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.
3204 * Return -ENOENT if we exhausted all instructions within given state.
3206 * It's legal to have a bit of a looping with the same starting and ending
3207 * insn index within the same state, e.g.: 3->4->5->3, so just because current
3208 * instruction index is the same as state's first_idx doesn't mean we are
3209 * done. If there is still some jump history left, we should keep going. We
3210 * need to take into account that we might have a jump history between given
3211 * state's parent and itself, due to checkpointing. In this case, we'll have
3212 * history entry recording a jump from last instruction of parent state and
3213 * first instruction of given state.
3215 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
3220 if (i == st->first_insn_idx) {
3223 if (cnt == 1 && st->jmp_history[0].idx == i)
3227 if (cnt && st->jmp_history[cnt - 1].idx == i) {
3228 i = st->jmp_history[cnt - 1].prev_idx;
3236 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
3238 const struct btf_type *func;
3239 struct btf *desc_btf;
3241 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
3244 desc_btf = find_kfunc_desc_btf(data, insn->off);
3245 if (IS_ERR(desc_btf))
3248 func = btf_type_by_id(desc_btf, insn->imm);
3249 return btf_name_by_offset(desc_btf, func->name_off);
3252 static inline void bt_init(struct backtrack_state *bt, u32 frame)
3257 static inline void bt_reset(struct backtrack_state *bt)
3259 struct bpf_verifier_env *env = bt->env;
3261 memset(bt, 0, sizeof(*bt));
3265 static inline u32 bt_empty(struct backtrack_state *bt)
3270 for (i = 0; i <= bt->frame; i++)
3271 mask |= bt->reg_masks[i] | bt->stack_masks[i];
3276 static inline int bt_subprog_enter(struct backtrack_state *bt)
3278 if (bt->frame == MAX_CALL_FRAMES - 1) {
3279 verbose(bt->env, "BUG subprog enter from frame %d\n", bt->frame);
3280 WARN_ONCE(1, "verifier backtracking bug");
3287 static inline int bt_subprog_exit(struct backtrack_state *bt)
3289 if (bt->frame == 0) {
3290 verbose(bt->env, "BUG subprog exit from frame 0\n");
3291 WARN_ONCE(1, "verifier backtracking bug");
3298 static inline void bt_set_frame_reg(struct backtrack_state *bt, u32 frame, u32 reg)
3300 bt->reg_masks[frame] |= 1 << reg;
3303 static inline void bt_clear_frame_reg(struct backtrack_state *bt, u32 frame, u32 reg)
3305 bt->reg_masks[frame] &= ~(1 << reg);
3308 static inline void bt_set_reg(struct backtrack_state *bt, u32 reg)
3310 bt_set_frame_reg(bt, bt->frame, reg);
3313 static inline void bt_clear_reg(struct backtrack_state *bt, u32 reg)
3315 bt_clear_frame_reg(bt, bt->frame, reg);
3318 static inline void bt_set_frame_slot(struct backtrack_state *bt, u32 frame, u32 slot)
3320 bt->stack_masks[frame] |= 1ull << slot;
3323 static inline void bt_clear_frame_slot(struct backtrack_state *bt, u32 frame, u32 slot)
3325 bt->stack_masks[frame] &= ~(1ull << slot);
3328 static inline void bt_set_slot(struct backtrack_state *bt, u32 slot)
3330 bt_set_frame_slot(bt, bt->frame, slot);
3333 static inline void bt_clear_slot(struct backtrack_state *bt, u32 slot)
3335 bt_clear_frame_slot(bt, bt->frame, slot);
3338 static inline u32 bt_frame_reg_mask(struct backtrack_state *bt, u32 frame)
3340 return bt->reg_masks[frame];
3343 static inline u32 bt_reg_mask(struct backtrack_state *bt)
3345 return bt->reg_masks[bt->frame];
3348 static inline u64 bt_frame_stack_mask(struct backtrack_state *bt, u32 frame)
3350 return bt->stack_masks[frame];
3353 static inline u64 bt_stack_mask(struct backtrack_state *bt)
3355 return bt->stack_masks[bt->frame];
3358 static inline bool bt_is_reg_set(struct backtrack_state *bt, u32 reg)
3360 return bt->reg_masks[bt->frame] & (1 << reg);
3363 static inline bool bt_is_slot_set(struct backtrack_state *bt, u32 slot)
3365 return bt->stack_masks[bt->frame] & (1ull << slot);
3368 /* format registers bitmask, e.g., "r0,r2,r4" for 0x15 mask */
3369 static void fmt_reg_mask(char *buf, ssize_t buf_sz, u32 reg_mask)
3371 DECLARE_BITMAP(mask, 64);
3377 bitmap_from_u64(mask, reg_mask);
3378 for_each_set_bit(i, mask, 32) {
3379 n = snprintf(buf, buf_sz, "%sr%d", first ? "" : ",", i);
3387 /* format stack slots bitmask, e.g., "-8,-24,-40" for 0x15 mask */
3388 static void fmt_stack_mask(char *buf, ssize_t buf_sz, u64 stack_mask)
3390 DECLARE_BITMAP(mask, 64);
3396 bitmap_from_u64(mask, stack_mask);
3397 for_each_set_bit(i, mask, 64) {
3398 n = snprintf(buf, buf_sz, "%s%d", first ? "" : ",", -(i + 1) * 8);
3407 /* For given verifier state backtrack_insn() is called from the last insn to
3408 * the first insn. Its purpose is to compute a bitmask of registers and
3409 * stack slots that needs precision in the parent verifier state.
3411 * @idx is an index of the instruction we are currently processing;
3412 * @subseq_idx is an index of the subsequent instruction that:
3413 * - *would be* executed next, if jump history is viewed in forward order;
3414 * - *was* processed previously during backtracking.
3416 static int backtrack_insn(struct bpf_verifier_env *env, int idx, int subseq_idx,
3417 struct backtrack_state *bt)
3419 const struct bpf_insn_cbs cbs = {
3420 .cb_call = disasm_kfunc_name,
3421 .cb_print = verbose,
3422 .private_data = env,
3424 struct bpf_insn *insn = env->prog->insnsi + idx;
3425 u8 class = BPF_CLASS(insn->code);
3426 u8 opcode = BPF_OP(insn->code);
3427 u8 mode = BPF_MODE(insn->code);
3428 u32 dreg = insn->dst_reg;
3429 u32 sreg = insn->src_reg;
3432 if (insn->code == 0)
3434 if (env->log.level & BPF_LOG_LEVEL2) {
3435 fmt_reg_mask(env->tmp_str_buf, TMP_STR_BUF_LEN, bt_reg_mask(bt));
3436 verbose(env, "mark_precise: frame%d: regs=%s ",
3437 bt->frame, env->tmp_str_buf);
3438 fmt_stack_mask(env->tmp_str_buf, TMP_STR_BUF_LEN, bt_stack_mask(bt));
3439 verbose(env, "stack=%s before ", env->tmp_str_buf);
3440 verbose(env, "%d: ", idx);
3441 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
3444 if (class == BPF_ALU || class == BPF_ALU64) {
3445 if (!bt_is_reg_set(bt, dreg))
3447 if (opcode == BPF_MOV) {
3448 if (BPF_SRC(insn->code) == BPF_X) {
3449 /* dreg = sreg or dreg = (s8, s16, s32)sreg
3450 * dreg needs precision after this insn
3451 * sreg needs precision before this insn
3453 bt_clear_reg(bt, dreg);
3454 bt_set_reg(bt, sreg);
3457 * dreg needs precision after this insn.
3458 * Corresponding register is already marked
3459 * as precise=true in this verifier state.
3460 * No further markings in parent are necessary
3462 bt_clear_reg(bt, dreg);
3465 if (BPF_SRC(insn->code) == BPF_X) {
3467 * both dreg and sreg need precision
3470 bt_set_reg(bt, sreg);
3472 * dreg still needs precision before this insn
3475 } else if (class == BPF_LDX) {
3476 if (!bt_is_reg_set(bt, dreg))
3478 bt_clear_reg(bt, dreg);
3480 /* scalars can only be spilled into stack w/o losing precision.
3481 * Load from any other memory can be zero extended.
3482 * The desire to keep that precision is already indicated
3483 * by 'precise' mark in corresponding register of this state.
3484 * No further tracking necessary.
3486 if (insn->src_reg != BPF_REG_FP)
3489 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
3490 * that [fp - off] slot contains scalar that needs to be
3491 * tracked with precision
3493 spi = (-insn->off - 1) / BPF_REG_SIZE;
3495 verbose(env, "BUG spi %d\n", spi);
3496 WARN_ONCE(1, "verifier backtracking bug");
3499 bt_set_slot(bt, spi);
3500 } else if (class == BPF_STX || class == BPF_ST) {
3501 if (bt_is_reg_set(bt, dreg))
3502 /* stx & st shouldn't be using _scalar_ dst_reg
3503 * to access memory. It means backtracking
3504 * encountered a case of pointer subtraction.
3507 /* scalars can only be spilled into stack */
3508 if (insn->dst_reg != BPF_REG_FP)
3510 spi = (-insn->off - 1) / BPF_REG_SIZE;
3512 verbose(env, "BUG spi %d\n", spi);
3513 WARN_ONCE(1, "verifier backtracking bug");
3516 if (!bt_is_slot_set(bt, spi))
3518 bt_clear_slot(bt, spi);
3519 if (class == BPF_STX)
3520 bt_set_reg(bt, sreg);
3521 } else if (class == BPF_JMP || class == BPF_JMP32) {
3522 if (bpf_pseudo_call(insn)) {
3523 int subprog_insn_idx, subprog;
3525 subprog_insn_idx = idx + insn->imm + 1;
3526 subprog = find_subprog(env, subprog_insn_idx);
3530 if (subprog_is_global(env, subprog)) {
3531 /* check that jump history doesn't have any
3532 * extra instructions from subprog; the next
3533 * instruction after call to global subprog
3534 * should be literally next instruction in
3537 WARN_ONCE(idx + 1 != subseq_idx, "verifier backtracking bug");
3538 /* r1-r5 are invalidated after subprog call,
3539 * so for global func call it shouldn't be set
3542 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3543 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3544 WARN_ONCE(1, "verifier backtracking bug");
3547 /* global subprog always sets R0 */
3548 bt_clear_reg(bt, BPF_REG_0);
3551 /* static subprog call instruction, which
3552 * means that we are exiting current subprog,
3553 * so only r1-r5 could be still requested as
3554 * precise, r0 and r6-r10 or any stack slot in
3555 * the current frame should be zero by now
3557 if (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) {
3558 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3559 WARN_ONCE(1, "verifier backtracking bug");
3562 /* we don't track register spills perfectly,
3563 * so fallback to force-precise instead of failing */
3564 if (bt_stack_mask(bt) != 0)
3566 /* propagate r1-r5 to the caller */
3567 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
3568 if (bt_is_reg_set(bt, i)) {
3569 bt_clear_reg(bt, i);
3570 bt_set_frame_reg(bt, bt->frame - 1, i);
3573 if (bt_subprog_exit(bt))
3577 } else if ((bpf_helper_call(insn) &&
3578 is_callback_calling_function(insn->imm) &&
3579 !is_async_callback_calling_function(insn->imm)) ||
3580 (bpf_pseudo_kfunc_call(insn) && is_callback_calling_kfunc(insn->imm))) {
3581 /* callback-calling helper or kfunc call, which means
3582 * we are exiting from subprog, but unlike the subprog
3583 * call handling above, we shouldn't propagate
3584 * precision of r1-r5 (if any requested), as they are
3585 * not actually arguments passed directly to callback
3588 if (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) {
3589 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3590 WARN_ONCE(1, "verifier backtracking bug");
3593 if (bt_stack_mask(bt) != 0)
3595 /* clear r1-r5 in callback subprog's mask */
3596 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
3597 bt_clear_reg(bt, i);
3598 if (bt_subprog_exit(bt))
3601 } else if (opcode == BPF_CALL) {
3602 /* kfunc with imm==0 is invalid and fixup_kfunc_call will
3603 * catch this error later. Make backtracking conservative
3606 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && insn->imm == 0)
3608 /* regular helper call sets R0 */
3609 bt_clear_reg(bt, BPF_REG_0);
3610 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3611 /* if backtracing was looking for registers R1-R5
3612 * they should have been found already.
3614 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3615 WARN_ONCE(1, "verifier backtracking bug");
3618 } else if (opcode == BPF_EXIT) {
3621 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3622 /* if backtracing was looking for registers R1-R5
3623 * they should have been found already.
3625 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3626 WARN_ONCE(1, "verifier backtracking bug");
3630 /* BPF_EXIT in subprog or callback always returns
3631 * right after the call instruction, so by checking
3632 * whether the instruction at subseq_idx-1 is subprog
3633 * call or not we can distinguish actual exit from
3634 * *subprog* from exit from *callback*. In the former
3635 * case, we need to propagate r0 precision, if
3636 * necessary. In the former we never do that.
3638 r0_precise = subseq_idx - 1 >= 0 &&
3639 bpf_pseudo_call(&env->prog->insnsi[subseq_idx - 1]) &&
3640 bt_is_reg_set(bt, BPF_REG_0);
3642 bt_clear_reg(bt, BPF_REG_0);
3643 if (bt_subprog_enter(bt))
3647 bt_set_reg(bt, BPF_REG_0);
3648 /* r6-r9 and stack slots will stay set in caller frame
3649 * bitmasks until we return back from callee(s)
3652 } else if (BPF_SRC(insn->code) == BPF_X) {
3653 if (!bt_is_reg_set(bt, dreg) && !bt_is_reg_set(bt, sreg))
3656 * Both dreg and sreg need precision before
3657 * this insn. If only sreg was marked precise
3658 * before it would be equally necessary to
3659 * propagate it to dreg.
3661 bt_set_reg(bt, dreg);
3662 bt_set_reg(bt, sreg);
3663 /* else dreg <cond> K
3664 * Only dreg still needs precision before
3665 * this insn, so for the K-based conditional
3666 * there is nothing new to be marked.
3669 } else if (class == BPF_LD) {
3670 if (!bt_is_reg_set(bt, dreg))
3672 bt_clear_reg(bt, dreg);
3673 /* It's ld_imm64 or ld_abs or ld_ind.
3674 * For ld_imm64 no further tracking of precision
3675 * into parent is necessary
3677 if (mode == BPF_IND || mode == BPF_ABS)
3678 /* to be analyzed */
3684 /* the scalar precision tracking algorithm:
3685 * . at the start all registers have precise=false.
3686 * . scalar ranges are tracked as normal through alu and jmp insns.
3687 * . once precise value of the scalar register is used in:
3688 * . ptr + scalar alu
3689 * . if (scalar cond K|scalar)
3690 * . helper_call(.., scalar, ...) where ARG_CONST is expected
3691 * backtrack through the verifier states and mark all registers and
3692 * stack slots with spilled constants that these scalar regisers
3693 * should be precise.
3694 * . during state pruning two registers (or spilled stack slots)
3695 * are equivalent if both are not precise.
3697 * Note the verifier cannot simply walk register parentage chain,
3698 * since many different registers and stack slots could have been
3699 * used to compute single precise scalar.
3701 * The approach of starting with precise=true for all registers and then
3702 * backtrack to mark a register as not precise when the verifier detects
3703 * that program doesn't care about specific value (e.g., when helper
3704 * takes register as ARG_ANYTHING parameter) is not safe.
3706 * It's ok to walk single parentage chain of the verifier states.
3707 * It's possible that this backtracking will go all the way till 1st insn.
3708 * All other branches will be explored for needing precision later.
3710 * The backtracking needs to deal with cases like:
3711 * 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)
3714 * if r5 > 0x79f goto pc+7
3715 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
3718 * call bpf_perf_event_output#25
3719 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
3723 * call foo // uses callee's r6 inside to compute r0
3727 * to track above reg_mask/stack_mask needs to be independent for each frame.
3729 * Also if parent's curframe > frame where backtracking started,
3730 * the verifier need to mark registers in both frames, otherwise callees
3731 * may incorrectly prune callers. This is similar to
3732 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
3734 * For now backtracking falls back into conservative marking.
3736 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
3737 struct bpf_verifier_state *st)
3739 struct bpf_func_state *func;
3740 struct bpf_reg_state *reg;
3743 if (env->log.level & BPF_LOG_LEVEL2) {
3744 verbose(env, "mark_precise: frame%d: falling back to forcing all scalars precise\n",
3748 /* big hammer: mark all scalars precise in this path.
3749 * pop_stack may still get !precise scalars.
3750 * We also skip current state and go straight to first parent state,
3751 * because precision markings in current non-checkpointed state are
3752 * not needed. See why in the comment in __mark_chain_precision below.
3754 for (st = st->parent; st; st = st->parent) {
3755 for (i = 0; i <= st->curframe; i++) {
3756 func = st->frame[i];
3757 for (j = 0; j < BPF_REG_FP; j++) {
3758 reg = &func->regs[j];
3759 if (reg->type != SCALAR_VALUE || reg->precise)
3761 reg->precise = true;
3762 if (env->log.level & BPF_LOG_LEVEL2) {
3763 verbose(env, "force_precise: frame%d: forcing r%d to be precise\n",
3767 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
3768 if (!is_spilled_reg(&func->stack[j]))
3770 reg = &func->stack[j].spilled_ptr;
3771 if (reg->type != SCALAR_VALUE || reg->precise)
3773 reg->precise = true;
3774 if (env->log.level & BPF_LOG_LEVEL2) {
3775 verbose(env, "force_precise: frame%d: forcing fp%d to be precise\n",
3783 static void mark_all_scalars_imprecise(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
3785 struct bpf_func_state *func;
3786 struct bpf_reg_state *reg;
3789 for (i = 0; i <= st->curframe; i++) {
3790 func = st->frame[i];
3791 for (j = 0; j < BPF_REG_FP; j++) {
3792 reg = &func->regs[j];
3793 if (reg->type != SCALAR_VALUE)
3795 reg->precise = false;
3797 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
3798 if (!is_spilled_reg(&func->stack[j]))
3800 reg = &func->stack[j].spilled_ptr;
3801 if (reg->type != SCALAR_VALUE)
3803 reg->precise = false;
3808 static bool idset_contains(struct bpf_idset *s, u32 id)
3812 for (i = 0; i < s->count; ++i)
3813 if (s->ids[i] == id)
3819 static int idset_push(struct bpf_idset *s, u32 id)
3821 if (WARN_ON_ONCE(s->count >= ARRAY_SIZE(s->ids)))
3823 s->ids[s->count++] = id;
3827 static void idset_reset(struct bpf_idset *s)
3832 /* Collect a set of IDs for all registers currently marked as precise in env->bt.
3833 * Mark all registers with these IDs as precise.
3835 static int mark_precise_scalar_ids(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
3837 struct bpf_idset *precise_ids = &env->idset_scratch;
3838 struct backtrack_state *bt = &env->bt;
3839 struct bpf_func_state *func;
3840 struct bpf_reg_state *reg;
3841 DECLARE_BITMAP(mask, 64);
3844 idset_reset(precise_ids);
3846 for (fr = bt->frame; fr >= 0; fr--) {
3847 func = st->frame[fr];
3849 bitmap_from_u64(mask, bt_frame_reg_mask(bt, fr));
3850 for_each_set_bit(i, mask, 32) {
3851 reg = &func->regs[i];
3852 if (!reg->id || reg->type != SCALAR_VALUE)
3854 if (idset_push(precise_ids, reg->id))
3858 bitmap_from_u64(mask, bt_frame_stack_mask(bt, fr));
3859 for_each_set_bit(i, mask, 64) {
3860 if (i >= func->allocated_stack / BPF_REG_SIZE)
3862 if (!is_spilled_scalar_reg(&func->stack[i]))
3864 reg = &func->stack[i].spilled_ptr;
3867 if (idset_push(precise_ids, reg->id))
3872 for (fr = 0; fr <= st->curframe; ++fr) {
3873 func = st->frame[fr];
3875 for (i = BPF_REG_0; i < BPF_REG_10; ++i) {
3876 reg = &func->regs[i];
3879 if (!idset_contains(precise_ids, reg->id))
3881 bt_set_frame_reg(bt, fr, i);
3883 for (i = 0; i < func->allocated_stack / BPF_REG_SIZE; ++i) {
3884 if (!is_spilled_scalar_reg(&func->stack[i]))
3886 reg = &func->stack[i].spilled_ptr;
3889 if (!idset_contains(precise_ids, reg->id))
3891 bt_set_frame_slot(bt, fr, i);
3899 * __mark_chain_precision() backtracks BPF program instruction sequence and
3900 * chain of verifier states making sure that register *regno* (if regno >= 0)
3901 * and/or stack slot *spi* (if spi >= 0) are marked as precisely tracked
3902 * SCALARS, as well as any other registers and slots that contribute to
3903 * a tracked state of given registers/stack slots, depending on specific BPF
3904 * assembly instructions (see backtrack_insns() for exact instruction handling
3905 * logic). This backtracking relies on recorded jmp_history and is able to
3906 * traverse entire chain of parent states. This process ends only when all the
3907 * necessary registers/slots and their transitive dependencies are marked as
3910 * One important and subtle aspect is that precise marks *do not matter* in
3911 * the currently verified state (current state). It is important to understand
3912 * why this is the case.
3914 * First, note that current state is the state that is not yet "checkpointed",
3915 * i.e., it is not yet put into env->explored_states, and it has no children
3916 * states as well. It's ephemeral, and can end up either a) being discarded if
3917 * compatible explored state is found at some point or BPF_EXIT instruction is
3918 * reached or b) checkpointed and put into env->explored_states, branching out
3919 * into one or more children states.
3921 * In the former case, precise markings in current state are completely
3922 * ignored by state comparison code (see regsafe() for details). Only
3923 * checkpointed ("old") state precise markings are important, and if old
3924 * state's register/slot is precise, regsafe() assumes current state's
3925 * register/slot as precise and checks value ranges exactly and precisely. If
3926 * states turn out to be compatible, current state's necessary precise
3927 * markings and any required parent states' precise markings are enforced
3928 * after the fact with propagate_precision() logic, after the fact. But it's
3929 * important to realize that in this case, even after marking current state
3930 * registers/slots as precise, we immediately discard current state. So what
3931 * actually matters is any of the precise markings propagated into current
3932 * state's parent states, which are always checkpointed (due to b) case above).
3933 * As such, for scenario a) it doesn't matter if current state has precise
3934 * markings set or not.
3936 * Now, for the scenario b), checkpointing and forking into child(ren)
3937 * state(s). Note that before current state gets to checkpointing step, any
3938 * processed instruction always assumes precise SCALAR register/slot
3939 * knowledge: if precise value or range is useful to prune jump branch, BPF
3940 * verifier takes this opportunity enthusiastically. Similarly, when
3941 * register's value is used to calculate offset or memory address, exact
3942 * knowledge of SCALAR range is assumed, checked, and enforced. So, similar to
3943 * what we mentioned above about state comparison ignoring precise markings
3944 * during state comparison, BPF verifier ignores and also assumes precise
3945 * markings *at will* during instruction verification process. But as verifier
3946 * assumes precision, it also propagates any precision dependencies across
3947 * parent states, which are not yet finalized, so can be further restricted
3948 * based on new knowledge gained from restrictions enforced by their children
3949 * states. This is so that once those parent states are finalized, i.e., when
3950 * they have no more active children state, state comparison logic in
3951 * is_state_visited() would enforce strict and precise SCALAR ranges, if
3952 * required for correctness.
3954 * To build a bit more intuition, note also that once a state is checkpointed,
3955 * the path we took to get to that state is not important. This is crucial
3956 * property for state pruning. When state is checkpointed and finalized at
3957 * some instruction index, it can be correctly and safely used to "short
3958 * circuit" any *compatible* state that reaches exactly the same instruction
3959 * index. I.e., if we jumped to that instruction from a completely different
3960 * code path than original finalized state was derived from, it doesn't
3961 * matter, current state can be discarded because from that instruction
3962 * forward having a compatible state will ensure we will safely reach the
3963 * exit. States describe preconditions for further exploration, but completely
3964 * forget the history of how we got here.
3966 * This also means that even if we needed precise SCALAR range to get to
3967 * finalized state, but from that point forward *that same* SCALAR register is
3968 * never used in a precise context (i.e., it's precise value is not needed for
3969 * correctness), it's correct and safe to mark such register as "imprecise"
3970 * (i.e., precise marking set to false). This is what we rely on when we do
3971 * not set precise marking in current state. If no child state requires
3972 * precision for any given SCALAR register, it's safe to dictate that it can
3973 * be imprecise. If any child state does require this register to be precise,
3974 * we'll mark it precise later retroactively during precise markings
3975 * propagation from child state to parent states.
3977 * Skipping precise marking setting in current state is a mild version of
3978 * relying on the above observation. But we can utilize this property even
3979 * more aggressively by proactively forgetting any precise marking in the
3980 * current state (which we inherited from the parent state), right before we
3981 * checkpoint it and branch off into new child state. This is done by
3982 * mark_all_scalars_imprecise() to hopefully get more permissive and generic
3983 * finalized states which help in short circuiting more future states.
3985 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno)
3987 struct backtrack_state *bt = &env->bt;
3988 struct bpf_verifier_state *st = env->cur_state;
3989 int first_idx = st->first_insn_idx;
3990 int last_idx = env->insn_idx;
3991 int subseq_idx = -1;
3992 struct bpf_func_state *func;
3993 struct bpf_reg_state *reg;
3994 bool skip_first = true;
3997 if (!env->bpf_capable)
4000 /* set frame number from which we are starting to backtrack */
4001 bt_init(bt, env->cur_state->curframe);
4003 /* Do sanity checks against current state of register and/or stack
4004 * slot, but don't set precise flag in current state, as precision
4005 * tracking in the current state is unnecessary.
4007 func = st->frame[bt->frame];
4009 reg = &func->regs[regno];
4010 if (reg->type != SCALAR_VALUE) {
4011 WARN_ONCE(1, "backtracing misuse");
4014 bt_set_reg(bt, regno);
4021 DECLARE_BITMAP(mask, 64);
4022 u32 history = st->jmp_history_cnt;
4024 if (env->log.level & BPF_LOG_LEVEL2) {
4025 verbose(env, "mark_precise: frame%d: last_idx %d first_idx %d subseq_idx %d \n",
4026 bt->frame, last_idx, first_idx, subseq_idx);
4029 /* If some register with scalar ID is marked as precise,
4030 * make sure that all registers sharing this ID are also precise.
4031 * This is needed to estimate effect of find_equal_scalars().
4032 * Do this at the last instruction of each state,
4033 * bpf_reg_state::id fields are valid for these instructions.
4035 * Allows to track precision in situation like below:
4037 * r2 = unknown value
4041 * r1 = r2 // r1 and r2 now share the same ID
4043 * --- state #1 {r1.id = A, r2.id = A} ---
4045 * if (r2 > 10) goto exit; // find_equal_scalars() assigns range to r1
4047 * --- state #2 {r1.id = A, r2.id = A} ---
4049 * r3 += r1 // need to mark both r1 and r2
4051 if (mark_precise_scalar_ids(env, st))
4055 /* we are at the entry into subprog, which
4056 * is expected for global funcs, but only if
4057 * requested precise registers are R1-R5
4058 * (which are global func's input arguments)
4060 if (st->curframe == 0 &&
4061 st->frame[0]->subprogno > 0 &&
4062 st->frame[0]->callsite == BPF_MAIN_FUNC &&
4063 bt_stack_mask(bt) == 0 &&
4064 (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) == 0) {
4065 bitmap_from_u64(mask, bt_reg_mask(bt));
4066 for_each_set_bit(i, mask, 32) {
4067 reg = &st->frame[0]->regs[i];
4068 bt_clear_reg(bt, i);
4069 if (reg->type == SCALAR_VALUE)
4070 reg->precise = true;
4075 verbose(env, "BUG backtracking func entry subprog %d reg_mask %x stack_mask %llx\n",
4076 st->frame[0]->subprogno, bt_reg_mask(bt), bt_stack_mask(bt));
4077 WARN_ONCE(1, "verifier backtracking bug");
4081 for (i = last_idx;;) {
4086 err = backtrack_insn(env, i, subseq_idx, bt);
4088 if (err == -ENOTSUPP) {
4089 mark_all_scalars_precise(env, env->cur_state);
4096 /* Found assignment(s) into tracked register in this state.
4097 * Since this state is already marked, just return.
4098 * Nothing to be tracked further in the parent state.
4102 i = get_prev_insn_idx(st, i, &history);
4105 if (i >= env->prog->len) {
4106 /* This can happen if backtracking reached insn 0
4107 * and there are still reg_mask or stack_mask
4109 * It means the backtracking missed the spot where
4110 * particular register was initialized with a constant.
4112 verbose(env, "BUG backtracking idx %d\n", i);
4113 WARN_ONCE(1, "verifier backtracking bug");
4121 for (fr = bt->frame; fr >= 0; fr--) {
4122 func = st->frame[fr];
4123 bitmap_from_u64(mask, bt_frame_reg_mask(bt, fr));
4124 for_each_set_bit(i, mask, 32) {
4125 reg = &func->regs[i];
4126 if (reg->type != SCALAR_VALUE) {
4127 bt_clear_frame_reg(bt, fr, i);
4131 bt_clear_frame_reg(bt, fr, i);
4133 reg->precise = true;
4136 bitmap_from_u64(mask, bt_frame_stack_mask(bt, fr));
4137 for_each_set_bit(i, mask, 64) {
4138 if (i >= func->allocated_stack / BPF_REG_SIZE) {
4139 /* the sequence of instructions:
4141 * 3: (7b) *(u64 *)(r3 -8) = r0
4142 * 4: (79) r4 = *(u64 *)(r10 -8)
4143 * doesn't contain jmps. It's backtracked
4144 * as a single block.
4145 * During backtracking insn 3 is not recognized as
4146 * stack access, so at the end of backtracking
4147 * stack slot fp-8 is still marked in stack_mask.
4148 * However the parent state may not have accessed
4149 * fp-8 and it's "unallocated" stack space.
4150 * In such case fallback to conservative.
4152 mark_all_scalars_precise(env, env->cur_state);
4157 if (!is_spilled_scalar_reg(&func->stack[i])) {
4158 bt_clear_frame_slot(bt, fr, i);
4161 reg = &func->stack[i].spilled_ptr;
4163 bt_clear_frame_slot(bt, fr, i);
4165 reg->precise = true;
4167 if (env->log.level & BPF_LOG_LEVEL2) {
4168 fmt_reg_mask(env->tmp_str_buf, TMP_STR_BUF_LEN,
4169 bt_frame_reg_mask(bt, fr));
4170 verbose(env, "mark_precise: frame%d: parent state regs=%s ",
4171 fr, env->tmp_str_buf);
4172 fmt_stack_mask(env->tmp_str_buf, TMP_STR_BUF_LEN,
4173 bt_frame_stack_mask(bt, fr));
4174 verbose(env, "stack=%s: ", env->tmp_str_buf);
4175 print_verifier_state(env, func, true);
4182 subseq_idx = first_idx;
4183 last_idx = st->last_insn_idx;
4184 first_idx = st->first_insn_idx;
4187 /* if we still have requested precise regs or slots, we missed
4188 * something (e.g., stack access through non-r10 register), so
4189 * fallback to marking all precise
4191 if (!bt_empty(bt)) {
4192 mark_all_scalars_precise(env, env->cur_state);
4199 int mark_chain_precision(struct bpf_verifier_env *env, int regno)
4201 return __mark_chain_precision(env, regno);
4204 /* mark_chain_precision_batch() assumes that env->bt is set in the caller to
4205 * desired reg and stack masks across all relevant frames
4207 static int mark_chain_precision_batch(struct bpf_verifier_env *env)
4209 return __mark_chain_precision(env, -1);
4212 static bool is_spillable_regtype(enum bpf_reg_type type)
4214 switch (base_type(type)) {
4215 case PTR_TO_MAP_VALUE:
4219 case PTR_TO_PACKET_META:
4220 case PTR_TO_PACKET_END:
4221 case PTR_TO_FLOW_KEYS:
4222 case CONST_PTR_TO_MAP:
4224 case PTR_TO_SOCK_COMMON:
4225 case PTR_TO_TCP_SOCK:
4226 case PTR_TO_XDP_SOCK:
4231 case PTR_TO_MAP_KEY:
4238 /* Does this register contain a constant zero? */
4239 static bool register_is_null(struct bpf_reg_state *reg)
4241 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
4244 static bool register_is_const(struct bpf_reg_state *reg)
4246 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
4249 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
4251 return tnum_is_unknown(reg->var_off) &&
4252 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
4253 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
4254 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
4255 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
4258 static bool register_is_bounded(struct bpf_reg_state *reg)
4260 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
4263 static bool __is_pointer_value(bool allow_ptr_leaks,
4264 const struct bpf_reg_state *reg)
4266 if (allow_ptr_leaks)
4269 return reg->type != SCALAR_VALUE;
4272 /* Copy src state preserving dst->parent and dst->live fields */
4273 static void copy_register_state(struct bpf_reg_state *dst, const struct bpf_reg_state *src)
4275 struct bpf_reg_state *parent = dst->parent;
4276 enum bpf_reg_liveness live = dst->live;
4279 dst->parent = parent;
4283 static void save_register_state(struct bpf_func_state *state,
4284 int spi, struct bpf_reg_state *reg,
4289 copy_register_state(&state->stack[spi].spilled_ptr, reg);
4290 if (size == BPF_REG_SIZE)
4291 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
4293 for (i = BPF_REG_SIZE; i > BPF_REG_SIZE - size; i--)
4294 state->stack[spi].slot_type[i - 1] = STACK_SPILL;
4296 /* size < 8 bytes spill */
4298 scrub_spilled_slot(&state->stack[spi].slot_type[i - 1]);
4301 static bool is_bpf_st_mem(struct bpf_insn *insn)
4303 return BPF_CLASS(insn->code) == BPF_ST && BPF_MODE(insn->code) == BPF_MEM;
4306 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
4307 * stack boundary and alignment are checked in check_mem_access()
4309 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
4310 /* stack frame we're writing to */
4311 struct bpf_func_state *state,
4312 int off, int size, int value_regno,
4315 struct bpf_func_state *cur; /* state of the current function */
4316 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
4317 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
4318 struct bpf_reg_state *reg = NULL;
4319 u32 dst_reg = insn->dst_reg;
4321 err = grow_stack_state(state, round_up(slot + 1, BPF_REG_SIZE));
4324 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
4325 * so it's aligned access and [off, off + size) are within stack limits
4327 if (!env->allow_ptr_leaks &&
4328 state->stack[spi].slot_type[0] == STACK_SPILL &&
4329 size != BPF_REG_SIZE) {
4330 verbose(env, "attempt to corrupt spilled pointer on stack\n");
4334 cur = env->cur_state->frame[env->cur_state->curframe];
4335 if (value_regno >= 0)
4336 reg = &cur->regs[value_regno];
4337 if (!env->bypass_spec_v4) {
4338 bool sanitize = reg && is_spillable_regtype(reg->type);
4340 for (i = 0; i < size; i++) {
4341 u8 type = state->stack[spi].slot_type[i];
4343 if (type != STACK_MISC && type != STACK_ZERO) {
4350 env->insn_aux_data[insn_idx].sanitize_stack_spill = true;
4353 err = destroy_if_dynptr_stack_slot(env, state, spi);
4357 mark_stack_slot_scratched(env, spi);
4358 if (reg && !(off % BPF_REG_SIZE) && register_is_bounded(reg) &&
4359 !register_is_null(reg) && env->bpf_capable) {
4360 if (dst_reg != BPF_REG_FP) {
4361 /* The backtracking logic can only recognize explicit
4362 * stack slot address like [fp - 8]. Other spill of
4363 * scalar via different register has to be conservative.
4364 * Backtrack from here and mark all registers as precise
4365 * that contributed into 'reg' being a constant.
4367 err = mark_chain_precision(env, value_regno);
4371 save_register_state(state, spi, reg, size);
4372 /* Break the relation on a narrowing spill. */
4373 if (fls64(reg->umax_value) > BITS_PER_BYTE * size)
4374 state->stack[spi].spilled_ptr.id = 0;
4375 } else if (!reg && !(off % BPF_REG_SIZE) && is_bpf_st_mem(insn) &&
4376 insn->imm != 0 && env->bpf_capable) {
4377 struct bpf_reg_state fake_reg = {};
4379 __mark_reg_known(&fake_reg, (u32)insn->imm);
4380 fake_reg.type = SCALAR_VALUE;
4381 save_register_state(state, spi, &fake_reg, size);
4382 } else if (reg && is_spillable_regtype(reg->type)) {
4383 /* register containing pointer is being spilled into stack */
4384 if (size != BPF_REG_SIZE) {
4385 verbose_linfo(env, insn_idx, "; ");
4386 verbose(env, "invalid size of register spill\n");
4389 if (state != cur && reg->type == PTR_TO_STACK) {
4390 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
4393 save_register_state(state, spi, reg, size);
4395 u8 type = STACK_MISC;
4397 /* regular write of data into stack destroys any spilled ptr */
4398 state->stack[spi].spilled_ptr.type = NOT_INIT;
4399 /* Mark slots as STACK_MISC if they belonged to spilled ptr/dynptr/iter. */
4400 if (is_stack_slot_special(&state->stack[spi]))
4401 for (i = 0; i < BPF_REG_SIZE; i++)
4402 scrub_spilled_slot(&state->stack[spi].slot_type[i]);
4404 /* only mark the slot as written if all 8 bytes were written
4405 * otherwise read propagation may incorrectly stop too soon
4406 * when stack slots are partially written.
4407 * This heuristic means that read propagation will be
4408 * conservative, since it will add reg_live_read marks
4409 * to stack slots all the way to first state when programs
4410 * writes+reads less than 8 bytes
4412 if (size == BPF_REG_SIZE)
4413 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
4415 /* when we zero initialize stack slots mark them as such */
4416 if ((reg && register_is_null(reg)) ||
4417 (!reg && is_bpf_st_mem(insn) && insn->imm == 0)) {
4418 /* backtracking doesn't work for STACK_ZERO yet. */
4419 err = mark_chain_precision(env, value_regno);
4425 /* Mark slots affected by this stack write. */
4426 for (i = 0; i < size; i++)
4427 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
4433 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
4434 * known to contain a variable offset.
4435 * This function checks whether the write is permitted and conservatively
4436 * tracks the effects of the write, considering that each stack slot in the
4437 * dynamic range is potentially written to.
4439 * 'off' includes 'regno->off'.
4440 * 'value_regno' can be -1, meaning that an unknown value is being written to
4443 * Spilled pointers in range are not marked as written because we don't know
4444 * what's going to be actually written. This means that read propagation for
4445 * future reads cannot be terminated by this write.
4447 * For privileged programs, uninitialized stack slots are considered
4448 * initialized by this write (even though we don't know exactly what offsets
4449 * are going to be written to). The idea is that we don't want the verifier to
4450 * reject future reads that access slots written to through variable offsets.
4452 static int check_stack_write_var_off(struct bpf_verifier_env *env,
4453 /* func where register points to */
4454 struct bpf_func_state *state,
4455 int ptr_regno, int off, int size,
4456 int value_regno, int insn_idx)
4458 struct bpf_func_state *cur; /* state of the current function */
4459 int min_off, max_off;
4461 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
4462 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
4463 bool writing_zero = false;
4464 /* set if the fact that we're writing a zero is used to let any
4465 * stack slots remain STACK_ZERO
4467 bool zero_used = false;
4469 cur = env->cur_state->frame[env->cur_state->curframe];
4470 ptr_reg = &cur->regs[ptr_regno];
4471 min_off = ptr_reg->smin_value + off;
4472 max_off = ptr_reg->smax_value + off + size;
4473 if (value_regno >= 0)
4474 value_reg = &cur->regs[value_regno];
4475 if ((value_reg && register_is_null(value_reg)) ||
4476 (!value_reg && is_bpf_st_mem(insn) && insn->imm == 0))
4477 writing_zero = true;
4479 err = grow_stack_state(state, round_up(-min_off, BPF_REG_SIZE));
4483 for (i = min_off; i < max_off; i++) {
4487 err = destroy_if_dynptr_stack_slot(env, state, spi);
4492 /* Variable offset writes destroy any spilled pointers in range. */
4493 for (i = min_off; i < max_off; i++) {
4494 u8 new_type, *stype;
4498 spi = slot / BPF_REG_SIZE;
4499 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
4500 mark_stack_slot_scratched(env, spi);
4502 if (!env->allow_ptr_leaks && *stype != STACK_MISC && *stype != STACK_ZERO) {
4503 /* Reject the write if range we may write to has not
4504 * been initialized beforehand. If we didn't reject
4505 * here, the ptr status would be erased below (even
4506 * though not all slots are actually overwritten),
4507 * possibly opening the door to leaks.
4509 * We do however catch STACK_INVALID case below, and
4510 * only allow reading possibly uninitialized memory
4511 * later for CAP_PERFMON, as the write may not happen to
4514 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
4519 /* Erase all spilled pointers. */
4520 state->stack[spi].spilled_ptr.type = NOT_INIT;
4522 /* Update the slot type. */
4523 new_type = STACK_MISC;
4524 if (writing_zero && *stype == STACK_ZERO) {
4525 new_type = STACK_ZERO;
4528 /* If the slot is STACK_INVALID, we check whether it's OK to
4529 * pretend that it will be initialized by this write. The slot
4530 * might not actually be written to, and so if we mark it as
4531 * initialized future reads might leak uninitialized memory.
4532 * For privileged programs, we will accept such reads to slots
4533 * that may or may not be written because, if we're reject
4534 * them, the error would be too confusing.
4536 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
4537 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
4544 /* backtracking doesn't work for STACK_ZERO yet. */
4545 err = mark_chain_precision(env, value_regno);
4552 /* When register 'dst_regno' is assigned some values from stack[min_off,
4553 * max_off), we set the register's type according to the types of the
4554 * respective stack slots. If all the stack values are known to be zeros, then
4555 * so is the destination reg. Otherwise, the register is considered to be
4556 * SCALAR. This function does not deal with register filling; the caller must
4557 * ensure that all spilled registers in the stack range have been marked as
4560 static void mark_reg_stack_read(struct bpf_verifier_env *env,
4561 /* func where src register points to */
4562 struct bpf_func_state *ptr_state,
4563 int min_off, int max_off, int dst_regno)
4565 struct bpf_verifier_state *vstate = env->cur_state;
4566 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4571 for (i = min_off; i < max_off; i++) {
4573 spi = slot / BPF_REG_SIZE;
4574 mark_stack_slot_scratched(env, spi);
4575 stype = ptr_state->stack[spi].slot_type;
4576 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
4580 if (zeros == max_off - min_off) {
4581 /* any access_size read into register is zero extended,
4582 * so the whole register == const_zero
4584 __mark_reg_const_zero(&state->regs[dst_regno]);
4585 /* backtracking doesn't support STACK_ZERO yet,
4586 * so mark it precise here, so that later
4587 * backtracking can stop here.
4588 * Backtracking may not need this if this register
4589 * doesn't participate in pointer adjustment.
4590 * Forward propagation of precise flag is not
4591 * necessary either. This mark is only to stop
4592 * backtracking. Any register that contributed
4593 * to const 0 was marked precise before spill.
4595 state->regs[dst_regno].precise = true;
4597 /* have read misc data from the stack */
4598 mark_reg_unknown(env, state->regs, dst_regno);
4600 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4603 /* Read the stack at 'off' and put the results into the register indicated by
4604 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
4607 * 'dst_regno' can be -1, meaning that the read value is not going to a
4610 * The access is assumed to be within the current stack bounds.
4612 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
4613 /* func where src register points to */
4614 struct bpf_func_state *reg_state,
4615 int off, int size, int dst_regno)
4617 struct bpf_verifier_state *vstate = env->cur_state;
4618 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4619 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
4620 struct bpf_reg_state *reg;
4623 stype = reg_state->stack[spi].slot_type;
4624 reg = ®_state->stack[spi].spilled_ptr;
4626 mark_stack_slot_scratched(env, spi);
4628 if (is_spilled_reg(®_state->stack[spi])) {
4631 for (i = BPF_REG_SIZE - 1; i > 0 && stype[i - 1] == STACK_SPILL; i--)
4634 if (size != BPF_REG_SIZE || spill_size != BPF_REG_SIZE) {
4635 if (reg->type != SCALAR_VALUE) {
4636 verbose_linfo(env, env->insn_idx, "; ");
4637 verbose(env, "invalid size of register fill\n");
4641 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4645 if (!(off % BPF_REG_SIZE) && size == spill_size) {
4646 /* The earlier check_reg_arg() has decided the
4647 * subreg_def for this insn. Save it first.
4649 s32 subreg_def = state->regs[dst_regno].subreg_def;
4651 copy_register_state(&state->regs[dst_regno], reg);
4652 state->regs[dst_regno].subreg_def = subreg_def;
4654 for (i = 0; i < size; i++) {
4655 type = stype[(slot - i) % BPF_REG_SIZE];
4656 if (type == STACK_SPILL)
4658 if (type == STACK_MISC)
4660 if (type == STACK_INVALID && env->allow_uninit_stack)
4662 verbose(env, "invalid read from stack off %d+%d size %d\n",
4666 mark_reg_unknown(env, state->regs, dst_regno);
4668 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4672 if (dst_regno >= 0) {
4673 /* restore register state from stack */
4674 copy_register_state(&state->regs[dst_regno], reg);
4675 /* mark reg as written since spilled pointer state likely
4676 * has its liveness marks cleared by is_state_visited()
4677 * which resets stack/reg liveness for state transitions
4679 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4680 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
4681 /* If dst_regno==-1, the caller is asking us whether
4682 * it is acceptable to use this value as a SCALAR_VALUE
4684 * We must not allow unprivileged callers to do that
4685 * with spilled pointers.
4687 verbose(env, "leaking pointer from stack off %d\n",
4691 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4693 for (i = 0; i < size; i++) {
4694 type = stype[(slot - i) % BPF_REG_SIZE];
4695 if (type == STACK_MISC)
4697 if (type == STACK_ZERO)
4699 if (type == STACK_INVALID && env->allow_uninit_stack)
4701 verbose(env, "invalid read from stack off %d+%d size %d\n",
4705 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4707 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
4712 enum bpf_access_src {
4713 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
4714 ACCESS_HELPER = 2, /* the access is performed by a helper */
4717 static int check_stack_range_initialized(struct bpf_verifier_env *env,
4718 int regno, int off, int access_size,
4719 bool zero_size_allowed,
4720 enum bpf_access_src type,
4721 struct bpf_call_arg_meta *meta);
4723 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
4725 return cur_regs(env) + regno;
4728 /* Read the stack at 'ptr_regno + off' and put the result into the register
4730 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
4731 * but not its variable offset.
4732 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
4734 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
4735 * filling registers (i.e. reads of spilled register cannot be detected when
4736 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
4737 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
4738 * offset; for a fixed offset check_stack_read_fixed_off should be used
4741 static int check_stack_read_var_off(struct bpf_verifier_env *env,
4742 int ptr_regno, int off, int size, int dst_regno)
4744 /* The state of the source register. */
4745 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
4746 struct bpf_func_state *ptr_state = func(env, reg);
4748 int min_off, max_off;
4750 /* Note that we pass a NULL meta, so raw access will not be permitted.
4752 err = check_stack_range_initialized(env, ptr_regno, off, size,
4753 false, ACCESS_DIRECT, NULL);
4757 min_off = reg->smin_value + off;
4758 max_off = reg->smax_value + off;
4759 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
4763 /* check_stack_read dispatches to check_stack_read_fixed_off or
4764 * check_stack_read_var_off.
4766 * The caller must ensure that the offset falls within the allocated stack
4769 * 'dst_regno' is a register which will receive the value from the stack. It
4770 * can be -1, meaning that the read value is not going to a register.
4772 static int check_stack_read(struct bpf_verifier_env *env,
4773 int ptr_regno, int off, int size,
4776 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
4777 struct bpf_func_state *state = func(env, reg);
4779 /* Some accesses are only permitted with a static offset. */
4780 bool var_off = !tnum_is_const(reg->var_off);
4782 /* The offset is required to be static when reads don't go to a
4783 * register, in order to not leak pointers (see
4784 * check_stack_read_fixed_off).
4786 if (dst_regno < 0 && var_off) {
4789 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4790 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
4794 /* Variable offset is prohibited for unprivileged mode for simplicity
4795 * since it requires corresponding support in Spectre masking for stack
4796 * ALU. See also retrieve_ptr_limit(). The check in
4797 * check_stack_access_for_ptr_arithmetic() called by
4798 * adjust_ptr_min_max_vals() prevents users from creating stack pointers
4799 * with variable offsets, therefore no check is required here. Further,
4800 * just checking it here would be insufficient as speculative stack
4801 * writes could still lead to unsafe speculative behaviour.
4804 off += reg->var_off.value;
4805 err = check_stack_read_fixed_off(env, state, off, size,
4808 /* Variable offset stack reads need more conservative handling
4809 * than fixed offset ones. Note that dst_regno >= 0 on this
4812 err = check_stack_read_var_off(env, ptr_regno, off, size,
4819 /* check_stack_write dispatches to check_stack_write_fixed_off or
4820 * check_stack_write_var_off.
4822 * 'ptr_regno' is the register used as a pointer into the stack.
4823 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
4824 * 'value_regno' is the register whose value we're writing to the stack. It can
4825 * be -1, meaning that we're not writing from a register.
4827 * The caller must ensure that the offset falls within the maximum stack size.
4829 static int check_stack_write(struct bpf_verifier_env *env,
4830 int ptr_regno, int off, int size,
4831 int value_regno, int insn_idx)
4833 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
4834 struct bpf_func_state *state = func(env, reg);
4837 if (tnum_is_const(reg->var_off)) {
4838 off += reg->var_off.value;
4839 err = check_stack_write_fixed_off(env, state, off, size,
4840 value_regno, insn_idx);
4842 /* Variable offset stack reads need more conservative handling
4843 * than fixed offset ones.
4845 err = check_stack_write_var_off(env, state,
4846 ptr_regno, off, size,
4847 value_regno, insn_idx);
4852 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
4853 int off, int size, enum bpf_access_type type)
4855 struct bpf_reg_state *regs = cur_regs(env);
4856 struct bpf_map *map = regs[regno].map_ptr;
4857 u32 cap = bpf_map_flags_to_cap(map);
4859 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
4860 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
4861 map->value_size, off, size);
4865 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
4866 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
4867 map->value_size, off, size);
4874 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
4875 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
4876 int off, int size, u32 mem_size,
4877 bool zero_size_allowed)
4879 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
4880 struct bpf_reg_state *reg;
4882 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
4885 reg = &cur_regs(env)[regno];
4886 switch (reg->type) {
4887 case PTR_TO_MAP_KEY:
4888 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
4889 mem_size, off, size);
4891 case PTR_TO_MAP_VALUE:
4892 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
4893 mem_size, off, size);
4896 case PTR_TO_PACKET_META:
4897 case PTR_TO_PACKET_END:
4898 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
4899 off, size, regno, reg->id, off, mem_size);
4903 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
4904 mem_size, off, size);
4910 /* check read/write into a memory region with possible variable offset */
4911 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
4912 int off, int size, u32 mem_size,
4913 bool zero_size_allowed)
4915 struct bpf_verifier_state *vstate = env->cur_state;
4916 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4917 struct bpf_reg_state *reg = &state->regs[regno];
4920 /* We may have adjusted the register pointing to memory region, so we
4921 * need to try adding each of min_value and max_value to off
4922 * to make sure our theoretical access will be safe.
4924 * The minimum value is only important with signed
4925 * comparisons where we can't assume the floor of a
4926 * value is 0. If we are using signed variables for our
4927 * index'es we need to make sure that whatever we use
4928 * will have a set floor within our range.
4930 if (reg->smin_value < 0 &&
4931 (reg->smin_value == S64_MIN ||
4932 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
4933 reg->smin_value + off < 0)) {
4934 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
4938 err = __check_mem_access(env, regno, reg->smin_value + off, size,
4939 mem_size, zero_size_allowed);
4941 verbose(env, "R%d min value is outside of the allowed memory range\n",
4946 /* If we haven't set a max value then we need to bail since we can't be
4947 * sure we won't do bad things.
4948 * If reg->umax_value + off could overflow, treat that as unbounded too.
4950 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
4951 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
4955 err = __check_mem_access(env, regno, reg->umax_value + off, size,
4956 mem_size, zero_size_allowed);
4958 verbose(env, "R%d max value is outside of the allowed memory range\n",
4966 static int __check_ptr_off_reg(struct bpf_verifier_env *env,
4967 const struct bpf_reg_state *reg, int regno,
4970 /* Access to this pointer-typed register or passing it to a helper
4971 * is only allowed in its original, unmodified form.
4975 verbose(env, "negative offset %s ptr R%d off=%d disallowed\n",
4976 reg_type_str(env, reg->type), regno, reg->off);
4980 if (!fixed_off_ok && reg->off) {
4981 verbose(env, "dereference of modified %s ptr R%d off=%d disallowed\n",
4982 reg_type_str(env, reg->type), regno, reg->off);
4986 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4989 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4990 verbose(env, "variable %s access var_off=%s disallowed\n",
4991 reg_type_str(env, reg->type), tn_buf);
4998 int check_ptr_off_reg(struct bpf_verifier_env *env,
4999 const struct bpf_reg_state *reg, int regno)
5001 return __check_ptr_off_reg(env, reg, regno, false);
5004 static int map_kptr_match_type(struct bpf_verifier_env *env,
5005 struct btf_field *kptr_field,
5006 struct bpf_reg_state *reg, u32 regno)
5008 const char *targ_name = btf_type_name(kptr_field->kptr.btf, kptr_field->kptr.btf_id);
5010 const char *reg_name = "";
5012 if (btf_is_kernel(reg->btf)) {
5013 perm_flags = PTR_MAYBE_NULL | PTR_TRUSTED | MEM_RCU;
5015 /* Only unreferenced case accepts untrusted pointers */
5016 if (kptr_field->type == BPF_KPTR_UNREF)
5017 perm_flags |= PTR_UNTRUSTED;
5019 perm_flags = PTR_MAYBE_NULL | MEM_ALLOC;
5022 if (base_type(reg->type) != PTR_TO_BTF_ID || (type_flag(reg->type) & ~perm_flags))
5025 /* We need to verify reg->type and reg->btf, before accessing reg->btf */
5026 reg_name = btf_type_name(reg->btf, reg->btf_id);
5028 /* For ref_ptr case, release function check should ensure we get one
5029 * referenced PTR_TO_BTF_ID, and that its fixed offset is 0. For the
5030 * normal store of unreferenced kptr, we must ensure var_off is zero.
5031 * Since ref_ptr cannot be accessed directly by BPF insns, checks for
5032 * reg->off and reg->ref_obj_id are not needed here.
5034 if (__check_ptr_off_reg(env, reg, regno, true))
5037 /* A full type match is needed, as BTF can be vmlinux, module or prog BTF, and
5038 * we also need to take into account the reg->off.
5040 * We want to support cases like:
5048 * v = func(); // PTR_TO_BTF_ID
5049 * val->foo = v; // reg->off is zero, btf and btf_id match type
5050 * val->bar = &v->br; // reg->off is still zero, but we need to retry with
5051 * // first member type of struct after comparison fails
5052 * val->baz = &v->bz; // reg->off is non-zero, so struct needs to be walked
5055 * In the kptr_ref case, check_func_arg_reg_off already ensures reg->off
5056 * is zero. We must also ensure that btf_struct_ids_match does not walk
5057 * the struct to match type against first member of struct, i.e. reject
5058 * second case from above. Hence, when type is BPF_KPTR_REF, we set
5059 * strict mode to true for type match.
5061 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
5062 kptr_field->kptr.btf, kptr_field->kptr.btf_id,
5063 kptr_field->type == BPF_KPTR_REF))
5067 verbose(env, "invalid kptr access, R%d type=%s%s ", regno,
5068 reg_type_str(env, reg->type), reg_name);
5069 verbose(env, "expected=%s%s", reg_type_str(env, PTR_TO_BTF_ID), targ_name);
5070 if (kptr_field->type == BPF_KPTR_UNREF)
5071 verbose(env, " or %s%s\n", reg_type_str(env, PTR_TO_BTF_ID | PTR_UNTRUSTED),
5078 /* The non-sleepable programs and sleepable programs with explicit bpf_rcu_read_lock()
5079 * can dereference RCU protected pointers and result is PTR_TRUSTED.
5081 static bool in_rcu_cs(struct bpf_verifier_env *env)
5083 return env->cur_state->active_rcu_lock ||
5084 env->cur_state->active_lock.ptr ||
5085 !env->prog->aux->sleepable;
5088 /* Once GCC supports btf_type_tag the following mechanism will be replaced with tag check */
5089 BTF_SET_START(rcu_protected_types)
5090 BTF_ID(struct, prog_test_ref_kfunc)
5091 BTF_ID(struct, cgroup)
5092 BTF_ID(struct, bpf_cpumask)
5093 BTF_ID(struct, task_struct)
5094 BTF_SET_END(rcu_protected_types)
5096 static bool rcu_protected_object(const struct btf *btf, u32 btf_id)
5098 if (!btf_is_kernel(btf))
5100 return btf_id_set_contains(&rcu_protected_types, btf_id);
5103 static bool rcu_safe_kptr(const struct btf_field *field)
5105 const struct btf_field_kptr *kptr = &field->kptr;
5107 return field->type == BPF_KPTR_REF && rcu_protected_object(kptr->btf, kptr->btf_id);
5110 static int check_map_kptr_access(struct bpf_verifier_env *env, u32 regno,
5111 int value_regno, int insn_idx,
5112 struct btf_field *kptr_field)
5114 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
5115 int class = BPF_CLASS(insn->code);
5116 struct bpf_reg_state *val_reg;
5118 /* Things we already checked for in check_map_access and caller:
5119 * - Reject cases where variable offset may touch kptr
5120 * - size of access (must be BPF_DW)
5121 * - tnum_is_const(reg->var_off)
5122 * - kptr_field->offset == off + reg->var_off.value
5124 /* Only BPF_[LDX,STX,ST] | BPF_MEM | BPF_DW is supported */
5125 if (BPF_MODE(insn->code) != BPF_MEM) {
5126 verbose(env, "kptr in map can only be accessed using BPF_MEM instruction mode\n");
5130 /* We only allow loading referenced kptr, since it will be marked as
5131 * untrusted, similar to unreferenced kptr.
5133 if (class != BPF_LDX && kptr_field->type == BPF_KPTR_REF) {
5134 verbose(env, "store to referenced kptr disallowed\n");
5138 if (class == BPF_LDX) {
5139 val_reg = reg_state(env, value_regno);
5140 /* We can simply mark the value_regno receiving the pointer
5141 * value from map as PTR_TO_BTF_ID, with the correct type.
5143 mark_btf_ld_reg(env, cur_regs(env), value_regno, PTR_TO_BTF_ID, kptr_field->kptr.btf,
5144 kptr_field->kptr.btf_id,
5145 rcu_safe_kptr(kptr_field) && in_rcu_cs(env) ?
5146 PTR_MAYBE_NULL | MEM_RCU :
5147 PTR_MAYBE_NULL | PTR_UNTRUSTED);
5148 /* For mark_ptr_or_null_reg */
5149 val_reg->id = ++env->id_gen;
5150 } else if (class == BPF_STX) {
5151 val_reg = reg_state(env, value_regno);
5152 if (!register_is_null(val_reg) &&
5153 map_kptr_match_type(env, kptr_field, val_reg, value_regno))
5155 } else if (class == BPF_ST) {
5157 verbose(env, "BPF_ST imm must be 0 when storing to kptr at off=%u\n",
5158 kptr_field->offset);
5162 verbose(env, "kptr in map can only be accessed using BPF_LDX/BPF_STX/BPF_ST\n");
5168 /* check read/write into a map element with possible variable offset */
5169 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
5170 int off, int size, bool zero_size_allowed,
5171 enum bpf_access_src src)
5173 struct bpf_verifier_state *vstate = env->cur_state;
5174 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5175 struct bpf_reg_state *reg = &state->regs[regno];
5176 struct bpf_map *map = reg->map_ptr;
5177 struct btf_record *rec;
5180 err = check_mem_region_access(env, regno, off, size, map->value_size,
5185 if (IS_ERR_OR_NULL(map->record))
5188 for (i = 0; i < rec->cnt; i++) {
5189 struct btf_field *field = &rec->fields[i];
5190 u32 p = field->offset;
5192 /* If any part of a field can be touched by load/store, reject
5193 * this program. To check that [x1, x2) overlaps with [y1, y2),
5194 * it is sufficient to check x1 < y2 && y1 < x2.
