1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
3 * Copyright (c) 2016 Facebook
4 * Copyright (c) 2018 Covalent IO, Inc. http://covalent.io
6 #include <uapi/linux/btf.h>
7 #include <linux/bpf-cgroup.h>
8 #include <linux/kernel.h>
9 #include <linux/types.h>
10 #include <linux/slab.h>
11 #include <linux/bpf.h>
12 #include <linux/btf.h>
13 #include <linux/bpf_verifier.h>
14 #include <linux/filter.h>
15 #include <net/netlink.h>
16 #include <linux/file.h>
17 #include <linux/vmalloc.h>
18 #include <linux/stringify.h>
19 #include <linux/bsearch.h>
20 #include <linux/sort.h>
21 #include <linux/perf_event.h>
22 #include <linux/ctype.h>
23 #include <linux/error-injection.h>
24 #include <linux/bpf_lsm.h>
25 #include <linux/btf_ids.h>
26 #include <linux/poison.h>
27 #include <linux/module.h>
28 #include <linux/cpumask.h>
33 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
34 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
35 [_id] = & _name ## _verifier_ops,
36 #define BPF_MAP_TYPE(_id, _ops)
37 #define BPF_LINK_TYPE(_id, _name)
38 #include <linux/bpf_types.h>
44 /* bpf_check() is a static code analyzer that walks eBPF program
45 * instruction by instruction and updates register/stack state.
46 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
48 * The first pass is depth-first-search to check that the program is a DAG.
49 * It rejects the following programs:
50 * - larger than BPF_MAXINSNS insns
51 * - if loop is present (detected via back-edge)
52 * - unreachable insns exist (shouldn't be a forest. program = one function)
53 * - out of bounds or malformed jumps
54 * The second pass is all possible path descent from the 1st insn.
55 * Since it's analyzing all paths through the program, the length of the
56 * analysis is limited to 64k insn, which may be hit even if total number of
57 * insn is less then 4K, but there are too many branches that change stack/regs.
58 * Number of 'branches to be analyzed' is limited to 1k
60 * On entry to each instruction, each register has a type, and the instruction
61 * changes the types of the registers depending on instruction semantics.
62 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
65 * All registers are 64-bit.
66 * R0 - return register
67 * R1-R5 argument passing registers
68 * R6-R9 callee saved registers
69 * R10 - frame pointer read-only
71 * At the start of BPF program the register R1 contains a pointer to bpf_context
72 * and has type PTR_TO_CTX.
74 * Verifier tracks arithmetic operations on pointers in case:
75 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
76 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
77 * 1st insn copies R10 (which has FRAME_PTR) type into R1
78 * and 2nd arithmetic instruction is pattern matched to recognize
79 * that it wants to construct a pointer to some element within stack.
80 * So after 2nd insn, the register R1 has type PTR_TO_STACK
81 * (and -20 constant is saved for further stack bounds checking).
82 * Meaning that this reg is a pointer to stack plus known immediate constant.
84 * Most of the time the registers have SCALAR_VALUE type, which
85 * means the register has some value, but it's not a valid pointer.
86 * (like pointer plus pointer becomes SCALAR_VALUE type)
88 * When verifier sees load or store instructions the type of base register
89 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
90 * four pointer types recognized by check_mem_access() function.
92 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
93 * and the range of [ptr, ptr + map's value_size) is accessible.
95 * registers used to pass values to function calls are checked against
96 * function argument constraints.
98 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
99 * It means that the register type passed to this function must be
100 * PTR_TO_STACK and it will be used inside the function as
101 * 'pointer to map element key'
103 * For example the argument constraints for bpf_map_lookup_elem():
104 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
105 * .arg1_type = ARG_CONST_MAP_PTR,
106 * .arg2_type = ARG_PTR_TO_MAP_KEY,
108 * ret_type says that this function returns 'pointer to map elem value or null'
109 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
110 * 2nd argument should be a pointer to stack, which will be used inside
111 * the helper function as a pointer to map element key.
113 * On the kernel side the helper function looks like:
114 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
116 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
117 * void *key = (void *) (unsigned long) r2;
120 * here kernel can access 'key' and 'map' pointers safely, knowing that
121 * [key, key + map->key_size) bytes are valid and were initialized on
122 * the stack of eBPF program.
125 * Corresponding eBPF program may look like:
126 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
127 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
128 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
129 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
130 * here verifier looks at prototype of map_lookup_elem() and sees:
131 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
132 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
134 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
135 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
136 * and were initialized prior to this call.
137 * If it's ok, then verifier allows this BPF_CALL insn and looks at
138 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
139 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
140 * returns either pointer to map value or NULL.
142 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
143 * insn, the register holding that pointer in the true branch changes state to
144 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
145 * branch. See check_cond_jmp_op().
147 * After the call R0 is set to return type of the function and registers R1-R5
148 * are set to NOT_INIT to indicate that they are no longer readable.
150 * The following reference types represent a potential reference to a kernel
151 * resource which, after first being allocated, must be checked and freed by
153 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
155 * When the verifier sees a helper call return a reference type, it allocates a
156 * pointer id for the reference and stores it in the current function state.
157 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
158 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
159 * passes through a NULL-check conditional. For the branch wherein the state is
160 * changed to CONST_IMM, the verifier releases the reference.
162 * For each helper function that allocates a reference, such as
163 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
164 * bpf_sk_release(). When a reference type passes into the release function,
165 * the verifier also releases the reference. If any unchecked or unreleased
166 * reference remains at the end of the program, the verifier rejects it.
169 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
170 struct bpf_verifier_stack_elem {
171 /* verifer state is 'st'
172 * before processing instruction 'insn_idx'
173 * and after processing instruction 'prev_insn_idx'
175 struct bpf_verifier_state st;
178 struct bpf_verifier_stack_elem *next;
179 /* length of verifier log at the time this state was pushed on stack */
183 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
184 #define BPF_COMPLEXITY_LIMIT_STATES 64
186 #define BPF_MAP_KEY_POISON (1ULL << 63)
187 #define BPF_MAP_KEY_SEEN (1ULL << 62)
189 #define BPF_MAP_PTR_UNPRIV 1UL
190 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
191 POISON_POINTER_DELTA))
192 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
194 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx);
195 static int release_reference(struct bpf_verifier_env *env, int ref_obj_id);
196 static void invalidate_non_owning_refs(struct bpf_verifier_env *env);
197 static bool in_rbtree_lock_required_cb(struct bpf_verifier_env *env);
198 static int ref_set_non_owning(struct bpf_verifier_env *env,
199 struct bpf_reg_state *reg);
200 static void specialize_kfunc(struct bpf_verifier_env *env,
201 u32 func_id, u16 offset, unsigned long *addr);
202 static bool is_trusted_reg(const struct bpf_reg_state *reg);
204 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
206 return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON;
209 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
211 return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV;
214 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
215 const struct bpf_map *map, bool unpriv)
217 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
218 unpriv |= bpf_map_ptr_unpriv(aux);
219 aux->map_ptr_state = (unsigned long)map |
220 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
223 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
225 return aux->map_key_state & BPF_MAP_KEY_POISON;
228 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
230 return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
233 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
235 return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
238 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
240 bool poisoned = bpf_map_key_poisoned(aux);
242 aux->map_key_state = state | BPF_MAP_KEY_SEEN |
243 (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
246 static bool bpf_helper_call(const struct bpf_insn *insn)
248 return insn->code == (BPF_JMP | BPF_CALL) &&
252 static bool bpf_pseudo_call(const struct bpf_insn *insn)
254 return insn->code == (BPF_JMP | BPF_CALL) &&
255 insn->src_reg == BPF_PSEUDO_CALL;
258 static bool bpf_pseudo_kfunc_call(const struct bpf_insn *insn)
260 return insn->code == (BPF_JMP | BPF_CALL) &&
261 insn->src_reg == BPF_PSEUDO_KFUNC_CALL;
264 struct bpf_call_arg_meta {
265 struct bpf_map *map_ptr;
282 struct btf_field *kptr_field;
285 struct bpf_kfunc_call_arg_meta {
290 const struct btf_type *func_proto;
291 const char *func_name;
304 /* arg_{btf,btf_id,owning_ref} are used by kfunc-specific handling,
305 * generally to pass info about user-defined local kptr types to later
308 * Record the local kptr type to be drop'd
309 * bpf_refcount_acquire (via KF_ARG_PTR_TO_REFCOUNTED_KPTR arg type)
310 * Record the local kptr type to be refcount_incr'd and use
311 * arg_owning_ref to determine whether refcount_acquire should be
319 struct btf_field *field;
322 struct btf_field *field;
325 enum bpf_dynptr_type type;
328 } initialized_dynptr;
336 struct btf *btf_vmlinux;
338 static DEFINE_MUTEX(bpf_verifier_lock);
340 static const struct bpf_line_info *
341 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
343 const struct bpf_line_info *linfo;
344 const struct bpf_prog *prog;
348 nr_linfo = prog->aux->nr_linfo;
350 if (!nr_linfo || insn_off >= prog->len)
353 linfo = prog->aux->linfo;
354 for (i = 1; i < nr_linfo; i++)
355 if (insn_off < linfo[i].insn_off)
358 return &linfo[i - 1];
361 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
363 struct bpf_verifier_env *env = private_data;
366 if (!bpf_verifier_log_needed(&env->log))
370 bpf_verifier_vlog(&env->log, fmt, args);
374 static const char *ltrim(const char *s)
382 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
384 const char *prefix_fmt, ...)
386 const struct bpf_line_info *linfo;
388 if (!bpf_verifier_log_needed(&env->log))
391 linfo = find_linfo(env, insn_off);
392 if (!linfo || linfo == env->prev_linfo)
398 va_start(args, prefix_fmt);
399 bpf_verifier_vlog(&env->log, prefix_fmt, args);
404 ltrim(btf_name_by_offset(env->prog->aux->btf,
407 env->prev_linfo = linfo;
410 static void verbose_invalid_scalar(struct bpf_verifier_env *env,
411 struct bpf_reg_state *reg,
412 struct tnum *range, const char *ctx,
413 const char *reg_name)
417 verbose(env, "At %s the register %s ", ctx, reg_name);
418 if (!tnum_is_unknown(reg->var_off)) {
419 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
420 verbose(env, "has value %s", tn_buf);
422 verbose(env, "has unknown scalar value");
424 tnum_strn(tn_buf, sizeof(tn_buf), *range);
425 verbose(env, " should have been in %s\n", tn_buf);
428 static bool type_is_pkt_pointer(enum bpf_reg_type type)
430 type = base_type(type);
431 return type == PTR_TO_PACKET ||
432 type == PTR_TO_PACKET_META;
435 static bool type_is_sk_pointer(enum bpf_reg_type type)
437 return type == PTR_TO_SOCKET ||
438 type == PTR_TO_SOCK_COMMON ||
439 type == PTR_TO_TCP_SOCK ||
440 type == PTR_TO_XDP_SOCK;
443 static bool type_may_be_null(u32 type)
445 return type & PTR_MAYBE_NULL;
448 static bool reg_not_null(const struct bpf_reg_state *reg)
450 enum bpf_reg_type type;
453 if (type_may_be_null(type))
456 type = base_type(type);
457 return type == PTR_TO_SOCKET ||
458 type == PTR_TO_TCP_SOCK ||
459 type == PTR_TO_MAP_VALUE ||
460 type == PTR_TO_MAP_KEY ||
461 type == PTR_TO_SOCK_COMMON ||
462 (type == PTR_TO_BTF_ID && is_trusted_reg(reg)) ||
466 static bool type_is_ptr_alloc_obj(u32 type)
468 return base_type(type) == PTR_TO_BTF_ID && type_flag(type) & MEM_ALLOC;
471 static bool type_is_non_owning_ref(u32 type)
473 return type_is_ptr_alloc_obj(type) && type_flag(type) & NON_OWN_REF;
476 static struct btf_record *reg_btf_record(const struct bpf_reg_state *reg)
478 struct btf_record *rec = NULL;
479 struct btf_struct_meta *meta;
481 if (reg->type == PTR_TO_MAP_VALUE) {
482 rec = reg->map_ptr->record;
483 } else if (type_is_ptr_alloc_obj(reg->type)) {
484 meta = btf_find_struct_meta(reg->btf, reg->btf_id);
491 static bool subprog_is_global(const struct bpf_verifier_env *env, int subprog)
493 struct bpf_func_info_aux *aux = env->prog->aux->func_info_aux;
495 return aux && aux[subprog].linkage == BTF_FUNC_GLOBAL;
498 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
500 return btf_record_has_field(reg_btf_record(reg), BPF_SPIN_LOCK);
503 static bool type_is_rdonly_mem(u32 type)
505 return type & MEM_RDONLY;
508 static bool is_acquire_function(enum bpf_func_id func_id,
509 const struct bpf_map *map)
511 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
513 if (func_id == BPF_FUNC_sk_lookup_tcp ||
514 func_id == BPF_FUNC_sk_lookup_udp ||
515 func_id == BPF_FUNC_skc_lookup_tcp ||
516 func_id == BPF_FUNC_ringbuf_reserve ||
517 func_id == BPF_FUNC_kptr_xchg)
520 if (func_id == BPF_FUNC_map_lookup_elem &&
521 (map_type == BPF_MAP_TYPE_SOCKMAP ||
522 map_type == BPF_MAP_TYPE_SOCKHASH))
528 static bool is_ptr_cast_function(enum bpf_func_id func_id)
530 return func_id == BPF_FUNC_tcp_sock ||
531 func_id == BPF_FUNC_sk_fullsock ||
532 func_id == BPF_FUNC_skc_to_tcp_sock ||
533 func_id == BPF_FUNC_skc_to_tcp6_sock ||
534 func_id == BPF_FUNC_skc_to_udp6_sock ||
535 func_id == BPF_FUNC_skc_to_mptcp_sock ||
536 func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
537 func_id == BPF_FUNC_skc_to_tcp_request_sock;
540 static bool is_dynptr_ref_function(enum bpf_func_id func_id)
542 return func_id == BPF_FUNC_dynptr_data;
545 static bool is_callback_calling_kfunc(u32 btf_id);
547 static bool is_callback_calling_function(enum bpf_func_id func_id)
549 return func_id == BPF_FUNC_for_each_map_elem ||
550 func_id == BPF_FUNC_timer_set_callback ||
551 func_id == BPF_FUNC_find_vma ||
552 func_id == BPF_FUNC_loop ||
553 func_id == BPF_FUNC_user_ringbuf_drain;
556 static bool is_async_callback_calling_function(enum bpf_func_id func_id)
558 return func_id == BPF_FUNC_timer_set_callback;
561 static bool is_storage_get_function(enum bpf_func_id func_id)
563 return func_id == BPF_FUNC_sk_storage_get ||
564 func_id == BPF_FUNC_inode_storage_get ||
565 func_id == BPF_FUNC_task_storage_get ||
566 func_id == BPF_FUNC_cgrp_storage_get;
569 static bool helper_multiple_ref_obj_use(enum bpf_func_id func_id,
570 const struct bpf_map *map)
572 int ref_obj_uses = 0;
574 if (is_ptr_cast_function(func_id))
576 if (is_acquire_function(func_id, map))
578 if (is_dynptr_ref_function(func_id))
581 return ref_obj_uses > 1;
584 static bool is_cmpxchg_insn(const struct bpf_insn *insn)
586 return BPF_CLASS(insn->code) == BPF_STX &&
587 BPF_MODE(insn->code) == BPF_ATOMIC &&
588 insn->imm == BPF_CMPXCHG;
591 /* string representation of 'enum bpf_reg_type'
593 * Note that reg_type_str() can not appear more than once in a single verbose()
596 static const char *reg_type_str(struct bpf_verifier_env *env,
597 enum bpf_reg_type type)
599 char postfix[16] = {0}, prefix[64] = {0};
600 static const char * const str[] = {
602 [SCALAR_VALUE] = "scalar",
603 [PTR_TO_CTX] = "ctx",
604 [CONST_PTR_TO_MAP] = "map_ptr",
605 [PTR_TO_MAP_VALUE] = "map_value",
606 [PTR_TO_STACK] = "fp",
607 [PTR_TO_PACKET] = "pkt",
608 [PTR_TO_PACKET_META] = "pkt_meta",
609 [PTR_TO_PACKET_END] = "pkt_end",
610 [PTR_TO_FLOW_KEYS] = "flow_keys",
611 [PTR_TO_SOCKET] = "sock",
612 [PTR_TO_SOCK_COMMON] = "sock_common",
613 [PTR_TO_TCP_SOCK] = "tcp_sock",
614 [PTR_TO_TP_BUFFER] = "tp_buffer",
615 [PTR_TO_XDP_SOCK] = "xdp_sock",
616 [PTR_TO_BTF_ID] = "ptr_",
617 [PTR_TO_MEM] = "mem",
618 [PTR_TO_BUF] = "buf",
619 [PTR_TO_FUNC] = "func",
620 [PTR_TO_MAP_KEY] = "map_key",
621 [CONST_PTR_TO_DYNPTR] = "dynptr_ptr",
624 if (type & PTR_MAYBE_NULL) {
625 if (base_type(type) == PTR_TO_BTF_ID)
626 strncpy(postfix, "or_null_", 16);
628 strncpy(postfix, "_or_null", 16);
631 snprintf(prefix, sizeof(prefix), "%s%s%s%s%s%s%s",
632 type & MEM_RDONLY ? "rdonly_" : "",
633 type & MEM_RINGBUF ? "ringbuf_" : "",
634 type & MEM_USER ? "user_" : "",
635 type & MEM_PERCPU ? "percpu_" : "",
636 type & MEM_RCU ? "rcu_" : "",
637 type & PTR_UNTRUSTED ? "untrusted_" : "",
638 type & PTR_TRUSTED ? "trusted_" : ""
641 snprintf(env->tmp_str_buf, TMP_STR_BUF_LEN, "%s%s%s",
642 prefix, str[base_type(type)], postfix);
643 return env->tmp_str_buf;
646 static char slot_type_char[] = {
647 [STACK_INVALID] = '?',
651 [STACK_DYNPTR] = 'd',
655 static void print_liveness(struct bpf_verifier_env *env,
656 enum bpf_reg_liveness live)
658 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
660 if (live & REG_LIVE_READ)
662 if (live & REG_LIVE_WRITTEN)
664 if (live & REG_LIVE_DONE)
668 static int __get_spi(s32 off)
670 return (-off - 1) / BPF_REG_SIZE;
673 static struct bpf_func_state *func(struct bpf_verifier_env *env,
674 const struct bpf_reg_state *reg)
676 struct bpf_verifier_state *cur = env->cur_state;
678 return cur->frame[reg->frameno];
681 static bool is_spi_bounds_valid(struct bpf_func_state *state, int spi, int nr_slots)
683 int allocated_slots = state->allocated_stack / BPF_REG_SIZE;
685 /* We need to check that slots between [spi - nr_slots + 1, spi] are
686 * within [0, allocated_stack).
688 * Please note that the spi grows downwards. For example, a dynptr
689 * takes the size of two stack slots; the first slot will be at
690 * spi and the second slot will be at spi - 1.
692 return spi - nr_slots + 1 >= 0 && spi < allocated_slots;
695 static int stack_slot_obj_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
696 const char *obj_kind, int nr_slots)
700 if (!tnum_is_const(reg->var_off)) {
701 verbose(env, "%s has to be at a constant offset\n", obj_kind);
705 off = reg->off + reg->var_off.value;
706 if (off % BPF_REG_SIZE) {
707 verbose(env, "cannot pass in %s at an offset=%d\n", obj_kind, off);
711 spi = __get_spi(off);
712 if (spi + 1 < nr_slots) {
713 verbose(env, "cannot pass in %s at an offset=%d\n", obj_kind, off);
717 if (!is_spi_bounds_valid(func(env, reg), spi, nr_slots))
722 static int dynptr_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
724 return stack_slot_obj_get_spi(env, reg, "dynptr", BPF_DYNPTR_NR_SLOTS);
727 static int iter_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg, int nr_slots)
729 return stack_slot_obj_get_spi(env, reg, "iter", nr_slots);
732 static const char *btf_type_name(const struct btf *btf, u32 id)
734 return btf_name_by_offset(btf, btf_type_by_id(btf, id)->name_off);
737 static const char *dynptr_type_str(enum bpf_dynptr_type type)
740 case BPF_DYNPTR_TYPE_LOCAL:
742 case BPF_DYNPTR_TYPE_RINGBUF:
744 case BPF_DYNPTR_TYPE_SKB:
746 case BPF_DYNPTR_TYPE_XDP:
748 case BPF_DYNPTR_TYPE_INVALID:
751 WARN_ONCE(1, "unknown dynptr type %d\n", type);
756 static const char *iter_type_str(const struct btf *btf, u32 btf_id)
758 if (!btf || btf_id == 0)
761 /* we already validated that type is valid and has conforming name */
762 return btf_type_name(btf, btf_id) + sizeof(ITER_PREFIX) - 1;
765 static const char *iter_state_str(enum bpf_iter_state state)
768 case BPF_ITER_STATE_ACTIVE:
770 case BPF_ITER_STATE_DRAINED:
772 case BPF_ITER_STATE_INVALID:
775 WARN_ONCE(1, "unknown iter state %d\n", state);
780 static void mark_reg_scratched(struct bpf_verifier_env *env, u32 regno)
782 env->scratched_regs |= 1U << regno;
785 static void mark_stack_slot_scratched(struct bpf_verifier_env *env, u32 spi)
787 env->scratched_stack_slots |= 1ULL << spi;
790 static bool reg_scratched(const struct bpf_verifier_env *env, u32 regno)
792 return (env->scratched_regs >> regno) & 1;
795 static bool stack_slot_scratched(const struct bpf_verifier_env *env, u64 regno)
797 return (env->scratched_stack_slots >> regno) & 1;
800 static bool verifier_state_scratched(const struct bpf_verifier_env *env)
802 return env->scratched_regs || env->scratched_stack_slots;
805 static void mark_verifier_state_clean(struct bpf_verifier_env *env)
807 env->scratched_regs = 0U;
808 env->scratched_stack_slots = 0ULL;
811 /* Used for printing the entire verifier state. */
812 static void mark_verifier_state_scratched(struct bpf_verifier_env *env)
814 env->scratched_regs = ~0U;
815 env->scratched_stack_slots = ~0ULL;
818 static enum bpf_dynptr_type arg_to_dynptr_type(enum bpf_arg_type arg_type)
820 switch (arg_type & DYNPTR_TYPE_FLAG_MASK) {
821 case DYNPTR_TYPE_LOCAL:
822 return BPF_DYNPTR_TYPE_LOCAL;
823 case DYNPTR_TYPE_RINGBUF:
824 return BPF_DYNPTR_TYPE_RINGBUF;
825 case DYNPTR_TYPE_SKB:
826 return BPF_DYNPTR_TYPE_SKB;
827 case DYNPTR_TYPE_XDP:
828 return BPF_DYNPTR_TYPE_XDP;
830 return BPF_DYNPTR_TYPE_INVALID;
834 static enum bpf_type_flag get_dynptr_type_flag(enum bpf_dynptr_type type)
837 case BPF_DYNPTR_TYPE_LOCAL:
838 return DYNPTR_TYPE_LOCAL;
839 case BPF_DYNPTR_TYPE_RINGBUF:
840 return DYNPTR_TYPE_RINGBUF;
841 case BPF_DYNPTR_TYPE_SKB:
842 return DYNPTR_TYPE_SKB;
843 case BPF_DYNPTR_TYPE_XDP:
844 return DYNPTR_TYPE_XDP;
850 static bool dynptr_type_refcounted(enum bpf_dynptr_type type)
852 return type == BPF_DYNPTR_TYPE_RINGBUF;
855 static void __mark_dynptr_reg(struct bpf_reg_state *reg,
856 enum bpf_dynptr_type type,
857 bool first_slot, int dynptr_id);
859 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
860 struct bpf_reg_state *reg);
862 static void mark_dynptr_stack_regs(struct bpf_verifier_env *env,
863 struct bpf_reg_state *sreg1,
864 struct bpf_reg_state *sreg2,
865 enum bpf_dynptr_type type)
867 int id = ++env->id_gen;
869 __mark_dynptr_reg(sreg1, type, true, id);
870 __mark_dynptr_reg(sreg2, type, false, id);
873 static void mark_dynptr_cb_reg(struct bpf_verifier_env *env,
874 struct bpf_reg_state *reg,
875 enum bpf_dynptr_type type)
877 __mark_dynptr_reg(reg, type, true, ++env->id_gen);
880 static int destroy_if_dynptr_stack_slot(struct bpf_verifier_env *env,
881 struct bpf_func_state *state, int spi);
883 static int mark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
884 enum bpf_arg_type arg_type, int insn_idx, int clone_ref_obj_id)
886 struct bpf_func_state *state = func(env, reg);
887 enum bpf_dynptr_type type;
890 spi = dynptr_get_spi(env, reg);
894 /* We cannot assume both spi and spi - 1 belong to the same dynptr,
895 * hence we need to call destroy_if_dynptr_stack_slot twice for both,
896 * to ensure that for the following example:
899 * So marking spi = 2 should lead to destruction of both d1 and d2. In
900 * case they do belong to same dynptr, second call won't see slot_type
901 * as STACK_DYNPTR and will simply skip destruction.
903 err = destroy_if_dynptr_stack_slot(env, state, spi);
906 err = destroy_if_dynptr_stack_slot(env, state, spi - 1);
910 for (i = 0; i < BPF_REG_SIZE; i++) {
911 state->stack[spi].slot_type[i] = STACK_DYNPTR;
912 state->stack[spi - 1].slot_type[i] = STACK_DYNPTR;
915 type = arg_to_dynptr_type(arg_type);
916 if (type == BPF_DYNPTR_TYPE_INVALID)
919 mark_dynptr_stack_regs(env, &state->stack[spi].spilled_ptr,
920 &state->stack[spi - 1].spilled_ptr, type);
922 if (dynptr_type_refcounted(type)) {
923 /* The id is used to track proper releasing */
926 if (clone_ref_obj_id)
927 id = clone_ref_obj_id;
929 id = acquire_reference_state(env, insn_idx);
934 state->stack[spi].spilled_ptr.ref_obj_id = id;
935 state->stack[spi - 1].spilled_ptr.ref_obj_id = id;
938 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
939 state->stack[spi - 1].spilled_ptr.live |= REG_LIVE_WRITTEN;
944 static void invalidate_dynptr(struct bpf_verifier_env *env, struct bpf_func_state *state, int spi)
948 for (i = 0; i < BPF_REG_SIZE; i++) {
949 state->stack[spi].slot_type[i] = STACK_INVALID;
950 state->stack[spi - 1].slot_type[i] = STACK_INVALID;
953 __mark_reg_not_init(env, &state->stack[spi].spilled_ptr);
954 __mark_reg_not_init(env, &state->stack[spi - 1].spilled_ptr);
956 /* Why do we need to set REG_LIVE_WRITTEN for STACK_INVALID slot?
958 * While we don't allow reading STACK_INVALID, it is still possible to
959 * do <8 byte writes marking some but not all slots as STACK_MISC. Then,
960 * helpers or insns can do partial read of that part without failing,
961 * but check_stack_range_initialized, check_stack_read_var_off, and
962 * check_stack_read_fixed_off will do mark_reg_read for all 8-bytes of
963 * the slot conservatively. Hence we need to prevent those liveness
966 * This was not a problem before because STACK_INVALID is only set by
967 * default (where the default reg state has its reg->parent as NULL), or
968 * in clean_live_states after REG_LIVE_DONE (at which point
969 * mark_reg_read won't walk reg->parent chain), but not randomly during
970 * verifier state exploration (like we did above). Hence, for our case
971 * parentage chain will still be live (i.e. reg->parent may be
972 * non-NULL), while earlier reg->parent was NULL, so we need
973 * REG_LIVE_WRITTEN to screen off read marker propagation when it is
974 * done later on reads or by mark_dynptr_read as well to unnecessary
975 * mark registers in verifier state.
977 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
978 state->stack[spi - 1].spilled_ptr.live |= REG_LIVE_WRITTEN;
981 static int unmark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
983 struct bpf_func_state *state = func(env, reg);
984 int spi, ref_obj_id, i;
986 spi = dynptr_get_spi(env, reg);
990 if (!dynptr_type_refcounted(state->stack[spi].spilled_ptr.dynptr.type)) {
991 invalidate_dynptr(env, state, spi);
995 ref_obj_id = state->stack[spi].spilled_ptr.ref_obj_id;
997 /* If the dynptr has a ref_obj_id, then we need to invalidate
1000 * 1) Any dynptrs with a matching ref_obj_id (clones)
1001 * 2) Any slices derived from this dynptr.
1004 /* Invalidate any slices associated with this dynptr */
1005 WARN_ON_ONCE(release_reference(env, ref_obj_id));
1007 /* Invalidate any dynptr clones */
1008 for (i = 1; i < state->allocated_stack / BPF_REG_SIZE; i++) {
1009 if (state->stack[i].spilled_ptr.ref_obj_id != ref_obj_id)
1012 /* it should always be the case that if the ref obj id
1013 * matches then the stack slot also belongs to a
1016 if (state->stack[i].slot_type[0] != STACK_DYNPTR) {
1017 verbose(env, "verifier internal error: misconfigured ref_obj_id\n");
1020 if (state->stack[i].spilled_ptr.dynptr.first_slot)
1021 invalidate_dynptr(env, state, i);
1027 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1028 struct bpf_reg_state *reg);
1030 static void mark_reg_invalid(const struct bpf_verifier_env *env, struct bpf_reg_state *reg)
1032 if (!env->allow_ptr_leaks)
1033 __mark_reg_not_init(env, reg);
1035 __mark_reg_unknown(env, reg);
1038 static int destroy_if_dynptr_stack_slot(struct bpf_verifier_env *env,
1039 struct bpf_func_state *state, int spi)
1041 struct bpf_func_state *fstate;
1042 struct bpf_reg_state *dreg;
1045 /* We always ensure that STACK_DYNPTR is never set partially,
1046 * hence just checking for slot_type[0] is enough. This is
1047 * different for STACK_SPILL, where it may be only set for
1048 * 1 byte, so code has to use is_spilled_reg.
1050 if (state->stack[spi].slot_type[0] != STACK_DYNPTR)
1053 /* Reposition spi to first slot */
1054 if (!state->stack[spi].spilled_ptr.dynptr.first_slot)
1057 if (dynptr_type_refcounted(state->stack[spi].spilled_ptr.dynptr.type)) {
1058 verbose(env, "cannot overwrite referenced dynptr\n");
1062 mark_stack_slot_scratched(env, spi);
1063 mark_stack_slot_scratched(env, spi - 1);
1065 /* Writing partially to one dynptr stack slot destroys both. */
1066 for (i = 0; i < BPF_REG_SIZE; i++) {
1067 state->stack[spi].slot_type[i] = STACK_INVALID;
1068 state->stack[spi - 1].slot_type[i] = STACK_INVALID;
1071 dynptr_id = state->stack[spi].spilled_ptr.id;
1072 /* Invalidate any slices associated with this dynptr */
1073 bpf_for_each_reg_in_vstate(env->cur_state, fstate, dreg, ({
1074 /* Dynptr slices are only PTR_TO_MEM_OR_NULL and PTR_TO_MEM */
1075 if (dreg->type != (PTR_TO_MEM | PTR_MAYBE_NULL) && dreg->type != PTR_TO_MEM)
1077 if (dreg->dynptr_id == dynptr_id)
1078 mark_reg_invalid(env, dreg);
1081 /* Do not release reference state, we are destroying dynptr on stack,
1082 * not using some helper to release it. Just reset register.
1084 __mark_reg_not_init(env, &state->stack[spi].spilled_ptr);
1085 __mark_reg_not_init(env, &state->stack[spi - 1].spilled_ptr);
1087 /* Same reason as unmark_stack_slots_dynptr above */
1088 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1089 state->stack[spi - 1].spilled_ptr.live |= REG_LIVE_WRITTEN;
1094 static bool is_dynptr_reg_valid_uninit(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
1098 if (reg->type == CONST_PTR_TO_DYNPTR)
1101 spi = dynptr_get_spi(env, reg);
1103 /* -ERANGE (i.e. spi not falling into allocated stack slots) isn't an
1104 * error because this just means the stack state hasn't been updated yet.
1105 * We will do check_mem_access to check and update stack bounds later.
1107 if (spi < 0 && spi != -ERANGE)
1110 /* We don't need to check if the stack slots are marked by previous
1111 * dynptr initializations because we allow overwriting existing unreferenced
1112 * STACK_DYNPTR slots, see mark_stack_slots_dynptr which calls
1113 * destroy_if_dynptr_stack_slot to ensure dynptr objects at the slots we are
1114 * touching are completely destructed before we reinitialize them for a new
1115 * one. For referenced ones, destroy_if_dynptr_stack_slot returns an error early
1116 * instead of delaying it until the end where the user will get "Unreleased
1122 static bool is_dynptr_reg_valid_init(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
1124 struct bpf_func_state *state = func(env, reg);
1127 /* This already represents first slot of initialized bpf_dynptr.
1129 * CONST_PTR_TO_DYNPTR already has fixed and var_off as 0 due to
1130 * check_func_arg_reg_off's logic, so we don't need to check its
1131 * offset and alignment.
1133 if (reg->type == CONST_PTR_TO_DYNPTR)
1136 spi = dynptr_get_spi(env, reg);
1139 if (!state->stack[spi].spilled_ptr.dynptr.first_slot)
1142 for (i = 0; i < BPF_REG_SIZE; i++) {
1143 if (state->stack[spi].slot_type[i] != STACK_DYNPTR ||
1144 state->stack[spi - 1].slot_type[i] != STACK_DYNPTR)
1151 static bool is_dynptr_type_expected(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
1152 enum bpf_arg_type arg_type)
1154 struct bpf_func_state *state = func(env, reg);
1155 enum bpf_dynptr_type dynptr_type;
1158 /* ARG_PTR_TO_DYNPTR takes any type of dynptr */
1159 if (arg_type == ARG_PTR_TO_DYNPTR)
1162 dynptr_type = arg_to_dynptr_type(arg_type);
1163 if (reg->type == CONST_PTR_TO_DYNPTR) {
1164 return reg->dynptr.type == dynptr_type;
1166 spi = dynptr_get_spi(env, reg);
1169 return state->stack[spi].spilled_ptr.dynptr.type == dynptr_type;
1173 static void __mark_reg_known_zero(struct bpf_reg_state *reg);
1175 static int mark_stack_slots_iter(struct bpf_verifier_env *env,
1176 struct bpf_reg_state *reg, int insn_idx,
1177 struct btf *btf, u32 btf_id, int nr_slots)
1179 struct bpf_func_state *state = func(env, reg);
1182 spi = iter_get_spi(env, reg, nr_slots);
1186 id = acquire_reference_state(env, insn_idx);
1190 for (i = 0; i < nr_slots; i++) {
1191 struct bpf_stack_state *slot = &state->stack[spi - i];
1192 struct bpf_reg_state *st = &slot->spilled_ptr;
1194 __mark_reg_known_zero(st);
1195 st->type = PTR_TO_STACK; /* we don't have dedicated reg type */
1196 st->live |= REG_LIVE_WRITTEN;
1197 st->ref_obj_id = i == 0 ? id : 0;
1199 st->iter.btf_id = btf_id;
1200 st->iter.state = BPF_ITER_STATE_ACTIVE;
1203 for (j = 0; j < BPF_REG_SIZE; j++)
1204 slot->slot_type[j] = STACK_ITER;
1206 mark_stack_slot_scratched(env, spi - i);
1212 static int unmark_stack_slots_iter(struct bpf_verifier_env *env,
1213 struct bpf_reg_state *reg, int nr_slots)
1215 struct bpf_func_state *state = func(env, reg);
1218 spi = iter_get_spi(env, reg, nr_slots);
1222 for (i = 0; i < nr_slots; i++) {
1223 struct bpf_stack_state *slot = &state->stack[spi - i];
1224 struct bpf_reg_state *st = &slot->spilled_ptr;
1227 WARN_ON_ONCE(release_reference(env, st->ref_obj_id));
1229 __mark_reg_not_init(env, st);
1231 /* see unmark_stack_slots_dynptr() for why we need to set REG_LIVE_WRITTEN */
1232 st->live |= REG_LIVE_WRITTEN;
1234 for (j = 0; j < BPF_REG_SIZE; j++)
1235 slot->slot_type[j] = STACK_INVALID;
1237 mark_stack_slot_scratched(env, spi - i);
1243 static bool is_iter_reg_valid_uninit(struct bpf_verifier_env *env,
1244 struct bpf_reg_state *reg, int nr_slots)
1246 struct bpf_func_state *state = func(env, reg);
1249 /* For -ERANGE (i.e. spi not falling into allocated stack slots), we
1250 * will do check_mem_access to check and update stack bounds later, so
1251 * return true for that case.
1253 spi = iter_get_spi(env, reg, nr_slots);
1259 for (i = 0; i < nr_slots; i++) {
1260 struct bpf_stack_state *slot = &state->stack[spi - i];
1262 for (j = 0; j < BPF_REG_SIZE; j++)
1263 if (slot->slot_type[j] == STACK_ITER)
1270 static bool is_iter_reg_valid_init(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
1271 struct btf *btf, u32 btf_id, int nr_slots)
1273 struct bpf_func_state *state = func(env, reg);
1276 spi = iter_get_spi(env, reg, nr_slots);
1280 for (i = 0; i < nr_slots; i++) {
1281 struct bpf_stack_state *slot = &state->stack[spi - i];
1282 struct bpf_reg_state *st = &slot->spilled_ptr;
1284 /* only main (first) slot has ref_obj_id set */
1285 if (i == 0 && !st->ref_obj_id)
1287 if (i != 0 && st->ref_obj_id)
1289 if (st->iter.btf != btf || st->iter.btf_id != btf_id)
1292 for (j = 0; j < BPF_REG_SIZE; j++)
1293 if (slot->slot_type[j] != STACK_ITER)
1300 /* Check if given stack slot is "special":
1301 * - spilled register state (STACK_SPILL);
1302 * - dynptr state (STACK_DYNPTR);
1303 * - iter state (STACK_ITER).
1305 static bool is_stack_slot_special(const struct bpf_stack_state *stack)
1307 enum bpf_stack_slot_type type = stack->slot_type[BPF_REG_SIZE - 1];
1319 WARN_ONCE(1, "unknown stack slot type %d\n", type);
1324 /* The reg state of a pointer or a bounded scalar was saved when
1325 * it was spilled to the stack.
1327 static bool is_spilled_reg(const struct bpf_stack_state *stack)
1329 return stack->slot_type[BPF_REG_SIZE - 1] == STACK_SPILL;
1332 static bool is_spilled_scalar_reg(const struct bpf_stack_state *stack)
1334 return stack->slot_type[BPF_REG_SIZE - 1] == STACK_SPILL &&
1335 stack->spilled_ptr.type == SCALAR_VALUE;
1338 static void scrub_spilled_slot(u8 *stype)
1340 if (*stype != STACK_INVALID)
1341 *stype = STACK_MISC;
1344 static void print_verifier_state(struct bpf_verifier_env *env,
1345 const struct bpf_func_state *state,
1348 const struct bpf_reg_state *reg;
1349 enum bpf_reg_type t;
1353 verbose(env, " frame%d:", state->frameno);
1354 for (i = 0; i < MAX_BPF_REG; i++) {
1355 reg = &state->regs[i];
1359 if (!print_all && !reg_scratched(env, i))
1361 verbose(env, " R%d", i);
1362 print_liveness(env, reg->live);
1364 if (t == SCALAR_VALUE && reg->precise)
1366 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
1367 tnum_is_const(reg->var_off)) {
1368 /* reg->off should be 0 for SCALAR_VALUE */
1369 verbose(env, "%s", t == SCALAR_VALUE ? "" : reg_type_str(env, t));
1370 verbose(env, "%lld", reg->var_off.value + reg->off);
1372 const char *sep = "";
1374 verbose(env, "%s", reg_type_str(env, t));
1375 if (base_type(t) == PTR_TO_BTF_ID)
1376 verbose(env, "%s", btf_type_name(reg->btf, reg->btf_id));
1379 * _a stands for append, was shortened to avoid multiline statements below.
1380 * This macro is used to output a comma separated list of attributes.
1382 #define verbose_a(fmt, ...) ({ verbose(env, "%s" fmt, sep, __VA_ARGS__); sep = ","; })
1385 verbose_a("id=%d", reg->id);
1386 if (reg->ref_obj_id)
1387 verbose_a("ref_obj_id=%d", reg->ref_obj_id);
1388 if (type_is_non_owning_ref(reg->type))
1389 verbose_a("%s", "non_own_ref");
1390 if (t != SCALAR_VALUE)
1391 verbose_a("off=%d", reg->off);
1392 if (type_is_pkt_pointer(t))
1393 verbose_a("r=%d", reg->range);
1394 else if (base_type(t) == CONST_PTR_TO_MAP ||
1395 base_type(t) == PTR_TO_MAP_KEY ||
1396 base_type(t) == PTR_TO_MAP_VALUE)
1397 verbose_a("ks=%d,vs=%d",
1398 reg->map_ptr->key_size,
1399 reg->map_ptr->value_size);
1400 if (tnum_is_const(reg->var_off)) {
1401 /* Typically an immediate SCALAR_VALUE, but
1402 * could be a pointer whose offset is too big
1405 verbose_a("imm=%llx", reg->var_off.value);
1407 if (reg->smin_value != reg->umin_value &&
1408 reg->smin_value != S64_MIN)
1409 verbose_a("smin=%lld", (long long)reg->smin_value);
1410 if (reg->smax_value != reg->umax_value &&
1411 reg->smax_value != S64_MAX)
1412 verbose_a("smax=%lld", (long long)reg->smax_value);
1413 if (reg->umin_value != 0)
1414 verbose_a("umin=%llu", (unsigned long long)reg->umin_value);
1415 if (reg->umax_value != U64_MAX)
1416 verbose_a("umax=%llu", (unsigned long long)reg->umax_value);
1417 if (!tnum_is_unknown(reg->var_off)) {
1420 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1421 verbose_a("var_off=%s", tn_buf);
1423 if (reg->s32_min_value != reg->smin_value &&
1424 reg->s32_min_value != S32_MIN)
1425 verbose_a("s32_min=%d", (int)(reg->s32_min_value));
1426 if (reg->s32_max_value != reg->smax_value &&
1427 reg->s32_max_value != S32_MAX)
1428 verbose_a("s32_max=%d", (int)(reg->s32_max_value));
1429 if (reg->u32_min_value != reg->umin_value &&
1430 reg->u32_min_value != U32_MIN)
1431 verbose_a("u32_min=%d", (int)(reg->u32_min_value));
1432 if (reg->u32_max_value != reg->umax_value &&
1433 reg->u32_max_value != U32_MAX)
1434 verbose_a("u32_max=%d", (int)(reg->u32_max_value));
1441 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
1442 char types_buf[BPF_REG_SIZE + 1];
1446 for (j = 0; j < BPF_REG_SIZE; j++) {
1447 if (state->stack[i].slot_type[j] != STACK_INVALID)
1449 types_buf[j] = slot_type_char[state->stack[i].slot_type[j]];
1451 types_buf[BPF_REG_SIZE] = 0;
1454 if (!print_all && !stack_slot_scratched(env, i))
1456 switch (state->stack[i].slot_type[BPF_REG_SIZE - 1]) {
1458 reg = &state->stack[i].spilled_ptr;
1461 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
1462 print_liveness(env, reg->live);
1463 verbose(env, "=%s", t == SCALAR_VALUE ? "" : reg_type_str(env, t));
1464 if (t == SCALAR_VALUE && reg->precise)
1466 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
1467 verbose(env, "%lld", reg->var_off.value + reg->off);
1470 i += BPF_DYNPTR_NR_SLOTS - 1;
1471 reg = &state->stack[i].spilled_ptr;
1473 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
1474 print_liveness(env, reg->live);
1475 verbose(env, "=dynptr_%s", dynptr_type_str(reg->dynptr.type));
1476 if (reg->ref_obj_id)
1477 verbose(env, "(ref_id=%d)", reg->ref_obj_id);
1480 /* only main slot has ref_obj_id set; skip others */
1481 reg = &state->stack[i].spilled_ptr;
1482 if (!reg->ref_obj_id)
1485 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
1486 print_liveness(env, reg->live);
1487 verbose(env, "=iter_%s(ref_id=%d,state=%s,depth=%u)",
1488 iter_type_str(reg->iter.btf, reg->iter.btf_id),
1489 reg->ref_obj_id, iter_state_str(reg->iter.state),
1495 reg = &state->stack[i].spilled_ptr;
1497 for (j = 0; j < BPF_REG_SIZE; j++)
1498 types_buf[j] = slot_type_char[state->stack[i].slot_type[j]];
1499 types_buf[BPF_REG_SIZE] = 0;
1501 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
1502 print_liveness(env, reg->live);
1503 verbose(env, "=%s", types_buf);
1507 if (state->acquired_refs && state->refs[0].id) {
1508 verbose(env, " refs=%d", state->refs[0].id);
1509 for (i = 1; i < state->acquired_refs; i++)
1510 if (state->refs[i].id)
1511 verbose(env, ",%d", state->refs[i].id);
1513 if (state->in_callback_fn)
1514 verbose(env, " cb");
1515 if (state->in_async_callback_fn)
1516 verbose(env, " async_cb");
1519 mark_verifier_state_clean(env);
1522 static inline u32 vlog_alignment(u32 pos)
1524 return round_up(max(pos + BPF_LOG_MIN_ALIGNMENT / 2, BPF_LOG_ALIGNMENT),
1525 BPF_LOG_MIN_ALIGNMENT) - pos - 1;
1528 static void print_insn_state(struct bpf_verifier_env *env,
1529 const struct bpf_func_state *state)
1531 if (env->prev_log_pos && env->prev_log_pos == env->log.end_pos) {
1532 /* remove new line character */
1533 bpf_vlog_reset(&env->log, env->prev_log_pos - 1);
1534 verbose(env, "%*c;", vlog_alignment(env->prev_insn_print_pos), ' ');
1536 verbose(env, "%d:", env->insn_idx);
1538 print_verifier_state(env, state, false);
1541 /* copy array src of length n * size bytes to dst. dst is reallocated if it's too
1542 * small to hold src. This is different from krealloc since we don't want to preserve
1543 * the contents of dst.
1545 * Leaves dst untouched if src is NULL or length is zero. Returns NULL if memory could
1548 static void *copy_array(void *dst, const void *src, size_t n, size_t size, gfp_t flags)
1554 if (ZERO_OR_NULL_PTR(src))
1557 if (unlikely(check_mul_overflow(n, size, &bytes)))
1560 alloc_bytes = max(ksize(orig), kmalloc_size_roundup(bytes));
1561 dst = krealloc(orig, alloc_bytes, flags);
1567 memcpy(dst, src, bytes);
1569 return dst ? dst : ZERO_SIZE_PTR;
1572 /* resize an array from old_n items to new_n items. the array is reallocated if it's too
1573 * small to hold new_n items. new items are zeroed out if the array grows.
1575 * Contrary to krealloc_array, does not free arr if new_n is zero.
1577 static void *realloc_array(void *arr, size_t old_n, size_t new_n, size_t size)
1582 if (!new_n || old_n == new_n)
1585 alloc_size = kmalloc_size_roundup(size_mul(new_n, size));
1586 new_arr = krealloc(arr, alloc_size, GFP_KERNEL);
1594 memset(arr + old_n * size, 0, (new_n - old_n) * size);
1597 return arr ? arr : ZERO_SIZE_PTR;
1600 static int copy_reference_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
1602 dst->refs = copy_array(dst->refs, src->refs, src->acquired_refs,
1603 sizeof(struct bpf_reference_state), GFP_KERNEL);
1607 dst->acquired_refs = src->acquired_refs;
1611 static int copy_stack_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
1613 size_t n = src->allocated_stack / BPF_REG_SIZE;
1615 dst->stack = copy_array(dst->stack, src->stack, n, sizeof(struct bpf_stack_state),
1620 dst->allocated_stack = src->allocated_stack;
1624 static int resize_reference_state(struct bpf_func_state *state, size_t n)
1626 state->refs = realloc_array(state->refs, state->acquired_refs, n,
1627 sizeof(struct bpf_reference_state));
1631 state->acquired_refs = n;
1635 static int grow_stack_state(struct bpf_func_state *state, int size)
1637 size_t old_n = state->allocated_stack / BPF_REG_SIZE, n = size / BPF_REG_SIZE;
1642 state->stack = realloc_array(state->stack, old_n, n, sizeof(struct bpf_stack_state));
1646 state->allocated_stack = size;
1650 /* Acquire a pointer id from the env and update the state->refs to include
1651 * this new pointer reference.
1652 * On success, returns a valid pointer id to associate with the register
1653 * On failure, returns a negative errno.
1655 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
1657 struct bpf_func_state *state = cur_func(env);
1658 int new_ofs = state->acquired_refs;
1661 err = resize_reference_state(state, state->acquired_refs + 1);
1665 state->refs[new_ofs].id = id;
1666 state->refs[new_ofs].insn_idx = insn_idx;
1667 state->refs[new_ofs].callback_ref = state->in_callback_fn ? state->frameno : 0;
1672 /* release function corresponding to acquire_reference_state(). Idempotent. */
1673 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
1677 last_idx = state->acquired_refs - 1;
1678 for (i = 0; i < state->acquired_refs; i++) {
1679 if (state->refs[i].id == ptr_id) {
1680 /* Cannot release caller references in callbacks */
1681 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno)
1683 if (last_idx && i != last_idx)
1684 memcpy(&state->refs[i], &state->refs[last_idx],
1685 sizeof(*state->refs));
1686 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
1687 state->acquired_refs--;
1694 static void free_func_state(struct bpf_func_state *state)
1699 kfree(state->stack);
1703 static void clear_jmp_history(struct bpf_verifier_state *state)
1705 kfree(state->jmp_history);
1706 state->jmp_history = NULL;
1707 state->jmp_history_cnt = 0;
1710 static void free_verifier_state(struct bpf_verifier_state *state,
1715 for (i = 0; i <= state->curframe; i++) {
1716 free_func_state(state->frame[i]);
1717 state->frame[i] = NULL;
1719 clear_jmp_history(state);
1724 /* copy verifier state from src to dst growing dst stack space
1725 * when necessary to accommodate larger src stack
1727 static int copy_func_state(struct bpf_func_state *dst,
1728 const struct bpf_func_state *src)
1732 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
1733 err = copy_reference_state(dst, src);
1736 return copy_stack_state(dst, src);
1739 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
1740 const struct bpf_verifier_state *src)
1742 struct bpf_func_state *dst;
1745 dst_state->jmp_history = copy_array(dst_state->jmp_history, src->jmp_history,
1746 src->jmp_history_cnt, sizeof(struct bpf_idx_pair),
1748 if (!dst_state->jmp_history)
1750 dst_state->jmp_history_cnt = src->jmp_history_cnt;
1752 /* if dst has more stack frames then src frame, free them */
1753 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
1754 free_func_state(dst_state->frame[i]);
1755 dst_state->frame[i] = NULL;
1757 dst_state->speculative = src->speculative;
1758 dst_state->active_rcu_lock = src->active_rcu_lock;
1759 dst_state->curframe = src->curframe;
1760 dst_state->active_lock.ptr = src->active_lock.ptr;
1761 dst_state->active_lock.id = src->active_lock.id;
1762 dst_state->branches = src->branches;
1763 dst_state->parent = src->parent;
1764 dst_state->first_insn_idx = src->first_insn_idx;
1765 dst_state->last_insn_idx = src->last_insn_idx;
1766 for (i = 0; i <= src->curframe; i++) {
1767 dst = dst_state->frame[i];
1769 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
1772 dst_state->frame[i] = dst;
1774 err = copy_func_state(dst, src->frame[i]);
1781 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
1784 u32 br = --st->branches;
1786 /* WARN_ON(br > 1) technically makes sense here,
1787 * but see comment in push_stack(), hence:
1789 WARN_ONCE((int)br < 0,
1790 "BUG update_branch_counts:branches_to_explore=%d\n",
1798 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
1799 int *insn_idx, bool pop_log)
1801 struct bpf_verifier_state *cur = env->cur_state;
1802 struct bpf_verifier_stack_elem *elem, *head = env->head;
1805 if (env->head == NULL)
1809 err = copy_verifier_state(cur, &head->st);
1814 bpf_vlog_reset(&env->log, head->log_pos);
1816 *insn_idx = head->insn_idx;
1818 *prev_insn_idx = head->prev_insn_idx;
1820 free_verifier_state(&head->st, false);
1827 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
1828 int insn_idx, int prev_insn_idx,
1831 struct bpf_verifier_state *cur = env->cur_state;
1832 struct bpf_verifier_stack_elem *elem;
1835 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1839 elem->insn_idx = insn_idx;
1840 elem->prev_insn_idx = prev_insn_idx;
1841 elem->next = env->head;
1842 elem->log_pos = env->log.end_pos;
1845 err = copy_verifier_state(&elem->st, cur);
1848 elem->st.speculative |= speculative;
1849 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1850 verbose(env, "The sequence of %d jumps is too complex.\n",
1854 if (elem->st.parent) {
1855 ++elem->st.parent->branches;
1856 /* WARN_ON(branches > 2) technically makes sense here,
1858 * 1. speculative states will bump 'branches' for non-branch
1860 * 2. is_state_visited() heuristics may decide not to create
1861 * a new state for a sequence of branches and all such current
1862 * and cloned states will be pointing to a single parent state
1863 * which might have large 'branches' count.
1868 free_verifier_state(env->cur_state, true);
1869 env->cur_state = NULL;
1870 /* pop all elements and return */
1871 while (!pop_stack(env, NULL, NULL, false));
1875 #define CALLER_SAVED_REGS 6
1876 static const int caller_saved[CALLER_SAVED_REGS] = {
1877 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
1880 /* This helper doesn't clear reg->id */
1881 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1883 reg->var_off = tnum_const(imm);
1884 reg->smin_value = (s64)imm;
1885 reg->smax_value = (s64)imm;
1886 reg->umin_value = imm;
1887 reg->umax_value = imm;
1889 reg->s32_min_value = (s32)imm;
1890 reg->s32_max_value = (s32)imm;
1891 reg->u32_min_value = (u32)imm;
1892 reg->u32_max_value = (u32)imm;
1895 /* Mark the unknown part of a register (variable offset or scalar value) as
1896 * known to have the value @imm.
1898 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1900 /* Clear off and union(map_ptr, range) */
1901 memset(((u8 *)reg) + sizeof(reg->type), 0,
1902 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1904 reg->ref_obj_id = 0;
1905 ___mark_reg_known(reg, imm);
1908 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1910 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1911 reg->s32_min_value = (s32)imm;
1912 reg->s32_max_value = (s32)imm;
1913 reg->u32_min_value = (u32)imm;
1914 reg->u32_max_value = (u32)imm;
1917 /* Mark the 'variable offset' part of a register as zero. This should be
1918 * used only on registers holding a pointer type.
1920 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1922 __mark_reg_known(reg, 0);
1925 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1927 __mark_reg_known(reg, 0);
1928 reg->type = SCALAR_VALUE;
1931 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1932 struct bpf_reg_state *regs, u32 regno)
1934 if (WARN_ON(regno >= MAX_BPF_REG)) {
1935 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1936 /* Something bad happened, let's kill all regs */
1937 for (regno = 0; regno < MAX_BPF_REG; regno++)
1938 __mark_reg_not_init(env, regs + regno);
1941 __mark_reg_known_zero(regs + regno);
1944 static void __mark_dynptr_reg(struct bpf_reg_state *reg, enum bpf_dynptr_type type,
1945 bool first_slot, int dynptr_id)
1947 /* reg->type has no meaning for STACK_DYNPTR, but when we set reg for
1948 * callback arguments, it does need to be CONST_PTR_TO_DYNPTR, so simply
1949 * set it unconditionally as it is ignored for STACK_DYNPTR anyway.
1951 __mark_reg_known_zero(reg);
1952 reg->type = CONST_PTR_TO_DYNPTR;
1953 /* Give each dynptr a unique id to uniquely associate slices to it. */
1954 reg->id = dynptr_id;
1955 reg->dynptr.type = type;
1956 reg->dynptr.first_slot = first_slot;
1959 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
1961 if (base_type(reg->type) == PTR_TO_MAP_VALUE) {
1962 const struct bpf_map *map = reg->map_ptr;
1964 if (map->inner_map_meta) {
1965 reg->type = CONST_PTR_TO_MAP;
1966 reg->map_ptr = map->inner_map_meta;
1967 /* transfer reg's id which is unique for every map_lookup_elem
1968 * as UID of the inner map.
1970 if (btf_record_has_field(map->inner_map_meta->record, BPF_TIMER))
1971 reg->map_uid = reg->id;
1972 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
1973 reg->type = PTR_TO_XDP_SOCK;
1974 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
1975 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
1976 reg->type = PTR_TO_SOCKET;
1978 reg->type = PTR_TO_MAP_VALUE;
1983 reg->type &= ~PTR_MAYBE_NULL;
1986 static void mark_reg_graph_node(struct bpf_reg_state *regs, u32 regno,
1987 struct btf_field_graph_root *ds_head)
1989 __mark_reg_known_zero(®s[regno]);
1990 regs[regno].type = PTR_TO_BTF_ID | MEM_ALLOC;
1991 regs[regno].btf = ds_head->btf;
1992 regs[regno].btf_id = ds_head->value_btf_id;
1993 regs[regno].off = ds_head->node_offset;
1996 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1998 return type_is_pkt_pointer(reg->type);
2001 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
2003 return reg_is_pkt_pointer(reg) ||
2004 reg->type == PTR_TO_PACKET_END;
2007 static bool reg_is_dynptr_slice_pkt(const struct bpf_reg_state *reg)
2009 return base_type(reg->type) == PTR_TO_MEM &&
2010 (reg->type & DYNPTR_TYPE_SKB || reg->type & DYNPTR_TYPE_XDP);
2013 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
2014 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
2015 enum bpf_reg_type which)
2017 /* The register can already have a range from prior markings.
2018 * This is fine as long as it hasn't been advanced from its
2021 return reg->type == which &&
2024 tnum_equals_const(reg->var_off, 0);
2027 /* Reset the min/max bounds of a register */
2028 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
2030 reg->smin_value = S64_MIN;
2031 reg->smax_value = S64_MAX;
2032 reg->umin_value = 0;
2033 reg->umax_value = U64_MAX;
2035 reg->s32_min_value = S32_MIN;
2036 reg->s32_max_value = S32_MAX;
2037 reg->u32_min_value = 0;
2038 reg->u32_max_value = U32_MAX;
2041 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
2043 reg->smin_value = S64_MIN;
2044 reg->smax_value = S64_MAX;
2045 reg->umin_value = 0;
2046 reg->umax_value = U64_MAX;
2049 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
2051 reg->s32_min_value = S32_MIN;
2052 reg->s32_max_value = S32_MAX;
2053 reg->u32_min_value = 0;
2054 reg->u32_max_value = U32_MAX;
2057 static void __update_reg32_bounds(struct bpf_reg_state *reg)
2059 struct tnum var32_off = tnum_subreg(reg->var_off);
2061 /* min signed is max(sign bit) | min(other bits) */
2062 reg->s32_min_value = max_t(s32, reg->s32_min_value,
2063 var32_off.value | (var32_off.mask & S32_MIN));
2064 /* max signed is min(sign bit) | max(other bits) */
2065 reg->s32_max_value = min_t(s32, reg->s32_max_value,
2066 var32_off.value | (var32_off.mask & S32_MAX));
2067 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
2068 reg->u32_max_value = min(reg->u32_max_value,
2069 (u32)(var32_off.value | var32_off.mask));
2072 static void __update_reg64_bounds(struct bpf_reg_state *reg)
2074 /* min signed is max(sign bit) | min(other bits) */
2075 reg->smin_value = max_t(s64, reg->smin_value,
2076 reg->var_off.value | (reg->var_off.mask & S64_MIN));
2077 /* max signed is min(sign bit) | max(other bits) */
2078 reg->smax_value = min_t(s64, reg->smax_value,
2079 reg->var_off.value | (reg->var_off.mask & S64_MAX));
2080 reg->umin_value = max(reg->umin_value, reg->var_off.value);
2081 reg->umax_value = min(reg->umax_value,
2082 reg->var_off.value | reg->var_off.mask);
2085 static void __update_reg_bounds(struct bpf_reg_state *reg)
2087 __update_reg32_bounds(reg);
2088 __update_reg64_bounds(reg);
2091 /* Uses signed min/max values to inform unsigned, and vice-versa */
2092 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
2094 /* Learn sign from signed bounds.
2095 * If we cannot cross the sign boundary, then signed and unsigned bounds
2096 * are the same, so combine. This works even in the negative case, e.g.
2097 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
2099 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
2100 reg->s32_min_value = reg->u32_min_value =
2101 max_t(u32, reg->s32_min_value, reg->u32_min_value);
2102 reg->s32_max_value = reg->u32_max_value =
2103 min_t(u32, reg->s32_max_value, reg->u32_max_value);
2106 /* Learn sign from unsigned bounds. Signed bounds cross the sign
2107 * boundary, so we must be careful.
2109 if ((s32)reg->u32_max_value >= 0) {
2110 /* Positive. We can't learn anything from the smin, but smax
2111 * is positive, hence safe.
2113 reg->s32_min_value = reg->u32_min_value;
2114 reg->s32_max_value = reg->u32_max_value =
2115 min_t(u32, reg->s32_max_value, reg->u32_max_value);
2116 } else if ((s32)reg->u32_min_value < 0) {
2117 /* Negative. We can't learn anything from the smax, but smin
2118 * is negative, hence safe.
2120 reg->s32_min_value = reg->u32_min_value =
2121 max_t(u32, reg->s32_min_value, reg->u32_min_value);
2122 reg->s32_max_value = reg->u32_max_value;
2126 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
2128 /* Learn sign from signed bounds.
2129 * If we cannot cross the sign boundary, then signed and unsigned bounds
2130 * are the same, so combine. This works even in the negative case, e.g.
2131 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
2133 if (reg->smin_value >= 0 || reg->smax_value < 0) {
2134 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
2136 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
2140 /* Learn sign from unsigned bounds. Signed bounds cross the sign
2141 * boundary, so we must be careful.
2143 if ((s64)reg->umax_value >= 0) {
2144 /* Positive. We can't learn anything from the smin, but smax
2145 * is positive, hence safe.
2147 reg->smin_value = reg->umin_value;
2148 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
2150 } else if ((s64)reg->umin_value < 0) {
2151 /* Negative. We can't learn anything from the smax, but smin
2152 * is negative, hence safe.
2154 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
2156 reg->smax_value = reg->umax_value;
2160 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
2162 __reg32_deduce_bounds(reg);
2163 __reg64_deduce_bounds(reg);
2166 /* Attempts to improve var_off based on unsigned min/max information */
2167 static void __reg_bound_offset(struct bpf_reg_state *reg)
2169 struct tnum var64_off = tnum_intersect(reg->var_off,
2170 tnum_range(reg->umin_value,
2172 struct tnum var32_off = tnum_intersect(tnum_subreg(var64_off),
2173 tnum_range(reg->u32_min_value,
2174 reg->u32_max_value));
2176 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
2179 static void reg_bounds_sync(struct bpf_reg_state *reg)
2181 /* We might have learned new bounds from the var_off. */
2182 __update_reg_bounds(reg);
2183 /* We might have learned something about the sign bit. */
2184 __reg_deduce_bounds(reg);
2185 /* We might have learned some bits from the bounds. */
2186 __reg_bound_offset(reg);
2187 /* Intersecting with the old var_off might have improved our bounds
2188 * slightly, e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2189 * then new var_off is (0; 0x7f...fc) which improves our umax.
2191 __update_reg_bounds(reg);
2194 static bool __reg32_bound_s64(s32 a)
2196 return a >= 0 && a <= S32_MAX;
2199 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
2201 reg->umin_value = reg->u32_min_value;
2202 reg->umax_value = reg->u32_max_value;
2204 /* Attempt to pull 32-bit signed bounds into 64-bit bounds but must
2205 * be positive otherwise set to worse case bounds and refine later
2208 if (__reg32_bound_s64(reg->s32_min_value) &&
2209 __reg32_bound_s64(reg->s32_max_value)) {
2210 reg->smin_value = reg->s32_min_value;
2211 reg->smax_value = reg->s32_max_value;
2213 reg->smin_value = 0;
2214 reg->smax_value = U32_MAX;
2218 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
2220 /* special case when 64-bit register has upper 32-bit register
2221 * zeroed. Typically happens after zext or <<32, >>32 sequence
2222 * allowing us to use 32-bit bounds directly,
2224 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
2225 __reg_assign_32_into_64(reg);
2227 /* Otherwise the best we can do is push lower 32bit known and
2228 * unknown bits into register (var_off set from jmp logic)
2229 * then learn as much as possible from the 64-bit tnum
2230 * known and unknown bits. The previous smin/smax bounds are
2231 * invalid here because of jmp32 compare so mark them unknown
2232 * so they do not impact tnum bounds calculation.
2234 __mark_reg64_unbounded(reg);
2236 reg_bounds_sync(reg);
2239 static bool __reg64_bound_s32(s64 a)
2241 return a >= S32_MIN && a <= S32_MAX;
2244 static bool __reg64_bound_u32(u64 a)
2246 return a >= U32_MIN && a <= U32_MAX;
2249 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
2251 __mark_reg32_unbounded(reg);
2252 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
2253 reg->s32_min_value = (s32)reg->smin_value;
2254 reg->s32_max_value = (s32)reg->smax_value;
2256 if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
2257 reg->u32_min_value = (u32)reg->umin_value;
2258 reg->u32_max_value = (u32)reg->umax_value;
2260 reg_bounds_sync(reg);
2263 /* Mark a register as having a completely unknown (scalar) value. */
2264 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
2265 struct bpf_reg_state *reg)
2268 * Clear type, off, and union(map_ptr, range) and
2269 * padding between 'type' and union
2271 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
2272 reg->type = SCALAR_VALUE;
2274 reg->ref_obj_id = 0;
2275 reg->var_off = tnum_unknown;
2277 reg->precise = !env->bpf_capable;
2278 __mark_reg_unbounded(reg);
2281 static void mark_reg_unknown(struct bpf_verifier_env *env,
2282 struct bpf_reg_state *regs, u32 regno)
2284 if (WARN_ON(regno >= MAX_BPF_REG)) {
2285 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
2286 /* Something bad happened, let's kill all regs except FP */
2287 for (regno = 0; regno < BPF_REG_FP; regno++)
2288 __mark_reg_not_init(env, regs + regno);
2291 __mark_reg_unknown(env, regs + regno);
2294 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
2295 struct bpf_reg_state *reg)
2297 __mark_reg_unknown(env, reg);
2298 reg->type = NOT_INIT;
2301 static void mark_reg_not_init(struct bpf_verifier_env *env,
2302 struct bpf_reg_state *regs, u32 regno)
2304 if (WARN_ON(regno >= MAX_BPF_REG)) {
2305 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
2306 /* Something bad happened, let's kill all regs except FP */
2307 for (regno = 0; regno < BPF_REG_FP; regno++)
2308 __mark_reg_not_init(env, regs + regno);
2311 __mark_reg_not_init(env, regs + regno);
2314 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
2315 struct bpf_reg_state *regs, u32 regno,
2316 enum bpf_reg_type reg_type,
2317 struct btf *btf, u32 btf_id,
2318 enum bpf_type_flag flag)
2320 if (reg_type == SCALAR_VALUE) {
2321 mark_reg_unknown(env, regs, regno);
2324 mark_reg_known_zero(env, regs, regno);
2325 regs[regno].type = PTR_TO_BTF_ID | flag;
2326 regs[regno].btf = btf;
2327 regs[regno].btf_id = btf_id;
2330 #define DEF_NOT_SUBREG (0)
2331 static void init_reg_state(struct bpf_verifier_env *env,
2332 struct bpf_func_state *state)
2334 struct bpf_reg_state *regs = state->regs;
2337 for (i = 0; i < MAX_BPF_REG; i++) {
2338 mark_reg_not_init(env, regs, i);
2339 regs[i].live = REG_LIVE_NONE;
2340 regs[i].parent = NULL;
2341 regs[i].subreg_def = DEF_NOT_SUBREG;
2345 regs[BPF_REG_FP].type = PTR_TO_STACK;
2346 mark_reg_known_zero(env, regs, BPF_REG_FP);
2347 regs[BPF_REG_FP].frameno = state->frameno;
2350 #define BPF_MAIN_FUNC (-1)
2351 static void init_func_state(struct bpf_verifier_env *env,
2352 struct bpf_func_state *state,
2353 int callsite, int frameno, int subprogno)
2355 state->callsite = callsite;
2356 state->frameno = frameno;
2357 state->subprogno = subprogno;
2358 state->callback_ret_range = tnum_range(0, 0);
2359 init_reg_state(env, state);
2360 mark_verifier_state_scratched(env);
2363 /* Similar to push_stack(), but for async callbacks */
2364 static struct bpf_verifier_state *push_async_cb(struct bpf_verifier_env *env,
2365 int insn_idx, int prev_insn_idx,
2368 struct bpf_verifier_stack_elem *elem;
2369 struct bpf_func_state *frame;
2371 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
2375 elem->insn_idx = insn_idx;
2376 elem->prev_insn_idx = prev_insn_idx;
2377 elem->next = env->head;
2378 elem->log_pos = env->log.end_pos;
2381 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
2383 "The sequence of %d jumps is too complex for async cb.\n",
2387 /* Unlike push_stack() do not copy_verifier_state().
2388 * The caller state doesn't matter.
2389 * This is async callback. It starts in a fresh stack.
2390 * Initialize it similar to do_check_common().
2392 elem->st.branches = 1;
2393 frame = kzalloc(sizeof(*frame), GFP_KERNEL);
2396 init_func_state(env, frame,
2397 BPF_MAIN_FUNC /* callsite */,
2398 0 /* frameno within this callchain */,
2399 subprog /* subprog number within this prog */);
2400 elem->st.frame[0] = frame;
2403 free_verifier_state(env->cur_state, true);
2404 env->cur_state = NULL;
2405 /* pop all elements and return */
2406 while (!pop_stack(env, NULL, NULL, false));
2412 SRC_OP, /* register is used as source operand */
2413 DST_OP, /* register is used as destination operand */
2414 DST_OP_NO_MARK /* same as above, check only, don't mark */
2417 static int cmp_subprogs(const void *a, const void *b)
2419 return ((struct bpf_subprog_info *)a)->start -
2420 ((struct bpf_subprog_info *)b)->start;
2423 static int find_subprog(struct bpf_verifier_env *env, int off)
2425 struct bpf_subprog_info *p;
2427 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
2428 sizeof(env->subprog_info[0]), cmp_subprogs);
2431 return p - env->subprog_info;
2435 static int add_subprog(struct bpf_verifier_env *env, int off)
2437 int insn_cnt = env->prog->len;
2440 if (off >= insn_cnt || off < 0) {
2441 verbose(env, "call to invalid destination\n");
2444 ret = find_subprog(env, off);
2447 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
2448 verbose(env, "too many subprograms\n");
2451 /* determine subprog starts. The end is one before the next starts */
2452 env->subprog_info[env->subprog_cnt++].start = off;
2453 sort(env->subprog_info, env->subprog_cnt,
2454 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
2455 return env->subprog_cnt - 1;
2458 #define MAX_KFUNC_DESCS 256
2459 #define MAX_KFUNC_BTFS 256
2461 struct bpf_kfunc_desc {
2462 struct btf_func_model func_model;
2469 struct bpf_kfunc_btf {
2471 struct module *module;
2475 struct bpf_kfunc_desc_tab {
2476 /* Sorted by func_id (BTF ID) and offset (fd_array offset) during
2477 * verification. JITs do lookups by bpf_insn, where func_id may not be
2478 * available, therefore at the end of verification do_misc_fixups()
2479 * sorts this by imm and offset.
