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>
32 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
33 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
34 [_id] = & _name ## _verifier_ops,
35 #define BPF_MAP_TYPE(_id, _ops)
36 #define BPF_LINK_TYPE(_id, _name)
37 #include <linux/bpf_types.h>
43 /* bpf_check() is a static code analyzer that walks eBPF program
44 * instruction by instruction and updates register/stack state.
45 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
47 * The first pass is depth-first-search to check that the program is a DAG.
48 * It rejects the following programs:
49 * - larger than BPF_MAXINSNS insns
50 * - if loop is present (detected via back-edge)
51 * - unreachable insns exist (shouldn't be a forest. program = one function)
52 * - out of bounds or malformed jumps
53 * The second pass is all possible path descent from the 1st insn.
54 * Since it's analyzing all paths through the program, the length of the
55 * analysis is limited to 64k insn, which may be hit even if total number of
56 * insn is less then 4K, but there are too many branches that change stack/regs.
57 * Number of 'branches to be analyzed' is limited to 1k
59 * On entry to each instruction, each register has a type, and the instruction
60 * changes the types of the registers depending on instruction semantics.
61 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
64 * All registers are 64-bit.
65 * R0 - return register
66 * R1-R5 argument passing registers
67 * R6-R9 callee saved registers
68 * R10 - frame pointer read-only
70 * At the start of BPF program the register R1 contains a pointer to bpf_context
71 * and has type PTR_TO_CTX.
73 * Verifier tracks arithmetic operations on pointers in case:
74 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
75 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
76 * 1st insn copies R10 (which has FRAME_PTR) type into R1
77 * and 2nd arithmetic instruction is pattern matched to recognize
78 * that it wants to construct a pointer to some element within stack.
79 * So after 2nd insn, the register R1 has type PTR_TO_STACK
80 * (and -20 constant is saved for further stack bounds checking).
81 * Meaning that this reg is a pointer to stack plus known immediate constant.
83 * Most of the time the registers have SCALAR_VALUE type, which
84 * means the register has some value, but it's not a valid pointer.
85 * (like pointer plus pointer becomes SCALAR_VALUE type)
87 * When verifier sees load or store instructions the type of base register
88 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
89 * four pointer types recognized by check_mem_access() function.
91 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
92 * and the range of [ptr, ptr + map's value_size) is accessible.
94 * registers used to pass values to function calls are checked against
95 * function argument constraints.
97 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
98 * It means that the register type passed to this function must be
99 * PTR_TO_STACK and it will be used inside the function as
100 * 'pointer to map element key'
102 * For example the argument constraints for bpf_map_lookup_elem():
103 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
104 * .arg1_type = ARG_CONST_MAP_PTR,
105 * .arg2_type = ARG_PTR_TO_MAP_KEY,
107 * ret_type says that this function returns 'pointer to map elem value or null'
108 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
109 * 2nd argument should be a pointer to stack, which will be used inside
110 * the helper function as a pointer to map element key.
112 * On the kernel side the helper function looks like:
113 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
115 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
116 * void *key = (void *) (unsigned long) r2;
119 * here kernel can access 'key' and 'map' pointers safely, knowing that
120 * [key, key + map->key_size) bytes are valid and were initialized on
121 * the stack of eBPF program.
124 * Corresponding eBPF program may look like:
125 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
126 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
127 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
128 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
129 * here verifier looks at prototype of map_lookup_elem() and sees:
130 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
131 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
133 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
134 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
135 * and were initialized prior to this call.
136 * If it's ok, then verifier allows this BPF_CALL insn and looks at
137 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
138 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
139 * returns either pointer to map value or NULL.
141 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
142 * insn, the register holding that pointer in the true branch changes state to
143 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
144 * branch. See check_cond_jmp_op().
146 * After the call R0 is set to return type of the function and registers R1-R5
147 * are set to NOT_INIT to indicate that they are no longer readable.
149 * The following reference types represent a potential reference to a kernel
150 * resource which, after first being allocated, must be checked and freed by
152 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
154 * When the verifier sees a helper call return a reference type, it allocates a
155 * pointer id for the reference and stores it in the current function state.
156 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
157 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
158 * passes through a NULL-check conditional. For the branch wherein the state is
159 * changed to CONST_IMM, the verifier releases the reference.
161 * For each helper function that allocates a reference, such as
162 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
163 * bpf_sk_release(). When a reference type passes into the release function,
164 * the verifier also releases the reference. If any unchecked or unreleased
165 * reference remains at the end of the program, the verifier rejects it.
168 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
169 struct bpf_verifier_stack_elem {
170 /* verifer state is 'st'
171 * before processing instruction 'insn_idx'
172 * and after processing instruction 'prev_insn_idx'
174 struct bpf_verifier_state st;
177 struct bpf_verifier_stack_elem *next;
178 /* length of verifier log at the time this state was pushed on stack */
182 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
183 #define BPF_COMPLEXITY_LIMIT_STATES 64
185 #define BPF_MAP_KEY_POISON (1ULL << 63)
186 #define BPF_MAP_KEY_SEEN (1ULL << 62)
188 #define BPF_MAP_PTR_UNPRIV 1UL
189 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
190 POISON_POINTER_DELTA))
191 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
193 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx);
194 static int release_reference(struct bpf_verifier_env *env, int ref_obj_id);
195 static void invalidate_non_owning_refs(struct bpf_verifier_env *env);
196 static bool in_rbtree_lock_required_cb(struct bpf_verifier_env *env);
197 static int ref_set_non_owning(struct bpf_verifier_env *env,
198 struct bpf_reg_state *reg);
199 static void specialize_kfunc(struct bpf_verifier_env *env,
200 u32 func_id, u16 offset, unsigned long *addr);
201 static bool is_trusted_reg(const struct bpf_reg_state *reg);
203 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
205 return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON;
208 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
210 return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV;
213 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
214 const struct bpf_map *map, bool unpriv)
216 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
217 unpriv |= bpf_map_ptr_unpriv(aux);
218 aux->map_ptr_state = (unsigned long)map |
219 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
222 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
224 return aux->map_key_state & BPF_MAP_KEY_POISON;
227 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
229 return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
232 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
234 return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
237 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
239 bool poisoned = bpf_map_key_poisoned(aux);
241 aux->map_key_state = state | BPF_MAP_KEY_SEEN |
242 (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
245 static bool bpf_helper_call(const struct bpf_insn *insn)
247 return insn->code == (BPF_JMP | BPF_CALL) &&
251 static bool bpf_pseudo_call(const struct bpf_insn *insn)
253 return insn->code == (BPF_JMP | BPF_CALL) &&
254 insn->src_reg == BPF_PSEUDO_CALL;
257 static bool bpf_pseudo_kfunc_call(const struct bpf_insn *insn)
259 return insn->code == (BPF_JMP | BPF_CALL) &&
260 insn->src_reg == BPF_PSEUDO_KFUNC_CALL;
263 struct bpf_call_arg_meta {
264 struct bpf_map *map_ptr;
281 struct btf_field *kptr_field;
284 struct bpf_kfunc_call_arg_meta {
289 const struct btf_type *func_proto;
290 const char *func_name;
303 /* arg_{btf,btf_id,owning_ref} are used by kfunc-specific handling,
304 * generally to pass info about user-defined local kptr types to later
307 * Record the local kptr type to be drop'd
308 * bpf_refcount_acquire (via KF_ARG_PTR_TO_REFCOUNTED_KPTR arg type)
309 * Record the local kptr type to be refcount_incr'd and use
310 * arg_owning_ref to determine whether refcount_acquire should be
318 struct btf_field *field;
321 struct btf_field *field;
324 enum bpf_dynptr_type type;
327 } initialized_dynptr;
335 struct btf *btf_vmlinux;
337 static DEFINE_MUTEX(bpf_verifier_lock);
339 static const struct bpf_line_info *
340 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
342 const struct bpf_line_info *linfo;
343 const struct bpf_prog *prog;
347 nr_linfo = prog->aux->nr_linfo;
349 if (!nr_linfo || insn_off >= prog->len)
352 linfo = prog->aux->linfo;
353 for (i = 1; i < nr_linfo; i++)
354 if (insn_off < linfo[i].insn_off)
357 return &linfo[i - 1];
360 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
362 struct bpf_verifier_env *env = private_data;
365 if (!bpf_verifier_log_needed(&env->log))
369 bpf_verifier_vlog(&env->log, fmt, args);
373 static const char *ltrim(const char *s)
381 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
383 const char *prefix_fmt, ...)
385 const struct bpf_line_info *linfo;
387 if (!bpf_verifier_log_needed(&env->log))
390 linfo = find_linfo(env, insn_off);
391 if (!linfo || linfo == env->prev_linfo)
397 va_start(args, prefix_fmt);
398 bpf_verifier_vlog(&env->log, prefix_fmt, args);
403 ltrim(btf_name_by_offset(env->prog->aux->btf,
406 env->prev_linfo = linfo;
409 static void verbose_invalid_scalar(struct bpf_verifier_env *env,
410 struct bpf_reg_state *reg,
411 struct tnum *range, const char *ctx,
412 const char *reg_name)
416 verbose(env, "At %s the register %s ", ctx, reg_name);
417 if (!tnum_is_unknown(reg->var_off)) {
418 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
419 verbose(env, "has value %s", tn_buf);
421 verbose(env, "has unknown scalar value");
423 tnum_strn(tn_buf, sizeof(tn_buf), *range);
424 verbose(env, " should have been in %s\n", tn_buf);
427 static bool type_is_pkt_pointer(enum bpf_reg_type type)
429 type = base_type(type);
430 return type == PTR_TO_PACKET ||
431 type == PTR_TO_PACKET_META;
434 static bool type_is_sk_pointer(enum bpf_reg_type type)
436 return type == PTR_TO_SOCKET ||
437 type == PTR_TO_SOCK_COMMON ||
438 type == PTR_TO_TCP_SOCK ||
439 type == PTR_TO_XDP_SOCK;
442 static bool type_may_be_null(u32 type)
444 return type & PTR_MAYBE_NULL;
447 static bool reg_not_null(const struct bpf_reg_state *reg)
449 enum bpf_reg_type type;
452 if (type_may_be_null(type))
455 type = base_type(type);
456 return type == PTR_TO_SOCKET ||
457 type == PTR_TO_TCP_SOCK ||
458 type == PTR_TO_MAP_VALUE ||
459 type == PTR_TO_MAP_KEY ||
460 type == PTR_TO_SOCK_COMMON ||
461 (type == PTR_TO_BTF_ID && is_trusted_reg(reg)) ||
465 static bool type_is_ptr_alloc_obj(u32 type)
467 return base_type(type) == PTR_TO_BTF_ID && type_flag(type) & MEM_ALLOC;
470 static bool type_is_non_owning_ref(u32 type)
472 return type_is_ptr_alloc_obj(type) && type_flag(type) & NON_OWN_REF;
475 static struct btf_record *reg_btf_record(const struct bpf_reg_state *reg)
477 struct btf_record *rec = NULL;
478 struct btf_struct_meta *meta;
480 if (reg->type == PTR_TO_MAP_VALUE) {
481 rec = reg->map_ptr->record;
482 } else if (type_is_ptr_alloc_obj(reg->type)) {
483 meta = btf_find_struct_meta(reg->btf, reg->btf_id);
490 static bool subprog_is_global(const struct bpf_verifier_env *env, int subprog)
492 struct bpf_func_info_aux *aux = env->prog->aux->func_info_aux;
494 return aux && aux[subprog].linkage == BTF_FUNC_GLOBAL;
497 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
499 return btf_record_has_field(reg_btf_record(reg), BPF_SPIN_LOCK);
502 static bool type_is_rdonly_mem(u32 type)
504 return type & MEM_RDONLY;
507 static bool is_acquire_function(enum bpf_func_id func_id,
508 const struct bpf_map *map)
510 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
512 if (func_id == BPF_FUNC_sk_lookup_tcp ||
513 func_id == BPF_FUNC_sk_lookup_udp ||
514 func_id == BPF_FUNC_skc_lookup_tcp ||
515 func_id == BPF_FUNC_ringbuf_reserve ||
516 func_id == BPF_FUNC_kptr_xchg)
519 if (func_id == BPF_FUNC_map_lookup_elem &&
520 (map_type == BPF_MAP_TYPE_SOCKMAP ||
521 map_type == BPF_MAP_TYPE_SOCKHASH))
527 static bool is_ptr_cast_function(enum bpf_func_id func_id)
529 return func_id == BPF_FUNC_tcp_sock ||
530 func_id == BPF_FUNC_sk_fullsock ||
531 func_id == BPF_FUNC_skc_to_tcp_sock ||
532 func_id == BPF_FUNC_skc_to_tcp6_sock ||
533 func_id == BPF_FUNC_skc_to_udp6_sock ||
534 func_id == BPF_FUNC_skc_to_mptcp_sock ||
535 func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
536 func_id == BPF_FUNC_skc_to_tcp_request_sock;
539 static bool is_dynptr_ref_function(enum bpf_func_id func_id)
541 return func_id == BPF_FUNC_dynptr_data;
544 static bool is_callback_calling_kfunc(u32 btf_id);
546 static bool is_callback_calling_function(enum bpf_func_id func_id)
548 return func_id == BPF_FUNC_for_each_map_elem ||
549 func_id == BPF_FUNC_timer_set_callback ||
550 func_id == BPF_FUNC_find_vma ||
551 func_id == BPF_FUNC_loop ||
552 func_id == BPF_FUNC_user_ringbuf_drain;
555 static bool is_async_callback_calling_function(enum bpf_func_id func_id)
557 return func_id == BPF_FUNC_timer_set_callback;
560 static bool is_storage_get_function(enum bpf_func_id func_id)
562 return func_id == BPF_FUNC_sk_storage_get ||
563 func_id == BPF_FUNC_inode_storage_get ||
564 func_id == BPF_FUNC_task_storage_get ||
565 func_id == BPF_FUNC_cgrp_storage_get;
568 static bool helper_multiple_ref_obj_use(enum bpf_func_id func_id,
569 const struct bpf_map *map)
571 int ref_obj_uses = 0;
573 if (is_ptr_cast_function(func_id))
575 if (is_acquire_function(func_id, map))
577 if (is_dynptr_ref_function(func_id))
580 return ref_obj_uses > 1;
583 static bool is_cmpxchg_insn(const struct bpf_insn *insn)
585 return BPF_CLASS(insn->code) == BPF_STX &&
586 BPF_MODE(insn->code) == BPF_ATOMIC &&
587 insn->imm == BPF_CMPXCHG;
590 /* string representation of 'enum bpf_reg_type'
592 * Note that reg_type_str() can not appear more than once in a single verbose()
595 static const char *reg_type_str(struct bpf_verifier_env *env,
596 enum bpf_reg_type type)
598 char postfix[16] = {0}, prefix[64] = {0};
599 static const char * const str[] = {
601 [SCALAR_VALUE] = "scalar",
602 [PTR_TO_CTX] = "ctx",
603 [CONST_PTR_TO_MAP] = "map_ptr",
604 [PTR_TO_MAP_VALUE] = "map_value",
605 [PTR_TO_STACK] = "fp",
606 [PTR_TO_PACKET] = "pkt",
607 [PTR_TO_PACKET_META] = "pkt_meta",
608 [PTR_TO_PACKET_END] = "pkt_end",
609 [PTR_TO_FLOW_KEYS] = "flow_keys",
610 [PTR_TO_SOCKET] = "sock",
611 [PTR_TO_SOCK_COMMON] = "sock_common",
612 [PTR_TO_TCP_SOCK] = "tcp_sock",
613 [PTR_TO_TP_BUFFER] = "tp_buffer",
614 [PTR_TO_XDP_SOCK] = "xdp_sock",
615 [PTR_TO_BTF_ID] = "ptr_",
616 [PTR_TO_MEM] = "mem",
617 [PTR_TO_BUF] = "buf",
618 [PTR_TO_FUNC] = "func",
619 [PTR_TO_MAP_KEY] = "map_key",
620 [CONST_PTR_TO_DYNPTR] = "dynptr_ptr",
623 if (type & PTR_MAYBE_NULL) {
624 if (base_type(type) == PTR_TO_BTF_ID)
625 strncpy(postfix, "or_null_", 16);
627 strncpy(postfix, "_or_null", 16);
630 snprintf(prefix, sizeof(prefix), "%s%s%s%s%s%s%s",
631 type & MEM_RDONLY ? "rdonly_" : "",
632 type & MEM_RINGBUF ? "ringbuf_" : "",
633 type & MEM_USER ? "user_" : "",
634 type & MEM_PERCPU ? "percpu_" : "",
635 type & MEM_RCU ? "rcu_" : "",
636 type & PTR_UNTRUSTED ? "untrusted_" : "",
637 type & PTR_TRUSTED ? "trusted_" : ""
640 snprintf(env->tmp_str_buf, TMP_STR_BUF_LEN, "%s%s%s",
641 prefix, str[base_type(type)], postfix);
642 return env->tmp_str_buf;
645 static char slot_type_char[] = {
646 [STACK_INVALID] = '?',
650 [STACK_DYNPTR] = 'd',
654 static void print_liveness(struct bpf_verifier_env *env,
655 enum bpf_reg_liveness live)
657 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
659 if (live & REG_LIVE_READ)
661 if (live & REG_LIVE_WRITTEN)
663 if (live & REG_LIVE_DONE)
667 static int __get_spi(s32 off)
669 return (-off - 1) / BPF_REG_SIZE;
672 static struct bpf_func_state *func(struct bpf_verifier_env *env,
673 const struct bpf_reg_state *reg)
675 struct bpf_verifier_state *cur = env->cur_state;
677 return cur->frame[reg->frameno];
680 static bool is_spi_bounds_valid(struct bpf_func_state *state, int spi, int nr_slots)
682 int allocated_slots = state->allocated_stack / BPF_REG_SIZE;
684 /* We need to check that slots between [spi - nr_slots + 1, spi] are
685 * within [0, allocated_stack).