5196 if (reg->smin_value + off < p + btf_field_type_size(field->type) &&
5197 p < reg->umax_value + off + size) {
5198 switch (field->type) {
5199 case BPF_KPTR_UNREF:
5201 if (src != ACCESS_DIRECT) {
5202 verbose(env, "kptr cannot be accessed indirectly by helper\n");
5205 if (!tnum_is_const(reg->var_off)) {
5206 verbose(env, "kptr access cannot have variable offset\n");
5209 if (p != off + reg->var_off.value) {
5210 verbose(env, "kptr access misaligned expected=%u off=%llu\n",
5211 p, off + reg->var_off.value);
5214 if (size != bpf_size_to_bytes(BPF_DW)) {
5215 verbose(env, "kptr access size must be BPF_DW\n");
5220 verbose(env, "%s cannot be accessed directly by load/store\n",
5221 btf_field_type_name(field->type));
5229 #define MAX_PACKET_OFF 0xffff
5231 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
5232 const struct bpf_call_arg_meta *meta,
5233 enum bpf_access_type t)
5235 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
5237 switch (prog_type) {
5238 /* Program types only with direct read access go here! */
5239 case BPF_PROG_TYPE_LWT_IN:
5240 case BPF_PROG_TYPE_LWT_OUT:
5241 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
5242 case BPF_PROG_TYPE_SK_REUSEPORT:
5243 case BPF_PROG_TYPE_FLOW_DISSECTOR:
5244 case BPF_PROG_TYPE_CGROUP_SKB:
5249 /* Program types with direct read + write access go here! */
5250 case BPF_PROG_TYPE_SCHED_CLS:
5251 case BPF_PROG_TYPE_SCHED_ACT:
5252 case BPF_PROG_TYPE_XDP:
5253 case BPF_PROG_TYPE_LWT_XMIT:
5254 case BPF_PROG_TYPE_SK_SKB:
5255 case BPF_PROG_TYPE_SK_MSG:
5257 return meta->pkt_access;
5259 env->seen_direct_write = true;
5262 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
5264 env->seen_direct_write = true;
5273 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
5274 int size, bool zero_size_allowed)
5276 struct bpf_reg_state *regs = cur_regs(env);
5277 struct bpf_reg_state *reg = ®s[regno];
5280 /* We may have added a variable offset to the packet pointer; but any
5281 * reg->range we have comes after that. We are only checking the fixed
5285 /* We don't allow negative numbers, because we aren't tracking enough
5286 * detail to prove they're safe.
5288 if (reg->smin_value < 0) {
5289 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
5294 err = reg->range < 0 ? -EINVAL :
5295 __check_mem_access(env, regno, off, size, reg->range,
5298 verbose(env, "R%d offset is outside of the packet\n", regno);
5302 /* __check_mem_access has made sure "off + size - 1" is within u16.
5303 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
5304 * otherwise find_good_pkt_pointers would have refused to set range info
5305 * that __check_mem_access would have rejected this pkt access.
5306 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
5308 env->prog->aux->max_pkt_offset =
5309 max_t(u32, env->prog->aux->max_pkt_offset,
5310 off + reg->umax_value + size - 1);
5315 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
5316 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
5317 enum bpf_access_type t, enum bpf_reg_type *reg_type,
5318 struct btf **btf, u32 *btf_id)
5320 struct bpf_insn_access_aux info = {
5321 .reg_type = *reg_type,
5325 if (env->ops->is_valid_access &&
5326 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
5327 /* A non zero info.ctx_field_size indicates that this field is a
5328 * candidate for later verifier transformation to load the whole
5329 * field and then apply a mask when accessed with a narrower
5330 * access than actual ctx access size. A zero info.ctx_field_size
5331 * will only allow for whole field access and rejects any other
5332 * type of narrower access.
5334 *reg_type = info.reg_type;
5336 if (base_type(*reg_type) == PTR_TO_BTF_ID) {
5338 *btf_id = info.btf_id;
5340 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
5342 /* remember the offset of last byte accessed in ctx */
5343 if (env->prog->aux->max_ctx_offset < off + size)
5344 env->prog->aux->max_ctx_offset = off + size;
5348 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
5352 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
5355 if (size < 0 || off < 0 ||
5356 (u64)off + size > sizeof(struct bpf_flow_keys)) {
5357 verbose(env, "invalid access to flow keys off=%d size=%d\n",
5364 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
5365 u32 regno, int off, int size,
5366 enum bpf_access_type t)
5368 struct bpf_reg_state *regs = cur_regs(env);
5369 struct bpf_reg_state *reg = ®s[regno];
5370 struct bpf_insn_access_aux info = {};
5373 if (reg->smin_value < 0) {
5374 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
5379 switch (reg->type) {
5380 case PTR_TO_SOCK_COMMON:
5381 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
5384 valid = bpf_sock_is_valid_access(off, size, t, &info);
5386 case PTR_TO_TCP_SOCK:
5387 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
5389 case PTR_TO_XDP_SOCK:
5390 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
5398 env->insn_aux_data[insn_idx].ctx_field_size =
5399 info.ctx_field_size;
5403 verbose(env, "R%d invalid %s access off=%d size=%d\n",
5404 regno, reg_type_str(env, reg->type), off, size);
5409 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
5411 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
5414 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
5416 const struct bpf_reg_state *reg = reg_state(env, regno);
5418 return reg->type == PTR_TO_CTX;
5421 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
5423 const struct bpf_reg_state *reg = reg_state(env, regno);
5425 return type_is_sk_pointer(reg->type);
5428 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
5430 const struct bpf_reg_state *reg = reg_state(env, regno);
5432 return type_is_pkt_pointer(reg->type);
5435 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
5437 const struct bpf_reg_state *reg = reg_state(env, regno);
5439 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
5440 return reg->type == PTR_TO_FLOW_KEYS;
5443 static u32 *reg2btf_ids[__BPF_REG_TYPE_MAX] = {
5445 [PTR_TO_SOCKET] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK],
5446 [PTR_TO_SOCK_COMMON] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
5447 [PTR_TO_TCP_SOCK] = &btf_sock_ids[BTF_SOCK_TYPE_TCP],
5449 [CONST_PTR_TO_MAP] = btf_bpf_map_id,
5452 static bool is_trusted_reg(const struct bpf_reg_state *reg)
5454 /* A referenced register is always trusted. */
5455 if (reg->ref_obj_id)
5458 /* Types listed in the reg2btf_ids are always trusted */
5459 if (reg2btf_ids[base_type(reg->type)])
5462 /* If a register is not referenced, it is trusted if it has the
5463 * MEM_ALLOC or PTR_TRUSTED type modifiers, and no others. Some of the
5464 * other type modifiers may be safe, but we elect to take an opt-in
5465 * approach here as some (e.g. PTR_UNTRUSTED and PTR_MAYBE_NULL) are
5468 * Eventually, we should make PTR_TRUSTED the single source of truth
5469 * for whether a register is trusted.
5471 return type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS &&
5472 !bpf_type_has_unsafe_modifiers(reg->type);
5475 static bool is_rcu_reg(const struct bpf_reg_state *reg)
5477 return reg->type & MEM_RCU;
5480 static void clear_trusted_flags(enum bpf_type_flag *flag)
5482 *flag &= ~(BPF_REG_TRUSTED_MODIFIERS | MEM_RCU);
5485 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
5486 const struct bpf_reg_state *reg,
5487 int off, int size, bool strict)
5489 struct tnum reg_off;
5492 /* Byte size accesses are always allowed. */
5493 if (!strict || size == 1)
5496 /* For platforms that do not have a Kconfig enabling
5497 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
5498 * NET_IP_ALIGN is universally set to '2'. And on platforms
5499 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
5500 * to this code only in strict mode where we want to emulate
5501 * the NET_IP_ALIGN==2 checking. Therefore use an
5502 * unconditional IP align value of '2'.
5506 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
5507 if (!tnum_is_aligned(reg_off, size)) {
5510 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5512 "misaligned packet access off %d+%s+%d+%d size %d\n",
5513 ip_align, tn_buf, reg->off, off, size);
5520 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
5521 const struct bpf_reg_state *reg,
5522 const char *pointer_desc,
5523 int off, int size, bool strict)
5525 struct tnum reg_off;
5527 /* Byte size accesses are always allowed. */
5528 if (!strict || size == 1)
5531 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
5532 if (!tnum_is_aligned(reg_off, size)) {
5535 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5536 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
5537 pointer_desc, tn_buf, reg->off, off, size);
5544 static int check_ptr_alignment(struct bpf_verifier_env *env,
5545 const struct bpf_reg_state *reg, int off,
5546 int size, bool strict_alignment_once)
5548 bool strict = env->strict_alignment || strict_alignment_once;
5549 const char *pointer_desc = "";
5551 switch (reg->type) {
5553 case PTR_TO_PACKET_META:
5554 /* Special case, because of NET_IP_ALIGN. Given metadata sits
5555 * right in front, treat it the very same way.
5557 return check_pkt_ptr_alignment(env, reg, off, size, strict);
5558 case PTR_TO_FLOW_KEYS:
5559 pointer_desc = "flow keys ";
5561 case PTR_TO_MAP_KEY:
5562 pointer_desc = "key ";
5564 case PTR_TO_MAP_VALUE:
5565 pointer_desc = "value ";
5568 pointer_desc = "context ";
5571 pointer_desc = "stack ";
5572 /* The stack spill tracking logic in check_stack_write_fixed_off()
5573 * and check_stack_read_fixed_off() relies on stack accesses being
5579 pointer_desc = "sock ";
5581 case PTR_TO_SOCK_COMMON:
5582 pointer_desc = "sock_common ";
5584 case PTR_TO_TCP_SOCK:
5585 pointer_desc = "tcp_sock ";
5587 case PTR_TO_XDP_SOCK:
5588 pointer_desc = "xdp_sock ";
5593 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
5597 static int update_stack_depth(struct bpf_verifier_env *env,
5598 const struct bpf_func_state *func,
5601 u16 stack = env->subprog_info[func->subprogno].stack_depth;
5606 /* update known max for given subprogram */
5607 env->subprog_info[func->subprogno].stack_depth = -off;
5611 /* starting from main bpf function walk all instructions of the function
5612 * and recursively walk all callees that given function can call.
5613 * Ignore jump and exit insns.
5614 * Since recursion is prevented by check_cfg() this algorithm
5615 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
5617 static int check_max_stack_depth_subprog(struct bpf_verifier_env *env, int idx)
5619 struct bpf_subprog_info *subprog = env->subprog_info;
5620 struct bpf_insn *insn = env->prog->insnsi;
5621 int depth = 0, frame = 0, i, subprog_end;
5622 bool tail_call_reachable = false;
5623 int ret_insn[MAX_CALL_FRAMES];
5624 int ret_prog[MAX_CALL_FRAMES];
5627 i = subprog[idx].start;
5629 /* protect against potential stack overflow that might happen when
5630 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
5631 * depth for such case down to 256 so that the worst case scenario
5632 * would result in 8k stack size (32 which is tailcall limit * 256 =
5635 * To get the idea what might happen, see an example:
5636 * func1 -> sub rsp, 128
5637 * subfunc1 -> sub rsp, 256
5638 * tailcall1 -> add rsp, 256
5639 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
5640 * subfunc2 -> sub rsp, 64
5641 * subfunc22 -> sub rsp, 128
5642 * tailcall2 -> add rsp, 128
5643 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
5645 * tailcall will unwind the current stack frame but it will not get rid
5646 * of caller's stack as shown on the example above.
5648 if (idx && subprog[idx].has_tail_call && depth >= 256) {
5650 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
5654 /* round up to 32-bytes, since this is granularity
5655 * of interpreter stack size
5657 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
5658 if (depth > MAX_BPF_STACK) {
5659 verbose(env, "combined stack size of %d calls is %d. Too large\n",
5664 subprog_end = subprog[idx + 1].start;
5665 for (; i < subprog_end; i++) {
5666 int next_insn, sidx;
5668 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
5670 /* remember insn and function to return to */
5671 ret_insn[frame] = i + 1;
5672 ret_prog[frame] = idx;
5674 /* find the callee */
5675 next_insn = i + insn[i].imm + 1;
5676 sidx = find_subprog(env, next_insn);
5678 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5682 if (subprog[sidx].is_async_cb) {
5683 if (subprog[sidx].has_tail_call) {
5684 verbose(env, "verifier bug. subprog has tail_call and async cb\n");
5687 /* async callbacks don't increase bpf prog stack size unless called directly */
5688 if (!bpf_pseudo_call(insn + i))
5694 if (subprog[idx].has_tail_call)
5695 tail_call_reachable = true;
5698 if (frame >= MAX_CALL_FRAMES) {
5699 verbose(env, "the call stack of %d frames is too deep !\n",
5705 /* if tail call got detected across bpf2bpf calls then mark each of the
5706 * currently present subprog frames as tail call reachable subprogs;
5707 * this info will be utilized by JIT so that we will be preserving the
5708 * tail call counter throughout bpf2bpf calls combined with tailcalls
5710 if (tail_call_reachable)
5711 for (j = 0; j < frame; j++)
5712 subprog[ret_prog[j]].tail_call_reachable = true;
5713 if (subprog[0].tail_call_reachable)
5714 env->prog->aux->tail_call_reachable = true;
5716 /* end of for() loop means the last insn of the 'subprog'
5717 * was reached. Doesn't matter whether it was JA or EXIT
5721 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
5723 i = ret_insn[frame];
5724 idx = ret_prog[frame];
5728 static int check_max_stack_depth(struct bpf_verifier_env *env)
5730 struct bpf_subprog_info *si = env->subprog_info;
5733 for (int i = 0; i < env->subprog_cnt; i++) {
5734 if (!i || si[i].is_async_cb) {
5735 ret = check_max_stack_depth_subprog(env, i);
5744 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
5745 static int get_callee_stack_depth(struct bpf_verifier_env *env,
5746 const struct bpf_insn *insn, int idx)
5748 int start = idx + insn->imm + 1, subprog;
5750 subprog = find_subprog(env, start);
5752 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5756 return env->subprog_info[subprog].stack_depth;
5760 static int __check_buffer_access(struct bpf_verifier_env *env,
5761 const char *buf_info,
5762 const struct bpf_reg_state *reg,
5763 int regno, int off, int size)
5767 "R%d invalid %s buffer access: off=%d, size=%d\n",
5768 regno, buf_info, off, size);
5771 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
5774 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5776 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
5777 regno, off, tn_buf);
5784 static int check_tp_buffer_access(struct bpf_verifier_env *env,
5785 const struct bpf_reg_state *reg,
5786 int regno, int off, int size)
5790 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
5794 if (off + size > env->prog->aux->max_tp_access)
5795 env->prog->aux->max_tp_access = off + size;
5800 static int check_buffer_access(struct bpf_verifier_env *env,
5801 const struct bpf_reg_state *reg,
5802 int regno, int off, int size,
5803 bool zero_size_allowed,
5806 const char *buf_info = type_is_rdonly_mem(reg->type) ? "rdonly" : "rdwr";
5809 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
5813 if (off + size > *max_access)
5814 *max_access = off + size;
5819 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
5820 static void zext_32_to_64(struct bpf_reg_state *reg)
5822 reg->var_off = tnum_subreg(reg->var_off);
5823 __reg_assign_32_into_64(reg);
5826 /* truncate register to smaller size (in bytes)
5827 * must be called with size < BPF_REG_SIZE
5829 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
5833 /* clear high bits in bit representation */
5834 reg->var_off = tnum_cast(reg->var_off, size);
5836 /* fix arithmetic bounds */
5837 mask = ((u64)1 << (size * 8)) - 1;
5838 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
5839 reg->umin_value &= mask;
5840 reg->umax_value &= mask;
5842 reg->umin_value = 0;
5843 reg->umax_value = mask;
5845 reg->smin_value = reg->umin_value;
5846 reg->smax_value = reg->umax_value;
5848 /* If size is smaller than 32bit register the 32bit register
5849 * values are also truncated so we push 64-bit bounds into
5850 * 32-bit bounds. Above were truncated < 32-bits already.
5854 __reg_combine_64_into_32(reg);
5857 static void set_sext64_default_val(struct bpf_reg_state *reg, int size)
5860 reg->smin_value = reg->s32_min_value = S8_MIN;
5861 reg->smax_value = reg->s32_max_value = S8_MAX;
5862 } else if (size == 2) {
5863 reg->smin_value = reg->s32_min_value = S16_MIN;
5864 reg->smax_value = reg->s32_max_value = S16_MAX;
5867 reg->smin_value = reg->s32_min_value = S32_MIN;
5868 reg->smax_value = reg->s32_max_value = S32_MAX;
5870 reg->umin_value = reg->u32_min_value = 0;
5871 reg->umax_value = U64_MAX;
5872 reg->u32_max_value = U32_MAX;
5873 reg->var_off = tnum_unknown;
5876 static void coerce_reg_to_size_sx(struct bpf_reg_state *reg, int size)
5878 s64 init_s64_max, init_s64_min, s64_max, s64_min, u64_cval;
5879 u64 top_smax_value, top_smin_value;
5880 u64 num_bits = size * 8;
5882 if (tnum_is_const(reg->var_off)) {
5883 u64_cval = reg->var_off.value;
5885 reg->var_off = tnum_const((s8)u64_cval);
5887 reg->var_off = tnum_const((s16)u64_cval);
5890 reg->var_off = tnum_const((s32)u64_cval);
5892 u64_cval = reg->var_off.value;
5893 reg->smax_value = reg->smin_value = u64_cval;
5894 reg->umax_value = reg->umin_value = u64_cval;
5895 reg->s32_max_value = reg->s32_min_value = u64_cval;
5896 reg->u32_max_value = reg->u32_min_value = u64_cval;
5900 top_smax_value = ((u64)reg->smax_value >> num_bits) << num_bits;
5901 top_smin_value = ((u64)reg->smin_value >> num_bits) << num_bits;
5903 if (top_smax_value != top_smin_value)
5906 /* find the s64_min and s64_min after sign extension */
5908 init_s64_max = (s8)reg->smax_value;
5909 init_s64_min = (s8)reg->smin_value;
5910 } else if (size == 2) {
5911 init_s64_max = (s16)reg->smax_value;
5912 init_s64_min = (s16)reg->smin_value;
5914 init_s64_max = (s32)reg->smax_value;
5915 init_s64_min = (s32)reg->smin_value;
5918 s64_max = max(init_s64_max, init_s64_min);
5919 s64_min = min(init_s64_max, init_s64_min);
5921 /* both of s64_max/s64_min positive or negative */
5922 if ((s64_max >= 0) == (s64_min >= 0)) {
5923 reg->smin_value = reg->s32_min_value = s64_min;
5924 reg->smax_value = reg->s32_max_value = s64_max;
5925 reg->umin_value = reg->u32_min_value = s64_min;
5926 reg->umax_value = reg->u32_max_value = s64_max;
5927 reg->var_off = tnum_range(s64_min, s64_max);
5932 set_sext64_default_val(reg, size);
5935 static void set_sext32_default_val(struct bpf_reg_state *reg, int size)
5938 reg->s32_min_value = S8_MIN;
5939 reg->s32_max_value = S8_MAX;
5942 reg->s32_min_value = S16_MIN;
5943 reg->s32_max_value = S16_MAX;
5945 reg->u32_min_value = 0;
5946 reg->u32_max_value = U32_MAX;
5949 static void coerce_subreg_to_size_sx(struct bpf_reg_state *reg, int size)
5951 s32 init_s32_max, init_s32_min, s32_max, s32_min, u32_val;
5952 u32 top_smax_value, top_smin_value;
5953 u32 num_bits = size * 8;
5955 if (tnum_is_const(reg->var_off)) {
5956 u32_val = reg->var_off.value;
5958 reg->var_off = tnum_const((s8)u32_val);
5960 reg->var_off = tnum_const((s16)u32_val);
5962 u32_val = reg->var_off.value;
5963 reg->s32_min_value = reg->s32_max_value = u32_val;
5964 reg->u32_min_value = reg->u32_max_value = u32_val;
5968 top_smax_value = ((u32)reg->s32_max_value >> num_bits) << num_bits;
5969 top_smin_value = ((u32)reg->s32_min_value >> num_bits) << num_bits;
5971 if (top_smax_value != top_smin_value)
5974 /* find the s32_min and s32_min after sign extension */
5976 init_s32_max = (s8)reg->s32_max_value;
5977 init_s32_min = (s8)reg->s32_min_value;
5980 init_s32_max = (s16)reg->s32_max_value;
5981 init_s32_min = (s16)reg->s32_min_value;
5983 s32_max = max(init_s32_max, init_s32_min);
5984 s32_min = min(init_s32_max, init_s32_min);
5986 if ((s32_min >= 0) == (s32_max >= 0)) {
5987 reg->s32_min_value = s32_min;
5988 reg->s32_max_value = s32_max;
5989 reg->u32_min_value = (u32)s32_min;
5990 reg->u32_max_value = (u32)s32_max;
5995 set_sext32_default_val(reg, size);
5998 static bool bpf_map_is_rdonly(const struct bpf_map *map)
6000 /* A map is considered read-only if the following condition are true:
6002 * 1) BPF program side cannot change any of the map content. The
6003 * BPF_F_RDONLY_PROG flag is throughout the lifetime of a map
6004 * and was set at map creation time.
6005 * 2) The map value(s) have been initialized from user space by a
6006 * loader and then "frozen", such that no new map update/delete
6007 * operations from syscall side are possible for the rest of
6008 * the map's lifetime from that point onwards.
6009 * 3) Any parallel/pending map update/delete operations from syscall
6010 * side have been completed. Only after that point, it's safe to
6011 * assume that map value(s) are immutable.
6013 return (map->map_flags & BPF_F_RDONLY_PROG) &&
6014 READ_ONCE(map->frozen) &&
6015 !bpf_map_write_active(map);
6018 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val,
6025 err = map->ops->map_direct_value_addr(map, &addr, off);
6028 ptr = (void *)(long)addr + off;
6032 *val = is_ldsx ? (s64)*(s8 *)ptr : (u64)*(u8 *)ptr;
6035 *val = is_ldsx ? (s64)*(s16 *)ptr : (u64)*(u16 *)ptr;
6038 *val = is_ldsx ? (s64)*(s32 *)ptr : (u64)*(u32 *)ptr;
6049 #define BTF_TYPE_SAFE_RCU(__type) __PASTE(__type, __safe_rcu)
6050 #define BTF_TYPE_SAFE_RCU_OR_NULL(__type) __PASTE(__type, __safe_rcu_or_null)
6051 #define BTF_TYPE_SAFE_TRUSTED(__type) __PASTE(__type, __safe_trusted)
6054 * Allow list few fields as RCU trusted or full trusted.
6055 * This logic doesn't allow mix tagging and will be removed once GCC supports
6059 /* RCU trusted: these fields are trusted in RCU CS and never NULL */
6060 BTF_TYPE_SAFE_RCU(struct task_struct) {
6061 const cpumask_t *cpus_ptr;
6062 struct css_set __rcu *cgroups;
6063 struct task_struct __rcu *real_parent;
6064 struct task_struct *group_leader;
6067 BTF_TYPE_SAFE_RCU(struct cgroup) {
6068 /* cgrp->kn is always accessible as documented in kernel/cgroup/cgroup.c */
6069 struct kernfs_node *kn;
6072 BTF_TYPE_SAFE_RCU(struct css_set) {
6073 struct cgroup *dfl_cgrp;
6076 /* RCU trusted: these fields are trusted in RCU CS and can be NULL */
6077 BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct) {
6078 struct file __rcu *exe_file;
6081 /* skb->sk, req->sk are not RCU protected, but we mark them as such
6082 * because bpf prog accessible sockets are SOCK_RCU_FREE.
6084 BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff) {
6088 BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock) {
6092 /* full trusted: these fields are trusted even outside of RCU CS and never NULL */
6093 BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta) {
6094 struct seq_file *seq;
6097 BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task) {
6098 struct bpf_iter_meta *meta;
6099 struct task_struct *task;
6102 BTF_TYPE_SAFE_TRUSTED(struct linux_binprm) {
6106 BTF_TYPE_SAFE_TRUSTED(struct file) {
6107 struct inode *f_inode;
6110 BTF_TYPE_SAFE_TRUSTED(struct dentry) {
6111 /* no negative dentry-s in places where bpf can see it */
6112 struct inode *d_inode;
6115 BTF_TYPE_SAFE_TRUSTED(struct socket) {
6119 static bool type_is_rcu(struct bpf_verifier_env *env,
6120 struct bpf_reg_state *reg,
6121 const char *field_name, u32 btf_id)
6123 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct task_struct));
6124 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct cgroup));
6125 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct css_set));
6127 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_rcu");
6130 static bool type_is_rcu_or_null(struct bpf_verifier_env *env,
6131 struct bpf_reg_state *reg,
6132 const char *field_name, u32 btf_id)
6134 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct));
6135 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff));
6136 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock));
6138 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_rcu_or_null");
6141 static bool type_is_trusted(struct bpf_verifier_env *env,
6142 struct bpf_reg_state *reg,
6143 const char *field_name, u32 btf_id)
6145 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta));
6146 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task));
6147 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct linux_binprm));
6148 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct file));
6149 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct dentry));
6150 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct socket));
6152 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_trusted");
6155 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
6156 struct bpf_reg_state *regs,
6157 int regno, int off, int size,
6158 enum bpf_access_type atype,
6161 struct bpf_reg_state *reg = regs + regno;
6162 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
6163 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
6164 const char *field_name = NULL;
6165 enum bpf_type_flag flag = 0;
6169 if (!env->allow_ptr_leaks) {
6171 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
6175 if (!env->prog->gpl_compatible && btf_is_kernel(reg->btf)) {
6177 "Cannot access kernel 'struct %s' from non-GPL compatible program\n",
6183 "R%d is ptr_%s invalid negative access: off=%d\n",
6187 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
6190 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6192 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
6193 regno, tname, off, tn_buf);
6197 if (reg->type & MEM_USER) {
6199 "R%d is ptr_%s access user memory: off=%d\n",
6204 if (reg->type & MEM_PERCPU) {
6206 "R%d is ptr_%s access percpu memory: off=%d\n",
6211 if (env->ops->btf_struct_access && !type_is_alloc(reg->type) && atype == BPF_WRITE) {
6212 if (!btf_is_kernel(reg->btf)) {
6213 verbose(env, "verifier internal error: reg->btf must be kernel btf\n");
6216 ret = env->ops->btf_struct_access(&env->log, reg, off, size);
6218 /* Writes are permitted with default btf_struct_access for
6219 * program allocated objects (which always have ref_obj_id > 0),
6220 * but not for untrusted PTR_TO_BTF_ID | MEM_ALLOC.
6222 if (atype != BPF_READ && !type_is_ptr_alloc_obj(reg->type)) {
6223 verbose(env, "only read is supported\n");
6227 if (type_is_alloc(reg->type) && !type_is_non_owning_ref(reg->type) &&
6229 verbose(env, "verifier internal error: ref_obj_id for allocated object must be non-zero\n");
6233 ret = btf_struct_access(&env->log, reg, off, size, atype, &btf_id, &flag, &field_name);
6239 if (ret != PTR_TO_BTF_ID) {
6242 } else if (type_flag(reg->type) & PTR_UNTRUSTED) {
6243 /* If this is an untrusted pointer, all pointers formed by walking it
6244 * also inherit the untrusted flag.
6246 flag = PTR_UNTRUSTED;
6248 } else if (is_trusted_reg(reg) || is_rcu_reg(reg)) {
6249 /* By default any pointer obtained from walking a trusted pointer is no
6250 * longer trusted, unless the field being accessed has explicitly been
6251 * marked as inheriting its parent's state of trust (either full or RCU).
6253 * 'cgroups' pointer is untrusted if task->cgroups dereference
6254 * happened in a sleepable program outside of bpf_rcu_read_lock()
6255 * section. In a non-sleepable program it's trusted while in RCU CS (aka MEM_RCU).
6256 * Note bpf_rcu_read_unlock() converts MEM_RCU pointers to PTR_UNTRUSTED.
6258 * A regular RCU-protected pointer with __rcu tag can also be deemed
6259 * trusted if we are in an RCU CS. Such pointer can be NULL.
6261 if (type_is_trusted(env, reg, field_name, btf_id)) {
6262 flag |= PTR_TRUSTED;
6263 } else if (in_rcu_cs(env) && !type_may_be_null(reg->type)) {
6264 if (type_is_rcu(env, reg, field_name, btf_id)) {
6265 /* ignore __rcu tag and mark it MEM_RCU */
6267 } else if (flag & MEM_RCU ||
6268 type_is_rcu_or_null(env, reg, field_name, btf_id)) {
6269 /* __rcu tagged pointers can be NULL */
6270 flag |= MEM_RCU | PTR_MAYBE_NULL;
6272 /* We always trust them */
6273 if (type_is_rcu_or_null(env, reg, field_name, btf_id) &&
6274 flag & PTR_UNTRUSTED)
6275 flag &= ~PTR_UNTRUSTED;
6276 } else if (flag & (MEM_PERCPU | MEM_USER)) {
6279 /* walking unknown pointers yields old deprecated PTR_TO_BTF_ID */
6280 clear_trusted_flags(&flag);
6284 * If not in RCU CS or MEM_RCU pointer can be NULL then
6285 * aggressively mark as untrusted otherwise such
6286 * pointers will be plain PTR_TO_BTF_ID without flags
6287 * and will be allowed to be passed into helpers for
6290 flag = PTR_UNTRUSTED;
6293 /* Old compat. Deprecated */
6294 clear_trusted_flags(&flag);
6297 if (atype == BPF_READ && value_regno >= 0)
6298 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id, flag);
6303 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
6304 struct bpf_reg_state *regs,
6305 int regno, int off, int size,
6306 enum bpf_access_type atype,
6309 struct bpf_reg_state *reg = regs + regno;
6310 struct bpf_map *map = reg->map_ptr;
6311 struct bpf_reg_state map_reg;
6312 enum bpf_type_flag flag = 0;
6313 const struct btf_type *t;
6319 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
6323 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
6324 verbose(env, "map_ptr access not supported for map type %d\n",
6329 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
6330 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
6332 if (!env->allow_ptr_leaks) {
6334 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
6340 verbose(env, "R%d is %s invalid negative access: off=%d\n",
6345 if (atype != BPF_READ) {
6346 verbose(env, "only read from %s is supported\n", tname);
6350 /* Simulate access to a PTR_TO_BTF_ID */
6351 memset(&map_reg, 0, sizeof(map_reg));
6352 mark_btf_ld_reg(env, &map_reg, 0, PTR_TO_BTF_ID, btf_vmlinux, *map->ops->map_btf_id, 0);
6353 ret = btf_struct_access(&env->log, &map_reg, off, size, atype, &btf_id, &flag, NULL);
6357 if (value_regno >= 0)
6358 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id, flag);
6363 /* Check that the stack access at the given offset is within bounds. The
6364 * maximum valid offset is -1.
6366 * The minimum valid offset is -MAX_BPF_STACK for writes, and
6367 * -state->allocated_stack for reads.
6369 static int check_stack_slot_within_bounds(int off,
6370 struct bpf_func_state *state,
6371 enum bpf_access_type t)
6376 min_valid_off = -MAX_BPF_STACK;
6378 min_valid_off = -state->allocated_stack;
6380 if (off < min_valid_off || off > -1)
6385 /* Check that the stack access at 'regno + off' falls within the maximum stack
6388 * 'off' includes `regno->offset`, but not its dynamic part (if any).
6390 static int check_stack_access_within_bounds(
6391 struct bpf_verifier_env *env,
6392 int regno, int off, int access_size,
6393 enum bpf_access_src src, enum bpf_access_type type)
6395 struct bpf_reg_state *regs = cur_regs(env);
6396 struct bpf_reg_state *reg = regs + regno;
6397 struct bpf_func_state *state = func(env, reg);
6398 int min_off, max_off;
6402 if (src == ACCESS_HELPER)
6403 /* We don't know if helpers are reading or writing (or both). */
6404 err_extra = " indirect access to";
6405 else if (type == BPF_READ)
6406 err_extra = " read from";
6408 err_extra = " write to";
6410 if (tnum_is_const(reg->var_off)) {
6411 min_off = reg->var_off.value + off;
6412 if (access_size > 0)
6413 max_off = min_off + access_size - 1;
6417 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
6418 reg->smin_value <= -BPF_MAX_VAR_OFF) {
6419 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
6423 min_off = reg->smin_value + off;
6424 if (access_size > 0)
6425 max_off = reg->smax_value + off + access_size - 1;
6430 err = check_stack_slot_within_bounds(min_off, state, type);
6432 err = check_stack_slot_within_bounds(max_off, state, type);
6435 if (tnum_is_const(reg->var_off)) {
6436 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
6437 err_extra, regno, off, access_size);
6441 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6442 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
6443 err_extra, regno, tn_buf, access_size);
6449 /* check whether memory at (regno + off) is accessible for t = (read | write)
6450 * if t==write, value_regno is a register which value is stored into memory
6451 * if t==read, value_regno is a register which will receive the value from memory
6452 * if t==write && value_regno==-1, some unknown value is stored into memory
6453 * if t==read && value_regno==-1, don't care what we read from memory
6455 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
6456 int off, int bpf_size, enum bpf_access_type t,
6457 int value_regno, bool strict_alignment_once, bool is_ldsx)
6459 struct bpf_reg_state *regs = cur_regs(env);
6460 struct bpf_reg_state *reg = regs + regno;
6461 struct bpf_func_state *state;
6464 size = bpf_size_to_bytes(bpf_size);
6468 /* alignment checks will add in reg->off themselves */
6469 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
6473 /* for access checks, reg->off is just part of off */
6476 if (reg->type == PTR_TO_MAP_KEY) {
6477 if (t == BPF_WRITE) {
6478 verbose(env, "write to change key R%d not allowed\n", regno);
6482 err = check_mem_region_access(env, regno, off, size,
6483 reg->map_ptr->key_size, false);
6486 if (value_regno >= 0)
6487 mark_reg_unknown(env, regs, value_regno);
6488 } else if (reg->type == PTR_TO_MAP_VALUE) {
6489 struct btf_field *kptr_field = NULL;
6491 if (t == BPF_WRITE && value_regno >= 0 &&
6492 is_pointer_value(env, value_regno)) {
6493 verbose(env, "R%d leaks addr into map\n", value_regno);
6496 err = check_map_access_type(env, regno, off, size, t);
6499 err = check_map_access(env, regno, off, size, false, ACCESS_DIRECT);
6502 if (tnum_is_const(reg->var_off))
6503 kptr_field = btf_record_find(reg->map_ptr->record,
6504 off + reg->var_off.value, BPF_KPTR);
6506 err = check_map_kptr_access(env, regno, value_regno, insn_idx, kptr_field);
6507 } else if (t == BPF_READ && value_regno >= 0) {
6508 struct bpf_map *map = reg->map_ptr;
6510 /* if map is read-only, track its contents as scalars */
6511 if (tnum_is_const(reg->var_off) &&
6512 bpf_map_is_rdonly(map) &&
6513 map->ops->map_direct_value_addr) {
6514 int map_off = off + reg->var_off.value;
6517 err = bpf_map_direct_read(map, map_off, size,
6522 regs[value_regno].type = SCALAR_VALUE;
6523 __mark_reg_known(®s[value_regno], val);
6525 mark_reg_unknown(env, regs, value_regno);
6528 } else if (base_type(reg->type) == PTR_TO_MEM) {
6529 bool rdonly_mem = type_is_rdonly_mem(reg->type);
6531 if (type_may_be_null(reg->type)) {
6532 verbose(env, "R%d invalid mem access '%s'\n", regno,
6533 reg_type_str(env, reg->type));
6537 if (t == BPF_WRITE && rdonly_mem) {
6538 verbose(env, "R%d cannot write into %s\n",
6539 regno, reg_type_str(env, reg->type));
6543 if (t == BPF_WRITE && value_regno >= 0 &&
6544 is_pointer_value(env, value_regno)) {
6545 verbose(env, "R%d leaks addr into mem\n", value_regno);
6549 err = check_mem_region_access(env, regno, off, size,
6550 reg->mem_size, false);
6551 if (!err && value_regno >= 0 && (t == BPF_READ || rdonly_mem))
6552 mark_reg_unknown(env, regs, value_regno);
6553 } else if (reg->type == PTR_TO_CTX) {
6554 enum bpf_reg_type reg_type = SCALAR_VALUE;
6555 struct btf *btf = NULL;
6558 if (t == BPF_WRITE && value_regno >= 0 &&
6559 is_pointer_value(env, value_regno)) {
6560 verbose(env, "R%d leaks addr into ctx\n", value_regno);
6564 err = check_ptr_off_reg(env, reg, regno);
6568 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf,
6571 verbose_linfo(env, insn_idx, "; ");
6572 if (!err && t == BPF_READ && value_regno >= 0) {
6573 /* ctx access returns either a scalar, or a
6574 * PTR_TO_PACKET[_META,_END]. In the latter
6575 * case, we know the offset is zero.
6577 if (reg_type == SCALAR_VALUE) {
6578 mark_reg_unknown(env, regs, value_regno);
6580 mark_reg_known_zero(env, regs,
6582 if (type_may_be_null(reg_type))
6583 regs[value_regno].id = ++env->id_gen;
6584 /* A load of ctx field could have different
6585 * actual load size with the one encoded in the
6586 * insn. When the dst is PTR, it is for sure not
6589 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
6590 if (base_type(reg_type) == PTR_TO_BTF_ID) {
6591 regs[value_regno].btf = btf;
6592 regs[value_regno].btf_id = btf_id;
6595 regs[value_regno].type = reg_type;
6598 } else if (reg->type == PTR_TO_STACK) {
6599 /* Basic bounds checks. */
6600 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
6604 state = func(env, reg);
6605 err = update_stack_depth(env, state, off);
6610 err = check_stack_read(env, regno, off, size,
6613 err = check_stack_write(env, regno, off, size,
6614 value_regno, insn_idx);
6615 } else if (reg_is_pkt_pointer(reg)) {
6616 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
6617 verbose(env, "cannot write into packet\n");
6620 if (t == BPF_WRITE && value_regno >= 0 &&
6621 is_pointer_value(env, value_regno)) {
6622 verbose(env, "R%d leaks addr into packet\n",
6626 err = check_packet_access(env, regno, off, size, false);
6627 if (!err && t == BPF_READ && value_regno >= 0)
6628 mark_reg_unknown(env, regs, value_regno);
6629 } else if (reg->type == PTR_TO_FLOW_KEYS) {
6630 if (t == BPF_WRITE && value_regno >= 0 &&
6631 is_pointer_value(env, value_regno)) {
6632 verbose(env, "R%d leaks addr into flow keys\n",
6637 err = check_flow_keys_access(env, off, size);
6638 if (!err && t == BPF_READ && value_regno >= 0)
6639 mark_reg_unknown(env, regs, value_regno);
6640 } else if (type_is_sk_pointer(reg->type)) {
6641 if (t == BPF_WRITE) {
6642 verbose(env, "R%d cannot write into %s\n",
6643 regno, reg_type_str(env, reg->type));
6646 err = check_sock_access(env, insn_idx, regno, off, size, t);
6647 if (!err && value_regno >= 0)
6648 mark_reg_unknown(env, regs, value_regno);
6649 } else if (reg->type == PTR_TO_TP_BUFFER) {
6650 err = check_tp_buffer_access(env, reg, regno, off, size);
6651 if (!err && t == BPF_READ && value_regno >= 0)
6652 mark_reg_unknown(env, regs, value_regno);
6653 } else if (base_type(reg->type) == PTR_TO_BTF_ID &&
6654 !type_may_be_null(reg->type)) {
6655 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
6657 } else if (reg->type == CONST_PTR_TO_MAP) {
6658 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
6660 } else if (base_type(reg->type) == PTR_TO_BUF) {
6661 bool rdonly_mem = type_is_rdonly_mem(reg->type);
6665 if (t == BPF_WRITE) {
6666 verbose(env, "R%d cannot write into %s\n",
6667 regno, reg_type_str(env, reg->type));
6670 max_access = &env->prog->aux->max_rdonly_access;
6672 max_access = &env->prog->aux->max_rdwr_access;
6675 err = check_buffer_access(env, reg, regno, off, size, false,
6678 if (!err && value_regno >= 0 && (rdonly_mem || t == BPF_READ))
6679 mark_reg_unknown(env, regs, value_regno);
6681 verbose(env, "R%d invalid mem access '%s'\n", regno,
6682 reg_type_str(env, reg->type));
6686 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
6687 regs[value_regno].type == SCALAR_VALUE) {
6689 /* b/h/w load zero-extends, mark upper bits as known 0 */
6690 coerce_reg_to_size(®s[value_regno], size);
6692 coerce_reg_to_size_sx(®s[value_regno], size);
6697 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
6702 switch (insn->imm) {
6704 case BPF_ADD | BPF_FETCH:
6706 case BPF_AND | BPF_FETCH:
6708 case BPF_OR | BPF_FETCH:
6710 case BPF_XOR | BPF_FETCH:
6715 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
6719 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
6720 verbose(env, "invalid atomic operand size\n");
6724 /* check src1 operand */
6725 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6729 /* check src2 operand */
6730 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6734 if (insn->imm == BPF_CMPXCHG) {
6735 /* Check comparison of R0 with memory location */
6736 const u32 aux_reg = BPF_REG_0;
6738 err = check_reg_arg(env, aux_reg, SRC_OP);
6742 if (is_pointer_value(env, aux_reg)) {
6743 verbose(env, "R%d leaks addr into mem\n", aux_reg);
6748 if (is_pointer_value(env, insn->src_reg)) {
6749 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
6753 if (is_ctx_reg(env, insn->dst_reg) ||
6754 is_pkt_reg(env, insn->dst_reg) ||
6755 is_flow_key_reg(env, insn->dst_reg) ||
6756 is_sk_reg(env, insn->dst_reg)) {
6757 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
6759 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
6763 if (insn->imm & BPF_FETCH) {
6764 if (insn->imm == BPF_CMPXCHG)
6765 load_reg = BPF_REG_0;
6767 load_reg = insn->src_reg;
6769 /* check and record load of old value */
6770 err = check_reg_arg(env, load_reg, DST_OP);
6774 /* This instruction accesses a memory location but doesn't
6775 * actually load it into a register.
6780 /* Check whether we can read the memory, with second call for fetch
6781 * case to simulate the register fill.
6783 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
6784 BPF_SIZE(insn->code), BPF_READ, -1, true, false);
6785 if (!err && load_reg >= 0)
6786 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
6787 BPF_SIZE(insn->code), BPF_READ, load_reg,
6792 /* Check whether we can write into the same memory. */
6793 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
6794 BPF_SIZE(insn->code), BPF_WRITE, -1, true, false);
6801 /* When register 'regno' is used to read the stack (either directly or through
6802 * a helper function) make sure that it's within stack boundary and, depending
6803 * on the access type, that all elements of the stack are initialized.
6805 * 'off' includes 'regno->off', but not its dynamic part (if any).
6807 * All registers that have been spilled on the stack in the slots within the
6808 * read offsets are marked as read.
6810 static int check_stack_range_initialized(
6811 struct bpf_verifier_env *env, int regno, int off,
6812 int access_size, bool zero_size_allowed,
6813 enum bpf_access_src type, struct bpf_call_arg_meta *meta)
6815 struct bpf_reg_state *reg = reg_state(env, regno);
6816 struct bpf_func_state *state = func(env, reg);
6817 int err, min_off, max_off, i, j, slot, spi;
6818 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
6819 enum bpf_access_type bounds_check_type;
6820 /* Some accesses can write anything into the stack, others are
6823 bool clobber = false;
6825 if (access_size == 0 && !zero_size_allowed) {
6826 verbose(env, "invalid zero-sized read\n");
6830 if (type == ACCESS_HELPER) {
6831 /* The bounds checks for writes are more permissive than for
6832 * reads. However, if raw_mode is not set, we'll do extra
6835 bounds_check_type = BPF_WRITE;
6838 bounds_check_type = BPF_READ;
6840 err = check_stack_access_within_bounds(env, regno, off, access_size,
6841 type, bounds_check_type);
6846 if (tnum_is_const(reg->var_off)) {
6847 min_off = max_off = reg->var_off.value + off;
6849 /* Variable offset is prohibited for unprivileged mode for
6850 * simplicity since it requires corresponding support in
6851 * Spectre masking for stack ALU.
6852 * See also retrieve_ptr_limit().
6854 if (!env->bypass_spec_v1) {
6857 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6858 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
6859 regno, err_extra, tn_buf);
6862 /* Only initialized buffer on stack is allowed to be accessed
6863 * with variable offset. With uninitialized buffer it's hard to
6864 * guarantee that whole memory is marked as initialized on
6865 * helper return since specific bounds are unknown what may
6866 * cause uninitialized stack leaking.
6868 if (meta && meta->raw_mode)
6871 min_off = reg->smin_value + off;
6872 max_off = reg->smax_value + off;
6875 if (meta && meta->raw_mode) {
6876 /* Ensure we won't be overwriting dynptrs when simulating byte
6877 * by byte access in check_helper_call using meta.access_size.
6878 * This would be a problem if we have a helper in the future
6881 * helper(uninit_mem, len, dynptr)
6883 * Now, uninint_mem may overlap with dynptr pointer. Hence, it
6884 * may end up writing to dynptr itself when touching memory from
6885 * arg 1. This can be relaxed on a case by case basis for known
6886 * safe cases, but reject due to the possibilitiy of aliasing by
6889 for (i = min_off; i < max_off + access_size; i++) {
6890 int stack_off = -i - 1;
6893 /* raw_mode may write past allocated_stack */
6894 if (state->allocated_stack <= stack_off)
6896 if (state->stack[spi].slot_type[stack_off % BPF_REG_SIZE] == STACK_DYNPTR) {
6897 verbose(env, "potential write to dynptr at off=%d disallowed\n", i);
6901 meta->access_size = access_size;
6902 meta->regno = regno;
6906 for (i = min_off; i < max_off + access_size; i++) {
6910 spi = slot / BPF_REG_SIZE;
6911 if (state->allocated_stack <= slot)
6913 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
6914 if (*stype == STACK_MISC)
6916 if ((*stype == STACK_ZERO) ||
6917 (*stype == STACK_INVALID && env->allow_uninit_stack)) {
6919 /* helper can write anything into the stack */
6920 *stype = STACK_MISC;
6925 if (is_spilled_reg(&state->stack[spi]) &&
6926 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
6927 env->allow_ptr_leaks)) {
6929 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
6930 for (j = 0; j < BPF_REG_SIZE; j++)
6931 scrub_spilled_slot(&state->stack[spi].slot_type[j]);
6937 if (tnum_is_const(reg->var_off)) {
6938 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
6939 err_extra, regno, min_off, i - min_off, access_size);
6943 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6944 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
6945 err_extra, regno, tn_buf, i - min_off, access_size);
6949 /* reading any byte out of 8-byte 'spill_slot' will cause
6950 * the whole slot to be marked as 'read'
6952 mark_reg_read(env, &state->stack[spi].spilled_ptr,
6953 state->stack[spi].spilled_ptr.parent,
6955 /* We do not set REG_LIVE_WRITTEN for stack slot, as we can not
6956 * be sure that whether stack slot is written to or not. Hence,
6957 * we must still conservatively propagate reads upwards even if
6958 * helper may write to the entire memory range.
6961 return update_stack_depth(env, state, min_off);
6964 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
6965 int access_size, bool zero_size_allowed,
6966 struct bpf_call_arg_meta *meta)
6968 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
6971 switch (base_type(reg->type)) {
6973 case PTR_TO_PACKET_META:
6974 return check_packet_access(env, regno, reg->off, access_size,
6976 case PTR_TO_MAP_KEY:
6977 if (meta && meta->raw_mode) {
6978 verbose(env, "R%d cannot write into %s\n", regno,
6979 reg_type_str(env, reg->type));
6982 return check_mem_region_access(env, regno, reg->off, access_size,
6983 reg->map_ptr->key_size, false);
6984 case PTR_TO_MAP_VALUE:
6985 if (check_map_access_type(env, regno, reg->off, access_size,
6986 meta && meta->raw_mode ? BPF_WRITE :
6989 return check_map_access(env, regno, reg->off, access_size,
6990 zero_size_allowed, ACCESS_HELPER);
6992 if (type_is_rdonly_mem(reg->type)) {
6993 if (meta && meta->raw_mode) {
6994 verbose(env, "R%d cannot write into %s\n", regno,
6995 reg_type_str(env, reg->type));
6999 return check_mem_region_access(env, regno, reg->off,
7000 access_size, reg->mem_size,
7003 if (type_is_rdonly_mem(reg->type)) {
7004 if (meta && meta->raw_mode) {
7005 verbose(env, "R%d cannot write into %s\n", regno,
7006 reg_type_str(env, reg->type));
7010 max_access = &env->prog->aux->max_rdonly_access;
7012 max_access = &env->prog->aux->max_rdwr_access;
7014 return check_buffer_access(env, reg, regno, reg->off,
7015 access_size, zero_size_allowed,
7018 return check_stack_range_initialized(
7020 regno, reg->off, access_size,
7021 zero_size_allowed, ACCESS_HELPER, meta);
7023 return check_ptr_to_btf_access(env, regs, regno, reg->off,
7024 access_size, BPF_READ, -1);
7026 /* in case the function doesn't know how to access the context,
7027 * (because we are in a program of type SYSCALL for example), we
7028 * can not statically check its size.
7029 * Dynamically check it now.
7031 if (!env->ops->convert_ctx_access) {
7032 enum bpf_access_type atype = meta && meta->raw_mode ? BPF_WRITE : BPF_READ;
7033 int offset = access_size - 1;
7035 /* Allow zero-byte read from PTR_TO_CTX */
7036 if (access_size == 0)
7037 return zero_size_allowed ? 0 : -EACCES;
7039 return check_mem_access(env, env->insn_idx, regno, offset, BPF_B,
7040 atype, -1, false, false);
7044 default: /* scalar_value or invalid ptr */
7045 /* Allow zero-byte read from NULL, regardless of pointer type */
7046 if (zero_size_allowed && access_size == 0 &&
7047 register_is_null(reg))
7050 verbose(env, "R%d type=%s ", regno,
7051 reg_type_str(env, reg->type));
7052 verbose(env, "expected=%s\n", reg_type_str(env, PTR_TO_STACK));
7057 static int check_mem_size_reg(struct bpf_verifier_env *env,
7058 struct bpf_reg_state *reg, u32 regno,
7059 bool zero_size_allowed,
7060 struct bpf_call_arg_meta *meta)
7064 /* This is used to refine r0 return value bounds for helpers
7065 * that enforce this value as an upper bound on return values.
7066 * See do_refine_retval_range() for helpers that can refine
7067 * the return value. C type of helper is u32 so we pull register
7068 * bound from umax_value however, if negative verifier errors
7069 * out. Only upper bounds can be learned because retval is an
7070 * int type and negative retvals are allowed.
7072 meta->msize_max_value = reg->umax_value;
7074 /* The register is SCALAR_VALUE; the access check
7075 * happens using its boundaries.
7077 if (!tnum_is_const(reg->var_off))
7078 /* For unprivileged variable accesses, disable raw
7079 * mode so that the program is required to
7080 * initialize all the memory that the helper could
7081 * just partially fill up.
7085 if (reg->smin_value < 0) {
7086 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
7091 if (reg->umin_value == 0) {
7092 err = check_helper_mem_access(env, regno - 1, 0,
7099 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
7100 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
7104 err = check_helper_mem_access(env, regno - 1,
7106 zero_size_allowed, meta);
7108 err = mark_chain_precision(env, regno);
7112 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
7113 u32 regno, u32 mem_size)
7115 bool may_be_null = type_may_be_null(reg->type);
7116 struct bpf_reg_state saved_reg;
7117 struct bpf_call_arg_meta meta;
7120 if (register_is_null(reg))
7123 memset(&meta, 0, sizeof(meta));
7124 /* Assuming that the register contains a value check if the memory
7125 * access is safe. Temporarily save and restore the register's state as
7126 * the conversion shouldn't be visible to a caller.
7130 mark_ptr_not_null_reg(reg);
7133 err = check_helper_mem_access(env, regno, mem_size, true, &meta);
7134 /* Check access for BPF_WRITE */
7135 meta.raw_mode = true;
7136 err = err ?: check_helper_mem_access(env, regno, mem_size, true, &meta);
7144 static int check_kfunc_mem_size_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
7147 struct bpf_reg_state *mem_reg = &cur_regs(env)[regno - 1];
7148 bool may_be_null = type_may_be_null(mem_reg->type);
7149 struct bpf_reg_state saved_reg;
7150 struct bpf_call_arg_meta meta;
7153 WARN_ON_ONCE(regno < BPF_REG_2 || regno > BPF_REG_5);
7155 memset(&meta, 0, sizeof(meta));
7158 saved_reg = *mem_reg;
7159 mark_ptr_not_null_reg(mem_reg);
7162 err = check_mem_size_reg(env, reg, regno, true, &meta);
7163 /* Check access for BPF_WRITE */
7164 meta.raw_mode = true;
7165 err = err ?: check_mem_size_reg(env, reg, regno, true, &meta);
7168 *mem_reg = saved_reg;
7172 /* Implementation details:
7173 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL.
7174 * bpf_obj_new returns PTR_TO_BTF_ID | MEM_ALLOC | PTR_MAYBE_NULL.
7175 * Two bpf_map_lookups (even with the same key) will have different reg->id.
7176 * Two separate bpf_obj_new will also have different reg->id.
7177 * For traditional PTR_TO_MAP_VALUE or PTR_TO_BTF_ID | MEM_ALLOC, the verifier
7178 * clears reg->id after value_or_null->value transition, since the verifier only
7179 * cares about the range of access to valid map value pointer and doesn't care
7180 * about actual address of the map element.
7181 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
7182 * reg->id > 0 after value_or_null->value transition. By doing so
7183 * two bpf_map_lookups will be considered two different pointers that
7184 * point to different bpf_spin_locks. Likewise for pointers to allocated objects
7185 * returned from bpf_obj_new.
7186 * The verifier allows taking only one bpf_spin_lock at a time to avoid
7188 * Since only one bpf_spin_lock is allowed the checks are simpler than
7189 * reg_is_refcounted() logic. The verifier needs to remember only
7190 * one spin_lock instead of array of acquired_refs.
7191 * cur_state->active_lock remembers which map value element or allocated
7192 * object got locked and clears it after bpf_spin_unlock.
7194 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
7197 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7198 struct bpf_verifier_state *cur = env->cur_state;
7199 bool is_const = tnum_is_const(reg->var_off);
7200 u64 val = reg->var_off.value;
7201 struct bpf_map *map = NULL;
7202 struct btf *btf = NULL;
7203 struct btf_record *rec;
7207 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
7211 if (reg->type == PTR_TO_MAP_VALUE) {
7215 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
7223 rec = reg_btf_record(reg);
7224 if (!btf_record_has_field(rec, BPF_SPIN_LOCK)) {
7225 verbose(env, "%s '%s' has no valid bpf_spin_lock\n", map ? "map" : "local",
7226 map ? map->name : "kptr");
7229 if (rec->spin_lock_off != val + reg->off) {
7230 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock' that is at %d\n",
7231 val + reg->off, rec->spin_lock_off);
7235 if (cur->active_lock.ptr) {
7237 "Locking two bpf_spin_locks are not allowed\n");
7241 cur->active_lock.ptr = map;
7243 cur->active_lock.ptr = btf;
7244 cur->active_lock.id = reg->id;
7253 if (!cur->active_lock.ptr) {
7254 verbose(env, "bpf_spin_unlock without taking a lock\n");
7257 if (cur->active_lock.ptr != ptr ||
7258 cur->active_lock.id != reg->id) {
7259 verbose(env, "bpf_spin_unlock of different lock\n");
7263 invalidate_non_owning_refs(env);
7265 cur->active_lock.ptr = NULL;
7266 cur->active_lock.id = 0;
7271 static int process_timer_func(struct bpf_verifier_env *env, int regno,
7272 struct bpf_call_arg_meta *meta)
7274 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7275 bool is_const = tnum_is_const(reg->var_off);
7276 struct bpf_map *map = reg->map_ptr;
7277 u64 val = reg->var_off.value;
7281 "R%d doesn't have constant offset. bpf_timer has to be at the constant offset\n",
7286 verbose(env, "map '%s' has to have BTF in order to use bpf_timer\n",
7290 if (!btf_record_has_field(map->record, BPF_TIMER)) {
7291 verbose(env, "map '%s' has no valid bpf_timer\n", map->name);
7294 if (map->record->timer_off != val + reg->off) {
7295 verbose(env, "off %lld doesn't point to 'struct bpf_timer' that is at %d\n",
7296 val + reg->off, map->record->timer_off);
7299 if (meta->map_ptr) {
7300 verbose(env, "verifier bug. Two map pointers in a timer helper\n");
7303 meta->map_uid = reg->map_uid;
7304 meta->map_ptr = map;
7308 static int process_kptr_func(struct bpf_verifier_env *env, int regno,
7309 struct bpf_call_arg_meta *meta)
7311 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7312 struct bpf_map *map_ptr = reg->map_ptr;
7313 struct btf_field *kptr_field;
7316 if (!tnum_is_const(reg->var_off)) {
7318 "R%d doesn't have constant offset. kptr has to be at the constant offset\n",
7322 if (!map_ptr->btf) {
7323 verbose(env, "map '%s' has to have BTF in order to use bpf_kptr_xchg\n",
7327 if (!btf_record_has_field(map_ptr->record, BPF_KPTR)) {
7328 verbose(env, "map '%s' has no valid kptr\n", map_ptr->name);
7332 meta->map_ptr = map_ptr;
7333 kptr_off = reg->off + reg->var_off.value;
7334 kptr_field = btf_record_find(map_ptr->record, kptr_off, BPF_KPTR);
7336 verbose(env, "off=%d doesn't point to kptr\n", kptr_off);
7339 if (kptr_field->type != BPF_KPTR_REF) {
7340 verbose(env, "off=%d kptr isn't referenced kptr\n", kptr_off);
7343 meta->kptr_field = kptr_field;
7347 /* There are two register types representing a bpf_dynptr, one is PTR_TO_STACK
7348 * which points to a stack slot, and the other is CONST_PTR_TO_DYNPTR.
7350 * In both cases we deal with the first 8 bytes, but need to mark the next 8
7351 * bytes as STACK_DYNPTR in case of PTR_TO_STACK. In case of
7352 * CONST_PTR_TO_DYNPTR, we are guaranteed to get the beginning of the object.
7354 * Mutability of bpf_dynptr is at two levels, one is at the level of struct
7355 * bpf_dynptr itself, i.e. whether the helper is receiving a pointer to struct
7356 * bpf_dynptr or pointer to const struct bpf_dynptr. In the former case, it can
7357 * mutate the view of the dynptr and also possibly destroy it. In the latter
7358 * case, it cannot mutate the bpf_dynptr itself but it can still mutate the
7359 * memory that dynptr points to.
7361 * The verifier will keep track both levels of mutation (bpf_dynptr's in
7362 * reg->type and the memory's in reg->dynptr.type), but there is no support for
7363 * readonly dynptr view yet, hence only the first case is tracked and checked.