2481 struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
2485 struct bpf_kfunc_btf_tab {
2486 struct bpf_kfunc_btf descs[MAX_KFUNC_BTFS];
2490 static int kfunc_desc_cmp_by_id_off(const void *a, const void *b)
2492 const struct bpf_kfunc_desc *d0 = a;
2493 const struct bpf_kfunc_desc *d1 = b;
2495 /* func_id is not greater than BTF_MAX_TYPE */
2496 return d0->func_id - d1->func_id ?: d0->offset - d1->offset;
2499 static int kfunc_btf_cmp_by_off(const void *a, const void *b)
2501 const struct bpf_kfunc_btf *d0 = a;
2502 const struct bpf_kfunc_btf *d1 = b;
2504 return d0->offset - d1->offset;
2507 static const struct bpf_kfunc_desc *
2508 find_kfunc_desc(const struct bpf_prog *prog, u32 func_id, u16 offset)
2510 struct bpf_kfunc_desc desc = {
2514 struct bpf_kfunc_desc_tab *tab;
2516 tab = prog->aux->kfunc_tab;
2517 return bsearch(&desc, tab->descs, tab->nr_descs,
2518 sizeof(tab->descs[0]), kfunc_desc_cmp_by_id_off);
2521 int bpf_get_kfunc_addr(const struct bpf_prog *prog, u32 func_id,
2522 u16 btf_fd_idx, u8 **func_addr)
2524 const struct bpf_kfunc_desc *desc;
2526 desc = find_kfunc_desc(prog, func_id, btf_fd_idx);
2530 *func_addr = (u8 *)desc->addr;
2534 static struct btf *__find_kfunc_desc_btf(struct bpf_verifier_env *env,
2537 struct bpf_kfunc_btf kf_btf = { .offset = offset };
2538 struct bpf_kfunc_btf_tab *tab;
2539 struct bpf_kfunc_btf *b;
2544 tab = env->prog->aux->kfunc_btf_tab;
2545 b = bsearch(&kf_btf, tab->descs, tab->nr_descs,
2546 sizeof(tab->descs[0]), kfunc_btf_cmp_by_off);
2548 if (tab->nr_descs == MAX_KFUNC_BTFS) {
2549 verbose(env, "too many different module BTFs\n");
2550 return ERR_PTR(-E2BIG);
2553 if (bpfptr_is_null(env->fd_array)) {
2554 verbose(env, "kfunc offset > 0 without fd_array is invalid\n");
2555 return ERR_PTR(-EPROTO);
2558 if (copy_from_bpfptr_offset(&btf_fd, env->fd_array,
2559 offset * sizeof(btf_fd),
2561 return ERR_PTR(-EFAULT);
2563 btf = btf_get_by_fd(btf_fd);
2565 verbose(env, "invalid module BTF fd specified\n");
2569 if (!btf_is_module(btf)) {
2570 verbose(env, "BTF fd for kfunc is not a module BTF\n");
2572 return ERR_PTR(-EINVAL);
2575 mod = btf_try_get_module(btf);
2578 return ERR_PTR(-ENXIO);
2581 b = &tab->descs[tab->nr_descs++];
2586 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2587 kfunc_btf_cmp_by_off, NULL);
2592 void bpf_free_kfunc_btf_tab(struct bpf_kfunc_btf_tab *tab)
2597 while (tab->nr_descs--) {
2598 module_put(tab->descs[tab->nr_descs].module);
2599 btf_put(tab->descs[tab->nr_descs].btf);
2604 static struct btf *find_kfunc_desc_btf(struct bpf_verifier_env *env, s16 offset)
2608 /* In the future, this can be allowed to increase limit
2609 * of fd index into fd_array, interpreted as u16.
2611 verbose(env, "negative offset disallowed for kernel module function call\n");
2612 return ERR_PTR(-EINVAL);
2615 return __find_kfunc_desc_btf(env, offset);
2617 return btf_vmlinux ?: ERR_PTR(-ENOENT);
2620 static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id, s16 offset)
2622 const struct btf_type *func, *func_proto;
2623 struct bpf_kfunc_btf_tab *btf_tab;
2624 struct bpf_kfunc_desc_tab *tab;
2625 struct bpf_prog_aux *prog_aux;
2626 struct bpf_kfunc_desc *desc;
2627 const char *func_name;
2628 struct btf *desc_btf;
2629 unsigned long call_imm;
2633 prog_aux = env->prog->aux;
2634 tab = prog_aux->kfunc_tab;
2635 btf_tab = prog_aux->kfunc_btf_tab;
2638 verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
2642 if (!env->prog->jit_requested) {
2643 verbose(env, "JIT is required for calling kernel function\n");
2647 if (!bpf_jit_supports_kfunc_call()) {
2648 verbose(env, "JIT does not support calling kernel function\n");
2652 if (!env->prog->gpl_compatible) {
2653 verbose(env, "cannot call kernel function from non-GPL compatible program\n");
2657 tab = kzalloc(sizeof(*tab), GFP_KERNEL);
2660 prog_aux->kfunc_tab = tab;
2663 /* func_id == 0 is always invalid, but instead of returning an error, be
2664 * conservative and wait until the code elimination pass before returning
2665 * error, so that invalid calls that get pruned out can be in BPF programs
2666 * loaded from userspace. It is also required that offset be untouched
2669 if (!func_id && !offset)
2672 if (!btf_tab && offset) {
2673 btf_tab = kzalloc(sizeof(*btf_tab), GFP_KERNEL);
2676 prog_aux->kfunc_btf_tab = btf_tab;
2679 desc_btf = find_kfunc_desc_btf(env, offset);
2680 if (IS_ERR(desc_btf)) {
2681 verbose(env, "failed to find BTF for kernel function\n");
2682 return PTR_ERR(desc_btf);
2685 if (find_kfunc_desc(env->prog, func_id, offset))
2688 if (tab->nr_descs == MAX_KFUNC_DESCS) {
2689 verbose(env, "too many different kernel function calls\n");
2693 func = btf_type_by_id(desc_btf, func_id);
2694 if (!func || !btf_type_is_func(func)) {
2695 verbose(env, "kernel btf_id %u is not a function\n",
2699 func_proto = btf_type_by_id(desc_btf, func->type);
2700 if (!func_proto || !btf_type_is_func_proto(func_proto)) {
2701 verbose(env, "kernel function btf_id %u does not have a valid func_proto\n",
2706 func_name = btf_name_by_offset(desc_btf, func->name_off);
2707 addr = kallsyms_lookup_name(func_name);
2709 verbose(env, "cannot find address for kernel function %s\n",
2713 specialize_kfunc(env, func_id, offset, &addr);
2715 if (bpf_jit_supports_far_kfunc_call()) {
2718 call_imm = BPF_CALL_IMM(addr);
2719 /* Check whether the relative offset overflows desc->imm */
2720 if ((unsigned long)(s32)call_imm != call_imm) {
2721 verbose(env, "address of kernel function %s is out of range\n",
2727 if (bpf_dev_bound_kfunc_id(func_id)) {
2728 err = bpf_dev_bound_kfunc_check(&env->log, prog_aux);
2733 desc = &tab->descs[tab->nr_descs++];
2734 desc->func_id = func_id;
2735 desc->imm = call_imm;
2736 desc->offset = offset;
2738 err = btf_distill_func_proto(&env->log, desc_btf,
2739 func_proto, func_name,
2742 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2743 kfunc_desc_cmp_by_id_off, NULL);
2747 static int kfunc_desc_cmp_by_imm_off(const void *a, const void *b)
2749 const struct bpf_kfunc_desc *d0 = a;
2750 const struct bpf_kfunc_desc *d1 = b;
2752 if (d0->imm != d1->imm)
2753 return d0->imm < d1->imm ? -1 : 1;
2754 if (d0->offset != d1->offset)
2755 return d0->offset < d1->offset ? -1 : 1;
2759 static void sort_kfunc_descs_by_imm_off(struct bpf_prog *prog)
2761 struct bpf_kfunc_desc_tab *tab;
2763 tab = prog->aux->kfunc_tab;
2767 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2768 kfunc_desc_cmp_by_imm_off, NULL);
2771 bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
2773 return !!prog->aux->kfunc_tab;
2776 const struct btf_func_model *
2777 bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
2778 const struct bpf_insn *insn)
2780 const struct bpf_kfunc_desc desc = {
2782 .offset = insn->off,
2784 const struct bpf_kfunc_desc *res;
2785 struct bpf_kfunc_desc_tab *tab;
2787 tab = prog->aux->kfunc_tab;
2788 res = bsearch(&desc, tab->descs, tab->nr_descs,
2789 sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm_off);
2791 return res ? &res->func_model : NULL;
2794 static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
2796 struct bpf_subprog_info *subprog = env->subprog_info;
2797 struct bpf_insn *insn = env->prog->insnsi;
2798 int i, ret, insn_cnt = env->prog->len;
2800 /* Add entry function. */
2801 ret = add_subprog(env, 0);
2805 for (i = 0; i < insn_cnt; i++, insn++) {
2806 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
2807 !bpf_pseudo_kfunc_call(insn))
2810 if (!env->bpf_capable) {
2811 verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
2815 if (bpf_pseudo_func(insn) || bpf_pseudo_call(insn))
2816 ret = add_subprog(env, i + insn->imm + 1);
2818 ret = add_kfunc_call(env, insn->imm, insn->off);
2824 /* Add a fake 'exit' subprog which could simplify subprog iteration
2825 * logic. 'subprog_cnt' should not be increased.
2827 subprog[env->subprog_cnt].start = insn_cnt;
2829 if (env->log.level & BPF_LOG_LEVEL2)
2830 for (i = 0; i < env->subprog_cnt; i++)
2831 verbose(env, "func#%d @%d\n", i, subprog[i].start);
2836 static int check_subprogs(struct bpf_verifier_env *env)
2838 int i, subprog_start, subprog_end, off, cur_subprog = 0;
2839 struct bpf_subprog_info *subprog = env->subprog_info;
2840 struct bpf_insn *insn = env->prog->insnsi;
2841 int insn_cnt = env->prog->len;
2843 /* now check that all jumps are within the same subprog */
2844 subprog_start = subprog[cur_subprog].start;
2845 subprog_end = subprog[cur_subprog + 1].start;
2846 for (i = 0; i < insn_cnt; i++) {
2847 u8 code = insn[i].code;
2849 if (code == (BPF_JMP | BPF_CALL) &&
2850 insn[i].src_reg == 0 &&
2851 insn[i].imm == BPF_FUNC_tail_call)
2852 subprog[cur_subprog].has_tail_call = true;
2853 if (BPF_CLASS(code) == BPF_LD &&
2854 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
2855 subprog[cur_subprog].has_ld_abs = true;
2856 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
2858 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
2860 if (code == (BPF_JMP32 | BPF_JA))
2861 off = i + insn[i].imm + 1;
2863 off = i + insn[i].off + 1;
2864 if (off < subprog_start || off >= subprog_end) {
2865 verbose(env, "jump out of range from insn %d to %d\n", i, off);
2869 if (i == subprog_end - 1) {
2870 /* to avoid fall-through from one subprog into another
2871 * the last insn of the subprog should be either exit
2872 * or unconditional jump back
2874 if (code != (BPF_JMP | BPF_EXIT) &&
2875 code != (BPF_JMP32 | BPF_JA) &&
2876 code != (BPF_JMP | BPF_JA)) {
2877 verbose(env, "last insn is not an exit or jmp\n");
2880 subprog_start = subprog_end;
2882 if (cur_subprog < env->subprog_cnt)
2883 subprog_end = subprog[cur_subprog + 1].start;
2889 /* Parentage chain of this register (or stack slot) should take care of all
2890 * issues like callee-saved registers, stack slot allocation time, etc.
2892 static int mark_reg_read(struct bpf_verifier_env *env,
2893 const struct bpf_reg_state *state,
2894 struct bpf_reg_state *parent, u8 flag)
2896 bool writes = parent == state->parent; /* Observe write marks */
2900 /* if read wasn't screened by an earlier write ... */
2901 if (writes && state->live & REG_LIVE_WRITTEN)
2903 if (parent->live & REG_LIVE_DONE) {
2904 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
2905 reg_type_str(env, parent->type),
2906 parent->var_off.value, parent->off);
2909 /* The first condition is more likely to be true than the
2910 * second, checked it first.
2912 if ((parent->live & REG_LIVE_READ) == flag ||
2913 parent->live & REG_LIVE_READ64)
2914 /* The parentage chain never changes and
2915 * this parent was already marked as LIVE_READ.
2916 * There is no need to keep walking the chain again and
2917 * keep re-marking all parents as LIVE_READ.
2918 * This case happens when the same register is read
2919 * multiple times without writes into it in-between.
2920 * Also, if parent has the stronger REG_LIVE_READ64 set,
2921 * then no need to set the weak REG_LIVE_READ32.
2924 /* ... then we depend on parent's value */
2925 parent->live |= flag;
2926 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
2927 if (flag == REG_LIVE_READ64)
2928 parent->live &= ~REG_LIVE_READ32;
2930 parent = state->parent;
2935 if (env->longest_mark_read_walk < cnt)
2936 env->longest_mark_read_walk = cnt;
2940 static int mark_dynptr_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
2942 struct bpf_func_state *state = func(env, reg);
2945 /* For CONST_PTR_TO_DYNPTR, it must have already been done by
2946 * check_reg_arg in check_helper_call and mark_btf_func_reg_size in
2949 if (reg->type == CONST_PTR_TO_DYNPTR)
2951 spi = dynptr_get_spi(env, reg);
2954 /* Caller ensures dynptr is valid and initialized, which means spi is in
2955 * bounds and spi is the first dynptr slot. Simply mark stack slot as
2958 ret = mark_reg_read(env, &state->stack[spi].spilled_ptr,
2959 state->stack[spi].spilled_ptr.parent, REG_LIVE_READ64);
2962 return mark_reg_read(env, &state->stack[spi - 1].spilled_ptr,
2963 state->stack[spi - 1].spilled_ptr.parent, REG_LIVE_READ64);
2966 static int mark_iter_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
2967 int spi, int nr_slots)
2969 struct bpf_func_state *state = func(env, reg);
2972 for (i = 0; i < nr_slots; i++) {
2973 struct bpf_reg_state *st = &state->stack[spi - i].spilled_ptr;
2975 err = mark_reg_read(env, st, st->parent, REG_LIVE_READ64);
2979 mark_stack_slot_scratched(env, spi - i);
2985 /* This function is supposed to be used by the following 32-bit optimization
2986 * code only. It returns TRUE if the source or destination register operates
2987 * on 64-bit, otherwise return FALSE.
2989 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
2990 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
2995 class = BPF_CLASS(code);
2997 if (class == BPF_JMP) {
2998 /* BPF_EXIT for "main" will reach here. Return TRUE
3003 if (op == BPF_CALL) {
3004 /* BPF to BPF call will reach here because of marking
3005 * caller saved clobber with DST_OP_NO_MARK for which we
3006 * don't care the register def because they are anyway
3007 * marked as NOT_INIT already.
3009 if (insn->src_reg == BPF_PSEUDO_CALL)
3011 /* Helper call will reach here because of arg type
3012 * check, conservatively return TRUE.
3021 if (class == BPF_ALU64 && op == BPF_END && (insn->imm == 16 || insn->imm == 32))
3024 if (class == BPF_ALU64 || class == BPF_JMP ||
3025 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
3028 if (class == BPF_ALU || class == BPF_JMP32)
3031 if (class == BPF_LDX) {
3033 return BPF_SIZE(code) == BPF_DW;
3034 /* LDX source must be ptr. */
3038 if (class == BPF_STX) {
3039 /* BPF_STX (including atomic variants) has multiple source
3040 * operands, one of which is a ptr. Check whether the caller is
3043 if (t == SRC_OP && reg->type != SCALAR_VALUE)
3045 return BPF_SIZE(code) == BPF_DW;
3048 if (class == BPF_LD) {
3049 u8 mode = BPF_MODE(code);
3052 if (mode == BPF_IMM)
3055 /* Both LD_IND and LD_ABS return 32-bit data. */
3059 /* Implicit ctx ptr. */
3060 if (regno == BPF_REG_6)
3063 /* Explicit source could be any width. */
3067 if (class == BPF_ST)
3068 /* The only source register for BPF_ST is a ptr. */
3071 /* Conservatively return true at default. */
3075 /* Return the regno defined by the insn, or -1. */
3076 static int insn_def_regno(const struct bpf_insn *insn)
3078 switch (BPF_CLASS(insn->code)) {
3084 if (BPF_MODE(insn->code) == BPF_ATOMIC &&
3085 (insn->imm & BPF_FETCH)) {
3086 if (insn->imm == BPF_CMPXCHG)
3089 return insn->src_reg;
3094 return insn->dst_reg;
3098 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
3099 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
3101 int dst_reg = insn_def_regno(insn);
3106 return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
3109 static void mark_insn_zext(struct bpf_verifier_env *env,
3110 struct bpf_reg_state *reg)
3112 s32 def_idx = reg->subreg_def;
3114 if (def_idx == DEF_NOT_SUBREG)
3117 env->insn_aux_data[def_idx - 1].zext_dst = true;
3118 /* The dst will be zero extended, so won't be sub-register anymore. */
3119 reg->subreg_def = DEF_NOT_SUBREG;
3122 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
3123 enum reg_arg_type t)
3125 struct bpf_verifier_state *vstate = env->cur_state;
3126 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3127 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
3128 struct bpf_reg_state *reg, *regs = state->regs;
3131 if (regno >= MAX_BPF_REG) {
3132 verbose(env, "R%d is invalid\n", regno);
3136 mark_reg_scratched(env, regno);
3139 rw64 = is_reg64(env, insn, regno, reg, t);
3141 /* check whether register used as source operand can be read */
3142 if (reg->type == NOT_INIT) {
3143 verbose(env, "R%d !read_ok\n", regno);
3146 /* We don't need to worry about FP liveness because it's read-only */
3147 if (regno == BPF_REG_FP)
3151 mark_insn_zext(env, reg);
3153 return mark_reg_read(env, reg, reg->parent,
3154 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
3156 /* check whether register used as dest operand can be written to */
3157 if (regno == BPF_REG_FP) {
3158 verbose(env, "frame pointer is read only\n");
3161 reg->live |= REG_LIVE_WRITTEN;
3162 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
3164 mark_reg_unknown(env, regs, regno);
3169 static void mark_jmp_point(struct bpf_verifier_env *env, int idx)
3171 env->insn_aux_data[idx].jmp_point = true;
3174 static bool is_jmp_point(struct bpf_verifier_env *env, int insn_idx)
3176 return env->insn_aux_data[insn_idx].jmp_point;
3179 /* for any branch, call, exit record the history of jmps in the given state */
3180 static int push_jmp_history(struct bpf_verifier_env *env,
3181 struct bpf_verifier_state *cur)
3183 u32 cnt = cur->jmp_history_cnt;
3184 struct bpf_idx_pair *p;
3187 if (!is_jmp_point(env, env->insn_idx))
3191 alloc_size = kmalloc_size_roundup(size_mul(cnt, sizeof(*p)));
3192 p = krealloc(cur->jmp_history, alloc_size, GFP_USER);
3195 p[cnt - 1].idx = env->insn_idx;
3196 p[cnt - 1].prev_idx = env->prev_insn_idx;
3197 cur->jmp_history = p;
3198 cur->jmp_history_cnt = cnt;
3202 /* Backtrack one insn at a time. If idx is not at the top of recorded
3203 * history then previous instruction came from straight line execution.
3205 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
3210 if (cnt && st->jmp_history[cnt - 1].idx == i) {
3211 i = st->jmp_history[cnt - 1].prev_idx;
3219 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
3221 const struct btf_type *func;
3222 struct btf *desc_btf;
3224 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
3227 desc_btf = find_kfunc_desc_btf(data, insn->off);
3228 if (IS_ERR(desc_btf))
3231 func = btf_type_by_id(desc_btf, insn->imm);
3232 return btf_name_by_offset(desc_btf, func->name_off);
3235 static inline void bt_init(struct backtrack_state *bt, u32 frame)
3240 static inline void bt_reset(struct backtrack_state *bt)
3242 struct bpf_verifier_env *env = bt->env;
3244 memset(bt, 0, sizeof(*bt));
3248 static inline u32 bt_empty(struct backtrack_state *bt)
3253 for (i = 0; i <= bt->frame; i++)
3254 mask |= bt->reg_masks[i] | bt->stack_masks[i];
3259 static inline int bt_subprog_enter(struct backtrack_state *bt)
3261 if (bt->frame == MAX_CALL_FRAMES - 1) {
3262 verbose(bt->env, "BUG subprog enter from frame %d\n", bt->frame);
3263 WARN_ONCE(1, "verifier backtracking bug");
3270 static inline int bt_subprog_exit(struct backtrack_state *bt)
3272 if (bt->frame == 0) {
3273 verbose(bt->env, "BUG subprog exit from frame 0\n");
3274 WARN_ONCE(1, "verifier backtracking bug");
3281 static inline void bt_set_frame_reg(struct backtrack_state *bt, u32 frame, u32 reg)
3283 bt->reg_masks[frame] |= 1 << reg;
3286 static inline void bt_clear_frame_reg(struct backtrack_state *bt, u32 frame, u32 reg)
3288 bt->reg_masks[frame] &= ~(1 << reg);
3291 static inline void bt_set_reg(struct backtrack_state *bt, u32 reg)
3293 bt_set_frame_reg(bt, bt->frame, reg);
3296 static inline void bt_clear_reg(struct backtrack_state *bt, u32 reg)
3298 bt_clear_frame_reg(bt, bt->frame, reg);
3301 static inline void bt_set_frame_slot(struct backtrack_state *bt, u32 frame, u32 slot)
3303 bt->stack_masks[frame] |= 1ull << slot;
3306 static inline void bt_clear_frame_slot(struct backtrack_state *bt, u32 frame, u32 slot)
3308 bt->stack_masks[frame] &= ~(1ull << slot);
3311 static inline void bt_set_slot(struct backtrack_state *bt, u32 slot)
3313 bt_set_frame_slot(bt, bt->frame, slot);
3316 static inline void bt_clear_slot(struct backtrack_state *bt, u32 slot)
3318 bt_clear_frame_slot(bt, bt->frame, slot);
3321 static inline u32 bt_frame_reg_mask(struct backtrack_state *bt, u32 frame)
3323 return bt->reg_masks[frame];
3326 static inline u32 bt_reg_mask(struct backtrack_state *bt)
3328 return bt->reg_masks[bt->frame];
3331 static inline u64 bt_frame_stack_mask(struct backtrack_state *bt, u32 frame)
3333 return bt->stack_masks[frame];
3336 static inline u64 bt_stack_mask(struct backtrack_state *bt)
3338 return bt->stack_masks[bt->frame];
3341 static inline bool bt_is_reg_set(struct backtrack_state *bt, u32 reg)
3343 return bt->reg_masks[bt->frame] & (1 << reg);
3346 static inline bool bt_is_slot_set(struct backtrack_state *bt, u32 slot)
3348 return bt->stack_masks[bt->frame] & (1ull << slot);
3351 /* format registers bitmask, e.g., "r0,r2,r4" for 0x15 mask */
3352 static void fmt_reg_mask(char *buf, ssize_t buf_sz, u32 reg_mask)
3354 DECLARE_BITMAP(mask, 64);
3360 bitmap_from_u64(mask, reg_mask);
3361 for_each_set_bit(i, mask, 32) {
3362 n = snprintf(buf, buf_sz, "%sr%d", first ? "" : ",", i);
3370 /* format stack slots bitmask, e.g., "-8,-24,-40" for 0x15 mask */
3371 static void fmt_stack_mask(char *buf, ssize_t buf_sz, u64 stack_mask)
3373 DECLARE_BITMAP(mask, 64);
3379 bitmap_from_u64(mask, stack_mask);
3380 for_each_set_bit(i, mask, 64) {
3381 n = snprintf(buf, buf_sz, "%s%d", first ? "" : ",", -(i + 1) * 8);
3390 /* For given verifier state backtrack_insn() is called from the last insn to
3391 * the first insn. Its purpose is to compute a bitmask of registers and
3392 * stack slots that needs precision in the parent verifier state.
3394 * @idx is an index of the instruction we are currently processing;
3395 * @subseq_idx is an index of the subsequent instruction that:
3396 * - *would be* executed next, if jump history is viewed in forward order;
3397 * - *was* processed previously during backtracking.
3399 static int backtrack_insn(struct bpf_verifier_env *env, int idx, int subseq_idx,
3400 struct backtrack_state *bt)
3402 const struct bpf_insn_cbs cbs = {
3403 .cb_call = disasm_kfunc_name,
3404 .cb_print = verbose,
3405 .private_data = env,
3407 struct bpf_insn *insn = env->prog->insnsi + idx;
3408 u8 class = BPF_CLASS(insn->code);
3409 u8 opcode = BPF_OP(insn->code);
3410 u8 mode = BPF_MODE(insn->code);
3411 u32 dreg = insn->dst_reg;
3412 u32 sreg = insn->src_reg;
3415 if (insn->code == 0)
3417 if (env->log.level & BPF_LOG_LEVEL2) {
3418 fmt_reg_mask(env->tmp_str_buf, TMP_STR_BUF_LEN, bt_reg_mask(bt));
3419 verbose(env, "mark_precise: frame%d: regs=%s ",
3420 bt->frame, env->tmp_str_buf);
3421 fmt_stack_mask(env->tmp_str_buf, TMP_STR_BUF_LEN, bt_stack_mask(bt));
3422 verbose(env, "stack=%s before ", env->tmp_str_buf);
3423 verbose(env, "%d: ", idx);
3424 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
3427 if (class == BPF_ALU || class == BPF_ALU64) {
3428 if (!bt_is_reg_set(bt, dreg))
3430 if (opcode == BPF_MOV) {
3431 if (BPF_SRC(insn->code) == BPF_X) {
3432 /* dreg = sreg or dreg = (s8, s16, s32)sreg
3433 * dreg needs precision after this insn
3434 * sreg needs precision before this insn
3436 bt_clear_reg(bt, dreg);
3437 bt_set_reg(bt, sreg);
3440 * dreg needs precision after this insn.
3441 * Corresponding register is already marked
3442 * as precise=true in this verifier state.
3443 * No further markings in parent are necessary
3445 bt_clear_reg(bt, dreg);
3448 if (BPF_SRC(insn->code) == BPF_X) {
3450 * both dreg and sreg need precision
3453 bt_set_reg(bt, sreg);
3455 * dreg still needs precision before this insn
3458 } else if (class == BPF_LDX) {
3459 if (!bt_is_reg_set(bt, dreg))
3461 bt_clear_reg(bt, dreg);
3463 /* scalars can only be spilled into stack w/o losing precision.
3464 * Load from any other memory can be zero extended.
3465 * The desire to keep that precision is already indicated
3466 * by 'precise' mark in corresponding register of this state.
3467 * No further tracking necessary.
3469 if (insn->src_reg != BPF_REG_FP)
3472 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
3473 * that [fp - off] slot contains scalar that needs to be
3474 * tracked with precision
3476 spi = (-insn->off - 1) / BPF_REG_SIZE;
3478 verbose(env, "BUG spi %d\n", spi);
3479 WARN_ONCE(1, "verifier backtracking bug");
3482 bt_set_slot(bt, spi);
3483 } else if (class == BPF_STX || class == BPF_ST) {
3484 if (bt_is_reg_set(bt, dreg))
3485 /* stx & st shouldn't be using _scalar_ dst_reg
3486 * to access memory. It means backtracking
3487 * encountered a case of pointer subtraction.
3490 /* scalars can only be spilled into stack */
3491 if (insn->dst_reg != BPF_REG_FP)
3493 spi = (-insn->off - 1) / BPF_REG_SIZE;
3495 verbose(env, "BUG spi %d\n", spi);
3496 WARN_ONCE(1, "verifier backtracking bug");
3499 if (!bt_is_slot_set(bt, spi))
3501 bt_clear_slot(bt, spi);
3502 if (class == BPF_STX)
3503 bt_set_reg(bt, sreg);
3504 } else if (class == BPF_JMP || class == BPF_JMP32) {
3505 if (bpf_pseudo_call(insn)) {
3506 int subprog_insn_idx, subprog;
3508 subprog_insn_idx = idx + insn->imm + 1;
3509 subprog = find_subprog(env, subprog_insn_idx);
3513 if (subprog_is_global(env, subprog)) {
3514 /* check that jump history doesn't have any
3515 * extra instructions from subprog; the next
3516 * instruction after call to global subprog
3517 * should be literally next instruction in
3520 WARN_ONCE(idx + 1 != subseq_idx, "verifier backtracking bug");
3521 /* r1-r5 are invalidated after subprog call,
3522 * so for global func call it shouldn't be set
3525 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3526 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3527 WARN_ONCE(1, "verifier backtracking bug");
3530 /* global subprog always sets R0 */
3531 bt_clear_reg(bt, BPF_REG_0);
3534 /* static subprog call instruction, which
3535 * means that we are exiting current subprog,
3536 * so only r1-r5 could be still requested as
3537 * precise, r0 and r6-r10 or any stack slot in
3538 * the current frame should be zero by now
3540 if (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) {
3541 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3542 WARN_ONCE(1, "verifier backtracking bug");
3545 /* we don't track register spills perfectly,
3546 * so fallback to force-precise instead of failing */
3547 if (bt_stack_mask(bt) != 0)
3549 /* propagate r1-r5 to the caller */
3550 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
3551 if (bt_is_reg_set(bt, i)) {
3552 bt_clear_reg(bt, i);
3553 bt_set_frame_reg(bt, bt->frame - 1, i);
3556 if (bt_subprog_exit(bt))
3560 } else if ((bpf_helper_call(insn) &&
3561 is_callback_calling_function(insn->imm) &&
3562 !is_async_callback_calling_function(insn->imm)) ||
3563 (bpf_pseudo_kfunc_call(insn) && is_callback_calling_kfunc(insn->imm))) {
3564 /* callback-calling helper or kfunc call, which means
3565 * we are exiting from subprog, but unlike the subprog
3566 * call handling above, we shouldn't propagate
3567 * precision of r1-r5 (if any requested), as they are
3568 * not actually arguments passed directly to callback
3571 if (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) {
3572 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3573 WARN_ONCE(1, "verifier backtracking bug");
3576 if (bt_stack_mask(bt) != 0)
3578 /* clear r1-r5 in callback subprog's mask */
3579 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
3580 bt_clear_reg(bt, i);
3581 if (bt_subprog_exit(bt))
3584 } else if (opcode == BPF_CALL) {
3585 /* kfunc with imm==0 is invalid and fixup_kfunc_call will
3586 * catch this error later. Make backtracking conservative
3589 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && insn->imm == 0)
3591 /* regular helper call sets R0 */
3592 bt_clear_reg(bt, BPF_REG_0);
3593 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3594 /* if backtracing was looking for registers R1-R5
3595 * they should have been found already.
3597 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3598 WARN_ONCE(1, "verifier backtracking bug");
3601 } else if (opcode == BPF_EXIT) {
3604 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3605 /* if backtracing was looking for registers R1-R5
3606 * they should have been found already.
3608 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3609 WARN_ONCE(1, "verifier backtracking bug");
3613 /* BPF_EXIT in subprog or callback always returns
3614 * right after the call instruction, so by checking
3615 * whether the instruction at subseq_idx-1 is subprog
3616 * call or not we can distinguish actual exit from
3617 * *subprog* from exit from *callback*. In the former
3618 * case, we need to propagate r0 precision, if
3619 * necessary. In the former we never do that.
3621 r0_precise = subseq_idx - 1 >= 0 &&
3622 bpf_pseudo_call(&env->prog->insnsi[subseq_idx - 1]) &&
3623 bt_is_reg_set(bt, BPF_REG_0);
3625 bt_clear_reg(bt, BPF_REG_0);
3626 if (bt_subprog_enter(bt))
3630 bt_set_reg(bt, BPF_REG_0);
3631 /* r6-r9 and stack slots will stay set in caller frame
3632 * bitmasks until we return back from callee(s)
3635 } else if (BPF_SRC(insn->code) == BPF_X) {
3636 if (!bt_is_reg_set(bt, dreg) && !bt_is_reg_set(bt, sreg))
3639 * Both dreg and sreg need precision before
3640 * this insn. If only sreg was marked precise
3641 * before it would be equally necessary to
3642 * propagate it to dreg.
3644 bt_set_reg(bt, dreg);
3645 bt_set_reg(bt, sreg);
3646 /* else dreg <cond> K
3647 * Only dreg still needs precision before
3648 * this insn, so for the K-based conditional
3649 * there is nothing new to be marked.
3652 } else if (class == BPF_LD) {
3653 if (!bt_is_reg_set(bt, dreg))
3655 bt_clear_reg(bt, dreg);
3656 /* It's ld_imm64 or ld_abs or ld_ind.
3657 * For ld_imm64 no further tracking of precision
3658 * into parent is necessary
3660 if (mode == BPF_IND || mode == BPF_ABS)
3661 /* to be analyzed */
3667 /* the scalar precision tracking algorithm:
3668 * . at the start all registers have precise=false.
3669 * . scalar ranges are tracked as normal through alu and jmp insns.
3670 * . once precise value of the scalar register is used in:
3671 * . ptr + scalar alu
3672 * . if (scalar cond K|scalar)
3673 * . helper_call(.., scalar, ...) where ARG_CONST is expected
3674 * backtrack through the verifier states and mark all registers and
3675 * stack slots with spilled constants that these scalar regisers
3676 * should be precise.
3677 * . during state pruning two registers (or spilled stack slots)
3678 * are equivalent if both are not precise.
3680 * Note the verifier cannot simply walk register parentage chain,
3681 * since many different registers and stack slots could have been
3682 * used to compute single precise scalar.
3684 * The approach of starting with precise=true for all registers and then
3685 * backtrack to mark a register as not precise when the verifier detects
3686 * that program doesn't care about specific value (e.g., when helper
3687 * takes register as ARG_ANYTHING parameter) is not safe.
3689 * It's ok to walk single parentage chain of the verifier states.
3690 * It's possible that this backtracking will go all the way till 1st insn.
3691 * All other branches will be explored for needing precision later.
3693 * The backtracking needs to deal with cases like:
3694 * 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)
3697 * if r5 > 0x79f goto pc+7
3698 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
3701 * call bpf_perf_event_output#25
3702 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
3706 * call foo // uses callee's r6 inside to compute r0
3710 * to track above reg_mask/stack_mask needs to be independent for each frame.
3712 * Also if parent's curframe > frame where backtracking started,
3713 * the verifier need to mark registers in both frames, otherwise callees
3714 * may incorrectly prune callers. This is similar to
3715 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
3717 * For now backtracking falls back into conservative marking.
3719 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
3720 struct bpf_verifier_state *st)
3722 struct bpf_func_state *func;
3723 struct bpf_reg_state *reg;
3726 if (env->log.level & BPF_LOG_LEVEL2) {
3727 verbose(env, "mark_precise: frame%d: falling back to forcing all scalars precise\n",
3731 /* big hammer: mark all scalars precise in this path.
3732 * pop_stack may still get !precise scalars.
3733 * We also skip current state and go straight to first parent state,
3734 * because precision markings in current non-checkpointed state are
3735 * not needed. See why in the comment in __mark_chain_precision below.
3737 for (st = st->parent; st; st = st->parent) {
3738 for (i = 0; i <= st->curframe; i++) {
3739 func = st->frame[i];
3740 for (j = 0; j < BPF_REG_FP; j++) {
3741 reg = &func->regs[j];
3742 if (reg->type != SCALAR_VALUE || reg->precise)
3744 reg->precise = true;
3745 if (env->log.level & BPF_LOG_LEVEL2) {
3746 verbose(env, "force_precise: frame%d: forcing r%d to be precise\n",
3750 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
3751 if (!is_spilled_reg(&func->stack[j]))
3753 reg = &func->stack[j].spilled_ptr;
3754 if (reg->type != SCALAR_VALUE || reg->precise)
3756 reg->precise = true;
3757 if (env->log.level & BPF_LOG_LEVEL2) {
3758 verbose(env, "force_precise: frame%d: forcing fp%d to be precise\n",
3766 static void mark_all_scalars_imprecise(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
3768 struct bpf_func_state *func;
3769 struct bpf_reg_state *reg;
3772 for (i = 0; i <= st->curframe; i++) {
3773 func = st->frame[i];
3774 for (j = 0; j < BPF_REG_FP; j++) {
3775 reg = &func->regs[j];
3776 if (reg->type != SCALAR_VALUE)
3778 reg->precise = false;
3780 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
3781 if (!is_spilled_reg(&func->stack[j]))
3783 reg = &func->stack[j].spilled_ptr;
3784 if (reg->type != SCALAR_VALUE)
3786 reg->precise = false;
3791 static bool idset_contains(struct bpf_idset *s, u32 id)
3795 for (i = 0; i < s->count; ++i)
3796 if (s->ids[i] == id)
3802 static int idset_push(struct bpf_idset *s, u32 id)
3804 if (WARN_ON_ONCE(s->count >= ARRAY_SIZE(s->ids)))
3806 s->ids[s->count++] = id;
3810 static void idset_reset(struct bpf_idset *s)
3815 /* Collect a set of IDs for all registers currently marked as precise in env->bt.
3816 * Mark all registers with these IDs as precise.
3818 static int mark_precise_scalar_ids(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
3820 struct bpf_idset *precise_ids = &env->idset_scratch;
3821 struct backtrack_state *bt = &env->bt;
3822 struct bpf_func_state *func;
3823 struct bpf_reg_state *reg;
3824 DECLARE_BITMAP(mask, 64);
3827 idset_reset(precise_ids);
3829 for (fr = bt->frame; fr >= 0; fr--) {
3830 func = st->frame[fr];
3832 bitmap_from_u64(mask, bt_frame_reg_mask(bt, fr));
3833 for_each_set_bit(i, mask, 32) {
3834 reg = &func->regs[i];
3835 if (!reg->id || reg->type != SCALAR_VALUE)
3837 if (idset_push(precise_ids, reg->id))
3841 bitmap_from_u64(mask, bt_frame_stack_mask(bt, fr));
3842 for_each_set_bit(i, mask, 64) {
3843 if (i >= func->allocated_stack / BPF_REG_SIZE)
3845 if (!is_spilled_scalar_reg(&func->stack[i]))
3847 reg = &func->stack[i].spilled_ptr;
3850 if (idset_push(precise_ids, reg->id))
3855 for (fr = 0; fr <= st->curframe; ++fr) {
3856 func = st->frame[fr];
3858 for (i = BPF_REG_0; i < BPF_REG_10; ++i) {
3859 reg = &func->regs[i];
3862 if (!idset_contains(precise_ids, reg->id))
3864 bt_set_frame_reg(bt, fr, i);
3866 for (i = 0; i < func->allocated_stack / BPF_REG_SIZE; ++i) {
3867 if (!is_spilled_scalar_reg(&func->stack[i]))
3869 reg = &func->stack[i].spilled_ptr;
3872 if (!idset_contains(precise_ids, reg->id))
3874 bt_set_frame_slot(bt, fr, i);
3882 * __mark_chain_precision() backtracks BPF program instruction sequence and
3883 * chain of verifier states making sure that register *regno* (if regno >= 0)
3884 * and/or stack slot *spi* (if spi >= 0) are marked as precisely tracked
3885 * SCALARS, as well as any other registers and slots that contribute to
3886 * a tracked state of given registers/stack slots, depending on specific BPF
3887 * assembly instructions (see backtrack_insns() for exact instruction handling
3888 * logic). This backtracking relies on recorded jmp_history and is able to
3889 * traverse entire chain of parent states. This process ends only when all the
3890 * necessary registers/slots and their transitive dependencies are marked as
3893 * One important and subtle aspect is that precise marks *do not matter* in
3894 * the currently verified state (current state). It is important to understand
3895 * why this is the case.
3897 * First, note that current state is the state that is not yet "checkpointed",
3898 * i.e., it is not yet put into env->explored_states, and it has no children
3899 * states as well. It's ephemeral, and can end up either a) being discarded if
3900 * compatible explored state is found at some point or BPF_EXIT instruction is
3901 * reached or b) checkpointed and put into env->explored_states, branching out
3902 * into one or more children states.
3904 * In the former case, precise markings in current state are completely
3905 * ignored by state comparison code (see regsafe() for details). Only
3906 * checkpointed ("old") state precise markings are important, and if old
3907 * state's register/slot is precise, regsafe() assumes current state's
3908 * register/slot as precise and checks value ranges exactly and precisely. If
3909 * states turn out to be compatible, current state's necessary precise
3910 * markings and any required parent states' precise markings are enforced
3911 * after the fact with propagate_precision() logic, after the fact. But it's
3912 * important to realize that in this case, even after marking current state
3913 * registers/slots as precise, we immediately discard current state. So what
3914 * actually matters is any of the precise markings propagated into current
3915 * state's parent states, which are always checkpointed (due to b) case above).
3916 * As such, for scenario a) it doesn't matter if current state has precise
3917 * markings set or not.
3919 * Now, for the scenario b), checkpointing and forking into child(ren)
3920 * state(s). Note that before current state gets to checkpointing step, any
3921 * processed instruction always assumes precise SCALAR register/slot
3922 * knowledge: if precise value or range is useful to prune jump branch, BPF
3923 * verifier takes this opportunity enthusiastically. Similarly, when
3924 * register's value is used to calculate offset or memory address, exact
3925 * knowledge of SCALAR range is assumed, checked, and enforced. So, similar to
3926 * what we mentioned above about state comparison ignoring precise markings
3927 * during state comparison, BPF verifier ignores and also assumes precise
3928 * markings *at will* during instruction verification process. But as verifier
3929 * assumes precision, it also propagates any precision dependencies across
3930 * parent states, which are not yet finalized, so can be further restricted
3931 * based on new knowledge gained from restrictions enforced by their children
3932 * states. This is so that once those parent states are finalized, i.e., when
3933 * they have no more active children state, state comparison logic in
3934 * is_state_visited() would enforce strict and precise SCALAR ranges, if
3935 * required for correctness.
3937 * To build a bit more intuition, note also that once a state is checkpointed,
3938 * the path we took to get to that state is not important. This is crucial
3939 * property for state pruning. When state is checkpointed and finalized at
3940 * some instruction index, it can be correctly and safely used to "short
3941 * circuit" any *compatible* state that reaches exactly the same instruction
3942 * index. I.e., if we jumped to that instruction from a completely different
3943 * code path than original finalized state was derived from, it doesn't
3944 * matter, current state can be discarded because from that instruction
3945 * forward having a compatible state will ensure we will safely reach the
3946 * exit. States describe preconditions for further exploration, but completely
3947 * forget the history of how we got here.
3949 * This also means that even if we needed precise SCALAR range to get to
3950 * finalized state, but from that point forward *that same* SCALAR register is
3951 * never used in a precise context (i.e., it's precise value is not needed for
3952 * correctness), it's correct and safe to mark such register as "imprecise"
3953 * (i.e., precise marking set to false). This is what we rely on when we do
3954 * not set precise marking in current state. If no child state requires
3955 * precision for any given SCALAR register, it's safe to dictate that it can
3956 * be imprecise. If any child state does require this register to be precise,
3957 * we'll mark it precise later retroactively during precise markings
3958 * propagation from child state to parent states.
3960 * Skipping precise marking setting in current state is a mild version of
3961 * relying on the above observation. But we can utilize this property even
3962 * more aggressively by proactively forgetting any precise marking in the
3963 * current state (which we inherited from the parent state), right before we
3964 * checkpoint it and branch off into new child state. This is done by
3965 * mark_all_scalars_imprecise() to hopefully get more permissive and generic
3966 * finalized states which help in short circuiting more future states.
3968 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno)
3970 struct backtrack_state *bt = &env->bt;
3971 struct bpf_verifier_state *st = env->cur_state;
3972 int first_idx = st->first_insn_idx;
3973 int last_idx = env->insn_idx;
3974 int subseq_idx = -1;
3975 struct bpf_func_state *func;
3976 struct bpf_reg_state *reg;
3977 bool skip_first = true;
3980 if (!env->bpf_capable)
3983 /* set frame number from which we are starting to backtrack */
3984 bt_init(bt, env->cur_state->curframe);
3986 /* Do sanity checks against current state of register and/or stack
3987 * slot, but don't set precise flag in current state, as precision
3988 * tracking in the current state is unnecessary.
3990 func = st->frame[bt->frame];
3992 reg = &func->regs[regno];
3993 if (reg->type != SCALAR_VALUE) {
3994 WARN_ONCE(1, "backtracing misuse");
3997 bt_set_reg(bt, regno);
4004 DECLARE_BITMAP(mask, 64);
4005 u32 history = st->jmp_history_cnt;
4007 if (env->log.level & BPF_LOG_LEVEL2) {
4008 verbose(env, "mark_precise: frame%d: last_idx %d first_idx %d subseq_idx %d \n",
4009 bt->frame, last_idx, first_idx, subseq_idx);
4012 /* If some register with scalar ID is marked as precise,
4013 * make sure that all registers sharing this ID are also precise.
4014 * This is needed to estimate effect of find_equal_scalars().
4015 * Do this at the last instruction of each state,
4016 * bpf_reg_state::id fields are valid for these instructions.
4018 * Allows to track precision in situation like below:
4020 * r2 = unknown value
4024 * r1 = r2 // r1 and r2 now share the same ID
4026 * --- state #1 {r1.id = A, r2.id = A} ---
4028 * if (r2 > 10) goto exit; // find_equal_scalars() assigns range to r1
4030 * --- state #2 {r1.id = A, r2.id = A} ---
4032 * r3 += r1 // need to mark both r1 and r2
4034 if (mark_precise_scalar_ids(env, st))
4038 /* we are at the entry into subprog, which
4039 * is expected for global funcs, but only if
4040 * requested precise registers are R1-R5
4041 * (which are global func's input arguments)
4043 if (st->curframe == 0 &&
4044 st->frame[0]->subprogno > 0 &&
4045 st->frame[0]->callsite == BPF_MAIN_FUNC &&
4046 bt_stack_mask(bt) == 0 &&
4047 (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) == 0) {
4048 bitmap_from_u64(mask, bt_reg_mask(bt));
4049 for_each_set_bit(i, mask, 32) {
4050 reg = &st->frame[0]->regs[i];
4051 bt_clear_reg(bt, i);
4052 if (reg->type == SCALAR_VALUE)
4053 reg->precise = true;
4058 verbose(env, "BUG backtracking func entry subprog %d reg_mask %x stack_mask %llx\n",
4059 st->frame[0]->subprogno, bt_reg_mask(bt), bt_stack_mask(bt));
4060 WARN_ONCE(1, "verifier backtracking bug");
4064 for (i = last_idx;;) {
4069 err = backtrack_insn(env, i, subseq_idx, bt);
4071 if (err == -ENOTSUPP) {
4072 mark_all_scalars_precise(env, env->cur_state);
4079 /* Found assignment(s) into tracked register in this state.
4080 * Since this state is already marked, just return.
4081 * Nothing to be tracked further in the parent state.
4087 i = get_prev_insn_idx(st, i, &history);
4088 if (i >= env->prog->len) {
4089 /* This can happen if backtracking reached insn 0
4090 * and there are still reg_mask or stack_mask
4092 * It means the backtracking missed the spot where
4093 * particular register was initialized with a constant.
4095 verbose(env, "BUG backtracking idx %d\n", i);
4096 WARN_ONCE(1, "verifier backtracking bug");
4104 for (fr = bt->frame; fr >= 0; fr--) {
4105 func = st->frame[fr];
4106 bitmap_from_u64(mask, bt_frame_reg_mask(bt, fr));
4107 for_each_set_bit(i, mask, 32) {
4108 reg = &func->regs[i];
4109 if (reg->type != SCALAR_VALUE) {
4110 bt_clear_frame_reg(bt, fr, i);
4114 bt_clear_frame_reg(bt, fr, i);
4116 reg->precise = true;
4119 bitmap_from_u64(mask, bt_frame_stack_mask(bt, fr));
4120 for_each_set_bit(i, mask, 64) {
4121 if (i >= func->allocated_stack / BPF_REG_SIZE) {
4122 /* the sequence of instructions:
4124 * 3: (7b) *(u64 *)(r3 -8) = r0
4125 * 4: (79) r4 = *(u64 *)(r10 -8)
4126 * doesn't contain jmps. It's backtracked
4127 * as a single block.
4128 * During backtracking insn 3 is not recognized as
4129 * stack access, so at the end of backtracking
4130 * stack slot fp-8 is still marked in stack_mask.
4131 * However the parent state may not have accessed
4132 * fp-8 and it's "unallocated" stack space.
4133 * In such case fallback to conservative.
4135 mark_all_scalars_precise(env, env->cur_state);
4140 if (!is_spilled_scalar_reg(&func->stack[i])) {
4141 bt_clear_frame_slot(bt, fr, i);
4144 reg = &func->stack[i].spilled_ptr;
4146 bt_clear_frame_slot(bt, fr, i);
4148 reg->precise = true;
4150 if (env->log.level & BPF_LOG_LEVEL2) {
4151 fmt_reg_mask(env->tmp_str_buf, TMP_STR_BUF_LEN,
4152 bt_frame_reg_mask(bt, fr));
4153 verbose(env, "mark_precise: frame%d: parent state regs=%s ",
4154 fr, env->tmp_str_buf);
4155 fmt_stack_mask(env->tmp_str_buf, TMP_STR_BUF_LEN,
4156 bt_frame_stack_mask(bt, fr));
4157 verbose(env, "stack=%s: ", env->tmp_str_buf);
4158 print_verifier_state(env, func, true);
4165 subseq_idx = first_idx;
4166 last_idx = st->last_insn_idx;
4167 first_idx = st->first_insn_idx;
4170 /* if we still have requested precise regs or slots, we missed
4171 * something (e.g., stack access through non-r10 register), so
4172 * fallback to marking all precise
4174 if (!bt_empty(bt)) {
4175 mark_all_scalars_precise(env, env->cur_state);
4182 int mark_chain_precision(struct bpf_verifier_env *env, int regno)
4184 return __mark_chain_precision(env, regno);
4187 /* mark_chain_precision_batch() assumes that env->bt is set in the caller to
4188 * desired reg and stack masks across all relevant frames
4190 static int mark_chain_precision_batch(struct bpf_verifier_env *env)
4192 return __mark_chain_precision(env, -1);
4195 static bool is_spillable_regtype(enum bpf_reg_type type)
4197 switch (base_type(type)) {
4198 case PTR_TO_MAP_VALUE:
4202 case PTR_TO_PACKET_META:
4203 case PTR_TO_PACKET_END:
4204 case PTR_TO_FLOW_KEYS:
4205 case CONST_PTR_TO_MAP:
4207 case PTR_TO_SOCK_COMMON:
4208 case PTR_TO_TCP_SOCK:
4209 case PTR_TO_XDP_SOCK:
4214 case PTR_TO_MAP_KEY:
4221 /* Does this register contain a constant zero? */
4222 static bool register_is_null(struct bpf_reg_state *reg)
4224 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
4227 static bool register_is_const(struct bpf_reg_state *reg)
4229 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
4232 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
4234 return tnum_is_unknown(reg->var_off) &&
4235 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
4236 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
4237 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
4238 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
4241 static bool register_is_bounded(struct bpf_reg_state *reg)
4243 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
4246 static bool __is_pointer_value(bool allow_ptr_leaks,
4247 const struct bpf_reg_state *reg)
4249 if (allow_ptr_leaks)
4252 return reg->type != SCALAR_VALUE;
4255 /* Copy src state preserving dst->parent and dst->live fields */
4256 static void copy_register_state(struct bpf_reg_state *dst, const struct bpf_reg_state *src)
4258 struct bpf_reg_state *parent = dst->parent;
4259 enum bpf_reg_liveness live = dst->live;
4262 dst->parent = parent;
4266 static void save_register_state(struct bpf_func_state *state,
4267 int spi, struct bpf_reg_state *reg,
4272 copy_register_state(&state->stack[spi].spilled_ptr, reg);
4273 if (size == BPF_REG_SIZE)
4274 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
4276 for (i = BPF_REG_SIZE; i > BPF_REG_SIZE - size; i--)
4277 state->stack[spi].slot_type[i - 1] = STACK_SPILL;
4279 /* size < 8 bytes spill */
4281 scrub_spilled_slot(&state->stack[spi].slot_type[i - 1]);
4284 static bool is_bpf_st_mem(struct bpf_insn *insn)
4286 return BPF_CLASS(insn->code) == BPF_ST && BPF_MODE(insn->code) == BPF_MEM;
4289 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
4290 * stack boundary and alignment are checked in check_mem_access()
4292 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
4293 /* stack frame we're writing to */
4294 struct bpf_func_state *state,
4295 int off, int size, int value_regno,
4298 struct bpf_func_state *cur; /* state of the current function */
4299 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
4300 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
4301 struct bpf_reg_state *reg = NULL;
4302 u32 dst_reg = insn->dst_reg;
4304 err = grow_stack_state(state, round_up(slot + 1, BPF_REG_SIZE));
4307 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
4308 * so it's aligned access and [off, off + size) are within stack limits
4310 if (!env->allow_ptr_leaks &&
4311 state->stack[spi].slot_type[0] == STACK_SPILL &&
4312 size != BPF_REG_SIZE) {
4313 verbose(env, "attempt to corrupt spilled pointer on stack\n");
4317 cur = env->cur_state->frame[env->cur_state->curframe];
4318 if (value_regno >= 0)
4319 reg = &cur->regs[value_regno];
4320 if (!env->bypass_spec_v4) {
4321 bool sanitize = reg && is_spillable_regtype(reg->type);
4323 for (i = 0; i < size; i++) {
4324 u8 type = state->stack[spi].slot_type[i];
4326 if (type != STACK_MISC && type != STACK_ZERO) {
4333 env->insn_aux_data[insn_idx].sanitize_stack_spill = true;
4336 err = destroy_if_dynptr_stack_slot(env, state, spi);
4340 mark_stack_slot_scratched(env, spi);
4341 if (reg && !(off % BPF_REG_SIZE) && register_is_bounded(reg) &&
4342 !register_is_null(reg) && env->bpf_capable) {
4343 if (dst_reg != BPF_REG_FP) {
4344 /* The backtracking logic can only recognize explicit
4345 * stack slot address like [fp - 8]. Other spill of
4346 * scalar via different register has to be conservative.
4347 * Backtrack from here and mark all registers as precise
4348 * that contributed into 'reg' being a constant.
4350 err = mark_chain_precision(env, value_regno);
4354 save_register_state(state, spi, reg, size);
4355 /* Break the relation on a narrowing spill. */
4356 if (fls64(reg->umax_value) > BITS_PER_BYTE * size)
4357 state->stack[spi].spilled_ptr.id = 0;
4358 } else if (!reg && !(off % BPF_REG_SIZE) && is_bpf_st_mem(insn) &&
4359 insn->imm != 0 && env->bpf_capable) {
4360 struct bpf_reg_state fake_reg = {};
4362 __mark_reg_known(&fake_reg, (u32)insn->imm);
4363 fake_reg.type = SCALAR_VALUE;
4364 save_register_state(state, spi, &fake_reg, size);
4365 } else if (reg && is_spillable_regtype(reg->type)) {
4366 /* register containing pointer is being spilled into stack */
4367 if (size != BPF_REG_SIZE) {
4368 verbose_linfo(env, insn_idx, "; ");
4369 verbose(env, "invalid size of register spill\n");
4372 if (state != cur && reg->type == PTR_TO_STACK) {
4373 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
4376 save_register_state(state, spi, reg, size);
4378 u8 type = STACK_MISC;
4380 /* regular write of data into stack destroys any spilled ptr */
4381 state->stack[spi].spilled_ptr.type = NOT_INIT;
4382 /* Mark slots as STACK_MISC if they belonged to spilled ptr/dynptr/iter. */
4383 if (is_stack_slot_special(&state->stack[spi]))
4384 for (i = 0; i < BPF_REG_SIZE; i++)
4385 scrub_spilled_slot(&state->stack[spi].slot_type[i]);
4387 /* only mark the slot as written if all 8 bytes were written
4388 * otherwise read propagation may incorrectly stop too soon
4389 * when stack slots are partially written.
4390 * This heuristic means that read propagation will be
4391 * conservative, since it will add reg_live_read marks
4392 * to stack slots all the way to first state when programs
4393 * writes+reads less than 8 bytes
4395 if (size == BPF_REG_SIZE)
4396 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
4398 /* when we zero initialize stack slots mark them as such */
4399 if ((reg && register_is_null(reg)) ||
4400 (!reg && is_bpf_st_mem(insn) && insn->imm == 0)) {
4401 /* backtracking doesn't work for STACK_ZERO yet. */
4402 err = mark_chain_precision(env, value_regno);
4408 /* Mark slots affected by this stack write. */
4409 for (i = 0; i < size; i++)
4410 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
4416 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
4417 * known to contain a variable offset.
4418 * This function checks whether the write is permitted and conservatively
4419 * tracks the effects of the write, considering that each stack slot in the
4420 * dynamic range is potentially written to.
4422 * 'off' includes 'regno->off'.
4423 * 'value_regno' can be -1, meaning that an unknown value is being written to
4426 * Spilled pointers in range are not marked as written because we don't know
4427 * what's going to be actually written. This means that read propagation for
4428 * future reads cannot be terminated by this write.
4430 * For privileged programs, uninitialized stack slots are considered
4431 * initialized by this write (even though we don't know exactly what offsets
4432 * are going to be written to). The idea is that we don't want the verifier to
4433 * reject future reads that access slots written to through variable offsets.
4435 static int check_stack_write_var_off(struct bpf_verifier_env *env,
4436 /* func where register points to */
4437 struct bpf_func_state *state,
4438 int ptr_regno, int off, int size,
4439 int value_regno, int insn_idx)
4441 struct bpf_func_state *cur; /* state of the current function */
4442 int min_off, max_off;
4444 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
4445 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
4446 bool writing_zero = false;
4447 /* set if the fact that we're writing a zero is used to let any
4448 * stack slots remain STACK_ZERO
4450 bool zero_used = false;
4452 cur = env->cur_state->frame[env->cur_state->curframe];
4453 ptr_reg = &cur->regs[ptr_regno];
4454 min_off = ptr_reg->smin_value + off;
4455 max_off = ptr_reg->smax_value + off + size;
4456 if (value_regno >= 0)
4457 value_reg = &cur->regs[value_regno];
4458 if ((value_reg && register_is_null(value_reg)) ||
4459 (!value_reg && is_bpf_st_mem(insn) && insn->imm == 0))
4460 writing_zero = true;
4462 err = grow_stack_state(state, round_up(-min_off, BPF_REG_SIZE));
4466 for (i = min_off; i < max_off; i++) {
4470 err = destroy_if_dynptr_stack_slot(env, state, spi);
4475 /* Variable offset writes destroy any spilled pointers in range. */
4476 for (i = min_off; i < max_off; i++) {
4477 u8 new_type, *stype;
4481 spi = slot / BPF_REG_SIZE;
4482 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
4483 mark_stack_slot_scratched(env, spi);
4485 if (!env->allow_ptr_leaks && *stype != STACK_MISC && *stype != STACK_ZERO) {
4486 /* Reject the write if range we may write to has not
4487 * been initialized beforehand. If we didn't reject
4488 * here, the ptr status would be erased below (even
4489 * though not all slots are actually overwritten),
4490 * possibly opening the door to leaks.
4492 * We do however catch STACK_INVALID case below, and
4493 * only allow reading possibly uninitialized memory
4494 * later for CAP_PERFMON, as the write may not happen to
4497 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
4502 /* Erase all spilled pointers. */
4503 state->stack[spi].spilled_ptr.type = NOT_INIT;
4505 /* Update the slot type. */
4506 new_type = STACK_MISC;
4507 if (writing_zero && *stype == STACK_ZERO) {
4508 new_type = STACK_ZERO;
4511 /* If the slot is STACK_INVALID, we check whether it's OK to
4512 * pretend that it will be initialized by this write. The slot
4513 * might not actually be written to, and so if we mark it as
4514 * initialized future reads might leak uninitialized memory.
4515 * For privileged programs, we will accept such reads to slots
4516 * that may or may not be written because, if we're reject
4517 * them, the error would be too confusing.
4519 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
4520 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
4527 /* backtracking doesn't work for STACK_ZERO yet. */
4528 err = mark_chain_precision(env, value_regno);
4535 /* When register 'dst_regno' is assigned some values from stack[min_off,
4536 * max_off), we set the register's type according to the types of the
4537 * respective stack slots. If all the stack values are known to be zeros, then
4538 * so is the destination reg. Otherwise, the register is considered to be
4539 * SCALAR. This function does not deal with register filling; the caller must
4540 * ensure that all spilled registers in the stack range have been marked as
4543 static void mark_reg_stack_read(struct bpf_verifier_env *env,
4544 /* func where src register points to */
4545 struct bpf_func_state *ptr_state,
4546 int min_off, int max_off, int dst_regno)
4548 struct bpf_verifier_state *vstate = env->cur_state;
4549 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4554 for (i = min_off; i < max_off; i++) {
4556 spi = slot / BPF_REG_SIZE;
4557 mark_stack_slot_scratched(env, spi);
4558 stype = ptr_state->stack[spi].slot_type;
4559 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
4563 if (zeros == max_off - min_off) {
4564 /* any access_size read into register is zero extended,
4565 * so the whole register == const_zero
4567 __mark_reg_const_zero(&state->regs[dst_regno]);
4568 /* backtracking doesn't support STACK_ZERO yet,
4569 * so mark it precise here, so that later
4570 * backtracking can stop here.
4571 * Backtracking may not need this if this register
4572 * doesn't participate in pointer adjustment.
4573 * Forward propagation of precise flag is not
4574 * necessary either. This mark is only to stop
4575 * backtracking. Any register that contributed
4576 * to const 0 was marked precise before spill.
4578 state->regs[dst_regno].precise = true;
4580 /* have read misc data from the stack */
4581 mark_reg_unknown(env, state->regs, dst_regno);
4583 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4586 /* Read the stack at 'off' and put the results into the register indicated by
4587 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
4590 * 'dst_regno' can be -1, meaning that the read value is not going to a
4593 * The access is assumed to be within the current stack bounds.
4595 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
4596 /* func where src register points to */
4597 struct bpf_func_state *reg_state,
4598 int off, int size, int dst_regno)
4600 struct bpf_verifier_state *vstate = env->cur_state;
4601 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4602 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
4603 struct bpf_reg_state *reg;
4606 stype = reg_state->stack[spi].slot_type;
4607 reg = ®_state->stack[spi].spilled_ptr;
4609 mark_stack_slot_scratched(env, spi);
4611 if (is_spilled_reg(®_state->stack[spi])) {
4614 for (i = BPF_REG_SIZE - 1; i > 0 && stype[i - 1] == STACK_SPILL; i--)
4617 if (size != BPF_REG_SIZE || spill_size != BPF_REG_SIZE) {
4618 if (reg->type != SCALAR_VALUE) {
4619 verbose_linfo(env, env->insn_idx, "; ");
4620 verbose(env, "invalid size of register fill\n");
4624 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4628 if (!(off % BPF_REG_SIZE) && size == spill_size) {
4629 /* The earlier check_reg_arg() has decided the
4630 * subreg_def for this insn. Save it first.
4632 s32 subreg_def = state->regs[dst_regno].subreg_def;
4634 copy_register_state(&state->regs[dst_regno], reg);
4635 state->regs[dst_regno].subreg_def = subreg_def;
4637 for (i = 0; i < size; i++) {
4638 type = stype[(slot - i) % BPF_REG_SIZE];
4639 if (type == STACK_SPILL)
4641 if (type == STACK_MISC)
4643 if (type == STACK_INVALID && env->allow_uninit_stack)
4645 verbose(env, "invalid read from stack off %d+%d size %d\n",
4649 mark_reg_unknown(env, state->regs, dst_regno);
4651 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4655 if (dst_regno >= 0) {
4656 /* restore register state from stack */
4657 copy_register_state(&state->regs[dst_regno], reg);
4658 /* mark reg as written since spilled pointer state likely
4659 * has its liveness marks cleared by is_state_visited()
4660 * which resets stack/reg liveness for state transitions
4662 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4663 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
4664 /* If dst_regno==-1, the caller is asking us whether
4665 * it is acceptable to use this value as a SCALAR_VALUE
4667 * We must not allow unprivileged callers to do that
4668 * with spilled pointers.
4670 verbose(env, "leaking pointer from stack off %d\n",
4674 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4676 for (i = 0; i < size; i++) {
4677 type = stype[(slot - i) % BPF_REG_SIZE];
4678 if (type == STACK_MISC)
4680 if (type == STACK_ZERO)
4682 if (type == STACK_INVALID && env->allow_uninit_stack)
4684 verbose(env, "invalid read from stack off %d+%d size %d\n",
4688 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4690 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
4695 enum bpf_access_src {
4696 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
4697 ACCESS_HELPER = 2, /* the access is performed by a helper */
4700 static int check_stack_range_initialized(struct bpf_verifier_env *env,
4701 int regno, int off, int access_size,
4702 bool zero_size_allowed,
4703 enum bpf_access_src type,
4704 struct bpf_call_arg_meta *meta);
4706 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
4708 return cur_regs(env) + regno;
4711 /* Read the stack at 'ptr_regno + off' and put the result into the register
4713 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
4714 * but not its variable offset.
4715 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
4717 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
4718 * filling registers (i.e. reads of spilled register cannot be detected when
4719 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
4720 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
4721 * offset; for a fixed offset check_stack_read_fixed_off should be used
4724 static int check_stack_read_var_off(struct bpf_verifier_env *env,
4725 int ptr_regno, int off, int size, int dst_regno)
4727 /* The state of the source register. */
4728 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
4729 struct bpf_func_state *ptr_state = func(env, reg);
4731 int min_off, max_off;
4733 /* Note that we pass a NULL meta, so raw access will not be permitted.
4735 err = check_stack_range_initialized(env, ptr_regno, off, size,
4736 false, ACCESS_DIRECT, NULL);
4740 min_off = reg->smin_value + off;
4741 max_off = reg->smax_value + off;
4742 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
4746 /* check_stack_read dispatches to check_stack_read_fixed_off or
4747 * check_stack_read_var_off.
4749 * The caller must ensure that the offset falls within the allocated stack
4752 * 'dst_regno' is a register which will receive the value from the stack. It
4753 * can be -1, meaning that the read value is not going to a register.
4755 static int check_stack_read(struct bpf_verifier_env *env,
4756 int ptr_regno, int off, int size,
4759 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
4760 struct bpf_func_state *state = func(env, reg);
4762 /* Some accesses are only permitted with a static offset. */
4763 bool var_off = !tnum_is_const(reg->var_off);
4765 /* The offset is required to be static when reads don't go to a
4766 * register, in order to not leak pointers (see
4767 * check_stack_read_fixed_off).
4769 if (dst_regno < 0 && var_off) {
4772 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4773 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
4777 /* Variable offset is prohibited for unprivileged mode for simplicity
4778 * since it requires corresponding support in Spectre masking for stack
4779 * ALU. See also retrieve_ptr_limit(). The check in
4780 * check_stack_access_for_ptr_arithmetic() called by
4781 * adjust_ptr_min_max_vals() prevents users from creating stack pointers
4782 * with variable offsets, therefore no check is required here. Further,
4783 * just checking it here would be insufficient as speculative stack
4784 * writes could still lead to unsafe speculative behaviour.
4787 off += reg->var_off.value;
4788 err = check_stack_read_fixed_off(env, state, off, size,
4791 /* Variable offset stack reads need more conservative handling
4792 * than fixed offset ones. Note that dst_regno >= 0 on this
4795 err = check_stack_read_var_off(env, ptr_regno, off, size,
4802 /* check_stack_write dispatches to check_stack_write_fixed_off or
4803 * check_stack_write_var_off.
4805 * 'ptr_regno' is the register used as a pointer into the stack.
4806 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
4807 * 'value_regno' is the register whose value we're writing to the stack. It can
4808 * be -1, meaning that we're not writing from a register.
4810 * The caller must ensure that the offset falls within the maximum stack size.
4812 static int check_stack_write(struct bpf_verifier_env *env,
4813 int ptr_regno, int off, int size,
4814 int value_regno, int insn_idx)
4816 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
4817 struct bpf_func_state *state = func(env, reg);
4820 if (tnum_is_const(reg->var_off)) {
4821 off += reg->var_off.value;
4822 err = check_stack_write_fixed_off(env, state, off, size,
4823 value_regno, insn_idx);
4825 /* Variable offset stack reads need more conservative handling
4826 * than fixed offset ones.
4828 err = check_stack_write_var_off(env, state,
4829 ptr_regno, off, size,
4830 value_regno, insn_idx);
4835 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
4836 int off, int size, enum bpf_access_type type)
4838 struct bpf_reg_state *regs = cur_regs(env);
4839 struct bpf_map *map = regs[regno].map_ptr;
4840 u32 cap = bpf_map_flags_to_cap(map);
4842 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
4843 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
4844 map->value_size, off, size);
4848 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
4849 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
4850 map->value_size, off, size);
4857 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
4858 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
4859 int off, int size, u32 mem_size,
4860 bool zero_size_allowed)
4862 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
4863 struct bpf_reg_state *reg;
4865 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
4868 reg = &cur_regs(env)[regno];
4869 switch (reg->type) {
4870 case PTR_TO_MAP_KEY:
4871 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
4872 mem_size, off, size);
4874 case PTR_TO_MAP_VALUE:
4875 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
4876 mem_size, off, size);
4879 case PTR_TO_PACKET_META:
4880 case PTR_TO_PACKET_END:
4881 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
4882 off, size, regno, reg->id, off, mem_size);
4886 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
4887 mem_size, off, size);
4893 /* check read/write into a memory region with possible variable offset */
4894 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
4895 int off, int size, u32 mem_size,
4896 bool zero_size_allowed)
4898 struct bpf_verifier_state *vstate = env->cur_state;
4899 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4900 struct bpf_reg_state *reg = &state->regs[regno];
4903 /* We may have adjusted the register pointing to memory region, so we
4904 * need to try adding each of min_value and max_value to off
4905 * to make sure our theoretical access will be safe.
4907 * The minimum value is only important with signed
4908 * comparisons where we can't assume the floor of a
4909 * value is 0. If we are using signed variables for our
4910 * index'es we need to make sure that whatever we use
4911 * will have a set floor within our range.