687 * Please note that the spi grows downwards. For example, a dynptr
688 * takes the size of two stack slots; the first slot will be at
689 * spi and the second slot will be at spi - 1.
691 return spi - nr_slots + 1 >= 0 && spi < allocated_slots;
694 static int stack_slot_obj_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
695 const char *obj_kind, int nr_slots)
699 if (!tnum_is_const(reg->var_off)) {
700 verbose(env, "%s has to be at a constant offset\n", obj_kind);
704 off = reg->off + reg->var_off.value;
705 if (off % BPF_REG_SIZE) {
706 verbose(env, "cannot pass in %s at an offset=%d\n", obj_kind, off);
710 spi = __get_spi(off);
711 if (spi + 1 < nr_slots) {
712 verbose(env, "cannot pass in %s at an offset=%d\n", obj_kind, off);
716 if (!is_spi_bounds_valid(func(env, reg), spi, nr_slots))
721 static int dynptr_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
723 return stack_slot_obj_get_spi(env, reg, "dynptr", BPF_DYNPTR_NR_SLOTS);
726 static int iter_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg, int nr_slots)
728 return stack_slot_obj_get_spi(env, reg, "iter", nr_slots);
731 static const char *btf_type_name(const struct btf *btf, u32 id)
733 return btf_name_by_offset(btf, btf_type_by_id(btf, id)->name_off);
736 static const char *dynptr_type_str(enum bpf_dynptr_type type)
739 case BPF_DYNPTR_TYPE_LOCAL:
741 case BPF_DYNPTR_TYPE_RINGBUF:
743 case BPF_DYNPTR_TYPE_SKB:
745 case BPF_DYNPTR_TYPE_XDP:
747 case BPF_DYNPTR_TYPE_INVALID:
750 WARN_ONCE(1, "unknown dynptr type %d\n", type);
755 static const char *iter_type_str(const struct btf *btf, u32 btf_id)
757 if (!btf || btf_id == 0)
760 /* we already validated that type is valid and has conforming name */
761 return btf_type_name(btf, btf_id) + sizeof(ITER_PREFIX) - 1;
764 static const char *iter_state_str(enum bpf_iter_state state)
767 case BPF_ITER_STATE_ACTIVE:
769 case BPF_ITER_STATE_DRAINED:
771 case BPF_ITER_STATE_INVALID:
774 WARN_ONCE(1, "unknown iter state %d\n", state);
779 static void mark_reg_scratched(struct bpf_verifier_env *env, u32 regno)
781 env->scratched_regs |= 1U << regno;
784 static void mark_stack_slot_scratched(struct bpf_verifier_env *env, u32 spi)
786 env->scratched_stack_slots |= 1ULL << spi;
789 static bool reg_scratched(const struct bpf_verifier_env *env, u32 regno)
791 return (env->scratched_regs >> regno) & 1;
794 static bool stack_slot_scratched(const struct bpf_verifier_env *env, u64 regno)
796 return (env->scratched_stack_slots >> regno) & 1;
799 static bool verifier_state_scratched(const struct bpf_verifier_env *env)
801 return env->scratched_regs || env->scratched_stack_slots;
804 static void mark_verifier_state_clean(struct bpf_verifier_env *env)
806 env->scratched_regs = 0U;
807 env->scratched_stack_slots = 0ULL;
810 /* Used for printing the entire verifier state. */
811 static void mark_verifier_state_scratched(struct bpf_verifier_env *env)
813 env->scratched_regs = ~0U;
814 env->scratched_stack_slots = ~0ULL;
817 static enum bpf_dynptr_type arg_to_dynptr_type(enum bpf_arg_type arg_type)
819 switch (arg_type & DYNPTR_TYPE_FLAG_MASK) {
820 case DYNPTR_TYPE_LOCAL:
821 return BPF_DYNPTR_TYPE_LOCAL;
822 case DYNPTR_TYPE_RINGBUF:
823 return BPF_DYNPTR_TYPE_RINGBUF;
824 case DYNPTR_TYPE_SKB:
825 return BPF_DYNPTR_TYPE_SKB;
826 case DYNPTR_TYPE_XDP:
827 return BPF_DYNPTR_TYPE_XDP;
829 return BPF_DYNPTR_TYPE_INVALID;
833 static enum bpf_type_flag get_dynptr_type_flag(enum bpf_dynptr_type type)
836 case BPF_DYNPTR_TYPE_LOCAL:
837 return DYNPTR_TYPE_LOCAL;
838 case BPF_DYNPTR_TYPE_RINGBUF:
839 return DYNPTR_TYPE_RINGBUF;
840 case BPF_DYNPTR_TYPE_SKB:
841 return DYNPTR_TYPE_SKB;
842 case BPF_DYNPTR_TYPE_XDP:
843 return DYNPTR_TYPE_XDP;
849 static bool dynptr_type_refcounted(enum bpf_dynptr_type type)
851 return type == BPF_DYNPTR_TYPE_RINGBUF;
854 static void __mark_dynptr_reg(struct bpf_reg_state *reg,
855 enum bpf_dynptr_type type,
856 bool first_slot, int dynptr_id);
858 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
859 struct bpf_reg_state *reg);
861 static void mark_dynptr_stack_regs(struct bpf_verifier_env *env,
862 struct bpf_reg_state *sreg1,
863 struct bpf_reg_state *sreg2,
864 enum bpf_dynptr_type type)
866 int id = ++env->id_gen;
868 __mark_dynptr_reg(sreg1, type, true, id);
869 __mark_dynptr_reg(sreg2, type, false, id);
872 static void mark_dynptr_cb_reg(struct bpf_verifier_env *env,
873 struct bpf_reg_state *reg,
874 enum bpf_dynptr_type type)
876 __mark_dynptr_reg(reg, type, true, ++env->id_gen);
879 static int destroy_if_dynptr_stack_slot(struct bpf_verifier_env *env,
880 struct bpf_func_state *state, int spi);
882 static int mark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
883 enum bpf_arg_type arg_type, int insn_idx, int clone_ref_obj_id)
885 struct bpf_func_state *state = func(env, reg);
886 enum bpf_dynptr_type type;
889 spi = dynptr_get_spi(env, reg);
893 /* We cannot assume both spi and spi - 1 belong to the same dynptr,
894 * hence we need to call destroy_if_dynptr_stack_slot twice for both,
895 * to ensure that for the following example:
898 * So marking spi = 2 should lead to destruction of both d1 and d2. In
899 * case they do belong to same dynptr, second call won't see slot_type
900 * as STACK_DYNPTR and will simply skip destruction.
902 err = destroy_if_dynptr_stack_slot(env, state, spi);
905 err = destroy_if_dynptr_stack_slot(env, state, spi - 1);
909 for (i = 0; i < BPF_REG_SIZE; i++) {
910 state->stack[spi].slot_type[i] = STACK_DYNPTR;
911 state->stack[spi - 1].slot_type[i] = STACK_DYNPTR;
914 type = arg_to_dynptr_type(arg_type);
915 if (type == BPF_DYNPTR_TYPE_INVALID)
918 mark_dynptr_stack_regs(env, &state->stack[spi].spilled_ptr,
919 &state->stack[spi - 1].spilled_ptr, type);
921 if (dynptr_type_refcounted(type)) {
922 /* The id is used to track proper releasing */
925 if (clone_ref_obj_id)
926 id = clone_ref_obj_id;
928 id = acquire_reference_state(env, insn_idx);
933 state->stack[spi].spilled_ptr.ref_obj_id = id;
934 state->stack[spi - 1].spilled_ptr.ref_obj_id = id;
937 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
938 state->stack[spi - 1].spilled_ptr.live |= REG_LIVE_WRITTEN;
943 static void invalidate_dynptr(struct bpf_verifier_env *env, struct bpf_func_state *state, int spi)
947 for (i = 0; i < BPF_REG_SIZE; i++) {
948 state->stack[spi].slot_type[i] = STACK_INVALID;
949 state->stack[spi - 1].slot_type[i] = STACK_INVALID;
952 __mark_reg_not_init(env, &state->stack[spi].spilled_ptr);
953 __mark_reg_not_init(env, &state->stack[spi - 1].spilled_ptr);
955 /* Why do we need to set REG_LIVE_WRITTEN for STACK_INVALID slot?
957 * While we don't allow reading STACK_INVALID, it is still possible to
958 * do <8 byte writes marking some but not all slots as STACK_MISC. Then,
959 * helpers or insns can do partial read of that part without failing,
960 * but check_stack_range_initialized, check_stack_read_var_off, and
961 * check_stack_read_fixed_off will do mark_reg_read for all 8-bytes of
962 * the slot conservatively. Hence we need to prevent those liveness
965 * This was not a problem before because STACK_INVALID is only set by
966 * default (where the default reg state has its reg->parent as NULL), or
967 * in clean_live_states after REG_LIVE_DONE (at which point
968 * mark_reg_read won't walk reg->parent chain), but not randomly during
969 * verifier state exploration (like we did above). Hence, for our case
970 * parentage chain will still be live (i.e. reg->parent may be
971 * non-NULL), while earlier reg->parent was NULL, so we need
972 * REG_LIVE_WRITTEN to screen off read marker propagation when it is
973 * done later on reads or by mark_dynptr_read as well to unnecessary
974 * mark registers in verifier state.
976 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
977 state->stack[spi - 1].spilled_ptr.live |= REG_LIVE_WRITTEN;
980 static int unmark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
982 struct bpf_func_state *state = func(env, reg);
983 int spi, ref_obj_id, i;
985 spi = dynptr_get_spi(env, reg);
989 if (!dynptr_type_refcounted(state->stack[spi].spilled_ptr.dynptr.type)) {
990 invalidate_dynptr(env, state, spi);
994 ref_obj_id = state->stack[spi].spilled_ptr.ref_obj_id;
996 /* If the dynptr has a ref_obj_id, then we need to invalidate
999 * 1) Any dynptrs with a matching ref_obj_id (clones)
1000 * 2) Any slices derived from this dynptr.
1003 /* Invalidate any slices associated with this dynptr */
1004 WARN_ON_ONCE(release_reference(env, ref_obj_id));
1006 /* Invalidate any dynptr clones */
1007 for (i = 1; i < state->allocated_stack / BPF_REG_SIZE; i++) {
1008 if (state->stack[i].spilled_ptr.ref_obj_id != ref_obj_id)
1011 /* it should always be the case that if the ref obj id
1012 * matches then the stack slot also belongs to a
1015 if (state->stack[i].slot_type[0] != STACK_DYNPTR) {
1016 verbose(env, "verifier internal error: misconfigured ref_obj_id\n");
1019 if (state->stack[i].spilled_ptr.dynptr.first_slot)
1020 invalidate_dynptr(env, state, i);
1026 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1027 struct bpf_reg_state *reg);
1029 static void mark_reg_invalid(const struct bpf_verifier_env *env, struct bpf_reg_state *reg)
1031 if (!env->allow_ptr_leaks)
1032 __mark_reg_not_init(env, reg);
1034 __mark_reg_unknown(env, reg);
1037 static int destroy_if_dynptr_stack_slot(struct bpf_verifier_env *env,
1038 struct bpf_func_state *state, int spi)
1040 struct bpf_func_state *fstate;
1041 struct bpf_reg_state *dreg;
1044 /* We always ensure that STACK_DYNPTR is never set partially,
1045 * hence just checking for slot_type[0] is enough. This is
1046 * different for STACK_SPILL, where it may be only set for
1047 * 1 byte, so code has to use is_spilled_reg.