7365 * This is consistent with how C applies the const modifier to a struct object,
7366 * where the pointer itself inside bpf_dynptr becomes const but not what it
7369 * Helpers which do not mutate the bpf_dynptr set MEM_RDONLY in their argument
7370 * type, and declare it as 'const struct bpf_dynptr *' in their prototype.
7372 static int process_dynptr_func(struct bpf_verifier_env *env, int regno, int insn_idx,
7373 enum bpf_arg_type arg_type, int clone_ref_obj_id)
7375 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7378 /* MEM_UNINIT and MEM_RDONLY are exclusive, when applied to an
7379 * ARG_PTR_TO_DYNPTR (or ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_*):
7381 if ((arg_type & (MEM_UNINIT | MEM_RDONLY)) == (MEM_UNINIT | MEM_RDONLY)) {
7382 verbose(env, "verifier internal error: misconfigured dynptr helper type flags\n");
7386 /* MEM_UNINIT - Points to memory that is an appropriate candidate for
7387 * constructing a mutable bpf_dynptr object.
7389 * Currently, this is only possible with PTR_TO_STACK
7390 * pointing to a region of at least 16 bytes which doesn't
7391 * contain an existing bpf_dynptr.
7393 * MEM_RDONLY - Points to a initialized bpf_dynptr that will not be
7394 * mutated or destroyed. However, the memory it points to
7397 * None - Points to a initialized dynptr that can be mutated and
7398 * destroyed, including mutation of the memory it points
7401 if (arg_type & MEM_UNINIT) {
7404 if (!is_dynptr_reg_valid_uninit(env, reg)) {
7405 verbose(env, "Dynptr has to be an uninitialized dynptr\n");
7409 /* we write BPF_DW bits (8 bytes) at a time */
7410 for (i = 0; i < BPF_DYNPTR_SIZE; i += 8) {
7411 err = check_mem_access(env, insn_idx, regno,
7412 i, BPF_DW, BPF_WRITE, -1, false, false);
7417 err = mark_stack_slots_dynptr(env, reg, arg_type, insn_idx, clone_ref_obj_id);
7418 } else /* MEM_RDONLY and None case from above */ {
7419 /* For the reg->type == PTR_TO_STACK case, bpf_dynptr is never const */
7420 if (reg->type == CONST_PTR_TO_DYNPTR && !(arg_type & MEM_RDONLY)) {
7421 verbose(env, "cannot pass pointer to const bpf_dynptr, the helper mutates it\n");
7425 if (!is_dynptr_reg_valid_init(env, reg)) {
7427 "Expected an initialized dynptr as arg #%d\n",
7432 /* Fold modifiers (in this case, MEM_RDONLY) when checking expected type */
7433 if (!is_dynptr_type_expected(env, reg, arg_type & ~MEM_RDONLY)) {
7435 "Expected a dynptr of type %s as arg #%d\n",
7436 dynptr_type_str(arg_to_dynptr_type(arg_type)), regno);
7440 err = mark_dynptr_read(env, reg);
7445 static u32 iter_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg, int spi)
7447 struct bpf_func_state *state = func(env, reg);
7449 return state->stack[spi].spilled_ptr.ref_obj_id;
7452 static bool is_iter_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7454 return meta->kfunc_flags & (KF_ITER_NEW | KF_ITER_NEXT | KF_ITER_DESTROY);
7457 static bool is_iter_new_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7459 return meta->kfunc_flags & KF_ITER_NEW;
7462 static bool is_iter_next_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7464 return meta->kfunc_flags & KF_ITER_NEXT;
7467 static bool is_iter_destroy_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7469 return meta->kfunc_flags & KF_ITER_DESTROY;
7472 static bool is_kfunc_arg_iter(struct bpf_kfunc_call_arg_meta *meta, int arg)
7474 /* btf_check_iter_kfuncs() guarantees that first argument of any iter
7475 * kfunc is iter state pointer
7477 return arg == 0 && is_iter_kfunc(meta);
7480 static int process_iter_arg(struct bpf_verifier_env *env, int regno, int insn_idx,
7481 struct bpf_kfunc_call_arg_meta *meta)
7483 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7484 const struct btf_type *t;
7485 const struct btf_param *arg;
7486 int spi, err, i, nr_slots;
7489 /* btf_check_iter_kfuncs() ensures we don't need to validate anything here */
7490 arg = &btf_params(meta->func_proto)[0];
7491 t = btf_type_skip_modifiers(meta->btf, arg->type, NULL); /* PTR */
7492 t = btf_type_skip_modifiers(meta->btf, t->type, &btf_id); /* STRUCT */
7493 nr_slots = t->size / BPF_REG_SIZE;
7495 if (is_iter_new_kfunc(meta)) {
7496 /* bpf_iter_<type>_new() expects pointer to uninit iter state */
7497 if (!is_iter_reg_valid_uninit(env, reg, nr_slots)) {
7498 verbose(env, "expected uninitialized iter_%s as arg #%d\n",
7499 iter_type_str(meta->btf, btf_id), regno);
7503 for (i = 0; i < nr_slots * 8; i += BPF_REG_SIZE) {
7504 err = check_mem_access(env, insn_idx, regno,
7505 i, BPF_DW, BPF_WRITE, -1, false, false);
7510 err = mark_stack_slots_iter(env, reg, insn_idx, meta->btf, btf_id, nr_slots);
7514 /* iter_next() or iter_destroy() expect initialized iter state*/
7515 if (!is_iter_reg_valid_init(env, reg, meta->btf, btf_id, nr_slots)) {
7516 verbose(env, "expected an initialized iter_%s as arg #%d\n",
7517 iter_type_str(meta->btf, btf_id), regno);
7521 spi = iter_get_spi(env, reg, nr_slots);
7525 err = mark_iter_read(env, reg, spi, nr_slots);
7529 /* remember meta->iter info for process_iter_next_call() */
7530 meta->iter.spi = spi;
7531 meta->iter.frameno = reg->frameno;
7532 meta->ref_obj_id = iter_ref_obj_id(env, reg, spi);
7534 if (is_iter_destroy_kfunc(meta)) {
7535 err = unmark_stack_slots_iter(env, reg, nr_slots);
7544 /* process_iter_next_call() is called when verifier gets to iterator's next
7545 * "method" (e.g., bpf_iter_num_next() for numbers iterator) call. We'll refer
7546 * to it as just "iter_next()" in comments below.
7548 * BPF verifier relies on a crucial contract for any iter_next()
7549 * implementation: it should *eventually* return NULL, and once that happens
7550 * it should keep returning NULL. That is, once iterator exhausts elements to
7551 * iterate, it should never reset or spuriously return new elements.
7553 * With the assumption of such contract, process_iter_next_call() simulates
7554 * a fork in the verifier state to validate loop logic correctness and safety
7555 * without having to simulate infinite amount of iterations.
7557 * In current state, we first assume that iter_next() returned NULL and
7558 * iterator state is set to DRAINED (BPF_ITER_STATE_DRAINED). In such
7559 * conditions we should not form an infinite loop and should eventually reach
7562 * Besides that, we also fork current state and enqueue it for later
7563 * verification. In a forked state we keep iterator state as ACTIVE
7564 * (BPF_ITER_STATE_ACTIVE) and assume non-NULL return from iter_next(). We
7565 * also bump iteration depth to prevent erroneous infinite loop detection
7566 * later on (see iter_active_depths_differ() comment for details). In this
7567 * state we assume that we'll eventually loop back to another iter_next()
7568 * calls (it could be in exactly same location or in some other instruction,
7569 * it doesn't matter, we don't make any unnecessary assumptions about this,
7570 * everything revolves around iterator state in a stack slot, not which
7571 * instruction is calling iter_next()). When that happens, we either will come
7572 * to iter_next() with equivalent state and can conclude that next iteration
7573 * will proceed in exactly the same way as we just verified, so it's safe to
7574 * assume that loop converges. If not, we'll go on another iteration
7575 * simulation with a different input state, until all possible starting states
7576 * are validated or we reach maximum number of instructions limit.
7578 * This way, we will either exhaustively discover all possible input states
7579 * that iterator loop can start with and eventually will converge, or we'll
7580 * effectively regress into bounded loop simulation logic and either reach
7581 * maximum number of instructions if loop is not provably convergent, or there
7582 * is some statically known limit on number of iterations (e.g., if there is
7583 * an explicit `if n > 100 then break;` statement somewhere in the loop).
7585 * One very subtle but very important aspect is that we *always* simulate NULL
7586 * condition first (as the current state) before we simulate non-NULL case.
7587 * This has to do with intricacies of scalar precision tracking. By simulating
7588 * "exit condition" of iter_next() returning NULL first, we make sure all the
7589 * relevant precision marks *that will be set **after** we exit iterator loop*
7590 * are propagated backwards to common parent state of NULL and non-NULL
7591 * branches. Thanks to that, state equivalence checks done later in forked
7592 * state, when reaching iter_next() for ACTIVE iterator, can assume that
7593 * precision marks are finalized and won't change. Because simulating another
7594 * ACTIVE iterator iteration won't change them (because given same input
7595 * states we'll end up with exactly same output states which we are currently
7596 * comparing; and verification after the loop already propagated back what
7597 * needs to be **additionally** tracked as precise). It's subtle, grok
7598 * precision tracking for more intuitive understanding.
7600 static int process_iter_next_call(struct bpf_verifier_env *env, int insn_idx,
7601 struct bpf_kfunc_call_arg_meta *meta)
7603 struct bpf_verifier_state *cur_st = env->cur_state, *queued_st;
7604 struct bpf_func_state *cur_fr = cur_st->frame[cur_st->curframe], *queued_fr;
7605 struct bpf_reg_state *cur_iter, *queued_iter;
7606 int iter_frameno = meta->iter.frameno;
7607 int iter_spi = meta->iter.spi;
7609 BTF_TYPE_EMIT(struct bpf_iter);
7611 cur_iter = &env->cur_state->frame[iter_frameno]->stack[iter_spi].spilled_ptr;
7613 if (cur_iter->iter.state != BPF_ITER_STATE_ACTIVE &&
7614 cur_iter->iter.state != BPF_ITER_STATE_DRAINED) {
7615 verbose(env, "verifier internal error: unexpected iterator state %d (%s)\n",
7616 cur_iter->iter.state, iter_state_str(cur_iter->iter.state));
7620 if (cur_iter->iter.state == BPF_ITER_STATE_ACTIVE) {
7621 /* branch out active iter state */
7622 queued_st = push_stack(env, insn_idx + 1, insn_idx, false);
7626 queued_iter = &queued_st->frame[iter_frameno]->stack[iter_spi].spilled_ptr;
7627 queued_iter->iter.state = BPF_ITER_STATE_ACTIVE;
7628 queued_iter->iter.depth++;
7630 queued_fr = queued_st->frame[queued_st->curframe];
7631 mark_ptr_not_null_reg(&queued_fr->regs[BPF_REG_0]);
7634 /* switch to DRAINED state, but keep the depth unchanged */
7635 /* mark current iter state as drained and assume returned NULL */
7636 cur_iter->iter.state = BPF_ITER_STATE_DRAINED;
7637 __mark_reg_const_zero(&cur_fr->regs[BPF_REG_0]);
7642 static bool arg_type_is_mem_size(enum bpf_arg_type type)
7644 return type == ARG_CONST_SIZE ||
7645 type == ARG_CONST_SIZE_OR_ZERO;
7648 static bool arg_type_is_release(enum bpf_arg_type type)
7650 return type & OBJ_RELEASE;
7653 static bool arg_type_is_dynptr(enum bpf_arg_type type)
7655 return base_type(type) == ARG_PTR_TO_DYNPTR;
7658 static int int_ptr_type_to_size(enum bpf_arg_type type)
7660 if (type == ARG_PTR_TO_INT)
7662 else if (type == ARG_PTR_TO_LONG)
7668 static int resolve_map_arg_type(struct bpf_verifier_env *env,
7669 const struct bpf_call_arg_meta *meta,
7670 enum bpf_arg_type *arg_type)
7672 if (!meta->map_ptr) {
7673 /* kernel subsystem misconfigured verifier */
7674 verbose(env, "invalid map_ptr to access map->type\n");
7678 switch (meta->map_ptr->map_type) {
7679 case BPF_MAP_TYPE_SOCKMAP:
7680 case BPF_MAP_TYPE_SOCKHASH:
7681 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
7682 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
7684 verbose(env, "invalid arg_type for sockmap/sockhash\n");
7688 case BPF_MAP_TYPE_BLOOM_FILTER:
7689 if (meta->func_id == BPF_FUNC_map_peek_elem)
7690 *arg_type = ARG_PTR_TO_MAP_VALUE;
7698 struct bpf_reg_types {
7699 const enum bpf_reg_type types[10];
7703 static const struct bpf_reg_types sock_types = {
7713 static const struct bpf_reg_types btf_id_sock_common_types = {
7720 PTR_TO_BTF_ID | PTR_TRUSTED,
7722 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
7726 static const struct bpf_reg_types mem_types = {
7734 PTR_TO_MEM | MEM_RINGBUF,
7736 PTR_TO_BTF_ID | PTR_TRUSTED,
7740 static const struct bpf_reg_types int_ptr_types = {
7750 static const struct bpf_reg_types spin_lock_types = {
7753 PTR_TO_BTF_ID | MEM_ALLOC,
7757 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
7758 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
7759 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
7760 static const struct bpf_reg_types ringbuf_mem_types = { .types = { PTR_TO_MEM | MEM_RINGBUF } };
7761 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
7762 static const struct bpf_reg_types btf_ptr_types = {
7765 PTR_TO_BTF_ID | PTR_TRUSTED,
7766 PTR_TO_BTF_ID | MEM_RCU,
7769 static const struct bpf_reg_types percpu_btf_ptr_types = {
7771 PTR_TO_BTF_ID | MEM_PERCPU,
7772 PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED,
7775 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
7776 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
7777 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
7778 static const struct bpf_reg_types timer_types = { .types = { PTR_TO_MAP_VALUE } };
7779 static const struct bpf_reg_types kptr_types = { .types = { PTR_TO_MAP_VALUE } };
7780 static const struct bpf_reg_types dynptr_types = {
7783 CONST_PTR_TO_DYNPTR,
7787 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
7788 [ARG_PTR_TO_MAP_KEY] = &mem_types,
7789 [ARG_PTR_TO_MAP_VALUE] = &mem_types,
7790 [ARG_CONST_SIZE] = &scalar_types,
7791 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
7792 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
7793 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
7794 [ARG_PTR_TO_CTX] = &context_types,
7795 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
7797 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
7799 [ARG_PTR_TO_SOCKET] = &fullsock_types,
7800 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
7801 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
7802 [ARG_PTR_TO_MEM] = &mem_types,
7803 [ARG_PTR_TO_RINGBUF_MEM] = &ringbuf_mem_types,
7804 [ARG_PTR_TO_INT] = &int_ptr_types,
7805 [ARG_PTR_TO_LONG] = &int_ptr_types,
7806 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
7807 [ARG_PTR_TO_FUNC] = &func_ptr_types,
7808 [ARG_PTR_TO_STACK] = &stack_ptr_types,
7809 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
7810 [ARG_PTR_TO_TIMER] = &timer_types,
7811 [ARG_PTR_TO_KPTR] = &kptr_types,
7812 [ARG_PTR_TO_DYNPTR] = &dynptr_types,
7815 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
7816 enum bpf_arg_type arg_type,
7817 const u32 *arg_btf_id,
7818 struct bpf_call_arg_meta *meta)
7820 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7821 enum bpf_reg_type expected, type = reg->type;
7822 const struct bpf_reg_types *compatible;
7825 compatible = compatible_reg_types[base_type(arg_type)];
7827 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
7831 /* ARG_PTR_TO_MEM + RDONLY is compatible with PTR_TO_MEM and PTR_TO_MEM + RDONLY,
7832 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM and NOT with PTR_TO_MEM + RDONLY
7834 * Same for MAYBE_NULL:
7836 * ARG_PTR_TO_MEM + MAYBE_NULL is compatible with PTR_TO_MEM and PTR_TO_MEM + MAYBE_NULL,
7837 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM but NOT with PTR_TO_MEM + MAYBE_NULL
7839 * ARG_PTR_TO_MEM is compatible with PTR_TO_MEM that is tagged with a dynptr type.
7841 * Therefore we fold these flags depending on the arg_type before comparison.
7843 if (arg_type & MEM_RDONLY)
7844 type &= ~MEM_RDONLY;
7845 if (arg_type & PTR_MAYBE_NULL)
7846 type &= ~PTR_MAYBE_NULL;
7847 if (base_type(arg_type) == ARG_PTR_TO_MEM)
7848 type &= ~DYNPTR_TYPE_FLAG_MASK;
7850 if (meta->func_id == BPF_FUNC_kptr_xchg && type_is_alloc(type))
7853 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
7854 expected = compatible->types[i];
7855 if (expected == NOT_INIT)
7858 if (type == expected)
7862 verbose(env, "R%d type=%s expected=", regno, reg_type_str(env, reg->type));
7863 for (j = 0; j + 1 < i; j++)
7864 verbose(env, "%s, ", reg_type_str(env, compatible->types[j]));
7865 verbose(env, "%s\n", reg_type_str(env, compatible->types[j]));
7869 if (base_type(reg->type) != PTR_TO_BTF_ID)
7872 if (compatible == &mem_types) {
7873 if (!(arg_type & MEM_RDONLY)) {
7875 "%s() may write into memory pointed by R%d type=%s\n",
7876 func_id_name(meta->func_id),
7877 regno, reg_type_str(env, reg->type));
7883 switch ((int)reg->type) {
7885 case PTR_TO_BTF_ID | PTR_TRUSTED:
7886 case PTR_TO_BTF_ID | MEM_RCU:
7887 case PTR_TO_BTF_ID | PTR_MAYBE_NULL:
7888 case PTR_TO_BTF_ID | PTR_MAYBE_NULL | MEM_RCU:
7890 /* For bpf_sk_release, it needs to match against first member
7891 * 'struct sock_common', hence make an exception for it. This
7892 * allows bpf_sk_release to work for multiple socket types.
7894 bool strict_type_match = arg_type_is_release(arg_type) &&
7895 meta->func_id != BPF_FUNC_sk_release;
7897 if (type_may_be_null(reg->type) &&
7898 (!type_may_be_null(arg_type) || arg_type_is_release(arg_type))) {
7899 verbose(env, "Possibly NULL pointer passed to helper arg%d\n", regno);
7904 if (!compatible->btf_id) {
7905 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
7908 arg_btf_id = compatible->btf_id;
7911 if (meta->func_id == BPF_FUNC_kptr_xchg) {
7912 if (map_kptr_match_type(env, meta->kptr_field, reg, regno))
7915 if (arg_btf_id == BPF_PTR_POISON) {
7916 verbose(env, "verifier internal error:");
7917 verbose(env, "R%d has non-overwritten BPF_PTR_POISON type\n",
7922 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
7923 btf_vmlinux, *arg_btf_id,
7924 strict_type_match)) {
7925 verbose(env, "R%d is of type %s but %s is expected\n",
7926 regno, btf_type_name(reg->btf, reg->btf_id),
7927 btf_type_name(btf_vmlinux, *arg_btf_id));
7933 case PTR_TO_BTF_ID | MEM_ALLOC:
7934 if (meta->func_id != BPF_FUNC_spin_lock && meta->func_id != BPF_FUNC_spin_unlock &&
7935 meta->func_id != BPF_FUNC_kptr_xchg) {
7936 verbose(env, "verifier internal error: unimplemented handling of MEM_ALLOC\n");
7939 if (meta->func_id == BPF_FUNC_kptr_xchg) {
7940 if (map_kptr_match_type(env, meta->kptr_field, reg, regno))
7944 case PTR_TO_BTF_ID | MEM_PERCPU:
7945 case PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED:
7946 /* Handled by helper specific checks */
7949 verbose(env, "verifier internal error: invalid PTR_TO_BTF_ID register for type match\n");
7955 static struct btf_field *
7956 reg_find_field_offset(const struct bpf_reg_state *reg, s32 off, u32 fields)
7958 struct btf_field *field;
7959 struct btf_record *rec;
7961 rec = reg_btf_record(reg);
7965 field = btf_record_find(rec, off, fields);
7972 int check_func_arg_reg_off(struct bpf_verifier_env *env,
7973 const struct bpf_reg_state *reg, int regno,
7974 enum bpf_arg_type arg_type)
7976 u32 type = reg->type;
7978 /* When referenced register is passed to release function, its fixed
7981 * We will check arg_type_is_release reg has ref_obj_id when storing
7982 * meta->release_regno.
7984 if (arg_type_is_release(arg_type)) {
7985 /* ARG_PTR_TO_DYNPTR with OBJ_RELEASE is a bit special, as it
7986 * may not directly point to the object being released, but to
7987 * dynptr pointing to such object, which might be at some offset
7988 * on the stack. In that case, we simply to fallback to the
7991 if (arg_type_is_dynptr(arg_type) && type == PTR_TO_STACK)
7994 /* Doing check_ptr_off_reg check for the offset will catch this
7995 * because fixed_off_ok is false, but checking here allows us
7996 * to give the user a better error message.
7999 verbose(env, "R%d must have zero offset when passed to release func or trusted arg to kfunc\n",
8003 return __check_ptr_off_reg(env, reg, regno, false);
8007 /* Pointer types where both fixed and variable offset is explicitly allowed: */
8010 case PTR_TO_PACKET_META:
8011 case PTR_TO_MAP_KEY:
8012 case PTR_TO_MAP_VALUE:
8014 case PTR_TO_MEM | MEM_RDONLY:
8015 case PTR_TO_MEM | MEM_RINGBUF:
8017 case PTR_TO_BUF | MEM_RDONLY:
8020 /* All the rest must be rejected, except PTR_TO_BTF_ID which allows
8024 case PTR_TO_BTF_ID | MEM_ALLOC:
8025 case PTR_TO_BTF_ID | PTR_TRUSTED:
8026 case PTR_TO_BTF_ID | MEM_RCU:
8027 case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF:
8028 case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF | MEM_RCU:
8029 /* When referenced PTR_TO_BTF_ID is passed to release function,
8030 * its fixed offset must be 0. In the other cases, fixed offset
8031 * can be non-zero. This was already checked above. So pass
8032 * fixed_off_ok as true to allow fixed offset for all other
8033 * cases. var_off always must be 0 for PTR_TO_BTF_ID, hence we
8034 * still need to do checks instead of returning.
8036 return __check_ptr_off_reg(env, reg, regno, true);
8038 return __check_ptr_off_reg(env, reg, regno, false);
8042 static struct bpf_reg_state *get_dynptr_arg_reg(struct bpf_verifier_env *env,
8043 const struct bpf_func_proto *fn,
8044 struct bpf_reg_state *regs)
8046 struct bpf_reg_state *state = NULL;
8049 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++)
8050 if (arg_type_is_dynptr(fn->arg_type[i])) {
8052 verbose(env, "verifier internal error: multiple dynptr args\n");
8055 state = ®s[BPF_REG_1 + i];
8059 verbose(env, "verifier internal error: no dynptr arg found\n");
8064 static int dynptr_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
8066 struct bpf_func_state *state = func(env, reg);
8069 if (reg->type == CONST_PTR_TO_DYNPTR)
8071 spi = dynptr_get_spi(env, reg);
8074 return state->stack[spi].spilled_ptr.id;
8077 static int dynptr_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
8079 struct bpf_func_state *state = func(env, reg);
8082 if (reg->type == CONST_PTR_TO_DYNPTR)
8083 return reg->ref_obj_id;
8084 spi = dynptr_get_spi(env, reg);
8087 return state->stack[spi].spilled_ptr.ref_obj_id;
8090 static enum bpf_dynptr_type dynptr_get_type(struct bpf_verifier_env *env,
8091 struct bpf_reg_state *reg)
8093 struct bpf_func_state *state = func(env, reg);
8096 if (reg->type == CONST_PTR_TO_DYNPTR)
8097 return reg->dynptr.type;
8099 spi = __get_spi(reg->off);
8101 verbose(env, "verifier internal error: invalid spi when querying dynptr type\n");
8102 return BPF_DYNPTR_TYPE_INVALID;
8105 return state->stack[spi].spilled_ptr.dynptr.type;
8108 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
8109 struct bpf_call_arg_meta *meta,
8110 const struct bpf_func_proto *fn,
8113 u32 regno = BPF_REG_1 + arg;
8114 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
8115 enum bpf_arg_type arg_type = fn->arg_type[arg];
8116 enum bpf_reg_type type = reg->type;
8117 u32 *arg_btf_id = NULL;
8120 if (arg_type == ARG_DONTCARE)
8123 err = check_reg_arg(env, regno, SRC_OP);
8127 if (arg_type == ARG_ANYTHING) {
8128 if (is_pointer_value(env, regno)) {
8129 verbose(env, "R%d leaks addr into helper function\n",
8136 if (type_is_pkt_pointer(type) &&
8137 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
8138 verbose(env, "helper access to the packet is not allowed\n");
8142 if (base_type(arg_type) == ARG_PTR_TO_MAP_VALUE) {
8143 err = resolve_map_arg_type(env, meta, &arg_type);
8148 if (register_is_null(reg) && type_may_be_null(arg_type))
8149 /* A NULL register has a SCALAR_VALUE type, so skip
8152 goto skip_type_check;
8154 /* arg_btf_id and arg_size are in a union. */
8155 if (base_type(arg_type) == ARG_PTR_TO_BTF_ID ||
8156 base_type(arg_type) == ARG_PTR_TO_SPIN_LOCK)
8157 arg_btf_id = fn->arg_btf_id[arg];
8159 err = check_reg_type(env, regno, arg_type, arg_btf_id, meta);
8163 err = check_func_arg_reg_off(env, reg, regno, arg_type);
8168 if (arg_type_is_release(arg_type)) {
8169 if (arg_type_is_dynptr(arg_type)) {
8170 struct bpf_func_state *state = func(env, reg);
8173 /* Only dynptr created on stack can be released, thus
8174 * the get_spi and stack state checks for spilled_ptr
8175 * should only be done before process_dynptr_func for
8178 if (reg->type == PTR_TO_STACK) {
8179 spi = dynptr_get_spi(env, reg);
8180 if (spi < 0 || !state->stack[spi].spilled_ptr.ref_obj_id) {
8181 verbose(env, "arg %d is an unacquired reference\n", regno);
8185 verbose(env, "cannot release unowned const bpf_dynptr\n");
8188 } else if (!reg->ref_obj_id && !register_is_null(reg)) {
8189 verbose(env, "R%d must be referenced when passed to release function\n",
8193 if (meta->release_regno) {
8194 verbose(env, "verifier internal error: more than one release argument\n");
8197 meta->release_regno = regno;
8200 if (reg->ref_obj_id) {
8201 if (meta->ref_obj_id) {
8202 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
8203 regno, reg->ref_obj_id,
8207 meta->ref_obj_id = reg->ref_obj_id;
8210 switch (base_type(arg_type)) {
8211 case ARG_CONST_MAP_PTR:
8212 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
8213 if (meta->map_ptr) {
8214 /* Use map_uid (which is unique id of inner map) to reject:
8215 * inner_map1 = bpf_map_lookup_elem(outer_map, key1)
8216 * inner_map2 = bpf_map_lookup_elem(outer_map, key2)
8217 * if (inner_map1 && inner_map2) {
8218 * timer = bpf_map_lookup_elem(inner_map1);
8220 * // mismatch would have been allowed
8221 * bpf_timer_init(timer, inner_map2);
8224 * Comparing map_ptr is enough to distinguish normal and outer maps.
8226 if (meta->map_ptr != reg->map_ptr ||
8227 meta->map_uid != reg->map_uid) {
8229 "timer pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
8230 meta->map_uid, reg->map_uid);
8234 meta->map_ptr = reg->map_ptr;
8235 meta->map_uid = reg->map_uid;
8237 case ARG_PTR_TO_MAP_KEY:
8238 /* bpf_map_xxx(..., map_ptr, ..., key) call:
8239 * check that [key, key + map->key_size) are within
8240 * stack limits and initialized
8242 if (!meta->map_ptr) {
8243 /* in function declaration map_ptr must come before
8244 * map_key, so that it's verified and known before
8245 * we have to check map_key here. Otherwise it means
8246 * that kernel subsystem misconfigured verifier
8248 verbose(env, "invalid map_ptr to access map->key\n");
8251 err = check_helper_mem_access(env, regno,
8252 meta->map_ptr->key_size, false,
8255 case ARG_PTR_TO_MAP_VALUE:
8256 if (type_may_be_null(arg_type) && register_is_null(reg))
8259 /* bpf_map_xxx(..., map_ptr, ..., value) call:
8260 * check [value, value + map->value_size) validity
8262 if (!meta->map_ptr) {
8263 /* kernel subsystem misconfigured verifier */
8264 verbose(env, "invalid map_ptr to access map->value\n");
8267 meta->raw_mode = arg_type & MEM_UNINIT;
8268 err = check_helper_mem_access(env, regno,
8269 meta->map_ptr->value_size, false,
8272 case ARG_PTR_TO_PERCPU_BTF_ID:
8274 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
8277 meta->ret_btf = reg->btf;
8278 meta->ret_btf_id = reg->btf_id;
8280 case ARG_PTR_TO_SPIN_LOCK:
8281 if (in_rbtree_lock_required_cb(env)) {
8282 verbose(env, "can't spin_{lock,unlock} in rbtree cb\n");
8285 if (meta->func_id == BPF_FUNC_spin_lock) {
8286 err = process_spin_lock(env, regno, true);
8289 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
8290 err = process_spin_lock(env, regno, false);
8294 verbose(env, "verifier internal error\n");
8298 case ARG_PTR_TO_TIMER:
8299 err = process_timer_func(env, regno, meta);
8303 case ARG_PTR_TO_FUNC:
8304 meta->subprogno = reg->subprogno;
8306 case ARG_PTR_TO_MEM:
8307 /* The access to this pointer is only checked when we hit the
8308 * next is_mem_size argument below.
8310 meta->raw_mode = arg_type & MEM_UNINIT;
8311 if (arg_type & MEM_FIXED_SIZE) {
8312 err = check_helper_mem_access(env, regno,
8313 fn->arg_size[arg], false,
8317 case ARG_CONST_SIZE:
8318 err = check_mem_size_reg(env, reg, regno, false, meta);
8320 case ARG_CONST_SIZE_OR_ZERO:
8321 err = check_mem_size_reg(env, reg, regno, true, meta);
8323 case ARG_PTR_TO_DYNPTR:
8324 err = process_dynptr_func(env, regno, insn_idx, arg_type, 0);
8328 case ARG_CONST_ALLOC_SIZE_OR_ZERO:
8329 if (!tnum_is_const(reg->var_off)) {
8330 verbose(env, "R%d is not a known constant'\n",
8334 meta->mem_size = reg->var_off.value;
8335 err = mark_chain_precision(env, regno);
8339 case ARG_PTR_TO_INT:
8340 case ARG_PTR_TO_LONG:
8342 int size = int_ptr_type_to_size(arg_type);
8344 err = check_helper_mem_access(env, regno, size, false, meta);
8347 err = check_ptr_alignment(env, reg, 0, size, true);
8350 case ARG_PTR_TO_CONST_STR:
8352 struct bpf_map *map = reg->map_ptr;
8357 if (!bpf_map_is_rdonly(map)) {
8358 verbose(env, "R%d does not point to a readonly map'\n", regno);
8362 if (!tnum_is_const(reg->var_off)) {
8363 verbose(env, "R%d is not a constant address'\n", regno);
8367 if (!map->ops->map_direct_value_addr) {
8368 verbose(env, "no direct value access support for this map type\n");
8372 err = check_map_access(env, regno, reg->off,
8373 map->value_size - reg->off, false,
8378 map_off = reg->off + reg->var_off.value;
8379 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
8381 verbose(env, "direct value access on string failed\n");
8385 str_ptr = (char *)(long)(map_addr);
8386 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
8387 verbose(env, "string is not zero-terminated\n");
8392 case ARG_PTR_TO_KPTR:
8393 err = process_kptr_func(env, regno, meta);
8402 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
8404 enum bpf_attach_type eatype = env->prog->expected_attach_type;
8405 enum bpf_prog_type type = resolve_prog_type(env->prog);
8407 if (func_id != BPF_FUNC_map_update_elem)
8410 /* It's not possible to get access to a locked struct sock in these
8411 * contexts, so updating is safe.
8414 case BPF_PROG_TYPE_TRACING:
8415 if (eatype == BPF_TRACE_ITER)
8418 case BPF_PROG_TYPE_SOCKET_FILTER:
8419 case BPF_PROG_TYPE_SCHED_CLS:
8420 case BPF_PROG_TYPE_SCHED_ACT:
8421 case BPF_PROG_TYPE_XDP:
8422 case BPF_PROG_TYPE_SK_REUSEPORT:
8423 case BPF_PROG_TYPE_FLOW_DISSECTOR:
8424 case BPF_PROG_TYPE_SK_LOOKUP:
8430 verbose(env, "cannot update sockmap in this context\n");
8434 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
8436 return env->prog->jit_requested &&
8437 bpf_jit_supports_subprog_tailcalls();
8440 static int check_map_func_compatibility(struct bpf_verifier_env *env,
8441 struct bpf_map *map, int func_id)
8446 /* We need a two way check, first is from map perspective ... */
8447 switch (map->map_type) {
8448 case BPF_MAP_TYPE_PROG_ARRAY:
8449 if (func_id != BPF_FUNC_tail_call)
8452 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
8453 if (func_id != BPF_FUNC_perf_event_read &&
8454 func_id != BPF_FUNC_perf_event_output &&
8455 func_id != BPF_FUNC_skb_output &&
8456 func_id != BPF_FUNC_perf_event_read_value &&
8457 func_id != BPF_FUNC_xdp_output)
8460 case BPF_MAP_TYPE_RINGBUF:
8461 if (func_id != BPF_FUNC_ringbuf_output &&
8462 func_id != BPF_FUNC_ringbuf_reserve &&
8463 func_id != BPF_FUNC_ringbuf_query &&
8464 func_id != BPF_FUNC_ringbuf_reserve_dynptr &&
8465 func_id != BPF_FUNC_ringbuf_submit_dynptr &&
8466 func_id != BPF_FUNC_ringbuf_discard_dynptr)
8469 case BPF_MAP_TYPE_USER_RINGBUF:
8470 if (func_id != BPF_FUNC_user_ringbuf_drain)
8473 case BPF_MAP_TYPE_STACK_TRACE:
8474 if (func_id != BPF_FUNC_get_stackid)
8477 case BPF_MAP_TYPE_CGROUP_ARRAY:
8478 if (func_id != BPF_FUNC_skb_under_cgroup &&
8479 func_id != BPF_FUNC_current_task_under_cgroup)
8482 case BPF_MAP_TYPE_CGROUP_STORAGE:
8483 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
8484 if (func_id != BPF_FUNC_get_local_storage)
8487 case BPF_MAP_TYPE_DEVMAP:
8488 case BPF_MAP_TYPE_DEVMAP_HASH:
8489 if (func_id != BPF_FUNC_redirect_map &&
8490 func_id != BPF_FUNC_map_lookup_elem)
8493 /* Restrict bpf side of cpumap and xskmap, open when use-cases
8496 case BPF_MAP_TYPE_CPUMAP:
8497 if (func_id != BPF_FUNC_redirect_map)
8500 case BPF_MAP_TYPE_XSKMAP:
8501 if (func_id != BPF_FUNC_redirect_map &&
8502 func_id != BPF_FUNC_map_lookup_elem)
8505 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
8506 case BPF_MAP_TYPE_HASH_OF_MAPS:
8507 if (func_id != BPF_FUNC_map_lookup_elem)
8510 case BPF_MAP_TYPE_SOCKMAP:
8511 if (func_id != BPF_FUNC_sk_redirect_map &&
8512 func_id != BPF_FUNC_sock_map_update &&
8513 func_id != BPF_FUNC_map_delete_elem &&
8514 func_id != BPF_FUNC_msg_redirect_map &&
8515 func_id != BPF_FUNC_sk_select_reuseport &&
8516 func_id != BPF_FUNC_map_lookup_elem &&
8517 !may_update_sockmap(env, func_id))
8520 case BPF_MAP_TYPE_SOCKHASH:
8521 if (func_id != BPF_FUNC_sk_redirect_hash &&
8522 func_id != BPF_FUNC_sock_hash_update &&
8523 func_id != BPF_FUNC_map_delete_elem &&
8524 func_id != BPF_FUNC_msg_redirect_hash &&
8525 func_id != BPF_FUNC_sk_select_reuseport &&
8526 func_id != BPF_FUNC_map_lookup_elem &&
8527 !may_update_sockmap(env, func_id))
8530 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
8531 if (func_id != BPF_FUNC_sk_select_reuseport)
8534 case BPF_MAP_TYPE_QUEUE:
8535 case BPF_MAP_TYPE_STACK:
8536 if (func_id != BPF_FUNC_map_peek_elem &&
8537 func_id != BPF_FUNC_map_pop_elem &&
8538 func_id != BPF_FUNC_map_push_elem)
8541 case BPF_MAP_TYPE_SK_STORAGE:
8542 if (func_id != BPF_FUNC_sk_storage_get &&
8543 func_id != BPF_FUNC_sk_storage_delete &&
8544 func_id != BPF_FUNC_kptr_xchg)
8547 case BPF_MAP_TYPE_INODE_STORAGE:
8548 if (func_id != BPF_FUNC_inode_storage_get &&
8549 func_id != BPF_FUNC_inode_storage_delete &&
8550 func_id != BPF_FUNC_kptr_xchg)
8553 case BPF_MAP_TYPE_TASK_STORAGE:
8554 if (func_id != BPF_FUNC_task_storage_get &&
8555 func_id != BPF_FUNC_task_storage_delete &&
8556 func_id != BPF_FUNC_kptr_xchg)
8559 case BPF_MAP_TYPE_CGRP_STORAGE:
8560 if (func_id != BPF_FUNC_cgrp_storage_get &&
8561 func_id != BPF_FUNC_cgrp_storage_delete &&
8562 func_id != BPF_FUNC_kptr_xchg)
8565 case BPF_MAP_TYPE_BLOOM_FILTER:
8566 if (func_id != BPF_FUNC_map_peek_elem &&
8567 func_id != BPF_FUNC_map_push_elem)
8574 /* ... and second from the function itself. */
8576 case BPF_FUNC_tail_call:
8577 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
8579 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
8580 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
8584 case BPF_FUNC_perf_event_read:
8585 case BPF_FUNC_perf_event_output:
8586 case BPF_FUNC_perf_event_read_value:
8587 case BPF_FUNC_skb_output:
8588 case BPF_FUNC_xdp_output:
8589 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
8592 case BPF_FUNC_ringbuf_output:
8593 case BPF_FUNC_ringbuf_reserve:
8594 case BPF_FUNC_ringbuf_query:
8595 case BPF_FUNC_ringbuf_reserve_dynptr:
8596 case BPF_FUNC_ringbuf_submit_dynptr:
8597 case BPF_FUNC_ringbuf_discard_dynptr:
8598 if (map->map_type != BPF_MAP_TYPE_RINGBUF)
8601 case BPF_FUNC_user_ringbuf_drain:
8602 if (map->map_type != BPF_MAP_TYPE_USER_RINGBUF)
8605 case BPF_FUNC_get_stackid:
8606 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
8609 case BPF_FUNC_current_task_under_cgroup:
8610 case BPF_FUNC_skb_under_cgroup:
8611 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
8614 case BPF_FUNC_redirect_map:
8615 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
8616 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
8617 map->map_type != BPF_MAP_TYPE_CPUMAP &&
8618 map->map_type != BPF_MAP_TYPE_XSKMAP)
8621 case BPF_FUNC_sk_redirect_map:
8622 case BPF_FUNC_msg_redirect_map:
8623 case BPF_FUNC_sock_map_update:
8624 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
8627 case BPF_FUNC_sk_redirect_hash:
8628 case BPF_FUNC_msg_redirect_hash:
8629 case BPF_FUNC_sock_hash_update:
8630 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
8633 case BPF_FUNC_get_local_storage:
8634 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
8635 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
8638 case BPF_FUNC_sk_select_reuseport:
8639 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
8640 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
8641 map->map_type != BPF_MAP_TYPE_SOCKHASH)
8644 case BPF_FUNC_map_pop_elem:
8645 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
8646 map->map_type != BPF_MAP_TYPE_STACK)
8649 case BPF_FUNC_map_peek_elem:
8650 case BPF_FUNC_map_push_elem:
8651 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
8652 map->map_type != BPF_MAP_TYPE_STACK &&
8653 map->map_type != BPF_MAP_TYPE_BLOOM_FILTER)
8656 case BPF_FUNC_map_lookup_percpu_elem:
8657 if (map->map_type != BPF_MAP_TYPE_PERCPU_ARRAY &&
8658 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
8659 map->map_type != BPF_MAP_TYPE_LRU_PERCPU_HASH)
8662 case BPF_FUNC_sk_storage_get:
8663 case BPF_FUNC_sk_storage_delete:
8664 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
8667 case BPF_FUNC_inode_storage_get:
8668 case BPF_FUNC_inode_storage_delete:
8669 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
8672 case BPF_FUNC_task_storage_get:
8673 case BPF_FUNC_task_storage_delete:
8674 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
8677 case BPF_FUNC_cgrp_storage_get:
8678 case BPF_FUNC_cgrp_storage_delete:
8679 if (map->map_type != BPF_MAP_TYPE_CGRP_STORAGE)
8688 verbose(env, "cannot pass map_type %d into func %s#%d\n",
8689 map->map_type, func_id_name(func_id), func_id);
8693 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
8697 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
8699 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
8701 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
8703 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
8705 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
8708 /* We only support one arg being in raw mode at the moment,
8709 * which is sufficient for the helper functions we have
8715 static bool check_args_pair_invalid(const struct bpf_func_proto *fn, int arg)
8717 bool is_fixed = fn->arg_type[arg] & MEM_FIXED_SIZE;
8718 bool has_size = fn->arg_size[arg] != 0;
8719 bool is_next_size = false;
8721 if (arg + 1 < ARRAY_SIZE(fn->arg_type))
8722 is_next_size = arg_type_is_mem_size(fn->arg_type[arg + 1]);
8724 if (base_type(fn->arg_type[arg]) != ARG_PTR_TO_MEM)
8725 return is_next_size;
8727 return has_size == is_next_size || is_next_size == is_fixed;
8730 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
8732 /* bpf_xxx(..., buf, len) call will access 'len'
8733 * bytes from memory 'buf'. Both arg types need
8734 * to be paired, so make sure there's no buggy
8735 * helper function specification.
8737 if (arg_type_is_mem_size(fn->arg1_type) ||
8738 check_args_pair_invalid(fn, 0) ||
8739 check_args_pair_invalid(fn, 1) ||
8740 check_args_pair_invalid(fn, 2) ||
8741 check_args_pair_invalid(fn, 3) ||
8742 check_args_pair_invalid(fn, 4))
8748 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
8752 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
8753 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_BTF_ID)
8754 return !!fn->arg_btf_id[i];
8755 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_SPIN_LOCK)
8756 return fn->arg_btf_id[i] == BPF_PTR_POISON;
8757 if (base_type(fn->arg_type[i]) != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i] &&
8758 /* arg_btf_id and arg_size are in a union. */
8759 (base_type(fn->arg_type[i]) != ARG_PTR_TO_MEM ||
8760 !(fn->arg_type[i] & MEM_FIXED_SIZE)))
8767 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
8769 return check_raw_mode_ok(fn) &&
8770 check_arg_pair_ok(fn) &&
8771 check_btf_id_ok(fn) ? 0 : -EINVAL;
8774 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
8775 * are now invalid, so turn them into unknown SCALAR_VALUE.
8777 * This also applies to dynptr slices belonging to skb and xdp dynptrs,
8778 * since these slices point to packet data.
8780 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
8782 struct bpf_func_state *state;
8783 struct bpf_reg_state *reg;
8785 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
8786 if (reg_is_pkt_pointer_any(reg) || reg_is_dynptr_slice_pkt(reg))
8787 mark_reg_invalid(env, reg);
8793 BEYOND_PKT_END = -2,
8796 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
8798 struct bpf_func_state *state = vstate->frame[vstate->curframe];
8799 struct bpf_reg_state *reg = &state->regs[regn];
8801 if (reg->type != PTR_TO_PACKET)
8802 /* PTR_TO_PACKET_META is not supported yet */
8805 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
8806 * How far beyond pkt_end it goes is unknown.
8807 * if (!range_open) it's the case of pkt >= pkt_end
8808 * if (range_open) it's the case of pkt > pkt_end
8809 * hence this pointer is at least 1 byte bigger than pkt_end
8812 reg->range = BEYOND_PKT_END;
8814 reg->range = AT_PKT_END;
8817 /* The pointer with the specified id has released its reference to kernel
8818 * resources. Identify all copies of the same pointer and clear the reference.
8820 static int release_reference(struct bpf_verifier_env *env,
8823 struct bpf_func_state *state;
8824 struct bpf_reg_state *reg;
8827 err = release_reference_state(cur_func(env), ref_obj_id);
8831 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
8832 if (reg->ref_obj_id == ref_obj_id)
8833 mark_reg_invalid(env, reg);
8839 static void invalidate_non_owning_refs(struct bpf_verifier_env *env)
8841 struct bpf_func_state *unused;
8842 struct bpf_reg_state *reg;
8844 bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
8845 if (type_is_non_owning_ref(reg->type))
8846 mark_reg_invalid(env, reg);
8850 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
8851 struct bpf_reg_state *regs)
8855 /* after the call registers r0 - r5 were scratched */
8856 for (i = 0; i < CALLER_SAVED_REGS; i++) {
8857 mark_reg_not_init(env, regs, caller_saved[i]);
8858 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
8862 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
8863 struct bpf_func_state *caller,
8864 struct bpf_func_state *callee,
8867 static int set_callee_state(struct bpf_verifier_env *env,
8868 struct bpf_func_state *caller,
8869 struct bpf_func_state *callee, int insn_idx);
8871 static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
8872 int *insn_idx, int subprog,
8873 set_callee_state_fn set_callee_state_cb)
8875 struct bpf_verifier_state *state = env->cur_state;
8876 struct bpf_func_state *caller, *callee;
8879 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
8880 verbose(env, "the call stack of %d frames is too deep\n",
8881 state->curframe + 2);
8885 caller = state->frame[state->curframe];
8886 if (state->frame[state->curframe + 1]) {
8887 verbose(env, "verifier bug. Frame %d already allocated\n",
8888 state->curframe + 1);
8892 err = btf_check_subprog_call(env, subprog, caller->regs);
8895 if (subprog_is_global(env, subprog)) {
8897 verbose(env, "Caller passes invalid args into func#%d\n",
8901 if (env->log.level & BPF_LOG_LEVEL)
8903 "Func#%d is global and valid. Skipping.\n",
8905 clear_caller_saved_regs(env, caller->regs);
8907 /* All global functions return a 64-bit SCALAR_VALUE */
8908 mark_reg_unknown(env, caller->regs, BPF_REG_0);
8909 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
8911 /* continue with next insn after call */
8916 /* set_callee_state is used for direct subprog calls, but we are
8917 * interested in validating only BPF helpers that can call subprogs as
8920 if (set_callee_state_cb != set_callee_state) {
8921 if (bpf_pseudo_kfunc_call(insn) &&
8922 !is_callback_calling_kfunc(insn->imm)) {
8923 verbose(env, "verifier bug: kfunc %s#%d not marked as callback-calling\n",
8924 func_id_name(insn->imm), insn->imm);
8926 } else if (!bpf_pseudo_kfunc_call(insn) &&
8927 !is_callback_calling_function(insn->imm)) { /* helper */
8928 verbose(env, "verifier bug: helper %s#%d not marked as callback-calling\n",
8929 func_id_name(insn->imm), insn->imm);
8934 if (insn->code == (BPF_JMP | BPF_CALL) &&
8935 insn->src_reg == 0 &&
8936 insn->imm == BPF_FUNC_timer_set_callback) {
8937 struct bpf_verifier_state *async_cb;
8939 /* there is no real recursion here. timer callbacks are async */
8940 env->subprog_info[subprog].is_async_cb = true;
8941 async_cb = push_async_cb(env, env->subprog_info[subprog].start,
8942 *insn_idx, subprog);
8945 callee = async_cb->frame[0];
8946 callee->async_entry_cnt = caller->async_entry_cnt + 1;
8948 /* Convert bpf_timer_set_callback() args into timer callback args */
8949 err = set_callee_state_cb(env, caller, callee, *insn_idx);
8953 clear_caller_saved_regs(env, caller->regs);
8954 mark_reg_unknown(env, caller->regs, BPF_REG_0);
8955 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
8956 /* continue with next insn after call */
8960 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
8963 state->frame[state->curframe + 1] = callee;
8965 /* callee cannot access r0, r6 - r9 for reading and has to write
8966 * into its own stack before reading from it.
8967 * callee can read/write into caller's stack
8969 init_func_state(env, callee,
8970 /* remember the callsite, it will be used by bpf_exit */
8971 *insn_idx /* callsite */,
8972 state->curframe + 1 /* frameno within this callchain */,
8973 subprog /* subprog number within this prog */);
8975 /* Transfer references to the callee */
8976 err = copy_reference_state(callee, caller);
8980 err = set_callee_state_cb(env, caller, callee, *insn_idx);
8984 clear_caller_saved_regs(env, caller->regs);
8986 /* only increment it after check_reg_arg() finished */
8989 /* and go analyze first insn of the callee */
8990 *insn_idx = env->subprog_info[subprog].start - 1;
8992 if (env->log.level & BPF_LOG_LEVEL) {
8993 verbose(env, "caller:\n");
8994 print_verifier_state(env, caller, true);
8995 verbose(env, "callee:\n");
8996 print_verifier_state(env, callee, true);
9001 free_func_state(callee);
9002 state->frame[state->curframe + 1] = NULL;
9006 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
9007 struct bpf_func_state *caller,
9008 struct bpf_func_state *callee)
9010 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
9011 * void *callback_ctx, u64 flags);
9012 * callback_fn(struct bpf_map *map, void *key, void *value,
9013 * void *callback_ctx);
9015 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
9017 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
9018 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
9019 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
9021 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
9022 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
9023 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
9025 /* pointer to stack or null */
9026 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
9029 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9033 static int set_callee_state(struct bpf_verifier_env *env,
9034 struct bpf_func_state *caller,
9035 struct bpf_func_state *callee, int insn_idx)
9039 /* copy r1 - r5 args that callee can access. The copy includes parent
9040 * pointers, which connects us up to the liveness chain
9042 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
9043 callee->regs[i] = caller->regs[i];
9047 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
9050 int subprog, target_insn;
9052 target_insn = *insn_idx + insn->imm + 1;
9053 subprog = find_subprog(env, target_insn);
9055 verbose(env, "verifier bug. No program starts at insn %d\n",
9060 return __check_func_call(env, insn, insn_idx, subprog, set_callee_state);
9063 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
9064 struct bpf_func_state *caller,
9065 struct bpf_func_state *callee,
9068 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
9069 struct bpf_map *map;
9072 if (bpf_map_ptr_poisoned(insn_aux)) {
9073 verbose(env, "tail_call abusing map_ptr\n");
9077 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
9078 if (!map->ops->map_set_for_each_callback_args ||
9079 !map->ops->map_for_each_callback) {
9080 verbose(env, "callback function not allowed for map\n");
9084 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
9088 callee->in_callback_fn = true;
9089 callee->callback_ret_range = tnum_range(0, 1);
9093 static int set_loop_callback_state(struct bpf_verifier_env *env,
9094 struct bpf_func_state *caller,
9095 struct bpf_func_state *callee,
9098 /* bpf_loop(u32 nr_loops, void *callback_fn, void *callback_ctx,
9100 * callback_fn(u32 index, void *callback_ctx);
9102 callee->regs[BPF_REG_1].type = SCALAR_VALUE;
9103 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
9106 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
9107 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9108 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9110 callee->in_callback_fn = true;
9111 callee->callback_ret_range = tnum_range(0, 1);
9115 static int set_timer_callback_state(struct bpf_verifier_env *env,
9116 struct bpf_func_state *caller,
9117 struct bpf_func_state *callee,
9120 struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
9122 /* bpf_timer_set_callback(struct bpf_timer *timer, void *callback_fn);
9123 * callback_fn(struct bpf_map *map, void *key, void *value);
9125 callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP;
9126 __mark_reg_known_zero(&callee->regs[BPF_REG_1]);
9127 callee->regs[BPF_REG_1].map_ptr = map_ptr;
9129 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
9130 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
9131 callee->regs[BPF_REG_2].map_ptr = map_ptr;
9133 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
9134 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
9135 callee->regs[BPF_REG_3].map_ptr = map_ptr;
9138 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9139 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9140 callee->in_async_callback_fn = true;
9141 callee->callback_ret_range = tnum_range(0, 1);
9145 static int set_find_vma_callback_state(struct bpf_verifier_env *env,
9146 struct bpf_func_state *caller,
9147 struct bpf_func_state *callee,
9150 /* bpf_find_vma(struct task_struct *task, u64 addr,
9151 * void *callback_fn, void *callback_ctx, u64 flags)
9152 * (callback_fn)(struct task_struct *task,
9153 * struct vm_area_struct *vma, void *callback_ctx);
9155 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
9157 callee->regs[BPF_REG_2].type = PTR_TO_BTF_ID;
9158 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
9159 callee->regs[BPF_REG_2].btf = btf_vmlinux;
9160 callee->regs[BPF_REG_2].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_VMA],
9162 /* pointer to stack or null */
9163 callee->regs[BPF_REG_3] = caller->regs[BPF_REG_4];
9166 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9167 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9168 callee->in_callback_fn = true;
9169 callee->callback_ret_range = tnum_range(0, 1);
9173 static int set_user_ringbuf_callback_state(struct bpf_verifier_env *env,
9174 struct bpf_func_state *caller,
9175 struct bpf_func_state *callee,
9178 /* bpf_user_ringbuf_drain(struct bpf_map *map, void *callback_fn, void
9179 * callback_ctx, u64 flags);
9180 * callback_fn(const struct bpf_dynptr_t* dynptr, void *callback_ctx);
9182 __mark_reg_not_init(env, &callee->regs[BPF_REG_0]);
9183 mark_dynptr_cb_reg(env, &callee->regs[BPF_REG_1], BPF_DYNPTR_TYPE_LOCAL);
9184 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
9187 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
9188 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9189 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9191 callee->in_callback_fn = true;
9192 callee->callback_ret_range = tnum_range(0, 1);
9196 static int set_rbtree_add_callback_state(struct bpf_verifier_env *env,
9197 struct bpf_func_state *caller,
9198 struct bpf_func_state *callee,
9201 /* void bpf_rbtree_add_impl(struct bpf_rb_root *root, struct bpf_rb_node *node,
9202 * bool (less)(struct bpf_rb_node *a, const struct bpf_rb_node *b));
9204 * 'struct bpf_rb_node *node' arg to bpf_rbtree_add_impl is the same PTR_TO_BTF_ID w/ offset
9205 * that 'less' callback args will be receiving. However, 'node' arg was release_reference'd
9206 * by this point, so look at 'root'
9208 struct btf_field *field;
9210 field = reg_find_field_offset(&caller->regs[BPF_REG_1], caller->regs[BPF_REG_1].off,
9212 if (!field || !field->graph_root.value_btf_id)
9215 mark_reg_graph_node(callee->regs, BPF_REG_1, &field->graph_root);
9216 ref_set_non_owning(env, &callee->regs[BPF_REG_1]);
9217 mark_reg_graph_node(callee->regs, BPF_REG_2, &field->graph_root);
9218 ref_set_non_owning(env, &callee->regs[BPF_REG_2]);
9220 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
9221 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9222 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9223 callee->in_callback_fn = true;
9224 callee->callback_ret_range = tnum_range(0, 1);
9228 static bool is_rbtree_lock_required_kfunc(u32 btf_id);
9230 /* Are we currently verifying the callback for a rbtree helper that must
9231 * be called with lock held? If so, no need to complain about unreleased
9234 static bool in_rbtree_lock_required_cb(struct bpf_verifier_env *env)
9236 struct bpf_verifier_state *state = env->cur_state;
9237 struct bpf_insn *insn = env->prog->insnsi;
9238 struct bpf_func_state *callee;
9241 if (!state->curframe)
9244 callee = state->frame[state->curframe];
9246 if (!callee->in_callback_fn)
9249 kfunc_btf_id = insn[callee->callsite].imm;
9250 return is_rbtree_lock_required_kfunc(kfunc_btf_id);
9253 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
9255 struct bpf_verifier_state *state = env->cur_state;
9256 struct bpf_func_state *caller, *callee;
9257 struct bpf_reg_state *r0;
9260 callee = state->frame[state->curframe];
9261 r0 = &callee->regs[BPF_REG_0];
9262 if (r0->type == PTR_TO_STACK) {
9263 /* technically it's ok to return caller's stack pointer
9264 * (or caller's caller's pointer) back to the caller,
9265 * since these pointers are valid. Only current stack
9266 * pointer will be invalid as soon as function exits,
9267 * but let's be conservative
9269 verbose(env, "cannot return stack pointer to the caller\n");
9273 caller = state->frame[state->curframe - 1];
9274 if (callee->in_callback_fn) {
9275 /* enforce R0 return value range [0, 1]. */
9276 struct tnum range = callee->callback_ret_range;
9278 if (r0->type != SCALAR_VALUE) {
9279 verbose(env, "R0 not a scalar value\n");
9282 if (!tnum_in(range, r0->var_off)) {
9283 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
9287 /* return to the caller whatever r0 had in the callee */
9288 caller->regs[BPF_REG_0] = *r0;
9291 /* callback_fn frame should have released its own additions to parent's
9292 * reference state at this point, or check_reference_leak would
9293 * complain, hence it must be the same as the caller. There is no need
9296 if (!callee->in_callback_fn) {
9297 /* Transfer references to the caller */
9298 err = copy_reference_state(caller, callee);
9303 *insn_idx = callee->callsite + 1;
9304 if (env->log.level & BPF_LOG_LEVEL) {
9305 verbose(env, "returning from callee:\n");
9306 print_verifier_state(env, callee, true);
9307 verbose(env, "to caller at %d:\n", *insn_idx);
9308 print_verifier_state(env, caller, true);
9310 /* clear everything in the callee */
9311 free_func_state(callee);
9312 state->frame[state->curframe--] = NULL;
9316 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
9318 struct bpf_call_arg_meta *meta)
9320 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
9322 if (ret_type != RET_INTEGER)
9326 case BPF_FUNC_get_stack:
9327 case BPF_FUNC_get_task_stack:
9328 case BPF_FUNC_probe_read_str:
9329 case BPF_FUNC_probe_read_kernel_str:
9330 case BPF_FUNC_probe_read_user_str:
9331 ret_reg->smax_value = meta->msize_max_value;
9332 ret_reg->s32_max_value = meta->msize_max_value;
9333 ret_reg->smin_value = -MAX_ERRNO;
9334 ret_reg->s32_min_value = -MAX_ERRNO;
9335 reg_bounds_sync(ret_reg);
9337 case BPF_FUNC_get_smp_processor_id:
9338 ret_reg->umax_value = nr_cpu_ids - 1;
9339 ret_reg->u32_max_value = nr_cpu_ids - 1;
9340 ret_reg->smax_value = nr_cpu_ids - 1;
9341 ret_reg->s32_max_value = nr_cpu_ids - 1;
9342 ret_reg->umin_value = 0;
9343 ret_reg->u32_min_value = 0;
9344 ret_reg->smin_value = 0;
9345 ret_reg->s32_min_value = 0;
9346 reg_bounds_sync(ret_reg);
9352 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
9353 int func_id, int insn_idx)
9355 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
9356 struct bpf_map *map = meta->map_ptr;
9358 if (func_id != BPF_FUNC_tail_call &&
9359 func_id != BPF_FUNC_map_lookup_elem &&
9360 func_id != BPF_FUNC_map_update_elem &&
9361 func_id != BPF_FUNC_map_delete_elem &&
9362 func_id != BPF_FUNC_map_push_elem &&
9363 func_id != BPF_FUNC_map_pop_elem &&
9364 func_id != BPF_FUNC_map_peek_elem &&
9365 func_id != BPF_FUNC_for_each_map_elem &&
9366 func_id != BPF_FUNC_redirect_map &&
9367 func_id != BPF_FUNC_map_lookup_percpu_elem)
9371 verbose(env, "kernel subsystem misconfigured verifier\n");
9375 /* In case of read-only, some additional restrictions
9376 * need to be applied in order to prevent altering the
9377 * state of the map from program side.