4913 if (reg->smin_value < 0 &&
4914 (reg->smin_value == S64_MIN ||
4915 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
4916 reg->smin_value + off < 0)) {
4917 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
4921 err = __check_mem_access(env, regno, reg->smin_value + off, size,
4922 mem_size, zero_size_allowed);
4924 verbose(env, "R%d min value is outside of the allowed memory range\n",
4929 /* If we haven't set a max value then we need to bail since we can't be
4930 * sure we won't do bad things.
4931 * If reg->umax_value + off could overflow, treat that as unbounded too.
4933 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
4934 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
4938 err = __check_mem_access(env, regno, reg->umax_value + off, size,
4939 mem_size, zero_size_allowed);
4941 verbose(env, "R%d max value is outside of the allowed memory range\n",
4949 static int __check_ptr_off_reg(struct bpf_verifier_env *env,
4950 const struct bpf_reg_state *reg, int regno,
4953 /* Access to this pointer-typed register or passing it to a helper
4954 * is only allowed in its original, unmodified form.
4958 verbose(env, "negative offset %s ptr R%d off=%d disallowed\n",
4959 reg_type_str(env, reg->type), regno, reg->off);
4963 if (!fixed_off_ok && reg->off) {
4964 verbose(env, "dereference of modified %s ptr R%d off=%d disallowed\n",
4965 reg_type_str(env, reg->type), regno, reg->off);
4969 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4972 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4973 verbose(env, "variable %s access var_off=%s disallowed\n",
4974 reg_type_str(env, reg->type), tn_buf);
4981 int check_ptr_off_reg(struct bpf_verifier_env *env,
4982 const struct bpf_reg_state *reg, int regno)
4984 return __check_ptr_off_reg(env, reg, regno, false);
4987 static int map_kptr_match_type(struct bpf_verifier_env *env,
4988 struct btf_field *kptr_field,
4989 struct bpf_reg_state *reg, u32 regno)
4991 const char *targ_name = btf_type_name(kptr_field->kptr.btf, kptr_field->kptr.btf_id);
4993 const char *reg_name = "";
4995 if (btf_is_kernel(reg->btf)) {
4996 perm_flags = PTR_MAYBE_NULL | PTR_TRUSTED | MEM_RCU;
4998 /* Only unreferenced case accepts untrusted pointers */
4999 if (kptr_field->type == BPF_KPTR_UNREF)
5000 perm_flags |= PTR_UNTRUSTED;
5002 perm_flags = PTR_MAYBE_NULL | MEM_ALLOC;
5005 if (base_type(reg->type) != PTR_TO_BTF_ID || (type_flag(reg->type) & ~perm_flags))
5008 /* We need to verify reg->type and reg->btf, before accessing reg->btf */
5009 reg_name = btf_type_name(reg->btf, reg->btf_id);
5011 /* For ref_ptr case, release function check should ensure we get one
5012 * referenced PTR_TO_BTF_ID, and that its fixed offset is 0. For the
5013 * normal store of unreferenced kptr, we must ensure var_off is zero.
5014 * Since ref_ptr cannot be accessed directly by BPF insns, checks for
5015 * reg->off and reg->ref_obj_id are not needed here.
5017 if (__check_ptr_off_reg(env, reg, regno, true))
5020 /* A full type match is needed, as BTF can be vmlinux, module or prog BTF, and
5021 * we also need to take into account the reg->off.
5023 * We want to support cases like:
5031 * v = func(); // PTR_TO_BTF_ID
5032 * val->foo = v; // reg->off is zero, btf and btf_id match type
5033 * val->bar = &v->br; // reg->off is still zero, but we need to retry with
5034 * // first member type of struct after comparison fails
5035 * val->baz = &v->bz; // reg->off is non-zero, so struct needs to be walked
5038 * In the kptr_ref case, check_func_arg_reg_off already ensures reg->off
5039 * is zero. We must also ensure that btf_struct_ids_match does not walk
5040 * the struct to match type against first member of struct, i.e. reject
5041 * second case from above. Hence, when type is BPF_KPTR_REF, we set
5042 * strict mode to true for type match.
5044 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
5045 kptr_field->kptr.btf, kptr_field->kptr.btf_id,
5046 kptr_field->type == BPF_KPTR_REF))
5050 verbose(env, "invalid kptr access, R%d type=%s%s ", regno,
5051 reg_type_str(env, reg->type), reg_name);
5052 verbose(env, "expected=%s%s", reg_type_str(env, PTR_TO_BTF_ID), targ_name);
5053 if (kptr_field->type == BPF_KPTR_UNREF)
5054 verbose(env, " or %s%s\n", reg_type_str(env, PTR_TO_BTF_ID | PTR_UNTRUSTED),
5061 /* The non-sleepable programs and sleepable programs with explicit bpf_rcu_read_lock()
5062 * can dereference RCU protected pointers and result is PTR_TRUSTED.
5064 static bool in_rcu_cs(struct bpf_verifier_env *env)
5066 return env->cur_state->active_rcu_lock ||
5067 env->cur_state->active_lock.ptr ||
5068 !env->prog->aux->sleepable;
5071 /* Once GCC supports btf_type_tag the following mechanism will be replaced with tag check */
5072 BTF_SET_START(rcu_protected_types)
5073 BTF_ID(struct, prog_test_ref_kfunc)
5074 BTF_ID(struct, cgroup)
5075 BTF_ID(struct, bpf_cpumask)
5076 BTF_ID(struct, task_struct)
5077 BTF_SET_END(rcu_protected_types)
5079 static bool rcu_protected_object(const struct btf *btf, u32 btf_id)
5081 if (!btf_is_kernel(btf))
5083 return btf_id_set_contains(&rcu_protected_types, btf_id);
5086 static bool rcu_safe_kptr(const struct btf_field *field)
5088 const struct btf_field_kptr *kptr = &field->kptr;
5090 return field->type == BPF_KPTR_REF && rcu_protected_object(kptr->btf, kptr->btf_id);
5093 static int check_map_kptr_access(struct bpf_verifier_env *env, u32 regno,
5094 int value_regno, int insn_idx,
5095 struct btf_field *kptr_field)
5097 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
5098 int class = BPF_CLASS(insn->code);
5099 struct bpf_reg_state *val_reg;
5101 /* Things we already checked for in check_map_access and caller:
5102 * - Reject cases where variable offset may touch kptr
5103 * - size of access (must be BPF_DW)
5104 * - tnum_is_const(reg->var_off)
5105 * - kptr_field->offset == off + reg->var_off.value
5107 /* Only BPF_[LDX,STX,ST] | BPF_MEM | BPF_DW is supported */
5108 if (BPF_MODE(insn->code) != BPF_MEM) {
5109 verbose(env, "kptr in map can only be accessed using BPF_MEM instruction mode\n");
5113 /* We only allow loading referenced kptr, since it will be marked as
5114 * untrusted, similar to unreferenced kptr.
5116 if (class != BPF_LDX && kptr_field->type == BPF_KPTR_REF) {
5117 verbose(env, "store to referenced kptr disallowed\n");
5121 if (class == BPF_LDX) {
5122 val_reg = reg_state(env, value_regno);
5123 /* We can simply mark the value_regno receiving the pointer
5124 * value from map as PTR_TO_BTF_ID, with the correct type.
5126 mark_btf_ld_reg(env, cur_regs(env), value_regno, PTR_TO_BTF_ID, kptr_field->kptr.btf,
5127 kptr_field->kptr.btf_id,
5128 rcu_safe_kptr(kptr_field) && in_rcu_cs(env) ?
5129 PTR_MAYBE_NULL | MEM_RCU :
5130 PTR_MAYBE_NULL | PTR_UNTRUSTED);
5131 /* For mark_ptr_or_null_reg */
5132 val_reg->id = ++env->id_gen;
5133 } else if (class == BPF_STX) {
5134 val_reg = reg_state(env, value_regno);
5135 if (!register_is_null(val_reg) &&
5136 map_kptr_match_type(env, kptr_field, val_reg, value_regno))
5138 } else if (class == BPF_ST) {
5140 verbose(env, "BPF_ST imm must be 0 when storing to kptr at off=%u\n",
5141 kptr_field->offset);
5145 verbose(env, "kptr in map can only be accessed using BPF_LDX/BPF_STX/BPF_ST\n");
5151 /* check read/write into a map element with possible variable offset */
5152 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
5153 int off, int size, bool zero_size_allowed,
5154 enum bpf_access_src src)
5156 struct bpf_verifier_state *vstate = env->cur_state;
5157 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5158 struct bpf_reg_state *reg = &state->regs[regno];
5159 struct bpf_map *map = reg->map_ptr;
5160 struct btf_record *rec;
5163 err = check_mem_region_access(env, regno, off, size, map->value_size,
5168 if (IS_ERR_OR_NULL(map->record))
5171 for (i = 0; i < rec->cnt; i++) {
5172 struct btf_field *field = &rec->fields[i];
5173 u32 p = field->offset;
5175 /* If any part of a field can be touched by load/store, reject
5176 * this program. To check that [x1, x2) overlaps with [y1, y2),
5177 * it is sufficient to check x1 < y2 && y1 < x2.
5179 if (reg->smin_value + off < p + btf_field_type_size(field->type) &&
5180 p < reg->umax_value + off + size) {
5181 switch (field->type) {
5182 case BPF_KPTR_UNREF:
5184 if (src != ACCESS_DIRECT) {
5185 verbose(env, "kptr cannot be accessed indirectly by helper\n");
5188 if (!tnum_is_const(reg->var_off)) {
5189 verbose(env, "kptr access cannot have variable offset\n");
5192 if (p != off + reg->var_off.value) {
5193 verbose(env, "kptr access misaligned expected=%u off=%llu\n",
5194 p, off + reg->var_off.value);
5197 if (size != bpf_size_to_bytes(BPF_DW)) {
5198 verbose(env, "kptr access size must be BPF_DW\n");
5203 verbose(env, "%s cannot be accessed directly by load/store\n",
5204 btf_field_type_name(field->type));
5212 #define MAX_PACKET_OFF 0xffff
5214 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
5215 const struct bpf_call_arg_meta *meta,
5216 enum bpf_access_type t)
5218 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
5220 switch (prog_type) {
5221 /* Program types only with direct read access go here! */
5222 case BPF_PROG_TYPE_LWT_IN:
5223 case BPF_PROG_TYPE_LWT_OUT:
5224 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
5225 case BPF_PROG_TYPE_SK_REUSEPORT:
5226 case BPF_PROG_TYPE_FLOW_DISSECTOR:
5227 case BPF_PROG_TYPE_CGROUP_SKB:
5232 /* Program types with direct read + write access go here! */
5233 case BPF_PROG_TYPE_SCHED_CLS:
5234 case BPF_PROG_TYPE_SCHED_ACT:
5235 case BPF_PROG_TYPE_XDP:
5236 case BPF_PROG_TYPE_LWT_XMIT:
5237 case BPF_PROG_TYPE_SK_SKB:
5238 case BPF_PROG_TYPE_SK_MSG:
5240 return meta->pkt_access;
5242 env->seen_direct_write = true;
5245 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
5247 env->seen_direct_write = true;
5256 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
5257 int size, bool zero_size_allowed)
5259 struct bpf_reg_state *regs = cur_regs(env);
5260 struct bpf_reg_state *reg = ®s[regno];
5263 /* We may have added a variable offset to the packet pointer; but any
5264 * reg->range we have comes after that. We are only checking the fixed
5268 /* We don't allow negative numbers, because we aren't tracking enough
5269 * detail to prove they're safe.
5271 if (reg->smin_value < 0) {
5272 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
5277 err = reg->range < 0 ? -EINVAL :
5278 __check_mem_access(env, regno, off, size, reg->range,
5281 verbose(env, "R%d offset is outside of the packet\n", regno);
5285 /* __check_mem_access has made sure "off + size - 1" is within u16.
5286 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
5287 * otherwise find_good_pkt_pointers would have refused to set range info
5288 * that __check_mem_access would have rejected this pkt access.
5289 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
5291 env->prog->aux->max_pkt_offset =
5292 max_t(u32, env->prog->aux->max_pkt_offset,
5293 off + reg->umax_value + size - 1);
5298 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
5299 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
5300 enum bpf_access_type t, enum bpf_reg_type *reg_type,
5301 struct btf **btf, u32 *btf_id)
5303 struct bpf_insn_access_aux info = {
5304 .reg_type = *reg_type,
5308 if (env->ops->is_valid_access &&
5309 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
5310 /* A non zero info.ctx_field_size indicates that this field is a
5311 * candidate for later verifier transformation to load the whole
5312 * field and then apply a mask when accessed with a narrower
5313 * access than actual ctx access size. A zero info.ctx_field_size
5314 * will only allow for whole field access and rejects any other
5315 * type of narrower access.
5317 *reg_type = info.reg_type;
5319 if (base_type(*reg_type) == PTR_TO_BTF_ID) {
5321 *btf_id = info.btf_id;
5323 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
5325 /* remember the offset of last byte accessed in ctx */
5326 if (env->prog->aux->max_ctx_offset < off + size)
5327 env->prog->aux->max_ctx_offset = off + size;
5331 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
5335 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
5338 if (size < 0 || off < 0 ||
5339 (u64)off + size > sizeof(struct bpf_flow_keys)) {
5340 verbose(env, "invalid access to flow keys off=%d size=%d\n",
5347 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
5348 u32 regno, int off, int size,
5349 enum bpf_access_type t)
5351 struct bpf_reg_state *regs = cur_regs(env);
5352 struct bpf_reg_state *reg = ®s[regno];
5353 struct bpf_insn_access_aux info = {};
5356 if (reg->smin_value < 0) {
5357 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
5362 switch (reg->type) {
5363 case PTR_TO_SOCK_COMMON:
5364 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
5367 valid = bpf_sock_is_valid_access(off, size, t, &info);
5369 case PTR_TO_TCP_SOCK:
5370 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
5372 case PTR_TO_XDP_SOCK:
5373 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
5381 env->insn_aux_data[insn_idx].ctx_field_size =
5382 info.ctx_field_size;
5386 verbose(env, "R%d invalid %s access off=%d size=%d\n",
5387 regno, reg_type_str(env, reg->type), off, size);
5392 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
5394 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
5397 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
5399 const struct bpf_reg_state *reg = reg_state(env, regno);
5401 return reg->type == PTR_TO_CTX;
5404 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
5406 const struct bpf_reg_state *reg = reg_state(env, regno);
5408 return type_is_sk_pointer(reg->type);
5411 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
5413 const struct bpf_reg_state *reg = reg_state(env, regno);
5415 return type_is_pkt_pointer(reg->type);
5418 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
5420 const struct bpf_reg_state *reg = reg_state(env, regno);
5422 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
5423 return reg->type == PTR_TO_FLOW_KEYS;
5426 static u32 *reg2btf_ids[__BPF_REG_TYPE_MAX] = {
5428 [PTR_TO_SOCKET] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK],
5429 [PTR_TO_SOCK_COMMON] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
5430 [PTR_TO_TCP_SOCK] = &btf_sock_ids[BTF_SOCK_TYPE_TCP],
5432 [CONST_PTR_TO_MAP] = btf_bpf_map_id,
5435 static bool is_trusted_reg(const struct bpf_reg_state *reg)
5437 /* A referenced register is always trusted. */
5438 if (reg->ref_obj_id)
5441 /* Types listed in the reg2btf_ids are always trusted */
5442 if (reg2btf_ids[base_type(reg->type)])
5445 /* If a register is not referenced, it is trusted if it has the
5446 * MEM_ALLOC or PTR_TRUSTED type modifiers, and no others. Some of the
5447 * other type modifiers may be safe, but we elect to take an opt-in
5448 * approach here as some (e.g. PTR_UNTRUSTED and PTR_MAYBE_NULL) are
5451 * Eventually, we should make PTR_TRUSTED the single source of truth
5452 * for whether a register is trusted.
5454 return type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS &&
5455 !bpf_type_has_unsafe_modifiers(reg->type);
5458 static bool is_rcu_reg(const struct bpf_reg_state *reg)
5460 return reg->type & MEM_RCU;
5463 static void clear_trusted_flags(enum bpf_type_flag *flag)
5465 *flag &= ~(BPF_REG_TRUSTED_MODIFIERS | MEM_RCU);
5468 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
5469 const struct bpf_reg_state *reg,
5470 int off, int size, bool strict)
5472 struct tnum reg_off;
5475 /* Byte size accesses are always allowed. */
5476 if (!strict || size == 1)
5479 /* For platforms that do not have a Kconfig enabling
5480 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
5481 * NET_IP_ALIGN is universally set to '2'. And on platforms
5482 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
5483 * to this code only in strict mode where we want to emulate
5484 * the NET_IP_ALIGN==2 checking. Therefore use an
5485 * unconditional IP align value of '2'.
5489 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
5490 if (!tnum_is_aligned(reg_off, size)) {
5493 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5495 "misaligned packet access off %d+%s+%d+%d size %d\n",
5496 ip_align, tn_buf, reg->off, off, size);
5503 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
5504 const struct bpf_reg_state *reg,
5505 const char *pointer_desc,
5506 int off, int size, bool strict)
5508 struct tnum reg_off;
5510 /* Byte size accesses are always allowed. */
5511 if (!strict || size == 1)
5514 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
5515 if (!tnum_is_aligned(reg_off, size)) {
5518 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5519 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
5520 pointer_desc, tn_buf, reg->off, off, size);
5527 static int check_ptr_alignment(struct bpf_verifier_env *env,
5528 const struct bpf_reg_state *reg, int off,
5529 int size, bool strict_alignment_once)
5531 bool strict = env->strict_alignment || strict_alignment_once;
5532 const char *pointer_desc = "";
5534 switch (reg->type) {
5536 case PTR_TO_PACKET_META:
5537 /* Special case, because of NET_IP_ALIGN. Given metadata sits
5538 * right in front, treat it the very same way.
5540 return check_pkt_ptr_alignment(env, reg, off, size, strict);
5541 case PTR_TO_FLOW_KEYS:
5542 pointer_desc = "flow keys ";
5544 case PTR_TO_MAP_KEY:
5545 pointer_desc = "key ";
5547 case PTR_TO_MAP_VALUE:
5548 pointer_desc = "value ";
5551 pointer_desc = "context ";
5554 pointer_desc = "stack ";
5555 /* The stack spill tracking logic in check_stack_write_fixed_off()
5556 * and check_stack_read_fixed_off() relies on stack accesses being
5562 pointer_desc = "sock ";
5564 case PTR_TO_SOCK_COMMON:
5565 pointer_desc = "sock_common ";
5567 case PTR_TO_TCP_SOCK:
5568 pointer_desc = "tcp_sock ";
5570 case PTR_TO_XDP_SOCK:
5571 pointer_desc = "xdp_sock ";
5576 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
5580 static int update_stack_depth(struct bpf_verifier_env *env,
5581 const struct bpf_func_state *func,
5584 u16 stack = env->subprog_info[func->subprogno].stack_depth;
5589 /* update known max for given subprogram */
5590 env->subprog_info[func->subprogno].stack_depth = -off;
5594 /* starting from main bpf function walk all instructions of the function
5595 * and recursively walk all callees that given function can call.
5596 * Ignore jump and exit insns.
5597 * Since recursion is prevented by check_cfg() this algorithm
5598 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
5600 static int check_max_stack_depth_subprog(struct bpf_verifier_env *env, int idx)
5602 struct bpf_subprog_info *subprog = env->subprog_info;
5603 struct bpf_insn *insn = env->prog->insnsi;
5604 int depth = 0, frame = 0, i, subprog_end;
5605 bool tail_call_reachable = false;
5606 int ret_insn[MAX_CALL_FRAMES];
5607 int ret_prog[MAX_CALL_FRAMES];
5610 i = subprog[idx].start;
5612 /* protect against potential stack overflow that might happen when
5613 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
5614 * depth for such case down to 256 so that the worst case scenario
5615 * would result in 8k stack size (32 which is tailcall limit * 256 =
5618 * To get the idea what might happen, see an example:
5619 * func1 -> sub rsp, 128
5620 * subfunc1 -> sub rsp, 256
5621 * tailcall1 -> add rsp, 256
5622 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
5623 * subfunc2 -> sub rsp, 64
5624 * subfunc22 -> sub rsp, 128
5625 * tailcall2 -> add rsp, 128
5626 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
5628 * tailcall will unwind the current stack frame but it will not get rid
5629 * of caller's stack as shown on the example above.
5631 if (idx && subprog[idx].has_tail_call && depth >= 256) {
5633 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
5637 /* round up to 32-bytes, since this is granularity
5638 * of interpreter stack size
5640 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
5641 if (depth > MAX_BPF_STACK) {
5642 verbose(env, "combined stack size of %d calls is %d. Too large\n",
5647 subprog_end = subprog[idx + 1].start;
5648 for (; i < subprog_end; i++) {
5649 int next_insn, sidx;
5651 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
5653 /* remember insn and function to return to */
5654 ret_insn[frame] = i + 1;
5655 ret_prog[frame] = idx;
5657 /* find the callee */
5658 next_insn = i + insn[i].imm + 1;
5659 sidx = find_subprog(env, next_insn);
5661 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5665 if (subprog[sidx].is_async_cb) {
5666 if (subprog[sidx].has_tail_call) {
5667 verbose(env, "verifier bug. subprog has tail_call and async cb\n");
5670 /* async callbacks don't increase bpf prog stack size unless called directly */
5671 if (!bpf_pseudo_call(insn + i))
5677 if (subprog[idx].has_tail_call)
5678 tail_call_reachable = true;
5681 if (frame >= MAX_CALL_FRAMES) {
5682 verbose(env, "the call stack of %d frames is too deep !\n",
5688 /* if tail call got detected across bpf2bpf calls then mark each of the
5689 * currently present subprog frames as tail call reachable subprogs;
5690 * this info will be utilized by JIT so that we will be preserving the
5691 * tail call counter throughout bpf2bpf calls combined with tailcalls
5693 if (tail_call_reachable)
5694 for (j = 0; j < frame; j++)
5695 subprog[ret_prog[j]].tail_call_reachable = true;
5696 if (subprog[0].tail_call_reachable)
5697 env->prog->aux->tail_call_reachable = true;
5699 /* end of for() loop means the last insn of the 'subprog'
5700 * was reached. Doesn't matter whether it was JA or EXIT
5704 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
5706 i = ret_insn[frame];
5707 idx = ret_prog[frame];
5711 static int check_max_stack_depth(struct bpf_verifier_env *env)
5713 struct bpf_subprog_info *si = env->subprog_info;
5716 for (int i = 0; i < env->subprog_cnt; i++) {
5717 if (!i || si[i].is_async_cb) {
5718 ret = check_max_stack_depth_subprog(env, i);
5727 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
5728 static int get_callee_stack_depth(struct bpf_verifier_env *env,
5729 const struct bpf_insn *insn, int idx)
5731 int start = idx + insn->imm + 1, subprog;
5733 subprog = find_subprog(env, start);
5735 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5739 return env->subprog_info[subprog].stack_depth;
5743 static int __check_buffer_access(struct bpf_verifier_env *env,
5744 const char *buf_info,
5745 const struct bpf_reg_state *reg,
5746 int regno, int off, int size)
5750 "R%d invalid %s buffer access: off=%d, size=%d\n",
5751 regno, buf_info, off, size);
5754 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
5757 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5759 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
5760 regno, off, tn_buf);
5767 static int check_tp_buffer_access(struct bpf_verifier_env *env,
5768 const struct bpf_reg_state *reg,
5769 int regno, int off, int size)
5773 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
5777 if (off + size > env->prog->aux->max_tp_access)
5778 env->prog->aux->max_tp_access = off + size;
5783 static int check_buffer_access(struct bpf_verifier_env *env,
5784 const struct bpf_reg_state *reg,
5785 int regno, int off, int size,
5786 bool zero_size_allowed,
5789 const char *buf_info = type_is_rdonly_mem(reg->type) ? "rdonly" : "rdwr";
5792 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
5796 if (off + size > *max_access)
5797 *max_access = off + size;
5802 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
5803 static void zext_32_to_64(struct bpf_reg_state *reg)
5805 reg->var_off = tnum_subreg(reg->var_off);
5806 __reg_assign_32_into_64(reg);
5809 /* truncate register to smaller size (in bytes)
5810 * must be called with size < BPF_REG_SIZE
5812 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
5816 /* clear high bits in bit representation */
5817 reg->var_off = tnum_cast(reg->var_off, size);
5819 /* fix arithmetic bounds */
5820 mask = ((u64)1 << (size * 8)) - 1;
5821 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
5822 reg->umin_value &= mask;
5823 reg->umax_value &= mask;
5825 reg->umin_value = 0;
5826 reg->umax_value = mask;
5828 reg->smin_value = reg->umin_value;
5829 reg->smax_value = reg->umax_value;
5831 /* If size is smaller than 32bit register the 32bit register
5832 * values are also truncated so we push 64-bit bounds into
5833 * 32-bit bounds. Above were truncated < 32-bits already.
5837 __reg_combine_64_into_32(reg);
5840 static void set_sext64_default_val(struct bpf_reg_state *reg, int size)
5843 reg->smin_value = reg->s32_min_value = S8_MIN;
5844 reg->smax_value = reg->s32_max_value = S8_MAX;
5845 } else if (size == 2) {
5846 reg->smin_value = reg->s32_min_value = S16_MIN;
5847 reg->smax_value = reg->s32_max_value = S16_MAX;
5850 reg->smin_value = reg->s32_min_value = S32_MIN;
5851 reg->smax_value = reg->s32_max_value = S32_MAX;
5853 reg->umin_value = reg->u32_min_value = 0;
5854 reg->umax_value = U64_MAX;
5855 reg->u32_max_value = U32_MAX;
5856 reg->var_off = tnum_unknown;
5859 static void coerce_reg_to_size_sx(struct bpf_reg_state *reg, int size)
5861 s64 init_s64_max, init_s64_min, s64_max, s64_min, u64_cval;
5862 u64 top_smax_value, top_smin_value;
5863 u64 num_bits = size * 8;
5865 if (tnum_is_const(reg->var_off)) {
5866 u64_cval = reg->var_off.value;
5868 reg->var_off = tnum_const((s8)u64_cval);
5870 reg->var_off = tnum_const((s16)u64_cval);
5873 reg->var_off = tnum_const((s32)u64_cval);
5875 u64_cval = reg->var_off.value;
5876 reg->smax_value = reg->smin_value = u64_cval;
5877 reg->umax_value = reg->umin_value = u64_cval;
5878 reg->s32_max_value = reg->s32_min_value = u64_cval;
5879 reg->u32_max_value = reg->u32_min_value = u64_cval;
5883 top_smax_value = ((u64)reg->smax_value >> num_bits) << num_bits;
5884 top_smin_value = ((u64)reg->smin_value >> num_bits) << num_bits;
5886 if (top_smax_value != top_smin_value)
5889 /* find the s64_min and s64_min after sign extension */
5891 init_s64_max = (s8)reg->smax_value;
5892 init_s64_min = (s8)reg->smin_value;
5893 } else if (size == 2) {
5894 init_s64_max = (s16)reg->smax_value;
5895 init_s64_min = (s16)reg->smin_value;
5897 init_s64_max = (s32)reg->smax_value;
5898 init_s64_min = (s32)reg->smin_value;
5901 s64_max = max(init_s64_max, init_s64_min);
5902 s64_min = min(init_s64_max, init_s64_min);
5904 /* both of s64_max/s64_min positive or negative */
5905 if ((s64_max >= 0) == (s64_min >= 0)) {
5906 reg->smin_value = reg->s32_min_value = s64_min;
5907 reg->smax_value = reg->s32_max_value = s64_max;
5908 reg->umin_value = reg->u32_min_value = s64_min;
5909 reg->umax_value = reg->u32_max_value = s64_max;
5910 reg->var_off = tnum_range(s64_min, s64_max);
5915 set_sext64_default_val(reg, size);
5918 static void set_sext32_default_val(struct bpf_reg_state *reg, int size)
5921 reg->s32_min_value = S8_MIN;
5922 reg->s32_max_value = S8_MAX;
5925 reg->s32_min_value = S16_MIN;
5926 reg->s32_max_value = S16_MAX;
5928 reg->u32_min_value = 0;
5929 reg->u32_max_value = U32_MAX;
5932 static void coerce_subreg_to_size_sx(struct bpf_reg_state *reg, int size)
5934 s32 init_s32_max, init_s32_min, s32_max, s32_min, u32_val;
5935 u32 top_smax_value, top_smin_value;
5936 u32 num_bits = size * 8;
5938 if (tnum_is_const(reg->var_off)) {
5939 u32_val = reg->var_off.value;
5941 reg->var_off = tnum_const((s8)u32_val);
5943 reg->var_off = tnum_const((s16)u32_val);
5945 u32_val = reg->var_off.value;
5946 reg->s32_min_value = reg->s32_max_value = u32_val;
5947 reg->u32_min_value = reg->u32_max_value = u32_val;
5951 top_smax_value = ((u32)reg->s32_max_value >> num_bits) << num_bits;
5952 top_smin_value = ((u32)reg->s32_min_value >> num_bits) << num_bits;
5954 if (top_smax_value != top_smin_value)
5957 /* find the s32_min and s32_min after sign extension */
5959 init_s32_max = (s8)reg->s32_max_value;
5960 init_s32_min = (s8)reg->s32_min_value;
5963 init_s32_max = (s16)reg->s32_max_value;
5964 init_s32_min = (s16)reg->s32_min_value;
5966 s32_max = max(init_s32_max, init_s32_min);
5967 s32_min = min(init_s32_max, init_s32_min);
5969 if ((s32_min >= 0) == (s32_max >= 0)) {
5970 reg->s32_min_value = s32_min;
5971 reg->s32_max_value = s32_max;
5972 reg->u32_min_value = (u32)s32_min;
5973 reg->u32_max_value = (u32)s32_max;
5978 set_sext32_default_val(reg, size);
5981 static bool bpf_map_is_rdonly(const struct bpf_map *map)
5983 /* A map is considered read-only if the following condition are true:
5985 * 1) BPF program side cannot change any of the map content. The
5986 * BPF_F_RDONLY_PROG flag is throughout the lifetime of a map
5987 * and was set at map creation time.
5988 * 2) The map value(s) have been initialized from user space by a
5989 * loader and then "frozen", such that no new map update/delete
5990 * operations from syscall side are possible for the rest of
5991 * the map's lifetime from that point onwards.
5992 * 3) Any parallel/pending map update/delete operations from syscall
5993 * side have been completed. Only after that point, it's safe to
5994 * assume that map value(s) are immutable.
5996 return (map->map_flags & BPF_F_RDONLY_PROG) &&
5997 READ_ONCE(map->frozen) &&
5998 !bpf_map_write_active(map);
6001 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val,
6008 err = map->ops->map_direct_value_addr(map, &addr, off);
6011 ptr = (void *)(long)addr + off;
6015 *val = is_ldsx ? (s64)*(s8 *)ptr : (u64)*(u8 *)ptr;
6018 *val = is_ldsx ? (s64)*(s16 *)ptr : (u64)*(u16 *)ptr;
6021 *val = is_ldsx ? (s64)*(s32 *)ptr : (u64)*(u32 *)ptr;
6032 #define BTF_TYPE_SAFE_RCU(__type) __PASTE(__type, __safe_rcu)
6033 #define BTF_TYPE_SAFE_RCU_OR_NULL(__type) __PASTE(__type, __safe_rcu_or_null)
6034 #define BTF_TYPE_SAFE_TRUSTED(__type) __PASTE(__type, __safe_trusted)
6037 * Allow list few fields as RCU trusted or full trusted.
6038 * This logic doesn't allow mix tagging and will be removed once GCC supports
6042 /* RCU trusted: these fields are trusted in RCU CS and never NULL */
6043 BTF_TYPE_SAFE_RCU(struct task_struct) {
6044 const cpumask_t *cpus_ptr;
6045 struct css_set __rcu *cgroups;
6046 struct task_struct __rcu *real_parent;
6047 struct task_struct *group_leader;
6050 BTF_TYPE_SAFE_RCU(struct cgroup) {
6051 /* cgrp->kn is always accessible as documented in kernel/cgroup/cgroup.c */
6052 struct kernfs_node *kn;
6055 BTF_TYPE_SAFE_RCU(struct css_set) {
6056 struct cgroup *dfl_cgrp;
6059 /* RCU trusted: these fields are trusted in RCU CS and can be NULL */
6060 BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct) {
6061 struct file __rcu *exe_file;
6064 /* skb->sk, req->sk are not RCU protected, but we mark them as such
6065 * because bpf prog accessible sockets are SOCK_RCU_FREE.
6067 BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff) {
6071 BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock) {
6075 /* full trusted: these fields are trusted even outside of RCU CS and never NULL */
6076 BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta) {
6077 struct seq_file *seq;
6080 BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task) {
6081 struct bpf_iter_meta *meta;
6082 struct task_struct *task;
6085 BTF_TYPE_SAFE_TRUSTED(struct linux_binprm) {
6089 BTF_TYPE_SAFE_TRUSTED(struct file) {
6090 struct inode *f_inode;
6093 BTF_TYPE_SAFE_TRUSTED(struct dentry) {
6094 /* no negative dentry-s in places where bpf can see it */
6095 struct inode *d_inode;
6098 BTF_TYPE_SAFE_TRUSTED(struct socket) {
6102 static bool type_is_rcu(struct bpf_verifier_env *env,
6103 struct bpf_reg_state *reg,
6104 const char *field_name, u32 btf_id)
6106 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct task_struct));
6107 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct cgroup));
6108 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct css_set));
6110 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_rcu");
6113 static bool type_is_rcu_or_null(struct bpf_verifier_env *env,
6114 struct bpf_reg_state *reg,
6115 const char *field_name, u32 btf_id)
6117 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct));
6118 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff));
6119 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock));
6121 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_rcu_or_null");
6124 static bool type_is_trusted(struct bpf_verifier_env *env,
6125 struct bpf_reg_state *reg,
6126 const char *field_name, u32 btf_id)
6128 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta));
6129 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task));
6130 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct linux_binprm));
6131 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct file));
6132 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct dentry));
6133 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct socket));
6135 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_trusted");
6138 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
6139 struct bpf_reg_state *regs,
6140 int regno, int off, int size,
6141 enum bpf_access_type atype,
6144 struct bpf_reg_state *reg = regs + regno;
6145 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
6146 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
6147 const char *field_name = NULL;
6148 enum bpf_type_flag flag = 0;
6152 if (!env->allow_ptr_leaks) {
6154 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
6158 if (!env->prog->gpl_compatible && btf_is_kernel(reg->btf)) {
6160 "Cannot access kernel 'struct %s' from non-GPL compatible program\n",
6166 "R%d is ptr_%s invalid negative access: off=%d\n",
6170 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
6173 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6175 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
6176 regno, tname, off, tn_buf);
6180 if (reg->type & MEM_USER) {
6182 "R%d is ptr_%s access user memory: off=%d\n",
6187 if (reg->type & MEM_PERCPU) {
6189 "R%d is ptr_%s access percpu memory: off=%d\n",
6194 if (env->ops->btf_struct_access && !type_is_alloc(reg->type) && atype == BPF_WRITE) {
6195 if (!btf_is_kernel(reg->btf)) {
6196 verbose(env, "verifier internal error: reg->btf must be kernel btf\n");
6199 ret = env->ops->btf_struct_access(&env->log, reg, off, size);
6201 /* Writes are permitted with default btf_struct_access for
6202 * program allocated objects (which always have ref_obj_id > 0),
6203 * but not for untrusted PTR_TO_BTF_ID | MEM_ALLOC.
6205 if (atype != BPF_READ && !type_is_ptr_alloc_obj(reg->type)) {
6206 verbose(env, "only read is supported\n");
6210 if (type_is_alloc(reg->type) && !type_is_non_owning_ref(reg->type) &&
6212 verbose(env, "verifier internal error: ref_obj_id for allocated object must be non-zero\n");
6216 ret = btf_struct_access(&env->log, reg, off, size, atype, &btf_id, &flag, &field_name);
6222 if (ret != PTR_TO_BTF_ID) {
6225 } else if (type_flag(reg->type) & PTR_UNTRUSTED) {
6226 /* If this is an untrusted pointer, all pointers formed by walking it
6227 * also inherit the untrusted flag.
6229 flag = PTR_UNTRUSTED;
6231 } else if (is_trusted_reg(reg) || is_rcu_reg(reg)) {
6232 /* By default any pointer obtained from walking a trusted pointer is no
6233 * longer trusted, unless the field being accessed has explicitly been
6234 * marked as inheriting its parent's state of trust (either full or RCU).
6236 * 'cgroups' pointer is untrusted if task->cgroups dereference
6237 * happened in a sleepable program outside of bpf_rcu_read_lock()
6238 * section. In a non-sleepable program it's trusted while in RCU CS (aka MEM_RCU).
6239 * Note bpf_rcu_read_unlock() converts MEM_RCU pointers to PTR_UNTRUSTED.
6241 * A regular RCU-protected pointer with __rcu tag can also be deemed
6242 * trusted if we are in an RCU CS. Such pointer can be NULL.
6244 if (type_is_trusted(env, reg, field_name, btf_id)) {
6245 flag |= PTR_TRUSTED;
6246 } else if (in_rcu_cs(env) && !type_may_be_null(reg->type)) {
6247 if (type_is_rcu(env, reg, field_name, btf_id)) {
6248 /* ignore __rcu tag and mark it MEM_RCU */
6250 } else if (flag & MEM_RCU ||
6251 type_is_rcu_or_null(env, reg, field_name, btf_id)) {
6252 /* __rcu tagged pointers can be NULL */
6253 flag |= MEM_RCU | PTR_MAYBE_NULL;
6255 /* We always trust them */
6256 if (type_is_rcu_or_null(env, reg, field_name, btf_id) &&
6257 flag & PTR_UNTRUSTED)
6258 flag &= ~PTR_UNTRUSTED;
6259 } else if (flag & (MEM_PERCPU | MEM_USER)) {
6262 /* walking unknown pointers yields old deprecated PTR_TO_BTF_ID */
6263 clear_trusted_flags(&flag);
6267 * If not in RCU CS or MEM_RCU pointer can be NULL then
6268 * aggressively mark as untrusted otherwise such
6269 * pointers will be plain PTR_TO_BTF_ID without flags
6270 * and will be allowed to be passed into helpers for
6273 flag = PTR_UNTRUSTED;
6276 /* Old compat. Deprecated */
6277 clear_trusted_flags(&flag);
6280 if (atype == BPF_READ && value_regno >= 0)
6281 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id, flag);
6286 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
6287 struct bpf_reg_state *regs,
6288 int regno, int off, int size,
6289 enum bpf_access_type atype,
6292 struct bpf_reg_state *reg = regs + regno;
6293 struct bpf_map *map = reg->map_ptr;
6294 struct bpf_reg_state map_reg;
6295 enum bpf_type_flag flag = 0;
6296 const struct btf_type *t;
6302 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
6306 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
6307 verbose(env, "map_ptr access not supported for map type %d\n",
6312 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
6313 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
6315 if (!env->allow_ptr_leaks) {
6317 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
6323 verbose(env, "R%d is %s invalid negative access: off=%d\n",
6328 if (atype != BPF_READ) {
6329 verbose(env, "only read from %s is supported\n", tname);
6333 /* Simulate access to a PTR_TO_BTF_ID */
6334 memset(&map_reg, 0, sizeof(map_reg));
6335 mark_btf_ld_reg(env, &map_reg, 0, PTR_TO_BTF_ID, btf_vmlinux, *map->ops->map_btf_id, 0);
6336 ret = btf_struct_access(&env->log, &map_reg, off, size, atype, &btf_id, &flag, NULL);
6340 if (value_regno >= 0)
6341 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id, flag);
6346 /* Check that the stack access at the given offset is within bounds. The
6347 * maximum valid offset is -1.
6349 * The minimum valid offset is -MAX_BPF_STACK for writes, and
6350 * -state->allocated_stack for reads.
6352 static int check_stack_slot_within_bounds(int off,
6353 struct bpf_func_state *state,
6354 enum bpf_access_type t)
6359 min_valid_off = -MAX_BPF_STACK;
6361 min_valid_off = -state->allocated_stack;
6363 if (off < min_valid_off || off > -1)
6368 /* Check that the stack access at 'regno + off' falls within the maximum stack
6371 * 'off' includes `regno->offset`, but not its dynamic part (if any).
6373 static int check_stack_access_within_bounds(
6374 struct bpf_verifier_env *env,
6375 int regno, int off, int access_size,
6376 enum bpf_access_src src, enum bpf_access_type type)
6378 struct bpf_reg_state *regs = cur_regs(env);
6379 struct bpf_reg_state *reg = regs + regno;
6380 struct bpf_func_state *state = func(env, reg);
6381 int min_off, max_off;
6385 if (src == ACCESS_HELPER)
6386 /* We don't know if helpers are reading or writing (or both). */
6387 err_extra = " indirect access to";
6388 else if (type == BPF_READ)
6389 err_extra = " read from";
6391 err_extra = " write to";
6393 if (tnum_is_const(reg->var_off)) {
6394 min_off = reg->var_off.value + off;
6395 if (access_size > 0)
6396 max_off = min_off + access_size - 1;
6400 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
6401 reg->smin_value <= -BPF_MAX_VAR_OFF) {
6402 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
6406 min_off = reg->smin_value + off;
6407 if (access_size > 0)
6408 max_off = reg->smax_value + off + access_size - 1;
6413 err = check_stack_slot_within_bounds(min_off, state, type);
6415 err = check_stack_slot_within_bounds(max_off, state, type);
6418 if (tnum_is_const(reg->var_off)) {
6419 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
6420 err_extra, regno, off, access_size);
6424 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6425 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
6426 err_extra, regno, tn_buf, access_size);
6432 /* check whether memory at (regno + off) is accessible for t = (read | write)
6433 * if t==write, value_regno is a register which value is stored into memory
6434 * if t==read, value_regno is a register which will receive the value from memory
6435 * if t==write && value_regno==-1, some unknown value is stored into memory
6436 * if t==read && value_regno==-1, don't care what we read from memory
6438 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
6439 int off, int bpf_size, enum bpf_access_type t,
6440 int value_regno, bool strict_alignment_once, bool is_ldsx)
6442 struct bpf_reg_state *regs = cur_regs(env);
6443 struct bpf_reg_state *reg = regs + regno;
6444 struct bpf_func_state *state;
6447 size = bpf_size_to_bytes(bpf_size);
6451 /* alignment checks will add in reg->off themselves */
6452 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
6456 /* for access checks, reg->off is just part of off */
6459 if (reg->type == PTR_TO_MAP_KEY) {
6460 if (t == BPF_WRITE) {
6461 verbose(env, "write to change key R%d not allowed\n", regno);
6465 err = check_mem_region_access(env, regno, off, size,
6466 reg->map_ptr->key_size, false);
6469 if (value_regno >= 0)
6470 mark_reg_unknown(env, regs, value_regno);
6471 } else if (reg->type == PTR_TO_MAP_VALUE) {
6472 struct btf_field *kptr_field = NULL;
6474 if (t == BPF_WRITE && value_regno >= 0 &&
6475 is_pointer_value(env, value_regno)) {
6476 verbose(env, "R%d leaks addr into map\n", value_regno);
6479 err = check_map_access_type(env, regno, off, size, t);
6482 err = check_map_access(env, regno, off, size, false, ACCESS_DIRECT);
6485 if (tnum_is_const(reg->var_off))
6486 kptr_field = btf_record_find(reg->map_ptr->record,
6487 off + reg->var_off.value, BPF_KPTR);
6489 err = check_map_kptr_access(env, regno, value_regno, insn_idx, kptr_field);
6490 } else if (t == BPF_READ && value_regno >= 0) {
6491 struct bpf_map *map = reg->map_ptr;
6493 /* if map is read-only, track its contents as scalars */
6494 if (tnum_is_const(reg->var_off) &&
6495 bpf_map_is_rdonly(map) &&
6496 map->ops->map_direct_value_addr) {
6497 int map_off = off + reg->var_off.value;
6500 err = bpf_map_direct_read(map, map_off, size,
6505 regs[value_regno].type = SCALAR_VALUE;
6506 __mark_reg_known(®s[value_regno], val);
6508 mark_reg_unknown(env, regs, value_regno);
6511 } else if (base_type(reg->type) == PTR_TO_MEM) {
6512 bool rdonly_mem = type_is_rdonly_mem(reg->type);
6514 if (type_may_be_null(reg->type)) {
6515 verbose(env, "R%d invalid mem access '%s'\n", regno,
6516 reg_type_str(env, reg->type));
6520 if (t == BPF_WRITE && rdonly_mem) {
6521 verbose(env, "R%d cannot write into %s\n",
6522 regno, reg_type_str(env, reg->type));
6526 if (t == BPF_WRITE && value_regno >= 0 &&
6527 is_pointer_value(env, value_regno)) {
6528 verbose(env, "R%d leaks addr into mem\n", value_regno);
6532 err = check_mem_region_access(env, regno, off, size,
6533 reg->mem_size, false);
6534 if (!err && value_regno >= 0 && (t == BPF_READ || rdonly_mem))
6535 mark_reg_unknown(env, regs, value_regno);
6536 } else if (reg->type == PTR_TO_CTX) {
6537 enum bpf_reg_type reg_type = SCALAR_VALUE;
6538 struct btf *btf = NULL;
6541 if (t == BPF_WRITE && value_regno >= 0 &&
6542 is_pointer_value(env, value_regno)) {
6543 verbose(env, "R%d leaks addr into ctx\n", value_regno);
6547 err = check_ptr_off_reg(env, reg, regno);
6551 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf,
6554 verbose_linfo(env, insn_idx, "; ");
6555 if (!err && t == BPF_READ && value_regno >= 0) {
6556 /* ctx access returns either a scalar, or a
6557 * PTR_TO_PACKET[_META,_END]. In the latter
6558 * case, we know the offset is zero.
6560 if (reg_type == SCALAR_VALUE) {
6561 mark_reg_unknown(env, regs, value_regno);
6563 mark_reg_known_zero(env, regs,
6565 if (type_may_be_null(reg_type))
6566 regs[value_regno].id = ++env->id_gen;
6567 /* A load of ctx field could have different
6568 * actual load size with the one encoded in the
6569 * insn. When the dst is PTR, it is for sure not
6572 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
6573 if (base_type(reg_type) == PTR_TO_BTF_ID) {
6574 regs[value_regno].btf = btf;
6575 regs[value_regno].btf_id = btf_id;
6578 regs[value_regno].type = reg_type;
6581 } else if (reg->type == PTR_TO_STACK) {
6582 /* Basic bounds checks. */
6583 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
6587 state = func(env, reg);
6588 err = update_stack_depth(env, state, off);
6593 err = check_stack_read(env, regno, off, size,
6596 err = check_stack_write(env, regno, off, size,
6597 value_regno, insn_idx);
6598 } else if (reg_is_pkt_pointer(reg)) {
6599 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
6600 verbose(env, "cannot write into packet\n");
6603 if (t == BPF_WRITE && value_regno >= 0 &&
6604 is_pointer_value(env, value_regno)) {
6605 verbose(env, "R%d leaks addr into packet\n",
6609 err = check_packet_access(env, regno, off, size, false);
6610 if (!err && t == BPF_READ && value_regno >= 0)
6611 mark_reg_unknown(env, regs, value_regno);
6612 } else if (reg->type == PTR_TO_FLOW_KEYS) {
6613 if (t == BPF_WRITE && value_regno >= 0 &&
6614 is_pointer_value(env, value_regno)) {
6615 verbose(env, "R%d leaks addr into flow keys\n",
6620 err = check_flow_keys_access(env, off, size);
6621 if (!err && t == BPF_READ && value_regno >= 0)
6622 mark_reg_unknown(env, regs, value_regno);
6623 } else if (type_is_sk_pointer(reg->type)) {
6624 if (t == BPF_WRITE) {
6625 verbose(env, "R%d cannot write into %s\n",
6626 regno, reg_type_str(env, reg->type));
6629 err = check_sock_access(env, insn_idx, regno, off, size, t);
6630 if (!err && value_regno >= 0)
6631 mark_reg_unknown(env, regs, value_regno);
6632 } else if (reg->type == PTR_TO_TP_BUFFER) {
6633 err = check_tp_buffer_access(env, reg, regno, off, size);
6634 if (!err && t == BPF_READ && value_regno >= 0)
6635 mark_reg_unknown(env, regs, value_regno);
6636 } else if (base_type(reg->type) == PTR_TO_BTF_ID &&
6637 !type_may_be_null(reg->type)) {
6638 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
6640 } else if (reg->type == CONST_PTR_TO_MAP) {
6641 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
6643 } else if (base_type(reg->type) == PTR_TO_BUF) {
6644 bool rdonly_mem = type_is_rdonly_mem(reg->type);
6648 if (t == BPF_WRITE) {
6649 verbose(env, "R%d cannot write into %s\n",
6650 regno, reg_type_str(env, reg->type));
6653 max_access = &env->prog->aux->max_rdonly_access;
6655 max_access = &env->prog->aux->max_rdwr_access;
6658 err = check_buffer_access(env, reg, regno, off, size, false,
6661 if (!err && value_regno >= 0 && (rdonly_mem || t == BPF_READ))
6662 mark_reg_unknown(env, regs, value_regno);
6664 verbose(env, "R%d invalid mem access '%s'\n", regno,
6665 reg_type_str(env, reg->type));
6669 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
6670 regs[value_regno].type == SCALAR_VALUE) {
6672 /* b/h/w load zero-extends, mark upper bits as known 0 */
6673 coerce_reg_to_size(®s[value_regno], size);
6675 coerce_reg_to_size_sx(®s[value_regno], size);
6680 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
6685 switch (insn->imm) {
6687 case BPF_ADD | BPF_FETCH:
6689 case BPF_AND | BPF_FETCH:
6691 case BPF_OR | BPF_FETCH:
6693 case BPF_XOR | BPF_FETCH:
6698 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
6702 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
6703 verbose(env, "invalid atomic operand size\n");
6707 /* check src1 operand */
6708 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6712 /* check src2 operand */
6713 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6717 if (insn->imm == BPF_CMPXCHG) {
6718 /* Check comparison of R0 with memory location */
6719 const u32 aux_reg = BPF_REG_0;
6721 err = check_reg_arg(env, aux_reg, SRC_OP);
6725 if (is_pointer_value(env, aux_reg)) {
6726 verbose(env, "R%d leaks addr into mem\n", aux_reg);
6731 if (is_pointer_value(env, insn->src_reg)) {
6732 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
6736 if (is_ctx_reg(env, insn->dst_reg) ||
6737 is_pkt_reg(env, insn->dst_reg) ||
6738 is_flow_key_reg(env, insn->dst_reg) ||
6739 is_sk_reg(env, insn->dst_reg)) {
6740 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
6742 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
6746 if (insn->imm & BPF_FETCH) {
6747 if (insn->imm == BPF_CMPXCHG)
6748 load_reg = BPF_REG_0;
6750 load_reg = insn->src_reg;
6752 /* check and record load of old value */
6753 err = check_reg_arg(env, load_reg, DST_OP);
6757 /* This instruction accesses a memory location but doesn't
6758 * actually load it into a register.
6763 /* Check whether we can read the memory, with second call for fetch
6764 * case to simulate the register fill.
6766 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
6767 BPF_SIZE(insn->code), BPF_READ, -1, true, false);
6768 if (!err && load_reg >= 0)
6769 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
6770 BPF_SIZE(insn->code), BPF_READ, load_reg,
6775 /* Check whether we can write into the same memory. */
6776 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
6777 BPF_SIZE(insn->code), BPF_WRITE, -1, true, false);
6784 /* When register 'regno' is used to read the stack (either directly or through
6785 * a helper function) make sure that it's within stack boundary and, depending
6786 * on the access type, that all elements of the stack are initialized.
6788 * 'off' includes 'regno->off', but not its dynamic part (if any).
6790 * All registers that have been spilled on the stack in the slots within the
6791 * read offsets are marked as read.
6793 static int check_stack_range_initialized(
6794 struct bpf_verifier_env *env, int regno, int off,
6795 int access_size, bool zero_size_allowed,
6796 enum bpf_access_src type, struct bpf_call_arg_meta *meta)
6798 struct bpf_reg_state *reg = reg_state(env, regno);
6799 struct bpf_func_state *state = func(env, reg);
6800 int err, min_off, max_off, i, j, slot, spi;
6801 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
6802 enum bpf_access_type bounds_check_type;
6803 /* Some accesses can write anything into the stack, others are
6806 bool clobber = false;
6808 if (access_size == 0 && !zero_size_allowed) {
6809 verbose(env, "invalid zero-sized read\n");
6813 if (type == ACCESS_HELPER) {
6814 /* The bounds checks for writes are more permissive than for
6815 * reads. However, if raw_mode is not set, we'll do extra
6818 bounds_check_type = BPF_WRITE;
6821 bounds_check_type = BPF_READ;
6823 err = check_stack_access_within_bounds(env, regno, off, access_size,
6824 type, bounds_check_type);
6829 if (tnum_is_const(reg->var_off)) {
6830 min_off = max_off = reg->var_off.value + off;
6832 /* Variable offset is prohibited for unprivileged mode for
6833 * simplicity since it requires corresponding support in
6834 * Spectre masking for stack ALU.
6835 * See also retrieve_ptr_limit().
6837 if (!env->bypass_spec_v1) {
6840 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6841 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
6842 regno, err_extra, tn_buf);
6845 /* Only initialized buffer on stack is allowed to be accessed
6846 * with variable offset. With uninitialized buffer it's hard to
6847 * guarantee that whole memory is marked as initialized on
6848 * helper return since specific bounds are unknown what may
6849 * cause uninitialized stack leaking.
6851 if (meta && meta->raw_mode)
6854 min_off = reg->smin_value + off;
6855 max_off = reg->smax_value + off;
6858 if (meta && meta->raw_mode) {
6859 /* Ensure we won't be overwriting dynptrs when simulating byte
6860 * by byte access in check_helper_call using meta.access_size.
6861 * This would be a problem if we have a helper in the future
6864 * helper(uninit_mem, len, dynptr)
6866 * Now, uninint_mem may overlap with dynptr pointer. Hence, it
6867 * may end up writing to dynptr itself when touching memory from
6868 * arg 1. This can be relaxed on a case by case basis for known
6869 * safe cases, but reject due to the possibilitiy of aliasing by
6872 for (i = min_off; i < max_off + access_size; i++) {
6873 int stack_off = -i - 1;
6876 /* raw_mode may write past allocated_stack */
6877 if (state->allocated_stack <= stack_off)
6879 if (state->stack[spi].slot_type[stack_off % BPF_REG_SIZE] == STACK_DYNPTR) {
6880 verbose(env, "potential write to dynptr at off=%d disallowed\n", i);
6884 meta->access_size = access_size;
6885 meta->regno = regno;
6889 for (i = min_off; i < max_off + access_size; i++) {
6893 spi = slot / BPF_REG_SIZE;
6894 if (state->allocated_stack <= slot)
6896 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
6897 if (*stype == STACK_MISC)
6899 if ((*stype == STACK_ZERO) ||
6900 (*stype == STACK_INVALID && env->allow_uninit_stack)) {
6902 /* helper can write anything into the stack */
6903 *stype = STACK_MISC;
6908 if (is_spilled_reg(&state->stack[spi]) &&
6909 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
6910 env->allow_ptr_leaks)) {
6912 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
6913 for (j = 0; j < BPF_REG_SIZE; j++)
6914 scrub_spilled_slot(&state->stack[spi].slot_type[j]);
6920 if (tnum_is_const(reg->var_off)) {
6921 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
6922 err_extra, regno, min_off, i - min_off, access_size);
6926 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6927 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
6928 err_extra, regno, tn_buf, i - min_off, access_size);
6932 /* reading any byte out of 8-byte 'spill_slot' will cause
6933 * the whole slot to be marked as 'read'
6935 mark_reg_read(env, &state->stack[spi].spilled_ptr,
6936 state->stack[spi].spilled_ptr.parent,
6938 /* We do not set REG_LIVE_WRITTEN for stack slot, as we can not
6939 * be sure that whether stack slot is written to or not. Hence,
6940 * we must still conservatively propagate reads upwards even if
6941 * helper may write to the entire memory range.
6944 return update_stack_depth(env, state, min_off);
6947 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
6948 int access_size, bool zero_size_allowed,
6949 struct bpf_call_arg_meta *meta)
6951 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
6954 switch (base_type(reg->type)) {
6956 case PTR_TO_PACKET_META:
6957 return check_packet_access(env, regno, reg->off, access_size,
6959 case PTR_TO_MAP_KEY:
6960 if (meta && meta->raw_mode) {
6961 verbose(env, "R%d cannot write into %s\n", regno,
6962 reg_type_str(env, reg->type));
6965 return check_mem_region_access(env, regno, reg->off, access_size,
6966 reg->map_ptr->key_size, false);
6967 case PTR_TO_MAP_VALUE:
6968 if (check_map_access_type(env, regno, reg->off, access_size,
6969 meta && meta->raw_mode ? BPF_WRITE :
6972 return check_map_access(env, regno, reg->off, access_size,
6973 zero_size_allowed, ACCESS_HELPER);
6975 if (type_is_rdonly_mem(reg->type)) {
6976 if (meta && meta->raw_mode) {
6977 verbose(env, "R%d cannot write into %s\n", regno,
6978 reg_type_str(env, reg->type));
6982 return check_mem_region_access(env, regno, reg->off,
6983 access_size, reg->mem_size,
6986 if (type_is_rdonly_mem(reg->type)) {
6987 if (meta && meta->raw_mode) {
6988 verbose(env, "R%d cannot write into %s\n", regno,
6989 reg_type_str(env, reg->type));
6993 max_access = &env->prog->aux->max_rdonly_access;
6995 max_access = &env->prog->aux->max_rdwr_access;
6997 return check_buffer_access(env, reg, regno, reg->off,
6998 access_size, zero_size_allowed,
7001 return check_stack_range_initialized(
7003 regno, reg->off, access_size,
7004 zero_size_allowed, ACCESS_HELPER, meta);
7006 return check_ptr_to_btf_access(env, regs, regno, reg->off,
7007 access_size, BPF_READ, -1);
7009 /* in case the function doesn't know how to access the context,
7010 * (because we are in a program of type SYSCALL for example), we
7011 * can not statically check its size.
7012 * Dynamically check it now.
7014 if (!env->ops->convert_ctx_access) {
7015 enum bpf_access_type atype = meta && meta->raw_mode ? BPF_WRITE : BPF_READ;
7016 int offset = access_size - 1;
7018 /* Allow zero-byte read from PTR_TO_CTX */
7019 if (access_size == 0)
7020 return zero_size_allowed ? 0 : -EACCES;
7022 return check_mem_access(env, env->insn_idx, regno, offset, BPF_B,
7023 atype, -1, false, false);
7027 default: /* scalar_value or invalid ptr */
7028 /* Allow zero-byte read from NULL, regardless of pointer type */
7029 if (zero_size_allowed && access_size == 0 &&
7030 register_is_null(reg))
7033 verbose(env, "R%d type=%s ", regno,
7034 reg_type_str(env, reg->type));
7035 verbose(env, "expected=%s\n", reg_type_str(env, PTR_TO_STACK));
7040 static int check_mem_size_reg(struct bpf_verifier_env *env,
7041 struct bpf_reg_state *reg, u32 regno,
7042 bool zero_size_allowed,
7043 struct bpf_call_arg_meta *meta)
7047 /* This is used to refine r0 return value bounds for helpers
7048 * that enforce this value as an upper bound on return values.
7049 * See do_refine_retval_range() for helpers that can refine
7050 * the return value. C type of helper is u32 so we pull register
7051 * bound from umax_value however, if negative verifier errors
7052 * out. Only upper bounds can be learned because retval is an
7053 * int type and negative retvals are allowed.
7055 meta->msize_max_value = reg->umax_value;
7057 /* The register is SCALAR_VALUE; the access check
7058 * happens using its boundaries.
7060 if (!tnum_is_const(reg->var_off))
7061 /* For unprivileged variable accesses, disable raw
7062 * mode so that the program is required to
7063 * initialize all the memory that the helper could
7064 * just partially fill up.
7068 if (reg->smin_value < 0) {
7069 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
7074 if (reg->umin_value == 0) {
7075 err = check_helper_mem_access(env, regno - 1, 0,
7082 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
7083 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
7087 err = check_helper_mem_access(env, regno - 1,
7089 zero_size_allowed, meta);
7091 err = mark_chain_precision(env, regno);
7095 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
7096 u32 regno, u32 mem_size)
7098 bool may_be_null = type_may_be_null(reg->type);
7099 struct bpf_reg_state saved_reg;
7100 struct bpf_call_arg_meta meta;
7103 if (register_is_null(reg))
7106 memset(&meta, 0, sizeof(meta));
7107 /* Assuming that the register contains a value check if the memory
7108 * access is safe. Temporarily save and restore the register's state as
7109 * the conversion shouldn't be visible to a caller.
7113 mark_ptr_not_null_reg(reg);
7116 err = check_helper_mem_access(env, regno, mem_size, true, &meta);
7117 /* Check access for BPF_WRITE */
7118 meta.raw_mode = true;
7119 err = err ?: check_helper_mem_access(env, regno, mem_size, true, &meta);
7127 static int check_kfunc_mem_size_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
7130 struct bpf_reg_state *mem_reg = &cur_regs(env)[regno - 1];
7131 bool may_be_null = type_may_be_null(mem_reg->type);
7132 struct bpf_reg_state saved_reg;
7133 struct bpf_call_arg_meta meta;
7136 WARN_ON_ONCE(regno < BPF_REG_2 || regno > BPF_REG_5);
7138 memset(&meta, 0, sizeof(meta));
7141 saved_reg = *mem_reg;
7142 mark_ptr_not_null_reg(mem_reg);
7145 err = check_mem_size_reg(env, reg, regno, true, &meta);
7146 /* Check access for BPF_WRITE */
7147 meta.raw_mode = true;
7148 err = err ?: check_mem_size_reg(env, reg, regno, true, &meta);
7151 *mem_reg = saved_reg;
7155 /* Implementation details:
7156 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL.
7157 * bpf_obj_new returns PTR_TO_BTF_ID | MEM_ALLOC | PTR_MAYBE_NULL.
7158 * Two bpf_map_lookups (even with the same key) will have different reg->id.
7159 * Two separate bpf_obj_new will also have different reg->id.
7160 * For traditional PTR_TO_MAP_VALUE or PTR_TO_BTF_ID | MEM_ALLOC, the verifier
7161 * clears reg->id after value_or_null->value transition, since the verifier only
7162 * cares about the range of access to valid map value pointer and doesn't care
7163 * about actual address of the map element.
7164 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
7165 * reg->id > 0 after value_or_null->value transition. By doing so
7166 * two bpf_map_lookups will be considered two different pointers that
7167 * point to different bpf_spin_locks. Likewise for pointers to allocated objects
7168 * returned from bpf_obj_new.
7169 * The verifier allows taking only one bpf_spin_lock at a time to avoid
7171 * Since only one bpf_spin_lock is allowed the checks are simpler than
7172 * reg_is_refcounted() logic. The verifier needs to remember only
7173 * one spin_lock instead of array of acquired_refs.
7174 * cur_state->active_lock remembers which map value element or allocated
7175 * object got locked and clears it after bpf_spin_unlock.
7177 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
7180 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7181 struct bpf_verifier_state *cur = env->cur_state;
7182 bool is_const = tnum_is_const(reg->var_off);
7183 u64 val = reg->var_off.value;
7184 struct bpf_map *map = NULL;
7185 struct btf *btf = NULL;
7186 struct btf_record *rec;
7190 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
7194 if (reg->type == PTR_TO_MAP_VALUE) {
7198 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
7206 rec = reg_btf_record(reg);
7207 if (!btf_record_has_field(rec, BPF_SPIN_LOCK)) {
7208 verbose(env, "%s '%s' has no valid bpf_spin_lock\n", map ? "map" : "local",
7209 map ? map->name : "kptr");
7212 if (rec->spin_lock_off != val + reg->off) {
7213 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock' that is at %d\n",
7214 val + reg->off, rec->spin_lock_off);
7218 if (cur->active_lock.ptr) {
7220 "Locking two bpf_spin_locks are not allowed\n");
7224 cur->active_lock.ptr = map;
7226 cur->active_lock.ptr = btf;
7227 cur->active_lock.id = reg->id;
7236 if (!cur->active_lock.ptr) {
7237 verbose(env, "bpf_spin_unlock without taking a lock\n");
7240 if (cur->active_lock.ptr != ptr ||
7241 cur->active_lock.id != reg->id) {
7242 verbose(env, "bpf_spin_unlock of different lock\n");
7246 invalidate_non_owning_refs(env);
7248 cur->active_lock.ptr = NULL;
7249 cur->active_lock.id = 0;
7254 static int process_timer_func(struct bpf_verifier_env *env, int regno,
7255 struct bpf_call_arg_meta *meta)
7257 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7258 bool is_const = tnum_is_const(reg->var_off);
7259 struct bpf_map *map = reg->map_ptr;
7260 u64 val = reg->var_off.value;
7264 "R%d doesn't have constant offset. bpf_timer has to be at the constant offset\n",
7269 verbose(env, "map '%s' has to have BTF in order to use bpf_timer\n",
7273 if (!btf_record_has_field(map->record, BPF_TIMER)) {
7274 verbose(env, "map '%s' has no valid bpf_timer\n", map->name);
7277 if (map->record->timer_off != val + reg->off) {
7278 verbose(env, "off %lld doesn't point to 'struct bpf_timer' that is at %d\n",
7279 val + reg->off, map->record->timer_off);
7282 if (meta->map_ptr) {
7283 verbose(env, "verifier bug. Two map pointers in a timer helper\n");
7286 meta->map_uid = reg->map_uid;
7287 meta->map_ptr = map;
7291 static int process_kptr_func(struct bpf_verifier_env *env, int regno,
7292 struct bpf_call_arg_meta *meta)
7294 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7295 struct bpf_map *map_ptr = reg->map_ptr;
7296 struct btf_field *kptr_field;
7299 if (!tnum_is_const(reg->var_off)) {
7301 "R%d doesn't have constant offset. kptr has to be at the constant offset\n",
7305 if (!map_ptr->btf) {
7306 verbose(env, "map '%s' has to have BTF in order to use bpf_kptr_xchg\n",
7310 if (!btf_record_has_field(map_ptr->record, BPF_KPTR)) {
7311 verbose(env, "map '%s' has no valid kptr\n", map_ptr->name);
7315 meta->map_ptr = map_ptr;
7316 kptr_off = reg->off + reg->var_off.value;
7317 kptr_field = btf_record_find(map_ptr->record, kptr_off, BPF_KPTR);
7319 verbose(env, "off=%d doesn't point to kptr\n", kptr_off);
7322 if (kptr_field->type != BPF_KPTR_REF) {
7323 verbose(env, "off=%d kptr isn't referenced kptr\n", kptr_off);
7326 meta->kptr_field = kptr_field;
7330 /* There are two register types representing a bpf_dynptr, one is PTR_TO_STACK
7331 * which points to a stack slot, and the other is CONST_PTR_TO_DYNPTR.
7333 * In both cases we deal with the first 8 bytes, but need to mark the next 8
7334 * bytes as STACK_DYNPTR in case of PTR_TO_STACK. In case of
7335 * CONST_PTR_TO_DYNPTR, we are guaranteed to get the beginning of the object.
7337 * Mutability of bpf_dynptr is at two levels, one is at the level of struct
7338 * bpf_dynptr itself, i.e. whether the helper is receiving a pointer to struct
7339 * bpf_dynptr or pointer to const struct bpf_dynptr. In the former case, it can
7340 * mutate the view of the dynptr and also possibly destroy it. In the latter
7341 * case, it cannot mutate the bpf_dynptr itself but it can still mutate the
7342 * memory that dynptr points to.
7344 * The verifier will keep track both levels of mutation (bpf_dynptr's in
7345 * reg->type and the memory's in reg->dynptr.type), but there is no support for
7346 * readonly dynptr view yet, hence only the first case is tracked and checked.
7348 * This is consistent with how C applies the const modifier to a struct object,
7349 * where the pointer itself inside bpf_dynptr becomes const but not what it
7352 * Helpers which do not mutate the bpf_dynptr set MEM_RDONLY in their argument
7353 * type, and declare it as 'const struct bpf_dynptr *' in their prototype.
7355 static int process_dynptr_func(struct bpf_verifier_env *env, int regno, int insn_idx,
7356 enum bpf_arg_type arg_type, int clone_ref_obj_id)
7358 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7361 /* MEM_UNINIT and MEM_RDONLY are exclusive, when applied to an
7362 * ARG_PTR_TO_DYNPTR (or ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_*):
7364 if ((arg_type & (MEM_UNINIT | MEM_RDONLY)) == (MEM_UNINIT | MEM_RDONLY)) {
7365 verbose(env, "verifier internal error: misconfigured dynptr helper type flags\n");
7369 /* MEM_UNINIT - Points to memory that is an appropriate candidate for
7370 * constructing a mutable bpf_dynptr object.
7372 * Currently, this is only possible with PTR_TO_STACK
7373 * pointing to a region of at least 16 bytes which doesn't
7374 * contain an existing bpf_dynptr.