1049 if (state->stack[spi].slot_type[0] != STACK_DYNPTR)
1052 /* Reposition spi to first slot */
1053 if (!state->stack[spi].spilled_ptr.dynptr.first_slot)
1056 if (dynptr_type_refcounted(state->stack[spi].spilled_ptr.dynptr.type)) {
1057 verbose(env, "cannot overwrite referenced dynptr\n");
1061 mark_stack_slot_scratched(env, spi);
1062 mark_stack_slot_scratched(env, spi - 1);
1064 /* Writing partially to one dynptr stack slot destroys both. */
1065 for (i = 0; i < BPF_REG_SIZE; i++) {
1066 state->stack[spi].slot_type[i] = STACK_INVALID;
1067 state->stack[spi - 1].slot_type[i] = STACK_INVALID;
1070 dynptr_id = state->stack[spi].spilled_ptr.id;
1071 /* Invalidate any slices associated with this dynptr */
1072 bpf_for_each_reg_in_vstate(env->cur_state, fstate, dreg, ({
1073 /* Dynptr slices are only PTR_TO_MEM_OR_NULL and PTR_TO_MEM */
1074 if (dreg->type != (PTR_TO_MEM | PTR_MAYBE_NULL) && dreg->type != PTR_TO_MEM)
1076 if (dreg->dynptr_id == dynptr_id)
1077 mark_reg_invalid(env, dreg);
1080 /* Do not release reference state, we are destroying dynptr on stack,
1081 * not using some helper to release it. Just reset register.
1083 __mark_reg_not_init(env, &state->stack[spi].spilled_ptr);
1084 __mark_reg_not_init(env, &state->stack[spi - 1].spilled_ptr);
1086 /* Same reason as unmark_stack_slots_dynptr above */
1087 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1088 state->stack[spi - 1].spilled_ptr.live |= REG_LIVE_WRITTEN;
1093 static bool is_dynptr_reg_valid_uninit(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
1097 if (reg->type == CONST_PTR_TO_DYNPTR)
1100 spi = dynptr_get_spi(env, reg);
1102 /* -ERANGE (i.e. spi not falling into allocated stack slots) isn't an
1103 * error because this just means the stack state hasn't been updated yet.
1104 * We will do check_mem_access to check and update stack bounds later.
1106 if (spi < 0 && spi != -ERANGE)
1109 /* We don't need to check if the stack slots are marked by previous
1110 * dynptr initializations because we allow overwriting existing unreferenced
1111 * STACK_DYNPTR slots, see mark_stack_slots_dynptr which calls
1112 * destroy_if_dynptr_stack_slot to ensure dynptr objects at the slots we are
1113 * touching are completely destructed before we reinitialize them for a new
1114 * one. For referenced ones, destroy_if_dynptr_stack_slot returns an error early
1115 * instead of delaying it until the end where the user will get "Unreleased
1121 static bool is_dynptr_reg_valid_init(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
1123 struct bpf_func_state *state = func(env, reg);
1126 /* This already represents first slot of initialized bpf_dynptr.
1128 * CONST_PTR_TO_DYNPTR already has fixed and var_off as 0 due to
1129 * check_func_arg_reg_off's logic, so we don't need to check its
1130 * offset and alignment.
1132 if (reg->type == CONST_PTR_TO_DYNPTR)
1135 spi = dynptr_get_spi(env, reg);
1138 if (!state->stack[spi].spilled_ptr.dynptr.first_slot)
1141 for (i = 0; i < BPF_REG_SIZE; i++) {
1142 if (state->stack[spi].slot_type[i] != STACK_DYNPTR ||
1143 state->stack[spi - 1].slot_type[i] != STACK_DYNPTR)
1150 static bool is_dynptr_type_expected(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
1151 enum bpf_arg_type arg_type)
1153 struct bpf_func_state *state = func(env, reg);
1154 enum bpf_dynptr_type dynptr_type;
1157 /* ARG_PTR_TO_DYNPTR takes any type of dynptr */
1158 if (arg_type == ARG_PTR_TO_DYNPTR)
1161 dynptr_type = arg_to_dynptr_type(arg_type);
1162 if (reg->type == CONST_PTR_TO_DYNPTR) {
1163 return reg->dynptr.type == dynptr_type;
1165 spi = dynptr_get_spi(env, reg);
1168 return state->stack[spi].spilled_ptr.dynptr.type == dynptr_type;
1172 static void __mark_reg_known_zero(struct bpf_reg_state *reg);
1174 static int mark_stack_slots_iter(struct bpf_verifier_env *env,
1175 struct bpf_reg_state *reg, int insn_idx,
1176 struct btf *btf, u32 btf_id, int nr_slots)
1178 struct bpf_func_state *state = func(env, reg);
1181 spi = iter_get_spi(env, reg, nr_slots);
1185 id = acquire_reference_state(env, insn_idx);
1189 for (i = 0; i < nr_slots; i++) {
1190 struct bpf_stack_state *slot = &state->stack[spi - i];
1191 struct bpf_reg_state *st = &slot->spilled_ptr;
1193 __mark_reg_known_zero(st);
1194 st->type = PTR_TO_STACK; /* we don't have dedicated reg type */
1195 st->live |= REG_LIVE_WRITTEN;
1196 st->ref_obj_id = i == 0 ? id : 0;
1198 st->iter.btf_id = btf_id;
1199 st->iter.state = BPF_ITER_STATE_ACTIVE;
1202 for (j = 0; j < BPF_REG_SIZE; j++)
1203 slot->slot_type[j] = STACK_ITER;
1205 mark_stack_slot_scratched(env, spi - i);
1211 static int unmark_stack_slots_iter(struct bpf_verifier_env *env,
1212 struct bpf_reg_state *reg, int nr_slots)
1214 struct bpf_func_state *state = func(env, reg);
1217 spi = iter_get_spi(env, reg, nr_slots);
1221 for (i = 0; i < nr_slots; i++) {
1222 struct bpf_stack_state *slot = &state->stack[spi - i];
1223 struct bpf_reg_state *st = &slot->spilled_ptr;
1226 WARN_ON_ONCE(release_reference(env, st->ref_obj_id));
1228 __mark_reg_not_init(env, st);
1230 /* see unmark_stack_slots_dynptr() for why we need to set REG_LIVE_WRITTEN */
1231 st->live |= REG_LIVE_WRITTEN;
1233 for (j = 0; j < BPF_REG_SIZE; j++)
1234 slot->slot_type[j] = STACK_INVALID;
1236 mark_stack_slot_scratched(env, spi - i);
1242 static bool is_iter_reg_valid_uninit(struct bpf_verifier_env *env,
1243 struct bpf_reg_state *reg, int nr_slots)
1245 struct bpf_func_state *state = func(env, reg);
1248 /* For -ERANGE (i.e. spi not falling into allocated stack slots), we
1249 * will do check_mem_access to check and update stack bounds later, so
1250 * return true for that case.
1252 spi = iter_get_spi(env, reg, nr_slots);
1258 for (i = 0; i < nr_slots; i++) {
1259 struct bpf_stack_state *slot = &state->stack[spi - i];
1261 for (j = 0; j < BPF_REG_SIZE; j++)
1262 if (slot->slot_type[j] == STACK_ITER)
1269 static bool is_iter_reg_valid_init(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
1270 struct btf *btf, u32 btf_id, int nr_slots)
1272 struct bpf_func_state *state = func(env, reg);
1275 spi = iter_get_spi(env, reg, nr_slots);
1279 for (i = 0; i < nr_slots; i++) {
1280 struct bpf_stack_state *slot = &state->stack[spi - i];
1281 struct bpf_reg_state *st = &slot->spilled_ptr;
1283 /* only main (first) slot has ref_obj_id set */
1284 if (i == 0 && !st->ref_obj_id)
1286 if (i != 0 && st->ref_obj_id)
1288 if (st->iter.btf != btf || st->iter.btf_id != btf_id)
1291 for (j = 0; j < BPF_REG_SIZE; j++)
1292 if (slot->slot_type[j] != STACK_ITER)
1299 /* Check if given stack slot is "special":
1300 * - spilled register state (STACK_SPILL);
1301 * - dynptr state (STACK_DYNPTR);
1302 * - iter state (STACK_ITER).
1304 static bool is_stack_slot_special(const struct bpf_stack_state *stack)
1306 enum bpf_stack_slot_type type = stack->slot_type[BPF_REG_SIZE - 1];
1318 WARN_ONCE(1, "unknown stack slot type %d\n", type);
1323 /* The reg state of a pointer or a bounded scalar was saved when
1324 * it was spilled to the stack.
1326 static bool is_spilled_reg(const struct bpf_stack_state *stack)
1328 return stack->slot_type[BPF_REG_SIZE - 1] == STACK_SPILL;
1331 static bool is_spilled_scalar_reg(const struct bpf_stack_state *stack)
1333 return stack->slot_type[BPF_REG_SIZE - 1] == STACK_SPILL &&
1334 stack->spilled_ptr.type == SCALAR_VALUE;
1337 static void scrub_spilled_slot(u8 *stype)
1339 if (*stype != STACK_INVALID)
1340 *stype = STACK_MISC;
1343 static void print_verifier_state(struct bpf_verifier_env *env,
1344 const struct bpf_func_state *state,
1347 const struct bpf_reg_state *reg;
1348 enum bpf_reg_type t;
1352 verbose(env, " frame%d:", state->frameno);
1353 for (i = 0; i < MAX_BPF_REG; i++) {
1354 reg = &state->regs[i];
1358 if (!print_all && !reg_scratched(env, i))
1360 verbose(env, " R%d", i);
1361 print_liveness(env, reg->live);
1363 if (t == SCALAR_VALUE && reg->precise)
1365 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
1366 tnum_is_const(reg->var_off)) {
1367 /* reg->off should be 0 for SCALAR_VALUE */
1368 verbose(env, "%s", t == SCALAR_VALUE ? "" : reg_type_str(env, t));
1369 verbose(env, "%lld", reg->var_off.value + reg->off);
1371 const char *sep = "";
1373 verbose(env, "%s", reg_type_str(env, t));
1374 if (base_type(t) == PTR_TO_BTF_ID)
1375 verbose(env, "%s", btf_type_name(reg->btf, reg->btf_id));
1378 * _a stands for append, was shortened to avoid multiline statements below.
1379 * This macro is used to output a comma separated list of attributes.
1381 #define verbose_a(fmt, ...) ({ verbose(env, "%s" fmt, sep, __VA_ARGS__); sep = ","; })
1384 verbose_a("id=%d", reg->id);
1385 if (reg->ref_obj_id)
1386 verbose_a("ref_obj_id=%d", reg->ref_obj_id);
1387 if (type_is_non_owning_ref(reg->type))
1388 verbose_a("%s", "non_own_ref");
1389 if (t != SCALAR_VALUE)
1390 verbose_a("off=%d", reg->off);
1391 if (type_is_pkt_pointer(t))
1392 verbose_a("r=%d", reg->range);
1393 else if (base_type(t) == CONST_PTR_TO_MAP ||
1394 base_type(t) == PTR_TO_MAP_KEY ||
1395 base_type(t) == PTR_TO_MAP_VALUE)
1396 verbose_a("ks=%d,vs=%d",
1397 reg->map_ptr->key_size,
1398 reg->map_ptr->value_size);
1399 if (tnum_is_const(reg->var_off)) {
1400 /* Typically an immediate SCALAR_VALUE, but
1401 * could be a pointer whose offset is too big
1404 verbose_a("imm=%llx", reg->var_off.value);
1406 if (reg->smin_value != reg->umin_value &&
1407 reg->smin_value != S64_MIN)
1408 verbose_a("smin=%lld", (long long)reg->smin_value);
1409 if (reg->smax_value != reg->umax_value &&
1410 reg->smax_value != S64_MAX)
1411 verbose_a("smax=%lld", (long long)reg->smax_value);
1412 if (reg->umin_value != 0)
1413 verbose_a("umin=%llu", (unsigned long long)reg->umin_value);
1414 if (reg->umax_value != U64_MAX)
1415 verbose_a("umax=%llu", (unsigned long long)reg->umax_value);
1416 if (!tnum_is_unknown(reg->var_off)) {
1419 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1420 verbose_a("var_off=%s", tn_buf);
1422 if (reg->s32_min_value != reg->smin_value &&
1423 reg->s32_min_value != S32_MIN)
1424 verbose_a("s32_min=%d", (int)(reg->s32_min_value));
1425 if (reg->s32_max_value != reg->smax_value &&
1426 reg->s32_max_value != S32_MAX)
1427 verbose_a("s32_max=%d", (int)(reg->s32_max_value));
1428 if (reg->u32_min_value != reg->umin_value &&
1429 reg->u32_min_value != U32_MIN)
1430 verbose_a("u32_min=%d", (int)(reg->u32_min_value));
1431 if (reg->u32_max_value != reg->umax_value &&
1432 reg->u32_max_value != U32_MAX)
1433 verbose_a("u32_max=%d", (int)(reg->u32_max_value));
1440 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
1441 char types_buf[BPF_REG_SIZE + 1];
1445 for (j = 0; j < BPF_REG_SIZE; j++) {
1446 if (state->stack[i].slot_type[j] != STACK_INVALID)
1448 types_buf[j] = slot_type_char[state->stack[i].slot_type[j]];
1450 types_buf[BPF_REG_SIZE] = 0;
1453 if (!print_all && !stack_slot_scratched(env, i))
1455 switch (state->stack[i].slot_type[BPF_REG_SIZE - 1]) {
1457 reg = &state->stack[i].spilled_ptr;
1460 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
1461 print_liveness(env, reg->live);
1462 verbose(env, "=%s", t == SCALAR_VALUE ? "" : reg_type_str(env, t));
1463 if (t == SCALAR_VALUE && reg->precise)
1465 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
1466 verbose(env, "%lld", reg->var_off.value + reg->off);
1469 i += BPF_DYNPTR_NR_SLOTS - 1;
1470 reg = &state->stack[i].spilled_ptr;
1472 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
1473 print_liveness(env, reg->live);
1474 verbose(env, "=dynptr_%s", dynptr_type_str(reg->dynptr.type));
1475 if (reg->ref_obj_id)
1476 verbose(env, "(ref_id=%d)", reg->ref_obj_id);
1479 /* only main slot has ref_obj_id set; skip others */
1480 reg = &state->stack[i].spilled_ptr;
1481 if (!reg->ref_obj_id)
1484 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
1485 print_liveness(env, reg->live);
1486 verbose(env, "=iter_%s(ref_id=%d,state=%s,depth=%u)",
1487 iter_type_str(reg->iter.btf, reg->iter.btf_id),
1488 reg->ref_obj_id, iter_state_str(reg->iter.state),
1494 reg = &state->stack[i].spilled_ptr;
1496 for (j = 0; j < BPF_REG_SIZE; j++)
1497 types_buf[j] = slot_type_char[state->stack[i].slot_type[j]];
1498 types_buf[BPF_REG_SIZE] = 0;
1500 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
1501 print_liveness(env, reg->live);
1502 verbose(env, "=%s", types_buf);
1506 if (state->acquired_refs && state->refs[0].id) {
1507 verbose(env, " refs=%d", state->refs[0].id);
1508 for (i = 1; i < state->acquired_refs; i++)
1509 if (state->refs[i].id)
1510 verbose(env, ",%d", state->refs[i].id);
1512 if (state->in_callback_fn)
1513 verbose(env, " cb");
1514 if (state->in_async_callback_fn)
1515 verbose(env, " async_cb");
1517 mark_verifier_state_clean(env);
1520 static inline u32 vlog_alignment(u32 pos)
1522 return round_up(max(pos + BPF_LOG_MIN_ALIGNMENT / 2, BPF_LOG_ALIGNMENT),
1523 BPF_LOG_MIN_ALIGNMENT) - pos - 1;
1526 static void print_insn_state(struct bpf_verifier_env *env,
1527 const struct bpf_func_state *state)
1529 if (env->prev_log_pos && env->prev_log_pos == env->log.end_pos) {
1530 /* remove new line character */
1531 bpf_vlog_reset(&env->log, env->prev_log_pos - 1);
1532 verbose(env, "%*c;", vlog_alignment(env->prev_insn_print_pos), ' ');
1534 verbose(env, "%d:", env->insn_idx);
1536 print_verifier_state(env, state, false);
1539 /* copy array src of length n * size bytes to dst. dst is reallocated if it's too
1540 * small to hold src. This is different from krealloc since we don't want to preserve
1541 * the contents of dst.
