2521 u16 btf_fd_idx, u8 **func_addr)
2523 const struct bpf_kfunc_desc *desc;
2525 desc = find_kfunc_desc(prog, func_id, btf_fd_idx);
2529 *func_addr = (u8 *)desc->addr;
2533 static struct btf *__find_kfunc_desc_btf(struct bpf_verifier_env *env,
2536 struct bpf_kfunc_btf kf_btf = { .offset = offset };
2537 struct bpf_kfunc_btf_tab *tab;
2538 struct bpf_kfunc_btf *b;
2543 tab = env->prog->aux->kfunc_btf_tab;
2544 b = bsearch(&kf_btf, tab->descs, tab->nr_descs,
2545 sizeof(tab->descs[0]), kfunc_btf_cmp_by_off);
2547 if (tab->nr_descs == MAX_KFUNC_BTFS) {
2548 verbose(env, "too many different module BTFs\n");
2549 return ERR_PTR(-E2BIG);
2552 if (bpfptr_is_null(env->fd_array)) {
2553 verbose(env, "kfunc offset > 0 without fd_array is invalid\n");
2554 return ERR_PTR(-EPROTO);
2557 if (copy_from_bpfptr_offset(&btf_fd, env->fd_array,
2558 offset * sizeof(btf_fd),
2560 return ERR_PTR(-EFAULT);
2562 btf = btf_get_by_fd(btf_fd);
2564 verbose(env, "invalid module BTF fd specified\n");
2568 if (!btf_is_module(btf)) {
2569 verbose(env, "BTF fd for kfunc is not a module BTF\n");
2571 return ERR_PTR(-EINVAL);
2574 mod = btf_try_get_module(btf);
2577 return ERR_PTR(-ENXIO);
2580 b = &tab->descs[tab->nr_descs++];
2585 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2586 kfunc_btf_cmp_by_off, NULL);
2591 void bpf_free_kfunc_btf_tab(struct bpf_kfunc_btf_tab *tab)
2596 while (tab->nr_descs--) {
2597 module_put(tab->descs[tab->nr_descs].module);
2598 btf_put(tab->descs[tab->nr_descs].btf);
2603 static struct btf *find_kfunc_desc_btf(struct bpf_verifier_env *env, s16 offset)
2607 /* In the future, this can be allowed to increase limit
2608 * of fd index into fd_array, interpreted as u16.
2610 verbose(env, "negative offset disallowed for kernel module function call\n");
2611 return ERR_PTR(-EINVAL);
2614 return __find_kfunc_desc_btf(env, offset);
2616 return btf_vmlinux ?: ERR_PTR(-ENOENT);
2619 static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id, s16 offset)
2621 const struct btf_type *func, *func_proto;
2622 struct bpf_kfunc_btf_tab *btf_tab;
2623 struct bpf_kfunc_desc_tab *tab;
2624 struct bpf_prog_aux *prog_aux;
2625 struct bpf_kfunc_desc *desc;
2626 const char *func_name;
2627 struct btf *desc_btf;
2628 unsigned long call_imm;
2632 prog_aux = env->prog->aux;
2633 tab = prog_aux->kfunc_tab;
2634 btf_tab = prog_aux->kfunc_btf_tab;
2637 verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
2641 if (!env->prog->jit_requested) {
2642 verbose(env, "JIT is required for calling kernel function\n");
2646 if (!bpf_jit_supports_kfunc_call()) {
2647 verbose(env, "JIT does not support calling kernel function\n");
2651 if (!env->prog->gpl_compatible) {
2652 verbose(env, "cannot call kernel function from non-GPL compatible program\n");
2656 tab = kzalloc(sizeof(*tab), GFP_KERNEL);
2659 prog_aux->kfunc_tab = tab;
2662 /* func_id == 0 is always invalid, but instead of returning an error, be
2663 * conservative and wait until the code elimination pass before returning
2664 * error, so that invalid calls that get pruned out can be in BPF programs
2665 * loaded from userspace. It is also required that offset be untouched
2668 if (!func_id && !offset)
2671 if (!btf_tab && offset) {
2672 btf_tab = kzalloc(sizeof(*btf_tab), GFP_KERNEL);
2675 prog_aux->kfunc_btf_tab = btf_tab;
2678 desc_btf = find_kfunc_desc_btf(env, offset);
2679 if (IS_ERR(desc_btf)) {
2680 verbose(env, "failed to find BTF for kernel function\n");
2681 return PTR_ERR(desc_btf);
2684 if (find_kfunc_desc(env->prog, func_id, offset))
2687 if (tab->nr_descs == MAX_KFUNC_DESCS) {
2688 verbose(env, "too many different kernel function calls\n");
2692 func = btf_type_by_id(desc_btf, func_id);
2693 if (!func || !btf_type_is_func(func)) {
2694 verbose(env, "kernel btf_id %u is not a function\n",
2698 func_proto = btf_type_by_id(desc_btf, func->type);
2699 if (!func_proto || !btf_type_is_func_proto(func_proto)) {
2700 verbose(env, "kernel function btf_id %u does not have a valid func_proto\n",
2705 func_name = btf_name_by_offset(desc_btf, func->name_off);
2706 addr = kallsyms_lookup_name(func_name);
2708 verbose(env, "cannot find address for kernel function %s\n",
2712 specialize_kfunc(env, func_id, offset, &addr);
2714 if (bpf_jit_supports_far_kfunc_call()) {
2717 call_imm = BPF_CALL_IMM(addr);
2718 /* Check whether the relative offset overflows desc->imm */
2719 if ((unsigned long)(s32)call_imm != call_imm) {
2720 verbose(env, "address of kernel function %s is out of range\n",
2726 if (bpf_dev_bound_kfunc_id(func_id)) {
2727 err = bpf_dev_bound_kfunc_check(&env->log, prog_aux);
2732 desc = &tab->descs[tab->nr_descs++];
2733 desc->func_id = func_id;
2734 desc->imm = call_imm;
2735 desc->offset = offset;
2737 err = btf_distill_func_proto(&env->log, desc_btf,
2738 func_proto, func_name,
2741 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2742 kfunc_desc_cmp_by_id_off, NULL);
2746 static int kfunc_desc_cmp_by_imm_off(const void *a, const void *b)
2748 const struct bpf_kfunc_desc *d0 = a;
2749 const struct bpf_kfunc_desc *d1 = b;
2751 if (d0->imm != d1->imm)
2752 return d0->imm < d1->imm ? -1 : 1;
2753 if (d0->offset != d1->offset)
2754 return d0->offset < d1->offset ? -1 : 1;
2758 static void sort_kfunc_descs_by_imm_off(struct bpf_prog *prog)
2760 struct bpf_kfunc_desc_tab *tab;
2762 tab = prog->aux->kfunc_tab;
2766 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2767 kfunc_desc_cmp_by_imm_off, NULL);
2770 bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
2772 return !!prog->aux->kfunc_tab;
2775 const struct btf_func_model *
2776 bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
2777 const struct bpf_insn *insn)
2779 const struct bpf_kfunc_desc desc = {
2781 .offset = insn->off,
2783 const struct bpf_kfunc_desc *res;
2784 struct bpf_kfunc_desc_tab *tab;
2786 tab = prog->aux->kfunc_tab;
2787 res = bsearch(&desc, tab->descs, tab->nr_descs,
2788 sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm_off);
2790 return res ? &res->func_model : NULL;
2793 static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
2795 struct bpf_subprog_info *subprog = env->subprog_info;
2796 struct bpf_insn *insn = env->prog->insnsi;
2797 int i, ret, insn_cnt = env->prog->len;
2799 /* Add entry function. */
2800 ret = add_subprog(env, 0);
2804 for (i = 0; i < insn_cnt; i++, insn++) {
2805 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
2806 !bpf_pseudo_kfunc_call(insn))
2809 if (!env->bpf_capable) {
2810 verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
2814 if (bpf_pseudo_func(insn) || bpf_pseudo_call(insn))
2815 ret = add_subprog(env, i + insn->imm + 1);
2817 ret = add_kfunc_call(env, insn->imm, insn->off);
2823 /* Add a fake 'exit' subprog which could simplify subprog iteration
2824 * logic. 'subprog_cnt' should not be increased.
2826 subprog[env->subprog_cnt].start = insn_cnt;
2828 if (env->log.level & BPF_LOG_LEVEL2)
2829 for (i = 0; i < env->subprog_cnt; i++)
2830 verbose(env, "func#%d @%d\n", i, subprog[i].start);
2835 static int check_subprogs(struct bpf_verifier_env *env)
2837 int i, subprog_start, subprog_end, off, cur_subprog = 0;
2838 struct bpf_subprog_info *subprog = env->subprog_info;
2839 struct bpf_insn *insn = env->prog->insnsi;
2840 int insn_cnt = env->prog->len;
2842 /* now check that all jumps are within the same subprog */
2843 subprog_start = subprog[cur_subprog].start;
2844 subprog_end = subprog[cur_subprog + 1].start;
2845 for (i = 0; i < insn_cnt; i++) {
2846 u8 code = insn[i].code;
2848 if (code == (BPF_JMP | BPF_CALL) &&
2849 insn[i].src_reg == 0 &&
2850 insn[i].imm == BPF_FUNC_tail_call)
2851 subprog[cur_subprog].has_tail_call = true;
2852 if (BPF_CLASS(code) == BPF_LD &&
2853 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
2854 subprog[cur_subprog].has_ld_abs = true;
2855 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
2857 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
2859 if (code == (BPF_JMP32 | BPF_JA))
2860 off = i + insn[i].imm + 1;
2862 off = i + insn[i].off + 1;
2863 if (off < subprog_start || off >= subprog_end) {
2864 verbose(env, "jump out of range from insn %d to %d\n", i, off);
2868 if (i == subprog_end - 1) {
2869 /* to avoid fall-through from one subprog into another
2870 * the last insn of the subprog should be either exit
2871 * or unconditional jump back
2873 if (code != (BPF_JMP | BPF_EXIT) &&
2874 code != (BPF_JMP32 | BPF_JA) &&
2875 code != (BPF_JMP | BPF_JA)) {
2876 verbose(env, "last insn is not an exit or jmp\n");
2879 subprog_start = subprog_end;
2881 if (cur_subprog < env->subprog_cnt)
2882 subprog_end = subprog[cur_subprog + 1].start;
2888 /* Parentage chain of this register (or stack slot) should take care of all
2889 * issues like callee-saved registers, stack slot allocation time, etc.
2891 static int mark_reg_read(struct bpf_verifier_env *env,
2892 const struct bpf_reg_state *state,
2893 struct bpf_reg_state *parent, u8 flag)
2895 bool writes = parent == state->parent; /* Observe write marks */
2899 /* if read wasn't screened by an earlier write ... */
2900 if (writes && state->live & REG_LIVE_WRITTEN)
2902 if (parent->live & REG_LIVE_DONE) {
2903 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
2904 reg_type_str(env, parent->type),
2905 parent->var_off.value, parent->off);
2908 /* The first condition is more likely to be true than the
2909 * second, checked it first.
2911 if ((parent->live & REG_LIVE_READ) == flag ||
2912 parent->live & REG_LIVE_READ64)
2913 /* The parentage chain never changes and
2914 * this parent was already marked as LIVE_READ.
2915 * There is no need to keep walking the chain again and
2916 * keep re-marking all parents as LIVE_READ.
2917 * This case happens when the same register is read
2918 * multiple times without writes into it in-between.
2919 * Also, if parent has the stronger REG_LIVE_READ64 set,
2920 * then no need to set the weak REG_LIVE_READ32.
2923 /* ... then we depend on parent's value */
2924 parent->live |= flag;
2925 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
2926 if (flag == REG_LIVE_READ64)
2927 parent->live &= ~REG_LIVE_READ32;
2929 parent = state->parent;
2934 if (env->longest_mark_read_walk < cnt)
2935 env->longest_mark_read_walk = cnt;
2939 static int mark_dynptr_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
2941 struct bpf_func_state *state = func(env, reg);
2944 /* For CONST_PTR_TO_DYNPTR, it must have already been done by
2945 * check_reg_arg in check_helper_call and mark_btf_func_reg_size in
2948 if (reg->type == CONST_PTR_TO_DYNPTR)
2950 spi = dynptr_get_spi(env, reg);
2953 /* Caller ensures dynptr is valid and initialized, which means spi is in
2954 * bounds and spi is the first dynptr slot. Simply mark stack slot as
2957 ret = mark_reg_read(env, &state->stack[spi].spilled_ptr,
2958 state->stack[spi].spilled_ptr.parent, REG_LIVE_READ64);
2961 return mark_reg_read(env, &state->stack[spi - 1].spilled_ptr,
2962 state->stack[spi - 1].spilled_ptr.parent, REG_LIVE_READ64);
2965 static int mark_iter_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
2966 int spi, int nr_slots)
2968 struct bpf_func_state *state = func(env, reg);
2971 for (i = 0; i < nr_slots; i++) {
2972 struct bpf_reg_state *st = &state->stack[spi - i].spilled_ptr;
2974 err = mark_reg_read(env, st, st->parent, REG_LIVE_READ64);
2978 mark_stack_slot_scratched(env, spi - i);
2984 /* This function is supposed to be used by the following 32-bit optimization
2985 * code only. It returns TRUE if the source or destination register operates
2986 * on 64-bit, otherwise return FALSE.
2988 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
2989 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
2994 class = BPF_CLASS(code);
2996 if (class == BPF_JMP) {
2997 /* BPF_EXIT for "main" will reach here. Return TRUE
3002 if (op == BPF_CALL) {
3003 /* BPF to BPF call will reach here because of marking
3004 * caller saved clobber with DST_OP_NO_MARK for which we
3005 * don't care the register def because they are anyway
3006 * marked as NOT_INIT already.
3008 if (insn->src_reg == BPF_PSEUDO_CALL)
3010 /* Helper call will reach here because of arg type
3011 * check, conservatively return TRUE.
3020 if (class == BPF_ALU64 && op == BPF_END && (insn->imm == 16 || insn->imm == 32))
3023 if (class == BPF_ALU64 || class == BPF_JMP ||
3024 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
3027 if (class == BPF_ALU || class == BPF_JMP32)
3030 if (class == BPF_LDX) {
3032 return BPF_SIZE(code) == BPF_DW;
3033 /* LDX source must be ptr. */
3037 if (class == BPF_STX) {
3038 /* BPF_STX (including atomic variants) has multiple source
3039 * operands, one of which is a ptr. Check whether the caller is
3042 if (t == SRC_OP && reg->type != SCALAR_VALUE)
3044 return BPF_SIZE(code) == BPF_DW;
3047 if (class == BPF_LD) {
3048 u8 mode = BPF_MODE(code);
3051 if (mode == BPF_IMM)
3054 /* Both LD_IND and LD_ABS return 32-bit data. */
3058 /* Implicit ctx ptr. */
3059 if (regno == BPF_REG_6)
3062 /* Explicit source could be any width. */
3066 if (class == BPF_ST)
3067 /* The only source register for BPF_ST is a ptr. */
3070 /* Conservatively return true at default. */
3074 /* Return the regno defined by the insn, or -1. */
3075 static int insn_def_regno(const struct bpf_insn *insn)
3077 switch (BPF_CLASS(insn->code)) {
3083 if (BPF_MODE(insn->code) == BPF_ATOMIC &&
3084 (insn->imm & BPF_FETCH)) {
3085 if (insn->imm == BPF_CMPXCHG)
3088 return insn->src_reg;
3093 return insn->dst_reg;
3097 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
3098 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
3100 int dst_reg = insn_def_regno(insn);
3105 return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
3108 static void mark_insn_zext(struct bpf_verifier_env *env,
3109 struct bpf_reg_state *reg)
3111 s32 def_idx = reg->subreg_def;
3113 if (def_idx == DEF_NOT_SUBREG)
3116 env->insn_aux_data[def_idx - 1].zext_dst = true;
3117 /* The dst will be zero extended, so won't be sub-register anymore. */
3118 reg->subreg_def = DEF_NOT_SUBREG;
3121 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
3122 enum reg_arg_type t)
3124 struct bpf_verifier_state *vstate = env->cur_state;
3125 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3126 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
3127 struct bpf_reg_state *reg, *regs = state->regs;
3130 if (regno >= MAX_BPF_REG) {
3131 verbose(env, "R%d is invalid\n", regno);
3135 mark_reg_scratched(env, regno);
3138 rw64 = is_reg64(env, insn, regno, reg, t);
3140 /* check whether register used as source operand can be read */
3141 if (reg->type == NOT_INIT) {
3142 verbose(env, "R%d !read_ok\n", regno);
3145 /* We don't need to worry about FP liveness because it's read-only */
3146 if (regno == BPF_REG_FP)
3150 mark_insn_zext(env, reg);
3152 return mark_reg_read(env, reg, reg->parent,
3153 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
3155 /* check whether register used as dest operand can be written to */
3156 if (regno == BPF_REG_FP) {
3157 verbose(env, "frame pointer is read only\n");
3160 reg->live |= REG_LIVE_WRITTEN;
3161 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
3163 mark_reg_unknown(env, regs, regno);
3168 static void mark_jmp_point(struct bpf_verifier_env *env, int idx)
3170 env->insn_aux_data[idx].jmp_point = true;
3173 static bool is_jmp_point(struct bpf_verifier_env *env, int insn_idx)
3175 return env->insn_aux_data[insn_idx].jmp_point;
3178 /* for any branch, call, exit record the history of jmps in the given state */
3179 static int push_jmp_history(struct bpf_verifier_env *env,
3180 struct bpf_verifier_state *cur)
3182 u32 cnt = cur->jmp_history_cnt;
3183 struct bpf_idx_pair *p;
3186 if (!is_jmp_point(env, env->insn_idx))
3190 alloc_size = kmalloc_size_roundup(size_mul(cnt, sizeof(*p)));
3191 p = krealloc(cur->jmp_history, alloc_size, GFP_USER);
3194 p[cnt - 1].idx = env->insn_idx;
3195 p[cnt - 1].prev_idx = env->prev_insn_idx;
3196 cur->jmp_history = p;
3197 cur->jmp_history_cnt = cnt;
3201 /* Backtrack one insn at a time. If idx is not at the top of recorded
3202 * history then previous instruction came from straight line execution.
3204 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
3209 if (cnt && st->jmp_history[cnt - 1].idx == i) {
3210 i = st->jmp_history[cnt - 1].prev_idx;
3218 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
3220 const struct btf_type *func;
3221 struct btf *desc_btf;
3223 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
3226 desc_btf = find_kfunc_desc_btf(data, insn->off);
3227 if (IS_ERR(desc_btf))
3230 func = btf_type_by_id(desc_btf, insn->imm);
3231 return btf_name_by_offset(desc_btf, func->name_off);
3234 static inline void bt_init(struct backtrack_state *bt, u32 frame)
3239 static inline void bt_reset(struct backtrack_state *bt)
3241 struct bpf_verifier_env *env = bt->env;
3243 memset(bt, 0, sizeof(*bt));
3247 static inline u32 bt_empty(struct backtrack_state *bt)
3252 for (i = 0; i <= bt->frame; i++)
3253 mask |= bt->reg_masks[i] | bt->stack_masks[i];
3258 static inline int bt_subprog_enter(struct backtrack_state *bt)
3260 if (bt->frame == MAX_CALL_FRAMES - 1) {
3261 verbose(bt->env, "BUG subprog enter from frame %d\n", bt->frame);
3262 WARN_ONCE(1, "verifier backtracking bug");
3269 static inline int bt_subprog_exit(struct backtrack_state *bt)
3271 if (bt->frame == 0) {
3272 verbose(bt->env, "BUG subprog exit from frame 0\n");
3273 WARN_ONCE(1, "verifier backtracking bug");
3280 static inline void bt_set_frame_reg(struct backtrack_state *bt, u32 frame, u32 reg)
3282 bt->reg_masks[frame] |= 1 << reg;
3285 static inline void bt_clear_frame_reg(struct backtrack_state *bt, u32 frame, u32 reg)
3287 bt->reg_masks[frame] &= ~(1 << reg);
3290 static inline void bt_set_reg(struct backtrack_state *bt, u32 reg)
3292 bt_set_frame_reg(bt, bt->frame, reg);
3295 static inline void bt_clear_reg(struct backtrack_state *bt, u32 reg)
3297 bt_clear_frame_reg(bt, bt->frame, reg);
3300 static inline void bt_set_frame_slot(struct backtrack_state *bt, u32 frame, u32 slot)
3302 bt->stack_masks[frame] |= 1ull << slot;
3305 static inline void bt_clear_frame_slot(struct backtrack_state *bt, u32 frame, u32 slot)
3307 bt->stack_masks[frame] &= ~(1ull << slot);
3310 static inline void bt_set_slot(struct backtrack_state *bt, u32 slot)
3312 bt_set_frame_slot(bt, bt->frame, slot);
3315 static inline void bt_clear_slot(struct backtrack_state *bt, u32 slot)
3317 bt_clear_frame_slot(bt, bt->frame, slot);
3320 static inline u32 bt_frame_reg_mask(struct backtrack_state *bt, u32 frame)
3322 return bt->reg_masks[frame];
3325 static inline u32 bt_reg_mask(struct backtrack_state *bt)
3327 return bt->reg_masks[bt->frame];
3330 static inline u64 bt_frame_stack_mask(struct backtrack_state *bt, u32 frame)
3332 return bt->stack_masks[frame];
3335 static inline u64 bt_stack_mask(struct backtrack_state *bt)
3337 return bt->stack_masks[bt->frame];
3340 static inline bool bt_is_reg_set(struct backtrack_state *bt, u32 reg)
3342 return bt->reg_masks[bt->frame] & (1 << reg);
3345 static inline bool bt_is_slot_set(struct backtrack_state *bt, u32 slot)
3347 return bt->stack_masks[bt->frame] & (1ull << slot);
3350 /* format registers bitmask, e.g., "r0,r2,r4" for 0x15 mask */
3351 static void fmt_reg_mask(char *buf, ssize_t buf_sz, u32 reg_mask)
3353 DECLARE_BITMAP(mask, 64);
3359 bitmap_from_u64(mask, reg_mask);
3360 for_each_set_bit(i, mask, 32) {
3361 n = snprintf(buf, buf_sz, "%sr%d", first ? "" : ",", i);
3369 /* format stack slots bitmask, e.g., "-8,-24,-40" for 0x15 mask */
3370 static void fmt_stack_mask(char *buf, ssize_t buf_sz, u64 stack_mask)
3372 DECLARE_BITMAP(mask, 64);
3378 bitmap_from_u64(mask, stack_mask);
3379 for_each_set_bit(i, mask, 64) {
3380 n = snprintf(buf, buf_sz, "%s%d", first ? "" : ",", -(i + 1) * 8);
3389 /* For given verifier state backtrack_insn() is called from the last insn to
3390 * the first insn. Its purpose is to compute a bitmask of registers and
3391 * stack slots that needs precision in the parent verifier state.
3393 * @idx is an index of the instruction we are currently processing;
3394 * @subseq_idx is an index of the subsequent instruction that:
3395 * - *would be* executed next, if jump history is viewed in forward order;
3396 * - *was* processed previously during backtracking.
3398 static int backtrack_insn(struct bpf_verifier_env *env, int idx, int subseq_idx,
3399 struct backtrack_state *bt)
3401 const struct bpf_insn_cbs cbs = {
3402 .cb_call = disasm_kfunc_name,
3403 .cb_print = verbose,
3404 .private_data = env,
3406 struct bpf_insn *insn = env->prog->insnsi + idx;
3407 u8 class = BPF_CLASS(insn->code);
3408 u8 opcode = BPF_OP(insn->code);
3409 u8 mode = BPF_MODE(insn->code);
3410 u32 dreg = insn->dst_reg;
3411 u32 sreg = insn->src_reg;
3414 if (insn->code == 0)
3416 if (env->log.level & BPF_LOG_LEVEL2) {
3417 fmt_reg_mask(env->tmp_str_buf, TMP_STR_BUF_LEN, bt_reg_mask(bt));
3418 verbose(env, "mark_precise: frame%d: regs=%s ",
3419 bt->frame, env->tmp_str_buf);
3420 fmt_stack_mask(env->tmp_str_buf, TMP_STR_BUF_LEN, bt_stack_mask(bt));
3421 verbose(env, "stack=%s before ", env->tmp_str_buf);
3422 verbose(env, "%d: ", idx);
3423 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
3426 if (class == BPF_ALU || class == BPF_ALU64) {
3427 if (!bt_is_reg_set(bt, dreg))
3429 if (opcode == BPF_MOV) {
3430 if (BPF_SRC(insn->code) == BPF_X) {
3431 /* dreg = sreg or dreg = (s8, s16, s32)sreg
3432 * dreg needs precision after this insn
3433 * sreg needs precision before this insn
3435 bt_clear_reg(bt, dreg);
3436 bt_set_reg(bt, sreg);
3439 * dreg needs precision after this insn.
3440 * Corresponding register is already marked
3441 * as precise=true in this verifier state.
3442 * No further markings in parent are necessary
3444 bt_clear_reg(bt, dreg);
3447 if (BPF_SRC(insn->code) == BPF_X) {
3449 * both dreg and sreg need precision
3452 bt_set_reg(bt, sreg);
3454 * dreg still needs precision before this insn
3457 } else if (class == BPF_LDX) {
3458 if (!bt_is_reg_set(bt, dreg))
3460 bt_clear_reg(bt, dreg);
3462 /* scalars can only be spilled into stack w/o losing precision.
3463 * Load from any other memory can be zero extended.
3464 * The desire to keep that precision is already indicated
3465 * by 'precise' mark in corresponding register of this state.
3466 * No further tracking necessary.
3468 if (insn->src_reg != BPF_REG_FP)
3471 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
3472 * that [fp - off] slot contains scalar that needs to be
3473 * tracked with precision
3475 spi = (-insn->off - 1) / BPF_REG_SIZE;
3477 verbose(env, "BUG spi %d\n", spi);
3478 WARN_ONCE(1, "verifier backtracking bug");
3481 bt_set_slot(bt, spi);
3482 } else if (class == BPF_STX || class == BPF_ST) {
3483 if (bt_is_reg_set(bt, dreg))
3484 /* stx & st shouldn't be using _scalar_ dst_reg
3485 * to access memory. It means backtracking
3486 * encountered a case of pointer subtraction.
3489 /* scalars can only be spilled into stack */
3490 if (insn->dst_reg != BPF_REG_FP)
3492 spi = (-insn->off - 1) / BPF_REG_SIZE;
3494 verbose(env, "BUG spi %d\n", spi);
3495 WARN_ONCE(1, "verifier backtracking bug");
3498 if (!bt_is_slot_set(bt, spi))
3500 bt_clear_slot(bt, spi);
3501 if (class == BPF_STX)
3502 bt_set_reg(bt, sreg);
3503 } else if (class == BPF_JMP || class == BPF_JMP32) {
3504 if (bpf_pseudo_call(insn)) {
3505 int subprog_insn_idx, subprog;
3507 subprog_insn_idx = idx + insn->imm + 1;
3508 subprog = find_subprog(env, subprog_insn_idx);
3512 if (subprog_is_global(env, subprog)) {
3513 /* check that jump history doesn't have any
3514 * extra instructions from subprog; the next
3515 * instruction after call to global subprog
3516 * should be literally next instruction in
3519 WARN_ONCE(idx + 1 != subseq_idx, "verifier backtracking bug");
3520 /* r1-r5 are invalidated after subprog call,
3521 * so for global func call it shouldn't be set
3524 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3525 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3526 WARN_ONCE(1, "verifier backtracking bug");
3529 /* global subprog always sets R0 */
3530 bt_clear_reg(bt, BPF_REG_0);
3533 /* static subprog call instruction, which
3534 * means that we are exiting current subprog,
3535 * so only r1-r5 could be still requested as
3536 * precise, r0 and r6-r10 or any stack slot in
3537 * the current frame should be zero by now
3539 if (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) {
3540 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3541 WARN_ONCE(1, "verifier backtracking bug");
3544 /* we don't track register spills perfectly,
3545 * so fallback to force-precise instead of failing */
3546 if (bt_stack_mask(bt) != 0)
3548 /* propagate r1-r5 to the caller */
3549 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
3550 if (bt_is_reg_set(bt, i)) {
3551 bt_clear_reg(bt, i);
3552 bt_set_frame_reg(bt, bt->frame - 1, i);
3555 if (bt_subprog_exit(bt))
3559 } else if ((bpf_helper_call(insn) &&
3560 is_callback_calling_function(insn->imm) &&
3561 !is_async_callback_calling_function(insn->imm)) ||
3562 (bpf_pseudo_kfunc_call(insn) && is_callback_calling_kfunc(insn->imm))) {
3563 /* callback-calling helper or kfunc call, which means
3564 * we are exiting from subprog, but unlike the subprog
3565 * call handling above, we shouldn't propagate
3566 * precision of r1-r5 (if any requested), as they are
3567 * not actually arguments passed directly to callback
3570 if (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) {
3571 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3572 WARN_ONCE(1, "verifier backtracking bug");
3575 if (bt_stack_mask(bt) != 0)
3577 /* clear r1-r5 in callback subprog's mask */
3578 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
3579 bt_clear_reg(bt, i);
3580 if (bt_subprog_exit(bt))
3583 } else if (opcode == BPF_CALL) {
3584 /* kfunc with imm==0 is invalid and fixup_kfunc_call will
3585 * catch this error later. Make backtracking conservative
3588 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && insn->imm == 0)
3590 /* regular helper call sets R0 */
3591 bt_clear_reg(bt, BPF_REG_0);
3592 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3593 /* if backtracing was looking for registers R1-R5
3594 * they should have been found already.
3596 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3597 WARN_ONCE(1, "verifier backtracking bug");
3600 } else if (opcode == BPF_EXIT) {
3603 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3604 /* if backtracing was looking for registers R1-R5
3605 * they should have been found already.
3607 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3608 WARN_ONCE(1, "verifier backtracking bug");
3612 /* BPF_EXIT in subprog or callback always returns
3613 * right after the call instruction, so by checking
3614 * whether the instruction at subseq_idx-1 is subprog
3615 * call or not we can distinguish actual exit from
3616 * *subprog* from exit from *callback*. In the former
3617 * case, we need to propagate r0 precision, if
3618 * necessary. In the former we never do that.
3620 r0_precise = subseq_idx - 1 >= 0 &&
3621 bpf_pseudo_call(&env->prog->insnsi[subseq_idx - 1]) &&
3622 bt_is_reg_set(bt, BPF_REG_0);
3624 bt_clear_reg(bt, BPF_REG_0);
3625 if (bt_subprog_enter(bt))
3629 bt_set_reg(bt, BPF_REG_0);
3630 /* r6-r9 and stack slots will stay set in caller frame
3631 * bitmasks until we return back from callee(s)
3634 } else if (BPF_SRC(insn->code) == BPF_X) {
3635 if (!bt_is_reg_set(bt, dreg) && !bt_is_reg_set(bt, sreg))
3638 * Both dreg and sreg need precision before
3639 * this insn. If only sreg was marked precise
3640 * before it would be equally necessary to
3641 * propagate it to dreg.
3643 bt_set_reg(bt, dreg);
3644 bt_set_reg(bt, sreg);
3645 /* else dreg <cond> K
3646 * Only dreg still needs precision before
3647 * this insn, so for the K-based conditional
3648 * there is nothing new to be marked.
3651 } else if (class == BPF_LD) {
3652 if (!bt_is_reg_set(bt, dreg))
3654 bt_clear_reg(bt, dreg);
3655 /* It's ld_imm64 or ld_abs or ld_ind.
3656 * For ld_imm64 no further tracking of precision
3657 * into parent is necessary
3659 if (mode == BPF_IND || mode == BPF_ABS)
3660 /* to be analyzed */
3666 /* the scalar precision tracking algorithm:
3667 * . at the start all registers have precise=false.
3668 * . scalar ranges are tracked as normal through alu and jmp insns.
3669 * . once precise value of the scalar register is used in:
3670 * . ptr + scalar alu
3671 * . if (scalar cond K|scalar)
3672 * . helper_call(.., scalar, ...) where ARG_CONST is expected
3673 * backtrack through the verifier states and mark all registers and
3674 * stack slots with spilled constants that these scalar regisers
3675 * should be precise.
3676 * . during state pruning two registers (or spilled stack slots)
3677 * are equivalent if both are not precise.
3679 * Note the verifier cannot simply walk register parentage chain,
3680 * since many different registers and stack slots could have been
3681 * used to compute single precise scalar.
3683 * The approach of starting with precise=true for all registers and then
3684 * backtrack to mark a register as not precise when the verifier detects
3685 * that program doesn't care about specific value (e.g., when helper
3686 * takes register as ARG_ANYTHING parameter) is not safe.
3688 * It's ok to walk single parentage chain of the verifier states.
3689 * It's possible that this backtracking will go all the way till 1st insn.
3690 * All other branches will be explored for needing precision later.
3692 * The backtracking needs to deal with cases like:
3693 * 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)
3696 * if r5 > 0x79f goto pc+7
3697 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
3700 * call bpf_perf_event_output#25
3701 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
3705 * call foo // uses callee's r6 inside to compute r0
3709 * to track above reg_mask/stack_mask needs to be independent for each frame.
3711 * Also if parent's curframe > frame where backtracking started,
3712 * the verifier need to mark registers in both frames, otherwise callees
3713 * may incorrectly prune callers. This is similar to
3714 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
3716 * For now backtracking falls back into conservative marking.
3718 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
3719 struct bpf_verifier_state *st)
3721 struct bpf_func_state *func;
3722 struct bpf_reg_state *reg;
3725 if (env->log.level & BPF_LOG_LEVEL2) {
3726 verbose(env, "mark_precise: frame%d: falling back to forcing all scalars precise\n",
3730 /* big hammer: mark all scalars precise in this path.
3731 * pop_stack may still get !precise scalars.
3732 * We also skip current state and go straight to first parent state,
3733 * because precision markings in current non-checkpointed state are
3734 * not needed. See why in the comment in __mark_chain_precision below.
3736 for (st = st->parent; st; st = st->parent) {
3737 for (i = 0; i <= st->curframe; i++) {
3738 func = st->frame[i];
3739 for (j = 0; j < BPF_REG_FP; j++) {
3740 reg = &func->regs[j];
3741 if (reg->type != SCALAR_VALUE || reg->precise)
3743 reg->precise = true;
3744 if (env->log.level & BPF_LOG_LEVEL2) {
3745 verbose(env, "force_precise: frame%d: forcing r%d to be precise\n",
3749 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
3750 if (!is_spilled_reg(&func->stack[j]))
3752 reg = &func->stack[j].spilled_ptr;
3753 if (reg->type != SCALAR_VALUE || reg->precise)
3755 reg->precise = true;
3756 if (env->log.level & BPF_LOG_LEVEL2) {
3757 verbose(env, "force_precise: frame%d: forcing fp%d to be precise\n",
3765 static void mark_all_scalars_imprecise(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
3767 struct bpf_func_state *func;
3768 struct bpf_reg_state *reg;
3771 for (i = 0; i <= st->curframe; i++) {
3772 func = st->frame[i];
3773 for (j = 0; j < BPF_REG_FP; j++) {
3774 reg = &func->regs[j];
3775 if (reg->type != SCALAR_VALUE)
3777 reg->precise = false;
3779 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
3780 if (!is_spilled_reg(&func->stack[j]))
3782 reg = &func->stack[j].spilled_ptr;
3783 if (reg->type != SCALAR_VALUE)
3785 reg->precise = false;
3790 static bool idset_contains(struct bpf_idset *s, u32 id)
3794 for (i = 0; i < s->count; ++i)
3795 if (s->ids[i] == id)
3801 static int idset_push(struct bpf_idset *s, u32 id)
3803 if (WARN_ON_ONCE(s->count >= ARRAY_SIZE(s->ids)))
3805 s->ids[s->count++] = id;
3809 static void idset_reset(struct bpf_idset *s)
3814 /* Collect a set of IDs for all registers currently marked as precise in env->bt.
3815 * Mark all registers with these IDs as precise.
3817 static int mark_precise_scalar_ids(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
3819 struct bpf_idset *precise_ids = &env->idset_scratch;
3820 struct backtrack_state *bt = &env->bt;
3821 struct bpf_func_state *func;
3822 struct bpf_reg_state *reg;
3823 DECLARE_BITMAP(mask, 64);
3826 idset_reset(precise_ids);
3828 for (fr = bt->frame; fr >= 0; fr--) {
3829 func = st->frame[fr];
3831 bitmap_from_u64(mask, bt_frame_reg_mask(bt, fr));
3832 for_each_set_bit(i, mask, 32) {
3833 reg = &func->regs[i];
3834 if (!reg->id || reg->type != SCALAR_VALUE)
3836 if (idset_push(precise_ids, reg->id))
3840 bitmap_from_u64(mask, bt_frame_stack_mask(bt, fr));
3841 for_each_set_bit(i, mask, 64) {
3842 if (i >= func->allocated_stack / BPF_REG_SIZE)
3844 if (!is_spilled_scalar_reg(&func->stack[i]))
3846 reg = &func->stack[i].spilled_ptr;
3849 if (idset_push(precise_ids, reg->id))
3854 for (fr = 0; fr <= st->curframe; ++fr) {
3855 func = st->frame[fr];
3857 for (i = BPF_REG_0; i < BPF_REG_10; ++i) {
3858 reg = &func->regs[i];
3861 if (!idset_contains(precise_ids, reg->id))
3863 bt_set_frame_reg(bt, fr, i);
3865 for (i = 0; i < func->allocated_stack / BPF_REG_SIZE; ++i) {
3866 if (!is_spilled_scalar_reg(&func->stack[i]))
3868 reg = &func->stack[i].spilled_ptr;
3871 if (!idset_contains(precise_ids, reg->id))
3873 bt_set_frame_slot(bt, fr, i);
3881 * __mark_chain_precision() backtracks BPF program instruction sequence and
3882 * chain of verifier states making sure that register *regno* (if regno >= 0)
3883 * and/or stack slot *spi* (if spi >= 0) are marked as precisely tracked
3884 * SCALARS, as well as any other registers and slots that contribute to
3885 * a tracked state of given registers/stack slots, depending on specific BPF
3886 * assembly instructions (see backtrack_insns() for exact instruction handling
3887 * logic). This backtracking relies on recorded jmp_history and is able to
3888 * traverse entire chain of parent states. This process ends only when all the
3889 * necessary registers/slots and their transitive dependencies are marked as
3892 * One important and subtle aspect is that precise marks *do not matter* in
3893 * the currently verified state (current state). It is important to understand
3894 * why this is the case.
3896 * First, note that current state is the state that is not yet "checkpointed",
3897 * i.e., it is not yet put into env->explored_states, and it has no children
3898 * states as well. It's ephemeral, and can end up either a) being discarded if
3899 * compatible explored state is found at some point or BPF_EXIT instruction is
3900 * reached or b) checkpointed and put into env->explored_states, branching out
3901 * into one or more children states.
3903 * In the former case, precise markings in current state are completely
3904 * ignored by state comparison code (see regsafe() for details). Only
3905 * checkpointed ("old") state precise markings are important, and if old
3906 * state's register/slot is precise, regsafe() assumes current state's
3907 * register/slot as precise and checks value ranges exactly and precisely. If
3908 * states turn out to be compatible, current state's necessary precise
3909 * markings and any required parent states' precise markings are enforced
3910 * after the fact with propagate_precision() logic, after the fact. But it's
3911 * important to realize that in this case, even after marking current state
3912 * registers/slots as precise, we immediately discard current state. So what
3913 * actually matters is any of the precise markings propagated into current
3914 * state's parent states, which are always checkpointed (due to b) case above).
3915 * As such, for scenario a) it doesn't matter if current state has precise
3916 * markings set or not.
3918 * Now, for the scenario b), checkpointing and forking into child(ren)
3919 * state(s). Note that before current state gets to checkpointing step, any
3920 * processed instruction always assumes precise SCALAR register/slot
3921 * knowledge: if precise value or range is useful to prune jump branch, BPF
3922 * verifier takes this opportunity enthusiastically. Similarly, when
3923 * register's value is used to calculate offset or memory address, exact
3924 * knowledge of SCALAR range is assumed, checked, and enforced. So, similar to
3925 * what we mentioned above about state comparison ignoring precise markings
3926 * during state comparison, BPF verifier ignores and also assumes precise
3927 * markings *at will* during instruction verification process. But as verifier
3928 * assumes precision, it also propagates any precision dependencies across
3929 * parent states, which are not yet finalized, so can be further restricted
3930 * based on new knowledge gained from restrictions enforced by their children
3931 * states. This is so that once those parent states are finalized, i.e., when
3932 * they have no more active children state, state comparison logic in
3933 * is_state_visited() would enforce strict and precise SCALAR ranges, if
3934 * required for correctness.
3936 * To build a bit more intuition, note also that once a state is checkpointed,
3937 * the path we took to get to that state is not important. This is crucial
3938 * property for state pruning. When state is checkpointed and finalized at
3939 * some instruction index, it can be correctly and safely used to "short
3940 * circuit" any *compatible* state that reaches exactly the same instruction
3941 * index. I.e., if we jumped to that instruction from a completely different
3942 * code path than original finalized state was derived from, it doesn't
3943 * matter, current state can be discarded because from that instruction
3944 * forward having a compatible state will ensure we will safely reach the
3945 * exit. States describe preconditions for further exploration, but completely
3946 * forget the history of how we got here.
3948 * This also means that even if we needed precise SCALAR range to get to
3949 * finalized state, but from that point forward *that same* SCALAR register is
3950 * never used in a precise context (i.e., it's precise value is not needed for
3951 * correctness), it's correct and safe to mark such register as "imprecise"
3952 * (i.e., precise marking set to false). This is what we rely on when we do
3953 * not set precise marking in current state. If no child state requires
3954 * precision for any given SCALAR register, it's safe to dictate that it can
3955 * be imprecise. If any child state does require this register to be precise,
3956 * we'll mark it precise later retroactively during precise markings
3957 * propagation from child state to parent states.
3959 * Skipping precise marking setting in current state is a mild version of
3960 * relying on the above observation. But we can utilize this property even
3961 * more aggressively by proactively forgetting any precise marking in the
3962 * current state (which we inherited from the parent state), right before we
3963 * checkpoint it and branch off into new child state. This is done by
3964 * mark_all_scalars_imprecise() to hopefully get more permissive and generic
3965 * finalized states which help in short circuiting more future states.
3967 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno)
3969 struct backtrack_state *bt = &env->bt;
3970 struct bpf_verifier_state *st = env->cur_state;
3971 int first_idx = st->first_insn_idx;
3972 int last_idx = env->insn_idx;
3973 int subseq_idx = -1;
3974 struct bpf_func_state *func;
3975 struct bpf_reg_state *reg;
3976 bool skip_first = true;
3979 if (!env->bpf_capable)
3982 /* set frame number from which we are starting to backtrack */
3983 bt_init(bt, env->cur_state->curframe);
3985 /* Do sanity checks against current state of register and/or stack
3986 * slot, but don't set precise flag in current state, as precision
3987 * tracking in the current state is unnecessary.
3989 func = st->frame[bt->frame];
3991 reg = &func->regs[regno];
3992 if (reg->type != SCALAR_VALUE) {
3993 WARN_ONCE(1, "backtracing misuse");
3996 bt_set_reg(bt, regno);
4003 DECLARE_BITMAP(mask, 64);
4004 u32 history = st->jmp_history_cnt;
4006 if (env->log.level & BPF_LOG_LEVEL2) {
4007 verbose(env, "mark_precise: frame%d: last_idx %d first_idx %d subseq_idx %d \n",
4008 bt->frame, last_idx, first_idx, subseq_idx);
4011 /* If some register with scalar ID is marked as precise,
4012 * make sure that all registers sharing this ID are also precise.
4013 * This is needed to estimate effect of find_equal_scalars().
4014 * Do this at the last instruction of each state,
4015 * bpf_reg_state::id fields are valid for these instructions.
4017 * Allows to track precision in situation like below:
4019 * r2 = unknown value
4023 * r1 = r2 // r1 and r2 now share the same ID
4025 * --- state #1 {r1.id = A, r2.id = A} ---
4027 * if (r2 > 10) goto exit; // find_equal_scalars() assigns range to r1
4029 * --- state #2 {r1.id = A, r2.id = A} ---
4031 * r3 += r1 // need to mark both r1 and r2
4033 if (mark_precise_scalar_ids(env, st))
4037 /* we are at the entry into subprog, which
4038 * is expected for global funcs, but only if
4039 * requested precise registers are R1-R5
4040 * (which are global func's input arguments)
4042 if (st->curframe == 0 &&
4043 st->frame[0]->subprogno > 0 &&
4044 st->frame[0]->callsite == BPF_MAIN_FUNC &&
4045 bt_stack_mask(bt) == 0 &&
4046 (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) == 0) {
4047 bitmap_from_u64(mask, bt_reg_mask(bt));
4048 for_each_set_bit(i, mask, 32) {
4049 reg = &st->frame[0]->regs[i];
4050 if (reg->type != SCALAR_VALUE) {
4051 bt_clear_reg(bt, i);
4054 reg->precise = true;
4059 verbose(env, "BUG backtracking func entry subprog %d reg_mask %x stack_mask %llx\n",
4060 st->frame[0]->subprogno, bt_reg_mask(bt), bt_stack_mask(bt));
4061 WARN_ONCE(1, "verifier backtracking bug");
4065 for (i = last_idx;;) {
4070 err = backtrack_insn(env, i, subseq_idx, bt);
4072 if (err == -ENOTSUPP) {
4073 mark_all_scalars_precise(env, env->cur_state);
4080 /* Found assignment(s) into tracked register in this state.
4081 * Since this state is already marked, just return.
4082 * Nothing to be tracked further in the parent state.
4088 i = get_prev_insn_idx(st, i, &history);
4089 if (i >= env->prog->len) {
4090 /* This can happen if backtracking reached insn 0
4091 * and there are still reg_mask or stack_mask
4093 * It means the backtracking missed the spot where
4094 * particular register was initialized with a constant.
4096 verbose(env, "BUG backtracking idx %d\n", i);
4097 WARN_ONCE(1, "verifier backtracking bug");
4105 for (fr = bt->frame; fr >= 0; fr--) {
4106 func = st->frame[fr];
4107 bitmap_from_u64(mask, bt_frame_reg_mask(bt, fr));
4108 for_each_set_bit(i, mask, 32) {
4109 reg = &func->regs[i];
4110 if (reg->type != SCALAR_VALUE) {
4111 bt_clear_frame_reg(bt, fr, i);
4115 bt_clear_frame_reg(bt, fr, i);
4117 reg->precise = true;
4120 bitmap_from_u64(mask, bt_frame_stack_mask(bt, fr));
4121 for_each_set_bit(i, mask, 64) {
4122 if (i >= func->allocated_stack / BPF_REG_SIZE) {
4123 /* the sequence of instructions:
4125 * 3: (7b) *(u64 *)(r3 -8) = r0
4126 * 4: (79) r4 = *(u64 *)(r10 -8)
4127 * doesn't contain jmps. It's backtracked
4128 * as a single block.
4129 * During backtracking insn 3 is not recognized as
4130 * stack access, so at the end of backtracking
4131 * stack slot fp-8 is still marked in stack_mask.
4132 * However the parent state may not have accessed
4133 * fp-8 and it's "unallocated" stack space.
4134 * In such case fallback to conservative.
4136 mark_all_scalars_precise(env, env->cur_state);
4141 if (!is_spilled_scalar_reg(&func->stack[i])) {
4142 bt_clear_frame_slot(bt, fr, i);
4145 reg = &func->stack[i].spilled_ptr;
4147 bt_clear_frame_slot(bt, fr, i);
4149 reg->precise = true;
4151 if (env->log.level & BPF_LOG_LEVEL2) {
4152 fmt_reg_mask(env->tmp_str_buf, TMP_STR_BUF_LEN,
4153 bt_frame_reg_mask(bt, fr));
4154 verbose(env, "mark_precise: frame%d: parent state regs=%s ",
4155 fr, env->tmp_str_buf);
4156 fmt_stack_mask(env->tmp_str_buf, TMP_STR_BUF_LEN,
4157 bt_frame_stack_mask(bt, fr));
4158 verbose(env, "stack=%s: ", env->tmp_str_buf);
4159 print_verifier_state(env, func, true);
4166 subseq_idx = first_idx;
4167 last_idx = st->last_insn_idx;
4168 first_idx = st->first_insn_idx;
4171 /* if we still have requested precise regs or slots, we missed
4172 * something (e.g., stack access through non-r10 register), so
4173 * fallback to marking all precise
4175 if (!bt_empty(bt)) {
4176 mark_all_scalars_precise(env, env->cur_state);
4183 int mark_chain_precision(struct bpf_verifier_env *env, int regno)
4185 return __mark_chain_precision(env, regno);
4188 /* mark_chain_precision_batch() assumes that env->bt is set in the caller to
4189 * desired reg and stack masks across all relevant frames
4191 static int mark_chain_precision_batch(struct bpf_verifier_env *env)
4193 return __mark_chain_precision(env, -1);
4196 static bool is_spillable_regtype(enum bpf_reg_type type)
4198 switch (base_type(type)) {
4199 case PTR_TO_MAP_VALUE:
4203 case PTR_TO_PACKET_META:
4204 case PTR_TO_PACKET_END:
4205 case PTR_TO_FLOW_KEYS:
4206 case CONST_PTR_TO_MAP:
4208 case PTR_TO_SOCK_COMMON:
4209 case PTR_TO_TCP_SOCK:
4210 case PTR_TO_XDP_SOCK:
4215 case PTR_TO_MAP_KEY:
4222 /* Does this register contain a constant zero? */
4223 static bool register_is_null(struct bpf_reg_state *reg)
4225 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
4228 static bool register_is_const(struct bpf_reg_state *reg)
4230 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
4233 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
4235 return tnum_is_unknown(reg->var_off) &&
4236 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
4237 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
4238 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
4239 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
4242 static bool register_is_bounded(struct bpf_reg_state *reg)
4244 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
4247 static bool __is_pointer_value(bool allow_ptr_leaks,
4248 const struct bpf_reg_state *reg)
4250 if (allow_ptr_leaks)
4253 return reg->type != SCALAR_VALUE;
4256 /* Copy src state preserving dst->parent and dst->live fields */
4257 static void copy_register_state(struct bpf_reg_state *dst, const struct bpf_reg_state *src)
4259 struct bpf_reg_state *parent = dst->parent;
4260 enum bpf_reg_liveness live = dst->live;
4263 dst->parent = parent;
4267 static void save_register_state(struct bpf_func_state *state,
4268 int spi, struct bpf_reg_state *reg,
4273 copy_register_state(&state->stack[spi].spilled_ptr, reg);
4274 if (size == BPF_REG_SIZE)
4275 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
4277 for (i = BPF_REG_SIZE; i > BPF_REG_SIZE - size; i--)
4278 state->stack[spi].slot_type[i - 1] = STACK_SPILL;
4280 /* size < 8 bytes spill */
4282 scrub_spilled_slot(&state->stack[spi].slot_type[i - 1]);
4285 static bool is_bpf_st_mem(struct bpf_insn *insn)
4287 return BPF_CLASS(insn->code) == BPF_ST && BPF_MODE(insn->code) == BPF_MEM;
4290 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
4291 * stack boundary and alignment are checked in check_mem_access()
4293 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
4294 /* stack frame we're writing to */
4295 struct bpf_func_state *state,
4296 int off, int size, int value_regno,
4299 struct bpf_func_state *cur; /* state of the current function */
4300 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
4301 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
4302 struct bpf_reg_state *reg = NULL;
4303 u32 dst_reg = insn->dst_reg;
4305 err = grow_stack_state(state, round_up(slot + 1, BPF_REG_SIZE));
4308 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
4309 * so it's aligned access and [off, off + size) are within stack limits
4311 if (!env->allow_ptr_leaks &&
4312 state->stack[spi].slot_type[0] == STACK_SPILL &&
4313 size != BPF_REG_SIZE) {
4314 verbose(env, "attempt to corrupt spilled pointer on stack\n");
4318 cur = env->cur_state->frame[env->cur_state->curframe];
4319 if (value_regno >= 0)
4320 reg = &cur->regs[value_regno];
4321 if (!env->bypass_spec_v4) {
4322 bool sanitize = reg && is_spillable_regtype(reg->type);
4324 for (i = 0; i < size; i++) {
4325 u8 type = state->stack[spi].slot_type[i];
4327 if (type != STACK_MISC && type != STACK_ZERO) {
4334 env->insn_aux_data[insn_idx].sanitize_stack_spill = true;
4337 err = destroy_if_dynptr_stack_slot(env, state, spi);
4341 mark_stack_slot_scratched(env, spi);
4342 if (reg && !(off % BPF_REG_SIZE) && register_is_bounded(reg) &&
4343 !register_is_null(reg) && env->bpf_capable) {
4344 if (dst_reg != BPF_REG_FP) {
4345 /* The backtracking logic can only recognize explicit
4346 * stack slot address like [fp - 8]. Other spill of
4347 * scalar via different register has to be conservative.
4348 * Backtrack from here and mark all registers as precise
4349 * that contributed into 'reg' being a constant.
4351 err = mark_chain_precision(env, value_regno);
4355 save_register_state(state, spi, reg, size);
4356 /* Break the relation on a narrowing spill. */
4357 if (fls64(reg->umax_value) > BITS_PER_BYTE * size)
4358 state->stack[spi].spilled_ptr.id = 0;
4359 } else if (!reg && !(off % BPF_REG_SIZE) && is_bpf_st_mem(insn) &&
4360 insn->imm != 0 && env->bpf_capable) {
4361 struct bpf_reg_state fake_reg = {};
4363 __mark_reg_known(&fake_reg, (u32)insn->imm);
4364 fake_reg.type = SCALAR_VALUE;
4365 save_register_state(state, spi, &fake_reg, size);
4366 } else if (reg && is_spillable_regtype(reg->type)) {
4367 /* register containing pointer is being spilled into stack */
4368 if (size != BPF_REG_SIZE) {
4369 verbose_linfo(env, insn_idx, "; ");
4370 verbose(env, "invalid size of register spill\n");
4373 if (state != cur && reg->type == PTR_TO_STACK) {
4374 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
4377 save_register_state(state, spi, reg, size);
4379 u8 type = STACK_MISC;
4381 /* regular write of data into stack destroys any spilled ptr */
4382 state->stack[spi].spilled_ptr.type = NOT_INIT;
4383 /* Mark slots as STACK_MISC if they belonged to spilled ptr/dynptr/iter. */
4384 if (is_stack_slot_special(&state->stack[spi]))
4385 for (i = 0; i < BPF_REG_SIZE; i++)
4386 scrub_spilled_slot(&state->stack[spi].slot_type[i]);
4388 /* only mark the slot as written if all 8 bytes were written
4389 * otherwise read propagation may incorrectly stop too soon
4390 * when stack slots are partially written.
4391 * This heuristic means that read propagation will be
4392 * conservative, since it will add reg_live_read marks
4393 * to stack slots all the way to first state when programs
4394 * writes+reads less than 8 bytes
4396 if (size == BPF_REG_SIZE)
4397 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
4399 /* when we zero initialize stack slots mark them as such */
4400 if ((reg && register_is_null(reg)) ||
4401 (!reg && is_bpf_st_mem(insn) && insn->imm == 0)) {
4402 /* backtracking doesn't work for STACK_ZERO yet. */
4403 err = mark_chain_precision(env, value_regno);
4409 /* Mark slots affected by this stack write. */
4410 for (i = 0; i < size; i++)
4411 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
4417 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
4418 * known to contain a variable offset.
4419 * This function checks whether the write is permitted and conservatively
4420 * tracks the effects of the write, considering that each stack slot in the
4421 * dynamic range is potentially written to.
4423 * 'off' includes 'regno->off'.
4424 * 'value_regno' can be -1, meaning that an unknown value is being written to
4427 * Spilled pointers in range are not marked as written because we don't know
4428 * what's going to be actually written. This means that read propagation for
4429 * future reads cannot be terminated by this write.
4431 * For privileged programs, uninitialized stack slots are considered
4432 * initialized by this write (even though we don't know exactly what offsets
4433 * are going to be written to). The idea is that we don't want the verifier to
4434 * reject future reads that access slots written to through variable offsets.
4436 static int check_stack_write_var_off(struct bpf_verifier_env *env,
4437 /* func where register points to */
4438 struct bpf_func_state *state,
4439 int ptr_regno, int off, int size,
4440 int value_regno, int insn_idx)
4442 struct bpf_func_state *cur; /* state of the current function */
4443 int min_off, max_off;
4445 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
4446 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
4447 bool writing_zero = false;
4448 /* set if the fact that we're writing a zero is used to let any
4449 * stack slots remain STACK_ZERO
4451 bool zero_used = false;
4453 cur = env->cur_state->frame[env->cur_state->curframe];
4454 ptr_reg = &cur->regs[ptr_regno];
4455 min_off = ptr_reg->smin_value + off;
4456 max_off = ptr_reg->smax_value + off + size;
4457 if (value_regno >= 0)
4458 value_reg = &cur->regs[value_regno];
4459 if ((value_reg && register_is_null(value_reg)) ||
4460 (!value_reg && is_bpf_st_mem(insn) && insn->imm == 0))
4461 writing_zero = true;
4463 err = grow_stack_state(state, round_up(-min_off, BPF_REG_SIZE));
4467 for (i = min_off; i < max_off; i++) {
4471 err = destroy_if_dynptr_stack_slot(env, state, spi);
4476 /* Variable offset writes destroy any spilled pointers in range. */
4477 for (i = min_off; i < max_off; i++) {
4478 u8 new_type, *stype;
4482 spi = slot / BPF_REG_SIZE;
4483 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
4484 mark_stack_slot_scratched(env, spi);
4486 if (!env->allow_ptr_leaks && *stype != STACK_MISC && *stype != STACK_ZERO) {
4487 /* Reject the write if range we may write to has not
4488 * been initialized beforehand. If we didn't reject
4489 * here, the ptr status would be erased below (even
4490 * though not all slots are actually overwritten),
4491 * possibly opening the door to leaks.
4493 * We do however catch STACK_INVALID case below, and
4494 * only allow reading possibly uninitialized memory
4495 * later for CAP_PERFMON, as the write may not happen to
4498 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
4503 /* Erase all spilled pointers. */
4504 state->stack[spi].spilled_ptr.type = NOT_INIT;
4506 /* Update the slot type. */
4507 new_type = STACK_MISC;
4508 if (writing_zero && *stype == STACK_ZERO) {
4509 new_type = STACK_ZERO;
4512 /* If the slot is STACK_INVALID, we check whether it's OK to
4513 * pretend that it will be initialized by this write. The slot
4514 * might not actually be written to, and so if we mark it as
4515 * initialized future reads might leak uninitialized memory.
4516 * For privileged programs, we will accept such reads to slots
4517 * that may or may not be written because, if we're reject
4518 * them, the error would be too confusing.
4520 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
4521 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
4528 /* backtracking doesn't work for STACK_ZERO yet. */
4529 err = mark_chain_precision(env, value_regno);
4536 /* When register 'dst_regno' is assigned some values from stack[min_off,
4537 * max_off), we set the register's type according to the types of the
4538 * respective stack slots. If all the stack values are known to be zeros, then
4539 * so is the destination reg. Otherwise, the register is considered to be
4540 * SCALAR. This function does not deal with register filling; the caller must
4541 * ensure that all spilled registers in the stack range have been marked as
4544 static void mark_reg_stack_read(struct bpf_verifier_env *env,
4545 /* func where src register points to */
4546 struct bpf_func_state *ptr_state,
4547 int min_off, int max_off, int dst_regno)
4549 struct bpf_verifier_state *vstate = env->cur_state;
4550 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4555 for (i = min_off; i < max_off; i++) {
4557 spi = slot / BPF_REG_SIZE;
4558 mark_stack_slot_scratched(env, spi);
4559 stype = ptr_state->stack[spi].slot_type;
4560 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
4564 if (zeros == max_off - min_off) {
4565 /* any access_size read into register is zero extended,
4566 * so the whole register == const_zero
4568 __mark_reg_const_zero(&state->regs[dst_regno]);
4569 /* backtracking doesn't support STACK_ZERO yet,
4570 * so mark it precise here, so that later
4571 * backtracking can stop here.
4572 * Backtracking may not need this if this register
4573 * doesn't participate in pointer adjustment.
4574 * Forward propagation of precise flag is not
4575 * necessary either. This mark is only to stop
4576 * backtracking. Any register that contributed
4577 * to const 0 was marked precise before spill.
4579 state->regs[dst_regno].precise = true;
4581 /* have read misc data from the stack */
4582 mark_reg_unknown(env, state->regs, dst_regno);
4584 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4587 /* Read the stack at 'off' and put the results into the register indicated by
4588 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
4591 * 'dst_regno' can be -1, meaning that the read value is not going to a
4594 * The access is assumed to be within the current stack bounds.
4596 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
4597 /* func where src register points to */
4598 struct bpf_func_state *reg_state,
4599 int off, int size, int dst_regno)
4601 struct bpf_verifier_state *vstate = env->cur_state;
4602 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4603 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
4604 struct bpf_reg_state *reg;
4607 stype = reg_state->stack[spi].slot_type;
4608 reg = ®_state->stack[spi].spilled_ptr;
4610 mark_stack_slot_scratched(env, spi);
4612 if (is_spilled_reg(®_state->stack[spi])) {
4615 for (i = BPF_REG_SIZE - 1; i > 0 && stype[i - 1] == STACK_SPILL; i--)
4618 if (size != BPF_REG_SIZE || spill_size != BPF_REG_SIZE) {
4619 if (reg->type != SCALAR_VALUE) {
4620 verbose_linfo(env, env->insn_idx, "; ");
4621 verbose(env, "invalid size of register fill\n");
4625 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4629 if (!(off % BPF_REG_SIZE) && size == spill_size) {
4630 /* The earlier check_reg_arg() has decided the
4631 * subreg_def for this insn. Save it first.