9379 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
9380 (func_id == BPF_FUNC_map_delete_elem ||
9381 func_id == BPF_FUNC_map_update_elem ||
9382 func_id == BPF_FUNC_map_push_elem ||
9383 func_id == BPF_FUNC_map_pop_elem)) {
9384 verbose(env, "write into map forbidden\n");
9388 if (!BPF_MAP_PTR(aux->map_ptr_state))
9389 bpf_map_ptr_store(aux, meta->map_ptr,
9390 !meta->map_ptr->bypass_spec_v1);
9391 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
9392 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
9393 !meta->map_ptr->bypass_spec_v1);
9398 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
9399 int func_id, int insn_idx)
9401 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
9402 struct bpf_reg_state *regs = cur_regs(env), *reg;
9403 struct bpf_map *map = meta->map_ptr;
9407 if (func_id != BPF_FUNC_tail_call)
9409 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
9410 verbose(env, "kernel subsystem misconfigured verifier\n");
9414 reg = ®s[BPF_REG_3];
9415 val = reg->var_off.value;
9416 max = map->max_entries;
9418 if (!(register_is_const(reg) && val < max)) {
9419 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
9423 err = mark_chain_precision(env, BPF_REG_3);
9426 if (bpf_map_key_unseen(aux))
9427 bpf_map_key_store(aux, val);
9428 else if (!bpf_map_key_poisoned(aux) &&
9429 bpf_map_key_immediate(aux) != val)
9430 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
9434 static int check_reference_leak(struct bpf_verifier_env *env)
9436 struct bpf_func_state *state = cur_func(env);
9437 bool refs_lingering = false;
9440 if (state->frameno && !state->in_callback_fn)
9443 for (i = 0; i < state->acquired_refs; i++) {
9444 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno)
9446 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
9447 state->refs[i].id, state->refs[i].insn_idx);
9448 refs_lingering = true;
9450 return refs_lingering ? -EINVAL : 0;
9453 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
9454 struct bpf_reg_state *regs)
9456 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
9457 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
9458 struct bpf_map *fmt_map = fmt_reg->map_ptr;
9459 struct bpf_bprintf_data data = {};
9460 int err, fmt_map_off, num_args;
9464 /* data must be an array of u64 */
9465 if (data_len_reg->var_off.value % 8)
9467 num_args = data_len_reg->var_off.value / 8;
9469 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
9470 * and map_direct_value_addr is set.
9472 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
9473 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
9476 verbose(env, "verifier bug\n");
9479 fmt = (char *)(long)fmt_addr + fmt_map_off;
9481 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
9482 * can focus on validating the format specifiers.
9484 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, num_args, &data);
9486 verbose(env, "Invalid format string\n");
9491 static int check_get_func_ip(struct bpf_verifier_env *env)
9493 enum bpf_prog_type type = resolve_prog_type(env->prog);
9494 int func_id = BPF_FUNC_get_func_ip;
9496 if (type == BPF_PROG_TYPE_TRACING) {
9497 if (!bpf_prog_has_trampoline(env->prog)) {
9498 verbose(env, "func %s#%d supported only for fentry/fexit/fmod_ret programs\n",
9499 func_id_name(func_id), func_id);
9503 } else if (type == BPF_PROG_TYPE_KPROBE) {
9507 verbose(env, "func %s#%d not supported for program type %d\n",
9508 func_id_name(func_id), func_id, type);
9512 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
9514 return &env->insn_aux_data[env->insn_idx];
9517 static bool loop_flag_is_zero(struct bpf_verifier_env *env)
9519 struct bpf_reg_state *regs = cur_regs(env);
9520 struct bpf_reg_state *reg = ®s[BPF_REG_4];
9521 bool reg_is_null = register_is_null(reg);
9524 mark_chain_precision(env, BPF_REG_4);
9529 static void update_loop_inline_state(struct bpf_verifier_env *env, u32 subprogno)
9531 struct bpf_loop_inline_state *state = &cur_aux(env)->loop_inline_state;
9533 if (!state->initialized) {
9534 state->initialized = 1;
9535 state->fit_for_inline = loop_flag_is_zero(env);
9536 state->callback_subprogno = subprogno;
9540 if (!state->fit_for_inline)
9543 state->fit_for_inline = (loop_flag_is_zero(env) &&
9544 state->callback_subprogno == subprogno);
9547 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
9550 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
9551 const struct bpf_func_proto *fn = NULL;
9552 enum bpf_return_type ret_type;
9553 enum bpf_type_flag ret_flag;
9554 struct bpf_reg_state *regs;
9555 struct bpf_call_arg_meta meta;
9556 int insn_idx = *insn_idx_p;
9558 int i, err, func_id;
9560 /* find function prototype */
9561 func_id = insn->imm;
9562 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
9563 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
9568 if (env->ops->get_func_proto)
9569 fn = env->ops->get_func_proto(func_id, env->prog);
9571 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
9576 /* eBPF programs must be GPL compatible to use GPL-ed functions */
9577 if (!env->prog->gpl_compatible && fn->gpl_only) {
9578 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
9582 if (fn->allowed && !fn->allowed(env->prog)) {
9583 verbose(env, "helper call is not allowed in probe\n");
9587 if (!env->prog->aux->sleepable && fn->might_sleep) {
9588 verbose(env, "helper call might sleep in a non-sleepable prog\n");
9592 /* With LD_ABS/IND some JITs save/restore skb from r1. */
9593 changes_data = bpf_helper_changes_pkt_data(fn->func);
9594 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
9595 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
9596 func_id_name(func_id), func_id);
9600 memset(&meta, 0, sizeof(meta));
9601 meta.pkt_access = fn->pkt_access;
9603 err = check_func_proto(fn, func_id);
9605 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
9606 func_id_name(func_id), func_id);
9610 if (env->cur_state->active_rcu_lock) {
9611 if (fn->might_sleep) {
9612 verbose(env, "sleepable helper %s#%d in rcu_read_lock region\n",
9613 func_id_name(func_id), func_id);
9617 if (env->prog->aux->sleepable && is_storage_get_function(func_id))
9618 env->insn_aux_data[insn_idx].storage_get_func_atomic = true;
9621 meta.func_id = func_id;
9623 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
9624 err = check_func_arg(env, i, &meta, fn, insn_idx);
9629 err = record_func_map(env, &meta, func_id, insn_idx);
9633 err = record_func_key(env, &meta, func_id, insn_idx);
9637 /* Mark slots with STACK_MISC in case of raw mode, stack offset
9638 * is inferred from register state.
9640 for (i = 0; i < meta.access_size; i++) {
9641 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
9642 BPF_WRITE, -1, false, false);
9647 regs = cur_regs(env);
9649 if (meta.release_regno) {
9651 /* This can only be set for PTR_TO_STACK, as CONST_PTR_TO_DYNPTR cannot
9652 * be released by any dynptr helper. Hence, unmark_stack_slots_dynptr
9653 * is safe to do directly.
9655 if (arg_type_is_dynptr(fn->arg_type[meta.release_regno - BPF_REG_1])) {
9656 if (regs[meta.release_regno].type == CONST_PTR_TO_DYNPTR) {
9657 verbose(env, "verifier internal error: CONST_PTR_TO_DYNPTR cannot be released\n");
9660 err = unmark_stack_slots_dynptr(env, ®s[meta.release_regno]);
9661 } else if (meta.ref_obj_id) {
9662 err = release_reference(env, meta.ref_obj_id);
9663 } else if (register_is_null(®s[meta.release_regno])) {
9664 /* meta.ref_obj_id can only be 0 if register that is meant to be
9665 * released is NULL, which must be > R0.
9670 verbose(env, "func %s#%d reference has not been acquired before\n",
9671 func_id_name(func_id), func_id);
9677 case BPF_FUNC_tail_call:
9678 err = check_reference_leak(env);
9680 verbose(env, "tail_call would lead to reference leak\n");
9684 case BPF_FUNC_get_local_storage:
9685 /* check that flags argument in get_local_storage(map, flags) is 0,
9686 * this is required because get_local_storage() can't return an error.
9688 if (!register_is_null(®s[BPF_REG_2])) {
9689 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
9693 case BPF_FUNC_for_each_map_elem:
9694 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9695 set_map_elem_callback_state);
9697 case BPF_FUNC_timer_set_callback:
9698 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9699 set_timer_callback_state);
9701 case BPF_FUNC_find_vma:
9702 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9703 set_find_vma_callback_state);
9705 case BPF_FUNC_snprintf:
9706 err = check_bpf_snprintf_call(env, regs);
9709 update_loop_inline_state(env, meta.subprogno);
9710 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9711 set_loop_callback_state);
9713 case BPF_FUNC_dynptr_from_mem:
9714 if (regs[BPF_REG_1].type != PTR_TO_MAP_VALUE) {
9715 verbose(env, "Unsupported reg type %s for bpf_dynptr_from_mem data\n",
9716 reg_type_str(env, regs[BPF_REG_1].type));
9720 case BPF_FUNC_set_retval:
9721 if (prog_type == BPF_PROG_TYPE_LSM &&
9722 env->prog->expected_attach_type == BPF_LSM_CGROUP) {
9723 if (!env->prog->aux->attach_func_proto->type) {
9724 /* Make sure programs that attach to void
9725 * hooks don't try to modify return value.
9727 verbose(env, "BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
9732 case BPF_FUNC_dynptr_data:
9734 struct bpf_reg_state *reg;
9737 reg = get_dynptr_arg_reg(env, fn, regs);
9742 if (meta.dynptr_id) {
9743 verbose(env, "verifier internal error: meta.dynptr_id already set\n");
9746 if (meta.ref_obj_id) {
9747 verbose(env, "verifier internal error: meta.ref_obj_id already set\n");
9751 id = dynptr_id(env, reg);
9753 verbose(env, "verifier internal error: failed to obtain dynptr id\n");
9757 ref_obj_id = dynptr_ref_obj_id(env, reg);
9758 if (ref_obj_id < 0) {
9759 verbose(env, "verifier internal error: failed to obtain dynptr ref_obj_id\n");
9763 meta.dynptr_id = id;
9764 meta.ref_obj_id = ref_obj_id;
9768 case BPF_FUNC_dynptr_write:
9770 enum bpf_dynptr_type dynptr_type;
9771 struct bpf_reg_state *reg;
9773 reg = get_dynptr_arg_reg(env, fn, regs);
9777 dynptr_type = dynptr_get_type(env, reg);
9778 if (dynptr_type == BPF_DYNPTR_TYPE_INVALID)
9781 if (dynptr_type == BPF_DYNPTR_TYPE_SKB)
9782 /* this will trigger clear_all_pkt_pointers(), which will
9783 * invalidate all dynptr slices associated with the skb
9785 changes_data = true;
9789 case BPF_FUNC_user_ringbuf_drain:
9790 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9791 set_user_ringbuf_callback_state);
9798 /* reset caller saved regs */
9799 for (i = 0; i < CALLER_SAVED_REGS; i++) {
9800 mark_reg_not_init(env, regs, caller_saved[i]);
9801 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
9804 /* helper call returns 64-bit value. */
9805 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
9807 /* update return register (already marked as written above) */
9808 ret_type = fn->ret_type;
9809 ret_flag = type_flag(ret_type);
9811 switch (base_type(ret_type)) {
9813 /* sets type to SCALAR_VALUE */
9814 mark_reg_unknown(env, regs, BPF_REG_0);
9817 regs[BPF_REG_0].type = NOT_INIT;
9819 case RET_PTR_TO_MAP_VALUE:
9820 /* There is no offset yet applied, variable or fixed */
9821 mark_reg_known_zero(env, regs, BPF_REG_0);
9822 /* remember map_ptr, so that check_map_access()
9823 * can check 'value_size' boundary of memory access
9824 * to map element returned from bpf_map_lookup_elem()
9826 if (meta.map_ptr == NULL) {
9828 "kernel subsystem misconfigured verifier\n");
9831 regs[BPF_REG_0].map_ptr = meta.map_ptr;
9832 regs[BPF_REG_0].map_uid = meta.map_uid;
9833 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE | ret_flag;
9834 if (!type_may_be_null(ret_type) &&
9835 btf_record_has_field(meta.map_ptr->record, BPF_SPIN_LOCK)) {
9836 regs[BPF_REG_0].id = ++env->id_gen;
9839 case RET_PTR_TO_SOCKET:
9840 mark_reg_known_zero(env, regs, BPF_REG_0);
9841 regs[BPF_REG_0].type = PTR_TO_SOCKET | ret_flag;
9843 case RET_PTR_TO_SOCK_COMMON:
9844 mark_reg_known_zero(env, regs, BPF_REG_0);
9845 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON | ret_flag;
9847 case RET_PTR_TO_TCP_SOCK:
9848 mark_reg_known_zero(env, regs, BPF_REG_0);
9849 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK | ret_flag;
9851 case RET_PTR_TO_MEM:
9852 mark_reg_known_zero(env, regs, BPF_REG_0);
9853 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
9854 regs[BPF_REG_0].mem_size = meta.mem_size;
9856 case RET_PTR_TO_MEM_OR_BTF_ID:
9858 const struct btf_type *t;
9860 mark_reg_known_zero(env, regs, BPF_REG_0);
9861 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
9862 if (!btf_type_is_struct(t)) {
9864 const struct btf_type *ret;
9867 /* resolve the type size of ksym. */
9868 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
9870 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
9871 verbose(env, "unable to resolve the size of type '%s': %ld\n",
9872 tname, PTR_ERR(ret));
9875 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
9876 regs[BPF_REG_0].mem_size = tsize;
9878 /* MEM_RDONLY may be carried from ret_flag, but it
9879 * doesn't apply on PTR_TO_BTF_ID. Fold it, otherwise
9880 * it will confuse the check of PTR_TO_BTF_ID in
9881 * check_mem_access().
9883 ret_flag &= ~MEM_RDONLY;
9885 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
9886 regs[BPF_REG_0].btf = meta.ret_btf;
9887 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
9891 case RET_PTR_TO_BTF_ID:
9893 struct btf *ret_btf;
9896 mark_reg_known_zero(env, regs, BPF_REG_0);
9897 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
9898 if (func_id == BPF_FUNC_kptr_xchg) {
9899 ret_btf = meta.kptr_field->kptr.btf;
9900 ret_btf_id = meta.kptr_field->kptr.btf_id;
9901 if (!btf_is_kernel(ret_btf))
9902 regs[BPF_REG_0].type |= MEM_ALLOC;
9904 if (fn->ret_btf_id == BPF_PTR_POISON) {
9905 verbose(env, "verifier internal error:");
9906 verbose(env, "func %s has non-overwritten BPF_PTR_POISON return type\n",
9907 func_id_name(func_id));
9910 ret_btf = btf_vmlinux;
9911 ret_btf_id = *fn->ret_btf_id;
9913 if (ret_btf_id == 0) {
9914 verbose(env, "invalid return type %u of func %s#%d\n",
9915 base_type(ret_type), func_id_name(func_id),
9919 regs[BPF_REG_0].btf = ret_btf;
9920 regs[BPF_REG_0].btf_id = ret_btf_id;
9924 verbose(env, "unknown return type %u of func %s#%d\n",
9925 base_type(ret_type), func_id_name(func_id), func_id);
9929 if (type_may_be_null(regs[BPF_REG_0].type))
9930 regs[BPF_REG_0].id = ++env->id_gen;
9932 if (helper_multiple_ref_obj_use(func_id, meta.map_ptr)) {
9933 verbose(env, "verifier internal error: func %s#%d sets ref_obj_id more than once\n",
9934 func_id_name(func_id), func_id);
9938 if (is_dynptr_ref_function(func_id))
9939 regs[BPF_REG_0].dynptr_id = meta.dynptr_id;
9941 if (is_ptr_cast_function(func_id) || is_dynptr_ref_function(func_id)) {
9942 /* For release_reference() */
9943 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
9944 } else if (is_acquire_function(func_id, meta.map_ptr)) {
9945 int id = acquire_reference_state(env, insn_idx);
9949 /* For mark_ptr_or_null_reg() */
9950 regs[BPF_REG_0].id = id;
9951 /* For release_reference() */
9952 regs[BPF_REG_0].ref_obj_id = id;
9955 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
9957 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
9961 if ((func_id == BPF_FUNC_get_stack ||
9962 func_id == BPF_FUNC_get_task_stack) &&
9963 !env->prog->has_callchain_buf) {
9964 const char *err_str;
9966 #ifdef CONFIG_PERF_EVENTS
9967 err = get_callchain_buffers(sysctl_perf_event_max_stack);
9968 err_str = "cannot get callchain buffer for func %s#%d\n";
9971 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
9974 verbose(env, err_str, func_id_name(func_id), func_id);
9978 env->prog->has_callchain_buf = true;
9981 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
9982 env->prog->call_get_stack = true;
9984 if (func_id == BPF_FUNC_get_func_ip) {
9985 if (check_get_func_ip(env))
9987 env->prog->call_get_func_ip = true;
9991 clear_all_pkt_pointers(env);
9995 /* mark_btf_func_reg_size() is used when the reg size is determined by
9996 * the BTF func_proto's return value size and argument.
9998 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
10001 struct bpf_reg_state *reg = &cur_regs(env)[regno];
10003 if (regno == BPF_REG_0) {
10004 /* Function return value */
10005 reg->live |= REG_LIVE_WRITTEN;
10006 reg->subreg_def = reg_size == sizeof(u64) ?
10007 DEF_NOT_SUBREG : env->insn_idx + 1;
10009 /* Function argument */
10010 if (reg_size == sizeof(u64)) {
10011 mark_insn_zext(env, reg);
10012 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
10014 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
10019 static bool is_kfunc_acquire(struct bpf_kfunc_call_arg_meta *meta)
10021 return meta->kfunc_flags & KF_ACQUIRE;
10024 static bool is_kfunc_release(struct bpf_kfunc_call_arg_meta *meta)
10026 return meta->kfunc_flags & KF_RELEASE;
10029 static bool is_kfunc_trusted_args(struct bpf_kfunc_call_arg_meta *meta)
10031 return (meta->kfunc_flags & KF_TRUSTED_ARGS) || is_kfunc_release(meta);
10034 static bool is_kfunc_sleepable(struct bpf_kfunc_call_arg_meta *meta)
10036 return meta->kfunc_flags & KF_SLEEPABLE;
10039 static bool is_kfunc_destructive(struct bpf_kfunc_call_arg_meta *meta)
10041 return meta->kfunc_flags & KF_DESTRUCTIVE;
10044 static bool is_kfunc_rcu(struct bpf_kfunc_call_arg_meta *meta)
10046 return meta->kfunc_flags & KF_RCU;
10049 static bool __kfunc_param_match_suffix(const struct btf *btf,
10050 const struct btf_param *arg,
10051 const char *suffix)
10053 int suffix_len = strlen(suffix), len;
10054 const char *param_name;
10056 /* In the future, this can be ported to use BTF tagging */
10057 param_name = btf_name_by_offset(btf, arg->name_off);
10058 if (str_is_empty(param_name))
10060 len = strlen(param_name);
10061 if (len < suffix_len)
10063 param_name += len - suffix_len;
10064 return !strncmp(param_name, suffix, suffix_len);
10067 static bool is_kfunc_arg_mem_size(const struct btf *btf,
10068 const struct btf_param *arg,
10069 const struct bpf_reg_state *reg)
10071 const struct btf_type *t;
10073 t = btf_type_skip_modifiers(btf, arg->type, NULL);
10074 if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
10077 return __kfunc_param_match_suffix(btf, arg, "__sz");
10080 static bool is_kfunc_arg_const_mem_size(const struct btf *btf,
10081 const struct btf_param *arg,
10082 const struct bpf_reg_state *reg)
10084 const struct btf_type *t;
10086 t = btf_type_skip_modifiers(btf, arg->type, NULL);
10087 if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
10090 return __kfunc_param_match_suffix(btf, arg, "__szk");
10093 static bool is_kfunc_arg_optional(const struct btf *btf, const struct btf_param *arg)
10095 return __kfunc_param_match_suffix(btf, arg, "__opt");
10098 static bool is_kfunc_arg_constant(const struct btf *btf, const struct btf_param *arg)
10100 return __kfunc_param_match_suffix(btf, arg, "__k");
10103 static bool is_kfunc_arg_ignore(const struct btf *btf, const struct btf_param *arg)
10105 return __kfunc_param_match_suffix(btf, arg, "__ign");
10108 static bool is_kfunc_arg_alloc_obj(const struct btf *btf, const struct btf_param *arg)
10110 return __kfunc_param_match_suffix(btf, arg, "__alloc");
10113 static bool is_kfunc_arg_uninit(const struct btf *btf, const struct btf_param *arg)
10115 return __kfunc_param_match_suffix(btf, arg, "__uninit");
10118 static bool is_kfunc_arg_refcounted_kptr(const struct btf *btf, const struct btf_param *arg)
10120 return __kfunc_param_match_suffix(btf, arg, "__refcounted_kptr");
10123 static bool is_kfunc_arg_scalar_with_name(const struct btf *btf,
10124 const struct btf_param *arg,
10127 int len, target_len = strlen(name);
10128 const char *param_name;
10130 param_name = btf_name_by_offset(btf, arg->name_off);
10131 if (str_is_empty(param_name))
10133 len = strlen(param_name);
10134 if (len != target_len)
10136 if (strcmp(param_name, name))
10144 KF_ARG_LIST_HEAD_ID,
10145 KF_ARG_LIST_NODE_ID,
10150 BTF_ID_LIST(kf_arg_btf_ids)
10151 BTF_ID(struct, bpf_dynptr_kern)
10152 BTF_ID(struct, bpf_list_head)
10153 BTF_ID(struct, bpf_list_node)
10154 BTF_ID(struct, bpf_rb_root)
10155 BTF_ID(struct, bpf_rb_node)
10157 static bool __is_kfunc_ptr_arg_type(const struct btf *btf,
10158 const struct btf_param *arg, int type)
10160 const struct btf_type *t;
10163 t = btf_type_skip_modifiers(btf, arg->type, NULL);
10166 if (!btf_type_is_ptr(t))
10168 t = btf_type_skip_modifiers(btf, t->type, &res_id);
10171 return btf_types_are_same(btf, res_id, btf_vmlinux, kf_arg_btf_ids[type]);
10174 static bool is_kfunc_arg_dynptr(const struct btf *btf, const struct btf_param *arg)
10176 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_DYNPTR_ID);
10179 static bool is_kfunc_arg_list_head(const struct btf *btf, const struct btf_param *arg)
10181 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_HEAD_ID);
10184 static bool is_kfunc_arg_list_node(const struct btf *btf, const struct btf_param *arg)
10186 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_NODE_ID);
10189 static bool is_kfunc_arg_rbtree_root(const struct btf *btf, const struct btf_param *arg)
10191 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RB_ROOT_ID);
10194 static bool is_kfunc_arg_rbtree_node(const struct btf *btf, const struct btf_param *arg)
10196 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RB_NODE_ID);
10199 static bool is_kfunc_arg_callback(struct bpf_verifier_env *env, const struct btf *btf,
10200 const struct btf_param *arg)
10202 const struct btf_type *t;
10204 t = btf_type_resolve_func_ptr(btf, arg->type, NULL);
10211 /* Returns true if struct is composed of scalars, 4 levels of nesting allowed */
10212 static bool __btf_type_is_scalar_struct(struct bpf_verifier_env *env,
10213 const struct btf *btf,
10214 const struct btf_type *t, int rec)
10216 const struct btf_type *member_type;
10217 const struct btf_member *member;
10220 if (!btf_type_is_struct(t))
10223 for_each_member(i, t, member) {
10224 const struct btf_array *array;
10226 member_type = btf_type_skip_modifiers(btf, member->type, NULL);
10227 if (btf_type_is_struct(member_type)) {
10229 verbose(env, "max struct nesting depth exceeded\n");
10232 if (!__btf_type_is_scalar_struct(env, btf, member_type, rec + 1))
10236 if (btf_type_is_array(member_type)) {
10237 array = btf_array(member_type);
10238 if (!array->nelems)
10240 member_type = btf_type_skip_modifiers(btf, array->type, NULL);
10241 if (!btf_type_is_scalar(member_type))
10245 if (!btf_type_is_scalar(member_type))
10251 enum kfunc_ptr_arg_type {
10253 KF_ARG_PTR_TO_ALLOC_BTF_ID, /* Allocated object */
10254 KF_ARG_PTR_TO_REFCOUNTED_KPTR, /* Refcounted local kptr */
10255 KF_ARG_PTR_TO_DYNPTR,
10256 KF_ARG_PTR_TO_ITER,
10257 KF_ARG_PTR_TO_LIST_HEAD,
10258 KF_ARG_PTR_TO_LIST_NODE,
10259 KF_ARG_PTR_TO_BTF_ID, /* Also covers reg2btf_ids conversions */
10261 KF_ARG_PTR_TO_MEM_SIZE, /* Size derived from next argument, skip it */
10262 KF_ARG_PTR_TO_CALLBACK,
10263 KF_ARG_PTR_TO_RB_ROOT,
10264 KF_ARG_PTR_TO_RB_NODE,
10267 enum special_kfunc_type {
10268 KF_bpf_obj_new_impl,
10269 KF_bpf_obj_drop_impl,
10270 KF_bpf_refcount_acquire_impl,
10271 KF_bpf_list_push_front_impl,
10272 KF_bpf_list_push_back_impl,
10273 KF_bpf_list_pop_front,
10274 KF_bpf_list_pop_back,
10275 KF_bpf_cast_to_kern_ctx,
10276 KF_bpf_rdonly_cast,
10277 KF_bpf_rcu_read_lock,
10278 KF_bpf_rcu_read_unlock,
10279 KF_bpf_rbtree_remove,
10280 KF_bpf_rbtree_add_impl,
10281 KF_bpf_rbtree_first,
10282 KF_bpf_dynptr_from_skb,
10283 KF_bpf_dynptr_from_xdp,
10284 KF_bpf_dynptr_slice,
10285 KF_bpf_dynptr_slice_rdwr,
10286 KF_bpf_dynptr_clone,
10289 BTF_SET_START(special_kfunc_set)
10290 BTF_ID(func, bpf_obj_new_impl)
10291 BTF_ID(func, bpf_obj_drop_impl)
10292 BTF_ID(func, bpf_refcount_acquire_impl)
10293 BTF_ID(func, bpf_list_push_front_impl)
10294 BTF_ID(func, bpf_list_push_back_impl)
10295 BTF_ID(func, bpf_list_pop_front)
10296 BTF_ID(func, bpf_list_pop_back)
10297 BTF_ID(func, bpf_cast_to_kern_ctx)
10298 BTF_ID(func, bpf_rdonly_cast)
10299 BTF_ID(func, bpf_rbtree_remove)
10300 BTF_ID(func, bpf_rbtree_add_impl)
10301 BTF_ID(func, bpf_rbtree_first)
10302 BTF_ID(func, bpf_dynptr_from_skb)
10303 BTF_ID(func, bpf_dynptr_from_xdp)
10304 BTF_ID(func, bpf_dynptr_slice)
10305 BTF_ID(func, bpf_dynptr_slice_rdwr)
10306 BTF_ID(func, bpf_dynptr_clone)
10307 BTF_SET_END(special_kfunc_set)
10309 BTF_ID_LIST(special_kfunc_list)
10310 BTF_ID(func, bpf_obj_new_impl)
10311 BTF_ID(func, bpf_obj_drop_impl)
10312 BTF_ID(func, bpf_refcount_acquire_impl)
10313 BTF_ID(func, bpf_list_push_front_impl)
10314 BTF_ID(func, bpf_list_push_back_impl)
10315 BTF_ID(func, bpf_list_pop_front)
10316 BTF_ID(func, bpf_list_pop_back)
10317 BTF_ID(func, bpf_cast_to_kern_ctx)
10318 BTF_ID(func, bpf_rdonly_cast)
10319 BTF_ID(func, bpf_rcu_read_lock)
10320 BTF_ID(func, bpf_rcu_read_unlock)
10321 BTF_ID(func, bpf_rbtree_remove)
10322 BTF_ID(func, bpf_rbtree_add_impl)
10323 BTF_ID(func, bpf_rbtree_first)
10324 BTF_ID(func, bpf_dynptr_from_skb)
10325 BTF_ID(func, bpf_dynptr_from_xdp)
10326 BTF_ID(func, bpf_dynptr_slice)
10327 BTF_ID(func, bpf_dynptr_slice_rdwr)
10328 BTF_ID(func, bpf_dynptr_clone)
10330 static bool is_kfunc_ret_null(struct bpf_kfunc_call_arg_meta *meta)
10332 if (meta->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl] &&
10333 meta->arg_owning_ref) {
10337 return meta->kfunc_flags & KF_RET_NULL;
10340 static bool is_kfunc_bpf_rcu_read_lock(struct bpf_kfunc_call_arg_meta *meta)
10342 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_lock];
10345 static bool is_kfunc_bpf_rcu_read_unlock(struct bpf_kfunc_call_arg_meta *meta)
10347 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_unlock];
10350 static enum kfunc_ptr_arg_type
10351 get_kfunc_ptr_arg_type(struct bpf_verifier_env *env,
10352 struct bpf_kfunc_call_arg_meta *meta,
10353 const struct btf_type *t, const struct btf_type *ref_t,
10354 const char *ref_tname, const struct btf_param *args,
10355 int argno, int nargs)
10357 u32 regno = argno + 1;
10358 struct bpf_reg_state *regs = cur_regs(env);
10359 struct bpf_reg_state *reg = ®s[regno];
10360 bool arg_mem_size = false;
10362 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx])
10363 return KF_ARG_PTR_TO_CTX;
10365 /* In this function, we verify the kfunc's BTF as per the argument type,
10366 * leaving the rest of the verification with respect to the register
10367 * type to our caller. When a set of conditions hold in the BTF type of
10368 * arguments, we resolve it to a known kfunc_ptr_arg_type.
10370 if (btf_get_prog_ctx_type(&env->log, meta->btf, t, resolve_prog_type(env->prog), argno))
10371 return KF_ARG_PTR_TO_CTX;
10373 if (is_kfunc_arg_alloc_obj(meta->btf, &args[argno]))
10374 return KF_ARG_PTR_TO_ALLOC_BTF_ID;
10376 if (is_kfunc_arg_refcounted_kptr(meta->btf, &args[argno]))
10377 return KF_ARG_PTR_TO_REFCOUNTED_KPTR;
10379 if (is_kfunc_arg_dynptr(meta->btf, &args[argno]))
10380 return KF_ARG_PTR_TO_DYNPTR;
10382 if (is_kfunc_arg_iter(meta, argno))
10383 return KF_ARG_PTR_TO_ITER;
10385 if (is_kfunc_arg_list_head(meta->btf, &args[argno]))
10386 return KF_ARG_PTR_TO_LIST_HEAD;
10388 if (is_kfunc_arg_list_node(meta->btf, &args[argno]))
10389 return KF_ARG_PTR_TO_LIST_NODE;
10391 if (is_kfunc_arg_rbtree_root(meta->btf, &args[argno]))
10392 return KF_ARG_PTR_TO_RB_ROOT;
10394 if (is_kfunc_arg_rbtree_node(meta->btf, &args[argno]))
10395 return KF_ARG_PTR_TO_RB_NODE;
10397 if ((base_type(reg->type) == PTR_TO_BTF_ID || reg2btf_ids[base_type(reg->type)])) {
10398 if (!btf_type_is_struct(ref_t)) {
10399 verbose(env, "kernel function %s args#%d pointer type %s %s is not supported\n",
10400 meta->func_name, argno, btf_type_str(ref_t), ref_tname);
10403 return KF_ARG_PTR_TO_BTF_ID;
10406 if (is_kfunc_arg_callback(env, meta->btf, &args[argno]))
10407 return KF_ARG_PTR_TO_CALLBACK;
10410 if (argno + 1 < nargs &&
10411 (is_kfunc_arg_mem_size(meta->btf, &args[argno + 1], ®s[regno + 1]) ||
10412 is_kfunc_arg_const_mem_size(meta->btf, &args[argno + 1], ®s[regno + 1])))
10413 arg_mem_size = true;
10415 /* This is the catch all argument type of register types supported by
10416 * check_helper_mem_access. However, we only allow when argument type is
10417 * pointer to scalar, or struct composed (recursively) of scalars. When
10418 * arg_mem_size is true, the pointer can be void *.
10420 if (!btf_type_is_scalar(ref_t) && !__btf_type_is_scalar_struct(env, meta->btf, ref_t, 0) &&
10421 (arg_mem_size ? !btf_type_is_void(ref_t) : 1)) {
10422 verbose(env, "arg#%d pointer type %s %s must point to %sscalar, or struct with scalar\n",
10423 argno, btf_type_str(ref_t), ref_tname, arg_mem_size ? "void, " : "");
10426 return arg_mem_size ? KF_ARG_PTR_TO_MEM_SIZE : KF_ARG_PTR_TO_MEM;
10429 static int process_kf_arg_ptr_to_btf_id(struct bpf_verifier_env *env,
10430 struct bpf_reg_state *reg,
10431 const struct btf_type *ref_t,
10432 const char *ref_tname, u32 ref_id,
10433 struct bpf_kfunc_call_arg_meta *meta,
10436 const struct btf_type *reg_ref_t;
10437 bool strict_type_match = false;
10438 const struct btf *reg_btf;
10439 const char *reg_ref_tname;
10442 if (base_type(reg->type) == PTR_TO_BTF_ID) {
10443 reg_btf = reg->btf;
10444 reg_ref_id = reg->btf_id;
10446 reg_btf = btf_vmlinux;
10447 reg_ref_id = *reg2btf_ids[base_type(reg->type)];
10450 /* Enforce strict type matching for calls to kfuncs that are acquiring
10451 * or releasing a reference, or are no-cast aliases. We do _not_
10452 * enforce strict matching for plain KF_TRUSTED_ARGS kfuncs by default,
10453 * as we want to enable BPF programs to pass types that are bitwise
10454 * equivalent without forcing them to explicitly cast with something
10455 * like bpf_cast_to_kern_ctx().
10457 * For example, say we had a type like the following:
10459 * struct bpf_cpumask {
10460 * cpumask_t cpumask;
10461 * refcount_t usage;
10464 * Note that as specified in <linux/cpumask.h>, cpumask_t is typedef'ed
10465 * to a struct cpumask, so it would be safe to pass a struct
10466 * bpf_cpumask * to a kfunc expecting a struct cpumask *.
10468 * The philosophy here is similar to how we allow scalars of different
10469 * types to be passed to kfuncs as long as the size is the same. The
10470 * only difference here is that we're simply allowing
10471 * btf_struct_ids_match() to walk the struct at the 0th offset, and
10474 if (is_kfunc_acquire(meta) ||
10475 (is_kfunc_release(meta) && reg->ref_obj_id) ||
10476 btf_type_ids_nocast_alias(&env->log, reg_btf, reg_ref_id, meta->btf, ref_id))
10477 strict_type_match = true;
10479 WARN_ON_ONCE(is_kfunc_trusted_args(meta) && reg->off);
10481 reg_ref_t = btf_type_skip_modifiers(reg_btf, reg_ref_id, ®_ref_id);
10482 reg_ref_tname = btf_name_by_offset(reg_btf, reg_ref_t->name_off);
10483 if (!btf_struct_ids_match(&env->log, reg_btf, reg_ref_id, reg->off, meta->btf, ref_id, strict_type_match)) {
10484 verbose(env, "kernel function %s args#%d expected pointer to %s %s but R%d has a pointer to %s %s\n",
10485 meta->func_name, argno, btf_type_str(ref_t), ref_tname, argno + 1,
10486 btf_type_str(reg_ref_t), reg_ref_tname);
10492 static int ref_set_non_owning(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
10494 struct bpf_verifier_state *state = env->cur_state;
10495 struct btf_record *rec = reg_btf_record(reg);
10497 if (!state->active_lock.ptr) {
10498 verbose(env, "verifier internal error: ref_set_non_owning w/o active lock\n");
10502 if (type_flag(reg->type) & NON_OWN_REF) {
10503 verbose(env, "verifier internal error: NON_OWN_REF already set\n");
10507 reg->type |= NON_OWN_REF;
10508 if (rec->refcount_off >= 0)
10509 reg->type |= MEM_RCU;
10514 static int ref_convert_owning_non_owning(struct bpf_verifier_env *env, u32 ref_obj_id)
10516 struct bpf_func_state *state, *unused;
10517 struct bpf_reg_state *reg;
10520 state = cur_func(env);
10523 verbose(env, "verifier internal error: ref_obj_id is zero for "
10524 "owning -> non-owning conversion\n");
10528 for (i = 0; i < state->acquired_refs; i++) {
10529 if (state->refs[i].id != ref_obj_id)
10532 /* Clear ref_obj_id here so release_reference doesn't clobber
10535 bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
10536 if (reg->ref_obj_id == ref_obj_id) {
10537 reg->ref_obj_id = 0;
10538 ref_set_non_owning(env, reg);
10544 verbose(env, "verifier internal error: ref state missing for ref_obj_id\n");
10548 /* Implementation details:
10550 * Each register points to some region of memory, which we define as an
10551 * allocation. Each allocation may embed a bpf_spin_lock which protects any
10552 * special BPF objects (bpf_list_head, bpf_rb_root, etc.) part of the same
10553 * allocation. The lock and the data it protects are colocated in the same
10556 * Hence, everytime a register holds a pointer value pointing to such
10557 * allocation, the verifier preserves a unique reg->id for it.
10559 * The verifier remembers the lock 'ptr' and the lock 'id' whenever
10560 * bpf_spin_lock is called.
10562 * To enable this, lock state in the verifier captures two values:
10563 * active_lock.ptr = Register's type specific pointer
10564 * active_lock.id = A unique ID for each register pointer value
10566 * Currently, PTR_TO_MAP_VALUE and PTR_TO_BTF_ID | MEM_ALLOC are the two
10567 * supported register types.
10569 * The active_lock.ptr in case of map values is the reg->map_ptr, and in case of
10570 * allocated objects is the reg->btf pointer.
10572 * The active_lock.id is non-unique for maps supporting direct_value_addr, as we
10573 * can establish the provenance of the map value statically for each distinct
10574 * lookup into such maps. They always contain a single map value hence unique
10575 * IDs for each pseudo load pessimizes the algorithm and rejects valid programs.
10577 * So, in case of global variables, they use array maps with max_entries = 1,
10578 * hence their active_lock.ptr becomes map_ptr and id = 0 (since they all point
10579 * into the same map value as max_entries is 1, as described above).
10581 * In case of inner map lookups, the inner map pointer has same map_ptr as the
10582 * outer map pointer (in verifier context), but each lookup into an inner map
10583 * assigns a fresh reg->id to the lookup, so while lookups into distinct inner
10584 * maps from the same outer map share the same map_ptr as active_lock.ptr, they
10585 * will get different reg->id assigned to each lookup, hence different
10588 * In case of allocated objects, active_lock.ptr is the reg->btf, and the
10589 * reg->id is a unique ID preserved after the NULL pointer check on the pointer
10590 * returned from bpf_obj_new. Each allocation receives a new reg->id.
10592 static int check_reg_allocation_locked(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
10597 switch ((int)reg->type) {
10598 case PTR_TO_MAP_VALUE:
10599 ptr = reg->map_ptr;
10601 case PTR_TO_BTF_ID | MEM_ALLOC:
10605 verbose(env, "verifier internal error: unknown reg type for lock check\n");
10610 if (!env->cur_state->active_lock.ptr)
10612 if (env->cur_state->active_lock.ptr != ptr ||
10613 env->cur_state->active_lock.id != id) {
10614 verbose(env, "held lock and object are not in the same allocation\n");
10620 static bool is_bpf_list_api_kfunc(u32 btf_id)
10622 return btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
10623 btf_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
10624 btf_id == special_kfunc_list[KF_bpf_list_pop_front] ||
10625 btf_id == special_kfunc_list[KF_bpf_list_pop_back];
10628 static bool is_bpf_rbtree_api_kfunc(u32 btf_id)
10630 return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl] ||
10631 btf_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
10632 btf_id == special_kfunc_list[KF_bpf_rbtree_first];
10635 static bool is_bpf_graph_api_kfunc(u32 btf_id)
10637 return is_bpf_list_api_kfunc(btf_id) || is_bpf_rbtree_api_kfunc(btf_id) ||
10638 btf_id == special_kfunc_list[KF_bpf_refcount_acquire_impl];
10641 static bool is_callback_calling_kfunc(u32 btf_id)
10643 return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl];
10646 static bool is_rbtree_lock_required_kfunc(u32 btf_id)
10648 return is_bpf_rbtree_api_kfunc(btf_id);
10651 static bool check_kfunc_is_graph_root_api(struct bpf_verifier_env *env,
10652 enum btf_field_type head_field_type,
10657 switch (head_field_type) {
10658 case BPF_LIST_HEAD:
10659 ret = is_bpf_list_api_kfunc(kfunc_btf_id);
10662 ret = is_bpf_rbtree_api_kfunc(kfunc_btf_id);
10665 verbose(env, "verifier internal error: unexpected graph root argument type %s\n",
10666 btf_field_type_name(head_field_type));
10671 verbose(env, "verifier internal error: %s head arg for unknown kfunc\n",
10672 btf_field_type_name(head_field_type));
10676 static bool check_kfunc_is_graph_node_api(struct bpf_verifier_env *env,
10677 enum btf_field_type node_field_type,
10682 switch (node_field_type) {
10683 case BPF_LIST_NODE:
10684 ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
10685 kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_back_impl]);
10688 ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
10689 kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl]);
10692 verbose(env, "verifier internal error: unexpected graph node argument type %s\n",
10693 btf_field_type_name(node_field_type));
10698 verbose(env, "verifier internal error: %s node arg for unknown kfunc\n",
10699 btf_field_type_name(node_field_type));
10704 __process_kf_arg_ptr_to_graph_root(struct bpf_verifier_env *env,
10705 struct bpf_reg_state *reg, u32 regno,
10706 struct bpf_kfunc_call_arg_meta *meta,
10707 enum btf_field_type head_field_type,
10708 struct btf_field **head_field)
10710 const char *head_type_name;
10711 struct btf_field *field;
10712 struct btf_record *rec;
10715 if (meta->btf != btf_vmlinux) {
10716 verbose(env, "verifier internal error: unexpected btf mismatch in kfunc call\n");
10720 if (!check_kfunc_is_graph_root_api(env, head_field_type, meta->func_id))
10723 head_type_name = btf_field_type_name(head_field_type);
10724 if (!tnum_is_const(reg->var_off)) {
10726 "R%d doesn't have constant offset. %s has to be at the constant offset\n",
10727 regno, head_type_name);
10731 rec = reg_btf_record(reg);
10732 head_off = reg->off + reg->var_off.value;
10733 field = btf_record_find(rec, head_off, head_field_type);
10735 verbose(env, "%s not found at offset=%u\n", head_type_name, head_off);
10739 /* All functions require bpf_list_head to be protected using a bpf_spin_lock */
10740 if (check_reg_allocation_locked(env, reg)) {
10741 verbose(env, "bpf_spin_lock at off=%d must be held for %s\n",
10742 rec->spin_lock_off, head_type_name);
10747 verbose(env, "verifier internal error: repeating %s arg\n", head_type_name);
10750 *head_field = field;
10754 static int process_kf_arg_ptr_to_list_head(struct bpf_verifier_env *env,
10755 struct bpf_reg_state *reg, u32 regno,
10756 struct bpf_kfunc_call_arg_meta *meta)
10758 return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, BPF_LIST_HEAD,
10759 &meta->arg_list_head.field);
10762 static int process_kf_arg_ptr_to_rbtree_root(struct bpf_verifier_env *env,
10763 struct bpf_reg_state *reg, u32 regno,
10764 struct bpf_kfunc_call_arg_meta *meta)
10766 return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, BPF_RB_ROOT,
10767 &meta->arg_rbtree_root.field);
10771 __process_kf_arg_ptr_to_graph_node(struct bpf_verifier_env *env,
10772 struct bpf_reg_state *reg, u32 regno,
10773 struct bpf_kfunc_call_arg_meta *meta,
10774 enum btf_field_type head_field_type,
10775 enum btf_field_type node_field_type,
10776 struct btf_field **node_field)
10778 const char *node_type_name;
10779 const struct btf_type *et, *t;
10780 struct btf_field *field;
10783 if (meta->btf != btf_vmlinux) {
10784 verbose(env, "verifier internal error: unexpected btf mismatch in kfunc call\n");
10788 if (!check_kfunc_is_graph_node_api(env, node_field_type, meta->func_id))
10791 node_type_name = btf_field_type_name(node_field_type);
10792 if (!tnum_is_const(reg->var_off)) {
10794 "R%d doesn't have constant offset. %s has to be at the constant offset\n",
10795 regno, node_type_name);
10799 node_off = reg->off + reg->var_off.value;
10800 field = reg_find_field_offset(reg, node_off, node_field_type);
10801 if (!field || field->offset != node_off) {
10802 verbose(env, "%s not found at offset=%u\n", node_type_name, node_off);
10806 field = *node_field;
10808 et = btf_type_by_id(field->graph_root.btf, field->graph_root.value_btf_id);
10809 t = btf_type_by_id(reg->btf, reg->btf_id);
10810 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, 0, field->graph_root.btf,
10811 field->graph_root.value_btf_id, true)) {
10812 verbose(env, "operation on %s expects arg#1 %s at offset=%d "
10813 "in struct %s, but arg is at offset=%d in struct %s\n",
10814 btf_field_type_name(head_field_type),
10815 btf_field_type_name(node_field_type),
10816 field->graph_root.node_offset,
10817 btf_name_by_offset(field->graph_root.btf, et->name_off),
10818 node_off, btf_name_by_offset(reg->btf, t->name_off));
10821 meta->arg_btf = reg->btf;
10822 meta->arg_btf_id = reg->btf_id;
10824 if (node_off != field->graph_root.node_offset) {
10825 verbose(env, "arg#1 offset=%d, but expected %s at offset=%d in struct %s\n",
10826 node_off, btf_field_type_name(node_field_type),
10827 field->graph_root.node_offset,
10828 btf_name_by_offset(field->graph_root.btf, et->name_off));
10835 static int process_kf_arg_ptr_to_list_node(struct bpf_verifier_env *env,
10836 struct bpf_reg_state *reg, u32 regno,
10837 struct bpf_kfunc_call_arg_meta *meta)
10839 return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta,
10840 BPF_LIST_HEAD, BPF_LIST_NODE,
10841 &meta->arg_list_head.field);
10844 static int process_kf_arg_ptr_to_rbtree_node(struct bpf_verifier_env *env,
10845 struct bpf_reg_state *reg, u32 regno,
10846 struct bpf_kfunc_call_arg_meta *meta)
10848 return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta,
10849 BPF_RB_ROOT, BPF_RB_NODE,
10850 &meta->arg_rbtree_root.field);
10853 static int check_kfunc_args(struct bpf_verifier_env *env, struct bpf_kfunc_call_arg_meta *meta,
10856 const char *func_name = meta->func_name, *ref_tname;
10857 const struct btf *btf = meta->btf;
10858 const struct btf_param *args;
10859 struct btf_record *rec;
10863 args = (const struct btf_param *)(meta->func_proto + 1);
10864 nargs = btf_type_vlen(meta->func_proto);
10865 if (nargs > MAX_BPF_FUNC_REG_ARGS) {
10866 verbose(env, "Function %s has %d > %d args\n", func_name, nargs,
10867 MAX_BPF_FUNC_REG_ARGS);
10871 /* Check that BTF function arguments match actual types that the
10874 for (i = 0; i < nargs; i++) {
10875 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[i + 1];
10876 const struct btf_type *t, *ref_t, *resolve_ret;
10877 enum bpf_arg_type arg_type = ARG_DONTCARE;
10878 u32 regno = i + 1, ref_id, type_size;
10879 bool is_ret_buf_sz = false;
10882 t = btf_type_skip_modifiers(btf, args[i].type, NULL);
10884 if (is_kfunc_arg_ignore(btf, &args[i]))
10887 if (btf_type_is_scalar(t)) {
10888 if (reg->type != SCALAR_VALUE) {
10889 verbose(env, "R%d is not a scalar\n", regno);
10893 if (is_kfunc_arg_constant(meta->btf, &args[i])) {
10894 if (meta->arg_constant.found) {
10895 verbose(env, "verifier internal error: only one constant argument permitted\n");
10898 if (!tnum_is_const(reg->var_off)) {
10899 verbose(env, "R%d must be a known constant\n", regno);
10902 ret = mark_chain_precision(env, regno);
10905 meta->arg_constant.found = true;
10906 meta->arg_constant.value = reg->var_off.value;
10907 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdonly_buf_size")) {
10908 meta->r0_rdonly = true;
10909 is_ret_buf_sz = true;
10910 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdwr_buf_size")) {
10911 is_ret_buf_sz = true;
10914 if (is_ret_buf_sz) {
10915 if (meta->r0_size) {
10916 verbose(env, "2 or more rdonly/rdwr_buf_size parameters for kfunc");
10920 if (!tnum_is_const(reg->var_off)) {
10921 verbose(env, "R%d is not a const\n", regno);
10925 meta->r0_size = reg->var_off.value;
10926 ret = mark_chain_precision(env, regno);
10933 if (!btf_type_is_ptr(t)) {
10934 verbose(env, "Unrecognized arg#%d type %s\n", i, btf_type_str(t));
10938 if ((is_kfunc_trusted_args(meta) || is_kfunc_rcu(meta)) &&
10939 (register_is_null(reg) || type_may_be_null(reg->type))) {
10940 verbose(env, "Possibly NULL pointer passed to trusted arg%d\n", i);
10944 if (reg->ref_obj_id) {
10945 if (is_kfunc_release(meta) && meta->ref_obj_id) {
10946 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
10947 regno, reg->ref_obj_id,
10951 meta->ref_obj_id = reg->ref_obj_id;
10952 if (is_kfunc_release(meta))
10953 meta->release_regno = regno;
10956 ref_t = btf_type_skip_modifiers(btf, t->type, &ref_id);
10957 ref_tname = btf_name_by_offset(btf, ref_t->name_off);
10959 kf_arg_type = get_kfunc_ptr_arg_type(env, meta, t, ref_t, ref_tname, args, i, nargs);
10960 if (kf_arg_type < 0)
10961 return kf_arg_type;
10963 switch (kf_arg_type) {
10964 case KF_ARG_PTR_TO_ALLOC_BTF_ID:
10965 case KF_ARG_PTR_TO_BTF_ID:
10966 if (!is_kfunc_trusted_args(meta) && !is_kfunc_rcu(meta))
10969 if (!is_trusted_reg(reg)) {
10970 if (!is_kfunc_rcu(meta)) {
10971 verbose(env, "R%d must be referenced or trusted\n", regno);
10974 if (!is_rcu_reg(reg)) {
10975 verbose(env, "R%d must be a rcu pointer\n", regno);
10981 case KF_ARG_PTR_TO_CTX:
10982 /* Trusted arguments have the same offset checks as release arguments */
10983 arg_type |= OBJ_RELEASE;
10985 case KF_ARG_PTR_TO_DYNPTR:
10986 case KF_ARG_PTR_TO_ITER:
10987 case KF_ARG_PTR_TO_LIST_HEAD:
10988 case KF_ARG_PTR_TO_LIST_NODE:
10989 case KF_ARG_PTR_TO_RB_ROOT:
10990 case KF_ARG_PTR_TO_RB_NODE:
10991 case KF_ARG_PTR_TO_MEM:
10992 case KF_ARG_PTR_TO_MEM_SIZE:
10993 case KF_ARG_PTR_TO_CALLBACK:
10994 case KF_ARG_PTR_TO_REFCOUNTED_KPTR:
10995 /* Trusted by default */
11002 if (is_kfunc_release(meta) && reg->ref_obj_id)
11003 arg_type |= OBJ_RELEASE;
11004 ret = check_func_arg_reg_off(env, reg, regno, arg_type);
11008 switch (kf_arg_type) {
11009 case KF_ARG_PTR_TO_CTX:
11010 if (reg->type != PTR_TO_CTX) {
11011 verbose(env, "arg#%d expected pointer to ctx, but got %s\n", i, btf_type_str(t));
11015 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
11016 ret = get_kern_ctx_btf_id(&env->log, resolve_prog_type(env->prog));
11019 meta->ret_btf_id = ret;
11022 case KF_ARG_PTR_TO_ALLOC_BTF_ID:
11023 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11024 verbose(env, "arg#%d expected pointer to allocated object\n", i);
11027 if (!reg->ref_obj_id) {
11028 verbose(env, "allocated object must be referenced\n");
11031 if (meta->btf == btf_vmlinux &&
11032 meta->func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
11033 meta->arg_btf = reg->btf;
11034 meta->arg_btf_id = reg->btf_id;
11037 case KF_ARG_PTR_TO_DYNPTR:
11039 enum bpf_arg_type dynptr_arg_type = ARG_PTR_TO_DYNPTR;
11040 int clone_ref_obj_id = 0;
11042 if (reg->type != PTR_TO_STACK &&
11043 reg->type != CONST_PTR_TO_DYNPTR) {
11044 verbose(env, "arg#%d expected pointer to stack or dynptr_ptr\n", i);
11048 if (reg->type == CONST_PTR_TO_DYNPTR)
11049 dynptr_arg_type |= MEM_RDONLY;
11051 if (is_kfunc_arg_uninit(btf, &args[i]))
11052 dynptr_arg_type |= MEM_UNINIT;
11054 if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) {
11055 dynptr_arg_type |= DYNPTR_TYPE_SKB;
11056 } else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_xdp]) {
11057 dynptr_arg_type |= DYNPTR_TYPE_XDP;
11058 } else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_clone] &&
11059 (dynptr_arg_type & MEM_UNINIT)) {
11060 enum bpf_dynptr_type parent_type = meta->initialized_dynptr.type;
11062 if (parent_type == BPF_DYNPTR_TYPE_INVALID) {
11063 verbose(env, "verifier internal error: no dynptr type for parent of clone\n");
11067 dynptr_arg_type |= (unsigned int)get_dynptr_type_flag(parent_type);
11068 clone_ref_obj_id = meta->initialized_dynptr.ref_obj_id;
11069 if (dynptr_type_refcounted(parent_type) && !clone_ref_obj_id) {
11070 verbose(env, "verifier internal error: missing ref obj id for parent of clone\n");
11075 ret = process_dynptr_func(env, regno, insn_idx, dynptr_arg_type, clone_ref_obj_id);
11079 if (!(dynptr_arg_type & MEM_UNINIT)) {
11080 int id = dynptr_id(env, reg);
11083 verbose(env, "verifier internal error: failed to obtain dynptr id\n");
11086 meta->initialized_dynptr.id = id;
11087 meta->initialized_dynptr.type = dynptr_get_type(env, reg);
11088 meta->initialized_dynptr.ref_obj_id = dynptr_ref_obj_id(env, reg);
11093 case KF_ARG_PTR_TO_ITER:
11094 ret = process_iter_arg(env, regno, insn_idx, meta);
11098 case KF_ARG_PTR_TO_LIST_HEAD:
11099 if (reg->type != PTR_TO_MAP_VALUE &&
11100 reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11101 verbose(env, "arg#%d expected pointer to map value or allocated object\n", i);
11104 if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
11105 verbose(env, "allocated object must be referenced\n");
11108 ret = process_kf_arg_ptr_to_list_head(env, reg, regno, meta);
11112 case KF_ARG_PTR_TO_RB_ROOT:
11113 if (reg->type != PTR_TO_MAP_VALUE &&
11114 reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11115 verbose(env, "arg#%d expected pointer to map value or allocated object\n", i);
11118 if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
11119 verbose(env, "allocated object must be referenced\n");
11122 ret = process_kf_arg_ptr_to_rbtree_root(env, reg, regno, meta);
11126 case KF_ARG_PTR_TO_LIST_NODE:
11127 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11128 verbose(env, "arg#%d expected pointer to allocated object\n", i);
11131 if (!reg->ref_obj_id) {
11132 verbose(env, "allocated object must be referenced\n");
11135 ret = process_kf_arg_ptr_to_list_node(env, reg, regno, meta);
11139 case KF_ARG_PTR_TO_RB_NODE:
11140 if (meta->func_id == special_kfunc_list[KF_bpf_rbtree_remove]) {
11141 if (!type_is_non_owning_ref(reg->type) || reg->ref_obj_id) {
11142 verbose(env, "rbtree_remove node input must be non-owning ref\n");
11145 if (in_rbtree_lock_required_cb(env)) {
11146 verbose(env, "rbtree_remove not allowed in rbtree cb\n");
11150 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11151 verbose(env, "arg#%d expected pointer to allocated object\n", i);
11154 if (!reg->ref_obj_id) {
11155 verbose(env, "allocated object must be referenced\n");
11160 ret = process_kf_arg_ptr_to_rbtree_node(env, reg, regno, meta);
11164 case KF_ARG_PTR_TO_BTF_ID:
11165 /* Only base_type is checked, further checks are done here */
11166 if ((base_type(reg->type) != PTR_TO_BTF_ID ||
11167 (bpf_type_has_unsafe_modifiers(reg->type) && !is_rcu_reg(reg))) &&
11168 !reg2btf_ids[base_type(reg->type)]) {
11169 verbose(env, "arg#%d is %s ", i, reg_type_str(env, reg->type));
11170 verbose(env, "expected %s or socket\n",
11171 reg_type_str(env, base_type(reg->type) |
11172 (type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS)));
11175 ret = process_kf_arg_ptr_to_btf_id(env, reg, ref_t, ref_tname, ref_id, meta, i);
11179 case KF_ARG_PTR_TO_MEM:
11180 resolve_ret = btf_resolve_size(btf, ref_t, &type_size);
11181 if (IS_ERR(resolve_ret)) {
11182 verbose(env, "arg#%d reference type('%s %s') size cannot be determined: %ld\n",
11183 i, btf_type_str(ref_t), ref_tname, PTR_ERR(resolve_ret));
11186 ret = check_mem_reg(env, reg, regno, type_size);
11190 case KF_ARG_PTR_TO_MEM_SIZE:
11192 struct bpf_reg_state *buff_reg = ®s[regno];
11193 const struct btf_param *buff_arg = &args[i];
11194 struct bpf_reg_state *size_reg = ®s[regno + 1];
11195 const struct btf_param *size_arg = &args[i + 1];
11197 if (!register_is_null(buff_reg) || !is_kfunc_arg_optional(meta->btf, buff_arg)) {
11198 ret = check_kfunc_mem_size_reg(env, size_reg, regno + 1);
11200 verbose(env, "arg#%d arg#%d memory, len pair leads to invalid memory access\n", i, i + 1);
11205 if (is_kfunc_arg_const_mem_size(meta->btf, size_arg, size_reg)) {
11206 if (meta->arg_constant.found) {
11207 verbose(env, "verifier internal error: only one constant argument permitted\n");
11210 if (!tnum_is_const(size_reg->var_off)) {
11211 verbose(env, "R%d must be a known constant\n", regno + 1);
11214 meta->arg_constant.found = true;
11215 meta->arg_constant.value = size_reg->var_off.value;
11218 /* Skip next '__sz' or '__szk' argument */
11222 case KF_ARG_PTR_TO_CALLBACK:
11223 if (reg->type != PTR_TO_FUNC) {
11224 verbose(env, "arg%d expected pointer to func\n", i);
11227 meta->subprogno = reg->subprogno;
11229 case KF_ARG_PTR_TO_REFCOUNTED_KPTR:
11230 if (!type_is_ptr_alloc_obj(reg->type)) {
11231 verbose(env, "arg#%d is neither owning or non-owning ref\n", i);
11234 if (!type_is_non_owning_ref(reg->type))
11235 meta->arg_owning_ref = true;
11237 rec = reg_btf_record(reg);
11239 verbose(env, "verifier internal error: Couldn't find btf_record\n");
11243 if (rec->refcount_off < 0) {
11244 verbose(env, "arg#%d doesn't point to a type with bpf_refcount field\n", i);
11248 meta->arg_btf = reg->btf;
11249 meta->arg_btf_id = reg->btf_id;
11254 if (is_kfunc_release(meta) && !meta->release_regno) {
11255 verbose(env, "release kernel function %s expects refcounted PTR_TO_BTF_ID\n",
11263 static int fetch_kfunc_meta(struct bpf_verifier_env *env,
11264 struct bpf_insn *insn,
11265 struct bpf_kfunc_call_arg_meta *meta,
11266 const char **kfunc_name)
11268 const struct btf_type *func, *func_proto;
11269 u32 func_id, *kfunc_flags;
11270 const char *func_name;
11271 struct btf *desc_btf;
11274 *kfunc_name = NULL;
11279 desc_btf = find_kfunc_desc_btf(env, insn->off);
11280 if (IS_ERR(desc_btf))
11281 return PTR_ERR(desc_btf);
11283 func_id = insn->imm;
11284 func = btf_type_by_id(desc_btf, func_id);
11285 func_name = btf_name_by_offset(desc_btf, func->name_off);
11287 *kfunc_name = func_name;
11288 func_proto = btf_type_by_id(desc_btf, func->type);
11290 kfunc_flags = btf_kfunc_id_set_contains(desc_btf, func_id, env->prog);
11291 if (!kfunc_flags) {
11295 memset(meta, 0, sizeof(*meta));
11296 meta->btf = desc_btf;
11297 meta->func_id = func_id;
11298 meta->kfunc_flags = *kfunc_flags;
11299 meta->func_proto = func_proto;
11300 meta->func_name = func_name;
11305 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
11308 const struct btf_type *t, *ptr_type;
11309 u32 i, nargs, ptr_type_id, release_ref_obj_id;
11310 struct bpf_reg_state *regs = cur_regs(env);
11311 const char *func_name, *ptr_type_name;
11312 bool sleepable, rcu_lock, rcu_unlock;
11313 struct bpf_kfunc_call_arg_meta meta;
11314 struct bpf_insn_aux_data *insn_aux;
11315 int err, insn_idx = *insn_idx_p;
11316 const struct btf_param *args;
11317 const struct btf_type *ret_t;
11318 struct btf *desc_btf;
11320 /* skip for now, but return error when we find this in fixup_kfunc_call */
11324 err = fetch_kfunc_meta(env, insn, &meta, &func_name);
11325 if (err == -EACCES && func_name)
11326 verbose(env, "calling kernel function %s is not allowed\n", func_name);
11329 desc_btf = meta.btf;
11330 insn_aux = &env->insn_aux_data[insn_idx];
11332 insn_aux->is_iter_next = is_iter_next_kfunc(&meta);
11334 if (is_kfunc_destructive(&meta) && !capable(CAP_SYS_BOOT)) {
11335 verbose(env, "destructive kfunc calls require CAP_SYS_BOOT capability\n");
11339 sleepable = is_kfunc_sleepable(&meta);
11340 if (sleepable && !env->prog->aux->sleepable) {
11341 verbose(env, "program must be sleepable to call sleepable kfunc %s\n", func_name);
11345 rcu_lock = is_kfunc_bpf_rcu_read_lock(&meta);
11346 rcu_unlock = is_kfunc_bpf_rcu_read_unlock(&meta);
11348 if (env->cur_state->active_rcu_lock) {
11349 struct bpf_func_state *state;
11350 struct bpf_reg_state *reg;
11352 if (in_rbtree_lock_required_cb(env) && (rcu_lock || rcu_unlock)) {
11353 verbose(env, "Calling bpf_rcu_read_{lock,unlock} in unnecessary rbtree callback\n");
11358 verbose(env, "nested rcu read lock (kernel function %s)\n", func_name);
11360 } else if (rcu_unlock) {
11361 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
11362 if (reg->type & MEM_RCU) {
11363 reg->type &= ~(MEM_RCU | PTR_MAYBE_NULL);
11364 reg->type |= PTR_UNTRUSTED;
11367 env->cur_state->active_rcu_lock = false;
11368 } else if (sleepable) {
11369 verbose(env, "kernel func %s is sleepable within rcu_read_lock region\n", func_name);
11372 } else if (rcu_lock) {
11373 env->cur_state->active_rcu_lock = true;
11374 } else if (rcu_unlock) {
11375 verbose(env, "unmatched rcu read unlock (kernel function %s)\n", func_name);
11379 /* Check the arguments */
11380 err = check_kfunc_args(env, &meta, insn_idx);
11383 /* In case of release function, we get register number of refcounted
11384 * PTR_TO_BTF_ID in bpf_kfunc_arg_meta, do the release now.