1543 * Leaves dst untouched if src is NULL or length is zero. Returns NULL if memory could
1546 static void *copy_array(void *dst, const void *src, size_t n, size_t size, gfp_t flags)
1552 if (ZERO_OR_NULL_PTR(src))
1555 if (unlikely(check_mul_overflow(n, size, &bytes)))
1558 alloc_bytes = max(ksize(orig), kmalloc_size_roundup(bytes));
1559 dst = krealloc(orig, alloc_bytes, flags);
1565 memcpy(dst, src, bytes);
1567 return dst ? dst : ZERO_SIZE_PTR;
1570 /* resize an array from old_n items to new_n items. the array is reallocated if it's too
1571 * small to hold new_n items. new items are zeroed out if the array grows.
1573 * Contrary to krealloc_array, does not free arr if new_n is zero.
1575 static void *realloc_array(void *arr, size_t old_n, size_t new_n, size_t size)
1580 if (!new_n || old_n == new_n)
1583 alloc_size = kmalloc_size_roundup(size_mul(new_n, size));
1584 new_arr = krealloc(arr, alloc_size, GFP_KERNEL);
1592 memset(arr + old_n * size, 0, (new_n - old_n) * size);
1595 return arr ? arr : ZERO_SIZE_PTR;
1598 static int copy_reference_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
1600 dst->refs = copy_array(dst->refs, src->refs, src->acquired_refs,
1601 sizeof(struct bpf_reference_state), GFP_KERNEL);
1605 dst->acquired_refs = src->acquired_refs;
1609 static int copy_stack_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
1611 size_t n = src->allocated_stack / BPF_REG_SIZE;
1613 dst->stack = copy_array(dst->stack, src->stack, n, sizeof(struct bpf_stack_state),
1618 dst->allocated_stack = src->allocated_stack;
1622 static int resize_reference_state(struct bpf_func_state *state, size_t n)
1624 state->refs = realloc_array(state->refs, state->acquired_refs, n,
1625 sizeof(struct bpf_reference_state));
1629 state->acquired_refs = n;
1633 static int grow_stack_state(struct bpf_func_state *state, int size)
1635 size_t old_n = state->allocated_stack / BPF_REG_SIZE, n = size / BPF_REG_SIZE;
1640 state->stack = realloc_array(state->stack, old_n, n, sizeof(struct bpf_stack_state));
1644 state->allocated_stack = size;
1648 /* Acquire a pointer id from the env and update the state->refs to include
1649 * this new pointer reference.
1650 * On success, returns a valid pointer id to associate with the register
1651 * On failure, returns a negative errno.
1653 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
1655 struct bpf_func_state *state = cur_func(env);
1656 int new_ofs = state->acquired_refs;
1659 err = resize_reference_state(state, state->acquired_refs + 1);
1663 state->refs[new_ofs].id = id;
1664 state->refs[new_ofs].insn_idx = insn_idx;
1665 state->refs[new_ofs].callback_ref = state->in_callback_fn ? state->frameno : 0;
1670 /* release function corresponding to acquire_reference_state(). Idempotent. */
1671 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
1675 last_idx = state->acquired_refs - 1;
1676 for (i = 0; i < state->acquired_refs; i++) {
1677 if (state->refs[i].id == ptr_id) {
1678 /* Cannot release caller references in callbacks */
1679 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno)
1681 if (last_idx && i != last_idx)
1682 memcpy(&state->refs[i], &state->refs[last_idx],
1683 sizeof(*state->refs));
1684 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
1685 state->acquired_refs--;
1692 static void free_func_state(struct bpf_func_state *state)
1697 kfree(state->stack);
1701 static void clear_jmp_history(struct bpf_verifier_state *state)
1703 kfree(state->jmp_history);
1704 state->jmp_history = NULL;
1705 state->jmp_history_cnt = 0;
1708 static void free_verifier_state(struct bpf_verifier_state *state,
1713 for (i = 0; i <= state->curframe; i++) {
1714 free_func_state(state->frame[i]);
1715 state->frame[i] = NULL;
1717 clear_jmp_history(state);
1722 /* copy verifier state from src to dst growing dst stack space
1723 * when necessary to accommodate larger src stack
1725 static int copy_func_state(struct bpf_func_state *dst,
1726 const struct bpf_func_state *src)
1730 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
1731 err = copy_reference_state(dst, src);
1734 return copy_stack_state(dst, src);
1737 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
1738 const struct bpf_verifier_state *src)
1740 struct bpf_func_state *dst;
1743 dst_state->jmp_history = copy_array(dst_state->jmp_history, src->jmp_history,
1744 src->jmp_history_cnt, sizeof(struct bpf_idx_pair),
1746 if (!dst_state->jmp_history)
1748 dst_state->jmp_history_cnt = src->jmp_history_cnt;
1750 /* if dst has more stack frames then src frame, free them */
1751 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
1752 free_func_state(dst_state->frame[i]);
1753 dst_state->frame[i] = NULL;
1755 dst_state->speculative = src->speculative;
1756 dst_state->active_rcu_lock = src->active_rcu_lock;
1757 dst_state->curframe = src->curframe;
1758 dst_state->active_lock.ptr = src->active_lock.ptr;
1759 dst_state->active_lock.id = src->active_lock.id;
1760 dst_state->branches = src->branches;
1761 dst_state->parent = src->parent;
1762 dst_state->first_insn_idx = src->first_insn_idx;
1763 dst_state->last_insn_idx = src->last_insn_idx;
1764 for (i = 0; i <= src->curframe; i++) {
1765 dst = dst_state->frame[i];
1767 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
1770 dst_state->frame[i] = dst;
1772 err = copy_func_state(dst, src->frame[i]);
1779 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
1782 u32 br = --st->branches;
1784 /* WARN_ON(br > 1) technically makes sense here,
1785 * but see comment in push_stack(), hence:
1787 WARN_ONCE((int)br < 0,
1788 "BUG update_branch_counts:branches_to_explore=%d\n",
1796 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
1797 int *insn_idx, bool pop_log)
1799 struct bpf_verifier_state *cur = env->cur_state;
1800 struct bpf_verifier_stack_elem *elem, *head = env->head;
1803 if (env->head == NULL)
1807 err = copy_verifier_state(cur, &head->st);
1812 bpf_vlog_reset(&env->log, head->log_pos);
1814 *insn_idx = head->insn_idx;
1816 *prev_insn_idx = head->prev_insn_idx;
1818 free_verifier_state(&head->st, false);
1825 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
1826 int insn_idx, int prev_insn_idx,
1829 struct bpf_verifier_state *cur = env->cur_state;
1830 struct bpf_verifier_stack_elem *elem;
1833 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1837 elem->insn_idx = insn_idx;
1838 elem->prev_insn_idx = prev_insn_idx;
1839 elem->next = env->head;
1840 elem->log_pos = env->log.end_pos;
1843 err = copy_verifier_state(&elem->st, cur);
1846 elem->st.speculative |= speculative;
1847 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1848 verbose(env, "The sequence of %d jumps is too complex.\n",
1852 if (elem->st.parent) {
1853 ++elem->st.parent->branches;
1854 /* WARN_ON(branches > 2) technically makes sense here,
1856 * 1. speculative states will bump 'branches' for non-branch
1858 * 2. is_state_visited() heuristics may decide not to create
1859 * a new state for a sequence of branches and all such current
1860 * and cloned states will be pointing to a single parent state
1861 * which might have large 'branches' count.
1866 free_verifier_state(env->cur_state, true);
1867 env->cur_state = NULL;
1868 /* pop all elements and return */
1869 while (!pop_stack(env, NULL, NULL, false));
1873 #define CALLER_SAVED_REGS 6
1874 static const int caller_saved[CALLER_SAVED_REGS] = {
1875 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
1878 /* This helper doesn't clear reg->id */
1879 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1881 reg->var_off = tnum_const(imm);
1882 reg->smin_value = (s64)imm;
1883 reg->smax_value = (s64)imm;
1884 reg->umin_value = imm;
1885 reg->umax_value = imm;
1887 reg->s32_min_value = (s32)imm;
1888 reg->s32_max_value = (s32)imm;
1889 reg->u32_min_value = (u32)imm;
1890 reg->u32_max_value = (u32)imm;
1893 /* Mark the unknown part of a register (variable offset or scalar value) as
1894 * known to have the value @imm.
1896 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1898 /* Clear off and union(map_ptr, range) */
1899 memset(((u8 *)reg) + sizeof(reg->type), 0,
1900 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1902 reg->ref_obj_id = 0;
1903 ___mark_reg_known(reg, imm);
1906 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1908 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1909 reg->s32_min_value = (s32)imm;
1910 reg->s32_max_value = (s32)imm;
1911 reg->u32_min_value = (u32)imm;
1912 reg->u32_max_value = (u32)imm;
1915 /* Mark the 'variable offset' part of a register as zero. This should be
1916 * used only on registers holding a pointer type.
1918 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1920 __mark_reg_known(reg, 0);
1923 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1925 __mark_reg_known(reg, 0);
1926 reg->type = SCALAR_VALUE;
1929 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1930 struct bpf_reg_state *regs, u32 regno)
1932 if (WARN_ON(regno >= MAX_BPF_REG)) {
1933 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1934 /* Something bad happened, let's kill all regs */
1935 for (regno = 0; regno < MAX_BPF_REG; regno++)
1936 __mark_reg_not_init(env, regs + regno);
1939 __mark_reg_known_zero(regs + regno);
1942 static void __mark_dynptr_reg(struct bpf_reg_state *reg, enum bpf_dynptr_type type,
1943 bool first_slot, int dynptr_id)
1945 /* reg->type has no meaning for STACK_DYNPTR, but when we set reg for
1946 * callback arguments, it does need to be CONST_PTR_TO_DYNPTR, so simply
1947 * set it unconditionally as it is ignored for STACK_DYNPTR anyway.
1949 __mark_reg_known_zero(reg);
1950 reg->type = CONST_PTR_TO_DYNPTR;
1951 /* Give each dynptr a unique id to uniquely associate slices to it. */
1952 reg->id = dynptr_id;
1953 reg->dynptr.type = type;
1954 reg->dynptr.first_slot = first_slot;
1957 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
1959 if (base_type(reg->type) == PTR_TO_MAP_VALUE) {
1960 const struct bpf_map *map = reg->map_ptr;
1962 if (map->inner_map_meta) {
1963 reg->type = CONST_PTR_TO_MAP;
1964 reg->map_ptr = map->inner_map_meta;
1965 /* transfer reg's id which is unique for every map_lookup_elem
1966 * as UID of the inner map.
1968 if (btf_record_has_field(map->inner_map_meta->record, BPF_TIMER))
1969 reg->map_uid = reg->id;
1970 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
1971 reg->type = PTR_TO_XDP_SOCK;
1972 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
1973 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
1974 reg->type = PTR_TO_SOCKET;
1976 reg->type = PTR_TO_MAP_VALUE;
1981 reg->type &= ~PTR_MAYBE_NULL;
1984 static void mark_reg_graph_node(struct bpf_reg_state *regs, u32 regno,
1985 struct btf_field_graph_root *ds_head)
1987 __mark_reg_known_zero(®s[regno]);
1988 regs[regno].type = PTR_TO_BTF_ID | MEM_ALLOC;
1989 regs[regno].btf = ds_head->btf;
1990 regs[regno].btf_id = ds_head->value_btf_id;
1991 regs[regno].off = ds_head->node_offset;
1994 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1996 return type_is_pkt_pointer(reg->type);
1999 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
2001 return reg_is_pkt_pointer(reg) ||
2002 reg->type == PTR_TO_PACKET_END;
2005 static bool reg_is_dynptr_slice_pkt(const struct bpf_reg_state *reg)
2007 return base_type(reg->type) == PTR_TO_MEM &&
2008 (reg->type & DYNPTR_TYPE_SKB || reg->type & DYNPTR_TYPE_XDP);
2011 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
2012 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
2013 enum bpf_reg_type which)
2015 /* The register can already have a range from prior markings.