4633 s32 subreg_def = state->regs[dst_regno].subreg_def;
4635 copy_register_state(&state->regs[dst_regno], reg);
4636 state->regs[dst_regno].subreg_def = subreg_def;
4638 for (i = 0; i < size; i++) {
4639 type = stype[(slot - i) % BPF_REG_SIZE];
4640 if (type == STACK_SPILL)
4642 if (type == STACK_MISC)
4644 if (type == STACK_INVALID && env->allow_uninit_stack)
4646 verbose(env, "invalid read from stack off %d+%d size %d\n",
4650 mark_reg_unknown(env, state->regs, dst_regno);
4652 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4656 if (dst_regno >= 0) {
4657 /* restore register state from stack */
4658 copy_register_state(&state->regs[dst_regno], reg);
4659 /* mark reg as written since spilled pointer state likely
4660 * has its liveness marks cleared by is_state_visited()
4661 * which resets stack/reg liveness for state transitions
4663 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4664 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
4665 /* If dst_regno==-1, the caller is asking us whether
4666 * it is acceptable to use this value as a SCALAR_VALUE
4668 * We must not allow unprivileged callers to do that
4669 * with spilled pointers.
4671 verbose(env, "leaking pointer from stack off %d\n",
4675 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4677 for (i = 0; i < size; i++) {
4678 type = stype[(slot - i) % BPF_REG_SIZE];
4679 if (type == STACK_MISC)
4681 if (type == STACK_ZERO)
4683 if (type == STACK_INVALID && env->allow_uninit_stack)
4685 verbose(env, "invalid read from stack off %d+%d size %d\n",
4689 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4691 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
4696 enum bpf_access_src {
4697 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
4698 ACCESS_HELPER = 2, /* the access is performed by a helper */
4701 static int check_stack_range_initialized(struct bpf_verifier_env *env,
4702 int regno, int off, int access_size,
4703 bool zero_size_allowed,
4704 enum bpf_access_src type,
4705 struct bpf_call_arg_meta *meta);
4707 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
4709 return cur_regs(env) + regno;
4712 /* Read the stack at 'ptr_regno + off' and put the result into the register
4714 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
4715 * but not its variable offset.
4716 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
4718 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
4719 * filling registers (i.e. reads of spilled register cannot be detected when
4720 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
4721 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
4722 * offset; for a fixed offset check_stack_read_fixed_off should be used
4725 static int check_stack_read_var_off(struct bpf_verifier_env *env,
4726 int ptr_regno, int off, int size, int dst_regno)
4728 /* The state of the source register. */
4729 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
4730 struct bpf_func_state *ptr_state = func(env, reg);
4732 int min_off, max_off;
4734 /* Note that we pass a NULL meta, so raw access will not be permitted.
4736 err = check_stack_range_initialized(env, ptr_regno, off, size,
4737 false, ACCESS_DIRECT, NULL);
4741 min_off = reg->smin_value + off;
4742 max_off = reg->smax_value + off;
4743 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
4747 /* check_stack_read dispatches to check_stack_read_fixed_off or
4748 * check_stack_read_var_off.
4750 * The caller must ensure that the offset falls within the allocated stack
4753 * 'dst_regno' is a register which will receive the value from the stack. It
4754 * can be -1, meaning that the read value is not going to a register.
4756 static int check_stack_read(struct bpf_verifier_env *env,
4757 int ptr_regno, int off, int size,
4760 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
4761 struct bpf_func_state *state = func(env, reg);
4763 /* Some accesses are only permitted with a static offset. */
4764 bool var_off = !tnum_is_const(reg->var_off);
4766 /* The offset is required to be static when reads don't go to a
4767 * register, in order to not leak pointers (see
4768 * check_stack_read_fixed_off).
4770 if (dst_regno < 0 && var_off) {
4773 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4774 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
4778 /* Variable offset is prohibited for unprivileged mode for simplicity
4779 * since it requires corresponding support in Spectre masking for stack
4780 * ALU. See also retrieve_ptr_limit(). The check in
4781 * check_stack_access_for_ptr_arithmetic() called by
4782 * adjust_ptr_min_max_vals() prevents users from creating stack pointers
4783 * with variable offsets, therefore no check is required here. Further,
4784 * just checking it here would be insufficient as speculative stack
4785 * writes could still lead to unsafe speculative behaviour.
4788 off += reg->var_off.value;
4789 err = check_stack_read_fixed_off(env, state, off, size,
4792 /* Variable offset stack reads need more conservative handling
4793 * than fixed offset ones. Note that dst_regno >= 0 on this
4796 err = check_stack_read_var_off(env, ptr_regno, off, size,
4803 /* check_stack_write dispatches to check_stack_write_fixed_off or
4804 * check_stack_write_var_off.
4806 * 'ptr_regno' is the register used as a pointer into the stack.
4807 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
4808 * 'value_regno' is the register whose value we're writing to the stack. It can
4809 * be -1, meaning that we're not writing from a register.
4811 * The caller must ensure that the offset falls within the maximum stack size.
4813 static int check_stack_write(struct bpf_verifier_env *env,
4814 int ptr_regno, int off, int size,
4815 int value_regno, int insn_idx)
4817 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
4818 struct bpf_func_state *state = func(env, reg);
4821 if (tnum_is_const(reg->var_off)) {
4822 off += reg->var_off.value;
4823 err = check_stack_write_fixed_off(env, state, off, size,
4824 value_regno, insn_idx);
4826 /* Variable offset stack reads need more conservative handling
4827 * than fixed offset ones.
4829 err = check_stack_write_var_off(env, state,
4830 ptr_regno, off, size,
4831 value_regno, insn_idx);
4836 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
4837 int off, int size, enum bpf_access_type type)
4839 struct bpf_reg_state *regs = cur_regs(env);
4840 struct bpf_map *map = regs[regno].map_ptr;
4841 u32 cap = bpf_map_flags_to_cap(map);
4843 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
4844 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
4845 map->value_size, off, size);
4849 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
4850 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
4851 map->value_size, off, size);
4858 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
4859 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
4860 int off, int size, u32 mem_size,
4861 bool zero_size_allowed)
4863 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
4864 struct bpf_reg_state *reg;
4866 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
4869 reg = &cur_regs(env)[regno];
4870 switch (reg->type) {
4871 case PTR_TO_MAP_KEY:
4872 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
4873 mem_size, off, size);
4875 case PTR_TO_MAP_VALUE:
4876 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
4877 mem_size, off, size);
4880 case PTR_TO_PACKET_META:
4881 case PTR_TO_PACKET_END:
4882 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
4883 off, size, regno, reg->id, off, mem_size);
4887 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
4888 mem_size, off, size);
4894 /* check read/write into a memory region with possible variable offset */
4895 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
4896 int off, int size, u32 mem_size,
4897 bool zero_size_allowed)
4899 struct bpf_verifier_state *vstate = env->cur_state;
4900 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4901 struct bpf_reg_state *reg = &state->regs[regno];
4904 /* We may have adjusted the register pointing to memory region, so we
4905 * need to try adding each of min_value and max_value to off
4906 * to make sure our theoretical access will be safe.
4908 * The minimum value is only important with signed
4909 * comparisons where we can't assume the floor of a
4910 * value is 0. If we are using signed variables for our
4911 * index'es we need to make sure that whatever we use
4912 * will have a set floor within our range.
4914 if (reg->smin_value < 0 &&
4915 (reg->smin_value == S64_MIN ||
4916 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
4917 reg->smin_value + off < 0)) {
4918 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
4922 err = __check_mem_access(env, regno, reg->smin_value + off, size,
4923 mem_size, zero_size_allowed);
4925 verbose(env, "R%d min value is outside of the allowed memory range\n",
4930 /* If we haven't set a max value then we need to bail since we can't be
4931 * sure we won't do bad things.
4932 * If reg->umax_value + off could overflow, treat that as unbounded too.
4934 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
4935 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
4939 err = __check_mem_access(env, regno, reg->umax_value + off, size,
4940 mem_size, zero_size_allowed);
4942 verbose(env, "R%d max value is outside of the allowed memory range\n",
4950 static int __check_ptr_off_reg(struct bpf_verifier_env *env,
4951 const struct bpf_reg_state *reg, int regno,
4954 /* Access to this pointer-typed register or passing it to a helper
4955 * is only allowed in its original, unmodified form.
4959 verbose(env, "negative offset %s ptr R%d off=%d disallowed\n",
4960 reg_type_str(env, reg->type), regno, reg->off);
4964 if (!fixed_off_ok && reg->off) {
4965 verbose(env, "dereference of modified %s ptr R%d off=%d disallowed\n",
4966 reg_type_str(env, reg->type), regno, reg->off);
4970 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4973 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4974 verbose(env, "variable %s access var_off=%s disallowed\n",
4975 reg_type_str(env, reg->type), tn_buf);
4982 int check_ptr_off_reg(struct bpf_verifier_env *env,
4983 const struct bpf_reg_state *reg, int regno)
4985 return __check_ptr_off_reg(env, reg, regno, false);
4988 static int map_kptr_match_type(struct bpf_verifier_env *env,
4989 struct btf_field *kptr_field,
4990 struct bpf_reg_state *reg, u32 regno)
4992 const char *targ_name = btf_type_name(kptr_field->kptr.btf, kptr_field->kptr.btf_id);
4994 const char *reg_name = "";
4996 if (btf_is_kernel(reg->btf)) {
4997 perm_flags = PTR_MAYBE_NULL | PTR_TRUSTED | MEM_RCU;
4999 /* Only unreferenced case accepts untrusted pointers */
5000 if (kptr_field->type == BPF_KPTR_UNREF)
5001 perm_flags |= PTR_UNTRUSTED;
5003 perm_flags = PTR_MAYBE_NULL | MEM_ALLOC;
5006 if (base_type(reg->type) != PTR_TO_BTF_ID || (type_flag(reg->type) & ~perm_flags))
5009 /* We need to verify reg->type and reg->btf, before accessing reg->btf */
5010 reg_name = btf_type_name(reg->btf, reg->btf_id);
5012 /* For ref_ptr case, release function check should ensure we get one
5013 * referenced PTR_TO_BTF_ID, and that its fixed offset is 0. For the
5014 * normal store of unreferenced kptr, we must ensure var_off is zero.
5015 * Since ref_ptr cannot be accessed directly by BPF insns, checks for
5016 * reg->off and reg->ref_obj_id are not needed here.
5018 if (__check_ptr_off_reg(env, reg, regno, true))
5021 /* A full type match is needed, as BTF can be vmlinux, module or prog BTF, and
5022 * we also need to take into account the reg->off.
5024 * We want to support cases like:
5032 * v = func(); // PTR_TO_BTF_ID
5033 * val->foo = v; // reg->off is zero, btf and btf_id match type
5034 * val->bar = &v->br; // reg->off is still zero, but we need to retry with
5035 * // first member type of struct after comparison fails
5036 * val->baz = &v->bz; // reg->off is non-zero, so struct needs to be walked
5039 * In the kptr_ref case, check_func_arg_reg_off already ensures reg->off
5040 * is zero. We must also ensure that btf_struct_ids_match does not walk
5041 * the struct to match type against first member of struct, i.e. reject
5042 * second case from above. Hence, when type is BPF_KPTR_REF, we set
5043 * strict mode to true for type match.
5045 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
5046 kptr_field->kptr.btf, kptr_field->kptr.btf_id,
5047 kptr_field->type == BPF_KPTR_REF))
5051 verbose(env, "invalid kptr access, R%d type=%s%s ", regno,
5052 reg_type_str(env, reg->type), reg_name);
5053 verbose(env, "expected=%s%s", reg_type_str(env, PTR_TO_BTF_ID), targ_name);
5054 if (kptr_field->type == BPF_KPTR_UNREF)
5055 verbose(env, " or %s%s\n", reg_type_str(env, PTR_TO_BTF_ID | PTR_UNTRUSTED),
5062 /* The non-sleepable programs and sleepable programs with explicit bpf_rcu_read_lock()
5063 * can dereference RCU protected pointers and result is PTR_TRUSTED.
5065 static bool in_rcu_cs(struct bpf_verifier_env *env)
5067 return env->cur_state->active_rcu_lock ||
5068 env->cur_state->active_lock.ptr ||
5069 !env->prog->aux->sleepable;
5072 /* Once GCC supports btf_type_tag the following mechanism will be replaced with tag check */
5073 BTF_SET_START(rcu_protected_types)
5074 BTF_ID(struct, prog_test_ref_kfunc)
5075 BTF_ID(struct, cgroup)
5076 BTF_ID(struct, bpf_cpumask)
5077 BTF_ID(struct, task_struct)
5078 BTF_SET_END(rcu_protected_types)
5080 static bool rcu_protected_object(const struct btf *btf, u32 btf_id)
5082 if (!btf_is_kernel(btf))
5084 return btf_id_set_contains(&rcu_protected_types, btf_id);
5087 static bool rcu_safe_kptr(const struct btf_field *field)
5089 const struct btf_field_kptr *kptr = &field->kptr;
5091 return field->type == BPF_KPTR_REF && rcu_protected_object(kptr->btf, kptr->btf_id);
5094 static int check_map_kptr_access(struct bpf_verifier_env *env, u32 regno,
5095 int value_regno, int insn_idx,
5096 struct btf_field *kptr_field)
5098 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
5099 int class = BPF_CLASS(insn->code);
5100 struct bpf_reg_state *val_reg;
5102 /* Things we already checked for in check_map_access and caller:
5103 * - Reject cases where variable offset may touch kptr
5104 * - size of access (must be BPF_DW)
5105 * - tnum_is_const(reg->var_off)
5106 * - kptr_field->offset == off + reg->var_off.value
5108 /* Only BPF_[LDX,STX,ST] | BPF_MEM | BPF_DW is supported */
5109 if (BPF_MODE(insn->code) != BPF_MEM) {
5110 verbose(env, "kptr in map can only be accessed using BPF_MEM instruction mode\n");
5114 /* We only allow loading referenced kptr, since it will be marked as
5115 * untrusted, similar to unreferenced kptr.
5117 if (class != BPF_LDX && kptr_field->type == BPF_KPTR_REF) {
5118 verbose(env, "store to referenced kptr disallowed\n");
5122 if (class == BPF_LDX) {
5123 val_reg = reg_state(env, value_regno);
5124 /* We can simply mark the value_regno receiving the pointer
5125 * value from map as PTR_TO_BTF_ID, with the correct type.
5127 mark_btf_ld_reg(env, cur_regs(env), value_regno, PTR_TO_BTF_ID, kptr_field->kptr.btf,
5128 kptr_field->kptr.btf_id,
5129 rcu_safe_kptr(kptr_field) && in_rcu_cs(env) ?
5130 PTR_MAYBE_NULL | MEM_RCU :
5131 PTR_MAYBE_NULL | PTR_UNTRUSTED);
5132 /* For mark_ptr_or_null_reg */
5133 val_reg->id = ++env->id_gen;
5134 } else if (class == BPF_STX) {
5135 val_reg = reg_state(env, value_regno);
5136 if (!register_is_null(val_reg) &&
5137 map_kptr_match_type(env, kptr_field, val_reg, value_regno))
5139 } else if (class == BPF_ST) {
5141 verbose(env, "BPF_ST imm must be 0 when storing to kptr at off=%u\n",
5142 kptr_field->offset);
5146 verbose(env, "kptr in map can only be accessed using BPF_LDX/BPF_STX/BPF_ST\n");
5152 /* check read/write into a map element with possible variable offset */
5153 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
5154 int off, int size, bool zero_size_allowed,
5155 enum bpf_access_src src)
5157 struct bpf_verifier_state *vstate = env->cur_state;
5158 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5159 struct bpf_reg_state *reg = &state->regs[regno];
5160 struct bpf_map *map = reg->map_ptr;
5161 struct btf_record *rec;
5164 err = check_mem_region_access(env, regno, off, size, map->value_size,
5169 if (IS_ERR_OR_NULL(map->record))
5172 for (i = 0; i < rec->cnt; i++) {
5173 struct btf_field *field = &rec->fields[i];
5174 u32 p = field->offset;
5176 /* If any part of a field can be touched by load/store, reject
5177 * this program. To check that [x1, x2) overlaps with [y1, y2),
5178 * it is sufficient to check x1 < y2 && y1 < x2.
5180 if (reg->smin_value + off < p + btf_field_type_size(field->type) &&
5181 p < reg->umax_value + off + size) {
5182 switch (field->type) {
5183 case BPF_KPTR_UNREF:
5185 if (src != ACCESS_DIRECT) {
5186 verbose(env, "kptr cannot be accessed indirectly by helper\n");
5189 if (!tnum_is_const(reg->var_off)) {
5190 verbose(env, "kptr access cannot have variable offset\n");
5193 if (p != off + reg->var_off.value) {
5194 verbose(env, "kptr access misaligned expected=%u off=%llu\n",
5195 p, off + reg->var_off.value);
5198 if (size != bpf_size_to_bytes(BPF_DW)) {
5199 verbose(env, "kptr access size must be BPF_DW\n");
5204 verbose(env, "%s cannot be accessed directly by load/store\n",
5205 btf_field_type_name(field->type));
5213 #define MAX_PACKET_OFF 0xffff
5215 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
5216 const struct bpf_call_arg_meta *meta,
5217 enum bpf_access_type t)
5219 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
5221 switch (prog_type) {
5222 /* Program types only with direct read access go here! */
5223 case BPF_PROG_TYPE_LWT_IN:
5224 case BPF_PROG_TYPE_LWT_OUT:
5225 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
5226 case BPF_PROG_TYPE_SK_REUSEPORT:
5227 case BPF_PROG_TYPE_FLOW_DISSECTOR:
5228 case BPF_PROG_TYPE_CGROUP_SKB:
5233 /* Program types with direct read + write access go here! */
5234 case BPF_PROG_TYPE_SCHED_CLS:
5235 case BPF_PROG_TYPE_SCHED_ACT:
5236 case BPF_PROG_TYPE_XDP:
5237 case BPF_PROG_TYPE_LWT_XMIT:
5238 case BPF_PROG_TYPE_SK_SKB:
5239 case BPF_PROG_TYPE_SK_MSG:
5241 return meta->pkt_access;
5243 env->seen_direct_write = true;
5246 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
5248 env->seen_direct_write = true;
5257 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
5258 int size, bool zero_size_allowed)
5260 struct bpf_reg_state *regs = cur_regs(env);
5261 struct bpf_reg_state *reg = ®s[regno];
5264 /* We may have added a variable offset to the packet pointer; but any
5265 * reg->range we have comes after that. We are only checking the fixed
5269 /* We don't allow negative numbers, because we aren't tracking enough
5270 * detail to prove they're safe.
5272 if (reg->smin_value < 0) {
5273 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
5278 err = reg->range < 0 ? -EINVAL :
5279 __check_mem_access(env, regno, off, size, reg->range,
5282 verbose(env, "R%d offset is outside of the packet\n", regno);
5286 /* __check_mem_access has made sure "off + size - 1" is within u16.
5287 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
5288 * otherwise find_good_pkt_pointers would have refused to set range info
5289 * that __check_mem_access would have rejected this pkt access.
5290 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
5292 env->prog->aux->max_pkt_offset =
5293 max_t(u32, env->prog->aux->max_pkt_offset,
5294 off + reg->umax_value + size - 1);
5299 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
5300 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
5301 enum bpf_access_type t, enum bpf_reg_type *reg_type,
5302 struct btf **btf, u32 *btf_id)
5304 struct bpf_insn_access_aux info = {
5305 .reg_type = *reg_type,
5309 if (env->ops->is_valid_access &&
5310 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
5311 /* A non zero info.ctx_field_size indicates that this field is a
5312 * candidate for later verifier transformation to load the whole
5313 * field and then apply a mask when accessed with a narrower
5314 * access than actual ctx access size. A zero info.ctx_field_size
5315 * will only allow for whole field access and rejects any other
5316 * type of narrower access.
5318 *reg_type = info.reg_type;
5320 if (base_type(*reg_type) == PTR_TO_BTF_ID) {
5322 *btf_id = info.btf_id;
5324 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
5326 /* remember the offset of last byte accessed in ctx */
5327 if (env->prog->aux->max_ctx_offset < off + size)
5328 env->prog->aux->max_ctx_offset = off + size;
5332 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
5336 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
5339 if (size < 0 || off < 0 ||
5340 (u64)off + size > sizeof(struct bpf_flow_keys)) {
5341 verbose(env, "invalid access to flow keys off=%d size=%d\n",
5348 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
5349 u32 regno, int off, int size,
5350 enum bpf_access_type t)
5352 struct bpf_reg_state *regs = cur_regs(env);
5353 struct bpf_reg_state *reg = ®s[regno];
5354 struct bpf_insn_access_aux info = {};
5357 if (reg->smin_value < 0) {
5358 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
5363 switch (reg->type) {
5364 case PTR_TO_SOCK_COMMON:
5365 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
5368 valid = bpf_sock_is_valid_access(off, size, t, &info);
5370 case PTR_TO_TCP_SOCK:
5371 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
5373 case PTR_TO_XDP_SOCK:
5374 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
5382 env->insn_aux_data[insn_idx].ctx_field_size =
5383 info.ctx_field_size;
5387 verbose(env, "R%d invalid %s access off=%d size=%d\n",
5388 regno, reg_type_str(env, reg->type), off, size);
5393 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
5395 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
5398 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
5400 const struct bpf_reg_state *reg = reg_state(env, regno);
5402 return reg->type == PTR_TO_CTX;
5405 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
5407 const struct bpf_reg_state *reg = reg_state(env, regno);
5409 return type_is_sk_pointer(reg->type);
5412 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
5414 const struct bpf_reg_state *reg = reg_state(env, regno);
5416 return type_is_pkt_pointer(reg->type);
5419 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
5421 const struct bpf_reg_state *reg = reg_state(env, regno);
5423 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
5424 return reg->type == PTR_TO_FLOW_KEYS;
5427 static u32 *reg2btf_ids[__BPF_REG_TYPE_MAX] = {
5429 [PTR_TO_SOCKET] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK],
5430 [PTR_TO_SOCK_COMMON] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
5431 [PTR_TO_TCP_SOCK] = &btf_sock_ids[BTF_SOCK_TYPE_TCP],
5433 [CONST_PTR_TO_MAP] = btf_bpf_map_id,
5436 static bool is_trusted_reg(const struct bpf_reg_state *reg)
5438 /* A referenced register is always trusted. */
5439 if (reg->ref_obj_id)
5442 /* Types listed in the reg2btf_ids are always trusted */
5443 if (reg2btf_ids[base_type(reg->type)])
5446 /* If a register is not referenced, it is trusted if it has the
5447 * MEM_ALLOC or PTR_TRUSTED type modifiers, and no others. Some of the
5448 * other type modifiers may be safe, but we elect to take an opt-in
5449 * approach here as some (e.g. PTR_UNTRUSTED and PTR_MAYBE_NULL) are
5452 * Eventually, we should make PTR_TRUSTED the single source of truth
5453 * for whether a register is trusted.
5455 return type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS &&
5456 !bpf_type_has_unsafe_modifiers(reg->type);
5459 static bool is_rcu_reg(const struct bpf_reg_state *reg)
5461 return reg->type & MEM_RCU;
5464 static void clear_trusted_flags(enum bpf_type_flag *flag)
5466 *flag &= ~(BPF_REG_TRUSTED_MODIFIERS | MEM_RCU);
5469 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
5470 const struct bpf_reg_state *reg,
5471 int off, int size, bool strict)
5473 struct tnum reg_off;
5476 /* Byte size accesses are always allowed. */
5477 if (!strict || size == 1)
5480 /* For platforms that do not have a Kconfig enabling
5481 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
5482 * NET_IP_ALIGN is universally set to '2'. And on platforms
5483 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
5484 * to this code only in strict mode where we want to emulate
5485 * the NET_IP_ALIGN==2 checking. Therefore use an
5486 * unconditional IP align value of '2'.
5490 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
5491 if (!tnum_is_aligned(reg_off, size)) {
5494 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5496 "misaligned packet access off %d+%s+%d+%d size %d\n",
5497 ip_align, tn_buf, reg->off, off, size);
5504 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
5505 const struct bpf_reg_state *reg,
5506 const char *pointer_desc,
5507 int off, int size, bool strict)
5509 struct tnum reg_off;
5511 /* Byte size accesses are always allowed. */
5512 if (!strict || size == 1)
5515 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
5516 if (!tnum_is_aligned(reg_off, size)) {
5519 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5520 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
5521 pointer_desc, tn_buf, reg->off, off, size);
5528 static int check_ptr_alignment(struct bpf_verifier_env *env,
5529 const struct bpf_reg_state *reg, int off,
5530 int size, bool strict_alignment_once)
5532 bool strict = env->strict_alignment || strict_alignment_once;
5533 const char *pointer_desc = "";
5535 switch (reg->type) {
5537 case PTR_TO_PACKET_META:
5538 /* Special case, because of NET_IP_ALIGN. Given metadata sits
5539 * right in front, treat it the very same way.
5541 return check_pkt_ptr_alignment(env, reg, off, size, strict);
5542 case PTR_TO_FLOW_KEYS:
5543 pointer_desc = "flow keys ";
5545 case PTR_TO_MAP_KEY:
5546 pointer_desc = "key ";
5548 case PTR_TO_MAP_VALUE:
5549 pointer_desc = "value ";
5552 pointer_desc = "context ";
5555 pointer_desc = "stack ";
5556 /* The stack spill tracking logic in check_stack_write_fixed_off()
5557 * and check_stack_read_fixed_off() relies on stack accesses being
5563 pointer_desc = "sock ";
5565 case PTR_TO_SOCK_COMMON:
5566 pointer_desc = "sock_common ";
5568 case PTR_TO_TCP_SOCK:
5569 pointer_desc = "tcp_sock ";
5571 case PTR_TO_XDP_SOCK:
5572 pointer_desc = "xdp_sock ";
5577 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
5581 static int update_stack_depth(struct bpf_verifier_env *env,
5582 const struct bpf_func_state *func,
5585 u16 stack = env->subprog_info[func->subprogno].stack_depth;
5590 /* update known max for given subprogram */
5591 env->subprog_info[func->subprogno].stack_depth = -off;
5595 /* starting from main bpf function walk all instructions of the function
5596 * and recursively walk all callees that given function can call.
5597 * Ignore jump and exit insns.
5598 * Since recursion is prevented by check_cfg() this algorithm
5599 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
5601 static int check_max_stack_depth_subprog(struct bpf_verifier_env *env, int idx)
5603 struct bpf_subprog_info *subprog = env->subprog_info;
5604 struct bpf_insn *insn = env->prog->insnsi;
5605 int depth = 0, frame = 0, i, subprog_end;
5606 bool tail_call_reachable = false;
5607 int ret_insn[MAX_CALL_FRAMES];
5608 int ret_prog[MAX_CALL_FRAMES];
5611 i = subprog[idx].start;
5613 /* protect against potential stack overflow that might happen when
5614 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
5615 * depth for such case down to 256 so that the worst case scenario
5616 * would result in 8k stack size (32 which is tailcall limit * 256 =
5619 * To get the idea what might happen, see an example:
5620 * func1 -> sub rsp, 128
5621 * subfunc1 -> sub rsp, 256
5622 * tailcall1 -> add rsp, 256
5623 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
5624 * subfunc2 -> sub rsp, 64
5625 * subfunc22 -> sub rsp, 128
5626 * tailcall2 -> add rsp, 128
5627 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
5629 * tailcall will unwind the current stack frame but it will not get rid
5630 * of caller's stack as shown on the example above.
5632 if (idx && subprog[idx].has_tail_call && depth >= 256) {
5634 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
5638 /* round up to 32-bytes, since this is granularity
5639 * of interpreter stack size
5641 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
5642 if (depth > MAX_BPF_STACK) {
5643 verbose(env, "combined stack size of %d calls is %d. Too large\n",
5648 subprog_end = subprog[idx + 1].start;
5649 for (; i < subprog_end; i++) {
5650 int next_insn, sidx;
5652 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
5654 /* remember insn and function to return to */
5655 ret_insn[frame] = i + 1;
5656 ret_prog[frame] = idx;
5658 /* find the callee */
5659 next_insn = i + insn[i].imm + 1;
5660 sidx = find_subprog(env, next_insn);
5662 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5666 if (subprog[sidx].is_async_cb) {
5667 if (subprog[sidx].has_tail_call) {
5668 verbose(env, "verifier bug. subprog has tail_call and async cb\n");
5671 /* async callbacks don't increase bpf prog stack size unless called directly */
5672 if (!bpf_pseudo_call(insn + i))
5678 if (subprog[idx].has_tail_call)
5679 tail_call_reachable = true;
5682 if (frame >= MAX_CALL_FRAMES) {
5683 verbose(env, "the call stack of %d frames is too deep !\n",
5689 /* if tail call got detected across bpf2bpf calls then mark each of the
5690 * currently present subprog frames as tail call reachable subprogs;
5691 * this info will be utilized by JIT so that we will be preserving the
5692 * tail call counter throughout bpf2bpf calls combined with tailcalls
5694 if (tail_call_reachable)
5695 for (j = 0; j < frame; j++)
5696 subprog[ret_prog[j]].tail_call_reachable = true;
5697 if (subprog[0].tail_call_reachable)
5698 env->prog->aux->tail_call_reachable = true;
5700 /* end of for() loop means the last insn of the 'subprog'
5701 * was reached. Doesn't matter whether it was JA or EXIT
5705 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
5707 i = ret_insn[frame];
5708 idx = ret_prog[frame];
5712 static int check_max_stack_depth(struct bpf_verifier_env *env)
5714 struct bpf_subprog_info *si = env->subprog_info;
5717 for (int i = 0; i < env->subprog_cnt; i++) {
5718 if (!i || si[i].is_async_cb) {
5719 ret = check_max_stack_depth_subprog(env, i);
5728 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
5729 static int get_callee_stack_depth(struct bpf_verifier_env *env,
5730 const struct bpf_insn *insn, int idx)
5732 int start = idx + insn->imm + 1, subprog;
5734 subprog = find_subprog(env, start);
5736 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5740 return env->subprog_info[subprog].stack_depth;
5744 static int __check_buffer_access(struct bpf_verifier_env *env,
5745 const char *buf_info,
5746 const struct bpf_reg_state *reg,
5747 int regno, int off, int size)
5751 "R%d invalid %s buffer access: off=%d, size=%d\n",
5752 regno, buf_info, off, size);
5755 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
5758 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5760 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
5761 regno, off, tn_buf);
5768 static int check_tp_buffer_access(struct bpf_verifier_env *env,
5769 const struct bpf_reg_state *reg,
5770 int regno, int off, int size)
5774 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
5778 if (off + size > env->prog->aux->max_tp_access)
5779 env->prog->aux->max_tp_access = off + size;
5784 static int check_buffer_access(struct bpf_verifier_env *env,
5785 const struct bpf_reg_state *reg,
5786 int regno, int off, int size,
5787 bool zero_size_allowed,
5790 const char *buf_info = type_is_rdonly_mem(reg->type) ? "rdonly" : "rdwr";
5793 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
5797 if (off + size > *max_access)
5798 *max_access = off + size;
5803 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
5804 static void zext_32_to_64(struct bpf_reg_state *reg)
5806 reg->var_off = tnum_subreg(reg->var_off);
5807 __reg_assign_32_into_64(reg);
5810 /* truncate register to smaller size (in bytes)
5811 * must be called with size < BPF_REG_SIZE
5813 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
5817 /* clear high bits in bit representation */
5818 reg->var_off = tnum_cast(reg->var_off, size);
5820 /* fix arithmetic bounds */
5821 mask = ((u64)1 << (size * 8)) - 1;
5822 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
5823 reg->umin_value &= mask;
5824 reg->umax_value &= mask;
5826 reg->umin_value = 0;
5827 reg->umax_value = mask;
5829 reg->smin_value = reg->umin_value;
5830 reg->smax_value = reg->umax_value;
5832 /* If size is smaller than 32bit register the 32bit register
5833 * values are also truncated so we push 64-bit bounds into
5834 * 32-bit bounds. Above were truncated < 32-bits already.
5838 __reg_combine_64_into_32(reg);
5841 static void set_sext64_default_val(struct bpf_reg_state *reg, int size)
5844 reg->smin_value = reg->s32_min_value = S8_MIN;
5845 reg->smax_value = reg->s32_max_value = S8_MAX;
5846 } else if (size == 2) {
5847 reg->smin_value = reg->s32_min_value = S16_MIN;
5848 reg->smax_value = reg->s32_max_value = S16_MAX;
5851 reg->smin_value = reg->s32_min_value = S32_MIN;
5852 reg->smax_value = reg->s32_max_value = S32_MAX;
5854 reg->umin_value = reg->u32_min_value = 0;
5855 reg->umax_value = U64_MAX;
5856 reg->u32_max_value = U32_MAX;
5857 reg->var_off = tnum_unknown;
5860 static void coerce_reg_to_size_sx(struct bpf_reg_state *reg, int size)
5862 s64 init_s64_max, init_s64_min, s64_max, s64_min, u64_cval;
5863 u64 top_smax_value, top_smin_value;
5864 u64 num_bits = size * 8;
5866 if (tnum_is_const(reg->var_off)) {
5867 u64_cval = reg->var_off.value;
5869 reg->var_off = tnum_const((s8)u64_cval);
5871 reg->var_off = tnum_const((s16)u64_cval);
5874 reg->var_off = tnum_const((s32)u64_cval);
5876 u64_cval = reg->var_off.value;
5877 reg->smax_value = reg->smin_value = u64_cval;
5878 reg->umax_value = reg->umin_value = u64_cval;
5879 reg->s32_max_value = reg->s32_min_value = u64_cval;
5880 reg->u32_max_value = reg->u32_min_value = u64_cval;
5884 top_smax_value = ((u64)reg->smax_value >> num_bits) << num_bits;
5885 top_smin_value = ((u64)reg->smin_value >> num_bits) << num_bits;
5887 if (top_smax_value != top_smin_value)
5890 /* find the s64_min and s64_min after sign extension */
5892 init_s64_max = (s8)reg->smax_value;
5893 init_s64_min = (s8)reg->smin_value;
5894 } else if (size == 2) {
5895 init_s64_max = (s16)reg->smax_value;
5896 init_s64_min = (s16)reg->smin_value;
5898 init_s64_max = (s32)reg->smax_value;
5899 init_s64_min = (s32)reg->smin_value;
5902 s64_max = max(init_s64_max, init_s64_min);
5903 s64_min = min(init_s64_max, init_s64_min);
5905 /* both of s64_max/s64_min positive or negative */
5906 if ((s64_max >= 0) == (s64_min >= 0)) {
5907 reg->smin_value = reg->s32_min_value = s64_min;
5908 reg->smax_value = reg->s32_max_value = s64_max;
5909 reg->umin_value = reg->u32_min_value = s64_min;
5910 reg->umax_value = reg->u32_max_value = s64_max;
5911 reg->var_off = tnum_range(s64_min, s64_max);
5916 set_sext64_default_val(reg, size);
5919 static void set_sext32_default_val(struct bpf_reg_state *reg, int size)
5922 reg->s32_min_value = S8_MIN;
5923 reg->s32_max_value = S8_MAX;
5926 reg->s32_min_value = S16_MIN;
5927 reg->s32_max_value = S16_MAX;
5929 reg->u32_min_value = 0;
5930 reg->u32_max_value = U32_MAX;
5933 static void coerce_subreg_to_size_sx(struct bpf_reg_state *reg, int size)
5935 s32 init_s32_max, init_s32_min, s32_max, s32_min, u32_val;
5936 u32 top_smax_value, top_smin_value;
5937 u32 num_bits = size * 8;
5939 if (tnum_is_const(reg->var_off)) {
5940 u32_val = reg->var_off.value;
5942 reg->var_off = tnum_const((s8)u32_val);
5944 reg->var_off = tnum_const((s16)u32_val);
5946 u32_val = reg->var_off.value;
5947 reg->s32_min_value = reg->s32_max_value = u32_val;
5948 reg->u32_min_value = reg->u32_max_value = u32_val;
5952 top_smax_value = ((u32)reg->s32_max_value >> num_bits) << num_bits;
5953 top_smin_value = ((u32)reg->s32_min_value >> num_bits) << num_bits;
5955 if (top_smax_value != top_smin_value)
5958 /* find the s32_min and s32_min after sign extension */
5960 init_s32_max = (s8)reg->s32_max_value;
5961 init_s32_min = (s8)reg->s32_min_value;
5964 init_s32_max = (s16)reg->s32_max_value;
5965 init_s32_min = (s16)reg->s32_min_value;
5967 s32_max = max(init_s32_max, init_s32_min);
5968 s32_min = min(init_s32_max, init_s32_min);
5970 if ((s32_min >= 0) == (s32_max >= 0)) {
5971 reg->s32_min_value = s32_min;
5972 reg->s32_max_value = s32_max;
5973 reg->u32_min_value = (u32)s32_min;
5974 reg->u32_max_value = (u32)s32_max;
5979 set_sext32_default_val(reg, size);
5982 static bool bpf_map_is_rdonly(const struct bpf_map *map)
5984 /* A map is considered read-only if the following condition are true:
5986 * 1) BPF program side cannot change any of the map content. The
5987 * BPF_F_RDONLY_PROG flag is throughout the lifetime of a map
5988 * and was set at map creation time.
5989 * 2) The map value(s) have been initialized from user space by a
5990 * loader and then "frozen", such that no new map update/delete
5991 * operations from syscall side are possible for the rest of
5992 * the map's lifetime from that point onwards.
5993 * 3) Any parallel/pending map update/delete operations from syscall
5994 * side have been completed. Only after that point, it's safe to
5995 * assume that map value(s) are immutable.
5997 return (map->map_flags & BPF_F_RDONLY_PROG) &&
5998 READ_ONCE(map->frozen) &&
5999 !bpf_map_write_active(map);
6002 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val,
6009 err = map->ops->map_direct_value_addr(map, &addr, off);
6012 ptr = (void *)(long)addr + off;
6016 *val = is_ldsx ? (s64)*(s8 *)ptr : (u64)*(u8 *)ptr;
6019 *val = is_ldsx ? (s64)*(s16 *)ptr : (u64)*(u16 *)ptr;
6022 *val = is_ldsx ? (s64)*(s32 *)ptr : (u64)*(u32 *)ptr;
6033 #define BTF_TYPE_SAFE_RCU(__type) __PASTE(__type, __safe_rcu)
6034 #define BTF_TYPE_SAFE_RCU_OR_NULL(__type) __PASTE(__type, __safe_rcu_or_null)
6035 #define BTF_TYPE_SAFE_TRUSTED(__type) __PASTE(__type, __safe_trusted)
6038 * Allow list few fields as RCU trusted or full trusted.
6039 * This logic doesn't allow mix tagging and will be removed once GCC supports
6043 /* RCU trusted: these fields are trusted in RCU CS and never NULL */
6044 BTF_TYPE_SAFE_RCU(struct task_struct) {
6045 const cpumask_t *cpus_ptr;
6046 struct css_set __rcu *cgroups;
6047 struct task_struct __rcu *real_parent;
6048 struct task_struct *group_leader;
6051 BTF_TYPE_SAFE_RCU(struct cgroup) {
6052 /* cgrp->kn is always accessible as documented in kernel/cgroup/cgroup.c */
6053 struct kernfs_node *kn;
6056 BTF_TYPE_SAFE_RCU(struct css_set) {
6057 struct cgroup *dfl_cgrp;
6060 /* RCU trusted: these fields are trusted in RCU CS and can be NULL */
6061 BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct) {
6062 struct file __rcu *exe_file;
6065 /* skb->sk, req->sk are not RCU protected, but we mark them as such
6066 * because bpf prog accessible sockets are SOCK_RCU_FREE.
6068 BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff) {
6072 BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock) {
6076 /* full trusted: these fields are trusted even outside of RCU CS and never NULL */
6077 BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta) {
6078 struct seq_file *seq;
6081 BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task) {
6082 struct bpf_iter_meta *meta;
6083 struct task_struct *task;
6086 BTF_TYPE_SAFE_TRUSTED(struct linux_binprm) {
6090 BTF_TYPE_SAFE_TRUSTED(struct file) {
6091 struct inode *f_inode;
6094 BTF_TYPE_SAFE_TRUSTED(struct dentry) {
6095 /* no negative dentry-s in places where bpf can see it */
6096 struct inode *d_inode;
6099 BTF_TYPE_SAFE_TRUSTED(struct socket) {
6103 static bool type_is_rcu(struct bpf_verifier_env *env,
6104 struct bpf_reg_state *reg,
6105 const char *field_name, u32 btf_id)
6107 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct task_struct));
6108 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct cgroup));
6109 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct css_set));
6111 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_rcu");
6114 static bool type_is_rcu_or_null(struct bpf_verifier_env *env,
6115 struct bpf_reg_state *reg,
6116 const char *field_name, u32 btf_id)
6118 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct));
6119 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff));
6120 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock));
6122 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_rcu_or_null");
6125 static bool type_is_trusted(struct bpf_verifier_env *env,
6126 struct bpf_reg_state *reg,
6127 const char *field_name, u32 btf_id)
6129 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta));
6130 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task));
6131 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct linux_binprm));
6132 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct file));
6133 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct dentry));
6134 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct socket));
6136 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_trusted");
6139 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
6140 struct bpf_reg_state *regs,
6141 int regno, int off, int size,
6142 enum bpf_access_type atype,
6145 struct bpf_reg_state *reg = regs + regno;
6146 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
6147 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
6148 const char *field_name = NULL;
6149 enum bpf_type_flag flag = 0;
6153 if (!env->allow_ptr_leaks) {
6155 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
6159 if (!env->prog->gpl_compatible && btf_is_kernel(reg->btf)) {
6161 "Cannot access kernel 'struct %s' from non-GPL compatible program\n",
6167 "R%d is ptr_%s invalid negative access: off=%d\n",
6171 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
6174 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6176 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
6177 regno, tname, off, tn_buf);
6181 if (reg->type & MEM_USER) {
6183 "R%d is ptr_%s access user memory: off=%d\n",
6188 if (reg->type & MEM_PERCPU) {
6190 "R%d is ptr_%s access percpu memory: off=%d\n",
6195 if (env->ops->btf_struct_access && !type_is_alloc(reg->type) && atype == BPF_WRITE) {
6196 if (!btf_is_kernel(reg->btf)) {
6197 verbose(env, "verifier internal error: reg->btf must be kernel btf\n");
6200 ret = env->ops->btf_struct_access(&env->log, reg, off, size);
6202 /* Writes are permitted with default btf_struct_access for
6203 * program allocated objects (which always have ref_obj_id > 0),
6204 * but not for untrusted PTR_TO_BTF_ID | MEM_ALLOC.
6206 if (atype != BPF_READ && !type_is_ptr_alloc_obj(reg->type)) {
6207 verbose(env, "only read is supported\n");
6211 if (type_is_alloc(reg->type) && !type_is_non_owning_ref(reg->type) &&
6213 verbose(env, "verifier internal error: ref_obj_id for allocated object must be non-zero\n");
6217 ret = btf_struct_access(&env->log, reg, off, size, atype, &btf_id, &flag, &field_name);
6223 if (ret != PTR_TO_BTF_ID) {
6226 } else if (type_flag(reg->type) & PTR_UNTRUSTED) {
6227 /* If this is an untrusted pointer, all pointers formed by walking it
6228 * also inherit the untrusted flag.
6230 flag = PTR_UNTRUSTED;
6232 } else if (is_trusted_reg(reg) || is_rcu_reg(reg)) {
6233 /* By default any pointer obtained from walking a trusted pointer is no
6234 * longer trusted, unless the field being accessed has explicitly been
6235 * marked as inheriting its parent's state of trust (either full or RCU).
6237 * 'cgroups' pointer is untrusted if task->cgroups dereference
6238 * happened in a sleepable program outside of bpf_rcu_read_lock()
6239 * section. In a non-sleepable program it's trusted while in RCU CS (aka MEM_RCU).
6240 * Note bpf_rcu_read_unlock() converts MEM_RCU pointers to PTR_UNTRUSTED.
6242 * A regular RCU-protected pointer with __rcu tag can also be deemed
6243 * trusted if we are in an RCU CS. Such pointer can be NULL.
6245 if (type_is_trusted(env, reg, field_name, btf_id)) {
6246 flag |= PTR_TRUSTED;
6247 } else if (in_rcu_cs(env) && !type_may_be_null(reg->type)) {
6248 if (type_is_rcu(env, reg, field_name, btf_id)) {
6249 /* ignore __rcu tag and mark it MEM_RCU */
6251 } else if (flag & MEM_RCU ||
6252 type_is_rcu_or_null(env, reg, field_name, btf_id)) {
6253 /* __rcu tagged pointers can be NULL */
6254 flag |= MEM_RCU | PTR_MAYBE_NULL;
6256 /* We always trust them */
6257 if (type_is_rcu_or_null(env, reg, field_name, btf_id) &&
6258 flag & PTR_UNTRUSTED)
6259 flag &= ~PTR_UNTRUSTED;
6260 } else if (flag & (MEM_PERCPU | MEM_USER)) {
6263 /* walking unknown pointers yields old deprecated PTR_TO_BTF_ID */
6264 clear_trusted_flags(&flag);
6268 * If not in RCU CS or MEM_RCU pointer can be NULL then
6269 * aggressively mark as untrusted otherwise such
6270 * pointers will be plain PTR_TO_BTF_ID without flags
6271 * and will be allowed to be passed into helpers for
6274 flag = PTR_UNTRUSTED;
6277 /* Old compat. Deprecated */
6278 clear_trusted_flags(&flag);
6281 if (atype == BPF_READ && value_regno >= 0)
6282 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id, flag);
6287 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
6288 struct bpf_reg_state *regs,
6289 int regno, int off, int size,
6290 enum bpf_access_type atype,
6293 struct bpf_reg_state *reg = regs + regno;
6294 struct bpf_map *map = reg->map_ptr;
6295 struct bpf_reg_state map_reg;
6296 enum bpf_type_flag flag = 0;
6297 const struct btf_type *t;
6303 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
6307 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
6308 verbose(env, "map_ptr access not supported for map type %d\n",
6313 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
6314 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
6316 if (!env->allow_ptr_leaks) {
6318 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
6324 verbose(env, "R%d is %s invalid negative access: off=%d\n",
6329 if (atype != BPF_READ) {
6330 verbose(env, "only read from %s is supported\n", tname);
6334 /* Simulate access to a PTR_TO_BTF_ID */
6335 memset(&map_reg, 0, sizeof(map_reg));
6336 mark_btf_ld_reg(env, &map_reg, 0, PTR_TO_BTF_ID, btf_vmlinux, *map->ops->map_btf_id, 0);
6337 ret = btf_struct_access(&env->log, &map_reg, off, size, atype, &btf_id, &flag, NULL);
6341 if (value_regno >= 0)
6342 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id, flag);
6347 /* Check that the stack access at the given offset is within bounds. The
6348 * maximum valid offset is -1.
6350 * The minimum valid offset is -MAX_BPF_STACK for writes, and
6351 * -state->allocated_stack for reads.
6353 static int check_stack_slot_within_bounds(int off,
6354 struct bpf_func_state *state,
6355 enum bpf_access_type t)
6360 min_valid_off = -MAX_BPF_STACK;
6362 min_valid_off = -state->allocated_stack;
6364 if (off < min_valid_off || off > -1)
6369 /* Check that the stack access at 'regno + off' falls within the maximum stack
6372 * 'off' includes `regno->offset`, but not its dynamic part (if any).
6374 static int check_stack_access_within_bounds(
6375 struct bpf_verifier_env *env,
6376 int regno, int off, int access_size,
6377 enum bpf_access_src src, enum bpf_access_type type)
6379 struct bpf_reg_state *regs = cur_regs(env);
6380 struct bpf_reg_state *reg = regs + regno;
6381 struct bpf_func_state *state = func(env, reg);
6382 int min_off, max_off;
6386 if (src == ACCESS_HELPER)
6387 /* We don't know if helpers are reading or writing (or both). */
6388 err_extra = " indirect access to";
6389 else if (type == BPF_READ)
6390 err_extra = " read from";
6392 err_extra = " write to";
6394 if (tnum_is_const(reg->var_off)) {
6395 min_off = reg->var_off.value + off;
6396 if (access_size > 0)
6397 max_off = min_off + access_size - 1;
6401 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
6402 reg->smin_value <= -BPF_MAX_VAR_OFF) {
6403 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
6407 min_off = reg->smin_value + off;
6408 if (access_size > 0)
6409 max_off = reg->smax_value + off + access_size - 1;
6414 err = check_stack_slot_within_bounds(min_off, state, type);
6416 err = check_stack_slot_within_bounds(max_off, state, type);
6419 if (tnum_is_const(reg->var_off)) {
6420 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
6421 err_extra, regno, off, access_size);
6425 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6426 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
6427 err_extra, regno, tn_buf, access_size);
6433 /* check whether memory at (regno + off) is accessible for t = (read | write)
6434 * if t==write, value_regno is a register which value is stored into memory
6435 * if t==read, value_regno is a register which will receive the value from memory
6436 * if t==write && value_regno==-1, some unknown value is stored into memory
6437 * if t==read && value_regno==-1, don't care what we read from memory
6439 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
6440 int off, int bpf_size, enum bpf_access_type t,
6441 int value_regno, bool strict_alignment_once, bool is_ldsx)
6443 struct bpf_reg_state *regs = cur_regs(env);
6444 struct bpf_reg_state *reg = regs + regno;
6445 struct bpf_func_state *state;
6448 size = bpf_size_to_bytes(bpf_size);
6452 /* alignment checks will add in reg->off themselves */
6453 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
6457 /* for access checks, reg->off is just part of off */
6460 if (reg->type == PTR_TO_MAP_KEY) {
6461 if (t == BPF_WRITE) {
6462 verbose(env, "write to change key R%d not allowed\n", regno);
6466 err = check_mem_region_access(env, regno, off, size,
6467 reg->map_ptr->key_size, false);
6470 if (value_regno >= 0)
6471 mark_reg_unknown(env, regs, value_regno);
6472 } else if (reg->type == PTR_TO_MAP_VALUE) {
6473 struct btf_field *kptr_field = NULL;
6475 if (t == BPF_WRITE && value_regno >= 0 &&
6476 is_pointer_value(env, value_regno)) {
6477 verbose(env, "R%d leaks addr into map\n", value_regno);
6480 err = check_map_access_type(env, regno, off, size, t);
6483 err = check_map_access(env, regno, off, size, false, ACCESS_DIRECT);
6486 if (tnum_is_const(reg->var_off))
6487 kptr_field = btf_record_find(reg->map_ptr->record,
6488 off + reg->var_off.value, BPF_KPTR);
6490 err = check_map_kptr_access(env, regno, value_regno, insn_idx, kptr_field);
6491 } else if (t == BPF_READ && value_regno >= 0) {
6492 struct bpf_map *map = reg->map_ptr;
6494 /* if map is read-only, track its contents as scalars */
6495 if (tnum_is_const(reg->var_off) &&
6496 bpf_map_is_rdonly(map) &&
6497 map->ops->map_direct_value_addr) {
6498 int map_off = off + reg->var_off.value;
6501 err = bpf_map_direct_read(map, map_off, size,
6506 regs[value_regno].type = SCALAR_VALUE;
6507 __mark_reg_known(®s[value_regno], val);
6509 mark_reg_unknown(env, regs, value_regno);
6512 } else if (base_type(reg->type) == PTR_TO_MEM) {
6513 bool rdonly_mem = type_is_rdonly_mem(reg->type);
6515 if (type_may_be_null(reg->type)) {
6516 verbose(env, "R%d invalid mem access '%s'\n", regno,
6517 reg_type_str(env, reg->type));
6521 if (t == BPF_WRITE && rdonly_mem) {
6522 verbose(env, "R%d cannot write into %s\n",
6523 regno, reg_type_str(env, reg->type));
6527 if (t == BPF_WRITE && value_regno >= 0 &&
6528 is_pointer_value(env, value_regno)) {
6529 verbose(env, "R%d leaks addr into mem\n", value_regno);
6533 err = check_mem_region_access(env, regno, off, size,
6534 reg->mem_size, false);
6535 if (!err && value_regno >= 0 && (t == BPF_READ || rdonly_mem))
6536 mark_reg_unknown(env, regs, value_regno);
6537 } else if (reg->type == PTR_TO_CTX) {
6538 enum bpf_reg_type reg_type = SCALAR_VALUE;
6539 struct btf *btf = NULL;
6542 if (t == BPF_WRITE && value_regno >= 0 &&
6543 is_pointer_value(env, value_regno)) {
6544 verbose(env, "R%d leaks addr into ctx\n", value_regno);
6548 err = check_ptr_off_reg(env, reg, regno);
6552 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf,
6555 verbose_linfo(env, insn_idx, "; ");
6556 if (!err && t == BPF_READ && value_regno >= 0) {
6557 /* ctx access returns either a scalar, or a
6558 * PTR_TO_PACKET[_META,_END]. In the latter
6559 * case, we know the offset is zero.
6561 if (reg_type == SCALAR_VALUE) {
6562 mark_reg_unknown(env, regs, value_regno);
6564 mark_reg_known_zero(env, regs,
6566 if (type_may_be_null(reg_type))
6567 regs[value_regno].id = ++env->id_gen;
6568 /* A load of ctx field could have different
6569 * actual load size with the one encoded in the
6570 * insn. When the dst is PTR, it is for sure not
6573 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
6574 if (base_type(reg_type) == PTR_TO_BTF_ID) {
6575 regs[value_regno].btf = btf;
6576 regs[value_regno].btf_id = btf_id;
6579 regs[value_regno].type = reg_type;
6582 } else if (reg->type == PTR_TO_STACK) {
6583 /* Basic bounds checks. */
6584 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
6588 state = func(env, reg);
6589 err = update_stack_depth(env, state, off);
6594 err = check_stack_read(env, regno, off, size,
6597 err = check_stack_write(env, regno, off, size,
6598 value_regno, insn_idx);
6599 } else if (reg_is_pkt_pointer(reg)) {
6600 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
6601 verbose(env, "cannot write into packet\n");
6604 if (t == BPF_WRITE && value_regno >= 0 &&
6605 is_pointer_value(env, value_regno)) {
6606 verbose(env, "R%d leaks addr into packet\n",
6610 err = check_packet_access(env, regno, off, size, false);
6611 if (!err && t == BPF_READ && value_regno >= 0)
6612 mark_reg_unknown(env, regs, value_regno);
6613 } else if (reg->type == PTR_TO_FLOW_KEYS) {
6614 if (t == BPF_WRITE && value_regno >= 0 &&
6615 is_pointer_value(env, value_regno)) {
6616 verbose(env, "R%d leaks addr into flow keys\n",
6621 err = check_flow_keys_access(env, off, size);
6622 if (!err && t == BPF_READ && value_regno >= 0)
6623 mark_reg_unknown(env, regs, value_regno);
6624 } else if (type_is_sk_pointer(reg->type)) {
6625 if (t == BPF_WRITE) {
6626 verbose(env, "R%d cannot write into %s\n",
6627 regno, reg_type_str(env, reg->type));
6630 err = check_sock_access(env, insn_idx, regno, off, size, t);
6631 if (!err && value_regno >= 0)
6632 mark_reg_unknown(env, regs, value_regno);
6633 } else if (reg->type == PTR_TO_TP_BUFFER) {
6634 err = check_tp_buffer_access(env, reg, regno, off, size);
6635 if (!err && t == BPF_READ && value_regno >= 0)
6636 mark_reg_unknown(env, regs, value_regno);
6637 } else if (base_type(reg->type) == PTR_TO_BTF_ID &&
6638 !type_may_be_null(reg->type)) {
6639 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
6641 } else if (reg->type == CONST_PTR_TO_MAP) {
6642 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
6644 } else if (base_type(reg->type) == PTR_TO_BUF) {
6645 bool rdonly_mem = type_is_rdonly_mem(reg->type);
6649 if (t == BPF_WRITE) {
6650 verbose(env, "R%d cannot write into %s\n",
6651 regno, reg_type_str(env, reg->type));
6654 max_access = &env->prog->aux->max_rdonly_access;
6656 max_access = &env->prog->aux->max_rdwr_access;
6659 err = check_buffer_access(env, reg, regno, off, size, false,
6662 if (!err && value_regno >= 0 && (rdonly_mem || t == BPF_READ))
6663 mark_reg_unknown(env, regs, value_regno);
6665 verbose(env, "R%d invalid mem access '%s'\n", regno,
6666 reg_type_str(env, reg->type));
6670 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
6671 regs[value_regno].type == SCALAR_VALUE) {
6673 /* b/h/w load zero-extends, mark upper bits as known 0 */
6674 coerce_reg_to_size(®s[value_regno], size);
6676 coerce_reg_to_size_sx(®s[value_regno], size);
6681 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
6686 switch (insn->imm) {
6688 case BPF_ADD | BPF_FETCH:
6690 case BPF_AND | BPF_FETCH:
6692 case BPF_OR | BPF_FETCH:
6694 case BPF_XOR | BPF_FETCH:
6699 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
6703 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
6704 verbose(env, "invalid atomic operand size\n");
6708 /* check src1 operand */
6709 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6713 /* check src2 operand */
6714 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6718 if (insn->imm == BPF_CMPXCHG) {
6719 /* Check comparison of R0 with memory location */
6720 const u32 aux_reg = BPF_REG_0;
6722 err = check_reg_arg(env, aux_reg, SRC_OP);
6726 if (is_pointer_value(env, aux_reg)) {
6727 verbose(env, "R%d leaks addr into mem\n", aux_reg);
6732 if (is_pointer_value(env, insn->src_reg)) {
6733 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
6737 if (is_ctx_reg(env, insn->dst_reg) ||
6738 is_pkt_reg(env, insn->dst_reg) ||
6739 is_flow_key_reg(env, insn->dst_reg) ||
6740 is_sk_reg(env, insn->dst_reg)) {
6741 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
6743 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
6747 if (insn->imm & BPF_FETCH) {
6748 if (insn->imm == BPF_CMPXCHG)
6749 load_reg = BPF_REG_0;
6751 load_reg = insn->src_reg;
6753 /* check and record load of old value */
6754 err = check_reg_arg(env, load_reg, DST_OP);
6758 /* This instruction accesses a memory location but doesn't
6759 * actually load it into a register.
6764 /* Check whether we can read the memory, with second call for fetch
6765 * case to simulate the register fill.
6767 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
6768 BPF_SIZE(insn->code), BPF_READ, -1, true, false);
6769 if (!err && load_reg >= 0)
6770 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
6771 BPF_SIZE(insn->code), BPF_READ, load_reg,
6776 /* Check whether we can write into the same memory. */
6777 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
6778 BPF_SIZE(insn->code), BPF_WRITE, -1, true, false);
6785 /* When register 'regno' is used to read the stack (either directly or through
6786 * a helper function) make sure that it's within stack boundary and, depending
6787 * on the access type, that all elements of the stack are initialized.
6789 * 'off' includes 'regno->off', but not its dynamic part (if any).
6791 * All registers that have been spilled on the stack in the slots within the
6792 * read offsets are marked as read.
6794 static int check_stack_range_initialized(
6795 struct bpf_verifier_env *env, int regno, int off,
6796 int access_size, bool zero_size_allowed,
6797 enum bpf_access_src type, struct bpf_call_arg_meta *meta)
6799 struct bpf_reg_state *reg = reg_state(env, regno);
6800 struct bpf_func_state *state = func(env, reg);
6801 int err, min_off, max_off, i, j, slot, spi;
6802 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
6803 enum bpf_access_type bounds_check_type;
6804 /* Some accesses can write anything into the stack, others are
6807 bool clobber = false;
6809 if (access_size == 0 && !zero_size_allowed) {
6810 verbose(env, "invalid zero-sized read\n");
6814 if (type == ACCESS_HELPER) {
6815 /* The bounds checks for writes are more permissive than for
6816 * reads. However, if raw_mode is not set, we'll do extra
6819 bounds_check_type = BPF_WRITE;
6822 bounds_check_type = BPF_READ;
6824 err = check_stack_access_within_bounds(env, regno, off, access_size,
6825 type, bounds_check_type);
6830 if (tnum_is_const(reg->var_off)) {
6831 min_off = max_off = reg->var_off.value + off;
6833 /* Variable offset is prohibited for unprivileged mode for
6834 * simplicity since it requires corresponding support in
6835 * Spectre masking for stack ALU.
6836 * See also retrieve_ptr_limit().
6838 if (!env->bypass_spec_v1) {
6841 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6842 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
6843 regno, err_extra, tn_buf);
6846 /* Only initialized buffer on stack is allowed to be accessed
6847 * with variable offset. With uninitialized buffer it's hard to
6848 * guarantee that whole memory is marked as initialized on
6849 * helper return since specific bounds are unknown what may
6850 * cause uninitialized stack leaking.
6852 if (meta && meta->raw_mode)
6855 min_off = reg->smin_value + off;
6856 max_off = reg->smax_value + off;
6859 if (meta && meta->raw_mode) {
6860 /* Ensure we won't be overwriting dynptrs when simulating byte
6861 * by byte access in check_helper_call using meta.access_size.