11386 if (meta.release_regno) {
11387 err = release_reference(env, regs[meta.release_regno].ref_obj_id);
11389 verbose(env, "kfunc %s#%d reference has not been acquired before\n",
11390 func_name, meta.func_id);
11395 if (meta.func_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
11396 meta.func_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
11397 meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
11398 release_ref_obj_id = regs[BPF_REG_2].ref_obj_id;
11399 insn_aux->insert_off = regs[BPF_REG_2].off;
11400 insn_aux->kptr_struct_meta = btf_find_struct_meta(meta.arg_btf, meta.arg_btf_id);
11401 err = ref_convert_owning_non_owning(env, release_ref_obj_id);
11403 verbose(env, "kfunc %s#%d conversion of owning ref to non-owning failed\n",
11404 func_name, meta.func_id);
11408 err = release_reference(env, release_ref_obj_id);
11410 verbose(env, "kfunc %s#%d reference has not been acquired before\n",
11411 func_name, meta.func_id);
11416 if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
11417 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
11418 set_rbtree_add_callback_state);
11420 verbose(env, "kfunc %s#%d failed callback verification\n",
11421 func_name, meta.func_id);
11426 for (i = 0; i < CALLER_SAVED_REGS; i++)
11427 mark_reg_not_init(env, regs, caller_saved[i]);
11429 /* Check return type */
11430 t = btf_type_skip_modifiers(desc_btf, meta.func_proto->type, NULL);
11432 if (is_kfunc_acquire(&meta) && !btf_type_is_struct_ptr(meta.btf, t)) {
11433 /* Only exception is bpf_obj_new_impl */
11434 if (meta.btf != btf_vmlinux ||
11435 (meta.func_id != special_kfunc_list[KF_bpf_obj_new_impl] &&
11436 meta.func_id != special_kfunc_list[KF_bpf_refcount_acquire_impl])) {
11437 verbose(env, "acquire kernel function does not return PTR_TO_BTF_ID\n");
11442 if (btf_type_is_scalar(t)) {
11443 mark_reg_unknown(env, regs, BPF_REG_0);
11444 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
11445 } else if (btf_type_is_ptr(t)) {
11446 ptr_type = btf_type_skip_modifiers(desc_btf, t->type, &ptr_type_id);
11448 if (meta.btf == btf_vmlinux && btf_id_set_contains(&special_kfunc_set, meta.func_id)) {
11449 if (meta.func_id == special_kfunc_list[KF_bpf_obj_new_impl]) {
11450 struct btf *ret_btf;
11453 if (unlikely(!bpf_global_ma_set))
11456 if (((u64)(u32)meta.arg_constant.value) != meta.arg_constant.value) {
11457 verbose(env, "local type ID argument must be in range [0, U32_MAX]\n");
11461 ret_btf = env->prog->aux->btf;
11462 ret_btf_id = meta.arg_constant.value;
11464 /* This may be NULL due to user not supplying a BTF */
11466 verbose(env, "bpf_obj_new requires prog BTF\n");
11470 ret_t = btf_type_by_id(ret_btf, ret_btf_id);
11471 if (!ret_t || !__btf_type_is_struct(ret_t)) {
11472 verbose(env, "bpf_obj_new type ID argument must be of a struct\n");
11476 mark_reg_known_zero(env, regs, BPF_REG_0);
11477 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
11478 regs[BPF_REG_0].btf = ret_btf;
11479 regs[BPF_REG_0].btf_id = ret_btf_id;
11481 insn_aux->obj_new_size = ret_t->size;
11482 insn_aux->kptr_struct_meta =
11483 btf_find_struct_meta(ret_btf, ret_btf_id);
11484 } else if (meta.func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) {
11485 mark_reg_known_zero(env, regs, BPF_REG_0);
11486 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
11487 regs[BPF_REG_0].btf = meta.arg_btf;
11488 regs[BPF_REG_0].btf_id = meta.arg_btf_id;
11490 insn_aux->kptr_struct_meta =
11491 btf_find_struct_meta(meta.arg_btf,
11493 } else if (meta.func_id == special_kfunc_list[KF_bpf_list_pop_front] ||
11494 meta.func_id == special_kfunc_list[KF_bpf_list_pop_back]) {
11495 struct btf_field *field = meta.arg_list_head.field;
11497 mark_reg_graph_node(regs, BPF_REG_0, &field->graph_root);
11498 } else if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
11499 meta.func_id == special_kfunc_list[KF_bpf_rbtree_first]) {
11500 struct btf_field *field = meta.arg_rbtree_root.field;
11502 mark_reg_graph_node(regs, BPF_REG_0, &field->graph_root);
11503 } else if (meta.func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
11504 mark_reg_known_zero(env, regs, BPF_REG_0);
11505 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_TRUSTED;
11506 regs[BPF_REG_0].btf = desc_btf;
11507 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
11508 } else if (meta.func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
11509 ret_t = btf_type_by_id(desc_btf, meta.arg_constant.value);
11510 if (!ret_t || !btf_type_is_struct(ret_t)) {
11512 "kfunc bpf_rdonly_cast type ID argument must be of a struct\n");
11516 mark_reg_known_zero(env, regs, BPF_REG_0);
11517 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
11518 regs[BPF_REG_0].btf = desc_btf;
11519 regs[BPF_REG_0].btf_id = meta.arg_constant.value;
11520 } else if (meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice] ||
11521 meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice_rdwr]) {
11522 enum bpf_type_flag type_flag = get_dynptr_type_flag(meta.initialized_dynptr.type);
11524 mark_reg_known_zero(env, regs, BPF_REG_0);
11526 if (!meta.arg_constant.found) {
11527 verbose(env, "verifier internal error: bpf_dynptr_slice(_rdwr) no constant size\n");
11531 regs[BPF_REG_0].mem_size = meta.arg_constant.value;
11533 /* PTR_MAYBE_NULL will be added when is_kfunc_ret_null is checked */
11534 regs[BPF_REG_0].type = PTR_TO_MEM | type_flag;
11536 if (meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice]) {
11537 regs[BPF_REG_0].type |= MEM_RDONLY;
11539 /* this will set env->seen_direct_write to true */
11540 if (!may_access_direct_pkt_data(env, NULL, BPF_WRITE)) {
11541 verbose(env, "the prog does not allow writes to packet data\n");
11546 if (!meta.initialized_dynptr.id) {
11547 verbose(env, "verifier internal error: no dynptr id\n");
11550 regs[BPF_REG_0].dynptr_id = meta.initialized_dynptr.id;
11552 /* we don't need to set BPF_REG_0's ref obj id
11553 * because packet slices are not refcounted (see
11554 * dynptr_type_refcounted)
11557 verbose(env, "kernel function %s unhandled dynamic return type\n",
11561 } else if (!__btf_type_is_struct(ptr_type)) {
11562 if (!meta.r0_size) {
11565 if (!IS_ERR(btf_resolve_size(desc_btf, ptr_type, &sz))) {
11567 meta.r0_rdonly = true;
11570 if (!meta.r0_size) {
11571 ptr_type_name = btf_name_by_offset(desc_btf,
11572 ptr_type->name_off);
11574 "kernel function %s returns pointer type %s %s is not supported\n",
11576 btf_type_str(ptr_type),
11581 mark_reg_known_zero(env, regs, BPF_REG_0);
11582 regs[BPF_REG_0].type = PTR_TO_MEM;
11583 regs[BPF_REG_0].mem_size = meta.r0_size;
11585 if (meta.r0_rdonly)
11586 regs[BPF_REG_0].type |= MEM_RDONLY;
11588 /* Ensures we don't access the memory after a release_reference() */
11589 if (meta.ref_obj_id)
11590 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
11592 mark_reg_known_zero(env, regs, BPF_REG_0);
11593 regs[BPF_REG_0].btf = desc_btf;
11594 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
11595 regs[BPF_REG_0].btf_id = ptr_type_id;
11598 if (is_kfunc_ret_null(&meta)) {
11599 regs[BPF_REG_0].type |= PTR_MAYBE_NULL;
11600 /* For mark_ptr_or_null_reg, see 93c230e3f5bd6 */
11601 regs[BPF_REG_0].id = ++env->id_gen;
11603 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
11604 if (is_kfunc_acquire(&meta)) {
11605 int id = acquire_reference_state(env, insn_idx);
11609 if (is_kfunc_ret_null(&meta))
11610 regs[BPF_REG_0].id = id;
11611 regs[BPF_REG_0].ref_obj_id = id;
11612 } else if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_first]) {
11613 ref_set_non_owning(env, ®s[BPF_REG_0]);
11616 if (reg_may_point_to_spin_lock(®s[BPF_REG_0]) && !regs[BPF_REG_0].id)
11617 regs[BPF_REG_0].id = ++env->id_gen;
11618 } else if (btf_type_is_void(t)) {
11619 if (meta.btf == btf_vmlinux && btf_id_set_contains(&special_kfunc_set, meta.func_id)) {
11620 if (meta.func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
11621 insn_aux->kptr_struct_meta =
11622 btf_find_struct_meta(meta.arg_btf,
11628 nargs = btf_type_vlen(meta.func_proto);
11629 args = (const struct btf_param *)(meta.func_proto + 1);
11630 for (i = 0; i < nargs; i++) {
11633 t = btf_type_skip_modifiers(desc_btf, args[i].type, NULL);
11634 if (btf_type_is_ptr(t))
11635 mark_btf_func_reg_size(env, regno, sizeof(void *));
11637 /* scalar. ensured by btf_check_kfunc_arg_match() */
11638 mark_btf_func_reg_size(env, regno, t->size);
11641 if (is_iter_next_kfunc(&meta)) {
11642 err = process_iter_next_call(env, insn_idx, &meta);
11650 static bool signed_add_overflows(s64 a, s64 b)
11652 /* Do the add in u64, where overflow is well-defined */
11653 s64 res = (s64)((u64)a + (u64)b);
11660 static bool signed_add32_overflows(s32 a, s32 b)
11662 /* Do the add in u32, where overflow is well-defined */
11663 s32 res = (s32)((u32)a + (u32)b);
11670 static bool signed_sub_overflows(s64 a, s64 b)
11672 /* Do the sub in u64, where overflow is well-defined */
11673 s64 res = (s64)((u64)a - (u64)b);
11680 static bool signed_sub32_overflows(s32 a, s32 b)
11682 /* Do the sub in u32, where overflow is well-defined */
11683 s32 res = (s32)((u32)a - (u32)b);
11690 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
11691 const struct bpf_reg_state *reg,
11692 enum bpf_reg_type type)
11694 bool known = tnum_is_const(reg->var_off);
11695 s64 val = reg->var_off.value;
11696 s64 smin = reg->smin_value;
11698 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
11699 verbose(env, "math between %s pointer and %lld is not allowed\n",
11700 reg_type_str(env, type), val);
11704 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
11705 verbose(env, "%s pointer offset %d is not allowed\n",
11706 reg_type_str(env, type), reg->off);
11710 if (smin == S64_MIN) {
11711 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
11712 reg_type_str(env, type));
11716 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
11717 verbose(env, "value %lld makes %s pointer be out of bounds\n",
11718 smin, reg_type_str(env, type));
11726 REASON_BOUNDS = -1,
11733 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
11734 u32 *alu_limit, bool mask_to_left)
11736 u32 max = 0, ptr_limit = 0;
11738 switch (ptr_reg->type) {
11740 /* Offset 0 is out-of-bounds, but acceptable start for the
11741 * left direction, see BPF_REG_FP. Also, unknown scalar
11742 * offset where we would need to deal with min/max bounds is
11743 * currently prohibited for unprivileged.
11745 max = MAX_BPF_STACK + mask_to_left;
11746 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
11748 case PTR_TO_MAP_VALUE:
11749 max = ptr_reg->map_ptr->value_size;
11750 ptr_limit = (mask_to_left ?
11751 ptr_reg->smin_value :
11752 ptr_reg->umax_value) + ptr_reg->off;
11755 return REASON_TYPE;
11758 if (ptr_limit >= max)
11759 return REASON_LIMIT;
11760 *alu_limit = ptr_limit;
11764 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
11765 const struct bpf_insn *insn)
11767 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
11770 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
11771 u32 alu_state, u32 alu_limit)
11773 /* If we arrived here from different branches with different
11774 * state or limits to sanitize, then this won't work.
11776 if (aux->alu_state &&
11777 (aux->alu_state != alu_state ||
11778 aux->alu_limit != alu_limit))
11779 return REASON_PATHS;
11781 /* Corresponding fixup done in do_misc_fixups(). */
11782 aux->alu_state = alu_state;
11783 aux->alu_limit = alu_limit;
11787 static int sanitize_val_alu(struct bpf_verifier_env *env,
11788 struct bpf_insn *insn)
11790 struct bpf_insn_aux_data *aux = cur_aux(env);
11792 if (can_skip_alu_sanitation(env, insn))
11795 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
11798 static bool sanitize_needed(u8 opcode)
11800 return opcode == BPF_ADD || opcode == BPF_SUB;
11803 struct bpf_sanitize_info {
11804 struct bpf_insn_aux_data aux;
11808 static struct bpf_verifier_state *
11809 sanitize_speculative_path(struct bpf_verifier_env *env,
11810 const struct bpf_insn *insn,
11811 u32 next_idx, u32 curr_idx)
11813 struct bpf_verifier_state *branch;
11814 struct bpf_reg_state *regs;
11816 branch = push_stack(env, next_idx, curr_idx, true);
11817 if (branch && insn) {
11818 regs = branch->frame[branch->curframe]->regs;
11819 if (BPF_SRC(insn->code) == BPF_K) {
11820 mark_reg_unknown(env, regs, insn->dst_reg);
11821 } else if (BPF_SRC(insn->code) == BPF_X) {
11822 mark_reg_unknown(env, regs, insn->dst_reg);
11823 mark_reg_unknown(env, regs, insn->src_reg);
11829 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
11830 struct bpf_insn *insn,
11831 const struct bpf_reg_state *ptr_reg,
11832 const struct bpf_reg_state *off_reg,
11833 struct bpf_reg_state *dst_reg,
11834 struct bpf_sanitize_info *info,
11835 const bool commit_window)
11837 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
11838 struct bpf_verifier_state *vstate = env->cur_state;
11839 bool off_is_imm = tnum_is_const(off_reg->var_off);
11840 bool off_is_neg = off_reg->smin_value < 0;
11841 bool ptr_is_dst_reg = ptr_reg == dst_reg;
11842 u8 opcode = BPF_OP(insn->code);
11843 u32 alu_state, alu_limit;
11844 struct bpf_reg_state tmp;
11848 if (can_skip_alu_sanitation(env, insn))
11851 /* We already marked aux for masking from non-speculative
11852 * paths, thus we got here in the first place. We only care
11853 * to explore bad access from here.
11855 if (vstate->speculative)
11858 if (!commit_window) {
11859 if (!tnum_is_const(off_reg->var_off) &&
11860 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
11861 return REASON_BOUNDS;
11863 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
11864 (opcode == BPF_SUB && !off_is_neg);
11867 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
11871 if (commit_window) {
11872 /* In commit phase we narrow the masking window based on
11873 * the observed pointer move after the simulated operation.
11875 alu_state = info->aux.alu_state;
11876 alu_limit = abs(info->aux.alu_limit - alu_limit);
11878 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
11879 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
11880 alu_state |= ptr_is_dst_reg ?
11881 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
11883 /* Limit pruning on unknown scalars to enable deep search for
11884 * potential masking differences from other program paths.
11887 env->explore_alu_limits = true;
11890 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
11894 /* If we're in commit phase, we're done here given we already
11895 * pushed the truncated dst_reg into the speculative verification
11898 * Also, when register is a known constant, we rewrite register-based
11899 * operation to immediate-based, and thus do not need masking (and as
11900 * a consequence, do not need to simulate the zero-truncation either).
11902 if (commit_window || off_is_imm)
11905 /* Simulate and find potential out-of-bounds access under
11906 * speculative execution from truncation as a result of
11907 * masking when off was not within expected range. If off
11908 * sits in dst, then we temporarily need to move ptr there
11909 * to simulate dst (== 0) +/-= ptr. Needed, for example,
11910 * for cases where we use K-based arithmetic in one direction
11911 * and truncated reg-based in the other in order to explore
11914 if (!ptr_is_dst_reg) {
11916 copy_register_state(dst_reg, ptr_reg);
11918 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
11920 if (!ptr_is_dst_reg && ret)
11922 return !ret ? REASON_STACK : 0;
11925 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
11927 struct bpf_verifier_state *vstate = env->cur_state;
11929 /* If we simulate paths under speculation, we don't update the
11930 * insn as 'seen' such that when we verify unreachable paths in
11931 * the non-speculative domain, sanitize_dead_code() can still
11932 * rewrite/sanitize them.
11934 if (!vstate->speculative)
11935 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
11938 static int sanitize_err(struct bpf_verifier_env *env,
11939 const struct bpf_insn *insn, int reason,
11940 const struct bpf_reg_state *off_reg,
11941 const struct bpf_reg_state *dst_reg)
11943 static const char *err = "pointer arithmetic with it prohibited for !root";
11944 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
11945 u32 dst = insn->dst_reg, src = insn->src_reg;
11948 case REASON_BOUNDS:
11949 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
11950 off_reg == dst_reg ? dst : src, err);
11953 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
11954 off_reg == dst_reg ? src : dst, err);
11957 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
11961 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
11965 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
11969 verbose(env, "verifier internal error: unknown reason (%d)\n",
11977 /* check that stack access falls within stack limits and that 'reg' doesn't
11978 * have a variable offset.
11980 * Variable offset is prohibited for unprivileged mode for simplicity since it
11981 * requires corresponding support in Spectre masking for stack ALU. See also
11982 * retrieve_ptr_limit().
11985 * 'off' includes 'reg->off'.
11987 static int check_stack_access_for_ptr_arithmetic(
11988 struct bpf_verifier_env *env,
11990 const struct bpf_reg_state *reg,
11993 if (!tnum_is_const(reg->var_off)) {
11996 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
11997 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
11998 regno, tn_buf, off);
12002 if (off >= 0 || off < -MAX_BPF_STACK) {
12003 verbose(env, "R%d stack pointer arithmetic goes out of range, "
12004 "prohibited for !root; off=%d\n", regno, off);
12011 static int sanitize_check_bounds(struct bpf_verifier_env *env,
12012 const struct bpf_insn *insn,
12013 const struct bpf_reg_state *dst_reg)
12015 u32 dst = insn->dst_reg;
12017 /* For unprivileged we require that resulting offset must be in bounds
12018 * in order to be able to sanitize access later on.
12020 if (env->bypass_spec_v1)
12023 switch (dst_reg->type) {
12025 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
12026 dst_reg->off + dst_reg->var_off.value))
12029 case PTR_TO_MAP_VALUE:
12030 if (check_map_access(env, dst, dst_reg->off, 1, false, ACCESS_HELPER)) {
12031 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
12032 "prohibited for !root\n", dst);
12043 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
12044 * Caller should also handle BPF_MOV case separately.
12045 * If we return -EACCES, caller may want to try again treating pointer as a
12046 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
12048 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
12049 struct bpf_insn *insn,
12050 const struct bpf_reg_state *ptr_reg,
12051 const struct bpf_reg_state *off_reg)
12053 struct bpf_verifier_state *vstate = env->cur_state;
12054 struct bpf_func_state *state = vstate->frame[vstate->curframe];
12055 struct bpf_reg_state *regs = state->regs, *dst_reg;
12056 bool known = tnum_is_const(off_reg->var_off);
12057 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
12058 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
12059 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
12060 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
12061 struct bpf_sanitize_info info = {};
12062 u8 opcode = BPF_OP(insn->code);
12063 u32 dst = insn->dst_reg;
12066 dst_reg = ®s[dst];
12068 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
12069 smin_val > smax_val || umin_val > umax_val) {
12070 /* Taint dst register if offset had invalid bounds derived from
12071 * e.g. dead branches.
12073 __mark_reg_unknown(env, dst_reg);
12077 if (BPF_CLASS(insn->code) != BPF_ALU64) {
12078 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
12079 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
12080 __mark_reg_unknown(env, dst_reg);
12085 "R%d 32-bit pointer arithmetic prohibited\n",
12090 if (ptr_reg->type & PTR_MAYBE_NULL) {
12091 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
12092 dst, reg_type_str(env, ptr_reg->type));
12096 switch (base_type(ptr_reg->type)) {
12097 case CONST_PTR_TO_MAP:
12098 /* smin_val represents the known value */
12099 if (known && smin_val == 0 && opcode == BPF_ADD)
12102 case PTR_TO_PACKET_END:
12103 case PTR_TO_SOCKET:
12104 case PTR_TO_SOCK_COMMON:
12105 case PTR_TO_TCP_SOCK:
12106 case PTR_TO_XDP_SOCK:
12107 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
12108 dst, reg_type_str(env, ptr_reg->type));
12114 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
12115 * The id may be overwritten later if we create a new variable offset.
12117 dst_reg->type = ptr_reg->type;
12118 dst_reg->id = ptr_reg->id;
12120 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
12121 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
12124 /* pointer types do not carry 32-bit bounds at the moment. */
12125 __mark_reg32_unbounded(dst_reg);
12127 if (sanitize_needed(opcode)) {
12128 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
12131 return sanitize_err(env, insn, ret, off_reg, dst_reg);
12136 /* We can take a fixed offset as long as it doesn't overflow
12137 * the s32 'off' field
12139 if (known && (ptr_reg->off + smin_val ==
12140 (s64)(s32)(ptr_reg->off + smin_val))) {
12141 /* pointer += K. Accumulate it into fixed offset */
12142 dst_reg->smin_value = smin_ptr;
12143 dst_reg->smax_value = smax_ptr;
12144 dst_reg->umin_value = umin_ptr;
12145 dst_reg->umax_value = umax_ptr;
12146 dst_reg->var_off = ptr_reg->var_off;
12147 dst_reg->off = ptr_reg->off + smin_val;
12148 dst_reg->raw = ptr_reg->raw;
12151 /* A new variable offset is created. Note that off_reg->off
12152 * == 0, since it's a scalar.
12153 * dst_reg gets the pointer type and since some positive
12154 * integer value was added to the pointer, give it a new 'id'
12155 * if it's a PTR_TO_PACKET.
12156 * this creates a new 'base' pointer, off_reg (variable) gets
12157 * added into the variable offset, and we copy the fixed offset
12160 if (signed_add_overflows(smin_ptr, smin_val) ||
12161 signed_add_overflows(smax_ptr, smax_val)) {
12162 dst_reg->smin_value = S64_MIN;
12163 dst_reg->smax_value = S64_MAX;
12165 dst_reg->smin_value = smin_ptr + smin_val;
12166 dst_reg->smax_value = smax_ptr + smax_val;
12168 if (umin_ptr + umin_val < umin_ptr ||
12169 umax_ptr + umax_val < umax_ptr) {
12170 dst_reg->umin_value = 0;
12171 dst_reg->umax_value = U64_MAX;
12173 dst_reg->umin_value = umin_ptr + umin_val;
12174 dst_reg->umax_value = umax_ptr + umax_val;
12176 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
12177 dst_reg->off = ptr_reg->off;
12178 dst_reg->raw = ptr_reg->raw;
12179 if (reg_is_pkt_pointer(ptr_reg)) {
12180 dst_reg->id = ++env->id_gen;
12181 /* something was added to pkt_ptr, set range to zero */
12182 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
12186 if (dst_reg == off_reg) {
12187 /* scalar -= pointer. Creates an unknown scalar */
12188 verbose(env, "R%d tried to subtract pointer from scalar\n",
12192 /* We don't allow subtraction from FP, because (according to
12193 * test_verifier.c test "invalid fp arithmetic", JITs might not
12194 * be able to deal with it.
12196 if (ptr_reg->type == PTR_TO_STACK) {
12197 verbose(env, "R%d subtraction from stack pointer prohibited\n",
12201 if (known && (ptr_reg->off - smin_val ==
12202 (s64)(s32)(ptr_reg->off - smin_val))) {
12203 /* pointer -= K. Subtract it from fixed offset */
12204 dst_reg->smin_value = smin_ptr;
12205 dst_reg->smax_value = smax_ptr;
12206 dst_reg->umin_value = umin_ptr;
12207 dst_reg->umax_value = umax_ptr;
12208 dst_reg->var_off = ptr_reg->var_off;
12209 dst_reg->id = ptr_reg->id;
12210 dst_reg->off = ptr_reg->off - smin_val;
12211 dst_reg->raw = ptr_reg->raw;
12214 /* A new variable offset is created. If the subtrahend is known
12215 * nonnegative, then any reg->range we had before is still good.
12217 if (signed_sub_overflows(smin_ptr, smax_val) ||
12218 signed_sub_overflows(smax_ptr, smin_val)) {
12219 /* Overflow possible, we know nothing */
12220 dst_reg->smin_value = S64_MIN;
12221 dst_reg->smax_value = S64_MAX;
12223 dst_reg->smin_value = smin_ptr - smax_val;
12224 dst_reg->smax_value = smax_ptr - smin_val;
12226 if (umin_ptr < umax_val) {
12227 /* Overflow possible, we know nothing */
12228 dst_reg->umin_value = 0;
12229 dst_reg->umax_value = U64_MAX;
12231 /* Cannot overflow (as long as bounds are consistent) */
12232 dst_reg->umin_value = umin_ptr - umax_val;
12233 dst_reg->umax_value = umax_ptr - umin_val;
12235 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
12236 dst_reg->off = ptr_reg->off;
12237 dst_reg->raw = ptr_reg->raw;
12238 if (reg_is_pkt_pointer(ptr_reg)) {
12239 dst_reg->id = ++env->id_gen;
12240 /* something was added to pkt_ptr, set range to zero */
12242 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
12248 /* bitwise ops on pointers are troublesome, prohibit. */
12249 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
12250 dst, bpf_alu_string[opcode >> 4]);
12253 /* other operators (e.g. MUL,LSH) produce non-pointer results */
12254 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
12255 dst, bpf_alu_string[opcode >> 4]);
12259 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
12261 reg_bounds_sync(dst_reg);
12262 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
12264 if (sanitize_needed(opcode)) {
12265 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
12268 return sanitize_err(env, insn, ret, off_reg, dst_reg);
12274 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
12275 struct bpf_reg_state *src_reg)
12277 s32 smin_val = src_reg->s32_min_value;
12278 s32 smax_val = src_reg->s32_max_value;
12279 u32 umin_val = src_reg->u32_min_value;
12280 u32 umax_val = src_reg->u32_max_value;
12282 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
12283 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
12284 dst_reg->s32_min_value = S32_MIN;
12285 dst_reg->s32_max_value = S32_MAX;
12287 dst_reg->s32_min_value += smin_val;
12288 dst_reg->s32_max_value += smax_val;
12290 if (dst_reg->u32_min_value + umin_val < umin_val ||
12291 dst_reg->u32_max_value + umax_val < umax_val) {
12292 dst_reg->u32_min_value = 0;
12293 dst_reg->u32_max_value = U32_MAX;
12295 dst_reg->u32_min_value += umin_val;
12296 dst_reg->u32_max_value += umax_val;
12300 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
12301 struct bpf_reg_state *src_reg)
12303 s64 smin_val = src_reg->smin_value;
12304 s64 smax_val = src_reg->smax_value;
12305 u64 umin_val = src_reg->umin_value;
12306 u64 umax_val = src_reg->umax_value;
12308 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
12309 signed_add_overflows(dst_reg->smax_value, smax_val)) {
12310 dst_reg->smin_value = S64_MIN;
12311 dst_reg->smax_value = S64_MAX;
12313 dst_reg->smin_value += smin_val;
12314 dst_reg->smax_value += smax_val;
12316 if (dst_reg->umin_value + umin_val < umin_val ||
12317 dst_reg->umax_value + umax_val < umax_val) {
12318 dst_reg->umin_value = 0;
12319 dst_reg->umax_value = U64_MAX;
12321 dst_reg->umin_value += umin_val;
12322 dst_reg->umax_value += umax_val;
12326 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
12327 struct bpf_reg_state *src_reg)
12329 s32 smin_val = src_reg->s32_min_value;
12330 s32 smax_val = src_reg->s32_max_value;
12331 u32 umin_val = src_reg->u32_min_value;
12332 u32 umax_val = src_reg->u32_max_value;
12334 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
12335 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
12336 /* Overflow possible, we know nothing */
12337 dst_reg->s32_min_value = S32_MIN;
12338 dst_reg->s32_max_value = S32_MAX;
12340 dst_reg->s32_min_value -= smax_val;
12341 dst_reg->s32_max_value -= smin_val;
12343 if (dst_reg->u32_min_value < umax_val) {
12344 /* Overflow possible, we know nothing */
12345 dst_reg->u32_min_value = 0;
12346 dst_reg->u32_max_value = U32_MAX;
12348 /* Cannot overflow (as long as bounds are consistent) */
12349 dst_reg->u32_min_value -= umax_val;
12350 dst_reg->u32_max_value -= umin_val;
12354 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
12355 struct bpf_reg_state *src_reg)
12357 s64 smin_val = src_reg->smin_value;
12358 s64 smax_val = src_reg->smax_value;
12359 u64 umin_val = src_reg->umin_value;
12360 u64 umax_val = src_reg->umax_value;
12362 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
12363 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
12364 /* Overflow possible, we know nothing */
12365 dst_reg->smin_value = S64_MIN;
12366 dst_reg->smax_value = S64_MAX;
12368 dst_reg->smin_value -= smax_val;
12369 dst_reg->smax_value -= smin_val;
12371 if (dst_reg->umin_value < umax_val) {
12372 /* Overflow possible, we know nothing */
12373 dst_reg->umin_value = 0;
12374 dst_reg->umax_value = U64_MAX;
12376 /* Cannot overflow (as long as bounds are consistent) */
12377 dst_reg->umin_value -= umax_val;
12378 dst_reg->umax_value -= umin_val;
12382 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
12383 struct bpf_reg_state *src_reg)
12385 s32 smin_val = src_reg->s32_min_value;
12386 u32 umin_val = src_reg->u32_min_value;
12387 u32 umax_val = src_reg->u32_max_value;
12389 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
12390 /* Ain't nobody got time to multiply that sign */
12391 __mark_reg32_unbounded(dst_reg);
12394 /* Both values are positive, so we can work with unsigned and
12395 * copy the result to signed (unless it exceeds S32_MAX).
12397 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
12398 /* Potential overflow, we know nothing */
12399 __mark_reg32_unbounded(dst_reg);
12402 dst_reg->u32_min_value *= umin_val;
12403 dst_reg->u32_max_value *= umax_val;
12404 if (dst_reg->u32_max_value > S32_MAX) {
12405 /* Overflow possible, we know nothing */
12406 dst_reg->s32_min_value = S32_MIN;
12407 dst_reg->s32_max_value = S32_MAX;
12409 dst_reg->s32_min_value = dst_reg->u32_min_value;
12410 dst_reg->s32_max_value = dst_reg->u32_max_value;
12414 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
12415 struct bpf_reg_state *src_reg)
12417 s64 smin_val = src_reg->smin_value;
12418 u64 umin_val = src_reg->umin_value;
12419 u64 umax_val = src_reg->umax_value;
12421 if (smin_val < 0 || dst_reg->smin_value < 0) {
12422 /* Ain't nobody got time to multiply that sign */
12423 __mark_reg64_unbounded(dst_reg);
12426 /* Both values are positive, so we can work with unsigned and
12427 * copy the result to signed (unless it exceeds S64_MAX).
12429 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
12430 /* Potential overflow, we know nothing */
12431 __mark_reg64_unbounded(dst_reg);
12434 dst_reg->umin_value *= umin_val;
12435 dst_reg->umax_value *= umax_val;
12436 if (dst_reg->umax_value > S64_MAX) {
12437 /* Overflow possible, we know nothing */
12438 dst_reg->smin_value = S64_MIN;
12439 dst_reg->smax_value = S64_MAX;
12441 dst_reg->smin_value = dst_reg->umin_value;
12442 dst_reg->smax_value = dst_reg->umax_value;
12446 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
12447 struct bpf_reg_state *src_reg)
12449 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12450 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12451 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12452 s32 smin_val = src_reg->s32_min_value;
12453 u32 umax_val = src_reg->u32_max_value;
12455 if (src_known && dst_known) {
12456 __mark_reg32_known(dst_reg, var32_off.value);
12460 /* We get our minimum from the var_off, since that's inherently
12461 * bitwise. Our maximum is the minimum of the operands' maxima.
12463 dst_reg->u32_min_value = var32_off.value;
12464 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
12465 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
12466 /* Lose signed bounds when ANDing negative numbers,
12467 * ain't nobody got time for that.
12469 dst_reg->s32_min_value = S32_MIN;
12470 dst_reg->s32_max_value = S32_MAX;
12472 /* ANDing two positives gives a positive, so safe to
12473 * cast result into s64.
12475 dst_reg->s32_min_value = dst_reg->u32_min_value;
12476 dst_reg->s32_max_value = dst_reg->u32_max_value;
12480 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
12481 struct bpf_reg_state *src_reg)
12483 bool src_known = tnum_is_const(src_reg->var_off);
12484 bool dst_known = tnum_is_const(dst_reg->var_off);
12485 s64 smin_val = src_reg->smin_value;
12486 u64 umax_val = src_reg->umax_value;
12488 if (src_known && dst_known) {
12489 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12493 /* We get our minimum from the var_off, since that's inherently
12494 * bitwise. Our maximum is the minimum of the operands' maxima.
12496 dst_reg->umin_value = dst_reg->var_off.value;
12497 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
12498 if (dst_reg->smin_value < 0 || smin_val < 0) {
12499 /* Lose signed bounds when ANDing negative numbers,
12500 * ain't nobody got time for that.
12502 dst_reg->smin_value = S64_MIN;
12503 dst_reg->smax_value = S64_MAX;
12505 /* ANDing two positives gives a positive, so safe to
12506 * cast result into s64.
12508 dst_reg->smin_value = dst_reg->umin_value;
12509 dst_reg->smax_value = dst_reg->umax_value;
12511 /* We may learn something more from the var_off */
12512 __update_reg_bounds(dst_reg);
12515 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
12516 struct bpf_reg_state *src_reg)
12518 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12519 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12520 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12521 s32 smin_val = src_reg->s32_min_value;
12522 u32 umin_val = src_reg->u32_min_value;
12524 if (src_known && dst_known) {
12525 __mark_reg32_known(dst_reg, var32_off.value);
12529 /* We get our maximum from the var_off, and our minimum is the
12530 * maximum of the operands' minima
12532 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
12533 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
12534 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
12535 /* Lose signed bounds when ORing negative numbers,
12536 * ain't nobody got time for that.
12538 dst_reg->s32_min_value = S32_MIN;
12539 dst_reg->s32_max_value = S32_MAX;
12541 /* ORing two positives gives a positive, so safe to
12542 * cast result into s64.
12544 dst_reg->s32_min_value = dst_reg->u32_min_value;
12545 dst_reg->s32_max_value = dst_reg->u32_max_value;
12549 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
12550 struct bpf_reg_state *src_reg)
12552 bool src_known = tnum_is_const(src_reg->var_off);
12553 bool dst_known = tnum_is_const(dst_reg->var_off);
12554 s64 smin_val = src_reg->smin_value;
12555 u64 umin_val = src_reg->umin_value;
12557 if (src_known && dst_known) {
12558 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12562 /* We get our maximum from the var_off, and our minimum is the
12563 * maximum of the operands' minima
12565 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
12566 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
12567 if (dst_reg->smin_value < 0 || smin_val < 0) {
12568 /* Lose signed bounds when ORing negative numbers,
12569 * ain't nobody got time for that.
12571 dst_reg->smin_value = S64_MIN;
12572 dst_reg->smax_value = S64_MAX;
12574 /* ORing two positives gives a positive, so safe to
12575 * cast result into s64.
12577 dst_reg->smin_value = dst_reg->umin_value;
12578 dst_reg->smax_value = dst_reg->umax_value;
12580 /* We may learn something more from the var_off */
12581 __update_reg_bounds(dst_reg);
12584 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
12585 struct bpf_reg_state *src_reg)
12587 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12588 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12589 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12590 s32 smin_val = src_reg->s32_min_value;
12592 if (src_known && dst_known) {
12593 __mark_reg32_known(dst_reg, var32_off.value);
12597 /* We get both minimum and maximum from the var32_off. */
12598 dst_reg->u32_min_value = var32_off.value;
12599 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
12601 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
12602 /* XORing two positive sign numbers gives a positive,
12603 * so safe to cast u32 result into s32.
12605 dst_reg->s32_min_value = dst_reg->u32_min_value;
12606 dst_reg->s32_max_value = dst_reg->u32_max_value;
12608 dst_reg->s32_min_value = S32_MIN;
12609 dst_reg->s32_max_value = S32_MAX;
12613 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
12614 struct bpf_reg_state *src_reg)
12616 bool src_known = tnum_is_const(src_reg->var_off);
12617 bool dst_known = tnum_is_const(dst_reg->var_off);
12618 s64 smin_val = src_reg->smin_value;
12620 if (src_known && dst_known) {
12621 /* dst_reg->var_off.value has been updated earlier */
12622 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12626 /* We get both minimum and maximum from the var_off. */
12627 dst_reg->umin_value = dst_reg->var_off.value;
12628 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
12630 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
12631 /* XORing two positive sign numbers gives a positive,
12632 * so safe to cast u64 result into s64.
12634 dst_reg->smin_value = dst_reg->umin_value;
12635 dst_reg->smax_value = dst_reg->umax_value;
12637 dst_reg->smin_value = S64_MIN;
12638 dst_reg->smax_value = S64_MAX;
12641 __update_reg_bounds(dst_reg);
12644 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
12645 u64 umin_val, u64 umax_val)
12647 /* We lose all sign bit information (except what we can pick
12650 dst_reg->s32_min_value = S32_MIN;
12651 dst_reg->s32_max_value = S32_MAX;
12652 /* If we might shift our top bit out, then we know nothing */
12653 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
12654 dst_reg->u32_min_value = 0;
12655 dst_reg->u32_max_value = U32_MAX;
12657 dst_reg->u32_min_value <<= umin_val;
12658 dst_reg->u32_max_value <<= umax_val;
12662 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
12663 struct bpf_reg_state *src_reg)
12665 u32 umax_val = src_reg->u32_max_value;
12666 u32 umin_val = src_reg->u32_min_value;
12667 /* u32 alu operation will zext upper bits */
12668 struct tnum subreg = tnum_subreg(dst_reg->var_off);
12670 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
12671 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
12672 /* Not required but being careful mark reg64 bounds as unknown so
12673 * that we are forced to pick them up from tnum and zext later and
12674 * if some path skips this step we are still safe.
12676 __mark_reg64_unbounded(dst_reg);
12677 __update_reg32_bounds(dst_reg);
12680 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
12681 u64 umin_val, u64 umax_val)
12683 /* Special case <<32 because it is a common compiler pattern to sign
12684 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
12685 * positive we know this shift will also be positive so we can track
12686 * bounds correctly. Otherwise we lose all sign bit information except
12687 * what we can pick up from var_off. Perhaps we can generalize this
12688 * later to shifts of any length.
12690 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
12691 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
12693 dst_reg->smax_value = S64_MAX;
12695 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
12696 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
12698 dst_reg->smin_value = S64_MIN;
12700 /* If we might shift our top bit out, then we know nothing */
12701 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
12702 dst_reg->umin_value = 0;
12703 dst_reg->umax_value = U64_MAX;
12705 dst_reg->umin_value <<= umin_val;
12706 dst_reg->umax_value <<= umax_val;
12710 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
12711 struct bpf_reg_state *src_reg)
12713 u64 umax_val = src_reg->umax_value;
12714 u64 umin_val = src_reg->umin_value;
12716 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
12717 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
12718 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
12720 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
12721 /* We may learn something more from the var_off */
12722 __update_reg_bounds(dst_reg);
12725 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
12726 struct bpf_reg_state *src_reg)
12728 struct tnum subreg = tnum_subreg(dst_reg->var_off);
12729 u32 umax_val = src_reg->u32_max_value;
12730 u32 umin_val = src_reg->u32_min_value;
12732 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
12733 * be negative, then either:
12734 * 1) src_reg might be zero, so the sign bit of the result is
12735 * unknown, so we lose our signed bounds
12736 * 2) it's known negative, thus the unsigned bounds capture the
12738 * 3) the signed bounds cross zero, so they tell us nothing
12740 * If the value in dst_reg is known nonnegative, then again the
12741 * unsigned bounds capture the signed bounds.
12742 * Thus, in all cases it suffices to blow away our signed bounds
12743 * and rely on inferring new ones from the unsigned bounds and
12744 * var_off of the result.
12746 dst_reg->s32_min_value = S32_MIN;
12747 dst_reg->s32_max_value = S32_MAX;
12749 dst_reg->var_off = tnum_rshift(subreg, umin_val);
12750 dst_reg->u32_min_value >>= umax_val;
12751 dst_reg->u32_max_value >>= umin_val;
12753 __mark_reg64_unbounded(dst_reg);
12754 __update_reg32_bounds(dst_reg);
12757 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
12758 struct bpf_reg_state *src_reg)
12760 u64 umax_val = src_reg->umax_value;
12761 u64 umin_val = src_reg->umin_value;
12763 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
12764 * be negative, then either:
12765 * 1) src_reg might be zero, so the sign bit of the result is
12766 * unknown, so we lose our signed bounds
12767 * 2) it's known negative, thus the unsigned bounds capture the
12769 * 3) the signed bounds cross zero, so they tell us nothing
12771 * If the value in dst_reg is known nonnegative, then again the
12772 * unsigned bounds capture the signed bounds.
12773 * Thus, in all cases it suffices to blow away our signed bounds
12774 * and rely on inferring new ones from the unsigned bounds and
12775 * var_off of the result.
12777 dst_reg->smin_value = S64_MIN;
12778 dst_reg->smax_value = S64_MAX;
12779 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
12780 dst_reg->umin_value >>= umax_val;
12781 dst_reg->umax_value >>= umin_val;
12783 /* Its not easy to operate on alu32 bounds here because it depends
12784 * on bits being shifted in. Take easy way out and mark unbounded
12785 * so we can recalculate later from tnum.
12787 __mark_reg32_unbounded(dst_reg);
12788 __update_reg_bounds(dst_reg);
12791 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
12792 struct bpf_reg_state *src_reg)
12794 u64 umin_val = src_reg->u32_min_value;
12796 /* Upon reaching here, src_known is true and
12797 * umax_val is equal to umin_val.
12799 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
12800 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
12802 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
12804 /* blow away the dst_reg umin_value/umax_value and rely on
12805 * dst_reg var_off to refine the result.
12807 dst_reg->u32_min_value = 0;
12808 dst_reg->u32_max_value = U32_MAX;
12810 __mark_reg64_unbounded(dst_reg);
12811 __update_reg32_bounds(dst_reg);
12814 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
12815 struct bpf_reg_state *src_reg)
12817 u64 umin_val = src_reg->umin_value;
12819 /* Upon reaching here, src_known is true and umax_val is equal
12822 dst_reg->smin_value >>= umin_val;
12823 dst_reg->smax_value >>= umin_val;
12825 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
12827 /* blow away the dst_reg umin_value/umax_value and rely on
12828 * dst_reg var_off to refine the result.
12830 dst_reg->umin_value = 0;
12831 dst_reg->umax_value = U64_MAX;
12833 /* Its not easy to operate on alu32 bounds here because it depends
12834 * on bits being shifted in from upper 32-bits. Take easy way out
12835 * and mark unbounded so we can recalculate later from tnum.
12837 __mark_reg32_unbounded(dst_reg);
12838 __update_reg_bounds(dst_reg);
12841 /* WARNING: This function does calculations on 64-bit values, but the actual
12842 * execution may occur on 32-bit values. Therefore, things like bitshifts
12843 * need extra checks in the 32-bit case.
12845 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
12846 struct bpf_insn *insn,
12847 struct bpf_reg_state *dst_reg,
12848 struct bpf_reg_state src_reg)
12850 struct bpf_reg_state *regs = cur_regs(env);
12851 u8 opcode = BPF_OP(insn->code);
12853 s64 smin_val, smax_val;
12854 u64 umin_val, umax_val;
12855 s32 s32_min_val, s32_max_val;
12856 u32 u32_min_val, u32_max_val;
12857 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
12858 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
12861 smin_val = src_reg.smin_value;
12862 smax_val = src_reg.smax_value;
12863 umin_val = src_reg.umin_value;
12864 umax_val = src_reg.umax_value;
12866 s32_min_val = src_reg.s32_min_value;
12867 s32_max_val = src_reg.s32_max_value;
12868 u32_min_val = src_reg.u32_min_value;
12869 u32_max_val = src_reg.u32_max_value;
12872 src_known = tnum_subreg_is_const(src_reg.var_off);
12874 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
12875 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
12876 /* Taint dst register if offset had invalid bounds
12877 * derived from e.g. dead branches.
12879 __mark_reg_unknown(env, dst_reg);
12883 src_known = tnum_is_const(src_reg.var_off);
12885 (smin_val != smax_val || umin_val != umax_val)) ||
12886 smin_val > smax_val || umin_val > umax_val) {
12887 /* Taint dst register if offset had invalid bounds
12888 * derived from e.g. dead branches.
12890 __mark_reg_unknown(env, dst_reg);
12896 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
12897 __mark_reg_unknown(env, dst_reg);
12901 if (sanitize_needed(opcode)) {
12902 ret = sanitize_val_alu(env, insn);
12904 return sanitize_err(env, insn, ret, NULL, NULL);
12907 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
12908 * There are two classes of instructions: The first class we track both
12909 * alu32 and alu64 sign/unsigned bounds independently this provides the
12910 * greatest amount of precision when alu operations are mixed with jmp32
12911 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
12912 * and BPF_OR. This is possible because these ops have fairly easy to
12913 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
12914 * See alu32 verifier tests for examples. The second class of
12915 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
12916 * with regards to tracking sign/unsigned bounds because the bits may
12917 * cross subreg boundaries in the alu64 case. When this happens we mark
12918 * the reg unbounded in the subreg bound space and use the resulting
12919 * tnum to calculate an approximation of the sign/unsigned bounds.
12923 scalar32_min_max_add(dst_reg, &src_reg);
12924 scalar_min_max_add(dst_reg, &src_reg);
12925 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
12928 scalar32_min_max_sub(dst_reg, &src_reg);
12929 scalar_min_max_sub(dst_reg, &src_reg);
12930 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
12933 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
12934 scalar32_min_max_mul(dst_reg, &src_reg);
12935 scalar_min_max_mul(dst_reg, &src_reg);
12938 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
12939 scalar32_min_max_and(dst_reg, &src_reg);
12940 scalar_min_max_and(dst_reg, &src_reg);
12943 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
12944 scalar32_min_max_or(dst_reg, &src_reg);
12945 scalar_min_max_or(dst_reg, &src_reg);
12948 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
12949 scalar32_min_max_xor(dst_reg, &src_reg);
12950 scalar_min_max_xor(dst_reg, &src_reg);
12953 if (umax_val >= insn_bitness) {
12954 /* Shifts greater than 31 or 63 are undefined.
12955 * This includes shifts by a negative number.
12957 mark_reg_unknown(env, regs, insn->dst_reg);
12961 scalar32_min_max_lsh(dst_reg, &src_reg);
12963 scalar_min_max_lsh(dst_reg, &src_reg);
12966 if (umax_val >= insn_bitness) {
12967 /* Shifts greater than 31 or 63 are undefined.
12968 * This includes shifts by a negative number.
12970 mark_reg_unknown(env, regs, insn->dst_reg);
12974 scalar32_min_max_rsh(dst_reg, &src_reg);
12976 scalar_min_max_rsh(dst_reg, &src_reg);
12979 if (umax_val >= insn_bitness) {
12980 /* Shifts greater than 31 or 63 are undefined.
12981 * This includes shifts by a negative number.
12983 mark_reg_unknown(env, regs, insn->dst_reg);
12987 scalar32_min_max_arsh(dst_reg, &src_reg);
12989 scalar_min_max_arsh(dst_reg, &src_reg);
12992 mark_reg_unknown(env, regs, insn->dst_reg);
12996 /* ALU32 ops are zero extended into 64bit register */
12998 zext_32_to_64(dst_reg);
12999 reg_bounds_sync(dst_reg);
13003 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
13006 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
13007 struct bpf_insn *insn)
13009 struct bpf_verifier_state *vstate = env->cur_state;
13010 struct bpf_func_state *state = vstate->frame[vstate->curframe];
13011 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
13012 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
13013 u8 opcode = BPF_OP(insn->code);
13016 dst_reg = ®s[insn->dst_reg];
13018 if (dst_reg->type != SCALAR_VALUE)
13021 /* Make sure ID is cleared otherwise dst_reg min/max could be
13022 * incorrectly propagated into other registers by find_equal_scalars()
13025 if (BPF_SRC(insn->code) == BPF_X) {
13026 src_reg = ®s[insn->src_reg];
13027 if (src_reg->type != SCALAR_VALUE) {
13028 if (dst_reg->type != SCALAR_VALUE) {
13029 /* Combining two pointers by any ALU op yields
13030 * an arbitrary scalar. Disallow all math except
13031 * pointer subtraction
13033 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
13034 mark_reg_unknown(env, regs, insn->dst_reg);
13037 verbose(env, "R%d pointer %s pointer prohibited\n",
13039 bpf_alu_string[opcode >> 4]);
13042 /* scalar += pointer
13043 * This is legal, but we have to reverse our
13044 * src/dest handling in computing the range
13046 err = mark_chain_precision(env, insn->dst_reg);
13049 return adjust_ptr_min_max_vals(env, insn,
13052 } else if (ptr_reg) {
13053 /* pointer += scalar */
13054 err = mark_chain_precision(env, insn->src_reg);
13057 return adjust_ptr_min_max_vals(env, insn,
13059 } else if (dst_reg->precise) {
13060 /* if dst_reg is precise, src_reg should be precise as well */
13061 err = mark_chain_precision(env, insn->src_reg);
13066 /* Pretend the src is a reg with a known value, since we only
13067 * need to be able to read from this state.
13069 off_reg.type = SCALAR_VALUE;
13070 __mark_reg_known(&off_reg, insn->imm);
13071 src_reg = &off_reg;
13072 if (ptr_reg) /* pointer += K */
13073 return adjust_ptr_min_max_vals(env, insn,
13077 /* Got here implies adding two SCALAR_VALUEs */
13078 if (WARN_ON_ONCE(ptr_reg)) {
13079 print_verifier_state(env, state, true);
13080 verbose(env, "verifier internal error: unexpected ptr_reg\n");
13083 if (WARN_ON(!src_reg)) {
13084 print_verifier_state(env, state, true);
13085 verbose(env, "verifier internal error: no src_reg\n");
13088 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
13091 /* check validity of 32-bit and 64-bit arithmetic operations */
13092 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
13094 struct bpf_reg_state *regs = cur_regs(env);
13095 u8 opcode = BPF_OP(insn->code);
13098 if (opcode == BPF_END || opcode == BPF_NEG) {
13099 if (opcode == BPF_NEG) {
13100 if (BPF_SRC(insn->code) != BPF_K ||
13101 insn->src_reg != BPF_REG_0 ||
13102 insn->off != 0 || insn->imm != 0) {
13103 verbose(env, "BPF_NEG uses reserved fields\n");
13107 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
13108 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
13109 (BPF_CLASS(insn->code) == BPF_ALU64 &&
13110 BPF_SRC(insn->code) != BPF_TO_LE)) {
13111 verbose(env, "BPF_END uses reserved fields\n");
13116 /* check src operand */
13117 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
13121 if (is_pointer_value(env, insn->dst_reg)) {
13122 verbose(env, "R%d pointer arithmetic prohibited\n",
13127 /* check dest operand */
13128 err = check_reg_arg(env, insn->dst_reg, DST_OP);
13132 } else if (opcode == BPF_MOV) {
13134 if (BPF_SRC(insn->code) == BPF_X) {
13135 if (insn->imm != 0) {
13136 verbose(env, "BPF_MOV uses reserved fields\n");
13140 if (BPF_CLASS(insn->code) == BPF_ALU) {
13141 if (insn->off != 0 && insn->off != 8 && insn->off != 16) {
13142 verbose(env, "BPF_MOV uses reserved fields\n");
13146 if (insn->off != 0 && insn->off != 8 && insn->off != 16 &&
13148 verbose(env, "BPF_MOV uses reserved fields\n");
13153 /* check src operand */
13154 err = check_reg_arg(env, insn->src_reg, SRC_OP);
13158 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
13159 verbose(env, "BPF_MOV uses reserved fields\n");
13164 /* check dest operand, mark as required later */
13165 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
13169 if (BPF_SRC(insn->code) == BPF_X) {
13170 struct bpf_reg_state *src_reg = regs + insn->src_reg;
13171 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
13172 bool need_id = src_reg->type == SCALAR_VALUE && !src_reg->id &&
13173 !tnum_is_const(src_reg->var_off);
13175 if (BPF_CLASS(insn->code) == BPF_ALU64) {
13176 if (insn->off == 0) {
13178 * copy register state to dest reg
13181 /* Assign src and dst registers the same ID
13182 * that will be used by find_equal_scalars()
13183 * to propagate min/max range.