7376 * MEM_RDONLY - Points to a initialized bpf_dynptr that will not be
7377 * mutated or destroyed. However, the memory it points to
7380 * None - Points to a initialized dynptr that can be mutated and
7381 * destroyed, including mutation of the memory it points
7384 if (arg_type & MEM_UNINIT) {
7387 if (!is_dynptr_reg_valid_uninit(env, reg)) {
7388 verbose(env, "Dynptr has to be an uninitialized dynptr\n");
7392 /* we write BPF_DW bits (8 bytes) at a time */
7393 for (i = 0; i < BPF_DYNPTR_SIZE; i += 8) {
7394 err = check_mem_access(env, insn_idx, regno,
7395 i, BPF_DW, BPF_WRITE, -1, false, false);
7400 err = mark_stack_slots_dynptr(env, reg, arg_type, insn_idx, clone_ref_obj_id);
7401 } else /* MEM_RDONLY and None case from above */ {
7402 /* For the reg->type == PTR_TO_STACK case, bpf_dynptr is never const */
7403 if (reg->type == CONST_PTR_TO_DYNPTR && !(arg_type & MEM_RDONLY)) {
7404 verbose(env, "cannot pass pointer to const bpf_dynptr, the helper mutates it\n");
7408 if (!is_dynptr_reg_valid_init(env, reg)) {
7410 "Expected an initialized dynptr as arg #%d\n",
7415 /* Fold modifiers (in this case, MEM_RDONLY) when checking expected type */
7416 if (!is_dynptr_type_expected(env, reg, arg_type & ~MEM_RDONLY)) {
7418 "Expected a dynptr of type %s as arg #%d\n",
7419 dynptr_type_str(arg_to_dynptr_type(arg_type)), regno);
7423 err = mark_dynptr_read(env, reg);
7428 static u32 iter_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg, int spi)
7430 struct bpf_func_state *state = func(env, reg);
7432 return state->stack[spi].spilled_ptr.ref_obj_id;
7435 static bool is_iter_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7437 return meta->kfunc_flags & (KF_ITER_NEW | KF_ITER_NEXT | KF_ITER_DESTROY);
7440 static bool is_iter_new_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7442 return meta->kfunc_flags & KF_ITER_NEW;
7445 static bool is_iter_next_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7447 return meta->kfunc_flags & KF_ITER_NEXT;
7450 static bool is_iter_destroy_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7452 return meta->kfunc_flags & KF_ITER_DESTROY;
7455 static bool is_kfunc_arg_iter(struct bpf_kfunc_call_arg_meta *meta, int arg)
7457 /* btf_check_iter_kfuncs() guarantees that first argument of any iter
7458 * kfunc is iter state pointer
7460 return arg == 0 && is_iter_kfunc(meta);
7463 static int process_iter_arg(struct bpf_verifier_env *env, int regno, int insn_idx,
7464 struct bpf_kfunc_call_arg_meta *meta)
7466 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7467 const struct btf_type *t;
7468 const struct btf_param *arg;
7469 int spi, err, i, nr_slots;
7472 /* btf_check_iter_kfuncs() ensures we don't need to validate anything here */
7473 arg = &btf_params(meta->func_proto)[0];
7474 t = btf_type_skip_modifiers(meta->btf, arg->type, NULL); /* PTR */
7475 t = btf_type_skip_modifiers(meta->btf, t->type, &btf_id); /* STRUCT */
7476 nr_slots = t->size / BPF_REG_SIZE;
7478 if (is_iter_new_kfunc(meta)) {
7479 /* bpf_iter_<type>_new() expects pointer to uninit iter state */
7480 if (!is_iter_reg_valid_uninit(env, reg, nr_slots)) {
7481 verbose(env, "expected uninitialized iter_%s as arg #%d\n",
7482 iter_type_str(meta->btf, btf_id), regno);
7486 for (i = 0; i < nr_slots * 8; i += BPF_REG_SIZE) {
7487 err = check_mem_access(env, insn_idx, regno,
7488 i, BPF_DW, BPF_WRITE, -1, false, false);
7493 err = mark_stack_slots_iter(env, reg, insn_idx, meta->btf, btf_id, nr_slots);
7497 /* iter_next() or iter_destroy() expect initialized iter state*/
7498 if (!is_iter_reg_valid_init(env, reg, meta->btf, btf_id, nr_slots)) {
7499 verbose(env, "expected an initialized iter_%s as arg #%d\n",
7500 iter_type_str(meta->btf, btf_id), regno);
7504 spi = iter_get_spi(env, reg, nr_slots);
7508 err = mark_iter_read(env, reg, spi, nr_slots);
7512 /* remember meta->iter info for process_iter_next_call() */
7513 meta->iter.spi = spi;
7514 meta->iter.frameno = reg->frameno;
7515 meta->ref_obj_id = iter_ref_obj_id(env, reg, spi);
7517 if (is_iter_destroy_kfunc(meta)) {
7518 err = unmark_stack_slots_iter(env, reg, nr_slots);
7527 /* process_iter_next_call() is called when verifier gets to iterator's next
7528 * "method" (e.g., bpf_iter_num_next() for numbers iterator) call. We'll refer
7529 * to it as just "iter_next()" in comments below.
7531 * BPF verifier relies on a crucial contract for any iter_next()
7532 * implementation: it should *eventually* return NULL, and once that happens
7533 * it should keep returning NULL. That is, once iterator exhausts elements to
7534 * iterate, it should never reset or spuriously return new elements.
7536 * With the assumption of such contract, process_iter_next_call() simulates
7537 * a fork in the verifier state to validate loop logic correctness and safety
7538 * without having to simulate infinite amount of iterations.
7540 * In current state, we first assume that iter_next() returned NULL and
7541 * iterator state is set to DRAINED (BPF_ITER_STATE_DRAINED). In such
7542 * conditions we should not form an infinite loop and should eventually reach
7545 * Besides that, we also fork current state and enqueue it for later
7546 * verification. In a forked state we keep iterator state as ACTIVE
7547 * (BPF_ITER_STATE_ACTIVE) and assume non-NULL return from iter_next(). We
7548 * also bump iteration depth to prevent erroneous infinite loop detection
7549 * later on (see iter_active_depths_differ() comment for details). In this
7550 * state we assume that we'll eventually loop back to another iter_next()
7551 * calls (it could be in exactly same location or in some other instruction,
7552 * it doesn't matter, we don't make any unnecessary assumptions about this,
7553 * everything revolves around iterator state in a stack slot, not which
7554 * instruction is calling iter_next()). When that happens, we either will come
7555 * to iter_next() with equivalent state and can conclude that next iteration
7556 * will proceed in exactly the same way as we just verified, so it's safe to
7557 * assume that loop converges. If not, we'll go on another iteration
7558 * simulation with a different input state, until all possible starting states
7559 * are validated or we reach maximum number of instructions limit.
7561 * This way, we will either exhaustively discover all possible input states
7562 * that iterator loop can start with and eventually will converge, or we'll
7563 * effectively regress into bounded loop simulation logic and either reach
7564 * maximum number of instructions if loop is not provably convergent, or there
7565 * is some statically known limit on number of iterations (e.g., if there is
7566 * an explicit `if n > 100 then break;` statement somewhere in the loop).
7568 * One very subtle but very important aspect is that we *always* simulate NULL
7569 * condition first (as the current state) before we simulate non-NULL case.
7570 * This has to do with intricacies of scalar precision tracking. By simulating
7571 * "exit condition" of iter_next() returning NULL first, we make sure all the
7572 * relevant precision marks *that will be set **after** we exit iterator loop*
7573 * are propagated backwards to common parent state of NULL and non-NULL
7574 * branches. Thanks to that, state equivalence checks done later in forked
7575 * state, when reaching iter_next() for ACTIVE iterator, can assume that
7576 * precision marks are finalized and won't change. Because simulating another
7577 * ACTIVE iterator iteration won't change them (because given same input
7578 * states we'll end up with exactly same output states which we are currently
7579 * comparing; and verification after the loop already propagated back what
7580 * needs to be **additionally** tracked as precise). It's subtle, grok
7581 * precision tracking for more intuitive understanding.
7583 static int process_iter_next_call(struct bpf_verifier_env *env, int insn_idx,
7584 struct bpf_kfunc_call_arg_meta *meta)
7586 struct bpf_verifier_state *cur_st = env->cur_state, *queued_st;
7587 struct bpf_func_state *cur_fr = cur_st->frame[cur_st->curframe], *queued_fr;
7588 struct bpf_reg_state *cur_iter, *queued_iter;
7589 int iter_frameno = meta->iter.frameno;
7590 int iter_spi = meta->iter.spi;
7592 BTF_TYPE_EMIT(struct bpf_iter);
7594 cur_iter = &env->cur_state->frame[iter_frameno]->stack[iter_spi].spilled_ptr;
7596 if (cur_iter->iter.state != BPF_ITER_STATE_ACTIVE &&
7597 cur_iter->iter.state != BPF_ITER_STATE_DRAINED) {
7598 verbose(env, "verifier internal error: unexpected iterator state %d (%s)\n",
7599 cur_iter->iter.state, iter_state_str(cur_iter->iter.state));
7603 if (cur_iter->iter.state == BPF_ITER_STATE_ACTIVE) {
7604 /* branch out active iter state */
7605 queued_st = push_stack(env, insn_idx + 1, insn_idx, false);
7609 queued_iter = &queued_st->frame[iter_frameno]->stack[iter_spi].spilled_ptr;
7610 queued_iter->iter.state = BPF_ITER_STATE_ACTIVE;
7611 queued_iter->iter.depth++;
7613 queued_fr = queued_st->frame[queued_st->curframe];
7614 mark_ptr_not_null_reg(&queued_fr->regs[BPF_REG_0]);
7617 /* switch to DRAINED state, but keep the depth unchanged */
7618 /* mark current iter state as drained and assume returned NULL */
7619 cur_iter->iter.state = BPF_ITER_STATE_DRAINED;
7620 __mark_reg_const_zero(&cur_fr->regs[BPF_REG_0]);
7625 static bool arg_type_is_mem_size(enum bpf_arg_type type)
7627 return type == ARG_CONST_SIZE ||
7628 type == ARG_CONST_SIZE_OR_ZERO;
7631 static bool arg_type_is_release(enum bpf_arg_type type)
7633 return type & OBJ_RELEASE;
7636 static bool arg_type_is_dynptr(enum bpf_arg_type type)
7638 return base_type(type) == ARG_PTR_TO_DYNPTR;
7641 static int int_ptr_type_to_size(enum bpf_arg_type type)
7643 if (type == ARG_PTR_TO_INT)
7645 else if (type == ARG_PTR_TO_LONG)
7651 static int resolve_map_arg_type(struct bpf_verifier_env *env,
7652 const struct bpf_call_arg_meta *meta,
7653 enum bpf_arg_type *arg_type)
7655 if (!meta->map_ptr) {
7656 /* kernel subsystem misconfigured verifier */
7657 verbose(env, "invalid map_ptr to access map->type\n");
7661 switch (meta->map_ptr->map_type) {
7662 case BPF_MAP_TYPE_SOCKMAP:
7663 case BPF_MAP_TYPE_SOCKHASH:
7664 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
7665 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
7667 verbose(env, "invalid arg_type for sockmap/sockhash\n");
7671 case BPF_MAP_TYPE_BLOOM_FILTER:
7672 if (meta->func_id == BPF_FUNC_map_peek_elem)
7673 *arg_type = ARG_PTR_TO_MAP_VALUE;
7681 struct bpf_reg_types {
7682 const enum bpf_reg_type types[10];
7686 static const struct bpf_reg_types sock_types = {
7696 static const struct bpf_reg_types btf_id_sock_common_types = {
7703 PTR_TO_BTF_ID | PTR_TRUSTED,
7705 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
7709 static const struct bpf_reg_types mem_types = {
7717 PTR_TO_MEM | MEM_RINGBUF,
7719 PTR_TO_BTF_ID | PTR_TRUSTED,
7723 static const struct bpf_reg_types int_ptr_types = {
7733 static const struct bpf_reg_types spin_lock_types = {
7736 PTR_TO_BTF_ID | MEM_ALLOC,
7740 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
7741 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
7742 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
7743 static const struct bpf_reg_types ringbuf_mem_types = { .types = { PTR_TO_MEM | MEM_RINGBUF } };
7744 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
7745 static const struct bpf_reg_types btf_ptr_types = {
7748 PTR_TO_BTF_ID | PTR_TRUSTED,
7749 PTR_TO_BTF_ID | MEM_RCU,
7752 static const struct bpf_reg_types percpu_btf_ptr_types = {
7754 PTR_TO_BTF_ID | MEM_PERCPU,
7755 PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED,
7758 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
7759 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
7760 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
7761 static const struct bpf_reg_types timer_types = { .types = { PTR_TO_MAP_VALUE } };
7762 static const struct bpf_reg_types kptr_types = { .types = { PTR_TO_MAP_VALUE } };
7763 static const struct bpf_reg_types dynptr_types = {
7766 CONST_PTR_TO_DYNPTR,
7770 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
7771 [ARG_PTR_TO_MAP_KEY] = &mem_types,
7772 [ARG_PTR_TO_MAP_VALUE] = &mem_types,
7773 [ARG_CONST_SIZE] = &scalar_types,
7774 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
7775 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
7776 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
7777 [ARG_PTR_TO_CTX] = &context_types,
7778 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
7780 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
7782 [ARG_PTR_TO_SOCKET] = &fullsock_types,
7783 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
7784 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
7785 [ARG_PTR_TO_MEM] = &mem_types,
7786 [ARG_PTR_TO_RINGBUF_MEM] = &ringbuf_mem_types,
7787 [ARG_PTR_TO_INT] = &int_ptr_types,
7788 [ARG_PTR_TO_LONG] = &int_ptr_types,
7789 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
7790 [ARG_PTR_TO_FUNC] = &func_ptr_types,
7791 [ARG_PTR_TO_STACK] = &stack_ptr_types,
7792 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
7793 [ARG_PTR_TO_TIMER] = &timer_types,
7794 [ARG_PTR_TO_KPTR] = &kptr_types,
7795 [ARG_PTR_TO_DYNPTR] = &dynptr_types,
7798 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
7799 enum bpf_arg_type arg_type,
7800 const u32 *arg_btf_id,
7801 struct bpf_call_arg_meta *meta)
7803 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7804 enum bpf_reg_type expected, type = reg->type;
7805 const struct bpf_reg_types *compatible;
7808 compatible = compatible_reg_types[base_type(arg_type)];
7810 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
7814 /* ARG_PTR_TO_MEM + RDONLY is compatible with PTR_TO_MEM and PTR_TO_MEM + RDONLY,
7815 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM and NOT with PTR_TO_MEM + RDONLY
7817 * Same for MAYBE_NULL:
7819 * ARG_PTR_TO_MEM + MAYBE_NULL is compatible with PTR_TO_MEM and PTR_TO_MEM + MAYBE_NULL,
7820 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM but NOT with PTR_TO_MEM + MAYBE_NULL
7822 * ARG_PTR_TO_MEM is compatible with PTR_TO_MEM that is tagged with a dynptr type.
7824 * Therefore we fold these flags depending on the arg_type before comparison.
7826 if (arg_type & MEM_RDONLY)
7827 type &= ~MEM_RDONLY;
7828 if (arg_type & PTR_MAYBE_NULL)
7829 type &= ~PTR_MAYBE_NULL;
7830 if (base_type(arg_type) == ARG_PTR_TO_MEM)
7831 type &= ~DYNPTR_TYPE_FLAG_MASK;
7833 if (meta->func_id == BPF_FUNC_kptr_xchg && type_is_alloc(type))
7836 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
7837 expected = compatible->types[i];
7838 if (expected == NOT_INIT)
7841 if (type == expected)
7845 verbose(env, "R%d type=%s expected=", regno, reg_type_str(env, reg->type));
7846 for (j = 0; j + 1 < i; j++)
7847 verbose(env, "%s, ", reg_type_str(env, compatible->types[j]));
7848 verbose(env, "%s\n", reg_type_str(env, compatible->types[j]));
7852 if (base_type(reg->type) != PTR_TO_BTF_ID)
7855 if (compatible == &mem_types) {
7856 if (!(arg_type & MEM_RDONLY)) {
7858 "%s() may write into memory pointed by R%d type=%s\n",
7859 func_id_name(meta->func_id),
7860 regno, reg_type_str(env, reg->type));
7866 switch ((int)reg->type) {
7868 case PTR_TO_BTF_ID | PTR_TRUSTED:
7869 case PTR_TO_BTF_ID | MEM_RCU:
7870 case PTR_TO_BTF_ID | PTR_MAYBE_NULL:
7871 case PTR_TO_BTF_ID | PTR_MAYBE_NULL | MEM_RCU:
7873 /* For bpf_sk_release, it needs to match against first member
7874 * 'struct sock_common', hence make an exception for it. This
7875 * allows bpf_sk_release to work for multiple socket types.
7877 bool strict_type_match = arg_type_is_release(arg_type) &&
7878 meta->func_id != BPF_FUNC_sk_release;
7880 if (type_may_be_null(reg->type) &&
7881 (!type_may_be_null(arg_type) || arg_type_is_release(arg_type))) {
7882 verbose(env, "Possibly NULL pointer passed to helper arg%d\n", regno);
7887 if (!compatible->btf_id) {
7888 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
7891 arg_btf_id = compatible->btf_id;
7894 if (meta->func_id == BPF_FUNC_kptr_xchg) {
7895 if (map_kptr_match_type(env, meta->kptr_field, reg, regno))
7898 if (arg_btf_id == BPF_PTR_POISON) {
7899 verbose(env, "verifier internal error:");
7900 verbose(env, "R%d has non-overwritten BPF_PTR_POISON type\n",
7905 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
7906 btf_vmlinux, *arg_btf_id,
7907 strict_type_match)) {
7908 verbose(env, "R%d is of type %s but %s is expected\n",
7909 regno, btf_type_name(reg->btf, reg->btf_id),
7910 btf_type_name(btf_vmlinux, *arg_btf_id));
7916 case PTR_TO_BTF_ID | MEM_ALLOC:
7917 if (meta->func_id != BPF_FUNC_spin_lock && meta->func_id != BPF_FUNC_spin_unlock &&
7918 meta->func_id != BPF_FUNC_kptr_xchg) {
7919 verbose(env, "verifier internal error: unimplemented handling of MEM_ALLOC\n");
7922 if (meta->func_id == BPF_FUNC_kptr_xchg) {
7923 if (map_kptr_match_type(env, meta->kptr_field, reg, regno))
7927 case PTR_TO_BTF_ID | MEM_PERCPU:
7928 case PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED:
7929 /* Handled by helper specific checks */
7932 verbose(env, "verifier internal error: invalid PTR_TO_BTF_ID register for type match\n");
7938 static struct btf_field *
7939 reg_find_field_offset(const struct bpf_reg_state *reg, s32 off, u32 fields)
7941 struct btf_field *field;
7942 struct btf_record *rec;
7944 rec = reg_btf_record(reg);
7948 field = btf_record_find(rec, off, fields);
7955 int check_func_arg_reg_off(struct bpf_verifier_env *env,
7956 const struct bpf_reg_state *reg, int regno,
7957 enum bpf_arg_type arg_type)
7959 u32 type = reg->type;
7961 /* When referenced register is passed to release function, its fixed
7964 * We will check arg_type_is_release reg has ref_obj_id when storing
7965 * meta->release_regno.
7967 if (arg_type_is_release(arg_type)) {
7968 /* ARG_PTR_TO_DYNPTR with OBJ_RELEASE is a bit special, as it
7969 * may not directly point to the object being released, but to
7970 * dynptr pointing to such object, which might be at some offset
7971 * on the stack. In that case, we simply to fallback to the
7974 if (arg_type_is_dynptr(arg_type) && type == PTR_TO_STACK)
7977 /* Doing check_ptr_off_reg check for the offset will catch this
7978 * because fixed_off_ok is false, but checking here allows us
7979 * to give the user a better error message.
7982 verbose(env, "R%d must have zero offset when passed to release func or trusted arg to kfunc\n",
7986 return __check_ptr_off_reg(env, reg, regno, false);
7990 /* Pointer types where both fixed and variable offset is explicitly allowed: */
7993 case PTR_TO_PACKET_META:
7994 case PTR_TO_MAP_KEY:
7995 case PTR_TO_MAP_VALUE:
7997 case PTR_TO_MEM | MEM_RDONLY:
7998 case PTR_TO_MEM | MEM_RINGBUF:
8000 case PTR_TO_BUF | MEM_RDONLY:
8003 /* All the rest must be rejected, except PTR_TO_BTF_ID which allows
8007 case PTR_TO_BTF_ID | MEM_ALLOC:
8008 case PTR_TO_BTF_ID | PTR_TRUSTED:
8009 case PTR_TO_BTF_ID | MEM_RCU:
8010 case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF:
8011 case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF | MEM_RCU:
8012 /* When referenced PTR_TO_BTF_ID is passed to release function,
8013 * its fixed offset must be 0. In the other cases, fixed offset
8014 * can be non-zero. This was already checked above. So pass
8015 * fixed_off_ok as true to allow fixed offset for all other
8016 * cases. var_off always must be 0 for PTR_TO_BTF_ID, hence we
8017 * still need to do checks instead of returning.
8019 return __check_ptr_off_reg(env, reg, regno, true);
8021 return __check_ptr_off_reg(env, reg, regno, false);
8025 static struct bpf_reg_state *get_dynptr_arg_reg(struct bpf_verifier_env *env,
8026 const struct bpf_func_proto *fn,
8027 struct bpf_reg_state *regs)
8029 struct bpf_reg_state *state = NULL;
8032 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++)
8033 if (arg_type_is_dynptr(fn->arg_type[i])) {
8035 verbose(env, "verifier internal error: multiple dynptr args\n");
8038 state = ®s[BPF_REG_1 + i];
8042 verbose(env, "verifier internal error: no dynptr arg found\n");
8047 static int dynptr_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
8049 struct bpf_func_state *state = func(env, reg);
8052 if (reg->type == CONST_PTR_TO_DYNPTR)
8054 spi = dynptr_get_spi(env, reg);
8057 return state->stack[spi].spilled_ptr.id;
8060 static int dynptr_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
8062 struct bpf_func_state *state = func(env, reg);
8065 if (reg->type == CONST_PTR_TO_DYNPTR)
8066 return reg->ref_obj_id;
8067 spi = dynptr_get_spi(env, reg);
8070 return state->stack[spi].spilled_ptr.ref_obj_id;
8073 static enum bpf_dynptr_type dynptr_get_type(struct bpf_verifier_env *env,
8074 struct bpf_reg_state *reg)
8076 struct bpf_func_state *state = func(env, reg);
8079 if (reg->type == CONST_PTR_TO_DYNPTR)
8080 return reg->dynptr.type;
8082 spi = __get_spi(reg->off);
8084 verbose(env, "verifier internal error: invalid spi when querying dynptr type\n");
8085 return BPF_DYNPTR_TYPE_INVALID;
8088 return state->stack[spi].spilled_ptr.dynptr.type;
8091 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
8092 struct bpf_call_arg_meta *meta,
8093 const struct bpf_func_proto *fn,
8096 u32 regno = BPF_REG_1 + arg;
8097 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
8098 enum bpf_arg_type arg_type = fn->arg_type[arg];
8099 enum bpf_reg_type type = reg->type;
8100 u32 *arg_btf_id = NULL;
8103 if (arg_type == ARG_DONTCARE)
8106 err = check_reg_arg(env, regno, SRC_OP);
8110 if (arg_type == ARG_ANYTHING) {
8111 if (is_pointer_value(env, regno)) {
8112 verbose(env, "R%d leaks addr into helper function\n",
8119 if (type_is_pkt_pointer(type) &&
8120 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
8121 verbose(env, "helper access to the packet is not allowed\n");
8125 if (base_type(arg_type) == ARG_PTR_TO_MAP_VALUE) {
8126 err = resolve_map_arg_type(env, meta, &arg_type);
8131 if (register_is_null(reg) && type_may_be_null(arg_type))
8132 /* A NULL register has a SCALAR_VALUE type, so skip
8135 goto skip_type_check;
8137 /* arg_btf_id and arg_size are in a union. */
8138 if (base_type(arg_type) == ARG_PTR_TO_BTF_ID ||
8139 base_type(arg_type) == ARG_PTR_TO_SPIN_LOCK)
8140 arg_btf_id = fn->arg_btf_id[arg];
8142 err = check_reg_type(env, regno, arg_type, arg_btf_id, meta);
8146 err = check_func_arg_reg_off(env, reg, regno, arg_type);
8151 if (arg_type_is_release(arg_type)) {
8152 if (arg_type_is_dynptr(arg_type)) {
8153 struct bpf_func_state *state = func(env, reg);
8156 /* Only dynptr created on stack can be released, thus
8157 * the get_spi and stack state checks for spilled_ptr
8158 * should only be done before process_dynptr_func for
8161 if (reg->type == PTR_TO_STACK) {
8162 spi = dynptr_get_spi(env, reg);
8163 if (spi < 0 || !state->stack[spi].spilled_ptr.ref_obj_id) {
8164 verbose(env, "arg %d is an unacquired reference\n", regno);
8168 verbose(env, "cannot release unowned const bpf_dynptr\n");
8171 } else if (!reg->ref_obj_id && !register_is_null(reg)) {
8172 verbose(env, "R%d must be referenced when passed to release function\n",
8176 if (meta->release_regno) {
8177 verbose(env, "verifier internal error: more than one release argument\n");
8180 meta->release_regno = regno;
8183 if (reg->ref_obj_id) {
8184 if (meta->ref_obj_id) {
8185 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
8186 regno, reg->ref_obj_id,
8190 meta->ref_obj_id = reg->ref_obj_id;
8193 switch (base_type(arg_type)) {
8194 case ARG_CONST_MAP_PTR:
8195 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
8196 if (meta->map_ptr) {
8197 /* Use map_uid (which is unique id of inner map) to reject:
8198 * inner_map1 = bpf_map_lookup_elem(outer_map, key1)
8199 * inner_map2 = bpf_map_lookup_elem(outer_map, key2)
8200 * if (inner_map1 && inner_map2) {
8201 * timer = bpf_map_lookup_elem(inner_map1);
8203 * // mismatch would have been allowed
8204 * bpf_timer_init(timer, inner_map2);
8207 * Comparing map_ptr is enough to distinguish normal and outer maps.
8209 if (meta->map_ptr != reg->map_ptr ||
8210 meta->map_uid != reg->map_uid) {
8212 "timer pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
8213 meta->map_uid, reg->map_uid);
8217 meta->map_ptr = reg->map_ptr;
8218 meta->map_uid = reg->map_uid;
8220 case ARG_PTR_TO_MAP_KEY:
8221 /* bpf_map_xxx(..., map_ptr, ..., key) call:
8222 * check that [key, key + map->key_size) are within
8223 * stack limits and initialized
8225 if (!meta->map_ptr) {
8226 /* in function declaration map_ptr must come before
8227 * map_key, so that it's verified and known before
8228 * we have to check map_key here. Otherwise it means
8229 * that kernel subsystem misconfigured verifier
8231 verbose(env, "invalid map_ptr to access map->key\n");
8234 err = check_helper_mem_access(env, regno,
8235 meta->map_ptr->key_size, false,
8238 case ARG_PTR_TO_MAP_VALUE:
8239 if (type_may_be_null(arg_type) && register_is_null(reg))
8242 /* bpf_map_xxx(..., map_ptr, ..., value) call:
8243 * check [value, value + map->value_size) validity
8245 if (!meta->map_ptr) {
8246 /* kernel subsystem misconfigured verifier */
8247 verbose(env, "invalid map_ptr to access map->value\n");
8250 meta->raw_mode = arg_type & MEM_UNINIT;
8251 err = check_helper_mem_access(env, regno,
8252 meta->map_ptr->value_size, false,
8255 case ARG_PTR_TO_PERCPU_BTF_ID:
8257 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
8260 meta->ret_btf = reg->btf;
8261 meta->ret_btf_id = reg->btf_id;
8263 case ARG_PTR_TO_SPIN_LOCK:
8264 if (in_rbtree_lock_required_cb(env)) {
8265 verbose(env, "can't spin_{lock,unlock} in rbtree cb\n");
8268 if (meta->func_id == BPF_FUNC_spin_lock) {
8269 err = process_spin_lock(env, regno, true);
8272 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
8273 err = process_spin_lock(env, regno, false);
8277 verbose(env, "verifier internal error\n");
8281 case ARG_PTR_TO_TIMER:
8282 err = process_timer_func(env, regno, meta);
8286 case ARG_PTR_TO_FUNC:
8287 meta->subprogno = reg->subprogno;
8289 case ARG_PTR_TO_MEM:
8290 /* The access to this pointer is only checked when we hit the
8291 * next is_mem_size argument below.
8293 meta->raw_mode = arg_type & MEM_UNINIT;
8294 if (arg_type & MEM_FIXED_SIZE) {
8295 err = check_helper_mem_access(env, regno,
8296 fn->arg_size[arg], false,
8300 case ARG_CONST_SIZE:
8301 err = check_mem_size_reg(env, reg, regno, false, meta);
8303 case ARG_CONST_SIZE_OR_ZERO:
8304 err = check_mem_size_reg(env, reg, regno, true, meta);
8306 case ARG_PTR_TO_DYNPTR:
8307 err = process_dynptr_func(env, regno, insn_idx, arg_type, 0);
8311 case ARG_CONST_ALLOC_SIZE_OR_ZERO:
8312 if (!tnum_is_const(reg->var_off)) {
8313 verbose(env, "R%d is not a known constant'\n",
8317 meta->mem_size = reg->var_off.value;
8318 err = mark_chain_precision(env, regno);
8322 case ARG_PTR_TO_INT:
8323 case ARG_PTR_TO_LONG:
8325 int size = int_ptr_type_to_size(arg_type);
8327 err = check_helper_mem_access(env, regno, size, false, meta);
8330 err = check_ptr_alignment(env, reg, 0, size, true);
8333 case ARG_PTR_TO_CONST_STR:
8335 struct bpf_map *map = reg->map_ptr;
8340 if (!bpf_map_is_rdonly(map)) {
8341 verbose(env, "R%d does not point to a readonly map'\n", regno);
8345 if (!tnum_is_const(reg->var_off)) {
8346 verbose(env, "R%d is not a constant address'\n", regno);
8350 if (!map->ops->map_direct_value_addr) {
8351 verbose(env, "no direct value access support for this map type\n");
8355 err = check_map_access(env, regno, reg->off,
8356 map->value_size - reg->off, false,
8361 map_off = reg->off + reg->var_off.value;
8362 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
8364 verbose(env, "direct value access on string failed\n");
8368 str_ptr = (char *)(long)(map_addr);
8369 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
8370 verbose(env, "string is not zero-terminated\n");
8375 case ARG_PTR_TO_KPTR:
8376 err = process_kptr_func(env, regno, meta);
8385 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
8387 enum bpf_attach_type eatype = env->prog->expected_attach_type;
8388 enum bpf_prog_type type = resolve_prog_type(env->prog);
8390 if (func_id != BPF_FUNC_map_update_elem)
8393 /* It's not possible to get access to a locked struct sock in these
8394 * contexts, so updating is safe.
8397 case BPF_PROG_TYPE_TRACING:
8398 if (eatype == BPF_TRACE_ITER)
8401 case BPF_PROG_TYPE_SOCKET_FILTER:
8402 case BPF_PROG_TYPE_SCHED_CLS:
8403 case BPF_PROG_TYPE_SCHED_ACT:
8404 case BPF_PROG_TYPE_XDP:
8405 case BPF_PROG_TYPE_SK_REUSEPORT:
8406 case BPF_PROG_TYPE_FLOW_DISSECTOR:
8407 case BPF_PROG_TYPE_SK_LOOKUP:
8413 verbose(env, "cannot update sockmap in this context\n");
8417 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
8419 return env->prog->jit_requested &&
8420 bpf_jit_supports_subprog_tailcalls();
8423 static int check_map_func_compatibility(struct bpf_verifier_env *env,
8424 struct bpf_map *map, int func_id)
8429 /* We need a two way check, first is from map perspective ... */
8430 switch (map->map_type) {
8431 case BPF_MAP_TYPE_PROG_ARRAY:
8432 if (func_id != BPF_FUNC_tail_call)
8435 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
8436 if (func_id != BPF_FUNC_perf_event_read &&
8437 func_id != BPF_FUNC_perf_event_output &&
8438 func_id != BPF_FUNC_skb_output &&
8439 func_id != BPF_FUNC_perf_event_read_value &&
8440 func_id != BPF_FUNC_xdp_output)
8443 case BPF_MAP_TYPE_RINGBUF:
8444 if (func_id != BPF_FUNC_ringbuf_output &&
8445 func_id != BPF_FUNC_ringbuf_reserve &&
8446 func_id != BPF_FUNC_ringbuf_query &&
8447 func_id != BPF_FUNC_ringbuf_reserve_dynptr &&
8448 func_id != BPF_FUNC_ringbuf_submit_dynptr &&
8449 func_id != BPF_FUNC_ringbuf_discard_dynptr)
8452 case BPF_MAP_TYPE_USER_RINGBUF:
8453 if (func_id != BPF_FUNC_user_ringbuf_drain)
8456 case BPF_MAP_TYPE_STACK_TRACE:
8457 if (func_id != BPF_FUNC_get_stackid)
8460 case BPF_MAP_TYPE_CGROUP_ARRAY:
8461 if (func_id != BPF_FUNC_skb_under_cgroup &&
8462 func_id != BPF_FUNC_current_task_under_cgroup)
8465 case BPF_MAP_TYPE_CGROUP_STORAGE:
8466 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
8467 if (func_id != BPF_FUNC_get_local_storage)
8470 case BPF_MAP_TYPE_DEVMAP:
8471 case BPF_MAP_TYPE_DEVMAP_HASH:
8472 if (func_id != BPF_FUNC_redirect_map &&
8473 func_id != BPF_FUNC_map_lookup_elem)
8476 /* Restrict bpf side of cpumap and xskmap, open when use-cases
8479 case BPF_MAP_TYPE_CPUMAP:
8480 if (func_id != BPF_FUNC_redirect_map)
8483 case BPF_MAP_TYPE_XSKMAP:
8484 if (func_id != BPF_FUNC_redirect_map &&
8485 func_id != BPF_FUNC_map_lookup_elem)
8488 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
8489 case BPF_MAP_TYPE_HASH_OF_MAPS:
8490 if (func_id != BPF_FUNC_map_lookup_elem)
8493 case BPF_MAP_TYPE_SOCKMAP:
8494 if (func_id != BPF_FUNC_sk_redirect_map &&
8495 func_id != BPF_FUNC_sock_map_update &&
8496 func_id != BPF_FUNC_map_delete_elem &&
8497 func_id != BPF_FUNC_msg_redirect_map &&
8498 func_id != BPF_FUNC_sk_select_reuseport &&
8499 func_id != BPF_FUNC_map_lookup_elem &&
8500 !may_update_sockmap(env, func_id))
8503 case BPF_MAP_TYPE_SOCKHASH:
8504 if (func_id != BPF_FUNC_sk_redirect_hash &&
8505 func_id != BPF_FUNC_sock_hash_update &&
8506 func_id != BPF_FUNC_map_delete_elem &&
8507 func_id != BPF_FUNC_msg_redirect_hash &&
8508 func_id != BPF_FUNC_sk_select_reuseport &&
8509 func_id != BPF_FUNC_map_lookup_elem &&
8510 !may_update_sockmap(env, func_id))
8513 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
8514 if (func_id != BPF_FUNC_sk_select_reuseport)
8517 case BPF_MAP_TYPE_QUEUE:
8518 case BPF_MAP_TYPE_STACK:
8519 if (func_id != BPF_FUNC_map_peek_elem &&
8520 func_id != BPF_FUNC_map_pop_elem &&
8521 func_id != BPF_FUNC_map_push_elem)
8524 case BPF_MAP_TYPE_SK_STORAGE:
8525 if (func_id != BPF_FUNC_sk_storage_get &&
8526 func_id != BPF_FUNC_sk_storage_delete &&
8527 func_id != BPF_FUNC_kptr_xchg)
8530 case BPF_MAP_TYPE_INODE_STORAGE:
8531 if (func_id != BPF_FUNC_inode_storage_get &&
8532 func_id != BPF_FUNC_inode_storage_delete &&
8533 func_id != BPF_FUNC_kptr_xchg)
8536 case BPF_MAP_TYPE_TASK_STORAGE:
8537 if (func_id != BPF_FUNC_task_storage_get &&
8538 func_id != BPF_FUNC_task_storage_delete &&
8539 func_id != BPF_FUNC_kptr_xchg)
8542 case BPF_MAP_TYPE_CGRP_STORAGE:
8543 if (func_id != BPF_FUNC_cgrp_storage_get &&
8544 func_id != BPF_FUNC_cgrp_storage_delete &&
8545 func_id != BPF_FUNC_kptr_xchg)
8548 case BPF_MAP_TYPE_BLOOM_FILTER:
8549 if (func_id != BPF_FUNC_map_peek_elem &&
8550 func_id != BPF_FUNC_map_push_elem)
8557 /* ... and second from the function itself. */
8559 case BPF_FUNC_tail_call:
8560 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
8562 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
8563 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
8567 case BPF_FUNC_perf_event_read:
8568 case BPF_FUNC_perf_event_output:
8569 case BPF_FUNC_perf_event_read_value:
8570 case BPF_FUNC_skb_output:
8571 case BPF_FUNC_xdp_output:
8572 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
8575 case BPF_FUNC_ringbuf_output:
8576 case BPF_FUNC_ringbuf_reserve:
8577 case BPF_FUNC_ringbuf_query:
8578 case BPF_FUNC_ringbuf_reserve_dynptr:
8579 case BPF_FUNC_ringbuf_submit_dynptr:
8580 case BPF_FUNC_ringbuf_discard_dynptr:
8581 if (map->map_type != BPF_MAP_TYPE_RINGBUF)
8584 case BPF_FUNC_user_ringbuf_drain:
8585 if (map->map_type != BPF_MAP_TYPE_USER_RINGBUF)
8588 case BPF_FUNC_get_stackid:
8589 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
8592 case BPF_FUNC_current_task_under_cgroup:
8593 case BPF_FUNC_skb_under_cgroup:
8594 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
8597 case BPF_FUNC_redirect_map:
8598 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
8599 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
8600 map->map_type != BPF_MAP_TYPE_CPUMAP &&
8601 map->map_type != BPF_MAP_TYPE_XSKMAP)
8604 case BPF_FUNC_sk_redirect_map:
8605 case BPF_FUNC_msg_redirect_map:
8606 case BPF_FUNC_sock_map_update:
8607 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
8610 case BPF_FUNC_sk_redirect_hash:
8611 case BPF_FUNC_msg_redirect_hash:
8612 case BPF_FUNC_sock_hash_update:
8613 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
8616 case BPF_FUNC_get_local_storage:
8617 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
8618 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
8621 case BPF_FUNC_sk_select_reuseport:
8622 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
8623 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
8624 map->map_type != BPF_MAP_TYPE_SOCKHASH)
8627 case BPF_FUNC_map_pop_elem:
8628 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
8629 map->map_type != BPF_MAP_TYPE_STACK)
8632 case BPF_FUNC_map_peek_elem:
8633 case BPF_FUNC_map_push_elem:
8634 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
8635 map->map_type != BPF_MAP_TYPE_STACK &&
8636 map->map_type != BPF_MAP_TYPE_BLOOM_FILTER)
8639 case BPF_FUNC_map_lookup_percpu_elem:
8640 if (map->map_type != BPF_MAP_TYPE_PERCPU_ARRAY &&
8641 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
8642 map->map_type != BPF_MAP_TYPE_LRU_PERCPU_HASH)
8645 case BPF_FUNC_sk_storage_get:
8646 case BPF_FUNC_sk_storage_delete:
8647 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
8650 case BPF_FUNC_inode_storage_get:
8651 case BPF_FUNC_inode_storage_delete:
8652 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
8655 case BPF_FUNC_task_storage_get:
8656 case BPF_FUNC_task_storage_delete:
8657 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
8660 case BPF_FUNC_cgrp_storage_get:
8661 case BPF_FUNC_cgrp_storage_delete:
8662 if (map->map_type != BPF_MAP_TYPE_CGRP_STORAGE)
8671 verbose(env, "cannot pass map_type %d into func %s#%d\n",
8672 map->map_type, func_id_name(func_id), func_id);
8676 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
8680 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
8682 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
8684 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
8686 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
8688 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
8691 /* We only support one arg being in raw mode at the moment,
8692 * which is sufficient for the helper functions we have
8698 static bool check_args_pair_invalid(const struct bpf_func_proto *fn, int arg)
8700 bool is_fixed = fn->arg_type[arg] & MEM_FIXED_SIZE;
8701 bool has_size = fn->arg_size[arg] != 0;
8702 bool is_next_size = false;
8704 if (arg + 1 < ARRAY_SIZE(fn->arg_type))
8705 is_next_size = arg_type_is_mem_size(fn->arg_type[arg + 1]);
8707 if (base_type(fn->arg_type[arg]) != ARG_PTR_TO_MEM)
8708 return is_next_size;
8710 return has_size == is_next_size || is_next_size == is_fixed;
8713 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
8715 /* bpf_xxx(..., buf, len) call will access 'len'
8716 * bytes from memory 'buf'. Both arg types need
8717 * to be paired, so make sure there's no buggy
8718 * helper function specification.
8720 if (arg_type_is_mem_size(fn->arg1_type) ||
8721 check_args_pair_invalid(fn, 0) ||
8722 check_args_pair_invalid(fn, 1) ||
8723 check_args_pair_invalid(fn, 2) ||
8724 check_args_pair_invalid(fn, 3) ||
8725 check_args_pair_invalid(fn, 4))
8731 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
8735 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
8736 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_BTF_ID)
8737 return !!fn->arg_btf_id[i];
8738 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_SPIN_LOCK)
8739 return fn->arg_btf_id[i] == BPF_PTR_POISON;
8740 if (base_type(fn->arg_type[i]) != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i] &&
8741 /* arg_btf_id and arg_size are in a union. */
8742 (base_type(fn->arg_type[i]) != ARG_PTR_TO_MEM ||
8743 !(fn->arg_type[i] & MEM_FIXED_SIZE)))
8750 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
8752 return check_raw_mode_ok(fn) &&
8753 check_arg_pair_ok(fn) &&
8754 check_btf_id_ok(fn) ? 0 : -EINVAL;
8757 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
8758 * are now invalid, so turn them into unknown SCALAR_VALUE.
8760 * This also applies to dynptr slices belonging to skb and xdp dynptrs,
8761 * since these slices point to packet data.
8763 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
8765 struct bpf_func_state *state;
8766 struct bpf_reg_state *reg;
8768 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
8769 if (reg_is_pkt_pointer_any(reg) || reg_is_dynptr_slice_pkt(reg))
8770 mark_reg_invalid(env, reg);
8776 BEYOND_PKT_END = -2,
8779 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
8781 struct bpf_func_state *state = vstate->frame[vstate->curframe];
8782 struct bpf_reg_state *reg = &state->regs[regn];
8784 if (reg->type != PTR_TO_PACKET)
8785 /* PTR_TO_PACKET_META is not supported yet */
8788 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
8789 * How far beyond pkt_end it goes is unknown.
8790 * if (!range_open) it's the case of pkt >= pkt_end
8791 * if (range_open) it's the case of pkt > pkt_end
8792 * hence this pointer is at least 1 byte bigger than pkt_end
8795 reg->range = BEYOND_PKT_END;
8797 reg->range = AT_PKT_END;
8800 /* The pointer with the specified id has released its reference to kernel
8801 * resources. Identify all copies of the same pointer and clear the reference.
8803 static int release_reference(struct bpf_verifier_env *env,
8806 struct bpf_func_state *state;
8807 struct bpf_reg_state *reg;
8810 err = release_reference_state(cur_func(env), ref_obj_id);
8814 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
8815 if (reg->ref_obj_id == ref_obj_id)
8816 mark_reg_invalid(env, reg);
8822 static void invalidate_non_owning_refs(struct bpf_verifier_env *env)
8824 struct bpf_func_state *unused;
8825 struct bpf_reg_state *reg;
8827 bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
8828 if (type_is_non_owning_ref(reg->type))
8829 mark_reg_invalid(env, reg);
8833 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
8834 struct bpf_reg_state *regs)
8838 /* after the call registers r0 - r5 were scratched */
8839 for (i = 0; i < CALLER_SAVED_REGS; i++) {
8840 mark_reg_not_init(env, regs, caller_saved[i]);
8841 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
8845 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
8846 struct bpf_func_state *caller,
8847 struct bpf_func_state *callee,
8850 static int set_callee_state(struct bpf_verifier_env *env,
8851 struct bpf_func_state *caller,
8852 struct bpf_func_state *callee, int insn_idx);
8854 static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
8855 int *insn_idx, int subprog,
8856 set_callee_state_fn set_callee_state_cb)
8858 struct bpf_verifier_state *state = env->cur_state;
8859 struct bpf_func_state *caller, *callee;
8862 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
8863 verbose(env, "the call stack of %d frames is too deep\n",
8864 state->curframe + 2);
8868 caller = state->frame[state->curframe];
8869 if (state->frame[state->curframe + 1]) {
8870 verbose(env, "verifier bug. Frame %d already allocated\n",
8871 state->curframe + 1);
8875 err = btf_check_subprog_call(env, subprog, caller->regs);
8878 if (subprog_is_global(env, subprog)) {
8880 verbose(env, "Caller passes invalid args into func#%d\n",
8884 if (env->log.level & BPF_LOG_LEVEL)
8886 "Func#%d is global and valid. Skipping.\n",
8888 clear_caller_saved_regs(env, caller->regs);
8890 /* All global functions return a 64-bit SCALAR_VALUE */
8891 mark_reg_unknown(env, caller->regs, BPF_REG_0);
8892 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
8894 /* continue with next insn after call */
8899 /* set_callee_state is used for direct subprog calls, but we are
8900 * interested in validating only BPF helpers that can call subprogs as
8903 if (set_callee_state_cb != set_callee_state) {
8904 if (bpf_pseudo_kfunc_call(insn) &&
8905 !is_callback_calling_kfunc(insn->imm)) {
8906 verbose(env, "verifier bug: kfunc %s#%d not marked as callback-calling\n",
8907 func_id_name(insn->imm), insn->imm);
8909 } else if (!bpf_pseudo_kfunc_call(insn) &&
8910 !is_callback_calling_function(insn->imm)) { /* helper */
8911 verbose(env, "verifier bug: helper %s#%d not marked as callback-calling\n",
8912 func_id_name(insn->imm), insn->imm);
8917 if (insn->code == (BPF_JMP | BPF_CALL) &&
8918 insn->src_reg == 0 &&
8919 insn->imm == BPF_FUNC_timer_set_callback) {
8920 struct bpf_verifier_state *async_cb;
8922 /* there is no real recursion here. timer callbacks are async */
8923 env->subprog_info[subprog].is_async_cb = true;
8924 async_cb = push_async_cb(env, env->subprog_info[subprog].start,
8925 *insn_idx, subprog);
8928 callee = async_cb->frame[0];
8929 callee->async_entry_cnt = caller->async_entry_cnt + 1;
8931 /* Convert bpf_timer_set_callback() args into timer callback args */
8932 err = set_callee_state_cb(env, caller, callee, *insn_idx);
8936 clear_caller_saved_regs(env, caller->regs);
8937 mark_reg_unknown(env, caller->regs, BPF_REG_0);
8938 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
8939 /* continue with next insn after call */
8943 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
8946 state->frame[state->curframe + 1] = callee;
8948 /* callee cannot access r0, r6 - r9 for reading and has to write
8949 * into its own stack before reading from it.
8950 * callee can read/write into caller's stack
8952 init_func_state(env, callee,
8953 /* remember the callsite, it will be used by bpf_exit */
8954 *insn_idx /* callsite */,
8955 state->curframe + 1 /* frameno within this callchain */,
8956 subprog /* subprog number within this prog */);
8958 /* Transfer references to the callee */
8959 err = copy_reference_state(callee, caller);
8963 err = set_callee_state_cb(env, caller, callee, *insn_idx);
8967 clear_caller_saved_regs(env, caller->regs);
8969 /* only increment it after check_reg_arg() finished */
8972 /* and go analyze first insn of the callee */
8973 *insn_idx = env->subprog_info[subprog].start - 1;
8975 if (env->log.level & BPF_LOG_LEVEL) {
8976 verbose(env, "caller:\n");
8977 print_verifier_state(env, caller, true);
8978 verbose(env, "callee:\n");
8979 print_verifier_state(env, callee, true);
8984 free_func_state(callee);
8985 state->frame[state->curframe + 1] = NULL;
8989 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
8990 struct bpf_func_state *caller,
8991 struct bpf_func_state *callee)
8993 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
8994 * void *callback_ctx, u64 flags);
8995 * callback_fn(struct bpf_map *map, void *key, void *value,
8996 * void *callback_ctx);
8998 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
9000 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
9001 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
9002 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
9004 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
9005 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
9006 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
9008 /* pointer to stack or null */
9009 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
9012 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9016 static int set_callee_state(struct bpf_verifier_env *env,
9017 struct bpf_func_state *caller,
9018 struct bpf_func_state *callee, int insn_idx)
9022 /* copy r1 - r5 args that callee can access. The copy includes parent
9023 * pointers, which connects us up to the liveness chain
9025 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
9026 callee->regs[i] = caller->regs[i];
9030 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
9033 int subprog, target_insn;
9035 target_insn = *insn_idx + insn->imm + 1;
9036 subprog = find_subprog(env, target_insn);
9038 verbose(env, "verifier bug. No program starts at insn %d\n",
9043 return __check_func_call(env, insn, insn_idx, subprog, set_callee_state);
9046 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
9047 struct bpf_func_state *caller,
9048 struct bpf_func_state *callee,
9051 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
9052 struct bpf_map *map;
9055 if (bpf_map_ptr_poisoned(insn_aux)) {
9056 verbose(env, "tail_call abusing map_ptr\n");
9060 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
9061 if (!map->ops->map_set_for_each_callback_args ||
9062 !map->ops->map_for_each_callback) {
9063 verbose(env, "callback function not allowed for map\n");
9067 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
9071 callee->in_callback_fn = true;
9072 callee->callback_ret_range = tnum_range(0, 1);
9076 static int set_loop_callback_state(struct bpf_verifier_env *env,
9077 struct bpf_func_state *caller,
9078 struct bpf_func_state *callee,
9081 /* bpf_loop(u32 nr_loops, void *callback_fn, void *callback_ctx,
9083 * callback_fn(u32 index, void *callback_ctx);
9085 callee->regs[BPF_REG_1].type = SCALAR_VALUE;
9086 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
9089 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
9090 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9091 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9093 callee->in_callback_fn = true;
9094 callee->callback_ret_range = tnum_range(0, 1);
9098 static int set_timer_callback_state(struct bpf_verifier_env *env,
9099 struct bpf_func_state *caller,
9100 struct bpf_func_state *callee,
9103 struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
9105 /* bpf_timer_set_callback(struct bpf_timer *timer, void *callback_fn);
9106 * callback_fn(struct bpf_map *map, void *key, void *value);
9108 callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP;
9109 __mark_reg_known_zero(&callee->regs[BPF_REG_1]);
9110 callee->regs[BPF_REG_1].map_ptr = map_ptr;
9112 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
9113 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
9114 callee->regs[BPF_REG_2].map_ptr = map_ptr;
9116 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
9117 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
9118 callee->regs[BPF_REG_3].map_ptr = map_ptr;
9121 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9122 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9123 callee->in_async_callback_fn = true;
9124 callee->callback_ret_range = tnum_range(0, 1);
9128 static int set_find_vma_callback_state(struct bpf_verifier_env *env,
9129 struct bpf_func_state *caller,
9130 struct bpf_func_state *callee,
9133 /* bpf_find_vma(struct task_struct *task, u64 addr,
9134 * void *callback_fn, void *callback_ctx, u64 flags)
9135 * (callback_fn)(struct task_struct *task,
9136 * struct vm_area_struct *vma, void *callback_ctx);
9138 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
9140 callee->regs[BPF_REG_2].type = PTR_TO_BTF_ID;
9141 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
9142 callee->regs[BPF_REG_2].btf = btf_vmlinux;
9143 callee->regs[BPF_REG_2].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_VMA],
9145 /* pointer to stack or null */
9146 callee->regs[BPF_REG_3] = caller->regs[BPF_REG_4];
9149 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9150 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9151 callee->in_callback_fn = true;
9152 callee->callback_ret_range = tnum_range(0, 1);
9156 static int set_user_ringbuf_callback_state(struct bpf_verifier_env *env,
9157 struct bpf_func_state *caller,
9158 struct bpf_func_state *callee,
9161 /* bpf_user_ringbuf_drain(struct bpf_map *map, void *callback_fn, void
9162 * callback_ctx, u64 flags);
9163 * callback_fn(const struct bpf_dynptr_t* dynptr, void *callback_ctx);
9165 __mark_reg_not_init(env, &callee->regs[BPF_REG_0]);
9166 mark_dynptr_cb_reg(env, &callee->regs[BPF_REG_1], BPF_DYNPTR_TYPE_LOCAL);
9167 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
9170 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
9171 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9172 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9174 callee->in_callback_fn = true;
9175 callee->callback_ret_range = tnum_range(0, 1);
9179 static int set_rbtree_add_callback_state(struct bpf_verifier_env *env,
9180 struct bpf_func_state *caller,
9181 struct bpf_func_state *callee,
9184 /* void bpf_rbtree_add_impl(struct bpf_rb_root *root, struct bpf_rb_node *node,
9185 * bool (less)(struct bpf_rb_node *a, const struct bpf_rb_node *b));
9187 * 'struct bpf_rb_node *node' arg to bpf_rbtree_add_impl is the same PTR_TO_BTF_ID w/ offset
9188 * that 'less' callback args will be receiving. However, 'node' arg was release_reference'd
9189 * by this point, so look at 'root'
9191 struct btf_field *field;
9193 field = reg_find_field_offset(&caller->regs[BPF_REG_1], caller->regs[BPF_REG_1].off,
9195 if (!field || !field->graph_root.value_btf_id)
9198 mark_reg_graph_node(callee->regs, BPF_REG_1, &field->graph_root);
9199 ref_set_non_owning(env, &callee->regs[BPF_REG_1]);
9200 mark_reg_graph_node(callee->regs, BPF_REG_2, &field->graph_root);
9201 ref_set_non_owning(env, &callee->regs[BPF_REG_2]);
9203 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
9204 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9205 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9206 callee->in_callback_fn = true;
9207 callee->callback_ret_range = tnum_range(0, 1);
9211 static bool is_rbtree_lock_required_kfunc(u32 btf_id);
9213 /* Are we currently verifying the callback for a rbtree helper that must
9214 * be called with lock held? If so, no need to complain about unreleased
9217 static bool in_rbtree_lock_required_cb(struct bpf_verifier_env *env)
9219 struct bpf_verifier_state *state = env->cur_state;
9220 struct bpf_insn *insn = env->prog->insnsi;
9221 struct bpf_func_state *callee;
9224 if (!state->curframe)
9227 callee = state->frame[state->curframe];
9229 if (!callee->in_callback_fn)
9232 kfunc_btf_id = insn[callee->callsite].imm;
9233 return is_rbtree_lock_required_kfunc(kfunc_btf_id);
9236 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
9238 struct bpf_verifier_state *state = env->cur_state;
9239 struct bpf_func_state *caller, *callee;
9240 struct bpf_reg_state *r0;
9243 callee = state->frame[state->curframe];
9244 r0 = &callee->regs[BPF_REG_0];
9245 if (r0->type == PTR_TO_STACK) {
9246 /* technically it's ok to return caller's stack pointer
9247 * (or caller's caller's pointer) back to the caller,
9248 * since these pointers are valid. Only current stack
9249 * pointer will be invalid as soon as function exits,
9250 * but let's be conservative
9252 verbose(env, "cannot return stack pointer to the caller\n");
9256 caller = state->frame[state->curframe - 1];
9257 if (callee->in_callback_fn) {
9258 /* enforce R0 return value range [0, 1]. */
9259 struct tnum range = callee->callback_ret_range;
9261 if (r0->type != SCALAR_VALUE) {
9262 verbose(env, "R0 not a scalar value\n");
9265 if (!tnum_in(range, r0->var_off)) {
9266 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
9270 /* return to the caller whatever r0 had in the callee */
9271 caller->regs[BPF_REG_0] = *r0;
9274 /* callback_fn frame should have released its own additions to parent's
9275 * reference state at this point, or check_reference_leak would
9276 * complain, hence it must be the same as the caller. There is no need
9279 if (!callee->in_callback_fn) {
9280 /* Transfer references to the caller */
9281 err = copy_reference_state(caller, callee);
9286 *insn_idx = callee->callsite + 1;
9287 if (env->log.level & BPF_LOG_LEVEL) {
9288 verbose(env, "returning from callee:\n");
9289 print_verifier_state(env, callee, true);
9290 verbose(env, "to caller at %d:\n", *insn_idx);
9291 print_verifier_state(env, caller, true);
9293 /* clear everything in the callee */
9294 free_func_state(callee);
9295 state->frame[state->curframe--] = NULL;
9299 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
9301 struct bpf_call_arg_meta *meta)
9303 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
9305 if (ret_type != RET_INTEGER)
9309 case BPF_FUNC_get_stack:
9310 case BPF_FUNC_get_task_stack:
9311 case BPF_FUNC_probe_read_str:
9312 case BPF_FUNC_probe_read_kernel_str:
9313 case BPF_FUNC_probe_read_user_str:
9314 ret_reg->smax_value = meta->msize_max_value;
9315 ret_reg->s32_max_value = meta->msize_max_value;
9316 ret_reg->smin_value = -MAX_ERRNO;
9317 ret_reg->s32_min_value = -MAX_ERRNO;
9318 reg_bounds_sync(ret_reg);
9320 case BPF_FUNC_get_smp_processor_id:
9321 ret_reg->umax_value = nr_cpu_ids - 1;
9322 ret_reg->u32_max_value = nr_cpu_ids - 1;
9323 ret_reg->smax_value = nr_cpu_ids - 1;
9324 ret_reg->s32_max_value = nr_cpu_ids - 1;
9325 ret_reg->umin_value = 0;
9326 ret_reg->u32_min_value = 0;
9327 ret_reg->smin_value = 0;
9328 ret_reg->s32_min_value = 0;
9329 reg_bounds_sync(ret_reg);
9335 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
9336 int func_id, int insn_idx)
9338 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
9339 struct bpf_map *map = meta->map_ptr;
9341 if (func_id != BPF_FUNC_tail_call &&
9342 func_id != BPF_FUNC_map_lookup_elem &&
9343 func_id != BPF_FUNC_map_update_elem &&
9344 func_id != BPF_FUNC_map_delete_elem &&
9345 func_id != BPF_FUNC_map_push_elem &&
9346 func_id != BPF_FUNC_map_pop_elem &&
9347 func_id != BPF_FUNC_map_peek_elem &&
9348 func_id != BPF_FUNC_for_each_map_elem &&
9349 func_id != BPF_FUNC_redirect_map &&
9350 func_id != BPF_FUNC_map_lookup_percpu_elem)
9354 verbose(env, "kernel subsystem misconfigured verifier\n");
9358 /* In case of read-only, some additional restrictions
9359 * need to be applied in order to prevent altering the
9360 * state of the map from program side.
9362 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
9363 (func_id == BPF_FUNC_map_delete_elem ||
9364 func_id == BPF_FUNC_map_update_elem ||
9365 func_id == BPF_FUNC_map_push_elem ||
9366 func_id == BPF_FUNC_map_pop_elem)) {
9367 verbose(env, "write into map forbidden\n");
9371 if (!BPF_MAP_PTR(aux->map_ptr_state))
9372 bpf_map_ptr_store(aux, meta->map_ptr,
9373 !meta->map_ptr->bypass_spec_v1);
9374 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
9375 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
9376 !meta->map_ptr->bypass_spec_v1);
9381 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
9382 int func_id, int insn_idx)
9384 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
9385 struct bpf_reg_state *regs = cur_regs(env), *reg;
9386 struct bpf_map *map = meta->map_ptr;
9390 if (func_id != BPF_FUNC_tail_call)
9392 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
9393 verbose(env, "kernel subsystem misconfigured verifier\n");
9397 reg = ®s[BPF_REG_3];
9398 val = reg->var_off.value;
9399 max = map->max_entries;
9401 if (!(register_is_const(reg) && val < max)) {
9402 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
9406 err = mark_chain_precision(env, BPF_REG_3);
9409 if (bpf_map_key_unseen(aux))
9410 bpf_map_key_store(aux, val);
9411 else if (!bpf_map_key_poisoned(aux) &&
9412 bpf_map_key_immediate(aux) != val)
9413 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
9417 static int check_reference_leak(struct bpf_verifier_env *env)
9419 struct bpf_func_state *state = cur_func(env);
9420 bool refs_lingering = false;
9423 if (state->frameno && !state->in_callback_fn)
9426 for (i = 0; i < state->acquired_refs; i++) {
9427 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno)
9429 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
9430 state->refs[i].id, state->refs[i].insn_idx);
9431 refs_lingering = true;
9433 return refs_lingering ? -EINVAL : 0;
9436 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
9437 struct bpf_reg_state *regs)
9439 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
9440 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
9441 struct bpf_map *fmt_map = fmt_reg->map_ptr;
9442 struct bpf_bprintf_data data = {};
9443 int err, fmt_map_off, num_args;
9447 /* data must be an array of u64 */
9448 if (data_len_reg->var_off.value % 8)
9450 num_args = data_len_reg->var_off.value / 8;
9452 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
9453 * and map_direct_value_addr is set.
9455 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
9456 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
9459 verbose(env, "verifier bug\n");
9462 fmt = (char *)(long)fmt_addr + fmt_map_off;
9464 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
9465 * can focus on validating the format specifiers.
9467 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, num_args, &data);
9469 verbose(env, "Invalid format string\n");
9474 static int check_get_func_ip(struct bpf_verifier_env *env)
9476 enum bpf_prog_type type = resolve_prog_type(env->prog);
9477 int func_id = BPF_FUNC_get_func_ip;
9479 if (type == BPF_PROG_TYPE_TRACING) {
9480 if (!bpf_prog_has_trampoline(env->prog)) {
9481 verbose(env, "func %s#%d supported only for fentry/fexit/fmod_ret programs\n",
9482 func_id_name(func_id), func_id);
9486 } else if (type == BPF_PROG_TYPE_KPROBE) {
9490 verbose(env, "func %s#%d not supported for program type %d\n",
9491 func_id_name(func_id), func_id, type);
9495 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
9497 return &env->insn_aux_data[env->insn_idx];
9500 static bool loop_flag_is_zero(struct bpf_verifier_env *env)
9502 struct bpf_reg_state *regs = cur_regs(env);
9503 struct bpf_reg_state *reg = ®s[BPF_REG_4];
9504 bool reg_is_null = register_is_null(reg);
9507 mark_chain_precision(env, BPF_REG_4);
9512 static void update_loop_inline_state(struct bpf_verifier_env *env, u32 subprogno)
9514 struct bpf_loop_inline_state *state = &cur_aux(env)->loop_inline_state;
9516 if (!state->initialized) {
9517 state->initialized = 1;
9518 state->fit_for_inline = loop_flag_is_zero(env);
9519 state->callback_subprogno = subprogno;
9523 if (!state->fit_for_inline)
9526 state->fit_for_inline = (loop_flag_is_zero(env) &&
9527 state->callback_subprogno == subprogno);
9530 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
9533 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
9534 const struct bpf_func_proto *fn = NULL;
9535 enum bpf_return_type ret_type;
9536 enum bpf_type_flag ret_flag;
9537 struct bpf_reg_state *regs;
9538 struct bpf_call_arg_meta meta;
9539 int insn_idx = *insn_idx_p;
9541 int i, err, func_id;
9543 /* find function prototype */
9544 func_id = insn->imm;
9545 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
9546 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
9551 if (env->ops->get_func_proto)
9552 fn = env->ops->get_func_proto(func_id, env->prog);
9554 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
9559 /* eBPF programs must be GPL compatible to use GPL-ed functions */
9560 if (!env->prog->gpl_compatible && fn->gpl_only) {
9561 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
9565 if (fn->allowed && !fn->allowed(env->prog)) {
9566 verbose(env, "helper call is not allowed in probe\n");
9570 if (!env->prog->aux->sleepable && fn->might_sleep) {
9571 verbose(env, "helper call might sleep in a non-sleepable prog\n");
9575 /* With LD_ABS/IND some JITs save/restore skb from r1. */
9576 changes_data = bpf_helper_changes_pkt_data(fn->func);
9577 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
9578 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
9579 func_id_name(func_id), func_id);
9583 memset(&meta, 0, sizeof(meta));
9584 meta.pkt_access = fn->pkt_access;
9586 err = check_func_proto(fn, func_id);
9588 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
9589 func_id_name(func_id), func_id);
9593 if (env->cur_state->active_rcu_lock) {
9594 if (fn->might_sleep) {
9595 verbose(env, "sleepable helper %s#%d in rcu_read_lock region\n",
9596 func_id_name(func_id), func_id);
9600 if (env->prog->aux->sleepable && is_storage_get_function(func_id))
9601 env->insn_aux_data[insn_idx].storage_get_func_atomic = true;
9604 meta.func_id = func_id;
9606 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
9607 err = check_func_arg(env, i, &meta, fn, insn_idx);
9612 err = record_func_map(env, &meta, func_id, insn_idx);
9616 err = record_func_key(env, &meta, func_id, insn_idx);
9620 /* Mark slots with STACK_MISC in case of raw mode, stack offset
9621 * is inferred from register state.
9623 for (i = 0; i < meta.access_size; i++) {
9624 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
9625 BPF_WRITE, -1, false, false);
9630 regs = cur_regs(env);
9632 if (meta.release_regno) {
9634 /* This can only be set for PTR_TO_STACK, as CONST_PTR_TO_DYNPTR cannot
9635 * be released by any dynptr helper. Hence, unmark_stack_slots_dynptr
9636 * is safe to do directly.
9638 if (arg_type_is_dynptr(fn->arg_type[meta.release_regno - BPF_REG_1])) {
9639 if (regs[meta.release_regno].type == CONST_PTR_TO_DYNPTR) {
9640 verbose(env, "verifier internal error: CONST_PTR_TO_DYNPTR cannot be released\n");
9643 err = unmark_stack_slots_dynptr(env, ®s[meta.release_regno]);
9644 } else if (meta.ref_obj_id) {
9645 err = release_reference(env, meta.ref_obj_id);
9646 } else if (register_is_null(®s[meta.release_regno])) {
9647 /* meta.ref_obj_id can only be 0 if register that is meant to be
9648 * released is NULL, which must be > R0.
9653 verbose(env, "func %s#%d reference has not been acquired before\n",
9654 func_id_name(func_id), func_id);
9660 case BPF_FUNC_tail_call:
9661 err = check_reference_leak(env);
9663 verbose(env, "tail_call would lead to reference leak\n");
9667 case BPF_FUNC_get_local_storage:
9668 /* check that flags argument in get_local_storage(map, flags) is 0,
9669 * this is required because get_local_storage() can't return an error.
9671 if (!register_is_null(®s[BPF_REG_2])) {
9672 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
9676 case BPF_FUNC_for_each_map_elem:
9677 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9678 set_map_elem_callback_state);
9680 case BPF_FUNC_timer_set_callback:
9681 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9682 set_timer_callback_state);
9684 case BPF_FUNC_find_vma:
9685 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9686 set_find_vma_callback_state);
9688 case BPF_FUNC_snprintf:
9689 err = check_bpf_snprintf_call(env, regs);
9692 update_loop_inline_state(env, meta.subprogno);
9693 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9694 set_loop_callback_state);
9696 case BPF_FUNC_dynptr_from_mem:
9697 if (regs[BPF_REG_1].type != PTR_TO_MAP_VALUE) {
9698 verbose(env, "Unsupported reg type %s for bpf_dynptr_from_mem data\n",
9699 reg_type_str(env, regs[BPF_REG_1].type));
9703 case BPF_FUNC_set_retval:
9704 if (prog_type == BPF_PROG_TYPE_LSM &&
9705 env->prog->expected_attach_type == BPF_LSM_CGROUP) {
9706 if (!env->prog->aux->attach_func_proto->type) {
9707 /* Make sure programs that attach to void
9708 * hooks don't try to modify return value.
9710 verbose(env, "BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
9715 case BPF_FUNC_dynptr_data:
9717 struct bpf_reg_state *reg;
9720 reg = get_dynptr_arg_reg(env, fn, regs);
9725 if (meta.dynptr_id) {
9726 verbose(env, "verifier internal error: meta.dynptr_id already set\n");
9729 if (meta.ref_obj_id) {
9730 verbose(env, "verifier internal error: meta.ref_obj_id already set\n");
9734 id = dynptr_id(env, reg);
9736 verbose(env, "verifier internal error: failed to obtain dynptr id\n");
9740 ref_obj_id = dynptr_ref_obj_id(env, reg);
9741 if (ref_obj_id < 0) {
9742 verbose(env, "verifier internal error: failed to obtain dynptr ref_obj_id\n");
9746 meta.dynptr_id = id;
9747 meta.ref_obj_id = ref_obj_id;
9751 case BPF_FUNC_dynptr_write:
9753 enum bpf_dynptr_type dynptr_type;
9754 struct bpf_reg_state *reg;
9756 reg = get_dynptr_arg_reg(env, fn, regs);
9760 dynptr_type = dynptr_get_type(env, reg);
9761 if (dynptr_type == BPF_DYNPTR_TYPE_INVALID)
9764 if (dynptr_type == BPF_DYNPTR_TYPE_SKB)
9765 /* this will trigger clear_all_pkt_pointers(), which will
9766 * invalidate all dynptr slices associated with the skb
9768 changes_data = true;
9772 case BPF_FUNC_user_ringbuf_drain:
9773 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9774 set_user_ringbuf_callback_state);
9781 /* reset caller saved regs */
9782 for (i = 0; i < CALLER_SAVED_REGS; i++) {
9783 mark_reg_not_init(env, regs, caller_saved[i]);
9784 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
9787 /* helper call returns 64-bit value. */
9788 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
9790 /* update return register (already marked as written above) */
9791 ret_type = fn->ret_type;
9792 ret_flag = type_flag(ret_type);
9794 switch (base_type(ret_type)) {
9796 /* sets type to SCALAR_VALUE */
9797 mark_reg_unknown(env, regs, BPF_REG_0);
9800 regs[BPF_REG_0].type = NOT_INIT;
9802 case RET_PTR_TO_MAP_VALUE:
9803 /* There is no offset yet applied, variable or fixed */
9804 mark_reg_known_zero(env, regs, BPF_REG_0);
9805 /* remember map_ptr, so that check_map_access()
9806 * can check 'value_size' boundary of memory access
9807 * to map element returned from bpf_map_lookup_elem()
9809 if (meta.map_ptr == NULL) {
9811 "kernel subsystem misconfigured verifier\n");
9814 regs[BPF_REG_0].map_ptr = meta.map_ptr;
9815 regs[BPF_REG_0].map_uid = meta.map_uid;
9816 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE | ret_flag;
9817 if (!type_may_be_null(ret_type) &&
9818 btf_record_has_field(meta.map_ptr->record, BPF_SPIN_LOCK)) {
9819 regs[BPF_REG_0].id = ++env->id_gen;
9822 case RET_PTR_TO_SOCKET:
9823 mark_reg_known_zero(env, regs, BPF_REG_0);
9824 regs[BPF_REG_0].type = PTR_TO_SOCKET | ret_flag;
9826 case RET_PTR_TO_SOCK_COMMON:
9827 mark_reg_known_zero(env, regs, BPF_REG_0);
9828 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON | ret_flag;
9830 case RET_PTR_TO_TCP_SOCK:
9831 mark_reg_known_zero(env, regs, BPF_REG_0);
9832 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK | ret_flag;
9834 case RET_PTR_TO_MEM:
9835 mark_reg_known_zero(env, regs, BPF_REG_0);
9836 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
9837 regs[BPF_REG_0].mem_size = meta.mem_size;
9839 case RET_PTR_TO_MEM_OR_BTF_ID:
9841 const struct btf_type *t;
9843 mark_reg_known_zero(env, regs, BPF_REG_0);
9844 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
9845 if (!btf_type_is_struct(t)) {
9847 const struct btf_type *ret;
9850 /* resolve the type size of ksym. */
9851 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
9853 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
9854 verbose(env, "unable to resolve the size of type '%s': %ld\n",
9855 tname, PTR_ERR(ret));
9858 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
9859 regs[BPF_REG_0].mem_size = tsize;
9861 /* MEM_RDONLY may be carried from ret_flag, but it
9862 * doesn't apply on PTR_TO_BTF_ID. Fold it, otherwise
9863 * it will confuse the check of PTR_TO_BTF_ID in
9864 * check_mem_access().