6862 * This would be a problem if we have a helper in the future
6865 * helper(uninit_mem, len, dynptr)
6867 * Now, uninint_mem may overlap with dynptr pointer. Hence, it
6868 * may end up writing to dynptr itself when touching memory from
6869 * arg 1. This can be relaxed on a case by case basis for known
6870 * safe cases, but reject due to the possibilitiy of aliasing by
6873 for (i = min_off; i < max_off + access_size; i++) {
6874 int stack_off = -i - 1;
6877 /* raw_mode may write past allocated_stack */
6878 if (state->allocated_stack <= stack_off)
6880 if (state->stack[spi].slot_type[stack_off % BPF_REG_SIZE] == STACK_DYNPTR) {
6881 verbose(env, "potential write to dynptr at off=%d disallowed\n", i);
6885 meta->access_size = access_size;
6886 meta->regno = regno;
6890 for (i = min_off; i < max_off + access_size; i++) {
6894 spi = slot / BPF_REG_SIZE;
6895 if (state->allocated_stack <= slot)
6897 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
6898 if (*stype == STACK_MISC)
6900 if ((*stype == STACK_ZERO) ||
6901 (*stype == STACK_INVALID && env->allow_uninit_stack)) {
6903 /* helper can write anything into the stack */
6904 *stype = STACK_MISC;
6909 if (is_spilled_reg(&state->stack[spi]) &&
6910 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
6911 env->allow_ptr_leaks)) {
6913 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
6914 for (j = 0; j < BPF_REG_SIZE; j++)
6915 scrub_spilled_slot(&state->stack[spi].slot_type[j]);
6921 if (tnum_is_const(reg->var_off)) {
6922 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
6923 err_extra, regno, min_off, i - min_off, access_size);
6927 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6928 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
6929 err_extra, regno, tn_buf, i - min_off, access_size);
6933 /* reading any byte out of 8-byte 'spill_slot' will cause
6934 * the whole slot to be marked as 'read'
6936 mark_reg_read(env, &state->stack[spi].spilled_ptr,
6937 state->stack[spi].spilled_ptr.parent,
6939 /* We do not set REG_LIVE_WRITTEN for stack slot, as we can not
6940 * be sure that whether stack slot is written to or not. Hence,
6941 * we must still conservatively propagate reads upwards even if
6942 * helper may write to the entire memory range.
6945 return update_stack_depth(env, state, min_off);
6948 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
6949 int access_size, bool zero_size_allowed,
6950 struct bpf_call_arg_meta *meta)
6952 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
6955 switch (base_type(reg->type)) {
6957 case PTR_TO_PACKET_META:
6958 return check_packet_access(env, regno, reg->off, access_size,
6960 case PTR_TO_MAP_KEY:
6961 if (meta && meta->raw_mode) {
6962 verbose(env, "R%d cannot write into %s\n", regno,
6963 reg_type_str(env, reg->type));
6966 return check_mem_region_access(env, regno, reg->off, access_size,
6967 reg->map_ptr->key_size, false);
6968 case PTR_TO_MAP_VALUE:
6969 if (check_map_access_type(env, regno, reg->off, access_size,
6970 meta && meta->raw_mode ? BPF_WRITE :
6973 return check_map_access(env, regno, reg->off, access_size,
6974 zero_size_allowed, ACCESS_HELPER);
6976 if (type_is_rdonly_mem(reg->type)) {
6977 if (meta && meta->raw_mode) {
6978 verbose(env, "R%d cannot write into %s\n", regno,
6979 reg_type_str(env, reg->type));
6983 return check_mem_region_access(env, regno, reg->off,
6984 access_size, reg->mem_size,
6987 if (type_is_rdonly_mem(reg->type)) {
6988 if (meta && meta->raw_mode) {
6989 verbose(env, "R%d cannot write into %s\n", regno,
6990 reg_type_str(env, reg->type));
6994 max_access = &env->prog->aux->max_rdonly_access;
6996 max_access = &env->prog->aux->max_rdwr_access;
6998 return check_buffer_access(env, reg, regno, reg->off,
6999 access_size, zero_size_allowed,
7002 return check_stack_range_initialized(
7004 regno, reg->off, access_size,
7005 zero_size_allowed, ACCESS_HELPER, meta);
7007 return check_ptr_to_btf_access(env, regs, regno, reg->off,
7008 access_size, BPF_READ, -1);
7010 /* in case the function doesn't know how to access the context,
7011 * (because we are in a program of type SYSCALL for example), we
7012 * can not statically check its size.
7013 * Dynamically check it now.
7015 if (!env->ops->convert_ctx_access) {
7016 enum bpf_access_type atype = meta && meta->raw_mode ? BPF_WRITE : BPF_READ;
7017 int offset = access_size - 1;
7019 /* Allow zero-byte read from PTR_TO_CTX */
7020 if (access_size == 0)
7021 return zero_size_allowed ? 0 : -EACCES;
7023 return check_mem_access(env, env->insn_idx, regno, offset, BPF_B,
7024 atype, -1, false, false);
7028 default: /* scalar_value or invalid ptr */
7029 /* Allow zero-byte read from NULL, regardless of pointer type */
7030 if (zero_size_allowed && access_size == 0 &&
7031 register_is_null(reg))
7034 verbose(env, "R%d type=%s ", regno,
7035 reg_type_str(env, reg->type));
7036 verbose(env, "expected=%s\n", reg_type_str(env, PTR_TO_STACK));
7041 static int check_mem_size_reg(struct bpf_verifier_env *env,
7042 struct bpf_reg_state *reg, u32 regno,
7043 bool zero_size_allowed,
7044 struct bpf_call_arg_meta *meta)
7048 /* This is used to refine r0 return value bounds for helpers
7049 * that enforce this value as an upper bound on return values.
7050 * See do_refine_retval_range() for helpers that can refine
7051 * the return value. C type of helper is u32 so we pull register
7052 * bound from umax_value however, if negative verifier errors
7053 * out. Only upper bounds can be learned because retval is an
7054 * int type and negative retvals are allowed.
7056 meta->msize_max_value = reg->umax_value;
7058 /* The register is SCALAR_VALUE; the access check
7059 * happens using its boundaries.
7061 if (!tnum_is_const(reg->var_off))
7062 /* For unprivileged variable accesses, disable raw
7063 * mode so that the program is required to
7064 * initialize all the memory that the helper could
7065 * just partially fill up.
7069 if (reg->smin_value < 0) {
7070 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
7075 if (reg->umin_value == 0) {
7076 err = check_helper_mem_access(env, regno - 1, 0,
7083 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
7084 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
7088 err = check_helper_mem_access(env, regno - 1,
7090 zero_size_allowed, meta);
7092 err = mark_chain_precision(env, regno);
7096 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
7097 u32 regno, u32 mem_size)
7099 bool may_be_null = type_may_be_null(reg->type);
7100 struct bpf_reg_state saved_reg;
7101 struct bpf_call_arg_meta meta;
7104 if (register_is_null(reg))
7107 memset(&meta, 0, sizeof(meta));
7108 /* Assuming that the register contains a value check if the memory
7109 * access is safe. Temporarily save and restore the register's state as
7110 * the conversion shouldn't be visible to a caller.
7114 mark_ptr_not_null_reg(reg);
7117 err = check_helper_mem_access(env, regno, mem_size, true, &meta);
7118 /* Check access for BPF_WRITE */
7119 meta.raw_mode = true;
7120 err = err ?: check_helper_mem_access(env, regno, mem_size, true, &meta);
7128 static int check_kfunc_mem_size_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
7131 struct bpf_reg_state *mem_reg = &cur_regs(env)[regno - 1];
7132 bool may_be_null = type_may_be_null(mem_reg->type);
7133 struct bpf_reg_state saved_reg;
7134 struct bpf_call_arg_meta meta;
7137 WARN_ON_ONCE(regno < BPF_REG_2 || regno > BPF_REG_5);
7139 memset(&meta, 0, sizeof(meta));
7142 saved_reg = *mem_reg;
7143 mark_ptr_not_null_reg(mem_reg);
7146 err = check_mem_size_reg(env, reg, regno, true, &meta);
7147 /* Check access for BPF_WRITE */
7148 meta.raw_mode = true;
7149 err = err ?: check_mem_size_reg(env, reg, regno, true, &meta);
7152 *mem_reg = saved_reg;
7156 /* Implementation details:
7157 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL.
7158 * bpf_obj_new returns PTR_TO_BTF_ID | MEM_ALLOC | PTR_MAYBE_NULL.
7159 * Two bpf_map_lookups (even with the same key) will have different reg->id.
7160 * Two separate bpf_obj_new will also have different reg->id.
7161 * For traditional PTR_TO_MAP_VALUE or PTR_TO_BTF_ID | MEM_ALLOC, the verifier
7162 * clears reg->id after value_or_null->value transition, since the verifier only
7163 * cares about the range of access to valid map value pointer and doesn't care
7164 * about actual address of the map element.
7165 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
7166 * reg->id > 0 after value_or_null->value transition. By doing so
7167 * two bpf_map_lookups will be considered two different pointers that
7168 * point to different bpf_spin_locks. Likewise for pointers to allocated objects
7169 * returned from bpf_obj_new.
7170 * The verifier allows taking only one bpf_spin_lock at a time to avoid
7172 * Since only one bpf_spin_lock is allowed the checks are simpler than
7173 * reg_is_refcounted() logic. The verifier needs to remember only
7174 * one spin_lock instead of array of acquired_refs.
7175 * cur_state->active_lock remembers which map value element or allocated
7176 * object got locked and clears it after bpf_spin_unlock.
7178 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
7181 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7182 struct bpf_verifier_state *cur = env->cur_state;
7183 bool is_const = tnum_is_const(reg->var_off);
7184 u64 val = reg->var_off.value;
7185 struct bpf_map *map = NULL;
7186 struct btf *btf = NULL;
7187 struct btf_record *rec;
7191 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
7195 if (reg->type == PTR_TO_MAP_VALUE) {
7199 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
7207 rec = reg_btf_record(reg);
7208 if (!btf_record_has_field(rec, BPF_SPIN_LOCK)) {
7209 verbose(env, "%s '%s' has no valid bpf_spin_lock\n", map ? "map" : "local",
7210 map ? map->name : "kptr");
7213 if (rec->spin_lock_off != val + reg->off) {
7214 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock' that is at %d\n",
7215 val + reg->off, rec->spin_lock_off);
7219 if (cur->active_lock.ptr) {
7221 "Locking two bpf_spin_locks are not allowed\n");
7225 cur->active_lock.ptr = map;
7227 cur->active_lock.ptr = btf;
7228 cur->active_lock.id = reg->id;
7237 if (!cur->active_lock.ptr) {
7238 verbose(env, "bpf_spin_unlock without taking a lock\n");
7241 if (cur->active_lock.ptr != ptr ||
7242 cur->active_lock.id != reg->id) {
7243 verbose(env, "bpf_spin_unlock of different lock\n");
7247 invalidate_non_owning_refs(env);
7249 cur->active_lock.ptr = NULL;
7250 cur->active_lock.id = 0;
7255 static int process_timer_func(struct bpf_verifier_env *env, int regno,
7256 struct bpf_call_arg_meta *meta)
7258 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7259 bool is_const = tnum_is_const(reg->var_off);
7260 struct bpf_map *map = reg->map_ptr;
7261 u64 val = reg->var_off.value;
7265 "R%d doesn't have constant offset. bpf_timer has to be at the constant offset\n",
7270 verbose(env, "map '%s' has to have BTF in order to use bpf_timer\n",
7274 if (!btf_record_has_field(map->record, BPF_TIMER)) {
7275 verbose(env, "map '%s' has no valid bpf_timer\n", map->name);
7278 if (map->record->timer_off != val + reg->off) {
7279 verbose(env, "off %lld doesn't point to 'struct bpf_timer' that is at %d\n",
7280 val + reg->off, map->record->timer_off);
7283 if (meta->map_ptr) {
7284 verbose(env, "verifier bug. Two map pointers in a timer helper\n");
7287 meta->map_uid = reg->map_uid;
7288 meta->map_ptr = map;
7292 static int process_kptr_func(struct bpf_verifier_env *env, int regno,
7293 struct bpf_call_arg_meta *meta)
7295 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7296 struct bpf_map *map_ptr = reg->map_ptr;
7297 struct btf_field *kptr_field;
7300 if (!tnum_is_const(reg->var_off)) {
7302 "R%d doesn't have constant offset. kptr has to be at the constant offset\n",
7306 if (!map_ptr->btf) {
7307 verbose(env, "map '%s' has to have BTF in order to use bpf_kptr_xchg\n",
7311 if (!btf_record_has_field(map_ptr->record, BPF_KPTR)) {
7312 verbose(env, "map '%s' has no valid kptr\n", map_ptr->name);
7316 meta->map_ptr = map_ptr;
7317 kptr_off = reg->off + reg->var_off.value;
7318 kptr_field = btf_record_find(map_ptr->record, kptr_off, BPF_KPTR);
7320 verbose(env, "off=%d doesn't point to kptr\n", kptr_off);
7323 if (kptr_field->type != BPF_KPTR_REF) {
7324 verbose(env, "off=%d kptr isn't referenced kptr\n", kptr_off);
7327 meta->kptr_field = kptr_field;
7331 /* There are two register types representing a bpf_dynptr, one is PTR_TO_STACK
7332 * which points to a stack slot, and the other is CONST_PTR_TO_DYNPTR.
7334 * In both cases we deal with the first 8 bytes, but need to mark the next 8
7335 * bytes as STACK_DYNPTR in case of PTR_TO_STACK. In case of
7336 * CONST_PTR_TO_DYNPTR, we are guaranteed to get the beginning of the object.
7338 * Mutability of bpf_dynptr is at two levels, one is at the level of struct
7339 * bpf_dynptr itself, i.e. whether the helper is receiving a pointer to struct
7340 * bpf_dynptr or pointer to const struct bpf_dynptr. In the former case, it can
7341 * mutate the view of the dynptr and also possibly destroy it. In the latter
7342 * case, it cannot mutate the bpf_dynptr itself but it can still mutate the
7343 * memory that dynptr points to.
7345 * The verifier will keep track both levels of mutation (bpf_dynptr's in
7346 * reg->type and the memory's in reg->dynptr.type), but there is no support for
7347 * readonly dynptr view yet, hence only the first case is tracked and checked.
7349 * This is consistent with how C applies the const modifier to a struct object,
7350 * where the pointer itself inside bpf_dynptr becomes const but not what it
7353 * Helpers which do not mutate the bpf_dynptr set MEM_RDONLY in their argument
7354 * type, and declare it as 'const struct bpf_dynptr *' in their prototype.
7356 static int process_dynptr_func(struct bpf_verifier_env *env, int regno, int insn_idx,
7357 enum bpf_arg_type arg_type, int clone_ref_obj_id)
7359 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7362 /* MEM_UNINIT and MEM_RDONLY are exclusive, when applied to an
7363 * ARG_PTR_TO_DYNPTR (or ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_*):
7365 if ((arg_type & (MEM_UNINIT | MEM_RDONLY)) == (MEM_UNINIT | MEM_RDONLY)) {
7366 verbose(env, "verifier internal error: misconfigured dynptr helper type flags\n");
7370 /* MEM_UNINIT - Points to memory that is an appropriate candidate for
7371 * constructing a mutable bpf_dynptr object.
7373 * Currently, this is only possible with PTR_TO_STACK
7374 * pointing to a region of at least 16 bytes which doesn't
7375 * contain an existing bpf_dynptr.
7377 * MEM_RDONLY - Points to a initialized bpf_dynptr that will not be
7378 * mutated or destroyed. However, the memory it points to
7381 * None - Points to a initialized dynptr that can be mutated and
7382 * destroyed, including mutation of the memory it points
7385 if (arg_type & MEM_UNINIT) {
7388 if (!is_dynptr_reg_valid_uninit(env, reg)) {
7389 verbose(env, "Dynptr has to be an uninitialized dynptr\n");
7393 /* we write BPF_DW bits (8 bytes) at a time */
7394 for (i = 0; i < BPF_DYNPTR_SIZE; i += 8) {
7395 err = check_mem_access(env, insn_idx, regno,
7396 i, BPF_DW, BPF_WRITE, -1, false, false);
7401 err = mark_stack_slots_dynptr(env, reg, arg_type, insn_idx, clone_ref_obj_id);
7402 } else /* MEM_RDONLY and None case from above */ {
7403 /* For the reg->type == PTR_TO_STACK case, bpf_dynptr is never const */
7404 if (reg->type == CONST_PTR_TO_DYNPTR && !(arg_type & MEM_RDONLY)) {
7405 verbose(env, "cannot pass pointer to const bpf_dynptr, the helper mutates it\n");
7409 if (!is_dynptr_reg_valid_init(env, reg)) {
7411 "Expected an initialized dynptr as arg #%d\n",
7416 /* Fold modifiers (in this case, MEM_RDONLY) when checking expected type */
7417 if (!is_dynptr_type_expected(env, reg, arg_type & ~MEM_RDONLY)) {
7419 "Expected a dynptr of type %s as arg #%d\n",
7420 dynptr_type_str(arg_to_dynptr_type(arg_type)), regno);
7424 err = mark_dynptr_read(env, reg);
7429 static u32 iter_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg, int spi)
7431 struct bpf_func_state *state = func(env, reg);
7433 return state->stack[spi].spilled_ptr.ref_obj_id;
7436 static bool is_iter_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7438 return meta->kfunc_flags & (KF_ITER_NEW | KF_ITER_NEXT | KF_ITER_DESTROY);
7441 static bool is_iter_new_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7443 return meta->kfunc_flags & KF_ITER_NEW;
7446 static bool is_iter_next_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7448 return meta->kfunc_flags & KF_ITER_NEXT;
7451 static bool is_iter_destroy_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7453 return meta->kfunc_flags & KF_ITER_DESTROY;
7456 static bool is_kfunc_arg_iter(struct bpf_kfunc_call_arg_meta *meta, int arg)
7458 /* btf_check_iter_kfuncs() guarantees that first argument of any iter
7459 * kfunc is iter state pointer
7461 return arg == 0 && is_iter_kfunc(meta);
7464 static int process_iter_arg(struct bpf_verifier_env *env, int regno, int insn_idx,
7465 struct bpf_kfunc_call_arg_meta *meta)
7467 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7468 const struct btf_type *t;
7469 const struct btf_param *arg;
7470 int spi, err, i, nr_slots;
7473 /* btf_check_iter_kfuncs() ensures we don't need to validate anything here */
7474 arg = &btf_params(meta->func_proto)[0];
7475 t = btf_type_skip_modifiers(meta->btf, arg->type, NULL); /* PTR */
7476 t = btf_type_skip_modifiers(meta->btf, t->type, &btf_id); /* STRUCT */
7477 nr_slots = t->size / BPF_REG_SIZE;
7479 if (is_iter_new_kfunc(meta)) {
7480 /* bpf_iter_<type>_new() expects pointer to uninit iter state */
7481 if (!is_iter_reg_valid_uninit(env, reg, nr_slots)) {
7482 verbose(env, "expected uninitialized iter_%s as arg #%d\n",
7483 iter_type_str(meta->btf, btf_id), regno);
7487 for (i = 0; i < nr_slots * 8; i += BPF_REG_SIZE) {
7488 err = check_mem_access(env, insn_idx, regno,
7489 i, BPF_DW, BPF_WRITE, -1, false, false);
7494 err = mark_stack_slots_iter(env, reg, insn_idx, meta->btf, btf_id, nr_slots);
7498 /* iter_next() or iter_destroy() expect initialized iter state*/
7499 if (!is_iter_reg_valid_init(env, reg, meta->btf, btf_id, nr_slots)) {
7500 verbose(env, "expected an initialized iter_%s as arg #%d\n",
7501 iter_type_str(meta->btf, btf_id), regno);
7505 spi = iter_get_spi(env, reg, nr_slots);
7509 err = mark_iter_read(env, reg, spi, nr_slots);
7513 /* remember meta->iter info for process_iter_next_call() */
7514 meta->iter.spi = spi;
7515 meta->iter.frameno = reg->frameno;
7516 meta->ref_obj_id = iter_ref_obj_id(env, reg, spi);
7518 if (is_iter_destroy_kfunc(meta)) {
7519 err = unmark_stack_slots_iter(env, reg, nr_slots);
7528 /* process_iter_next_call() is called when verifier gets to iterator's next
7529 * "method" (e.g., bpf_iter_num_next() for numbers iterator) call. We'll refer
7530 * to it as just "iter_next()" in comments below.
7532 * BPF verifier relies on a crucial contract for any iter_next()
7533 * implementation: it should *eventually* return NULL, and once that happens
7534 * it should keep returning NULL. That is, once iterator exhausts elements to
7535 * iterate, it should never reset or spuriously return new elements.
7537 * With the assumption of such contract, process_iter_next_call() simulates
7538 * a fork in the verifier state to validate loop logic correctness and safety
7539 * without having to simulate infinite amount of iterations.
7541 * In current state, we first assume that iter_next() returned NULL and
7542 * iterator state is set to DRAINED (BPF_ITER_STATE_DRAINED). In such
7543 * conditions we should not form an infinite loop and should eventually reach
7546 * Besides that, we also fork current state and enqueue it for later
7547 * verification. In a forked state we keep iterator state as ACTIVE
7548 * (BPF_ITER_STATE_ACTIVE) and assume non-NULL return from iter_next(). We
7549 * also bump iteration depth to prevent erroneous infinite loop detection
7550 * later on (see iter_active_depths_differ() comment for details). In this
7551 * state we assume that we'll eventually loop back to another iter_next()
7552 * calls (it could be in exactly same location or in some other instruction,
7553 * it doesn't matter, we don't make any unnecessary assumptions about this,
7554 * everything revolves around iterator state in a stack slot, not which
7555 * instruction is calling iter_next()). When that happens, we either will come
7556 * to iter_next() with equivalent state and can conclude that next iteration
7557 * will proceed in exactly the same way as we just verified, so it's safe to
7558 * assume that loop converges. If not, we'll go on another iteration
7559 * simulation with a different input state, until all possible starting states
7560 * are validated or we reach maximum number of instructions limit.
7562 * This way, we will either exhaustively discover all possible input states
7563 * that iterator loop can start with and eventually will converge, or we'll
7564 * effectively regress into bounded loop simulation logic and either reach
7565 * maximum number of instructions if loop is not provably convergent, or there
7566 * is some statically known limit on number of iterations (e.g., if there is
7567 * an explicit `if n > 100 then break;` statement somewhere in the loop).
7569 * One very subtle but very important aspect is that we *always* simulate NULL
7570 * condition first (as the current state) before we simulate non-NULL case.
7571 * This has to do with intricacies of scalar precision tracking. By simulating
7572 * "exit condition" of iter_next() returning NULL first, we make sure all the
7573 * relevant precision marks *that will be set **after** we exit iterator loop*
7574 * are propagated backwards to common parent state of NULL and non-NULL
7575 * branches. Thanks to that, state equivalence checks done later in forked
7576 * state, when reaching iter_next() for ACTIVE iterator, can assume that
7577 * precision marks are finalized and won't change. Because simulating another
7578 * ACTIVE iterator iteration won't change them (because given same input
7579 * states we'll end up with exactly same output states which we are currently
7580 * comparing; and verification after the loop already propagated back what
7581 * needs to be **additionally** tracked as precise). It's subtle, grok
7582 * precision tracking for more intuitive understanding.
7584 static int process_iter_next_call(struct bpf_verifier_env *env, int insn_idx,
7585 struct bpf_kfunc_call_arg_meta *meta)
7587 struct bpf_verifier_state *cur_st = env->cur_state, *queued_st;
7588 struct bpf_func_state *cur_fr = cur_st->frame[cur_st->curframe], *queued_fr;
7589 struct bpf_reg_state *cur_iter, *queued_iter;
7590 int iter_frameno = meta->iter.frameno;
7591 int iter_spi = meta->iter.spi;
7593 BTF_TYPE_EMIT(struct bpf_iter);
7595 cur_iter = &env->cur_state->frame[iter_frameno]->stack[iter_spi].spilled_ptr;
7597 if (cur_iter->iter.state != BPF_ITER_STATE_ACTIVE &&
7598 cur_iter->iter.state != BPF_ITER_STATE_DRAINED) {
7599 verbose(env, "verifier internal error: unexpected iterator state %d (%s)\n",
7600 cur_iter->iter.state, iter_state_str(cur_iter->iter.state));
7604 if (cur_iter->iter.state == BPF_ITER_STATE_ACTIVE) {
7605 /* branch out active iter state */
7606 queued_st = push_stack(env, insn_idx + 1, insn_idx, false);
7610 queued_iter = &queued_st->frame[iter_frameno]->stack[iter_spi].spilled_ptr;
7611 queued_iter->iter.state = BPF_ITER_STATE_ACTIVE;
7612 queued_iter->iter.depth++;
7614 queued_fr = queued_st->frame[queued_st->curframe];
7615 mark_ptr_not_null_reg(&queued_fr->regs[BPF_REG_0]);
7618 /* switch to DRAINED state, but keep the depth unchanged */
7619 /* mark current iter state as drained and assume returned NULL */
7620 cur_iter->iter.state = BPF_ITER_STATE_DRAINED;
7621 __mark_reg_const_zero(&cur_fr->regs[BPF_REG_0]);
7626 static bool arg_type_is_mem_size(enum bpf_arg_type type)
7628 return type == ARG_CONST_SIZE ||
7629 type == ARG_CONST_SIZE_OR_ZERO;
7632 static bool arg_type_is_release(enum bpf_arg_type type)
7634 return type & OBJ_RELEASE;
7637 static bool arg_type_is_dynptr(enum bpf_arg_type type)
7639 return base_type(type) == ARG_PTR_TO_DYNPTR;
7642 static int int_ptr_type_to_size(enum bpf_arg_type type)
7644 if (type == ARG_PTR_TO_INT)
7646 else if (type == ARG_PTR_TO_LONG)
7652 static int resolve_map_arg_type(struct bpf_verifier_env *env,
7653 const struct bpf_call_arg_meta *meta,
7654 enum bpf_arg_type *arg_type)
7656 if (!meta->map_ptr) {
7657 /* kernel subsystem misconfigured verifier */
7658 verbose(env, "invalid map_ptr to access map->type\n");
7662 switch (meta->map_ptr->map_type) {
7663 case BPF_MAP_TYPE_SOCKMAP:
7664 case BPF_MAP_TYPE_SOCKHASH:
7665 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
7666 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
7668 verbose(env, "invalid arg_type for sockmap/sockhash\n");
7672 case BPF_MAP_TYPE_BLOOM_FILTER:
7673 if (meta->func_id == BPF_FUNC_map_peek_elem)
7674 *arg_type = ARG_PTR_TO_MAP_VALUE;
7682 struct bpf_reg_types {
7683 const enum bpf_reg_type types[10];
7687 static const struct bpf_reg_types sock_types = {
7697 static const struct bpf_reg_types btf_id_sock_common_types = {
7704 PTR_TO_BTF_ID | PTR_TRUSTED,
7706 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
7710 static const struct bpf_reg_types mem_types = {
7718 PTR_TO_MEM | MEM_RINGBUF,
7720 PTR_TO_BTF_ID | PTR_TRUSTED,
7724 static const struct bpf_reg_types int_ptr_types = {
7734 static const struct bpf_reg_types spin_lock_types = {
7737 PTR_TO_BTF_ID | MEM_ALLOC,
7741 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
7742 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
7743 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
7744 static const struct bpf_reg_types ringbuf_mem_types = { .types = { PTR_TO_MEM | MEM_RINGBUF } };
7745 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
7746 static const struct bpf_reg_types btf_ptr_types = {
7749 PTR_TO_BTF_ID | PTR_TRUSTED,
7750 PTR_TO_BTF_ID | MEM_RCU,
7753 static const struct bpf_reg_types percpu_btf_ptr_types = {
7755 PTR_TO_BTF_ID | MEM_PERCPU,
7756 PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED,
7759 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
7760 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
7761 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
7762 static const struct bpf_reg_types timer_types = { .types = { PTR_TO_MAP_VALUE } };
7763 static const struct bpf_reg_types kptr_types = { .types = { PTR_TO_MAP_VALUE } };
7764 static const struct bpf_reg_types dynptr_types = {
7767 CONST_PTR_TO_DYNPTR,
7771 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
7772 [ARG_PTR_TO_MAP_KEY] = &mem_types,
7773 [ARG_PTR_TO_MAP_VALUE] = &mem_types,
7774 [ARG_CONST_SIZE] = &scalar_types,
7775 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
7776 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
7777 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
7778 [ARG_PTR_TO_CTX] = &context_types,
7779 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
7781 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
7783 [ARG_PTR_TO_SOCKET] = &fullsock_types,
7784 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
7785 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
7786 [ARG_PTR_TO_MEM] = &mem_types,
7787 [ARG_PTR_TO_RINGBUF_MEM] = &ringbuf_mem_types,
7788 [ARG_PTR_TO_INT] = &int_ptr_types,
7789 [ARG_PTR_TO_LONG] = &int_ptr_types,
7790 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
7791 [ARG_PTR_TO_FUNC] = &func_ptr_types,
7792 [ARG_PTR_TO_STACK] = &stack_ptr_types,
7793 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
7794 [ARG_PTR_TO_TIMER] = &timer_types,
7795 [ARG_PTR_TO_KPTR] = &kptr_types,
7796 [ARG_PTR_TO_DYNPTR] = &dynptr_types,
7799 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
7800 enum bpf_arg_type arg_type,
7801 const u32 *arg_btf_id,
7802 struct bpf_call_arg_meta *meta)
7804 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7805 enum bpf_reg_type expected, type = reg->type;
7806 const struct bpf_reg_types *compatible;
7809 compatible = compatible_reg_types[base_type(arg_type)];
7811 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
7815 /* ARG_PTR_TO_MEM + RDONLY is compatible with PTR_TO_MEM and PTR_TO_MEM + RDONLY,
7816 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM and NOT with PTR_TO_MEM + RDONLY
7818 * Same for MAYBE_NULL:
7820 * ARG_PTR_TO_MEM + MAYBE_NULL is compatible with PTR_TO_MEM and PTR_TO_MEM + MAYBE_NULL,
7821 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM but NOT with PTR_TO_MEM + MAYBE_NULL
7823 * ARG_PTR_TO_MEM is compatible with PTR_TO_MEM that is tagged with a dynptr type.
7825 * Therefore we fold these flags depending on the arg_type before comparison.
7827 if (arg_type & MEM_RDONLY)
7828 type &= ~MEM_RDONLY;
7829 if (arg_type & PTR_MAYBE_NULL)
7830 type &= ~PTR_MAYBE_NULL;
7831 if (base_type(arg_type) == ARG_PTR_TO_MEM)
7832 type &= ~DYNPTR_TYPE_FLAG_MASK;
7834 if (meta->func_id == BPF_FUNC_kptr_xchg && type_is_alloc(type))
7837 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
7838 expected = compatible->types[i];
7839 if (expected == NOT_INIT)
7842 if (type == expected)
7846 verbose(env, "R%d type=%s expected=", regno, reg_type_str(env, reg->type));
7847 for (j = 0; j + 1 < i; j++)
7848 verbose(env, "%s, ", reg_type_str(env, compatible->types[j]));
7849 verbose(env, "%s\n", reg_type_str(env, compatible->types[j]));
7853 if (base_type(reg->type) != PTR_TO_BTF_ID)
7856 if (compatible == &mem_types) {
7857 if (!(arg_type & MEM_RDONLY)) {
7859 "%s() may write into memory pointed by R%d type=%s\n",
7860 func_id_name(meta->func_id),
7861 regno, reg_type_str(env, reg->type));
7867 switch ((int)reg->type) {
7869 case PTR_TO_BTF_ID | PTR_TRUSTED:
7870 case PTR_TO_BTF_ID | MEM_RCU:
7871 case PTR_TO_BTF_ID | PTR_MAYBE_NULL:
7872 case PTR_TO_BTF_ID | PTR_MAYBE_NULL | MEM_RCU:
7874 /* For bpf_sk_release, it needs to match against first member
7875 * 'struct sock_common', hence make an exception for it. This
7876 * allows bpf_sk_release to work for multiple socket types.
7878 bool strict_type_match = arg_type_is_release(arg_type) &&
7879 meta->func_id != BPF_FUNC_sk_release;
7881 if (type_may_be_null(reg->type) &&
7882 (!type_may_be_null(arg_type) || arg_type_is_release(arg_type))) {
7883 verbose(env, "Possibly NULL pointer passed to helper arg%d\n", regno);
7888 if (!compatible->btf_id) {
7889 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
7892 arg_btf_id = compatible->btf_id;
7895 if (meta->func_id == BPF_FUNC_kptr_xchg) {
7896 if (map_kptr_match_type(env, meta->kptr_field, reg, regno))
7899 if (arg_btf_id == BPF_PTR_POISON) {
7900 verbose(env, "verifier internal error:");
7901 verbose(env, "R%d has non-overwritten BPF_PTR_POISON type\n",
7906 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
7907 btf_vmlinux, *arg_btf_id,
7908 strict_type_match)) {
7909 verbose(env, "R%d is of type %s but %s is expected\n",
7910 regno, btf_type_name(reg->btf, reg->btf_id),
7911 btf_type_name(btf_vmlinux, *arg_btf_id));
7917 case PTR_TO_BTF_ID | MEM_ALLOC:
7918 if (meta->func_id != BPF_FUNC_spin_lock && meta->func_id != BPF_FUNC_spin_unlock &&
7919 meta->func_id != BPF_FUNC_kptr_xchg) {
7920 verbose(env, "verifier internal error: unimplemented handling of MEM_ALLOC\n");
7923 if (meta->func_id == BPF_FUNC_kptr_xchg) {
7924 if (map_kptr_match_type(env, meta->kptr_field, reg, regno))
7928 case PTR_TO_BTF_ID | MEM_PERCPU:
7929 case PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED:
7930 /* Handled by helper specific checks */
7933 verbose(env, "verifier internal error: invalid PTR_TO_BTF_ID register for type match\n");
7939 static struct btf_field *
7940 reg_find_field_offset(const struct bpf_reg_state *reg, s32 off, u32 fields)
7942 struct btf_field *field;
7943 struct btf_record *rec;
7945 rec = reg_btf_record(reg);
7949 field = btf_record_find(rec, off, fields);
7956 int check_func_arg_reg_off(struct bpf_verifier_env *env,
7957 const struct bpf_reg_state *reg, int regno,
7958 enum bpf_arg_type arg_type)
7960 u32 type = reg->type;
7962 /* When referenced register is passed to release function, its fixed
7965 * We will check arg_type_is_release reg has ref_obj_id when storing
7966 * meta->release_regno.
7968 if (arg_type_is_release(arg_type)) {
7969 /* ARG_PTR_TO_DYNPTR with OBJ_RELEASE is a bit special, as it
7970 * may not directly point to the object being released, but to
7971 * dynptr pointing to such object, which might be at some offset
7972 * on the stack. In that case, we simply to fallback to the
7975 if (arg_type_is_dynptr(arg_type) && type == PTR_TO_STACK)
7978 /* Doing check_ptr_off_reg check for the offset will catch this
7979 * because fixed_off_ok is false, but checking here allows us
7980 * to give the user a better error message.
7983 verbose(env, "R%d must have zero offset when passed to release func or trusted arg to kfunc\n",
7987 return __check_ptr_off_reg(env, reg, regno, false);
7991 /* Pointer types where both fixed and variable offset is explicitly allowed: */
7994 case PTR_TO_PACKET_META:
7995 case PTR_TO_MAP_KEY:
7996 case PTR_TO_MAP_VALUE:
7998 case PTR_TO_MEM | MEM_RDONLY:
7999 case PTR_TO_MEM | MEM_RINGBUF:
8001 case PTR_TO_BUF | MEM_RDONLY:
8004 /* All the rest must be rejected, except PTR_TO_BTF_ID which allows
8008 case PTR_TO_BTF_ID | MEM_ALLOC:
8009 case PTR_TO_BTF_ID | PTR_TRUSTED:
8010 case PTR_TO_BTF_ID | MEM_RCU:
8011 case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF:
8012 case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF | MEM_RCU:
8013 /* When referenced PTR_TO_BTF_ID is passed to release function,
8014 * its fixed offset must be 0. In the other cases, fixed offset
8015 * can be non-zero. This was already checked above. So pass
8016 * fixed_off_ok as true to allow fixed offset for all other
8017 * cases. var_off always must be 0 for PTR_TO_BTF_ID, hence we
8018 * still need to do checks instead of returning.
8020 return __check_ptr_off_reg(env, reg, regno, true);
8022 return __check_ptr_off_reg(env, reg, regno, false);
8026 static struct bpf_reg_state *get_dynptr_arg_reg(struct bpf_verifier_env *env,
8027 const struct bpf_func_proto *fn,
8028 struct bpf_reg_state *regs)
8030 struct bpf_reg_state *state = NULL;
8033 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++)
8034 if (arg_type_is_dynptr(fn->arg_type[i])) {
8036 verbose(env, "verifier internal error: multiple dynptr args\n");
8039 state = ®s[BPF_REG_1 + i];
8043 verbose(env, "verifier internal error: no dynptr arg found\n");
8048 static int dynptr_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
8050 struct bpf_func_state *state = func(env, reg);
8053 if (reg->type == CONST_PTR_TO_DYNPTR)
8055 spi = dynptr_get_spi(env, reg);
8058 return state->stack[spi].spilled_ptr.id;
8061 static int dynptr_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
8063 struct bpf_func_state *state = func(env, reg);
8066 if (reg->type == CONST_PTR_TO_DYNPTR)
8067 return reg->ref_obj_id;
8068 spi = dynptr_get_spi(env, reg);
8071 return state->stack[spi].spilled_ptr.ref_obj_id;
8074 static enum bpf_dynptr_type dynptr_get_type(struct bpf_verifier_env *env,
8075 struct bpf_reg_state *reg)
8077 struct bpf_func_state *state = func(env, reg);
8080 if (reg->type == CONST_PTR_TO_DYNPTR)
8081 return reg->dynptr.type;
8083 spi = __get_spi(reg->off);
8085 verbose(env, "verifier internal error: invalid spi when querying dynptr type\n");
8086 return BPF_DYNPTR_TYPE_INVALID;
8089 return state->stack[spi].spilled_ptr.dynptr.type;
8092 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
8093 struct bpf_call_arg_meta *meta,
8094 const struct bpf_func_proto *fn,
8097 u32 regno = BPF_REG_1 + arg;
8098 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
8099 enum bpf_arg_type arg_type = fn->arg_type[arg];
8100 enum bpf_reg_type type = reg->type;
8101 u32 *arg_btf_id = NULL;
8104 if (arg_type == ARG_DONTCARE)
8107 err = check_reg_arg(env, regno, SRC_OP);
8111 if (arg_type == ARG_ANYTHING) {
8112 if (is_pointer_value(env, regno)) {
8113 verbose(env, "R%d leaks addr into helper function\n",
8120 if (type_is_pkt_pointer(type) &&
8121 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
8122 verbose(env, "helper access to the packet is not allowed\n");
8126 if (base_type(arg_type) == ARG_PTR_TO_MAP_VALUE) {
8127 err = resolve_map_arg_type(env, meta, &arg_type);
8132 if (register_is_null(reg) && type_may_be_null(arg_type))
8133 /* A NULL register has a SCALAR_VALUE type, so skip
8136 goto skip_type_check;
8138 /* arg_btf_id and arg_size are in a union. */
8139 if (base_type(arg_type) == ARG_PTR_TO_BTF_ID ||
8140 base_type(arg_type) == ARG_PTR_TO_SPIN_LOCK)
8141 arg_btf_id = fn->arg_btf_id[arg];
8143 err = check_reg_type(env, regno, arg_type, arg_btf_id, meta);
8147 err = check_func_arg_reg_off(env, reg, regno, arg_type);
8152 if (arg_type_is_release(arg_type)) {
8153 if (arg_type_is_dynptr(arg_type)) {
8154 struct bpf_func_state *state = func(env, reg);
8157 /* Only dynptr created on stack can be released, thus
8158 * the get_spi and stack state checks for spilled_ptr
8159 * should only be done before process_dynptr_func for
8162 if (reg->type == PTR_TO_STACK) {
8163 spi = dynptr_get_spi(env, reg);
8164 if (spi < 0 || !state->stack[spi].spilled_ptr.ref_obj_id) {
8165 verbose(env, "arg %d is an unacquired reference\n", regno);
8169 verbose(env, "cannot release unowned const bpf_dynptr\n");
8172 } else if (!reg->ref_obj_id && !register_is_null(reg)) {
8173 verbose(env, "R%d must be referenced when passed to release function\n",
8177 if (meta->release_regno) {
8178 verbose(env, "verifier internal error: more than one release argument\n");
8181 meta->release_regno = regno;
8184 if (reg->ref_obj_id) {
8185 if (meta->ref_obj_id) {
8186 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
8187 regno, reg->ref_obj_id,
8191 meta->ref_obj_id = reg->ref_obj_id;
8194 switch (base_type(arg_type)) {
8195 case ARG_CONST_MAP_PTR:
8196 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
8197 if (meta->map_ptr) {
8198 /* Use map_uid (which is unique id of inner map) to reject:
8199 * inner_map1 = bpf_map_lookup_elem(outer_map, key1)
8200 * inner_map2 = bpf_map_lookup_elem(outer_map, key2)
8201 * if (inner_map1 && inner_map2) {
8202 * timer = bpf_map_lookup_elem(inner_map1);
8204 * // mismatch would have been allowed
8205 * bpf_timer_init(timer, inner_map2);
8208 * Comparing map_ptr is enough to distinguish normal and outer maps.
8210 if (meta->map_ptr != reg->map_ptr ||
8211 meta->map_uid != reg->map_uid) {
8213 "timer pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
8214 meta->map_uid, reg->map_uid);
8218 meta->map_ptr = reg->map_ptr;
8219 meta->map_uid = reg->map_uid;
8221 case ARG_PTR_TO_MAP_KEY:
8222 /* bpf_map_xxx(..., map_ptr, ..., key) call:
8223 * check that [key, key + map->key_size) are within
8224 * stack limits and initialized
8226 if (!meta->map_ptr) {
8227 /* in function declaration map_ptr must come before
8228 * map_key, so that it's verified and known before
8229 * we have to check map_key here. Otherwise it means
8230 * that kernel subsystem misconfigured verifier
8232 verbose(env, "invalid map_ptr to access map->key\n");
8235 err = check_helper_mem_access(env, regno,
8236 meta->map_ptr->key_size, false,
8239 case ARG_PTR_TO_MAP_VALUE:
8240 if (type_may_be_null(arg_type) && register_is_null(reg))
8243 /* bpf_map_xxx(..., map_ptr, ..., value) call:
8244 * check [value, value + map->value_size) validity
8246 if (!meta->map_ptr) {
8247 /* kernel subsystem misconfigured verifier */
8248 verbose(env, "invalid map_ptr to access map->value\n");
8251 meta->raw_mode = arg_type & MEM_UNINIT;
8252 err = check_helper_mem_access(env, regno,
8253 meta->map_ptr->value_size, false,
8256 case ARG_PTR_TO_PERCPU_BTF_ID:
8258 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
8261 meta->ret_btf = reg->btf;
8262 meta->ret_btf_id = reg->btf_id;
8264 case ARG_PTR_TO_SPIN_LOCK:
8265 if (in_rbtree_lock_required_cb(env)) {
8266 verbose(env, "can't spin_{lock,unlock} in rbtree cb\n");
8269 if (meta->func_id == BPF_FUNC_spin_lock) {
8270 err = process_spin_lock(env, regno, true);
8273 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
8274 err = process_spin_lock(env, regno, false);
8278 verbose(env, "verifier internal error\n");
8282 case ARG_PTR_TO_TIMER:
8283 err = process_timer_func(env, regno, meta);
8287 case ARG_PTR_TO_FUNC:
8288 meta->subprogno = reg->subprogno;
8290 case ARG_PTR_TO_MEM:
8291 /* The access to this pointer is only checked when we hit the
8292 * next is_mem_size argument below.
8294 meta->raw_mode = arg_type & MEM_UNINIT;
8295 if (arg_type & MEM_FIXED_SIZE) {
8296 err = check_helper_mem_access(env, regno,
8297 fn->arg_size[arg], false,
8301 case ARG_CONST_SIZE:
8302 err = check_mem_size_reg(env, reg, regno, false, meta);
8304 case ARG_CONST_SIZE_OR_ZERO:
8305 err = check_mem_size_reg(env, reg, regno, true, meta);
8307 case ARG_PTR_TO_DYNPTR:
8308 err = process_dynptr_func(env, regno, insn_idx, arg_type, 0);
8312 case ARG_CONST_ALLOC_SIZE_OR_ZERO:
8313 if (!tnum_is_const(reg->var_off)) {
8314 verbose(env, "R%d is not a known constant'\n",
8318 meta->mem_size = reg->var_off.value;
8319 err = mark_chain_precision(env, regno);
8323 case ARG_PTR_TO_INT:
8324 case ARG_PTR_TO_LONG:
8326 int size = int_ptr_type_to_size(arg_type);
8328 err = check_helper_mem_access(env, regno, size, false, meta);
8331 err = check_ptr_alignment(env, reg, 0, size, true);
8334 case ARG_PTR_TO_CONST_STR:
8336 struct bpf_map *map = reg->map_ptr;
8341 if (!bpf_map_is_rdonly(map)) {
8342 verbose(env, "R%d does not point to a readonly map'\n", regno);
8346 if (!tnum_is_const(reg->var_off)) {
8347 verbose(env, "R%d is not a constant address'\n", regno);
8351 if (!map->ops->map_direct_value_addr) {
8352 verbose(env, "no direct value access support for this map type\n");
8356 err = check_map_access(env, regno, reg->off,
8357 map->value_size - reg->off, false,
8362 map_off = reg->off + reg->var_off.value;
8363 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
8365 verbose(env, "direct value access on string failed\n");
8369 str_ptr = (char *)(long)(map_addr);
8370 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
8371 verbose(env, "string is not zero-terminated\n");
8376 case ARG_PTR_TO_KPTR:
8377 err = process_kptr_func(env, regno, meta);
8386 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
8388 enum bpf_attach_type eatype = env->prog->expected_attach_type;
8389 enum bpf_prog_type type = resolve_prog_type(env->prog);
8391 if (func_id != BPF_FUNC_map_update_elem)
8394 /* It's not possible to get access to a locked struct sock in these
8395 * contexts, so updating is safe.
8398 case BPF_PROG_TYPE_TRACING:
8399 if (eatype == BPF_TRACE_ITER)
8402 case BPF_PROG_TYPE_SOCKET_FILTER:
8403 case BPF_PROG_TYPE_SCHED_CLS:
8404 case BPF_PROG_TYPE_SCHED_ACT:
8405 case BPF_PROG_TYPE_XDP:
8406 case BPF_PROG_TYPE_SK_REUSEPORT:
8407 case BPF_PROG_TYPE_FLOW_DISSECTOR:
8408 case BPF_PROG_TYPE_SK_LOOKUP:
8414 verbose(env, "cannot update sockmap in this context\n");
8418 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
8420 return env->prog->jit_requested &&
8421 bpf_jit_supports_subprog_tailcalls();
8424 static int check_map_func_compatibility(struct bpf_verifier_env *env,
8425 struct bpf_map *map, int func_id)
8430 /* We need a two way check, first is from map perspective ... */
8431 switch (map->map_type) {
8432 case BPF_MAP_TYPE_PROG_ARRAY:
8433 if (func_id != BPF_FUNC_tail_call)
8436 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
8437 if (func_id != BPF_FUNC_perf_event_read &&
8438 func_id != BPF_FUNC_perf_event_output &&
8439 func_id != BPF_FUNC_skb_output &&
8440 func_id != BPF_FUNC_perf_event_read_value &&
8441 func_id != BPF_FUNC_xdp_output)
8444 case BPF_MAP_TYPE_RINGBUF:
8445 if (func_id != BPF_FUNC_ringbuf_output &&
8446 func_id != BPF_FUNC_ringbuf_reserve &&
8447 func_id != BPF_FUNC_ringbuf_query &&
8448 func_id != BPF_FUNC_ringbuf_reserve_dynptr &&
8449 func_id != BPF_FUNC_ringbuf_submit_dynptr &&
8450 func_id != BPF_FUNC_ringbuf_discard_dynptr)
8453 case BPF_MAP_TYPE_USER_RINGBUF:
8454 if (func_id != BPF_FUNC_user_ringbuf_drain)
8457 case BPF_MAP_TYPE_STACK_TRACE:
8458 if (func_id != BPF_FUNC_get_stackid)
8461 case BPF_MAP_TYPE_CGROUP_ARRAY:
8462 if (func_id != BPF_FUNC_skb_under_cgroup &&
8463 func_id != BPF_FUNC_current_task_under_cgroup)
8466 case BPF_MAP_TYPE_CGROUP_STORAGE:
8467 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
8468 if (func_id != BPF_FUNC_get_local_storage)
8471 case BPF_MAP_TYPE_DEVMAP:
8472 case BPF_MAP_TYPE_DEVMAP_HASH:
8473 if (func_id != BPF_FUNC_redirect_map &&
8474 func_id != BPF_FUNC_map_lookup_elem)
8477 /* Restrict bpf side of cpumap and xskmap, open when use-cases
8480 case BPF_MAP_TYPE_CPUMAP:
8481 if (func_id != BPF_FUNC_redirect_map)
8484 case BPF_MAP_TYPE_XSKMAP:
8485 if (func_id != BPF_FUNC_redirect_map &&
8486 func_id != BPF_FUNC_map_lookup_elem)
8489 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
8490 case BPF_MAP_TYPE_HASH_OF_MAPS:
8491 if (func_id != BPF_FUNC_map_lookup_elem)
8494 case BPF_MAP_TYPE_SOCKMAP:
8495 if (func_id != BPF_FUNC_sk_redirect_map &&
8496 func_id != BPF_FUNC_sock_map_update &&
8497 func_id != BPF_FUNC_map_delete_elem &&
8498 func_id != BPF_FUNC_msg_redirect_map &&
8499 func_id != BPF_FUNC_sk_select_reuseport &&
8500 func_id != BPF_FUNC_map_lookup_elem &&
8501 !may_update_sockmap(env, func_id))
8504 case BPF_MAP_TYPE_SOCKHASH:
8505 if (func_id != BPF_FUNC_sk_redirect_hash &&
8506 func_id != BPF_FUNC_sock_hash_update &&
8507 func_id != BPF_FUNC_map_delete_elem &&
8508 func_id != BPF_FUNC_msg_redirect_hash &&
8509 func_id != BPF_FUNC_sk_select_reuseport &&
8510 func_id != BPF_FUNC_map_lookup_elem &&
8511 !may_update_sockmap(env, func_id))
8514 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
8515 if (func_id != BPF_FUNC_sk_select_reuseport)
8518 case BPF_MAP_TYPE_QUEUE:
8519 case BPF_MAP_TYPE_STACK:
8520 if (func_id != BPF_FUNC_map_peek_elem &&
8521 func_id != BPF_FUNC_map_pop_elem &&
8522 func_id != BPF_FUNC_map_push_elem)
8525 case BPF_MAP_TYPE_SK_STORAGE:
8526 if (func_id != BPF_FUNC_sk_storage_get &&
8527 func_id != BPF_FUNC_sk_storage_delete &&
8528 func_id != BPF_FUNC_kptr_xchg)
8531 case BPF_MAP_TYPE_INODE_STORAGE:
8532 if (func_id != BPF_FUNC_inode_storage_get &&
8533 func_id != BPF_FUNC_inode_storage_delete &&
8534 func_id != BPF_FUNC_kptr_xchg)
8537 case BPF_MAP_TYPE_TASK_STORAGE:
8538 if (func_id != BPF_FUNC_task_storage_get &&
8539 func_id != BPF_FUNC_task_storage_delete &&
8540 func_id != BPF_FUNC_kptr_xchg)
8543 case BPF_MAP_TYPE_CGRP_STORAGE:
8544 if (func_id != BPF_FUNC_cgrp_storage_get &&
8545 func_id != BPF_FUNC_cgrp_storage_delete &&
8546 func_id != BPF_FUNC_kptr_xchg)
8549 case BPF_MAP_TYPE_BLOOM_FILTER:
8550 if (func_id != BPF_FUNC_map_peek_elem &&
8551 func_id != BPF_FUNC_map_push_elem)
8558 /* ... and second from the function itself. */
8560 case BPF_FUNC_tail_call:
8561 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
8563 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
8564 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
8568 case BPF_FUNC_perf_event_read:
8569 case BPF_FUNC_perf_event_output:
8570 case BPF_FUNC_perf_event_read_value:
8571 case BPF_FUNC_skb_output:
8572 case BPF_FUNC_xdp_output:
8573 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
8576 case BPF_FUNC_ringbuf_output:
8577 case BPF_FUNC_ringbuf_reserve:
8578 case BPF_FUNC_ringbuf_query:
8579 case BPF_FUNC_ringbuf_reserve_dynptr:
8580 case BPF_FUNC_ringbuf_submit_dynptr:
8581 case BPF_FUNC_ringbuf_discard_dynptr:
8582 if (map->map_type != BPF_MAP_TYPE_RINGBUF)
8585 case BPF_FUNC_user_ringbuf_drain:
8586 if (map->map_type != BPF_MAP_TYPE_USER_RINGBUF)
8589 case BPF_FUNC_get_stackid:
8590 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
8593 case BPF_FUNC_current_task_under_cgroup:
8594 case BPF_FUNC_skb_under_cgroup:
8595 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
8598 case BPF_FUNC_redirect_map:
8599 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
8600 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
8601 map->map_type != BPF_MAP_TYPE_CPUMAP &&
8602 map->map_type != BPF_MAP_TYPE_XSKMAP)
8605 case BPF_FUNC_sk_redirect_map:
8606 case BPF_FUNC_msg_redirect_map:
8607 case BPF_FUNC_sock_map_update:
8608 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
8611 case BPF_FUNC_sk_redirect_hash:
8612 case BPF_FUNC_msg_redirect_hash:
8613 case BPF_FUNC_sock_hash_update:
8614 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
8617 case BPF_FUNC_get_local_storage:
8618 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
8619 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
8622 case BPF_FUNC_sk_select_reuseport:
8623 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
8624 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
8625 map->map_type != BPF_MAP_TYPE_SOCKHASH)
8628 case BPF_FUNC_map_pop_elem:
8629 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
8630 map->map_type != BPF_MAP_TYPE_STACK)
8633 case BPF_FUNC_map_peek_elem:
8634 case BPF_FUNC_map_push_elem:
8635 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
8636 map->map_type != BPF_MAP_TYPE_STACK &&
8637 map->map_type != BPF_MAP_TYPE_BLOOM_FILTER)
8640 case BPF_FUNC_map_lookup_percpu_elem:
8641 if (map->map_type != BPF_MAP_TYPE_PERCPU_ARRAY &&
8642 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
8643 map->map_type != BPF_MAP_TYPE_LRU_PERCPU_HASH)
8646 case BPF_FUNC_sk_storage_get:
8647 case BPF_FUNC_sk_storage_delete:
8648 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
8651 case BPF_FUNC_inode_storage_get:
8652 case BPF_FUNC_inode_storage_delete:
8653 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
8656 case BPF_FUNC_task_storage_get:
8657 case BPF_FUNC_task_storage_delete:
8658 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
8661 case BPF_FUNC_cgrp_storage_get:
8662 case BPF_FUNC_cgrp_storage_delete:
8663 if (map->map_type != BPF_MAP_TYPE_CGRP_STORAGE)
8672 verbose(env, "cannot pass map_type %d into func %s#%d\n",
8673 map->map_type, func_id_name(func_id), func_id);
8677 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
8681 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
8683 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
8685 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
8687 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
8689 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
8692 /* We only support one arg being in raw mode at the moment,
8693 * which is sufficient for the helper functions we have
8699 static bool check_args_pair_invalid(const struct bpf_func_proto *fn, int arg)
8701 bool is_fixed = fn->arg_type[arg] & MEM_FIXED_SIZE;
8702 bool has_size = fn->arg_size[arg] != 0;
8703 bool is_next_size = false;
8705 if (arg + 1 < ARRAY_SIZE(fn->arg_type))
8706 is_next_size = arg_type_is_mem_size(fn->arg_type[arg + 1]);
8708 if (base_type(fn->arg_type[arg]) != ARG_PTR_TO_MEM)
8709 return is_next_size;
8711 return has_size == is_next_size || is_next_size == is_fixed;
8714 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
8716 /* bpf_xxx(..., buf, len) call will access 'len'
8717 * bytes from memory 'buf'. Both arg types need
8718 * to be paired, so make sure there's no buggy
8719 * helper function specification.
8721 if (arg_type_is_mem_size(fn->arg1_type) ||
8722 check_args_pair_invalid(fn, 0) ||
8723 check_args_pair_invalid(fn, 1) ||
8724 check_args_pair_invalid(fn, 2) ||
8725 check_args_pair_invalid(fn, 3) ||
8726 check_args_pair_invalid(fn, 4))
8732 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
8736 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
8737 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_BTF_ID)
8738 return !!fn->arg_btf_id[i];
8739 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_SPIN_LOCK)
8740 return fn->arg_btf_id[i] == BPF_PTR_POISON;
8741 if (base_type(fn->arg_type[i]) != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i] &&
8742 /* arg_btf_id and arg_size are in a union. */
8743 (base_type(fn->arg_type[i]) != ARG_PTR_TO_MEM ||
8744 !(fn->arg_type[i] & MEM_FIXED_SIZE)))
8751 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
8753 return check_raw_mode_ok(fn) &&
8754 check_arg_pair_ok(fn) &&
8755 check_btf_id_ok(fn) ? 0 : -EINVAL;
8758 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
8759 * are now invalid, so turn them into unknown SCALAR_VALUE.
8761 * This also applies to dynptr slices belonging to skb and xdp dynptrs,
8762 * since these slices point to packet data.
8764 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
8766 struct bpf_func_state *state;
8767 struct bpf_reg_state *reg;
8769 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
8770 if (reg_is_pkt_pointer_any(reg) || reg_is_dynptr_slice_pkt(reg))
8771 mark_reg_invalid(env, reg);
8777 BEYOND_PKT_END = -2,
8780 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
8782 struct bpf_func_state *state = vstate->frame[vstate->curframe];
8783 struct bpf_reg_state *reg = &state->regs[regn];
8785 if (reg->type != PTR_TO_PACKET)
8786 /* PTR_TO_PACKET_META is not supported yet */
8789 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
8790 * How far beyond pkt_end it goes is unknown.
8791 * if (!range_open) it's the case of pkt >= pkt_end
8792 * if (range_open) it's the case of pkt > pkt_end
8793 * hence this pointer is at least 1 byte bigger than pkt_end
8796 reg->range = BEYOND_PKT_END;
8798 reg->range = AT_PKT_END;
8801 /* The pointer with the specified id has released its reference to kernel
8802 * resources. Identify all copies of the same pointer and clear the reference.
8804 static int release_reference(struct bpf_verifier_env *env,
8807 struct bpf_func_state *state;
8808 struct bpf_reg_state *reg;
8811 err = release_reference_state(cur_func(env), ref_obj_id);
8815 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
8816 if (reg->ref_obj_id == ref_obj_id)
8817 mark_reg_invalid(env, reg);
8823 static void invalidate_non_owning_refs(struct bpf_verifier_env *env)
8825 struct bpf_func_state *unused;
8826 struct bpf_reg_state *reg;
8828 bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
8829 if (type_is_non_owning_ref(reg->type))
8830 mark_reg_invalid(env, reg);
8834 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
8835 struct bpf_reg_state *regs)
8839 /* after the call registers r0 - r5 were scratched */
8840 for (i = 0; i < CALLER_SAVED_REGS; i++) {
8841 mark_reg_not_init(env, regs, caller_saved[i]);
8842 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
8846 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
8847 struct bpf_func_state *caller,
8848 struct bpf_func_state *callee,
8851 static int set_callee_state(struct bpf_verifier_env *env,
8852 struct bpf_func_state *caller,
8853 struct bpf_func_state *callee, int insn_idx);
8855 static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
8856 int *insn_idx, int subprog,
8857 set_callee_state_fn set_callee_state_cb)
8859 struct bpf_verifier_state *state = env->cur_state;
8860 struct bpf_func_state *caller, *callee;
8863 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
8864 verbose(env, "the call stack of %d frames is too deep\n",
8865 state->curframe + 2);
8869 caller = state->frame[state->curframe];
8870 if (state->frame[state->curframe + 1]) {
8871 verbose(env, "verifier bug. Frame %d already allocated\n",
8872 state->curframe + 1);
8876 err = btf_check_subprog_call(env, subprog, caller->regs);
8879 if (subprog_is_global(env, subprog)) {
8881 verbose(env, "Caller passes invalid args into func#%d\n",
8885 if (env->log.level & BPF_LOG_LEVEL)
8887 "Func#%d is global and valid. Skipping.\n",
8889 clear_caller_saved_regs(env, caller->regs);
8891 /* All global functions return a 64-bit SCALAR_VALUE */
8892 mark_reg_unknown(env, caller->regs, BPF_REG_0);
8893 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
8895 /* continue with next insn after call */
8900 /* set_callee_state is used for direct subprog calls, but we are
8901 * interested in validating only BPF helpers that can call subprogs as
8904 if (set_callee_state_cb != set_callee_state) {
8905 if (bpf_pseudo_kfunc_call(insn) &&
8906 !is_callback_calling_kfunc(insn->imm)) {
8907 verbose(env, "verifier bug: kfunc %s#%d not marked as callback-calling\n",
8908 func_id_name(insn->imm), insn->imm);
8910 } else if (!bpf_pseudo_kfunc_call(insn) &&
8911 !is_callback_calling_function(insn->imm)) { /* helper */
8912 verbose(env, "verifier bug: helper %s#%d not marked as callback-calling\n",
8913 func_id_name(insn->imm), insn->imm);
8918 if (insn->code == (BPF_JMP | BPF_CALL) &&
8919 insn->src_reg == 0 &&
8920 insn->imm == BPF_FUNC_timer_set_callback) {
8921 struct bpf_verifier_state *async_cb;
8923 /* there is no real recursion here. timer callbacks are async */
8924 env->subprog_info[subprog].is_async_cb = true;
8925 async_cb = push_async_cb(env, env->subprog_info[subprog].start,
8926 *insn_idx, subprog);
8929 callee = async_cb->frame[0];
8930 callee->async_entry_cnt = caller->async_entry_cnt + 1;
8932 /* Convert bpf_timer_set_callback() args into timer callback args */
8933 err = set_callee_state_cb(env, caller, callee, *insn_idx);
8937 clear_caller_saved_regs(env, caller->regs);
8938 mark_reg_unknown(env, caller->regs, BPF_REG_0);
8939 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
8940 /* continue with next insn after call */
8944 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
8947 state->frame[state->curframe + 1] = callee;
8949 /* callee cannot access r0, r6 - r9 for reading and has to write
8950 * into its own stack before reading from it.