13185 src_reg->id = ++env->id_gen;
13186 copy_register_state(dst_reg, src_reg);
13187 dst_reg->live |= REG_LIVE_WRITTEN;
13188 dst_reg->subreg_def = DEF_NOT_SUBREG;
13190 /* case: R1 = (s8, s16 s32)R2 */
13191 if (is_pointer_value(env, insn->src_reg)) {
13193 "R%d sign-extension part of pointer\n",
13196 } else if (src_reg->type == SCALAR_VALUE) {
13199 no_sext = src_reg->umax_value < (1ULL << (insn->off - 1));
13200 if (no_sext && need_id)
13201 src_reg->id = ++env->id_gen;
13202 copy_register_state(dst_reg, src_reg);
13205 coerce_reg_to_size_sx(dst_reg, insn->off >> 3);
13206 dst_reg->live |= REG_LIVE_WRITTEN;
13207 dst_reg->subreg_def = DEF_NOT_SUBREG;
13209 mark_reg_unknown(env, regs, insn->dst_reg);
13213 /* R1 = (u32) R2 */
13214 if (is_pointer_value(env, insn->src_reg)) {
13216 "R%d partial copy of pointer\n",
13219 } else if (src_reg->type == SCALAR_VALUE) {
13220 if (insn->off == 0) {
13221 bool is_src_reg_u32 = src_reg->umax_value <= U32_MAX;
13223 if (is_src_reg_u32 && need_id)
13224 src_reg->id = ++env->id_gen;
13225 copy_register_state(dst_reg, src_reg);
13226 /* Make sure ID is cleared if src_reg is not in u32
13227 * range otherwise dst_reg min/max could be incorrectly
13228 * propagated into src_reg by find_equal_scalars()
13230 if (!is_src_reg_u32)
13232 dst_reg->live |= REG_LIVE_WRITTEN;
13233 dst_reg->subreg_def = env->insn_idx + 1;
13235 /* case: W1 = (s8, s16)W2 */
13236 bool no_sext = src_reg->umax_value < (1ULL << (insn->off - 1));
13238 if (no_sext && need_id)
13239 src_reg->id = ++env->id_gen;
13240 copy_register_state(dst_reg, src_reg);
13243 dst_reg->live |= REG_LIVE_WRITTEN;
13244 dst_reg->subreg_def = env->insn_idx + 1;
13245 coerce_subreg_to_size_sx(dst_reg, insn->off >> 3);
13248 mark_reg_unknown(env, regs,
13251 zext_32_to_64(dst_reg);
13252 reg_bounds_sync(dst_reg);
13256 * remember the value we stored into this reg
13258 /* clear any state __mark_reg_known doesn't set */
13259 mark_reg_unknown(env, regs, insn->dst_reg);
13260 regs[insn->dst_reg].type = SCALAR_VALUE;
13261 if (BPF_CLASS(insn->code) == BPF_ALU64) {
13262 __mark_reg_known(regs + insn->dst_reg,
13265 __mark_reg_known(regs + insn->dst_reg,
13270 } else if (opcode > BPF_END) {
13271 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
13274 } else { /* all other ALU ops: and, sub, xor, add, ... */
13276 if (BPF_SRC(insn->code) == BPF_X) {
13277 if (insn->imm != 0 || insn->off > 1 ||
13278 (insn->off == 1 && opcode != BPF_MOD && opcode != BPF_DIV)) {
13279 verbose(env, "BPF_ALU uses reserved fields\n");
13282 /* check src1 operand */
13283 err = check_reg_arg(env, insn->src_reg, SRC_OP);
13287 if (insn->src_reg != BPF_REG_0 || insn->off > 1 ||
13288 (insn->off == 1 && opcode != BPF_MOD && opcode != BPF_DIV)) {
13289 verbose(env, "BPF_ALU uses reserved fields\n");
13294 /* check src2 operand */
13295 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
13299 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
13300 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
13301 verbose(env, "div by zero\n");
13305 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
13306 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
13307 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
13309 if (insn->imm < 0 || insn->imm >= size) {
13310 verbose(env, "invalid shift %d\n", insn->imm);
13315 /* check dest operand */
13316 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
13320 return adjust_reg_min_max_vals(env, insn);
13326 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
13327 struct bpf_reg_state *dst_reg,
13328 enum bpf_reg_type type,
13329 bool range_right_open)
13331 struct bpf_func_state *state;
13332 struct bpf_reg_state *reg;
13335 if (dst_reg->off < 0 ||
13336 (dst_reg->off == 0 && range_right_open))
13337 /* This doesn't give us any range */
13340 if (dst_reg->umax_value > MAX_PACKET_OFF ||
13341 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
13342 /* Risk of overflow. For instance, ptr + (1<<63) may be less
13343 * than pkt_end, but that's because it's also less than pkt.
13347 new_range = dst_reg->off;
13348 if (range_right_open)
13351 /* Examples for register markings:
13353 * pkt_data in dst register:
13357 * if (r2 > pkt_end) goto <handle exception>
13362 * if (r2 < pkt_end) goto <access okay>
13363 * <handle exception>
13366 * r2 == dst_reg, pkt_end == src_reg
13367 * r2=pkt(id=n,off=8,r=0)
13368 * r3=pkt(id=n,off=0,r=0)
13370 * pkt_data in src register:
13374 * if (pkt_end >= r2) goto <access okay>
13375 * <handle exception>
13379 * if (pkt_end <= r2) goto <handle exception>
13383 * pkt_end == dst_reg, r2 == src_reg
13384 * r2=pkt(id=n,off=8,r=0)
13385 * r3=pkt(id=n,off=0,r=0)
13387 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
13388 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
13389 * and [r3, r3 + 8-1) respectively is safe to access depending on
13393 /* If our ids match, then we must have the same max_value. And we
13394 * don't care about the other reg's fixed offset, since if it's too big
13395 * the range won't allow anything.
13396 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
13398 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
13399 if (reg->type == type && reg->id == dst_reg->id)
13400 /* keep the maximum range already checked */
13401 reg->range = max(reg->range, new_range);
13405 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
13407 struct tnum subreg = tnum_subreg(reg->var_off);
13408 s32 sval = (s32)val;
13412 if (tnum_is_const(subreg))
13413 return !!tnum_equals_const(subreg, val);
13414 else if (val < reg->u32_min_value || val > reg->u32_max_value)
13418 if (tnum_is_const(subreg))
13419 return !tnum_equals_const(subreg, val);
13420 else if (val < reg->u32_min_value || val > reg->u32_max_value)
13424 if ((~subreg.mask & subreg.value) & val)
13426 if (!((subreg.mask | subreg.value) & val))
13430 if (reg->u32_min_value > val)
13432 else if (reg->u32_max_value <= val)
13436 if (reg->s32_min_value > sval)
13438 else if (reg->s32_max_value <= sval)
13442 if (reg->u32_max_value < val)
13444 else if (reg->u32_min_value >= val)
13448 if (reg->s32_max_value < sval)
13450 else if (reg->s32_min_value >= sval)
13454 if (reg->u32_min_value >= val)
13456 else if (reg->u32_max_value < val)
13460 if (reg->s32_min_value >= sval)
13462 else if (reg->s32_max_value < sval)
13466 if (reg->u32_max_value <= val)
13468 else if (reg->u32_min_value > val)
13472 if (reg->s32_max_value <= sval)
13474 else if (reg->s32_min_value > sval)
13483 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
13485 s64 sval = (s64)val;
13489 if (tnum_is_const(reg->var_off))
13490 return !!tnum_equals_const(reg->var_off, val);
13491 else if (val < reg->umin_value || val > reg->umax_value)
13495 if (tnum_is_const(reg->var_off))
13496 return !tnum_equals_const(reg->var_off, val);
13497 else if (val < reg->umin_value || val > reg->umax_value)
13501 if ((~reg->var_off.mask & reg->var_off.value) & val)
13503 if (!((reg->var_off.mask | reg->var_off.value) & val))
13507 if (reg->umin_value > val)
13509 else if (reg->umax_value <= val)
13513 if (reg->smin_value > sval)
13515 else if (reg->smax_value <= sval)
13519 if (reg->umax_value < val)
13521 else if (reg->umin_value >= val)
13525 if (reg->smax_value < sval)
13527 else if (reg->smin_value >= sval)
13531 if (reg->umin_value >= val)
13533 else if (reg->umax_value < val)
13537 if (reg->smin_value >= sval)
13539 else if (reg->smax_value < sval)
13543 if (reg->umax_value <= val)
13545 else if (reg->umin_value > val)
13549 if (reg->smax_value <= sval)
13551 else if (reg->smin_value > sval)
13559 /* compute branch direction of the expression "if (reg opcode val) goto target;"
13561 * 1 - branch will be taken and "goto target" will be executed
13562 * 0 - branch will not be taken and fall-through to next insn
13563 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
13566 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
13569 if (__is_pointer_value(false, reg)) {
13570 if (!reg_not_null(reg))
13573 /* If pointer is valid tests against zero will fail so we can
13574 * use this to direct branch taken.
13590 return is_branch32_taken(reg, val, opcode);
13591 return is_branch64_taken(reg, val, opcode);
13594 static int flip_opcode(u32 opcode)
13596 /* How can we transform "a <op> b" into "b <op> a"? */
13597 static const u8 opcode_flip[16] = {
13598 /* these stay the same */
13599 [BPF_JEQ >> 4] = BPF_JEQ,
13600 [BPF_JNE >> 4] = BPF_JNE,
13601 [BPF_JSET >> 4] = BPF_JSET,
13602 /* these swap "lesser" and "greater" (L and G in the opcodes) */
13603 [BPF_JGE >> 4] = BPF_JLE,
13604 [BPF_JGT >> 4] = BPF_JLT,
13605 [BPF_JLE >> 4] = BPF_JGE,
13606 [BPF_JLT >> 4] = BPF_JGT,
13607 [BPF_JSGE >> 4] = BPF_JSLE,
13608 [BPF_JSGT >> 4] = BPF_JSLT,
13609 [BPF_JSLE >> 4] = BPF_JSGE,
13610 [BPF_JSLT >> 4] = BPF_JSGT
13612 return opcode_flip[opcode >> 4];
13615 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
13616 struct bpf_reg_state *src_reg,
13619 struct bpf_reg_state *pkt;
13621 if (src_reg->type == PTR_TO_PACKET_END) {
13623 } else if (dst_reg->type == PTR_TO_PACKET_END) {
13625 opcode = flip_opcode(opcode);
13630 if (pkt->range >= 0)
13635 /* pkt <= pkt_end */
13638 /* pkt > pkt_end */
13639 if (pkt->range == BEYOND_PKT_END)
13640 /* pkt has at last one extra byte beyond pkt_end */
13641 return opcode == BPF_JGT;
13644 /* pkt < pkt_end */
13647 /* pkt >= pkt_end */
13648 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
13649 return opcode == BPF_JGE;
13655 /* Adjusts the register min/max values in the case that the dst_reg is the
13656 * variable register that we are working on, and src_reg is a constant or we're
13657 * simply doing a BPF_K check.
13658 * In JEQ/JNE cases we also adjust the var_off values.
13660 static void reg_set_min_max(struct bpf_reg_state *true_reg,
13661 struct bpf_reg_state *false_reg,
13662 u64 val, u32 val32,
13663 u8 opcode, bool is_jmp32)
13665 struct tnum false_32off = tnum_subreg(false_reg->var_off);
13666 struct tnum false_64off = false_reg->var_off;
13667 struct tnum true_32off = tnum_subreg(true_reg->var_off);
13668 struct tnum true_64off = true_reg->var_off;
13669 s64 sval = (s64)val;
13670 s32 sval32 = (s32)val32;
13672 /* If the dst_reg is a pointer, we can't learn anything about its
13673 * variable offset from the compare (unless src_reg were a pointer into
13674 * the same object, but we don't bother with that.
13675 * Since false_reg and true_reg have the same type by construction, we
13676 * only need to check one of them for pointerness.
13678 if (__is_pointer_value(false, false_reg))
13682 /* JEQ/JNE comparison doesn't change the register equivalence.
13685 * if (r1 == 42) goto label;
13687 * label: // here both r1 and r2 are known to be 42.
13689 * Hence when marking register as known preserve it's ID.
13693 __mark_reg32_known(true_reg, val32);
13694 true_32off = tnum_subreg(true_reg->var_off);
13696 ___mark_reg_known(true_reg, val);
13697 true_64off = true_reg->var_off;
13702 __mark_reg32_known(false_reg, val32);
13703 false_32off = tnum_subreg(false_reg->var_off);
13705 ___mark_reg_known(false_reg, val);
13706 false_64off = false_reg->var_off;
13711 false_32off = tnum_and(false_32off, tnum_const(~val32));
13712 if (is_power_of_2(val32))
13713 true_32off = tnum_or(true_32off,
13714 tnum_const(val32));
13716 false_64off = tnum_and(false_64off, tnum_const(~val));
13717 if (is_power_of_2(val))
13718 true_64off = tnum_or(true_64off,
13726 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
13727 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
13729 false_reg->u32_max_value = min(false_reg->u32_max_value,
13731 true_reg->u32_min_value = max(true_reg->u32_min_value,
13734 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
13735 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
13737 false_reg->umax_value = min(false_reg->umax_value, false_umax);
13738 true_reg->umin_value = max(true_reg->umin_value, true_umin);
13746 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
13747 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
13749 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
13750 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
13752 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
13753 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
13755 false_reg->smax_value = min(false_reg->smax_value, false_smax);
13756 true_reg->smin_value = max(true_reg->smin_value, true_smin);
13764 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
13765 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
13767 false_reg->u32_min_value = max(false_reg->u32_min_value,
13769 true_reg->u32_max_value = min(true_reg->u32_max_value,
13772 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
13773 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
13775 false_reg->umin_value = max(false_reg->umin_value, false_umin);
13776 true_reg->umax_value = min(true_reg->umax_value, true_umax);
13784 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
13785 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
13787 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
13788 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
13790 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
13791 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
13793 false_reg->smin_value = max(false_reg->smin_value, false_smin);
13794 true_reg->smax_value = min(true_reg->smax_value, true_smax);
13803 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
13804 tnum_subreg(false_32off));
13805 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
13806 tnum_subreg(true_32off));
13807 __reg_combine_32_into_64(false_reg);
13808 __reg_combine_32_into_64(true_reg);
13810 false_reg->var_off = false_64off;
13811 true_reg->var_off = true_64off;
13812 __reg_combine_64_into_32(false_reg);
13813 __reg_combine_64_into_32(true_reg);
13817 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
13818 * the variable reg.
13820 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
13821 struct bpf_reg_state *false_reg,
13822 u64 val, u32 val32,
13823 u8 opcode, bool is_jmp32)
13825 opcode = flip_opcode(opcode);
13826 /* This uses zero as "not present in table"; luckily the zero opcode,
13827 * BPF_JA, can't get here.
13830 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
13833 /* Regs are known to be equal, so intersect their min/max/var_off */
13834 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
13835 struct bpf_reg_state *dst_reg)
13837 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
13838 dst_reg->umin_value);
13839 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
13840 dst_reg->umax_value);
13841 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
13842 dst_reg->smin_value);
13843 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
13844 dst_reg->smax_value);
13845 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
13847 reg_bounds_sync(src_reg);
13848 reg_bounds_sync(dst_reg);
13851 static void reg_combine_min_max(struct bpf_reg_state *true_src,
13852 struct bpf_reg_state *true_dst,
13853 struct bpf_reg_state *false_src,
13854 struct bpf_reg_state *false_dst,
13859 __reg_combine_min_max(true_src, true_dst);
13862 __reg_combine_min_max(false_src, false_dst);
13867 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
13868 struct bpf_reg_state *reg, u32 id,
13871 if (type_may_be_null(reg->type) && reg->id == id &&
13872 (is_rcu_reg(reg) || !WARN_ON_ONCE(!reg->id))) {
13873 /* Old offset (both fixed and variable parts) should have been
13874 * known-zero, because we don't allow pointer arithmetic on
13875 * pointers that might be NULL. If we see this happening, don't
13876 * convert the register.
13878 * But in some cases, some helpers that return local kptrs
13879 * advance offset for the returned pointer. In those cases, it
13880 * is fine to expect to see reg->off.
13882 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value || !tnum_equals_const(reg->var_off, 0)))
13884 if (!(type_is_ptr_alloc_obj(reg->type) || type_is_non_owning_ref(reg->type)) &&
13885 WARN_ON_ONCE(reg->off))
13889 reg->type = SCALAR_VALUE;
13890 /* We don't need id and ref_obj_id from this point
13891 * onwards anymore, thus we should better reset it,
13892 * so that state pruning has chances to take effect.
13895 reg->ref_obj_id = 0;
13900 mark_ptr_not_null_reg(reg);
13902 if (!reg_may_point_to_spin_lock(reg)) {
13903 /* For not-NULL ptr, reg->ref_obj_id will be reset
13904 * in release_reference().
13906 * reg->id is still used by spin_lock ptr. Other
13907 * than spin_lock ptr type, reg->id can be reset.
13914 /* The logic is similar to find_good_pkt_pointers(), both could eventually
13915 * be folded together at some point.
13917 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
13920 struct bpf_func_state *state = vstate->frame[vstate->curframe];
13921 struct bpf_reg_state *regs = state->regs, *reg;
13922 u32 ref_obj_id = regs[regno].ref_obj_id;
13923 u32 id = regs[regno].id;
13925 if (ref_obj_id && ref_obj_id == id && is_null)
13926 /* regs[regno] is in the " == NULL" branch.
13927 * No one could have freed the reference state before
13928 * doing the NULL check.
13930 WARN_ON_ONCE(release_reference_state(state, id));
13932 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
13933 mark_ptr_or_null_reg(state, reg, id, is_null);
13937 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
13938 struct bpf_reg_state *dst_reg,
13939 struct bpf_reg_state *src_reg,
13940 struct bpf_verifier_state *this_branch,
13941 struct bpf_verifier_state *other_branch)
13943 if (BPF_SRC(insn->code) != BPF_X)
13946 /* Pointers are always 64-bit. */
13947 if (BPF_CLASS(insn->code) == BPF_JMP32)
13950 switch (BPF_OP(insn->code)) {
13952 if ((dst_reg->type == PTR_TO_PACKET &&
13953 src_reg->type == PTR_TO_PACKET_END) ||
13954 (dst_reg->type == PTR_TO_PACKET_META &&
13955 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13956 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
13957 find_good_pkt_pointers(this_branch, dst_reg,
13958 dst_reg->type, false);
13959 mark_pkt_end(other_branch, insn->dst_reg, true);
13960 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
13961 src_reg->type == PTR_TO_PACKET) ||
13962 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
13963 src_reg->type == PTR_TO_PACKET_META)) {
13964 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
13965 find_good_pkt_pointers(other_branch, src_reg,
13966 src_reg->type, true);
13967 mark_pkt_end(this_branch, insn->src_reg, false);
13973 if ((dst_reg->type == PTR_TO_PACKET &&
13974 src_reg->type == PTR_TO_PACKET_END) ||
13975 (dst_reg->type == PTR_TO_PACKET_META &&
13976 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13977 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
13978 find_good_pkt_pointers(other_branch, dst_reg,
13979 dst_reg->type, true);
13980 mark_pkt_end(this_branch, insn->dst_reg, false);
13981 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
13982 src_reg->type == PTR_TO_PACKET) ||
13983 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
13984 src_reg->type == PTR_TO_PACKET_META)) {
13985 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
13986 find_good_pkt_pointers(this_branch, src_reg,
13987 src_reg->type, false);
13988 mark_pkt_end(other_branch, insn->src_reg, true);
13994 if ((dst_reg->type == PTR_TO_PACKET &&
13995 src_reg->type == PTR_TO_PACKET_END) ||
13996 (dst_reg->type == PTR_TO_PACKET_META &&
13997 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13998 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
13999 find_good_pkt_pointers(this_branch, dst_reg,
14000 dst_reg->type, true);
14001 mark_pkt_end(other_branch, insn->dst_reg, false);
14002 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
14003 src_reg->type == PTR_TO_PACKET) ||
14004 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
14005 src_reg->type == PTR_TO_PACKET_META)) {
14006 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
14007 find_good_pkt_pointers(other_branch, src_reg,
14008 src_reg->type, false);
14009 mark_pkt_end(this_branch, insn->src_reg, true);
14015 if ((dst_reg->type == PTR_TO_PACKET &&
14016 src_reg->type == PTR_TO_PACKET_END) ||
14017 (dst_reg->type == PTR_TO_PACKET_META &&
14018 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
14019 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
14020 find_good_pkt_pointers(other_branch, dst_reg,
14021 dst_reg->type, false);
14022 mark_pkt_end(this_branch, insn->dst_reg, true);
14023 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
14024 src_reg->type == PTR_TO_PACKET) ||
14025 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
14026 src_reg->type == PTR_TO_PACKET_META)) {
14027 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
14028 find_good_pkt_pointers(this_branch, src_reg,
14029 src_reg->type, true);
14030 mark_pkt_end(other_branch, insn->src_reg, false);
14042 static void find_equal_scalars(struct bpf_verifier_state *vstate,
14043 struct bpf_reg_state *known_reg)
14045 struct bpf_func_state *state;
14046 struct bpf_reg_state *reg;
14048 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
14049 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
14050 copy_register_state(reg, known_reg);
14054 static int check_cond_jmp_op(struct bpf_verifier_env *env,
14055 struct bpf_insn *insn, int *insn_idx)
14057 struct bpf_verifier_state *this_branch = env->cur_state;
14058 struct bpf_verifier_state *other_branch;
14059 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
14060 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
14061 struct bpf_reg_state *eq_branch_regs;
14062 u8 opcode = BPF_OP(insn->code);
14067 /* Only conditional jumps are expected to reach here. */
14068 if (opcode == BPF_JA || opcode > BPF_JSLE) {
14069 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
14073 /* check src2 operand */
14074 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
14078 dst_reg = ®s[insn->dst_reg];
14079 if (BPF_SRC(insn->code) == BPF_X) {
14080 if (insn->imm != 0) {
14081 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
14085 /* check src1 operand */
14086 err = check_reg_arg(env, insn->src_reg, SRC_OP);
14090 src_reg = ®s[insn->src_reg];
14091 if (!(reg_is_pkt_pointer_any(dst_reg) && reg_is_pkt_pointer_any(src_reg)) &&
14092 is_pointer_value(env, insn->src_reg)) {
14093 verbose(env, "R%d pointer comparison prohibited\n",
14098 if (insn->src_reg != BPF_REG_0) {
14099 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
14104 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
14106 if (BPF_SRC(insn->code) == BPF_K) {
14107 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
14108 } else if (src_reg->type == SCALAR_VALUE &&
14109 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
14110 pred = is_branch_taken(dst_reg,
14111 tnum_subreg(src_reg->var_off).value,
14114 } else if (src_reg->type == SCALAR_VALUE &&
14115 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
14116 pred = is_branch_taken(dst_reg,
14117 src_reg->var_off.value,
14120 } else if (dst_reg->type == SCALAR_VALUE &&
14121 is_jmp32 && tnum_is_const(tnum_subreg(dst_reg->var_off))) {
14122 pred = is_branch_taken(src_reg,
14123 tnum_subreg(dst_reg->var_off).value,
14124 flip_opcode(opcode),
14126 } else if (dst_reg->type == SCALAR_VALUE &&
14127 !is_jmp32 && tnum_is_const(dst_reg->var_off)) {
14128 pred = is_branch_taken(src_reg,
14129 dst_reg->var_off.value,
14130 flip_opcode(opcode),
14132 } else if (reg_is_pkt_pointer_any(dst_reg) &&
14133 reg_is_pkt_pointer_any(src_reg) &&
14135 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
14139 /* If we get here with a dst_reg pointer type it is because
14140 * above is_branch_taken() special cased the 0 comparison.
14142 if (!__is_pointer_value(false, dst_reg))
14143 err = mark_chain_precision(env, insn->dst_reg);
14144 if (BPF_SRC(insn->code) == BPF_X && !err &&
14145 !__is_pointer_value(false, src_reg))
14146 err = mark_chain_precision(env, insn->src_reg);
14152 /* Only follow the goto, ignore fall-through. If needed, push
14153 * the fall-through branch for simulation under speculative
14156 if (!env->bypass_spec_v1 &&
14157 !sanitize_speculative_path(env, insn, *insn_idx + 1,
14160 if (env->log.level & BPF_LOG_LEVEL)
14161 print_insn_state(env, this_branch->frame[this_branch->curframe]);
14162 *insn_idx += insn->off;
14164 } else if (pred == 0) {
14165 /* Only follow the fall-through branch, since that's where the
14166 * program will go. If needed, push the goto branch for
14167 * simulation under speculative execution.
14169 if (!env->bypass_spec_v1 &&
14170 !sanitize_speculative_path(env, insn,
14171 *insn_idx + insn->off + 1,
14174 if (env->log.level & BPF_LOG_LEVEL)
14175 print_insn_state(env, this_branch->frame[this_branch->curframe]);
14179 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
14183 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
14185 /* detect if we are comparing against a constant value so we can adjust
14186 * our min/max values for our dst register.
14187 * this is only legit if both are scalars (or pointers to the same
14188 * object, I suppose, see the PTR_MAYBE_NULL related if block below),
14189 * because otherwise the different base pointers mean the offsets aren't
14192 if (BPF_SRC(insn->code) == BPF_X) {
14193 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
14195 if (dst_reg->type == SCALAR_VALUE &&
14196 src_reg->type == SCALAR_VALUE) {
14197 if (tnum_is_const(src_reg->var_off) ||
14199 tnum_is_const(tnum_subreg(src_reg->var_off))))
14200 reg_set_min_max(&other_branch_regs[insn->dst_reg],
14202 src_reg->var_off.value,
14203 tnum_subreg(src_reg->var_off).value,
14205 else if (tnum_is_const(dst_reg->var_off) ||
14207 tnum_is_const(tnum_subreg(dst_reg->var_off))))
14208 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
14210 dst_reg->var_off.value,
14211 tnum_subreg(dst_reg->var_off).value,
14213 else if (!is_jmp32 &&
14214 (opcode == BPF_JEQ || opcode == BPF_JNE))
14215 /* Comparing for equality, we can combine knowledge */
14216 reg_combine_min_max(&other_branch_regs[insn->src_reg],
14217 &other_branch_regs[insn->dst_reg],
14218 src_reg, dst_reg, opcode);
14220 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
14221 find_equal_scalars(this_branch, src_reg);
14222 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
14226 } else if (dst_reg->type == SCALAR_VALUE) {
14227 reg_set_min_max(&other_branch_regs[insn->dst_reg],
14228 dst_reg, insn->imm, (u32)insn->imm,
14232 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
14233 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
14234 find_equal_scalars(this_branch, dst_reg);
14235 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
14238 /* if one pointer register is compared to another pointer
14239 * register check if PTR_MAYBE_NULL could be lifted.
14240 * E.g. register A - maybe null
14241 * register B - not null
14242 * for JNE A, B, ... - A is not null in the false branch;
14243 * for JEQ A, B, ... - A is not null in the true branch.
14245 * Since PTR_TO_BTF_ID points to a kernel struct that does
14246 * not need to be null checked by the BPF program, i.e.,
14247 * could be null even without PTR_MAYBE_NULL marking, so
14248 * only propagate nullness when neither reg is that type.
14250 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_X &&
14251 __is_pointer_value(false, src_reg) && __is_pointer_value(false, dst_reg) &&
14252 type_may_be_null(src_reg->type) != type_may_be_null(dst_reg->type) &&
14253 base_type(src_reg->type) != PTR_TO_BTF_ID &&
14254 base_type(dst_reg->type) != PTR_TO_BTF_ID) {
14255 eq_branch_regs = NULL;
14258 eq_branch_regs = other_branch_regs;
14261 eq_branch_regs = regs;
14267 if (eq_branch_regs) {
14268 if (type_may_be_null(src_reg->type))
14269 mark_ptr_not_null_reg(&eq_branch_regs[insn->src_reg]);
14271 mark_ptr_not_null_reg(&eq_branch_regs[insn->dst_reg]);
14275 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
14276 * NOTE: these optimizations below are related with pointer comparison
14277 * which will never be JMP32.
14279 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
14280 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
14281 type_may_be_null(dst_reg->type)) {
14282 /* Mark all identical registers in each branch as either
14283 * safe or unknown depending R == 0 or R != 0 conditional.
14285 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
14286 opcode == BPF_JNE);
14287 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
14288 opcode == BPF_JEQ);
14289 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
14290 this_branch, other_branch) &&
14291 is_pointer_value(env, insn->dst_reg)) {
14292 verbose(env, "R%d pointer comparison prohibited\n",
14296 if (env->log.level & BPF_LOG_LEVEL)
14297 print_insn_state(env, this_branch->frame[this_branch->curframe]);
14301 /* verify BPF_LD_IMM64 instruction */
14302 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
14304 struct bpf_insn_aux_data *aux = cur_aux(env);
14305 struct bpf_reg_state *regs = cur_regs(env);
14306 struct bpf_reg_state *dst_reg;
14307 struct bpf_map *map;
14310 if (BPF_SIZE(insn->code) != BPF_DW) {
14311 verbose(env, "invalid BPF_LD_IMM insn\n");
14314 if (insn->off != 0) {
14315 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
14319 err = check_reg_arg(env, insn->dst_reg, DST_OP);
14323 dst_reg = ®s[insn->dst_reg];
14324 if (insn->src_reg == 0) {
14325 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
14327 dst_reg->type = SCALAR_VALUE;
14328 __mark_reg_known(®s[insn->dst_reg], imm);
14332 /* All special src_reg cases are listed below. From this point onwards
14333 * we either succeed and assign a corresponding dst_reg->type after
14334 * zeroing the offset, or fail and reject the program.
14336 mark_reg_known_zero(env, regs, insn->dst_reg);
14338 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
14339 dst_reg->type = aux->btf_var.reg_type;
14340 switch (base_type(dst_reg->type)) {
14342 dst_reg->mem_size = aux->btf_var.mem_size;
14344 case PTR_TO_BTF_ID:
14345 dst_reg->btf = aux->btf_var.btf;
14346 dst_reg->btf_id = aux->btf_var.btf_id;
14349 verbose(env, "bpf verifier is misconfigured\n");
14355 if (insn->src_reg == BPF_PSEUDO_FUNC) {
14356 struct bpf_prog_aux *aux = env->prog->aux;
14357 u32 subprogno = find_subprog(env,
14358 env->insn_idx + insn->imm + 1);
14360 if (!aux->func_info) {
14361 verbose(env, "missing btf func_info\n");
14364 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
14365 verbose(env, "callback function not static\n");
14369 dst_reg->type = PTR_TO_FUNC;
14370 dst_reg->subprogno = subprogno;
14374 map = env->used_maps[aux->map_index];
14375 dst_reg->map_ptr = map;
14377 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
14378 insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
14379 dst_reg->type = PTR_TO_MAP_VALUE;
14380 dst_reg->off = aux->map_off;
14381 WARN_ON_ONCE(map->max_entries != 1);
14382 /* We want reg->id to be same (0) as map_value is not distinct */
14383 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
14384 insn->src_reg == BPF_PSEUDO_MAP_IDX) {
14385 dst_reg->type = CONST_PTR_TO_MAP;
14387 verbose(env, "bpf verifier is misconfigured\n");
14394 static bool may_access_skb(enum bpf_prog_type type)
14397 case BPF_PROG_TYPE_SOCKET_FILTER:
14398 case BPF_PROG_TYPE_SCHED_CLS:
14399 case BPF_PROG_TYPE_SCHED_ACT:
14406 /* verify safety of LD_ABS|LD_IND instructions:
14407 * - they can only appear in the programs where ctx == skb
14408 * - since they are wrappers of function calls, they scratch R1-R5 registers,
14409 * preserve R6-R9, and store return value into R0
14412 * ctx == skb == R6 == CTX
14415 * SRC == any register
14416 * IMM == 32-bit immediate
14419 * R0 - 8/16/32-bit skb data converted to cpu endianness
14421 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
14423 struct bpf_reg_state *regs = cur_regs(env);
14424 static const int ctx_reg = BPF_REG_6;
14425 u8 mode = BPF_MODE(insn->code);
14428 if (!may_access_skb(resolve_prog_type(env->prog))) {
14429 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
14433 if (!env->ops->gen_ld_abs) {
14434 verbose(env, "bpf verifier is misconfigured\n");
14438 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
14439 BPF_SIZE(insn->code) == BPF_DW ||
14440 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
14441 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
14445 /* check whether implicit source operand (register R6) is readable */
14446 err = check_reg_arg(env, ctx_reg, SRC_OP);
14450 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
14451 * gen_ld_abs() may terminate the program at runtime, leading to
14454 err = check_reference_leak(env);
14456 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
14460 if (env->cur_state->active_lock.ptr) {
14461 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
14465 if (env->cur_state->active_rcu_lock) {
14466 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_rcu_read_lock-ed region\n");
14470 if (regs[ctx_reg].type != PTR_TO_CTX) {
14472 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
14476 if (mode == BPF_IND) {
14477 /* check explicit source operand */
14478 err = check_reg_arg(env, insn->src_reg, SRC_OP);
14483 err = check_ptr_off_reg(env, ®s[ctx_reg], ctx_reg);
14487 /* reset caller saved regs to unreadable */
14488 for (i = 0; i < CALLER_SAVED_REGS; i++) {
14489 mark_reg_not_init(env, regs, caller_saved[i]);
14490 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
14493 /* mark destination R0 register as readable, since it contains
14494 * the value fetched from the packet.
14495 * Already marked as written above.
14497 mark_reg_unknown(env, regs, BPF_REG_0);
14498 /* ld_abs load up to 32-bit skb data. */
14499 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
14503 static int check_return_code(struct bpf_verifier_env *env)
14505 struct tnum enforce_attach_type_range = tnum_unknown;
14506 const struct bpf_prog *prog = env->prog;
14507 struct bpf_reg_state *reg;
14508 struct tnum range = tnum_range(0, 1), const_0 = tnum_const(0);
14509 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
14511 struct bpf_func_state *frame = env->cur_state->frame[0];
14512 const bool is_subprog = frame->subprogno;
14514 /* LSM and struct_ops func-ptr's return type could be "void" */
14516 switch (prog_type) {
14517 case BPF_PROG_TYPE_LSM:
14518 if (prog->expected_attach_type == BPF_LSM_CGROUP)
14519 /* See below, can be 0 or 0-1 depending on hook. */
14522 case BPF_PROG_TYPE_STRUCT_OPS:
14523 if (!prog->aux->attach_func_proto->type)
14531 /* eBPF calling convention is such that R0 is used
14532 * to return the value from eBPF program.
14533 * Make sure that it's readable at this time
14534 * of bpf_exit, which means that program wrote
14535 * something into it earlier
14537 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
14541 if (is_pointer_value(env, BPF_REG_0)) {
14542 verbose(env, "R0 leaks addr as return value\n");
14546 reg = cur_regs(env) + BPF_REG_0;
14548 if (frame->in_async_callback_fn) {
14549 /* enforce return zero from async callbacks like timer */
14550 if (reg->type != SCALAR_VALUE) {
14551 verbose(env, "In async callback the register R0 is not a known value (%s)\n",
14552 reg_type_str(env, reg->type));
14556 if (!tnum_in(const_0, reg->var_off)) {
14557 verbose_invalid_scalar(env, reg, &const_0, "async callback", "R0");
14564 if (reg->type != SCALAR_VALUE) {
14565 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
14566 reg_type_str(env, reg->type));
14572 switch (prog_type) {
14573 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
14574 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
14575 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
14576 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
14577 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
14578 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
14579 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
14580 range = tnum_range(1, 1);
14581 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
14582 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
14583 range = tnum_range(0, 3);
14585 case BPF_PROG_TYPE_CGROUP_SKB:
14586 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
14587 range = tnum_range(0, 3);
14588 enforce_attach_type_range = tnum_range(2, 3);
14591 case BPF_PROG_TYPE_CGROUP_SOCK:
14592 case BPF_PROG_TYPE_SOCK_OPS:
14593 case BPF_PROG_TYPE_CGROUP_DEVICE:
14594 case BPF_PROG_TYPE_CGROUP_SYSCTL:
14595 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
14597 case BPF_PROG_TYPE_RAW_TRACEPOINT:
14598 if (!env->prog->aux->attach_btf_id)
14600 range = tnum_const(0);
14602 case BPF_PROG_TYPE_TRACING:
14603 switch (env->prog->expected_attach_type) {
14604 case BPF_TRACE_FENTRY:
14605 case BPF_TRACE_FEXIT:
14606 range = tnum_const(0);
14608 case BPF_TRACE_RAW_TP:
14609 case BPF_MODIFY_RETURN:
14611 case BPF_TRACE_ITER:
14617 case BPF_PROG_TYPE_SK_LOOKUP:
14618 range = tnum_range(SK_DROP, SK_PASS);
14621 case BPF_PROG_TYPE_LSM:
14622 if (env->prog->expected_attach_type != BPF_LSM_CGROUP) {
14623 /* Regular BPF_PROG_TYPE_LSM programs can return
14628 if (!env->prog->aux->attach_func_proto->type) {
14629 /* Make sure programs that attach to void
14630 * hooks don't try to modify return value.
14632 range = tnum_range(1, 1);
14636 case BPF_PROG_TYPE_NETFILTER:
14637 range = tnum_range(NF_DROP, NF_ACCEPT);
14639 case BPF_PROG_TYPE_EXT:
14640 /* freplace program can return anything as its return value
14641 * depends on the to-be-replaced kernel func or bpf program.
14647 if (reg->type != SCALAR_VALUE) {
14648 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
14649 reg_type_str(env, reg->type));
14653 if (!tnum_in(range, reg->var_off)) {
14654 verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
14655 if (prog->expected_attach_type == BPF_LSM_CGROUP &&
14656 prog_type == BPF_PROG_TYPE_LSM &&
14657 !prog->aux->attach_func_proto->type)
14658 verbose(env, "Note, BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
14662 if (!tnum_is_unknown(enforce_attach_type_range) &&
14663 tnum_in(enforce_attach_type_range, reg->var_off))
14664 env->prog->enforce_expected_attach_type = 1;
14668 /* non-recursive DFS pseudo code
14669 * 1 procedure DFS-iterative(G,v):
14670 * 2 label v as discovered
14671 * 3 let S be a stack
14673 * 5 while S is not empty
14675 * 7 if t is what we're looking for:
14677 * 9 for all edges e in G.adjacentEdges(t) do
14678 * 10 if edge e is already labelled
14679 * 11 continue with the next edge
14680 * 12 w <- G.adjacentVertex(t,e)
14681 * 13 if vertex w is not discovered and not explored
14682 * 14 label e as tree-edge
14683 * 15 label w as discovered
14686 * 18 else if vertex w is discovered
14687 * 19 label e as back-edge
14689 * 21 // vertex w is explored
14690 * 22 label e as forward- or cross-edge
14691 * 23 label t as explored
14695 * 0x10 - discovered
14696 * 0x11 - discovered and fall-through edge labelled
14697 * 0x12 - discovered and fall-through and branch edges labelled
14708 static u32 state_htab_size(struct bpf_verifier_env *env)
14710 return env->prog->len;
14713 static struct bpf_verifier_state_list **explored_state(
14714 struct bpf_verifier_env *env,
14717 struct bpf_verifier_state *cur = env->cur_state;
14718 struct bpf_func_state *state = cur->frame[cur->curframe];
14720 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
14723 static void mark_prune_point(struct bpf_verifier_env *env, int idx)
14725 env->insn_aux_data[idx].prune_point = true;
14728 static bool is_prune_point(struct bpf_verifier_env *env, int insn_idx)
14730 return env->insn_aux_data[insn_idx].prune_point;
14733 static void mark_force_checkpoint(struct bpf_verifier_env *env, int idx)
14735 env->insn_aux_data[idx].force_checkpoint = true;
14738 static bool is_force_checkpoint(struct bpf_verifier_env *env, int insn_idx)
14740 return env->insn_aux_data[insn_idx].force_checkpoint;
14745 DONE_EXPLORING = 0,
14746 KEEP_EXPLORING = 1,
14749 /* t, w, e - match pseudo-code above:
14750 * t - index of current instruction
14751 * w - next instruction
14754 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
14757 int *insn_stack = env->cfg.insn_stack;
14758 int *insn_state = env->cfg.insn_state;
14760 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
14761 return DONE_EXPLORING;
14763 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
14764 return DONE_EXPLORING;
14766 if (w < 0 || w >= env->prog->len) {
14767 verbose_linfo(env, t, "%d: ", t);
14768 verbose(env, "jump out of range from insn %d to %d\n", t, w);
14773 /* mark branch target for state pruning */
14774 mark_prune_point(env, w);
14775 mark_jmp_point(env, w);
14778 if (insn_state[w] == 0) {
14780 insn_state[t] = DISCOVERED | e;
14781 insn_state[w] = DISCOVERED;
14782 if (env->cfg.cur_stack >= env->prog->len)
14784 insn_stack[env->cfg.cur_stack++] = w;
14785 return KEEP_EXPLORING;
14786 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
14787 if (loop_ok && env->bpf_capable)
14788 return DONE_EXPLORING;
14789 verbose_linfo(env, t, "%d: ", t);
14790 verbose_linfo(env, w, "%d: ", w);
14791 verbose(env, "back-edge from insn %d to %d\n", t, w);
14793 } else if (insn_state[w] == EXPLORED) {
14794 /* forward- or cross-edge */
14795 insn_state[t] = DISCOVERED | e;
14797 verbose(env, "insn state internal bug\n");
14800 return DONE_EXPLORING;
14803 static int visit_func_call_insn(int t, struct bpf_insn *insns,
14804 struct bpf_verifier_env *env,
14809 insn_sz = bpf_is_ldimm64(&insns[t]) ? 2 : 1;
14810 ret = push_insn(t, t + insn_sz, FALLTHROUGH, env, false);
14814 mark_prune_point(env, t + insn_sz);
14815 /* when we exit from subprog, we need to record non-linear history */
14816 mark_jmp_point(env, t + insn_sz);
14818 if (visit_callee) {
14819 mark_prune_point(env, t);
14820 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env,
14821 /* It's ok to allow recursion from CFG point of
14822 * view. __check_func_call() will do the actual
14825 bpf_pseudo_func(insns + t));
14830 /* Visits the instruction at index t and returns one of the following:
14831 * < 0 - an error occurred
14832 * DONE_EXPLORING - the instruction was fully explored
14833 * KEEP_EXPLORING - there is still work to be done before it is fully explored
14835 static int visit_insn(int t, struct bpf_verifier_env *env)
14837 struct bpf_insn *insns = env->prog->insnsi, *insn = &insns[t];
14838 int ret, off, insn_sz;
14840 if (bpf_pseudo_func(insn))
14841 return visit_func_call_insn(t, insns, env, true);
14843 /* All non-branch instructions have a single fall-through edge. */
14844 if (BPF_CLASS(insn->code) != BPF_JMP &&
14845 BPF_CLASS(insn->code) != BPF_JMP32) {
14846 insn_sz = bpf_is_ldimm64(insn) ? 2 : 1;
14847 return push_insn(t, t + insn_sz, FALLTHROUGH, env, false);
14850 switch (BPF_OP(insn->code)) {
14852 return DONE_EXPLORING;
14855 if (insn->src_reg == 0 && insn->imm == BPF_FUNC_timer_set_callback)
14856 /* Mark this call insn as a prune point to trigger
14857 * is_state_visited() check before call itself is
14858 * processed by __check_func_call(). Otherwise new
14859 * async state will be pushed for further exploration.
14861 mark_prune_point(env, t);
14862 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
14863 struct bpf_kfunc_call_arg_meta meta;
14865 ret = fetch_kfunc_meta(env, insn, &meta, NULL);
14866 if (ret == 0 && is_iter_next_kfunc(&meta)) {
14867 mark_prune_point(env, t);
14868 /* Checking and saving state checkpoints at iter_next() call
14869 * is crucial for fast convergence of open-coded iterator loop
14870 * logic, so we need to force it. If we don't do that,
14871 * is_state_visited() might skip saving a checkpoint, causing
14872 * unnecessarily long sequence of not checkpointed
14873 * instructions and jumps, leading to exhaustion of jump
14874 * history buffer, and potentially other undesired outcomes.
14875 * It is expected that with correct open-coded iterators
14876 * convergence will happen quickly, so we don't run a risk of
14877 * exhausting memory.
14879 mark_force_checkpoint(env, t);
14882 return visit_func_call_insn(t, insns, env, insn->src_reg == BPF_PSEUDO_CALL);
14885 if (BPF_SRC(insn->code) != BPF_K)
14888 if (BPF_CLASS(insn->code) == BPF_JMP)
14893 /* unconditional jump with single edge */
14894 ret = push_insn(t, t + off + 1, FALLTHROUGH, env,
14899 mark_prune_point(env, t + off + 1);
14900 mark_jmp_point(env, t + off + 1);
14905 /* conditional jump with two edges */
14906 mark_prune_point(env, t);
14908 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
14912 return push_insn(t, t + insn->off + 1, BRANCH, env, true);
14916 /* non-recursive depth-first-search to detect loops in BPF program
14917 * loop == back-edge in directed graph
14919 static int check_cfg(struct bpf_verifier_env *env)
14921 int insn_cnt = env->prog->len;
14922 int *insn_stack, *insn_state;
14926 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
14930 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
14932 kvfree(insn_state);
14936 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
14937 insn_stack[0] = 0; /* 0 is the first instruction */
14938 env->cfg.cur_stack = 1;
14940 while (env->cfg.cur_stack > 0) {
14941 int t = insn_stack[env->cfg.cur_stack - 1];
14943 ret = visit_insn(t, env);
14945 case DONE_EXPLORING:
14946 insn_state[t] = EXPLORED;
14947 env->cfg.cur_stack--;
14949 case KEEP_EXPLORING:
14953 verbose(env, "visit_insn internal bug\n");
14960 if (env->cfg.cur_stack < 0) {
14961 verbose(env, "pop stack internal bug\n");
14966 for (i = 0; i < insn_cnt; i++) {
14967 struct bpf_insn *insn = &env->prog->insnsi[i];
14969 if (insn_state[i] != EXPLORED) {
14970 verbose(env, "unreachable insn %d\n", i);
14974 if (bpf_is_ldimm64(insn)) {
14975 if (insn_state[i + 1] != 0) {
14976 verbose(env, "jump into the middle of ldimm64 insn %d\n", i);
14980 i++; /* skip second half of ldimm64 */
14983 ret = 0; /* cfg looks good */
14986 kvfree(insn_state);
14987 kvfree(insn_stack);
14988 env->cfg.insn_state = env->cfg.insn_stack = NULL;
14992 static int check_abnormal_return(struct bpf_verifier_env *env)
14996 for (i = 1; i < env->subprog_cnt; i++) {
14997 if (env->subprog_info[i].has_ld_abs) {
14998 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
15001 if (env->subprog_info[i].has_tail_call) {
15002 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
15009 /* The minimum supported BTF func info size */
15010 #define MIN_BPF_FUNCINFO_SIZE 8
15011 #define MAX_FUNCINFO_REC_SIZE 252
15013 static int check_btf_func(struct bpf_verifier_env *env,
15014 const union bpf_attr *attr,
15017 const struct btf_type *type, *func_proto, *ret_type;
15018 u32 i, nfuncs, urec_size, min_size;
15019 u32 krec_size = sizeof(struct bpf_func_info);
15020 struct bpf_func_info *krecord;
15021 struct bpf_func_info_aux *info_aux = NULL;
15022 struct bpf_prog *prog;
15023 const struct btf *btf;
15025 u32 prev_offset = 0;
15026 bool scalar_return;
15029 nfuncs = attr->func_info_cnt;
15031 if (check_abnormal_return(env))
15036 if (nfuncs != env->subprog_cnt) {
15037 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
15041 urec_size = attr->func_info_rec_size;
15042 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
15043 urec_size > MAX_FUNCINFO_REC_SIZE ||
15044 urec_size % sizeof(u32)) {
15045 verbose(env, "invalid func info rec size %u\n", urec_size);
15050 btf = prog->aux->btf;
15052 urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
15053 min_size = min_t(u32, krec_size, urec_size);
15055 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
15058 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
15062 for (i = 0; i < nfuncs; i++) {
15063 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
15065 if (ret == -E2BIG) {
15066 verbose(env, "nonzero tailing record in func info");
15067 /* set the size kernel expects so loader can zero
15068 * out the rest of the record.
15070 if (copy_to_bpfptr_offset(uattr,
15071 offsetof(union bpf_attr, func_info_rec_size),
15072 &min_size, sizeof(min_size)))
15078 if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
15083 /* check insn_off */
15086 if (krecord[i].insn_off) {
15088 "nonzero insn_off %u for the first func info record",
15089 krecord[i].insn_off);
15092 } else if (krecord[i].insn_off <= prev_offset) {
15094 "same or smaller insn offset (%u) than previous func info record (%u)",
15095 krecord[i].insn_off, prev_offset);
15099 if (env->subprog_info[i].start != krecord[i].insn_off) {
15100 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
15104 /* check type_id */
15105 type = btf_type_by_id(btf, krecord[i].type_id);
15106 if (!type || !btf_type_is_func(type)) {
15107 verbose(env, "invalid type id %d in func info",
15108 krecord[i].type_id);
15111 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
15113 func_proto = btf_type_by_id(btf, type->type);
15114 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
15115 /* btf_func_check() already verified it during BTF load */
15117 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
15119 btf_type_is_small_int(ret_type) || btf_is_any_enum(ret_type);
15120 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
15121 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
15124 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
15125 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
15129 prev_offset = krecord[i].insn_off;
15130 bpfptr_add(&urecord, urec_size);
15133 prog->aux->func_info = krecord;
15134 prog->aux->func_info_cnt = nfuncs;
15135 prog->aux->func_info_aux = info_aux;
15144 static void adjust_btf_func(struct bpf_verifier_env *env)
15146 struct bpf_prog_aux *aux = env->prog->aux;
15149 if (!aux->func_info)
15152 for (i = 0; i < env->subprog_cnt; i++)
15153 aux->func_info[i].insn_off = env->subprog_info[i].start;
15156 #define MIN_BPF_LINEINFO_SIZE offsetofend(struct bpf_line_info, line_col)
15157 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
15159 static int check_btf_line(struct bpf_verifier_env *env,
15160 const union bpf_attr *attr,
15163 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
15164 struct bpf_subprog_info *sub;
15165 struct bpf_line_info *linfo;
15166 struct bpf_prog *prog;
15167 const struct btf *btf;
15171 nr_linfo = attr->line_info_cnt;
15174 if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info))
15177 rec_size = attr->line_info_rec_size;
15178 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
15179 rec_size > MAX_LINEINFO_REC_SIZE ||
15180 rec_size & (sizeof(u32) - 1))
15183 /* Need to zero it in case the userspace may
15184 * pass in a smaller bpf_line_info object.
15186 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
15187 GFP_KERNEL | __GFP_NOWARN);
15192 btf = prog->aux->btf;
15195 sub = env->subprog_info;
15196 ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
15197 expected_size = sizeof(struct bpf_line_info);
15198 ncopy = min_t(u32, expected_size, rec_size);
15199 for (i = 0; i < nr_linfo; i++) {
15200 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
15202 if (err == -E2BIG) {
15203 verbose(env, "nonzero tailing record in line_info");
15204 if (copy_to_bpfptr_offset(uattr,
15205 offsetof(union bpf_attr, line_info_rec_size),
15206 &expected_size, sizeof(expected_size)))
15212 if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
15218 * Check insn_off to ensure
15219 * 1) strictly increasing AND
15220 * 2) bounded by prog->len
15222 * The linfo[0].insn_off == 0 check logically falls into
15223 * the later "missing bpf_line_info for func..." case
15224 * because the first linfo[0].insn_off must be the
15225 * first sub also and the first sub must have
15226 * subprog_info[0].start == 0.
15228 if ((i && linfo[i].insn_off <= prev_offset) ||
15229 linfo[i].insn_off >= prog->len) {
15230 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
15231 i, linfo[i].insn_off, prev_offset,
15237 if (!prog->insnsi[linfo[i].insn_off].code) {
15239 "Invalid insn code at line_info[%u].insn_off\n",
15245 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
15246 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
15247 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
15252 if (s != env->subprog_cnt) {
15253 if (linfo[i].insn_off == sub[s].start) {
15254 sub[s].linfo_idx = i;
15256 } else if (sub[s].start < linfo[i].insn_off) {
15257 verbose(env, "missing bpf_line_info for func#%u\n", s);
15263 prev_offset = linfo[i].insn_off;
15264 bpfptr_add(&ulinfo, rec_size);
15267 if (s != env->subprog_cnt) {
15268 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
15269 env->subprog_cnt - s, s);
15274 prog->aux->linfo = linfo;
15275 prog->aux->nr_linfo = nr_linfo;
15284 #define MIN_CORE_RELO_SIZE sizeof(struct bpf_core_relo)
15285 #define MAX_CORE_RELO_SIZE MAX_FUNCINFO_REC_SIZE
15287 static int check_core_relo(struct bpf_verifier_env *env,
15288 const union bpf_attr *attr,
15291 u32 i, nr_core_relo, ncopy, expected_size, rec_size;
15292 struct bpf_core_relo core_relo = {};
15293 struct bpf_prog *prog = env->prog;
15294 const struct btf *btf = prog->aux->btf;
15295 struct bpf_core_ctx ctx = {
15299 bpfptr_t u_core_relo;
15302 nr_core_relo = attr->core_relo_cnt;
15305 if (nr_core_relo > INT_MAX / sizeof(struct bpf_core_relo))
15308 rec_size = attr->core_relo_rec_size;
15309 if (rec_size < MIN_CORE_RELO_SIZE ||
15310 rec_size > MAX_CORE_RELO_SIZE ||
15311 rec_size % sizeof(u32))
15314 u_core_relo = make_bpfptr(attr->core_relos, uattr.is_kernel);
15315 expected_size = sizeof(struct bpf_core_relo);
15316 ncopy = min_t(u32, expected_size, rec_size);
15318 /* Unlike func_info and line_info, copy and apply each CO-RE
15319 * relocation record one at a time.
15321 for (i = 0; i < nr_core_relo; i++) {
15322 /* future proofing when sizeof(bpf_core_relo) changes */
15323 err = bpf_check_uarg_tail_zero(u_core_relo, expected_size, rec_size);
15325 if (err == -E2BIG) {
15326 verbose(env, "nonzero tailing record in core_relo");
15327 if (copy_to_bpfptr_offset(uattr,
15328 offsetof(union bpf_attr, core_relo_rec_size),
15329 &expected_size, sizeof(expected_size)))
15335 if (copy_from_bpfptr(&core_relo, u_core_relo, ncopy)) {
15340 if (core_relo.insn_off % 8 || core_relo.insn_off / 8 >= prog->len) {
15341 verbose(env, "Invalid core_relo[%u].insn_off:%u prog->len:%u\n",
15342 i, core_relo.insn_off, prog->len);
15347 err = bpf_core_apply(&ctx, &core_relo, i,
15348 &prog->insnsi[core_relo.insn_off / 8]);
15351 bpfptr_add(&u_core_relo, rec_size);
15356 static int check_btf_info(struct bpf_verifier_env *env,
15357 const union bpf_attr *attr,
15363 if (!attr->func_info_cnt && !attr->line_info_cnt) {
15364 if (check_abnormal_return(env))
15369 btf = btf_get_by_fd(attr->prog_btf_fd);
15371 return PTR_ERR(btf);
15372 if (btf_is_kernel(btf)) {
15376 env->prog->aux->btf = btf;
15378 err = check_btf_func(env, attr, uattr);
15382 err = check_btf_line(env, attr, uattr);
15386 err = check_core_relo(env, attr, uattr);
15393 /* check %cur's range satisfies %old's */
15394 static bool range_within(struct bpf_reg_state *old,
15395 struct bpf_reg_state *cur)
15397 return old->umin_value <= cur->umin_value &&
15398 old->umax_value >= cur->umax_value &&
15399 old->smin_value <= cur->smin_value &&
15400 old->smax_value >= cur->smax_value &&
15401 old->u32_min_value <= cur->u32_min_value &&
15402 old->u32_max_value >= cur->u32_max_value &&
15403 old->s32_min_value <= cur->s32_min_value &&
15404 old->s32_max_value >= cur->s32_max_value;
15407 /* If in the old state two registers had the same id, then they need to have
15408 * the same id in the new state as well. But that id could be different from
15409 * the old state, so we need to track the mapping from old to new ids.
15410 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
15411 * regs with old id 5 must also have new id 9 for the new state to be safe. But
15412 * regs with a different old id could still have new id 9, we don't care about
15414 * So we look through our idmap to see if this old id has been seen before. If
15415 * so, we require the new id to match; otherwise, we add the id pair to the map.
15417 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap)
15419 struct bpf_id_pair *map = idmap->map;
15422 /* either both IDs should be set or both should be zero */
15423 if (!!old_id != !!cur_id)
15426 if (old_id == 0) /* cur_id == 0 as well */
15429 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
15431 /* Reached an empty slot; haven't seen this id before */
15432 map[i].old = old_id;
15433 map[i].cur = cur_id;
15436 if (map[i].old == old_id)
15437 return map[i].cur == cur_id;
15438 if (map[i].cur == cur_id)
15441 /* We ran out of idmap slots, which should be impossible */
15446 /* Similar to check_ids(), but allocate a unique temporary ID
15447 * for 'old_id' or 'cur_id' of zero.
15448 * This makes pairs like '0 vs unique ID', 'unique ID vs 0' valid.
15450 static bool check_scalar_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap)
15452 old_id = old_id ? old_id : ++idmap->tmp_id_gen;
15453 cur_id = cur_id ? cur_id : ++idmap->tmp_id_gen;
15455 return check_ids(old_id, cur_id, idmap);
15458 static void clean_func_state(struct bpf_verifier_env *env,
15459 struct bpf_func_state *st)
15461 enum bpf_reg_liveness live;
15464 for (i = 0; i < BPF_REG_FP; i++) {
15465 live = st->regs[i].live;
15466 /* liveness must not touch this register anymore */
15467 st->regs[i].live |= REG_LIVE_DONE;
15468 if (!(live & REG_LIVE_READ))
15469 /* since the register is unused, clear its state
15470 * to make further comparison simpler
15472 __mark_reg_not_init(env, &st->regs[i]);
15475 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
15476 live = st->stack[i].spilled_ptr.live;
15477 /* liveness must not touch this stack slot anymore */
15478 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
15479 if (!(live & REG_LIVE_READ)) {
15480 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
15481 for (j = 0; j < BPF_REG_SIZE; j++)
15482 st->stack[i].slot_type[j] = STACK_INVALID;
15487 static void clean_verifier_state(struct bpf_verifier_env *env,
15488 struct bpf_verifier_state *st)
15492 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
15493 /* all regs in this state in all frames were already marked */
15496 for (i = 0; i <= st->curframe; i++)
15497 clean_func_state(env, st->frame[i]);
15500 /* the parentage chains form a tree.
15501 * the verifier states are added to state lists at given insn and
15502 * pushed into state stack for future exploration.
15503 * when the verifier reaches bpf_exit insn some of the verifer states
15504 * stored in the state lists have their final liveness state already,
15505 * but a lot of states will get revised from liveness point of view when
15506 * the verifier explores other branches.
15509 * 2: if r1 == 100 goto pc+1
15512 * when the verifier reaches exit insn the register r0 in the state list of
15513 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
15514 * of insn 2 and goes exploring further. At the insn 4 it will walk the
15515 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
15517 * Since the verifier pushes the branch states as it sees them while exploring
15518 * the program the condition of walking the branch instruction for the second
15519 * time means that all states below this branch were already explored and
15520 * their final liveness marks are already propagated.
15521 * Hence when the verifier completes the search of state list in is_state_visited()
15522 * we can call this clean_live_states() function to mark all liveness states
15523 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
15524 * will not be used.
15525 * This function also clears the registers and stack for states that !READ
15526 * to simplify state merging.
15528 * Important note here that walking the same branch instruction in the callee
15529 * doesn't meant that the states are DONE. The verifier has to compare
15532 static void clean_live_states(struct bpf_verifier_env *env, int insn,
15533 struct bpf_verifier_state *cur)
15535 struct bpf_verifier_state_list *sl;
15538 sl = *explored_state(env, insn);
15540 if (sl->state.branches)
15542 if (sl->state.insn_idx != insn ||
15543 sl->state.curframe != cur->curframe)
15545 for (i = 0; i <= cur->curframe; i++)
15546 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
15548 clean_verifier_state(env, &sl->state);
15554 static bool regs_exact(const struct bpf_reg_state *rold,
15555 const struct bpf_reg_state *rcur,
15556 struct bpf_idmap *idmap)
15558 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
15559 check_ids(rold->id, rcur->id, idmap) &&
15560 check_ids(rold->ref_obj_id, rcur->ref_obj_id, idmap);
15563 /* Returns true if (rold safe implies rcur safe) */
15564 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
15565 struct bpf_reg_state *rcur, struct bpf_idmap *idmap)
15567 if (!(rold->live & REG_LIVE_READ))
15568 /* explored state didn't use this */
15570 if (rold->type == NOT_INIT)
15571 /* explored state can't have used this */
15573 if (rcur->type == NOT_INIT)
15576 /* Enforce that register types have to match exactly, including their
15577 * modifiers (like PTR_MAYBE_NULL, MEM_RDONLY, etc), as a general
15580 * One can make a point that using a pointer register as unbounded
15581 * SCALAR would be technically acceptable, but this could lead to
15582 * pointer leaks because scalars are allowed to leak while pointers
15583 * are not. We could make this safe in special cases if root is
15584 * calling us, but it's probably not worth the hassle.