2016 * This is fine as long as it hasn't been advanced from its
2019 return reg->type == which &&
2022 tnum_equals_const(reg->var_off, 0);
2025 /* Reset the min/max bounds of a register */
2026 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
2028 reg->smin_value = S64_MIN;
2029 reg->smax_value = S64_MAX;
2030 reg->umin_value = 0;
2031 reg->umax_value = U64_MAX;
2033 reg->s32_min_value = S32_MIN;
2034 reg->s32_max_value = S32_MAX;
2035 reg->u32_min_value = 0;
2036 reg->u32_max_value = U32_MAX;
2039 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
2041 reg->smin_value = S64_MIN;
2042 reg->smax_value = S64_MAX;
2043 reg->umin_value = 0;
2044 reg->umax_value = U64_MAX;
2047 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
2049 reg->s32_min_value = S32_MIN;
2050 reg->s32_max_value = S32_MAX;
2051 reg->u32_min_value = 0;
2052 reg->u32_max_value = U32_MAX;
2055 static void __update_reg32_bounds(struct bpf_reg_state *reg)
2057 struct tnum var32_off = tnum_subreg(reg->var_off);
2059 /* min signed is max(sign bit) | min(other bits) */
2060 reg->s32_min_value = max_t(s32, reg->s32_min_value,
2061 var32_off.value | (var32_off.mask & S32_MIN));
2062 /* max signed is min(sign bit) | max(other bits) */
2063 reg->s32_max_value = min_t(s32, reg->s32_max_value,
2064 var32_off.value | (var32_off.mask & S32_MAX));
2065 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
2066 reg->u32_max_value = min(reg->u32_max_value,
2067 (u32)(var32_off.value | var32_off.mask));
2070 static void __update_reg64_bounds(struct bpf_reg_state *reg)
2072 /* min signed is max(sign bit) | min(other bits) */
2073 reg->smin_value = max_t(s64, reg->smin_value,
2074 reg->var_off.value | (reg->var_off.mask & S64_MIN));
2075 /* max signed is min(sign bit) | max(other bits) */
2076 reg->smax_value = min_t(s64, reg->smax_value,
2077 reg->var_off.value | (reg->var_off.mask & S64_MAX));
2078 reg->umin_value = max(reg->umin_value, reg->var_off.value);
2079 reg->umax_value = min(reg->umax_value,
2080 reg->var_off.value | reg->var_off.mask);
2083 static void __update_reg_bounds(struct bpf_reg_state *reg)
2085 __update_reg32_bounds(reg);
2086 __update_reg64_bounds(reg);
2089 /* Uses signed min/max values to inform unsigned, and vice-versa */
2090 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
2092 /* Learn sign from signed bounds.
2093 * If we cannot cross the sign boundary, then signed and unsigned bounds
2094 * are the same, so combine. This works even in the negative case, e.g.
2095 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
2097 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
2098 reg->s32_min_value = reg->u32_min_value =
2099 max_t(u32, reg->s32_min_value, reg->u32_min_value);
2100 reg->s32_max_value = reg->u32_max_value =
2101 min_t(u32, reg->s32_max_value, reg->u32_max_value);
2104 /* Learn sign from unsigned bounds. Signed bounds cross the sign
2105 * boundary, so we must be careful.
2107 if ((s32)reg->u32_max_value >= 0) {
2108 /* Positive. We can't learn anything from the smin, but smax
2109 * is positive, hence safe.
2111 reg->s32_min_value = reg->u32_min_value;
2112 reg->s32_max_value = reg->u32_max_value =
2113 min_t(u32, reg->s32_max_value, reg->u32_max_value);
2114 } else if ((s32)reg->u32_min_value < 0) {
2115 /* Negative. We can't learn anything from the smax, but smin
2116 * is negative, hence safe.
2118 reg->s32_min_value = reg->u32_min_value =
2119 max_t(u32, reg->s32_min_value, reg->u32_min_value);
2120 reg->s32_max_value = reg->u32_max_value;
2124 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
2126 /* Learn sign from signed bounds.
2127 * If we cannot cross the sign boundary, then signed and unsigned bounds
2128 * are the same, so combine. This works even in the negative case, e.g.
2129 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
2131 if (reg->smin_value >= 0 || reg->smax_value < 0) {
2132 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
2134 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
2138 /* Learn sign from unsigned bounds. Signed bounds cross the sign
2139 * boundary, so we must be careful.
2141 if ((s64)reg->umax_value >= 0) {
2142 /* Positive. We can't learn anything from the smin, but smax
2143 * is positive, hence safe.
2145 reg->smin_value = reg->umin_value;
2146 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
2148 } else if ((s64)reg->umin_value < 0) {
2149 /* Negative. We can't learn anything from the smax, but smin
2150 * is negative, hence safe.
2152 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
2154 reg->smax_value = reg->umax_value;
2158 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
2160 __reg32_deduce_bounds(reg);
2161 __reg64_deduce_bounds(reg);
2164 /* Attempts to improve var_off based on unsigned min/max information */
2165 static void __reg_bound_offset(struct bpf_reg_state *reg)
2167 struct tnum var64_off = tnum_intersect(reg->var_off,
2168 tnum_range(reg->umin_value,
2170 struct tnum var32_off = tnum_intersect(tnum_subreg(var64_off),
2171 tnum_range(reg->u32_min_value,
2172 reg->u32_max_value));
2174 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
2177 static void reg_bounds_sync(struct bpf_reg_state *reg)
2179 /* We might have learned new bounds from the var_off. */
2180 __update_reg_bounds(reg);
2181 /* We might have learned something about the sign bit. */
2182 __reg_deduce_bounds(reg);
2183 /* We might have learned some bits from the bounds. */
2184 __reg_bound_offset(reg);
2185 /* Intersecting with the old var_off might have improved our bounds
2186 * slightly, e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2187 * then new var_off is (0; 0x7f...fc) which improves our umax.
2189 __update_reg_bounds(reg);
2192 static bool __reg32_bound_s64(s32 a)
2194 return a >= 0 && a <= S32_MAX;
2197 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
2199 reg->umin_value = reg->u32_min_value;
2200 reg->umax_value = reg->u32_max_value;
2202 /* Attempt to pull 32-bit signed bounds into 64-bit bounds but must
2203 * be positive otherwise set to worse case bounds and refine later
2206 if (__reg32_bound_s64(reg->s32_min_value) &&
2207 __reg32_bound_s64(reg->s32_max_value)) {
2208 reg->smin_value = reg->s32_min_value;
2209 reg->smax_value = reg->s32_max_value;
2211 reg->smin_value = 0;
2212 reg->smax_value = U32_MAX;
2216 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
2218 /* special case when 64-bit register has upper 32-bit register
2219 * zeroed. Typically happens after zext or <<32, >>32 sequence
2220 * allowing us to use 32-bit bounds directly,
2222 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
2223 __reg_assign_32_into_64(reg);
2225 /* Otherwise the best we can do is push lower 32bit known and
2226 * unknown bits into register (var_off set from jmp logic)
2227 * then learn as much as possible from the 64-bit tnum
2228 * known and unknown bits. The previous smin/smax bounds are
2229 * invalid here because of jmp32 compare so mark them unknown
2230 * so they do not impact tnum bounds calculation.
2232 __mark_reg64_unbounded(reg);
2234 reg_bounds_sync(reg);
2237 static bool __reg64_bound_s32(s64 a)
2239 return a >= S32_MIN && a <= S32_MAX;
2242 static bool __reg64_bound_u32(u64 a)
2244 return a >= U32_MIN && a <= U32_MAX;
2247 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
2249 __mark_reg32_unbounded(reg);
2250 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
2251 reg->s32_min_value = (s32)reg->smin_value;
2252 reg->s32_max_value = (s32)reg->smax_value;
2254 if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
2255 reg->u32_min_value = (u32)reg->umin_value;
2256 reg->u32_max_value = (u32)reg->umax_value;
2258 reg_bounds_sync(reg);
2261 /* Mark a register as having a completely unknown (scalar) value. */
2262 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
2263 struct bpf_reg_state *reg)
2266 * Clear type, off, and union(map_ptr, range) and
2267 * padding between 'type' and union
2269 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
2270 reg->type = SCALAR_VALUE;
2272 reg->ref_obj_id = 0;
2273 reg->var_off = tnum_unknown;
2275 reg->precise = !env->bpf_capable;
2276 __mark_reg_unbounded(reg);
2279 static void mark_reg_unknown(struct bpf_verifier_env *env,
2280 struct bpf_reg_state *regs, u32 regno)
2282 if (WARN_ON(regno >= MAX_BPF_REG)) {
2283 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
2284 /* Something bad happened, let's kill all regs except FP */
2285 for (regno = 0; regno < BPF_REG_FP; regno++)
2286 __mark_reg_not_init(env, regs + regno);
2289 __mark_reg_unknown(env, regs + regno);
2292 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
2293 struct bpf_reg_state *reg)
2295 __mark_reg_unknown(env, reg);
2296 reg->type = NOT_INIT;
2299 static void mark_reg_not_init(struct bpf_verifier_env *env,
2300 struct bpf_reg_state *regs, u32 regno)
2302 if (WARN_ON(regno >= MAX_BPF_REG)) {
2303 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
2304 /* Something bad happened, let's kill all regs except FP */
2305 for (regno = 0; regno < BPF_REG_FP; regno++)
2306 __mark_reg_not_init(env, regs + regno);
2309 __mark_reg_not_init(env, regs + regno);
2312 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
2313 struct bpf_reg_state *regs, u32 regno,
2314 enum bpf_reg_type reg_type,
2315 struct btf *btf, u32 btf_id,
2316 enum bpf_type_flag flag)
2318 if (reg_type == SCALAR_VALUE) {
2319 mark_reg_unknown(env, regs, regno);
2322 mark_reg_known_zero(env, regs, regno);
2323 regs[regno].type = PTR_TO_BTF_ID | flag;
2324 regs[regno].btf = btf;
2325 regs[regno].btf_id = btf_id;
2328 #define DEF_NOT_SUBREG (0)
2329 static void init_reg_state(struct bpf_verifier_env *env,
2330 struct bpf_func_state *state)
2332 struct bpf_reg_state *regs = state->regs;
2335 for (i = 0; i < MAX_BPF_REG; i++) {
2336 mark_reg_not_init(env, regs, i);
2337 regs[i].live = REG_LIVE_NONE;
2338 regs[i].parent = NULL;
2339 regs[i].subreg_def = DEF_NOT_SUBREG;
2343 regs[BPF_REG_FP].type = PTR_TO_STACK;
2344 mark_reg_known_zero(env, regs, BPF_REG_FP);
2345 regs[BPF_REG_FP].frameno = state->frameno;
2348 #define BPF_MAIN_FUNC (-1)
2349 static void init_func_state(struct bpf_verifier_env *env,
2350 struct bpf_func_state *state,
2351 int callsite, int frameno, int subprogno)
2353 state->callsite = callsite;
2354 state->frameno = frameno;
2355 state->subprogno = subprogno;
2356 state->callback_ret_range = tnum_range(0, 0);
2357 init_reg_state(env, state);
2358 mark_verifier_state_scratched(env);
2361 /* Similar to push_stack(), but for async callbacks */
2362 static struct bpf_verifier_state *push_async_cb(struct bpf_verifier_env *env,
2363 int insn_idx, int prev_insn_idx,
2366 struct bpf_verifier_stack_elem *elem;
2367 struct bpf_func_state *frame;
2369 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
2373 elem->insn_idx = insn_idx;
2374 elem->prev_insn_idx = prev_insn_idx;
2375 elem->next = env->head;
2376 elem->log_pos = env->log.end_pos;
2379 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
2381 "The sequence of %d jumps is too complex for async cb.\n",
2385 /* Unlike push_stack() do not copy_verifier_state().
2386 * The caller state doesn't matter.
2387 * This is async callback. It starts in a fresh stack.
2388 * Initialize it similar to do_check_common().
2390 elem->st.branches = 1;
2391 frame = kzalloc(sizeof(*frame), GFP_KERNEL);
2394 init_func_state(env, frame,
2395 BPF_MAIN_FUNC /* callsite */,
2396 0 /* frameno within this callchain */,
2397 subprog /* subprog number within this prog */);
2398 elem->st.frame[0] = frame;
2401 free_verifier_state(env->cur_state, true);
2402 env->cur_state = NULL;
2403 /* pop all elements and return */
2404 while (!pop_stack(env, NULL, NULL, false));
2410 SRC_OP, /* register is used as source operand */
2411 DST_OP, /* register is used as destination operand */
2412 DST_OP_NO_MARK /* same as above, check only, don't mark */
2415 static int cmp_subprogs(const void *a, const void *b)
2417 return ((struct bpf_subprog_info *)a)->start -
2418 ((struct bpf_subprog_info *)b)->start;
2421 static int find_subprog(struct bpf_verifier_env *env, int off)
2423 struct bpf_subprog_info *p;
2425 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
2426 sizeof(env->subprog_info[0]), cmp_subprogs);
2429 return p - env->subprog_info;
2433 static int add_subprog(struct bpf_verifier_env *env, int off)
2435 int insn_cnt = env->prog->len;
2438 if (off >= insn_cnt || off < 0) {
2439 verbose(env, "call to invalid destination\n");
2442 ret = find_subprog(env, off);
2445 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
2446 verbose(env, "too many subprograms\n");
2449 /* determine subprog starts. The end is one before the next starts */
2450 env->subprog_info[env->subprog_cnt++].start = off;
2451 sort(env->subprog_info, env->subprog_cnt,
2452 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
2453 return env->subprog_cnt - 1;
2456 #define MAX_KFUNC_DESCS 256
2457 #define MAX_KFUNC_BTFS 256
2459 struct bpf_kfunc_desc {
2460 struct btf_func_model func_model;
2467 struct bpf_kfunc_btf {
2469 struct module *module;
2473 struct bpf_kfunc_desc_tab {
2474 /* Sorted by func_id (BTF ID) and offset (fd_array offset) during
2475 * verification. JITs do lookups by bpf_insn, where func_id may not be
2476 * available, therefore at the end of verification do_misc_fixups()
2477 * sorts this by imm and offset.