8951 * callee can read/write into caller's stack
8953 init_func_state(env, callee,
8954 /* remember the callsite, it will be used by bpf_exit */
8955 *insn_idx /* callsite */,
8956 state->curframe + 1 /* frameno within this callchain */,
8957 subprog /* subprog number within this prog */);
8959 /* Transfer references to the callee */
8960 err = copy_reference_state(callee, caller);
8964 err = set_callee_state_cb(env, caller, callee, *insn_idx);
8968 clear_caller_saved_regs(env, caller->regs);
8970 /* only increment it after check_reg_arg() finished */
8973 /* and go analyze first insn of the callee */
8974 *insn_idx = env->subprog_info[subprog].start - 1;
8976 if (env->log.level & BPF_LOG_LEVEL) {
8977 verbose(env, "caller:\n");
8978 print_verifier_state(env, caller, true);
8979 verbose(env, "callee:\n");
8980 print_verifier_state(env, callee, true);
8985 free_func_state(callee);
8986 state->frame[state->curframe + 1] = NULL;
8990 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
8991 struct bpf_func_state *caller,
8992 struct bpf_func_state *callee)
8994 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
8995 * void *callback_ctx, u64 flags);
8996 * callback_fn(struct bpf_map *map, void *key, void *value,
8997 * void *callback_ctx);
8999 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
9001 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
9002 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
9003 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
9005 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
9006 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
9007 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
9009 /* pointer to stack or null */
9010 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
9013 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9017 static int set_callee_state(struct bpf_verifier_env *env,
9018 struct bpf_func_state *caller,
9019 struct bpf_func_state *callee, int insn_idx)
9023 /* copy r1 - r5 args that callee can access. The copy includes parent
9024 * pointers, which connects us up to the liveness chain
9026 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
9027 callee->regs[i] = caller->regs[i];
9031 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
9034 int subprog, target_insn;
9036 target_insn = *insn_idx + insn->imm + 1;
9037 subprog = find_subprog(env, target_insn);
9039 verbose(env, "verifier bug. No program starts at insn %d\n",
9044 return __check_func_call(env, insn, insn_idx, subprog, set_callee_state);
9047 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
9048 struct bpf_func_state *caller,
9049 struct bpf_func_state *callee,
9052 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
9053 struct bpf_map *map;
9056 if (bpf_map_ptr_poisoned(insn_aux)) {
9057 verbose(env, "tail_call abusing map_ptr\n");
9061 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
9062 if (!map->ops->map_set_for_each_callback_args ||
9063 !map->ops->map_for_each_callback) {
9064 verbose(env, "callback function not allowed for map\n");
9068 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
9072 callee->in_callback_fn = true;
9073 callee->callback_ret_range = tnum_range(0, 1);
9077 static int set_loop_callback_state(struct bpf_verifier_env *env,
9078 struct bpf_func_state *caller,
9079 struct bpf_func_state *callee,
9082 /* bpf_loop(u32 nr_loops, void *callback_fn, void *callback_ctx,
9084 * callback_fn(u32 index, void *callback_ctx);
9086 callee->regs[BPF_REG_1].type = SCALAR_VALUE;
9087 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
9090 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
9091 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9092 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9094 callee->in_callback_fn = true;
9095 callee->callback_ret_range = tnum_range(0, 1);
9099 static int set_timer_callback_state(struct bpf_verifier_env *env,
9100 struct bpf_func_state *caller,
9101 struct bpf_func_state *callee,
9104 struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
9106 /* bpf_timer_set_callback(struct bpf_timer *timer, void *callback_fn);
9107 * callback_fn(struct bpf_map *map, void *key, void *value);
9109 callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP;
9110 __mark_reg_known_zero(&callee->regs[BPF_REG_1]);
9111 callee->regs[BPF_REG_1].map_ptr = map_ptr;
9113 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
9114 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
9115 callee->regs[BPF_REG_2].map_ptr = map_ptr;
9117 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
9118 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
9119 callee->regs[BPF_REG_3].map_ptr = map_ptr;
9122 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9123 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9124 callee->in_async_callback_fn = true;
9125 callee->callback_ret_range = tnum_range(0, 1);
9129 static int set_find_vma_callback_state(struct bpf_verifier_env *env,
9130 struct bpf_func_state *caller,
9131 struct bpf_func_state *callee,
9134 /* bpf_find_vma(struct task_struct *task, u64 addr,
9135 * void *callback_fn, void *callback_ctx, u64 flags)
9136 * (callback_fn)(struct task_struct *task,
9137 * struct vm_area_struct *vma, void *callback_ctx);
9139 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
9141 callee->regs[BPF_REG_2].type = PTR_TO_BTF_ID;
9142 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
9143 callee->regs[BPF_REG_2].btf = btf_vmlinux;
9144 callee->regs[BPF_REG_2].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_VMA],
9146 /* pointer to stack or null */
9147 callee->regs[BPF_REG_3] = caller->regs[BPF_REG_4];
9150 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9151 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9152 callee->in_callback_fn = true;
9153 callee->callback_ret_range = tnum_range(0, 1);
9157 static int set_user_ringbuf_callback_state(struct bpf_verifier_env *env,
9158 struct bpf_func_state *caller,
9159 struct bpf_func_state *callee,
9162 /* bpf_user_ringbuf_drain(struct bpf_map *map, void *callback_fn, void
9163 * callback_ctx, u64 flags);
9164 * callback_fn(const struct bpf_dynptr_t* dynptr, void *callback_ctx);
9166 __mark_reg_not_init(env, &callee->regs[BPF_REG_0]);
9167 mark_dynptr_cb_reg(env, &callee->regs[BPF_REG_1], BPF_DYNPTR_TYPE_LOCAL);
9168 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
9171 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
9172 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9173 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9175 callee->in_callback_fn = true;
9176 callee->callback_ret_range = tnum_range(0, 1);
9180 static int set_rbtree_add_callback_state(struct bpf_verifier_env *env,
9181 struct bpf_func_state *caller,
9182 struct bpf_func_state *callee,
9185 /* void bpf_rbtree_add_impl(struct bpf_rb_root *root, struct bpf_rb_node *node,
9186 * bool (less)(struct bpf_rb_node *a, const struct bpf_rb_node *b));
9188 * 'struct bpf_rb_node *node' arg to bpf_rbtree_add_impl is the same PTR_TO_BTF_ID w/ offset
9189 * that 'less' callback args will be receiving. However, 'node' arg was release_reference'd
9190 * by this point, so look at 'root'
9192 struct btf_field *field;
9194 field = reg_find_field_offset(&caller->regs[BPF_REG_1], caller->regs[BPF_REG_1].off,
9196 if (!field || !field->graph_root.value_btf_id)
9199 mark_reg_graph_node(callee->regs, BPF_REG_1, &field->graph_root);
9200 ref_set_non_owning(env, &callee->regs[BPF_REG_1]);
9201 mark_reg_graph_node(callee->regs, BPF_REG_2, &field->graph_root);
9202 ref_set_non_owning(env, &callee->regs[BPF_REG_2]);
9204 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
9205 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9206 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9207 callee->in_callback_fn = true;
9208 callee->callback_ret_range = tnum_range(0, 1);
9212 static bool is_rbtree_lock_required_kfunc(u32 btf_id);
9214 /* Are we currently verifying the callback for a rbtree helper that must
9215 * be called with lock held? If so, no need to complain about unreleased
9218 static bool in_rbtree_lock_required_cb(struct bpf_verifier_env *env)
9220 struct bpf_verifier_state *state = env->cur_state;
9221 struct bpf_insn *insn = env->prog->insnsi;
9222 struct bpf_func_state *callee;
9225 if (!state->curframe)
9228 callee = state->frame[state->curframe];
9230 if (!callee->in_callback_fn)
9233 kfunc_btf_id = insn[callee->callsite].imm;
9234 return is_rbtree_lock_required_kfunc(kfunc_btf_id);
9237 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
9239 struct bpf_verifier_state *state = env->cur_state;
9240 struct bpf_func_state *caller, *callee;
9241 struct bpf_reg_state *r0;
9244 callee = state->frame[state->curframe];
9245 r0 = &callee->regs[BPF_REG_0];
9246 if (r0->type == PTR_TO_STACK) {
9247 /* technically it's ok to return caller's stack pointer
9248 * (or caller's caller's pointer) back to the caller,
9249 * since these pointers are valid. Only current stack
9250 * pointer will be invalid as soon as function exits,
9251 * but let's be conservative
9253 verbose(env, "cannot return stack pointer to the caller\n");
9257 caller = state->frame[state->curframe - 1];
9258 if (callee->in_callback_fn) {
9259 /* enforce R0 return value range [0, 1]. */
9260 struct tnum range = callee->callback_ret_range;
9262 if (r0->type != SCALAR_VALUE) {
9263 verbose(env, "R0 not a scalar value\n");
9266 if (!tnum_in(range, r0->var_off)) {
9267 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
9271 /* return to the caller whatever r0 had in the callee */
9272 caller->regs[BPF_REG_0] = *r0;
9275 /* callback_fn frame should have released its own additions to parent's
9276 * reference state at this point, or check_reference_leak would
9277 * complain, hence it must be the same as the caller. There is no need
9280 if (!callee->in_callback_fn) {
9281 /* Transfer references to the caller */
9282 err = copy_reference_state(caller, callee);
9287 *insn_idx = callee->callsite + 1;
9288 if (env->log.level & BPF_LOG_LEVEL) {
9289 verbose(env, "returning from callee:\n");
9290 print_verifier_state(env, callee, true);
9291 verbose(env, "to caller at %d:\n", *insn_idx);
9292 print_verifier_state(env, caller, true);
9294 /* clear everything in the callee */
9295 free_func_state(callee);
9296 state->frame[state->curframe--] = NULL;
9300 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
9302 struct bpf_call_arg_meta *meta)
9304 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
9306 if (ret_type != RET_INTEGER)
9310 case BPF_FUNC_get_stack:
9311 case BPF_FUNC_get_task_stack:
9312 case BPF_FUNC_probe_read_str:
9313 case BPF_FUNC_probe_read_kernel_str:
9314 case BPF_FUNC_probe_read_user_str:
9315 ret_reg->smax_value = meta->msize_max_value;
9316 ret_reg->s32_max_value = meta->msize_max_value;
9317 ret_reg->smin_value = -MAX_ERRNO;
9318 ret_reg->s32_min_value = -MAX_ERRNO;
9319 reg_bounds_sync(ret_reg);
9321 case BPF_FUNC_get_smp_processor_id:
9322 ret_reg->umax_value = nr_cpu_ids - 1;
9323 ret_reg->u32_max_value = nr_cpu_ids - 1;
9324 ret_reg->smax_value = nr_cpu_ids - 1;
9325 ret_reg->s32_max_value = nr_cpu_ids - 1;
9326 ret_reg->umin_value = 0;
9327 ret_reg->u32_min_value = 0;
9328 ret_reg->smin_value = 0;
9329 ret_reg->s32_min_value = 0;
9330 reg_bounds_sync(ret_reg);
9336 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
9337 int func_id, int insn_idx)
9339 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
9340 struct bpf_map *map = meta->map_ptr;
9342 if (func_id != BPF_FUNC_tail_call &&
9343 func_id != BPF_FUNC_map_lookup_elem &&
9344 func_id != BPF_FUNC_map_update_elem &&
9345 func_id != BPF_FUNC_map_delete_elem &&
9346 func_id != BPF_FUNC_map_push_elem &&
9347 func_id != BPF_FUNC_map_pop_elem &&
9348 func_id != BPF_FUNC_map_peek_elem &&
9349 func_id != BPF_FUNC_for_each_map_elem &&
9350 func_id != BPF_FUNC_redirect_map &&
9351 func_id != BPF_FUNC_map_lookup_percpu_elem)
9355 verbose(env, "kernel subsystem misconfigured verifier\n");
9359 /* In case of read-only, some additional restrictions
9360 * need to be applied in order to prevent altering the
9361 * state of the map from program side.
9363 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
9364 (func_id == BPF_FUNC_map_delete_elem ||
9365 func_id == BPF_FUNC_map_update_elem ||
9366 func_id == BPF_FUNC_map_push_elem ||
9367 func_id == BPF_FUNC_map_pop_elem)) {
9368 verbose(env, "write into map forbidden\n");
9372 if (!BPF_MAP_PTR(aux->map_ptr_state))
9373 bpf_map_ptr_store(aux, meta->map_ptr,
9374 !meta->map_ptr->bypass_spec_v1);
9375 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
9376 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
9377 !meta->map_ptr->bypass_spec_v1);
9382 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
9383 int func_id, int insn_idx)
9385 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
9386 struct bpf_reg_state *regs = cur_regs(env), *reg;
9387 struct bpf_map *map = meta->map_ptr;
9391 if (func_id != BPF_FUNC_tail_call)
9393 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
9394 verbose(env, "kernel subsystem misconfigured verifier\n");
9398 reg = ®s[BPF_REG_3];
9399 val = reg->var_off.value;
9400 max = map->max_entries;
9402 if (!(register_is_const(reg) && val < max)) {
9403 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
9407 err = mark_chain_precision(env, BPF_REG_3);
9410 if (bpf_map_key_unseen(aux))
9411 bpf_map_key_store(aux, val);
9412 else if (!bpf_map_key_poisoned(aux) &&
9413 bpf_map_key_immediate(aux) != val)
9414 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
9418 static int check_reference_leak(struct bpf_verifier_env *env)
9420 struct bpf_func_state *state = cur_func(env);
9421 bool refs_lingering = false;
9424 if (state->frameno && !state->in_callback_fn)
9427 for (i = 0; i < state->acquired_refs; i++) {
9428 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno)
9430 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
9431 state->refs[i].id, state->refs[i].insn_idx);
9432 refs_lingering = true;
9434 return refs_lingering ? -EINVAL : 0;
9437 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
9438 struct bpf_reg_state *regs)
9440 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
9441 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
9442 struct bpf_map *fmt_map = fmt_reg->map_ptr;
9443 struct bpf_bprintf_data data = {};
9444 int err, fmt_map_off, num_args;
9448 /* data must be an array of u64 */
9449 if (data_len_reg->var_off.value % 8)
9451 num_args = data_len_reg->var_off.value / 8;
9453 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
9454 * and map_direct_value_addr is set.
9456 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
9457 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
9460 verbose(env, "verifier bug\n");
9463 fmt = (char *)(long)fmt_addr + fmt_map_off;
9465 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
9466 * can focus on validating the format specifiers.
9468 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, num_args, &data);
9470 verbose(env, "Invalid format string\n");
9475 static int check_get_func_ip(struct bpf_verifier_env *env)
9477 enum bpf_prog_type type = resolve_prog_type(env->prog);
9478 int func_id = BPF_FUNC_get_func_ip;
9480 if (type == BPF_PROG_TYPE_TRACING) {
9481 if (!bpf_prog_has_trampoline(env->prog)) {
9482 verbose(env, "func %s#%d supported only for fentry/fexit/fmod_ret programs\n",
9483 func_id_name(func_id), func_id);
9487 } else if (type == BPF_PROG_TYPE_KPROBE) {
9491 verbose(env, "func %s#%d not supported for program type %d\n",
9492 func_id_name(func_id), func_id, type);
9496 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
9498 return &env->insn_aux_data[env->insn_idx];
9501 static bool loop_flag_is_zero(struct bpf_verifier_env *env)
9503 struct bpf_reg_state *regs = cur_regs(env);
9504 struct bpf_reg_state *reg = ®s[BPF_REG_4];
9505 bool reg_is_null = register_is_null(reg);
9508 mark_chain_precision(env, BPF_REG_4);
9513 static void update_loop_inline_state(struct bpf_verifier_env *env, u32 subprogno)
9515 struct bpf_loop_inline_state *state = &cur_aux(env)->loop_inline_state;
9517 if (!state->initialized) {
9518 state->initialized = 1;
9519 state->fit_for_inline = loop_flag_is_zero(env);
9520 state->callback_subprogno = subprogno;
9524 if (!state->fit_for_inline)
9527 state->fit_for_inline = (loop_flag_is_zero(env) &&
9528 state->callback_subprogno == subprogno);
9531 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
9534 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
9535 const struct bpf_func_proto *fn = NULL;
9536 enum bpf_return_type ret_type;
9537 enum bpf_type_flag ret_flag;
9538 struct bpf_reg_state *regs;
9539 struct bpf_call_arg_meta meta;
9540 int insn_idx = *insn_idx_p;
9542 int i, err, func_id;
9544 /* find function prototype */
9545 func_id = insn->imm;
9546 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
9547 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
9552 if (env->ops->get_func_proto)
9553 fn = env->ops->get_func_proto(func_id, env->prog);
9555 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
9560 /* eBPF programs must be GPL compatible to use GPL-ed functions */
9561 if (!env->prog->gpl_compatible && fn->gpl_only) {
9562 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
9566 if (fn->allowed && !fn->allowed(env->prog)) {
9567 verbose(env, "helper call is not allowed in probe\n");
9571 if (!env->prog->aux->sleepable && fn->might_sleep) {
9572 verbose(env, "helper call might sleep in a non-sleepable prog\n");
9576 /* With LD_ABS/IND some JITs save/restore skb from r1. */
9577 changes_data = bpf_helper_changes_pkt_data(fn->func);
9578 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
9579 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
9580 func_id_name(func_id), func_id);
9584 memset(&meta, 0, sizeof(meta));
9585 meta.pkt_access = fn->pkt_access;
9587 err = check_func_proto(fn, func_id);
9589 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
9590 func_id_name(func_id), func_id);
9594 if (env->cur_state->active_rcu_lock) {
9595 if (fn->might_sleep) {
9596 verbose(env, "sleepable helper %s#%d in rcu_read_lock region\n",
9597 func_id_name(func_id), func_id);
9601 if (env->prog->aux->sleepable && is_storage_get_function(func_id))
9602 env->insn_aux_data[insn_idx].storage_get_func_atomic = true;
9605 meta.func_id = func_id;
9607 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
9608 err = check_func_arg(env, i, &meta, fn, insn_idx);
9613 err = record_func_map(env, &meta, func_id, insn_idx);
9617 err = record_func_key(env, &meta, func_id, insn_idx);
9621 /* Mark slots with STACK_MISC in case of raw mode, stack offset
9622 * is inferred from register state.
9624 for (i = 0; i < meta.access_size; i++) {
9625 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
9626 BPF_WRITE, -1, false, false);
9631 regs = cur_regs(env);
9633 if (meta.release_regno) {
9635 /* This can only be set for PTR_TO_STACK, as CONST_PTR_TO_DYNPTR cannot
9636 * be released by any dynptr helper. Hence, unmark_stack_slots_dynptr
9637 * is safe to do directly.
9639 if (arg_type_is_dynptr(fn->arg_type[meta.release_regno - BPF_REG_1])) {
9640 if (regs[meta.release_regno].type == CONST_PTR_TO_DYNPTR) {
9641 verbose(env, "verifier internal error: CONST_PTR_TO_DYNPTR cannot be released\n");
9644 err = unmark_stack_slots_dynptr(env, ®s[meta.release_regno]);
9645 } else if (meta.ref_obj_id) {
9646 err = release_reference(env, meta.ref_obj_id);
9647 } else if (register_is_null(®s[meta.release_regno])) {
9648 /* meta.ref_obj_id can only be 0 if register that is meant to be
9649 * released is NULL, which must be > R0.
9654 verbose(env, "func %s#%d reference has not been acquired before\n",
9655 func_id_name(func_id), func_id);
9661 case BPF_FUNC_tail_call:
9662 err = check_reference_leak(env);
9664 verbose(env, "tail_call would lead to reference leak\n");
9668 case BPF_FUNC_get_local_storage:
9669 /* check that flags argument in get_local_storage(map, flags) is 0,
9670 * this is required because get_local_storage() can't return an error.
9672 if (!register_is_null(®s[BPF_REG_2])) {
9673 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
9677 case BPF_FUNC_for_each_map_elem:
9678 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9679 set_map_elem_callback_state);
9681 case BPF_FUNC_timer_set_callback:
9682 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9683 set_timer_callback_state);
9685 case BPF_FUNC_find_vma:
9686 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9687 set_find_vma_callback_state);
9689 case BPF_FUNC_snprintf:
9690 err = check_bpf_snprintf_call(env, regs);
9693 update_loop_inline_state(env, meta.subprogno);
9694 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9695 set_loop_callback_state);
9697 case BPF_FUNC_dynptr_from_mem:
9698 if (regs[BPF_REG_1].type != PTR_TO_MAP_VALUE) {
9699 verbose(env, "Unsupported reg type %s for bpf_dynptr_from_mem data\n",
9700 reg_type_str(env, regs[BPF_REG_1].type));
9704 case BPF_FUNC_set_retval:
9705 if (prog_type == BPF_PROG_TYPE_LSM &&
9706 env->prog->expected_attach_type == BPF_LSM_CGROUP) {
9707 if (!env->prog->aux->attach_func_proto->type) {
9708 /* Make sure programs that attach to void
9709 * hooks don't try to modify return value.
9711 verbose(env, "BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
9716 case BPF_FUNC_dynptr_data:
9718 struct bpf_reg_state *reg;
9721 reg = get_dynptr_arg_reg(env, fn, regs);
9726 if (meta.dynptr_id) {
9727 verbose(env, "verifier internal error: meta.dynptr_id already set\n");
9730 if (meta.ref_obj_id) {
9731 verbose(env, "verifier internal error: meta.ref_obj_id already set\n");
9735 id = dynptr_id(env, reg);
9737 verbose(env, "verifier internal error: failed to obtain dynptr id\n");
9741 ref_obj_id = dynptr_ref_obj_id(env, reg);
9742 if (ref_obj_id < 0) {
9743 verbose(env, "verifier internal error: failed to obtain dynptr ref_obj_id\n");
9747 meta.dynptr_id = id;
9748 meta.ref_obj_id = ref_obj_id;
9752 case BPF_FUNC_dynptr_write:
9754 enum bpf_dynptr_type dynptr_type;
9755 struct bpf_reg_state *reg;
9757 reg = get_dynptr_arg_reg(env, fn, regs);
9761 dynptr_type = dynptr_get_type(env, reg);
9762 if (dynptr_type == BPF_DYNPTR_TYPE_INVALID)
9765 if (dynptr_type == BPF_DYNPTR_TYPE_SKB)
9766 /* this will trigger clear_all_pkt_pointers(), which will
9767 * invalidate all dynptr slices associated with the skb
9769 changes_data = true;
9773 case BPF_FUNC_user_ringbuf_drain:
9774 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9775 set_user_ringbuf_callback_state);
9782 /* reset caller saved regs */
9783 for (i = 0; i < CALLER_SAVED_REGS; i++) {
9784 mark_reg_not_init(env, regs, caller_saved[i]);
9785 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
9788 /* helper call returns 64-bit value. */
9789 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
9791 /* update return register (already marked as written above) */
9792 ret_type = fn->ret_type;
9793 ret_flag = type_flag(ret_type);
9795 switch (base_type(ret_type)) {
9797 /* sets type to SCALAR_VALUE */
9798 mark_reg_unknown(env, regs, BPF_REG_0);
9801 regs[BPF_REG_0].type = NOT_INIT;
9803 case RET_PTR_TO_MAP_VALUE:
9804 /* There is no offset yet applied, variable or fixed */
9805 mark_reg_known_zero(env, regs, BPF_REG_0);
9806 /* remember map_ptr, so that check_map_access()
9807 * can check 'value_size' boundary of memory access
9808 * to map element returned from bpf_map_lookup_elem()
9810 if (meta.map_ptr == NULL) {
9812 "kernel subsystem misconfigured verifier\n");
9815 regs[BPF_REG_0].map_ptr = meta.map_ptr;
9816 regs[BPF_REG_0].map_uid = meta.map_uid;
9817 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE | ret_flag;
9818 if (!type_may_be_null(ret_type) &&
9819 btf_record_has_field(meta.map_ptr->record, BPF_SPIN_LOCK)) {
9820 regs[BPF_REG_0].id = ++env->id_gen;
9823 case RET_PTR_TO_SOCKET:
9824 mark_reg_known_zero(env, regs, BPF_REG_0);
9825 regs[BPF_REG_0].type = PTR_TO_SOCKET | ret_flag;
9827 case RET_PTR_TO_SOCK_COMMON:
9828 mark_reg_known_zero(env, regs, BPF_REG_0);
9829 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON | ret_flag;
9831 case RET_PTR_TO_TCP_SOCK:
9832 mark_reg_known_zero(env, regs, BPF_REG_0);
9833 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK | ret_flag;
9835 case RET_PTR_TO_MEM:
9836 mark_reg_known_zero(env, regs, BPF_REG_0);
9837 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
9838 regs[BPF_REG_0].mem_size = meta.mem_size;
9840 case RET_PTR_TO_MEM_OR_BTF_ID:
9842 const struct btf_type *t;
9844 mark_reg_known_zero(env, regs, BPF_REG_0);
9845 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
9846 if (!btf_type_is_struct(t)) {
9848 const struct btf_type *ret;
9851 /* resolve the type size of ksym. */
9852 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
9854 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
9855 verbose(env, "unable to resolve the size of type '%s': %ld\n",
9856 tname, PTR_ERR(ret));
9859 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
9860 regs[BPF_REG_0].mem_size = tsize;
9862 /* MEM_RDONLY may be carried from ret_flag, but it
9863 * doesn't apply on PTR_TO_BTF_ID. Fold it, otherwise
9864 * it will confuse the check of PTR_TO_BTF_ID in
9865 * check_mem_access().
9867 ret_flag &= ~MEM_RDONLY;
9869 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
9870 regs[BPF_REG_0].btf = meta.ret_btf;
9871 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
9875 case RET_PTR_TO_BTF_ID:
9877 struct btf *ret_btf;
9880 mark_reg_known_zero(env, regs, BPF_REG_0);
9881 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
9882 if (func_id == BPF_FUNC_kptr_xchg) {
9883 ret_btf = meta.kptr_field->kptr.btf;
9884 ret_btf_id = meta.kptr_field->kptr.btf_id;
9885 if (!btf_is_kernel(ret_btf))
9886 regs[BPF_REG_0].type |= MEM_ALLOC;
9888 if (fn->ret_btf_id == BPF_PTR_POISON) {
9889 verbose(env, "verifier internal error:");
9890 verbose(env, "func %s has non-overwritten BPF_PTR_POISON return type\n",
9891 func_id_name(func_id));
9894 ret_btf = btf_vmlinux;
9895 ret_btf_id = *fn->ret_btf_id;
9897 if (ret_btf_id == 0) {
9898 verbose(env, "invalid return type %u of func %s#%d\n",
9899 base_type(ret_type), func_id_name(func_id),
9903 regs[BPF_REG_0].btf = ret_btf;
9904 regs[BPF_REG_0].btf_id = ret_btf_id;
9908 verbose(env, "unknown return type %u of func %s#%d\n",
9909 base_type(ret_type), func_id_name(func_id), func_id);
9913 if (type_may_be_null(regs[BPF_REG_0].type))
9914 regs[BPF_REG_0].id = ++env->id_gen;
9916 if (helper_multiple_ref_obj_use(func_id, meta.map_ptr)) {
9917 verbose(env, "verifier internal error: func %s#%d sets ref_obj_id more than once\n",
9918 func_id_name(func_id), func_id);
9922 if (is_dynptr_ref_function(func_id))
9923 regs[BPF_REG_0].dynptr_id = meta.dynptr_id;
9925 if (is_ptr_cast_function(func_id) || is_dynptr_ref_function(func_id)) {
9926 /* For release_reference() */
9927 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
9928 } else if (is_acquire_function(func_id, meta.map_ptr)) {
9929 int id = acquire_reference_state(env, insn_idx);
9933 /* For mark_ptr_or_null_reg() */
9934 regs[BPF_REG_0].id = id;
9935 /* For release_reference() */
9936 regs[BPF_REG_0].ref_obj_id = id;
9939 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
9941 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
9945 if ((func_id == BPF_FUNC_get_stack ||
9946 func_id == BPF_FUNC_get_task_stack) &&
9947 !env->prog->has_callchain_buf) {
9948 const char *err_str;
9950 #ifdef CONFIG_PERF_EVENTS
9951 err = get_callchain_buffers(sysctl_perf_event_max_stack);
9952 err_str = "cannot get callchain buffer for func %s#%d\n";
9955 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
9958 verbose(env, err_str, func_id_name(func_id), func_id);
9962 env->prog->has_callchain_buf = true;
9965 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
9966 env->prog->call_get_stack = true;
9968 if (func_id == BPF_FUNC_get_func_ip) {
9969 if (check_get_func_ip(env))
9971 env->prog->call_get_func_ip = true;
9975 clear_all_pkt_pointers(env);
9979 /* mark_btf_func_reg_size() is used when the reg size is determined by
9980 * the BTF func_proto's return value size and argument.
9982 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
9985 struct bpf_reg_state *reg = &cur_regs(env)[regno];
9987 if (regno == BPF_REG_0) {
9988 /* Function return value */
9989 reg->live |= REG_LIVE_WRITTEN;
9990 reg->subreg_def = reg_size == sizeof(u64) ?
9991 DEF_NOT_SUBREG : env->insn_idx + 1;
9993 /* Function argument */
9994 if (reg_size == sizeof(u64)) {
9995 mark_insn_zext(env, reg);
9996 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
9998 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
10003 static bool is_kfunc_acquire(struct bpf_kfunc_call_arg_meta *meta)
10005 return meta->kfunc_flags & KF_ACQUIRE;
10008 static bool is_kfunc_release(struct bpf_kfunc_call_arg_meta *meta)
10010 return meta->kfunc_flags & KF_RELEASE;
10013 static bool is_kfunc_trusted_args(struct bpf_kfunc_call_arg_meta *meta)
10015 return (meta->kfunc_flags & KF_TRUSTED_ARGS) || is_kfunc_release(meta);
10018 static bool is_kfunc_sleepable(struct bpf_kfunc_call_arg_meta *meta)
10020 return meta->kfunc_flags & KF_SLEEPABLE;
10023 static bool is_kfunc_destructive(struct bpf_kfunc_call_arg_meta *meta)
10025 return meta->kfunc_flags & KF_DESTRUCTIVE;
10028 static bool is_kfunc_rcu(struct bpf_kfunc_call_arg_meta *meta)
10030 return meta->kfunc_flags & KF_RCU;
10033 static bool __kfunc_param_match_suffix(const struct btf *btf,
10034 const struct btf_param *arg,
10035 const char *suffix)
10037 int suffix_len = strlen(suffix), len;
10038 const char *param_name;
10040 /* In the future, this can be ported to use BTF tagging */
10041 param_name = btf_name_by_offset(btf, arg->name_off);
10042 if (str_is_empty(param_name))
10044 len = strlen(param_name);
10045 if (len < suffix_len)
10047 param_name += len - suffix_len;
10048 return !strncmp(param_name, suffix, suffix_len);
10051 static bool is_kfunc_arg_mem_size(const struct btf *btf,
10052 const struct btf_param *arg,
10053 const struct bpf_reg_state *reg)
10055 const struct btf_type *t;
10057 t = btf_type_skip_modifiers(btf, arg->type, NULL);
10058 if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
10061 return __kfunc_param_match_suffix(btf, arg, "__sz");
10064 static bool is_kfunc_arg_const_mem_size(const struct btf *btf,
10065 const struct btf_param *arg,
10066 const struct bpf_reg_state *reg)
10068 const struct btf_type *t;
10070 t = btf_type_skip_modifiers(btf, arg->type, NULL);
10071 if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
10074 return __kfunc_param_match_suffix(btf, arg, "__szk");
10077 static bool is_kfunc_arg_optional(const struct btf *btf, const struct btf_param *arg)
10079 return __kfunc_param_match_suffix(btf, arg, "__opt");
10082 static bool is_kfunc_arg_constant(const struct btf *btf, const struct btf_param *arg)
10084 return __kfunc_param_match_suffix(btf, arg, "__k");
10087 static bool is_kfunc_arg_ignore(const struct btf *btf, const struct btf_param *arg)
10089 return __kfunc_param_match_suffix(btf, arg, "__ign");
10092 static bool is_kfunc_arg_alloc_obj(const struct btf *btf, const struct btf_param *arg)
10094 return __kfunc_param_match_suffix(btf, arg, "__alloc");
10097 static bool is_kfunc_arg_uninit(const struct btf *btf, const struct btf_param *arg)
10099 return __kfunc_param_match_suffix(btf, arg, "__uninit");
10102 static bool is_kfunc_arg_refcounted_kptr(const struct btf *btf, const struct btf_param *arg)
10104 return __kfunc_param_match_suffix(btf, arg, "__refcounted_kptr");
10107 static bool is_kfunc_arg_scalar_with_name(const struct btf *btf,
10108 const struct btf_param *arg,
10111 int len, target_len = strlen(name);
10112 const char *param_name;
10114 param_name = btf_name_by_offset(btf, arg->name_off);
10115 if (str_is_empty(param_name))
10117 len = strlen(param_name);
10118 if (len != target_len)
10120 if (strcmp(param_name, name))
10128 KF_ARG_LIST_HEAD_ID,
10129 KF_ARG_LIST_NODE_ID,
10134 BTF_ID_LIST(kf_arg_btf_ids)
10135 BTF_ID(struct, bpf_dynptr_kern)
10136 BTF_ID(struct, bpf_list_head)
10137 BTF_ID(struct, bpf_list_node)
10138 BTF_ID(struct, bpf_rb_root)
10139 BTF_ID(struct, bpf_rb_node)
10141 static bool __is_kfunc_ptr_arg_type(const struct btf *btf,
10142 const struct btf_param *arg, int type)
10144 const struct btf_type *t;
10147 t = btf_type_skip_modifiers(btf, arg->type, NULL);
10150 if (!btf_type_is_ptr(t))
10152 t = btf_type_skip_modifiers(btf, t->type, &res_id);
10155 return btf_types_are_same(btf, res_id, btf_vmlinux, kf_arg_btf_ids[type]);
10158 static bool is_kfunc_arg_dynptr(const struct btf *btf, const struct btf_param *arg)
10160 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_DYNPTR_ID);
10163 static bool is_kfunc_arg_list_head(const struct btf *btf, const struct btf_param *arg)
10165 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_HEAD_ID);
10168 static bool is_kfunc_arg_list_node(const struct btf *btf, const struct btf_param *arg)
10170 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_NODE_ID);
10173 static bool is_kfunc_arg_rbtree_root(const struct btf *btf, const struct btf_param *arg)
10175 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RB_ROOT_ID);
10178 static bool is_kfunc_arg_rbtree_node(const struct btf *btf, const struct btf_param *arg)
10180 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RB_NODE_ID);
10183 static bool is_kfunc_arg_callback(struct bpf_verifier_env *env, const struct btf *btf,
10184 const struct btf_param *arg)
10186 const struct btf_type *t;
10188 t = btf_type_resolve_func_ptr(btf, arg->type, NULL);
10195 /* Returns true if struct is composed of scalars, 4 levels of nesting allowed */
10196 static bool __btf_type_is_scalar_struct(struct bpf_verifier_env *env,
10197 const struct btf *btf,
10198 const struct btf_type *t, int rec)
10200 const struct btf_type *member_type;
10201 const struct btf_member *member;
10204 if (!btf_type_is_struct(t))
10207 for_each_member(i, t, member) {
10208 const struct btf_array *array;
10210 member_type = btf_type_skip_modifiers(btf, member->type, NULL);
10211 if (btf_type_is_struct(member_type)) {
10213 verbose(env, "max struct nesting depth exceeded\n");
10216 if (!__btf_type_is_scalar_struct(env, btf, member_type, rec + 1))
10220 if (btf_type_is_array(member_type)) {
10221 array = btf_array(member_type);
10222 if (!array->nelems)
10224 member_type = btf_type_skip_modifiers(btf, array->type, NULL);
10225 if (!btf_type_is_scalar(member_type))
10229 if (!btf_type_is_scalar(member_type))
10235 enum kfunc_ptr_arg_type {
10237 KF_ARG_PTR_TO_ALLOC_BTF_ID, /* Allocated object */
10238 KF_ARG_PTR_TO_REFCOUNTED_KPTR, /* Refcounted local kptr */
10239 KF_ARG_PTR_TO_DYNPTR,
10240 KF_ARG_PTR_TO_ITER,
10241 KF_ARG_PTR_TO_LIST_HEAD,
10242 KF_ARG_PTR_TO_LIST_NODE,
10243 KF_ARG_PTR_TO_BTF_ID, /* Also covers reg2btf_ids conversions */
10245 KF_ARG_PTR_TO_MEM_SIZE, /* Size derived from next argument, skip it */
10246 KF_ARG_PTR_TO_CALLBACK,
10247 KF_ARG_PTR_TO_RB_ROOT,
10248 KF_ARG_PTR_TO_RB_NODE,
10251 enum special_kfunc_type {
10252 KF_bpf_obj_new_impl,
10253 KF_bpf_obj_drop_impl,
10254 KF_bpf_refcount_acquire_impl,
10255 KF_bpf_list_push_front_impl,
10256 KF_bpf_list_push_back_impl,
10257 KF_bpf_list_pop_front,
10258 KF_bpf_list_pop_back,
10259 KF_bpf_cast_to_kern_ctx,
10260 KF_bpf_rdonly_cast,
10261 KF_bpf_rcu_read_lock,
10262 KF_bpf_rcu_read_unlock,
10263 KF_bpf_rbtree_remove,
10264 KF_bpf_rbtree_add_impl,
10265 KF_bpf_rbtree_first,
10266 KF_bpf_dynptr_from_skb,
10267 KF_bpf_dynptr_from_xdp,
10268 KF_bpf_dynptr_slice,
10269 KF_bpf_dynptr_slice_rdwr,
10270 KF_bpf_dynptr_clone,
10273 BTF_SET_START(special_kfunc_set)
10274 BTF_ID(func, bpf_obj_new_impl)
10275 BTF_ID(func, bpf_obj_drop_impl)
10276 BTF_ID(func, bpf_refcount_acquire_impl)
10277 BTF_ID(func, bpf_list_push_front_impl)
10278 BTF_ID(func, bpf_list_push_back_impl)
10279 BTF_ID(func, bpf_list_pop_front)
10280 BTF_ID(func, bpf_list_pop_back)
10281 BTF_ID(func, bpf_cast_to_kern_ctx)
10282 BTF_ID(func, bpf_rdonly_cast)
10283 BTF_ID(func, bpf_rbtree_remove)
10284 BTF_ID(func, bpf_rbtree_add_impl)
10285 BTF_ID(func, bpf_rbtree_first)
10286 BTF_ID(func, bpf_dynptr_from_skb)
10287 BTF_ID(func, bpf_dynptr_from_xdp)
10288 BTF_ID(func, bpf_dynptr_slice)
10289 BTF_ID(func, bpf_dynptr_slice_rdwr)
10290 BTF_ID(func, bpf_dynptr_clone)
10291 BTF_SET_END(special_kfunc_set)
10293 BTF_ID_LIST(special_kfunc_list)
10294 BTF_ID(func, bpf_obj_new_impl)
10295 BTF_ID(func, bpf_obj_drop_impl)
10296 BTF_ID(func, bpf_refcount_acquire_impl)
10297 BTF_ID(func, bpf_list_push_front_impl)
10298 BTF_ID(func, bpf_list_push_back_impl)
10299 BTF_ID(func, bpf_list_pop_front)
10300 BTF_ID(func, bpf_list_pop_back)
10301 BTF_ID(func, bpf_cast_to_kern_ctx)
10302 BTF_ID(func, bpf_rdonly_cast)
10303 BTF_ID(func, bpf_rcu_read_lock)
10304 BTF_ID(func, bpf_rcu_read_unlock)
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)
10314 static bool is_kfunc_ret_null(struct bpf_kfunc_call_arg_meta *meta)
10316 if (meta->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl] &&
10317 meta->arg_owning_ref) {
10321 return meta->kfunc_flags & KF_RET_NULL;
10324 static bool is_kfunc_bpf_rcu_read_lock(struct bpf_kfunc_call_arg_meta *meta)
10326 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_lock];
10329 static bool is_kfunc_bpf_rcu_read_unlock(struct bpf_kfunc_call_arg_meta *meta)
10331 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_unlock];
10334 static enum kfunc_ptr_arg_type
10335 get_kfunc_ptr_arg_type(struct bpf_verifier_env *env,
10336 struct bpf_kfunc_call_arg_meta *meta,
10337 const struct btf_type *t, const struct btf_type *ref_t,
10338 const char *ref_tname, const struct btf_param *args,
10339 int argno, int nargs)
10341 u32 regno = argno + 1;
10342 struct bpf_reg_state *regs = cur_regs(env);
10343 struct bpf_reg_state *reg = ®s[regno];
10344 bool arg_mem_size = false;
10346 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx])
10347 return KF_ARG_PTR_TO_CTX;
10349 /* In this function, we verify the kfunc's BTF as per the argument type,
10350 * leaving the rest of the verification with respect to the register
10351 * type to our caller. When a set of conditions hold in the BTF type of
10352 * arguments, we resolve it to a known kfunc_ptr_arg_type.
10354 if (btf_get_prog_ctx_type(&env->log, meta->btf, t, resolve_prog_type(env->prog), argno))
10355 return KF_ARG_PTR_TO_CTX;
10357 if (is_kfunc_arg_alloc_obj(meta->btf, &args[argno]))
10358 return KF_ARG_PTR_TO_ALLOC_BTF_ID;
10360 if (is_kfunc_arg_refcounted_kptr(meta->btf, &args[argno]))
10361 return KF_ARG_PTR_TO_REFCOUNTED_KPTR;
10363 if (is_kfunc_arg_dynptr(meta->btf, &args[argno]))
10364 return KF_ARG_PTR_TO_DYNPTR;
10366 if (is_kfunc_arg_iter(meta, argno))
10367 return KF_ARG_PTR_TO_ITER;
10369 if (is_kfunc_arg_list_head(meta->btf, &args[argno]))
10370 return KF_ARG_PTR_TO_LIST_HEAD;
10372 if (is_kfunc_arg_list_node(meta->btf, &args[argno]))
10373 return KF_ARG_PTR_TO_LIST_NODE;
10375 if (is_kfunc_arg_rbtree_root(meta->btf, &args[argno]))
10376 return KF_ARG_PTR_TO_RB_ROOT;
10378 if (is_kfunc_arg_rbtree_node(meta->btf, &args[argno]))
10379 return KF_ARG_PTR_TO_RB_NODE;
10381 if ((base_type(reg->type) == PTR_TO_BTF_ID || reg2btf_ids[base_type(reg->type)])) {
10382 if (!btf_type_is_struct(ref_t)) {
10383 verbose(env, "kernel function %s args#%d pointer type %s %s is not supported\n",
10384 meta->func_name, argno, btf_type_str(ref_t), ref_tname);
10387 return KF_ARG_PTR_TO_BTF_ID;
10390 if (is_kfunc_arg_callback(env, meta->btf, &args[argno]))
10391 return KF_ARG_PTR_TO_CALLBACK;
10394 if (argno + 1 < nargs &&
10395 (is_kfunc_arg_mem_size(meta->btf, &args[argno + 1], ®s[regno + 1]) ||
10396 is_kfunc_arg_const_mem_size(meta->btf, &args[argno + 1], ®s[regno + 1])))
10397 arg_mem_size = true;
10399 /* This is the catch all argument type of register types supported by
10400 * check_helper_mem_access. However, we only allow when argument type is
10401 * pointer to scalar, or struct composed (recursively) of scalars. When
10402 * arg_mem_size is true, the pointer can be void *.
10404 if (!btf_type_is_scalar(ref_t) && !__btf_type_is_scalar_struct(env, meta->btf, ref_t, 0) &&
10405 (arg_mem_size ? !btf_type_is_void(ref_t) : 1)) {
10406 verbose(env, "arg#%d pointer type %s %s must point to %sscalar, or struct with scalar\n",
10407 argno, btf_type_str(ref_t), ref_tname, arg_mem_size ? "void, " : "");
10410 return arg_mem_size ? KF_ARG_PTR_TO_MEM_SIZE : KF_ARG_PTR_TO_MEM;
10413 static int process_kf_arg_ptr_to_btf_id(struct bpf_verifier_env *env,
10414 struct bpf_reg_state *reg,
10415 const struct btf_type *ref_t,
10416 const char *ref_tname, u32 ref_id,
10417 struct bpf_kfunc_call_arg_meta *meta,
10420 const struct btf_type *reg_ref_t;
10421 bool strict_type_match = false;
10422 const struct btf *reg_btf;
10423 const char *reg_ref_tname;
10426 if (base_type(reg->type) == PTR_TO_BTF_ID) {
10427 reg_btf = reg->btf;
10428 reg_ref_id = reg->btf_id;
10430 reg_btf = btf_vmlinux;
10431 reg_ref_id = *reg2btf_ids[base_type(reg->type)];
10434 /* Enforce strict type matching for calls to kfuncs that are acquiring
10435 * or releasing a reference, or are no-cast aliases. We do _not_
10436 * enforce strict matching for plain KF_TRUSTED_ARGS kfuncs by default,
10437 * as we want to enable BPF programs to pass types that are bitwise
10438 * equivalent without forcing them to explicitly cast with something
10439 * like bpf_cast_to_kern_ctx().
10441 * For example, say we had a type like the following:
10443 * struct bpf_cpumask {
10444 * cpumask_t cpumask;
10445 * refcount_t usage;
10448 * Note that as specified in <linux/cpumask.h>, cpumask_t is typedef'ed
10449 * to a struct cpumask, so it would be safe to pass a struct
10450 * bpf_cpumask * to a kfunc expecting a struct cpumask *.
10452 * The philosophy here is similar to how we allow scalars of different
10453 * types to be passed to kfuncs as long as the size is the same. The
10454 * only difference here is that we're simply allowing
10455 * btf_struct_ids_match() to walk the struct at the 0th offset, and
10458 if (is_kfunc_acquire(meta) ||
10459 (is_kfunc_release(meta) && reg->ref_obj_id) ||
10460 btf_type_ids_nocast_alias(&env->log, reg_btf, reg_ref_id, meta->btf, ref_id))
10461 strict_type_match = true;
10463 WARN_ON_ONCE(is_kfunc_trusted_args(meta) && reg->off);
10465 reg_ref_t = btf_type_skip_modifiers(reg_btf, reg_ref_id, ®_ref_id);
10466 reg_ref_tname = btf_name_by_offset(reg_btf, reg_ref_t->name_off);
10467 if (!btf_struct_ids_match(&env->log, reg_btf, reg_ref_id, reg->off, meta->btf, ref_id, strict_type_match)) {
10468 verbose(env, "kernel function %s args#%d expected pointer to %s %s but R%d has a pointer to %s %s\n",
10469 meta->func_name, argno, btf_type_str(ref_t), ref_tname, argno + 1,
10470 btf_type_str(reg_ref_t), reg_ref_tname);
10476 static int ref_set_non_owning(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
10478 struct bpf_verifier_state *state = env->cur_state;
10479 struct btf_record *rec = reg_btf_record(reg);
10481 if (!state->active_lock.ptr) {
10482 verbose(env, "verifier internal error: ref_set_non_owning w/o active lock\n");
10486 if (type_flag(reg->type) & NON_OWN_REF) {
10487 verbose(env, "verifier internal error: NON_OWN_REF already set\n");
10491 reg->type |= NON_OWN_REF;
10492 if (rec->refcount_off >= 0)
10493 reg->type |= MEM_RCU;
10498 static int ref_convert_owning_non_owning(struct bpf_verifier_env *env, u32 ref_obj_id)
10500 struct bpf_func_state *state, *unused;
10501 struct bpf_reg_state *reg;
10504 state = cur_func(env);
10507 verbose(env, "verifier internal error: ref_obj_id is zero for "
10508 "owning -> non-owning conversion\n");
10512 for (i = 0; i < state->acquired_refs; i++) {
10513 if (state->refs[i].id != ref_obj_id)
10516 /* Clear ref_obj_id here so release_reference doesn't clobber
10519 bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
10520 if (reg->ref_obj_id == ref_obj_id) {
10521 reg->ref_obj_id = 0;
10522 ref_set_non_owning(env, reg);
10528 verbose(env, "verifier internal error: ref state missing for ref_obj_id\n");
10532 /* Implementation details:
10534 * Each register points to some region of memory, which we define as an
10535 * allocation. Each allocation may embed a bpf_spin_lock which protects any
10536 * special BPF objects (bpf_list_head, bpf_rb_root, etc.) part of the same
10537 * allocation. The lock and the data it protects are colocated in the same
10540 * Hence, everytime a register holds a pointer value pointing to such
10541 * allocation, the verifier preserves a unique reg->id for it.
10543 * The verifier remembers the lock 'ptr' and the lock 'id' whenever
10544 * bpf_spin_lock is called.
10546 * To enable this, lock state in the verifier captures two values:
10547 * active_lock.ptr = Register's type specific pointer
10548 * active_lock.id = A unique ID for each register pointer value
10550 * Currently, PTR_TO_MAP_VALUE and PTR_TO_BTF_ID | MEM_ALLOC are the two
10551 * supported register types.
10553 * The active_lock.ptr in case of map values is the reg->map_ptr, and in case of
10554 * allocated objects is the reg->btf pointer.
10556 * The active_lock.id is non-unique for maps supporting direct_value_addr, as we
10557 * can establish the provenance of the map value statically for each distinct
10558 * lookup into such maps. They always contain a single map value hence unique
10559 * IDs for each pseudo load pessimizes the algorithm and rejects valid programs.
10561 * So, in case of global variables, they use array maps with max_entries = 1,
10562 * hence their active_lock.ptr becomes map_ptr and id = 0 (since they all point
10563 * into the same map value as max_entries is 1, as described above).
10565 * In case of inner map lookups, the inner map pointer has same map_ptr as the
10566 * outer map pointer (in verifier context), but each lookup into an inner map
10567 * assigns a fresh reg->id to the lookup, so while lookups into distinct inner
10568 * maps from the same outer map share the same map_ptr as active_lock.ptr, they
10569 * will get different reg->id assigned to each lookup, hence different
10572 * In case of allocated objects, active_lock.ptr is the reg->btf, and the
10573 * reg->id is a unique ID preserved after the NULL pointer check on the pointer
10574 * returned from bpf_obj_new. Each allocation receives a new reg->id.