9866 ret_flag &= ~MEM_RDONLY;
9868 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
9869 regs[BPF_REG_0].btf = meta.ret_btf;
9870 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
9874 case RET_PTR_TO_BTF_ID:
9876 struct btf *ret_btf;
9879 mark_reg_known_zero(env, regs, BPF_REG_0);
9880 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
9881 if (func_id == BPF_FUNC_kptr_xchg) {
9882 ret_btf = meta.kptr_field->kptr.btf;
9883 ret_btf_id = meta.kptr_field->kptr.btf_id;
9884 if (!btf_is_kernel(ret_btf))
9885 regs[BPF_REG_0].type |= MEM_ALLOC;
9887 if (fn->ret_btf_id == BPF_PTR_POISON) {
9888 verbose(env, "verifier internal error:");
9889 verbose(env, "func %s has non-overwritten BPF_PTR_POISON return type\n",
9890 func_id_name(func_id));
9893 ret_btf = btf_vmlinux;
9894 ret_btf_id = *fn->ret_btf_id;
9896 if (ret_btf_id == 0) {
9897 verbose(env, "invalid return type %u of func %s#%d\n",
9898 base_type(ret_type), func_id_name(func_id),
9902 regs[BPF_REG_0].btf = ret_btf;
9903 regs[BPF_REG_0].btf_id = ret_btf_id;
9907 verbose(env, "unknown return type %u of func %s#%d\n",
9908 base_type(ret_type), func_id_name(func_id), func_id);
9912 if (type_may_be_null(regs[BPF_REG_0].type))
9913 regs[BPF_REG_0].id = ++env->id_gen;
9915 if (helper_multiple_ref_obj_use(func_id, meta.map_ptr)) {
9916 verbose(env, "verifier internal error: func %s#%d sets ref_obj_id more than once\n",
9917 func_id_name(func_id), func_id);
9921 if (is_dynptr_ref_function(func_id))
9922 regs[BPF_REG_0].dynptr_id = meta.dynptr_id;
9924 if (is_ptr_cast_function(func_id) || is_dynptr_ref_function(func_id)) {
9925 /* For release_reference() */
9926 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
9927 } else if (is_acquire_function(func_id, meta.map_ptr)) {
9928 int id = acquire_reference_state(env, insn_idx);
9932 /* For mark_ptr_or_null_reg() */
9933 regs[BPF_REG_0].id = id;
9934 /* For release_reference() */
9935 regs[BPF_REG_0].ref_obj_id = id;
9938 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
9940 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
9944 if ((func_id == BPF_FUNC_get_stack ||
9945 func_id == BPF_FUNC_get_task_stack) &&
9946 !env->prog->has_callchain_buf) {
9947 const char *err_str;
9949 #ifdef CONFIG_PERF_EVENTS
9950 err = get_callchain_buffers(sysctl_perf_event_max_stack);
9951 err_str = "cannot get callchain buffer for func %s#%d\n";
9954 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
9957 verbose(env, err_str, func_id_name(func_id), func_id);
9961 env->prog->has_callchain_buf = true;
9964 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
9965 env->prog->call_get_stack = true;
9967 if (func_id == BPF_FUNC_get_func_ip) {
9968 if (check_get_func_ip(env))
9970 env->prog->call_get_func_ip = true;
9974 clear_all_pkt_pointers(env);
9978 /* mark_btf_func_reg_size() is used when the reg size is determined by
9979 * the BTF func_proto's return value size and argument.
9981 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
9984 struct bpf_reg_state *reg = &cur_regs(env)[regno];
9986 if (regno == BPF_REG_0) {
9987 /* Function return value */
9988 reg->live |= REG_LIVE_WRITTEN;
9989 reg->subreg_def = reg_size == sizeof(u64) ?
9990 DEF_NOT_SUBREG : env->insn_idx + 1;
9992 /* Function argument */
9993 if (reg_size == sizeof(u64)) {
9994 mark_insn_zext(env, reg);
9995 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
9997 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
10002 static bool is_kfunc_acquire(struct bpf_kfunc_call_arg_meta *meta)
10004 return meta->kfunc_flags & KF_ACQUIRE;
10007 static bool is_kfunc_release(struct bpf_kfunc_call_arg_meta *meta)
10009 return meta->kfunc_flags & KF_RELEASE;
10012 static bool is_kfunc_trusted_args(struct bpf_kfunc_call_arg_meta *meta)
10014 return (meta->kfunc_flags & KF_TRUSTED_ARGS) || is_kfunc_release(meta);
10017 static bool is_kfunc_sleepable(struct bpf_kfunc_call_arg_meta *meta)
10019 return meta->kfunc_flags & KF_SLEEPABLE;
10022 static bool is_kfunc_destructive(struct bpf_kfunc_call_arg_meta *meta)
10024 return meta->kfunc_flags & KF_DESTRUCTIVE;
10027 static bool is_kfunc_rcu(struct bpf_kfunc_call_arg_meta *meta)
10029 return meta->kfunc_flags & KF_RCU;
10032 static bool __kfunc_param_match_suffix(const struct btf *btf,
10033 const struct btf_param *arg,
10034 const char *suffix)
10036 int suffix_len = strlen(suffix), len;
10037 const char *param_name;
10039 /* In the future, this can be ported to use BTF tagging */
10040 param_name = btf_name_by_offset(btf, arg->name_off);
10041 if (str_is_empty(param_name))
10043 len = strlen(param_name);
10044 if (len < suffix_len)
10046 param_name += len - suffix_len;
10047 return !strncmp(param_name, suffix, suffix_len);
10050 static bool is_kfunc_arg_mem_size(const struct btf *btf,
10051 const struct btf_param *arg,
10052 const struct bpf_reg_state *reg)
10054 const struct btf_type *t;
10056 t = btf_type_skip_modifiers(btf, arg->type, NULL);
10057 if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
10060 return __kfunc_param_match_suffix(btf, arg, "__sz");
10063 static bool is_kfunc_arg_const_mem_size(const struct btf *btf,
10064 const struct btf_param *arg,
10065 const struct bpf_reg_state *reg)
10067 const struct btf_type *t;
10069 t = btf_type_skip_modifiers(btf, arg->type, NULL);
10070 if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
10073 return __kfunc_param_match_suffix(btf, arg, "__szk");
10076 static bool is_kfunc_arg_optional(const struct btf *btf, const struct btf_param *arg)
10078 return __kfunc_param_match_suffix(btf, arg, "__opt");
10081 static bool is_kfunc_arg_constant(const struct btf *btf, const struct btf_param *arg)
10083 return __kfunc_param_match_suffix(btf, arg, "__k");
10086 static bool is_kfunc_arg_ignore(const struct btf *btf, const struct btf_param *arg)
10088 return __kfunc_param_match_suffix(btf, arg, "__ign");
10091 static bool is_kfunc_arg_alloc_obj(const struct btf *btf, const struct btf_param *arg)
10093 return __kfunc_param_match_suffix(btf, arg, "__alloc");
10096 static bool is_kfunc_arg_uninit(const struct btf *btf, const struct btf_param *arg)
10098 return __kfunc_param_match_suffix(btf, arg, "__uninit");
10101 static bool is_kfunc_arg_refcounted_kptr(const struct btf *btf, const struct btf_param *arg)
10103 return __kfunc_param_match_suffix(btf, arg, "__refcounted_kptr");
10106 static bool is_kfunc_arg_scalar_with_name(const struct btf *btf,
10107 const struct btf_param *arg,
10110 int len, target_len = strlen(name);
10111 const char *param_name;
10113 param_name = btf_name_by_offset(btf, arg->name_off);
10114 if (str_is_empty(param_name))
10116 len = strlen(param_name);
10117 if (len != target_len)
10119 if (strcmp(param_name, name))
10127 KF_ARG_LIST_HEAD_ID,
10128 KF_ARG_LIST_NODE_ID,
10133 BTF_ID_LIST(kf_arg_btf_ids)
10134 BTF_ID(struct, bpf_dynptr_kern)
10135 BTF_ID(struct, bpf_list_head)
10136 BTF_ID(struct, bpf_list_node)
10137 BTF_ID(struct, bpf_rb_root)
10138 BTF_ID(struct, bpf_rb_node)
10140 static bool __is_kfunc_ptr_arg_type(const struct btf *btf,
10141 const struct btf_param *arg, int type)
10143 const struct btf_type *t;
10146 t = btf_type_skip_modifiers(btf, arg->type, NULL);
10149 if (!btf_type_is_ptr(t))
10151 t = btf_type_skip_modifiers(btf, t->type, &res_id);
10154 return btf_types_are_same(btf, res_id, btf_vmlinux, kf_arg_btf_ids[type]);
10157 static bool is_kfunc_arg_dynptr(const struct btf *btf, const struct btf_param *arg)
10159 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_DYNPTR_ID);
10162 static bool is_kfunc_arg_list_head(const struct btf *btf, const struct btf_param *arg)
10164 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_HEAD_ID);
10167 static bool is_kfunc_arg_list_node(const struct btf *btf, const struct btf_param *arg)
10169 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_NODE_ID);
10172 static bool is_kfunc_arg_rbtree_root(const struct btf *btf, const struct btf_param *arg)
10174 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RB_ROOT_ID);
10177 static bool is_kfunc_arg_rbtree_node(const struct btf *btf, const struct btf_param *arg)
10179 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RB_NODE_ID);
10182 static bool is_kfunc_arg_callback(struct bpf_verifier_env *env, const struct btf *btf,
10183 const struct btf_param *arg)
10185 const struct btf_type *t;
10187 t = btf_type_resolve_func_ptr(btf, arg->type, NULL);
10194 /* Returns true if struct is composed of scalars, 4 levels of nesting allowed */
10195 static bool __btf_type_is_scalar_struct(struct bpf_verifier_env *env,
10196 const struct btf *btf,
10197 const struct btf_type *t, int rec)
10199 const struct btf_type *member_type;
10200 const struct btf_member *member;
10203 if (!btf_type_is_struct(t))
10206 for_each_member(i, t, member) {
10207 const struct btf_array *array;
10209 member_type = btf_type_skip_modifiers(btf, member->type, NULL);
10210 if (btf_type_is_struct(member_type)) {
10212 verbose(env, "max struct nesting depth exceeded\n");
10215 if (!__btf_type_is_scalar_struct(env, btf, member_type, rec + 1))
10219 if (btf_type_is_array(member_type)) {
10220 array = btf_array(member_type);
10221 if (!array->nelems)
10223 member_type = btf_type_skip_modifiers(btf, array->type, NULL);
10224 if (!btf_type_is_scalar(member_type))
10228 if (!btf_type_is_scalar(member_type))
10234 enum kfunc_ptr_arg_type {
10236 KF_ARG_PTR_TO_ALLOC_BTF_ID, /* Allocated object */
10237 KF_ARG_PTR_TO_REFCOUNTED_KPTR, /* Refcounted local kptr */
10238 KF_ARG_PTR_TO_DYNPTR,
10239 KF_ARG_PTR_TO_ITER,
10240 KF_ARG_PTR_TO_LIST_HEAD,
10241 KF_ARG_PTR_TO_LIST_NODE,
10242 KF_ARG_PTR_TO_BTF_ID, /* Also covers reg2btf_ids conversions */
10244 KF_ARG_PTR_TO_MEM_SIZE, /* Size derived from next argument, skip it */
10245 KF_ARG_PTR_TO_CALLBACK,
10246 KF_ARG_PTR_TO_RB_ROOT,
10247 KF_ARG_PTR_TO_RB_NODE,
10250 enum special_kfunc_type {
10251 KF_bpf_obj_new_impl,
10252 KF_bpf_obj_drop_impl,
10253 KF_bpf_refcount_acquire_impl,
10254 KF_bpf_list_push_front_impl,
10255 KF_bpf_list_push_back_impl,
10256 KF_bpf_list_pop_front,
10257 KF_bpf_list_pop_back,
10258 KF_bpf_cast_to_kern_ctx,
10259 KF_bpf_rdonly_cast,
10260 KF_bpf_rcu_read_lock,
10261 KF_bpf_rcu_read_unlock,
10262 KF_bpf_rbtree_remove,
10263 KF_bpf_rbtree_add_impl,
10264 KF_bpf_rbtree_first,
10265 KF_bpf_dynptr_from_skb,
10266 KF_bpf_dynptr_from_xdp,
10267 KF_bpf_dynptr_slice,
10268 KF_bpf_dynptr_slice_rdwr,
10269 KF_bpf_dynptr_clone,
10272 BTF_SET_START(special_kfunc_set)
10273 BTF_ID(func, bpf_obj_new_impl)
10274 BTF_ID(func, bpf_obj_drop_impl)
10275 BTF_ID(func, bpf_refcount_acquire_impl)
10276 BTF_ID(func, bpf_list_push_front_impl)
10277 BTF_ID(func, bpf_list_push_back_impl)
10278 BTF_ID(func, bpf_list_pop_front)
10279 BTF_ID(func, bpf_list_pop_back)
10280 BTF_ID(func, bpf_cast_to_kern_ctx)
10281 BTF_ID(func, bpf_rdonly_cast)
10282 BTF_ID(func, bpf_rbtree_remove)
10283 BTF_ID(func, bpf_rbtree_add_impl)
10284 BTF_ID(func, bpf_rbtree_first)
10285 BTF_ID(func, bpf_dynptr_from_skb)
10286 BTF_ID(func, bpf_dynptr_from_xdp)
10287 BTF_ID(func, bpf_dynptr_slice)
10288 BTF_ID(func, bpf_dynptr_slice_rdwr)
10289 BTF_ID(func, bpf_dynptr_clone)
10290 BTF_SET_END(special_kfunc_set)
10292 BTF_ID_LIST(special_kfunc_list)
10293 BTF_ID(func, bpf_obj_new_impl)
10294 BTF_ID(func, bpf_obj_drop_impl)
10295 BTF_ID(func, bpf_refcount_acquire_impl)
10296 BTF_ID(func, bpf_list_push_front_impl)
10297 BTF_ID(func, bpf_list_push_back_impl)
10298 BTF_ID(func, bpf_list_pop_front)
10299 BTF_ID(func, bpf_list_pop_back)
10300 BTF_ID(func, bpf_cast_to_kern_ctx)
10301 BTF_ID(func, bpf_rdonly_cast)
10302 BTF_ID(func, bpf_rcu_read_lock)
10303 BTF_ID(func, bpf_rcu_read_unlock)
10304 BTF_ID(func, bpf_rbtree_remove)
10305 BTF_ID(func, bpf_rbtree_add_impl)
10306 BTF_ID(func, bpf_rbtree_first)
10307 BTF_ID(func, bpf_dynptr_from_skb)
10308 BTF_ID(func, bpf_dynptr_from_xdp)
10309 BTF_ID(func, bpf_dynptr_slice)
10310 BTF_ID(func, bpf_dynptr_slice_rdwr)
10311 BTF_ID(func, bpf_dynptr_clone)
10313 static bool is_kfunc_ret_null(struct bpf_kfunc_call_arg_meta *meta)
10315 if (meta->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl] &&
10316 meta->arg_owning_ref) {
10320 return meta->kfunc_flags & KF_RET_NULL;
10323 static bool is_kfunc_bpf_rcu_read_lock(struct bpf_kfunc_call_arg_meta *meta)
10325 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_lock];
10328 static bool is_kfunc_bpf_rcu_read_unlock(struct bpf_kfunc_call_arg_meta *meta)
10330 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_unlock];
10333 static enum kfunc_ptr_arg_type
10334 get_kfunc_ptr_arg_type(struct bpf_verifier_env *env,
10335 struct bpf_kfunc_call_arg_meta *meta,
10336 const struct btf_type *t, const struct btf_type *ref_t,
10337 const char *ref_tname, const struct btf_param *args,
10338 int argno, int nargs)
10340 u32 regno = argno + 1;
10341 struct bpf_reg_state *regs = cur_regs(env);
10342 struct bpf_reg_state *reg = ®s[regno];
10343 bool arg_mem_size = false;
10345 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx])
10346 return KF_ARG_PTR_TO_CTX;
10348 /* In this function, we verify the kfunc's BTF as per the argument type,
10349 * leaving the rest of the verification with respect to the register
10350 * type to our caller. When a set of conditions hold in the BTF type of
10351 * arguments, we resolve it to a known kfunc_ptr_arg_type.
10353 if (btf_get_prog_ctx_type(&env->log, meta->btf, t, resolve_prog_type(env->prog), argno))
10354 return KF_ARG_PTR_TO_CTX;
10356 if (is_kfunc_arg_alloc_obj(meta->btf, &args[argno]))
10357 return KF_ARG_PTR_TO_ALLOC_BTF_ID;
10359 if (is_kfunc_arg_refcounted_kptr(meta->btf, &args[argno]))
10360 return KF_ARG_PTR_TO_REFCOUNTED_KPTR;
10362 if (is_kfunc_arg_dynptr(meta->btf, &args[argno]))
10363 return KF_ARG_PTR_TO_DYNPTR;
10365 if (is_kfunc_arg_iter(meta, argno))
10366 return KF_ARG_PTR_TO_ITER;
10368 if (is_kfunc_arg_list_head(meta->btf, &args[argno]))
10369 return KF_ARG_PTR_TO_LIST_HEAD;
10371 if (is_kfunc_arg_list_node(meta->btf, &args[argno]))
10372 return KF_ARG_PTR_TO_LIST_NODE;
10374 if (is_kfunc_arg_rbtree_root(meta->btf, &args[argno]))
10375 return KF_ARG_PTR_TO_RB_ROOT;
10377 if (is_kfunc_arg_rbtree_node(meta->btf, &args[argno]))
10378 return KF_ARG_PTR_TO_RB_NODE;
10380 if ((base_type(reg->type) == PTR_TO_BTF_ID || reg2btf_ids[base_type(reg->type)])) {
10381 if (!btf_type_is_struct(ref_t)) {
10382 verbose(env, "kernel function %s args#%d pointer type %s %s is not supported\n",
10383 meta->func_name, argno, btf_type_str(ref_t), ref_tname);
10386 return KF_ARG_PTR_TO_BTF_ID;
10389 if (is_kfunc_arg_callback(env, meta->btf, &args[argno]))
10390 return KF_ARG_PTR_TO_CALLBACK;
10393 if (argno + 1 < nargs &&
10394 (is_kfunc_arg_mem_size(meta->btf, &args[argno + 1], ®s[regno + 1]) ||
10395 is_kfunc_arg_const_mem_size(meta->btf, &args[argno + 1], ®s[regno + 1])))
10396 arg_mem_size = true;
10398 /* This is the catch all argument type of register types supported by
10399 * check_helper_mem_access. However, we only allow when argument type is
10400 * pointer to scalar, or struct composed (recursively) of scalars. When
10401 * arg_mem_size is true, the pointer can be void *.
10403 if (!btf_type_is_scalar(ref_t) && !__btf_type_is_scalar_struct(env, meta->btf, ref_t, 0) &&
10404 (arg_mem_size ? !btf_type_is_void(ref_t) : 1)) {
10405 verbose(env, "arg#%d pointer type %s %s must point to %sscalar, or struct with scalar\n",
10406 argno, btf_type_str(ref_t), ref_tname, arg_mem_size ? "void, " : "");
10409 return arg_mem_size ? KF_ARG_PTR_TO_MEM_SIZE : KF_ARG_PTR_TO_MEM;
10412 static int process_kf_arg_ptr_to_btf_id(struct bpf_verifier_env *env,
10413 struct bpf_reg_state *reg,
10414 const struct btf_type *ref_t,
10415 const char *ref_tname, u32 ref_id,
10416 struct bpf_kfunc_call_arg_meta *meta,
10419 const struct btf_type *reg_ref_t;
10420 bool strict_type_match = false;
10421 const struct btf *reg_btf;
10422 const char *reg_ref_tname;
10425 if (base_type(reg->type) == PTR_TO_BTF_ID) {
10426 reg_btf = reg->btf;
10427 reg_ref_id = reg->btf_id;
10429 reg_btf = btf_vmlinux;
10430 reg_ref_id = *reg2btf_ids[base_type(reg->type)];
10433 /* Enforce strict type matching for calls to kfuncs that are acquiring
10434 * or releasing a reference, or are no-cast aliases. We do _not_
10435 * enforce strict matching for plain KF_TRUSTED_ARGS kfuncs by default,
10436 * as we want to enable BPF programs to pass types that are bitwise
10437 * equivalent without forcing them to explicitly cast with something
10438 * like bpf_cast_to_kern_ctx().
10440 * For example, say we had a type like the following:
10442 * struct bpf_cpumask {
10443 * cpumask_t cpumask;
10444 * refcount_t usage;
10447 * Note that as specified in <linux/cpumask.h>, cpumask_t is typedef'ed
10448 * to a struct cpumask, so it would be safe to pass a struct
10449 * bpf_cpumask * to a kfunc expecting a struct cpumask *.
10451 * The philosophy here is similar to how we allow scalars of different
10452 * types to be passed to kfuncs as long as the size is the same. The
10453 * only difference here is that we're simply allowing
10454 * btf_struct_ids_match() to walk the struct at the 0th offset, and
10457 if (is_kfunc_acquire(meta) ||
10458 (is_kfunc_release(meta) && reg->ref_obj_id) ||
10459 btf_type_ids_nocast_alias(&env->log, reg_btf, reg_ref_id, meta->btf, ref_id))
10460 strict_type_match = true;
10462 WARN_ON_ONCE(is_kfunc_trusted_args(meta) && reg->off);
10464 reg_ref_t = btf_type_skip_modifiers(reg_btf, reg_ref_id, ®_ref_id);
10465 reg_ref_tname = btf_name_by_offset(reg_btf, reg_ref_t->name_off);
10466 if (!btf_struct_ids_match(&env->log, reg_btf, reg_ref_id, reg->off, meta->btf, ref_id, strict_type_match)) {
10467 verbose(env, "kernel function %s args#%d expected pointer to %s %s but R%d has a pointer to %s %s\n",
10468 meta->func_name, argno, btf_type_str(ref_t), ref_tname, argno + 1,
10469 btf_type_str(reg_ref_t), reg_ref_tname);
10475 static int ref_set_non_owning(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
10477 struct bpf_verifier_state *state = env->cur_state;
10478 struct btf_record *rec = reg_btf_record(reg);
10480 if (!state->active_lock.ptr) {
10481 verbose(env, "verifier internal error: ref_set_non_owning w/o active lock\n");
10485 if (type_flag(reg->type) & NON_OWN_REF) {
10486 verbose(env, "verifier internal error: NON_OWN_REF already set\n");
10490 reg->type |= NON_OWN_REF;
10491 if (rec->refcount_off >= 0)
10492 reg->type |= MEM_RCU;
10497 static int ref_convert_owning_non_owning(struct bpf_verifier_env *env, u32 ref_obj_id)
10499 struct bpf_func_state *state, *unused;
10500 struct bpf_reg_state *reg;
10503 state = cur_func(env);
10506 verbose(env, "verifier internal error: ref_obj_id is zero for "
10507 "owning -> non-owning conversion\n");
10511 for (i = 0; i < state->acquired_refs; i++) {
10512 if (state->refs[i].id != ref_obj_id)
10515 /* Clear ref_obj_id here so release_reference doesn't clobber
10518 bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
10519 if (reg->ref_obj_id == ref_obj_id) {
10520 reg->ref_obj_id = 0;
10521 ref_set_non_owning(env, reg);
10527 verbose(env, "verifier internal error: ref state missing for ref_obj_id\n");
10531 /* Implementation details:
10533 * Each register points to some region of memory, which we define as an
10534 * allocation. Each allocation may embed a bpf_spin_lock which protects any
10535 * special BPF objects (bpf_list_head, bpf_rb_root, etc.) part of the same
10536 * allocation. The lock and the data it protects are colocated in the same
10539 * Hence, everytime a register holds a pointer value pointing to such
10540 * allocation, the verifier preserves a unique reg->id for it.
10542 * The verifier remembers the lock 'ptr' and the lock 'id' whenever
10543 * bpf_spin_lock is called.
10545 * To enable this, lock state in the verifier captures two values:
10546 * active_lock.ptr = Register's type specific pointer
10547 * active_lock.id = A unique ID for each register pointer value
10549 * Currently, PTR_TO_MAP_VALUE and PTR_TO_BTF_ID | MEM_ALLOC are the two
10550 * supported register types.
10552 * The active_lock.ptr in case of map values is the reg->map_ptr, and in case of
10553 * allocated objects is the reg->btf pointer.
10555 * The active_lock.id is non-unique for maps supporting direct_value_addr, as we
10556 * can establish the provenance of the map value statically for each distinct
10557 * lookup into such maps. They always contain a single map value hence unique
10558 * IDs for each pseudo load pessimizes the algorithm and rejects valid programs.
10560 * So, in case of global variables, they use array maps with max_entries = 1,
10561 * hence their active_lock.ptr becomes map_ptr and id = 0 (since they all point
10562 * into the same map value as max_entries is 1, as described above).
10564 * In case of inner map lookups, the inner map pointer has same map_ptr as the
10565 * outer map pointer (in verifier context), but each lookup into an inner map
10566 * assigns a fresh reg->id to the lookup, so while lookups into distinct inner
10567 * maps from the same outer map share the same map_ptr as active_lock.ptr, they
10568 * will get different reg->id assigned to each lookup, hence different
10571 * In case of allocated objects, active_lock.ptr is the reg->btf, and the
10572 * reg->id is a unique ID preserved after the NULL pointer check on the pointer
10573 * returned from bpf_obj_new. Each allocation receives a new reg->id.
10575 static int check_reg_allocation_locked(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
10580 switch ((int)reg->type) {
10581 case PTR_TO_MAP_VALUE:
10582 ptr = reg->map_ptr;
10584 case PTR_TO_BTF_ID | MEM_ALLOC:
10588 verbose(env, "verifier internal error: unknown reg type for lock check\n");
10593 if (!env->cur_state->active_lock.ptr)
10595 if (env->cur_state->active_lock.ptr != ptr ||
10596 env->cur_state->active_lock.id != id) {
10597 verbose(env, "held lock and object are not in the same allocation\n");
10603 static bool is_bpf_list_api_kfunc(u32 btf_id)
10605 return btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
10606 btf_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
10607 btf_id == special_kfunc_list[KF_bpf_list_pop_front] ||
10608 btf_id == special_kfunc_list[KF_bpf_list_pop_back];
10611 static bool is_bpf_rbtree_api_kfunc(u32 btf_id)
10613 return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl] ||
10614 btf_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
10615 btf_id == special_kfunc_list[KF_bpf_rbtree_first];
10618 static bool is_bpf_graph_api_kfunc(u32 btf_id)
10620 return is_bpf_list_api_kfunc(btf_id) || is_bpf_rbtree_api_kfunc(btf_id) ||
10621 btf_id == special_kfunc_list[KF_bpf_refcount_acquire_impl];
10624 static bool is_callback_calling_kfunc(u32 btf_id)
10626 return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl];
10629 static bool is_rbtree_lock_required_kfunc(u32 btf_id)
10631 return is_bpf_rbtree_api_kfunc(btf_id);
10634 static bool check_kfunc_is_graph_root_api(struct bpf_verifier_env *env,
10635 enum btf_field_type head_field_type,
10640 switch (head_field_type) {
10641 case BPF_LIST_HEAD:
10642 ret = is_bpf_list_api_kfunc(kfunc_btf_id);
10645 ret = is_bpf_rbtree_api_kfunc(kfunc_btf_id);
10648 verbose(env, "verifier internal error: unexpected graph root argument type %s\n",
10649 btf_field_type_name(head_field_type));
10654 verbose(env, "verifier internal error: %s head arg for unknown kfunc\n",
10655 btf_field_type_name(head_field_type));
10659 static bool check_kfunc_is_graph_node_api(struct bpf_verifier_env *env,
10660 enum btf_field_type node_field_type,
10665 switch (node_field_type) {
10666 case BPF_LIST_NODE:
10667 ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
10668 kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_back_impl]);
10671 ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
10672 kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl]);
10675 verbose(env, "verifier internal error: unexpected graph node argument type %s\n",
10676 btf_field_type_name(node_field_type));
10681 verbose(env, "verifier internal error: %s node arg for unknown kfunc\n",
10682 btf_field_type_name(node_field_type));
10687 __process_kf_arg_ptr_to_graph_root(struct bpf_verifier_env *env,
10688 struct bpf_reg_state *reg, u32 regno,
10689 struct bpf_kfunc_call_arg_meta *meta,
10690 enum btf_field_type head_field_type,
10691 struct btf_field **head_field)
10693 const char *head_type_name;
10694 struct btf_field *field;
10695 struct btf_record *rec;
10698 if (meta->btf != btf_vmlinux) {
10699 verbose(env, "verifier internal error: unexpected btf mismatch in kfunc call\n");
10703 if (!check_kfunc_is_graph_root_api(env, head_field_type, meta->func_id))
10706 head_type_name = btf_field_type_name(head_field_type);
10707 if (!tnum_is_const(reg->var_off)) {
10709 "R%d doesn't have constant offset. %s has to be at the constant offset\n",
10710 regno, head_type_name);
10714 rec = reg_btf_record(reg);
10715 head_off = reg->off + reg->var_off.value;
10716 field = btf_record_find(rec, head_off, head_field_type);
10718 verbose(env, "%s not found at offset=%u\n", head_type_name, head_off);
10722 /* All functions require bpf_list_head to be protected using a bpf_spin_lock */
10723 if (check_reg_allocation_locked(env, reg)) {
10724 verbose(env, "bpf_spin_lock at off=%d must be held for %s\n",
10725 rec->spin_lock_off, head_type_name);
10730 verbose(env, "verifier internal error: repeating %s arg\n", head_type_name);
10733 *head_field = field;
10737 static int process_kf_arg_ptr_to_list_head(struct bpf_verifier_env *env,
10738 struct bpf_reg_state *reg, u32 regno,
10739 struct bpf_kfunc_call_arg_meta *meta)
10741 return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, BPF_LIST_HEAD,
10742 &meta->arg_list_head.field);
10745 static int process_kf_arg_ptr_to_rbtree_root(struct bpf_verifier_env *env,
10746 struct bpf_reg_state *reg, u32 regno,
10747 struct bpf_kfunc_call_arg_meta *meta)
10749 return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, BPF_RB_ROOT,
10750 &meta->arg_rbtree_root.field);
10754 __process_kf_arg_ptr_to_graph_node(struct bpf_verifier_env *env,
10755 struct bpf_reg_state *reg, u32 regno,
10756 struct bpf_kfunc_call_arg_meta *meta,
10757 enum btf_field_type head_field_type,
10758 enum btf_field_type node_field_type,
10759 struct btf_field **node_field)
10761 const char *node_type_name;
10762 const struct btf_type *et, *t;
10763 struct btf_field *field;
10766 if (meta->btf != btf_vmlinux) {
10767 verbose(env, "verifier internal error: unexpected btf mismatch in kfunc call\n");
10771 if (!check_kfunc_is_graph_node_api(env, node_field_type, meta->func_id))
10774 node_type_name = btf_field_type_name(node_field_type);
10775 if (!tnum_is_const(reg->var_off)) {
10777 "R%d doesn't have constant offset. %s has to be at the constant offset\n",
10778 regno, node_type_name);
10782 node_off = reg->off + reg->var_off.value;
10783 field = reg_find_field_offset(reg, node_off, node_field_type);
10784 if (!field || field->offset != node_off) {
10785 verbose(env, "%s not found at offset=%u\n", node_type_name, node_off);
10789 field = *node_field;
10791 et = btf_type_by_id(field->graph_root.btf, field->graph_root.value_btf_id);
10792 t = btf_type_by_id(reg->btf, reg->btf_id);
10793 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, 0, field->graph_root.btf,
10794 field->graph_root.value_btf_id, true)) {
10795 verbose(env, "operation on %s expects arg#1 %s at offset=%d "
10796 "in struct %s, but arg is at offset=%d in struct %s\n",
10797 btf_field_type_name(head_field_type),
10798 btf_field_type_name(node_field_type),
10799 field->graph_root.node_offset,
10800 btf_name_by_offset(field->graph_root.btf, et->name_off),
10801 node_off, btf_name_by_offset(reg->btf, t->name_off));
10804 meta->arg_btf = reg->btf;
10805 meta->arg_btf_id = reg->btf_id;
10807 if (node_off != field->graph_root.node_offset) {
10808 verbose(env, "arg#1 offset=%d, but expected %s at offset=%d in struct %s\n",
10809 node_off, btf_field_type_name(node_field_type),
10810 field->graph_root.node_offset,
10811 btf_name_by_offset(field->graph_root.btf, et->name_off));
10818 static int process_kf_arg_ptr_to_list_node(struct bpf_verifier_env *env,
10819 struct bpf_reg_state *reg, u32 regno,
10820 struct bpf_kfunc_call_arg_meta *meta)
10822 return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta,
10823 BPF_LIST_HEAD, BPF_LIST_NODE,
10824 &meta->arg_list_head.field);
10827 static int process_kf_arg_ptr_to_rbtree_node(struct bpf_verifier_env *env,
10828 struct bpf_reg_state *reg, u32 regno,
10829 struct bpf_kfunc_call_arg_meta *meta)
10831 return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta,
10832 BPF_RB_ROOT, BPF_RB_NODE,
10833 &meta->arg_rbtree_root.field);
10836 static int check_kfunc_args(struct bpf_verifier_env *env, struct bpf_kfunc_call_arg_meta *meta,
10839 const char *func_name = meta->func_name, *ref_tname;
10840 const struct btf *btf = meta->btf;
10841 const struct btf_param *args;
10842 struct btf_record *rec;
10846 args = (const struct btf_param *)(meta->func_proto + 1);
10847 nargs = btf_type_vlen(meta->func_proto);
10848 if (nargs > MAX_BPF_FUNC_REG_ARGS) {
10849 verbose(env, "Function %s has %d > %d args\n", func_name, nargs,
10850 MAX_BPF_FUNC_REG_ARGS);
10854 /* Check that BTF function arguments match actual types that the
10857 for (i = 0; i < nargs; i++) {
10858 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[i + 1];
10859 const struct btf_type *t, *ref_t, *resolve_ret;
10860 enum bpf_arg_type arg_type = ARG_DONTCARE;
10861 u32 regno = i + 1, ref_id, type_size;
10862 bool is_ret_buf_sz = false;
10865 t = btf_type_skip_modifiers(btf, args[i].type, NULL);
10867 if (is_kfunc_arg_ignore(btf, &args[i]))
10870 if (btf_type_is_scalar(t)) {
10871 if (reg->type != SCALAR_VALUE) {
10872 verbose(env, "R%d is not a scalar\n", regno);
10876 if (is_kfunc_arg_constant(meta->btf, &args[i])) {
10877 if (meta->arg_constant.found) {
10878 verbose(env, "verifier internal error: only one constant argument permitted\n");
10881 if (!tnum_is_const(reg->var_off)) {
10882 verbose(env, "R%d must be a known constant\n", regno);
10885 ret = mark_chain_precision(env, regno);
10888 meta->arg_constant.found = true;
10889 meta->arg_constant.value = reg->var_off.value;
10890 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdonly_buf_size")) {
10891 meta->r0_rdonly = true;
10892 is_ret_buf_sz = true;
10893 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdwr_buf_size")) {
10894 is_ret_buf_sz = true;
10897 if (is_ret_buf_sz) {
10898 if (meta->r0_size) {
10899 verbose(env, "2 or more rdonly/rdwr_buf_size parameters for kfunc");
10903 if (!tnum_is_const(reg->var_off)) {
10904 verbose(env, "R%d is not a const\n", regno);
10908 meta->r0_size = reg->var_off.value;
10909 ret = mark_chain_precision(env, regno);
10916 if (!btf_type_is_ptr(t)) {
10917 verbose(env, "Unrecognized arg#%d type %s\n", i, btf_type_str(t));
10921 if ((is_kfunc_trusted_args(meta) || is_kfunc_rcu(meta)) &&
10922 (register_is_null(reg) || type_may_be_null(reg->type))) {
10923 verbose(env, "Possibly NULL pointer passed to trusted arg%d\n", i);
10927 if (reg->ref_obj_id) {
10928 if (is_kfunc_release(meta) && meta->ref_obj_id) {
10929 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
10930 regno, reg->ref_obj_id,
10934 meta->ref_obj_id = reg->ref_obj_id;
10935 if (is_kfunc_release(meta))
10936 meta->release_regno = regno;
10939 ref_t = btf_type_skip_modifiers(btf, t->type, &ref_id);
10940 ref_tname = btf_name_by_offset(btf, ref_t->name_off);
10942 kf_arg_type = get_kfunc_ptr_arg_type(env, meta, t, ref_t, ref_tname, args, i, nargs);
10943 if (kf_arg_type < 0)
10944 return kf_arg_type;
10946 switch (kf_arg_type) {
10947 case KF_ARG_PTR_TO_ALLOC_BTF_ID:
10948 case KF_ARG_PTR_TO_BTF_ID:
10949 if (!is_kfunc_trusted_args(meta) && !is_kfunc_rcu(meta))
10952 if (!is_trusted_reg(reg)) {
10953 if (!is_kfunc_rcu(meta)) {
10954 verbose(env, "R%d must be referenced or trusted\n", regno);
10957 if (!is_rcu_reg(reg)) {
10958 verbose(env, "R%d must be a rcu pointer\n", regno);
10964 case KF_ARG_PTR_TO_CTX:
10965 /* Trusted arguments have the same offset checks as release arguments */
10966 arg_type |= OBJ_RELEASE;
10968 case KF_ARG_PTR_TO_DYNPTR:
10969 case KF_ARG_PTR_TO_ITER:
10970 case KF_ARG_PTR_TO_LIST_HEAD:
10971 case KF_ARG_PTR_TO_LIST_NODE:
10972 case KF_ARG_PTR_TO_RB_ROOT:
10973 case KF_ARG_PTR_TO_RB_NODE:
10974 case KF_ARG_PTR_TO_MEM:
10975 case KF_ARG_PTR_TO_MEM_SIZE:
10976 case KF_ARG_PTR_TO_CALLBACK:
10977 case KF_ARG_PTR_TO_REFCOUNTED_KPTR:
10978 /* Trusted by default */
10985 if (is_kfunc_release(meta) && reg->ref_obj_id)
10986 arg_type |= OBJ_RELEASE;
10987 ret = check_func_arg_reg_off(env, reg, regno, arg_type);
10991 switch (kf_arg_type) {
10992 case KF_ARG_PTR_TO_CTX:
10993 if (reg->type != PTR_TO_CTX) {
10994 verbose(env, "arg#%d expected pointer to ctx, but got %s\n", i, btf_type_str(t));
10998 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
10999 ret = get_kern_ctx_btf_id(&env->log, resolve_prog_type(env->prog));
11002 meta->ret_btf_id = ret;
11005 case KF_ARG_PTR_TO_ALLOC_BTF_ID:
11006 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11007 verbose(env, "arg#%d expected pointer to allocated object\n", i);
11010 if (!reg->ref_obj_id) {
11011 verbose(env, "allocated object must be referenced\n");
11014 if (meta->btf == btf_vmlinux &&
11015 meta->func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
11016 meta->arg_btf = reg->btf;
11017 meta->arg_btf_id = reg->btf_id;
11020 case KF_ARG_PTR_TO_DYNPTR:
11022 enum bpf_arg_type dynptr_arg_type = ARG_PTR_TO_DYNPTR;
11023 int clone_ref_obj_id = 0;
11025 if (reg->type != PTR_TO_STACK &&
11026 reg->type != CONST_PTR_TO_DYNPTR) {
11027 verbose(env, "arg#%d expected pointer to stack or dynptr_ptr\n", i);
11031 if (reg->type == CONST_PTR_TO_DYNPTR)
11032 dynptr_arg_type |= MEM_RDONLY;
11034 if (is_kfunc_arg_uninit(btf, &args[i]))
11035 dynptr_arg_type |= MEM_UNINIT;
11037 if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) {
11038 dynptr_arg_type |= DYNPTR_TYPE_SKB;
11039 } else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_xdp]) {
11040 dynptr_arg_type |= DYNPTR_TYPE_XDP;
11041 } else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_clone] &&
11042 (dynptr_arg_type & MEM_UNINIT)) {
11043 enum bpf_dynptr_type parent_type = meta->initialized_dynptr.type;
11045 if (parent_type == BPF_DYNPTR_TYPE_INVALID) {
11046 verbose(env, "verifier internal error: no dynptr type for parent of clone\n");
11050 dynptr_arg_type |= (unsigned int)get_dynptr_type_flag(parent_type);
11051 clone_ref_obj_id = meta->initialized_dynptr.ref_obj_id;
11052 if (dynptr_type_refcounted(parent_type) && !clone_ref_obj_id) {
11053 verbose(env, "verifier internal error: missing ref obj id for parent of clone\n");
11058 ret = process_dynptr_func(env, regno, insn_idx, dynptr_arg_type, clone_ref_obj_id);
11062 if (!(dynptr_arg_type & MEM_UNINIT)) {
11063 int id = dynptr_id(env, reg);
11066 verbose(env, "verifier internal error: failed to obtain dynptr id\n");
11069 meta->initialized_dynptr.id = id;
11070 meta->initialized_dynptr.type = dynptr_get_type(env, reg);
11071 meta->initialized_dynptr.ref_obj_id = dynptr_ref_obj_id(env, reg);
11076 case KF_ARG_PTR_TO_ITER:
11077 ret = process_iter_arg(env, regno, insn_idx, meta);
11081 case KF_ARG_PTR_TO_LIST_HEAD:
11082 if (reg->type != PTR_TO_MAP_VALUE &&
11083 reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11084 verbose(env, "arg#%d expected pointer to map value or allocated object\n", i);
11087 if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
11088 verbose(env, "allocated object must be referenced\n");
11091 ret = process_kf_arg_ptr_to_list_head(env, reg, regno, meta);
11095 case KF_ARG_PTR_TO_RB_ROOT:
11096 if (reg->type != PTR_TO_MAP_VALUE &&
11097 reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11098 verbose(env, "arg#%d expected pointer to map value or allocated object\n", i);
11101 if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
11102 verbose(env, "allocated object must be referenced\n");
11105 ret = process_kf_arg_ptr_to_rbtree_root(env, reg, regno, meta);
11109 case KF_ARG_PTR_TO_LIST_NODE:
11110 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11111 verbose(env, "arg#%d expected pointer to allocated object\n", i);
11114 if (!reg->ref_obj_id) {
11115 verbose(env, "allocated object must be referenced\n");
11118 ret = process_kf_arg_ptr_to_list_node(env, reg, regno, meta);
11122 case KF_ARG_PTR_TO_RB_NODE:
11123 if (meta->func_id == special_kfunc_list[KF_bpf_rbtree_remove]) {
11124 if (!type_is_non_owning_ref(reg->type) || reg->ref_obj_id) {
11125 verbose(env, "rbtree_remove node input must be non-owning ref\n");
11128 if (in_rbtree_lock_required_cb(env)) {
11129 verbose(env, "rbtree_remove not allowed in rbtree cb\n");
11133 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11134 verbose(env, "arg#%d expected pointer to allocated object\n", i);
11137 if (!reg->ref_obj_id) {
11138 verbose(env, "allocated object must be referenced\n");
11143 ret = process_kf_arg_ptr_to_rbtree_node(env, reg, regno, meta);
11147 case KF_ARG_PTR_TO_BTF_ID:
11148 /* Only base_type is checked, further checks are done here */
11149 if ((base_type(reg->type) != PTR_TO_BTF_ID ||
11150 (bpf_type_has_unsafe_modifiers(reg->type) && !is_rcu_reg(reg))) &&
11151 !reg2btf_ids[base_type(reg->type)]) {
11152 verbose(env, "arg#%d is %s ", i, reg_type_str(env, reg->type));
11153 verbose(env, "expected %s or socket\n",
11154 reg_type_str(env, base_type(reg->type) |
11155 (type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS)));
11158 ret = process_kf_arg_ptr_to_btf_id(env, reg, ref_t, ref_tname, ref_id, meta, i);
11162 case KF_ARG_PTR_TO_MEM:
11163 resolve_ret = btf_resolve_size(btf, ref_t, &type_size);
11164 if (IS_ERR(resolve_ret)) {
11165 verbose(env, "arg#%d reference type('%s %s') size cannot be determined: %ld\n",
11166 i, btf_type_str(ref_t), ref_tname, PTR_ERR(resolve_ret));
11169 ret = check_mem_reg(env, reg, regno, type_size);
11173 case KF_ARG_PTR_TO_MEM_SIZE:
11175 struct bpf_reg_state *buff_reg = ®s[regno];
11176 const struct btf_param *buff_arg = &args[i];
11177 struct bpf_reg_state *size_reg = ®s[regno + 1];
11178 const struct btf_param *size_arg = &args[i + 1];
11180 if (!register_is_null(buff_reg) || !is_kfunc_arg_optional(meta->btf, buff_arg)) {
11181 ret = check_kfunc_mem_size_reg(env, size_reg, regno + 1);
11183 verbose(env, "arg#%d arg#%d memory, len pair leads to invalid memory access\n", i, i + 1);
11188 if (is_kfunc_arg_const_mem_size(meta->btf, size_arg, size_reg)) {
11189 if (meta->arg_constant.found) {
11190 verbose(env, "verifier internal error: only one constant argument permitted\n");
11193 if (!tnum_is_const(size_reg->var_off)) {
11194 verbose(env, "R%d must be a known constant\n", regno + 1);
11197 meta->arg_constant.found = true;
11198 meta->arg_constant.value = size_reg->var_off.value;
11201 /* Skip next '__sz' or '__szk' argument */
11205 case KF_ARG_PTR_TO_CALLBACK:
11206 if (reg->type != PTR_TO_FUNC) {
11207 verbose(env, "arg%d expected pointer to func\n", i);
11210 meta->subprogno = reg->subprogno;
11212 case KF_ARG_PTR_TO_REFCOUNTED_KPTR:
11213 if (!type_is_ptr_alloc_obj(reg->type)) {
11214 verbose(env, "arg#%d is neither owning or non-owning ref\n", i);
11217 if (!type_is_non_owning_ref(reg->type))
11218 meta->arg_owning_ref = true;
11220 rec = reg_btf_record(reg);
11222 verbose(env, "verifier internal error: Couldn't find btf_record\n");
11226 if (rec->refcount_off < 0) {
11227 verbose(env, "arg#%d doesn't point to a type with bpf_refcount field\n", i);
11231 meta->arg_btf = reg->btf;
11232 meta->arg_btf_id = reg->btf_id;
11237 if (is_kfunc_release(meta) && !meta->release_regno) {
11238 verbose(env, "release kernel function %s expects refcounted PTR_TO_BTF_ID\n",
11246 static int fetch_kfunc_meta(struct bpf_verifier_env *env,
11247 struct bpf_insn *insn,
11248 struct bpf_kfunc_call_arg_meta *meta,
11249 const char **kfunc_name)
11251 const struct btf_type *func, *func_proto;
11252 u32 func_id, *kfunc_flags;
11253 const char *func_name;
11254 struct btf *desc_btf;
11257 *kfunc_name = NULL;
11262 desc_btf = find_kfunc_desc_btf(env, insn->off);
11263 if (IS_ERR(desc_btf))
11264 return PTR_ERR(desc_btf);
11266 func_id = insn->imm;
11267 func = btf_type_by_id(desc_btf, func_id);
11268 func_name = btf_name_by_offset(desc_btf, func->name_off);
11270 *kfunc_name = func_name;
11271 func_proto = btf_type_by_id(desc_btf, func->type);
11273 kfunc_flags = btf_kfunc_id_set_contains(desc_btf, func_id, env->prog);
11274 if (!kfunc_flags) {
11278 memset(meta, 0, sizeof(*meta));
11279 meta->btf = desc_btf;
11280 meta->func_id = func_id;
11281 meta->kfunc_flags = *kfunc_flags;
11282 meta->func_proto = func_proto;
11283 meta->func_name = func_name;
11288 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
11291 const struct btf_type *t, *ptr_type;
11292 u32 i, nargs, ptr_type_id, release_ref_obj_id;
11293 struct bpf_reg_state *regs = cur_regs(env);
11294 const char *func_name, *ptr_type_name;
11295 bool sleepable, rcu_lock, rcu_unlock;
11296 struct bpf_kfunc_call_arg_meta meta;
11297 struct bpf_insn_aux_data *insn_aux;
11298 int err, insn_idx = *insn_idx_p;
11299 const struct btf_param *args;
11300 const struct btf_type *ret_t;
11301 struct btf *desc_btf;
11303 /* skip for now, but return error when we find this in fixup_kfunc_call */
11307 err = fetch_kfunc_meta(env, insn, &meta, &func_name);
11308 if (err == -EACCES && func_name)
11309 verbose(env, "calling kernel function %s is not allowed\n", func_name);
11312 desc_btf = meta.btf;
11313 insn_aux = &env->insn_aux_data[insn_idx];
11315 insn_aux->is_iter_next = is_iter_next_kfunc(&meta);
11317 if (is_kfunc_destructive(&meta) && !capable(CAP_SYS_BOOT)) {
11318 verbose(env, "destructive kfunc calls require CAP_SYS_BOOT capability\n");
11322 sleepable = is_kfunc_sleepable(&meta);
11323 if (sleepable && !env->prog->aux->sleepable) {
11324 verbose(env, "program must be sleepable to call sleepable kfunc %s\n", func_name);
11328 rcu_lock = is_kfunc_bpf_rcu_read_lock(&meta);
11329 rcu_unlock = is_kfunc_bpf_rcu_read_unlock(&meta);
11331 if (env->cur_state->active_rcu_lock) {
11332 struct bpf_func_state *state;
11333 struct bpf_reg_state *reg;
11335 if (in_rbtree_lock_required_cb(env) && (rcu_lock || rcu_unlock)) {
11336 verbose(env, "Calling bpf_rcu_read_{lock,unlock} in unnecessary rbtree callback\n");
11341 verbose(env, "nested rcu read lock (kernel function %s)\n", func_name);
11343 } else if (rcu_unlock) {
11344 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
11345 if (reg->type & MEM_RCU) {
11346 reg->type &= ~(MEM_RCU | PTR_MAYBE_NULL);
11347 reg->type |= PTR_UNTRUSTED;
11350 env->cur_state->active_rcu_lock = false;
11351 } else if (sleepable) {
11352 verbose(env, "kernel func %s is sleepable within rcu_read_lock region\n", func_name);
11355 } else if (rcu_lock) {
11356 env->cur_state->active_rcu_lock = true;
11357 } else if (rcu_unlock) {
11358 verbose(env, "unmatched rcu read unlock (kernel function %s)\n", func_name);
11362 /* Check the arguments */
11363 err = check_kfunc_args(env, &meta, insn_idx);
11366 /* In case of release function, we get register number of refcounted
11367 * PTR_TO_BTF_ID in bpf_kfunc_arg_meta, do the release now.
11369 if (meta.release_regno) {
11370 err = release_reference(env, regs[meta.release_regno].ref_obj_id);
11372 verbose(env, "kfunc %s#%d reference has not been acquired before\n",
11373 func_name, meta.func_id);
11378 if (meta.func_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
11379 meta.func_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
11380 meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
11381 release_ref_obj_id = regs[BPF_REG_2].ref_obj_id;
11382 insn_aux->insert_off = regs[BPF_REG_2].off;
11383 insn_aux->kptr_struct_meta = btf_find_struct_meta(meta.arg_btf, meta.arg_btf_id);
11384 err = ref_convert_owning_non_owning(env, release_ref_obj_id);
11386 verbose(env, "kfunc %s#%d conversion of owning ref to non-owning failed\n",
11387 func_name, meta.func_id);
11391 err = release_reference(env, release_ref_obj_id);
11393 verbose(env, "kfunc %s#%d reference has not been acquired before\n",
11394 func_name, meta.func_id);
11399 if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
11400 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
11401 set_rbtree_add_callback_state);
11403 verbose(env, "kfunc %s#%d failed callback verification\n",
11404 func_name, meta.func_id);
11409 for (i = 0; i < CALLER_SAVED_REGS; i++)
11410 mark_reg_not_init(env, regs, caller_saved[i]);
11412 /* Check return type */
11413 t = btf_type_skip_modifiers(desc_btf, meta.func_proto->type, NULL);
11415 if (is_kfunc_acquire(&meta) && !btf_type_is_struct_ptr(meta.btf, t)) {
11416 /* Only exception is bpf_obj_new_impl */
11417 if (meta.btf != btf_vmlinux ||
11418 (meta.func_id != special_kfunc_list[KF_bpf_obj_new_impl] &&
11419 meta.func_id != special_kfunc_list[KF_bpf_refcount_acquire_impl])) {
11420 verbose(env, "acquire kernel function does not return PTR_TO_BTF_ID\n");
11425 if (btf_type_is_scalar(t)) {
11426 mark_reg_unknown(env, regs, BPF_REG_0);
11427 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
11428 } else if (btf_type_is_ptr(t)) {
11429 ptr_type = btf_type_skip_modifiers(desc_btf, t->type, &ptr_type_id);
11431 if (meta.btf == btf_vmlinux && btf_id_set_contains(&special_kfunc_set, meta.func_id)) {
11432 if (meta.func_id == special_kfunc_list[KF_bpf_obj_new_impl]) {
11433 struct btf *ret_btf;
11436 if (unlikely(!bpf_global_ma_set))
11439 if (((u64)(u32)meta.arg_constant.value) != meta.arg_constant.value) {
11440 verbose(env, "local type ID argument must be in range [0, U32_MAX]\n");
11444 ret_btf = env->prog->aux->btf;
11445 ret_btf_id = meta.arg_constant.value;
11447 /* This may be NULL due to user not supplying a BTF */
11449 verbose(env, "bpf_obj_new requires prog BTF\n");
11453 ret_t = btf_type_by_id(ret_btf, ret_btf_id);
11454 if (!ret_t || !__btf_type_is_struct(ret_t)) {
11455 verbose(env, "bpf_obj_new type ID argument must be of a struct\n");
11459 mark_reg_known_zero(env, regs, BPF_REG_0);
11460 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
11461 regs[BPF_REG_0].btf = ret_btf;
11462 regs[BPF_REG_0].btf_id = ret_btf_id;
11464 insn_aux->obj_new_size = ret_t->size;
11465 insn_aux->kptr_struct_meta =
11466 btf_find_struct_meta(ret_btf, ret_btf_id);
11467 } else if (meta.func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) {
11468 mark_reg_known_zero(env, regs, BPF_REG_0);
11469 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
11470 regs[BPF_REG_0].btf = meta.arg_btf;
11471 regs[BPF_REG_0].btf_id = meta.arg_btf_id;
11473 insn_aux->kptr_struct_meta =
11474 btf_find_struct_meta(meta.arg_btf,
11476 } else if (meta.func_id == special_kfunc_list[KF_bpf_list_pop_front] ||
11477 meta.func_id == special_kfunc_list[KF_bpf_list_pop_back]) {
11478 struct btf_field *field = meta.arg_list_head.field;
11480 mark_reg_graph_node(regs, BPF_REG_0, &field->graph_root);
11481 } else if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
11482 meta.func_id == special_kfunc_list[KF_bpf_rbtree_first]) {
11483 struct btf_field *field = meta.arg_rbtree_root.field;
11485 mark_reg_graph_node(regs, BPF_REG_0, &field->graph_root);
11486 } else if (meta.func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
11487 mark_reg_known_zero(env, regs, BPF_REG_0);
11488 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_TRUSTED;
11489 regs[BPF_REG_0].btf = desc_btf;
11490 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
11491 } else if (meta.func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
11492 ret_t = btf_type_by_id(desc_btf, meta.arg_constant.value);
11493 if (!ret_t || !btf_type_is_struct(ret_t)) {
11495 "kfunc bpf_rdonly_cast type ID argument must be of a struct\n");
11499 mark_reg_known_zero(env, regs, BPF_REG_0);
11500 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
11501 regs[BPF_REG_0].btf = desc_btf;
11502 regs[BPF_REG_0].btf_id = meta.arg_constant.value;
11503 } else if (meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice] ||
11504 meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice_rdwr]) {
11505 enum bpf_type_flag type_flag = get_dynptr_type_flag(meta.initialized_dynptr.type);
11507 mark_reg_known_zero(env, regs, BPF_REG_0);
11509 if (!meta.arg_constant.found) {
11510 verbose(env, "verifier internal error: bpf_dynptr_slice(_rdwr) no constant size\n");
11514 regs[BPF_REG_0].mem_size = meta.arg_constant.value;
11516 /* PTR_MAYBE_NULL will be added when is_kfunc_ret_null is checked */
11517 regs[BPF_REG_0].type = PTR_TO_MEM | type_flag;
11519 if (meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice]) {
11520 regs[BPF_REG_0].type |= MEM_RDONLY;
11522 /* this will set env->seen_direct_write to true */
11523 if (!may_access_direct_pkt_data(env, NULL, BPF_WRITE)) {
11524 verbose(env, "the prog does not allow writes to packet data\n");
11529 if (!meta.initialized_dynptr.id) {
11530 verbose(env, "verifier internal error: no dynptr id\n");
11533 regs[BPF_REG_0].dynptr_id = meta.initialized_dynptr.id;
11535 /* we don't need to set BPF_REG_0's ref obj id
11536 * because packet slices are not refcounted (see
11537 * dynptr_type_refcounted)
11540 verbose(env, "kernel function %s unhandled dynamic return type\n",
11544 } else if (!__btf_type_is_struct(ptr_type)) {
11545 if (!meta.r0_size) {
11548 if (!IS_ERR(btf_resolve_size(desc_btf, ptr_type, &sz))) {
11550 meta.r0_rdonly = true;
11553 if (!meta.r0_size) {
11554 ptr_type_name = btf_name_by_offset(desc_btf,
11555 ptr_type->name_off);
11557 "kernel function %s returns pointer type %s %s is not supported\n",
11559 btf_type_str(ptr_type),
11564 mark_reg_known_zero(env, regs, BPF_REG_0);
11565 regs[BPF_REG_0].type = PTR_TO_MEM;
11566 regs[BPF_REG_0].mem_size = meta.r0_size;
11568 if (meta.r0_rdonly)
11569 regs[BPF_REG_0].type |= MEM_RDONLY;
11571 /* Ensures we don't access the memory after a release_reference() */
11572 if (meta.ref_obj_id)
11573 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
11575 mark_reg_known_zero(env, regs, BPF_REG_0);
11576 regs[BPF_REG_0].btf = desc_btf;
11577 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
11578 regs[BPF_REG_0].btf_id = ptr_type_id;
11581 if (is_kfunc_ret_null(&meta)) {
11582 regs[BPF_REG_0].type |= PTR_MAYBE_NULL;
11583 /* For mark_ptr_or_null_reg, see 93c230e3f5bd6 */
11584 regs[BPF_REG_0].id = ++env->id_gen;
11586 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
11587 if (is_kfunc_acquire(&meta)) {
11588 int id = acquire_reference_state(env, insn_idx);
11592 if (is_kfunc_ret_null(&meta))
11593 regs[BPF_REG_0].id = id;
11594 regs[BPF_REG_0].ref_obj_id = id;
11595 } else if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_first]) {
11596 ref_set_non_owning(env, ®s[BPF_REG_0]);
11599 if (reg_may_point_to_spin_lock(®s[BPF_REG_0]) && !regs[BPF_REG_0].id)
11600 regs[BPF_REG_0].id = ++env->id_gen;
11601 } else if (btf_type_is_void(t)) {
11602 if (meta.btf == btf_vmlinux && btf_id_set_contains(&special_kfunc_set, meta.func_id)) {
11603 if (meta.func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
11604 insn_aux->kptr_struct_meta =
11605 btf_find_struct_meta(meta.arg_btf,
11611 nargs = btf_type_vlen(meta.func_proto);
11612 args = (const struct btf_param *)(meta.func_proto + 1);
11613 for (i = 0; i < nargs; i++) {
11616 t = btf_type_skip_modifiers(desc_btf, args[i].type, NULL);
11617 if (btf_type_is_ptr(t))
11618 mark_btf_func_reg_size(env, regno, sizeof(void *));
11620 /* scalar. ensured by btf_check_kfunc_arg_match() */
11621 mark_btf_func_reg_size(env, regno, t->size);
11624 if (is_iter_next_kfunc(&meta)) {
11625 err = process_iter_next_call(env, insn_idx, &meta);
11633 static bool signed_add_overflows(s64 a, s64 b)
11635 /* Do the add in u64, where overflow is well-defined */
11636 s64 res = (s64)((u64)a + (u64)b);
11643 static bool signed_add32_overflows(s32 a, s32 b)
11645 /* Do the add in u32, where overflow is well-defined */
11646 s32 res = (s32)((u32)a + (u32)b);
11653 static bool signed_sub_overflows(s64 a, s64 b)
11655 /* Do the sub in u64, where overflow is well-defined */
11656 s64 res = (s64)((u64)a - (u64)b);
11663 static bool signed_sub32_overflows(s32 a, s32 b)
11665 /* Do the sub in u32, where overflow is well-defined */
11666 s32 res = (s32)((u32)a - (u32)b);
11673 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
11674 const struct bpf_reg_state *reg,
11675 enum bpf_reg_type type)
11677 bool known = tnum_is_const(reg->var_off);
11678 s64 val = reg->var_off.value;
11679 s64 smin = reg->smin_value;
11681 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
11682 verbose(env, "math between %s pointer and %lld is not allowed\n",
11683 reg_type_str(env, type), val);
11687 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
11688 verbose(env, "%s pointer offset %d is not allowed\n",
11689 reg_type_str(env, type), reg->off);
11693 if (smin == S64_MIN) {
11694 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
11695 reg_type_str(env, type));
11699 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
11700 verbose(env, "value %lld makes %s pointer be out of bounds\n",
11701 smin, reg_type_str(env, type));
11709 REASON_BOUNDS = -1,
11716 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
11717 u32 *alu_limit, bool mask_to_left)
11719 u32 max = 0, ptr_limit = 0;
11721 switch (ptr_reg->type) {
11723 /* Offset 0 is out-of-bounds, but acceptable start for the
11724 * left direction, see BPF_REG_FP. Also, unknown scalar
11725 * offset where we would need to deal with min/max bounds is
11726 * currently prohibited for unprivileged.
11728 max = MAX_BPF_STACK + mask_to_left;
11729 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
11731 case PTR_TO_MAP_VALUE:
11732 max = ptr_reg->map_ptr->value_size;
11733 ptr_limit = (mask_to_left ?
11734 ptr_reg->smin_value :
11735 ptr_reg->umax_value) + ptr_reg->off;
11738 return REASON_TYPE;
11741 if (ptr_limit >= max)
11742 return REASON_LIMIT;
11743 *alu_limit = ptr_limit;
11747 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
11748 const struct bpf_insn *insn)
11750 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
11753 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
11754 u32 alu_state, u32 alu_limit)
11756 /* If we arrived here from different branches with different
11757 * state or limits to sanitize, then this won't work.
11759 if (aux->alu_state &&
11760 (aux->alu_state != alu_state ||
11761 aux->alu_limit != alu_limit))
11762 return REASON_PATHS;
11764 /* Corresponding fixup done in do_misc_fixups(). */
11765 aux->alu_state = alu_state;
11766 aux->alu_limit = alu_limit;
11770 static int sanitize_val_alu(struct bpf_verifier_env *env,
11771 struct bpf_insn *insn)
11773 struct bpf_insn_aux_data *aux = cur_aux(env);
11775 if (can_skip_alu_sanitation(env, insn))
11778 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
11781 static bool sanitize_needed(u8 opcode)
11783 return opcode == BPF_ADD || opcode == BPF_SUB;
11786 struct bpf_sanitize_info {
11787 struct bpf_insn_aux_data aux;
11791 static struct bpf_verifier_state *
11792 sanitize_speculative_path(struct bpf_verifier_env *env,
11793 const struct bpf_insn *insn,
11794 u32 next_idx, u32 curr_idx)
11796 struct bpf_verifier_state *branch;
11797 struct bpf_reg_state *regs;
11799 branch = push_stack(env, next_idx, curr_idx, true);
11800 if (branch && insn) {
11801 regs = branch->frame[branch->curframe]->regs;
11802 if (BPF_SRC(insn->code) == BPF_K) {
11803 mark_reg_unknown(env, regs, insn->dst_reg);
11804 } else if (BPF_SRC(insn->code) == BPF_X) {
11805 mark_reg_unknown(env, regs, insn->dst_reg);
11806 mark_reg_unknown(env, regs, insn->src_reg);
11812 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
11813 struct bpf_insn *insn,
11814 const struct bpf_reg_state *ptr_reg,
11815 const struct bpf_reg_state *off_reg,
11816 struct bpf_reg_state *dst_reg,
11817 struct bpf_sanitize_info *info,
11818 const bool commit_window)
11820 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
11821 struct bpf_verifier_state *vstate = env->cur_state;
11822 bool off_is_imm = tnum_is_const(off_reg->var_off);
11823 bool off_is_neg = off_reg->smin_value < 0;
11824 bool ptr_is_dst_reg = ptr_reg == dst_reg;
11825 u8 opcode = BPF_OP(insn->code);
11826 u32 alu_state, alu_limit;
11827 struct bpf_reg_state tmp;
11831 if (can_skip_alu_sanitation(env, insn))
11834 /* We already marked aux for masking from non-speculative
11835 * paths, thus we got here in the first place. We only care
11836 * to explore bad access from here.
11838 if (vstate->speculative)
11841 if (!commit_window) {
11842 if (!tnum_is_const(off_reg->var_off) &&
11843 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
11844 return REASON_BOUNDS;
11846 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
11847 (opcode == BPF_SUB && !off_is_neg);
11850 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
11854 if (commit_window) {
11855 /* In commit phase we narrow the masking window based on
11856 * the observed pointer move after the simulated operation.
11858 alu_state = info->aux.alu_state;
11859 alu_limit = abs(info->aux.alu_limit - alu_limit);
11861 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
11862 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
11863 alu_state |= ptr_is_dst_reg ?
11864 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
11866 /* Limit pruning on unknown scalars to enable deep search for
11867 * potential masking differences from other program paths.
11870 env->explore_alu_limits = true;
11873 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
11877 /* If we're in commit phase, we're done here given we already
11878 * pushed the truncated dst_reg into the speculative verification
11881 * Also, when register is a known constant, we rewrite register-based
11882 * operation to immediate-based, and thus do not need masking (and as
11883 * a consequence, do not need to simulate the zero-truncation either).
11885 if (commit_window || off_is_imm)
11888 /* Simulate and find potential out-of-bounds access under
11889 * speculative execution from truncation as a result of
11890 * masking when off was not within expected range. If off
11891 * sits in dst, then we temporarily need to move ptr there
11892 * to simulate dst (== 0) +/-= ptr. Needed, for example,
11893 * for cases where we use K-based arithmetic in one direction
11894 * and truncated reg-based in the other in order to explore
11897 if (!ptr_is_dst_reg) {
11899 copy_register_state(dst_reg, ptr_reg);
11901 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
11903 if (!ptr_is_dst_reg && ret)
11905 return !ret ? REASON_STACK : 0;
11908 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
11910 struct bpf_verifier_state *vstate = env->cur_state;
11912 /* If we simulate paths under speculation, we don't update the
11913 * insn as 'seen' such that when we verify unreachable paths in
11914 * the non-speculative domain, sanitize_dead_code() can still
11915 * rewrite/sanitize them.
11917 if (!vstate->speculative)
11918 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
11921 static int sanitize_err(struct bpf_verifier_env *env,
11922 const struct bpf_insn *insn, int reason,
11923 const struct bpf_reg_state *off_reg,
11924 const struct bpf_reg_state *dst_reg)
11926 static const char *err = "pointer arithmetic with it prohibited for !root";
11927 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
11928 u32 dst = insn->dst_reg, src = insn->src_reg;
11931 case REASON_BOUNDS:
11932 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
11933 off_reg == dst_reg ? dst : src, err);
11936 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
11937 off_reg == dst_reg ? src : dst, err);
11940 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
11944 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
11948 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
11952 verbose(env, "verifier internal error: unknown reason (%d)\n",
11960 /* check that stack access falls within stack limits and that 'reg' doesn't
11961 * have a variable offset.
11963 * Variable offset is prohibited for unprivileged mode for simplicity since it
11964 * requires corresponding support in Spectre masking for stack ALU. See also
11965 * retrieve_ptr_limit().
11968 * 'off' includes 'reg->off'.
11970 static int check_stack_access_for_ptr_arithmetic(
11971 struct bpf_verifier_env *env,
11973 const struct bpf_reg_state *reg,
11976 if (!tnum_is_const(reg->var_off)) {
11979 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
11980 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
11981 regno, tn_buf, off);
11985 if (off >= 0 || off < -MAX_BPF_STACK) {
11986 verbose(env, "R%d stack pointer arithmetic goes out of range, "
11987 "prohibited for !root; off=%d\n", regno, off);
11994 static int sanitize_check_bounds(struct bpf_verifier_env *env,
11995 const struct bpf_insn *insn,
11996 const struct bpf_reg_state *dst_reg)
11998 u32 dst = insn->dst_reg;
12000 /* For unprivileged we require that resulting offset must be in bounds
12001 * in order to be able to sanitize access later on.
12003 if (env->bypass_spec_v1)
12006 switch (dst_reg->type) {
12008 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
12009 dst_reg->off + dst_reg->var_off.value))
12012 case PTR_TO_MAP_VALUE:
12013 if (check_map_access(env, dst, dst_reg->off, 1, false, ACCESS_HELPER)) {
12014 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
12015 "prohibited for !root\n", dst);
12026 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
12027 * Caller should also handle BPF_MOV case separately.
12028 * If we return -EACCES, caller may want to try again treating pointer as a
12029 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
12031 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
12032 struct bpf_insn *insn,
12033 const struct bpf_reg_state *ptr_reg,
12034 const struct bpf_reg_state *off_reg)
12036 struct bpf_verifier_state *vstate = env->cur_state;
12037 struct bpf_func_state *state = vstate->frame[vstate->curframe];
12038 struct bpf_reg_state *regs = state->regs, *dst_reg;
12039 bool known = tnum_is_const(off_reg->var_off);
12040 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
12041 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
12042 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
12043 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
12044 struct bpf_sanitize_info info = {};
12045 u8 opcode = BPF_OP(insn->code);
12046 u32 dst = insn->dst_reg;
12049 dst_reg = ®s[dst];
12051 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
12052 smin_val > smax_val || umin_val > umax_val) {
12053 /* Taint dst register if offset had invalid bounds derived from
12054 * e.g. dead branches.
12056 __mark_reg_unknown(env, dst_reg);
12060 if (BPF_CLASS(insn->code) != BPF_ALU64) {
12061 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
12062 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
12063 __mark_reg_unknown(env, dst_reg);
12068 "R%d 32-bit pointer arithmetic prohibited\n",
12073 if (ptr_reg->type & PTR_MAYBE_NULL) {
12074 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
12075 dst, reg_type_str(env, ptr_reg->type));
12079 switch (base_type(ptr_reg->type)) {
12080 case CONST_PTR_TO_MAP:
12081 /* smin_val represents the known value */
12082 if (known && smin_val == 0 && opcode == BPF_ADD)
12085 case PTR_TO_PACKET_END:
12086 case PTR_TO_SOCKET:
12087 case PTR_TO_SOCK_COMMON:
12088 case PTR_TO_TCP_SOCK:
12089 case PTR_TO_XDP_SOCK:
12090 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
12091 dst, reg_type_str(env, ptr_reg->type));
12097 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
12098 * The id may be overwritten later if we create a new variable offset.