2479 struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
2483 struct bpf_kfunc_btf_tab {
2484 struct bpf_kfunc_btf descs[MAX_KFUNC_BTFS];
2488 static int kfunc_desc_cmp_by_id_off(const void *a, const void *b)
2490 const struct bpf_kfunc_desc *d0 = a;
2491 const struct bpf_kfunc_desc *d1 = b;
2493 /* func_id is not greater than BTF_MAX_TYPE */
2494 return d0->func_id - d1->func_id ?: d0->offset - d1->offset;
2497 static int kfunc_btf_cmp_by_off(const void *a, const void *b)
2499 const struct bpf_kfunc_btf *d0 = a;
2500 const struct bpf_kfunc_btf *d1 = b;
2502 return d0->offset - d1->offset;
2505 static const struct bpf_kfunc_desc *
2506 find_kfunc_desc(const struct bpf_prog *prog, u32 func_id, u16 offset)
2508 struct bpf_kfunc_desc desc = {
2512 struct bpf_kfunc_desc_tab *tab;
2514 tab = prog->aux->kfunc_tab;
2515 return bsearch(&desc, tab->descs, tab->nr_descs,
2516 sizeof(tab->descs[0]), kfunc_desc_cmp_by_id_off);
2519 int bpf_get_kfunc_addr(const struct bpf_prog *prog, u32 func_id,
2520 u16 btf_fd_idx, u8 **func_addr)
2522 const struct bpf_kfunc_desc *desc;
2524 desc = find_kfunc_desc(prog, func_id, btf_fd_idx);
2528 *func_addr = (u8 *)desc->addr;
2532 static struct btf *__find_kfunc_desc_btf(struct bpf_verifier_env *env,
2535 struct bpf_kfunc_btf kf_btf = { .offset = offset };
2536 struct bpf_kfunc_btf_tab *tab;
2537 struct bpf_kfunc_btf *b;
2542 tab = env->prog->aux->kfunc_btf_tab;
2543 b = bsearch(&kf_btf, tab->descs, tab->nr_descs,
2544 sizeof(tab->descs[0]), kfunc_btf_cmp_by_off);
2546 if (tab->nr_descs == MAX_KFUNC_BTFS) {
2547 verbose(env, "too many different module BTFs\n");
2548 return ERR_PTR(-E2BIG);
2551 if (bpfptr_is_null(env->fd_array)) {
2552 verbose(env, "kfunc offset > 0 without fd_array is invalid\n");
2553 return ERR_PTR(-EPROTO);
2556 if (copy_from_bpfptr_offset(&btf_fd, env->fd_array,
2557 offset * sizeof(btf_fd),
2559 return ERR_PTR(-EFAULT);
2561 btf = btf_get_by_fd(btf_fd);
2563 verbose(env, "invalid module BTF fd specified\n");
2567 if (!btf_is_module(btf)) {
2568 verbose(env, "BTF fd for kfunc is not a module BTF\n");
2570 return ERR_PTR(-EINVAL);
2573 mod = btf_try_get_module(btf);
2576 return ERR_PTR(-ENXIO);
2579 b = &tab->descs[tab->nr_descs++];
2584 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2585 kfunc_btf_cmp_by_off, NULL);
2590 void bpf_free_kfunc_btf_tab(struct bpf_kfunc_btf_tab *tab)
2595 while (tab->nr_descs--) {
2596 module_put(tab->descs[tab->nr_descs].module);
2597 btf_put(tab->descs[tab->nr_descs].btf);
2602 static struct btf *find_kfunc_desc_btf(struct bpf_verifier_env *env, s16 offset)
2606 /* In the future, this can be allowed to increase limit
2607 * of fd index into fd_array, interpreted as u16.
2609 verbose(env, "negative offset disallowed for kernel module function call\n");
2610 return ERR_PTR(-EINVAL);
2613 return __find_kfunc_desc_btf(env, offset);
2615 return btf_vmlinux ?: ERR_PTR(-ENOENT);
2618 static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id, s16 offset)
2620 const struct btf_type *func, *func_proto;
2621 struct bpf_kfunc_btf_tab *btf_tab;
2622 struct bpf_kfunc_desc_tab *tab;
2623 struct bpf_prog_aux *prog_aux;
2624 struct bpf_kfunc_desc *desc;
2625 const char *func_name;
2626 struct btf *desc_btf;
2627 unsigned long call_imm;
2631 prog_aux = env->prog->aux;
2632 tab = prog_aux->kfunc_tab;
2633 btf_tab = prog_aux->kfunc_btf_tab;
2636 verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
2640 if (!env->prog->jit_requested) {
2641 verbose(env, "JIT is required for calling kernel function\n");
2645 if (!bpf_jit_supports_kfunc_call()) {
2646 verbose(env, "JIT does not support calling kernel function\n");
2650 if (!env->prog->gpl_compatible) {
2651 verbose(env, "cannot call kernel function from non-GPL compatible program\n");
2655 tab = kzalloc(sizeof(*tab), GFP_KERNEL);
2658 prog_aux->kfunc_tab = tab;
2661 /* func_id == 0 is always invalid, but instead of returning an error, be
2662 * conservative and wait until the code elimination pass before returning
2663 * error, so that invalid calls that get pruned out can be in BPF programs
2664 * loaded from userspace. It is also required that offset be untouched
2667 if (!func_id && !offset)
2670 if (!btf_tab && offset) {
2671 btf_tab = kzalloc(sizeof(*btf_tab), GFP_KERNEL);
2674 prog_aux->kfunc_btf_tab = btf_tab;
2677 desc_btf = find_kfunc_desc_btf(env, offset);
2678 if (IS_ERR(desc_btf)) {
2679 verbose(env, "failed to find BTF for kernel function\n");
2680 return PTR_ERR(desc_btf);
2683 if (find_kfunc_desc(env->prog, func_id, offset))
2686 if (tab->nr_descs == MAX_KFUNC_DESCS) {
2687 verbose(env, "too many different kernel function calls\n");
2691 func = btf_type_by_id(desc_btf, func_id);
2692 if (!func || !btf_type_is_func(func)) {
2693 verbose(env, "kernel btf_id %u is not a function\n",
2697 func_proto = btf_type_by_id(desc_btf, func->type);
2698 if (!func_proto || !btf_type_is_func_proto(func_proto)) {
2699 verbose(env, "kernel function btf_id %u does not have a valid func_proto\n",
2704 func_name = btf_name_by_offset(desc_btf, func->name_off);
2705 addr = kallsyms_lookup_name(func_name);
2707 verbose(env, "cannot find address for kernel function %s\n",
2711 specialize_kfunc(env, func_id, offset, &addr);
2713 if (bpf_jit_supports_far_kfunc_call()) {
2716 call_imm = BPF_CALL_IMM(addr);
2717 /* Check whether the relative offset overflows desc->imm */
2718 if ((unsigned long)(s32)call_imm != call_imm) {
2719 verbose(env, "address of kernel function %s is out of range\n",
2725 if (bpf_dev_bound_kfunc_id(func_id)) {
2726 err = bpf_dev_bound_kfunc_check(&env->log, prog_aux);
2731 desc = &tab->descs[tab->nr_descs++];
2732 desc->func_id = func_id;
2733 desc->imm = call_imm;
2734 desc->offset = offset;
2736 err = btf_distill_func_proto(&env->log, desc_btf,
2737 func_proto, func_name,
2740 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2741 kfunc_desc_cmp_by_id_off, NULL);
2745 static int kfunc_desc_cmp_by_imm_off(const void *a, const void *b)
2747 const struct bpf_kfunc_desc *d0 = a;
2748 const struct bpf_kfunc_desc *d1 = b;
2750 if (d0->imm != d1->imm)
2751 return d0->imm < d1->imm ? -1 : 1;
2752 if (d0->offset != d1->offset)
2753 return d0->offset < d1->offset ? -1 : 1;
2757 static void sort_kfunc_descs_by_imm_off(struct bpf_prog *prog)
2759 struct bpf_kfunc_desc_tab *tab;
2761 tab = prog->aux->kfunc_tab;
2765 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2766 kfunc_desc_cmp_by_imm_off, NULL);
2769 bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
2771 return !!prog->aux->kfunc_tab;
2774 const struct btf_func_model *
2775 bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
2776 const struct bpf_insn *insn)
2778 const struct bpf_kfunc_desc desc = {
2780 .offset = insn->off,
2782 const struct bpf_kfunc_desc *res;
2783 struct bpf_kfunc_desc_tab *tab;
2785 tab = prog->aux->kfunc_tab;
2786 res = bsearch(&desc, tab->descs, tab->nr_descs,
2787 sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm_off);
2789 return res ? &res->func_model : NULL;
2792 static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
2794 struct bpf_subprog_info *subprog = env->subprog_info;
2795 struct bpf_insn *insn = env->prog->insnsi;
2796 int i, ret, insn_cnt = env->prog->len;
2798 /* Add entry function. */
2799 ret = add_subprog(env, 0);
2803 for (i = 0; i < insn_cnt; i++, insn++) {
2804 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
2805 !bpf_pseudo_kfunc_call(insn))
2808 if (!env->bpf_capable) {
2809 verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
2813 if (bpf_pseudo_func(insn) || bpf_pseudo_call(insn))
2814 ret = add_subprog(env, i + insn->imm + 1);
2816 ret = add_kfunc_call(env, insn->imm, insn->off);
2822 /* Add a fake 'exit' subprog which could simplify subprog iteration
2823 * logic. 'subprog_cnt' should not be increased.
2825 subprog[env->subprog_cnt].start = insn_cnt;
2827 if (env->log.level & BPF_LOG_LEVEL2)
2828 for (i = 0; i < env->subprog_cnt; i++)
2829 verbose(env, "func#%d @%d\n", i, subprog[i].start);
2834 static int check_subprogs(struct bpf_verifier_env *env)
2836 int i, subprog_start, subprog_end, off, cur_subprog = 0;
2837 struct bpf_subprog_info *subprog = env->subprog_info;
2838 struct bpf_insn *insn = env->prog->insnsi;
2839 int insn_cnt = env->prog->len;
2841 /* now check that all jumps are within the same subprog */
2842 subprog_start = subprog[cur_subprog].start;
2843 subprog_end = subprog[cur_subprog + 1].start;
2844 for (i = 0; i < insn_cnt; i++) {
2845 u8 code = insn[i].code;
2847 if (code == (BPF_JMP | BPF_CALL) &&
2848 insn[i].src_reg == 0 &&
2849 insn[i].imm == BPF_FUNC_tail_call)
2850 subprog[cur_subprog].has_tail_call = true;
2851 if (BPF_CLASS(code) == BPF_LD &&
2852 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
2853 subprog[cur_subprog].has_ld_abs = true;
2854 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
2856 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
2858 off = i + insn[i].off + 1;
2859 if (off < subprog_start || off >= subprog_end) {
2860 verbose(env, "jump out of range from insn %d to %d\n", i, off);
2864 if (i == subprog_end - 1) {
2865 /* to avoid fall-through from one subprog into another
2866 * the last insn of the subprog should be either exit
2867 * or unconditional jump back
2869 if (code != (BPF_JMP | BPF_EXIT) &&
2870 code != (BPF_JMP | BPF_JA)) {
2871 verbose(env, "last insn is not an exit or jmp\n");
2874 subprog_start = subprog_end;
2876 if (cur_subprog < env->subprog_cnt)
2877 subprog_end = subprog[cur_subprog + 1].start;
2883 /* Parentage chain of this register (or stack slot) should take care of all
2884 * issues like callee-saved registers, stack slot allocation time, etc.
2886 static int mark_reg_read(struct bpf_verifier_env *env,
2887 const struct bpf_reg_state *state,
2888 struct bpf_reg_state *parent, u8 flag)
2890 bool writes = parent == state->parent; /* Observe write marks */
2894 /* if read wasn't screened by an earlier write ... */
2895 if (writes && state->live & REG_LIVE_WRITTEN)
2897 if (parent->live & REG_LIVE_DONE) {
2898 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
2899 reg_type_str(env, parent->type),
2900 parent->var_off.value, parent->off);
2903 /* The first condition is more likely to be true than the
2904 * second, checked it first.
2906 if ((parent->live & REG_LIVE_READ) == flag ||
2907 parent->live & REG_LIVE_READ64)
2908 /* The parentage chain never changes and
2909 * this parent was already marked as LIVE_READ.
2910 * There is no need to keep walking the chain again and
2911 * keep re-marking all parents as LIVE_READ.
2912 * This case happens when the same register is read
2913 * multiple times without writes into it in-between.
2914 * Also, if parent has the stronger REG_LIVE_READ64 set,
2915 * then no need to set the weak REG_LIVE_READ32.
2918 /* ... then we depend on parent's value */
2919 parent->live |= flag;
2920 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
2921 if (flag == REG_LIVE_READ64)
2922 parent->live &= ~REG_LIVE_READ32;
2924 parent = state->parent;
2929 if (env->longest_mark_read_walk < cnt)
2930 env->longest_mark_read_walk = cnt;
2934 static int mark_dynptr_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
2936 struct bpf_func_state *state = func(env, reg);
2939 /* For CONST_PTR_TO_DYNPTR, it must have already been done by
2940 * check_reg_arg in check_helper_call and mark_btf_func_reg_size in
2943 if (reg->type == CONST_PTR_TO_DYNPTR)
2945 spi = dynptr_get_spi(env, reg);
2948 /* Caller ensures dynptr is valid and initialized, which means spi is in
2949 * bounds and spi is the first dynptr slot. Simply mark stack slot as
2952 ret = mark_reg_read(env, &state->stack[spi].spilled_ptr,
2953 state->stack[spi].spilled_ptr.parent, REG_LIVE_READ64);
2956 return mark_reg_read(env, &state->stack[spi - 1].spilled_ptr,
2957 state->stack[spi - 1].spilled_ptr.parent, REG_LIVE_READ64);
2960 static int mark_iter_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
2961 int spi, int nr_slots)
2963 struct bpf_func_state *state = func(env, reg);
2966 for (i = 0; i < nr_slots; i++) {
2967 struct bpf_reg_state *st = &state->stack[spi - i].spilled_ptr;
2969 err = mark_reg_read(env, st, st->parent, REG_LIVE_READ64);
2973 mark_stack_slot_scratched(env, spi - i);
2979 /* This function is supposed to be used by the following 32-bit optimization
2980 * code only. It returns TRUE if the source or destination register operates
2981 * on 64-bit, otherwise return FALSE.
2983 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
2984 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
2989 class = BPF_CLASS(code);
2991 if (class == BPF_JMP) {
2992 /* BPF_EXIT for "main" will reach here. Return TRUE
2997 if (op == BPF_CALL) {
2998 /* BPF to BPF call will reach here because of marking
2999 * caller saved clobber with DST_OP_NO_MARK for which we
3000 * don't care the register def because they are anyway
3001 * marked as NOT_INIT already.
3003 if (insn->src_reg == BPF_PSEUDO_CALL)
3005 /* Helper call will reach here because of arg type
3006 * check, conservatively return TRUE.