15586 * Also, register types that are *not* MAYBE_NULL could technically be
15587 * safe to use as their MAYBE_NULL variants (e.g., PTR_TO_MAP_VALUE
15588 * is safe to be used as PTR_TO_MAP_VALUE_OR_NULL, provided both point
15589 * to the same map).
15590 * However, if the old MAYBE_NULL register then got NULL checked,
15591 * doing so could have affected others with the same id, and we can't
15592 * check for that because we lost the id when we converted to
15593 * a non-MAYBE_NULL variant.
15594 * So, as a general rule we don't allow mixing MAYBE_NULL and
15595 * non-MAYBE_NULL registers as well.
15597 if (rold->type != rcur->type)
15600 switch (base_type(rold->type)) {
15602 if (env->explore_alu_limits) {
15603 /* explore_alu_limits disables tnum_in() and range_within()
15604 * logic and requires everything to be strict
15606 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
15607 check_scalar_ids(rold->id, rcur->id, idmap);
15609 if (!rold->precise)
15611 /* Why check_ids() for scalar registers?
15613 * Consider the following BPF code:
15614 * 1: r6 = ... unbound scalar, ID=a ...
15615 * 2: r7 = ... unbound scalar, ID=b ...
15616 * 3: if (r6 > r7) goto +1
15618 * 5: if (r6 > X) goto ...
15619 * 6: ... memory operation using r7 ...
15621 * First verification path is [1-6]:
15622 * - at (4) same bpf_reg_state::id (b) would be assigned to r6 and r7;
15623 * - at (5) r6 would be marked <= X, find_equal_scalars() would also mark
15624 * r7 <= X, because r6 and r7 share same id.
15625 * Next verification path is [1-4, 6].
15627 * Instruction (6) would be reached in two states:
15628 * I. r6{.id=b}, r7{.id=b} via path 1-6;
15629 * II. r6{.id=a}, r7{.id=b} via path 1-4, 6.
15631 * Use check_ids() to distinguish these states.
15633 * Also verify that new value satisfies old value range knowledge.
15635 return range_within(rold, rcur) &&
15636 tnum_in(rold->var_off, rcur->var_off) &&
15637 check_scalar_ids(rold->id, rcur->id, idmap);
15638 case PTR_TO_MAP_KEY:
15639 case PTR_TO_MAP_VALUE:
15642 case PTR_TO_TP_BUFFER:
15643 /* If the new min/max/var_off satisfy the old ones and
15644 * everything else matches, we are OK.
15646 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, var_off)) == 0 &&
15647 range_within(rold, rcur) &&
15648 tnum_in(rold->var_off, rcur->var_off) &&
15649 check_ids(rold->id, rcur->id, idmap) &&
15650 check_ids(rold->ref_obj_id, rcur->ref_obj_id, idmap);
15651 case PTR_TO_PACKET_META:
15652 case PTR_TO_PACKET:
15653 /* We must have at least as much range as the old ptr
15654 * did, so that any accesses which were safe before are
15655 * still safe. This is true even if old range < old off,
15656 * since someone could have accessed through (ptr - k), or
15657 * even done ptr -= k in a register, to get a safe access.
15659 if (rold->range > rcur->range)
15661 /* If the offsets don't match, we can't trust our alignment;
15662 * nor can we be sure that we won't fall out of range.
15664 if (rold->off != rcur->off)
15666 /* id relations must be preserved */
15667 if (!check_ids(rold->id, rcur->id, idmap))
15669 /* new val must satisfy old val knowledge */
15670 return range_within(rold, rcur) &&
15671 tnum_in(rold->var_off, rcur->var_off);
15673 /* two stack pointers are equal only if they're pointing to
15674 * the same stack frame, since fp-8 in foo != fp-8 in bar
15676 return regs_exact(rold, rcur, idmap) && rold->frameno == rcur->frameno;
15678 return regs_exact(rold, rcur, idmap);
15682 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
15683 struct bpf_func_state *cur, struct bpf_idmap *idmap)
15687 /* walk slots of the explored stack and ignore any additional
15688 * slots in the current stack, since explored(safe) state
15691 for (i = 0; i < old->allocated_stack; i++) {
15692 struct bpf_reg_state *old_reg, *cur_reg;
15694 spi = i / BPF_REG_SIZE;
15696 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
15697 i += BPF_REG_SIZE - 1;
15698 /* explored state didn't use this */
15702 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
15705 if (env->allow_uninit_stack &&
15706 old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC)
15709 /* explored stack has more populated slots than current stack
15710 * and these slots were used
15712 if (i >= cur->allocated_stack)
15715 /* if old state was safe with misc data in the stack
15716 * it will be safe with zero-initialized stack.
15717 * The opposite is not true
15719 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
15720 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
15722 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
15723 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
15724 /* Ex: old explored (safe) state has STACK_SPILL in
15725 * this stack slot, but current has STACK_MISC ->
15726 * this verifier states are not equivalent,
15727 * return false to continue verification of this path
15730 if (i % BPF_REG_SIZE != BPF_REG_SIZE - 1)
15732 /* Both old and cur are having same slot_type */
15733 switch (old->stack[spi].slot_type[BPF_REG_SIZE - 1]) {
15735 /* when explored and current stack slot are both storing
15736 * spilled registers, check that stored pointers types
15737 * are the same as well.
15738 * Ex: explored safe path could have stored
15739 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
15740 * but current path has stored:
15741 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
15742 * such verifier states are not equivalent.
15743 * return false to continue verification of this path
15745 if (!regsafe(env, &old->stack[spi].spilled_ptr,
15746 &cur->stack[spi].spilled_ptr, idmap))
15750 old_reg = &old->stack[spi].spilled_ptr;
15751 cur_reg = &cur->stack[spi].spilled_ptr;
15752 if (old_reg->dynptr.type != cur_reg->dynptr.type ||
15753 old_reg->dynptr.first_slot != cur_reg->dynptr.first_slot ||
15754 !check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap))
15758 old_reg = &old->stack[spi].spilled_ptr;
15759 cur_reg = &cur->stack[spi].spilled_ptr;
15760 /* iter.depth is not compared between states as it
15761 * doesn't matter for correctness and would otherwise
15762 * prevent convergence; we maintain it only to prevent
15763 * infinite loop check triggering, see
15764 * iter_active_depths_differ()
15766 if (old_reg->iter.btf != cur_reg->iter.btf ||
15767 old_reg->iter.btf_id != cur_reg->iter.btf_id ||
15768 old_reg->iter.state != cur_reg->iter.state ||
15769 /* ignore {old_reg,cur_reg}->iter.depth, see above */
15770 !check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap))
15775 case STACK_INVALID:
15777 /* Ensure that new unhandled slot types return false by default */
15785 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur,
15786 struct bpf_idmap *idmap)
15790 if (old->acquired_refs != cur->acquired_refs)
15793 for (i = 0; i < old->acquired_refs; i++) {
15794 if (!check_ids(old->refs[i].id, cur->refs[i].id, idmap))
15801 /* compare two verifier states
15803 * all states stored in state_list are known to be valid, since
15804 * verifier reached 'bpf_exit' instruction through them
15806 * this function is called when verifier exploring different branches of
15807 * execution popped from the state stack. If it sees an old state that has
15808 * more strict register state and more strict stack state then this execution
15809 * branch doesn't need to be explored further, since verifier already
15810 * concluded that more strict state leads to valid finish.
15812 * Therefore two states are equivalent if register state is more conservative
15813 * and explored stack state is more conservative than the current one.
15816 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
15817 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
15819 * In other words if current stack state (one being explored) has more
15820 * valid slots than old one that already passed validation, it means
15821 * the verifier can stop exploring and conclude that current state is valid too
15823 * Similarly with registers. If explored state has register type as invalid
15824 * whereas register type in current state is meaningful, it means that
15825 * the current state will reach 'bpf_exit' instruction safely
15827 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
15828 struct bpf_func_state *cur)
15832 for (i = 0; i < MAX_BPF_REG; i++)
15833 if (!regsafe(env, &old->regs[i], &cur->regs[i],
15834 &env->idmap_scratch))
15837 if (!stacksafe(env, old, cur, &env->idmap_scratch))
15840 if (!refsafe(old, cur, &env->idmap_scratch))
15846 static bool states_equal(struct bpf_verifier_env *env,
15847 struct bpf_verifier_state *old,
15848 struct bpf_verifier_state *cur)
15852 if (old->curframe != cur->curframe)
15855 env->idmap_scratch.tmp_id_gen = env->id_gen;
15856 memset(&env->idmap_scratch.map, 0, sizeof(env->idmap_scratch.map));
15858 /* Verification state from speculative execution simulation
15859 * must never prune a non-speculative execution one.
15861 if (old->speculative && !cur->speculative)
15864 if (old->active_lock.ptr != cur->active_lock.ptr)
15867 /* Old and cur active_lock's have to be either both present
15870 if (!!old->active_lock.id != !!cur->active_lock.id)
15873 if (old->active_lock.id &&
15874 !check_ids(old->active_lock.id, cur->active_lock.id, &env->idmap_scratch))
15877 if (old->active_rcu_lock != cur->active_rcu_lock)
15880 /* for states to be equal callsites have to be the same
15881 * and all frame states need to be equivalent
15883 for (i = 0; i <= old->curframe; i++) {
15884 if (old->frame[i]->callsite != cur->frame[i]->callsite)
15886 if (!func_states_equal(env, old->frame[i], cur->frame[i]))
15892 /* Return 0 if no propagation happened. Return negative error code if error
15893 * happened. Otherwise, return the propagated bit.
15895 static int propagate_liveness_reg(struct bpf_verifier_env *env,
15896 struct bpf_reg_state *reg,
15897 struct bpf_reg_state *parent_reg)
15899 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
15900 u8 flag = reg->live & REG_LIVE_READ;
15903 /* When comes here, read flags of PARENT_REG or REG could be any of
15904 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
15905 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
15907 if (parent_flag == REG_LIVE_READ64 ||
15908 /* Or if there is no read flag from REG. */
15910 /* Or if the read flag from REG is the same as PARENT_REG. */
15911 parent_flag == flag)
15914 err = mark_reg_read(env, reg, parent_reg, flag);
15921 /* A write screens off any subsequent reads; but write marks come from the
15922 * straight-line code between a state and its parent. When we arrive at an
15923 * equivalent state (jump target or such) we didn't arrive by the straight-line
15924 * code, so read marks in the state must propagate to the parent regardless
15925 * of the state's write marks. That's what 'parent == state->parent' comparison
15926 * in mark_reg_read() is for.
15928 static int propagate_liveness(struct bpf_verifier_env *env,
15929 const struct bpf_verifier_state *vstate,
15930 struct bpf_verifier_state *vparent)
15932 struct bpf_reg_state *state_reg, *parent_reg;
15933 struct bpf_func_state *state, *parent;
15934 int i, frame, err = 0;
15936 if (vparent->curframe != vstate->curframe) {
15937 WARN(1, "propagate_live: parent frame %d current frame %d\n",
15938 vparent->curframe, vstate->curframe);
15941 /* Propagate read liveness of registers... */
15942 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
15943 for (frame = 0; frame <= vstate->curframe; frame++) {
15944 parent = vparent->frame[frame];
15945 state = vstate->frame[frame];
15946 parent_reg = parent->regs;
15947 state_reg = state->regs;
15948 /* We don't need to worry about FP liveness, it's read-only */
15949 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
15950 err = propagate_liveness_reg(env, &state_reg[i],
15954 if (err == REG_LIVE_READ64)
15955 mark_insn_zext(env, &parent_reg[i]);
15958 /* Propagate stack slots. */
15959 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
15960 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
15961 parent_reg = &parent->stack[i].spilled_ptr;
15962 state_reg = &state->stack[i].spilled_ptr;
15963 err = propagate_liveness_reg(env, state_reg,
15972 /* find precise scalars in the previous equivalent state and
15973 * propagate them into the current state
15975 static int propagate_precision(struct bpf_verifier_env *env,
15976 const struct bpf_verifier_state *old)
15978 struct bpf_reg_state *state_reg;
15979 struct bpf_func_state *state;
15980 int i, err = 0, fr;
15983 for (fr = old->curframe; fr >= 0; fr--) {
15984 state = old->frame[fr];
15985 state_reg = state->regs;
15987 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
15988 if (state_reg->type != SCALAR_VALUE ||
15989 !state_reg->precise ||
15990 !(state_reg->live & REG_LIVE_READ))
15992 if (env->log.level & BPF_LOG_LEVEL2) {
15994 verbose(env, "frame %d: propagating r%d", fr, i);
15996 verbose(env, ",r%d", i);
15998 bt_set_frame_reg(&env->bt, fr, i);
16002 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
16003 if (!is_spilled_reg(&state->stack[i]))
16005 state_reg = &state->stack[i].spilled_ptr;
16006 if (state_reg->type != SCALAR_VALUE ||
16007 !state_reg->precise ||
16008 !(state_reg->live & REG_LIVE_READ))
16010 if (env->log.level & BPF_LOG_LEVEL2) {
16012 verbose(env, "frame %d: propagating fp%d",
16013 fr, (-i - 1) * BPF_REG_SIZE);
16015 verbose(env, ",fp%d", (-i - 1) * BPF_REG_SIZE);
16017 bt_set_frame_slot(&env->bt, fr, i);
16021 verbose(env, "\n");
16024 err = mark_chain_precision_batch(env);
16031 static bool states_maybe_looping(struct bpf_verifier_state *old,
16032 struct bpf_verifier_state *cur)
16034 struct bpf_func_state *fold, *fcur;
16035 int i, fr = cur->curframe;
16037 if (old->curframe != fr)
16040 fold = old->frame[fr];
16041 fcur = cur->frame[fr];
16042 for (i = 0; i < MAX_BPF_REG; i++)
16043 if (memcmp(&fold->regs[i], &fcur->regs[i],
16044 offsetof(struct bpf_reg_state, parent)))
16049 static bool is_iter_next_insn(struct bpf_verifier_env *env, int insn_idx)
16051 return env->insn_aux_data[insn_idx].is_iter_next;
16054 /* is_state_visited() handles iter_next() (see process_iter_next_call() for
16055 * terminology) calls specially: as opposed to bounded BPF loops, it *expects*
16056 * states to match, which otherwise would look like an infinite loop. So while
16057 * iter_next() calls are taken care of, we still need to be careful and
16058 * prevent erroneous and too eager declaration of "ininite loop", when
16059 * iterators are involved.
16061 * Here's a situation in pseudo-BPF assembly form:
16063 * 0: again: ; set up iter_next() call args
16064 * 1: r1 = &it ; <CHECKPOINT HERE>
16065 * 2: call bpf_iter_num_next ; this is iter_next() call
16066 * 3: if r0 == 0 goto done
16067 * 4: ... something useful here ...
16068 * 5: goto again ; another iteration
16071 * 8: call bpf_iter_num_destroy ; clean up iter state
16074 * This is a typical loop. Let's assume that we have a prune point at 1:,
16075 * before we get to `call bpf_iter_num_next` (e.g., because of that `goto
16076 * again`, assuming other heuristics don't get in a way).
16078 * When we first time come to 1:, let's say we have some state X. We proceed
16079 * to 2:, fork states, enqueue ACTIVE, validate NULL case successfully, exit.
16080 * Now we come back to validate that forked ACTIVE state. We proceed through
16081 * 3-5, come to goto, jump to 1:. Let's assume our state didn't change, so we
16082 * are converging. But the problem is that we don't know that yet, as this
16083 * convergence has to happen at iter_next() call site only. So if nothing is
16084 * done, at 1: verifier will use bounded loop logic and declare infinite
16085 * looping (and would be *technically* correct, if not for iterator's
16086 * "eventual sticky NULL" contract, see process_iter_next_call()). But we
16087 * don't want that. So what we do in process_iter_next_call() when we go on
16088 * another ACTIVE iteration, we bump slot->iter.depth, to mark that it's
16089 * a different iteration. So when we suspect an infinite loop, we additionally
16090 * check if any of the *ACTIVE* iterator states depths differ. If yes, we
16091 * pretend we are not looping and wait for next iter_next() call.
16093 * This only applies to ACTIVE state. In DRAINED state we don't expect to
16094 * loop, because that would actually mean infinite loop, as DRAINED state is
16095 * "sticky", and so we'll keep returning into the same instruction with the
16096 * same state (at least in one of possible code paths).
16098 * This approach allows to keep infinite loop heuristic even in the face of
16099 * active iterator. E.g., C snippet below is and will be detected as
16100 * inifintely looping:
16102 * struct bpf_iter_num it;
16105 * bpf_iter_num_new(&it, 0, 10);
16106 * while ((p = bpf_iter_num_next(&t))) {
16108 * while (x--) {} // <<-- infinite loop here
16112 static bool iter_active_depths_differ(struct bpf_verifier_state *old, struct bpf_verifier_state *cur)
16114 struct bpf_reg_state *slot, *cur_slot;
16115 struct bpf_func_state *state;
16118 for (fr = old->curframe; fr >= 0; fr--) {
16119 state = old->frame[fr];
16120 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
16121 if (state->stack[i].slot_type[0] != STACK_ITER)
16124 slot = &state->stack[i].spilled_ptr;
16125 if (slot->iter.state != BPF_ITER_STATE_ACTIVE)
16128 cur_slot = &cur->frame[fr]->stack[i].spilled_ptr;
16129 if (cur_slot->iter.depth != slot->iter.depth)
16136 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
16138 struct bpf_verifier_state_list *new_sl;
16139 struct bpf_verifier_state_list *sl, **pprev;
16140 struct bpf_verifier_state *cur = env->cur_state, *new;
16141 int i, j, err, states_cnt = 0;
16142 bool force_new_state = env->test_state_freq || is_force_checkpoint(env, insn_idx);
16143 bool add_new_state = force_new_state;
16145 /* bpf progs typically have pruning point every 4 instructions
16146 * http://vger.kernel.org/bpfconf2019.html#session-1
16147 * Do not add new state for future pruning if the verifier hasn't seen
16148 * at least 2 jumps and at least 8 instructions.
16149 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
16150 * In tests that amounts to up to 50% reduction into total verifier
16151 * memory consumption and 20% verifier time speedup.
16153 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
16154 env->insn_processed - env->prev_insn_processed >= 8)
16155 add_new_state = true;
16157 pprev = explored_state(env, insn_idx);
16160 clean_live_states(env, insn_idx, cur);
16164 if (sl->state.insn_idx != insn_idx)
16167 if (sl->state.branches) {
16168 struct bpf_func_state *frame = sl->state.frame[sl->state.curframe];
16170 if (frame->in_async_callback_fn &&
16171 frame->async_entry_cnt != cur->frame[cur->curframe]->async_entry_cnt) {
16172 /* Different async_entry_cnt means that the verifier is
16173 * processing another entry into async callback.
16174 * Seeing the same state is not an indication of infinite
16175 * loop or infinite recursion.
16176 * But finding the same state doesn't mean that it's safe
16177 * to stop processing the current state. The previous state
16178 * hasn't yet reached bpf_exit, since state.branches > 0.
16179 * Checking in_async_callback_fn alone is not enough either.
16180 * Since the verifier still needs to catch infinite loops
16181 * inside async callbacks.
16183 goto skip_inf_loop_check;
16185 /* BPF open-coded iterators loop detection is special.
16186 * states_maybe_looping() logic is too simplistic in detecting
16187 * states that *might* be equivalent, because it doesn't know
16188 * about ID remapping, so don't even perform it.
16189 * See process_iter_next_call() and iter_active_depths_differ()
16190 * for overview of the logic. When current and one of parent
16191 * states are detected as equivalent, it's a good thing: we prove
16192 * convergence and can stop simulating further iterations.
16193 * It's safe to assume that iterator loop will finish, taking into
16194 * account iter_next() contract of eventually returning
16195 * sticky NULL result.
16197 if (is_iter_next_insn(env, insn_idx)) {
16198 if (states_equal(env, &sl->state, cur)) {
16199 struct bpf_func_state *cur_frame;
16200 struct bpf_reg_state *iter_state, *iter_reg;
16203 cur_frame = cur->frame[cur->curframe];
16204 /* btf_check_iter_kfuncs() enforces that
16205 * iter state pointer is always the first arg
16207 iter_reg = &cur_frame->regs[BPF_REG_1];
16208 /* current state is valid due to states_equal(),
16209 * so we can assume valid iter and reg state,
16210 * no need for extra (re-)validations
16212 spi = __get_spi(iter_reg->off + iter_reg->var_off.value);
16213 iter_state = &func(env, iter_reg)->stack[spi].spilled_ptr;
16214 if (iter_state->iter.state == BPF_ITER_STATE_ACTIVE)
16217 goto skip_inf_loop_check;
16219 /* attempt to detect infinite loop to avoid unnecessary doomed work */
16220 if (states_maybe_looping(&sl->state, cur) &&
16221 states_equal(env, &sl->state, cur) &&
16222 !iter_active_depths_differ(&sl->state, cur)) {
16223 verbose_linfo(env, insn_idx, "; ");
16224 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
16227 /* if the verifier is processing a loop, avoid adding new state
16228 * too often, since different loop iterations have distinct
16229 * states and may not help future pruning.
16230 * This threshold shouldn't be too low to make sure that
16231 * a loop with large bound will be rejected quickly.
16232 * The most abusive loop will be:
16234 * if r1 < 1000000 goto pc-2
16235 * 1M insn_procssed limit / 100 == 10k peak states.
16236 * This threshold shouldn't be too high either, since states
16237 * at the end of the loop are likely to be useful in pruning.
16239 skip_inf_loop_check:
16240 if (!force_new_state &&
16241 env->jmps_processed - env->prev_jmps_processed < 20 &&
16242 env->insn_processed - env->prev_insn_processed < 100)
16243 add_new_state = false;
16246 if (states_equal(env, &sl->state, cur)) {
16249 /* reached equivalent register/stack state,
16250 * prune the search.
16251 * Registers read by the continuation are read by us.
16252 * If we have any write marks in env->cur_state, they
16253 * will prevent corresponding reads in the continuation
16254 * from reaching our parent (an explored_state). Our
16255 * own state will get the read marks recorded, but
16256 * they'll be immediately forgotten as we're pruning
16257 * this state and will pop a new one.
16259 err = propagate_liveness(env, &sl->state, cur);
16261 /* if previous state reached the exit with precision and
16262 * current state is equivalent to it (except precsion marks)
16263 * the precision needs to be propagated back in
16264 * the current state.
16266 err = err ? : push_jmp_history(env, cur);
16267 err = err ? : propagate_precision(env, &sl->state);
16273 /* when new state is not going to be added do not increase miss count.
16274 * Otherwise several loop iterations will remove the state
16275 * recorded earlier. The goal of these heuristics is to have
16276 * states from some iterations of the loop (some in the beginning
16277 * and some at the end) to help pruning.
16281 /* heuristic to determine whether this state is beneficial
16282 * to keep checking from state equivalence point of view.
16283 * Higher numbers increase max_states_per_insn and verification time,
16284 * but do not meaningfully decrease insn_processed.
16286 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
16287 /* the state is unlikely to be useful. Remove it to
16288 * speed up verification
16291 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
16292 u32 br = sl->state.branches;
16295 "BUG live_done but branches_to_explore %d\n",
16297 free_verifier_state(&sl->state, false);
16299 env->peak_states--;
16301 /* cannot free this state, since parentage chain may
16302 * walk it later. Add it for free_list instead to
16303 * be freed at the end of verification
16305 sl->next = env->free_list;
16306 env->free_list = sl;
16316 if (env->max_states_per_insn < states_cnt)
16317 env->max_states_per_insn = states_cnt;
16319 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
16322 if (!add_new_state)
16325 /* There were no equivalent states, remember the current one.
16326 * Technically the current state is not proven to be safe yet,
16327 * but it will either reach outer most bpf_exit (which means it's safe)
16328 * or it will be rejected. When there are no loops the verifier won't be
16329 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
16330 * again on the way to bpf_exit.
16331 * When looping the sl->state.branches will be > 0 and this state
16332 * will not be considered for equivalence until branches == 0.
16334 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
16337 env->total_states++;
16338 env->peak_states++;
16339 env->prev_jmps_processed = env->jmps_processed;
16340 env->prev_insn_processed = env->insn_processed;
16342 /* forget precise markings we inherited, see __mark_chain_precision */
16343 if (env->bpf_capable)
16344 mark_all_scalars_imprecise(env, cur);
16346 /* add new state to the head of linked list */
16347 new = &new_sl->state;
16348 err = copy_verifier_state(new, cur);
16350 free_verifier_state(new, false);
16354 new->insn_idx = insn_idx;
16355 WARN_ONCE(new->branches != 1,
16356 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
16359 cur->first_insn_idx = insn_idx;
16360 clear_jmp_history(cur);
16361 new_sl->next = *explored_state(env, insn_idx);
16362 *explored_state(env, insn_idx) = new_sl;
16363 /* connect new state to parentage chain. Current frame needs all
16364 * registers connected. Only r6 - r9 of the callers are alive (pushed
16365 * to the stack implicitly by JITs) so in callers' frames connect just
16366 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
16367 * the state of the call instruction (with WRITTEN set), and r0 comes
16368 * from callee with its full parentage chain, anyway.
16370 /* clear write marks in current state: the writes we did are not writes
16371 * our child did, so they don't screen off its reads from us.
16372 * (There are no read marks in current state, because reads always mark
16373 * their parent and current state never has children yet. Only
16374 * explored_states can get read marks.)
16376 for (j = 0; j <= cur->curframe; j++) {
16377 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
16378 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
16379 for (i = 0; i < BPF_REG_FP; i++)
16380 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
16383 /* all stack frames are accessible from callee, clear them all */
16384 for (j = 0; j <= cur->curframe; j++) {
16385 struct bpf_func_state *frame = cur->frame[j];
16386 struct bpf_func_state *newframe = new->frame[j];
16388 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
16389 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
16390 frame->stack[i].spilled_ptr.parent =
16391 &newframe->stack[i].spilled_ptr;
16397 /* Return true if it's OK to have the same insn return a different type. */
16398 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
16400 switch (base_type(type)) {
16402 case PTR_TO_SOCKET:
16403 case PTR_TO_SOCK_COMMON:
16404 case PTR_TO_TCP_SOCK:
16405 case PTR_TO_XDP_SOCK:
16406 case PTR_TO_BTF_ID:
16413 /* If an instruction was previously used with particular pointer types, then we
16414 * need to be careful to avoid cases such as the below, where it may be ok
16415 * for one branch accessing the pointer, but not ok for the other branch:
16420 * R1 = some_other_valid_ptr;
16423 * R2 = *(u32 *)(R1 + 0);
16425 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
16427 return src != prev && (!reg_type_mismatch_ok(src) ||
16428 !reg_type_mismatch_ok(prev));
16431 static int save_aux_ptr_type(struct bpf_verifier_env *env, enum bpf_reg_type type,
16432 bool allow_trust_missmatch)
16434 enum bpf_reg_type *prev_type = &env->insn_aux_data[env->insn_idx].ptr_type;
16436 if (*prev_type == NOT_INIT) {
16437 /* Saw a valid insn
16438 * dst_reg = *(u32 *)(src_reg + off)
16439 * save type to validate intersecting paths
16442 } else if (reg_type_mismatch(type, *prev_type)) {
16443 /* Abuser program is trying to use the same insn
16444 * dst_reg = *(u32*) (src_reg + off)
16445 * with different pointer types:
16446 * src_reg == ctx in one branch and
16447 * src_reg == stack|map in some other branch.
16450 if (allow_trust_missmatch &&
16451 base_type(type) == PTR_TO_BTF_ID &&
16452 base_type(*prev_type) == PTR_TO_BTF_ID) {
16454 * Have to support a use case when one path through
16455 * the program yields TRUSTED pointer while another
16456 * is UNTRUSTED. Fallback to UNTRUSTED to generate
16457 * BPF_PROBE_MEM/BPF_PROBE_MEMSX.
16459 *prev_type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
16461 verbose(env, "same insn cannot be used with different pointers\n");
16469 static int do_check(struct bpf_verifier_env *env)
16471 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
16472 struct bpf_verifier_state *state = env->cur_state;
16473 struct bpf_insn *insns = env->prog->insnsi;
16474 struct bpf_reg_state *regs;
16475 int insn_cnt = env->prog->len;
16476 bool do_print_state = false;
16477 int prev_insn_idx = -1;
16480 struct bpf_insn *insn;
16484 env->prev_insn_idx = prev_insn_idx;
16485 if (env->insn_idx >= insn_cnt) {
16486 verbose(env, "invalid insn idx %d insn_cnt %d\n",
16487 env->insn_idx, insn_cnt);
16491 insn = &insns[env->insn_idx];
16492 class = BPF_CLASS(insn->code);
16494 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
16496 "BPF program is too large. Processed %d insn\n",
16497 env->insn_processed);
16501 state->last_insn_idx = env->prev_insn_idx;
16503 if (is_prune_point(env, env->insn_idx)) {
16504 err = is_state_visited(env, env->insn_idx);
16508 /* found equivalent state, can prune the search */
16509 if (env->log.level & BPF_LOG_LEVEL) {
16510 if (do_print_state)
16511 verbose(env, "\nfrom %d to %d%s: safe\n",
16512 env->prev_insn_idx, env->insn_idx,
16513 env->cur_state->speculative ?
16514 " (speculative execution)" : "");
16516 verbose(env, "%d: safe\n", env->insn_idx);
16518 goto process_bpf_exit;
16522 if (is_jmp_point(env, env->insn_idx)) {
16523 err = push_jmp_history(env, state);
16528 if (signal_pending(current))
16531 if (need_resched())
16534 if (env->log.level & BPF_LOG_LEVEL2 && do_print_state) {
16535 verbose(env, "\nfrom %d to %d%s:",
16536 env->prev_insn_idx, env->insn_idx,
16537 env->cur_state->speculative ?
16538 " (speculative execution)" : "");
16539 print_verifier_state(env, state->frame[state->curframe], true);
16540 do_print_state = false;
16543 if (env->log.level & BPF_LOG_LEVEL) {
16544 const struct bpf_insn_cbs cbs = {
16545 .cb_call = disasm_kfunc_name,
16546 .cb_print = verbose,
16547 .private_data = env,
16550 if (verifier_state_scratched(env))
16551 print_insn_state(env, state->frame[state->curframe]);
16553 verbose_linfo(env, env->insn_idx, "; ");
16554 env->prev_log_pos = env->log.end_pos;
16555 verbose(env, "%d: ", env->insn_idx);
16556 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
16557 env->prev_insn_print_pos = env->log.end_pos - env->prev_log_pos;
16558 env->prev_log_pos = env->log.end_pos;
16561 if (bpf_prog_is_offloaded(env->prog->aux)) {
16562 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
16563 env->prev_insn_idx);
16568 regs = cur_regs(env);
16569 sanitize_mark_insn_seen(env);
16570 prev_insn_idx = env->insn_idx;
16572 if (class == BPF_ALU || class == BPF_ALU64) {
16573 err = check_alu_op(env, insn);
16577 } else if (class == BPF_LDX) {
16578 enum bpf_reg_type src_reg_type;
16580 /* check for reserved fields is already done */
16582 /* check src operand */
16583 err = check_reg_arg(env, insn->src_reg, SRC_OP);
16587 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
16591 src_reg_type = regs[insn->src_reg].type;
16593 /* check that memory (src_reg + off) is readable,
16594 * the state of dst_reg will be updated by this func
16596 err = check_mem_access(env, env->insn_idx, insn->src_reg,
16597 insn->off, BPF_SIZE(insn->code),
16598 BPF_READ, insn->dst_reg, false,
16599 BPF_MODE(insn->code) == BPF_MEMSX);
16603 err = save_aux_ptr_type(env, src_reg_type, true);
16606 } else if (class == BPF_STX) {
16607 enum bpf_reg_type dst_reg_type;
16609 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
16610 err = check_atomic(env, env->insn_idx, insn);
16617 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
16618 verbose(env, "BPF_STX uses reserved fields\n");
16622 /* check src1 operand */
16623 err = check_reg_arg(env, insn->src_reg, SRC_OP);
16626 /* check src2 operand */
16627 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
16631 dst_reg_type = regs[insn->dst_reg].type;
16633 /* check that memory (dst_reg + off) is writeable */
16634 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
16635 insn->off, BPF_SIZE(insn->code),
16636 BPF_WRITE, insn->src_reg, false, false);
16640 err = save_aux_ptr_type(env, dst_reg_type, false);
16643 } else if (class == BPF_ST) {
16644 enum bpf_reg_type dst_reg_type;
16646 if (BPF_MODE(insn->code) != BPF_MEM ||
16647 insn->src_reg != BPF_REG_0) {
16648 verbose(env, "BPF_ST uses reserved fields\n");
16651 /* check src operand */
16652 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
16656 dst_reg_type = regs[insn->dst_reg].type;
16658 /* check that memory (dst_reg + off) is writeable */
16659 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
16660 insn->off, BPF_SIZE(insn->code),
16661 BPF_WRITE, -1, false, false);
16665 err = save_aux_ptr_type(env, dst_reg_type, false);
16668 } else if (class == BPF_JMP || class == BPF_JMP32) {
16669 u8 opcode = BPF_OP(insn->code);
16671 env->jmps_processed++;
16672 if (opcode == BPF_CALL) {
16673 if (BPF_SRC(insn->code) != BPF_K ||
16674 (insn->src_reg != BPF_PSEUDO_KFUNC_CALL
16675 && insn->off != 0) ||
16676 (insn->src_reg != BPF_REG_0 &&
16677 insn->src_reg != BPF_PSEUDO_CALL &&
16678 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
16679 insn->dst_reg != BPF_REG_0 ||
16680 class == BPF_JMP32) {
16681 verbose(env, "BPF_CALL uses reserved fields\n");
16685 if (env->cur_state->active_lock.ptr) {
16686 if ((insn->src_reg == BPF_REG_0 && insn->imm != BPF_FUNC_spin_unlock) ||
16687 (insn->src_reg == BPF_PSEUDO_CALL) ||
16688 (insn->src_reg == BPF_PSEUDO_KFUNC_CALL &&
16689 (insn->off != 0 || !is_bpf_graph_api_kfunc(insn->imm)))) {
16690 verbose(env, "function calls are not allowed while holding a lock\n");
16694 if (insn->src_reg == BPF_PSEUDO_CALL)
16695 err = check_func_call(env, insn, &env->insn_idx);
16696 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
16697 err = check_kfunc_call(env, insn, &env->insn_idx);
16699 err = check_helper_call(env, insn, &env->insn_idx);
16703 mark_reg_scratched(env, BPF_REG_0);
16704 } else if (opcode == BPF_JA) {
16705 if (BPF_SRC(insn->code) != BPF_K ||
16706 insn->src_reg != BPF_REG_0 ||
16707 insn->dst_reg != BPF_REG_0 ||
16708 (class == BPF_JMP && insn->imm != 0) ||
16709 (class == BPF_JMP32 && insn->off != 0)) {
16710 verbose(env, "BPF_JA uses reserved fields\n");
16714 if (class == BPF_JMP)
16715 env->insn_idx += insn->off + 1;
16717 env->insn_idx += insn->imm + 1;
16720 } else if (opcode == BPF_EXIT) {
16721 if (BPF_SRC(insn->code) != BPF_K ||
16723 insn->src_reg != BPF_REG_0 ||
16724 insn->dst_reg != BPF_REG_0 ||
16725 class == BPF_JMP32) {
16726 verbose(env, "BPF_EXIT uses reserved fields\n");
16730 if (env->cur_state->active_lock.ptr &&
16731 !in_rbtree_lock_required_cb(env)) {
16732 verbose(env, "bpf_spin_unlock is missing\n");
16736 if (env->cur_state->active_rcu_lock &&
16737 !in_rbtree_lock_required_cb(env)) {
16738 verbose(env, "bpf_rcu_read_unlock is missing\n");
16742 /* We must do check_reference_leak here before
16743 * prepare_func_exit to handle the case when
16744 * state->curframe > 0, it may be a callback
16745 * function, for which reference_state must
16746 * match caller reference state when it exits.
16748 err = check_reference_leak(env);
16752 if (state->curframe) {
16753 /* exit from nested function */
16754 err = prepare_func_exit(env, &env->insn_idx);
16757 do_print_state = true;
16761 err = check_return_code(env);
16765 mark_verifier_state_scratched(env);
16766 update_branch_counts(env, env->cur_state);
16767 err = pop_stack(env, &prev_insn_idx,
16768 &env->insn_idx, pop_log);
16770 if (err != -ENOENT)
16774 do_print_state = true;
16778 err = check_cond_jmp_op(env, insn, &env->insn_idx);
16782 } else if (class == BPF_LD) {
16783 u8 mode = BPF_MODE(insn->code);
16785 if (mode == BPF_ABS || mode == BPF_IND) {
16786 err = check_ld_abs(env, insn);
16790 } else if (mode == BPF_IMM) {
16791 err = check_ld_imm(env, insn);
16796 sanitize_mark_insn_seen(env);
16798 verbose(env, "invalid BPF_LD mode\n");
16802 verbose(env, "unknown insn class %d\n", class);
16812 static int find_btf_percpu_datasec(struct btf *btf)
16814 const struct btf_type *t;
16819 * Both vmlinux and module each have their own ".data..percpu"
16820 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
16821 * types to look at only module's own BTF types.
16823 n = btf_nr_types(btf);
16824 if (btf_is_module(btf))
16825 i = btf_nr_types(btf_vmlinux);
16829 for(; i < n; i++) {
16830 t = btf_type_by_id(btf, i);
16831 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
16834 tname = btf_name_by_offset(btf, t->name_off);
16835 if (!strcmp(tname, ".data..percpu"))
16842 /* replace pseudo btf_id with kernel symbol address */
16843 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
16844 struct bpf_insn *insn,
16845 struct bpf_insn_aux_data *aux)
16847 const struct btf_var_secinfo *vsi;
16848 const struct btf_type *datasec;
16849 struct btf_mod_pair *btf_mod;
16850 const struct btf_type *t;
16851 const char *sym_name;
16852 bool percpu = false;
16853 u32 type, id = insn->imm;
16857 int i, btf_fd, err;
16859 btf_fd = insn[1].imm;
16861 btf = btf_get_by_fd(btf_fd);
16863 verbose(env, "invalid module BTF object FD specified.\n");
16867 if (!btf_vmlinux) {
16868 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
16875 t = btf_type_by_id(btf, id);
16877 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
16882 if (!btf_type_is_var(t) && !btf_type_is_func(t)) {
16883 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR or KIND_FUNC\n", id);
16888 sym_name = btf_name_by_offset(btf, t->name_off);
16889 addr = kallsyms_lookup_name(sym_name);
16891 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
16896 insn[0].imm = (u32)addr;
16897 insn[1].imm = addr >> 32;
16899 if (btf_type_is_func(t)) {
16900 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
16901 aux->btf_var.mem_size = 0;
16905 datasec_id = find_btf_percpu_datasec(btf);
16906 if (datasec_id > 0) {
16907 datasec = btf_type_by_id(btf, datasec_id);
16908 for_each_vsi(i, datasec, vsi) {
16909 if (vsi->type == id) {
16917 t = btf_type_skip_modifiers(btf, type, NULL);
16919 aux->btf_var.reg_type = PTR_TO_BTF_ID | MEM_PERCPU;
16920 aux->btf_var.btf = btf;
16921 aux->btf_var.btf_id = type;
16922 } else if (!btf_type_is_struct(t)) {
16923 const struct btf_type *ret;
16927 /* resolve the type size of ksym. */
16928 ret = btf_resolve_size(btf, t, &tsize);
16930 tname = btf_name_by_offset(btf, t->name_off);
16931 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
16932 tname, PTR_ERR(ret));
16936 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
16937 aux->btf_var.mem_size = tsize;
16939 aux->btf_var.reg_type = PTR_TO_BTF_ID;
16940 aux->btf_var.btf = btf;
16941 aux->btf_var.btf_id = type;
16944 /* check whether we recorded this BTF (and maybe module) already */
16945 for (i = 0; i < env->used_btf_cnt; i++) {
16946 if (env->used_btfs[i].btf == btf) {
16952 if (env->used_btf_cnt >= MAX_USED_BTFS) {
16957 btf_mod = &env->used_btfs[env->used_btf_cnt];
16958 btf_mod->btf = btf;
16959 btf_mod->module = NULL;
16961 /* if we reference variables from kernel module, bump its refcount */
16962 if (btf_is_module(btf)) {
16963 btf_mod->module = btf_try_get_module(btf);
16964 if (!btf_mod->module) {
16970 env->used_btf_cnt++;
16978 static bool is_tracing_prog_type(enum bpf_prog_type type)
16981 case BPF_PROG_TYPE_KPROBE:
16982 case BPF_PROG_TYPE_TRACEPOINT:
16983 case BPF_PROG_TYPE_PERF_EVENT:
16984 case BPF_PROG_TYPE_RAW_TRACEPOINT:
16985 case BPF_PROG_TYPE_RAW_TRACEPOINT_WRITABLE:
16992 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
16993 struct bpf_map *map,
16994 struct bpf_prog *prog)
16997 enum bpf_prog_type prog_type = resolve_prog_type(prog);
16999 if (btf_record_has_field(map->record, BPF_LIST_HEAD) ||
17000 btf_record_has_field(map->record, BPF_RB_ROOT)) {
17001 if (is_tracing_prog_type(prog_type)) {
17002 verbose(env, "tracing progs cannot use bpf_{list_head,rb_root} yet\n");
17007 if (btf_record_has_field(map->record, BPF_SPIN_LOCK)) {
17008 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
17009 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
17013 if (is_tracing_prog_type(prog_type)) {
17014 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
17019 if (btf_record_has_field(map->record, BPF_TIMER)) {
17020 if (is_tracing_prog_type(prog_type)) {
17021 verbose(env, "tracing progs cannot use bpf_timer yet\n");
17026 if ((bpf_prog_is_offloaded(prog->aux) || bpf_map_is_offloaded(map)) &&
17027 !bpf_offload_prog_map_match(prog, map)) {
17028 verbose(env, "offload device mismatch between prog and map\n");
17032 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
17033 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
17037 if (prog->aux->sleepable)
17038 switch (map->map_type) {
17039 case BPF_MAP_TYPE_HASH:
17040 case BPF_MAP_TYPE_LRU_HASH:
17041 case BPF_MAP_TYPE_ARRAY:
17042 case BPF_MAP_TYPE_PERCPU_HASH:
17043 case BPF_MAP_TYPE_PERCPU_ARRAY:
17044 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
17045 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
17046 case BPF_MAP_TYPE_HASH_OF_MAPS:
17047 case BPF_MAP_TYPE_RINGBUF:
17048 case BPF_MAP_TYPE_USER_RINGBUF:
17049 case BPF_MAP_TYPE_INODE_STORAGE:
17050 case BPF_MAP_TYPE_SK_STORAGE:
17051 case BPF_MAP_TYPE_TASK_STORAGE:
17052 case BPF_MAP_TYPE_CGRP_STORAGE:
17056 "Sleepable programs can only use array, hash, ringbuf and local storage maps\n");
17063 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
17065 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
17066 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
17069 /* find and rewrite pseudo imm in ld_imm64 instructions:
17071 * 1. if it accesses map FD, replace it with actual map pointer.
17072 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
17074 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
17076 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
17078 struct bpf_insn *insn = env->prog->insnsi;
17079 int insn_cnt = env->prog->len;
17082 err = bpf_prog_calc_tag(env->prog);
17086 for (i = 0; i < insn_cnt; i++, insn++) {
17087 if (BPF_CLASS(insn->code) == BPF_LDX &&
17088 ((BPF_MODE(insn->code) != BPF_MEM && BPF_MODE(insn->code) != BPF_MEMSX) ||
17090 verbose(env, "BPF_LDX uses reserved fields\n");
17094 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
17095 struct bpf_insn_aux_data *aux;
17096 struct bpf_map *map;
17101 if (i == insn_cnt - 1 || insn[1].code != 0 ||
17102 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
17103 insn[1].off != 0) {
17104 verbose(env, "invalid bpf_ld_imm64 insn\n");
17108 if (insn[0].src_reg == 0)
17109 /* valid generic load 64-bit imm */
17112 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
17113 aux = &env->insn_aux_data[i];
17114 err = check_pseudo_btf_id(env, insn, aux);
17120 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
17121 aux = &env->insn_aux_data[i];
17122 aux->ptr_type = PTR_TO_FUNC;
17126 /* In final convert_pseudo_ld_imm64() step, this is
17127 * converted into regular 64-bit imm load insn.
17129 switch (insn[0].src_reg) {
17130 case BPF_PSEUDO_MAP_VALUE:
17131 case BPF_PSEUDO_MAP_IDX_VALUE:
17133 case BPF_PSEUDO_MAP_FD:
17134 case BPF_PSEUDO_MAP_IDX:
17135 if (insn[1].imm == 0)
17139 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
17143 switch (insn[0].src_reg) {
17144 case BPF_PSEUDO_MAP_IDX_VALUE:
17145 case BPF_PSEUDO_MAP_IDX:
17146 if (bpfptr_is_null(env->fd_array)) {
17147 verbose(env, "fd_idx without fd_array is invalid\n");
17150 if (copy_from_bpfptr_offset(&fd, env->fd_array,
17151 insn[0].imm * sizeof(fd),
17161 map = __bpf_map_get(f);
17163 verbose(env, "fd %d is not pointing to valid bpf_map\n",
17165 return PTR_ERR(map);
17168 err = check_map_prog_compatibility(env, map, env->prog);
17174 aux = &env->insn_aux_data[i];
17175 if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
17176 insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
17177 addr = (unsigned long)map;
17179 u32 off = insn[1].imm;
17181 if (off >= BPF_MAX_VAR_OFF) {
17182 verbose(env, "direct value offset of %u is not allowed\n", off);
17187 if (!map->ops->map_direct_value_addr) {
17188 verbose(env, "no direct value access support for this map type\n");
17193 err = map->ops->map_direct_value_addr(map, &addr, off);
17195 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
17196 map->value_size, off);
17201 aux->map_off = off;
17205 insn[0].imm = (u32)addr;
17206 insn[1].imm = addr >> 32;
17208 /* check whether we recorded this map already */
17209 for (j = 0; j < env->used_map_cnt; j++) {
17210 if (env->used_maps[j] == map) {
17211 aux->map_index = j;
17217 if (env->used_map_cnt >= MAX_USED_MAPS) {
17222 /* hold the map. If the program is rejected by verifier,
17223 * the map will be released by release_maps() or it
17224 * will be used by the valid program until it's unloaded
17225 * and all maps are released in free_used_maps()
17229 aux->map_index = env->used_map_cnt;
17230 env->used_maps[env->used_map_cnt++] = map;
17232 if (bpf_map_is_cgroup_storage(map) &&
17233 bpf_cgroup_storage_assign(env->prog->aux, map)) {
17234 verbose(env, "only one cgroup storage of each type is allowed\n");
17246 /* Basic sanity check before we invest more work here. */
17247 if (!bpf_opcode_in_insntable(insn->code)) {
17248 verbose(env, "unknown opcode %02x\n", insn->code);
17253 /* now all pseudo BPF_LD_IMM64 instructions load valid
17254 * 'struct bpf_map *' into a register instead of user map_fd.
17255 * These pointers will be used later by verifier to validate map access.
17260 /* drop refcnt of maps used by the rejected program */
17261 static void release_maps(struct bpf_verifier_env *env)
17263 __bpf_free_used_maps(env->prog->aux, env->used_maps,
17264 env->used_map_cnt);
17267 /* drop refcnt of maps used by the rejected program */
17268 static void release_btfs(struct bpf_verifier_env *env)
17270 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
17271 env->used_btf_cnt);
17274 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
17275 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
17277 struct bpf_insn *insn = env->prog->insnsi;
17278 int insn_cnt = env->prog->len;
17281 for (i = 0; i < insn_cnt; i++, insn++) {
17282 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
17284 if (insn->src_reg == BPF_PSEUDO_FUNC)
17290 /* single env->prog->insni[off] instruction was replaced with the range
17291 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
17292 * [0, off) and [off, end) to new locations, so the patched range stays zero
17294 static void adjust_insn_aux_data(struct bpf_verifier_env *env,
17295 struct bpf_insn_aux_data *new_data,
17296 struct bpf_prog *new_prog, u32 off, u32 cnt)
17298 struct bpf_insn_aux_data *old_data = env->insn_aux_data;
17299 struct bpf_insn *insn = new_prog->insnsi;
17300 u32 old_seen = old_data[off].seen;
17304 /* aux info at OFF always needs adjustment, no matter fast path
17305 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
17306 * original insn at old prog.
17308 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
17312 prog_len = new_prog->len;
17314 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
17315 memcpy(new_data + off + cnt - 1, old_data + off,
17316 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
17317 for (i = off; i < off + cnt - 1; i++) {
17318 /* Expand insni[off]'s seen count to the patched range. */
17319 new_data[i].seen = old_seen;
17320 new_data[i].zext_dst = insn_has_def32(env, insn + i);
17322 env->insn_aux_data = new_data;
17326 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
17332 /* NOTE: fake 'exit' subprog should be updated as well. */
17333 for (i = 0; i <= env->subprog_cnt; i++) {
17334 if (env->subprog_info[i].start <= off)
17336 env->subprog_info[i].start += len - 1;
17340 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
17342 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
17343 int i, sz = prog->aux->size_poke_tab;
17344 struct bpf_jit_poke_descriptor *desc;
17346 for (i = 0; i < sz; i++) {
17348 if (desc->insn_idx <= off)
17350 desc->insn_idx += len - 1;
17354 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
17355 const struct bpf_insn *patch, u32 len)
17357 struct bpf_prog *new_prog;
17358 struct bpf_insn_aux_data *new_data = NULL;
17361 new_data = vzalloc(array_size(env->prog->len + len - 1,
17362 sizeof(struct bpf_insn_aux_data)));
17367 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
17368 if (IS_ERR(new_prog)) {
17369 if (PTR_ERR(new_prog) == -ERANGE)
17371 "insn %d cannot be patched due to 16-bit range\n",
17372 env->insn_aux_data[off].orig_idx);
17376 adjust_insn_aux_data(env, new_data, new_prog, off, len);
17377 adjust_subprog_starts(env, off, len);
17378 adjust_poke_descs(new_prog, off, len);
17382 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
17387 /* find first prog starting at or after off (first to remove) */
17388 for (i = 0; i < env->subprog_cnt; i++)
17389 if (env->subprog_info[i].start >= off)
17391 /* find first prog starting at or after off + cnt (first to stay) */
17392 for (j = i; j < env->subprog_cnt; j++)
17393 if (env->subprog_info[j].start >= off + cnt)
17395 /* if j doesn't start exactly at off + cnt, we are just removing
17396 * the front of previous prog
17398 if (env->subprog_info[j].start != off + cnt)
17402 struct bpf_prog_aux *aux = env->prog->aux;
17405 /* move fake 'exit' subprog as well */
17406 move = env->subprog_cnt + 1 - j;
17408 memmove(env->subprog_info + i,
17409 env->subprog_info + j,
17410 sizeof(*env->subprog_info) * move);
17411 env->subprog_cnt -= j - i;
17413 /* remove func_info */
17414 if (aux->func_info) {
17415 move = aux->func_info_cnt - j;
17417 memmove(aux->func_info + i,
17418 aux->func_info + j,
17419 sizeof(*aux->func_info) * move);
17420 aux->func_info_cnt -= j - i;
17421 /* func_info->insn_off is set after all code rewrites,
17422 * in adjust_btf_func() - no need to adjust
17426 /* convert i from "first prog to remove" to "first to adjust" */
17427 if (env->subprog_info[i].start == off)
17431 /* update fake 'exit' subprog as well */
17432 for (; i <= env->subprog_cnt; i++)
17433 env->subprog_info[i].start -= cnt;
17438 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
17441 struct bpf_prog *prog = env->prog;
17442 u32 i, l_off, l_cnt, nr_linfo;
17443 struct bpf_line_info *linfo;
17445 nr_linfo = prog->aux->nr_linfo;
17449 linfo = prog->aux->linfo;
17451 /* find first line info to remove, count lines to be removed */
17452 for (i = 0; i < nr_linfo; i++)
17453 if (linfo[i].insn_off >= off)
17458 for (; i < nr_linfo; i++)
17459 if (linfo[i].insn_off < off + cnt)
17464 /* First live insn doesn't match first live linfo, it needs to "inherit"
17465 * last removed linfo. prog is already modified, so prog->len == off
17466 * means no live instructions after (tail of the program was removed).
17468 if (prog->len != off && l_cnt &&
17469 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
17471 linfo[--i].insn_off = off + cnt;
17474 /* remove the line info which refer to the removed instructions */
17476 memmove(linfo + l_off, linfo + i,
17477 sizeof(*linfo) * (nr_linfo - i));
17479 prog->aux->nr_linfo -= l_cnt;
17480 nr_linfo = prog->aux->nr_linfo;
17483 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
17484 for (i = l_off; i < nr_linfo; i++)
17485 linfo[i].insn_off -= cnt;
17487 /* fix up all subprogs (incl. 'exit') which start >= off */
17488 for (i = 0; i <= env->subprog_cnt; i++)
17489 if (env->subprog_info[i].linfo_idx > l_off) {
17490 /* program may have started in the removed region but
17491 * may not be fully removed
17493 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
17494 env->subprog_info[i].linfo_idx -= l_cnt;
17496 env->subprog_info[i].linfo_idx = l_off;
17502 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
17504 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17505 unsigned int orig_prog_len = env->prog->len;
17508 if (bpf_prog_is_offloaded(env->prog->aux))
17509 bpf_prog_offload_remove_insns(env, off, cnt);
17511 err = bpf_remove_insns(env->prog, off, cnt);
17515 err = adjust_subprog_starts_after_remove(env, off, cnt);
17519 err = bpf_adj_linfo_after_remove(env, off, cnt);
17523 memmove(aux_data + off, aux_data + off + cnt,
17524 sizeof(*aux_data) * (orig_prog_len - off - cnt));
17529 /* The verifier does more data flow analysis than llvm and will not
17530 * explore branches that are dead at run time. Malicious programs can
17531 * have dead code too. Therefore replace all dead at-run-time code
17534 * Just nops are not optimal, e.g. if they would sit at the end of the
17535 * program and through another bug we would manage to jump there, then
17536 * we'd execute beyond program memory otherwise. Returning exception
17537 * code also wouldn't work since we can have subprogs where the dead
17538 * code could be located.
17540 static void sanitize_dead_code(struct bpf_verifier_env *env)
17542 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17543 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
17544 struct bpf_insn *insn = env->prog->insnsi;
17545 const int insn_cnt = env->prog->len;
17548 for (i = 0; i < insn_cnt; i++) {
17549 if (aux_data[i].seen)
17551 memcpy(insn + i, &trap, sizeof(trap));
17552 aux_data[i].zext_dst = false;
17556 static bool insn_is_cond_jump(u8 code)
17561 if (BPF_CLASS(code) == BPF_JMP32)
17562 return op != BPF_JA;
17564 if (BPF_CLASS(code) != BPF_JMP)
17567 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
17570 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
17572 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17573 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
17574 struct bpf_insn *insn = env->prog->insnsi;
17575 const int insn_cnt = env->prog->len;
17578 for (i = 0; i < insn_cnt; i++, insn++) {
17579 if (!insn_is_cond_jump(insn->code))
17582 if (!aux_data[i + 1].seen)
17583 ja.off = insn->off;
17584 else if (!aux_data[i + 1 + insn->off].seen)
17589 if (bpf_prog_is_offloaded(env->prog->aux))
17590 bpf_prog_offload_replace_insn(env, i, &ja);
17592 memcpy(insn, &ja, sizeof(ja));
17596 static int opt_remove_dead_code(struct bpf_verifier_env *env)
17598 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17599 int insn_cnt = env->prog->len;
17602 for (i = 0; i < insn_cnt; i++) {
17606 while (i + j < insn_cnt && !aux_data[i + j].seen)
17611 err = verifier_remove_insns(env, i, j);
17614 insn_cnt = env->prog->len;
17620 static int opt_remove_nops(struct bpf_verifier_env *env)
17622 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
17623 struct bpf_insn *insn = env->prog->insnsi;
17624 int insn_cnt = env->prog->len;
17627 for (i = 0; i < insn_cnt; i++) {
17628 if (memcmp(&insn[i], &ja, sizeof(ja)))
17631 err = verifier_remove_insns(env, i, 1);
17641 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
17642 const union bpf_attr *attr)
17644 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
17645 struct bpf_insn_aux_data *aux = env->insn_aux_data;
17646 int i, patch_len, delta = 0, len = env->prog->len;
17647 struct bpf_insn *insns = env->prog->insnsi;
17648 struct bpf_prog *new_prog;
17651 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
17652 zext_patch[1] = BPF_ZEXT_REG(0);
17653 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
17654 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
17655 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
17656 for (i = 0; i < len; i++) {
17657 int adj_idx = i + delta;
17658 struct bpf_insn insn;
17661 insn = insns[adj_idx];
17662 load_reg = insn_def_regno(&insn);
17663 if (!aux[adj_idx].zext_dst) {
17671 class = BPF_CLASS(code);
17672 if (load_reg == -1)
17675 /* NOTE: arg "reg" (the fourth one) is only used for
17676 * BPF_STX + SRC_OP, so it is safe to pass NULL
17679 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
17680 if (class == BPF_LD &&
17681 BPF_MODE(code) == BPF_IMM)
17686 /* ctx load could be transformed into wider load. */
17687 if (class == BPF_LDX &&
17688 aux[adj_idx].ptr_type == PTR_TO_CTX)
17691 imm_rnd = get_random_u32();
17692 rnd_hi32_patch[0] = insn;
17693 rnd_hi32_patch[1].imm = imm_rnd;
17694 rnd_hi32_patch[3].dst_reg = load_reg;
17695 patch = rnd_hi32_patch;
17697 goto apply_patch_buffer;
17700 /* Add in an zero-extend instruction if a) the JIT has requested
17701 * it or b) it's a CMPXCHG.
17703 * The latter is because: BPF_CMPXCHG always loads a value into
17704 * R0, therefore always zero-extends. However some archs'
17705 * equivalent instruction only does this load when the
17706 * comparison is successful. This detail of CMPXCHG is
17707 * orthogonal to the general zero-extension behaviour of the
17708 * CPU, so it's treated independently of bpf_jit_needs_zext.