12100 dst_reg->type = ptr_reg->type;
12101 dst_reg->id = ptr_reg->id;
12103 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
12104 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
12107 /* pointer types do not carry 32-bit bounds at the moment. */
12108 __mark_reg32_unbounded(dst_reg);
12110 if (sanitize_needed(opcode)) {
12111 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
12114 return sanitize_err(env, insn, ret, off_reg, dst_reg);
12119 /* We can take a fixed offset as long as it doesn't overflow
12120 * the s32 'off' field
12122 if (known && (ptr_reg->off + smin_val ==
12123 (s64)(s32)(ptr_reg->off + smin_val))) {
12124 /* pointer += K. Accumulate it into fixed offset */
12125 dst_reg->smin_value = smin_ptr;
12126 dst_reg->smax_value = smax_ptr;
12127 dst_reg->umin_value = umin_ptr;
12128 dst_reg->umax_value = umax_ptr;
12129 dst_reg->var_off = ptr_reg->var_off;
12130 dst_reg->off = ptr_reg->off + smin_val;
12131 dst_reg->raw = ptr_reg->raw;
12134 /* A new variable offset is created. Note that off_reg->off
12135 * == 0, since it's a scalar.
12136 * dst_reg gets the pointer type and since some positive
12137 * integer value was added to the pointer, give it a new 'id'
12138 * if it's a PTR_TO_PACKET.
12139 * this creates a new 'base' pointer, off_reg (variable) gets
12140 * added into the variable offset, and we copy the fixed offset
12143 if (signed_add_overflows(smin_ptr, smin_val) ||
12144 signed_add_overflows(smax_ptr, smax_val)) {
12145 dst_reg->smin_value = S64_MIN;
12146 dst_reg->smax_value = S64_MAX;
12148 dst_reg->smin_value = smin_ptr + smin_val;
12149 dst_reg->smax_value = smax_ptr + smax_val;
12151 if (umin_ptr + umin_val < umin_ptr ||
12152 umax_ptr + umax_val < umax_ptr) {
12153 dst_reg->umin_value = 0;
12154 dst_reg->umax_value = U64_MAX;
12156 dst_reg->umin_value = umin_ptr + umin_val;
12157 dst_reg->umax_value = umax_ptr + umax_val;
12159 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
12160 dst_reg->off = ptr_reg->off;
12161 dst_reg->raw = ptr_reg->raw;
12162 if (reg_is_pkt_pointer(ptr_reg)) {
12163 dst_reg->id = ++env->id_gen;
12164 /* something was added to pkt_ptr, set range to zero */
12165 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
12169 if (dst_reg == off_reg) {
12170 /* scalar -= pointer. Creates an unknown scalar */
12171 verbose(env, "R%d tried to subtract pointer from scalar\n",
12175 /* We don't allow subtraction from FP, because (according to
12176 * test_verifier.c test "invalid fp arithmetic", JITs might not
12177 * be able to deal with it.
12179 if (ptr_reg->type == PTR_TO_STACK) {
12180 verbose(env, "R%d subtraction from stack pointer prohibited\n",
12184 if (known && (ptr_reg->off - smin_val ==
12185 (s64)(s32)(ptr_reg->off - smin_val))) {
12186 /* pointer -= K. Subtract it from fixed offset */
12187 dst_reg->smin_value = smin_ptr;
12188 dst_reg->smax_value = smax_ptr;
12189 dst_reg->umin_value = umin_ptr;
12190 dst_reg->umax_value = umax_ptr;
12191 dst_reg->var_off = ptr_reg->var_off;
12192 dst_reg->id = ptr_reg->id;
12193 dst_reg->off = ptr_reg->off - smin_val;
12194 dst_reg->raw = ptr_reg->raw;
12197 /* A new variable offset is created. If the subtrahend is known
12198 * nonnegative, then any reg->range we had before is still good.
12200 if (signed_sub_overflows(smin_ptr, smax_val) ||
12201 signed_sub_overflows(smax_ptr, smin_val)) {
12202 /* Overflow possible, we know nothing */
12203 dst_reg->smin_value = S64_MIN;
12204 dst_reg->smax_value = S64_MAX;
12206 dst_reg->smin_value = smin_ptr - smax_val;
12207 dst_reg->smax_value = smax_ptr - smin_val;
12209 if (umin_ptr < umax_val) {
12210 /* Overflow possible, we know nothing */
12211 dst_reg->umin_value = 0;
12212 dst_reg->umax_value = U64_MAX;
12214 /* Cannot overflow (as long as bounds are consistent) */
12215 dst_reg->umin_value = umin_ptr - umax_val;
12216 dst_reg->umax_value = umax_ptr - umin_val;
12218 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
12219 dst_reg->off = ptr_reg->off;
12220 dst_reg->raw = ptr_reg->raw;
12221 if (reg_is_pkt_pointer(ptr_reg)) {
12222 dst_reg->id = ++env->id_gen;
12223 /* something was added to pkt_ptr, set range to zero */
12225 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
12231 /* bitwise ops on pointers are troublesome, prohibit. */
12232 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
12233 dst, bpf_alu_string[opcode >> 4]);
12236 /* other operators (e.g. MUL,LSH) produce non-pointer results */
12237 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
12238 dst, bpf_alu_string[opcode >> 4]);
12242 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
12244 reg_bounds_sync(dst_reg);
12245 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
12247 if (sanitize_needed(opcode)) {
12248 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
12251 return sanitize_err(env, insn, ret, off_reg, dst_reg);
12257 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
12258 struct bpf_reg_state *src_reg)
12260 s32 smin_val = src_reg->s32_min_value;
12261 s32 smax_val = src_reg->s32_max_value;
12262 u32 umin_val = src_reg->u32_min_value;
12263 u32 umax_val = src_reg->u32_max_value;
12265 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
12266 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
12267 dst_reg->s32_min_value = S32_MIN;
12268 dst_reg->s32_max_value = S32_MAX;
12270 dst_reg->s32_min_value += smin_val;
12271 dst_reg->s32_max_value += smax_val;
12273 if (dst_reg->u32_min_value + umin_val < umin_val ||
12274 dst_reg->u32_max_value + umax_val < umax_val) {
12275 dst_reg->u32_min_value = 0;
12276 dst_reg->u32_max_value = U32_MAX;
12278 dst_reg->u32_min_value += umin_val;
12279 dst_reg->u32_max_value += umax_val;
12283 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
12284 struct bpf_reg_state *src_reg)
12286 s64 smin_val = src_reg->smin_value;
12287 s64 smax_val = src_reg->smax_value;
12288 u64 umin_val = src_reg->umin_value;
12289 u64 umax_val = src_reg->umax_value;
12291 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
12292 signed_add_overflows(dst_reg->smax_value, smax_val)) {
12293 dst_reg->smin_value = S64_MIN;
12294 dst_reg->smax_value = S64_MAX;
12296 dst_reg->smin_value += smin_val;
12297 dst_reg->smax_value += smax_val;
12299 if (dst_reg->umin_value + umin_val < umin_val ||
12300 dst_reg->umax_value + umax_val < umax_val) {
12301 dst_reg->umin_value = 0;
12302 dst_reg->umax_value = U64_MAX;
12304 dst_reg->umin_value += umin_val;
12305 dst_reg->umax_value += umax_val;
12309 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
12310 struct bpf_reg_state *src_reg)
12312 s32 smin_val = src_reg->s32_min_value;
12313 s32 smax_val = src_reg->s32_max_value;
12314 u32 umin_val = src_reg->u32_min_value;
12315 u32 umax_val = src_reg->u32_max_value;
12317 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
12318 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
12319 /* Overflow possible, we know nothing */
12320 dst_reg->s32_min_value = S32_MIN;
12321 dst_reg->s32_max_value = S32_MAX;
12323 dst_reg->s32_min_value -= smax_val;
12324 dst_reg->s32_max_value -= smin_val;
12326 if (dst_reg->u32_min_value < umax_val) {
12327 /* Overflow possible, we know nothing */
12328 dst_reg->u32_min_value = 0;
12329 dst_reg->u32_max_value = U32_MAX;
12331 /* Cannot overflow (as long as bounds are consistent) */
12332 dst_reg->u32_min_value -= umax_val;
12333 dst_reg->u32_max_value -= umin_val;
12337 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
12338 struct bpf_reg_state *src_reg)
12340 s64 smin_val = src_reg->smin_value;
12341 s64 smax_val = src_reg->smax_value;
12342 u64 umin_val = src_reg->umin_value;
12343 u64 umax_val = src_reg->umax_value;
12345 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
12346 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
12347 /* Overflow possible, we know nothing */
12348 dst_reg->smin_value = S64_MIN;
12349 dst_reg->smax_value = S64_MAX;
12351 dst_reg->smin_value -= smax_val;
12352 dst_reg->smax_value -= smin_val;
12354 if (dst_reg->umin_value < umax_val) {
12355 /* Overflow possible, we know nothing */
12356 dst_reg->umin_value = 0;
12357 dst_reg->umax_value = U64_MAX;
12359 /* Cannot overflow (as long as bounds are consistent) */
12360 dst_reg->umin_value -= umax_val;
12361 dst_reg->umax_value -= umin_val;
12365 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
12366 struct bpf_reg_state *src_reg)
12368 s32 smin_val = src_reg->s32_min_value;
12369 u32 umin_val = src_reg->u32_min_value;
12370 u32 umax_val = src_reg->u32_max_value;
12372 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
12373 /* Ain't nobody got time to multiply that sign */
12374 __mark_reg32_unbounded(dst_reg);
12377 /* Both values are positive, so we can work with unsigned and
12378 * copy the result to signed (unless it exceeds S32_MAX).
12380 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
12381 /* Potential overflow, we know nothing */
12382 __mark_reg32_unbounded(dst_reg);
12385 dst_reg->u32_min_value *= umin_val;
12386 dst_reg->u32_max_value *= umax_val;
12387 if (dst_reg->u32_max_value > S32_MAX) {
12388 /* Overflow possible, we know nothing */
12389 dst_reg->s32_min_value = S32_MIN;
12390 dst_reg->s32_max_value = S32_MAX;
12392 dst_reg->s32_min_value = dst_reg->u32_min_value;
12393 dst_reg->s32_max_value = dst_reg->u32_max_value;
12397 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
12398 struct bpf_reg_state *src_reg)
12400 s64 smin_val = src_reg->smin_value;
12401 u64 umin_val = src_reg->umin_value;
12402 u64 umax_val = src_reg->umax_value;
12404 if (smin_val < 0 || dst_reg->smin_value < 0) {
12405 /* Ain't nobody got time to multiply that sign */
12406 __mark_reg64_unbounded(dst_reg);
12409 /* Both values are positive, so we can work with unsigned and
12410 * copy the result to signed (unless it exceeds S64_MAX).
12412 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
12413 /* Potential overflow, we know nothing */
12414 __mark_reg64_unbounded(dst_reg);
12417 dst_reg->umin_value *= umin_val;
12418 dst_reg->umax_value *= umax_val;
12419 if (dst_reg->umax_value > S64_MAX) {
12420 /* Overflow possible, we know nothing */
12421 dst_reg->smin_value = S64_MIN;
12422 dst_reg->smax_value = S64_MAX;
12424 dst_reg->smin_value = dst_reg->umin_value;
12425 dst_reg->smax_value = dst_reg->umax_value;
12429 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
12430 struct bpf_reg_state *src_reg)
12432 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12433 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12434 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12435 s32 smin_val = src_reg->s32_min_value;
12436 u32 umax_val = src_reg->u32_max_value;
12438 if (src_known && dst_known) {
12439 __mark_reg32_known(dst_reg, var32_off.value);
12443 /* We get our minimum from the var_off, since that's inherently
12444 * bitwise. Our maximum is the minimum of the operands' maxima.
12446 dst_reg->u32_min_value = var32_off.value;
12447 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
12448 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
12449 /* Lose signed bounds when ANDing negative numbers,
12450 * ain't nobody got time for that.
12452 dst_reg->s32_min_value = S32_MIN;
12453 dst_reg->s32_max_value = S32_MAX;
12455 /* ANDing two positives gives a positive, so safe to
12456 * cast result into s64.
12458 dst_reg->s32_min_value = dst_reg->u32_min_value;
12459 dst_reg->s32_max_value = dst_reg->u32_max_value;
12463 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
12464 struct bpf_reg_state *src_reg)
12466 bool src_known = tnum_is_const(src_reg->var_off);
12467 bool dst_known = tnum_is_const(dst_reg->var_off);
12468 s64 smin_val = src_reg->smin_value;
12469 u64 umax_val = src_reg->umax_value;
12471 if (src_known && dst_known) {
12472 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12476 /* We get our minimum from the var_off, since that's inherently
12477 * bitwise. Our maximum is the minimum of the operands' maxima.
12479 dst_reg->umin_value = dst_reg->var_off.value;
12480 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
12481 if (dst_reg->smin_value < 0 || smin_val < 0) {
12482 /* Lose signed bounds when ANDing negative numbers,
12483 * ain't nobody got time for that.
12485 dst_reg->smin_value = S64_MIN;
12486 dst_reg->smax_value = S64_MAX;
12488 /* ANDing two positives gives a positive, so safe to
12489 * cast result into s64.
12491 dst_reg->smin_value = dst_reg->umin_value;
12492 dst_reg->smax_value = dst_reg->umax_value;
12494 /* We may learn something more from the var_off */
12495 __update_reg_bounds(dst_reg);
12498 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
12499 struct bpf_reg_state *src_reg)
12501 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12502 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12503 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12504 s32 smin_val = src_reg->s32_min_value;
12505 u32 umin_val = src_reg->u32_min_value;
12507 if (src_known && dst_known) {
12508 __mark_reg32_known(dst_reg, var32_off.value);
12512 /* We get our maximum from the var_off, and our minimum is the
12513 * maximum of the operands' minima
12515 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
12516 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
12517 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
12518 /* Lose signed bounds when ORing negative numbers,
12519 * ain't nobody got time for that.
12521 dst_reg->s32_min_value = S32_MIN;
12522 dst_reg->s32_max_value = S32_MAX;
12524 /* ORing two positives gives a positive, so safe to
12525 * cast result into s64.
12527 dst_reg->s32_min_value = dst_reg->u32_min_value;
12528 dst_reg->s32_max_value = dst_reg->u32_max_value;
12532 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
12533 struct bpf_reg_state *src_reg)
12535 bool src_known = tnum_is_const(src_reg->var_off);
12536 bool dst_known = tnum_is_const(dst_reg->var_off);
12537 s64 smin_val = src_reg->smin_value;
12538 u64 umin_val = src_reg->umin_value;
12540 if (src_known && dst_known) {
12541 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12545 /* We get our maximum from the var_off, and our minimum is the
12546 * maximum of the operands' minima
12548 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
12549 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
12550 if (dst_reg->smin_value < 0 || smin_val < 0) {
12551 /* Lose signed bounds when ORing negative numbers,
12552 * ain't nobody got time for that.
12554 dst_reg->smin_value = S64_MIN;
12555 dst_reg->smax_value = S64_MAX;
12557 /* ORing two positives gives a positive, so safe to
12558 * cast result into s64.
12560 dst_reg->smin_value = dst_reg->umin_value;
12561 dst_reg->smax_value = dst_reg->umax_value;
12563 /* We may learn something more from the var_off */
12564 __update_reg_bounds(dst_reg);
12567 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
12568 struct bpf_reg_state *src_reg)
12570 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12571 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12572 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12573 s32 smin_val = src_reg->s32_min_value;
12575 if (src_known && dst_known) {
12576 __mark_reg32_known(dst_reg, var32_off.value);
12580 /* We get both minimum and maximum from the var32_off. */
12581 dst_reg->u32_min_value = var32_off.value;
12582 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
12584 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
12585 /* XORing two positive sign numbers gives a positive,
12586 * so safe to cast u32 result into s32.
12588 dst_reg->s32_min_value = dst_reg->u32_min_value;
12589 dst_reg->s32_max_value = dst_reg->u32_max_value;
12591 dst_reg->s32_min_value = S32_MIN;
12592 dst_reg->s32_max_value = S32_MAX;
12596 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
12597 struct bpf_reg_state *src_reg)
12599 bool src_known = tnum_is_const(src_reg->var_off);
12600 bool dst_known = tnum_is_const(dst_reg->var_off);
12601 s64 smin_val = src_reg->smin_value;
12603 if (src_known && dst_known) {
12604 /* dst_reg->var_off.value has been updated earlier */
12605 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12609 /* We get both minimum and maximum from the var_off. */
12610 dst_reg->umin_value = dst_reg->var_off.value;
12611 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
12613 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
12614 /* XORing two positive sign numbers gives a positive,
12615 * so safe to cast u64 result into s64.
12617 dst_reg->smin_value = dst_reg->umin_value;
12618 dst_reg->smax_value = dst_reg->umax_value;
12620 dst_reg->smin_value = S64_MIN;
12621 dst_reg->smax_value = S64_MAX;
12624 __update_reg_bounds(dst_reg);
12627 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
12628 u64 umin_val, u64 umax_val)
12630 /* We lose all sign bit information (except what we can pick
12633 dst_reg->s32_min_value = S32_MIN;
12634 dst_reg->s32_max_value = S32_MAX;
12635 /* If we might shift our top bit out, then we know nothing */
12636 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
12637 dst_reg->u32_min_value = 0;
12638 dst_reg->u32_max_value = U32_MAX;
12640 dst_reg->u32_min_value <<= umin_val;
12641 dst_reg->u32_max_value <<= umax_val;
12645 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
12646 struct bpf_reg_state *src_reg)
12648 u32 umax_val = src_reg->u32_max_value;
12649 u32 umin_val = src_reg->u32_min_value;
12650 /* u32 alu operation will zext upper bits */
12651 struct tnum subreg = tnum_subreg(dst_reg->var_off);
12653 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
12654 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
12655 /* Not required but being careful mark reg64 bounds as unknown so
12656 * that we are forced to pick them up from tnum and zext later and
12657 * if some path skips this step we are still safe.
12659 __mark_reg64_unbounded(dst_reg);
12660 __update_reg32_bounds(dst_reg);
12663 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
12664 u64 umin_val, u64 umax_val)
12666 /* Special case <<32 because it is a common compiler pattern to sign
12667 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
12668 * positive we know this shift will also be positive so we can track
12669 * bounds correctly. Otherwise we lose all sign bit information except
12670 * what we can pick up from var_off. Perhaps we can generalize this
12671 * later to shifts of any length.
12673 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
12674 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
12676 dst_reg->smax_value = S64_MAX;
12678 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
12679 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
12681 dst_reg->smin_value = S64_MIN;
12683 /* If we might shift our top bit out, then we know nothing */
12684 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
12685 dst_reg->umin_value = 0;
12686 dst_reg->umax_value = U64_MAX;
12688 dst_reg->umin_value <<= umin_val;
12689 dst_reg->umax_value <<= umax_val;
12693 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
12694 struct bpf_reg_state *src_reg)
12696 u64 umax_val = src_reg->umax_value;
12697 u64 umin_val = src_reg->umin_value;
12699 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
12700 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
12701 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
12703 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
12704 /* We may learn something more from the var_off */
12705 __update_reg_bounds(dst_reg);
12708 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
12709 struct bpf_reg_state *src_reg)
12711 struct tnum subreg = tnum_subreg(dst_reg->var_off);
12712 u32 umax_val = src_reg->u32_max_value;
12713 u32 umin_val = src_reg->u32_min_value;
12715 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
12716 * be negative, then either:
12717 * 1) src_reg might be zero, so the sign bit of the result is
12718 * unknown, so we lose our signed bounds
12719 * 2) it's known negative, thus the unsigned bounds capture the
12721 * 3) the signed bounds cross zero, so they tell us nothing
12723 * If the value in dst_reg is known nonnegative, then again the
12724 * unsigned bounds capture the signed bounds.
12725 * Thus, in all cases it suffices to blow away our signed bounds
12726 * and rely on inferring new ones from the unsigned bounds and
12727 * var_off of the result.
12729 dst_reg->s32_min_value = S32_MIN;
12730 dst_reg->s32_max_value = S32_MAX;
12732 dst_reg->var_off = tnum_rshift(subreg, umin_val);
12733 dst_reg->u32_min_value >>= umax_val;
12734 dst_reg->u32_max_value >>= umin_val;
12736 __mark_reg64_unbounded(dst_reg);
12737 __update_reg32_bounds(dst_reg);
12740 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
12741 struct bpf_reg_state *src_reg)
12743 u64 umax_val = src_reg->umax_value;
12744 u64 umin_val = src_reg->umin_value;
12746 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
12747 * be negative, then either:
12748 * 1) src_reg might be zero, so the sign bit of the result is
12749 * unknown, so we lose our signed bounds
12750 * 2) it's known negative, thus the unsigned bounds capture the
12752 * 3) the signed bounds cross zero, so they tell us nothing
12754 * If the value in dst_reg is known nonnegative, then again the
12755 * unsigned bounds capture the signed bounds.
12756 * Thus, in all cases it suffices to blow away our signed bounds
12757 * and rely on inferring new ones from the unsigned bounds and
12758 * var_off of the result.
12760 dst_reg->smin_value = S64_MIN;
12761 dst_reg->smax_value = S64_MAX;
12762 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
12763 dst_reg->umin_value >>= umax_val;
12764 dst_reg->umax_value >>= umin_val;
12766 /* Its not easy to operate on alu32 bounds here because it depends
12767 * on bits being shifted in. Take easy way out and mark unbounded
12768 * so we can recalculate later from tnum.
12770 __mark_reg32_unbounded(dst_reg);
12771 __update_reg_bounds(dst_reg);
12774 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
12775 struct bpf_reg_state *src_reg)
12777 u64 umin_val = src_reg->u32_min_value;
12779 /* Upon reaching here, src_known is true and
12780 * umax_val is equal to umin_val.
12782 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
12783 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
12785 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
12787 /* blow away the dst_reg umin_value/umax_value and rely on
12788 * dst_reg var_off to refine the result.
12790 dst_reg->u32_min_value = 0;
12791 dst_reg->u32_max_value = U32_MAX;
12793 __mark_reg64_unbounded(dst_reg);
12794 __update_reg32_bounds(dst_reg);
12797 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
12798 struct bpf_reg_state *src_reg)
12800 u64 umin_val = src_reg->umin_value;
12802 /* Upon reaching here, src_known is true and umax_val is equal
12805 dst_reg->smin_value >>= umin_val;
12806 dst_reg->smax_value >>= umin_val;
12808 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
12810 /* blow away the dst_reg umin_value/umax_value and rely on
12811 * dst_reg var_off to refine the result.
12813 dst_reg->umin_value = 0;
12814 dst_reg->umax_value = U64_MAX;
12816 /* Its not easy to operate on alu32 bounds here because it depends
12817 * on bits being shifted in from upper 32-bits. Take easy way out
12818 * and mark unbounded so we can recalculate later from tnum.
12820 __mark_reg32_unbounded(dst_reg);
12821 __update_reg_bounds(dst_reg);
12824 /* WARNING: This function does calculations on 64-bit values, but the actual
12825 * execution may occur on 32-bit values. Therefore, things like bitshifts
12826 * need extra checks in the 32-bit case.
12828 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
12829 struct bpf_insn *insn,
12830 struct bpf_reg_state *dst_reg,
12831 struct bpf_reg_state src_reg)
12833 struct bpf_reg_state *regs = cur_regs(env);
12834 u8 opcode = BPF_OP(insn->code);
12836 s64 smin_val, smax_val;
12837 u64 umin_val, umax_val;
12838 s32 s32_min_val, s32_max_val;
12839 u32 u32_min_val, u32_max_val;
12840 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
12841 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
12844 smin_val = src_reg.smin_value;
12845 smax_val = src_reg.smax_value;
12846 umin_val = src_reg.umin_value;
12847 umax_val = src_reg.umax_value;
12849 s32_min_val = src_reg.s32_min_value;
12850 s32_max_val = src_reg.s32_max_value;
12851 u32_min_val = src_reg.u32_min_value;
12852 u32_max_val = src_reg.u32_max_value;
12855 src_known = tnum_subreg_is_const(src_reg.var_off);
12857 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
12858 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
12859 /* Taint dst register if offset had invalid bounds
12860 * derived from e.g. dead branches.
12862 __mark_reg_unknown(env, dst_reg);
12866 src_known = tnum_is_const(src_reg.var_off);
12868 (smin_val != smax_val || umin_val != umax_val)) ||
12869 smin_val > smax_val || umin_val > umax_val) {
12870 /* Taint dst register if offset had invalid bounds
12871 * derived from e.g. dead branches.
12873 __mark_reg_unknown(env, dst_reg);
12879 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
12880 __mark_reg_unknown(env, dst_reg);
12884 if (sanitize_needed(opcode)) {
12885 ret = sanitize_val_alu(env, insn);
12887 return sanitize_err(env, insn, ret, NULL, NULL);
12890 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
12891 * There are two classes of instructions: The first class we track both
12892 * alu32 and alu64 sign/unsigned bounds independently this provides the
12893 * greatest amount of precision when alu operations are mixed with jmp32
12894 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
12895 * and BPF_OR. This is possible because these ops have fairly easy to
12896 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
12897 * See alu32 verifier tests for examples. The second class of
12898 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
12899 * with regards to tracking sign/unsigned bounds because the bits may
12900 * cross subreg boundaries in the alu64 case. When this happens we mark
12901 * the reg unbounded in the subreg bound space and use the resulting
12902 * tnum to calculate an approximation of the sign/unsigned bounds.
12906 scalar32_min_max_add(dst_reg, &src_reg);
12907 scalar_min_max_add(dst_reg, &src_reg);
12908 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
12911 scalar32_min_max_sub(dst_reg, &src_reg);
12912 scalar_min_max_sub(dst_reg, &src_reg);
12913 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
12916 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
12917 scalar32_min_max_mul(dst_reg, &src_reg);
12918 scalar_min_max_mul(dst_reg, &src_reg);
12921 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
12922 scalar32_min_max_and(dst_reg, &src_reg);
12923 scalar_min_max_and(dst_reg, &src_reg);
12926 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
12927 scalar32_min_max_or(dst_reg, &src_reg);
12928 scalar_min_max_or(dst_reg, &src_reg);
12931 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
12932 scalar32_min_max_xor(dst_reg, &src_reg);
12933 scalar_min_max_xor(dst_reg, &src_reg);
12936 if (umax_val >= insn_bitness) {
12937 /* Shifts greater than 31 or 63 are undefined.
12938 * This includes shifts by a negative number.
12940 mark_reg_unknown(env, regs, insn->dst_reg);
12944 scalar32_min_max_lsh(dst_reg, &src_reg);
12946 scalar_min_max_lsh(dst_reg, &src_reg);
12949 if (umax_val >= insn_bitness) {
12950 /* Shifts greater than 31 or 63 are undefined.
12951 * This includes shifts by a negative number.
12953 mark_reg_unknown(env, regs, insn->dst_reg);
12957 scalar32_min_max_rsh(dst_reg, &src_reg);
12959 scalar_min_max_rsh(dst_reg, &src_reg);
12962 if (umax_val >= insn_bitness) {
12963 /* Shifts greater than 31 or 63 are undefined.
12964 * This includes shifts by a negative number.
12966 mark_reg_unknown(env, regs, insn->dst_reg);
12970 scalar32_min_max_arsh(dst_reg, &src_reg);
12972 scalar_min_max_arsh(dst_reg, &src_reg);
12975 mark_reg_unknown(env, regs, insn->dst_reg);
12979 /* ALU32 ops are zero extended into 64bit register */
12981 zext_32_to_64(dst_reg);
12982 reg_bounds_sync(dst_reg);
12986 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
12989 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
12990 struct bpf_insn *insn)
12992 struct bpf_verifier_state *vstate = env->cur_state;
12993 struct bpf_func_state *state = vstate->frame[vstate->curframe];
12994 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
12995 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
12996 u8 opcode = BPF_OP(insn->code);
12999 dst_reg = ®s[insn->dst_reg];
13001 if (dst_reg->type != SCALAR_VALUE)
13004 /* Make sure ID is cleared otherwise dst_reg min/max could be
13005 * incorrectly propagated into other registers by find_equal_scalars()
13008 if (BPF_SRC(insn->code) == BPF_X) {
13009 src_reg = ®s[insn->src_reg];
13010 if (src_reg->type != SCALAR_VALUE) {
13011 if (dst_reg->type != SCALAR_VALUE) {
13012 /* Combining two pointers by any ALU op yields
13013 * an arbitrary scalar. Disallow all math except
13014 * pointer subtraction
13016 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
13017 mark_reg_unknown(env, regs, insn->dst_reg);
13020 verbose(env, "R%d pointer %s pointer prohibited\n",
13022 bpf_alu_string[opcode >> 4]);
13025 /* scalar += pointer
13026 * This is legal, but we have to reverse our
13027 * src/dest handling in computing the range
13029 err = mark_chain_precision(env, insn->dst_reg);
13032 return adjust_ptr_min_max_vals(env, insn,
13035 } else if (ptr_reg) {
13036 /* pointer += scalar */
13037 err = mark_chain_precision(env, insn->src_reg);
13040 return adjust_ptr_min_max_vals(env, insn,
13042 } else if (dst_reg->precise) {
13043 /* if dst_reg is precise, src_reg should be precise as well */
13044 err = mark_chain_precision(env, insn->src_reg);
13049 /* Pretend the src is a reg with a known value, since we only
13050 * need to be able to read from this state.
13052 off_reg.type = SCALAR_VALUE;
13053 __mark_reg_known(&off_reg, insn->imm);
13054 src_reg = &off_reg;
13055 if (ptr_reg) /* pointer += K */
13056 return adjust_ptr_min_max_vals(env, insn,
13060 /* Got here implies adding two SCALAR_VALUEs */
13061 if (WARN_ON_ONCE(ptr_reg)) {
13062 print_verifier_state(env, state, true);
13063 verbose(env, "verifier internal error: unexpected ptr_reg\n");
13066 if (WARN_ON(!src_reg)) {
13067 print_verifier_state(env, state, true);
13068 verbose(env, "verifier internal error: no src_reg\n");
13071 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
13074 /* check validity of 32-bit and 64-bit arithmetic operations */
13075 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
13077 struct bpf_reg_state *regs = cur_regs(env);
13078 u8 opcode = BPF_OP(insn->code);
13081 if (opcode == BPF_END || opcode == BPF_NEG) {
13082 if (opcode == BPF_NEG) {
13083 if (BPF_SRC(insn->code) != BPF_K ||
13084 insn->src_reg != BPF_REG_0 ||
13085 insn->off != 0 || insn->imm != 0) {
13086 verbose(env, "BPF_NEG uses reserved fields\n");
13090 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
13091 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
13092 (BPF_CLASS(insn->code) == BPF_ALU64 &&
13093 BPF_SRC(insn->code) != BPF_TO_LE)) {
13094 verbose(env, "BPF_END uses reserved fields\n");
13099 /* check src operand */
13100 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
13104 if (is_pointer_value(env, insn->dst_reg)) {
13105 verbose(env, "R%d pointer arithmetic prohibited\n",
13110 /* check dest operand */
13111 err = check_reg_arg(env, insn->dst_reg, DST_OP);
13115 } else if (opcode == BPF_MOV) {
13117 if (BPF_SRC(insn->code) == BPF_X) {
13118 if (insn->imm != 0) {
13119 verbose(env, "BPF_MOV uses reserved fields\n");
13123 if (BPF_CLASS(insn->code) == BPF_ALU) {
13124 if (insn->off != 0 && insn->off != 8 && insn->off != 16) {
13125 verbose(env, "BPF_MOV uses reserved fields\n");
13129 if (insn->off != 0 && insn->off != 8 && insn->off != 16 &&
13131 verbose(env, "BPF_MOV uses reserved fields\n");
13136 /* check src operand */
13137 err = check_reg_arg(env, insn->src_reg, SRC_OP);
13141 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
13142 verbose(env, "BPF_MOV uses reserved fields\n");
13147 /* check dest operand, mark as required later */
13148 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
13152 if (BPF_SRC(insn->code) == BPF_X) {
13153 struct bpf_reg_state *src_reg = regs + insn->src_reg;
13154 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
13155 bool need_id = src_reg->type == SCALAR_VALUE && !src_reg->id &&
13156 !tnum_is_const(src_reg->var_off);
13158 if (BPF_CLASS(insn->code) == BPF_ALU64) {
13159 if (insn->off == 0) {
13161 * copy register state to dest reg
13164 /* Assign src and dst registers the same ID
13165 * that will be used by find_equal_scalars()
13166 * to propagate min/max range.
13168 src_reg->id = ++env->id_gen;
13169 copy_register_state(dst_reg, src_reg);
13170 dst_reg->live |= REG_LIVE_WRITTEN;
13171 dst_reg->subreg_def = DEF_NOT_SUBREG;
13173 /* case: R1 = (s8, s16 s32)R2 */
13174 if (is_pointer_value(env, insn->src_reg)) {
13176 "R%d sign-extension part of pointer\n",
13179 } else if (src_reg->type == SCALAR_VALUE) {
13182 no_sext = src_reg->umax_value < (1ULL << (insn->off - 1));
13183 if (no_sext && need_id)
13184 src_reg->id = ++env->id_gen;
13185 copy_register_state(dst_reg, src_reg);
13188 coerce_reg_to_size_sx(dst_reg, insn->off >> 3);
13189 dst_reg->live |= REG_LIVE_WRITTEN;
13190 dst_reg->subreg_def = DEF_NOT_SUBREG;
13192 mark_reg_unknown(env, regs, insn->dst_reg);
13196 /* R1 = (u32) R2 */
13197 if (is_pointer_value(env, insn->src_reg)) {
13199 "R%d partial copy of pointer\n",
13202 } else if (src_reg->type == SCALAR_VALUE) {
13203 if (insn->off == 0) {
13204 bool is_src_reg_u32 = src_reg->umax_value <= U32_MAX;
13206 if (is_src_reg_u32 && need_id)
13207 src_reg->id = ++env->id_gen;
13208 copy_register_state(dst_reg, src_reg);
13209 /* Make sure ID is cleared if src_reg is not in u32
13210 * range otherwise dst_reg min/max could be incorrectly
13211 * propagated into src_reg by find_equal_scalars()
13213 if (!is_src_reg_u32)
13215 dst_reg->live |= REG_LIVE_WRITTEN;
13216 dst_reg->subreg_def = env->insn_idx + 1;
13218 /* case: W1 = (s8, s16)W2 */
13219 bool no_sext = src_reg->umax_value < (1ULL << (insn->off - 1));
13221 if (no_sext && need_id)
13222 src_reg->id = ++env->id_gen;
13223 copy_register_state(dst_reg, src_reg);
13226 dst_reg->live |= REG_LIVE_WRITTEN;
13227 dst_reg->subreg_def = env->insn_idx + 1;
13228 coerce_subreg_to_size_sx(dst_reg, insn->off >> 3);
13231 mark_reg_unknown(env, regs,
13234 zext_32_to_64(dst_reg);
13235 reg_bounds_sync(dst_reg);
13239 * remember the value we stored into this reg
13241 /* clear any state __mark_reg_known doesn't set */
13242 mark_reg_unknown(env, regs, insn->dst_reg);
13243 regs[insn->dst_reg].type = SCALAR_VALUE;
13244 if (BPF_CLASS(insn->code) == BPF_ALU64) {
13245 __mark_reg_known(regs + insn->dst_reg,
13248 __mark_reg_known(regs + insn->dst_reg,
13253 } else if (opcode > BPF_END) {
13254 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
13257 } else { /* all other ALU ops: and, sub, xor, add, ... */
13259 if (BPF_SRC(insn->code) == BPF_X) {
13260 if (insn->imm != 0 || insn->off > 1 ||
13261 (insn->off == 1 && opcode != BPF_MOD && opcode != BPF_DIV)) {
13262 verbose(env, "BPF_ALU uses reserved fields\n");
13265 /* check src1 operand */
13266 err = check_reg_arg(env, insn->src_reg, SRC_OP);
13270 if (insn->src_reg != BPF_REG_0 || insn->off > 1 ||
13271 (insn->off == 1 && opcode != BPF_MOD && opcode != BPF_DIV)) {
13272 verbose(env, "BPF_ALU uses reserved fields\n");
13277 /* check src2 operand */
13278 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
13282 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
13283 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
13284 verbose(env, "div by zero\n");
13288 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
13289 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
13290 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
13292 if (insn->imm < 0 || insn->imm >= size) {
13293 verbose(env, "invalid shift %d\n", insn->imm);
13298 /* check dest operand */
13299 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
13303 return adjust_reg_min_max_vals(env, insn);
13309 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
13310 struct bpf_reg_state *dst_reg,
13311 enum bpf_reg_type type,
13312 bool range_right_open)
13314 struct bpf_func_state *state;
13315 struct bpf_reg_state *reg;
13318 if (dst_reg->off < 0 ||
13319 (dst_reg->off == 0 && range_right_open))
13320 /* This doesn't give us any range */
13323 if (dst_reg->umax_value > MAX_PACKET_OFF ||
13324 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
13325 /* Risk of overflow. For instance, ptr + (1<<63) may be less
13326 * than pkt_end, but that's because it's also less than pkt.
13330 new_range = dst_reg->off;
13331 if (range_right_open)
13334 /* Examples for register markings:
13336 * pkt_data in dst register:
13340 * if (r2 > pkt_end) goto <handle exception>
13345 * if (r2 < pkt_end) goto <access okay>
13346 * <handle exception>
13349 * r2 == dst_reg, pkt_end == src_reg
13350 * r2=pkt(id=n,off=8,r=0)
13351 * r3=pkt(id=n,off=0,r=0)
13353 * pkt_data in src register:
13357 * if (pkt_end >= r2) goto <access okay>
13358 * <handle exception>
13362 * if (pkt_end <= r2) goto <handle exception>
13366 * pkt_end == dst_reg, r2 == src_reg
13367 * r2=pkt(id=n,off=8,r=0)
13368 * r3=pkt(id=n,off=0,r=0)
13370 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
13371 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
13372 * and [r3, r3 + 8-1) respectively is safe to access depending on
13376 /* If our ids match, then we must have the same max_value. And we
13377 * don't care about the other reg's fixed offset, since if it's too big
13378 * the range won't allow anything.
13379 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
13381 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
13382 if (reg->type == type && reg->id == dst_reg->id)
13383 /* keep the maximum range already checked */
13384 reg->range = max(reg->range, new_range);
13388 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
13390 struct tnum subreg = tnum_subreg(reg->var_off);
13391 s32 sval = (s32)val;
13395 if (tnum_is_const(subreg))
13396 return !!tnum_equals_const(subreg, val);
13397 else if (val < reg->u32_min_value || val > reg->u32_max_value)
13401 if (tnum_is_const(subreg))
13402 return !tnum_equals_const(subreg, val);
13403 else if (val < reg->u32_min_value || val > reg->u32_max_value)
13407 if ((~subreg.mask & subreg.value) & val)
13409 if (!((subreg.mask | subreg.value) & val))
13413 if (reg->u32_min_value > val)
13415 else if (reg->u32_max_value <= val)
13419 if (reg->s32_min_value > sval)
13421 else if (reg->s32_max_value <= sval)
13425 if (reg->u32_max_value < val)
13427 else if (reg->u32_min_value >= val)
13431 if (reg->s32_max_value < sval)
13433 else if (reg->s32_min_value >= sval)
13437 if (reg->u32_min_value >= val)
13439 else if (reg->u32_max_value < val)
13443 if (reg->s32_min_value >= sval)
13445 else if (reg->s32_max_value < sval)
13449 if (reg->u32_max_value <= val)
13451 else if (reg->u32_min_value > val)
13455 if (reg->s32_max_value <= sval)
13457 else if (reg->s32_min_value > sval)
13466 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
13468 s64 sval = (s64)val;
13472 if (tnum_is_const(reg->var_off))
13473 return !!tnum_equals_const(reg->var_off, val);
13474 else if (val < reg->umin_value || val > reg->umax_value)
13478 if (tnum_is_const(reg->var_off))
13479 return !tnum_equals_const(reg->var_off, val);
13480 else if (val < reg->umin_value || val > reg->umax_value)
13484 if ((~reg->var_off.mask & reg->var_off.value) & val)
13486 if (!((reg->var_off.mask | reg->var_off.value) & val))
13490 if (reg->umin_value > val)
13492 else if (reg->umax_value <= val)
13496 if (reg->smin_value > sval)
13498 else if (reg->smax_value <= sval)
13502 if (reg->umax_value < val)
13504 else if (reg->umin_value >= val)
13508 if (reg->smax_value < sval)
13510 else if (reg->smin_value >= sval)
13514 if (reg->umin_value >= val)
13516 else if (reg->umax_value < val)
13520 if (reg->smin_value >= sval)
13522 else if (reg->smax_value < sval)
13526 if (reg->umax_value <= val)
13528 else if (reg->umin_value > val)
13532 if (reg->smax_value <= sval)
13534 else if (reg->smin_value > sval)
13542 /* compute branch direction of the expression "if (reg opcode val) goto target;"
13544 * 1 - branch will be taken and "goto target" will be executed
13545 * 0 - branch will not be taken and fall-through to next insn
13546 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
13549 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
13552 if (__is_pointer_value(false, reg)) {
13553 if (!reg_not_null(reg))
13556 /* If pointer is valid tests against zero will fail so we can
13557 * use this to direct branch taken.
13573 return is_branch32_taken(reg, val, opcode);
13574 return is_branch64_taken(reg, val, opcode);
13577 static int flip_opcode(u32 opcode)
13579 /* How can we transform "a <op> b" into "b <op> a"? */
13580 static const u8 opcode_flip[16] = {
13581 /* these stay the same */
13582 [BPF_JEQ >> 4] = BPF_JEQ,
13583 [BPF_JNE >> 4] = BPF_JNE,
13584 [BPF_JSET >> 4] = BPF_JSET,
13585 /* these swap "lesser" and "greater" (L and G in the opcodes) */
13586 [BPF_JGE >> 4] = BPF_JLE,
13587 [BPF_JGT >> 4] = BPF_JLT,
13588 [BPF_JLE >> 4] = BPF_JGE,
13589 [BPF_JLT >> 4] = BPF_JGT,
13590 [BPF_JSGE >> 4] = BPF_JSLE,
13591 [BPF_JSGT >> 4] = BPF_JSLT,
13592 [BPF_JSLE >> 4] = BPF_JSGE,
13593 [BPF_JSLT >> 4] = BPF_JSGT
13595 return opcode_flip[opcode >> 4];
13598 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
13599 struct bpf_reg_state *src_reg,
13602 struct bpf_reg_state *pkt;
13604 if (src_reg->type == PTR_TO_PACKET_END) {
13606 } else if (dst_reg->type == PTR_TO_PACKET_END) {
13608 opcode = flip_opcode(opcode);
13613 if (pkt->range >= 0)
13618 /* pkt <= pkt_end */
13621 /* pkt > pkt_end */
13622 if (pkt->range == BEYOND_PKT_END)
13623 /* pkt has at last one extra byte beyond pkt_end */
13624 return opcode == BPF_JGT;
13627 /* pkt < pkt_end */
13630 /* pkt >= pkt_end */
13631 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
13632 return opcode == BPF_JGE;
13638 /* Adjusts the register min/max values in the case that the dst_reg is the
13639 * variable register that we are working on, and src_reg is a constant or we're
13640 * simply doing a BPF_K check.
13641 * In JEQ/JNE cases we also adjust the var_off values.
13643 static void reg_set_min_max(struct bpf_reg_state *true_reg,
13644 struct bpf_reg_state *false_reg,
13645 u64 val, u32 val32,
13646 u8 opcode, bool is_jmp32)
13648 struct tnum false_32off = tnum_subreg(false_reg->var_off);
13649 struct tnum false_64off = false_reg->var_off;
13650 struct tnum true_32off = tnum_subreg(true_reg->var_off);
13651 struct tnum true_64off = true_reg->var_off;
13652 s64 sval = (s64)val;
13653 s32 sval32 = (s32)val32;
13655 /* If the dst_reg is a pointer, we can't learn anything about its
13656 * variable offset from the compare (unless src_reg were a pointer into
13657 * the same object, but we don't bother with that.
13658 * Since false_reg and true_reg have the same type by construction, we
13659 * only need to check one of them for pointerness.
13661 if (__is_pointer_value(false, false_reg))
13665 /* JEQ/JNE comparison doesn't change the register equivalence.
13668 * if (r1 == 42) goto label;
13670 * label: // here both r1 and r2 are known to be 42.
13672 * Hence when marking register as known preserve it's ID.
13676 __mark_reg32_known(true_reg, val32);
13677 true_32off = tnum_subreg(true_reg->var_off);
13679 ___mark_reg_known(true_reg, val);
13680 true_64off = true_reg->var_off;
13685 __mark_reg32_known(false_reg, val32);
13686 false_32off = tnum_subreg(false_reg->var_off);
13688 ___mark_reg_known(false_reg, val);
13689 false_64off = false_reg->var_off;
13694 false_32off = tnum_and(false_32off, tnum_const(~val32));
13695 if (is_power_of_2(val32))
13696 true_32off = tnum_or(true_32off,
13697 tnum_const(val32));
13699 false_64off = tnum_and(false_64off, tnum_const(~val));
13700 if (is_power_of_2(val))
13701 true_64off = tnum_or(true_64off,
13709 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
13710 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
13712 false_reg->u32_max_value = min(false_reg->u32_max_value,
13714 true_reg->u32_min_value = max(true_reg->u32_min_value,
13717 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
13718 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
13720 false_reg->umax_value = min(false_reg->umax_value, false_umax);
13721 true_reg->umin_value = max(true_reg->umin_value, true_umin);
13729 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
13730 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
13732 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
13733 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
13735 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
13736 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
13738 false_reg->smax_value = min(false_reg->smax_value, false_smax);
13739 true_reg->smin_value = max(true_reg->smin_value, true_smin);
13747 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
13748 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
13750 false_reg->u32_min_value = max(false_reg->u32_min_value,
13752 true_reg->u32_max_value = min(true_reg->u32_max_value,
13755 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
13756 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
13758 false_reg->umin_value = max(false_reg->umin_value, false_umin);
13759 true_reg->umax_value = min(true_reg->umax_value, true_umax);
13767 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
13768 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
13770 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
13771 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
13773 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
13774 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
13776 false_reg->smin_value = max(false_reg->smin_value, false_smin);
13777 true_reg->smax_value = min(true_reg->smax_value, true_smax);
13786 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
13787 tnum_subreg(false_32off));
13788 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
13789 tnum_subreg(true_32off));
13790 __reg_combine_32_into_64(false_reg);
13791 __reg_combine_32_into_64(true_reg);
13793 false_reg->var_off = false_64off;
13794 true_reg->var_off = true_64off;
13795 __reg_combine_64_into_32(false_reg);
13796 __reg_combine_64_into_32(true_reg);
13800 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
13801 * the variable reg.
13803 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
13804 struct bpf_reg_state *false_reg,
13805 u64 val, u32 val32,
13806 u8 opcode, bool is_jmp32)
13808 opcode = flip_opcode(opcode);
13809 /* This uses zero as "not present in table"; luckily the zero opcode,
13810 * BPF_JA, can't get here.
13813 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
13816 /* Regs are known to be equal, so intersect their min/max/var_off */
13817 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
13818 struct bpf_reg_state *dst_reg)
13820 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
13821 dst_reg->umin_value);
13822 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
13823 dst_reg->umax_value);
13824 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
13825 dst_reg->smin_value);
13826 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
13827 dst_reg->smax_value);
13828 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
13830 reg_bounds_sync(src_reg);
13831 reg_bounds_sync(dst_reg);
13834 static void reg_combine_min_max(struct bpf_reg_state *true_src,
13835 struct bpf_reg_state *true_dst,
13836 struct bpf_reg_state *false_src,
13837 struct bpf_reg_state *false_dst,
13842 __reg_combine_min_max(true_src, true_dst);
13845 __reg_combine_min_max(false_src, false_dst);
13850 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
13851 struct bpf_reg_state *reg, u32 id,
13854 if (type_may_be_null(reg->type) && reg->id == id &&
13855 (is_rcu_reg(reg) || !WARN_ON_ONCE(!reg->id))) {
13856 /* Old offset (both fixed and variable parts) should have been
13857 * known-zero, because we don't allow pointer arithmetic on
13858 * pointers that might be NULL. If we see this happening, don't
13859 * convert the register.
13861 * But in some cases, some helpers that return local kptrs
13862 * advance offset for the returned pointer. In those cases, it
13863 * is fine to expect to see reg->off.
13865 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value || !tnum_equals_const(reg->var_off, 0)))
13867 if (!(type_is_ptr_alloc_obj(reg->type) || type_is_non_owning_ref(reg->type)) &&
13868 WARN_ON_ONCE(reg->off))
13872 reg->type = SCALAR_VALUE;
13873 /* We don't need id and ref_obj_id from this point
13874 * onwards anymore, thus we should better reset it,
13875 * so that state pruning has chances to take effect.
13878 reg->ref_obj_id = 0;
13883 mark_ptr_not_null_reg(reg);
13885 if (!reg_may_point_to_spin_lock(reg)) {
13886 /* For not-NULL ptr, reg->ref_obj_id will be reset
13887 * in release_reference().
13889 * reg->id is still used by spin_lock ptr. Other
13890 * than spin_lock ptr type, reg->id can be reset.
13897 /* The logic is similar to find_good_pkt_pointers(), both could eventually
13898 * be folded together at some point.
13900 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
13903 struct bpf_func_state *state = vstate->frame[vstate->curframe];
13904 struct bpf_reg_state *regs = state->regs, *reg;
13905 u32 ref_obj_id = regs[regno].ref_obj_id;
13906 u32 id = regs[regno].id;
13908 if (ref_obj_id && ref_obj_id == id && is_null)
13909 /* regs[regno] is in the " == NULL" branch.
13910 * No one could have freed the reference state before
13911 * doing the NULL check.
13913 WARN_ON_ONCE(release_reference_state(state, id));
13915 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
13916 mark_ptr_or_null_reg(state, reg, id, is_null);
13920 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
13921 struct bpf_reg_state *dst_reg,
13922 struct bpf_reg_state *src_reg,
13923 struct bpf_verifier_state *this_branch,
13924 struct bpf_verifier_state *other_branch)
13926 if (BPF_SRC(insn->code) != BPF_X)
13929 /* Pointers are always 64-bit. */
13930 if (BPF_CLASS(insn->code) == BPF_JMP32)
13933 switch (BPF_OP(insn->code)) {
13935 if ((dst_reg->type == PTR_TO_PACKET &&
13936 src_reg->type == PTR_TO_PACKET_END) ||
13937 (dst_reg->type == PTR_TO_PACKET_META &&
13938 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13939 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
13940 find_good_pkt_pointers(this_branch, dst_reg,
13941 dst_reg->type, false);
13942 mark_pkt_end(other_branch, insn->dst_reg, true);
13943 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
13944 src_reg->type == PTR_TO_PACKET) ||
13945 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
13946 src_reg->type == PTR_TO_PACKET_META)) {
13947 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
13948 find_good_pkt_pointers(other_branch, src_reg,
13949 src_reg->type, true);
13950 mark_pkt_end(this_branch, insn->src_reg, false);
13956 if ((dst_reg->type == PTR_TO_PACKET &&
13957 src_reg->type == PTR_TO_PACKET_END) ||
13958 (dst_reg->type == PTR_TO_PACKET_META &&
13959 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13960 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
13961 find_good_pkt_pointers(other_branch, dst_reg,
13962 dst_reg->type, true);
13963 mark_pkt_end(this_branch, insn->dst_reg, false);
13964 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
13965 src_reg->type == PTR_TO_PACKET) ||
13966 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
13967 src_reg->type == PTR_TO_PACKET_META)) {
13968 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
13969 find_good_pkt_pointers(this_branch, src_reg,
13970 src_reg->type, false);
13971 mark_pkt_end(other_branch, insn->src_reg, true);
13977 if ((dst_reg->type == PTR_TO_PACKET &&
13978 src_reg->type == PTR_TO_PACKET_END) ||
13979 (dst_reg->type == PTR_TO_PACKET_META &&
13980 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13981 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
13982 find_good_pkt_pointers(this_branch, dst_reg,
13983 dst_reg->type, true);
13984 mark_pkt_end(other_branch, insn->dst_reg, false);
13985 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
13986 src_reg->type == PTR_TO_PACKET) ||
13987 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
13988 src_reg->type == PTR_TO_PACKET_META)) {
13989 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
13990 find_good_pkt_pointers(other_branch, src_reg,
13991 src_reg->type, false);
13992 mark_pkt_end(this_branch, insn->src_reg, true);
13998 if ((dst_reg->type == PTR_TO_PACKET &&
13999 src_reg->type == PTR_TO_PACKET_END) ||
14000 (dst_reg->type == PTR_TO_PACKET_META &&
14001 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
14002 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
14003 find_good_pkt_pointers(other_branch, dst_reg,
14004 dst_reg->type, false);
14005 mark_pkt_end(this_branch, insn->dst_reg, true);
14006 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
14007 src_reg->type == PTR_TO_PACKET) ||
14008 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
14009 src_reg->type == PTR_TO_PACKET_META)) {
14010 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
14011 find_good_pkt_pointers(this_branch, src_reg,
14012 src_reg->type, true);
14013 mark_pkt_end(other_branch, insn->src_reg, false);
14025 static void find_equal_scalars(struct bpf_verifier_state *vstate,
14026 struct bpf_reg_state *known_reg)
14028 struct bpf_func_state *state;
14029 struct bpf_reg_state *reg;
14031 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
14032 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
14033 copy_register_state(reg, known_reg);
14037 static int check_cond_jmp_op(struct bpf_verifier_env *env,
14038 struct bpf_insn *insn, int *insn_idx)
14040 struct bpf_verifier_state *this_branch = env->cur_state;
14041 struct bpf_verifier_state *other_branch;
14042 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
14043 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
14044 struct bpf_reg_state *eq_branch_regs;
14045 u8 opcode = BPF_OP(insn->code);
14050 /* Only conditional jumps are expected to reach here. */
14051 if (opcode == BPF_JA || opcode > BPF_JSLE) {
14052 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
14056 /* check src2 operand */
14057 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
14061 dst_reg = ®s[insn->dst_reg];
14062 if (BPF_SRC(insn->code) == BPF_X) {
14063 if (insn->imm != 0) {
14064 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
14068 /* check src1 operand */
14069 err = check_reg_arg(env, insn->src_reg, SRC_OP);
14073 src_reg = ®s[insn->src_reg];
14074 if (!(reg_is_pkt_pointer_any(dst_reg) && reg_is_pkt_pointer_any(src_reg)) &&
14075 is_pointer_value(env, insn->src_reg)) {
14076 verbose(env, "R%d pointer comparison prohibited\n",
14081 if (insn->src_reg != BPF_REG_0) {
14082 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
14087 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
14089 if (BPF_SRC(insn->code) == BPF_K) {
14090 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
14091 } else if (src_reg->type == SCALAR_VALUE &&
14092 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
14093 pred = is_branch_taken(dst_reg,
14094 tnum_subreg(src_reg->var_off).value,
14097 } else if (src_reg->type == SCALAR_VALUE &&
14098 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
14099 pred = is_branch_taken(dst_reg,
14100 src_reg->var_off.value,
14103 } else if (dst_reg->type == SCALAR_VALUE &&
14104 is_jmp32 && tnum_is_const(tnum_subreg(dst_reg->var_off))) {
14105 pred = is_branch_taken(src_reg,
14106 tnum_subreg(dst_reg->var_off).value,
14107 flip_opcode(opcode),
14109 } else if (dst_reg->type == SCALAR_VALUE &&
14110 !is_jmp32 && tnum_is_const(dst_reg->var_off)) {
14111 pred = is_branch_taken(src_reg,
14112 dst_reg->var_off.value,
14113 flip_opcode(opcode),
14115 } else if (reg_is_pkt_pointer_any(dst_reg) &&
14116 reg_is_pkt_pointer_any(src_reg) &&
14118 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
14122 /* If we get here with a dst_reg pointer type it is because
14123 * above is_branch_taken() special cased the 0 comparison.
14125 if (!__is_pointer_value(false, dst_reg))
14126 err = mark_chain_precision(env, insn->dst_reg);
14127 if (BPF_SRC(insn->code) == BPF_X && !err &&
14128 !__is_pointer_value(false, src_reg))
14129 err = mark_chain_precision(env, insn->src_reg);
14135 /* Only follow the goto, ignore fall-through. If needed, push
14136 * the fall-through branch for simulation under speculative
14139 if (!env->bypass_spec_v1 &&
14140 !sanitize_speculative_path(env, insn, *insn_idx + 1,
14143 if (env->log.level & BPF_LOG_LEVEL)
14144 print_insn_state(env, this_branch->frame[this_branch->curframe]);
14145 *insn_idx += insn->off;
14147 } else if (pred == 0) {
14148 /* Only follow the fall-through branch, since that's where the
14149 * program will go. If needed, push the goto branch for
14150 * simulation under speculative execution.
14152 if (!env->bypass_spec_v1 &&
14153 !sanitize_speculative_path(env, insn,
14154 *insn_idx + insn->off + 1,
14157 if (env->log.level & BPF_LOG_LEVEL)
14158 print_insn_state(env, this_branch->frame[this_branch->curframe]);
14162 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
14166 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
14168 /* detect if we are comparing against a constant value so we can adjust
14169 * our min/max values for our dst register.
14170 * this is only legit if both are scalars (or pointers to the same
14171 * object, I suppose, see the PTR_MAYBE_NULL related if block below),
14172 * because otherwise the different base pointers mean the offsets aren't
14175 if (BPF_SRC(insn->code) == BPF_X) {
14176 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
14178 if (dst_reg->type == SCALAR_VALUE &&
14179 src_reg->type == SCALAR_VALUE) {
14180 if (tnum_is_const(src_reg->var_off) ||
14182 tnum_is_const(tnum_subreg(src_reg->var_off))))
14183 reg_set_min_max(&other_branch_regs[insn->dst_reg],
14185 src_reg->var_off.value,
14186 tnum_subreg(src_reg->var_off).value,
14188 else if (tnum_is_const(dst_reg->var_off) ||
14190 tnum_is_const(tnum_subreg(dst_reg->var_off))))
14191 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
14193 dst_reg->var_off.value,
14194 tnum_subreg(dst_reg->var_off).value,
14196 else if (!is_jmp32 &&
14197 (opcode == BPF_JEQ || opcode == BPF_JNE))
14198 /* Comparing for equality, we can combine knowledge */
14199 reg_combine_min_max(&other_branch_regs[insn->src_reg],
14200 &other_branch_regs[insn->dst_reg],
14201 src_reg, dst_reg, opcode);
14203 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
14204 find_equal_scalars(this_branch, src_reg);
14205 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
14209 } else if (dst_reg->type == SCALAR_VALUE) {
14210 reg_set_min_max(&other_branch_regs[insn->dst_reg],
14211 dst_reg, insn->imm, (u32)insn->imm,
14215 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
14216 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
14217 find_equal_scalars(this_branch, dst_reg);
14218 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
14221 /* if one pointer register is compared to another pointer
14222 * register check if PTR_MAYBE_NULL could be lifted.
14223 * E.g. register A - maybe null
14224 * register B - not null
14225 * for JNE A, B, ... - A is not null in the false branch;
14226 * for JEQ A, B, ... - A is not null in the true branch.
14228 * Since PTR_TO_BTF_ID points to a kernel struct that does
14229 * not need to be null checked by the BPF program, i.e.,
14230 * could be null even without PTR_MAYBE_NULL marking, so
14231 * only propagate nullness when neither reg is that type.
14233 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_X &&
14234 __is_pointer_value(false, src_reg) && __is_pointer_value(false, dst_reg) &&
14235 type_may_be_null(src_reg->type) != type_may_be_null(dst_reg->type) &&
14236 base_type(src_reg->type) != PTR_TO_BTF_ID &&
14237 base_type(dst_reg->type) != PTR_TO_BTF_ID) {
14238 eq_branch_regs = NULL;
14241 eq_branch_regs = other_branch_regs;
14244 eq_branch_regs = regs;
14250 if (eq_branch_regs) {
14251 if (type_may_be_null(src_reg->type))
14252 mark_ptr_not_null_reg(&eq_branch_regs[insn->src_reg]);
14254 mark_ptr_not_null_reg(&eq_branch_regs[insn->dst_reg]);
14258 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
14259 * NOTE: these optimizations below are related with pointer comparison
14260 * which will never be JMP32.
14262 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
14263 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
14264 type_may_be_null(dst_reg->type)) {
14265 /* Mark all identical registers in each branch as either
14266 * safe or unknown depending R == 0 or R != 0 conditional.
14268 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
14269 opcode == BPF_JNE);
14270 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
14271 opcode == BPF_JEQ);
14272 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
14273 this_branch, other_branch) &&
14274 is_pointer_value(env, insn->dst_reg)) {
14275 verbose(env, "R%d pointer comparison prohibited\n",
14279 if (env->log.level & BPF_LOG_LEVEL)
14280 print_insn_state(env, this_branch->frame[this_branch->curframe]);
14284 /* verify BPF_LD_IMM64 instruction */
14285 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
14287 struct bpf_insn_aux_data *aux = cur_aux(env);
14288 struct bpf_reg_state *regs = cur_regs(env);
14289 struct bpf_reg_state *dst_reg;
14290 struct bpf_map *map;
14293 if (BPF_SIZE(insn->code) != BPF_DW) {
14294 verbose(env, "invalid BPF_LD_IMM insn\n");
14297 if (insn->off != 0) {
14298 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
14302 err = check_reg_arg(env, insn->dst_reg, DST_OP);
14306 dst_reg = ®s[insn->dst_reg];
14307 if (insn->src_reg == 0) {
14308 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
14310 dst_reg->type = SCALAR_VALUE;
14311 __mark_reg_known(®s[insn->dst_reg], imm);
14315 /* All special src_reg cases are listed below. From this point onwards
14316 * we either succeed and assign a corresponding dst_reg->type after
14317 * zeroing the offset, or fail and reject the program.
14319 mark_reg_known_zero(env, regs, insn->dst_reg);
14321 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
14322 dst_reg->type = aux->btf_var.reg_type;
14323 switch (base_type(dst_reg->type)) {
14325 dst_reg->mem_size = aux->btf_var.mem_size;
14327 case PTR_TO_BTF_ID:
14328 dst_reg->btf = aux->btf_var.btf;
14329 dst_reg->btf_id = aux->btf_var.btf_id;
14332 verbose(env, "bpf verifier is misconfigured\n");
14338 if (insn->src_reg == BPF_PSEUDO_FUNC) {
14339 struct bpf_prog_aux *aux = env->prog->aux;
14340 u32 subprogno = find_subprog(env,
14341 env->insn_idx + insn->imm + 1);
14343 if (!aux->func_info) {
14344 verbose(env, "missing btf func_info\n");
14347 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
14348 verbose(env, "callback function not static\n");
14352 dst_reg->type = PTR_TO_FUNC;
14353 dst_reg->subprogno = subprogno;
14357 map = env->used_maps[aux->map_index];
14358 dst_reg->map_ptr = map;
14360 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
14361 insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
14362 dst_reg->type = PTR_TO_MAP_VALUE;
14363 dst_reg->off = aux->map_off;
14364 WARN_ON_ONCE(map->max_entries != 1);
14365 /* We want reg->id to be same (0) as map_value is not distinct */
14366 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
14367 insn->src_reg == BPF_PSEUDO_MAP_IDX) {
14368 dst_reg->type = CONST_PTR_TO_MAP;
14370 verbose(env, "bpf verifier is misconfigured\n");
14377 static bool may_access_skb(enum bpf_prog_type type)
14380 case BPF_PROG_TYPE_SOCKET_FILTER:
14381 case BPF_PROG_TYPE_SCHED_CLS:
14382 case BPF_PROG_TYPE_SCHED_ACT:
14389 /* verify safety of LD_ABS|LD_IND instructions:
14390 * - they can only appear in the programs where ctx == skb
14391 * - since they are wrappers of function calls, they scratch R1-R5 registers,
14392 * preserve R6-R9, and store return value into R0
14395 * ctx == skb == R6 == CTX
14398 * SRC == any register
14399 * IMM == 32-bit immediate
14402 * R0 - 8/16/32-bit skb data converted to cpu endianness
14404 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
14406 struct bpf_reg_state *regs = cur_regs(env);
14407 static const int ctx_reg = BPF_REG_6;
14408 u8 mode = BPF_MODE(insn->code);
14411 if (!may_access_skb(resolve_prog_type(env->prog))) {
14412 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
14416 if (!env->ops->gen_ld_abs) {
14417 verbose(env, "bpf verifier is misconfigured\n");
14421 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
14422 BPF_SIZE(insn->code) == BPF_DW ||
14423 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
14424 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
14428 /* check whether implicit source operand (register R6) is readable */
14429 err = check_reg_arg(env, ctx_reg, SRC_OP);
14433 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
14434 * gen_ld_abs() may terminate the program at runtime, leading to
14437 err = check_reference_leak(env);
14439 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
14443 if (env->cur_state->active_lock.ptr) {
14444 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
14448 if (env->cur_state->active_rcu_lock) {
14449 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_rcu_read_lock-ed region\n");
14453 if (regs[ctx_reg].type != PTR_TO_CTX) {
14455 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
14459 if (mode == BPF_IND) {
14460 /* check explicit source operand */
14461 err = check_reg_arg(env, insn->src_reg, SRC_OP);
14466 err = check_ptr_off_reg(env, ®s[ctx_reg], ctx_reg);
14470 /* reset caller saved regs to unreadable */
14471 for (i = 0; i < CALLER_SAVED_REGS; i++) {
14472 mark_reg_not_init(env, regs, caller_saved[i]);
14473 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
14476 /* mark destination R0 register as readable, since it contains
14477 * the value fetched from the packet.
14478 * Already marked as written above.
14480 mark_reg_unknown(env, regs, BPF_REG_0);
14481 /* ld_abs load up to 32-bit skb data. */
14482 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
14486 static int check_return_code(struct bpf_verifier_env *env)
14488 struct tnum enforce_attach_type_range = tnum_unknown;
14489 const struct bpf_prog *prog = env->prog;
14490 struct bpf_reg_state *reg;
14491 struct tnum range = tnum_range(0, 1), const_0 = tnum_const(0);
14492 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
14494 struct bpf_func_state *frame = env->cur_state->frame[0];
14495 const bool is_subprog = frame->subprogno;
14497 /* LSM and struct_ops func-ptr's return type could be "void" */
14499 switch (prog_type) {
14500 case BPF_PROG_TYPE_LSM:
14501 if (prog->expected_attach_type == BPF_LSM_CGROUP)
14502 /* See below, can be 0 or 0-1 depending on hook. */
14505 case BPF_PROG_TYPE_STRUCT_OPS:
14506 if (!prog->aux->attach_func_proto->type)
14514 /* eBPF calling convention is such that R0 is used
14515 * to return the value from eBPF program.
14516 * Make sure that it's readable at this time
14517 * of bpf_exit, which means that program wrote
14518 * something into it earlier
14520 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
14524 if (is_pointer_value(env, BPF_REG_0)) {
14525 verbose(env, "R0 leaks addr as return value\n");
14529 reg = cur_regs(env) + BPF_REG_0;
14531 if (frame->in_async_callback_fn) {
14532 /* enforce return zero from async callbacks like timer */
14533 if (reg->type != SCALAR_VALUE) {
14534 verbose(env, "In async callback the register R0 is not a known value (%s)\n",
14535 reg_type_str(env, reg->type));
14539 if (!tnum_in(const_0, reg->var_off)) {
14540 verbose_invalid_scalar(env, reg, &const_0, "async callback", "R0");
14547 if (reg->type != SCALAR_VALUE) {
14548 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
14549 reg_type_str(env, reg->type));
14555 switch (prog_type) {
14556 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
14557 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
14558 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
14559 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
14560 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
14561 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
14562 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
14563 range = tnum_range(1, 1);
14564 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
14565 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
14566 range = tnum_range(0, 3);
14568 case BPF_PROG_TYPE_CGROUP_SKB:
14569 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
14570 range = tnum_range(0, 3);
14571 enforce_attach_type_range = tnum_range(2, 3);
14574 case BPF_PROG_TYPE_CGROUP_SOCK:
14575 case BPF_PROG_TYPE_SOCK_OPS:
14576 case BPF_PROG_TYPE_CGROUP_DEVICE:
14577 case BPF_PROG_TYPE_CGROUP_SYSCTL:
14578 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
14580 case BPF_PROG_TYPE_RAW_TRACEPOINT:
14581 if (!env->prog->aux->attach_btf_id)
14583 range = tnum_const(0);
14585 case BPF_PROG_TYPE_TRACING:
14586 switch (env->prog->expected_attach_type) {
14587 case BPF_TRACE_FENTRY:
14588 case BPF_TRACE_FEXIT:
14589 range = tnum_const(0);
14591 case BPF_TRACE_RAW_TP:
14592 case BPF_MODIFY_RETURN:
14594 case BPF_TRACE_ITER:
14600 case BPF_PROG_TYPE_SK_LOOKUP:
14601 range = tnum_range(SK_DROP, SK_PASS);
14604 case BPF_PROG_TYPE_LSM:
14605 if (env->prog->expected_attach_type != BPF_LSM_CGROUP) {
14606 /* Regular BPF_PROG_TYPE_LSM programs can return
14611 if (!env->prog->aux->attach_func_proto->type) {
14612 /* Make sure programs that attach to void
14613 * hooks don't try to modify return value.
14615 range = tnum_range(1, 1);
14619 case BPF_PROG_TYPE_NETFILTER:
14620 range = tnum_range(NF_DROP, NF_ACCEPT);
14622 case BPF_PROG_TYPE_EXT:
14623 /* freplace program can return anything as its return value
14624 * depends on the to-be-replaced kernel func or bpf program.