3015 if (class == BPF_ALU64 || class == BPF_JMP ||
3016 /* BPF_END always use BPF_ALU class. */
3017 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
3020 if (class == BPF_ALU || class == BPF_JMP32)
3023 if (class == BPF_LDX) {
3025 return BPF_SIZE(code) == BPF_DW;
3026 /* LDX source must be ptr. */
3030 if (class == BPF_STX) {
3031 /* BPF_STX (including atomic variants) has multiple source
3032 * operands, one of which is a ptr. Check whether the caller is
3035 if (t == SRC_OP && reg->type != SCALAR_VALUE)
3037 return BPF_SIZE(code) == BPF_DW;
3040 if (class == BPF_LD) {
3041 u8 mode = BPF_MODE(code);
3044 if (mode == BPF_IMM)
3047 /* Both LD_IND and LD_ABS return 32-bit data. */
3051 /* Implicit ctx ptr. */
3052 if (regno == BPF_REG_6)
3055 /* Explicit source could be any width. */
3059 if (class == BPF_ST)
3060 /* The only source register for BPF_ST is a ptr. */
3063 /* Conservatively return true at default. */
3067 /* Return the regno defined by the insn, or -1. */
3068 static int insn_def_regno(const struct bpf_insn *insn)
3070 switch (BPF_CLASS(insn->code)) {
3076 if (BPF_MODE(insn->code) == BPF_ATOMIC &&
3077 (insn->imm & BPF_FETCH)) {
3078 if (insn->imm == BPF_CMPXCHG)
3081 return insn->src_reg;
3086 return insn->dst_reg;
3090 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
3091 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
3093 int dst_reg = insn_def_regno(insn);
3098 return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
3101 static void mark_insn_zext(struct bpf_verifier_env *env,
3102 struct bpf_reg_state *reg)
3104 s32 def_idx = reg->subreg_def;
3106 if (def_idx == DEF_NOT_SUBREG)
3109 env->insn_aux_data[def_idx - 1].zext_dst = true;
3110 /* The dst will be zero extended, so won't be sub-register anymore. */
3111 reg->subreg_def = DEF_NOT_SUBREG;
3114 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
3115 enum reg_arg_type t)
3117 struct bpf_verifier_state *vstate = env->cur_state;
3118 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3119 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
3120 struct bpf_reg_state *reg, *regs = state->regs;
3123 if (regno >= MAX_BPF_REG) {
3124 verbose(env, "R%d is invalid\n", regno);
3128 mark_reg_scratched(env, regno);
3131 rw64 = is_reg64(env, insn, regno, reg, t);
3133 /* check whether register used as source operand can be read */
3134 if (reg->type == NOT_INIT) {
3135 verbose(env, "R%d !read_ok\n", regno);
3138 /* We don't need to worry about FP liveness because it's read-only */
3139 if (regno == BPF_REG_FP)
3143 mark_insn_zext(env, reg);
3145 return mark_reg_read(env, reg, reg->parent,
3146 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
3148 /* check whether register used as dest operand can be written to */
3149 if (regno == BPF_REG_FP) {
3150 verbose(env, "frame pointer is read only\n");
3153 reg->live |= REG_LIVE_WRITTEN;
3154 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
3156 mark_reg_unknown(env, regs, regno);
3161 static void mark_jmp_point(struct bpf_verifier_env *env, int idx)
3163 env->insn_aux_data[idx].jmp_point = true;
3166 static bool is_jmp_point(struct bpf_verifier_env *env, int insn_idx)
3168 return env->insn_aux_data[insn_idx].jmp_point;
3171 /* for any branch, call, exit record the history of jmps in the given state */
3172 static int push_jmp_history(struct bpf_verifier_env *env,
3173 struct bpf_verifier_state *cur)
3175 u32 cnt = cur->jmp_history_cnt;
3176 struct bpf_idx_pair *p;
3179 if (!is_jmp_point(env, env->insn_idx))
3183 alloc_size = kmalloc_size_roundup(size_mul(cnt, sizeof(*p)));
3184 p = krealloc(cur->jmp_history, alloc_size, GFP_USER);
3187 p[cnt - 1].idx = env->insn_idx;
3188 p[cnt - 1].prev_idx = env->prev_insn_idx;
3189 cur->jmp_history = p;
3190 cur->jmp_history_cnt = cnt;
3194 /* Backtrack one insn at a time. If idx is not at the top of recorded
3195 * history then previous instruction came from straight line execution.
3197 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
3202 if (cnt && st->jmp_history[cnt - 1].idx == i) {
3203 i = st->jmp_history[cnt - 1].prev_idx;
3211 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
3213 const struct btf_type *func;
3214 struct btf *desc_btf;
3216 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
3219 desc_btf = find_kfunc_desc_btf(data, insn->off);
3220 if (IS_ERR(desc_btf))
3223 func = btf_type_by_id(desc_btf, insn->imm);
3224 return btf_name_by_offset(desc_btf, func->name_off);
3227 static inline void bt_init(struct backtrack_state *bt, u32 frame)
3232 static inline void bt_reset(struct backtrack_state *bt)
3234 struct bpf_verifier_env *env = bt->env;
3236 memset(bt, 0, sizeof(*bt));
3240 static inline u32 bt_empty(struct backtrack_state *bt)
3245 for (i = 0; i <= bt->frame; i++)
3246 mask |= bt->reg_masks[i] | bt->stack_masks[i];
3251 static inline int bt_subprog_enter(struct backtrack_state *bt)
3253 if (bt->frame == MAX_CALL_FRAMES - 1) {
3254 verbose(bt->env, "BUG subprog enter from frame %d\n", bt->frame);
3255 WARN_ONCE(1, "verifier backtracking bug");
3262 static inline int bt_subprog_exit(struct backtrack_state *bt)
3264 if (bt->frame == 0) {
3265 verbose(bt->env, "BUG subprog exit from frame 0\n");
3266 WARN_ONCE(1, "verifier backtracking bug");
3273 static inline void bt_set_frame_reg(struct backtrack_state *bt, u32 frame, u32 reg)
3275 bt->reg_masks[frame] |= 1 << reg;
3278 static inline void bt_clear_frame_reg(struct backtrack_state *bt, u32 frame, u32 reg)
3280 bt->reg_masks[frame] &= ~(1 << reg);
3283 static inline void bt_set_reg(struct backtrack_state *bt, u32 reg)
3285 bt_set_frame_reg(bt, bt->frame, reg);
3288 static inline void bt_clear_reg(struct backtrack_state *bt, u32 reg)
3290 bt_clear_frame_reg(bt, bt->frame, reg);
3293 static inline void bt_set_frame_slot(struct backtrack_state *bt, u32 frame, u32 slot)
3295 bt->stack_masks[frame] |= 1ull << slot;
3298 static inline void bt_clear_frame_slot(struct backtrack_state *bt, u32 frame, u32 slot)
3300 bt->stack_masks[frame] &= ~(1ull << slot);
3303 static inline void bt_set_slot(struct backtrack_state *bt, u32 slot)
3305 bt_set_frame_slot(bt, bt->frame, slot);
3308 static inline void bt_clear_slot(struct backtrack_state *bt, u32 slot)
3310 bt_clear_frame_slot(bt, bt->frame, slot);
3313 static inline u32 bt_frame_reg_mask(struct backtrack_state *bt, u32 frame)
3315 return bt->reg_masks[frame];
3318 static inline u32 bt_reg_mask(struct backtrack_state *bt)
3320 return bt->reg_masks[bt->frame];
3323 static inline u64 bt_frame_stack_mask(struct backtrack_state *bt, u32 frame)
3325 return bt->stack_masks[frame];
3328 static inline u64 bt_stack_mask(struct backtrack_state *bt)
3330 return bt->stack_masks[bt->frame];
3333 static inline bool bt_is_reg_set(struct backtrack_state *bt, u32 reg)
3335 return bt->reg_masks[bt->frame] & (1 << reg);
3338 static inline bool bt_is_slot_set(struct backtrack_state *bt, u32 slot)
3340 return bt->stack_masks[bt->frame] & (1ull << slot);
3343 /* format registers bitmask, e.g., "r0,r2,r4" for 0x15 mask */
3344 static void fmt_reg_mask(char *buf, ssize_t buf_sz, u32 reg_mask)
3346 DECLARE_BITMAP(mask, 64);
3352 bitmap_from_u64(mask, reg_mask);
3353 for_each_set_bit(i, mask, 32) {
3354 n = snprintf(buf, buf_sz, "%sr%d", first ? "" : ",", i);
3362 /* format stack slots bitmask, e.g., "-8,-24,-40" for 0x15 mask */
3363 static void fmt_stack_mask(char *buf, ssize_t buf_sz, u64 stack_mask)
3365 DECLARE_BITMAP(mask, 64);
3371 bitmap_from_u64(mask, stack_mask);
3372 for_each_set_bit(i, mask, 64) {
3373 n = snprintf(buf, buf_sz, "%s%d", first ? "" : ",", -(i + 1) * 8);
3382 /* For given verifier state backtrack_insn() is called from the last insn to
3383 * the first insn. Its purpose is to compute a bitmask of registers and
3384 * stack slots that needs precision in the parent verifier state.
3386 * @idx is an index of the instruction we are currently processing;
3387 * @subseq_idx is an index of the subsequent instruction that:
3388 * - *would be* executed next, if jump history is viewed in forward order;
3389 * - *was* processed previously during backtracking.
3391 static int backtrack_insn(struct bpf_verifier_env *env, int idx, int subseq_idx,
3392 struct backtrack_state *bt)
3394 const struct bpf_insn_cbs cbs = {
3395 .cb_call = disasm_kfunc_name,
3396 .cb_print = verbose,
3397 .private_data = env,
3399 struct bpf_insn *insn = env->prog->insnsi + idx;
3400 u8 class = BPF_CLASS(insn->code);
3401 u8 opcode = BPF_OP(insn->code);
3402 u8 mode = BPF_MODE(insn->code);
3403 u32 dreg = insn->dst_reg;
3404 u32 sreg = insn->src_reg;
3407 if (insn->code == 0)
3409 if (env->log.level & BPF_LOG_LEVEL2) {
3410 fmt_reg_mask(env->tmp_str_buf, TMP_STR_BUF_LEN, bt_reg_mask(bt));
3411 verbose(env, "mark_precise: frame%d: regs=%s ",
3412 bt->frame, env->tmp_str_buf);
3413 fmt_stack_mask(env->tmp_str_buf, TMP_STR_BUF_LEN, bt_stack_mask(bt));
3414 verbose(env, "stack=%s before ", env->tmp_str_buf);
3415 verbose(env, "%d: ", idx);
3416 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
3419 if (class == BPF_ALU || class == BPF_ALU64) {
3420 if (!bt_is_reg_set(bt, dreg))
3422 if (opcode == BPF_MOV) {
3423 if (BPF_SRC(insn->code) == BPF_X) {
3425 * dreg needs precision after this insn
3426 * sreg needs precision before this insn
3428 bt_clear_reg(bt, dreg);
3429 bt_set_reg(bt, sreg);
3432 * dreg needs precision after this insn.
3433 * Corresponding register is already marked
3434 * as precise=true in this verifier state.
3435 * No further markings in parent are necessary
3437 bt_clear_reg(bt, dreg);
3440 if (BPF_SRC(insn->code) == BPF_X) {
3442 * both dreg and sreg need precision
3445 bt_set_reg(bt, sreg);
3447 * dreg still needs precision before this insn
3450 } else if (class == BPF_LDX) {
3451 if (!bt_is_reg_set(bt, dreg))
3453 bt_clear_reg(bt, dreg);
3455 /* scalars can only be spilled into stack w/o losing precision.
3456 * Load from any other memory can be zero extended.
3457 * The desire to keep that precision is already indicated
3458 * by 'precise' mark in corresponding register of this state.
3459 * No further tracking necessary.
3461 if (insn->src_reg != BPF_REG_FP)
3464 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
3465 * that [fp - off] slot contains scalar that needs to be
3466 * tracked with precision
3468 spi = (-insn->off - 1) / BPF_REG_SIZE;
3470 verbose(env, "BUG spi %d\n", spi);
3471 WARN_ONCE(1, "verifier backtracking bug");
3474 bt_set_slot(bt, spi);
3475 } else if (class == BPF_STX || class == BPF_ST) {
3476 if (bt_is_reg_set(bt, dreg))
3477 /* stx & st shouldn't be using _scalar_ dst_reg
3478 * to access memory. It means backtracking
3479 * encountered a case of pointer subtraction.
3482 /* scalars can only be spilled into stack */
3483 if (insn->dst_reg != BPF_REG_FP)
3485 spi = (-insn->off - 1) / BPF_REG_SIZE;
3487 verbose(env, "BUG spi %d\n", spi);
3488 WARN_ONCE(1, "verifier backtracking bug");
3491 if (!bt_is_slot_set(bt, spi))
3493 bt_clear_slot(bt, spi);
3494 if (class == BPF_STX)
3495 bt_set_reg(bt, sreg);
3496 } else if (class == BPF_JMP || class == BPF_JMP32) {
3497 if (bpf_pseudo_call(insn)) {
3498 int subprog_insn_idx, subprog;
3500 subprog_insn_idx = idx + insn->imm + 1;
3501 subprog = find_subprog(env, subprog_insn_idx);
3505 if (subprog_is_global(env, subprog)) {
3506 /* check that jump history doesn't have any
3507 * extra instructions from subprog; the next
3508 * instruction after call to global subprog
3509 * should be literally next instruction in
3512 WARN_ONCE(idx + 1 != subseq_idx, "verifier backtracking bug");
3513 /* r1-r5 are invalidated after subprog call,
3514 * so for global func call it shouldn't be set
3517 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3518 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3519 WARN_ONCE(1, "verifier backtracking bug");
3522 /* global subprog always sets R0 */
3523 bt_clear_reg(bt, BPF_REG_0);
3526 /* static subprog call instruction, which
3527 * means that we are exiting current subprog,
3528 * so only r1-r5 could be still requested as
3529 * precise, r0 and r6-r10 or any stack slot in
3530 * the current frame should be zero by now
3532 if (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) {
3533 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3534 WARN_ONCE(1, "verifier backtracking bug");
3537 /* we don't track register spills perfectly,
3538 * so fallback to force-precise instead of failing */
3539 if (bt_stack_mask(bt) != 0)
3541 /* propagate r1-r5 to the caller */
3542 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
3543 if (bt_is_reg_set(bt, i)) {
3544 bt_clear_reg(bt, i);
3545 bt_set_frame_reg(bt, bt->frame - 1, i);
3548 if (bt_subprog_exit(bt))
3552 } else if ((bpf_helper_call(insn) &&
3553 is_callback_calling_function(insn->imm) &&
3554 !is_async_callback_calling_function(insn->imm)) ||
3555 (bpf_pseudo_kfunc_call(insn) && is_callback_calling_kfunc(insn->imm))) {
3556 /* callback-calling helper or kfunc call, which means
3557 * we are exiting from subprog, but unlike the subprog
3558 * call handling above, we shouldn't propagate
3559 * precision of r1-r5 (if any requested), as they are
3560 * not actually arguments passed directly to callback
3563 if (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) {
3564 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3565 WARN_ONCE(1, "verifier backtracking bug");
3568 if (bt_stack_mask(bt) != 0)
3570 /* clear r1-r5 in callback subprog's mask */
3571 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
3572 bt_clear_reg(bt, i);
3573 if (bt_subprog_exit(bt))
3576 } else if (opcode == BPF_CALL) {
3577 /* kfunc with imm==0 is invalid and fixup_kfunc_call will
3578 * catch this error later. Make backtracking conservative
3581 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && insn->imm == 0)
3583 /* regular helper call sets R0 */
3584 bt_clear_reg(bt, BPF_REG_0);
3585 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3586 /* if backtracing was looking for registers R1-R5
3587 * they should have been found already.