17710 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
17713 /* Zero-extension is done by the caller. */
17714 if (bpf_pseudo_kfunc_call(&insn))
17717 if (WARN_ON(load_reg == -1)) {
17718 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
17722 zext_patch[0] = insn;
17723 zext_patch[1].dst_reg = load_reg;
17724 zext_patch[1].src_reg = load_reg;
17725 patch = zext_patch;
17727 apply_patch_buffer:
17728 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
17731 env->prog = new_prog;
17732 insns = new_prog->insnsi;
17733 aux = env->insn_aux_data;
17734 delta += patch_len - 1;
17740 /* convert load instructions that access fields of a context type into a
17741 * sequence of instructions that access fields of the underlying structure:
17742 * struct __sk_buff -> struct sk_buff
17743 * struct bpf_sock_ops -> struct sock
17745 static int convert_ctx_accesses(struct bpf_verifier_env *env)
17747 const struct bpf_verifier_ops *ops = env->ops;
17748 int i, cnt, size, ctx_field_size, delta = 0;
17749 const int insn_cnt = env->prog->len;
17750 struct bpf_insn insn_buf[16], *insn;
17751 u32 target_size, size_default, off;
17752 struct bpf_prog *new_prog;
17753 enum bpf_access_type type;
17754 bool is_narrower_load;
17756 if (ops->gen_prologue || env->seen_direct_write) {
17757 if (!ops->gen_prologue) {
17758 verbose(env, "bpf verifier is misconfigured\n");
17761 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
17763 if (cnt >= ARRAY_SIZE(insn_buf)) {
17764 verbose(env, "bpf verifier is misconfigured\n");
17767 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
17771 env->prog = new_prog;
17776 if (bpf_prog_is_offloaded(env->prog->aux))
17779 insn = env->prog->insnsi + delta;
17781 for (i = 0; i < insn_cnt; i++, insn++) {
17782 bpf_convert_ctx_access_t convert_ctx_access;
17785 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
17786 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
17787 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
17788 insn->code == (BPF_LDX | BPF_MEM | BPF_DW) ||
17789 insn->code == (BPF_LDX | BPF_MEMSX | BPF_B) ||
17790 insn->code == (BPF_LDX | BPF_MEMSX | BPF_H) ||
17791 insn->code == (BPF_LDX | BPF_MEMSX | BPF_W)) {
17793 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
17794 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
17795 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
17796 insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
17797 insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
17798 insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
17799 insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
17800 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
17806 if (type == BPF_WRITE &&
17807 env->insn_aux_data[i + delta].sanitize_stack_spill) {
17808 struct bpf_insn patch[] = {
17813 cnt = ARRAY_SIZE(patch);
17814 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
17819 env->prog = new_prog;
17820 insn = new_prog->insnsi + i + delta;
17824 switch ((int)env->insn_aux_data[i + delta].ptr_type) {
17826 if (!ops->convert_ctx_access)
17828 convert_ctx_access = ops->convert_ctx_access;
17830 case PTR_TO_SOCKET:
17831 case PTR_TO_SOCK_COMMON:
17832 convert_ctx_access = bpf_sock_convert_ctx_access;
17834 case PTR_TO_TCP_SOCK:
17835 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
17837 case PTR_TO_XDP_SOCK:
17838 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
17840 case PTR_TO_BTF_ID:
17841 case PTR_TO_BTF_ID | PTR_UNTRUSTED:
17842 /* PTR_TO_BTF_ID | MEM_ALLOC always has a valid lifetime, unlike
17843 * PTR_TO_BTF_ID, and an active ref_obj_id, but the same cannot
17844 * be said once it is marked PTR_UNTRUSTED, hence we must handle
17845 * any faults for loads into such types. BPF_WRITE is disallowed
17848 case PTR_TO_BTF_ID | MEM_ALLOC | PTR_UNTRUSTED:
17849 if (type == BPF_READ) {
17850 if (BPF_MODE(insn->code) == BPF_MEM)
17851 insn->code = BPF_LDX | BPF_PROBE_MEM |
17852 BPF_SIZE((insn)->code);
17854 insn->code = BPF_LDX | BPF_PROBE_MEMSX |
17855 BPF_SIZE((insn)->code);
17856 env->prog->aux->num_exentries++;
17863 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
17864 size = BPF_LDST_BYTES(insn);
17865 mode = BPF_MODE(insn->code);
17867 /* If the read access is a narrower load of the field,
17868 * convert to a 4/8-byte load, to minimum program type specific
17869 * convert_ctx_access changes. If conversion is successful,
17870 * we will apply proper mask to the result.
17872 is_narrower_load = size < ctx_field_size;
17873 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
17875 if (is_narrower_load) {
17878 if (type == BPF_WRITE) {
17879 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
17884 if (ctx_field_size == 4)
17886 else if (ctx_field_size == 8)
17887 size_code = BPF_DW;
17889 insn->off = off & ~(size_default - 1);
17890 insn->code = BPF_LDX | BPF_MEM | size_code;
17894 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
17896 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
17897 (ctx_field_size && !target_size)) {
17898 verbose(env, "bpf verifier is misconfigured\n");
17902 if (is_narrower_load && size < target_size) {
17903 u8 shift = bpf_ctx_narrow_access_offset(
17904 off, size, size_default) * 8;
17905 if (shift && cnt + 1 >= ARRAY_SIZE(insn_buf)) {
17906 verbose(env, "bpf verifier narrow ctx load misconfigured\n");
17909 if (ctx_field_size <= 4) {
17911 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
17914 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
17915 (1 << size * 8) - 1);
17918 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
17921 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
17922 (1ULL << size * 8) - 1);
17925 if (mode == BPF_MEMSX)
17926 insn_buf[cnt++] = BPF_RAW_INSN(BPF_ALU64 | BPF_MOV | BPF_X,
17927 insn->dst_reg, insn->dst_reg,
17930 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
17936 /* keep walking new program and skip insns we just inserted */
17937 env->prog = new_prog;
17938 insn = new_prog->insnsi + i + delta;
17944 static int jit_subprogs(struct bpf_verifier_env *env)
17946 struct bpf_prog *prog = env->prog, **func, *tmp;
17947 int i, j, subprog_start, subprog_end = 0, len, subprog;
17948 struct bpf_map *map_ptr;
17949 struct bpf_insn *insn;
17950 void *old_bpf_func;
17951 int err, num_exentries;
17953 if (env->subprog_cnt <= 1)
17956 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
17957 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn))
17960 /* Upon error here we cannot fall back to interpreter but
17961 * need a hard reject of the program. Thus -EFAULT is
17962 * propagated in any case.
17964 subprog = find_subprog(env, i + insn->imm + 1);
17966 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
17967 i + insn->imm + 1);
17970 /* temporarily remember subprog id inside insn instead of
17971 * aux_data, since next loop will split up all insns into funcs
17973 insn->off = subprog;
17974 /* remember original imm in case JIT fails and fallback
17975 * to interpreter will be needed
17977 env->insn_aux_data[i].call_imm = insn->imm;
17978 /* point imm to __bpf_call_base+1 from JITs point of view */
17980 if (bpf_pseudo_func(insn))
17981 /* jit (e.g. x86_64) may emit fewer instructions
17982 * if it learns a u32 imm is the same as a u64 imm.
17983 * Force a non zero here.
17988 err = bpf_prog_alloc_jited_linfo(prog);
17990 goto out_undo_insn;
17993 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
17995 goto out_undo_insn;
17997 for (i = 0; i < env->subprog_cnt; i++) {
17998 subprog_start = subprog_end;
17999 subprog_end = env->subprog_info[i + 1].start;
18001 len = subprog_end - subprog_start;
18002 /* bpf_prog_run() doesn't call subprogs directly,
18003 * hence main prog stats include the runtime of subprogs.
18004 * subprogs don't have IDs and not reachable via prog_get_next_id
18005 * func[i]->stats will never be accessed and stays NULL
18007 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
18010 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
18011 len * sizeof(struct bpf_insn));
18012 func[i]->type = prog->type;
18013 func[i]->len = len;
18014 if (bpf_prog_calc_tag(func[i]))
18016 func[i]->is_func = 1;
18017 func[i]->aux->func_idx = i;
18018 /* Below members will be freed only at prog->aux */
18019 func[i]->aux->btf = prog->aux->btf;
18020 func[i]->aux->func_info = prog->aux->func_info;
18021 func[i]->aux->func_info_cnt = prog->aux->func_info_cnt;
18022 func[i]->aux->poke_tab = prog->aux->poke_tab;
18023 func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
18025 for (j = 0; j < prog->aux->size_poke_tab; j++) {
18026 struct bpf_jit_poke_descriptor *poke;
18028 poke = &prog->aux->poke_tab[j];
18029 if (poke->insn_idx < subprog_end &&
18030 poke->insn_idx >= subprog_start)
18031 poke->aux = func[i]->aux;
18034 func[i]->aux->name[0] = 'F';
18035 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
18036 func[i]->jit_requested = 1;
18037 func[i]->blinding_requested = prog->blinding_requested;
18038 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
18039 func[i]->aux->kfunc_btf_tab = prog->aux->kfunc_btf_tab;
18040 func[i]->aux->linfo = prog->aux->linfo;
18041 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
18042 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
18043 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
18045 insn = func[i]->insnsi;
18046 for (j = 0; j < func[i]->len; j++, insn++) {
18047 if (BPF_CLASS(insn->code) == BPF_LDX &&
18048 (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
18049 BPF_MODE(insn->code) == BPF_PROBE_MEMSX))
18052 func[i]->aux->num_exentries = num_exentries;
18053 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
18054 func[i] = bpf_int_jit_compile(func[i]);
18055 if (!func[i]->jited) {
18062 /* at this point all bpf functions were successfully JITed
18063 * now populate all bpf_calls with correct addresses and
18064 * run last pass of JIT
18066 for (i = 0; i < env->subprog_cnt; i++) {
18067 insn = func[i]->insnsi;
18068 for (j = 0; j < func[i]->len; j++, insn++) {
18069 if (bpf_pseudo_func(insn)) {
18070 subprog = insn->off;
18071 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
18072 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
18075 if (!bpf_pseudo_call(insn))
18077 subprog = insn->off;
18078 insn->imm = BPF_CALL_IMM(func[subprog]->bpf_func);
18081 /* we use the aux data to keep a list of the start addresses
18082 * of the JITed images for each function in the program
18084 * for some architectures, such as powerpc64, the imm field
18085 * might not be large enough to hold the offset of the start
18086 * address of the callee's JITed image from __bpf_call_base
18088 * in such cases, we can lookup the start address of a callee
18089 * by using its subprog id, available from the off field of
18090 * the call instruction, as an index for this list
18092 func[i]->aux->func = func;
18093 func[i]->aux->func_cnt = env->subprog_cnt;
18095 for (i = 0; i < env->subprog_cnt; i++) {
18096 old_bpf_func = func[i]->bpf_func;
18097 tmp = bpf_int_jit_compile(func[i]);
18098 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
18099 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
18106 /* finally lock prog and jit images for all functions and
18107 * populate kallsysm. Begin at the first subprogram, since
18108 * bpf_prog_load will add the kallsyms for the main program.
18110 for (i = 1; i < env->subprog_cnt; i++) {
18111 bpf_prog_lock_ro(func[i]);
18112 bpf_prog_kallsyms_add(func[i]);
18115 /* Last step: make now unused interpreter insns from main
18116 * prog consistent for later dump requests, so they can
18117 * later look the same as if they were interpreted only.
18119 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
18120 if (bpf_pseudo_func(insn)) {
18121 insn[0].imm = env->insn_aux_data[i].call_imm;
18122 insn[1].imm = insn->off;
18126 if (!bpf_pseudo_call(insn))
18128 insn->off = env->insn_aux_data[i].call_imm;
18129 subprog = find_subprog(env, i + insn->off + 1);
18130 insn->imm = subprog;
18134 prog->bpf_func = func[0]->bpf_func;
18135 prog->jited_len = func[0]->jited_len;
18136 prog->aux->extable = func[0]->aux->extable;
18137 prog->aux->num_exentries = func[0]->aux->num_exentries;
18138 prog->aux->func = func;
18139 prog->aux->func_cnt = env->subprog_cnt;
18140 bpf_prog_jit_attempt_done(prog);
18143 /* We failed JIT'ing, so at this point we need to unregister poke
18144 * descriptors from subprogs, so that kernel is not attempting to
18145 * patch it anymore as we're freeing the subprog JIT memory.
18147 for (i = 0; i < prog->aux->size_poke_tab; i++) {
18148 map_ptr = prog->aux->poke_tab[i].tail_call.map;
18149 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
18151 /* At this point we're guaranteed that poke descriptors are not
18152 * live anymore. We can just unlink its descriptor table as it's
18153 * released with the main prog.
18155 for (i = 0; i < env->subprog_cnt; i++) {
18158 func[i]->aux->poke_tab = NULL;
18159 bpf_jit_free(func[i]);
18163 /* cleanup main prog to be interpreted */
18164 prog->jit_requested = 0;
18165 prog->blinding_requested = 0;
18166 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
18167 if (!bpf_pseudo_call(insn))
18170 insn->imm = env->insn_aux_data[i].call_imm;
18172 bpf_prog_jit_attempt_done(prog);
18176 static int fixup_call_args(struct bpf_verifier_env *env)
18178 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
18179 struct bpf_prog *prog = env->prog;
18180 struct bpf_insn *insn = prog->insnsi;
18181 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
18186 if (env->prog->jit_requested &&
18187 !bpf_prog_is_offloaded(env->prog->aux)) {
18188 err = jit_subprogs(env);
18191 if (err == -EFAULT)
18194 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
18195 if (has_kfunc_call) {
18196 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
18199 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
18200 /* When JIT fails the progs with bpf2bpf calls and tail_calls
18201 * have to be rejected, since interpreter doesn't support them yet.
18203 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
18206 for (i = 0; i < prog->len; i++, insn++) {
18207 if (bpf_pseudo_func(insn)) {
18208 /* When JIT fails the progs with callback calls
18209 * have to be rejected, since interpreter doesn't support them yet.
18211 verbose(env, "callbacks are not allowed in non-JITed programs\n");
18215 if (!bpf_pseudo_call(insn))
18217 depth = get_callee_stack_depth(env, insn, i);
18220 bpf_patch_call_args(insn, depth);
18227 /* replace a generic kfunc with a specialized version if necessary */
18228 static void specialize_kfunc(struct bpf_verifier_env *env,
18229 u32 func_id, u16 offset, unsigned long *addr)
18231 struct bpf_prog *prog = env->prog;
18232 bool seen_direct_write;
18236 if (bpf_dev_bound_kfunc_id(func_id)) {
18237 xdp_kfunc = bpf_dev_bound_resolve_kfunc(prog, func_id);
18239 *addr = (unsigned long)xdp_kfunc;
18242 /* fallback to default kfunc when not supported by netdev */
18248 if (func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) {
18249 seen_direct_write = env->seen_direct_write;
18250 is_rdonly = !may_access_direct_pkt_data(env, NULL, BPF_WRITE);
18253 *addr = (unsigned long)bpf_dynptr_from_skb_rdonly;
18255 /* restore env->seen_direct_write to its original value, since
18256 * may_access_direct_pkt_data mutates it
18258 env->seen_direct_write = seen_direct_write;
18262 static void __fixup_collection_insert_kfunc(struct bpf_insn_aux_data *insn_aux,
18263 u16 struct_meta_reg,
18264 u16 node_offset_reg,
18265 struct bpf_insn *insn,
18266 struct bpf_insn *insn_buf,
18269 struct btf_struct_meta *kptr_struct_meta = insn_aux->kptr_struct_meta;
18270 struct bpf_insn addr[2] = { BPF_LD_IMM64(struct_meta_reg, (long)kptr_struct_meta) };
18272 insn_buf[0] = addr[0];
18273 insn_buf[1] = addr[1];
18274 insn_buf[2] = BPF_MOV64_IMM(node_offset_reg, insn_aux->insert_off);
18275 insn_buf[3] = *insn;
18279 static int fixup_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
18280 struct bpf_insn *insn_buf, int insn_idx, int *cnt)
18282 const struct bpf_kfunc_desc *desc;
18285 verbose(env, "invalid kernel function call not eliminated in verifier pass\n");
18291 /* insn->imm has the btf func_id. Replace it with an offset relative to
18292 * __bpf_call_base, unless the JIT needs to call functions that are
18293 * further than 32 bits away (bpf_jit_supports_far_kfunc_call()).
18295 desc = find_kfunc_desc(env->prog, insn->imm, insn->off);
18297 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
18302 if (!bpf_jit_supports_far_kfunc_call())
18303 insn->imm = BPF_CALL_IMM(desc->addr);
18306 if (desc->func_id == special_kfunc_list[KF_bpf_obj_new_impl]) {
18307 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18308 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
18309 u64 obj_new_size = env->insn_aux_data[insn_idx].obj_new_size;
18311 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_1, obj_new_size);
18312 insn_buf[1] = addr[0];
18313 insn_buf[2] = addr[1];
18314 insn_buf[3] = *insn;
18316 } else if (desc->func_id == special_kfunc_list[KF_bpf_obj_drop_impl] ||
18317 desc->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) {
18318 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18319 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
18321 if (desc->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl] &&
18322 !kptr_struct_meta) {
18323 verbose(env, "verifier internal error: kptr_struct_meta expected at insn_idx %d\n",
18328 insn_buf[0] = addr[0];
18329 insn_buf[1] = addr[1];
18330 insn_buf[2] = *insn;
18332 } else if (desc->func_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
18333 desc->func_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
18334 desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
18335 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18336 int struct_meta_reg = BPF_REG_3;
18337 int node_offset_reg = BPF_REG_4;
18339 /* rbtree_add has extra 'less' arg, so args-to-fixup are in diff regs */
18340 if (desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
18341 struct_meta_reg = BPF_REG_4;
18342 node_offset_reg = BPF_REG_5;
18345 if (!kptr_struct_meta) {
18346 verbose(env, "verifier internal error: kptr_struct_meta expected at insn_idx %d\n",
18351 __fixup_collection_insert_kfunc(&env->insn_aux_data[insn_idx], struct_meta_reg,
18352 node_offset_reg, insn, insn_buf, cnt);
18353 } else if (desc->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx] ||
18354 desc->func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
18355 insn_buf[0] = BPF_MOV64_REG(BPF_REG_0, BPF_REG_1);
18361 /* Do various post-verification rewrites in a single program pass.
18362 * These rewrites simplify JIT and interpreter implementations.
18364 static int do_misc_fixups(struct bpf_verifier_env *env)
18366 struct bpf_prog *prog = env->prog;
18367 enum bpf_attach_type eatype = prog->expected_attach_type;
18368 enum bpf_prog_type prog_type = resolve_prog_type(prog);
18369 struct bpf_insn *insn = prog->insnsi;
18370 const struct bpf_func_proto *fn;
18371 const int insn_cnt = prog->len;
18372 const struct bpf_map_ops *ops;
18373 struct bpf_insn_aux_data *aux;
18374 struct bpf_insn insn_buf[16];
18375 struct bpf_prog *new_prog;
18376 struct bpf_map *map_ptr;
18377 int i, ret, cnt, delta = 0;
18379 for (i = 0; i < insn_cnt; i++, insn++) {
18380 /* Make divide-by-zero exceptions impossible. */
18381 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
18382 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
18383 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
18384 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
18385 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
18386 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
18387 struct bpf_insn *patchlet;
18388 struct bpf_insn chk_and_div[] = {
18389 /* [R,W]x div 0 -> 0 */
18390 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
18391 BPF_JNE | BPF_K, insn->src_reg,
18393 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
18394 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
18397 struct bpf_insn chk_and_mod[] = {
18398 /* [R,W]x mod 0 -> [R,W]x */
18399 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
18400 BPF_JEQ | BPF_K, insn->src_reg,
18401 0, 1 + (is64 ? 0 : 1), 0),
18403 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
18404 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
18407 patchlet = isdiv ? chk_and_div : chk_and_mod;
18408 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
18409 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
18411 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
18416 env->prog = prog = new_prog;
18417 insn = new_prog->insnsi + i + delta;
18421 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
18422 if (BPF_CLASS(insn->code) == BPF_LD &&
18423 (BPF_MODE(insn->code) == BPF_ABS ||
18424 BPF_MODE(insn->code) == BPF_IND)) {
18425 cnt = env->ops->gen_ld_abs(insn, insn_buf);
18426 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
18427 verbose(env, "bpf verifier is misconfigured\n");
18431 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18436 env->prog = prog = new_prog;
18437 insn = new_prog->insnsi + i + delta;
18441 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
18442 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
18443 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
18444 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
18445 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
18446 struct bpf_insn *patch = &insn_buf[0];
18447 bool issrc, isneg, isimm;
18450 aux = &env->insn_aux_data[i + delta];
18451 if (!aux->alu_state ||
18452 aux->alu_state == BPF_ALU_NON_POINTER)
18455 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
18456 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
18457 BPF_ALU_SANITIZE_SRC;
18458 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
18460 off_reg = issrc ? insn->src_reg : insn->dst_reg;
18462 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
18465 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
18466 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
18467 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
18468 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
18469 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
18470 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
18471 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
18474 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
18475 insn->src_reg = BPF_REG_AX;
18477 insn->code = insn->code == code_add ?
18478 code_sub : code_add;
18480 if (issrc && isneg && !isimm)
18481 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
18482 cnt = patch - insn_buf;
18484 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18489 env->prog = prog = new_prog;
18490 insn = new_prog->insnsi + i + delta;
18494 if (insn->code != (BPF_JMP | BPF_CALL))
18496 if (insn->src_reg == BPF_PSEUDO_CALL)
18498 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
18499 ret = fixup_kfunc_call(env, insn, insn_buf, i + delta, &cnt);
18505 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18510 env->prog = prog = new_prog;
18511 insn = new_prog->insnsi + i + delta;
18515 if (insn->imm == BPF_FUNC_get_route_realm)
18516 prog->dst_needed = 1;
18517 if (insn->imm == BPF_FUNC_get_prandom_u32)
18518 bpf_user_rnd_init_once();
18519 if (insn->imm == BPF_FUNC_override_return)
18520 prog->kprobe_override = 1;
18521 if (insn->imm == BPF_FUNC_tail_call) {
18522 /* If we tail call into other programs, we
18523 * cannot make any assumptions since they can
18524 * be replaced dynamically during runtime in
18525 * the program array.
18527 prog->cb_access = 1;
18528 if (!allow_tail_call_in_subprogs(env))
18529 prog->aux->stack_depth = MAX_BPF_STACK;
18530 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
18532 /* mark bpf_tail_call as different opcode to avoid
18533 * conditional branch in the interpreter for every normal
18534 * call and to prevent accidental JITing by JIT compiler
18535 * that doesn't support bpf_tail_call yet
18538 insn->code = BPF_JMP | BPF_TAIL_CALL;
18540 aux = &env->insn_aux_data[i + delta];
18541 if (env->bpf_capable && !prog->blinding_requested &&
18542 prog->jit_requested &&
18543 !bpf_map_key_poisoned(aux) &&
18544 !bpf_map_ptr_poisoned(aux) &&
18545 !bpf_map_ptr_unpriv(aux)) {
18546 struct bpf_jit_poke_descriptor desc = {
18547 .reason = BPF_POKE_REASON_TAIL_CALL,
18548 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
18549 .tail_call.key = bpf_map_key_immediate(aux),
18550 .insn_idx = i + delta,
18553 ret = bpf_jit_add_poke_descriptor(prog, &desc);
18555 verbose(env, "adding tail call poke descriptor failed\n");
18559 insn->imm = ret + 1;
18563 if (!bpf_map_ptr_unpriv(aux))
18566 /* instead of changing every JIT dealing with tail_call
18567 * emit two extra insns:
18568 * if (index >= max_entries) goto out;
18569 * index &= array->index_mask;
18570 * to avoid out-of-bounds cpu speculation
18572 if (bpf_map_ptr_poisoned(aux)) {
18573 verbose(env, "tail_call abusing map_ptr\n");
18577 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
18578 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
18579 map_ptr->max_entries, 2);
18580 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
18581 container_of(map_ptr,
18584 insn_buf[2] = *insn;
18586 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18591 env->prog = prog = new_prog;
18592 insn = new_prog->insnsi + i + delta;
18596 if (insn->imm == BPF_FUNC_timer_set_callback) {
18597 /* The verifier will process callback_fn as many times as necessary
18598 * with different maps and the register states prepared by
18599 * set_timer_callback_state will be accurate.
18601 * The following use case is valid:
18602 * map1 is shared by prog1, prog2, prog3.
18603 * prog1 calls bpf_timer_init for some map1 elements
18604 * prog2 calls bpf_timer_set_callback for some map1 elements.
18605 * Those that were not bpf_timer_init-ed will return -EINVAL.
18606 * prog3 calls bpf_timer_start for some map1 elements.
18607 * Those that were not both bpf_timer_init-ed and
18608 * bpf_timer_set_callback-ed will return -EINVAL.
18610 struct bpf_insn ld_addrs[2] = {
18611 BPF_LD_IMM64(BPF_REG_3, (long)prog->aux),
18614 insn_buf[0] = ld_addrs[0];
18615 insn_buf[1] = ld_addrs[1];
18616 insn_buf[2] = *insn;
18619 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18624 env->prog = prog = new_prog;
18625 insn = new_prog->insnsi + i + delta;
18626 goto patch_call_imm;
18629 if (is_storage_get_function(insn->imm)) {
18630 if (!env->prog->aux->sleepable ||
18631 env->insn_aux_data[i + delta].storage_get_func_atomic)
18632 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_ATOMIC);
18634 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_KERNEL);
18635 insn_buf[1] = *insn;
18638 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18643 env->prog = prog = new_prog;
18644 insn = new_prog->insnsi + i + delta;
18645 goto patch_call_imm;
18648 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
18649 * and other inlining handlers are currently limited to 64 bit
18652 if (prog->jit_requested && BITS_PER_LONG == 64 &&
18653 (insn->imm == BPF_FUNC_map_lookup_elem ||
18654 insn->imm == BPF_FUNC_map_update_elem ||
18655 insn->imm == BPF_FUNC_map_delete_elem ||
18656 insn->imm == BPF_FUNC_map_push_elem ||
18657 insn->imm == BPF_FUNC_map_pop_elem ||
18658 insn->imm == BPF_FUNC_map_peek_elem ||
18659 insn->imm == BPF_FUNC_redirect_map ||
18660 insn->imm == BPF_FUNC_for_each_map_elem ||
18661 insn->imm == BPF_FUNC_map_lookup_percpu_elem)) {
18662 aux = &env->insn_aux_data[i + delta];
18663 if (bpf_map_ptr_poisoned(aux))
18664 goto patch_call_imm;
18666 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
18667 ops = map_ptr->ops;
18668 if (insn->imm == BPF_FUNC_map_lookup_elem &&
18669 ops->map_gen_lookup) {
18670 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
18671 if (cnt == -EOPNOTSUPP)
18672 goto patch_map_ops_generic;
18673 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
18674 verbose(env, "bpf verifier is misconfigured\n");
18678 new_prog = bpf_patch_insn_data(env, i + delta,
18684 env->prog = prog = new_prog;
18685 insn = new_prog->insnsi + i + delta;
18689 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
18690 (void *(*)(struct bpf_map *map, void *key))NULL));
18691 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
18692 (long (*)(struct bpf_map *map, void *key))NULL));
18693 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
18694 (long (*)(struct bpf_map *map, void *key, void *value,
18696 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
18697 (long (*)(struct bpf_map *map, void *value,
18699 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
18700 (long (*)(struct bpf_map *map, void *value))NULL));
18701 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
18702 (long (*)(struct bpf_map *map, void *value))NULL));
18703 BUILD_BUG_ON(!__same_type(ops->map_redirect,
18704 (long (*)(struct bpf_map *map, u64 index, u64 flags))NULL));
18705 BUILD_BUG_ON(!__same_type(ops->map_for_each_callback,
18706 (long (*)(struct bpf_map *map,
18707 bpf_callback_t callback_fn,
18708 void *callback_ctx,
18710 BUILD_BUG_ON(!__same_type(ops->map_lookup_percpu_elem,
18711 (void *(*)(struct bpf_map *map, void *key, u32 cpu))NULL));
18713 patch_map_ops_generic:
18714 switch (insn->imm) {
18715 case BPF_FUNC_map_lookup_elem:
18716 insn->imm = BPF_CALL_IMM(ops->map_lookup_elem);
18718 case BPF_FUNC_map_update_elem:
18719 insn->imm = BPF_CALL_IMM(ops->map_update_elem);
18721 case BPF_FUNC_map_delete_elem:
18722 insn->imm = BPF_CALL_IMM(ops->map_delete_elem);
18724 case BPF_FUNC_map_push_elem:
18725 insn->imm = BPF_CALL_IMM(ops->map_push_elem);
18727 case BPF_FUNC_map_pop_elem:
18728 insn->imm = BPF_CALL_IMM(ops->map_pop_elem);
18730 case BPF_FUNC_map_peek_elem:
18731 insn->imm = BPF_CALL_IMM(ops->map_peek_elem);
18733 case BPF_FUNC_redirect_map:
18734 insn->imm = BPF_CALL_IMM(ops->map_redirect);
18736 case BPF_FUNC_for_each_map_elem:
18737 insn->imm = BPF_CALL_IMM(ops->map_for_each_callback);
18739 case BPF_FUNC_map_lookup_percpu_elem:
18740 insn->imm = BPF_CALL_IMM(ops->map_lookup_percpu_elem);
18744 goto patch_call_imm;
18747 /* Implement bpf_jiffies64 inline. */
18748 if (prog->jit_requested && BITS_PER_LONG == 64 &&
18749 insn->imm == BPF_FUNC_jiffies64) {
18750 struct bpf_insn ld_jiffies_addr[2] = {
18751 BPF_LD_IMM64(BPF_REG_0,
18752 (unsigned long)&jiffies),
18755 insn_buf[0] = ld_jiffies_addr[0];
18756 insn_buf[1] = ld_jiffies_addr[1];
18757 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
18761 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
18767 env->prog = prog = new_prog;
18768 insn = new_prog->insnsi + i + delta;
18772 /* Implement bpf_get_func_arg inline. */
18773 if (prog_type == BPF_PROG_TYPE_TRACING &&
18774 insn->imm == BPF_FUNC_get_func_arg) {
18775 /* Load nr_args from ctx - 8 */
18776 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
18777 insn_buf[1] = BPF_JMP32_REG(BPF_JGE, BPF_REG_2, BPF_REG_0, 6);
18778 insn_buf[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_2, 3);
18779 insn_buf[3] = BPF_ALU64_REG(BPF_ADD, BPF_REG_2, BPF_REG_1);
18780 insn_buf[4] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_2, 0);
18781 insn_buf[5] = BPF_STX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
18782 insn_buf[6] = BPF_MOV64_IMM(BPF_REG_0, 0);
18783 insn_buf[7] = BPF_JMP_A(1);
18784 insn_buf[8] = BPF_MOV64_IMM(BPF_REG_0, -EINVAL);
18787 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18792 env->prog = prog = new_prog;
18793 insn = new_prog->insnsi + i + delta;
18797 /* Implement bpf_get_func_ret inline. */
18798 if (prog_type == BPF_PROG_TYPE_TRACING &&
18799 insn->imm == BPF_FUNC_get_func_ret) {
18800 if (eatype == BPF_TRACE_FEXIT ||
18801 eatype == BPF_MODIFY_RETURN) {
18802 /* Load nr_args from ctx - 8 */
18803 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
18804 insn_buf[1] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_0, 3);
18805 insn_buf[2] = BPF_ALU64_REG(BPF_ADD, BPF_REG_0, BPF_REG_1);
18806 insn_buf[3] = BPF_LDX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
18807 insn_buf[4] = BPF_STX_MEM(BPF_DW, BPF_REG_2, BPF_REG_3, 0);
18808 insn_buf[5] = BPF_MOV64_IMM(BPF_REG_0, 0);
18811 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, -EOPNOTSUPP);
18815 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18820 env->prog = prog = new_prog;
18821 insn = new_prog->insnsi + i + delta;
18825 /* Implement get_func_arg_cnt inline. */
18826 if (prog_type == BPF_PROG_TYPE_TRACING &&
18827 insn->imm == BPF_FUNC_get_func_arg_cnt) {
18828 /* Load nr_args from ctx - 8 */
18829 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
18831 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
18835 env->prog = prog = new_prog;
18836 insn = new_prog->insnsi + i + delta;
18840 /* Implement bpf_get_func_ip inline. */
18841 if (prog_type == BPF_PROG_TYPE_TRACING &&
18842 insn->imm == BPF_FUNC_get_func_ip) {
18843 /* Load IP address from ctx - 16 */
18844 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -16);
18846 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
18850 env->prog = prog = new_prog;
18851 insn = new_prog->insnsi + i + delta;
18856 fn = env->ops->get_func_proto(insn->imm, env->prog);
18857 /* all functions that have prototype and verifier allowed
18858 * programs to call them, must be real in-kernel functions
18862 "kernel subsystem misconfigured func %s#%d\n",
18863 func_id_name(insn->imm), insn->imm);
18866 insn->imm = fn->func - __bpf_call_base;
18869 /* Since poke tab is now finalized, publish aux to tracker. */
18870 for (i = 0; i < prog->aux->size_poke_tab; i++) {
18871 map_ptr = prog->aux->poke_tab[i].tail_call.map;
18872 if (!map_ptr->ops->map_poke_track ||
18873 !map_ptr->ops->map_poke_untrack ||
18874 !map_ptr->ops->map_poke_run) {
18875 verbose(env, "bpf verifier is misconfigured\n");
18879 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
18881 verbose(env, "tracking tail call prog failed\n");
18886 sort_kfunc_descs_by_imm_off(env->prog);
18891 static struct bpf_prog *inline_bpf_loop(struct bpf_verifier_env *env,
18894 u32 callback_subprogno,
18897 s32 r6_offset = stack_base + 0 * BPF_REG_SIZE;
18898 s32 r7_offset = stack_base + 1 * BPF_REG_SIZE;
18899 s32 r8_offset = stack_base + 2 * BPF_REG_SIZE;
18900 int reg_loop_max = BPF_REG_6;
18901 int reg_loop_cnt = BPF_REG_7;
18902 int reg_loop_ctx = BPF_REG_8;
18904 struct bpf_prog *new_prog;
18905 u32 callback_start;
18906 u32 call_insn_offset;
18907 s32 callback_offset;
18909 /* This represents an inlined version of bpf_iter.c:bpf_loop,
18910 * be careful to modify this code in sync.
18912 struct bpf_insn insn_buf[] = {
18913 /* Return error and jump to the end of the patch if
18914 * expected number of iterations is too big.
18916 BPF_JMP_IMM(BPF_JLE, BPF_REG_1, BPF_MAX_LOOPS, 2),
18917 BPF_MOV32_IMM(BPF_REG_0, -E2BIG),
18918 BPF_JMP_IMM(BPF_JA, 0, 0, 16),
18919 /* spill R6, R7, R8 to use these as loop vars */
18920 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_6, r6_offset),
18921 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_7, r7_offset),
18922 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_8, r8_offset),
18923 /* initialize loop vars */
18924 BPF_MOV64_REG(reg_loop_max, BPF_REG_1),
18925 BPF_MOV32_IMM(reg_loop_cnt, 0),
18926 BPF_MOV64_REG(reg_loop_ctx, BPF_REG_3),
18928 * if reg_loop_cnt >= reg_loop_max skip the loop body
18930 BPF_JMP_REG(BPF_JGE, reg_loop_cnt, reg_loop_max, 5),
18932 * correct callback offset would be set after patching
18934 BPF_MOV64_REG(BPF_REG_1, reg_loop_cnt),
18935 BPF_MOV64_REG(BPF_REG_2, reg_loop_ctx),
18937 /* increment loop counter */
18938 BPF_ALU64_IMM(BPF_ADD, reg_loop_cnt, 1),
18939 /* jump to loop header if callback returned 0 */
18940 BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, -6),
18941 /* return value of bpf_loop,
18942 * set R0 to the number of iterations
18944 BPF_MOV64_REG(BPF_REG_0, reg_loop_cnt),
18945 /* restore original values of R6, R7, R8 */
18946 BPF_LDX_MEM(BPF_DW, BPF_REG_6, BPF_REG_10, r6_offset),
18947 BPF_LDX_MEM(BPF_DW, BPF_REG_7, BPF_REG_10, r7_offset),
18948 BPF_LDX_MEM(BPF_DW, BPF_REG_8, BPF_REG_10, r8_offset),
18951 *cnt = ARRAY_SIZE(insn_buf);
18952 new_prog = bpf_patch_insn_data(env, position, insn_buf, *cnt);
18956 /* callback start is known only after patching */
18957 callback_start = env->subprog_info[callback_subprogno].start;
18958 /* Note: insn_buf[12] is an offset of BPF_CALL_REL instruction */
18959 call_insn_offset = position + 12;
18960 callback_offset = callback_start - call_insn_offset - 1;
18961 new_prog->insnsi[call_insn_offset].imm = callback_offset;
18966 static bool is_bpf_loop_call(struct bpf_insn *insn)
18968 return insn->code == (BPF_JMP | BPF_CALL) &&
18969 insn->src_reg == 0 &&
18970 insn->imm == BPF_FUNC_loop;
18973 /* For all sub-programs in the program (including main) check
18974 * insn_aux_data to see if there are bpf_loop calls that require
18975 * inlining. If such calls are found the calls are replaced with a
18976 * sequence of instructions produced by `inline_bpf_loop` function and
18977 * subprog stack_depth is increased by the size of 3 registers.
18978 * This stack space is used to spill values of the R6, R7, R8. These
18979 * registers are used to store the loop bound, counter and context
18982 static int optimize_bpf_loop(struct bpf_verifier_env *env)
18984 struct bpf_subprog_info *subprogs = env->subprog_info;
18985 int i, cur_subprog = 0, cnt, delta = 0;
18986 struct bpf_insn *insn = env->prog->insnsi;
18987 int insn_cnt = env->prog->len;
18988 u16 stack_depth = subprogs[cur_subprog].stack_depth;
18989 u16 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
18990 u16 stack_depth_extra = 0;
18992 for (i = 0; i < insn_cnt; i++, insn++) {
18993 struct bpf_loop_inline_state *inline_state =
18994 &env->insn_aux_data[i + delta].loop_inline_state;
18996 if (is_bpf_loop_call(insn) && inline_state->fit_for_inline) {
18997 struct bpf_prog *new_prog;
18999 stack_depth_extra = BPF_REG_SIZE * 3 + stack_depth_roundup;
19000 new_prog = inline_bpf_loop(env,
19002 -(stack_depth + stack_depth_extra),
19003 inline_state->callback_subprogno,
19009 env->prog = new_prog;
19010 insn = new_prog->insnsi + i + delta;
19013 if (subprogs[cur_subprog + 1].start == i + delta + 1) {
19014 subprogs[cur_subprog].stack_depth += stack_depth_extra;
19016 stack_depth = subprogs[cur_subprog].stack_depth;
19017 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
19018 stack_depth_extra = 0;
19022 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
19027 static void free_states(struct bpf_verifier_env *env)
19029 struct bpf_verifier_state_list *sl, *sln;
19032 sl = env->free_list;
19035 free_verifier_state(&sl->state, false);
19039 env->free_list = NULL;
19041 if (!env->explored_states)
19044 for (i = 0; i < state_htab_size(env); i++) {
19045 sl = env->explored_states[i];
19049 free_verifier_state(&sl->state, false);
19053 env->explored_states[i] = NULL;
19057 static int do_check_common(struct bpf_verifier_env *env, int subprog)
19059 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
19060 struct bpf_verifier_state *state;
19061 struct bpf_reg_state *regs;
19064 env->prev_linfo = NULL;
19067 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
19070 state->curframe = 0;
19071 state->speculative = false;
19072 state->branches = 1;
19073 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
19074 if (!state->frame[0]) {
19078 env->cur_state = state;
19079 init_func_state(env, state->frame[0],
19080 BPF_MAIN_FUNC /* callsite */,
19083 state->first_insn_idx = env->subprog_info[subprog].start;
19084 state->last_insn_idx = -1;
19086 regs = state->frame[state->curframe]->regs;
19087 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
19088 ret = btf_prepare_func_args(env, subprog, regs);
19091 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
19092 if (regs[i].type == PTR_TO_CTX)
19093 mark_reg_known_zero(env, regs, i);
19094 else if (regs[i].type == SCALAR_VALUE)
19095 mark_reg_unknown(env, regs, i);
19096 else if (base_type(regs[i].type) == PTR_TO_MEM) {
19097 const u32 mem_size = regs[i].mem_size;
19099 mark_reg_known_zero(env, regs, i);
19100 regs[i].mem_size = mem_size;
19101 regs[i].id = ++env->id_gen;
19105 /* 1st arg to a function */
19106 regs[BPF_REG_1].type = PTR_TO_CTX;
19107 mark_reg_known_zero(env, regs, BPF_REG_1);
19108 ret = btf_check_subprog_arg_match(env, subprog, regs);
19109 if (ret == -EFAULT)
19110 /* unlikely verifier bug. abort.
19111 * ret == 0 and ret < 0 are sadly acceptable for
19112 * main() function due to backward compatibility.
19113 * Like socket filter program may be written as:
19114 * int bpf_prog(struct pt_regs *ctx)
19115 * and never dereference that ctx in the program.
19116 * 'struct pt_regs' is a type mismatch for socket
19117 * filter that should be using 'struct __sk_buff'.
19122 ret = do_check(env);
19124 /* check for NULL is necessary, since cur_state can be freed inside
19125 * do_check() under memory pressure.
19127 if (env->cur_state) {
19128 free_verifier_state(env->cur_state, true);
19129 env->cur_state = NULL;
19131 while (!pop_stack(env, NULL, NULL, false));
19132 if (!ret && pop_log)
19133 bpf_vlog_reset(&env->log, 0);
19138 /* Verify all global functions in a BPF program one by one based on their BTF.
19139 * All global functions must pass verification. Otherwise the whole program is rejected.
19150 * foo() will be verified first for R1=any_scalar_value. During verification it
19151 * will be assumed that bar() already verified successfully and call to bar()
19152 * from foo() will be checked for type match only. Later bar() will be verified
19153 * independently to check that it's safe for R1=any_scalar_value.
19155 static int do_check_subprogs(struct bpf_verifier_env *env)
19157 struct bpf_prog_aux *aux = env->prog->aux;
19160 if (!aux->func_info)
19163 for (i = 1; i < env->subprog_cnt; i++) {
19164 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
19166 env->insn_idx = env->subprog_info[i].start;
19167 WARN_ON_ONCE(env->insn_idx == 0);
19168 ret = do_check_common(env, i);
19171 } else if (env->log.level & BPF_LOG_LEVEL) {
19173 "Func#%d is safe for any args that match its prototype\n",
19180 static int do_check_main(struct bpf_verifier_env *env)
19185 ret = do_check_common(env, 0);
19187 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
19192 static void print_verification_stats(struct bpf_verifier_env *env)
19196 if (env->log.level & BPF_LOG_STATS) {
19197 verbose(env, "verification time %lld usec\n",
19198 div_u64(env->verification_time, 1000));
19199 verbose(env, "stack depth ");
19200 for (i = 0; i < env->subprog_cnt; i++) {
19201 u32 depth = env->subprog_info[i].stack_depth;
19203 verbose(env, "%d", depth);
19204 if (i + 1 < env->subprog_cnt)
19207 verbose(env, "\n");
19209 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
19210 "total_states %d peak_states %d mark_read %d\n",
19211 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
19212 env->max_states_per_insn, env->total_states,
19213 env->peak_states, env->longest_mark_read_walk);
19216 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
19218 const struct btf_type *t, *func_proto;
19219 const struct bpf_struct_ops *st_ops;
19220 const struct btf_member *member;
19221 struct bpf_prog *prog = env->prog;
19222 u32 btf_id, member_idx;
19225 if (!prog->gpl_compatible) {
19226 verbose(env, "struct ops programs must have a GPL compatible license\n");
19230 btf_id = prog->aux->attach_btf_id;
19231 st_ops = bpf_struct_ops_find(btf_id);
19233 verbose(env, "attach_btf_id %u is not a supported struct\n",
19239 member_idx = prog->expected_attach_type;
19240 if (member_idx >= btf_type_vlen(t)) {
19241 verbose(env, "attach to invalid member idx %u of struct %s\n",
19242 member_idx, st_ops->name);
19246 member = &btf_type_member(t)[member_idx];
19247 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
19248 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
19251 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
19252 mname, member_idx, st_ops->name);
19256 if (st_ops->check_member) {
19257 int err = st_ops->check_member(t, member, prog);
19260 verbose(env, "attach to unsupported member %s of struct %s\n",
19261 mname, st_ops->name);
19266 prog->aux->attach_func_proto = func_proto;
19267 prog->aux->attach_func_name = mname;
19268 env->ops = st_ops->verifier_ops;
19272 #define SECURITY_PREFIX "security_"
19274 static int check_attach_modify_return(unsigned long addr, const char *func_name)
19276 if (within_error_injection_list(addr) ||
19277 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
19283 /* list of non-sleepable functions that are otherwise on
19284 * ALLOW_ERROR_INJECTION list
19286 BTF_SET_START(btf_non_sleepable_error_inject)
19287 /* Three functions below can be called from sleepable and non-sleepable context.
19288 * Assume non-sleepable from bpf safety point of view.
19290 BTF_ID(func, __filemap_add_folio)
19291 BTF_ID(func, should_fail_alloc_page)
19292 BTF_ID(func, should_failslab)
19293 BTF_SET_END(btf_non_sleepable_error_inject)
19295 static int check_non_sleepable_error_inject(u32 btf_id)
19297 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
19300 int bpf_check_attach_target(struct bpf_verifier_log *log,
19301 const struct bpf_prog *prog,
19302 const struct bpf_prog *tgt_prog,
19304 struct bpf_attach_target_info *tgt_info)
19306 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
19307 const char prefix[] = "btf_trace_";
19308 int ret = 0, subprog = -1, i;
19309 const struct btf_type *t;
19310 bool conservative = true;
19314 struct module *mod = NULL;
19317 bpf_log(log, "Tracing programs must provide btf_id\n");
19320 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
19323 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
19326 t = btf_type_by_id(btf, btf_id);
19328 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
19331 tname = btf_name_by_offset(btf, t->name_off);
19333 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
19337 struct bpf_prog_aux *aux = tgt_prog->aux;
19339 if (bpf_prog_is_dev_bound(prog->aux) &&
19340 !bpf_prog_dev_bound_match(prog, tgt_prog)) {
19341 bpf_log(log, "Target program bound device mismatch");
19345 for (i = 0; i < aux->func_info_cnt; i++)
19346 if (aux->func_info[i].type_id == btf_id) {
19350 if (subprog == -1) {
19351 bpf_log(log, "Subprog %s doesn't exist\n", tname);
19354 conservative = aux->func_info_aux[subprog].unreliable;
19355 if (prog_extension) {
19356 if (conservative) {
19358 "Cannot replace static functions\n");
19361 if (!prog->jit_requested) {
19363 "Extension programs should be JITed\n");
19367 if (!tgt_prog->jited) {
19368 bpf_log(log, "Can attach to only JITed progs\n");
19371 if (tgt_prog->type == prog->type) {
19372 /* Cannot fentry/fexit another fentry/fexit program.
19373 * Cannot attach program extension to another extension.
19374 * It's ok to attach fentry/fexit to extension program.
19376 bpf_log(log, "Cannot recursively attach\n");
19379 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
19381 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
19382 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
19383 /* Program extensions can extend all program types
19384 * except fentry/fexit. The reason is the following.
19385 * The fentry/fexit programs are used for performance
19386 * analysis, stats and can be attached to any program
19387 * type except themselves. When extension program is
19388 * replacing XDP function it is necessary to allow
19389 * performance analysis of all functions. Both original
19390 * XDP program and its program extension. Hence
19391 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
19392 * allowed. If extending of fentry/fexit was allowed it
19393 * would be possible to create long call chain
19394 * fentry->extension->fentry->extension beyond
19395 * reasonable stack size. Hence extending fentry is not
19398 bpf_log(log, "Cannot extend fentry/fexit\n");
19402 if (prog_extension) {
19403 bpf_log(log, "Cannot replace kernel functions\n");
19408 switch (prog->expected_attach_type) {
19409 case BPF_TRACE_RAW_TP:
19412 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
19415 if (!btf_type_is_typedef(t)) {
19416 bpf_log(log, "attach_btf_id %u is not a typedef\n",
19420 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
19421 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
19425 tname += sizeof(prefix) - 1;
19426 t = btf_type_by_id(btf, t->type);
19427 if (!btf_type_is_ptr(t))
19428 /* should never happen in valid vmlinux build */
19430 t = btf_type_by_id(btf, t->type);
19431 if (!btf_type_is_func_proto(t))
19432 /* should never happen in valid vmlinux build */
19436 case BPF_TRACE_ITER:
19437 if (!btf_type_is_func(t)) {
19438 bpf_log(log, "attach_btf_id %u is not a function\n",
19442 t = btf_type_by_id(btf, t->type);
19443 if (!btf_type_is_func_proto(t))
19445 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
19450 if (!prog_extension)
19453 case BPF_MODIFY_RETURN:
19455 case BPF_LSM_CGROUP:
19456 case BPF_TRACE_FENTRY:
19457 case BPF_TRACE_FEXIT:
19458 if (!btf_type_is_func(t)) {
19459 bpf_log(log, "attach_btf_id %u is not a function\n",
19463 if (prog_extension &&
19464 btf_check_type_match(log, prog, btf, t))
19466 t = btf_type_by_id(btf, t->type);
19467 if (!btf_type_is_func_proto(t))
19470 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
19471 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
19472 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
19475 if (tgt_prog && conservative)
19478 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
19484 addr = (long) tgt_prog->bpf_func;
19486 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
19488 if (btf_is_module(btf)) {
19489 mod = btf_try_get_module(btf);
19491 addr = find_kallsyms_symbol_value(mod, tname);
19495 addr = kallsyms_lookup_name(tname);
19500 "The address of function %s cannot be found\n",
19506 if (prog->aux->sleepable) {
19508 switch (prog->type) {
19509 case BPF_PROG_TYPE_TRACING:
19511 /* fentry/fexit/fmod_ret progs can be sleepable if they are
19512 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
19514 if (!check_non_sleepable_error_inject(btf_id) &&
19515 within_error_injection_list(addr))
19517 /* fentry/fexit/fmod_ret progs can also be sleepable if they are
19518 * in the fmodret id set with the KF_SLEEPABLE flag.
19521 u32 *flags = btf_kfunc_is_modify_return(btf, btf_id,
19524 if (flags && (*flags & KF_SLEEPABLE))
19528 case BPF_PROG_TYPE_LSM:
19529 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
19530 * Only some of them are sleepable.
19532 if (bpf_lsm_is_sleepable_hook(btf_id))
19540 bpf_log(log, "%s is not sleepable\n", tname);
19543 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
19546 bpf_log(log, "can't modify return codes of BPF programs\n");
19550 if (btf_kfunc_is_modify_return(btf, btf_id, prog) ||
19551 !check_attach_modify_return(addr, tname))
19555 bpf_log(log, "%s() is not modifiable\n", tname);
19562 tgt_info->tgt_addr = addr;
19563 tgt_info->tgt_name = tname;
19564 tgt_info->tgt_type = t;
19565 tgt_info->tgt_mod = mod;
19569 BTF_SET_START(btf_id_deny)
19572 BTF_ID(func, migrate_disable)
19573 BTF_ID(func, migrate_enable)
19575 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
19576 BTF_ID(func, rcu_read_unlock_strict)
19578 #if defined(CONFIG_DEBUG_PREEMPT) || defined(CONFIG_TRACE_PREEMPT_TOGGLE)
19579 BTF_ID(func, preempt_count_add)
19580 BTF_ID(func, preempt_count_sub)
19582 #ifdef CONFIG_PREEMPT_RCU
19583 BTF_ID(func, __rcu_read_lock)
19584 BTF_ID(func, __rcu_read_unlock)
19586 BTF_SET_END(btf_id_deny)
19588 static bool can_be_sleepable(struct bpf_prog *prog)
19590 if (prog->type == BPF_PROG_TYPE_TRACING) {
19591 switch (prog->expected_attach_type) {
19592 case BPF_TRACE_FENTRY:
19593 case BPF_TRACE_FEXIT:
19594 case BPF_MODIFY_RETURN:
19595 case BPF_TRACE_ITER:
19601 return prog->type == BPF_PROG_TYPE_LSM ||
19602 prog->type == BPF_PROG_TYPE_KPROBE /* only for uprobes */ ||
19603 prog->type == BPF_PROG_TYPE_STRUCT_OPS;
19606 static int check_attach_btf_id(struct bpf_verifier_env *env)
19608 struct bpf_prog *prog = env->prog;
19609 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
19610 struct bpf_attach_target_info tgt_info = {};
19611 u32 btf_id = prog->aux->attach_btf_id;
19612 struct bpf_trampoline *tr;
19616 if (prog->type == BPF_PROG_TYPE_SYSCALL) {
19617 if (prog->aux->sleepable)
19618 /* attach_btf_id checked to be zero already */
19620 verbose(env, "Syscall programs can only be sleepable\n");
19624 if (prog->aux->sleepable && !can_be_sleepable(prog)) {
19625 verbose(env, "Only fentry/fexit/fmod_ret, lsm, iter, uprobe, and struct_ops programs can be sleepable\n");
19629 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
19630 return check_struct_ops_btf_id(env);
19632 if (prog->type != BPF_PROG_TYPE_TRACING &&
19633 prog->type != BPF_PROG_TYPE_LSM &&
19634 prog->type != BPF_PROG_TYPE_EXT)
19637 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
19641 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
19642 /* to make freplace equivalent to their targets, they need to
19643 * inherit env->ops and expected_attach_type for the rest of the
19646 env->ops = bpf_verifier_ops[tgt_prog->type];
19647 prog->expected_attach_type = tgt_prog->expected_attach_type;
19650 /* store info about the attachment target that will be used later */
19651 prog->aux->attach_func_proto = tgt_info.tgt_type;
19652 prog->aux->attach_func_name = tgt_info.tgt_name;
19653 prog->aux->mod = tgt_info.tgt_mod;
19656 prog->aux->saved_dst_prog_type = tgt_prog->type;
19657 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
19660 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
19661 prog->aux->attach_btf_trace = true;
19663 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
19664 if (!bpf_iter_prog_supported(prog))
19669 if (prog->type == BPF_PROG_TYPE_LSM) {
19670 ret = bpf_lsm_verify_prog(&env->log, prog);
19673 } else if (prog->type == BPF_PROG_TYPE_TRACING &&
19674 btf_id_set_contains(&btf_id_deny, btf_id)) {
19678 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
19679 tr = bpf_trampoline_get(key, &tgt_info);
19683 if (tgt_prog && tgt_prog->aux->tail_call_reachable)
19684 tr->flags = BPF_TRAMP_F_TAIL_CALL_CTX;
19686 prog->aux->dst_trampoline = tr;
19690 struct btf *bpf_get_btf_vmlinux(void)
19692 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
19693 mutex_lock(&bpf_verifier_lock);
19695 btf_vmlinux = btf_parse_vmlinux();
19696 mutex_unlock(&bpf_verifier_lock);
19698 return btf_vmlinux;
19701 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr, __u32 uattr_size)
19703 u64 start_time = ktime_get_ns();
19704 struct bpf_verifier_env *env;
19705 int i, len, ret = -EINVAL, err;
19709 /* no program is valid */
19710 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
19713 /* 'struct bpf_verifier_env' can be global, but since it's not small,
19714 * allocate/free it every time bpf_check() is called
19716 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
19722 len = (*prog)->len;
19723 env->insn_aux_data =
19724 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
19726 if (!env->insn_aux_data)
19728 for (i = 0; i < len; i++)
19729 env->insn_aux_data[i].orig_idx = i;
19731 env->ops = bpf_verifier_ops[env->prog->type];
19732 env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
19733 is_priv = bpf_capable();
19735 bpf_get_btf_vmlinux();
19737 /* grab the mutex to protect few globals used by verifier */
19739 mutex_lock(&bpf_verifier_lock);
19741 /* user could have requested verbose verifier output
19742 * and supplied buffer to store the verification trace
19744 ret = bpf_vlog_init(&env->log, attr->log_level,
19745 (char __user *) (unsigned long) attr->log_buf,
19750 mark_verifier_state_clean(env);
19752 if (IS_ERR(btf_vmlinux)) {
19753 /* Either gcc or pahole or kernel are broken. */
19754 verbose(env, "in-kernel BTF is malformed\n");
19755 ret = PTR_ERR(btf_vmlinux);
19756 goto skip_full_check;
19759 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
19760 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
19761 env->strict_alignment = true;
19762 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
19763 env->strict_alignment = false;
19765 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
19766 env->allow_uninit_stack = bpf_allow_uninit_stack();
19767 env->bypass_spec_v1 = bpf_bypass_spec_v1();
19768 env->bypass_spec_v4 = bpf_bypass_spec_v4();
19769 env->bpf_capable = bpf_capable();
19772 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
19774 env->explored_states = kvcalloc(state_htab_size(env),
19775 sizeof(struct bpf_verifier_state_list *),
19778 if (!env->explored_states)
19779 goto skip_full_check;
19781 ret = add_subprog_and_kfunc(env);
19783 goto skip_full_check;
19785 ret = check_subprogs(env);
19787 goto skip_full_check;
19789 ret = check_btf_info(env, attr, uattr);
19791 goto skip_full_check;
19793 ret = check_attach_btf_id(env);
19795 goto skip_full_check;
19797 ret = resolve_pseudo_ldimm64(env);
19799 goto skip_full_check;
19801 if (bpf_prog_is_offloaded(env->prog->aux)) {
19802 ret = bpf_prog_offload_verifier_prep(env->prog);
19804 goto skip_full_check;
19807 ret = check_cfg(env);
19809 goto skip_full_check;
19811 ret = do_check_subprogs(env);
19812 ret = ret ?: do_check_main(env);
19814 if (ret == 0 && bpf_prog_is_offloaded(env->prog->aux))
19815 ret = bpf_prog_offload_finalize(env);
19818 kvfree(env->explored_states);
19821 ret = check_max_stack_depth(env);
19823 /* instruction rewrites happen after this point */
19825 ret = optimize_bpf_loop(env);
19829 opt_hard_wire_dead_code_branches(env);
19831 ret = opt_remove_dead_code(env);
19833 ret = opt_remove_nops(env);
19836 sanitize_dead_code(env);
19840 /* program is valid, convert *(u32*)(ctx + off) accesses */
19841 ret = convert_ctx_accesses(env);
19844 ret = do_misc_fixups(env);
19846 /* do 32-bit optimization after insn patching has done so those patched
19847 * insns could be handled correctly.
19849 if (ret == 0 && !bpf_prog_is_offloaded(env->prog->aux)) {
19850 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
19851 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
19856 ret = fixup_call_args(env);
19858 env->verification_time = ktime_get_ns() - start_time;
19859 print_verification_stats(env);
19860 env->prog->aux->verified_insns = env->insn_processed;
19862 /* preserve original error even if log finalization is successful */
19863 err = bpf_vlog_finalize(&env->log, &log_true_size);
19867 if (uattr_size >= offsetofend(union bpf_attr, log_true_size) &&
19868 copy_to_bpfptr_offset(uattr, offsetof(union bpf_attr, log_true_size),
19869 &log_true_size, sizeof(log_true_size))) {
19871 goto err_release_maps;
19875 goto err_release_maps;
19877 if (env->used_map_cnt) {
19878 /* if program passed verifier, update used_maps in bpf_prog_info */
19879 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
19880 sizeof(env->used_maps[0]),
19883 if (!env->prog->aux->used_maps) {
19885 goto err_release_maps;
19888 memcpy(env->prog->aux->used_maps, env->used_maps,
19889 sizeof(env->used_maps[0]) * env->used_map_cnt);
19890 env->prog->aux->used_map_cnt = env->used_map_cnt;
19892 if (env->used_btf_cnt) {
19893 /* if program passed verifier, update used_btfs in bpf_prog_aux */
19894 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
19895 sizeof(env->used_btfs[0]),
19897 if (!env->prog->aux->used_btfs) {
19899 goto err_release_maps;
19902 memcpy(env->prog->aux->used_btfs, env->used_btfs,
19903 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
19904 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
19906 if (env->used_map_cnt || env->used_btf_cnt) {
19907 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
19908 * bpf_ld_imm64 instructions
19910 convert_pseudo_ld_imm64(env);
19913 adjust_btf_func(env);
19916 if (!env->prog->aux->used_maps)
19917 /* if we didn't copy map pointers into bpf_prog_info, release
19918 * them now. Otherwise free_used_maps() will release them.
19921 if (!env->prog->aux->used_btfs)
19924 /* extension progs temporarily inherit the attach_type of their targets
19925 for verification purposes, so set it back to zero before returning
19927 if (env->prog->type == BPF_PROG_TYPE_EXT)
19928 env->prog->expected_attach_type = 0;
19933 mutex_unlock(&bpf_verifier_lock);
19934 vfree(env->insn_aux_data);