14630 if (reg->type != SCALAR_VALUE) {
14631 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
14632 reg_type_str(env, reg->type));
14636 if (!tnum_in(range, reg->var_off)) {
14637 verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
14638 if (prog->expected_attach_type == BPF_LSM_CGROUP &&
14639 prog_type == BPF_PROG_TYPE_LSM &&
14640 !prog->aux->attach_func_proto->type)
14641 verbose(env, "Note, BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
14645 if (!tnum_is_unknown(enforce_attach_type_range) &&
14646 tnum_in(enforce_attach_type_range, reg->var_off))
14647 env->prog->enforce_expected_attach_type = 1;
14651 /* non-recursive DFS pseudo code
14652 * 1 procedure DFS-iterative(G,v):
14653 * 2 label v as discovered
14654 * 3 let S be a stack
14656 * 5 while S is not empty
14658 * 7 if t is what we're looking for:
14660 * 9 for all edges e in G.adjacentEdges(t) do
14661 * 10 if edge e is already labelled
14662 * 11 continue with the next edge
14663 * 12 w <- G.adjacentVertex(t,e)
14664 * 13 if vertex w is not discovered and not explored
14665 * 14 label e as tree-edge
14666 * 15 label w as discovered
14669 * 18 else if vertex w is discovered
14670 * 19 label e as back-edge
14672 * 21 // vertex w is explored
14673 * 22 label e as forward- or cross-edge
14674 * 23 label t as explored
14678 * 0x10 - discovered
14679 * 0x11 - discovered and fall-through edge labelled
14680 * 0x12 - discovered and fall-through and branch edges labelled
14691 static u32 state_htab_size(struct bpf_verifier_env *env)
14693 return env->prog->len;
14696 static struct bpf_verifier_state_list **explored_state(
14697 struct bpf_verifier_env *env,
14700 struct bpf_verifier_state *cur = env->cur_state;
14701 struct bpf_func_state *state = cur->frame[cur->curframe];
14703 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
14706 static void mark_prune_point(struct bpf_verifier_env *env, int idx)
14708 env->insn_aux_data[idx].prune_point = true;
14711 static bool is_prune_point(struct bpf_verifier_env *env, int insn_idx)
14713 return env->insn_aux_data[insn_idx].prune_point;
14716 static void mark_force_checkpoint(struct bpf_verifier_env *env, int idx)
14718 env->insn_aux_data[idx].force_checkpoint = true;
14721 static bool is_force_checkpoint(struct bpf_verifier_env *env, int insn_idx)
14723 return env->insn_aux_data[insn_idx].force_checkpoint;
14728 DONE_EXPLORING = 0,
14729 KEEP_EXPLORING = 1,
14732 /* t, w, e - match pseudo-code above:
14733 * t - index of current instruction
14734 * w - next instruction
14737 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
14740 int *insn_stack = env->cfg.insn_stack;
14741 int *insn_state = env->cfg.insn_state;
14743 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
14744 return DONE_EXPLORING;
14746 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
14747 return DONE_EXPLORING;
14749 if (w < 0 || w >= env->prog->len) {
14750 verbose_linfo(env, t, "%d: ", t);
14751 verbose(env, "jump out of range from insn %d to %d\n", t, w);
14756 /* mark branch target for state pruning */
14757 mark_prune_point(env, w);
14758 mark_jmp_point(env, w);
14761 if (insn_state[w] == 0) {
14763 insn_state[t] = DISCOVERED | e;
14764 insn_state[w] = DISCOVERED;
14765 if (env->cfg.cur_stack >= env->prog->len)
14767 insn_stack[env->cfg.cur_stack++] = w;
14768 return KEEP_EXPLORING;
14769 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
14770 if (loop_ok && env->bpf_capable)
14771 return DONE_EXPLORING;
14772 verbose_linfo(env, t, "%d: ", t);
14773 verbose_linfo(env, w, "%d: ", w);
14774 verbose(env, "back-edge from insn %d to %d\n", t, w);
14776 } else if (insn_state[w] == EXPLORED) {
14777 /* forward- or cross-edge */
14778 insn_state[t] = DISCOVERED | e;
14780 verbose(env, "insn state internal bug\n");
14783 return DONE_EXPLORING;
14786 static int visit_func_call_insn(int t, struct bpf_insn *insns,
14787 struct bpf_verifier_env *env,
14792 insn_sz = bpf_is_ldimm64(&insns[t]) ? 2 : 1;
14793 ret = push_insn(t, t + insn_sz, FALLTHROUGH, env, false);
14797 mark_prune_point(env, t + insn_sz);
14798 /* when we exit from subprog, we need to record non-linear history */
14799 mark_jmp_point(env, t + insn_sz);
14801 if (visit_callee) {
14802 mark_prune_point(env, t);
14803 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env,
14804 /* It's ok to allow recursion from CFG point of
14805 * view. __check_func_call() will do the actual
14808 bpf_pseudo_func(insns + t));
14813 /* Visits the instruction at index t and returns one of the following:
14814 * < 0 - an error occurred
14815 * DONE_EXPLORING - the instruction was fully explored
14816 * KEEP_EXPLORING - there is still work to be done before it is fully explored
14818 static int visit_insn(int t, struct bpf_verifier_env *env)
14820 struct bpf_insn *insns = env->prog->insnsi, *insn = &insns[t];
14821 int ret, off, insn_sz;
14823 if (bpf_pseudo_func(insn))
14824 return visit_func_call_insn(t, insns, env, true);
14826 /* All non-branch instructions have a single fall-through edge. */
14827 if (BPF_CLASS(insn->code) != BPF_JMP &&
14828 BPF_CLASS(insn->code) != BPF_JMP32) {
14829 insn_sz = bpf_is_ldimm64(insn) ? 2 : 1;
14830 return push_insn(t, t + insn_sz, FALLTHROUGH, env, false);
14833 switch (BPF_OP(insn->code)) {
14835 return DONE_EXPLORING;
14838 if (insn->src_reg == 0 && insn->imm == BPF_FUNC_timer_set_callback)
14839 /* Mark this call insn as a prune point to trigger
14840 * is_state_visited() check before call itself is
14841 * processed by __check_func_call(). Otherwise new
14842 * async state will be pushed for further exploration.
14844 mark_prune_point(env, t);
14845 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
14846 struct bpf_kfunc_call_arg_meta meta;
14848 ret = fetch_kfunc_meta(env, insn, &meta, NULL);
14849 if (ret == 0 && is_iter_next_kfunc(&meta)) {
14850 mark_prune_point(env, t);
14851 /* Checking and saving state checkpoints at iter_next() call
14852 * is crucial for fast convergence of open-coded iterator loop
14853 * logic, so we need to force it. If we don't do that,
14854 * is_state_visited() might skip saving a checkpoint, causing
14855 * unnecessarily long sequence of not checkpointed
14856 * instructions and jumps, leading to exhaustion of jump
14857 * history buffer, and potentially other undesired outcomes.
14858 * It is expected that with correct open-coded iterators
14859 * convergence will happen quickly, so we don't run a risk of
14860 * exhausting memory.
14862 mark_force_checkpoint(env, t);
14865 return visit_func_call_insn(t, insns, env, insn->src_reg == BPF_PSEUDO_CALL);
14868 if (BPF_SRC(insn->code) != BPF_K)
14871 if (BPF_CLASS(insn->code) == BPF_JMP)
14876 /* unconditional jump with single edge */
14877 ret = push_insn(t, t + off + 1, FALLTHROUGH, env,
14882 mark_prune_point(env, t + off + 1);
14883 mark_jmp_point(env, t + off + 1);
14888 /* conditional jump with two edges */
14889 mark_prune_point(env, t);
14891 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
14895 return push_insn(t, t + insn->off + 1, BRANCH, env, true);
14899 /* non-recursive depth-first-search to detect loops in BPF program
14900 * loop == back-edge in directed graph
14902 static int check_cfg(struct bpf_verifier_env *env)
14904 int insn_cnt = env->prog->len;
14905 int *insn_stack, *insn_state;
14909 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
14913 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
14915 kvfree(insn_state);
14919 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
14920 insn_stack[0] = 0; /* 0 is the first instruction */
14921 env->cfg.cur_stack = 1;
14923 while (env->cfg.cur_stack > 0) {
14924 int t = insn_stack[env->cfg.cur_stack - 1];
14926 ret = visit_insn(t, env);
14928 case DONE_EXPLORING:
14929 insn_state[t] = EXPLORED;
14930 env->cfg.cur_stack--;
14932 case KEEP_EXPLORING:
14936 verbose(env, "visit_insn internal bug\n");
14943 if (env->cfg.cur_stack < 0) {
14944 verbose(env, "pop stack internal bug\n");
14949 for (i = 0; i < insn_cnt; i++) {
14950 struct bpf_insn *insn = &env->prog->insnsi[i];
14952 if (insn_state[i] != EXPLORED) {
14953 verbose(env, "unreachable insn %d\n", i);
14957 if (bpf_is_ldimm64(insn)) {
14958 if (insn_state[i + 1] != 0) {
14959 verbose(env, "jump into the middle of ldimm64 insn %d\n", i);
14963 i++; /* skip second half of ldimm64 */
14966 ret = 0; /* cfg looks good */
14969 kvfree(insn_state);
14970 kvfree(insn_stack);
14971 env->cfg.insn_state = env->cfg.insn_stack = NULL;
14975 static int check_abnormal_return(struct bpf_verifier_env *env)
14979 for (i = 1; i < env->subprog_cnt; i++) {
14980 if (env->subprog_info[i].has_ld_abs) {
14981 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
14984 if (env->subprog_info[i].has_tail_call) {
14985 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
14992 /* The minimum supported BTF func info size */
14993 #define MIN_BPF_FUNCINFO_SIZE 8
14994 #define MAX_FUNCINFO_REC_SIZE 252
14996 static int check_btf_func(struct bpf_verifier_env *env,
14997 const union bpf_attr *attr,
15000 const struct btf_type *type, *func_proto, *ret_type;
15001 u32 i, nfuncs, urec_size, min_size;
15002 u32 krec_size = sizeof(struct bpf_func_info);
15003 struct bpf_func_info *krecord;
15004 struct bpf_func_info_aux *info_aux = NULL;
15005 struct bpf_prog *prog;
15006 const struct btf *btf;
15008 u32 prev_offset = 0;
15009 bool scalar_return;
15012 nfuncs = attr->func_info_cnt;
15014 if (check_abnormal_return(env))
15019 if (nfuncs != env->subprog_cnt) {
15020 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
15024 urec_size = attr->func_info_rec_size;
15025 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
15026 urec_size > MAX_FUNCINFO_REC_SIZE ||
15027 urec_size % sizeof(u32)) {
15028 verbose(env, "invalid func info rec size %u\n", urec_size);
15033 btf = prog->aux->btf;
15035 urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
15036 min_size = min_t(u32, krec_size, urec_size);
15038 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
15041 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
15045 for (i = 0; i < nfuncs; i++) {
15046 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
15048 if (ret == -E2BIG) {
15049 verbose(env, "nonzero tailing record in func info");
15050 /* set the size kernel expects so loader can zero
15051 * out the rest of the record.
15053 if (copy_to_bpfptr_offset(uattr,
15054 offsetof(union bpf_attr, func_info_rec_size),
15055 &min_size, sizeof(min_size)))
15061 if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
15066 /* check insn_off */
15069 if (krecord[i].insn_off) {
15071 "nonzero insn_off %u for the first func info record",
15072 krecord[i].insn_off);
15075 } else if (krecord[i].insn_off <= prev_offset) {
15077 "same or smaller insn offset (%u) than previous func info record (%u)",
15078 krecord[i].insn_off, prev_offset);
15082 if (env->subprog_info[i].start != krecord[i].insn_off) {
15083 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
15087 /* check type_id */
15088 type = btf_type_by_id(btf, krecord[i].type_id);
15089 if (!type || !btf_type_is_func(type)) {
15090 verbose(env, "invalid type id %d in func info",
15091 krecord[i].type_id);
15094 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
15096 func_proto = btf_type_by_id(btf, type->type);
15097 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
15098 /* btf_func_check() already verified it during BTF load */
15100 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
15102 btf_type_is_small_int(ret_type) || btf_is_any_enum(ret_type);
15103 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
15104 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
15107 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
15108 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
15112 prev_offset = krecord[i].insn_off;
15113 bpfptr_add(&urecord, urec_size);
15116 prog->aux->func_info = krecord;
15117 prog->aux->func_info_cnt = nfuncs;
15118 prog->aux->func_info_aux = info_aux;
15127 static void adjust_btf_func(struct bpf_verifier_env *env)
15129 struct bpf_prog_aux *aux = env->prog->aux;
15132 if (!aux->func_info)
15135 for (i = 0; i < env->subprog_cnt; i++)
15136 aux->func_info[i].insn_off = env->subprog_info[i].start;
15139 #define MIN_BPF_LINEINFO_SIZE offsetofend(struct bpf_line_info, line_col)
15140 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
15142 static int check_btf_line(struct bpf_verifier_env *env,
15143 const union bpf_attr *attr,
15146 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
15147 struct bpf_subprog_info *sub;
15148 struct bpf_line_info *linfo;
15149 struct bpf_prog *prog;
15150 const struct btf *btf;
15154 nr_linfo = attr->line_info_cnt;
15157 if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info))
15160 rec_size = attr->line_info_rec_size;
15161 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
15162 rec_size > MAX_LINEINFO_REC_SIZE ||
15163 rec_size & (sizeof(u32) - 1))
15166 /* Need to zero it in case the userspace may
15167 * pass in a smaller bpf_line_info object.
15169 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
15170 GFP_KERNEL | __GFP_NOWARN);
15175 btf = prog->aux->btf;
15178 sub = env->subprog_info;
15179 ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
15180 expected_size = sizeof(struct bpf_line_info);
15181 ncopy = min_t(u32, expected_size, rec_size);
15182 for (i = 0; i < nr_linfo; i++) {
15183 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
15185 if (err == -E2BIG) {
15186 verbose(env, "nonzero tailing record in line_info");
15187 if (copy_to_bpfptr_offset(uattr,
15188 offsetof(union bpf_attr, line_info_rec_size),
15189 &expected_size, sizeof(expected_size)))
15195 if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
15201 * Check insn_off to ensure
15202 * 1) strictly increasing AND
15203 * 2) bounded by prog->len
15205 * The linfo[0].insn_off == 0 check logically falls into
15206 * the later "missing bpf_line_info for func..." case
15207 * because the first linfo[0].insn_off must be the
15208 * first sub also and the first sub must have
15209 * subprog_info[0].start == 0.
15211 if ((i && linfo[i].insn_off <= prev_offset) ||
15212 linfo[i].insn_off >= prog->len) {
15213 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
15214 i, linfo[i].insn_off, prev_offset,
15220 if (!prog->insnsi[linfo[i].insn_off].code) {
15222 "Invalid insn code at line_info[%u].insn_off\n",
15228 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
15229 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
15230 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
15235 if (s != env->subprog_cnt) {
15236 if (linfo[i].insn_off == sub[s].start) {
15237 sub[s].linfo_idx = i;
15239 } else if (sub[s].start < linfo[i].insn_off) {
15240 verbose(env, "missing bpf_line_info for func#%u\n", s);
15246 prev_offset = linfo[i].insn_off;
15247 bpfptr_add(&ulinfo, rec_size);
15250 if (s != env->subprog_cnt) {
15251 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
15252 env->subprog_cnt - s, s);
15257 prog->aux->linfo = linfo;
15258 prog->aux->nr_linfo = nr_linfo;
15267 #define MIN_CORE_RELO_SIZE sizeof(struct bpf_core_relo)
15268 #define MAX_CORE_RELO_SIZE MAX_FUNCINFO_REC_SIZE
15270 static int check_core_relo(struct bpf_verifier_env *env,
15271 const union bpf_attr *attr,
15274 u32 i, nr_core_relo, ncopy, expected_size, rec_size;
15275 struct bpf_core_relo core_relo = {};
15276 struct bpf_prog *prog = env->prog;
15277 const struct btf *btf = prog->aux->btf;
15278 struct bpf_core_ctx ctx = {
15282 bpfptr_t u_core_relo;
15285 nr_core_relo = attr->core_relo_cnt;
15288 if (nr_core_relo > INT_MAX / sizeof(struct bpf_core_relo))
15291 rec_size = attr->core_relo_rec_size;
15292 if (rec_size < MIN_CORE_RELO_SIZE ||
15293 rec_size > MAX_CORE_RELO_SIZE ||
15294 rec_size % sizeof(u32))
15297 u_core_relo = make_bpfptr(attr->core_relos, uattr.is_kernel);
15298 expected_size = sizeof(struct bpf_core_relo);
15299 ncopy = min_t(u32, expected_size, rec_size);
15301 /* Unlike func_info and line_info, copy and apply each CO-RE
15302 * relocation record one at a time.
15304 for (i = 0; i < nr_core_relo; i++) {
15305 /* future proofing when sizeof(bpf_core_relo) changes */
15306 err = bpf_check_uarg_tail_zero(u_core_relo, expected_size, rec_size);
15308 if (err == -E2BIG) {
15309 verbose(env, "nonzero tailing record in core_relo");
15310 if (copy_to_bpfptr_offset(uattr,
15311 offsetof(union bpf_attr, core_relo_rec_size),
15312 &expected_size, sizeof(expected_size)))
15318 if (copy_from_bpfptr(&core_relo, u_core_relo, ncopy)) {
15323 if (core_relo.insn_off % 8 || core_relo.insn_off / 8 >= prog->len) {
15324 verbose(env, "Invalid core_relo[%u].insn_off:%u prog->len:%u\n",
15325 i, core_relo.insn_off, prog->len);
15330 err = bpf_core_apply(&ctx, &core_relo, i,
15331 &prog->insnsi[core_relo.insn_off / 8]);
15334 bpfptr_add(&u_core_relo, rec_size);
15339 static int check_btf_info(struct bpf_verifier_env *env,
15340 const union bpf_attr *attr,
15346 if (!attr->func_info_cnt && !attr->line_info_cnt) {
15347 if (check_abnormal_return(env))
15352 btf = btf_get_by_fd(attr->prog_btf_fd);
15354 return PTR_ERR(btf);
15355 if (btf_is_kernel(btf)) {
15359 env->prog->aux->btf = btf;
15361 err = check_btf_func(env, attr, uattr);
15365 err = check_btf_line(env, attr, uattr);
15369 err = check_core_relo(env, attr, uattr);
15376 /* check %cur's range satisfies %old's */
15377 static bool range_within(struct bpf_reg_state *old,
15378 struct bpf_reg_state *cur)
15380 return old->umin_value <= cur->umin_value &&
15381 old->umax_value >= cur->umax_value &&
15382 old->smin_value <= cur->smin_value &&
15383 old->smax_value >= cur->smax_value &&
15384 old->u32_min_value <= cur->u32_min_value &&
15385 old->u32_max_value >= cur->u32_max_value &&
15386 old->s32_min_value <= cur->s32_min_value &&
15387 old->s32_max_value >= cur->s32_max_value;
15390 /* If in the old state two registers had the same id, then they need to have
15391 * the same id in the new state as well. But that id could be different from
15392 * the old state, so we need to track the mapping from old to new ids.
15393 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
15394 * regs with old id 5 must also have new id 9 for the new state to be safe. But
15395 * regs with a different old id could still have new id 9, we don't care about
15397 * So we look through our idmap to see if this old id has been seen before. If
15398 * so, we require the new id to match; otherwise, we add the id pair to the map.
15400 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap)
15402 struct bpf_id_pair *map = idmap->map;
15405 /* either both IDs should be set or both should be zero */
15406 if (!!old_id != !!cur_id)
15409 if (old_id == 0) /* cur_id == 0 as well */
15412 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
15414 /* Reached an empty slot; haven't seen this id before */
15415 map[i].old = old_id;
15416 map[i].cur = cur_id;
15419 if (map[i].old == old_id)
15420 return map[i].cur == cur_id;
15421 if (map[i].cur == cur_id)
15424 /* We ran out of idmap slots, which should be impossible */
15429 /* Similar to check_ids(), but allocate a unique temporary ID
15430 * for 'old_id' or 'cur_id' of zero.
15431 * This makes pairs like '0 vs unique ID', 'unique ID vs 0' valid.
15433 static bool check_scalar_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap)
15435 old_id = old_id ? old_id : ++idmap->tmp_id_gen;
15436 cur_id = cur_id ? cur_id : ++idmap->tmp_id_gen;
15438 return check_ids(old_id, cur_id, idmap);
15441 static void clean_func_state(struct bpf_verifier_env *env,
15442 struct bpf_func_state *st)
15444 enum bpf_reg_liveness live;
15447 for (i = 0; i < BPF_REG_FP; i++) {
15448 live = st->regs[i].live;
15449 /* liveness must not touch this register anymore */
15450 st->regs[i].live |= REG_LIVE_DONE;
15451 if (!(live & REG_LIVE_READ))
15452 /* since the register is unused, clear its state
15453 * to make further comparison simpler
15455 __mark_reg_not_init(env, &st->regs[i]);
15458 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
15459 live = st->stack[i].spilled_ptr.live;
15460 /* liveness must not touch this stack slot anymore */
15461 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
15462 if (!(live & REG_LIVE_READ)) {
15463 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
15464 for (j = 0; j < BPF_REG_SIZE; j++)
15465 st->stack[i].slot_type[j] = STACK_INVALID;
15470 static void clean_verifier_state(struct bpf_verifier_env *env,
15471 struct bpf_verifier_state *st)
15475 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
15476 /* all regs in this state in all frames were already marked */
15479 for (i = 0; i <= st->curframe; i++)
15480 clean_func_state(env, st->frame[i]);
15483 /* the parentage chains form a tree.
15484 * the verifier states are added to state lists at given insn and
15485 * pushed into state stack for future exploration.
15486 * when the verifier reaches bpf_exit insn some of the verifer states
15487 * stored in the state lists have their final liveness state already,
15488 * but a lot of states will get revised from liveness point of view when
15489 * the verifier explores other branches.
15492 * 2: if r1 == 100 goto pc+1
15495 * when the verifier reaches exit insn the register r0 in the state list of
15496 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
15497 * of insn 2 and goes exploring further. At the insn 4 it will walk the
15498 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
15500 * Since the verifier pushes the branch states as it sees them while exploring
15501 * the program the condition of walking the branch instruction for the second
15502 * time means that all states below this branch were already explored and
15503 * their final liveness marks are already propagated.
15504 * Hence when the verifier completes the search of state list in is_state_visited()
15505 * we can call this clean_live_states() function to mark all liveness states
15506 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
15507 * will not be used.
15508 * This function also clears the registers and stack for states that !READ
15509 * to simplify state merging.
15511 * Important note here that walking the same branch instruction in the callee
15512 * doesn't meant that the states are DONE. The verifier has to compare
15515 static void clean_live_states(struct bpf_verifier_env *env, int insn,
15516 struct bpf_verifier_state *cur)
15518 struct bpf_verifier_state_list *sl;
15521 sl = *explored_state(env, insn);
15523 if (sl->state.branches)
15525 if (sl->state.insn_idx != insn ||
15526 sl->state.curframe != cur->curframe)
15528 for (i = 0; i <= cur->curframe; i++)
15529 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
15531 clean_verifier_state(env, &sl->state);
15537 static bool regs_exact(const struct bpf_reg_state *rold,
15538 const struct bpf_reg_state *rcur,
15539 struct bpf_idmap *idmap)
15541 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
15542 check_ids(rold->id, rcur->id, idmap) &&
15543 check_ids(rold->ref_obj_id, rcur->ref_obj_id, idmap);
15546 /* Returns true if (rold safe implies rcur safe) */
15547 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
15548 struct bpf_reg_state *rcur, struct bpf_idmap *idmap)
15550 if (!(rold->live & REG_LIVE_READ))
15551 /* explored state didn't use this */
15553 if (rold->type == NOT_INIT)
15554 /* explored state can't have used this */
15556 if (rcur->type == NOT_INIT)
15559 /* Enforce that register types have to match exactly, including their
15560 * modifiers (like PTR_MAYBE_NULL, MEM_RDONLY, etc), as a general
15563 * One can make a point that using a pointer register as unbounded
15564 * SCALAR would be technically acceptable, but this could lead to
15565 * pointer leaks because scalars are allowed to leak while pointers
15566 * are not. We could make this safe in special cases if root is
15567 * calling us, but it's probably not worth the hassle.
15569 * Also, register types that are *not* MAYBE_NULL could technically be
15570 * safe to use as their MAYBE_NULL variants (e.g., PTR_TO_MAP_VALUE
15571 * is safe to be used as PTR_TO_MAP_VALUE_OR_NULL, provided both point
15572 * to the same map).
15573 * However, if the old MAYBE_NULL register then got NULL checked,
15574 * doing so could have affected others with the same id, and we can't
15575 * check for that because we lost the id when we converted to
15576 * a non-MAYBE_NULL variant.
15577 * So, as a general rule we don't allow mixing MAYBE_NULL and
15578 * non-MAYBE_NULL registers as well.
15580 if (rold->type != rcur->type)
15583 switch (base_type(rold->type)) {
15585 if (env->explore_alu_limits) {
15586 /* explore_alu_limits disables tnum_in() and range_within()
15587 * logic and requires everything to be strict
15589 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
15590 check_scalar_ids(rold->id, rcur->id, idmap);
15592 if (!rold->precise)
15594 /* Why check_ids() for scalar registers?
15596 * Consider the following BPF code:
15597 * 1: r6 = ... unbound scalar, ID=a ...
15598 * 2: r7 = ... unbound scalar, ID=b ...
15599 * 3: if (r6 > r7) goto +1
15601 * 5: if (r6 > X) goto ...
15602 * 6: ... memory operation using r7 ...
15604 * First verification path is [1-6]:
15605 * - at (4) same bpf_reg_state::id (b) would be assigned to r6 and r7;
15606 * - at (5) r6 would be marked <= X, find_equal_scalars() would also mark
15607 * r7 <= X, because r6 and r7 share same id.
15608 * Next verification path is [1-4, 6].
15610 * Instruction (6) would be reached in two states:
15611 * I. r6{.id=b}, r7{.id=b} via path 1-6;
15612 * II. r6{.id=a}, r7{.id=b} via path 1-4, 6.
15614 * Use check_ids() to distinguish these states.
15616 * Also verify that new value satisfies old value range knowledge.
15618 return range_within(rold, rcur) &&
15619 tnum_in(rold->var_off, rcur->var_off) &&
15620 check_scalar_ids(rold->id, rcur->id, idmap);
15621 case PTR_TO_MAP_KEY:
15622 case PTR_TO_MAP_VALUE:
15625 case PTR_TO_TP_BUFFER:
15626 /* If the new min/max/var_off satisfy the old ones and
15627 * everything else matches, we are OK.
15629 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, var_off)) == 0 &&
15630 range_within(rold, rcur) &&
15631 tnum_in(rold->var_off, rcur->var_off) &&
15632 check_ids(rold->id, rcur->id, idmap) &&
15633 check_ids(rold->ref_obj_id, rcur->ref_obj_id, idmap);
15634 case PTR_TO_PACKET_META:
15635 case PTR_TO_PACKET:
15636 /* We must have at least as much range as the old ptr
15637 * did, so that any accesses which were safe before are
15638 * still safe. This is true even if old range < old off,
15639 * since someone could have accessed through (ptr - k), or
15640 * even done ptr -= k in a register, to get a safe access.
15642 if (rold->range > rcur->range)
15644 /* If the offsets don't match, we can't trust our alignment;
15645 * nor can we be sure that we won't fall out of range.
15647 if (rold->off != rcur->off)
15649 /* id relations must be preserved */
15650 if (!check_ids(rold->id, rcur->id, idmap))
15652 /* new val must satisfy old val knowledge */
15653 return range_within(rold, rcur) &&
15654 tnum_in(rold->var_off, rcur->var_off);
15656 /* two stack pointers are equal only if they're pointing to
15657 * the same stack frame, since fp-8 in foo != fp-8 in bar
15659 return regs_exact(rold, rcur, idmap) && rold->frameno == rcur->frameno;
15661 return regs_exact(rold, rcur, idmap);
15665 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
15666 struct bpf_func_state *cur, struct bpf_idmap *idmap)
15670 /* walk slots of the explored stack and ignore any additional
15671 * slots in the current stack, since explored(safe) state
15674 for (i = 0; i < old->allocated_stack; i++) {
15675 struct bpf_reg_state *old_reg, *cur_reg;
15677 spi = i / BPF_REG_SIZE;
15679 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
15680 i += BPF_REG_SIZE - 1;
15681 /* explored state didn't use this */
15685 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
15688 if (env->allow_uninit_stack &&
15689 old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC)
15692 /* explored stack has more populated slots than current stack
15693 * and these slots were used
15695 if (i >= cur->allocated_stack)
15698 /* if old state was safe with misc data in the stack
15699 * it will be safe with zero-initialized stack.
15700 * The opposite is not true
15702 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
15703 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
15705 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
15706 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
15707 /* Ex: old explored (safe) state has STACK_SPILL in
15708 * this stack slot, but current has STACK_MISC ->
15709 * this verifier states are not equivalent,
15710 * return false to continue verification of this path
15713 if (i % BPF_REG_SIZE != BPF_REG_SIZE - 1)
15715 /* Both old and cur are having same slot_type */
15716 switch (old->stack[spi].slot_type[BPF_REG_SIZE - 1]) {
15718 /* when explored and current stack slot are both storing
15719 * spilled registers, check that stored pointers types
15720 * are the same as well.
15721 * Ex: explored safe path could have stored
15722 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
15723 * but current path has stored:
15724 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
15725 * such verifier states are not equivalent.
15726 * return false to continue verification of this path
15728 if (!regsafe(env, &old->stack[spi].spilled_ptr,
15729 &cur->stack[spi].spilled_ptr, idmap))
15733 old_reg = &old->stack[spi].spilled_ptr;
15734 cur_reg = &cur->stack[spi].spilled_ptr;
15735 if (old_reg->dynptr.type != cur_reg->dynptr.type ||
15736 old_reg->dynptr.first_slot != cur_reg->dynptr.first_slot ||
15737 !check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap))
15741 old_reg = &old->stack[spi].spilled_ptr;
15742 cur_reg = &cur->stack[spi].spilled_ptr;
15743 /* iter.depth is not compared between states as it
15744 * doesn't matter for correctness and would otherwise
15745 * prevent convergence; we maintain it only to prevent
15746 * infinite loop check triggering, see
15747 * iter_active_depths_differ()
15749 if (old_reg->iter.btf != cur_reg->iter.btf ||
15750 old_reg->iter.btf_id != cur_reg->iter.btf_id ||
15751 old_reg->iter.state != cur_reg->iter.state ||
15752 /* ignore {old_reg,cur_reg}->iter.depth, see above */
15753 !check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap))
15758 case STACK_INVALID:
15760 /* Ensure that new unhandled slot types return false by default */
15768 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur,
15769 struct bpf_idmap *idmap)
15773 if (old->acquired_refs != cur->acquired_refs)
15776 for (i = 0; i < old->acquired_refs; i++) {
15777 if (!check_ids(old->refs[i].id, cur->refs[i].id, idmap))
15784 /* compare two verifier states
15786 * all states stored in state_list are known to be valid, since
15787 * verifier reached 'bpf_exit' instruction through them
15789 * this function is called when verifier exploring different branches of
15790 * execution popped from the state stack. If it sees an old state that has
15791 * more strict register state and more strict stack state then this execution
15792 * branch doesn't need to be explored further, since verifier already
15793 * concluded that more strict state leads to valid finish.
15795 * Therefore two states are equivalent if register state is more conservative
15796 * and explored stack state is more conservative than the current one.
15799 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
15800 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
15802 * In other words if current stack state (one being explored) has more
15803 * valid slots than old one that already passed validation, it means
15804 * the verifier can stop exploring and conclude that current state is valid too
15806 * Similarly with registers. If explored state has register type as invalid
15807 * whereas register type in current state is meaningful, it means that
15808 * the current state will reach 'bpf_exit' instruction safely
15810 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
15811 struct bpf_func_state *cur)
15815 for (i = 0; i < MAX_BPF_REG; i++)
15816 if (!regsafe(env, &old->regs[i], &cur->regs[i],
15817 &env->idmap_scratch))
15820 if (!stacksafe(env, old, cur, &env->idmap_scratch))
15823 if (!refsafe(old, cur, &env->idmap_scratch))
15829 static bool states_equal(struct bpf_verifier_env *env,
15830 struct bpf_verifier_state *old,
15831 struct bpf_verifier_state *cur)
15835 if (old->curframe != cur->curframe)
15838 env->idmap_scratch.tmp_id_gen = env->id_gen;
15839 memset(&env->idmap_scratch.map, 0, sizeof(env->idmap_scratch.map));
15841 /* Verification state from speculative execution simulation
15842 * must never prune a non-speculative execution one.
15844 if (old->speculative && !cur->speculative)
15847 if (old->active_lock.ptr != cur->active_lock.ptr)
15850 /* Old and cur active_lock's have to be either both present
15853 if (!!old->active_lock.id != !!cur->active_lock.id)
15856 if (old->active_lock.id &&
15857 !check_ids(old->active_lock.id, cur->active_lock.id, &env->idmap_scratch))
15860 if (old->active_rcu_lock != cur->active_rcu_lock)
15863 /* for states to be equal callsites have to be the same
15864 * and all frame states need to be equivalent
15866 for (i = 0; i <= old->curframe; i++) {
15867 if (old->frame[i]->callsite != cur->frame[i]->callsite)
15869 if (!func_states_equal(env, old->frame[i], cur->frame[i]))
15875 /* Return 0 if no propagation happened. Return negative error code if error
15876 * happened. Otherwise, return the propagated bit.
15878 static int propagate_liveness_reg(struct bpf_verifier_env *env,
15879 struct bpf_reg_state *reg,
15880 struct bpf_reg_state *parent_reg)
15882 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
15883 u8 flag = reg->live & REG_LIVE_READ;
15886 /* When comes here, read flags of PARENT_REG or REG could be any of
15887 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
15888 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
15890 if (parent_flag == REG_LIVE_READ64 ||
15891 /* Or if there is no read flag from REG. */
15893 /* Or if the read flag from REG is the same as PARENT_REG. */
15894 parent_flag == flag)
15897 err = mark_reg_read(env, reg, parent_reg, flag);
15904 /* A write screens off any subsequent reads; but write marks come from the
15905 * straight-line code between a state and its parent. When we arrive at an
15906 * equivalent state (jump target or such) we didn't arrive by the straight-line
15907 * code, so read marks in the state must propagate to the parent regardless
15908 * of the state's write marks. That's what 'parent == state->parent' comparison
15909 * in mark_reg_read() is for.
15911 static int propagate_liveness(struct bpf_verifier_env *env,
15912 const struct bpf_verifier_state *vstate,
15913 struct bpf_verifier_state *vparent)
15915 struct bpf_reg_state *state_reg, *parent_reg;
15916 struct bpf_func_state *state, *parent;
15917 int i, frame, err = 0;
15919 if (vparent->curframe != vstate->curframe) {
15920 WARN(1, "propagate_live: parent frame %d current frame %d\n",
15921 vparent->curframe, vstate->curframe);
15924 /* Propagate read liveness of registers... */
15925 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
15926 for (frame = 0; frame <= vstate->curframe; frame++) {
15927 parent = vparent->frame[frame];
15928 state = vstate->frame[frame];
15929 parent_reg = parent->regs;
15930 state_reg = state->regs;
15931 /* We don't need to worry about FP liveness, it's read-only */
15932 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
15933 err = propagate_liveness_reg(env, &state_reg[i],
15937 if (err == REG_LIVE_READ64)
15938 mark_insn_zext(env, &parent_reg[i]);
15941 /* Propagate stack slots. */
15942 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
15943 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
15944 parent_reg = &parent->stack[i].spilled_ptr;
15945 state_reg = &state->stack[i].spilled_ptr;
15946 err = propagate_liveness_reg(env, state_reg,
15955 /* find precise scalars in the previous equivalent state and
15956 * propagate them into the current state
15958 static int propagate_precision(struct bpf_verifier_env *env,
15959 const struct bpf_verifier_state *old)
15961 struct bpf_reg_state *state_reg;
15962 struct bpf_func_state *state;
15963 int i, err = 0, fr;
15966 for (fr = old->curframe; fr >= 0; fr--) {
15967 state = old->frame[fr];
15968 state_reg = state->regs;
15970 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
15971 if (state_reg->type != SCALAR_VALUE ||
15972 !state_reg->precise ||
15973 !(state_reg->live & REG_LIVE_READ))
15975 if (env->log.level & BPF_LOG_LEVEL2) {
15977 verbose(env, "frame %d: propagating r%d", fr, i);
15979 verbose(env, ",r%d", i);
15981 bt_set_frame_reg(&env->bt, fr, i);
15985 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
15986 if (!is_spilled_reg(&state->stack[i]))
15988 state_reg = &state->stack[i].spilled_ptr;
15989 if (state_reg->type != SCALAR_VALUE ||
15990 !state_reg->precise ||
15991 !(state_reg->live & REG_LIVE_READ))
15993 if (env->log.level & BPF_LOG_LEVEL2) {
15995 verbose(env, "frame %d: propagating fp%d",
15996 fr, (-i - 1) * BPF_REG_SIZE);
15998 verbose(env, ",fp%d", (-i - 1) * BPF_REG_SIZE);
16000 bt_set_frame_slot(&env->bt, fr, i);
16004 verbose(env, "\n");
16007 err = mark_chain_precision_batch(env);
16014 static bool states_maybe_looping(struct bpf_verifier_state *old,
16015 struct bpf_verifier_state *cur)
16017 struct bpf_func_state *fold, *fcur;
16018 int i, fr = cur->curframe;
16020 if (old->curframe != fr)
16023 fold = old->frame[fr];
16024 fcur = cur->frame[fr];
16025 for (i = 0; i < MAX_BPF_REG; i++)
16026 if (memcmp(&fold->regs[i], &fcur->regs[i],
16027 offsetof(struct bpf_reg_state, parent)))
16032 static bool is_iter_next_insn(struct bpf_verifier_env *env, int insn_idx)
16034 return env->insn_aux_data[insn_idx].is_iter_next;
16037 /* is_state_visited() handles iter_next() (see process_iter_next_call() for
16038 * terminology) calls specially: as opposed to bounded BPF loops, it *expects*
16039 * states to match, which otherwise would look like an infinite loop. So while
16040 * iter_next() calls are taken care of, we still need to be careful and
16041 * prevent erroneous and too eager declaration of "ininite loop", when
16042 * iterators are involved.
16044 * Here's a situation in pseudo-BPF assembly form:
16046 * 0: again: ; set up iter_next() call args
16047 * 1: r1 = &it ; <CHECKPOINT HERE>
16048 * 2: call bpf_iter_num_next ; this is iter_next() call
16049 * 3: if r0 == 0 goto done
16050 * 4: ... something useful here ...
16051 * 5: goto again ; another iteration
16054 * 8: call bpf_iter_num_destroy ; clean up iter state
16057 * This is a typical loop. Let's assume that we have a prune point at 1:,
16058 * before we get to `call bpf_iter_num_next` (e.g., because of that `goto
16059 * again`, assuming other heuristics don't get in a way).
16061 * When we first time come to 1:, let's say we have some state X. We proceed
16062 * to 2:, fork states, enqueue ACTIVE, validate NULL case successfully, exit.
16063 * Now we come back to validate that forked ACTIVE state. We proceed through
16064 * 3-5, come to goto, jump to 1:. Let's assume our state didn't change, so we
16065 * are converging. But the problem is that we don't know that yet, as this
16066 * convergence has to happen at iter_next() call site only. So if nothing is
16067 * done, at 1: verifier will use bounded loop logic and declare infinite
16068 * looping (and would be *technically* correct, if not for iterator's
16069 * "eventual sticky NULL" contract, see process_iter_next_call()). But we
16070 * don't want that. So what we do in process_iter_next_call() when we go on
16071 * another ACTIVE iteration, we bump slot->iter.depth, to mark that it's
16072 * a different iteration. So when we suspect an infinite loop, we additionally
16073 * check if any of the *ACTIVE* iterator states depths differ. If yes, we
16074 * pretend we are not looping and wait for next iter_next() call.
16076 * This only applies to ACTIVE state. In DRAINED state we don't expect to
16077 * loop, because that would actually mean infinite loop, as DRAINED state is
16078 * "sticky", and so we'll keep returning into the same instruction with the
16079 * same state (at least in one of possible code paths).
16081 * This approach allows to keep infinite loop heuristic even in the face of
16082 * active iterator. E.g., C snippet below is and will be detected as
16083 * inifintely looping:
16085 * struct bpf_iter_num it;
16088 * bpf_iter_num_new(&it, 0, 10);
16089 * while ((p = bpf_iter_num_next(&t))) {
16091 * while (x--) {} // <<-- infinite loop here
16095 static bool iter_active_depths_differ(struct bpf_verifier_state *old, struct bpf_verifier_state *cur)
16097 struct bpf_reg_state *slot, *cur_slot;
16098 struct bpf_func_state *state;
16101 for (fr = old->curframe; fr >= 0; fr--) {
16102 state = old->frame[fr];
16103 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
16104 if (state->stack[i].slot_type[0] != STACK_ITER)
16107 slot = &state->stack[i].spilled_ptr;
16108 if (slot->iter.state != BPF_ITER_STATE_ACTIVE)
16111 cur_slot = &cur->frame[fr]->stack[i].spilled_ptr;
16112 if (cur_slot->iter.depth != slot->iter.depth)
16119 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
16121 struct bpf_verifier_state_list *new_sl;
16122 struct bpf_verifier_state_list *sl, **pprev;
16123 struct bpf_verifier_state *cur = env->cur_state, *new;
16124 int i, j, err, states_cnt = 0;
16125 bool force_new_state = env->test_state_freq || is_force_checkpoint(env, insn_idx);
16126 bool add_new_state = force_new_state;
16128 /* bpf progs typically have pruning point every 4 instructions
16129 * http://vger.kernel.org/bpfconf2019.html#session-1
16130 * Do not add new state for future pruning if the verifier hasn't seen
16131 * at least 2 jumps and at least 8 instructions.
16132 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
16133 * In tests that amounts to up to 50% reduction into total verifier
16134 * memory consumption and 20% verifier time speedup.
16136 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
16137 env->insn_processed - env->prev_insn_processed >= 8)
16138 add_new_state = true;
16140 pprev = explored_state(env, insn_idx);
16143 clean_live_states(env, insn_idx, cur);
16147 if (sl->state.insn_idx != insn_idx)
16150 if (sl->state.branches) {
16151 struct bpf_func_state *frame = sl->state.frame[sl->state.curframe];
16153 if (frame->in_async_callback_fn &&
16154 frame->async_entry_cnt != cur->frame[cur->curframe]->async_entry_cnt) {
16155 /* Different async_entry_cnt means that the verifier is
16156 * processing another entry into async callback.
16157 * Seeing the same state is not an indication of infinite
16158 * loop or infinite recursion.
16159 * But finding the same state doesn't mean that it's safe
16160 * to stop processing the current state. The previous state
16161 * hasn't yet reached bpf_exit, since state.branches > 0.
16162 * Checking in_async_callback_fn alone is not enough either.
16163 * Since the verifier still needs to catch infinite loops
16164 * inside async callbacks.
16166 goto skip_inf_loop_check;
16168 /* BPF open-coded iterators loop detection is special.
16169 * states_maybe_looping() logic is too simplistic in detecting
16170 * states that *might* be equivalent, because it doesn't know
16171 * about ID remapping, so don't even perform it.
16172 * See process_iter_next_call() and iter_active_depths_differ()
16173 * for overview of the logic. When current and one of parent
16174 * states are detected as equivalent, it's a good thing: we prove
16175 * convergence and can stop simulating further iterations.
16176 * It's safe to assume that iterator loop will finish, taking into
16177 * account iter_next() contract of eventually returning
16178 * sticky NULL result.
16180 if (is_iter_next_insn(env, insn_idx)) {
16181 if (states_equal(env, &sl->state, cur)) {
16182 struct bpf_func_state *cur_frame;
16183 struct bpf_reg_state *iter_state, *iter_reg;
16186 cur_frame = cur->frame[cur->curframe];
16187 /* btf_check_iter_kfuncs() enforces that
16188 * iter state pointer is always the first arg
16190 iter_reg = &cur_frame->regs[BPF_REG_1];
16191 /* current state is valid due to states_equal(),
16192 * so we can assume valid iter and reg state,
16193 * no need for extra (re-)validations
16195 spi = __get_spi(iter_reg->off + iter_reg->var_off.value);
16196 iter_state = &func(env, iter_reg)->stack[spi].spilled_ptr;
16197 if (iter_state->iter.state == BPF_ITER_STATE_ACTIVE)
16200 goto skip_inf_loop_check;
16202 /* attempt to detect infinite loop to avoid unnecessary doomed work */
16203 if (states_maybe_looping(&sl->state, cur) &&
16204 states_equal(env, &sl->state, cur) &&
16205 !iter_active_depths_differ(&sl->state, cur)) {
16206 verbose_linfo(env, insn_idx, "; ");
16207 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
16210 /* if the verifier is processing a loop, avoid adding new state
16211 * too often, since different loop iterations have distinct
16212 * states and may not help future pruning.
16213 * This threshold shouldn't be too low to make sure that
16214 * a loop with large bound will be rejected quickly.
16215 * The most abusive loop will be:
16217 * if r1 < 1000000 goto pc-2
16218 * 1M insn_procssed limit / 100 == 10k peak states.
16219 * This threshold shouldn't be too high either, since states
16220 * at the end of the loop are likely to be useful in pruning.
16222 skip_inf_loop_check:
16223 if (!force_new_state &&
16224 env->jmps_processed - env->prev_jmps_processed < 20 &&
16225 env->insn_processed - env->prev_insn_processed < 100)
16226 add_new_state = false;
16229 if (states_equal(env, &sl->state, cur)) {
16232 /* reached equivalent register/stack state,
16233 * prune the search.
16234 * Registers read by the continuation are read by us.
16235 * If we have any write marks in env->cur_state, they
16236 * will prevent corresponding reads in the continuation
16237 * from reaching our parent (an explored_state). Our
16238 * own state will get the read marks recorded, but
16239 * they'll be immediately forgotten as we're pruning
16240 * this state and will pop a new one.
16242 err = propagate_liveness(env, &sl->state, cur);
16244 /* if previous state reached the exit with precision and
16245 * current state is equivalent to it (except precsion marks)
16246 * the precision needs to be propagated back in
16247 * the current state.
16249 err = err ? : push_jmp_history(env, cur);
16250 err = err ? : propagate_precision(env, &sl->state);
16256 /* when new state is not going to be added do not increase miss count.
16257 * Otherwise several loop iterations will remove the state
16258 * recorded earlier. The goal of these heuristics is to have
16259 * states from some iterations of the loop (some in the beginning
16260 * and some at the end) to help pruning.
16264 /* heuristic to determine whether this state is beneficial
16265 * to keep checking from state equivalence point of view.
16266 * Higher numbers increase max_states_per_insn and verification time,
16267 * but do not meaningfully decrease insn_processed.
16269 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
16270 /* the state is unlikely to be useful. Remove it to
16271 * speed up verification
16274 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
16275 u32 br = sl->state.branches;
16278 "BUG live_done but branches_to_explore %d\n",
16280 free_verifier_state(&sl->state, false);
16282 env->peak_states--;
16284 /* cannot free this state, since parentage chain may
16285 * walk it later. Add it for free_list instead to
16286 * be freed at the end of verification
16288 sl->next = env->free_list;
16289 env->free_list = sl;
16299 if (env->max_states_per_insn < states_cnt)
16300 env->max_states_per_insn = states_cnt;
16302 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
16305 if (!add_new_state)
16308 /* There were no equivalent states, remember the current one.
16309 * Technically the current state is not proven to be safe yet,
16310 * but it will either reach outer most bpf_exit (which means it's safe)
16311 * or it will be rejected. When there are no loops the verifier won't be
16312 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
16313 * again on the way to bpf_exit.
16314 * When looping the sl->state.branches will be > 0 and this state
16315 * will not be considered for equivalence until branches == 0.
16317 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
16320 env->total_states++;
16321 env->peak_states++;
16322 env->prev_jmps_processed = env->jmps_processed;
16323 env->prev_insn_processed = env->insn_processed;
16325 /* forget precise markings we inherited, see __mark_chain_precision */
16326 if (env->bpf_capable)
16327 mark_all_scalars_imprecise(env, cur);
16329 /* add new state to the head of linked list */
16330 new = &new_sl->state;
16331 err = copy_verifier_state(new, cur);
16333 free_verifier_state(new, false);
16337 new->insn_idx = insn_idx;
16338 WARN_ONCE(new->branches != 1,
16339 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
16342 cur->first_insn_idx = insn_idx;
16343 clear_jmp_history(cur);
16344 new_sl->next = *explored_state(env, insn_idx);
16345 *explored_state(env, insn_idx) = new_sl;
16346 /* connect new state to parentage chain. Current frame needs all
16347 * registers connected. Only r6 - r9 of the callers are alive (pushed
16348 * to the stack implicitly by JITs) so in callers' frames connect just
16349 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
16350 * the state of the call instruction (with WRITTEN set), and r0 comes
16351 * from callee with its full parentage chain, anyway.
16353 /* clear write marks in current state: the writes we did are not writes
16354 * our child did, so they don't screen off its reads from us.
16355 * (There are no read marks in current state, because reads always mark
16356 * their parent and current state never has children yet. Only
16357 * explored_states can get read marks.)
16359 for (j = 0; j <= cur->curframe; j++) {
16360 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
16361 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
16362 for (i = 0; i < BPF_REG_FP; i++)
16363 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
16366 /* all stack frames are accessible from callee, clear them all */
16367 for (j = 0; j <= cur->curframe; j++) {
16368 struct bpf_func_state *frame = cur->frame[j];
16369 struct bpf_func_state *newframe = new->frame[j];
16371 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
16372 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
16373 frame->stack[i].spilled_ptr.parent =
16374 &newframe->stack[i].spilled_ptr;
16380 /* Return true if it's OK to have the same insn return a different type. */
16381 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
16383 switch (base_type(type)) {
16385 case PTR_TO_SOCKET:
16386 case PTR_TO_SOCK_COMMON:
16387 case PTR_TO_TCP_SOCK:
16388 case PTR_TO_XDP_SOCK:
16389 case PTR_TO_BTF_ID:
16396 /* If an instruction was previously used with particular pointer types, then we
16397 * need to be careful to avoid cases such as the below, where it may be ok
16398 * for one branch accessing the pointer, but not ok for the other branch:
16403 * R1 = some_other_valid_ptr;
16406 * R2 = *(u32 *)(R1 + 0);
16408 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
16410 return src != prev && (!reg_type_mismatch_ok(src) ||
16411 !reg_type_mismatch_ok(prev));
16414 static int save_aux_ptr_type(struct bpf_verifier_env *env, enum bpf_reg_type type,
16415 bool allow_trust_missmatch)
16417 enum bpf_reg_type *prev_type = &env->insn_aux_data[env->insn_idx].ptr_type;
16419 if (*prev_type == NOT_INIT) {
16420 /* Saw a valid insn
16421 * dst_reg = *(u32 *)(src_reg + off)
16422 * save type to validate intersecting paths
16425 } else if (reg_type_mismatch(type, *prev_type)) {
16426 /* Abuser program is trying to use the same insn
16427 * dst_reg = *(u32*) (src_reg + off)
16428 * with different pointer types:
16429 * src_reg == ctx in one branch and
16430 * src_reg == stack|map in some other branch.
16433 if (allow_trust_missmatch &&
16434 base_type(type) == PTR_TO_BTF_ID &&
16435 base_type(*prev_type) == PTR_TO_BTF_ID) {
16437 * Have to support a use case when one path through
16438 * the program yields TRUSTED pointer while another
16439 * is UNTRUSTED. Fallback to UNTRUSTED to generate
16440 * BPF_PROBE_MEM/BPF_PROBE_MEMSX.
16442 *prev_type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
16444 verbose(env, "same insn cannot be used with different pointers\n");
16452 static int do_check(struct bpf_verifier_env *env)
16454 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
16455 struct bpf_verifier_state *state = env->cur_state;
16456 struct bpf_insn *insns = env->prog->insnsi;
16457 struct bpf_reg_state *regs;
16458 int insn_cnt = env->prog->len;
16459 bool do_print_state = false;
16460 int prev_insn_idx = -1;
16463 struct bpf_insn *insn;
16467 env->prev_insn_idx = prev_insn_idx;
16468 if (env->insn_idx >= insn_cnt) {
16469 verbose(env, "invalid insn idx %d insn_cnt %d\n",
16470 env->insn_idx, insn_cnt);
16474 insn = &insns[env->insn_idx];
16475 class = BPF_CLASS(insn->code);
16477 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
16479 "BPF program is too large. Processed %d insn\n",
16480 env->insn_processed);
16484 state->last_insn_idx = env->prev_insn_idx;
16486 if (is_prune_point(env, env->insn_idx)) {
16487 err = is_state_visited(env, env->insn_idx);
16491 /* found equivalent state, can prune the search */
16492 if (env->log.level & BPF_LOG_LEVEL) {
16493 if (do_print_state)
16494 verbose(env, "\nfrom %d to %d%s: safe\n",
16495 env->prev_insn_idx, env->insn_idx,
16496 env->cur_state->speculative ?
16497 " (speculative execution)" : "");
16499 verbose(env, "%d: safe\n", env->insn_idx);
16501 goto process_bpf_exit;
16505 if (is_jmp_point(env, env->insn_idx)) {
16506 err = push_jmp_history(env, state);
16511 if (signal_pending(current))
16514 if (need_resched())
16517 if (env->log.level & BPF_LOG_LEVEL2 && do_print_state) {
16518 verbose(env, "\nfrom %d to %d%s:",
16519 env->prev_insn_idx, env->insn_idx,
16520 env->cur_state->speculative ?
16521 " (speculative execution)" : "");
16522 print_verifier_state(env, state->frame[state->curframe], true);
16523 do_print_state = false;
16526 if (env->log.level & BPF_LOG_LEVEL) {
16527 const struct bpf_insn_cbs cbs = {
16528 .cb_call = disasm_kfunc_name,
16529 .cb_print = verbose,
16530 .private_data = env,
16533 if (verifier_state_scratched(env))
16534 print_insn_state(env, state->frame[state->curframe]);
16536 verbose_linfo(env, env->insn_idx, "; ");
16537 env->prev_log_pos = env->log.end_pos;
16538 verbose(env, "%d: ", env->insn_idx);
16539 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
16540 env->prev_insn_print_pos = env->log.end_pos - env->prev_log_pos;
16541 env->prev_log_pos = env->log.end_pos;
16544 if (bpf_prog_is_offloaded(env->prog->aux)) {
16545 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
16546 env->prev_insn_idx);
16551 regs = cur_regs(env);
16552 sanitize_mark_insn_seen(env);
16553 prev_insn_idx = env->insn_idx;
16555 if (class == BPF_ALU || class == BPF_ALU64) {
16556 err = check_alu_op(env, insn);
16560 } else if (class == BPF_LDX) {
16561 enum bpf_reg_type src_reg_type;
16563 /* check for reserved fields is already done */
16565 /* check src operand */
16566 err = check_reg_arg(env, insn->src_reg, SRC_OP);
16570 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
16574 src_reg_type = regs[insn->src_reg].type;
16576 /* check that memory (src_reg + off) is readable,
16577 * the state of dst_reg will be updated by this func
16579 err = check_mem_access(env, env->insn_idx, insn->src_reg,
16580 insn->off, BPF_SIZE(insn->code),
16581 BPF_READ, insn->dst_reg, false,
16582 BPF_MODE(insn->code) == BPF_MEMSX);
16586 err = save_aux_ptr_type(env, src_reg_type, true);
16589 } else if (class == BPF_STX) {
16590 enum bpf_reg_type dst_reg_type;
16592 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
16593 err = check_atomic(env, env->insn_idx, insn);
16600 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
16601 verbose(env, "BPF_STX uses reserved fields\n");
16605 /* check src1 operand */
16606 err = check_reg_arg(env, insn->src_reg, SRC_OP);
16609 /* check src2 operand */
16610 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
16614 dst_reg_type = regs[insn->dst_reg].type;
16616 /* check that memory (dst_reg + off) is writeable */
16617 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
16618 insn->off, BPF_SIZE(insn->code),
16619 BPF_WRITE, insn->src_reg, false, false);
16623 err = save_aux_ptr_type(env, dst_reg_type, false);
16626 } else if (class == BPF_ST) {
16627 enum bpf_reg_type dst_reg_type;
16629 if (BPF_MODE(insn->code) != BPF_MEM ||
16630 insn->src_reg != BPF_REG_0) {
16631 verbose(env, "BPF_ST uses reserved fields\n");
16634 /* check src operand */
16635 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
16639 dst_reg_type = regs[insn->dst_reg].type;
16641 /* check that memory (dst_reg + off) is writeable */
16642 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
16643 insn->off, BPF_SIZE(insn->code),
16644 BPF_WRITE, -1, false, false);
16648 err = save_aux_ptr_type(env, dst_reg_type, false);
16651 } else if (class == BPF_JMP || class == BPF_JMP32) {
16652 u8 opcode = BPF_OP(insn->code);
16654 env->jmps_processed++;
16655 if (opcode == BPF_CALL) {
16656 if (BPF_SRC(insn->code) != BPF_K ||
16657 (insn->src_reg != BPF_PSEUDO_KFUNC_CALL
16658 && insn->off != 0) ||
16659 (insn->src_reg != BPF_REG_0 &&
16660 insn->src_reg != BPF_PSEUDO_CALL &&
16661 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
16662 insn->dst_reg != BPF_REG_0 ||
16663 class == BPF_JMP32) {
16664 verbose(env, "BPF_CALL uses reserved fields\n");
16668 if (env->cur_state->active_lock.ptr) {
16669 if ((insn->src_reg == BPF_REG_0 && insn->imm != BPF_FUNC_spin_unlock) ||
16670 (insn->src_reg == BPF_PSEUDO_CALL) ||
16671 (insn->src_reg == BPF_PSEUDO_KFUNC_CALL &&
16672 (insn->off != 0 || !is_bpf_graph_api_kfunc(insn->imm)))) {
16673 verbose(env, "function calls are not allowed while holding a lock\n");
16677 if (insn->src_reg == BPF_PSEUDO_CALL)
16678 err = check_func_call(env, insn, &env->insn_idx);
16679 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
16680 err = check_kfunc_call(env, insn, &env->insn_idx);
16682 err = check_helper_call(env, insn, &env->insn_idx);
16686 mark_reg_scratched(env, BPF_REG_0);
16687 } else if (opcode == BPF_JA) {
16688 if (BPF_SRC(insn->code) != BPF_K ||
16689 insn->src_reg != BPF_REG_0 ||
16690 insn->dst_reg != BPF_REG_0 ||
16691 (class == BPF_JMP && insn->imm != 0) ||
16692 (class == BPF_JMP32 && insn->off != 0)) {
16693 verbose(env, "BPF_JA uses reserved fields\n");
16697 if (class == BPF_JMP)
16698 env->insn_idx += insn->off + 1;
16700 env->insn_idx += insn->imm + 1;
16703 } else if (opcode == BPF_EXIT) {
16704 if (BPF_SRC(insn->code) != BPF_K ||
16706 insn->src_reg != BPF_REG_0 ||
16707 insn->dst_reg != BPF_REG_0 ||
16708 class == BPF_JMP32) {
16709 verbose(env, "BPF_EXIT uses reserved fields\n");
16713 if (env->cur_state->active_lock.ptr &&
16714 !in_rbtree_lock_required_cb(env)) {
16715 verbose(env, "bpf_spin_unlock is missing\n");
16719 if (env->cur_state->active_rcu_lock &&
16720 !in_rbtree_lock_required_cb(env)) {
16721 verbose(env, "bpf_rcu_read_unlock is missing\n");
16725 /* We must do check_reference_leak here before
16726 * prepare_func_exit to handle the case when
16727 * state->curframe > 0, it may be a callback
16728 * function, for which reference_state must
16729 * match caller reference state when it exits.
16731 err = check_reference_leak(env);
16735 if (state->curframe) {
16736 /* exit from nested function */
16737 err = prepare_func_exit(env, &env->insn_idx);
16740 do_print_state = true;
16744 err = check_return_code(env);
16748 mark_verifier_state_scratched(env);
16749 update_branch_counts(env, env->cur_state);
16750 err = pop_stack(env, &prev_insn_idx,
16751 &env->insn_idx, pop_log);
16753 if (err != -ENOENT)
16757 do_print_state = true;
16761 err = check_cond_jmp_op(env, insn, &env->insn_idx);
16765 } else if (class == BPF_LD) {
16766 u8 mode = BPF_MODE(insn->code);
16768 if (mode == BPF_ABS || mode == BPF_IND) {
16769 err = check_ld_abs(env, insn);
16773 } else if (mode == BPF_IMM) {
16774 err = check_ld_imm(env, insn);
16779 sanitize_mark_insn_seen(env);
16781 verbose(env, "invalid BPF_LD mode\n");
16785 verbose(env, "unknown insn class %d\n", class);
16795 static int find_btf_percpu_datasec(struct btf *btf)
16797 const struct btf_type *t;
16802 * Both vmlinux and module each have their own ".data..percpu"
16803 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
16804 * types to look at only module's own BTF types.
16806 n = btf_nr_types(btf);
16807 if (btf_is_module(btf))
16808 i = btf_nr_types(btf_vmlinux);
16812 for(; i < n; i++) {
16813 t = btf_type_by_id(btf, i);
16814 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
16817 tname = btf_name_by_offset(btf, t->name_off);
16818 if (!strcmp(tname, ".data..percpu"))
16825 /* replace pseudo btf_id with kernel symbol address */
16826 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
16827 struct bpf_insn *insn,
16828 struct bpf_insn_aux_data *aux)
16830 const struct btf_var_secinfo *vsi;
16831 const struct btf_type *datasec;
16832 struct btf_mod_pair *btf_mod;
16833 const struct btf_type *t;
16834 const char *sym_name;
16835 bool percpu = false;
16836 u32 type, id = insn->imm;
16840 int i, btf_fd, err;
16842 btf_fd = insn[1].imm;
16844 btf = btf_get_by_fd(btf_fd);
16846 verbose(env, "invalid module BTF object FD specified.\n");
16850 if (!btf_vmlinux) {
16851 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
16858 t = btf_type_by_id(btf, id);
16860 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
16865 if (!btf_type_is_var(t) && !btf_type_is_func(t)) {
16866 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR or KIND_FUNC\n", id);
16871 sym_name = btf_name_by_offset(btf, t->name_off);
16872 addr = kallsyms_lookup_name(sym_name);
16874 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
16879 insn[0].imm = (u32)addr;
16880 insn[1].imm = addr >> 32;
16882 if (btf_type_is_func(t)) {
16883 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
16884 aux->btf_var.mem_size = 0;
16888 datasec_id = find_btf_percpu_datasec(btf);
16889 if (datasec_id > 0) {
16890 datasec = btf_type_by_id(btf, datasec_id);
16891 for_each_vsi(i, datasec, vsi) {
16892 if (vsi->type == id) {
16900 t = btf_type_skip_modifiers(btf, type, NULL);
16902 aux->btf_var.reg_type = PTR_TO_BTF_ID | MEM_PERCPU;
16903 aux->btf_var.btf = btf;
16904 aux->btf_var.btf_id = type;
16905 } else if (!btf_type_is_struct(t)) {
16906 const struct btf_type *ret;
16910 /* resolve the type size of ksym. */
16911 ret = btf_resolve_size(btf, t, &tsize);
16913 tname = btf_name_by_offset(btf, t->name_off);
16914 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
16915 tname, PTR_ERR(ret));
16919 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
16920 aux->btf_var.mem_size = tsize;
16922 aux->btf_var.reg_type = PTR_TO_BTF_ID;
16923 aux->btf_var.btf = btf;
16924 aux->btf_var.btf_id = type;
16927 /* check whether we recorded this BTF (and maybe module) already */
16928 for (i = 0; i < env->used_btf_cnt; i++) {
16929 if (env->used_btfs[i].btf == btf) {
16935 if (env->used_btf_cnt >= MAX_USED_BTFS) {
16940 btf_mod = &env->used_btfs[env->used_btf_cnt];
16941 btf_mod->btf = btf;
16942 btf_mod->module = NULL;
16944 /* if we reference variables from kernel module, bump its refcount */
16945 if (btf_is_module(btf)) {
16946 btf_mod->module = btf_try_get_module(btf);
16947 if (!btf_mod->module) {
16953 env->used_btf_cnt++;
16961 static bool is_tracing_prog_type(enum bpf_prog_type type)
16964 case BPF_PROG_TYPE_KPROBE:
16965 case BPF_PROG_TYPE_TRACEPOINT:
16966 case BPF_PROG_TYPE_PERF_EVENT:
16967 case BPF_PROG_TYPE_RAW_TRACEPOINT:
16968 case BPF_PROG_TYPE_RAW_TRACEPOINT_WRITABLE:
16975 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
16976 struct bpf_map *map,
16977 struct bpf_prog *prog)
16980 enum bpf_prog_type prog_type = resolve_prog_type(prog);
16982 if (btf_record_has_field(map->record, BPF_LIST_HEAD) ||
16983 btf_record_has_field(map->record, BPF_RB_ROOT)) {
16984 if (is_tracing_prog_type(prog_type)) {
16985 verbose(env, "tracing progs cannot use bpf_{list_head,rb_root} yet\n");
16990 if (btf_record_has_field(map->record, BPF_SPIN_LOCK)) {
16991 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
16992 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
16996 if (is_tracing_prog_type(prog_type)) {
16997 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
17002 if (btf_record_has_field(map->record, BPF_TIMER)) {
17003 if (is_tracing_prog_type(prog_type)) {
17004 verbose(env, "tracing progs cannot use bpf_timer yet\n");
17009 if ((bpf_prog_is_offloaded(prog->aux) || bpf_map_is_offloaded(map)) &&
17010 !bpf_offload_prog_map_match(prog, map)) {
17011 verbose(env, "offload device mismatch between prog and map\n");
17015 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
17016 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
17020 if (prog->aux->sleepable)
17021 switch (map->map_type) {
17022 case BPF_MAP_TYPE_HASH:
17023 case BPF_MAP_TYPE_LRU_HASH:
17024 case BPF_MAP_TYPE_ARRAY:
17025 case BPF_MAP_TYPE_PERCPU_HASH:
17026 case BPF_MAP_TYPE_PERCPU_ARRAY:
17027 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
17028 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
17029 case BPF_MAP_TYPE_HASH_OF_MAPS:
17030 case BPF_MAP_TYPE_RINGBUF:
17031 case BPF_MAP_TYPE_USER_RINGBUF:
17032 case BPF_MAP_TYPE_INODE_STORAGE:
17033 case BPF_MAP_TYPE_SK_STORAGE:
17034 case BPF_MAP_TYPE_TASK_STORAGE:
17035 case BPF_MAP_TYPE_CGRP_STORAGE:
17039 "Sleepable programs can only use array, hash, ringbuf and local storage maps\n");
17046 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
17048 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
17049 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
17052 /* find and rewrite pseudo imm in ld_imm64 instructions:
17054 * 1. if it accesses map FD, replace it with actual map pointer.
17055 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
17057 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
17059 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
17061 struct bpf_insn *insn = env->prog->insnsi;
17062 int insn_cnt = env->prog->len;
17065 err = bpf_prog_calc_tag(env->prog);
17069 for (i = 0; i < insn_cnt; i++, insn++) {
17070 if (BPF_CLASS(insn->code) == BPF_LDX &&
17071 ((BPF_MODE(insn->code) != BPF_MEM && BPF_MODE(insn->code) != BPF_MEMSX) ||
17073 verbose(env, "BPF_LDX uses reserved fields\n");
17077 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
17078 struct bpf_insn_aux_data *aux;
17079 struct bpf_map *map;
17084 if (i == insn_cnt - 1 || insn[1].code != 0 ||
17085 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
17086 insn[1].off != 0) {
17087 verbose(env, "invalid bpf_ld_imm64 insn\n");
17091 if (insn[0].src_reg == 0)
17092 /* valid generic load 64-bit imm */
17095 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
17096 aux = &env->insn_aux_data[i];
17097 err = check_pseudo_btf_id(env, insn, aux);
17103 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
17104 aux = &env->insn_aux_data[i];
17105 aux->ptr_type = PTR_TO_FUNC;
17109 /* In final convert_pseudo_ld_imm64() step, this is
17110 * converted into regular 64-bit imm load insn.
17112 switch (insn[0].src_reg) {
17113 case BPF_PSEUDO_MAP_VALUE:
17114 case BPF_PSEUDO_MAP_IDX_VALUE:
17116 case BPF_PSEUDO_MAP_FD:
17117 case BPF_PSEUDO_MAP_IDX:
17118 if (insn[1].imm == 0)
17122 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
17126 switch (insn[0].src_reg) {
17127 case BPF_PSEUDO_MAP_IDX_VALUE:
17128 case BPF_PSEUDO_MAP_IDX:
17129 if (bpfptr_is_null(env->fd_array)) {
17130 verbose(env, "fd_idx without fd_array is invalid\n");
17133 if (copy_from_bpfptr_offset(&fd, env->fd_array,
17134 insn[0].imm * sizeof(fd),
17144 map = __bpf_map_get(f);
17146 verbose(env, "fd %d is not pointing to valid bpf_map\n",
17148 return PTR_ERR(map);
17151 err = check_map_prog_compatibility(env, map, env->prog);
17157 aux = &env->insn_aux_data[i];
17158 if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
17159 insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
17160 addr = (unsigned long)map;
17162 u32 off = insn[1].imm;
17164 if (off >= BPF_MAX_VAR_OFF) {
17165 verbose(env, "direct value offset of %u is not allowed\n", off);
17170 if (!map->ops->map_direct_value_addr) {
17171 verbose(env, "no direct value access support for this map type\n");
17176 err = map->ops->map_direct_value_addr(map, &addr, off);
17178 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
17179 map->value_size, off);
17184 aux->map_off = off;
17188 insn[0].imm = (u32)addr;
17189 insn[1].imm = addr >> 32;
17191 /* check whether we recorded this map already */
17192 for (j = 0; j < env->used_map_cnt; j++) {
17193 if (env->used_maps[j] == map) {
17194 aux->map_index = j;
17200 if (env->used_map_cnt >= MAX_USED_MAPS) {
17205 /* hold the map. If the program is rejected by verifier,
17206 * the map will be released by release_maps() or it
17207 * will be used by the valid program until it's unloaded
17208 * and all maps are released in free_used_maps()
17212 aux->map_index = env->used_map_cnt;
17213 env->used_maps[env->used_map_cnt++] = map;
17215 if (bpf_map_is_cgroup_storage(map) &&
17216 bpf_cgroup_storage_assign(env->prog->aux, map)) {
17217 verbose(env, "only one cgroup storage of each type is allowed\n");
17229 /* Basic sanity check before we invest more work here. */
17230 if (!bpf_opcode_in_insntable(insn->code)) {
17231 verbose(env, "unknown opcode %02x\n", insn->code);
17236 /* now all pseudo BPF_LD_IMM64 instructions load valid
17237 * 'struct bpf_map *' into a register instead of user map_fd.
17238 * These pointers will be used later by verifier to validate map access.
17243 /* drop refcnt of maps used by the rejected program */
17244 static void release_maps(struct bpf_verifier_env *env)
17246 __bpf_free_used_maps(env->prog->aux, env->used_maps,
17247 env->used_map_cnt);
17250 /* drop refcnt of maps used by the rejected program */
17251 static void release_btfs(struct bpf_verifier_env *env)
17253 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
17254 env->used_btf_cnt);
17257 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
17258 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
17260 struct bpf_insn *insn = env->prog->insnsi;
17261 int insn_cnt = env->prog->len;
17264 for (i = 0; i < insn_cnt; i++, insn++) {
17265 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
17267 if (insn->src_reg == BPF_PSEUDO_FUNC)
17273 /* single env->prog->insni[off] instruction was replaced with the range
17274 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
17275 * [0, off) and [off, end) to new locations, so the patched range stays zero
17277 static void adjust_insn_aux_data(struct bpf_verifier_env *env,
17278 struct bpf_insn_aux_data *new_data,
17279 struct bpf_prog *new_prog, u32 off, u32 cnt)
17281 struct bpf_insn_aux_data *old_data = env->insn_aux_data;
17282 struct bpf_insn *insn = new_prog->insnsi;
17283 u32 old_seen = old_data[off].seen;
17287 /* aux info at OFF always needs adjustment, no matter fast path
17288 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
17289 * original insn at old prog.