10576 static int check_reg_allocation_locked(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
10581 switch ((int)reg->type) {
10582 case PTR_TO_MAP_VALUE:
10583 ptr = reg->map_ptr;
10585 case PTR_TO_BTF_ID | MEM_ALLOC:
10589 verbose(env, "verifier internal error: unknown reg type for lock check\n");
10594 if (!env->cur_state->active_lock.ptr)
10596 if (env->cur_state->active_lock.ptr != ptr ||
10597 env->cur_state->active_lock.id != id) {
10598 verbose(env, "held lock and object are not in the same allocation\n");
10604 static bool is_bpf_list_api_kfunc(u32 btf_id)
10606 return btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
10607 btf_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
10608 btf_id == special_kfunc_list[KF_bpf_list_pop_front] ||
10609 btf_id == special_kfunc_list[KF_bpf_list_pop_back];
10612 static bool is_bpf_rbtree_api_kfunc(u32 btf_id)
10614 return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl] ||
10615 btf_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
10616 btf_id == special_kfunc_list[KF_bpf_rbtree_first];
10619 static bool is_bpf_graph_api_kfunc(u32 btf_id)
10621 return is_bpf_list_api_kfunc(btf_id) || is_bpf_rbtree_api_kfunc(btf_id) ||
10622 btf_id == special_kfunc_list[KF_bpf_refcount_acquire_impl];
10625 static bool is_callback_calling_kfunc(u32 btf_id)
10627 return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl];
10630 static bool is_rbtree_lock_required_kfunc(u32 btf_id)
10632 return is_bpf_rbtree_api_kfunc(btf_id);
10635 static bool check_kfunc_is_graph_root_api(struct bpf_verifier_env *env,
10636 enum btf_field_type head_field_type,
10641 switch (head_field_type) {
10642 case BPF_LIST_HEAD:
10643 ret = is_bpf_list_api_kfunc(kfunc_btf_id);
10646 ret = is_bpf_rbtree_api_kfunc(kfunc_btf_id);
10649 verbose(env, "verifier internal error: unexpected graph root argument type %s\n",
10650 btf_field_type_name(head_field_type));
10655 verbose(env, "verifier internal error: %s head arg for unknown kfunc\n",
10656 btf_field_type_name(head_field_type));
10660 static bool check_kfunc_is_graph_node_api(struct bpf_verifier_env *env,
10661 enum btf_field_type node_field_type,
10666 switch (node_field_type) {
10667 case BPF_LIST_NODE:
10668 ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
10669 kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_back_impl]);
10672 ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
10673 kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl]);
10676 verbose(env, "verifier internal error: unexpected graph node argument type %s\n",
10677 btf_field_type_name(node_field_type));
10682 verbose(env, "verifier internal error: %s node arg for unknown kfunc\n",
10683 btf_field_type_name(node_field_type));
10688 __process_kf_arg_ptr_to_graph_root(struct bpf_verifier_env *env,
10689 struct bpf_reg_state *reg, u32 regno,
10690 struct bpf_kfunc_call_arg_meta *meta,
10691 enum btf_field_type head_field_type,
10692 struct btf_field **head_field)
10694 const char *head_type_name;
10695 struct btf_field *field;
10696 struct btf_record *rec;
10699 if (meta->btf != btf_vmlinux) {
10700 verbose(env, "verifier internal error: unexpected btf mismatch in kfunc call\n");
10704 if (!check_kfunc_is_graph_root_api(env, head_field_type, meta->func_id))
10707 head_type_name = btf_field_type_name(head_field_type);
10708 if (!tnum_is_const(reg->var_off)) {
10710 "R%d doesn't have constant offset. %s has to be at the constant offset\n",
10711 regno, head_type_name);
10715 rec = reg_btf_record(reg);
10716 head_off = reg->off + reg->var_off.value;
10717 field = btf_record_find(rec, head_off, head_field_type);
10719 verbose(env, "%s not found at offset=%u\n", head_type_name, head_off);
10723 /* All functions require bpf_list_head to be protected using a bpf_spin_lock */
10724 if (check_reg_allocation_locked(env, reg)) {
10725 verbose(env, "bpf_spin_lock at off=%d must be held for %s\n",
10726 rec->spin_lock_off, head_type_name);
10731 verbose(env, "verifier internal error: repeating %s arg\n", head_type_name);
10734 *head_field = field;
10738 static int process_kf_arg_ptr_to_list_head(struct bpf_verifier_env *env,
10739 struct bpf_reg_state *reg, u32 regno,
10740 struct bpf_kfunc_call_arg_meta *meta)
10742 return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, BPF_LIST_HEAD,
10743 &meta->arg_list_head.field);
10746 static int process_kf_arg_ptr_to_rbtree_root(struct bpf_verifier_env *env,
10747 struct bpf_reg_state *reg, u32 regno,
10748 struct bpf_kfunc_call_arg_meta *meta)
10750 return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, BPF_RB_ROOT,
10751 &meta->arg_rbtree_root.field);
10755 __process_kf_arg_ptr_to_graph_node(struct bpf_verifier_env *env,
10756 struct bpf_reg_state *reg, u32 regno,
10757 struct bpf_kfunc_call_arg_meta *meta,
10758 enum btf_field_type head_field_type,
10759 enum btf_field_type node_field_type,
10760 struct btf_field **node_field)
10762 const char *node_type_name;
10763 const struct btf_type *et, *t;
10764 struct btf_field *field;
10767 if (meta->btf != btf_vmlinux) {
10768 verbose(env, "verifier internal error: unexpected btf mismatch in kfunc call\n");
10772 if (!check_kfunc_is_graph_node_api(env, node_field_type, meta->func_id))
10775 node_type_name = btf_field_type_name(node_field_type);
10776 if (!tnum_is_const(reg->var_off)) {
10778 "R%d doesn't have constant offset. %s has to be at the constant offset\n",
10779 regno, node_type_name);
10783 node_off = reg->off + reg->var_off.value;
10784 field = reg_find_field_offset(reg, node_off, node_field_type);
10785 if (!field || field->offset != node_off) {
10786 verbose(env, "%s not found at offset=%u\n", node_type_name, node_off);
10790 field = *node_field;
10792 et = btf_type_by_id(field->graph_root.btf, field->graph_root.value_btf_id);
10793 t = btf_type_by_id(reg->btf, reg->btf_id);
10794 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, 0, field->graph_root.btf,
10795 field->graph_root.value_btf_id, true)) {
10796 verbose(env, "operation on %s expects arg#1 %s at offset=%d "
10797 "in struct %s, but arg is at offset=%d in struct %s\n",
10798 btf_field_type_name(head_field_type),
10799 btf_field_type_name(node_field_type),
10800 field->graph_root.node_offset,
10801 btf_name_by_offset(field->graph_root.btf, et->name_off),
10802 node_off, btf_name_by_offset(reg->btf, t->name_off));
10805 meta->arg_btf = reg->btf;
10806 meta->arg_btf_id = reg->btf_id;
10808 if (node_off != field->graph_root.node_offset) {
10809 verbose(env, "arg#1 offset=%d, but expected %s at offset=%d in struct %s\n",
10810 node_off, btf_field_type_name(node_field_type),
10811 field->graph_root.node_offset,
10812 btf_name_by_offset(field->graph_root.btf, et->name_off));
10819 static int process_kf_arg_ptr_to_list_node(struct bpf_verifier_env *env,
10820 struct bpf_reg_state *reg, u32 regno,
10821 struct bpf_kfunc_call_arg_meta *meta)
10823 return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta,
10824 BPF_LIST_HEAD, BPF_LIST_NODE,
10825 &meta->arg_list_head.field);
10828 static int process_kf_arg_ptr_to_rbtree_node(struct bpf_verifier_env *env,
10829 struct bpf_reg_state *reg, u32 regno,
10830 struct bpf_kfunc_call_arg_meta *meta)
10832 return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta,
10833 BPF_RB_ROOT, BPF_RB_NODE,
10834 &meta->arg_rbtree_root.field);
10837 static int check_kfunc_args(struct bpf_verifier_env *env, struct bpf_kfunc_call_arg_meta *meta,
10840 const char *func_name = meta->func_name, *ref_tname;
10841 const struct btf *btf = meta->btf;
10842 const struct btf_param *args;
10843 struct btf_record *rec;
10847 args = (const struct btf_param *)(meta->func_proto + 1);
10848 nargs = btf_type_vlen(meta->func_proto);
10849 if (nargs > MAX_BPF_FUNC_REG_ARGS) {
10850 verbose(env, "Function %s has %d > %d args\n", func_name, nargs,
10851 MAX_BPF_FUNC_REG_ARGS);
10855 /* Check that BTF function arguments match actual types that the
10858 for (i = 0; i < nargs; i++) {
10859 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[i + 1];
10860 const struct btf_type *t, *ref_t, *resolve_ret;
10861 enum bpf_arg_type arg_type = ARG_DONTCARE;
10862 u32 regno = i + 1, ref_id, type_size;
10863 bool is_ret_buf_sz = false;
10866 t = btf_type_skip_modifiers(btf, args[i].type, NULL);
10868 if (is_kfunc_arg_ignore(btf, &args[i]))
10871 if (btf_type_is_scalar(t)) {
10872 if (reg->type != SCALAR_VALUE) {
10873 verbose(env, "R%d is not a scalar\n", regno);
10877 if (is_kfunc_arg_constant(meta->btf, &args[i])) {
10878 if (meta->arg_constant.found) {
10879 verbose(env, "verifier internal error: only one constant argument permitted\n");
10882 if (!tnum_is_const(reg->var_off)) {
10883 verbose(env, "R%d must be a known constant\n", regno);
10886 ret = mark_chain_precision(env, regno);
10889 meta->arg_constant.found = true;
10890 meta->arg_constant.value = reg->var_off.value;
10891 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdonly_buf_size")) {
10892 meta->r0_rdonly = true;
10893 is_ret_buf_sz = true;
10894 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdwr_buf_size")) {
10895 is_ret_buf_sz = true;
10898 if (is_ret_buf_sz) {
10899 if (meta->r0_size) {
10900 verbose(env, "2 or more rdonly/rdwr_buf_size parameters for kfunc");
10904 if (!tnum_is_const(reg->var_off)) {
10905 verbose(env, "R%d is not a const\n", regno);
10909 meta->r0_size = reg->var_off.value;
10910 ret = mark_chain_precision(env, regno);
10917 if (!btf_type_is_ptr(t)) {
10918 verbose(env, "Unrecognized arg#%d type %s\n", i, btf_type_str(t));
10922 if ((is_kfunc_trusted_args(meta) || is_kfunc_rcu(meta)) &&
10923 (register_is_null(reg) || type_may_be_null(reg->type))) {
10924 verbose(env, "Possibly NULL pointer passed to trusted arg%d\n", i);
10928 if (reg->ref_obj_id) {
10929 if (is_kfunc_release(meta) && meta->ref_obj_id) {
10930 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
10931 regno, reg->ref_obj_id,
10935 meta->ref_obj_id = reg->ref_obj_id;
10936 if (is_kfunc_release(meta))
10937 meta->release_regno = regno;
10940 ref_t = btf_type_skip_modifiers(btf, t->type, &ref_id);
10941 ref_tname = btf_name_by_offset(btf, ref_t->name_off);
10943 kf_arg_type = get_kfunc_ptr_arg_type(env, meta, t, ref_t, ref_tname, args, i, nargs);
10944 if (kf_arg_type < 0)
10945 return kf_arg_type;
10947 switch (kf_arg_type) {
10948 case KF_ARG_PTR_TO_ALLOC_BTF_ID:
10949 case KF_ARG_PTR_TO_BTF_ID:
10950 if (!is_kfunc_trusted_args(meta) && !is_kfunc_rcu(meta))
10953 if (!is_trusted_reg(reg)) {
10954 if (!is_kfunc_rcu(meta)) {
10955 verbose(env, "R%d must be referenced or trusted\n", regno);
10958 if (!is_rcu_reg(reg)) {
10959 verbose(env, "R%d must be a rcu pointer\n", regno);
10965 case KF_ARG_PTR_TO_CTX:
10966 /* Trusted arguments have the same offset checks as release arguments */
10967 arg_type |= OBJ_RELEASE;
10969 case KF_ARG_PTR_TO_DYNPTR:
10970 case KF_ARG_PTR_TO_ITER:
10971 case KF_ARG_PTR_TO_LIST_HEAD:
10972 case KF_ARG_PTR_TO_LIST_NODE:
10973 case KF_ARG_PTR_TO_RB_ROOT:
10974 case KF_ARG_PTR_TO_RB_NODE:
10975 case KF_ARG_PTR_TO_MEM:
10976 case KF_ARG_PTR_TO_MEM_SIZE:
10977 case KF_ARG_PTR_TO_CALLBACK:
10978 case KF_ARG_PTR_TO_REFCOUNTED_KPTR:
10979 /* Trusted by default */
10986 if (is_kfunc_release(meta) && reg->ref_obj_id)
10987 arg_type |= OBJ_RELEASE;
10988 ret = check_func_arg_reg_off(env, reg, regno, arg_type);
10992 switch (kf_arg_type) {
10993 case KF_ARG_PTR_TO_CTX:
10994 if (reg->type != PTR_TO_CTX) {
10995 verbose(env, "arg#%d expected pointer to ctx, but got %s\n", i, btf_type_str(t));
10999 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
11000 ret = get_kern_ctx_btf_id(&env->log, resolve_prog_type(env->prog));
11003 meta->ret_btf_id = ret;
11006 case KF_ARG_PTR_TO_ALLOC_BTF_ID:
11007 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11008 verbose(env, "arg#%d expected pointer to allocated object\n", i);
11011 if (!reg->ref_obj_id) {
11012 verbose(env, "allocated object must be referenced\n");
11015 if (meta->btf == btf_vmlinux &&
11016 meta->func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
11017 meta->arg_btf = reg->btf;
11018 meta->arg_btf_id = reg->btf_id;
11021 case KF_ARG_PTR_TO_DYNPTR:
11023 enum bpf_arg_type dynptr_arg_type = ARG_PTR_TO_DYNPTR;
11024 int clone_ref_obj_id = 0;
11026 if (reg->type != PTR_TO_STACK &&
11027 reg->type != CONST_PTR_TO_DYNPTR) {
11028 verbose(env, "arg#%d expected pointer to stack or dynptr_ptr\n", i);
11032 if (reg->type == CONST_PTR_TO_DYNPTR)
11033 dynptr_arg_type |= MEM_RDONLY;
11035 if (is_kfunc_arg_uninit(btf, &args[i]))
11036 dynptr_arg_type |= MEM_UNINIT;
11038 if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) {
11039 dynptr_arg_type |= DYNPTR_TYPE_SKB;
11040 } else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_xdp]) {
11041 dynptr_arg_type |= DYNPTR_TYPE_XDP;
11042 } else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_clone] &&
11043 (dynptr_arg_type & MEM_UNINIT)) {
11044 enum bpf_dynptr_type parent_type = meta->initialized_dynptr.type;
11046 if (parent_type == BPF_DYNPTR_TYPE_INVALID) {
11047 verbose(env, "verifier internal error: no dynptr type for parent of clone\n");
11051 dynptr_arg_type |= (unsigned int)get_dynptr_type_flag(parent_type);
11052 clone_ref_obj_id = meta->initialized_dynptr.ref_obj_id;
11053 if (dynptr_type_refcounted(parent_type) && !clone_ref_obj_id) {
11054 verbose(env, "verifier internal error: missing ref obj id for parent of clone\n");
11059 ret = process_dynptr_func(env, regno, insn_idx, dynptr_arg_type, clone_ref_obj_id);
11063 if (!(dynptr_arg_type & MEM_UNINIT)) {
11064 int id = dynptr_id(env, reg);
11067 verbose(env, "verifier internal error: failed to obtain dynptr id\n");
11070 meta->initialized_dynptr.id = id;
11071 meta->initialized_dynptr.type = dynptr_get_type(env, reg);
11072 meta->initialized_dynptr.ref_obj_id = dynptr_ref_obj_id(env, reg);
11077 case KF_ARG_PTR_TO_ITER:
11078 ret = process_iter_arg(env, regno, insn_idx, meta);
11082 case KF_ARG_PTR_TO_LIST_HEAD:
11083 if (reg->type != PTR_TO_MAP_VALUE &&
11084 reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11085 verbose(env, "arg#%d expected pointer to map value or allocated object\n", i);
11088 if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
11089 verbose(env, "allocated object must be referenced\n");
11092 ret = process_kf_arg_ptr_to_list_head(env, reg, regno, meta);
11096 case KF_ARG_PTR_TO_RB_ROOT:
11097 if (reg->type != PTR_TO_MAP_VALUE &&
11098 reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11099 verbose(env, "arg#%d expected pointer to map value or allocated object\n", i);
11102 if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
11103 verbose(env, "allocated object must be referenced\n");
11106 ret = process_kf_arg_ptr_to_rbtree_root(env, reg, regno, meta);
11110 case KF_ARG_PTR_TO_LIST_NODE:
11111 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11112 verbose(env, "arg#%d expected pointer to allocated object\n", i);
11115 if (!reg->ref_obj_id) {
11116 verbose(env, "allocated object must be referenced\n");
11119 ret = process_kf_arg_ptr_to_list_node(env, reg, regno, meta);
11123 case KF_ARG_PTR_TO_RB_NODE:
11124 if (meta->func_id == special_kfunc_list[KF_bpf_rbtree_remove]) {
11125 if (!type_is_non_owning_ref(reg->type) || reg->ref_obj_id) {
11126 verbose(env, "rbtree_remove node input must be non-owning ref\n");
11129 if (in_rbtree_lock_required_cb(env)) {
11130 verbose(env, "rbtree_remove not allowed in rbtree cb\n");
11134 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11135 verbose(env, "arg#%d expected pointer to allocated object\n", i);
11138 if (!reg->ref_obj_id) {
11139 verbose(env, "allocated object must be referenced\n");
11144 ret = process_kf_arg_ptr_to_rbtree_node(env, reg, regno, meta);
11148 case KF_ARG_PTR_TO_BTF_ID:
11149 /* Only base_type is checked, further checks are done here */
11150 if ((base_type(reg->type) != PTR_TO_BTF_ID ||
11151 (bpf_type_has_unsafe_modifiers(reg->type) && !is_rcu_reg(reg))) &&
11152 !reg2btf_ids[base_type(reg->type)]) {
11153 verbose(env, "arg#%d is %s ", i, reg_type_str(env, reg->type));
11154 verbose(env, "expected %s or socket\n",
11155 reg_type_str(env, base_type(reg->type) |
11156 (type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS)));
11159 ret = process_kf_arg_ptr_to_btf_id(env, reg, ref_t, ref_tname, ref_id, meta, i);
11163 case KF_ARG_PTR_TO_MEM:
11164 resolve_ret = btf_resolve_size(btf, ref_t, &type_size);
11165 if (IS_ERR(resolve_ret)) {
11166 verbose(env, "arg#%d reference type('%s %s') size cannot be determined: %ld\n",
11167 i, btf_type_str(ref_t), ref_tname, PTR_ERR(resolve_ret));
11170 ret = check_mem_reg(env, reg, regno, type_size);
11174 case KF_ARG_PTR_TO_MEM_SIZE:
11176 struct bpf_reg_state *buff_reg = ®s[regno];
11177 const struct btf_param *buff_arg = &args[i];
11178 struct bpf_reg_state *size_reg = ®s[regno + 1];
11179 const struct btf_param *size_arg = &args[i + 1];
11181 if (!register_is_null(buff_reg) || !is_kfunc_arg_optional(meta->btf, buff_arg)) {
11182 ret = check_kfunc_mem_size_reg(env, size_reg, regno + 1);
11184 verbose(env, "arg#%d arg#%d memory, len pair leads to invalid memory access\n", i, i + 1);
11189 if (is_kfunc_arg_const_mem_size(meta->btf, size_arg, size_reg)) {
11190 if (meta->arg_constant.found) {
11191 verbose(env, "verifier internal error: only one constant argument permitted\n");
11194 if (!tnum_is_const(size_reg->var_off)) {
11195 verbose(env, "R%d must be a known constant\n", regno + 1);
11198 meta->arg_constant.found = true;
11199 meta->arg_constant.value = size_reg->var_off.value;
11202 /* Skip next '__sz' or '__szk' argument */
11206 case KF_ARG_PTR_TO_CALLBACK:
11207 meta->subprogno = reg->subprogno;
11209 case KF_ARG_PTR_TO_REFCOUNTED_KPTR:
11210 if (!type_is_ptr_alloc_obj(reg->type)) {
11211 verbose(env, "arg#%d is neither owning or non-owning ref\n", i);
11214 if (!type_is_non_owning_ref(reg->type))
11215 meta->arg_owning_ref = true;
11217 rec = reg_btf_record(reg);
11219 verbose(env, "verifier internal error: Couldn't find btf_record\n");
11223 if (rec->refcount_off < 0) {
11224 verbose(env, "arg#%d doesn't point to a type with bpf_refcount field\n", i);
11228 meta->arg_btf = reg->btf;
11229 meta->arg_btf_id = reg->btf_id;
11234 if (is_kfunc_release(meta) && !meta->release_regno) {
11235 verbose(env, "release kernel function %s expects refcounted PTR_TO_BTF_ID\n",
11243 static int fetch_kfunc_meta(struct bpf_verifier_env *env,
11244 struct bpf_insn *insn,
11245 struct bpf_kfunc_call_arg_meta *meta,
11246 const char **kfunc_name)
11248 const struct btf_type *func, *func_proto;
11249 u32 func_id, *kfunc_flags;
11250 const char *func_name;
11251 struct btf *desc_btf;
11254 *kfunc_name = NULL;
11259 desc_btf = find_kfunc_desc_btf(env, insn->off);
11260 if (IS_ERR(desc_btf))
11261 return PTR_ERR(desc_btf);
11263 func_id = insn->imm;
11264 func = btf_type_by_id(desc_btf, func_id);
11265 func_name = btf_name_by_offset(desc_btf, func->name_off);
11267 *kfunc_name = func_name;
11268 func_proto = btf_type_by_id(desc_btf, func->type);
11270 kfunc_flags = btf_kfunc_id_set_contains(desc_btf, func_id, env->prog);
11271 if (!kfunc_flags) {
11275 memset(meta, 0, sizeof(*meta));
11276 meta->btf = desc_btf;
11277 meta->func_id = func_id;
11278 meta->kfunc_flags = *kfunc_flags;
11279 meta->func_proto = func_proto;
11280 meta->func_name = func_name;
11285 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
11288 const struct btf_type *t, *ptr_type;
11289 u32 i, nargs, ptr_type_id, release_ref_obj_id;
11290 struct bpf_reg_state *regs = cur_regs(env);
11291 const char *func_name, *ptr_type_name;
11292 bool sleepable, rcu_lock, rcu_unlock;
11293 struct bpf_kfunc_call_arg_meta meta;
11294 struct bpf_insn_aux_data *insn_aux;
11295 int err, insn_idx = *insn_idx_p;
11296 const struct btf_param *args;
11297 const struct btf_type *ret_t;
11298 struct btf *desc_btf;
11300 /* skip for now, but return error when we find this in fixup_kfunc_call */
11304 err = fetch_kfunc_meta(env, insn, &meta, &func_name);
11305 if (err == -EACCES && func_name)
11306 verbose(env, "calling kernel function %s is not allowed\n", func_name);
11309 desc_btf = meta.btf;
11310 insn_aux = &env->insn_aux_data[insn_idx];
11312 insn_aux->is_iter_next = is_iter_next_kfunc(&meta);
11314 if (is_kfunc_destructive(&meta) && !capable(CAP_SYS_BOOT)) {
11315 verbose(env, "destructive kfunc calls require CAP_SYS_BOOT capability\n");
11319 sleepable = is_kfunc_sleepable(&meta);
11320 if (sleepable && !env->prog->aux->sleepable) {
11321 verbose(env, "program must be sleepable to call sleepable kfunc %s\n", func_name);
11325 rcu_lock = is_kfunc_bpf_rcu_read_lock(&meta);
11326 rcu_unlock = is_kfunc_bpf_rcu_read_unlock(&meta);
11328 if (env->cur_state->active_rcu_lock) {
11329 struct bpf_func_state *state;
11330 struct bpf_reg_state *reg;
11332 if (in_rbtree_lock_required_cb(env) && (rcu_lock || rcu_unlock)) {
11333 verbose(env, "Calling bpf_rcu_read_{lock,unlock} in unnecessary rbtree callback\n");
11338 verbose(env, "nested rcu read lock (kernel function %s)\n", func_name);
11340 } else if (rcu_unlock) {
11341 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
11342 if (reg->type & MEM_RCU) {
11343 reg->type &= ~(MEM_RCU | PTR_MAYBE_NULL);
11344 reg->type |= PTR_UNTRUSTED;
11347 env->cur_state->active_rcu_lock = false;
11348 } else if (sleepable) {
11349 verbose(env, "kernel func %s is sleepable within rcu_read_lock region\n", func_name);
11352 } else if (rcu_lock) {
11353 env->cur_state->active_rcu_lock = true;
11354 } else if (rcu_unlock) {
11355 verbose(env, "unmatched rcu read unlock (kernel function %s)\n", func_name);
11359 /* Check the arguments */
11360 err = check_kfunc_args(env, &meta, insn_idx);
11363 /* In case of release function, we get register number of refcounted
11364 * PTR_TO_BTF_ID in bpf_kfunc_arg_meta, do the release now.
11366 if (meta.release_regno) {
11367 err = release_reference(env, regs[meta.release_regno].ref_obj_id);
11369 verbose(env, "kfunc %s#%d reference has not been acquired before\n",
11370 func_name, meta.func_id);
11375 if (meta.func_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
11376 meta.func_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
11377 meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
11378 release_ref_obj_id = regs[BPF_REG_2].ref_obj_id;
11379 insn_aux->insert_off = regs[BPF_REG_2].off;
11380 insn_aux->kptr_struct_meta = btf_find_struct_meta(meta.arg_btf, meta.arg_btf_id);
11381 err = ref_convert_owning_non_owning(env, release_ref_obj_id);
11383 verbose(env, "kfunc %s#%d conversion of owning ref to non-owning failed\n",
11384 func_name, meta.func_id);
11388 err = release_reference(env, release_ref_obj_id);
11390 verbose(env, "kfunc %s#%d reference has not been acquired before\n",
11391 func_name, meta.func_id);
11396 if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
11397 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
11398 set_rbtree_add_callback_state);
11400 verbose(env, "kfunc %s#%d failed callback verification\n",
11401 func_name, meta.func_id);
11406 for (i = 0; i < CALLER_SAVED_REGS; i++)
11407 mark_reg_not_init(env, regs, caller_saved[i]);
11409 /* Check return type */
11410 t = btf_type_skip_modifiers(desc_btf, meta.func_proto->type, NULL);
11412 if (is_kfunc_acquire(&meta) && !btf_type_is_struct_ptr(meta.btf, t)) {
11413 /* Only exception is bpf_obj_new_impl */
11414 if (meta.btf != btf_vmlinux ||
11415 (meta.func_id != special_kfunc_list[KF_bpf_obj_new_impl] &&
11416 meta.func_id != special_kfunc_list[KF_bpf_refcount_acquire_impl])) {
11417 verbose(env, "acquire kernel function does not return PTR_TO_BTF_ID\n");
11422 if (btf_type_is_scalar(t)) {
11423 mark_reg_unknown(env, regs, BPF_REG_0);
11424 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
11425 } else if (btf_type_is_ptr(t)) {
11426 ptr_type = btf_type_skip_modifiers(desc_btf, t->type, &ptr_type_id);
11428 if (meta.btf == btf_vmlinux && btf_id_set_contains(&special_kfunc_set, meta.func_id)) {
11429 if (meta.func_id == special_kfunc_list[KF_bpf_obj_new_impl]) {
11430 struct btf *ret_btf;
11433 if (unlikely(!bpf_global_ma_set))
11436 if (((u64)(u32)meta.arg_constant.value) != meta.arg_constant.value) {
11437 verbose(env, "local type ID argument must be in range [0, U32_MAX]\n");
11441 ret_btf = env->prog->aux->btf;
11442 ret_btf_id = meta.arg_constant.value;
11444 /* This may be NULL due to user not supplying a BTF */
11446 verbose(env, "bpf_obj_new requires prog BTF\n");
11450 ret_t = btf_type_by_id(ret_btf, ret_btf_id);
11451 if (!ret_t || !__btf_type_is_struct(ret_t)) {
11452 verbose(env, "bpf_obj_new type ID argument must be of a struct\n");
11456 mark_reg_known_zero(env, regs, BPF_REG_0);
11457 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
11458 regs[BPF_REG_0].btf = ret_btf;
11459 regs[BPF_REG_0].btf_id = ret_btf_id;
11461 insn_aux->obj_new_size = ret_t->size;
11462 insn_aux->kptr_struct_meta =
11463 btf_find_struct_meta(ret_btf, ret_btf_id);
11464 } else if (meta.func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) {
11465 mark_reg_known_zero(env, regs, BPF_REG_0);
11466 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
11467 regs[BPF_REG_0].btf = meta.arg_btf;
11468 regs[BPF_REG_0].btf_id = meta.arg_btf_id;
11470 insn_aux->kptr_struct_meta =
11471 btf_find_struct_meta(meta.arg_btf,
11473 } else if (meta.func_id == special_kfunc_list[KF_bpf_list_pop_front] ||
11474 meta.func_id == special_kfunc_list[KF_bpf_list_pop_back]) {
11475 struct btf_field *field = meta.arg_list_head.field;
11477 mark_reg_graph_node(regs, BPF_REG_0, &field->graph_root);
11478 } else if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
11479 meta.func_id == special_kfunc_list[KF_bpf_rbtree_first]) {
11480 struct btf_field *field = meta.arg_rbtree_root.field;
11482 mark_reg_graph_node(regs, BPF_REG_0, &field->graph_root);
11483 } else if (meta.func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
11484 mark_reg_known_zero(env, regs, BPF_REG_0);
11485 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_TRUSTED;
11486 regs[BPF_REG_0].btf = desc_btf;
11487 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
11488 } else if (meta.func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
11489 ret_t = btf_type_by_id(desc_btf, meta.arg_constant.value);
11490 if (!ret_t || !btf_type_is_struct(ret_t)) {
11492 "kfunc bpf_rdonly_cast type ID argument must be of a struct\n");
11496 mark_reg_known_zero(env, regs, BPF_REG_0);
11497 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
11498 regs[BPF_REG_0].btf = desc_btf;
11499 regs[BPF_REG_0].btf_id = meta.arg_constant.value;
11500 } else if (meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice] ||
11501 meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice_rdwr]) {
11502 enum bpf_type_flag type_flag = get_dynptr_type_flag(meta.initialized_dynptr.type);
11504 mark_reg_known_zero(env, regs, BPF_REG_0);
11506 if (!meta.arg_constant.found) {
11507 verbose(env, "verifier internal error: bpf_dynptr_slice(_rdwr) no constant size\n");
11511 regs[BPF_REG_0].mem_size = meta.arg_constant.value;
11513 /* PTR_MAYBE_NULL will be added when is_kfunc_ret_null is checked */
11514 regs[BPF_REG_0].type = PTR_TO_MEM | type_flag;
11516 if (meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice]) {
11517 regs[BPF_REG_0].type |= MEM_RDONLY;
11519 /* this will set env->seen_direct_write to true */
11520 if (!may_access_direct_pkt_data(env, NULL, BPF_WRITE)) {
11521 verbose(env, "the prog does not allow writes to packet data\n");
11526 if (!meta.initialized_dynptr.id) {
11527 verbose(env, "verifier internal error: no dynptr id\n");
11530 regs[BPF_REG_0].dynptr_id = meta.initialized_dynptr.id;
11532 /* we don't need to set BPF_REG_0's ref obj id
11533 * because packet slices are not refcounted (see
11534 * dynptr_type_refcounted)
11537 verbose(env, "kernel function %s unhandled dynamic return type\n",
11541 } else if (!__btf_type_is_struct(ptr_type)) {
11542 if (!meta.r0_size) {
11545 if (!IS_ERR(btf_resolve_size(desc_btf, ptr_type, &sz))) {
11547 meta.r0_rdonly = true;
11550 if (!meta.r0_size) {
11551 ptr_type_name = btf_name_by_offset(desc_btf,
11552 ptr_type->name_off);
11554 "kernel function %s returns pointer type %s %s is not supported\n",
11556 btf_type_str(ptr_type),
11561 mark_reg_known_zero(env, regs, BPF_REG_0);
11562 regs[BPF_REG_0].type = PTR_TO_MEM;
11563 regs[BPF_REG_0].mem_size = meta.r0_size;
11565 if (meta.r0_rdonly)
11566 regs[BPF_REG_0].type |= MEM_RDONLY;
11568 /* Ensures we don't access the memory after a release_reference() */
11569 if (meta.ref_obj_id)
11570 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
11572 mark_reg_known_zero(env, regs, BPF_REG_0);
11573 regs[BPF_REG_0].btf = desc_btf;
11574 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
11575 regs[BPF_REG_0].btf_id = ptr_type_id;
11578 if (is_kfunc_ret_null(&meta)) {
11579 regs[BPF_REG_0].type |= PTR_MAYBE_NULL;
11580 /* For mark_ptr_or_null_reg, see 93c230e3f5bd6 */
11581 regs[BPF_REG_0].id = ++env->id_gen;
11583 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
11584 if (is_kfunc_acquire(&meta)) {
11585 int id = acquire_reference_state(env, insn_idx);
11589 if (is_kfunc_ret_null(&meta))
11590 regs[BPF_REG_0].id = id;
11591 regs[BPF_REG_0].ref_obj_id = id;
11592 } else if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_first]) {
11593 ref_set_non_owning(env, ®s[BPF_REG_0]);
11596 if (reg_may_point_to_spin_lock(®s[BPF_REG_0]) && !regs[BPF_REG_0].id)
11597 regs[BPF_REG_0].id = ++env->id_gen;
11598 } else if (btf_type_is_void(t)) {
11599 if (meta.btf == btf_vmlinux && btf_id_set_contains(&special_kfunc_set, meta.func_id)) {
11600 if (meta.func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
11601 insn_aux->kptr_struct_meta =
11602 btf_find_struct_meta(meta.arg_btf,
11608 nargs = btf_type_vlen(meta.func_proto);
11609 args = (const struct btf_param *)(meta.func_proto + 1);
11610 for (i = 0; i < nargs; i++) {
11613 t = btf_type_skip_modifiers(desc_btf, args[i].type, NULL);
11614 if (btf_type_is_ptr(t))
11615 mark_btf_func_reg_size(env, regno, sizeof(void *));
11617 /* scalar. ensured by btf_check_kfunc_arg_match() */
11618 mark_btf_func_reg_size(env, regno, t->size);
11621 if (is_iter_next_kfunc(&meta)) {
11622 err = process_iter_next_call(env, insn_idx, &meta);
11630 static bool signed_add_overflows(s64 a, s64 b)
11632 /* Do the add in u64, where overflow is well-defined */
11633 s64 res = (s64)((u64)a + (u64)b);
11640 static bool signed_add32_overflows(s32 a, s32 b)
11642 /* Do the add in u32, where overflow is well-defined */
11643 s32 res = (s32)((u32)a + (u32)b);
11650 static bool signed_sub_overflows(s64 a, s64 b)
11652 /* Do the sub in u64, where overflow is well-defined */
11653 s64 res = (s64)((u64)a - (u64)b);
11660 static bool signed_sub32_overflows(s32 a, s32 b)
11662 /* Do the sub in u32, where overflow is well-defined */
11663 s32 res = (s32)((u32)a - (u32)b);
11670 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
11671 const struct bpf_reg_state *reg,
11672 enum bpf_reg_type type)
11674 bool known = tnum_is_const(reg->var_off);
11675 s64 val = reg->var_off.value;
11676 s64 smin = reg->smin_value;
11678 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
11679 verbose(env, "math between %s pointer and %lld is not allowed\n",
11680 reg_type_str(env, type), val);
11684 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
11685 verbose(env, "%s pointer offset %d is not allowed\n",
11686 reg_type_str(env, type), reg->off);
11690 if (smin == S64_MIN) {
11691 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
11692 reg_type_str(env, type));
11696 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
11697 verbose(env, "value %lld makes %s pointer be out of bounds\n",
11698 smin, reg_type_str(env, type));
11706 REASON_BOUNDS = -1,
11713 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
11714 u32 *alu_limit, bool mask_to_left)
11716 u32 max = 0, ptr_limit = 0;
11718 switch (ptr_reg->type) {
11720 /* Offset 0 is out-of-bounds, but acceptable start for the
11721 * left direction, see BPF_REG_FP. Also, unknown scalar
11722 * offset where we would need to deal with min/max bounds is
11723 * currently prohibited for unprivileged.
11725 max = MAX_BPF_STACK + mask_to_left;
11726 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
11728 case PTR_TO_MAP_VALUE:
11729 max = ptr_reg->map_ptr->value_size;
11730 ptr_limit = (mask_to_left ?
11731 ptr_reg->smin_value :
11732 ptr_reg->umax_value) + ptr_reg->off;
11735 return REASON_TYPE;
11738 if (ptr_limit >= max)
11739 return REASON_LIMIT;
11740 *alu_limit = ptr_limit;
11744 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
11745 const struct bpf_insn *insn)
11747 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
11750 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
11751 u32 alu_state, u32 alu_limit)
11753 /* If we arrived here from different branches with different
11754 * state or limits to sanitize, then this won't work.
11756 if (aux->alu_state &&
11757 (aux->alu_state != alu_state ||
11758 aux->alu_limit != alu_limit))
11759 return REASON_PATHS;
11761 /* Corresponding fixup done in do_misc_fixups(). */
11762 aux->alu_state = alu_state;
11763 aux->alu_limit = alu_limit;
11767 static int sanitize_val_alu(struct bpf_verifier_env *env,
11768 struct bpf_insn *insn)
11770 struct bpf_insn_aux_data *aux = cur_aux(env);
11772 if (can_skip_alu_sanitation(env, insn))
11775 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
11778 static bool sanitize_needed(u8 opcode)
11780 return opcode == BPF_ADD || opcode == BPF_SUB;
11783 struct bpf_sanitize_info {
11784 struct bpf_insn_aux_data aux;
11788 static struct bpf_verifier_state *
11789 sanitize_speculative_path(struct bpf_verifier_env *env,
11790 const struct bpf_insn *insn,
11791 u32 next_idx, u32 curr_idx)
11793 struct bpf_verifier_state *branch;
11794 struct bpf_reg_state *regs;
11796 branch = push_stack(env, next_idx, curr_idx, true);
11797 if (branch && insn) {
11798 regs = branch->frame[branch->curframe]->regs;
11799 if (BPF_SRC(insn->code) == BPF_K) {
11800 mark_reg_unknown(env, regs, insn->dst_reg);
11801 } else if (BPF_SRC(insn->code) == BPF_X) {
11802 mark_reg_unknown(env, regs, insn->dst_reg);
11803 mark_reg_unknown(env, regs, insn->src_reg);
11809 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
11810 struct bpf_insn *insn,
11811 const struct bpf_reg_state *ptr_reg,
11812 const struct bpf_reg_state *off_reg,
11813 struct bpf_reg_state *dst_reg,
11814 struct bpf_sanitize_info *info,
11815 const bool commit_window)
11817 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
11818 struct bpf_verifier_state *vstate = env->cur_state;
11819 bool off_is_imm = tnum_is_const(off_reg->var_off);
11820 bool off_is_neg = off_reg->smin_value < 0;
11821 bool ptr_is_dst_reg = ptr_reg == dst_reg;
11822 u8 opcode = BPF_OP(insn->code);
11823 u32 alu_state, alu_limit;
11824 struct bpf_reg_state tmp;
11828 if (can_skip_alu_sanitation(env, insn))
11831 /* We already marked aux for masking from non-speculative
11832 * paths, thus we got here in the first place. We only care
11833 * to explore bad access from here.
11835 if (vstate->speculative)
11838 if (!commit_window) {
11839 if (!tnum_is_const(off_reg->var_off) &&
11840 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
11841 return REASON_BOUNDS;
11843 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
11844 (opcode == BPF_SUB && !off_is_neg);
11847 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
11851 if (commit_window) {
11852 /* In commit phase we narrow the masking window based on
11853 * the observed pointer move after the simulated operation.
11855 alu_state = info->aux.alu_state;
11856 alu_limit = abs(info->aux.alu_limit - alu_limit);
11858 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
11859 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
11860 alu_state |= ptr_is_dst_reg ?
11861 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
11863 /* Limit pruning on unknown scalars to enable deep search for
11864 * potential masking differences from other program paths.
11867 env->explore_alu_limits = true;
11870 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
11874 /* If we're in commit phase, we're done here given we already
11875 * pushed the truncated dst_reg into the speculative verification
11878 * Also, when register is a known constant, we rewrite register-based
11879 * operation to immediate-based, and thus do not need masking (and as
11880 * a consequence, do not need to simulate the zero-truncation either).
11882 if (commit_window || off_is_imm)
11885 /* Simulate and find potential out-of-bounds access under
11886 * speculative execution from truncation as a result of
11887 * masking when off was not within expected range. If off
11888 * sits in dst, then we temporarily need to move ptr there
11889 * to simulate dst (== 0) +/-= ptr. Needed, for example,
11890 * for cases where we use K-based arithmetic in one direction
11891 * and truncated reg-based in the other in order to explore
11894 if (!ptr_is_dst_reg) {
11896 copy_register_state(dst_reg, ptr_reg);
11898 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
11900 if (!ptr_is_dst_reg && ret)
11902 return !ret ? REASON_STACK : 0;
11905 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
11907 struct bpf_verifier_state *vstate = env->cur_state;
11909 /* If we simulate paths under speculation, we don't update the
11910 * insn as 'seen' such that when we verify unreachable paths in
11911 * the non-speculative domain, sanitize_dead_code() can still
11912 * rewrite/sanitize them.
11914 if (!vstate->speculative)
11915 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
11918 static int sanitize_err(struct bpf_verifier_env *env,
11919 const struct bpf_insn *insn, int reason,
11920 const struct bpf_reg_state *off_reg,
11921 const struct bpf_reg_state *dst_reg)
11923 static const char *err = "pointer arithmetic with it prohibited for !root";
11924 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
11925 u32 dst = insn->dst_reg, src = insn->src_reg;
11928 case REASON_BOUNDS:
11929 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
11930 off_reg == dst_reg ? dst : src, err);
11933 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
11934 off_reg == dst_reg ? src : dst, err);
11937 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
11941 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
11945 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
11949 verbose(env, "verifier internal error: unknown reason (%d)\n",
11957 /* check that stack access falls within stack limits and that 'reg' doesn't
11958 * have a variable offset.
11960 * Variable offset is prohibited for unprivileged mode for simplicity since it
11961 * requires corresponding support in Spectre masking for stack ALU. See also
11962 * retrieve_ptr_limit().
11965 * 'off' includes 'reg->off'.
11967 static int check_stack_access_for_ptr_arithmetic(
11968 struct bpf_verifier_env *env,
11970 const struct bpf_reg_state *reg,
11973 if (!tnum_is_const(reg->var_off)) {
11976 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
11977 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
11978 regno, tn_buf, off);
11982 if (off >= 0 || off < -MAX_BPF_STACK) {
11983 verbose(env, "R%d stack pointer arithmetic goes out of range, "
11984 "prohibited for !root; off=%d\n", regno, off);
11991 static int sanitize_check_bounds(struct bpf_verifier_env *env,
11992 const struct bpf_insn *insn,
11993 const struct bpf_reg_state *dst_reg)
11995 u32 dst = insn->dst_reg;
11997 /* For unprivileged we require that resulting offset must be in bounds
11998 * in order to be able to sanitize access later on.
12000 if (env->bypass_spec_v1)
12003 switch (dst_reg->type) {
12005 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
12006 dst_reg->off + dst_reg->var_off.value))
12009 case PTR_TO_MAP_VALUE:
12010 if (check_map_access(env, dst, dst_reg->off, 1, false, ACCESS_HELPER)) {
12011 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
12012 "prohibited for !root\n", dst);
12023 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
12024 * Caller should also handle BPF_MOV case separately.
12025 * If we return -EACCES, caller may want to try again treating pointer as a
12026 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
12028 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
12029 struct bpf_insn *insn,
12030 const struct bpf_reg_state *ptr_reg,
12031 const struct bpf_reg_state *off_reg)
12033 struct bpf_verifier_state *vstate = env->cur_state;
12034 struct bpf_func_state *state = vstate->frame[vstate->curframe];
12035 struct bpf_reg_state *regs = state->regs, *dst_reg;
12036 bool known = tnum_is_const(off_reg->var_off);
12037 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
12038 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
12039 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
12040 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
12041 struct bpf_sanitize_info info = {};
12042 u8 opcode = BPF_OP(insn->code);
12043 u32 dst = insn->dst_reg;
12046 dst_reg = ®s[dst];
12048 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
12049 smin_val > smax_val || umin_val > umax_val) {
12050 /* Taint dst register if offset had invalid bounds derived from
12051 * e.g. dead branches.
12053 __mark_reg_unknown(env, dst_reg);
12057 if (BPF_CLASS(insn->code) != BPF_ALU64) {
12058 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
12059 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
12060 __mark_reg_unknown(env, dst_reg);
12065 "R%d 32-bit pointer arithmetic prohibited\n",
12070 if (ptr_reg->type & PTR_MAYBE_NULL) {
12071 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
12072 dst, reg_type_str(env, ptr_reg->type));
12076 switch (base_type(ptr_reg->type)) {
12077 case CONST_PTR_TO_MAP:
12078 /* smin_val represents the known value */
12079 if (known && smin_val == 0 && opcode == BPF_ADD)
12082 case PTR_TO_PACKET_END:
12083 case PTR_TO_SOCKET:
12084 case PTR_TO_SOCK_COMMON:
12085 case PTR_TO_TCP_SOCK:
12086 case PTR_TO_XDP_SOCK:
12087 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
12088 dst, reg_type_str(env, ptr_reg->type));
12094 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
12095 * The id may be overwritten later if we create a new variable offset.
12097 dst_reg->type = ptr_reg->type;
12098 dst_reg->id = ptr_reg->id;
12100 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
12101 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
12104 /* pointer types do not carry 32-bit bounds at the moment. */
12105 __mark_reg32_unbounded(dst_reg);
12107 if (sanitize_needed(opcode)) {
12108 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
12111 return sanitize_err(env, insn, ret, off_reg, dst_reg);
12116 /* We can take a fixed offset as long as it doesn't overflow
12117 * the s32 'off' field
12119 if (known && (ptr_reg->off + smin_val ==
12120 (s64)(s32)(ptr_reg->off + smin_val))) {
12121 /* pointer += K. Accumulate it into fixed offset */
12122 dst_reg->smin_value = smin_ptr;
12123 dst_reg->smax_value = smax_ptr;
12124 dst_reg->umin_value = umin_ptr;
12125 dst_reg->umax_value = umax_ptr;
12126 dst_reg->var_off = ptr_reg->var_off;
12127 dst_reg->off = ptr_reg->off + smin_val;
12128 dst_reg->raw = ptr_reg->raw;
12131 /* A new variable offset is created. Note that off_reg->off
12132 * == 0, since it's a scalar.
12133 * dst_reg gets the pointer type and since some positive
12134 * integer value was added to the pointer, give it a new 'id'
12135 * if it's a PTR_TO_PACKET.
12136 * this creates a new 'base' pointer, off_reg (variable) gets
12137 * added into the variable offset, and we copy the fixed offset
12140 if (signed_add_overflows(smin_ptr, smin_val) ||
12141 signed_add_overflows(smax_ptr, smax_val)) {
12142 dst_reg->smin_value = S64_MIN;
12143 dst_reg->smax_value = S64_MAX;
12145 dst_reg->smin_value = smin_ptr + smin_val;
12146 dst_reg->smax_value = smax_ptr + smax_val;
12148 if (umin_ptr + umin_val < umin_ptr ||
12149 umax_ptr + umax_val < umax_ptr) {
12150 dst_reg->umin_value = 0;
12151 dst_reg->umax_value = U64_MAX;
12153 dst_reg->umin_value = umin_ptr + umin_val;
12154 dst_reg->umax_value = umax_ptr + umax_val;
12156 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
12157 dst_reg->off = ptr_reg->off;
12158 dst_reg->raw = ptr_reg->raw;
12159 if (reg_is_pkt_pointer(ptr_reg)) {
12160 dst_reg->id = ++env->id_gen;
12161 /* something was added to pkt_ptr, set range to zero */
12162 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
12166 if (dst_reg == off_reg) {
12167 /* scalar -= pointer. Creates an unknown scalar */
12168 verbose(env, "R%d tried to subtract pointer from scalar\n",
12172 /* We don't allow subtraction from FP, because (according to
12173 * test_verifier.c test "invalid fp arithmetic", JITs might not
12174 * be able to deal with it.
12176 if (ptr_reg->type == PTR_TO_STACK) {
12177 verbose(env, "R%d subtraction from stack pointer prohibited\n",
12181 if (known && (ptr_reg->off - smin_val ==
12182 (s64)(s32)(ptr_reg->off - smin_val))) {
12183 /* pointer -= K. Subtract it from fixed offset */
12184 dst_reg->smin_value = smin_ptr;
12185 dst_reg->smax_value = smax_ptr;
12186 dst_reg->umin_value = umin_ptr;
12187 dst_reg->umax_value = umax_ptr;
12188 dst_reg->var_off = ptr_reg->var_off;
12189 dst_reg->id = ptr_reg->id;
12190 dst_reg->off = ptr_reg->off - smin_val;
12191 dst_reg->raw = ptr_reg->raw;
12194 /* A new variable offset is created. If the subtrahend is known
12195 * nonnegative, then any reg->range we had before is still good.
12197 if (signed_sub_overflows(smin_ptr, smax_val) ||
12198 signed_sub_overflows(smax_ptr, smin_val)) {
12199 /* Overflow possible, we know nothing */
12200 dst_reg->smin_value = S64_MIN;
12201 dst_reg->smax_value = S64_MAX;
12203 dst_reg->smin_value = smin_ptr - smax_val;
12204 dst_reg->smax_value = smax_ptr - smin_val;
12206 if (umin_ptr < umax_val) {
12207 /* Overflow possible, we know nothing */
12208 dst_reg->umin_value = 0;
12209 dst_reg->umax_value = U64_MAX;
12211 /* Cannot overflow (as long as bounds are consistent) */
12212 dst_reg->umin_value = umin_ptr - umax_val;
12213 dst_reg->umax_value = umax_ptr - umin_val;
12215 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
12216 dst_reg->off = ptr_reg->off;
12217 dst_reg->raw = ptr_reg->raw;
12218 if (reg_is_pkt_pointer(ptr_reg)) {
12219 dst_reg->id = ++env->id_gen;
12220 /* something was added to pkt_ptr, set range to zero */
12222 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
12228 /* bitwise ops on pointers are troublesome, prohibit. */
12229 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
12230 dst, bpf_alu_string[opcode >> 4]);
12233 /* other operators (e.g. MUL,LSH) produce non-pointer results */
12234 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
12235 dst, bpf_alu_string[opcode >> 4]);
12239 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
12241 reg_bounds_sync(dst_reg);
12242 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
12244 if (sanitize_needed(opcode)) {
12245 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
12248 return sanitize_err(env, insn, ret, off_reg, dst_reg);
12254 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
12255 struct bpf_reg_state *src_reg)
12257 s32 smin_val = src_reg->s32_min_value;
12258 s32 smax_val = src_reg->s32_max_value;
12259 u32 umin_val = src_reg->u32_min_value;
12260 u32 umax_val = src_reg->u32_max_value;
12262 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
12263 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
12264 dst_reg->s32_min_value = S32_MIN;
12265 dst_reg->s32_max_value = S32_MAX;
12267 dst_reg->s32_min_value += smin_val;
12268 dst_reg->s32_max_value += smax_val;
12270 if (dst_reg->u32_min_value + umin_val < umin_val ||
12271 dst_reg->u32_max_value + umax_val < umax_val) {
12272 dst_reg->u32_min_value = 0;
12273 dst_reg->u32_max_value = U32_MAX;
12275 dst_reg->u32_min_value += umin_val;
12276 dst_reg->u32_max_value += umax_val;
12280 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
12281 struct bpf_reg_state *src_reg)
12283 s64 smin_val = src_reg->smin_value;
12284 s64 smax_val = src_reg->smax_value;
12285 u64 umin_val = src_reg->umin_value;
12286 u64 umax_val = src_reg->umax_value;
12288 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
12289 signed_add_overflows(dst_reg->smax_value, smax_val)) {
12290 dst_reg->smin_value = S64_MIN;
12291 dst_reg->smax_value = S64_MAX;
12293 dst_reg->smin_value += smin_val;
12294 dst_reg->smax_value += smax_val;
12296 if (dst_reg->umin_value + umin_val < umin_val ||
12297 dst_reg->umax_value + umax_val < umax_val) {
12298 dst_reg->umin_value = 0;
12299 dst_reg->umax_value = U64_MAX;
12301 dst_reg->umin_value += umin_val;
12302 dst_reg->umax_value += umax_val;
12306 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
12307 struct bpf_reg_state *src_reg)
12309 s32 smin_val = src_reg->s32_min_value;
12310 s32 smax_val = src_reg->s32_max_value;
12311 u32 umin_val = src_reg->u32_min_value;
12312 u32 umax_val = src_reg->u32_max_value;
12314 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
12315 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
12316 /* Overflow possible, we know nothing */
12317 dst_reg->s32_min_value = S32_MIN;
12318 dst_reg->s32_max_value = S32_MAX;
12320 dst_reg->s32_min_value -= smax_val;
12321 dst_reg->s32_max_value -= smin_val;
12323 if (dst_reg->u32_min_value < umax_val) {
12324 /* Overflow possible, we know nothing */
12325 dst_reg->u32_min_value = 0;
12326 dst_reg->u32_max_value = U32_MAX;
12328 /* Cannot overflow (as long as bounds are consistent) */
12329 dst_reg->u32_min_value -= umax_val;
12330 dst_reg->u32_max_value -= umin_val;
12334 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
12335 struct bpf_reg_state *src_reg)
12337 s64 smin_val = src_reg->smin_value;
12338 s64 smax_val = src_reg->smax_value;
12339 u64 umin_val = src_reg->umin_value;
12340 u64 umax_val = src_reg->umax_value;
12342 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
12343 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
12344 /* Overflow possible, we know nothing */
12345 dst_reg->smin_value = S64_MIN;
12346 dst_reg->smax_value = S64_MAX;
12348 dst_reg->smin_value -= smax_val;
12349 dst_reg->smax_value -= smin_val;
12351 if (dst_reg->umin_value < umax_val) {
12352 /* Overflow possible, we know nothing */
12353 dst_reg->umin_value = 0;
12354 dst_reg->umax_value = U64_MAX;
12356 /* Cannot overflow (as long as bounds are consistent) */
12357 dst_reg->umin_value -= umax_val;
12358 dst_reg->umax_value -= umin_val;
12362 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
12363 struct bpf_reg_state *src_reg)
12365 s32 smin_val = src_reg->s32_min_value;
12366 u32 umin_val = src_reg->u32_min_value;
12367 u32 umax_val = src_reg->u32_max_value;
12369 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
12370 /* Ain't nobody got time to multiply that sign */
12371 __mark_reg32_unbounded(dst_reg);
12374 /* Both values are positive, so we can work with unsigned and
12375 * copy the result to signed (unless it exceeds S32_MAX).
12377 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
12378 /* Potential overflow, we know nothing */
12379 __mark_reg32_unbounded(dst_reg);
12382 dst_reg->u32_min_value *= umin_val;
12383 dst_reg->u32_max_value *= umax_val;
12384 if (dst_reg->u32_max_value > S32_MAX) {
12385 /* Overflow possible, we know nothing */
12386 dst_reg->s32_min_value = S32_MIN;
12387 dst_reg->s32_max_value = S32_MAX;
12389 dst_reg->s32_min_value = dst_reg->u32_min_value;
12390 dst_reg->s32_max_value = dst_reg->u32_max_value;
12394 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
12395 struct bpf_reg_state *src_reg)
12397 s64 smin_val = src_reg->smin_value;
12398 u64 umin_val = src_reg->umin_value;
12399 u64 umax_val = src_reg->umax_value;
12401 if (smin_val < 0 || dst_reg->smin_value < 0) {
12402 /* Ain't nobody got time to multiply that sign */
12403 __mark_reg64_unbounded(dst_reg);
12406 /* Both values are positive, so we can work with unsigned and
12407 * copy the result to signed (unless it exceeds S64_MAX).
12409 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
12410 /* Potential overflow, we know nothing */
12411 __mark_reg64_unbounded(dst_reg);
12414 dst_reg->umin_value *= umin_val;
12415 dst_reg->umax_value *= umax_val;
12416 if (dst_reg->umax_value > S64_MAX) {
12417 /* Overflow possible, we know nothing */
12418 dst_reg->smin_value = S64_MIN;
12419 dst_reg->smax_value = S64_MAX;
12421 dst_reg->smin_value = dst_reg->umin_value;
12422 dst_reg->smax_value = dst_reg->umax_value;
12426 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
12427 struct bpf_reg_state *src_reg)
12429 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12430 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12431 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12432 s32 smin_val = src_reg->s32_min_value;
12433 u32 umax_val = src_reg->u32_max_value;
12435 if (src_known && dst_known) {
12436 __mark_reg32_known(dst_reg, var32_off.value);
12440 /* We get our minimum from the var_off, since that's inherently
12441 * bitwise. Our maximum is the minimum of the operands' maxima.
12443 dst_reg->u32_min_value = var32_off.value;
12444 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
12445 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
12446 /* Lose signed bounds when ANDing negative numbers,
12447 * ain't nobody got time for that.
12449 dst_reg->s32_min_value = S32_MIN;
12450 dst_reg->s32_max_value = S32_MAX;
12452 /* ANDing two positives gives a positive, so safe to
12453 * cast result into s64.
12455 dst_reg->s32_min_value = dst_reg->u32_min_value;
12456 dst_reg->s32_max_value = dst_reg->u32_max_value;
12460 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
12461 struct bpf_reg_state *src_reg)
12463 bool src_known = tnum_is_const(src_reg->var_off);
12464 bool dst_known = tnum_is_const(dst_reg->var_off);
12465 s64 smin_val = src_reg->smin_value;
12466 u64 umax_val = src_reg->umax_value;
12468 if (src_known && dst_known) {
12469 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12473 /* We get our minimum from the var_off, since that's inherently
12474 * bitwise. Our maximum is the minimum of the operands' maxima.
12476 dst_reg->umin_value = dst_reg->var_off.value;
12477 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
12478 if (dst_reg->smin_value < 0 || smin_val < 0) {
12479 /* Lose signed bounds when ANDing negative numbers,
12480 * ain't nobody got time for that.
12482 dst_reg->smin_value = S64_MIN;
12483 dst_reg->smax_value = S64_MAX;
12485 /* ANDing two positives gives a positive, so safe to
12486 * cast result into s64.
12488 dst_reg->smin_value = dst_reg->umin_value;
12489 dst_reg->smax_value = dst_reg->umax_value;
12491 /* We may learn something more from the var_off */
12492 __update_reg_bounds(dst_reg);
12495 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
12496 struct bpf_reg_state *src_reg)
12498 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12499 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12500 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12501 s32 smin_val = src_reg->s32_min_value;
12502 u32 umin_val = src_reg->u32_min_value;
12504 if (src_known && dst_known) {
12505 __mark_reg32_known(dst_reg, var32_off.value);
12509 /* We get our maximum from the var_off, and our minimum is the
12510 * maximum of the operands' minima
12512 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
12513 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
12514 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
12515 /* Lose signed bounds when ORing negative numbers,
12516 * ain't nobody got time for that.
12518 dst_reg->s32_min_value = S32_MIN;
12519 dst_reg->s32_max_value = S32_MAX;
12521 /* ORing two positives gives a positive, so safe to
12522 * cast result into s64.
12524 dst_reg->s32_min_value = dst_reg->u32_min_value;
12525 dst_reg->s32_max_value = dst_reg->u32_max_value;
12529 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
12530 struct bpf_reg_state *src_reg)
12532 bool src_known = tnum_is_const(src_reg->var_off);
12533 bool dst_known = tnum_is_const(dst_reg->var_off);
12534 s64 smin_val = src_reg->smin_value;
12535 u64 umin_val = src_reg->umin_value;
12537 if (src_known && dst_known) {
12538 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12542 /* We get our maximum from the var_off, and our minimum is the
12543 * maximum of the operands' minima
12545 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
12546 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
12547 if (dst_reg->smin_value < 0 || smin_val < 0) {
12548 /* Lose signed bounds when ORing negative numbers,
12549 * ain't nobody got time for that.
12551 dst_reg->smin_value = S64_MIN;
12552 dst_reg->smax_value = S64_MAX;
12554 /* ORing two positives gives a positive, so safe to
12555 * cast result into s64.
12557 dst_reg->smin_value = dst_reg->umin_value;
12558 dst_reg->smax_value = dst_reg->umax_value;
12560 /* We may learn something more from the var_off */
12561 __update_reg_bounds(dst_reg);
12564 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
12565 struct bpf_reg_state *src_reg)
12567 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12568 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12569 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12570 s32 smin_val = src_reg->s32_min_value;
12572 if (src_known && dst_known) {
12573 __mark_reg32_known(dst_reg, var32_off.value);
12577 /* We get both minimum and maximum from the var32_off. */
12578 dst_reg->u32_min_value = var32_off.value;
12579 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
12581 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
12582 /* XORing two positive sign numbers gives a positive,
12583 * so safe to cast u32 result into s32.
12585 dst_reg->s32_min_value = dst_reg->u32_min_value;
12586 dst_reg->s32_max_value = dst_reg->u32_max_value;
12588 dst_reg->s32_min_value = S32_MIN;
12589 dst_reg->s32_max_value = S32_MAX;
12593 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
12594 struct bpf_reg_state *src_reg)
12596 bool src_known = tnum_is_const(src_reg->var_off);
12597 bool dst_known = tnum_is_const(dst_reg->var_off);
12598 s64 smin_val = src_reg->smin_value;
12600 if (src_known && dst_known) {
12601 /* dst_reg->var_off.value has been updated earlier */
12602 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12606 /* We get both minimum and maximum from the var_off. */
12607 dst_reg->umin_value = dst_reg->var_off.value;
12608 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
12610 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
12611 /* XORing two positive sign numbers gives a positive,
12612 * so safe to cast u64 result into s64.
12614 dst_reg->smin_value = dst_reg->umin_value;
12615 dst_reg->smax_value = dst_reg->umax_value;
12617 dst_reg->smin_value = S64_MIN;
12618 dst_reg->smax_value = S64_MAX;
12621 __update_reg_bounds(dst_reg);
12624 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
12625 u64 umin_val, u64 umax_val)
12627 /* We lose all sign bit information (except what we can pick
12630 dst_reg->s32_min_value = S32_MIN;
12631 dst_reg->s32_max_value = S32_MAX;
12632 /* If we might shift our top bit out, then we know nothing */
12633 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
12634 dst_reg->u32_min_value = 0;
12635 dst_reg->u32_max_value = U32_MAX;
12637 dst_reg->u32_min_value <<= umin_val;
12638 dst_reg->u32_max_value <<= umax_val;
12642 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
12643 struct bpf_reg_state *src_reg)
12645 u32 umax_val = src_reg->u32_max_value;
12646 u32 umin_val = src_reg->u32_min_value;
12647 /* u32 alu operation will zext upper bits */
12648 struct tnum subreg = tnum_subreg(dst_reg->var_off);
12650 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
12651 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
12652 /* Not required but being careful mark reg64 bounds as unknown so
12653 * that we are forced to pick them up from tnum and zext later and
12654 * if some path skips this step we are still safe.
12656 __mark_reg64_unbounded(dst_reg);
12657 __update_reg32_bounds(dst_reg);
12660 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
12661 u64 umin_val, u64 umax_val)
12663 /* Special case <<32 because it is a common compiler pattern to sign
12664 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
12665 * positive we know this shift will also be positive so we can track
12666 * bounds correctly. Otherwise we lose all sign bit information except
12667 * what we can pick up from var_off. Perhaps we can generalize this
12668 * later to shifts of any length.
12670 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
12671 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
12673 dst_reg->smax_value = S64_MAX;
12675 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
12676 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
12678 dst_reg->smin_value = S64_MIN;
12680 /* If we might shift our top bit out, then we know nothing */
12681 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
12682 dst_reg->umin_value = 0;
12683 dst_reg->umax_value = U64_MAX;
12685 dst_reg->umin_value <<= umin_val;
12686 dst_reg->umax_value <<= umax_val;
12690 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
12691 struct bpf_reg_state *src_reg)
12693 u64 umax_val = src_reg->umax_value;
12694 u64 umin_val = src_reg->umin_value;
12696 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
12697 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
12698 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
12700 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
12701 /* We may learn something more from the var_off */
12702 __update_reg_bounds(dst_reg);
12705 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
12706 struct bpf_reg_state *src_reg)
12708 struct tnum subreg = tnum_subreg(dst_reg->var_off);
12709 u32 umax_val = src_reg->u32_max_value;
12710 u32 umin_val = src_reg->u32_min_value;
12712 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
12713 * be negative, then either:
12714 * 1) src_reg might be zero, so the sign bit of the result is
12715 * unknown, so we lose our signed bounds
12716 * 2) it's known negative, thus the unsigned bounds capture the
12718 * 3) the signed bounds cross zero, so they tell us nothing
12720 * If the value in dst_reg is known nonnegative, then again the
12721 * unsigned bounds capture the signed bounds.
12722 * Thus, in all cases it suffices to blow away our signed bounds
12723 * and rely on inferring new ones from the unsigned bounds and
12724 * var_off of the result.
12726 dst_reg->s32_min_value = S32_MIN;
12727 dst_reg->s32_max_value = S32_MAX;
12729 dst_reg->var_off = tnum_rshift(subreg, umin_val);
12730 dst_reg->u32_min_value >>= umax_val;
12731 dst_reg->u32_max_value >>= umin_val;
12733 __mark_reg64_unbounded(dst_reg);
12734 __update_reg32_bounds(dst_reg);
12737 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
12738 struct bpf_reg_state *src_reg)
12740 u64 umax_val = src_reg->umax_value;
12741 u64 umin_val = src_reg->umin_value;
12743 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
12744 * be negative, then either:
12745 * 1) src_reg might be zero, so the sign bit of the result is
12746 * unknown, so we lose our signed bounds
12747 * 2) it's known negative, thus the unsigned bounds capture the
12749 * 3) the signed bounds cross zero, so they tell us nothing
12751 * If the value in dst_reg is known nonnegative, then again the
12752 * unsigned bounds capture the signed bounds.
12753 * Thus, in all cases it suffices to blow away our signed bounds
12754 * and rely on inferring new ones from the unsigned bounds and
12755 * var_off of the result.
12757 dst_reg->smin_value = S64_MIN;
12758 dst_reg->smax_value = S64_MAX;
12759 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
12760 dst_reg->umin_value >>= umax_val;
12761 dst_reg->umax_value >>= umin_val;
12763 /* Its not easy to operate on alu32 bounds here because it depends
12764 * on bits being shifted in. Take easy way out and mark unbounded
12765 * so we can recalculate later from tnum.
12767 __mark_reg32_unbounded(dst_reg);
12768 __update_reg_bounds(dst_reg);
12771 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
12772 struct bpf_reg_state *src_reg)
12774 u64 umin_val = src_reg->u32_min_value;
12776 /* Upon reaching here, src_known is true and
12777 * umax_val is equal to umin_val.
12779 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
12780 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
12782 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
12784 /* blow away the dst_reg umin_value/umax_value and rely on
12785 * dst_reg var_off to refine the result.
12787 dst_reg->u32_min_value = 0;
12788 dst_reg->u32_max_value = U32_MAX;
12790 __mark_reg64_unbounded(dst_reg);
12791 __update_reg32_bounds(dst_reg);
12794 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
12795 struct bpf_reg_state *src_reg)
12797 u64 umin_val = src_reg->umin_value;
12799 /* Upon reaching here, src_known is true and umax_val is equal
12802 dst_reg->smin_value >>= umin_val;
12803 dst_reg->smax_value >>= umin_val;
12805 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
12807 /* blow away the dst_reg umin_value/umax_value and rely on
12808 * dst_reg var_off to refine the result.
12810 dst_reg->umin_value = 0;
12811 dst_reg->umax_value = U64_MAX;
12813 /* Its not easy to operate on alu32 bounds here because it depends
12814 * on bits being shifted in from upper 32-bits. Take easy way out
12815 * and mark unbounded so we can recalculate later from tnum.
12817 __mark_reg32_unbounded(dst_reg);
12818 __update_reg_bounds(dst_reg);
12821 /* WARNING: This function does calculations on 64-bit values, but the actual
12822 * execution may occur on 32-bit values. Therefore, things like bitshifts
12823 * need extra checks in the 32-bit case.
12825 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
12826 struct bpf_insn *insn,
12827 struct bpf_reg_state *dst_reg,
12828 struct bpf_reg_state src_reg)
12830 struct bpf_reg_state *regs = cur_regs(env);
12831 u8 opcode = BPF_OP(insn->code);
12833 s64 smin_val, smax_val;
12834 u64 umin_val, umax_val;
12835 s32 s32_min_val, s32_max_val;
12836 u32 u32_min_val, u32_max_val;
12837 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
12838 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
12841 smin_val = src_reg.smin_value;
12842 smax_val = src_reg.smax_value;
12843 umin_val = src_reg.umin_value;
12844 umax_val = src_reg.umax_value;
12846 s32_min_val = src_reg.s32_min_value;
12847 s32_max_val = src_reg.s32_max_value;
12848 u32_min_val = src_reg.u32_min_value;
12849 u32_max_val = src_reg.u32_max_value;
12852 src_known = tnum_subreg_is_const(src_reg.var_off);
12854 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
12855 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
12856 /* Taint dst register if offset had invalid bounds
12857 * derived from e.g. dead branches.
12859 __mark_reg_unknown(env, dst_reg);
12863 src_known = tnum_is_const(src_reg.var_off);
12865 (smin_val != smax_val || umin_val != umax_val)) ||
12866 smin_val > smax_val || umin_val > umax_val) {
12867 /* Taint dst register if offset had invalid bounds
12868 * derived from e.g. dead branches.
12870 __mark_reg_unknown(env, dst_reg);
12876 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
12877 __mark_reg_unknown(env, dst_reg);
12881 if (sanitize_needed(opcode)) {
12882 ret = sanitize_val_alu(env, insn);
12884 return sanitize_err(env, insn, ret, NULL, NULL);
12887 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
12888 * There are two classes of instructions: The first class we track both
12889 * alu32 and alu64 sign/unsigned bounds independently this provides the
12890 * greatest amount of precision when alu operations are mixed with jmp32
12891 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
12892 * and BPF_OR. This is possible because these ops have fairly easy to
12893 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
12894 * See alu32 verifier tests for examples. The second class of
12895 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
12896 * with regards to tracking sign/unsigned bounds because the bits may
12897 * cross subreg boundaries in the alu64 case. When this happens we mark
12898 * the reg unbounded in the subreg bound space and use the resulting
12899 * tnum to calculate an approximation of the sign/unsigned bounds.
12903 scalar32_min_max_add(dst_reg, &src_reg);
12904 scalar_min_max_add(dst_reg, &src_reg);
12905 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
12908 scalar32_min_max_sub(dst_reg, &src_reg);
12909 scalar_min_max_sub(dst_reg, &src_reg);
12910 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
12913 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
12914 scalar32_min_max_mul(dst_reg, &src_reg);
12915 scalar_min_max_mul(dst_reg, &src_reg);
12918 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
12919 scalar32_min_max_and(dst_reg, &src_reg);
12920 scalar_min_max_and(dst_reg, &src_reg);
12923 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
12924 scalar32_min_max_or(dst_reg, &src_reg);
12925 scalar_min_max_or(dst_reg, &src_reg);
12928 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
12929 scalar32_min_max_xor(dst_reg, &src_reg);
12930 scalar_min_max_xor(dst_reg, &src_reg);
12933 if (umax_val >= insn_bitness) {
12934 /* Shifts greater than 31 or 63 are undefined.
12935 * This includes shifts by a negative number.
12937 mark_reg_unknown(env, regs, insn->dst_reg);
12941 scalar32_min_max_lsh(dst_reg, &src_reg);
12943 scalar_min_max_lsh(dst_reg, &src_reg);
12946 if (umax_val >= insn_bitness) {
12947 /* Shifts greater than 31 or 63 are undefined.
12948 * This includes shifts by a negative number.
12950 mark_reg_unknown(env, regs, insn->dst_reg);
12954 scalar32_min_max_rsh(dst_reg, &src_reg);
12956 scalar_min_max_rsh(dst_reg, &src_reg);
12959 if (umax_val >= insn_bitness) {
12960 /* Shifts greater than 31 or 63 are undefined.
12961 * This includes shifts by a negative number.
12963 mark_reg_unknown(env, regs, insn->dst_reg);
12967 scalar32_min_max_arsh(dst_reg, &src_reg);
12969 scalar_min_max_arsh(dst_reg, &src_reg);
12972 mark_reg_unknown(env, regs, insn->dst_reg);
12976 /* ALU32 ops are zero extended into 64bit register */
12978 zext_32_to_64(dst_reg);
12979 reg_bounds_sync(dst_reg);
12983 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
12986 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
12987 struct bpf_insn *insn)
12989 struct bpf_verifier_state *vstate = env->cur_state;
12990 struct bpf_func_state *state = vstate->frame[vstate->curframe];
12991 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
12992 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
12993 u8 opcode = BPF_OP(insn->code);
12996 dst_reg = ®s[insn->dst_reg];
12998 if (dst_reg->type != SCALAR_VALUE)
13001 /* Make sure ID is cleared otherwise dst_reg min/max could be
13002 * incorrectly propagated into other registers by find_equal_scalars()
13005 if (BPF_SRC(insn->code) == BPF_X) {
13006 src_reg = ®s[insn->src_reg];
13007 if (src_reg->type != SCALAR_VALUE) {
13008 if (dst_reg->type != SCALAR_VALUE) {
13009 /* Combining two pointers by any ALU op yields
13010 * an arbitrary scalar. Disallow all math except
13011 * pointer subtraction
13013 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
13014 mark_reg_unknown(env, regs, insn->dst_reg);
13017 verbose(env, "R%d pointer %s pointer prohibited\n",
13019 bpf_alu_string[opcode >> 4]);
13022 /* scalar += pointer
13023 * This is legal, but we have to reverse our
13024 * src/dest handling in computing the range
13026 err = mark_chain_precision(env, insn->dst_reg);
13029 return adjust_ptr_min_max_vals(env, insn,
13032 } else if (ptr_reg) {
13033 /* pointer += scalar */
13034 err = mark_chain_precision(env, insn->src_reg);
13037 return adjust_ptr_min_max_vals(env, insn,
13039 } else if (dst_reg->precise) {
13040 /* if dst_reg is precise, src_reg should be precise as well */
13041 err = mark_chain_precision(env, insn->src_reg);
13046 /* Pretend the src is a reg with a known value, since we only
13047 * need to be able to read from this state.
13049 off_reg.type = SCALAR_VALUE;
13050 __mark_reg_known(&off_reg, insn->imm);
13051 src_reg = &off_reg;
13052 if (ptr_reg) /* pointer += K */
13053 return adjust_ptr_min_max_vals(env, insn,
13057 /* Got here implies adding two SCALAR_VALUEs */
13058 if (WARN_ON_ONCE(ptr_reg)) {
13059 print_verifier_state(env, state, true);
13060 verbose(env, "verifier internal error: unexpected ptr_reg\n");
13063 if (WARN_ON(!src_reg)) {
13064 print_verifier_state(env, state, true);
13065 verbose(env, "verifier internal error: no src_reg\n");
13068 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
13071 /* check validity of 32-bit and 64-bit arithmetic operations */
13072 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
13074 struct bpf_reg_state *regs = cur_regs(env);
13075 u8 opcode = BPF_OP(insn->code);
13078 if (opcode == BPF_END || opcode == BPF_NEG) {
13079 if (opcode == BPF_NEG) {
13080 if (BPF_SRC(insn->code) != BPF_K ||
13081 insn->src_reg != BPF_REG_0 ||
13082 insn->off != 0 || insn->imm != 0) {
13083 verbose(env, "BPF_NEG uses reserved fields\n");
13087 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
13088 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
13089 (BPF_CLASS(insn->code) == BPF_ALU64 &&
13090 BPF_SRC(insn->code) != BPF_TO_LE)) {
13091 verbose(env, "BPF_END uses reserved fields\n");
13096 /* check src operand */
13097 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
13101 if (is_pointer_value(env, insn->dst_reg)) {
13102 verbose(env, "R%d pointer arithmetic prohibited\n",
13107 /* check dest operand */
13108 err = check_reg_arg(env, insn->dst_reg, DST_OP);
13112 } else if (opcode == BPF_MOV) {
13114 if (BPF_SRC(insn->code) == BPF_X) {
13115 if (insn->imm != 0) {
13116 verbose(env, "BPF_MOV uses reserved fields\n");
13120 if (BPF_CLASS(insn->code) == BPF_ALU) {
13121 if (insn->off != 0 && insn->off != 8 && insn->off != 16) {
13122 verbose(env, "BPF_MOV uses reserved fields\n");
13126 if (insn->off != 0 && insn->off != 8 && insn->off != 16 &&
13128 verbose(env, "BPF_MOV uses reserved fields\n");
13133 /* check src operand */
13134 err = check_reg_arg(env, insn->src_reg, SRC_OP);
13138 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
13139 verbose(env, "BPF_MOV uses reserved fields\n");
13144 /* check dest operand, mark as required later */
13145 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
13149 if (BPF_SRC(insn->code) == BPF_X) {
13150 struct bpf_reg_state *src_reg = regs + insn->src_reg;
13151 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
13152 bool need_id = src_reg->type == SCALAR_VALUE && !src_reg->id &&
13153 !tnum_is_const(src_reg->var_off);
13155 if (BPF_CLASS(insn->code) == BPF_ALU64) {
13156 if (insn->off == 0) {
13158 * copy register state to dest reg
13161 /* Assign src and dst registers the same ID
13162 * that will be used by find_equal_scalars()
13163 * to propagate min/max range.
13165 src_reg->id = ++env->id_gen;
13166 copy_register_state(dst_reg, src_reg);
13167 dst_reg->live |= REG_LIVE_WRITTEN;
13168 dst_reg->subreg_def = DEF_NOT_SUBREG;
13170 /* case: R1 = (s8, s16 s32)R2 */
13171 if (is_pointer_value(env, insn->src_reg)) {
13173 "R%d sign-extension part of pointer\n",
13176 } else if (src_reg->type == SCALAR_VALUE) {
13179 no_sext = src_reg->umax_value < (1ULL << (insn->off - 1));
13180 if (no_sext && need_id)
13181 src_reg->id = ++env->id_gen;
13182 copy_register_state(dst_reg, src_reg);
13185 coerce_reg_to_size_sx(dst_reg, insn->off >> 3);
13186 dst_reg->live |= REG_LIVE_WRITTEN;
13187 dst_reg->subreg_def = DEF_NOT_SUBREG;
13189 mark_reg_unknown(env, regs, insn->dst_reg);
13193 /* R1 = (u32) R2 */
13194 if (is_pointer_value(env, insn->src_reg)) {
13196 "R%d partial copy of pointer\n",
13199 } else if (src_reg->type == SCALAR_VALUE) {
13200 if (insn->off == 0) {
13201 bool is_src_reg_u32 = src_reg->umax_value <= U32_MAX;
13203 if (is_src_reg_u32 && need_id)
13204 src_reg->id = ++env->id_gen;
13205 copy_register_state(dst_reg, src_reg);
13206 /* Make sure ID is cleared if src_reg is not in u32
13207 * range otherwise dst_reg min/max could be incorrectly
13208 * propagated into src_reg by find_equal_scalars()
13210 if (!is_src_reg_u32)
13212 dst_reg->live |= REG_LIVE_WRITTEN;
13213 dst_reg->subreg_def = env->insn_idx + 1;
13215 /* case: W1 = (s8, s16)W2 */
13216 bool no_sext = src_reg->umax_value < (1ULL << (insn->off - 1));
13218 if (no_sext && need_id)
13219 src_reg->id = ++env->id_gen;
13220 copy_register_state(dst_reg, src_reg);
13223 dst_reg->live |= REG_LIVE_WRITTEN;
13224 dst_reg->subreg_def = env->insn_idx + 1;
13225 coerce_subreg_to_size_sx(dst_reg, insn->off >> 3);
13228 mark_reg_unknown(env, regs,
13231 zext_32_to_64(dst_reg);
13232 reg_bounds_sync(dst_reg);
13236 * remember the value we stored into this reg
13238 /* clear any state __mark_reg_known doesn't set */
13239 mark_reg_unknown(env, regs, insn->dst_reg);
13240 regs[insn->dst_reg].type = SCALAR_VALUE;
13241 if (BPF_CLASS(insn->code) == BPF_ALU64) {
13242 __mark_reg_known(regs + insn->dst_reg,
13245 __mark_reg_known(regs + insn->dst_reg,
13250 } else if (opcode > BPF_END) {
13251 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
13254 } else { /* all other ALU ops: and, sub, xor, add, ... */
13256 if (BPF_SRC(insn->code) == BPF_X) {
13257 if (insn->imm != 0 || insn->off > 1 ||
13258 (insn->off == 1 && opcode != BPF_MOD && opcode != BPF_DIV)) {
13259 verbose(env, "BPF_ALU uses reserved fields\n");
13262 /* check src1 operand */
13263 err = check_reg_arg(env, insn->src_reg, SRC_OP);
13267 if (insn->src_reg != BPF_REG_0 || insn->off > 1 ||
13268 (insn->off == 1 && opcode != BPF_MOD && opcode != BPF_DIV)) {
13269 verbose(env, "BPF_ALU uses reserved fields\n");
13274 /* check src2 operand */
13275 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
13279 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
13280 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
13281 verbose(env, "div by zero\n");
13285 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
13286 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
13287 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
13289 if (insn->imm < 0 || insn->imm >= size) {
13290 verbose(env, "invalid shift %d\n", insn->imm);
13295 /* check dest operand */
13296 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
13300 return adjust_reg_min_max_vals(env, insn);
13306 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
13307 struct bpf_reg_state *dst_reg,
13308 enum bpf_reg_type type,
13309 bool range_right_open)
13311 struct bpf_func_state *state;
13312 struct bpf_reg_state *reg;
13315 if (dst_reg->off < 0 ||
13316 (dst_reg->off == 0 && range_right_open))
13317 /* This doesn't give us any range */
13320 if (dst_reg->umax_value > MAX_PACKET_OFF ||
13321 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
13322 /* Risk of overflow. For instance, ptr + (1<<63) may be less
13323 * than pkt_end, but that's because it's also less than pkt.
13327 new_range = dst_reg->off;
13328 if (range_right_open)
13331 /* Examples for register markings:
13333 * pkt_data in dst register:
13337 * if (r2 > pkt_end) goto <handle exception>
13342 * if (r2 < pkt_end) goto <access okay>
13343 * <handle exception>
13346 * r2 == dst_reg, pkt_end == src_reg
13347 * r2=pkt(id=n,off=8,r=0)
13348 * r3=pkt(id=n,off=0,r=0)
13350 * pkt_data in src register:
13354 * if (pkt_end >= r2) goto <access okay>
13355 * <handle exception>
13359 * if (pkt_end <= r2) goto <handle exception>
13363 * pkt_end == dst_reg, r2 == src_reg
13364 * r2=pkt(id=n,off=8,r=0)
13365 * r3=pkt(id=n,off=0,r=0)
13367 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
13368 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
13369 * and [r3, r3 + 8-1) respectively is safe to access depending on
13373 /* If our ids match, then we must have the same max_value. And we
13374 * don't care about the other reg's fixed offset, since if it's too big
13375 * the range won't allow anything.
13376 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
13378 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
13379 if (reg->type == type && reg->id == dst_reg->id)
13380 /* keep the maximum range already checked */
13381 reg->range = max(reg->range, new_range);
13385 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
13387 struct tnum subreg = tnum_subreg(reg->var_off);
13388 s32 sval = (s32)val;
13392 if (tnum_is_const(subreg))
13393 return !!tnum_equals_const(subreg, val);
13394 else if (val < reg->u32_min_value || val > reg->u32_max_value)
13398 if (tnum_is_const(subreg))
13399 return !tnum_equals_const(subreg, val);
13400 else if (val < reg->u32_min_value || val > reg->u32_max_value)
13404 if ((~subreg.mask & subreg.value) & val)
13406 if (!((subreg.mask | subreg.value) & val))
13410 if (reg->u32_min_value > val)
13412 else if (reg->u32_max_value <= val)
13416 if (reg->s32_min_value > sval)
13418 else if (reg->s32_max_value <= sval)
13422 if (reg->u32_max_value < val)
13424 else if (reg->u32_min_value >= val)
13428 if (reg->s32_max_value < sval)
13430 else if (reg->s32_min_value >= sval)
13434 if (reg->u32_min_value >= val)
13436 else if (reg->u32_max_value < val)
13440 if (reg->s32_min_value >= sval)
13442 else if (reg->s32_max_value < sval)
13446 if (reg->u32_max_value <= val)
13448 else if (reg->u32_min_value > val)
13452 if (reg->s32_max_value <= sval)
13454 else if (reg->s32_min_value > sval)
13463 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
13465 s64 sval = (s64)val;
13469 if (tnum_is_const(reg->var_off))
13470 return !!tnum_equals_const(reg->var_off, val);
13471 else if (val < reg->umin_value || val > reg->umax_value)
13475 if (tnum_is_const(reg->var_off))
13476 return !tnum_equals_const(reg->var_off, val);
13477 else if (val < reg->umin_value || val > reg->umax_value)
13481 if ((~reg->var_off.mask & reg->var_off.value) & val)
13483 if (!((reg->var_off.mask | reg->var_off.value) & val))
13487 if (reg->umin_value > val)
13489 else if (reg->umax_value <= val)
13493 if (reg->smin_value > sval)
13495 else if (reg->smax_value <= sval)
13499 if (reg->umax_value < val)
13501 else if (reg->umin_value >= val)
13505 if (reg->smax_value < sval)
13507 else if (reg->smin_value >= sval)
13511 if (reg->umin_value >= val)
13513 else if (reg->umax_value < val)
13517 if (reg->smin_value >= sval)
13519 else if (reg->smax_value < sval)
13523 if (reg->umax_value <= val)
13525 else if (reg->umin_value > val)
13529 if (reg->smax_value <= sval)
13531 else if (reg->smin_value > sval)
13539 /* compute branch direction of the expression "if (reg opcode val) goto target;"
13541 * 1 - branch will be taken and "goto target" will be executed
13542 * 0 - branch will not be taken and fall-through to next insn
13543 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
13546 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
13549 if (__is_pointer_value(false, reg)) {
13550 if (!reg_not_null(reg))
13553 /* If pointer is valid tests against zero will fail so we can
13554 * use this to direct branch taken.
13570 return is_branch32_taken(reg, val, opcode);
13571 return is_branch64_taken(reg, val, opcode);
13574 static int flip_opcode(u32 opcode)
13576 /* How can we transform "a <op> b" into "b <op> a"? */
13577 static const u8 opcode_flip[16] = {
13578 /* these stay the same */
13579 [BPF_JEQ >> 4] = BPF_JEQ,
13580 [BPF_JNE >> 4] = BPF_JNE,
13581 [BPF_JSET >> 4] = BPF_JSET,
13582 /* these swap "lesser" and "greater" (L and G in the opcodes) */
13583 [BPF_JGE >> 4] = BPF_JLE,
13584 [BPF_JGT >> 4] = BPF_JLT,
13585 [BPF_JLE >> 4] = BPF_JGE,
13586 [BPF_JLT >> 4] = BPF_JGT,
13587 [BPF_JSGE >> 4] = BPF_JSLE,
13588 [BPF_JSGT >> 4] = BPF_JSLT,
13589 [BPF_JSLE >> 4] = BPF_JSGE,
13590 [BPF_JSLT >> 4] = BPF_JSGT
13592 return opcode_flip[opcode >> 4];
13595 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
13596 struct bpf_reg_state *src_reg,
13599 struct bpf_reg_state *pkt;
13601 if (src_reg->type == PTR_TO_PACKET_END) {
13603 } else if (dst_reg->type == PTR_TO_PACKET_END) {
13605 opcode = flip_opcode(opcode);
13610 if (pkt->range >= 0)
13615 /* pkt <= pkt_end */
13618 /* pkt > pkt_end */
13619 if (pkt->range == BEYOND_PKT_END)
13620 /* pkt has at last one extra byte beyond pkt_end */
13621 return opcode == BPF_JGT;
13624 /* pkt < pkt_end */
13627 /* pkt >= pkt_end */
13628 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
13629 return opcode == BPF_JGE;
13635 /* Adjusts the register min/max values in the case that the dst_reg is the
13636 * variable register that we are working on, and src_reg is a constant or we're
13637 * simply doing a BPF_K check.
13638 * In JEQ/JNE cases we also adjust the var_off values.
13640 static void reg_set_min_max(struct bpf_reg_state *true_reg,
13641 struct bpf_reg_state *false_reg,
13642 u64 val, u32 val32,
13643 u8 opcode, bool is_jmp32)
13645 struct tnum false_32off = tnum_subreg(false_reg->var_off);
13646 struct tnum false_64off = false_reg->var_off;
13647 struct tnum true_32off = tnum_subreg(true_reg->var_off);
13648 struct tnum true_64off = true_reg->var_off;
13649 s64 sval = (s64)val;
13650 s32 sval32 = (s32)val32;
13652 /* If the dst_reg is a pointer, we can't learn anything about its
13653 * variable offset from the compare (unless src_reg were a pointer into
13654 * the same object, but we don't bother with that.
13655 * Since false_reg and true_reg have the same type by construction, we
13656 * only need to check one of them for pointerness.
13658 if (__is_pointer_value(false, false_reg))
13662 /* JEQ/JNE comparison doesn't change the register equivalence.
13665 * if (r1 == 42) goto label;
13667 * label: // here both r1 and r2 are known to be 42.
13669 * Hence when marking register as known preserve it's ID.
13673 __mark_reg32_known(true_reg, val32);
13674 true_32off = tnum_subreg(true_reg->var_off);
13676 ___mark_reg_known(true_reg, val);
13677 true_64off = true_reg->var_off;
13682 __mark_reg32_known(false_reg, val32);
13683 false_32off = tnum_subreg(false_reg->var_off);
13685 ___mark_reg_known(false_reg, val);
13686 false_64off = false_reg->var_off;
13691 false_32off = tnum_and(false_32off, tnum_const(~val32));
13692 if (is_power_of_2(val32))
13693 true_32off = tnum_or(true_32off,
13694 tnum_const(val32));
13696 false_64off = tnum_and(false_64off, tnum_const(~val));
13697 if (is_power_of_2(val))
13698 true_64off = tnum_or(true_64off,
13706 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
13707 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
13709 false_reg->u32_max_value = min(false_reg->u32_max_value,
13711 true_reg->u32_min_value = max(true_reg->u32_min_value,
13714 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
13715 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
13717 false_reg->umax_value = min(false_reg->umax_value, false_umax);
13718 true_reg->umin_value = max(true_reg->umin_value, true_umin);
13726 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
13727 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
13729 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
13730 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
13732 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
13733 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
13735 false_reg->smax_value = min(false_reg->smax_value, false_smax);
13736 true_reg->smin_value = max(true_reg->smin_value, true_smin);
13744 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
13745 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
13747 false_reg->u32_min_value = max(false_reg->u32_min_value,
13749 true_reg->u32_max_value = min(true_reg->u32_max_value,
13752 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
13753 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
13755 false_reg->umin_value = max(false_reg->umin_value, false_umin);
13756 true_reg->umax_value = min(true_reg->umax_value, true_umax);
13764 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
13765 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
13767 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
13768 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
13770 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
13771 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
13773 false_reg->smin_value = max(false_reg->smin_value, false_smin);
13774 true_reg->smax_value = min(true_reg->smax_value, true_smax);
13783 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
13784 tnum_subreg(false_32off));
13785 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
13786 tnum_subreg(true_32off));
13787 __reg_combine_32_into_64(false_reg);
13788 __reg_combine_32_into_64(true_reg);
13790 false_reg->var_off = false_64off;
13791 true_reg->var_off = true_64off;
13792 __reg_combine_64_into_32(false_reg);
13793 __reg_combine_64_into_32(true_reg);
13797 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
13798 * the variable reg.
13800 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
13801 struct bpf_reg_state *false_reg,
13802 u64 val, u32 val32,
13803 u8 opcode, bool is_jmp32)
13805 opcode = flip_opcode(opcode);
13806 /* This uses zero as "not present in table"; luckily the zero opcode,
13807 * BPF_JA, can't get here.
13810 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
13813 /* Regs are known to be equal, so intersect their min/max/var_off */
13814 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
13815 struct bpf_reg_state *dst_reg)
13817 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
13818 dst_reg->umin_value);
13819 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
13820 dst_reg->umax_value);
13821 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
13822 dst_reg->smin_value);
13823 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
13824 dst_reg->smax_value);
13825 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
13827 reg_bounds_sync(src_reg);
13828 reg_bounds_sync(dst_reg);
13831 static void reg_combine_min_max(struct bpf_reg_state *true_src,
13832 struct bpf_reg_state *true_dst,
13833 struct bpf_reg_state *false_src,
13834 struct bpf_reg_state *false_dst,
13839 __reg_combine_min_max(true_src, true_dst);
13842 __reg_combine_min_max(false_src, false_dst);
13847 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
13848 struct bpf_reg_state *reg, u32 id,
13851 if (type_may_be_null(reg->type) && reg->id == id &&
13852 (is_rcu_reg(reg) || !WARN_ON_ONCE(!reg->id))) {
13853 /* Old offset (both fixed and variable parts) should have been
13854 * known-zero, because we don't allow pointer arithmetic on
13855 * pointers that might be NULL. If we see this happening, don't
13856 * convert the register.
13858 * But in some cases, some helpers that return local kptrs
13859 * advance offset for the returned pointer. In those cases, it
13860 * is fine to expect to see reg->off.
13862 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value || !tnum_equals_const(reg->var_off, 0)))
13864 if (!(type_is_ptr_alloc_obj(reg->type) || type_is_non_owning_ref(reg->type)) &&
13865 WARN_ON_ONCE(reg->off))
13869 reg->type = SCALAR_VALUE;
13870 /* We don't need id and ref_obj_id from this point
13871 * onwards anymore, thus we should better reset it,
13872 * so that state pruning has chances to take effect.
13875 reg->ref_obj_id = 0;
13880 mark_ptr_not_null_reg(reg);
13882 if (!reg_may_point_to_spin_lock(reg)) {
13883 /* For not-NULL ptr, reg->ref_obj_id will be reset
13884 * in release_reference().
13886 * reg->id is still used by spin_lock ptr. Other
13887 * than spin_lock ptr type, reg->id can be reset.
13894 /* The logic is similar to find_good_pkt_pointers(), both could eventually
13895 * be folded together at some point.
13897 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
13900 struct bpf_func_state *state = vstate->frame[vstate->curframe];
13901 struct bpf_reg_state *regs = state->regs, *reg;
13902 u32 ref_obj_id = regs[regno].ref_obj_id;
13903 u32 id = regs[regno].id;
13905 if (ref_obj_id && ref_obj_id == id && is_null)
13906 /* regs[regno] is in the " == NULL" branch.
13907 * No one could have freed the reference state before
13908 * doing the NULL check.
13910 WARN_ON_ONCE(release_reference_state(state, id));
13912 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
13913 mark_ptr_or_null_reg(state, reg, id, is_null);
13917 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
13918 struct bpf_reg_state *dst_reg,
13919 struct bpf_reg_state *src_reg,
13920 struct bpf_verifier_state *this_branch,
13921 struct bpf_verifier_state *other_branch)
13923 if (BPF_SRC(insn->code) != BPF_X)
13926 /* Pointers are always 64-bit. */
13927 if (BPF_CLASS(insn->code) == BPF_JMP32)
13930 switch (BPF_OP(insn->code)) {
13932 if ((dst_reg->type == PTR_TO_PACKET &&
13933 src_reg->type == PTR_TO_PACKET_END) ||
13934 (dst_reg->type == PTR_TO_PACKET_META &&
13935 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13936 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
13937 find_good_pkt_pointers(this_branch, dst_reg,
13938 dst_reg->type, false);
13939 mark_pkt_end(other_branch, insn->dst_reg, true);
13940 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
13941 src_reg->type == PTR_TO_PACKET) ||
13942 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
13943 src_reg->type == PTR_TO_PACKET_META)) {
13944 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
13945 find_good_pkt_pointers(other_branch, src_reg,
13946 src_reg->type, true);
13947 mark_pkt_end(this_branch, insn->src_reg, false);
13953 if ((dst_reg->type == PTR_TO_PACKET &&
13954 src_reg->type == PTR_TO_PACKET_END) ||
13955 (dst_reg->type == PTR_TO_PACKET_META &&
13956 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13957 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
13958 find_good_pkt_pointers(other_branch, dst_reg,
13959 dst_reg->type, true);
13960 mark_pkt_end(this_branch, insn->dst_reg, false);
13961 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
13962 src_reg->type == PTR_TO_PACKET) ||
13963 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
13964 src_reg->type == PTR_TO_PACKET_META)) {
13965 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
13966 find_good_pkt_pointers(this_branch, src_reg,
13967 src_reg->type, false);
13968 mark_pkt_end(other_branch, insn->src_reg, true);
13974 if ((dst_reg->type == PTR_TO_PACKET &&
13975 src_reg->type == PTR_TO_PACKET_END) ||
13976 (dst_reg->type == PTR_TO_PACKET_META &&
13977 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13978 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
13979 find_good_pkt_pointers(this_branch, dst_reg,
13980 dst_reg->type, true);
13981 mark_pkt_end(other_branch, insn->dst_reg, false);
13982 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
13983 src_reg->type == PTR_TO_PACKET) ||
13984 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
13985 src_reg->type == PTR_TO_PACKET_META)) {
13986 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
13987 find_good_pkt_pointers(other_branch, src_reg,
13988 src_reg->type, false);
13989 mark_pkt_end(this_branch, insn->src_reg, true);
13995 if ((dst_reg->type == PTR_TO_PACKET &&
13996 src_reg->type == PTR_TO_PACKET_END) ||
13997 (dst_reg->type == PTR_TO_PACKET_META &&
13998 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13999 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
14000 find_good_pkt_pointers(other_branch, dst_reg,
14001 dst_reg->type, false);
14002 mark_pkt_end(this_branch, insn->dst_reg, true);
14003 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
14004 src_reg->type == PTR_TO_PACKET) ||
14005 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
14006 src_reg->type == PTR_TO_PACKET_META)) {
14007 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
14008 find_good_pkt_pointers(this_branch, src_reg,
14009 src_reg->type, true);
14010 mark_pkt_end(other_branch, insn->src_reg, false);
14022 static void find_equal_scalars(struct bpf_verifier_state *vstate,
14023 struct bpf_reg_state *known_reg)
14025 struct bpf_func_state *state;
14026 struct bpf_reg_state *reg;
14028 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
14029 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
14030 copy_register_state(reg, known_reg);
14034 static int check_cond_jmp_op(struct bpf_verifier_env *env,
14035 struct bpf_insn *insn, int *insn_idx)
14037 struct bpf_verifier_state *this_branch = env->cur_state;
14038 struct bpf_verifier_state *other_branch;
14039 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
14040 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
14041 struct bpf_reg_state *eq_branch_regs;
14042 u8 opcode = BPF_OP(insn->code);
14047 /* Only conditional jumps are expected to reach here. */
14048 if (opcode == BPF_JA || opcode > BPF_JSLE) {
14049 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
14053 /* check src2 operand */
14054 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
14058 dst_reg = ®s[insn->dst_reg];
14059 if (BPF_SRC(insn->code) == BPF_X) {
14060 if (insn->imm != 0) {
14061 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
14065 /* check src1 operand */
14066 err = check_reg_arg(env, insn->src_reg, SRC_OP);
14070 src_reg = ®s[insn->src_reg];
14071 if (!(reg_is_pkt_pointer_any(dst_reg) && reg_is_pkt_pointer_any(src_reg)) &&
14072 is_pointer_value(env, insn->src_reg)) {
14073 verbose(env, "R%d pointer comparison prohibited\n",
14078 if (insn->src_reg != BPF_REG_0) {
14079 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
14084 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
14086 if (BPF_SRC(insn->code) == BPF_K) {
14087 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
14088 } else if (src_reg->type == SCALAR_VALUE &&
14089 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
14090 pred = is_branch_taken(dst_reg,
14091 tnum_subreg(src_reg->var_off).value,
14094 } else if (src_reg->type == SCALAR_VALUE &&
14095 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
14096 pred = is_branch_taken(dst_reg,
14097 src_reg->var_off.value,
14100 } else if (dst_reg->type == SCALAR_VALUE &&
14101 is_jmp32 && tnum_is_const(tnum_subreg(dst_reg->var_off))) {
14102 pred = is_branch_taken(src_reg,
14103 tnum_subreg(dst_reg->var_off).value,
14104 flip_opcode(opcode),
14106 } else if (dst_reg->type == SCALAR_VALUE &&
14107 !is_jmp32 && tnum_is_const(dst_reg->var_off)) {
14108 pred = is_branch_taken(src_reg,
14109 dst_reg->var_off.value,
14110 flip_opcode(opcode),
14112 } else if (reg_is_pkt_pointer_any(dst_reg) &&
14113 reg_is_pkt_pointer_any(src_reg) &&
14115 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
14119 /* If we get here with a dst_reg pointer type it is because
14120 * above is_branch_taken() special cased the 0 comparison.
14122 if (!__is_pointer_value(false, dst_reg))
14123 err = mark_chain_precision(env, insn->dst_reg);
14124 if (BPF_SRC(insn->code) == BPF_X && !err &&
14125 !__is_pointer_value(false, src_reg))
14126 err = mark_chain_precision(env, insn->src_reg);
14132 /* Only follow the goto, ignore fall-through. If needed, push
14133 * the fall-through branch for simulation under speculative
14136 if (!env->bypass_spec_v1 &&
14137 !sanitize_speculative_path(env, insn, *insn_idx + 1,
14140 *insn_idx += insn->off;
14142 } else if (pred == 0) {
14143 /* Only follow the fall-through branch, since that's where the
14144 * program will go. If needed, push the goto branch for
14145 * simulation under speculative execution.
14147 if (!env->bypass_spec_v1 &&
14148 !sanitize_speculative_path(env, insn,
14149 *insn_idx + insn->off + 1,
14155 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
14159 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
14161 /* detect if we are comparing against a constant value so we can adjust
14162 * our min/max values for our dst register.
14163 * this is only legit if both are scalars (or pointers to the same
14164 * object, I suppose, see the PTR_MAYBE_NULL related if block below),
14165 * because otherwise the different base pointers mean the offsets aren't
14168 if (BPF_SRC(insn->code) == BPF_X) {
14169 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
14171 if (dst_reg->type == SCALAR_VALUE &&
14172 src_reg->type == SCALAR_VALUE) {
14173 if (tnum_is_const(src_reg->var_off) ||
14175 tnum_is_const(tnum_subreg(src_reg->var_off))))
14176 reg_set_min_max(&other_branch_regs[insn->dst_reg],
14178 src_reg->var_off.value,
14179 tnum_subreg(src_reg->var_off).value,
14181 else if (tnum_is_const(dst_reg->var_off) ||
14183 tnum_is_const(tnum_subreg(dst_reg->var_off))))
14184 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
14186 dst_reg->var_off.value,
14187 tnum_subreg(dst_reg->var_off).value,
14189 else if (!is_jmp32 &&
14190 (opcode == BPF_JEQ || opcode == BPF_JNE))
14191 /* Comparing for equality, we can combine knowledge */
14192 reg_combine_min_max(&other_branch_regs[insn->src_reg],
14193 &other_branch_regs[insn->dst_reg],
14194 src_reg, dst_reg, opcode);
14196 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
14197 find_equal_scalars(this_branch, src_reg);
14198 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
14202 } else if (dst_reg->type == SCALAR_VALUE) {
14203 reg_set_min_max(&other_branch_regs[insn->dst_reg],
14204 dst_reg, insn->imm, (u32)insn->imm,
14208 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
14209 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
14210 find_equal_scalars(this_branch, dst_reg);
14211 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
14214 /* if one pointer register is compared to another pointer
14215 * register check if PTR_MAYBE_NULL could be lifted.
14216 * E.g. register A - maybe null
14217 * register B - not null
14218 * for JNE A, B, ... - A is not null in the false branch;
14219 * for JEQ A, B, ... - A is not null in the true branch.
14221 * Since PTR_TO_BTF_ID points to a kernel struct that does
14222 * not need to be null checked by the BPF program, i.e.,
14223 * could be null even without PTR_MAYBE_NULL marking, so
14224 * only propagate nullness when neither reg is that type.
14226 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_X &&
14227 __is_pointer_value(false, src_reg) && __is_pointer_value(false, dst_reg) &&
14228 type_may_be_null(src_reg->type) != type_may_be_null(dst_reg->type) &&
14229 base_type(src_reg->type) != PTR_TO_BTF_ID &&
14230 base_type(dst_reg->type) != PTR_TO_BTF_ID) {
14231 eq_branch_regs = NULL;
14234 eq_branch_regs = other_branch_regs;
14237 eq_branch_regs = regs;
14243 if (eq_branch_regs) {
14244 if (type_may_be_null(src_reg->type))
14245 mark_ptr_not_null_reg(&eq_branch_regs[insn->src_reg]);
14247 mark_ptr_not_null_reg(&eq_branch_regs[insn->dst_reg]);
14251 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
14252 * NOTE: these optimizations below are related with pointer comparison
14253 * which will never be JMP32.
14255 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
14256 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
14257 type_may_be_null(dst_reg->type)) {
14258 /* Mark all identical registers in each branch as either
14259 * safe or unknown depending R == 0 or R != 0 conditional.
14261 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
14262 opcode == BPF_JNE);
14263 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
14264 opcode == BPF_JEQ);
14265 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
14266 this_branch, other_branch) &&
14267 is_pointer_value(env, insn->dst_reg)) {
14268 verbose(env, "R%d pointer comparison prohibited\n",
14272 if (env->log.level & BPF_LOG_LEVEL)
14273 print_insn_state(env, this_branch->frame[this_branch->curframe]);
14277 /* verify BPF_LD_IMM64 instruction */
14278 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
14280 struct bpf_insn_aux_data *aux = cur_aux(env);
14281 struct bpf_reg_state *regs = cur_regs(env);
14282 struct bpf_reg_state *dst_reg;
14283 struct bpf_map *map;
14286 if (BPF_SIZE(insn->code) != BPF_DW) {
14287 verbose(env, "invalid BPF_LD_IMM insn\n");
14290 if (insn->off != 0) {
14291 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
14295 err = check_reg_arg(env, insn->dst_reg, DST_OP);
14299 dst_reg = ®s[insn->dst_reg];
14300 if (insn->src_reg == 0) {
14301 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
14303 dst_reg->type = SCALAR_VALUE;
14304 __mark_reg_known(®s[insn->dst_reg], imm);
14308 /* All special src_reg cases are listed below. From this point onwards
14309 * we either succeed and assign a corresponding dst_reg->type after
14310 * zeroing the offset, or fail and reject the program.
14312 mark_reg_known_zero(env, regs, insn->dst_reg);
14314 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
14315 dst_reg->type = aux->btf_var.reg_type;
14316 switch (base_type(dst_reg->type)) {
14318 dst_reg->mem_size = aux->btf_var.mem_size;
14320 case PTR_TO_BTF_ID:
14321 dst_reg->btf = aux->btf_var.btf;
14322 dst_reg->btf_id = aux->btf_var.btf_id;
14325 verbose(env, "bpf verifier is misconfigured\n");
14331 if (insn->src_reg == BPF_PSEUDO_FUNC) {
14332 struct bpf_prog_aux *aux = env->prog->aux;
14333 u32 subprogno = find_subprog(env,
14334 env->insn_idx + insn->imm + 1);
14336 if (!aux->func_info) {
14337 verbose(env, "missing btf func_info\n");
14340 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
14341 verbose(env, "callback function not static\n");
14345 dst_reg->type = PTR_TO_FUNC;
14346 dst_reg->subprogno = subprogno;
14350 map = env->used_maps[aux->map_index];
14351 dst_reg->map_ptr = map;
14353 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
14354 insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
14355 dst_reg->type = PTR_TO_MAP_VALUE;
14356 dst_reg->off = aux->map_off;
14357 WARN_ON_ONCE(map->max_entries != 1);
14358 /* We want reg->id to be same (0) as map_value is not distinct */
14359 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
14360 insn->src_reg == BPF_PSEUDO_MAP_IDX) {
14361 dst_reg->type = CONST_PTR_TO_MAP;
14363 verbose(env, "bpf verifier is misconfigured\n");
14370 static bool may_access_skb(enum bpf_prog_type type)
14373 case BPF_PROG_TYPE_SOCKET_FILTER:
14374 case BPF_PROG_TYPE_SCHED_CLS:
14375 case BPF_PROG_TYPE_SCHED_ACT:
14382 /* verify safety of LD_ABS|LD_IND instructions:
14383 * - they can only appear in the programs where ctx == skb
14384 * - since they are wrappers of function calls, they scratch R1-R5 registers,
14385 * preserve R6-R9, and store return value into R0
14388 * ctx == skb == R6 == CTX
14391 * SRC == any register
14392 * IMM == 32-bit immediate
14395 * R0 - 8/16/32-bit skb data converted to cpu endianness
14397 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
14399 struct bpf_reg_state *regs = cur_regs(env);
14400 static const int ctx_reg = BPF_REG_6;
14401 u8 mode = BPF_MODE(insn->code);
14404 if (!may_access_skb(resolve_prog_type(env->prog))) {
14405 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
14409 if (!env->ops->gen_ld_abs) {
14410 verbose(env, "bpf verifier is misconfigured\n");
14414 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
14415 BPF_SIZE(insn->code) == BPF_DW ||
14416 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
14417 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
14421 /* check whether implicit source operand (register R6) is readable */
14422 err = check_reg_arg(env, ctx_reg, SRC_OP);
14426 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
14427 * gen_ld_abs() may terminate the program at runtime, leading to
14430 err = check_reference_leak(env);
14432 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
14436 if (env->cur_state->active_lock.ptr) {
14437 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
14441 if (env->cur_state->active_rcu_lock) {
14442 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_rcu_read_lock-ed region\n");
14446 if (regs[ctx_reg].type != PTR_TO_CTX) {
14448 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
14452 if (mode == BPF_IND) {
14453 /* check explicit source operand */
14454 err = check_reg_arg(env, insn->src_reg, SRC_OP);
14459 err = check_ptr_off_reg(env, ®s[ctx_reg], ctx_reg);
14463 /* reset caller saved regs to unreadable */
14464 for (i = 0; i < CALLER_SAVED_REGS; i++) {
14465 mark_reg_not_init(env, regs, caller_saved[i]);
14466 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
14469 /* mark destination R0 register as readable, since it contains
14470 * the value fetched from the packet.
14471 * Already marked as written above.
14473 mark_reg_unknown(env, regs, BPF_REG_0);
14474 /* ld_abs load up to 32-bit skb data. */
14475 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
14479 static int check_return_code(struct bpf_verifier_env *env)
14481 struct tnum enforce_attach_type_range = tnum_unknown;
14482 const struct bpf_prog *prog = env->prog;
14483 struct bpf_reg_state *reg;
14484 struct tnum range = tnum_range(0, 1);
14485 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
14487 struct bpf_func_state *frame = env->cur_state->frame[0];
14488 const bool is_subprog = frame->subprogno;
14490 /* LSM and struct_ops func-ptr's return type could be "void" */
14492 switch (prog_type) {
14493 case BPF_PROG_TYPE_LSM:
14494 if (prog->expected_attach_type == BPF_LSM_CGROUP)
14495 /* See below, can be 0 or 0-1 depending on hook. */
14498 case BPF_PROG_TYPE_STRUCT_OPS:
14499 if (!prog->aux->attach_func_proto->type)
14507 /* eBPF calling convention is such that R0 is used
14508 * to return the value from eBPF program.
14509 * Make sure that it's readable at this time
14510 * of bpf_exit, which means that program wrote
14511 * something into it earlier
14513 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
14517 if (is_pointer_value(env, BPF_REG_0)) {
14518 verbose(env, "R0 leaks addr as return value\n");
14522 reg = cur_regs(env) + BPF_REG_0;
14524 if (frame->in_async_callback_fn) {
14525 /* enforce return zero from async callbacks like timer */
14526 if (reg->type != SCALAR_VALUE) {
14527 verbose(env, "In async callback the register R0 is not a known value (%s)\n",
14528 reg_type_str(env, reg->type));
14532 if (!tnum_in(tnum_const(0), reg->var_off)) {
14533 verbose_invalid_scalar(env, reg, &range, "async callback", "R0");
14540 if (reg->type != SCALAR_VALUE) {
14541 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
14542 reg_type_str(env, reg->type));
14548 switch (prog_type) {
14549 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
14550 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
14551 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
14552 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
14553 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
14554 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
14555 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
14556 range = tnum_range(1, 1);
14557 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
14558 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
14559 range = tnum_range(0, 3);
14561 case BPF_PROG_TYPE_CGROUP_SKB:
14562 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
14563 range = tnum_range(0, 3);
14564 enforce_attach_type_range = tnum_range(2, 3);
14567 case BPF_PROG_TYPE_CGROUP_SOCK:
14568 case BPF_PROG_TYPE_SOCK_OPS:
14569 case BPF_PROG_TYPE_CGROUP_DEVICE:
14570 case BPF_PROG_TYPE_CGROUP_SYSCTL:
14571 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
14573 case BPF_PROG_TYPE_RAW_TRACEPOINT:
14574 if (!env->prog->aux->attach_btf_id)
14576 range = tnum_const(0);
14578 case BPF_PROG_TYPE_TRACING:
14579 switch (env->prog->expected_attach_type) {
14580 case BPF_TRACE_FENTRY:
14581 case BPF_TRACE_FEXIT:
14582 range = tnum_const(0);
14584 case BPF_TRACE_RAW_TP:
14585 case BPF_MODIFY_RETURN:
14587 case BPF_TRACE_ITER:
14593 case BPF_PROG_TYPE_SK_LOOKUP:
14594 range = tnum_range(SK_DROP, SK_PASS);
14597 case BPF_PROG_TYPE_LSM:
14598 if (env->prog->expected_attach_type != BPF_LSM_CGROUP) {
14599 /* Regular BPF_PROG_TYPE_LSM programs can return
14604 if (!env->prog->aux->attach_func_proto->type) {
14605 /* Make sure programs that attach to void
14606 * hooks don't try to modify return value.
14608 range = tnum_range(1, 1);
14612 case BPF_PROG_TYPE_NETFILTER:
14613 range = tnum_range(NF_DROP, NF_ACCEPT);
14615 case BPF_PROG_TYPE_EXT:
14616 /* freplace program can return anything as its return value
14617 * depends on the to-be-replaced kernel func or bpf program.
14623 if (reg->type != SCALAR_VALUE) {
14624 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
14625 reg_type_str(env, reg->type));
14629 if (!tnum_in(range, reg->var_off)) {
14630 verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
14631 if (prog->expected_attach_type == BPF_LSM_CGROUP &&
14632 prog_type == BPF_PROG_TYPE_LSM &&
14633 !prog->aux->attach_func_proto->type)
14634 verbose(env, "Note, BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
14638 if (!tnum_is_unknown(enforce_attach_type_range) &&
14639 tnum_in(enforce_attach_type_range, reg->var_off))
14640 env->prog->enforce_expected_attach_type = 1;
14644 /* non-recursive DFS pseudo code
14645 * 1 procedure DFS-iterative(G,v):
14646 * 2 label v as discovered
14647 * 3 let S be a stack
14649 * 5 while S is not empty
14651 * 7 if t is what we're looking for:
14653 * 9 for all edges e in G.adjacentEdges(t) do
14654 * 10 if edge e is already labelled
14655 * 11 continue with the next edge
14656 * 12 w <- G.adjacentVertex(t,e)
14657 * 13 if vertex w is not discovered and not explored
14658 * 14 label e as tree-edge
14659 * 15 label w as discovered
14662 * 18 else if vertex w is discovered
14663 * 19 label e as back-edge
14665 * 21 // vertex w is explored
14666 * 22 label e as forward- or cross-edge
14667 * 23 label t as explored
14671 * 0x10 - discovered
14672 * 0x11 - discovered and fall-through edge labelled
14673 * 0x12 - discovered and fall-through and branch edges labelled
14684 static u32 state_htab_size(struct bpf_verifier_env *env)
14686 return env->prog->len;
14689 static struct bpf_verifier_state_list **explored_state(
14690 struct bpf_verifier_env *env,
14693 struct bpf_verifier_state *cur = env->cur_state;
14694 struct bpf_func_state *state = cur->frame[cur->curframe];
14696 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
14699 static void mark_prune_point(struct bpf_verifier_env *env, int idx)
14701 env->insn_aux_data[idx].prune_point = true;
14704 static bool is_prune_point(struct bpf_verifier_env *env, int insn_idx)
14706 return env->insn_aux_data[insn_idx].prune_point;
14709 static void mark_force_checkpoint(struct bpf_verifier_env *env, int idx)
14711 env->insn_aux_data[idx].force_checkpoint = true;
14714 static bool is_force_checkpoint(struct bpf_verifier_env *env, int insn_idx)
14716 return env->insn_aux_data[insn_idx].force_checkpoint;
14721 DONE_EXPLORING = 0,
14722 KEEP_EXPLORING = 1,
14725 /* t, w, e - match pseudo-code above:
14726 * t - index of current instruction
14727 * w - next instruction
14730 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
14733 int *insn_stack = env->cfg.insn_stack;
14734 int *insn_state = env->cfg.insn_state;
14736 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
14737 return DONE_EXPLORING;
14739 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
14740 return DONE_EXPLORING;
14742 if (w < 0 || w >= env->prog->len) {
14743 verbose_linfo(env, t, "%d: ", t);
14744 verbose(env, "jump out of range from insn %d to %d\n", t, w);
14749 /* mark branch target for state pruning */
14750 mark_prune_point(env, w);
14751 mark_jmp_point(env, w);
14754 if (insn_state[w] == 0) {
14756 insn_state[t] = DISCOVERED | e;
14757 insn_state[w] = DISCOVERED;
14758 if (env->cfg.cur_stack >= env->prog->len)
14760 insn_stack[env->cfg.cur_stack++] = w;
14761 return KEEP_EXPLORING;
14762 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
14763 if (loop_ok && env->bpf_capable)
14764 return DONE_EXPLORING;
14765 verbose_linfo(env, t, "%d: ", t);
14766 verbose_linfo(env, w, "%d: ", w);
14767 verbose(env, "back-edge from insn %d to %d\n", t, w);
14769 } else if (insn_state[w] == EXPLORED) {
14770 /* forward- or cross-edge */
14771 insn_state[t] = DISCOVERED | e;
14773 verbose(env, "insn state internal bug\n");
14776 return DONE_EXPLORING;
14779 static int visit_func_call_insn(int t, struct bpf_insn *insns,
14780 struct bpf_verifier_env *env,
14785 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
14789 mark_prune_point(env, t + 1);
14790 /* when we exit from subprog, we need to record non-linear history */
14791 mark_jmp_point(env, t + 1);
14793 if (visit_callee) {
14794 mark_prune_point(env, t);
14795 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env,
14796 /* It's ok to allow recursion from CFG point of
14797 * view. __check_func_call() will do the actual
14800 bpf_pseudo_func(insns + t));
14805 /* Visits the instruction at index t and returns one of the following:
14806 * < 0 - an error occurred
14807 * DONE_EXPLORING - the instruction was fully explored
14808 * KEEP_EXPLORING - there is still work to be done before it is fully explored
14810 static int visit_insn(int t, struct bpf_verifier_env *env)
14812 struct bpf_insn *insns = env->prog->insnsi, *insn = &insns[t];
14815 if (bpf_pseudo_func(insn))
14816 return visit_func_call_insn(t, insns, env, true);
14818 /* All non-branch instructions have a single fall-through edge. */
14819 if (BPF_CLASS(insn->code) != BPF_JMP &&
14820 BPF_CLASS(insn->code) != BPF_JMP32)
14821 return push_insn(t, t + 1, FALLTHROUGH, env, false);
14823 switch (BPF_OP(insn->code)) {
14825 return DONE_EXPLORING;
14828 if (insn->src_reg == 0 && insn->imm == BPF_FUNC_timer_set_callback)
14829 /* Mark this call insn as a prune point to trigger
14830 * is_state_visited() check before call itself is
14831 * processed by __check_func_call(). Otherwise new
14832 * async state will be pushed for further exploration.
14834 mark_prune_point(env, t);
14835 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
14836 struct bpf_kfunc_call_arg_meta meta;
14838 ret = fetch_kfunc_meta(env, insn, &meta, NULL);
14839 if (ret == 0 && is_iter_next_kfunc(&meta)) {
14840 mark_prune_point(env, t);
14841 /* Checking and saving state checkpoints at iter_next() call
14842 * is crucial for fast convergence of open-coded iterator loop
14843 * logic, so we need to force it. If we don't do that,
14844 * is_state_visited() might skip saving a checkpoint, causing
14845 * unnecessarily long sequence of not checkpointed
14846 * instructions and jumps, leading to exhaustion of jump
14847 * history buffer, and potentially other undesired outcomes.
14848 * It is expected that with correct open-coded iterators
14849 * convergence will happen quickly, so we don't run a risk of
14850 * exhausting memory.
14852 mark_force_checkpoint(env, t);
14855 return visit_func_call_insn(t, insns, env, insn->src_reg == BPF_PSEUDO_CALL);
14858 if (BPF_SRC(insn->code) != BPF_K)
14861 if (BPF_CLASS(insn->code) == BPF_JMP)
14866 /* unconditional jump with single edge */
14867 ret = push_insn(t, t + off + 1, FALLTHROUGH, env,
14872 mark_prune_point(env, t + off + 1);
14873 mark_jmp_point(env, t + off + 1);
14878 /* conditional jump with two edges */
14879 mark_prune_point(env, t);
14881 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
14885 return push_insn(t, t + insn->off + 1, BRANCH, env, true);
14889 /* non-recursive depth-first-search to detect loops in BPF program
14890 * loop == back-edge in directed graph
14892 static int check_cfg(struct bpf_verifier_env *env)
14894 int insn_cnt = env->prog->len;
14895 int *insn_stack, *insn_state;
14899 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
14903 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
14905 kvfree(insn_state);
14909 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
14910 insn_stack[0] = 0; /* 0 is the first instruction */
14911 env->cfg.cur_stack = 1;
14913 while (env->cfg.cur_stack > 0) {
14914 int t = insn_stack[env->cfg.cur_stack - 1];
14916 ret = visit_insn(t, env);
14918 case DONE_EXPLORING:
14919 insn_state[t] = EXPLORED;
14920 env->cfg.cur_stack--;
14922 case KEEP_EXPLORING:
14926 verbose(env, "visit_insn internal bug\n");
14933 if (env->cfg.cur_stack < 0) {
14934 verbose(env, "pop stack internal bug\n");
14939 for (i = 0; i < insn_cnt; i++) {
14940 if (insn_state[i] != EXPLORED) {
14941 verbose(env, "unreachable insn %d\n", i);
14946 ret = 0; /* cfg looks good */
14949 kvfree(insn_state);
14950 kvfree(insn_stack);
14951 env->cfg.insn_state = env->cfg.insn_stack = NULL;
14955 static int check_abnormal_return(struct bpf_verifier_env *env)
14959 for (i = 1; i < env->subprog_cnt; i++) {
14960 if (env->subprog_info[i].has_ld_abs) {
14961 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
14964 if (env->subprog_info[i].has_tail_call) {
14965 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
14972 /* The minimum supported BTF func info size */
14973 #define MIN_BPF_FUNCINFO_SIZE 8
14974 #define MAX_FUNCINFO_REC_SIZE 252
14976 static int check_btf_func(struct bpf_verifier_env *env,
14977 const union bpf_attr *attr,
14980 const struct btf_type *type, *func_proto, *ret_type;
14981 u32 i, nfuncs, urec_size, min_size;
14982 u32 krec_size = sizeof(struct bpf_func_info);
14983 struct bpf_func_info *krecord;
14984 struct bpf_func_info_aux *info_aux = NULL;
14985 struct bpf_prog *prog;
14986 const struct btf *btf;
14988 u32 prev_offset = 0;
14989 bool scalar_return;
14992 nfuncs = attr->func_info_cnt;
14994 if (check_abnormal_return(env))
14999 if (nfuncs != env->subprog_cnt) {
15000 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
15004 urec_size = attr->func_info_rec_size;
15005 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
15006 urec_size > MAX_FUNCINFO_REC_SIZE ||
15007 urec_size % sizeof(u32)) {
15008 verbose(env, "invalid func info rec size %u\n", urec_size);
15013 btf = prog->aux->btf;
15015 urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
15016 min_size = min_t(u32, krec_size, urec_size);
15018 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
15021 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
15025 for (i = 0; i < nfuncs; i++) {
15026 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
15028 if (ret == -E2BIG) {
15029 verbose(env, "nonzero tailing record in func info");
15030 /* set the size kernel expects so loader can zero
15031 * out the rest of the record.
15033 if (copy_to_bpfptr_offset(uattr,
15034 offsetof(union bpf_attr, func_info_rec_size),
15035 &min_size, sizeof(min_size)))
15041 if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
15046 /* check insn_off */
15049 if (krecord[i].insn_off) {
15051 "nonzero insn_off %u for the first func info record",
15052 krecord[i].insn_off);
15055 } else if (krecord[i].insn_off <= prev_offset) {
15057 "same or smaller insn offset (%u) than previous func info record (%u)",
15058 krecord[i].insn_off, prev_offset);
15062 if (env->subprog_info[i].start != krecord[i].insn_off) {
15063 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
15067 /* check type_id */
15068 type = btf_type_by_id(btf, krecord[i].type_id);
15069 if (!type || !btf_type_is_func(type)) {
15070 verbose(env, "invalid type id %d in func info",
15071 krecord[i].type_id);
15074 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
15076 func_proto = btf_type_by_id(btf, type->type);
15077 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
15078 /* btf_func_check() already verified it during BTF load */
15080 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
15082 btf_type_is_small_int(ret_type) || btf_is_any_enum(ret_type);
15083 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
15084 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
15087 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
15088 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
15092 prev_offset = krecord[i].insn_off;
15093 bpfptr_add(&urecord, urec_size);
15096 prog->aux->func_info = krecord;
15097 prog->aux->func_info_cnt = nfuncs;
15098 prog->aux->func_info_aux = info_aux;
15107 static void adjust_btf_func(struct bpf_verifier_env *env)
15109 struct bpf_prog_aux *aux = env->prog->aux;
15112 if (!aux->func_info)
15115 for (i = 0; i < env->subprog_cnt; i++)
15116 aux->func_info[i].insn_off = env->subprog_info[i].start;
15119 #define MIN_BPF_LINEINFO_SIZE offsetofend(struct bpf_line_info, line_col)
15120 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
15122 static int check_btf_line(struct bpf_verifier_env *env,
15123 const union bpf_attr *attr,
15126 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
15127 struct bpf_subprog_info *sub;
15128 struct bpf_line_info *linfo;
15129 struct bpf_prog *prog;
15130 const struct btf *btf;
15134 nr_linfo = attr->line_info_cnt;
15137 if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info))
15140 rec_size = attr->line_info_rec_size;
15141 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
15142 rec_size > MAX_LINEINFO_REC_SIZE ||
15143 rec_size & (sizeof(u32) - 1))
15146 /* Need to zero it in case the userspace may
15147 * pass in a smaller bpf_line_info object.
15149 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
15150 GFP_KERNEL | __GFP_NOWARN);
15155 btf = prog->aux->btf;
15158 sub = env->subprog_info;
15159 ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
15160 expected_size = sizeof(struct bpf_line_info);
15161 ncopy = min_t(u32, expected_size, rec_size);
15162 for (i = 0; i < nr_linfo; i++) {
15163 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
15165 if (err == -E2BIG) {
15166 verbose(env, "nonzero tailing record in line_info");
15167 if (copy_to_bpfptr_offset(uattr,
15168 offsetof(union bpf_attr, line_info_rec_size),
15169 &expected_size, sizeof(expected_size)))
15175 if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
15181 * Check insn_off to ensure
15182 * 1) strictly increasing AND
15183 * 2) bounded by prog->len
15185 * The linfo[0].insn_off == 0 check logically falls into
15186 * the later "missing bpf_line_info for func..." case
15187 * because the first linfo[0].insn_off must be the
15188 * first sub also and the first sub must have
15189 * subprog_info[0].start == 0.
15191 if ((i && linfo[i].insn_off <= prev_offset) ||
15192 linfo[i].insn_off >= prog->len) {
15193 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
15194 i, linfo[i].insn_off, prev_offset,
15200 if (!prog->insnsi[linfo[i].insn_off].code) {
15202 "Invalid insn code at line_info[%u].insn_off\n",
15208 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
15209 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
15210 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
15215 if (s != env->subprog_cnt) {
15216 if (linfo[i].insn_off == sub[s].start) {
15217 sub[s].linfo_idx = i;
15219 } else if (sub[s].start < linfo[i].insn_off) {
15220 verbose(env, "missing bpf_line_info for func#%u\n", s);
15226 prev_offset = linfo[i].insn_off;
15227 bpfptr_add(&ulinfo, rec_size);
15230 if (s != env->subprog_cnt) {
15231 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
15232 env->subprog_cnt - s, s);
15237 prog->aux->linfo = linfo;
15238 prog->aux->nr_linfo = nr_linfo;
15247 #define MIN_CORE_RELO_SIZE sizeof(struct bpf_core_relo)
15248 #define MAX_CORE_RELO_SIZE MAX_FUNCINFO_REC_SIZE
15250 static int check_core_relo(struct bpf_verifier_env *env,
15251 const union bpf_attr *attr,
15254 u32 i, nr_core_relo, ncopy, expected_size, rec_size;
15255 struct bpf_core_relo core_relo = {};
15256 struct bpf_prog *prog = env->prog;
15257 const struct btf *btf = prog->aux->btf;
15258 struct bpf_core_ctx ctx = {
15262 bpfptr_t u_core_relo;
15265 nr_core_relo = attr->core_relo_cnt;
15268 if (nr_core_relo > INT_MAX / sizeof(struct bpf_core_relo))
15271 rec_size = attr->core_relo_rec_size;
15272 if (rec_size < MIN_CORE_RELO_SIZE ||
15273 rec_size > MAX_CORE_RELO_SIZE ||
15274 rec_size % sizeof(u32))
15277 u_core_relo = make_bpfptr(attr->core_relos, uattr.is_kernel);
15278 expected_size = sizeof(struct bpf_core_relo);
15279 ncopy = min_t(u32, expected_size, rec_size);
15281 /* Unlike func_info and line_info, copy and apply each CO-RE
15282 * relocation record one at a time.
15284 for (i = 0; i < nr_core_relo; i++) {
15285 /* future proofing when sizeof(bpf_core_relo) changes */
15286 err = bpf_check_uarg_tail_zero(u_core_relo, expected_size, rec_size);
15288 if (err == -E2BIG) {
15289 verbose(env, "nonzero tailing record in core_relo");
15290 if (copy_to_bpfptr_offset(uattr,
15291 offsetof(union bpf_attr, core_relo_rec_size),
15292 &expected_size, sizeof(expected_size)))
15298 if (copy_from_bpfptr(&core_relo, u_core_relo, ncopy)) {
15303 if (core_relo.insn_off % 8 || core_relo.insn_off / 8 >= prog->len) {
15304 verbose(env, "Invalid core_relo[%u].insn_off:%u prog->len:%u\n",
15305 i, core_relo.insn_off, prog->len);
15310 err = bpf_core_apply(&ctx, &core_relo, i,
15311 &prog->insnsi[core_relo.insn_off / 8]);
15314 bpfptr_add(&u_core_relo, rec_size);
15319 static int check_btf_info(struct bpf_verifier_env *env,
15320 const union bpf_attr *attr,
15326 if (!attr->func_info_cnt && !attr->line_info_cnt) {
15327 if (check_abnormal_return(env))
15332 btf = btf_get_by_fd(attr->prog_btf_fd);
15334 return PTR_ERR(btf);
15335 if (btf_is_kernel(btf)) {
15339 env->prog->aux->btf = btf;
15341 err = check_btf_func(env, attr, uattr);
15345 err = check_btf_line(env, attr, uattr);
15349 err = check_core_relo(env, attr, uattr);
15356 /* check %cur's range satisfies %old's */
15357 static bool range_within(struct bpf_reg_state *old,
15358 struct bpf_reg_state *cur)
15360 return old->umin_value <= cur->umin_value &&
15361 old->umax_value >= cur->umax_value &&
15362 old->smin_value <= cur->smin_value &&
15363 old->smax_value >= cur->smax_value &&
15364 old->u32_min_value <= cur->u32_min_value &&
15365 old->u32_max_value >= cur->u32_max_value &&
15366 old->s32_min_value <= cur->s32_min_value &&
15367 old->s32_max_value >= cur->s32_max_value;
15370 /* If in the old state two registers had the same id, then they need to have
15371 * the same id in the new state as well. But that id could be different from
15372 * the old state, so we need to track the mapping from old to new ids.
15373 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
15374 * regs with old id 5 must also have new id 9 for the new state to be safe. But
15375 * regs with a different old id could still have new id 9, we don't care about
15377 * So we look through our idmap to see if this old id has been seen before. If
15378 * so, we require the new id to match; otherwise, we add the id pair to the map.
15380 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap)
15382 struct bpf_id_pair *map = idmap->map;
15385 /* either both IDs should be set or both should be zero */
15386 if (!!old_id != !!cur_id)
15389 if (old_id == 0) /* cur_id == 0 as well */
15392 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
15394 /* Reached an empty slot; haven't seen this id before */
15395 map[i].old = old_id;
15396 map[i].cur = cur_id;
15399 if (map[i].old == old_id)
15400 return map[i].cur == cur_id;
15401 if (map[i].cur == cur_id)
15404 /* We ran out of idmap slots, which should be impossible */
15409 /* Similar to check_ids(), but allocate a unique temporary ID
15410 * for 'old_id' or 'cur_id' of zero.