17291 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
17295 prog_len = new_prog->len;
17297 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
17298 memcpy(new_data + off + cnt - 1, old_data + off,
17299 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
17300 for (i = off; i < off + cnt - 1; i++) {
17301 /* Expand insni[off]'s seen count to the patched range. */
17302 new_data[i].seen = old_seen;
17303 new_data[i].zext_dst = insn_has_def32(env, insn + i);
17305 env->insn_aux_data = new_data;
17309 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
17315 /* NOTE: fake 'exit' subprog should be updated as well. */
17316 for (i = 0; i <= env->subprog_cnt; i++) {
17317 if (env->subprog_info[i].start <= off)
17319 env->subprog_info[i].start += len - 1;
17323 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
17325 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
17326 int i, sz = prog->aux->size_poke_tab;
17327 struct bpf_jit_poke_descriptor *desc;
17329 for (i = 0; i < sz; i++) {
17331 if (desc->insn_idx <= off)
17333 desc->insn_idx += len - 1;
17337 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
17338 const struct bpf_insn *patch, u32 len)
17340 struct bpf_prog *new_prog;
17341 struct bpf_insn_aux_data *new_data = NULL;
17344 new_data = vzalloc(array_size(env->prog->len + len - 1,
17345 sizeof(struct bpf_insn_aux_data)));
17350 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
17351 if (IS_ERR(new_prog)) {
17352 if (PTR_ERR(new_prog) == -ERANGE)
17354 "insn %d cannot be patched due to 16-bit range\n",
17355 env->insn_aux_data[off].orig_idx);
17359 adjust_insn_aux_data(env, new_data, new_prog, off, len);
17360 adjust_subprog_starts(env, off, len);
17361 adjust_poke_descs(new_prog, off, len);
17365 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
17370 /* find first prog starting at or after off (first to remove) */
17371 for (i = 0; i < env->subprog_cnt; i++)
17372 if (env->subprog_info[i].start >= off)
17374 /* find first prog starting at or after off + cnt (first to stay) */
17375 for (j = i; j < env->subprog_cnt; j++)
17376 if (env->subprog_info[j].start >= off + cnt)
17378 /* if j doesn't start exactly at off + cnt, we are just removing
17379 * the front of previous prog
17381 if (env->subprog_info[j].start != off + cnt)
17385 struct bpf_prog_aux *aux = env->prog->aux;
17388 /* move fake 'exit' subprog as well */
17389 move = env->subprog_cnt + 1 - j;
17391 memmove(env->subprog_info + i,
17392 env->subprog_info + j,
17393 sizeof(*env->subprog_info) * move);
17394 env->subprog_cnt -= j - i;
17396 /* remove func_info */
17397 if (aux->func_info) {
17398 move = aux->func_info_cnt - j;
17400 memmove(aux->func_info + i,
17401 aux->func_info + j,
17402 sizeof(*aux->func_info) * move);
17403 aux->func_info_cnt -= j - i;
17404 /* func_info->insn_off is set after all code rewrites,
17405 * in adjust_btf_func() - no need to adjust
17409 /* convert i from "first prog to remove" to "first to adjust" */
17410 if (env->subprog_info[i].start == off)
17414 /* update fake 'exit' subprog as well */
17415 for (; i <= env->subprog_cnt; i++)
17416 env->subprog_info[i].start -= cnt;
17421 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
17424 struct bpf_prog *prog = env->prog;
17425 u32 i, l_off, l_cnt, nr_linfo;
17426 struct bpf_line_info *linfo;
17428 nr_linfo = prog->aux->nr_linfo;
17432 linfo = prog->aux->linfo;
17434 /* find first line info to remove, count lines to be removed */
17435 for (i = 0; i < nr_linfo; i++)
17436 if (linfo[i].insn_off >= off)
17441 for (; i < nr_linfo; i++)
17442 if (linfo[i].insn_off < off + cnt)
17447 /* First live insn doesn't match first live linfo, it needs to "inherit"
17448 * last removed linfo. prog is already modified, so prog->len == off
17449 * means no live instructions after (tail of the program was removed).
17451 if (prog->len != off && l_cnt &&
17452 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
17454 linfo[--i].insn_off = off + cnt;
17457 /* remove the line info which refer to the removed instructions */
17459 memmove(linfo + l_off, linfo + i,
17460 sizeof(*linfo) * (nr_linfo - i));
17462 prog->aux->nr_linfo -= l_cnt;
17463 nr_linfo = prog->aux->nr_linfo;
17466 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
17467 for (i = l_off; i < nr_linfo; i++)
17468 linfo[i].insn_off -= cnt;
17470 /* fix up all subprogs (incl. 'exit') which start >= off */
17471 for (i = 0; i <= env->subprog_cnt; i++)
17472 if (env->subprog_info[i].linfo_idx > l_off) {
17473 /* program may have started in the removed region but
17474 * may not be fully removed
17476 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
17477 env->subprog_info[i].linfo_idx -= l_cnt;
17479 env->subprog_info[i].linfo_idx = l_off;
17485 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
17487 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17488 unsigned int orig_prog_len = env->prog->len;
17491 if (bpf_prog_is_offloaded(env->prog->aux))
17492 bpf_prog_offload_remove_insns(env, off, cnt);
17494 err = bpf_remove_insns(env->prog, off, cnt);
17498 err = adjust_subprog_starts_after_remove(env, off, cnt);
17502 err = bpf_adj_linfo_after_remove(env, off, cnt);
17506 memmove(aux_data + off, aux_data + off + cnt,
17507 sizeof(*aux_data) * (orig_prog_len - off - cnt));
17512 /* The verifier does more data flow analysis than llvm and will not
17513 * explore branches that are dead at run time. Malicious programs can
17514 * have dead code too. Therefore replace all dead at-run-time code
17517 * Just nops are not optimal, e.g. if they would sit at the end of the
17518 * program and through another bug we would manage to jump there, then
17519 * we'd execute beyond program memory otherwise. Returning exception
17520 * code also wouldn't work since we can have subprogs where the dead
17521 * code could be located.
17523 static void sanitize_dead_code(struct bpf_verifier_env *env)
17525 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17526 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
17527 struct bpf_insn *insn = env->prog->insnsi;
17528 const int insn_cnt = env->prog->len;
17531 for (i = 0; i < insn_cnt; i++) {
17532 if (aux_data[i].seen)
17534 memcpy(insn + i, &trap, sizeof(trap));
17535 aux_data[i].zext_dst = false;
17539 static bool insn_is_cond_jump(u8 code)
17544 if (BPF_CLASS(code) == BPF_JMP32)
17545 return op != BPF_JA;
17547 if (BPF_CLASS(code) != BPF_JMP)
17550 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
17553 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
17555 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17556 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
17557 struct bpf_insn *insn = env->prog->insnsi;
17558 const int insn_cnt = env->prog->len;
17561 for (i = 0; i < insn_cnt; i++, insn++) {
17562 if (!insn_is_cond_jump(insn->code))
17565 if (!aux_data[i + 1].seen)
17566 ja.off = insn->off;
17567 else if (!aux_data[i + 1 + insn->off].seen)
17572 if (bpf_prog_is_offloaded(env->prog->aux))
17573 bpf_prog_offload_replace_insn(env, i, &ja);
17575 memcpy(insn, &ja, sizeof(ja));
17579 static int opt_remove_dead_code(struct bpf_verifier_env *env)
17581 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17582 int insn_cnt = env->prog->len;
17585 for (i = 0; i < insn_cnt; i++) {
17589 while (i + j < insn_cnt && !aux_data[i + j].seen)
17594 err = verifier_remove_insns(env, i, j);
17597 insn_cnt = env->prog->len;
17603 static int opt_remove_nops(struct bpf_verifier_env *env)
17605 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
17606 struct bpf_insn *insn = env->prog->insnsi;
17607 int insn_cnt = env->prog->len;
17610 for (i = 0; i < insn_cnt; i++) {
17611 if (memcmp(&insn[i], &ja, sizeof(ja)))
17614 err = verifier_remove_insns(env, i, 1);
17624 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
17625 const union bpf_attr *attr)
17627 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
17628 struct bpf_insn_aux_data *aux = env->insn_aux_data;
17629 int i, patch_len, delta = 0, len = env->prog->len;
17630 struct bpf_insn *insns = env->prog->insnsi;
17631 struct bpf_prog *new_prog;
17634 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
17635 zext_patch[1] = BPF_ZEXT_REG(0);
17636 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
17637 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
17638 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
17639 for (i = 0; i < len; i++) {
17640 int adj_idx = i + delta;
17641 struct bpf_insn insn;
17644 insn = insns[adj_idx];
17645 load_reg = insn_def_regno(&insn);
17646 if (!aux[adj_idx].zext_dst) {
17654 class = BPF_CLASS(code);
17655 if (load_reg == -1)
17658 /* NOTE: arg "reg" (the fourth one) is only used for
17659 * BPF_STX + SRC_OP, so it is safe to pass NULL
17662 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
17663 if (class == BPF_LD &&
17664 BPF_MODE(code) == BPF_IMM)
17669 /* ctx load could be transformed into wider load. */
17670 if (class == BPF_LDX &&
17671 aux[adj_idx].ptr_type == PTR_TO_CTX)
17674 imm_rnd = get_random_u32();
17675 rnd_hi32_patch[0] = insn;
17676 rnd_hi32_patch[1].imm = imm_rnd;
17677 rnd_hi32_patch[3].dst_reg = load_reg;
17678 patch = rnd_hi32_patch;
17680 goto apply_patch_buffer;
17683 /* Add in an zero-extend instruction if a) the JIT has requested
17684 * it or b) it's a CMPXCHG.
17686 * The latter is because: BPF_CMPXCHG always loads a value into
17687 * R0, therefore always zero-extends. However some archs'
17688 * equivalent instruction only does this load when the
17689 * comparison is successful. This detail of CMPXCHG is
17690 * orthogonal to the general zero-extension behaviour of the
17691 * CPU, so it's treated independently of bpf_jit_needs_zext.
17693 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
17696 /* Zero-extension is done by the caller. */
17697 if (bpf_pseudo_kfunc_call(&insn))
17700 if (WARN_ON(load_reg == -1)) {
17701 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
17705 zext_patch[0] = insn;
17706 zext_patch[1].dst_reg = load_reg;
17707 zext_patch[1].src_reg = load_reg;
17708 patch = zext_patch;
17710 apply_patch_buffer:
17711 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
17714 env->prog = new_prog;
17715 insns = new_prog->insnsi;
17716 aux = env->insn_aux_data;
17717 delta += patch_len - 1;
17723 /* convert load instructions that access fields of a context type into a
17724 * sequence of instructions that access fields of the underlying structure:
17725 * struct __sk_buff -> struct sk_buff
17726 * struct bpf_sock_ops -> struct sock
17728 static int convert_ctx_accesses(struct bpf_verifier_env *env)
17730 const struct bpf_verifier_ops *ops = env->ops;
17731 int i, cnt, size, ctx_field_size, delta = 0;
17732 const int insn_cnt = env->prog->len;
17733 struct bpf_insn insn_buf[16], *insn;
17734 u32 target_size, size_default, off;
17735 struct bpf_prog *new_prog;
17736 enum bpf_access_type type;
17737 bool is_narrower_load;
17739 if (ops->gen_prologue || env->seen_direct_write) {
17740 if (!ops->gen_prologue) {
17741 verbose(env, "bpf verifier is misconfigured\n");
17744 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
17746 if (cnt >= ARRAY_SIZE(insn_buf)) {
17747 verbose(env, "bpf verifier is misconfigured\n");
17750 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
17754 env->prog = new_prog;
17759 if (bpf_prog_is_offloaded(env->prog->aux))
17762 insn = env->prog->insnsi + delta;
17764 for (i = 0; i < insn_cnt; i++, insn++) {
17765 bpf_convert_ctx_access_t convert_ctx_access;
17768 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
17769 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
17770 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
17771 insn->code == (BPF_LDX | BPF_MEM | BPF_DW) ||
17772 insn->code == (BPF_LDX | BPF_MEMSX | BPF_B) ||
17773 insn->code == (BPF_LDX | BPF_MEMSX | BPF_H) ||
17774 insn->code == (BPF_LDX | BPF_MEMSX | BPF_W)) {
17776 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
17777 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
17778 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
17779 insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
17780 insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
17781 insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
17782 insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
17783 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
17789 if (type == BPF_WRITE &&
17790 env->insn_aux_data[i + delta].sanitize_stack_spill) {
17791 struct bpf_insn patch[] = {
17796 cnt = ARRAY_SIZE(patch);
17797 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
17802 env->prog = new_prog;
17803 insn = new_prog->insnsi + i + delta;
17807 switch ((int)env->insn_aux_data[i + delta].ptr_type) {
17809 if (!ops->convert_ctx_access)
17811 convert_ctx_access = ops->convert_ctx_access;
17813 case PTR_TO_SOCKET:
17814 case PTR_TO_SOCK_COMMON:
17815 convert_ctx_access = bpf_sock_convert_ctx_access;
17817 case PTR_TO_TCP_SOCK:
17818 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
17820 case PTR_TO_XDP_SOCK:
17821 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
17823 case PTR_TO_BTF_ID:
17824 case PTR_TO_BTF_ID | PTR_UNTRUSTED:
17825 /* PTR_TO_BTF_ID | MEM_ALLOC always has a valid lifetime, unlike
17826 * PTR_TO_BTF_ID, and an active ref_obj_id, but the same cannot
17827 * be said once it is marked PTR_UNTRUSTED, hence we must handle
17828 * any faults for loads into such types. BPF_WRITE is disallowed
17831 case PTR_TO_BTF_ID | MEM_ALLOC | PTR_UNTRUSTED:
17832 if (type == BPF_READ) {
17833 if (BPF_MODE(insn->code) == BPF_MEM)
17834 insn->code = BPF_LDX | BPF_PROBE_MEM |
17835 BPF_SIZE((insn)->code);
17837 insn->code = BPF_LDX | BPF_PROBE_MEMSX |
17838 BPF_SIZE((insn)->code);
17839 env->prog->aux->num_exentries++;
17846 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
17847 size = BPF_LDST_BYTES(insn);
17848 mode = BPF_MODE(insn->code);
17850 /* If the read access is a narrower load of the field,
17851 * convert to a 4/8-byte load, to minimum program type specific
17852 * convert_ctx_access changes. If conversion is successful,
17853 * we will apply proper mask to the result.
17855 is_narrower_load = size < ctx_field_size;
17856 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
17858 if (is_narrower_load) {
17861 if (type == BPF_WRITE) {
17862 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
17867 if (ctx_field_size == 4)
17869 else if (ctx_field_size == 8)
17870 size_code = BPF_DW;
17872 insn->off = off & ~(size_default - 1);
17873 insn->code = BPF_LDX | BPF_MEM | size_code;
17877 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
17879 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
17880 (ctx_field_size && !target_size)) {
17881 verbose(env, "bpf verifier is misconfigured\n");
17885 if (is_narrower_load && size < target_size) {
17886 u8 shift = bpf_ctx_narrow_access_offset(
17887 off, size, size_default) * 8;
17888 if (shift && cnt + 1 >= ARRAY_SIZE(insn_buf)) {
17889 verbose(env, "bpf verifier narrow ctx load misconfigured\n");
17892 if (ctx_field_size <= 4) {
17894 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
17897 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
17898 (1 << size * 8) - 1);
17901 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
17904 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
17905 (1ULL << size * 8) - 1);
17908 if (mode == BPF_MEMSX)
17909 insn_buf[cnt++] = BPF_RAW_INSN(BPF_ALU64 | BPF_MOV | BPF_X,
17910 insn->dst_reg, insn->dst_reg,
17913 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
17919 /* keep walking new program and skip insns we just inserted */
17920 env->prog = new_prog;
17921 insn = new_prog->insnsi + i + delta;
17927 static int jit_subprogs(struct bpf_verifier_env *env)
17929 struct bpf_prog *prog = env->prog, **func, *tmp;
17930 int i, j, subprog_start, subprog_end = 0, len, subprog;
17931 struct bpf_map *map_ptr;
17932 struct bpf_insn *insn;
17933 void *old_bpf_func;
17934 int err, num_exentries;
17936 if (env->subprog_cnt <= 1)
17939 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
17940 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn))
17943 /* Upon error here we cannot fall back to interpreter but
17944 * need a hard reject of the program. Thus -EFAULT is
17945 * propagated in any case.
17947 subprog = find_subprog(env, i + insn->imm + 1);
17949 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
17950 i + insn->imm + 1);
17953 /* temporarily remember subprog id inside insn instead of
17954 * aux_data, since next loop will split up all insns into funcs
17956 insn->off = subprog;
17957 /* remember original imm in case JIT fails and fallback
17958 * to interpreter will be needed
17960 env->insn_aux_data[i].call_imm = insn->imm;
17961 /* point imm to __bpf_call_base+1 from JITs point of view */
17963 if (bpf_pseudo_func(insn))
17964 /* jit (e.g. x86_64) may emit fewer instructions
17965 * if it learns a u32 imm is the same as a u64 imm.
17966 * Force a non zero here.
17971 err = bpf_prog_alloc_jited_linfo(prog);
17973 goto out_undo_insn;
17976 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
17978 goto out_undo_insn;
17980 for (i = 0; i < env->subprog_cnt; i++) {
17981 subprog_start = subprog_end;
17982 subprog_end = env->subprog_info[i + 1].start;
17984 len = subprog_end - subprog_start;
17985 /* bpf_prog_run() doesn't call subprogs directly,
17986 * hence main prog stats include the runtime of subprogs.
17987 * subprogs don't have IDs and not reachable via prog_get_next_id
17988 * func[i]->stats will never be accessed and stays NULL
17990 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
17993 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
17994 len * sizeof(struct bpf_insn));
17995 func[i]->type = prog->type;
17996 func[i]->len = len;
17997 if (bpf_prog_calc_tag(func[i]))
17999 func[i]->is_func = 1;
18000 func[i]->aux->func_idx = i;
18001 /* Below members will be freed only at prog->aux */
18002 func[i]->aux->btf = prog->aux->btf;
18003 func[i]->aux->func_info = prog->aux->func_info;
18004 func[i]->aux->func_info_cnt = prog->aux->func_info_cnt;
18005 func[i]->aux->poke_tab = prog->aux->poke_tab;
18006 func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
18008 for (j = 0; j < prog->aux->size_poke_tab; j++) {
18009 struct bpf_jit_poke_descriptor *poke;
18011 poke = &prog->aux->poke_tab[j];
18012 if (poke->insn_idx < subprog_end &&
18013 poke->insn_idx >= subprog_start)
18014 poke->aux = func[i]->aux;
18017 func[i]->aux->name[0] = 'F';
18018 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
18019 func[i]->jit_requested = 1;
18020 func[i]->blinding_requested = prog->blinding_requested;
18021 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
18022 func[i]->aux->kfunc_btf_tab = prog->aux->kfunc_btf_tab;
18023 func[i]->aux->linfo = prog->aux->linfo;
18024 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
18025 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
18026 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
18028 insn = func[i]->insnsi;
18029 for (j = 0; j < func[i]->len; j++, insn++) {
18030 if (BPF_CLASS(insn->code) == BPF_LDX &&
18031 (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
18032 BPF_MODE(insn->code) == BPF_PROBE_MEMSX))
18035 func[i]->aux->num_exentries = num_exentries;
18036 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
18037 func[i] = bpf_int_jit_compile(func[i]);
18038 if (!func[i]->jited) {
18045 /* at this point all bpf functions were successfully JITed
18046 * now populate all bpf_calls with correct addresses and
18047 * run last pass of JIT
18049 for (i = 0; i < env->subprog_cnt; i++) {
18050 insn = func[i]->insnsi;
18051 for (j = 0; j < func[i]->len; j++, insn++) {
18052 if (bpf_pseudo_func(insn)) {
18053 subprog = insn->off;
18054 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
18055 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
18058 if (!bpf_pseudo_call(insn))
18060 subprog = insn->off;
18061 insn->imm = BPF_CALL_IMM(func[subprog]->bpf_func);
18064 /* we use the aux data to keep a list of the start addresses
18065 * of the JITed images for each function in the program
18067 * for some architectures, such as powerpc64, the imm field
18068 * might not be large enough to hold the offset of the start
18069 * address of the callee's JITed image from __bpf_call_base
18071 * in such cases, we can lookup the start address of a callee
18072 * by using its subprog id, available from the off field of
18073 * the call instruction, as an index for this list
18075 func[i]->aux->func = func;
18076 func[i]->aux->func_cnt = env->subprog_cnt;
18078 for (i = 0; i < env->subprog_cnt; i++) {
18079 old_bpf_func = func[i]->bpf_func;
18080 tmp = bpf_int_jit_compile(func[i]);
18081 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
18082 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
18089 /* finally lock prog and jit images for all functions and
18090 * populate kallsysm. Begin at the first subprogram, since
18091 * bpf_prog_load will add the kallsyms for the main program.
18093 for (i = 1; i < env->subprog_cnt; i++) {
18094 bpf_prog_lock_ro(func[i]);
18095 bpf_prog_kallsyms_add(func[i]);
18098 /* Last step: make now unused interpreter insns from main
18099 * prog consistent for later dump requests, so they can
18100 * later look the same as if they were interpreted only.
18102 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
18103 if (bpf_pseudo_func(insn)) {
18104 insn[0].imm = env->insn_aux_data[i].call_imm;
18105 insn[1].imm = insn->off;
18109 if (!bpf_pseudo_call(insn))
18111 insn->off = env->insn_aux_data[i].call_imm;
18112 subprog = find_subprog(env, i + insn->off + 1);
18113 insn->imm = subprog;
18117 prog->bpf_func = func[0]->bpf_func;
18118 prog->jited_len = func[0]->jited_len;
18119 prog->aux->extable = func[0]->aux->extable;
18120 prog->aux->num_exentries = func[0]->aux->num_exentries;
18121 prog->aux->func = func;
18122 prog->aux->func_cnt = env->subprog_cnt;
18123 bpf_prog_jit_attempt_done(prog);
18126 /* We failed JIT'ing, so at this point we need to unregister poke
18127 * descriptors from subprogs, so that kernel is not attempting to
18128 * patch it anymore as we're freeing the subprog JIT memory.
18130 for (i = 0; i < prog->aux->size_poke_tab; i++) {
18131 map_ptr = prog->aux->poke_tab[i].tail_call.map;
18132 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
18134 /* At this point we're guaranteed that poke descriptors are not
18135 * live anymore. We can just unlink its descriptor table as it's
18136 * released with the main prog.
18138 for (i = 0; i < env->subprog_cnt; i++) {
18141 func[i]->aux->poke_tab = NULL;
18142 bpf_jit_free(func[i]);
18146 /* cleanup main prog to be interpreted */
18147 prog->jit_requested = 0;
18148 prog->blinding_requested = 0;
18149 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
18150 if (!bpf_pseudo_call(insn))
18153 insn->imm = env->insn_aux_data[i].call_imm;
18155 bpf_prog_jit_attempt_done(prog);
18159 static int fixup_call_args(struct bpf_verifier_env *env)
18161 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
18162 struct bpf_prog *prog = env->prog;
18163 struct bpf_insn *insn = prog->insnsi;
18164 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
18169 if (env->prog->jit_requested &&
18170 !bpf_prog_is_offloaded(env->prog->aux)) {
18171 err = jit_subprogs(env);
18174 if (err == -EFAULT)
18177 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
18178 if (has_kfunc_call) {
18179 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
18182 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
18183 /* When JIT fails the progs with bpf2bpf calls and tail_calls
18184 * have to be rejected, since interpreter doesn't support them yet.
18186 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
18189 for (i = 0; i < prog->len; i++, insn++) {
18190 if (bpf_pseudo_func(insn)) {
18191 /* When JIT fails the progs with callback calls
18192 * have to be rejected, since interpreter doesn't support them yet.
18194 verbose(env, "callbacks are not allowed in non-JITed programs\n");
18198 if (!bpf_pseudo_call(insn))
18200 depth = get_callee_stack_depth(env, insn, i);
18203 bpf_patch_call_args(insn, depth);
18210 /* replace a generic kfunc with a specialized version if necessary */
18211 static void specialize_kfunc(struct bpf_verifier_env *env,
18212 u32 func_id, u16 offset, unsigned long *addr)
18214 struct bpf_prog *prog = env->prog;
18215 bool seen_direct_write;
18219 if (bpf_dev_bound_kfunc_id(func_id)) {
18220 xdp_kfunc = bpf_dev_bound_resolve_kfunc(prog, func_id);
18222 *addr = (unsigned long)xdp_kfunc;
18225 /* fallback to default kfunc when not supported by netdev */
18231 if (func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) {
18232 seen_direct_write = env->seen_direct_write;
18233 is_rdonly = !may_access_direct_pkt_data(env, NULL, BPF_WRITE);
18236 *addr = (unsigned long)bpf_dynptr_from_skb_rdonly;
18238 /* restore env->seen_direct_write to its original value, since
18239 * may_access_direct_pkt_data mutates it
18241 env->seen_direct_write = seen_direct_write;
18245 static void __fixup_collection_insert_kfunc(struct bpf_insn_aux_data *insn_aux,
18246 u16 struct_meta_reg,
18247 u16 node_offset_reg,
18248 struct bpf_insn *insn,
18249 struct bpf_insn *insn_buf,
18252 struct btf_struct_meta *kptr_struct_meta = insn_aux->kptr_struct_meta;
18253 struct bpf_insn addr[2] = { BPF_LD_IMM64(struct_meta_reg, (long)kptr_struct_meta) };
18255 insn_buf[0] = addr[0];
18256 insn_buf[1] = addr[1];
18257 insn_buf[2] = BPF_MOV64_IMM(node_offset_reg, insn_aux->insert_off);
18258 insn_buf[3] = *insn;
18262 static int fixup_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
18263 struct bpf_insn *insn_buf, int insn_idx, int *cnt)
18265 const struct bpf_kfunc_desc *desc;
18268 verbose(env, "invalid kernel function call not eliminated in verifier pass\n");
18274 /* insn->imm has the btf func_id. Replace it with an offset relative to
18275 * __bpf_call_base, unless the JIT needs to call functions that are
18276 * further than 32 bits away (bpf_jit_supports_far_kfunc_call()).
18278 desc = find_kfunc_desc(env->prog, insn->imm, insn->off);
18280 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
18285 if (!bpf_jit_supports_far_kfunc_call())
18286 insn->imm = BPF_CALL_IMM(desc->addr);
18289 if (desc->func_id == special_kfunc_list[KF_bpf_obj_new_impl]) {
18290 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18291 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
18292 u64 obj_new_size = env->insn_aux_data[insn_idx].obj_new_size;
18294 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_1, obj_new_size);
18295 insn_buf[1] = addr[0];
18296 insn_buf[2] = addr[1];
18297 insn_buf[3] = *insn;
18299 } else if (desc->func_id == special_kfunc_list[KF_bpf_obj_drop_impl] ||
18300 desc->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) {
18301 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18302 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
18304 if (desc->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl] &&
18305 !kptr_struct_meta) {
18306 verbose(env, "verifier internal error: kptr_struct_meta expected at insn_idx %d\n",
18311 insn_buf[0] = addr[0];
18312 insn_buf[1] = addr[1];
18313 insn_buf[2] = *insn;
18315 } else if (desc->func_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
18316 desc->func_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
18317 desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
18318 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18319 int struct_meta_reg = BPF_REG_3;
18320 int node_offset_reg = BPF_REG_4;
18322 /* rbtree_add has extra 'less' arg, so args-to-fixup are in diff regs */
18323 if (desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
18324 struct_meta_reg = BPF_REG_4;
18325 node_offset_reg = BPF_REG_5;
18328 if (!kptr_struct_meta) {
18329 verbose(env, "verifier internal error: kptr_struct_meta expected at insn_idx %d\n",
18334 __fixup_collection_insert_kfunc(&env->insn_aux_data[insn_idx], struct_meta_reg,
18335 node_offset_reg, insn, insn_buf, cnt);
18336 } else if (desc->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx] ||
18337 desc->func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
18338 insn_buf[0] = BPF_MOV64_REG(BPF_REG_0, BPF_REG_1);
18344 /* Do various post-verification rewrites in a single program pass.
18345 * These rewrites simplify JIT and interpreter implementations.
18347 static int do_misc_fixups(struct bpf_verifier_env *env)
18349 struct bpf_prog *prog = env->prog;
18350 enum bpf_attach_type eatype = prog->expected_attach_type;
18351 enum bpf_prog_type prog_type = resolve_prog_type(prog);
18352 struct bpf_insn *insn = prog->insnsi;
18353 const struct bpf_func_proto *fn;
18354 const int insn_cnt = prog->len;
18355 const struct bpf_map_ops *ops;
18356 struct bpf_insn_aux_data *aux;
18357 struct bpf_insn insn_buf[16];
18358 struct bpf_prog *new_prog;
18359 struct bpf_map *map_ptr;
18360 int i, ret, cnt, delta = 0;
18362 for (i = 0; i < insn_cnt; i++, insn++) {
18363 /* Make divide-by-zero exceptions impossible. */
18364 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
18365 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
18366 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
18367 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
18368 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
18369 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
18370 struct bpf_insn *patchlet;
18371 struct bpf_insn chk_and_div[] = {
18372 /* [R,W]x div 0 -> 0 */
18373 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
18374 BPF_JNE | BPF_K, insn->src_reg,
18376 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
18377 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
18380 struct bpf_insn chk_and_mod[] = {
18381 /* [R,W]x mod 0 -> [R,W]x */
18382 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
18383 BPF_JEQ | BPF_K, insn->src_reg,
18384 0, 1 + (is64 ? 0 : 1), 0),
18386 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
18387 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
18390 patchlet = isdiv ? chk_and_div : chk_and_mod;
18391 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
18392 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
18394 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
18399 env->prog = prog = new_prog;
18400 insn = new_prog->insnsi + i + delta;
18404 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
18405 if (BPF_CLASS(insn->code) == BPF_LD &&
18406 (BPF_MODE(insn->code) == BPF_ABS ||
18407 BPF_MODE(insn->code) == BPF_IND)) {
18408 cnt = env->ops->gen_ld_abs(insn, insn_buf);
18409 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
18410 verbose(env, "bpf verifier is misconfigured\n");
18414 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18419 env->prog = prog = new_prog;
18420 insn = new_prog->insnsi + i + delta;
18424 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
18425 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
18426 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
18427 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
18428 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
18429 struct bpf_insn *patch = &insn_buf[0];
18430 bool issrc, isneg, isimm;
18433 aux = &env->insn_aux_data[i + delta];
18434 if (!aux->alu_state ||
18435 aux->alu_state == BPF_ALU_NON_POINTER)
18438 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
18439 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
18440 BPF_ALU_SANITIZE_SRC;
18441 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
18443 off_reg = issrc ? insn->src_reg : insn->dst_reg;
18445 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
18448 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
18449 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
18450 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
18451 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
18452 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
18453 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
18454 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
18457 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
18458 insn->src_reg = BPF_REG_AX;
18460 insn->code = insn->code == code_add ?
18461 code_sub : code_add;
18463 if (issrc && isneg && !isimm)
18464 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
18465 cnt = patch - insn_buf;
18467 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18472 env->prog = prog = new_prog;
18473 insn = new_prog->insnsi + i + delta;
18477 if (insn->code != (BPF_JMP | BPF_CALL))
18479 if (insn->src_reg == BPF_PSEUDO_CALL)
18481 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
18482 ret = fixup_kfunc_call(env, insn, insn_buf, i + delta, &cnt);
18488 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18493 env->prog = prog = new_prog;
18494 insn = new_prog->insnsi + i + delta;
18498 if (insn->imm == BPF_FUNC_get_route_realm)
18499 prog->dst_needed = 1;
18500 if (insn->imm == BPF_FUNC_get_prandom_u32)
18501 bpf_user_rnd_init_once();
18502 if (insn->imm == BPF_FUNC_override_return)
18503 prog->kprobe_override = 1;
18504 if (insn->imm == BPF_FUNC_tail_call) {
18505 /* If we tail call into other programs, we
18506 * cannot make any assumptions since they can
18507 * be replaced dynamically during runtime in
18508 * the program array.
18510 prog->cb_access = 1;
18511 if (!allow_tail_call_in_subprogs(env))
18512 prog->aux->stack_depth = MAX_BPF_STACK;
18513 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
18515 /* mark bpf_tail_call as different opcode to avoid
18516 * conditional branch in the interpreter for every normal
18517 * call and to prevent accidental JITing by JIT compiler
18518 * that doesn't support bpf_tail_call yet
18521 insn->code = BPF_JMP | BPF_TAIL_CALL;
18523 aux = &env->insn_aux_data[i + delta];
18524 if (env->bpf_capable && !prog->blinding_requested &&
18525 prog->jit_requested &&
18526 !bpf_map_key_poisoned(aux) &&
18527 !bpf_map_ptr_poisoned(aux) &&
18528 !bpf_map_ptr_unpriv(aux)) {
18529 struct bpf_jit_poke_descriptor desc = {
18530 .reason = BPF_POKE_REASON_TAIL_CALL,
18531 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
18532 .tail_call.key = bpf_map_key_immediate(aux),
18533 .insn_idx = i + delta,
18536 ret = bpf_jit_add_poke_descriptor(prog, &desc);
18538 verbose(env, "adding tail call poke descriptor failed\n");
18542 insn->imm = ret + 1;
18546 if (!bpf_map_ptr_unpriv(aux))
18549 /* instead of changing every JIT dealing with tail_call
18550 * emit two extra insns:
18551 * if (index >= max_entries) goto out;
18552 * index &= array->index_mask;
18553 * to avoid out-of-bounds cpu speculation
18555 if (bpf_map_ptr_poisoned(aux)) {
18556 verbose(env, "tail_call abusing map_ptr\n");
18560 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
18561 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
18562 map_ptr->max_entries, 2);
18563 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
18564 container_of(map_ptr,
18567 insn_buf[2] = *insn;
18569 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18574 env->prog = prog = new_prog;
18575 insn = new_prog->insnsi + i + delta;
18579 if (insn->imm == BPF_FUNC_timer_set_callback) {
18580 /* The verifier will process callback_fn as many times as necessary
18581 * with different maps and the register states prepared by
18582 * set_timer_callback_state will be accurate.
18584 * The following use case is valid:
18585 * map1 is shared by prog1, prog2, prog3.
18586 * prog1 calls bpf_timer_init for some map1 elements
18587 * prog2 calls bpf_timer_set_callback for some map1 elements.
18588 * Those that were not bpf_timer_init-ed will return -EINVAL.
18589 * prog3 calls bpf_timer_start for some map1 elements.
18590 * Those that were not both bpf_timer_init-ed and
18591 * bpf_timer_set_callback-ed will return -EINVAL.
18593 struct bpf_insn ld_addrs[2] = {
18594 BPF_LD_IMM64(BPF_REG_3, (long)prog->aux),
18597 insn_buf[0] = ld_addrs[0];
18598 insn_buf[1] = ld_addrs[1];
18599 insn_buf[2] = *insn;
18602 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18607 env->prog = prog = new_prog;
18608 insn = new_prog->insnsi + i + delta;
18609 goto patch_call_imm;
18612 if (is_storage_get_function(insn->imm)) {
18613 if (!env->prog->aux->sleepable ||
18614 env->insn_aux_data[i + delta].storage_get_func_atomic)
18615 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_ATOMIC);
18617 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_KERNEL);
18618 insn_buf[1] = *insn;
18621 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18626 env->prog = prog = new_prog;
18627 insn = new_prog->insnsi + i + delta;
18628 goto patch_call_imm;
18631 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
18632 * and other inlining handlers are currently limited to 64 bit
18635 if (prog->jit_requested && BITS_PER_LONG == 64 &&
18636 (insn->imm == BPF_FUNC_map_lookup_elem ||
18637 insn->imm == BPF_FUNC_map_update_elem ||
18638 insn->imm == BPF_FUNC_map_delete_elem ||
18639 insn->imm == BPF_FUNC_map_push_elem ||
18640 insn->imm == BPF_FUNC_map_pop_elem ||
18641 insn->imm == BPF_FUNC_map_peek_elem ||
18642 insn->imm == BPF_FUNC_redirect_map ||
18643 insn->imm == BPF_FUNC_for_each_map_elem ||
18644 insn->imm == BPF_FUNC_map_lookup_percpu_elem)) {
18645 aux = &env->insn_aux_data[i + delta];
18646 if (bpf_map_ptr_poisoned(aux))
18647 goto patch_call_imm;
18649 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
18650 ops = map_ptr->ops;
18651 if (insn->imm == BPF_FUNC_map_lookup_elem &&
18652 ops->map_gen_lookup) {
18653 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
18654 if (cnt == -EOPNOTSUPP)
18655 goto patch_map_ops_generic;
18656 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
18657 verbose(env, "bpf verifier is misconfigured\n");
18661 new_prog = bpf_patch_insn_data(env, i + delta,
18667 env->prog = prog = new_prog;
18668 insn = new_prog->insnsi + i + delta;
18672 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
18673 (void *(*)(struct bpf_map *map, void *key))NULL));
18674 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
18675 (long (*)(struct bpf_map *map, void *key))NULL));
18676 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
18677 (long (*)(struct bpf_map *map, void *key, void *value,
18679 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
18680 (long (*)(struct bpf_map *map, void *value,
18682 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
18683 (long (*)(struct bpf_map *map, void *value))NULL));
18684 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
18685 (long (*)(struct bpf_map *map, void *value))NULL));
18686 BUILD_BUG_ON(!__same_type(ops->map_redirect,
18687 (long (*)(struct bpf_map *map, u64 index, u64 flags))NULL));
18688 BUILD_BUG_ON(!__same_type(ops->map_for_each_callback,
18689 (long (*)(struct bpf_map *map,
18690 bpf_callback_t callback_fn,
18691 void *callback_ctx,
18693 BUILD_BUG_ON(!__same_type(ops->map_lookup_percpu_elem,
18694 (void *(*)(struct bpf_map *map, void *key, u32 cpu))NULL));
18696 patch_map_ops_generic:
18697 switch (insn->imm) {
18698 case BPF_FUNC_map_lookup_elem:
18699 insn->imm = BPF_CALL_IMM(ops->map_lookup_elem);
18701 case BPF_FUNC_map_update_elem:
18702 insn->imm = BPF_CALL_IMM(ops->map_update_elem);
18704 case BPF_FUNC_map_delete_elem:
18705 insn->imm = BPF_CALL_IMM(ops->map_delete_elem);
18707 case BPF_FUNC_map_push_elem:
18708 insn->imm = BPF_CALL_IMM(ops->map_push_elem);
18710 case BPF_FUNC_map_pop_elem:
18711 insn->imm = BPF_CALL_IMM(ops->map_pop_elem);
18713 case BPF_FUNC_map_peek_elem:
18714 insn->imm = BPF_CALL_IMM(ops->map_peek_elem);
18716 case BPF_FUNC_redirect_map:
18717 insn->imm = BPF_CALL_IMM(ops->map_redirect);
18719 case BPF_FUNC_for_each_map_elem:
18720 insn->imm = BPF_CALL_IMM(ops->map_for_each_callback);
18722 case BPF_FUNC_map_lookup_percpu_elem:
18723 insn->imm = BPF_CALL_IMM(ops->map_lookup_percpu_elem);
18727 goto patch_call_imm;
18730 /* Implement bpf_jiffies64 inline. */
18731 if (prog->jit_requested && BITS_PER_LONG == 64 &&
18732 insn->imm == BPF_FUNC_jiffies64) {
18733 struct bpf_insn ld_jiffies_addr[2] = {
18734 BPF_LD_IMM64(BPF_REG_0,
18735 (unsigned long)&jiffies),
18738 insn_buf[0] = ld_jiffies_addr[0];
18739 insn_buf[1] = ld_jiffies_addr[1];
18740 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
18744 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
18750 env->prog = prog = new_prog;
18751 insn = new_prog->insnsi + i + delta;
18755 /* Implement bpf_get_func_arg inline. */
18756 if (prog_type == BPF_PROG_TYPE_TRACING &&
18757 insn->imm == BPF_FUNC_get_func_arg) {
18758 /* Load nr_args from ctx - 8 */
18759 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
18760 insn_buf[1] = BPF_JMP32_REG(BPF_JGE, BPF_REG_2, BPF_REG_0, 6);
18761 insn_buf[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_2, 3);
18762 insn_buf[3] = BPF_ALU64_REG(BPF_ADD, BPF_REG_2, BPF_REG_1);
18763 insn_buf[4] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_2, 0);
18764 insn_buf[5] = BPF_STX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
18765 insn_buf[6] = BPF_MOV64_IMM(BPF_REG_0, 0);
18766 insn_buf[7] = BPF_JMP_A(1);
18767 insn_buf[8] = BPF_MOV64_IMM(BPF_REG_0, -EINVAL);
18770 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18775 env->prog = prog = new_prog;
18776 insn = new_prog->insnsi + i + delta;
18780 /* Implement bpf_get_func_ret inline. */
18781 if (prog_type == BPF_PROG_TYPE_TRACING &&
18782 insn->imm == BPF_FUNC_get_func_ret) {
18783 if (eatype == BPF_TRACE_FEXIT ||
18784 eatype == BPF_MODIFY_RETURN) {
18785 /* Load nr_args from ctx - 8 */
18786 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
18787 insn_buf[1] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_0, 3);
18788 insn_buf[2] = BPF_ALU64_REG(BPF_ADD, BPF_REG_0, BPF_REG_1);
18789 insn_buf[3] = BPF_LDX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
18790 insn_buf[4] = BPF_STX_MEM(BPF_DW, BPF_REG_2, BPF_REG_3, 0);
18791 insn_buf[5] = BPF_MOV64_IMM(BPF_REG_0, 0);
18794 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, -EOPNOTSUPP);
18798 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18803 env->prog = prog = new_prog;
18804 insn = new_prog->insnsi + i + delta;
18808 /* Implement get_func_arg_cnt inline. */
18809 if (prog_type == BPF_PROG_TYPE_TRACING &&
18810 insn->imm == BPF_FUNC_get_func_arg_cnt) {
18811 /* Load nr_args from ctx - 8 */
18812 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
18814 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
18818 env->prog = prog = new_prog;
18819 insn = new_prog->insnsi + i + delta;
18823 /* Implement bpf_get_func_ip inline. */
18824 if (prog_type == BPF_PROG_TYPE_TRACING &&
18825 insn->imm == BPF_FUNC_get_func_ip) {
18826 /* Load IP address from ctx - 16 */
18827 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -16);
18829 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
18833 env->prog = prog = new_prog;
18834 insn = new_prog->insnsi + i + delta;
18839 fn = env->ops->get_func_proto(insn->imm, env->prog);
18840 /* all functions that have prototype and verifier allowed
18841 * programs to call them, must be real in-kernel functions
18845 "kernel subsystem misconfigured func %s#%d\n",
18846 func_id_name(insn->imm), insn->imm);
18849 insn->imm = fn->func - __bpf_call_base;
18852 /* Since poke tab is now finalized, publish aux to tracker. */
18853 for (i = 0; i < prog->aux->size_poke_tab; i++) {
18854 map_ptr = prog->aux->poke_tab[i].tail_call.map;
18855 if (!map_ptr->ops->map_poke_track ||
18856 !map_ptr->ops->map_poke_untrack ||
18857 !map_ptr->ops->map_poke_run) {
18858 verbose(env, "bpf verifier is misconfigured\n");
18862 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
18864 verbose(env, "tracking tail call prog failed\n");
18869 sort_kfunc_descs_by_imm_off(env->prog);
18874 static struct bpf_prog *inline_bpf_loop(struct bpf_verifier_env *env,
18877 u32 callback_subprogno,
18880 s32 r6_offset = stack_base + 0 * BPF_REG_SIZE;
18881 s32 r7_offset = stack_base + 1 * BPF_REG_SIZE;
18882 s32 r8_offset = stack_base + 2 * BPF_REG_SIZE;
18883 int reg_loop_max = BPF_REG_6;
18884 int reg_loop_cnt = BPF_REG_7;
18885 int reg_loop_ctx = BPF_REG_8;
18887 struct bpf_prog *new_prog;
18888 u32 callback_start;
18889 u32 call_insn_offset;
18890 s32 callback_offset;
18892 /* This represents an inlined version of bpf_iter.c:bpf_loop,
18893 * be careful to modify this code in sync.
18895 struct bpf_insn insn_buf[] = {
18896 /* Return error and jump to the end of the patch if
18897 * expected number of iterations is too big.
18899 BPF_JMP_IMM(BPF_JLE, BPF_REG_1, BPF_MAX_LOOPS, 2),
18900 BPF_MOV32_IMM(BPF_REG_0, -E2BIG),
18901 BPF_JMP_IMM(BPF_JA, 0, 0, 16),
18902 /* spill R6, R7, R8 to use these as loop vars */
18903 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_6, r6_offset),
18904 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_7, r7_offset),
18905 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_8, r8_offset),
18906 /* initialize loop vars */
18907 BPF_MOV64_REG(reg_loop_max, BPF_REG_1),
18908 BPF_MOV32_IMM(reg_loop_cnt, 0),
18909 BPF_MOV64_REG(reg_loop_ctx, BPF_REG_3),
18911 * if reg_loop_cnt >= reg_loop_max skip the loop body
18913 BPF_JMP_REG(BPF_JGE, reg_loop_cnt, reg_loop_max, 5),
18915 * correct callback offset would be set after patching
18917 BPF_MOV64_REG(BPF_REG_1, reg_loop_cnt),
18918 BPF_MOV64_REG(BPF_REG_2, reg_loop_ctx),
18920 /* increment loop counter */
18921 BPF_ALU64_IMM(BPF_ADD, reg_loop_cnt, 1),
18922 /* jump to loop header if callback returned 0 */
18923 BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, -6),
18924 /* return value of bpf_loop,
18925 * set R0 to the number of iterations
18927 BPF_MOV64_REG(BPF_REG_0, reg_loop_cnt),
18928 /* restore original values of R6, R7, R8 */
18929 BPF_LDX_MEM(BPF_DW, BPF_REG_6, BPF_REG_10, r6_offset),
18930 BPF_LDX_MEM(BPF_DW, BPF_REG_7, BPF_REG_10, r7_offset),
18931 BPF_LDX_MEM(BPF_DW, BPF_REG_8, BPF_REG_10, r8_offset),
18934 *cnt = ARRAY_SIZE(insn_buf);
18935 new_prog = bpf_patch_insn_data(env, position, insn_buf, *cnt);
18939 /* callback start is known only after patching */
18940 callback_start = env->subprog_info[callback_subprogno].start;
18941 /* Note: insn_buf[12] is an offset of BPF_CALL_REL instruction */
18942 call_insn_offset = position + 12;
18943 callback_offset = callback_start - call_insn_offset - 1;
18944 new_prog->insnsi[call_insn_offset].imm = callback_offset;
18949 static bool is_bpf_loop_call(struct bpf_insn *insn)
18951 return insn->code == (BPF_JMP | BPF_CALL) &&
18952 insn->src_reg == 0 &&
18953 insn->imm == BPF_FUNC_loop;
18956 /* For all sub-programs in the program (including main) check
18957 * insn_aux_data to see if there are bpf_loop calls that require
18958 * inlining. If such calls are found the calls are replaced with a
18959 * sequence of instructions produced by `inline_bpf_loop` function and
18960 * subprog stack_depth is increased by the size of 3 registers.
18961 * This stack space is used to spill values of the R6, R7, R8. These
18962 * registers are used to store the loop bound, counter and context
18965 static int optimize_bpf_loop(struct bpf_verifier_env *env)
18967 struct bpf_subprog_info *subprogs = env->subprog_info;
18968 int i, cur_subprog = 0, cnt, delta = 0;
18969 struct bpf_insn *insn = env->prog->insnsi;
18970 int insn_cnt = env->prog->len;
18971 u16 stack_depth = subprogs[cur_subprog].stack_depth;
18972 u16 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
18973 u16 stack_depth_extra = 0;
18975 for (i = 0; i < insn_cnt; i++, insn++) {
18976 struct bpf_loop_inline_state *inline_state =
18977 &env->insn_aux_data[i + delta].loop_inline_state;
18979 if (is_bpf_loop_call(insn) && inline_state->fit_for_inline) {
18980 struct bpf_prog *new_prog;
18982 stack_depth_extra = BPF_REG_SIZE * 3 + stack_depth_roundup;
18983 new_prog = inline_bpf_loop(env,
18985 -(stack_depth + stack_depth_extra),
18986 inline_state->callback_subprogno,
18992 env->prog = new_prog;
18993 insn = new_prog->insnsi + i + delta;
18996 if (subprogs[cur_subprog + 1].start == i + delta + 1) {
18997 subprogs[cur_subprog].stack_depth += stack_depth_extra;
18999 stack_depth = subprogs[cur_subprog].stack_depth;
19000 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
19001 stack_depth_extra = 0;
19005 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
19010 static void free_states(struct bpf_verifier_env *env)
19012 struct bpf_verifier_state_list *sl, *sln;
19015 sl = env->free_list;
19018 free_verifier_state(&sl->state, false);
19022 env->free_list = NULL;
19024 if (!env->explored_states)
19027 for (i = 0; i < state_htab_size(env); i++) {
19028 sl = env->explored_states[i];
19032 free_verifier_state(&sl->state, false);
19036 env->explored_states[i] = NULL;
19040 static int do_check_common(struct bpf_verifier_env *env, int subprog)
19042 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
19043 struct bpf_verifier_state *state;
19044 struct bpf_reg_state *regs;
19047 env->prev_linfo = NULL;
19050 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
19053 state->curframe = 0;
19054 state->speculative = false;
19055 state->branches = 1;
19056 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
19057 if (!state->frame[0]) {
19061 env->cur_state = state;
19062 init_func_state(env, state->frame[0],
19063 BPF_MAIN_FUNC /* callsite */,
19066 state->first_insn_idx = env->subprog_info[subprog].start;
19067 state->last_insn_idx = -1;
19069 regs = state->frame[state->curframe]->regs;
19070 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
19071 ret = btf_prepare_func_args(env, subprog, regs);
19074 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
19075 if (regs[i].type == PTR_TO_CTX)
19076 mark_reg_known_zero(env, regs, i);
19077 else if (regs[i].type == SCALAR_VALUE)
19078 mark_reg_unknown(env, regs, i);
19079 else if (base_type(regs[i].type) == PTR_TO_MEM) {
19080 const u32 mem_size = regs[i].mem_size;
19082 mark_reg_known_zero(env, regs, i);
19083 regs[i].mem_size = mem_size;
19084 regs[i].id = ++env->id_gen;
19088 /* 1st arg to a function */
19089 regs[BPF_REG_1].type = PTR_TO_CTX;
19090 mark_reg_known_zero(env, regs, BPF_REG_1);
19091 ret = btf_check_subprog_arg_match(env, subprog, regs);
19092 if (ret == -EFAULT)
19093 /* unlikely verifier bug. abort.
19094 * ret == 0 and ret < 0 are sadly acceptable for
19095 * main() function due to backward compatibility.
19096 * Like socket filter program may be written as:
19097 * int bpf_prog(struct pt_regs *ctx)
19098 * and never dereference that ctx in the program.
19099 * 'struct pt_regs' is a type mismatch for socket
19100 * filter that should be using 'struct __sk_buff'.
19105 ret = do_check(env);
19107 /* check for NULL is necessary, since cur_state can be freed inside
19108 * do_check() under memory pressure.
19110 if (env->cur_state) {
19111 free_verifier_state(env->cur_state, true);
19112 env->cur_state = NULL;
19114 while (!pop_stack(env, NULL, NULL, false));
19115 if (!ret && pop_log)
19116 bpf_vlog_reset(&env->log, 0);
19121 /* Verify all global functions in a BPF program one by one based on their BTF.
19122 * All global functions must pass verification. Otherwise the whole program is rejected.
19133 * foo() will be verified first for R1=any_scalar_value. During verification it
19134 * will be assumed that bar() already verified successfully and call to bar()
19135 * from foo() will be checked for type match only. Later bar() will be verified
19136 * independently to check that it's safe for R1=any_scalar_value.
19138 static int do_check_subprogs(struct bpf_verifier_env *env)
19140 struct bpf_prog_aux *aux = env->prog->aux;
19143 if (!aux->func_info)
19146 for (i = 1; i < env->subprog_cnt; i++) {
19147 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
19149 env->insn_idx = env->subprog_info[i].start;
19150 WARN_ON_ONCE(env->insn_idx == 0);
19151 ret = do_check_common(env, i);
19154 } else if (env->log.level & BPF_LOG_LEVEL) {
19156 "Func#%d is safe for any args that match its prototype\n",
19163 static int do_check_main(struct bpf_verifier_env *env)
19168 ret = do_check_common(env, 0);
19170 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
19175 static void print_verification_stats(struct bpf_verifier_env *env)
19179 if (env->log.level & BPF_LOG_STATS) {
19180 verbose(env, "verification time %lld usec\n",
19181 div_u64(env->verification_time, 1000));
19182 verbose(env, "stack depth ");
19183 for (i = 0; i < env->subprog_cnt; i++) {
19184 u32 depth = env->subprog_info[i].stack_depth;
19186 verbose(env, "%d", depth);
19187 if (i + 1 < env->subprog_cnt)
19190 verbose(env, "\n");
19192 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
19193 "total_states %d peak_states %d mark_read %d\n",
19194 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
19195 env->max_states_per_insn, env->total_states,
19196 env->peak_states, env->longest_mark_read_walk);
19199 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
19201 const struct btf_type *t, *func_proto;
19202 const struct bpf_struct_ops *st_ops;
19203 const struct btf_member *member;
19204 struct bpf_prog *prog = env->prog;
19205 u32 btf_id, member_idx;
19208 if (!prog->gpl_compatible) {
19209 verbose(env, "struct ops programs must have a GPL compatible license\n");
19213 btf_id = prog->aux->attach_btf_id;
19214 st_ops = bpf_struct_ops_find(btf_id);
19216 verbose(env, "attach_btf_id %u is not a supported struct\n",
19222 member_idx = prog->expected_attach_type;
19223 if (member_idx >= btf_type_vlen(t)) {
19224 verbose(env, "attach to invalid member idx %u of struct %s\n",
19225 member_idx, st_ops->name);
19229 member = &btf_type_member(t)[member_idx];
19230 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
19231 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
19234 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
19235 mname, member_idx, st_ops->name);
19239 if (st_ops->check_member) {
19240 int err = st_ops->check_member(t, member, prog);
19243 verbose(env, "attach to unsupported member %s of struct %s\n",
19244 mname, st_ops->name);
19249 prog->aux->attach_func_proto = func_proto;
19250 prog->aux->attach_func_name = mname;
19251 env->ops = st_ops->verifier_ops;
19255 #define SECURITY_PREFIX "security_"
19257 static int check_attach_modify_return(unsigned long addr, const char *func_name)
19259 if (within_error_injection_list(addr) ||
19260 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
19266 /* list of non-sleepable functions that are otherwise on
19267 * ALLOW_ERROR_INJECTION list
19269 BTF_SET_START(btf_non_sleepable_error_inject)
19270 /* Three functions below can be called from sleepable and non-sleepable context.
19271 * Assume non-sleepable from bpf safety point of view.
19273 BTF_ID(func, __filemap_add_folio)
19274 BTF_ID(func, should_fail_alloc_page)
19275 BTF_ID(func, should_failslab)
19276 BTF_SET_END(btf_non_sleepable_error_inject)
19278 static int check_non_sleepable_error_inject(u32 btf_id)
19280 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
19283 int bpf_check_attach_target(struct bpf_verifier_log *log,
19284 const struct bpf_prog *prog,
19285 const struct bpf_prog *tgt_prog,
19287 struct bpf_attach_target_info *tgt_info)
19289 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
19290 const char prefix[] = "btf_trace_";
19291 int ret = 0, subprog = -1, i;
19292 const struct btf_type *t;
19293 bool conservative = true;
19297 struct module *mod = NULL;
19300 bpf_log(log, "Tracing programs must provide btf_id\n");
19303 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
19306 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
19309 t = btf_type_by_id(btf, btf_id);
19311 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
19314 tname = btf_name_by_offset(btf, t->name_off);
19316 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
19320 struct bpf_prog_aux *aux = tgt_prog->aux;
19322 if (bpf_prog_is_dev_bound(prog->aux) &&
19323 !bpf_prog_dev_bound_match(prog, tgt_prog)) {
19324 bpf_log(log, "Target program bound device mismatch");
19328 for (i = 0; i < aux->func_info_cnt; i++)
19329 if (aux->func_info[i].type_id == btf_id) {
19333 if (subprog == -1) {
19334 bpf_log(log, "Subprog %s doesn't exist\n", tname);
19337 conservative = aux->func_info_aux[subprog].unreliable;
19338 if (prog_extension) {
19339 if (conservative) {
19341 "Cannot replace static functions\n");
19344 if (!prog->jit_requested) {
19346 "Extension programs should be JITed\n");
19350 if (!tgt_prog->jited) {
19351 bpf_log(log, "Can attach to only JITed progs\n");
19354 if (tgt_prog->type == prog->type) {
19355 /* Cannot fentry/fexit another fentry/fexit program.
19356 * Cannot attach program extension to another extension.
19357 * It's ok to attach fentry/fexit to extension program.
19359 bpf_log(log, "Cannot recursively attach\n");
19362 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
19364 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
19365 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
19366 /* Program extensions can extend all program types
19367 * except fentry/fexit. The reason is the following.
19368 * The fentry/fexit programs are used for performance
19369 * analysis, stats and can be attached to any program
19370 * type except themselves. When extension program is
19371 * replacing XDP function it is necessary to allow
19372 * performance analysis of all functions. Both original
19373 * XDP program and its program extension. Hence
19374 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
19375 * allowed. If extending of fentry/fexit was allowed it
19376 * would be possible to create long call chain
19377 * fentry->extension->fentry->extension beyond
19378 * reasonable stack size. Hence extending fentry is not
19381 bpf_log(log, "Cannot extend fentry/fexit\n");
19385 if (prog_extension) {
19386 bpf_log(log, "Cannot replace kernel functions\n");
19391 switch (prog->expected_attach_type) {
19392 case BPF_TRACE_RAW_TP:
19395 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
19398 if (!btf_type_is_typedef(t)) {
19399 bpf_log(log, "attach_btf_id %u is not a typedef\n",
19403 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
19404 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
19408 tname += sizeof(prefix) - 1;
19409 t = btf_type_by_id(btf, t->type);
19410 if (!btf_type_is_ptr(t))
19411 /* should never happen in valid vmlinux build */
19413 t = btf_type_by_id(btf, t->type);
19414 if (!btf_type_is_func_proto(t))
19415 /* should never happen in valid vmlinux build */
19419 case BPF_TRACE_ITER:
19420 if (!btf_type_is_func(t)) {
19421 bpf_log(log, "attach_btf_id %u is not a function\n",
19425 t = btf_type_by_id(btf, t->type);
19426 if (!btf_type_is_func_proto(t))
19428 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
19433 if (!prog_extension)
19436 case BPF_MODIFY_RETURN:
19438 case BPF_LSM_CGROUP:
19439 case BPF_TRACE_FENTRY:
19440 case BPF_TRACE_FEXIT:
19441 if (!btf_type_is_func(t)) {
19442 bpf_log(log, "attach_btf_id %u is not a function\n",
19446 if (prog_extension &&
19447 btf_check_type_match(log, prog, btf, t))
19449 t = btf_type_by_id(btf, t->type);
19450 if (!btf_type_is_func_proto(t))
19453 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
19454 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
19455 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
19458 if (tgt_prog && conservative)
19461 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
19467 addr = (long) tgt_prog->bpf_func;
19469 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
19471 if (btf_is_module(btf)) {
19472 mod = btf_try_get_module(btf);
19474 addr = find_kallsyms_symbol_value(mod, tname);
19478 addr = kallsyms_lookup_name(tname);
19483 "The address of function %s cannot be found\n",
19489 if (prog->aux->sleepable) {
19491 switch (prog->type) {
19492 case BPF_PROG_TYPE_TRACING:
19494 /* fentry/fexit/fmod_ret progs can be sleepable if they are
19495 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
19497 if (!check_non_sleepable_error_inject(btf_id) &&
19498 within_error_injection_list(addr))
19500 /* fentry/fexit/fmod_ret progs can also be sleepable if they are
19501 * in the fmodret id set with the KF_SLEEPABLE flag.
19504 u32 *flags = btf_kfunc_is_modify_return(btf, btf_id,
19507 if (flags && (*flags & KF_SLEEPABLE))
19511 case BPF_PROG_TYPE_LSM:
19512 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
19513 * Only some of them are sleepable.
19515 if (bpf_lsm_is_sleepable_hook(btf_id))
19523 bpf_log(log, "%s is not sleepable\n", tname);
19526 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
19529 bpf_log(log, "can't modify return codes of BPF programs\n");
19533 if (btf_kfunc_is_modify_return(btf, btf_id, prog) ||
19534 !check_attach_modify_return(addr, tname))
19538 bpf_log(log, "%s() is not modifiable\n", tname);
19545 tgt_info->tgt_addr = addr;
19546 tgt_info->tgt_name = tname;
19547 tgt_info->tgt_type = t;
19548 tgt_info->tgt_mod = mod;
19552 BTF_SET_START(btf_id_deny)
19555 BTF_ID(func, migrate_disable)
19556 BTF_ID(func, migrate_enable)
19558 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
19559 BTF_ID(func, rcu_read_unlock_strict)
19561 #if defined(CONFIG_DEBUG_PREEMPT) || defined(CONFIG_TRACE_PREEMPT_TOGGLE)
19562 BTF_ID(func, preempt_count_add)
19563 BTF_ID(func, preempt_count_sub)
19565 #ifdef CONFIG_PREEMPT_RCU
19566 BTF_ID(func, __rcu_read_lock)
19567 BTF_ID(func, __rcu_read_unlock)
19569 BTF_SET_END(btf_id_deny)
19571 static bool can_be_sleepable(struct bpf_prog *prog)
19573 if (prog->type == BPF_PROG_TYPE_TRACING) {
19574 switch (prog->expected_attach_type) {
19575 case BPF_TRACE_FENTRY:
19576 case BPF_TRACE_FEXIT:
19577 case BPF_MODIFY_RETURN:
19578 case BPF_TRACE_ITER:
19584 return prog->type == BPF_PROG_TYPE_LSM ||
19585 prog->type == BPF_PROG_TYPE_KPROBE /* only for uprobes */ ||
19586 prog->type == BPF_PROG_TYPE_STRUCT_OPS;
19589 static int check_attach_btf_id(struct bpf_verifier_env *env)
19591 struct bpf_prog *prog = env->prog;
19592 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
19593 struct bpf_attach_target_info tgt_info = {};
19594 u32 btf_id = prog->aux->attach_btf_id;
19595 struct bpf_trampoline *tr;
19599 if (prog->type == BPF_PROG_TYPE_SYSCALL) {
19600 if (prog->aux->sleepable)
19601 /* attach_btf_id checked to be zero already */
19603 verbose(env, "Syscall programs can only be sleepable\n");
19607 if (prog->aux->sleepable && !can_be_sleepable(prog)) {
19608 verbose(env, "Only fentry/fexit/fmod_ret, lsm, iter, uprobe, and struct_ops programs can be sleepable\n");
19612 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
19613 return check_struct_ops_btf_id(env);
19615 if (prog->type != BPF_PROG_TYPE_TRACING &&
19616 prog->type != BPF_PROG_TYPE_LSM &&
19617 prog->type != BPF_PROG_TYPE_EXT)
19620 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
19624 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
19625 /* to make freplace equivalent to their targets, they need to
19626 * inherit env->ops and expected_attach_type for the rest of the
19629 env->ops = bpf_verifier_ops[tgt_prog->type];
19630 prog->expected_attach_type = tgt_prog->expected_attach_type;
19633 /* store info about the attachment target that will be used later */
19634 prog->aux->attach_func_proto = tgt_info.tgt_type;
19635 prog->aux->attach_func_name = tgt_info.tgt_name;
19636 prog->aux->mod = tgt_info.tgt_mod;
19639 prog->aux->saved_dst_prog_type = tgt_prog->type;
19640 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
19643 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
19644 prog->aux->attach_btf_trace = true;
19646 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
19647 if (!bpf_iter_prog_supported(prog))
19652 if (prog->type == BPF_PROG_TYPE_LSM) {
19653 ret = bpf_lsm_verify_prog(&env->log, prog);
19656 } else if (prog->type == BPF_PROG_TYPE_TRACING &&
19657 btf_id_set_contains(&btf_id_deny, btf_id)) {
19661 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
19662 tr = bpf_trampoline_get(key, &tgt_info);
19666 if (tgt_prog && tgt_prog->aux->tail_call_reachable)
19667 tr->flags = BPF_TRAMP_F_TAIL_CALL_CTX;
19669 prog->aux->dst_trampoline = tr;
19673 struct btf *bpf_get_btf_vmlinux(void)
19675 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
19676 mutex_lock(&bpf_verifier_lock);
19678 btf_vmlinux = btf_parse_vmlinux();
19679 mutex_unlock(&bpf_verifier_lock);
19681 return btf_vmlinux;
19684 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr, __u32 uattr_size)
19686 u64 start_time = ktime_get_ns();
19687 struct bpf_verifier_env *env;
19688 int i, len, ret = -EINVAL, err;
19692 /* no program is valid */
19693 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
19696 /* 'struct bpf_verifier_env' can be global, but since it's not small,
19697 * allocate/free it every time bpf_check() is called
19699 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
19705 len = (*prog)->len;
19706 env->insn_aux_data =
19707 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
19709 if (!env->insn_aux_data)
19711 for (i = 0; i < len; i++)
19712 env->insn_aux_data[i].orig_idx = i;
19714 env->ops = bpf_verifier_ops[env->prog->type];
19715 env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
19716 is_priv = bpf_capable();
19718 bpf_get_btf_vmlinux();
19720 /* grab the mutex to protect few globals used by verifier */
19722 mutex_lock(&bpf_verifier_lock);
19724 /* user could have requested verbose verifier output
19725 * and supplied buffer to store the verification trace
19727 ret = bpf_vlog_init(&env->log, attr->log_level,
19728 (char __user *) (unsigned long) attr->log_buf,
19733 mark_verifier_state_clean(env);
19735 if (IS_ERR(btf_vmlinux)) {
19736 /* Either gcc or pahole or kernel are broken. */
19737 verbose(env, "in-kernel BTF is malformed\n");
19738 ret = PTR_ERR(btf_vmlinux);
19739 goto skip_full_check;
19742 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
19743 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
19744 env->strict_alignment = true;
19745 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
19746 env->strict_alignment = false;
19748 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
19749 env->allow_uninit_stack = bpf_allow_uninit_stack();
19750 env->bypass_spec_v1 = bpf_bypass_spec_v1();
19751 env->bypass_spec_v4 = bpf_bypass_spec_v4();
19752 env->bpf_capable = bpf_capable();
19755 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
19757 env->explored_states = kvcalloc(state_htab_size(env),
19758 sizeof(struct bpf_verifier_state_list *),
19761 if (!env->explored_states)
19762 goto skip_full_check;
19764 ret = add_subprog_and_kfunc(env);
19766 goto skip_full_check;
19768 ret = check_subprogs(env);
19770 goto skip_full_check;
19772 ret = check_btf_info(env, attr, uattr);
19774 goto skip_full_check;
19776 ret = check_attach_btf_id(env);
19778 goto skip_full_check;
19780 ret = resolve_pseudo_ldimm64(env);
19782 goto skip_full_check;
19784 if (bpf_prog_is_offloaded(env->prog->aux)) {
19785 ret = bpf_prog_offload_verifier_prep(env->prog);
19787 goto skip_full_check;
19790 ret = check_cfg(env);
19792 goto skip_full_check;
19794 ret = do_check_subprogs(env);
19795 ret = ret ?: do_check_main(env);
19797 if (ret == 0 && bpf_prog_is_offloaded(env->prog->aux))
19798 ret = bpf_prog_offload_finalize(env);
19801 kvfree(env->explored_states);
19804 ret = check_max_stack_depth(env);
19806 /* instruction rewrites happen after this point */
19808 ret = optimize_bpf_loop(env);
19812 opt_hard_wire_dead_code_branches(env);
19814 ret = opt_remove_dead_code(env);
19816 ret = opt_remove_nops(env);
19819 sanitize_dead_code(env);
19823 /* program is valid, convert *(u32*)(ctx + off) accesses */
19824 ret = convert_ctx_accesses(env);
19827 ret = do_misc_fixups(env);
19829 /* do 32-bit optimization after insn patching has done so those patched
19830 * insns could be handled correctly.
19832 if (ret == 0 && !bpf_prog_is_offloaded(env->prog->aux)) {
19833 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
19834 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
19839 ret = fixup_call_args(env);
19841 env->verification_time = ktime_get_ns() - start_time;
19842 print_verification_stats(env);
19843 env->prog->aux->verified_insns = env->insn_processed;
19845 /* preserve original error even if log finalization is successful */
19846 err = bpf_vlog_finalize(&env->log, &log_true_size);
19850 if (uattr_size >= offsetofend(union bpf_attr, log_true_size) &&
19851 copy_to_bpfptr_offset(uattr, offsetof(union bpf_attr, log_true_size),
19852 &log_true_size, sizeof(log_true_size))) {
19854 goto err_release_maps;
19858 goto err_release_maps;
19860 if (env->used_map_cnt) {
19861 /* if program passed verifier, update used_maps in bpf_prog_info */
19862 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
19863 sizeof(env->used_maps[0]),
19866 if (!env->prog->aux->used_maps) {
19868 goto err_release_maps;
19871 memcpy(env->prog->aux->used_maps, env->used_maps,
19872 sizeof(env->used_maps[0]) * env->used_map_cnt);
19873 env->prog->aux->used_map_cnt = env->used_map_cnt;
19875 if (env->used_btf_cnt) {
19876 /* if program passed verifier, update used_btfs in bpf_prog_aux */
19877 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
19878 sizeof(env->used_btfs[0]),
19880 if (!env->prog->aux->used_btfs) {
19882 goto err_release_maps;
19885 memcpy(env->prog->aux->used_btfs, env->used_btfs,
19886 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
19887 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
19889 if (env->used_map_cnt || env->used_btf_cnt) {
19890 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
19891 * bpf_ld_imm64 instructions
19893 convert_pseudo_ld_imm64(env);
19896 adjust_btf_func(env);
19899 if (!env->prog->aux->used_maps)
19900 /* if we didn't copy map pointers into bpf_prog_info, release
19901 * them now. Otherwise free_used_maps() will release them.
19904 if (!env->prog->aux->used_btfs)
19907 /* extension progs temporarily inherit the attach_type of their targets
19908 for verification purposes, so set it back to zero before returning
19910 if (env->prog->type == BPF_PROG_TYPE_EXT)
19911 env->prog->expected_attach_type = 0;
19916 mutex_unlock(&bpf_verifier_lock);
19917 vfree(env->insn_aux_data);