15411 * This makes pairs like '0 vs unique ID', 'unique ID vs 0' valid.
15413 static bool check_scalar_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap)
15415 old_id = old_id ? old_id : ++idmap->tmp_id_gen;
15416 cur_id = cur_id ? cur_id : ++idmap->tmp_id_gen;
15418 return check_ids(old_id, cur_id, idmap);
15421 static void clean_func_state(struct bpf_verifier_env *env,
15422 struct bpf_func_state *st)
15424 enum bpf_reg_liveness live;
15427 for (i = 0; i < BPF_REG_FP; i++) {
15428 live = st->regs[i].live;
15429 /* liveness must not touch this register anymore */
15430 st->regs[i].live |= REG_LIVE_DONE;
15431 if (!(live & REG_LIVE_READ))
15432 /* since the register is unused, clear its state
15433 * to make further comparison simpler
15435 __mark_reg_not_init(env, &st->regs[i]);
15438 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
15439 live = st->stack[i].spilled_ptr.live;
15440 /* liveness must not touch this stack slot anymore */
15441 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
15442 if (!(live & REG_LIVE_READ)) {
15443 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
15444 for (j = 0; j < BPF_REG_SIZE; j++)
15445 st->stack[i].slot_type[j] = STACK_INVALID;
15450 static void clean_verifier_state(struct bpf_verifier_env *env,
15451 struct bpf_verifier_state *st)
15455 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
15456 /* all regs in this state in all frames were already marked */
15459 for (i = 0; i <= st->curframe; i++)
15460 clean_func_state(env, st->frame[i]);
15463 /* the parentage chains form a tree.
15464 * the verifier states are added to state lists at given insn and
15465 * pushed into state stack for future exploration.
15466 * when the verifier reaches bpf_exit insn some of the verifer states
15467 * stored in the state lists have their final liveness state already,
15468 * but a lot of states will get revised from liveness point of view when
15469 * the verifier explores other branches.
15472 * 2: if r1 == 100 goto pc+1
15475 * when the verifier reaches exit insn the register r0 in the state list of
15476 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
15477 * of insn 2 and goes exploring further. At the insn 4 it will walk the
15478 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
15480 * Since the verifier pushes the branch states as it sees them while exploring
15481 * the program the condition of walking the branch instruction for the second
15482 * time means that all states below this branch were already explored and
15483 * their final liveness marks are already propagated.
15484 * Hence when the verifier completes the search of state list in is_state_visited()
15485 * we can call this clean_live_states() function to mark all liveness states
15486 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
15487 * will not be used.
15488 * This function also clears the registers and stack for states that !READ
15489 * to simplify state merging.
15491 * Important note here that walking the same branch instruction in the callee
15492 * doesn't meant that the states are DONE. The verifier has to compare
15495 static void clean_live_states(struct bpf_verifier_env *env, int insn,
15496 struct bpf_verifier_state *cur)
15498 struct bpf_verifier_state_list *sl;
15501 sl = *explored_state(env, insn);
15503 if (sl->state.branches)
15505 if (sl->state.insn_idx != insn ||
15506 sl->state.curframe != cur->curframe)
15508 for (i = 0; i <= cur->curframe; i++)
15509 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
15511 clean_verifier_state(env, &sl->state);
15517 static bool regs_exact(const struct bpf_reg_state *rold,
15518 const struct bpf_reg_state *rcur,
15519 struct bpf_idmap *idmap)
15521 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
15522 check_ids(rold->id, rcur->id, idmap) &&
15523 check_ids(rold->ref_obj_id, rcur->ref_obj_id, idmap);
15526 /* Returns true if (rold safe implies rcur safe) */
15527 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
15528 struct bpf_reg_state *rcur, struct bpf_idmap *idmap)
15530 if (!(rold->live & REG_LIVE_READ))
15531 /* explored state didn't use this */
15533 if (rold->type == NOT_INIT)
15534 /* explored state can't have used this */
15536 if (rcur->type == NOT_INIT)
15539 /* Enforce that register types have to match exactly, including their
15540 * modifiers (like PTR_MAYBE_NULL, MEM_RDONLY, etc), as a general
15543 * One can make a point that using a pointer register as unbounded
15544 * SCALAR would be technically acceptable, but this could lead to
15545 * pointer leaks because scalars are allowed to leak while pointers
15546 * are not. We could make this safe in special cases if root is
15547 * calling us, but it's probably not worth the hassle.
15549 * Also, register types that are *not* MAYBE_NULL could technically be
15550 * safe to use as their MAYBE_NULL variants (e.g., PTR_TO_MAP_VALUE
15551 * is safe to be used as PTR_TO_MAP_VALUE_OR_NULL, provided both point
15552 * to the same map).
15553 * However, if the old MAYBE_NULL register then got NULL checked,
15554 * doing so could have affected others with the same id, and we can't
15555 * check for that because we lost the id when we converted to
15556 * a non-MAYBE_NULL variant.
15557 * So, as a general rule we don't allow mixing MAYBE_NULL and
15558 * non-MAYBE_NULL registers as well.
15560 if (rold->type != rcur->type)
15563 switch (base_type(rold->type)) {
15565 if (env->explore_alu_limits) {
15566 /* explore_alu_limits disables tnum_in() and range_within()
15567 * logic and requires everything to be strict
15569 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
15570 check_scalar_ids(rold->id, rcur->id, idmap);
15572 if (!rold->precise)
15574 /* Why check_ids() for scalar registers?
15576 * Consider the following BPF code:
15577 * 1: r6 = ... unbound scalar, ID=a ...
15578 * 2: r7 = ... unbound scalar, ID=b ...
15579 * 3: if (r6 > r7) goto +1
15581 * 5: if (r6 > X) goto ...
15582 * 6: ... memory operation using r7 ...
15584 * First verification path is [1-6]:
15585 * - at (4) same bpf_reg_state::id (b) would be assigned to r6 and r7;
15586 * - at (5) r6 would be marked <= X, find_equal_scalars() would also mark
15587 * r7 <= X, because r6 and r7 share same id.
15588 * Next verification path is [1-4, 6].
15590 * Instruction (6) would be reached in two states:
15591 * I. r6{.id=b}, r7{.id=b} via path 1-6;
15592 * II. r6{.id=a}, r7{.id=b} via path 1-4, 6.
15594 * Use check_ids() to distinguish these states.
15596 * Also verify that new value satisfies old value range knowledge.
15598 return range_within(rold, rcur) &&
15599 tnum_in(rold->var_off, rcur->var_off) &&
15600 check_scalar_ids(rold->id, rcur->id, idmap);
15601 case PTR_TO_MAP_KEY:
15602 case PTR_TO_MAP_VALUE:
15605 case PTR_TO_TP_BUFFER:
15606 /* If the new min/max/var_off satisfy the old ones and
15607 * everything else matches, we are OK.
15609 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, var_off)) == 0 &&
15610 range_within(rold, rcur) &&
15611 tnum_in(rold->var_off, rcur->var_off) &&
15612 check_ids(rold->id, rcur->id, idmap) &&
15613 check_ids(rold->ref_obj_id, rcur->ref_obj_id, idmap);
15614 case PTR_TO_PACKET_META:
15615 case PTR_TO_PACKET:
15616 /* We must have at least as much range as the old ptr
15617 * did, so that any accesses which were safe before are
15618 * still safe. This is true even if old range < old off,
15619 * since someone could have accessed through (ptr - k), or
15620 * even done ptr -= k in a register, to get a safe access.
15622 if (rold->range > rcur->range)
15624 /* If the offsets don't match, we can't trust our alignment;
15625 * nor can we be sure that we won't fall out of range.
15627 if (rold->off != rcur->off)
15629 /* id relations must be preserved */
15630 if (!check_ids(rold->id, rcur->id, idmap))
15632 /* new val must satisfy old val knowledge */
15633 return range_within(rold, rcur) &&
15634 tnum_in(rold->var_off, rcur->var_off);
15636 /* two stack pointers are equal only if they're pointing to
15637 * the same stack frame, since fp-8 in foo != fp-8 in bar
15639 return regs_exact(rold, rcur, idmap) && rold->frameno == rcur->frameno;
15641 return regs_exact(rold, rcur, idmap);
15645 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
15646 struct bpf_func_state *cur, struct bpf_idmap *idmap)
15650 /* walk slots of the explored stack and ignore any additional
15651 * slots in the current stack, since explored(safe) state
15654 for (i = 0; i < old->allocated_stack; i++) {
15655 struct bpf_reg_state *old_reg, *cur_reg;
15657 spi = i / BPF_REG_SIZE;
15659 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
15660 i += BPF_REG_SIZE - 1;
15661 /* explored state didn't use this */
15665 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
15668 if (env->allow_uninit_stack &&
15669 old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC)
15672 /* explored stack has more populated slots than current stack
15673 * and these slots were used
15675 if (i >= cur->allocated_stack)
15678 /* if old state was safe with misc data in the stack
15679 * it will be safe with zero-initialized stack.
15680 * The opposite is not true
15682 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
15683 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
15685 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
15686 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
15687 /* Ex: old explored (safe) state has STACK_SPILL in
15688 * this stack slot, but current has STACK_MISC ->
15689 * this verifier states are not equivalent,
15690 * return false to continue verification of this path
15693 if (i % BPF_REG_SIZE != BPF_REG_SIZE - 1)
15695 /* Both old and cur are having same slot_type */
15696 switch (old->stack[spi].slot_type[BPF_REG_SIZE - 1]) {
15698 /* when explored and current stack slot are both storing
15699 * spilled registers, check that stored pointers types
15700 * are the same as well.
15701 * Ex: explored safe path could have stored
15702 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
15703 * but current path has stored:
15704 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
15705 * such verifier states are not equivalent.
15706 * return false to continue verification of this path
15708 if (!regsafe(env, &old->stack[spi].spilled_ptr,
15709 &cur->stack[spi].spilled_ptr, idmap))
15713 old_reg = &old->stack[spi].spilled_ptr;
15714 cur_reg = &cur->stack[spi].spilled_ptr;
15715 if (old_reg->dynptr.type != cur_reg->dynptr.type ||
15716 old_reg->dynptr.first_slot != cur_reg->dynptr.first_slot ||
15717 !check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap))
15721 old_reg = &old->stack[spi].spilled_ptr;
15722 cur_reg = &cur->stack[spi].spilled_ptr;
15723 /* iter.depth is not compared between states as it
15724 * doesn't matter for correctness and would otherwise
15725 * prevent convergence; we maintain it only to prevent
15726 * infinite loop check triggering, see
15727 * iter_active_depths_differ()
15729 if (old_reg->iter.btf != cur_reg->iter.btf ||
15730 old_reg->iter.btf_id != cur_reg->iter.btf_id ||
15731 old_reg->iter.state != cur_reg->iter.state ||
15732 /* ignore {old_reg,cur_reg}->iter.depth, see above */
15733 !check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap))
15738 case STACK_INVALID:
15740 /* Ensure that new unhandled slot types return false by default */
15748 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur,
15749 struct bpf_idmap *idmap)
15753 if (old->acquired_refs != cur->acquired_refs)
15756 for (i = 0; i < old->acquired_refs; i++) {
15757 if (!check_ids(old->refs[i].id, cur->refs[i].id, idmap))
15764 /* compare two verifier states
15766 * all states stored in state_list are known to be valid, since
15767 * verifier reached 'bpf_exit' instruction through them
15769 * this function is called when verifier exploring different branches of
15770 * execution popped from the state stack. If it sees an old state that has
15771 * more strict register state and more strict stack state then this execution
15772 * branch doesn't need to be explored further, since verifier already
15773 * concluded that more strict state leads to valid finish.
15775 * Therefore two states are equivalent if register state is more conservative
15776 * and explored stack state is more conservative than the current one.
15779 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
15780 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
15782 * In other words if current stack state (one being explored) has more
15783 * valid slots than old one that already passed validation, it means
15784 * the verifier can stop exploring and conclude that current state is valid too
15786 * Similarly with registers. If explored state has register type as invalid
15787 * whereas register type in current state is meaningful, it means that
15788 * the current state will reach 'bpf_exit' instruction safely
15790 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
15791 struct bpf_func_state *cur)
15795 for (i = 0; i < MAX_BPF_REG; i++)
15796 if (!regsafe(env, &old->regs[i], &cur->regs[i],
15797 &env->idmap_scratch))
15800 if (!stacksafe(env, old, cur, &env->idmap_scratch))
15803 if (!refsafe(old, cur, &env->idmap_scratch))
15809 static bool states_equal(struct bpf_verifier_env *env,
15810 struct bpf_verifier_state *old,
15811 struct bpf_verifier_state *cur)
15815 if (old->curframe != cur->curframe)
15818 env->idmap_scratch.tmp_id_gen = env->id_gen;
15819 memset(&env->idmap_scratch.map, 0, sizeof(env->idmap_scratch.map));
15821 /* Verification state from speculative execution simulation
15822 * must never prune a non-speculative execution one.
15824 if (old->speculative && !cur->speculative)
15827 if (old->active_lock.ptr != cur->active_lock.ptr)
15830 /* Old and cur active_lock's have to be either both present
15833 if (!!old->active_lock.id != !!cur->active_lock.id)
15836 if (old->active_lock.id &&
15837 !check_ids(old->active_lock.id, cur->active_lock.id, &env->idmap_scratch))
15840 if (old->active_rcu_lock != cur->active_rcu_lock)
15843 /* for states to be equal callsites have to be the same
15844 * and all frame states need to be equivalent
15846 for (i = 0; i <= old->curframe; i++) {
15847 if (old->frame[i]->callsite != cur->frame[i]->callsite)
15849 if (!func_states_equal(env, old->frame[i], cur->frame[i]))
15855 /* Return 0 if no propagation happened. Return negative error code if error
15856 * happened. Otherwise, return the propagated bit.
15858 static int propagate_liveness_reg(struct bpf_verifier_env *env,
15859 struct bpf_reg_state *reg,
15860 struct bpf_reg_state *parent_reg)
15862 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
15863 u8 flag = reg->live & REG_LIVE_READ;
15866 /* When comes here, read flags of PARENT_REG or REG could be any of
15867 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
15868 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
15870 if (parent_flag == REG_LIVE_READ64 ||
15871 /* Or if there is no read flag from REG. */
15873 /* Or if the read flag from REG is the same as PARENT_REG. */
15874 parent_flag == flag)
15877 err = mark_reg_read(env, reg, parent_reg, flag);
15884 /* A write screens off any subsequent reads; but write marks come from the
15885 * straight-line code between a state and its parent. When we arrive at an
15886 * equivalent state (jump target or such) we didn't arrive by the straight-line
15887 * code, so read marks in the state must propagate to the parent regardless
15888 * of the state's write marks. That's what 'parent == state->parent' comparison
15889 * in mark_reg_read() is for.
15891 static int propagate_liveness(struct bpf_verifier_env *env,
15892 const struct bpf_verifier_state *vstate,
15893 struct bpf_verifier_state *vparent)
15895 struct bpf_reg_state *state_reg, *parent_reg;
15896 struct bpf_func_state *state, *parent;
15897 int i, frame, err = 0;
15899 if (vparent->curframe != vstate->curframe) {
15900 WARN(1, "propagate_live: parent frame %d current frame %d\n",
15901 vparent->curframe, vstate->curframe);
15904 /* Propagate read liveness of registers... */
15905 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
15906 for (frame = 0; frame <= vstate->curframe; frame++) {
15907 parent = vparent->frame[frame];
15908 state = vstate->frame[frame];
15909 parent_reg = parent->regs;
15910 state_reg = state->regs;
15911 /* We don't need to worry about FP liveness, it's read-only */
15912 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
15913 err = propagate_liveness_reg(env, &state_reg[i],
15917 if (err == REG_LIVE_READ64)
15918 mark_insn_zext(env, &parent_reg[i]);
15921 /* Propagate stack slots. */
15922 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
15923 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
15924 parent_reg = &parent->stack[i].spilled_ptr;
15925 state_reg = &state->stack[i].spilled_ptr;
15926 err = propagate_liveness_reg(env, state_reg,
15935 /* find precise scalars in the previous equivalent state and
15936 * propagate them into the current state
15938 static int propagate_precision(struct bpf_verifier_env *env,
15939 const struct bpf_verifier_state *old)
15941 struct bpf_reg_state *state_reg;
15942 struct bpf_func_state *state;
15943 int i, err = 0, fr;
15946 for (fr = old->curframe; fr >= 0; fr--) {
15947 state = old->frame[fr];
15948 state_reg = state->regs;
15950 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
15951 if (state_reg->type != SCALAR_VALUE ||
15952 !state_reg->precise ||
15953 !(state_reg->live & REG_LIVE_READ))
15955 if (env->log.level & BPF_LOG_LEVEL2) {
15957 verbose(env, "frame %d: propagating r%d", fr, i);
15959 verbose(env, ",r%d", i);
15961 bt_set_frame_reg(&env->bt, fr, i);
15965 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
15966 if (!is_spilled_reg(&state->stack[i]))
15968 state_reg = &state->stack[i].spilled_ptr;
15969 if (state_reg->type != SCALAR_VALUE ||
15970 !state_reg->precise ||
15971 !(state_reg->live & REG_LIVE_READ))
15973 if (env->log.level & BPF_LOG_LEVEL2) {
15975 verbose(env, "frame %d: propagating fp%d",
15976 fr, (-i - 1) * BPF_REG_SIZE);
15978 verbose(env, ",fp%d", (-i - 1) * BPF_REG_SIZE);
15980 bt_set_frame_slot(&env->bt, fr, i);
15984 verbose(env, "\n");
15987 err = mark_chain_precision_batch(env);
15994 static bool states_maybe_looping(struct bpf_verifier_state *old,
15995 struct bpf_verifier_state *cur)
15997 struct bpf_func_state *fold, *fcur;
15998 int i, fr = cur->curframe;
16000 if (old->curframe != fr)
16003 fold = old->frame[fr];
16004 fcur = cur->frame[fr];
16005 for (i = 0; i < MAX_BPF_REG; i++)
16006 if (memcmp(&fold->regs[i], &fcur->regs[i],
16007 offsetof(struct bpf_reg_state, parent)))
16012 static bool is_iter_next_insn(struct bpf_verifier_env *env, int insn_idx)
16014 return env->insn_aux_data[insn_idx].is_iter_next;
16017 /* is_state_visited() handles iter_next() (see process_iter_next_call() for
16018 * terminology) calls specially: as opposed to bounded BPF loops, it *expects*
16019 * states to match, which otherwise would look like an infinite loop. So while
16020 * iter_next() calls are taken care of, we still need to be careful and
16021 * prevent erroneous and too eager declaration of "ininite loop", when
16022 * iterators are involved.
16024 * Here's a situation in pseudo-BPF assembly form:
16026 * 0: again: ; set up iter_next() call args
16027 * 1: r1 = &it ; <CHECKPOINT HERE>
16028 * 2: call bpf_iter_num_next ; this is iter_next() call
16029 * 3: if r0 == 0 goto done
16030 * 4: ... something useful here ...
16031 * 5: goto again ; another iteration
16034 * 8: call bpf_iter_num_destroy ; clean up iter state
16037 * This is a typical loop. Let's assume that we have a prune point at 1:,
16038 * before we get to `call bpf_iter_num_next` (e.g., because of that `goto
16039 * again`, assuming other heuristics don't get in a way).
16041 * When we first time come to 1:, let's say we have some state X. We proceed
16042 * to 2:, fork states, enqueue ACTIVE, validate NULL case successfully, exit.
16043 * Now we come back to validate that forked ACTIVE state. We proceed through
16044 * 3-5, come to goto, jump to 1:. Let's assume our state didn't change, so we
16045 * are converging. But the problem is that we don't know that yet, as this
16046 * convergence has to happen at iter_next() call site only. So if nothing is
16047 * done, at 1: verifier will use bounded loop logic and declare infinite
16048 * looping (and would be *technically* correct, if not for iterator's
16049 * "eventual sticky NULL" contract, see process_iter_next_call()). But we
16050 * don't want that. So what we do in process_iter_next_call() when we go on
16051 * another ACTIVE iteration, we bump slot->iter.depth, to mark that it's
16052 * a different iteration. So when we suspect an infinite loop, we additionally
16053 * check if any of the *ACTIVE* iterator states depths differ. If yes, we
16054 * pretend we are not looping and wait for next iter_next() call.
16056 * This only applies to ACTIVE state. In DRAINED state we don't expect to
16057 * loop, because that would actually mean infinite loop, as DRAINED state is
16058 * "sticky", and so we'll keep returning into the same instruction with the
16059 * same state (at least in one of possible code paths).
16061 * This approach allows to keep infinite loop heuristic even in the face of
16062 * active iterator. E.g., C snippet below is and will be detected as
16063 * inifintely looping:
16065 * struct bpf_iter_num it;
16068 * bpf_iter_num_new(&it, 0, 10);
16069 * while ((p = bpf_iter_num_next(&t))) {
16071 * while (x--) {} // <<-- infinite loop here
16075 static bool iter_active_depths_differ(struct bpf_verifier_state *old, struct bpf_verifier_state *cur)
16077 struct bpf_reg_state *slot, *cur_slot;
16078 struct bpf_func_state *state;
16081 for (fr = old->curframe; fr >= 0; fr--) {
16082 state = old->frame[fr];
16083 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
16084 if (state->stack[i].slot_type[0] != STACK_ITER)
16087 slot = &state->stack[i].spilled_ptr;
16088 if (slot->iter.state != BPF_ITER_STATE_ACTIVE)
16091 cur_slot = &cur->frame[fr]->stack[i].spilled_ptr;
16092 if (cur_slot->iter.depth != slot->iter.depth)
16099 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
16101 struct bpf_verifier_state_list *new_sl;
16102 struct bpf_verifier_state_list *sl, **pprev;
16103 struct bpf_verifier_state *cur = env->cur_state, *new;
16104 int i, j, err, states_cnt = 0;
16105 bool force_new_state = env->test_state_freq || is_force_checkpoint(env, insn_idx);
16106 bool add_new_state = force_new_state;
16108 /* bpf progs typically have pruning point every 4 instructions
16109 * http://vger.kernel.org/bpfconf2019.html#session-1
16110 * Do not add new state for future pruning if the verifier hasn't seen
16111 * at least 2 jumps and at least 8 instructions.
16112 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
16113 * In tests that amounts to up to 50% reduction into total verifier
16114 * memory consumption and 20% verifier time speedup.
16116 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
16117 env->insn_processed - env->prev_insn_processed >= 8)
16118 add_new_state = true;
16120 pprev = explored_state(env, insn_idx);
16123 clean_live_states(env, insn_idx, cur);
16127 if (sl->state.insn_idx != insn_idx)
16130 if (sl->state.branches) {
16131 struct bpf_func_state *frame = sl->state.frame[sl->state.curframe];
16133 if (frame->in_async_callback_fn &&
16134 frame->async_entry_cnt != cur->frame[cur->curframe]->async_entry_cnt) {
16135 /* Different async_entry_cnt means that the verifier is
16136 * processing another entry into async callback.
16137 * Seeing the same state is not an indication of infinite
16138 * loop or infinite recursion.
16139 * But finding the same state doesn't mean that it's safe
16140 * to stop processing the current state. The previous state
16141 * hasn't yet reached bpf_exit, since state.branches > 0.
16142 * Checking in_async_callback_fn alone is not enough either.
16143 * Since the verifier still needs to catch infinite loops
16144 * inside async callbacks.
16146 goto skip_inf_loop_check;
16148 /* BPF open-coded iterators loop detection is special.
16149 * states_maybe_looping() logic is too simplistic in detecting
16150 * states that *might* be equivalent, because it doesn't know
16151 * about ID remapping, so don't even perform it.
16152 * See process_iter_next_call() and iter_active_depths_differ()
16153 * for overview of the logic. When current and one of parent
16154 * states are detected as equivalent, it's a good thing: we prove
16155 * convergence and can stop simulating further iterations.
16156 * It's safe to assume that iterator loop will finish, taking into
16157 * account iter_next() contract of eventually returning
16158 * sticky NULL result.
16160 if (is_iter_next_insn(env, insn_idx)) {
16161 if (states_equal(env, &sl->state, cur)) {
16162 struct bpf_func_state *cur_frame;
16163 struct bpf_reg_state *iter_state, *iter_reg;
16166 cur_frame = cur->frame[cur->curframe];
16167 /* btf_check_iter_kfuncs() enforces that
16168 * iter state pointer is always the first arg
16170 iter_reg = &cur_frame->regs[BPF_REG_1];
16171 /* current state is valid due to states_equal(),
16172 * so we can assume valid iter and reg state,
16173 * no need for extra (re-)validations
16175 spi = __get_spi(iter_reg->off + iter_reg->var_off.value);
16176 iter_state = &func(env, iter_reg)->stack[spi].spilled_ptr;
16177 if (iter_state->iter.state == BPF_ITER_STATE_ACTIVE)
16180 goto skip_inf_loop_check;
16182 /* attempt to detect infinite loop to avoid unnecessary doomed work */
16183 if (states_maybe_looping(&sl->state, cur) &&
16184 states_equal(env, &sl->state, cur) &&
16185 !iter_active_depths_differ(&sl->state, cur)) {
16186 verbose_linfo(env, insn_idx, "; ");
16187 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
16190 /* if the verifier is processing a loop, avoid adding new state
16191 * too often, since different loop iterations have distinct
16192 * states and may not help future pruning.
16193 * This threshold shouldn't be too low to make sure that
16194 * a loop with large bound will be rejected quickly.
16195 * The most abusive loop will be:
16197 * if r1 < 1000000 goto pc-2
16198 * 1M insn_procssed limit / 100 == 10k peak states.
16199 * This threshold shouldn't be too high either, since states
16200 * at the end of the loop are likely to be useful in pruning.
16202 skip_inf_loop_check:
16203 if (!force_new_state &&
16204 env->jmps_processed - env->prev_jmps_processed < 20 &&
16205 env->insn_processed - env->prev_insn_processed < 100)
16206 add_new_state = false;
16209 if (states_equal(env, &sl->state, cur)) {
16212 /* reached equivalent register/stack state,
16213 * prune the search.
16214 * Registers read by the continuation are read by us.
16215 * If we have any write marks in env->cur_state, they
16216 * will prevent corresponding reads in the continuation
16217 * from reaching our parent (an explored_state). Our
16218 * own state will get the read marks recorded, but
16219 * they'll be immediately forgotten as we're pruning
16220 * this state and will pop a new one.
16222 err = propagate_liveness(env, &sl->state, cur);
16224 /* if previous state reached the exit with precision and
16225 * current state is equivalent to it (except precsion marks)
16226 * the precision needs to be propagated back in
16227 * the current state.
16229 err = err ? : push_jmp_history(env, cur);
16230 err = err ? : propagate_precision(env, &sl->state);
16236 /* when new state is not going to be added do not increase miss count.
16237 * Otherwise several loop iterations will remove the state
16238 * recorded earlier. The goal of these heuristics is to have
16239 * states from some iterations of the loop (some in the beginning
16240 * and some at the end) to help pruning.
16244 /* heuristic to determine whether this state is beneficial
16245 * to keep checking from state equivalence point of view.
16246 * Higher numbers increase max_states_per_insn and verification time,
16247 * but do not meaningfully decrease insn_processed.
16249 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
16250 /* the state is unlikely to be useful. Remove it to
16251 * speed up verification
16254 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
16255 u32 br = sl->state.branches;
16258 "BUG live_done but branches_to_explore %d\n",
16260 free_verifier_state(&sl->state, false);
16262 env->peak_states--;
16264 /* cannot free this state, since parentage chain may
16265 * walk it later. Add it for free_list instead to
16266 * be freed at the end of verification
16268 sl->next = env->free_list;
16269 env->free_list = sl;
16279 if (env->max_states_per_insn < states_cnt)
16280 env->max_states_per_insn = states_cnt;
16282 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
16285 if (!add_new_state)
16288 /* There were no equivalent states, remember the current one.
16289 * Technically the current state is not proven to be safe yet,
16290 * but it will either reach outer most bpf_exit (which means it's safe)
16291 * or it will be rejected. When there are no loops the verifier won't be
16292 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
16293 * again on the way to bpf_exit.
16294 * When looping the sl->state.branches will be > 0 and this state
16295 * will not be considered for equivalence until branches == 0.
16297 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
16300 env->total_states++;
16301 env->peak_states++;
16302 env->prev_jmps_processed = env->jmps_processed;
16303 env->prev_insn_processed = env->insn_processed;
16305 /* forget precise markings we inherited, see __mark_chain_precision */
16306 if (env->bpf_capable)
16307 mark_all_scalars_imprecise(env, cur);
16309 /* add new state to the head of linked list */
16310 new = &new_sl->state;
16311 err = copy_verifier_state(new, cur);
16313 free_verifier_state(new, false);
16317 new->insn_idx = insn_idx;
16318 WARN_ONCE(new->branches != 1,
16319 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
16322 cur->first_insn_idx = insn_idx;
16323 clear_jmp_history(cur);
16324 new_sl->next = *explored_state(env, insn_idx);
16325 *explored_state(env, insn_idx) = new_sl;
16326 /* connect new state to parentage chain. Current frame needs all
16327 * registers connected. Only r6 - r9 of the callers are alive (pushed
16328 * to the stack implicitly by JITs) so in callers' frames connect just
16329 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
16330 * the state of the call instruction (with WRITTEN set), and r0 comes
16331 * from callee with its full parentage chain, anyway.
16333 /* clear write marks in current state: the writes we did are not writes
16334 * our child did, so they don't screen off its reads from us.
16335 * (There are no read marks in current state, because reads always mark
16336 * their parent and current state never has children yet. Only
16337 * explored_states can get read marks.)
16339 for (j = 0; j <= cur->curframe; j++) {
16340 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
16341 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
16342 for (i = 0; i < BPF_REG_FP; i++)
16343 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
16346 /* all stack frames are accessible from callee, clear them all */
16347 for (j = 0; j <= cur->curframe; j++) {
16348 struct bpf_func_state *frame = cur->frame[j];
16349 struct bpf_func_state *newframe = new->frame[j];
16351 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
16352 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
16353 frame->stack[i].spilled_ptr.parent =
16354 &newframe->stack[i].spilled_ptr;
16360 /* Return true if it's OK to have the same insn return a different type. */
16361 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
16363 switch (base_type(type)) {
16365 case PTR_TO_SOCKET:
16366 case PTR_TO_SOCK_COMMON:
16367 case PTR_TO_TCP_SOCK:
16368 case PTR_TO_XDP_SOCK:
16369 case PTR_TO_BTF_ID:
16376 /* If an instruction was previously used with particular pointer types, then we
16377 * need to be careful to avoid cases such as the below, where it may be ok
16378 * for one branch accessing the pointer, but not ok for the other branch:
16383 * R1 = some_other_valid_ptr;
16386 * R2 = *(u32 *)(R1 + 0);
16388 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
16390 return src != prev && (!reg_type_mismatch_ok(src) ||
16391 !reg_type_mismatch_ok(prev));
16394 static int save_aux_ptr_type(struct bpf_verifier_env *env, enum bpf_reg_type type,
16395 bool allow_trust_missmatch)
16397 enum bpf_reg_type *prev_type = &env->insn_aux_data[env->insn_idx].ptr_type;
16399 if (*prev_type == NOT_INIT) {
16400 /* Saw a valid insn
16401 * dst_reg = *(u32 *)(src_reg + off)
16402 * save type to validate intersecting paths
16405 } else if (reg_type_mismatch(type, *prev_type)) {
16406 /* Abuser program is trying to use the same insn
16407 * dst_reg = *(u32*) (src_reg + off)
16408 * with different pointer types:
16409 * src_reg == ctx in one branch and
16410 * src_reg == stack|map in some other branch.
16413 if (allow_trust_missmatch &&
16414 base_type(type) == PTR_TO_BTF_ID &&
16415 base_type(*prev_type) == PTR_TO_BTF_ID) {
16417 * Have to support a use case when one path through
16418 * the program yields TRUSTED pointer while another
16419 * is UNTRUSTED. Fallback to UNTRUSTED to generate
16420 * BPF_PROBE_MEM/BPF_PROBE_MEMSX.
16422 *prev_type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
16424 verbose(env, "same insn cannot be used with different pointers\n");
16432 static int do_check(struct bpf_verifier_env *env)
16434 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
16435 struct bpf_verifier_state *state = env->cur_state;
16436 struct bpf_insn *insns = env->prog->insnsi;
16437 struct bpf_reg_state *regs;
16438 int insn_cnt = env->prog->len;
16439 bool do_print_state = false;
16440 int prev_insn_idx = -1;
16443 struct bpf_insn *insn;
16447 env->prev_insn_idx = prev_insn_idx;
16448 if (env->insn_idx >= insn_cnt) {
16449 verbose(env, "invalid insn idx %d insn_cnt %d\n",
16450 env->insn_idx, insn_cnt);
16454 insn = &insns[env->insn_idx];
16455 class = BPF_CLASS(insn->code);
16457 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
16459 "BPF program is too large. Processed %d insn\n",
16460 env->insn_processed);
16464 state->last_insn_idx = env->prev_insn_idx;
16466 if (is_prune_point(env, env->insn_idx)) {
16467 err = is_state_visited(env, env->insn_idx);
16471 /* found equivalent state, can prune the search */
16472 if (env->log.level & BPF_LOG_LEVEL) {
16473 if (do_print_state)
16474 verbose(env, "\nfrom %d to %d%s: safe\n",
16475 env->prev_insn_idx, env->insn_idx,
16476 env->cur_state->speculative ?
16477 " (speculative execution)" : "");
16479 verbose(env, "%d: safe\n", env->insn_idx);
16481 goto process_bpf_exit;
16485 if (is_jmp_point(env, env->insn_idx)) {
16486 err = push_jmp_history(env, state);
16491 if (signal_pending(current))
16494 if (need_resched())
16497 if (env->log.level & BPF_LOG_LEVEL2 && do_print_state) {
16498 verbose(env, "\nfrom %d to %d%s:",
16499 env->prev_insn_idx, env->insn_idx,
16500 env->cur_state->speculative ?
16501 " (speculative execution)" : "");
16502 print_verifier_state(env, state->frame[state->curframe], true);
16503 do_print_state = false;
16506 if (env->log.level & BPF_LOG_LEVEL) {
16507 const struct bpf_insn_cbs cbs = {
16508 .cb_call = disasm_kfunc_name,
16509 .cb_print = verbose,
16510 .private_data = env,
16513 if (verifier_state_scratched(env))
16514 print_insn_state(env, state->frame[state->curframe]);
16516 verbose_linfo(env, env->insn_idx, "; ");
16517 env->prev_log_pos = env->log.end_pos;
16518 verbose(env, "%d: ", env->insn_idx);
16519 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
16520 env->prev_insn_print_pos = env->log.end_pos - env->prev_log_pos;
16521 env->prev_log_pos = env->log.end_pos;
16524 if (bpf_prog_is_offloaded(env->prog->aux)) {
16525 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
16526 env->prev_insn_idx);
16531 regs = cur_regs(env);
16532 sanitize_mark_insn_seen(env);
16533 prev_insn_idx = env->insn_idx;
16535 if (class == BPF_ALU || class == BPF_ALU64) {
16536 err = check_alu_op(env, insn);
16540 } else if (class == BPF_LDX) {
16541 enum bpf_reg_type src_reg_type;
16543 /* check for reserved fields is already done */
16545 /* check src operand */
16546 err = check_reg_arg(env, insn->src_reg, SRC_OP);
16550 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
16554 src_reg_type = regs[insn->src_reg].type;
16556 /* check that memory (src_reg + off) is readable,
16557 * the state of dst_reg will be updated by this func
16559 err = check_mem_access(env, env->insn_idx, insn->src_reg,
16560 insn->off, BPF_SIZE(insn->code),
16561 BPF_READ, insn->dst_reg, false,
16562 BPF_MODE(insn->code) == BPF_MEMSX);
16566 err = save_aux_ptr_type(env, src_reg_type, true);
16569 } else if (class == BPF_STX) {
16570 enum bpf_reg_type dst_reg_type;
16572 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
16573 err = check_atomic(env, env->insn_idx, insn);
16580 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
16581 verbose(env, "BPF_STX uses reserved fields\n");
16585 /* check src1 operand */
16586 err = check_reg_arg(env, insn->src_reg, SRC_OP);
16589 /* check src2 operand */
16590 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
16594 dst_reg_type = regs[insn->dst_reg].type;
16596 /* check that memory (dst_reg + off) is writeable */
16597 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
16598 insn->off, BPF_SIZE(insn->code),
16599 BPF_WRITE, insn->src_reg, false, false);
16603 err = save_aux_ptr_type(env, dst_reg_type, false);
16606 } else if (class == BPF_ST) {
16607 enum bpf_reg_type dst_reg_type;
16609 if (BPF_MODE(insn->code) != BPF_MEM ||
16610 insn->src_reg != BPF_REG_0) {
16611 verbose(env, "BPF_ST uses reserved fields\n");
16614 /* check src 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, -1, false, false);
16628 err = save_aux_ptr_type(env, dst_reg_type, false);
16631 } else if (class == BPF_JMP || class == BPF_JMP32) {
16632 u8 opcode = BPF_OP(insn->code);
16634 env->jmps_processed++;
16635 if (opcode == BPF_CALL) {
16636 if (BPF_SRC(insn->code) != BPF_K ||
16637 (insn->src_reg != BPF_PSEUDO_KFUNC_CALL
16638 && insn->off != 0) ||
16639 (insn->src_reg != BPF_REG_0 &&
16640 insn->src_reg != BPF_PSEUDO_CALL &&
16641 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
16642 insn->dst_reg != BPF_REG_0 ||
16643 class == BPF_JMP32) {
16644 verbose(env, "BPF_CALL uses reserved fields\n");
16648 if (env->cur_state->active_lock.ptr) {
16649 if ((insn->src_reg == BPF_REG_0 && insn->imm != BPF_FUNC_spin_unlock) ||
16650 (insn->src_reg == BPF_PSEUDO_CALL) ||
16651 (insn->src_reg == BPF_PSEUDO_KFUNC_CALL &&
16652 (insn->off != 0 || !is_bpf_graph_api_kfunc(insn->imm)))) {
16653 verbose(env, "function calls are not allowed while holding a lock\n");
16657 if (insn->src_reg == BPF_PSEUDO_CALL)
16658 err = check_func_call(env, insn, &env->insn_idx);
16659 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
16660 err = check_kfunc_call(env, insn, &env->insn_idx);
16662 err = check_helper_call(env, insn, &env->insn_idx);
16666 mark_reg_scratched(env, BPF_REG_0);
16667 } else if (opcode == BPF_JA) {
16668 if (BPF_SRC(insn->code) != BPF_K ||
16669 insn->src_reg != BPF_REG_0 ||
16670 insn->dst_reg != BPF_REG_0 ||
16671 (class == BPF_JMP && insn->imm != 0) ||
16672 (class == BPF_JMP32 && insn->off != 0)) {
16673 verbose(env, "BPF_JA uses reserved fields\n");
16677 if (class == BPF_JMP)
16678 env->insn_idx += insn->off + 1;
16680 env->insn_idx += insn->imm + 1;
16683 } else if (opcode == BPF_EXIT) {
16684 if (BPF_SRC(insn->code) != BPF_K ||
16686 insn->src_reg != BPF_REG_0 ||
16687 insn->dst_reg != BPF_REG_0 ||
16688 class == BPF_JMP32) {
16689 verbose(env, "BPF_EXIT uses reserved fields\n");
16693 if (env->cur_state->active_lock.ptr &&
16694 !in_rbtree_lock_required_cb(env)) {
16695 verbose(env, "bpf_spin_unlock is missing\n");
16699 if (env->cur_state->active_rcu_lock &&
16700 !in_rbtree_lock_required_cb(env)) {
16701 verbose(env, "bpf_rcu_read_unlock is missing\n");
16705 /* We must do check_reference_leak here before
16706 * prepare_func_exit to handle the case when
16707 * state->curframe > 0, it may be a callback
16708 * function, for which reference_state must
16709 * match caller reference state when it exits.
16711 err = check_reference_leak(env);
16715 if (state->curframe) {
16716 /* exit from nested function */
16717 err = prepare_func_exit(env, &env->insn_idx);
16720 do_print_state = true;
16724 err = check_return_code(env);
16728 mark_verifier_state_scratched(env);
16729 update_branch_counts(env, env->cur_state);
16730 err = pop_stack(env, &prev_insn_idx,
16731 &env->insn_idx, pop_log);
16733 if (err != -ENOENT)
16737 do_print_state = true;
16741 err = check_cond_jmp_op(env, insn, &env->insn_idx);
16745 } else if (class == BPF_LD) {
16746 u8 mode = BPF_MODE(insn->code);
16748 if (mode == BPF_ABS || mode == BPF_IND) {
16749 err = check_ld_abs(env, insn);
16753 } else if (mode == BPF_IMM) {
16754 err = check_ld_imm(env, insn);
16759 sanitize_mark_insn_seen(env);
16761 verbose(env, "invalid BPF_LD mode\n");
16765 verbose(env, "unknown insn class %d\n", class);
16775 static int find_btf_percpu_datasec(struct btf *btf)
16777 const struct btf_type *t;
16782 * Both vmlinux and module each have their own ".data..percpu"
16783 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
16784 * types to look at only module's own BTF types.
16786 n = btf_nr_types(btf);
16787 if (btf_is_module(btf))
16788 i = btf_nr_types(btf_vmlinux);
16792 for(; i < n; i++) {
16793 t = btf_type_by_id(btf, i);
16794 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
16797 tname = btf_name_by_offset(btf, t->name_off);
16798 if (!strcmp(tname, ".data..percpu"))
16805 /* replace pseudo btf_id with kernel symbol address */
16806 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
16807 struct bpf_insn *insn,
16808 struct bpf_insn_aux_data *aux)
16810 const struct btf_var_secinfo *vsi;
16811 const struct btf_type *datasec;
16812 struct btf_mod_pair *btf_mod;
16813 const struct btf_type *t;
16814 const char *sym_name;
16815 bool percpu = false;
16816 u32 type, id = insn->imm;
16820 int i, btf_fd, err;
16822 btf_fd = insn[1].imm;
16824 btf = btf_get_by_fd(btf_fd);
16826 verbose(env, "invalid module BTF object FD specified.\n");
16830 if (!btf_vmlinux) {
16831 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
16838 t = btf_type_by_id(btf, id);
16840 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
16845 if (!btf_type_is_var(t) && !btf_type_is_func(t)) {
16846 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR or KIND_FUNC\n", id);
16851 sym_name = btf_name_by_offset(btf, t->name_off);
16852 addr = kallsyms_lookup_name(sym_name);
16854 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
16859 insn[0].imm = (u32)addr;
16860 insn[1].imm = addr >> 32;
16862 if (btf_type_is_func(t)) {
16863 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
16864 aux->btf_var.mem_size = 0;
16868 datasec_id = find_btf_percpu_datasec(btf);
16869 if (datasec_id > 0) {
16870 datasec = btf_type_by_id(btf, datasec_id);
16871 for_each_vsi(i, datasec, vsi) {
16872 if (vsi->type == id) {
16880 t = btf_type_skip_modifiers(btf, type, NULL);
16882 aux->btf_var.reg_type = PTR_TO_BTF_ID | MEM_PERCPU;
16883 aux->btf_var.btf = btf;
16884 aux->btf_var.btf_id = type;
16885 } else if (!btf_type_is_struct(t)) {
16886 const struct btf_type *ret;
16890 /* resolve the type size of ksym. */
16891 ret = btf_resolve_size(btf, t, &tsize);
16893 tname = btf_name_by_offset(btf, t->name_off);
16894 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
16895 tname, PTR_ERR(ret));
16899 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
16900 aux->btf_var.mem_size = tsize;
16902 aux->btf_var.reg_type = PTR_TO_BTF_ID;
16903 aux->btf_var.btf = btf;
16904 aux->btf_var.btf_id = type;
16907 /* check whether we recorded this BTF (and maybe module) already */
16908 for (i = 0; i < env->used_btf_cnt; i++) {
16909 if (env->used_btfs[i].btf == btf) {
16915 if (env->used_btf_cnt >= MAX_USED_BTFS) {
16920 btf_mod = &env->used_btfs[env->used_btf_cnt];
16921 btf_mod->btf = btf;
16922 btf_mod->module = NULL;
16924 /* if we reference variables from kernel module, bump its refcount */
16925 if (btf_is_module(btf)) {
16926 btf_mod->module = btf_try_get_module(btf);
16927 if (!btf_mod->module) {
16933 env->used_btf_cnt++;
16941 static bool is_tracing_prog_type(enum bpf_prog_type type)
16944 case BPF_PROG_TYPE_KPROBE:
16945 case BPF_PROG_TYPE_TRACEPOINT:
16946 case BPF_PROG_TYPE_PERF_EVENT:
16947 case BPF_PROG_TYPE_RAW_TRACEPOINT:
16948 case BPF_PROG_TYPE_RAW_TRACEPOINT_WRITABLE:
16955 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
16956 struct bpf_map *map,
16957 struct bpf_prog *prog)
16960 enum bpf_prog_type prog_type = resolve_prog_type(prog);
16962 if (btf_record_has_field(map->record, BPF_LIST_HEAD) ||
16963 btf_record_has_field(map->record, BPF_RB_ROOT)) {
16964 if (is_tracing_prog_type(prog_type)) {
16965 verbose(env, "tracing progs cannot use bpf_{list_head,rb_root} yet\n");
16970 if (btf_record_has_field(map->record, BPF_SPIN_LOCK)) {
16971 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
16972 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
16976 if (is_tracing_prog_type(prog_type)) {
16977 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
16982 if (btf_record_has_field(map->record, BPF_TIMER)) {
16983 if (is_tracing_prog_type(prog_type)) {
16984 verbose(env, "tracing progs cannot use bpf_timer yet\n");
16989 if ((bpf_prog_is_offloaded(prog->aux) || bpf_map_is_offloaded(map)) &&
16990 !bpf_offload_prog_map_match(prog, map)) {
16991 verbose(env, "offload device mismatch between prog and map\n");
16995 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
16996 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
17000 if (prog->aux->sleepable)
17001 switch (map->map_type) {
17002 case BPF_MAP_TYPE_HASH:
17003 case BPF_MAP_TYPE_LRU_HASH:
17004 case BPF_MAP_TYPE_ARRAY:
17005 case BPF_MAP_TYPE_PERCPU_HASH:
17006 case BPF_MAP_TYPE_PERCPU_ARRAY:
17007 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
17008 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
17009 case BPF_MAP_TYPE_HASH_OF_MAPS:
17010 case BPF_MAP_TYPE_RINGBUF:
17011 case BPF_MAP_TYPE_USER_RINGBUF:
17012 case BPF_MAP_TYPE_INODE_STORAGE:
17013 case BPF_MAP_TYPE_SK_STORAGE:
17014 case BPF_MAP_TYPE_TASK_STORAGE:
17015 case BPF_MAP_TYPE_CGRP_STORAGE:
17019 "Sleepable programs can only use array, hash, ringbuf and local storage maps\n");
17026 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
17028 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
17029 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
17032 /* find and rewrite pseudo imm in ld_imm64 instructions:
17034 * 1. if it accesses map FD, replace it with actual map pointer.
17035 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
17037 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
17039 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
17041 struct bpf_insn *insn = env->prog->insnsi;
17042 int insn_cnt = env->prog->len;
17045 err = bpf_prog_calc_tag(env->prog);
17049 for (i = 0; i < insn_cnt; i++, insn++) {
17050 if (BPF_CLASS(insn->code) == BPF_LDX &&
17051 ((BPF_MODE(insn->code) != BPF_MEM && BPF_MODE(insn->code) != BPF_MEMSX) ||
17053 verbose(env, "BPF_LDX uses reserved fields\n");
17057 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
17058 struct bpf_insn_aux_data *aux;
17059 struct bpf_map *map;
17064 if (i == insn_cnt - 1 || insn[1].code != 0 ||
17065 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
17066 insn[1].off != 0) {
17067 verbose(env, "invalid bpf_ld_imm64 insn\n");
17071 if (insn[0].src_reg == 0)
17072 /* valid generic load 64-bit imm */
17075 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
17076 aux = &env->insn_aux_data[i];
17077 err = check_pseudo_btf_id(env, insn, aux);
17083 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
17084 aux = &env->insn_aux_data[i];
17085 aux->ptr_type = PTR_TO_FUNC;
17089 /* In final convert_pseudo_ld_imm64() step, this is
17090 * converted into regular 64-bit imm load insn.
17092 switch (insn[0].src_reg) {
17093 case BPF_PSEUDO_MAP_VALUE:
17094 case BPF_PSEUDO_MAP_IDX_VALUE:
17096 case BPF_PSEUDO_MAP_FD:
17097 case BPF_PSEUDO_MAP_IDX:
17098 if (insn[1].imm == 0)
17102 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
17106 switch (insn[0].src_reg) {
17107 case BPF_PSEUDO_MAP_IDX_VALUE:
17108 case BPF_PSEUDO_MAP_IDX:
17109 if (bpfptr_is_null(env->fd_array)) {
17110 verbose(env, "fd_idx without fd_array is invalid\n");
17113 if (copy_from_bpfptr_offset(&fd, env->fd_array,
17114 insn[0].imm * sizeof(fd),
17124 map = __bpf_map_get(f);
17126 verbose(env, "fd %d is not pointing to valid bpf_map\n",
17128 return PTR_ERR(map);
17131 err = check_map_prog_compatibility(env, map, env->prog);
17137 aux = &env->insn_aux_data[i];
17138 if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
17139 insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
17140 addr = (unsigned long)map;
17142 u32 off = insn[1].imm;
17144 if (off >= BPF_MAX_VAR_OFF) {
17145 verbose(env, "direct value offset of %u is not allowed\n", off);
17150 if (!map->ops->map_direct_value_addr) {
17151 verbose(env, "no direct value access support for this map type\n");
17156 err = map->ops->map_direct_value_addr(map, &addr, off);
17158 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
17159 map->value_size, off);
17164 aux->map_off = off;
17168 insn[0].imm = (u32)addr;
17169 insn[1].imm = addr >> 32;
17171 /* check whether we recorded this map already */
17172 for (j = 0; j < env->used_map_cnt; j++) {
17173 if (env->used_maps[j] == map) {
17174 aux->map_index = j;
17180 if (env->used_map_cnt >= MAX_USED_MAPS) {
17185 /* hold the map. If the program is rejected by verifier,
17186 * the map will be released by release_maps() or it
17187 * will be used by the valid program until it's unloaded
17188 * and all maps are released in free_used_maps()
17192 aux->map_index = env->used_map_cnt;
17193 env->used_maps[env->used_map_cnt++] = map;
17195 if (bpf_map_is_cgroup_storage(map) &&
17196 bpf_cgroup_storage_assign(env->prog->aux, map)) {
17197 verbose(env, "only one cgroup storage of each type is allowed\n");
17209 /* Basic sanity check before we invest more work here. */
17210 if (!bpf_opcode_in_insntable(insn->code)) {
17211 verbose(env, "unknown opcode %02x\n", insn->code);
17216 /* now all pseudo BPF_LD_IMM64 instructions load valid
17217 * 'struct bpf_map *' into a register instead of user map_fd.
17218 * These pointers will be used later by verifier to validate map access.
17223 /* drop refcnt of maps used by the rejected program */
17224 static void release_maps(struct bpf_verifier_env *env)
17226 __bpf_free_used_maps(env->prog->aux, env->used_maps,
17227 env->used_map_cnt);
17230 /* drop refcnt of maps used by the rejected program */
17231 static void release_btfs(struct bpf_verifier_env *env)
17233 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
17234 env->used_btf_cnt);
17237 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
17238 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
17240 struct bpf_insn *insn = env->prog->insnsi;
17241 int insn_cnt = env->prog->len;
17244 for (i = 0; i < insn_cnt; i++, insn++) {
17245 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
17247 if (insn->src_reg == BPF_PSEUDO_FUNC)
17253 /* single env->prog->insni[off] instruction was replaced with the range
17254 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
17255 * [0, off) and [off, end) to new locations, so the patched range stays zero
17257 static void adjust_insn_aux_data(struct bpf_verifier_env *env,
17258 struct bpf_insn_aux_data *new_data,
17259 struct bpf_prog *new_prog, u32 off, u32 cnt)
17261 struct bpf_insn_aux_data *old_data = env->insn_aux_data;
17262 struct bpf_insn *insn = new_prog->insnsi;
17263 u32 old_seen = old_data[off].seen;
17267 /* aux info at OFF always needs adjustment, no matter fast path
17268 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
17269 * original insn at old prog.
17271 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
17275 prog_len = new_prog->len;
17277 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
17278 memcpy(new_data + off + cnt - 1, old_data + off,
17279 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
17280 for (i = off; i < off + cnt - 1; i++) {
17281 /* Expand insni[off]'s seen count to the patched range. */
17282 new_data[i].seen = old_seen;
17283 new_data[i].zext_dst = insn_has_def32(env, insn + i);
17285 env->insn_aux_data = new_data;
17289 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
17295 /* NOTE: fake 'exit' subprog should be updated as well. */
17296 for (i = 0; i <= env->subprog_cnt; i++) {
17297 if (env->subprog_info[i].start <= off)
17299 env->subprog_info[i].start += len - 1;
17303 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
17305 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
17306 int i, sz = prog->aux->size_poke_tab;
17307 struct bpf_jit_poke_descriptor *desc;
17309 for (i = 0; i < sz; i++) {
17311 if (desc->insn_idx <= off)
17313 desc->insn_idx += len - 1;
17317 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
17318 const struct bpf_insn *patch, u32 len)
17320 struct bpf_prog *new_prog;
17321 struct bpf_insn_aux_data *new_data = NULL;
17324 new_data = vzalloc(array_size(env->prog->len + len - 1,
17325 sizeof(struct bpf_insn_aux_data)));
17330 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
17331 if (IS_ERR(new_prog)) {
17332 if (PTR_ERR(new_prog) == -ERANGE)
17334 "insn %d cannot be patched due to 16-bit range\n",
17335 env->insn_aux_data[off].orig_idx);
17339 adjust_insn_aux_data(env, new_data, new_prog, off, len);
17340 adjust_subprog_starts(env, off, len);
17341 adjust_poke_descs(new_prog, off, len);
17345 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
17350 /* find first prog starting at or after off (first to remove) */
17351 for (i = 0; i < env->subprog_cnt; i++)
17352 if (env->subprog_info[i].start >= off)
17354 /* find first prog starting at or after off + cnt (first to stay) */
17355 for (j = i; j < env->subprog_cnt; j++)
17356 if (env->subprog_info[j].start >= off + cnt)
17358 /* if j doesn't start exactly at off + cnt, we are just removing
17359 * the front of previous prog
17361 if (env->subprog_info[j].start != off + cnt)
17365 struct bpf_prog_aux *aux = env->prog->aux;
17368 /* move fake 'exit' subprog as well */
17369 move = env->subprog_cnt + 1 - j;
17371 memmove(env->subprog_info + i,
17372 env->subprog_info + j,
17373 sizeof(*env->subprog_info) * move);
17374 env->subprog_cnt -= j - i;
17376 /* remove func_info */
17377 if (aux->func_info) {
17378 move = aux->func_info_cnt - j;
17380 memmove(aux->func_info + i,
17381 aux->func_info + j,
17382 sizeof(*aux->func_info) * move);
17383 aux->func_info_cnt -= j - i;
17384 /* func_info->insn_off is set after all code rewrites,
17385 * in adjust_btf_func() - no need to adjust
17389 /* convert i from "first prog to remove" to "first to adjust" */
17390 if (env->subprog_info[i].start == off)
17394 /* update fake 'exit' subprog as well */
17395 for (; i <= env->subprog_cnt; i++)
17396 env->subprog_info[i].start -= cnt;
17401 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
17404 struct bpf_prog *prog = env->prog;
17405 u32 i, l_off, l_cnt, nr_linfo;
17406 struct bpf_line_info *linfo;
17408 nr_linfo = prog->aux->nr_linfo;
17412 linfo = prog->aux->linfo;
17414 /* find first line info to remove, count lines to be removed */
17415 for (i = 0; i < nr_linfo; i++)
17416 if (linfo[i].insn_off >= off)
17421 for (; i < nr_linfo; i++)
17422 if (linfo[i].insn_off < off + cnt)
17427 /* First live insn doesn't match first live linfo, it needs to "inherit"
17428 * last removed linfo. prog is already modified, so prog->len == off
17429 * means no live instructions after (tail of the program was removed).
17431 if (prog->len != off && l_cnt &&
17432 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
17434 linfo[--i].insn_off = off + cnt;
17437 /* remove the line info which refer to the removed instructions */
17439 memmove(linfo + l_off, linfo + i,
17440 sizeof(*linfo) * (nr_linfo - i));
17442 prog->aux->nr_linfo -= l_cnt;
17443 nr_linfo = prog->aux->nr_linfo;
17446 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
17447 for (i = l_off; i < nr_linfo; i++)
17448 linfo[i].insn_off -= cnt;
17450 /* fix up all subprogs (incl. 'exit') which start >= off */
17451 for (i = 0; i <= env->subprog_cnt; i++)
17452 if (env->subprog_info[i].linfo_idx > l_off) {
17453 /* program may have started in the removed region but
17454 * may not be fully removed
17456 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
17457 env->subprog_info[i].linfo_idx -= l_cnt;
17459 env->subprog_info[i].linfo_idx = l_off;
17465 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
17467 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17468 unsigned int orig_prog_len = env->prog->len;
17471 if (bpf_prog_is_offloaded(env->prog->aux))
17472 bpf_prog_offload_remove_insns(env, off, cnt);
17474 err = bpf_remove_insns(env->prog, off, cnt);
17478 err = adjust_subprog_starts_after_remove(env, off, cnt);
17482 err = bpf_adj_linfo_after_remove(env, off, cnt);
17486 memmove(aux_data + off, aux_data + off + cnt,
17487 sizeof(*aux_data) * (orig_prog_len - off - cnt));
17492 /* The verifier does more data flow analysis than llvm and will not
17493 * explore branches that are dead at run time. Malicious programs can
17494 * have dead code too. Therefore replace all dead at-run-time code
17497 * Just nops are not optimal, e.g. if they would sit at the end of the
17498 * program and through another bug we would manage to jump there, then
17499 * we'd execute beyond program memory otherwise. Returning exception
17500 * code also wouldn't work since we can have subprogs where the dead
17501 * code could be located.
17503 static void sanitize_dead_code(struct bpf_verifier_env *env)
17505 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17506 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
17507 struct bpf_insn *insn = env->prog->insnsi;
17508 const int insn_cnt = env->prog->len;
17511 for (i = 0; i < insn_cnt; i++) {
17512 if (aux_data[i].seen)
17514 memcpy(insn + i, &trap, sizeof(trap));
17515 aux_data[i].zext_dst = false;
17519 static bool insn_is_cond_jump(u8 code)
17524 if (BPF_CLASS(code) == BPF_JMP32)
17525 return op != BPF_JA;
17527 if (BPF_CLASS(code) != BPF_JMP)
17530 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
17533 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
17535 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17536 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
17537 struct bpf_insn *insn = env->prog->insnsi;
17538 const int insn_cnt = env->prog->len;
17541 for (i = 0; i < insn_cnt; i++, insn++) {
17542 if (!insn_is_cond_jump(insn->code))
17545 if (!aux_data[i + 1].seen)
17546 ja.off = insn->off;
17547 else if (!aux_data[i + 1 + insn->off].seen)
17552 if (bpf_prog_is_offloaded(env->prog->aux))
17553 bpf_prog_offload_replace_insn(env, i, &ja);
17555 memcpy(insn, &ja, sizeof(ja));
17559 static int opt_remove_dead_code(struct bpf_verifier_env *env)
17561 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17562 int insn_cnt = env->prog->len;
17565 for (i = 0; i < insn_cnt; i++) {
17569 while (i + j < insn_cnt && !aux_data[i + j].seen)
17574 err = verifier_remove_insns(env, i, j);
17577 insn_cnt = env->prog->len;
17583 static int opt_remove_nops(struct bpf_verifier_env *env)
17585 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
17586 struct bpf_insn *insn = env->prog->insnsi;
17587 int insn_cnt = env->prog->len;
17590 for (i = 0; i < insn_cnt; i++) {
17591 if (memcmp(&insn[i], &ja, sizeof(ja)))
17594 err = verifier_remove_insns(env, i, 1);
17604 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
17605 const union bpf_attr *attr)
17607 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
17608 struct bpf_insn_aux_data *aux = env->insn_aux_data;
17609 int i, patch_len, delta = 0, len = env->prog->len;
17610 struct bpf_insn *insns = env->prog->insnsi;
17611 struct bpf_prog *new_prog;
17614 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
17615 zext_patch[1] = BPF_ZEXT_REG(0);
17616 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
17617 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
17618 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
17619 for (i = 0; i < len; i++) {
17620 int adj_idx = i + delta;
17621 struct bpf_insn insn;
17624 insn = insns[adj_idx];
17625 load_reg = insn_def_regno(&insn);
17626 if (!aux[adj_idx].zext_dst) {
17634 class = BPF_CLASS(code);
17635 if (load_reg == -1)
17638 /* NOTE: arg "reg" (the fourth one) is only used for
17639 * BPF_STX + SRC_OP, so it is safe to pass NULL
17642 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
17643 if (class == BPF_LD &&
17644 BPF_MODE(code) == BPF_IMM)
17649 /* ctx load could be transformed into wider load. */
17650 if (class == BPF_LDX &&
17651 aux[adj_idx].ptr_type == PTR_TO_CTX)
17654 imm_rnd = get_random_u32();
17655 rnd_hi32_patch[0] = insn;
17656 rnd_hi32_patch[1].imm = imm_rnd;
17657 rnd_hi32_patch[3].dst_reg = load_reg;
17658 patch = rnd_hi32_patch;
17660 goto apply_patch_buffer;
17663 /* Add in an zero-extend instruction if a) the JIT has requested
17664 * it or b) it's a CMPXCHG.
17666 * The latter is because: BPF_CMPXCHG always loads a value into
17667 * R0, therefore always zero-extends. However some archs'
17668 * equivalent instruction only does this load when the
17669 * comparison is successful. This detail of CMPXCHG is
17670 * orthogonal to the general zero-extension behaviour of the
17671 * CPU, so it's treated independently of bpf_jit_needs_zext.
17673 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
17676 /* Zero-extension is done by the caller. */
17677 if (bpf_pseudo_kfunc_call(&insn))
17680 if (WARN_ON(load_reg == -1)) {
17681 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
17685 zext_patch[0] = insn;
17686 zext_patch[1].dst_reg = load_reg;
17687 zext_patch[1].src_reg = load_reg;
17688 patch = zext_patch;
17690 apply_patch_buffer:
17691 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
17694 env->prog = new_prog;
17695 insns = new_prog->insnsi;
17696 aux = env->insn_aux_data;
17697 delta += patch_len - 1;
17703 /* convert load instructions that access fields of a context type into a
17704 * sequence of instructions that access fields of the underlying structure:
17705 * struct __sk_buff -> struct sk_buff
17706 * struct bpf_sock_ops -> struct sock
17708 static int convert_ctx_accesses(struct bpf_verifier_env *env)
17710 const struct bpf_verifier_ops *ops = env->ops;
17711 int i, cnt, size, ctx_field_size, delta = 0;
17712 const int insn_cnt = env->prog->len;
17713 struct bpf_insn insn_buf[16], *insn;
17714 u32 target_size, size_default, off;
17715 struct bpf_prog *new_prog;
17716 enum bpf_access_type type;
17717 bool is_narrower_load;
17719 if (ops->gen_prologue || env->seen_direct_write) {
17720 if (!ops->gen_prologue) {
17721 verbose(env, "bpf verifier is misconfigured\n");
17724 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
17726 if (cnt >= ARRAY_SIZE(insn_buf)) {
17727 verbose(env, "bpf verifier is misconfigured\n");
17730 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
17734 env->prog = new_prog;
17739 if (bpf_prog_is_offloaded(env->prog->aux))
17742 insn = env->prog->insnsi + delta;
17744 for (i = 0; i < insn_cnt; i++, insn++) {
17745 bpf_convert_ctx_access_t convert_ctx_access;
17748 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
17749 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
17750 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
17751 insn->code == (BPF_LDX | BPF_MEM | BPF_DW) ||
17752 insn->code == (BPF_LDX | BPF_MEMSX | BPF_B) ||
17753 insn->code == (BPF_LDX | BPF_MEMSX | BPF_H) ||
17754 insn->code == (BPF_LDX | BPF_MEMSX | BPF_W)) {
17756 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
17757 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
17758 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
17759 insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
17760 insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
17761 insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
17762 insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
17763 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
17769 if (type == BPF_WRITE &&
17770 env->insn_aux_data[i + delta].sanitize_stack_spill) {
17771 struct bpf_insn patch[] = {
17776 cnt = ARRAY_SIZE(patch);
17777 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
17782 env->prog = new_prog;
17783 insn = new_prog->insnsi + i + delta;
17787 switch ((int)env->insn_aux_data[i + delta].ptr_type) {
17789 if (!ops->convert_ctx_access)
17791 convert_ctx_access = ops->convert_ctx_access;
17793 case PTR_TO_SOCKET:
17794 case PTR_TO_SOCK_COMMON:
17795 convert_ctx_access = bpf_sock_convert_ctx_access;
17797 case PTR_TO_TCP_SOCK:
17798 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
17800 case PTR_TO_XDP_SOCK:
17801 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
17803 case PTR_TO_BTF_ID:
17804 case PTR_TO_BTF_ID | PTR_UNTRUSTED:
17805 /* PTR_TO_BTF_ID | MEM_ALLOC always has a valid lifetime, unlike
17806 * PTR_TO_BTF_ID, and an active ref_obj_id, but the same cannot
17807 * be said once it is marked PTR_UNTRUSTED, hence we must handle
17808 * any faults for loads into such types. BPF_WRITE is disallowed
17811 case PTR_TO_BTF_ID | MEM_ALLOC | PTR_UNTRUSTED:
17812 if (type == BPF_READ) {
17813 if (BPF_MODE(insn->code) == BPF_MEM)
17814 insn->code = BPF_LDX | BPF_PROBE_MEM |
17815 BPF_SIZE((insn)->code);
17817 insn->code = BPF_LDX | BPF_PROBE_MEMSX |
17818 BPF_SIZE((insn)->code);
17819 env->prog->aux->num_exentries++;
17826 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
17827 size = BPF_LDST_BYTES(insn);
17828 mode = BPF_MODE(insn->code);
17830 /* If the read access is a narrower load of the field,
17831 * convert to a 4/8-byte load, to minimum program type specific
17832 * convert_ctx_access changes. If conversion is successful,
17833 * we will apply proper mask to the result.
17835 is_narrower_load = size < ctx_field_size;
17836 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
17838 if (is_narrower_load) {
17841 if (type == BPF_WRITE) {
17842 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
17847 if (ctx_field_size == 4)
17849 else if (ctx_field_size == 8)
17850 size_code = BPF_DW;
17852 insn->off = off & ~(size_default - 1);
17853 insn->code = BPF_LDX | BPF_MEM | size_code;
17857 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
17859 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
17860 (ctx_field_size && !target_size)) {
17861 verbose(env, "bpf verifier is misconfigured\n");
17865 if (is_narrower_load && size < target_size) {
17866 u8 shift = bpf_ctx_narrow_access_offset(
17867 off, size, size_default) * 8;
17868 if (shift && cnt + 1 >= ARRAY_SIZE(insn_buf)) {
17869 verbose(env, "bpf verifier narrow ctx load misconfigured\n");
17872 if (ctx_field_size <= 4) {
17874 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
17877 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
17878 (1 << size * 8) - 1);
17881 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
17884 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
17885 (1ULL << size * 8) - 1);
17888 if (mode == BPF_MEMSX)
17889 insn_buf[cnt++] = BPF_RAW_INSN(BPF_ALU64 | BPF_MOV | BPF_X,
17890 insn->dst_reg, insn->dst_reg,
17893 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
17899 /* keep walking new program and skip insns we just inserted */
17900 env->prog = new_prog;
17901 insn = new_prog->insnsi + i + delta;
17907 static int jit_subprogs(struct bpf_verifier_env *env)
17909 struct bpf_prog *prog = env->prog, **func, *tmp;
17910 int i, j, subprog_start, subprog_end = 0, len, subprog;
17911 struct bpf_map *map_ptr;
17912 struct bpf_insn *insn;
17913 void *old_bpf_func;
17914 int err, num_exentries;
17916 if (env->subprog_cnt <= 1)
17919 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
17920 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn))
17923 /* Upon error here we cannot fall back to interpreter but
17924 * need a hard reject of the program. Thus -EFAULT is
17925 * propagated in any case.
17927 subprog = find_subprog(env, i + insn->imm + 1);
17929 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
17930 i + insn->imm + 1);
17933 /* temporarily remember subprog id inside insn instead of
17934 * aux_data, since next loop will split up all insns into funcs
17936 insn->off = subprog;
17937 /* remember original imm in case JIT fails and fallback
17938 * to interpreter will be needed
17940 env->insn_aux_data[i].call_imm = insn->imm;
17941 /* point imm to __bpf_call_base+1 from JITs point of view */
17943 if (bpf_pseudo_func(insn))
17944 /* jit (e.g. x86_64) may emit fewer instructions
17945 * if it learns a u32 imm is the same as a u64 imm.
17946 * Force a non zero here.
17951 err = bpf_prog_alloc_jited_linfo(prog);
17953 goto out_undo_insn;
17956 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
17958 goto out_undo_insn;
17960 for (i = 0; i < env->subprog_cnt; i++) {
17961 subprog_start = subprog_end;
17962 subprog_end = env->subprog_info[i + 1].start;
17964 len = subprog_end - subprog_start;
17965 /* bpf_prog_run() doesn't call subprogs directly,
17966 * hence main prog stats include the runtime of subprogs.
17967 * subprogs don't have IDs and not reachable via prog_get_next_id
17968 * func[i]->stats will never be accessed and stays NULL
17970 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
17973 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
17974 len * sizeof(struct bpf_insn));
17975 func[i]->type = prog->type;
17976 func[i]->len = len;
17977 if (bpf_prog_calc_tag(func[i]))
17979 func[i]->is_func = 1;
17980 func[i]->aux->func_idx = i;
17981 /* Below members will be freed only at prog->aux */
17982 func[i]->aux->btf = prog->aux->btf;
17983 func[i]->aux->func_info = prog->aux->func_info;
17984 func[i]->aux->func_info_cnt = prog->aux->func_info_cnt;
17985 func[i]->aux->poke_tab = prog->aux->poke_tab;
17986 func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
17988 for (j = 0; j < prog->aux->size_poke_tab; j++) {
17989 struct bpf_jit_poke_descriptor *poke;
17991 poke = &prog->aux->poke_tab[j];
17992 if (poke->insn_idx < subprog_end &&
17993 poke->insn_idx >= subprog_start)
17994 poke->aux = func[i]->aux;
17997 func[i]->aux->name[0] = 'F';
17998 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
17999 func[i]->jit_requested = 1;
18000 func[i]->blinding_requested = prog->blinding_requested;
18001 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
18002 func[i]->aux->kfunc_btf_tab = prog->aux->kfunc_btf_tab;
18003 func[i]->aux->linfo = prog->aux->linfo;
18004 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
18005 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
18006 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
18008 insn = func[i]->insnsi;
18009 for (j = 0; j < func[i]->len; j++, insn++) {
18010 if (BPF_CLASS(insn->code) == BPF_LDX &&
18011 (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
18012 BPF_MODE(insn->code) == BPF_PROBE_MEMSX))
18015 func[i]->aux->num_exentries = num_exentries;
18016 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
18017 func[i] = bpf_int_jit_compile(func[i]);
18018 if (!func[i]->jited) {
18025 /* at this point all bpf functions were successfully JITed
18026 * now populate all bpf_calls with correct addresses and
18027 * run last pass of JIT
18029 for (i = 0; i < env->subprog_cnt; i++) {
18030 insn = func[i]->insnsi;
18031 for (j = 0; j < func[i]->len; j++, insn++) {
18032 if (bpf_pseudo_func(insn)) {
18033 subprog = insn->off;
18034 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
18035 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
18038 if (!bpf_pseudo_call(insn))
18040 subprog = insn->off;
18041 insn->imm = BPF_CALL_IMM(func[subprog]->bpf_func);
18044 /* we use the aux data to keep a list of the start addresses
18045 * of the JITed images for each function in the program
18047 * for some architectures, such as powerpc64, the imm field
18048 * might not be large enough to hold the offset of the start
18049 * address of the callee's JITed image from __bpf_call_base
18051 * in such cases, we can lookup the start address of a callee
18052 * by using its subprog id, available from the off field of
18053 * the call instruction, as an index for this list
18055 func[i]->aux->func = func;
18056 func[i]->aux->func_cnt = env->subprog_cnt;
18058 for (i = 0; i < env->subprog_cnt; i++) {
18059 old_bpf_func = func[i]->bpf_func;
18060 tmp = bpf_int_jit_compile(func[i]);
18061 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
18062 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
18069 /* finally lock prog and jit images for all functions and
18070 * populate kallsysm. Begin at the first subprogram, since
18071 * bpf_prog_load will add the kallsyms for the main program.
18073 for (i = 1; i < env->subprog_cnt; i++) {
18074 bpf_prog_lock_ro(func[i]);
18075 bpf_prog_kallsyms_add(func[i]);
18078 /* Last step: make now unused interpreter insns from main
18079 * prog consistent for later dump requests, so they can
18080 * later look the same as if they were interpreted only.
18082 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
18083 if (bpf_pseudo_func(insn)) {
18084 insn[0].imm = env->insn_aux_data[i].call_imm;
18085 insn[1].imm = insn->off;
18089 if (!bpf_pseudo_call(insn))
18091 insn->off = env->insn_aux_data[i].call_imm;
18092 subprog = find_subprog(env, i + insn->off + 1);
18093 insn->imm = subprog;
18097 prog->bpf_func = func[0]->bpf_func;
18098 prog->jited_len = func[0]->jited_len;
18099 prog->aux->extable = func[0]->aux->extable;
18100 prog->aux->num_exentries = func[0]->aux->num_exentries;
18101 prog->aux->func = func;
18102 prog->aux->func_cnt = env->subprog_cnt;
18103 bpf_prog_jit_attempt_done(prog);
18106 /* We failed JIT'ing, so at this point we need to unregister poke
18107 * descriptors from subprogs, so that kernel is not attempting to
18108 * patch it anymore as we're freeing the subprog JIT memory.
18110 for (i = 0; i < prog->aux->size_poke_tab; i++) {
18111 map_ptr = prog->aux->poke_tab[i].tail_call.map;
18112 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
18114 /* At this point we're guaranteed that poke descriptors are not
18115 * live anymore. We can just unlink its descriptor table as it's
18116 * released with the main prog.
18118 for (i = 0; i < env->subprog_cnt; i++) {
18121 func[i]->aux->poke_tab = NULL;
18122 bpf_jit_free(func[i]);
18126 /* cleanup main prog to be interpreted */
18127 prog->jit_requested = 0;
18128 prog->blinding_requested = 0;
18129 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
18130 if (!bpf_pseudo_call(insn))
18133 insn->imm = env->insn_aux_data[i].call_imm;
18135 bpf_prog_jit_attempt_done(prog);
18139 static int fixup_call_args(struct bpf_verifier_env *env)
18141 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
18142 struct bpf_prog *prog = env->prog;
18143 struct bpf_insn *insn = prog->insnsi;
18144 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
18149 if (env->prog->jit_requested &&
18150 !bpf_prog_is_offloaded(env->prog->aux)) {
18151 err = jit_subprogs(env);
18154 if (err == -EFAULT)
18157 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
18158 if (has_kfunc_call) {
18159 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
18162 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
18163 /* When JIT fails the progs with bpf2bpf calls and tail_calls
18164 * have to be rejected, since interpreter doesn't support them yet.
18166 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
18169 for (i = 0; i < prog->len; i++, insn++) {
18170 if (bpf_pseudo_func(insn)) {
18171 /* When JIT fails the progs with callback calls
18172 * have to be rejected, since interpreter doesn't support them yet.
18174 verbose(env, "callbacks are not allowed in non-JITed programs\n");
18178 if (!bpf_pseudo_call(insn))
18180 depth = get_callee_stack_depth(env, insn, i);
18183 bpf_patch_call_args(insn, depth);
18190 /* replace a generic kfunc with a specialized version if necessary */
18191 static void specialize_kfunc(struct bpf_verifier_env *env,
18192 u32 func_id, u16 offset, unsigned long *addr)
18194 struct bpf_prog *prog = env->prog;
18195 bool seen_direct_write;
18199 if (bpf_dev_bound_kfunc_id(func_id)) {
18200 xdp_kfunc = bpf_dev_bound_resolve_kfunc(prog, func_id);
18202 *addr = (unsigned long)xdp_kfunc;
18205 /* fallback to default kfunc when not supported by netdev */
18211 if (func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) {
18212 seen_direct_write = env->seen_direct_write;
18213 is_rdonly = !may_access_direct_pkt_data(env, NULL, BPF_WRITE);
18216 *addr = (unsigned long)bpf_dynptr_from_skb_rdonly;
18218 /* restore env->seen_direct_write to its original value, since
18219 * may_access_direct_pkt_data mutates it
18221 env->seen_direct_write = seen_direct_write;
18225 static void __fixup_collection_insert_kfunc(struct bpf_insn_aux_data *insn_aux,
18226 u16 struct_meta_reg,
18227 u16 node_offset_reg,
18228 struct bpf_insn *insn,
18229 struct bpf_insn *insn_buf,
18232 struct btf_struct_meta *kptr_struct_meta = insn_aux->kptr_struct_meta;
18233 struct bpf_insn addr[2] = { BPF_LD_IMM64(struct_meta_reg, (long)kptr_struct_meta) };
18235 insn_buf[0] = addr[0];
18236 insn_buf[1] = addr[1];
18237 insn_buf[2] = BPF_MOV64_IMM(node_offset_reg, insn_aux->insert_off);
18238 insn_buf[3] = *insn;
18242 static int fixup_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
18243 struct bpf_insn *insn_buf, int insn_idx, int *cnt)
18245 const struct bpf_kfunc_desc *desc;
18248 verbose(env, "invalid kernel function call not eliminated in verifier pass\n");
18254 /* insn->imm has the btf func_id. Replace it with an offset relative to
18255 * __bpf_call_base, unless the JIT needs to call functions that are
18256 * further than 32 bits away (bpf_jit_supports_far_kfunc_call()).
18258 desc = find_kfunc_desc(env->prog, insn->imm, insn->off);
18260 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
18265 if (!bpf_jit_supports_far_kfunc_call())
18266 insn->imm = BPF_CALL_IMM(desc->addr);
18269 if (desc->func_id == special_kfunc_list[KF_bpf_obj_new_impl]) {
18270 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18271 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
18272 u64 obj_new_size = env->insn_aux_data[insn_idx].obj_new_size;
18274 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_1, obj_new_size);
18275 insn_buf[1] = addr[0];
18276 insn_buf[2] = addr[1];
18277 insn_buf[3] = *insn;
18279 } else if (desc->func_id == special_kfunc_list[KF_bpf_obj_drop_impl] ||
18280 desc->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) {
18281 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18282 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
18284 if (desc->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl] &&
18285 !kptr_struct_meta) {
18286 verbose(env, "verifier internal error: kptr_struct_meta expected at insn_idx %d\n",
18291 insn_buf[0] = addr[0];
18292 insn_buf[1] = addr[1];
18293 insn_buf[2] = *insn;
18295 } else if (desc->func_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
18296 desc->func_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
18297 desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
18298 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18299 int struct_meta_reg = BPF_REG_3;
18300 int node_offset_reg = BPF_REG_4;
18302 /* rbtree_add has extra 'less' arg, so args-to-fixup are in diff regs */
18303 if (desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
18304 struct_meta_reg = BPF_REG_4;
18305 node_offset_reg = BPF_REG_5;
18308 if (!kptr_struct_meta) {
18309 verbose(env, "verifier internal error: kptr_struct_meta expected at insn_idx %d\n",
18314 __fixup_collection_insert_kfunc(&env->insn_aux_data[insn_idx], struct_meta_reg,
18315 node_offset_reg, insn, insn_buf, cnt);
18316 } else if (desc->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx] ||
18317 desc->func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
18318 insn_buf[0] = BPF_MOV64_REG(BPF_REG_0, BPF_REG_1);
18324 /* Do various post-verification rewrites in a single program pass.
18325 * These rewrites simplify JIT and interpreter implementations.
18327 static int do_misc_fixups(struct bpf_verifier_env *env)
18329 struct bpf_prog *prog = env->prog;
18330 enum bpf_attach_type eatype = prog->expected_attach_type;
18331 enum bpf_prog_type prog_type = resolve_prog_type(prog);
18332 struct bpf_insn *insn = prog->insnsi;
18333 const struct bpf_func_proto *fn;
18334 const int insn_cnt = prog->len;
18335 const struct bpf_map_ops *ops;
18336 struct bpf_insn_aux_data *aux;
18337 struct bpf_insn insn_buf[16];
18338 struct bpf_prog *new_prog;
18339 struct bpf_map *map_ptr;
18340 int i, ret, cnt, delta = 0;
18342 for (i = 0; i < insn_cnt; i++, insn++) {
18343 /* Make divide-by-zero exceptions impossible. */
18344 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
18345 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
18346 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
18347 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
18348 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
18349 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
18350 struct bpf_insn *patchlet;
18351 struct bpf_insn chk_and_div[] = {
18352 /* [R,W]x div 0 -> 0 */
18353 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
18354 BPF_JNE | BPF_K, insn->src_reg,
18356 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
18357 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
18360 struct bpf_insn chk_and_mod[] = {
18361 /* [R,W]x mod 0 -> [R,W]x */
18362 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
18363 BPF_JEQ | BPF_K, insn->src_reg,
18364 0, 1 + (is64 ? 0 : 1), 0),
18366 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
18367 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
18370 patchlet = isdiv ? chk_and_div : chk_and_mod;
18371 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
18372 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
18374 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
18379 env->prog = prog = new_prog;
18380 insn = new_prog->insnsi + i + delta;
18384 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
18385 if (BPF_CLASS(insn->code) == BPF_LD &&
18386 (BPF_MODE(insn->code) == BPF_ABS ||
18387 BPF_MODE(insn->code) == BPF_IND)) {
18388 cnt = env->ops->gen_ld_abs(insn, insn_buf);
18389 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
18390 verbose(env, "bpf verifier is misconfigured\n");
18394 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18399 env->prog = prog = new_prog;
18400 insn = new_prog->insnsi + i + delta;
18404 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
18405 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
18406 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
18407 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
18408 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
18409 struct bpf_insn *patch = &insn_buf[0];
18410 bool issrc, isneg, isimm;
18413 aux = &env->insn_aux_data[i + delta];
18414 if (!aux->alu_state ||
18415 aux->alu_state == BPF_ALU_NON_POINTER)
18418 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
18419 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
18420 BPF_ALU_SANITIZE_SRC;
18421 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
18423 off_reg = issrc ? insn->src_reg : insn->dst_reg;
18425 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
18428 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
18429 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
18430 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
18431 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
18432 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
18433 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
18434 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
18437 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
18438 insn->src_reg = BPF_REG_AX;
18440 insn->code = insn->code == code_add ?
18441 code_sub : code_add;
18443 if (issrc && isneg && !isimm)
18444 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
18445 cnt = patch - insn_buf;
18447 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18452 env->prog = prog = new_prog;
18453 insn = new_prog->insnsi + i + delta;
18457 if (insn->code != (BPF_JMP | BPF_CALL))
18459 if (insn->src_reg == BPF_PSEUDO_CALL)
18461 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
18462 ret = fixup_kfunc_call(env, insn, insn_buf, i + delta, &cnt);
18468 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18473 env->prog = prog = new_prog;
18474 insn = new_prog->insnsi + i + delta;
18478 if (insn->imm == BPF_FUNC_get_route_realm)
18479 prog->dst_needed = 1;
18480 if (insn->imm == BPF_FUNC_get_prandom_u32)
18481 bpf_user_rnd_init_once();
18482 if (insn->imm == BPF_FUNC_override_return)
18483 prog->kprobe_override = 1;
18484 if (insn->imm == BPF_FUNC_tail_call) {
18485 /* If we tail call into other programs, we
18486 * cannot make any assumptions since they can
18487 * be replaced dynamically during runtime in
18488 * the program array.
18490 prog->cb_access = 1;
18491 if (!allow_tail_call_in_subprogs(env))
18492 prog->aux->stack_depth = MAX_BPF_STACK;
18493 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
18495 /* mark bpf_tail_call as different opcode to avoid
18496 * conditional branch in the interpreter for every normal
18497 * call and to prevent accidental JITing by JIT compiler
18498 * that doesn't support bpf_tail_call yet
18501 insn->code = BPF_JMP | BPF_TAIL_CALL;
18503 aux = &env->insn_aux_data[i + delta];
18504 if (env->bpf_capable && !prog->blinding_requested &&
18505 prog->jit_requested &&
18506 !bpf_map_key_poisoned(aux) &&
18507 !bpf_map_ptr_poisoned(aux) &&
18508 !bpf_map_ptr_unpriv(aux)) {
18509 struct bpf_jit_poke_descriptor desc = {
18510 .reason = BPF_POKE_REASON_TAIL_CALL,
18511 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
18512 .tail_call.key = bpf_map_key_immediate(aux),
18513 .insn_idx = i + delta,
18516 ret = bpf_jit_add_poke_descriptor(prog, &desc);
18518 verbose(env, "adding tail call poke descriptor failed\n");
18522 insn->imm = ret + 1;
18526 if (!bpf_map_ptr_unpriv(aux))
18529 /* instead of changing every JIT dealing with tail_call
18530 * emit two extra insns:
18531 * if (index >= max_entries) goto out;
18532 * index &= array->index_mask;
18533 * to avoid out-of-bounds cpu speculation
18535 if (bpf_map_ptr_poisoned(aux)) {
18536 verbose(env, "tail_call abusing map_ptr\n");
18540 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
18541 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
18542 map_ptr->max_entries, 2);
18543 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
18544 container_of(map_ptr,
18547 insn_buf[2] = *insn;
18549 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18554 env->prog = prog = new_prog;
18555 insn = new_prog->insnsi + i + delta;
18559 if (insn->imm == BPF_FUNC_timer_set_callback) {
18560 /* The verifier will process callback_fn as many times as necessary
18561 * with different maps and the register states prepared by
18562 * set_timer_callback_state will be accurate.
18564 * The following use case is valid:
18565 * map1 is shared by prog1, prog2, prog3.
18566 * prog1 calls bpf_timer_init for some map1 elements
18567 * prog2 calls bpf_timer_set_callback for some map1 elements.
18568 * Those that were not bpf_timer_init-ed will return -EINVAL.
18569 * prog3 calls bpf_timer_start for some map1 elements.
18570 * Those that were not both bpf_timer_init-ed and
18571 * bpf_timer_set_callback-ed will return -EINVAL.
18573 struct bpf_insn ld_addrs[2] = {
18574 BPF_LD_IMM64(BPF_REG_3, (long)prog->aux),
18577 insn_buf[0] = ld_addrs[0];
18578 insn_buf[1] = ld_addrs[1];
18579 insn_buf[2] = *insn;
18582 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18587 env->prog = prog = new_prog;
18588 insn = new_prog->insnsi + i + delta;
18589 goto patch_call_imm;
18592 if (is_storage_get_function(insn->imm)) {
18593 if (!env->prog->aux->sleepable ||
18594 env->insn_aux_data[i + delta].storage_get_func_atomic)
18595 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_ATOMIC);
18597 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_KERNEL);
18598 insn_buf[1] = *insn;
18601 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18606 env->prog = prog = new_prog;
18607 insn = new_prog->insnsi + i + delta;
18608 goto patch_call_imm;
18611 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
18612 * and other inlining handlers are currently limited to 64 bit
18615 if (prog->jit_requested && BITS_PER_LONG == 64 &&
18616 (insn->imm == BPF_FUNC_map_lookup_elem ||
18617 insn->imm == BPF_FUNC_map_update_elem ||
18618 insn->imm == BPF_FUNC_map_delete_elem ||
18619 insn->imm == BPF_FUNC_map_push_elem ||
18620 insn->imm == BPF_FUNC_map_pop_elem ||
18621 insn->imm == BPF_FUNC_map_peek_elem ||
18622 insn->imm == BPF_FUNC_redirect_map ||
18623 insn->imm == BPF_FUNC_for_each_map_elem ||
18624 insn->imm == BPF_FUNC_map_lookup_percpu_elem)) {
18625 aux = &env->insn_aux_data[i + delta];
18626 if (bpf_map_ptr_poisoned(aux))
18627 goto patch_call_imm;
18629 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
18630 ops = map_ptr->ops;
18631 if (insn->imm == BPF_FUNC_map_lookup_elem &&
18632 ops->map_gen_lookup) {
18633 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
18634 if (cnt == -EOPNOTSUPP)
18635 goto patch_map_ops_generic;
18636 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
18637 verbose(env, "bpf verifier is misconfigured\n");
18641 new_prog = bpf_patch_insn_data(env, i + delta,
18647 env->prog = prog = new_prog;
18648 insn = new_prog->insnsi + i + delta;
18652 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
18653 (void *(*)(struct bpf_map *map, void *key))NULL));
18654 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
18655 (long (*)(struct bpf_map *map, void *key))NULL));
18656 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
18657 (long (*)(struct bpf_map *map, void *key, void *value,
18659 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
18660 (long (*)(struct bpf_map *map, void *value,
18662 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
18663 (long (*)(struct bpf_map *map, void *value))NULL));
18664 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
18665 (long (*)(struct bpf_map *map, void *value))NULL));
18666 BUILD_BUG_ON(!__same_type(ops->map_redirect,
18667 (long (*)(struct bpf_map *map, u64 index, u64 flags))NULL));
18668 BUILD_BUG_ON(!__same_type(ops->map_for_each_callback,
18669 (long (*)(struct bpf_map *map,
18670 bpf_callback_t callback_fn,
18671 void *callback_ctx,
18673 BUILD_BUG_ON(!__same_type(ops->map_lookup_percpu_elem,
18674 (void *(*)(struct bpf_map *map, void *key, u32 cpu))NULL));
18676 patch_map_ops_generic:
18677 switch (insn->imm) {
18678 case BPF_FUNC_map_lookup_elem:
18679 insn->imm = BPF_CALL_IMM(ops->map_lookup_elem);
18681 case BPF_FUNC_map_update_elem:
18682 insn->imm = BPF_CALL_IMM(ops->map_update_elem);
18684 case BPF_FUNC_map_delete_elem:
18685 insn->imm = BPF_CALL_IMM(ops->map_delete_elem);
18687 case BPF_FUNC_map_push_elem:
18688 insn->imm = BPF_CALL_IMM(ops->map_push_elem);
18690 case BPF_FUNC_map_pop_elem:
18691 insn->imm = BPF_CALL_IMM(ops->map_pop_elem);
18693 case BPF_FUNC_map_peek_elem:
18694 insn->imm = BPF_CALL_IMM(ops->map_peek_elem);
18696 case BPF_FUNC_redirect_map:
18697 insn->imm = BPF_CALL_IMM(ops->map_redirect);
18699 case BPF_FUNC_for_each_map_elem:
18700 insn->imm = BPF_CALL_IMM(ops->map_for_each_callback);
18702 case BPF_FUNC_map_lookup_percpu_elem:
18703 insn->imm = BPF_CALL_IMM(ops->map_lookup_percpu_elem);
18707 goto patch_call_imm;
18710 /* Implement bpf_jiffies64 inline. */
18711 if (prog->jit_requested && BITS_PER_LONG == 64 &&
18712 insn->imm == BPF_FUNC_jiffies64) {
18713 struct bpf_insn ld_jiffies_addr[2] = {
18714 BPF_LD_IMM64(BPF_REG_0,
18715 (unsigned long)&jiffies),
18718 insn_buf[0] = ld_jiffies_addr[0];
18719 insn_buf[1] = ld_jiffies_addr[1];
18720 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
18724 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
18730 env->prog = prog = new_prog;
18731 insn = new_prog->insnsi + i + delta;
18735 /* Implement bpf_get_func_arg inline. */
18736 if (prog_type == BPF_PROG_TYPE_TRACING &&
18737 insn->imm == BPF_FUNC_get_func_arg) {
18738 /* Load nr_args from ctx - 8 */
18739 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
18740 insn_buf[1] = BPF_JMP32_REG(BPF_JGE, BPF_REG_2, BPF_REG_0, 6);
18741 insn_buf[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_2, 3);
18742 insn_buf[3] = BPF_ALU64_REG(BPF_ADD, BPF_REG_2, BPF_REG_1);
18743 insn_buf[4] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_2, 0);
18744 insn_buf[5] = BPF_STX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
18745 insn_buf[6] = BPF_MOV64_IMM(BPF_REG_0, 0);
18746 insn_buf[7] = BPF_JMP_A(1);
18747 insn_buf[8] = BPF_MOV64_IMM(BPF_REG_0, -EINVAL);
18750 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18755 env->prog = prog = new_prog;
18756 insn = new_prog->insnsi + i + delta;
18760 /* Implement bpf_get_func_ret inline. */
18761 if (prog_type == BPF_PROG_TYPE_TRACING &&
18762 insn->imm == BPF_FUNC_get_func_ret) {
18763 if (eatype == BPF_TRACE_FEXIT ||
18764 eatype == BPF_MODIFY_RETURN) {
18765 /* Load nr_args from ctx - 8 */
18766 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
18767 insn_buf[1] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_0, 3);
18768 insn_buf[2] = BPF_ALU64_REG(BPF_ADD, BPF_REG_0, BPF_REG_1);
18769 insn_buf[3] = BPF_LDX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
18770 insn_buf[4] = BPF_STX_MEM(BPF_DW, BPF_REG_2, BPF_REG_3, 0);
18771 insn_buf[5] = BPF_MOV64_IMM(BPF_REG_0, 0);
18774 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, -EOPNOTSUPP);
18778 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18783 env->prog = prog = new_prog;
18784 insn = new_prog->insnsi + i + delta;
18788 /* Implement get_func_arg_cnt inline. */
18789 if (prog_type == BPF_PROG_TYPE_TRACING &&
18790 insn->imm == BPF_FUNC_get_func_arg_cnt) {
18791 /* Load nr_args from ctx - 8 */
18792 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
18794 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
18798 env->prog = prog = new_prog;
18799 insn = new_prog->insnsi + i + delta;
18803 /* Implement bpf_get_func_ip inline. */
18804 if (prog_type == BPF_PROG_TYPE_TRACING &&
18805 insn->imm == BPF_FUNC_get_func_ip) {
18806 /* Load IP address from ctx - 16 */
18807 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -16);
18809 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
18813 env->prog = prog = new_prog;
18814 insn = new_prog->insnsi + i + delta;
18819 fn = env->ops->get_func_proto(insn->imm, env->prog);
18820 /* all functions that have prototype and verifier allowed
18821 * programs to call them, must be real in-kernel functions
18825 "kernel subsystem misconfigured func %s#%d\n",
18826 func_id_name(insn->imm), insn->imm);
18829 insn->imm = fn->func - __bpf_call_base;
18832 /* Since poke tab is now finalized, publish aux to tracker. */
18833 for (i = 0; i < prog->aux->size_poke_tab; i++) {
18834 map_ptr = prog->aux->poke_tab[i].tail_call.map;
18835 if (!map_ptr->ops->map_poke_track ||
18836 !map_ptr->ops->map_poke_untrack ||
18837 !map_ptr->ops->map_poke_run) {
18838 verbose(env, "bpf verifier is misconfigured\n");
18842 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
18844 verbose(env, "tracking tail call prog failed\n");
18849 sort_kfunc_descs_by_imm_off(env->prog);
18854 static struct bpf_prog *inline_bpf_loop(struct bpf_verifier_env *env,
18857 u32 callback_subprogno,
18860 s32 r6_offset = stack_base + 0 * BPF_REG_SIZE;
18861 s32 r7_offset = stack_base + 1 * BPF_REG_SIZE;
18862 s32 r8_offset = stack_base + 2 * BPF_REG_SIZE;
18863 int reg_loop_max = BPF_REG_6;
18864 int reg_loop_cnt = BPF_REG_7;
18865 int reg_loop_ctx = BPF_REG_8;
18867 struct bpf_prog *new_prog;
18868 u32 callback_start;
18869 u32 call_insn_offset;
18870 s32 callback_offset;
18872 /* This represents an inlined version of bpf_iter.c:bpf_loop,
18873 * be careful to modify this code in sync.
18875 struct bpf_insn insn_buf[] = {
18876 /* Return error and jump to the end of the patch if
18877 * expected number of iterations is too big.
18879 BPF_JMP_IMM(BPF_JLE, BPF_REG_1, BPF_MAX_LOOPS, 2),
18880 BPF_MOV32_IMM(BPF_REG_0, -E2BIG),
18881 BPF_JMP_IMM(BPF_JA, 0, 0, 16),
18882 /* spill R6, R7, R8 to use these as loop vars */
18883 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_6, r6_offset),
18884 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_7, r7_offset),
18885 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_8, r8_offset),
18886 /* initialize loop vars */
18887 BPF_MOV64_REG(reg_loop_max, BPF_REG_1),
18888 BPF_MOV32_IMM(reg_loop_cnt, 0),
18889 BPF_MOV64_REG(reg_loop_ctx, BPF_REG_3),
18891 * if reg_loop_cnt >= reg_loop_max skip the loop body
18893 BPF_JMP_REG(BPF_JGE, reg_loop_cnt, reg_loop_max, 5),
18895 * correct callback offset would be set after patching
18897 BPF_MOV64_REG(BPF_REG_1, reg_loop_cnt),
18898 BPF_MOV64_REG(BPF_REG_2, reg_loop_ctx),
18900 /* increment loop counter */
18901 BPF_ALU64_IMM(BPF_ADD, reg_loop_cnt, 1),
18902 /* jump to loop header if callback returned 0 */
18903 BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, -6),
18904 /* return value of bpf_loop,
18905 * set R0 to the number of iterations
18907 BPF_MOV64_REG(BPF_REG_0, reg_loop_cnt),
18908 /* restore original values of R6, R7, R8 */
18909 BPF_LDX_MEM(BPF_DW, BPF_REG_6, BPF_REG_10, r6_offset),
18910 BPF_LDX_MEM(BPF_DW, BPF_REG_7, BPF_REG_10, r7_offset),
18911 BPF_LDX_MEM(BPF_DW, BPF_REG_8, BPF_REG_10, r8_offset),
18914 *cnt = ARRAY_SIZE(insn_buf);
18915 new_prog = bpf_patch_insn_data(env, position, insn_buf, *cnt);
18919 /* callback start is known only after patching */
18920 callback_start = env->subprog_info[callback_subprogno].start;
18921 /* Note: insn_buf[12] is an offset of BPF_CALL_REL instruction */
18922 call_insn_offset = position + 12;
18923 callback_offset = callback_start - call_insn_offset - 1;
18924 new_prog->insnsi[call_insn_offset].imm = callback_offset;
18929 static bool is_bpf_loop_call(struct bpf_insn *insn)
18931 return insn->code == (BPF_JMP | BPF_CALL) &&
18932 insn->src_reg == 0 &&
18933 insn->imm == BPF_FUNC_loop;
18936 /* For all sub-programs in the program (including main) check
18937 * insn_aux_data to see if there are bpf_loop calls that require
18938 * inlining. If such calls are found the calls are replaced with a
18939 * sequence of instructions produced by `inline_bpf_loop` function and
18940 * subprog stack_depth is increased by the size of 3 registers.
18941 * This stack space is used to spill values of the R6, R7, R8. These
18942 * registers are used to store the loop bound, counter and context
18945 static int optimize_bpf_loop(struct bpf_verifier_env *env)
18947 struct bpf_subprog_info *subprogs = env->subprog_info;
18948 int i, cur_subprog = 0, cnt, delta = 0;
18949 struct bpf_insn *insn = env->prog->insnsi;
18950 int insn_cnt = env->prog->len;
18951 u16 stack_depth = subprogs[cur_subprog].stack_depth;
18952 u16 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
18953 u16 stack_depth_extra = 0;
18955 for (i = 0; i < insn_cnt; i++, insn++) {
18956 struct bpf_loop_inline_state *inline_state =
18957 &env->insn_aux_data[i + delta].loop_inline_state;
18959 if (is_bpf_loop_call(insn) && inline_state->fit_for_inline) {
18960 struct bpf_prog *new_prog;
18962 stack_depth_extra = BPF_REG_SIZE * 3 + stack_depth_roundup;
18963 new_prog = inline_bpf_loop(env,
18965 -(stack_depth + stack_depth_extra),
18966 inline_state->callback_subprogno,
18972 env->prog = new_prog;
18973 insn = new_prog->insnsi + i + delta;
18976 if (subprogs[cur_subprog + 1].start == i + delta + 1) {
18977 subprogs[cur_subprog].stack_depth += stack_depth_extra;
18979 stack_depth = subprogs[cur_subprog].stack_depth;
18980 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
18981 stack_depth_extra = 0;
18985 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
18990 static void free_states(struct bpf_verifier_env *env)
18992 struct bpf_verifier_state_list *sl, *sln;
18995 sl = env->free_list;
18998 free_verifier_state(&sl->state, false);
19002 env->free_list = NULL;
19004 if (!env->explored_states)
19007 for (i = 0; i < state_htab_size(env); i++) {
19008 sl = env->explored_states[i];
19012 free_verifier_state(&sl->state, false);
19016 env->explored_states[i] = NULL;
19020 static int do_check_common(struct bpf_verifier_env *env, int subprog)
19022 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
19023 struct bpf_verifier_state *state;
19024 struct bpf_reg_state *regs;
19027 env->prev_linfo = NULL;
19030 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
19033 state->curframe = 0;
19034 state->speculative = false;
19035 state->branches = 1;
19036 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
19037 if (!state->frame[0]) {
19041 env->cur_state = state;
19042 init_func_state(env, state->frame[0],
19043 BPF_MAIN_FUNC /* callsite */,
19046 state->first_insn_idx = env->subprog_info[subprog].start;
19047 state->last_insn_idx = -1;
19049 regs = state->frame[state->curframe]->regs;
19050 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
19051 ret = btf_prepare_func_args(env, subprog, regs);
19054 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
19055 if (regs[i].type == PTR_TO_CTX)
19056 mark_reg_known_zero(env, regs, i);
19057 else if (regs[i].type == SCALAR_VALUE)
19058 mark_reg_unknown(env, regs, i);
19059 else if (base_type(regs[i].type) == PTR_TO_MEM) {
19060 const u32 mem_size = regs[i].mem_size;
19062 mark_reg_known_zero(env, regs, i);
19063 regs[i].mem_size = mem_size;
19064 regs[i].id = ++env->id_gen;
19068 /* 1st arg to a function */
19069 regs[BPF_REG_1].type = PTR_TO_CTX;
19070 mark_reg_known_zero(env, regs, BPF_REG_1);
19071 ret = btf_check_subprog_arg_match(env, subprog, regs);
19072 if (ret == -EFAULT)
19073 /* unlikely verifier bug. abort.
19074 * ret == 0 and ret < 0 are sadly acceptable for
19075 * main() function due to backward compatibility.
19076 * Like socket filter program may be written as:
19077 * int bpf_prog(struct pt_regs *ctx)
19078 * and never dereference that ctx in the program.
19079 * 'struct pt_regs' is a type mismatch for socket
19080 * filter that should be using 'struct __sk_buff'.
19085 ret = do_check(env);
19087 /* check for NULL is necessary, since cur_state can be freed inside
19088 * do_check() under memory pressure.
19090 if (env->cur_state) {
19091 free_verifier_state(env->cur_state, true);
19092 env->cur_state = NULL;
19094 while (!pop_stack(env, NULL, NULL, false));
19095 if (!ret && pop_log)
19096 bpf_vlog_reset(&env->log, 0);
19101 /* Verify all global functions in a BPF program one by one based on their BTF.
19102 * All global functions must pass verification. Otherwise the whole program is rejected.
19113 * foo() will be verified first for R1=any_scalar_value. During verification it
19114 * will be assumed that bar() already verified successfully and call to bar()
19115 * from foo() will be checked for type match only. Later bar() will be verified
19116 * independently to check that it's safe for R1=any_scalar_value.
19118 static int do_check_subprogs(struct bpf_verifier_env *env)
19120 struct bpf_prog_aux *aux = env->prog->aux;
19123 if (!aux->func_info)
19126 for (i = 1; i < env->subprog_cnt; i++) {
19127 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
19129 env->insn_idx = env->subprog_info[i].start;
19130 WARN_ON_ONCE(env->insn_idx == 0);
19131 ret = do_check_common(env, i);
19134 } else if (env->log.level & BPF_LOG_LEVEL) {
19136 "Func#%d is safe for any args that match its prototype\n",
19143 static int do_check_main(struct bpf_verifier_env *env)
19148 ret = do_check_common(env, 0);
19150 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
19155 static void print_verification_stats(struct bpf_verifier_env *env)
19159 if (env->log.level & BPF_LOG_STATS) {
19160 verbose(env, "verification time %lld usec\n",
19161 div_u64(env->verification_time, 1000));
19162 verbose(env, "stack depth ");
19163 for (i = 0; i < env->subprog_cnt; i++) {
19164 u32 depth = env->subprog_info[i].stack_depth;
19166 verbose(env, "%d", depth);
19167 if (i + 1 < env->subprog_cnt)
19170 verbose(env, "\n");
19172 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
19173 "total_states %d peak_states %d mark_read %d\n",
19174 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
19175 env->max_states_per_insn, env->total_states,
19176 env->peak_states, env->longest_mark_read_walk);
19179 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
19181 const struct btf_type *t, *func_proto;
19182 const struct bpf_struct_ops *st_ops;
19183 const struct btf_member *member;
19184 struct bpf_prog *prog = env->prog;
19185 u32 btf_id, member_idx;
19188 if (!prog->gpl_compatible) {
19189 verbose(env, "struct ops programs must have a GPL compatible license\n");
19193 btf_id = prog->aux->attach_btf_id;
19194 st_ops = bpf_struct_ops_find(btf_id);
19196 verbose(env, "attach_btf_id %u is not a supported struct\n",
19202 member_idx = prog->expected_attach_type;
19203 if (member_idx >= btf_type_vlen(t)) {
19204 verbose(env, "attach to invalid member idx %u of struct %s\n",
19205 member_idx, st_ops->name);
19209 member = &btf_type_member(t)[member_idx];
19210 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
19211 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
19214 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
19215 mname, member_idx, st_ops->name);
19219 if (st_ops->check_member) {
19220 int err = st_ops->check_member(t, member, prog);
19223 verbose(env, "attach to unsupported member %s of struct %s\n",
19224 mname, st_ops->name);
19229 prog->aux->attach_func_proto = func_proto;
19230 prog->aux->attach_func_name = mname;
19231 env->ops = st_ops->verifier_ops;
19235 #define SECURITY_PREFIX "security_"
19237 static int check_attach_modify_return(unsigned long addr, const char *func_name)
19239 if (within_error_injection_list(addr) ||
19240 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
19246 /* list of non-sleepable functions that are otherwise on
19247 * ALLOW_ERROR_INJECTION list
19249 BTF_SET_START(btf_non_sleepable_error_inject)
19250 /* Three functions below can be called from sleepable and non-sleepable context.
19251 * Assume non-sleepable from bpf safety point of view.
19253 BTF_ID(func, __filemap_add_folio)
19254 BTF_ID(func, should_fail_alloc_page)
19255 BTF_ID(func, should_failslab)
19256 BTF_SET_END(btf_non_sleepable_error_inject)
19258 static int check_non_sleepable_error_inject(u32 btf_id)
19260 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
19263 int bpf_check_attach_target(struct bpf_verifier_log *log,
19264 const struct bpf_prog *prog,
19265 const struct bpf_prog *tgt_prog,
19267 struct bpf_attach_target_info *tgt_info)
19269 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
19270 const char prefix[] = "btf_trace_";
19271 int ret = 0, subprog = -1, i;
19272 const struct btf_type *t;
19273 bool conservative = true;
19277 struct module *mod = NULL;
19280 bpf_log(log, "Tracing programs must provide btf_id\n");
19283 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
19286 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
19289 t = btf_type_by_id(btf, btf_id);
19291 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
19294 tname = btf_name_by_offset(btf, t->name_off);
19296 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
19300 struct bpf_prog_aux *aux = tgt_prog->aux;
19302 if (bpf_prog_is_dev_bound(prog->aux) &&
19303 !bpf_prog_dev_bound_match(prog, tgt_prog)) {
19304 bpf_log(log, "Target program bound device mismatch");
19308 for (i = 0; i < aux->func_info_cnt; i++)
19309 if (aux->func_info[i].type_id == btf_id) {
19313 if (subprog == -1) {
19314 bpf_log(log, "Subprog %s doesn't exist\n", tname);
19317 conservative = aux->func_info_aux[subprog].unreliable;
19318 if (prog_extension) {
19319 if (conservative) {
19321 "Cannot replace static functions\n");
19324 if (!prog->jit_requested) {
19326 "Extension programs should be JITed\n");
19330 if (!tgt_prog->jited) {
19331 bpf_log(log, "Can attach to only JITed progs\n");
19334 if (tgt_prog->type == prog->type) {
19335 /* Cannot fentry/fexit another fentry/fexit program.
19336 * Cannot attach program extension to another extension.
19337 * It's ok to attach fentry/fexit to extension program.
19339 bpf_log(log, "Cannot recursively attach\n");
19342 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
19344 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
19345 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
19346 /* Program extensions can extend all program types
19347 * except fentry/fexit. The reason is the following.
19348 * The fentry/fexit programs are used for performance
19349 * analysis, stats and can be attached to any program
19350 * type except themselves. When extension program is
19351 * replacing XDP function it is necessary to allow
19352 * performance analysis of all functions. Both original
19353 * XDP program and its program extension. Hence
19354 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
19355 * allowed. If extending of fentry/fexit was allowed it
19356 * would be possible to create long call chain
19357 * fentry->extension->fentry->extension beyond
19358 * reasonable stack size. Hence extending fentry is not
19361 bpf_log(log, "Cannot extend fentry/fexit\n");
19365 if (prog_extension) {
19366 bpf_log(log, "Cannot replace kernel functions\n");
19371 switch (prog->expected_attach_type) {
19372 case BPF_TRACE_RAW_TP:
19375 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
19378 if (!btf_type_is_typedef(t)) {
19379 bpf_log(log, "attach_btf_id %u is not a typedef\n",
19383 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
19384 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
19388 tname += sizeof(prefix) - 1;
19389 t = btf_type_by_id(btf, t->type);
19390 if (!btf_type_is_ptr(t))
19391 /* should never happen in valid vmlinux build */
19393 t = btf_type_by_id(btf, t->type);
19394 if (!btf_type_is_func_proto(t))
19395 /* should never happen in valid vmlinux build */
19399 case BPF_TRACE_ITER:
19400 if (!btf_type_is_func(t)) {
19401 bpf_log(log, "attach_btf_id %u is not a function\n",
19405 t = btf_type_by_id(btf, t->type);
19406 if (!btf_type_is_func_proto(t))
19408 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
19413 if (!prog_extension)
19416 case BPF_MODIFY_RETURN:
19418 case BPF_LSM_CGROUP:
19419 case BPF_TRACE_FENTRY:
19420 case BPF_TRACE_FEXIT:
19421 if (!btf_type_is_func(t)) {
19422 bpf_log(log, "attach_btf_id %u is not a function\n",
19426 if (prog_extension &&
19427 btf_check_type_match(log, prog, btf, t))
19429 t = btf_type_by_id(btf, t->type);
19430 if (!btf_type_is_func_proto(t))
19433 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
19434 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
19435 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
19438 if (tgt_prog && conservative)
19441 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
19447 addr = (long) tgt_prog->bpf_func;
19449 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
19451 if (btf_is_module(btf)) {
19452 mod = btf_try_get_module(btf);
19454 addr = find_kallsyms_symbol_value(mod, tname);
19458 addr = kallsyms_lookup_name(tname);
19463 "The address of function %s cannot be found\n",
19469 if (prog->aux->sleepable) {
19471 switch (prog->type) {
19472 case BPF_PROG_TYPE_TRACING:
19474 /* fentry/fexit/fmod_ret progs can be sleepable if they are
19475 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
19477 if (!check_non_sleepable_error_inject(btf_id) &&
19478 within_error_injection_list(addr))
19480 /* fentry/fexit/fmod_ret progs can also be sleepable if they are
19481 * in the fmodret id set with the KF_SLEEPABLE flag.
19484 u32 *flags = btf_kfunc_is_modify_return(btf, btf_id,
19487 if (flags && (*flags & KF_SLEEPABLE))
19491 case BPF_PROG_TYPE_LSM:
19492 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
19493 * Only some of them are sleepable.
19495 if (bpf_lsm_is_sleepable_hook(btf_id))
19503 bpf_log(log, "%s is not sleepable\n", tname);
19506 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
19509 bpf_log(log, "can't modify return codes of BPF programs\n");
19513 if (btf_kfunc_is_modify_return(btf, btf_id, prog) ||
19514 !check_attach_modify_return(addr, tname))
19518 bpf_log(log, "%s() is not modifiable\n", tname);
19525 tgt_info->tgt_addr = addr;
19526 tgt_info->tgt_name = tname;
19527 tgt_info->tgt_type = t;
19528 tgt_info->tgt_mod = mod;
19532 BTF_SET_START(btf_id_deny)
19535 BTF_ID(func, migrate_disable)
19536 BTF_ID(func, migrate_enable)
19538 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
19539 BTF_ID(func, rcu_read_unlock_strict)
19541 #if defined(CONFIG_DEBUG_PREEMPT) || defined(CONFIG_TRACE_PREEMPT_TOGGLE)
19542 BTF_ID(func, preempt_count_add)
19543 BTF_ID(func, preempt_count_sub)
19545 #ifdef CONFIG_PREEMPT_RCU
19546 BTF_ID(func, __rcu_read_lock)
19547 BTF_ID(func, __rcu_read_unlock)
19549 BTF_SET_END(btf_id_deny)
19551 static bool can_be_sleepable(struct bpf_prog *prog)
19553 if (prog->type == BPF_PROG_TYPE_TRACING) {
19554 switch (prog->expected_attach_type) {
19555 case BPF_TRACE_FENTRY:
19556 case BPF_TRACE_FEXIT:
19557 case BPF_MODIFY_RETURN:
19558 case BPF_TRACE_ITER:
19564 return prog->type == BPF_PROG_TYPE_LSM ||
19565 prog->type == BPF_PROG_TYPE_KPROBE /* only for uprobes */ ||
19566 prog->type == BPF_PROG_TYPE_STRUCT_OPS;
19569 static int check_attach_btf_id(struct bpf_verifier_env *env)
19571 struct bpf_prog *prog = env->prog;
19572 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
19573 struct bpf_attach_target_info tgt_info = {};
19574 u32 btf_id = prog->aux->attach_btf_id;
19575 struct bpf_trampoline *tr;
19579 if (prog->type == BPF_PROG_TYPE_SYSCALL) {
19580 if (prog->aux->sleepable)
19581 /* attach_btf_id checked to be zero already */
19583 verbose(env, "Syscall programs can only be sleepable\n");
19587 if (prog->aux->sleepable && !can_be_sleepable(prog)) {
19588 verbose(env, "Only fentry/fexit/fmod_ret, lsm, iter, uprobe, and struct_ops programs can be sleepable\n");
19592 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
19593 return check_struct_ops_btf_id(env);
19595 if (prog->type != BPF_PROG_TYPE_TRACING &&
19596 prog->type != BPF_PROG_TYPE_LSM &&
19597 prog->type != BPF_PROG_TYPE_EXT)
19600 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
19604 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
19605 /* to make freplace equivalent to their targets, they need to
19606 * inherit env->ops and expected_attach_type for the rest of the
19609 env->ops = bpf_verifier_ops[tgt_prog->type];
19610 prog->expected_attach_type = tgt_prog->expected_attach_type;
19613 /* store info about the attachment target that will be used later */
19614 prog->aux->attach_func_proto = tgt_info.tgt_type;
19615 prog->aux->attach_func_name = tgt_info.tgt_name;
19616 prog->aux->mod = tgt_info.tgt_mod;
19619 prog->aux->saved_dst_prog_type = tgt_prog->type;
19620 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
19623 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
19624 prog->aux->attach_btf_trace = true;
19626 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
19627 if (!bpf_iter_prog_supported(prog))
19632 if (prog->type == BPF_PROG_TYPE_LSM) {
19633 ret = bpf_lsm_verify_prog(&env->log, prog);
19636 } else if (prog->type == BPF_PROG_TYPE_TRACING &&
19637 btf_id_set_contains(&btf_id_deny, btf_id)) {
19641 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
19642 tr = bpf_trampoline_get(key, &tgt_info);
19646 prog->aux->dst_trampoline = tr;
19650 struct btf *bpf_get_btf_vmlinux(void)
19652 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
19653 mutex_lock(&bpf_verifier_lock);
19655 btf_vmlinux = btf_parse_vmlinux();
19656 mutex_unlock(&bpf_verifier_lock);
19658 return btf_vmlinux;
19661 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr, __u32 uattr_size)
19663 u64 start_time = ktime_get_ns();
19664 struct bpf_verifier_env *env;
19665 int i, len, ret = -EINVAL, err;
19669 /* no program is valid */
19670 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
19673 /* 'struct bpf_verifier_env' can be global, but since it's not small,
19674 * allocate/free it every time bpf_check() is called
19676 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
19682 len = (*prog)->len;
19683 env->insn_aux_data =
19684 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
19686 if (!env->insn_aux_data)
19688 for (i = 0; i < len; i++)
19689 env->insn_aux_data[i].orig_idx = i;
19691 env->ops = bpf_verifier_ops[env->prog->type];
19692 env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
19693 is_priv = bpf_capable();
19695 bpf_get_btf_vmlinux();
19697 /* grab the mutex to protect few globals used by verifier */
19699 mutex_lock(&bpf_verifier_lock);
19701 /* user could have requested verbose verifier output
19702 * and supplied buffer to store the verification trace
19704 ret = bpf_vlog_init(&env->log, attr->log_level,
19705 (char __user *) (unsigned long) attr->log_buf,
19710 mark_verifier_state_clean(env);
19712 if (IS_ERR(btf_vmlinux)) {
19713 /* Either gcc or pahole or kernel are broken. */
19714 verbose(env, "in-kernel BTF is malformed\n");
19715 ret = PTR_ERR(btf_vmlinux);
19716 goto skip_full_check;
19719 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
19720 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
19721 env->strict_alignment = true;
19722 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
19723 env->strict_alignment = false;
19725 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
19726 env->allow_uninit_stack = bpf_allow_uninit_stack();
19727 env->bypass_spec_v1 = bpf_bypass_spec_v1();
19728 env->bypass_spec_v4 = bpf_bypass_spec_v4();
19729 env->bpf_capable = bpf_capable();
19732 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
19734 env->explored_states = kvcalloc(state_htab_size(env),
19735 sizeof(struct bpf_verifier_state_list *),
19738 if (!env->explored_states)
19739 goto skip_full_check;
19741 ret = add_subprog_and_kfunc(env);
19743 goto skip_full_check;
19745 ret = check_subprogs(env);
19747 goto skip_full_check;
19749 ret = check_btf_info(env, attr, uattr);
19751 goto skip_full_check;
19753 ret = check_attach_btf_id(env);
19755 goto skip_full_check;
19757 ret = resolve_pseudo_ldimm64(env);
19759 goto skip_full_check;
19761 if (bpf_prog_is_offloaded(env->prog->aux)) {
19762 ret = bpf_prog_offload_verifier_prep(env->prog);
19764 goto skip_full_check;
19767 ret = check_cfg(env);
19769 goto skip_full_check;
19771 ret = do_check_subprogs(env);
19772 ret = ret ?: do_check_main(env);
19774 if (ret == 0 && bpf_prog_is_offloaded(env->prog->aux))
19775 ret = bpf_prog_offload_finalize(env);
19778 kvfree(env->explored_states);
19781 ret = check_max_stack_depth(env);
19783 /* instruction rewrites happen after this point */
19785 ret = optimize_bpf_loop(env);
19789 opt_hard_wire_dead_code_branches(env);
19791 ret = opt_remove_dead_code(env);
19793 ret = opt_remove_nops(env);
19796 sanitize_dead_code(env);
19800 /* program is valid, convert *(u32*)(ctx + off) accesses */
19801 ret = convert_ctx_accesses(env);
19804 ret = do_misc_fixups(env);
19806 /* do 32-bit optimization after insn patching has done so those patched
19807 * insns could be handled correctly.
19809 if (ret == 0 && !bpf_prog_is_offloaded(env->prog->aux)) {
19810 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
19811 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
19816 ret = fixup_call_args(env);
19818 env->verification_time = ktime_get_ns() - start_time;
19819 print_verification_stats(env);
19820 env->prog->aux->verified_insns = env->insn_processed;
19822 /* preserve original error even if log finalization is successful */
19823 err = bpf_vlog_finalize(&env->log, &log_true_size);
19827 if (uattr_size >= offsetofend(union bpf_attr, log_true_size) &&
19828 copy_to_bpfptr_offset(uattr, offsetof(union bpf_attr, log_true_size),
19829 &log_true_size, sizeof(log_true_size))) {
19831 goto err_release_maps;
19835 goto err_release_maps;
19837 if (env->used_map_cnt) {
19838 /* if program passed verifier, update used_maps in bpf_prog_info */
19839 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
19840 sizeof(env->used_maps[0]),
19843 if (!env->prog->aux->used_maps) {
19845 goto err_release_maps;
19848 memcpy(env->prog->aux->used_maps, env->used_maps,
19849 sizeof(env->used_maps[0]) * env->used_map_cnt);
19850 env->prog->aux->used_map_cnt = env->used_map_cnt;
19852 if (env->used_btf_cnt) {
19853 /* if program passed verifier, update used_btfs in bpf_prog_aux */
19854 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
19855 sizeof(env->used_btfs[0]),
19857 if (!env->prog->aux->used_btfs) {
19859 goto err_release_maps;
19862 memcpy(env->prog->aux->used_btfs, env->used_btfs,
19863 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
19864 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
19866 if (env->used_map_cnt || env->used_btf_cnt) {
19867 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
19868 * bpf_ld_imm64 instructions
19870 convert_pseudo_ld_imm64(env);
19873 adjust_btf_func(env);
19876 if (!env->prog->aux->used_maps)
19877 /* if we didn't copy map pointers into bpf_prog_info, release
19878 * them now. Otherwise free_used_maps() will release them.
19881 if (!env->prog->aux->used_btfs)
19884 /* extension progs temporarily inherit the attach_type of their targets
19885 for verification purposes, so set it back to zero before returning
19887 if (env->prog->type == BPF_PROG_TYPE_EXT)
19888 env->prog->expected_attach_type = 0;
19893 mutex_unlock(&bpf_verifier_lock);
19894 vfree(env->insn_aux_data);