3589 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3590 WARN_ONCE(1, "verifier backtracking bug");
3593 } else if (opcode == BPF_EXIT) {
3596 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3597 /* if backtracing was looking for registers R1-R5
3598 * they should have been found already.
3600 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3601 WARN_ONCE(1, "verifier backtracking bug");
3605 /* BPF_EXIT in subprog or callback always returns
3606 * right after the call instruction, so by checking
3607 * whether the instruction at subseq_idx-1 is subprog
3608 * call or not we can distinguish actual exit from
3609 * *subprog* from exit from *callback*. In the former
3610 * case, we need to propagate r0 precision, if
3611 * necessary. In the former we never do that.
3613 r0_precise = subseq_idx - 1 >= 0 &&
3614 bpf_pseudo_call(&env->prog->insnsi[subseq_idx - 1]) &&
3615 bt_is_reg_set(bt, BPF_REG_0);
3617 bt_clear_reg(bt, BPF_REG_0);
3618 if (bt_subprog_enter(bt))
3622 bt_set_reg(bt, BPF_REG_0);
3623 /* r6-r9 and stack slots will stay set in caller frame
3624 * bitmasks until we return back from callee(s)
3627 } else if (BPF_SRC(insn->code) == BPF_X) {
3628 if (!bt_is_reg_set(bt, dreg) && !bt_is_reg_set(bt, sreg))
3631 * Both dreg and sreg need precision before
3632 * this insn. If only sreg was marked precise
3633 * before it would be equally necessary to
3634 * propagate it to dreg.
3636 bt_set_reg(bt, dreg);
3637 bt_set_reg(bt, sreg);
3638 /* else dreg <cond> K
3639 * Only dreg still needs precision before
3640 * this insn, so for the K-based conditional
3641 * there is nothing new to be marked.
3644 } else if (class == BPF_LD) {
3645 if (!bt_is_reg_set(bt, dreg))
3647 bt_clear_reg(bt, dreg);
3648 /* It's ld_imm64 or ld_abs or ld_ind.
3649 * For ld_imm64 no further tracking of precision
3650 * into parent is necessary
3652 if (mode == BPF_IND || mode == BPF_ABS)
3653 /* to be analyzed */
3659 /* the scalar precision tracking algorithm:
3660 * . at the start all registers have precise=false.
3661 * . scalar ranges are tracked as normal through alu and jmp insns.
3662 * . once precise value of the scalar register is used in:
3663 * . ptr + scalar alu
3664 * . if (scalar cond K|scalar)
3665 * . helper_call(.., scalar, ...) where ARG_CONST is expected
3666 * backtrack through the verifier states and mark all registers and
3667 * stack slots with spilled constants that these scalar regisers
3668 * should be precise.
3669 * . during state pruning two registers (or spilled stack slots)
3670 * are equivalent if both are not precise.
3672 * Note the verifier cannot simply walk register parentage chain,
3673 * since many different registers and stack slots could have been
3674 * used to compute single precise scalar.
3676 * The approach of starting with precise=true for all registers and then
3677 * backtrack to mark a register as not precise when the verifier detects
3678 * that program doesn't care about specific value (e.g., when helper
3679 * takes register as ARG_ANYTHING parameter) is not safe.
3681 * It's ok to walk single parentage chain of the verifier states.
3682 * It's possible that this backtracking will go all the way till 1st insn.
3683 * All other branches will be explored for needing precision later.
3685 * The backtracking needs to deal with cases like:
3686 * 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)
3689 * if r5 > 0x79f goto pc+7
3690 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
3693 * call bpf_perf_event_output#25
3694 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
3698 * call foo // uses callee's r6 inside to compute r0
3702 * to track above reg_mask/stack_mask needs to be independent for each frame.
3704 * Also if parent's curframe > frame where backtracking started,
3705 * the verifier need to mark registers in both frames, otherwise callees
3706 * may incorrectly prune callers. This is similar to
3707 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
3709 * For now backtracking falls back into conservative marking.
3711 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
3712 struct bpf_verifier_state *st)
3714 struct bpf_func_state *func;
3715 struct bpf_reg_state *reg;
3718 if (env->log.level & BPF_LOG_LEVEL2) {
3719 verbose(env, "mark_precise: frame%d: falling back to forcing all scalars precise\n",
3723 /* big hammer: mark all scalars precise in this path.
3724 * pop_stack may still get !precise scalars.
3725 * We also skip current state and go straight to first parent state,
3726 * because precision markings in current non-checkpointed state are
3727 * not needed. See why in the comment in __mark_chain_precision below.
3729 for (st = st->parent; st; st = st->parent) {
3730 for (i = 0; i <= st->curframe; i++) {
3731 func = st->frame[i];
3732 for (j = 0; j < BPF_REG_FP; j++) {
3733 reg = &func->regs[j];
3734 if (reg->type != SCALAR_VALUE || reg->precise)
3736 reg->precise = true;
3737 if (env->log.level & BPF_LOG_LEVEL2) {
3738 verbose(env, "force_precise: frame%d: forcing r%d to be precise\n",
3742 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
3743 if (!is_spilled_reg(&func->stack[j]))
3745 reg = &func->stack[j].spilled_ptr;
3746 if (reg->type != SCALAR_VALUE || reg->precise)
3748 reg->precise = true;
3749 if (env->log.level & BPF_LOG_LEVEL2) {
3750 verbose(env, "force_precise: frame%d: forcing fp%d to be precise\n",
3758 static void mark_all_scalars_imprecise(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
3760 struct bpf_func_state *func;
3761 struct bpf_reg_state *reg;
3764 for (i = 0; i <= st->curframe; i++) {
3765 func = st->frame[i];
3766 for (j = 0; j < BPF_REG_FP; j++) {
3767 reg = &func->regs[j];
3768 if (reg->type != SCALAR_VALUE)
3770 reg->precise = false;
3772 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
3773 if (!is_spilled_reg(&func->stack[j]))
3775 reg = &func->stack[j].spilled_ptr;
3776 if (reg->type != SCALAR_VALUE)
3778 reg->precise = false;
3783 static bool idset_contains(struct bpf_idset *s, u32 id)
3787 for (i = 0; i < s->count; ++i)
3788 if (s->ids[i] == id)
3794 static int idset_push(struct bpf_idset *s, u32 id)
3796 if (WARN_ON_ONCE(s->count >= ARRAY_SIZE(s->ids)))
3798 s->ids[s->count++] = id;
3802 static void idset_reset(struct bpf_idset *s)
3807 /* Collect a set of IDs for all registers currently marked as precise in env->bt.
3808 * Mark all registers with these IDs as precise.
3810 static int mark_precise_scalar_ids(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
3812 struct bpf_idset *precise_ids = &env->idset_scratch;
3813 struct backtrack_state *bt = &env->bt;
3814 struct bpf_func_state *func;
3815 struct bpf_reg_state *reg;
3816 DECLARE_BITMAP(mask, 64);
3819 idset_reset(precise_ids);
3821 for (fr = bt->frame; fr >= 0; fr--) {
3822 func = st->frame[fr];
3824 bitmap_from_u64(mask, bt_frame_reg_mask(bt, fr));
3825 for_each_set_bit(i, mask, 32) {
3826 reg = &func->regs[i];
3827 if (!reg->id || reg->type != SCALAR_VALUE)
3829 if (idset_push(precise_ids, reg->id))
3833 bitmap_from_u64(mask, bt_frame_stack_mask(bt, fr));
3834 for_each_set_bit(i, mask, 64) {
3835 if (i >= func->allocated_stack / BPF_REG_SIZE)
3837 if (!is_spilled_scalar_reg(&func->stack[i]))
3839 reg = &func->stack[i].spilled_ptr;
3842 if (idset_push(precise_ids, reg->id))
3847 for (fr = 0; fr <= st->curframe; ++fr) {
3848 func = st->frame[fr];
3850 for (i = BPF_REG_0; i < BPF_REG_10; ++i) {
3851 reg = &func->regs[i];
3854 if (!idset_contains(precise_ids, reg->id))
3856 bt_set_frame_reg(bt, fr, i);
3858 for (i = 0; i < func->allocated_stack / BPF_REG_SIZE; ++i) {
3859 if (!is_spilled_scalar_reg(&func->stack[i]))
3861 reg = &func->stack[i].spilled_ptr;
3864 if (!idset_contains(precise_ids, reg->id))
3866 bt_set_frame_slot(bt, fr, i);
3874 * __mark_chain_precision() backtracks BPF program instruction sequence and
3875 * chain of verifier states making sure that register *regno* (if regno >= 0)
3876 * and/or stack slot *spi* (if spi >= 0) are marked as precisely tracked
3877 * SCALARS, as well as any other registers and slots that contribute to
3878 * a tracked state of given registers/stack slots, depending on specific BPF
3879 * assembly instructions (see backtrack_insns() for exact instruction handling
3880 * logic). This backtracking relies on recorded jmp_history and is able to
3881 * traverse entire chain of parent states. This process ends only when all the
3882 * necessary registers/slots and their transitive dependencies are marked as
3885 * One important and subtle aspect is that precise marks *do not matter* in
3886 * the currently verified state (current state). It is important to understand
3887 * why this is the case.
3889 * First, note that current state is the state that is not yet "checkpointed",
3890 * i.e., it is not yet put into env->explored_states, and it has no children
3891 * states as well. It's ephemeral, and can end up either a) being discarded if
3892 * compatible explored state is found at some point or BPF_EXIT instruction is
3893 * reached or b) checkpointed and put into env->explored_states, branching out
3894 * into one or more children states.
3896 * In the former case, precise markings in current state are completely
3897 * ignored by state comparison code (see regsafe() for details). Only
3898 * checkpointed ("old") state precise markings are important, and if old
3899 * state's register/slot is precise, regsafe() assumes current state's
3900 * register/slot as precise and checks value ranges exactly and precisely. If
3901 * states turn out to be compatible, current state's necessary precise
3902 * markings and any required parent states' precise markings are enforced
3903 * after the fact with propagate_precision() logic, after the fact. But it's
3904 * important to realize that in this case, even after marking current state
3905 * registers/slots as precise, we immediately discard current state. So what
3906 * actually matters is any of the precise markings propagated into current
3907 * state's parent states, which are always checkpointed (due to b) case above).
3908 * As such, for scenario a) it doesn't matter if current state has precise
3909 * markings set or not.
3911 * Now, for the scenario b), checkpointing and forking into child(ren)
3912 * state(s). Note that before current state gets to checkpointing step, any
3913 * processed instruction always assumes precise SCALAR register/slot
3914 * knowledge: if precise value or range is useful to prune jump branch, BPF
3915 * verifier takes this opportunity enthusiastically. Similarly, when
3916 * register's value is used to calculate offset or memory address, exact
3917 * knowledge of SCALAR range is assumed, checked, and enforced. So, similar to
3918 * what we mentioned above about state comparison ignoring precise markings
3919 * during state comparison, BPF verifier ignores and also assumes precise
3920 * markings *at will* during instruction verification process. But as verifier
3921 * assumes precision, it also propagates any precision dependencies across
3922 * parent states, which are not yet finalized, so can be further restricted
3923 * based on new knowledge gained from restrictions enforced by their children
3924 * states. This is so that once those parent states are finalized, i.e., when
3925 * they have no more active children state, state comparison logic in
3926 * is_state_visited() would enforce strict and precise SCALAR ranges, if
3927 * required for correctness.
3929 * To build a bit more intuition, note also that once a state is checkpointed,
3930 * the path we took to get to that state is not important. This is crucial
3931 * property for state pruning. When state is checkpointed and finalized at
3932 * some instruction index, it can be correctly and safely used to "short
3933 * circuit" any *compatible* state that reaches exactly the same instruction
3934 * index. I.e., if we jumped to that instruction from a completely different
3935 * code path than original finalized state was derived from, it doesn't
3936 * matter, current state can be discarded because from that instruction
3937 * forward having a compatible state will ensure we will safely reach the
3938 * exit. States describe preconditions for further exploration, but completely
3939 * forget the history of how we got here.
3941 * This also means that even if we needed precise SCALAR range to get to
3942 * finalized state, but from that point forward *that same* SCALAR register is
3943 * never used in a precise context (i.e., it's precise value is not needed for
3944 * correctness), it's correct and safe to mark such register as "imprecise"
3945 * (i.e., precise marking set to false). This is what we rely on when we do
3946 * not set precise marking in current state. If no child state requires
3947 * precision for any given SCALAR register, it's safe to dictate that it can
3948 * be imprecise. If any child state does require this register to be precise,
3949 * we'll mark it precise later retroactively during precise markings
3950 * propagation from child state to parent states.
3952 * Skipping precise marking setting in current state is a mild version of
3953 * relying on the above observation. But we can utilize this property even
3954 * more aggressively by proactively forgetting any precise marking in the
3955 * current state (which we inherited from the parent state), right before we
3956 * checkpoint it and branch off into new child state. This is done by
3957 * mark_all_scalars_imprecise() to hopefully get more permissive and generic
3958 * finalized states which help in short circuiting more future states.
3960 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno)
3962 struct backtrack_state *bt = &env->bt;
3963 struct bpf_verifier_state *st = env->cur_state;
3964 int first_idx = st->first_insn_idx;
3965 int last_idx = env->insn_idx;
3966 int subseq_idx = -1;
3967 struct bpf_func_state *func;
3968 struct bpf_reg_state *reg;
3969 bool skip_first = true;
3972 if (!env->bpf_capable)
3975 /* set frame number from which we are starting to backtrack */
3976 bt_init(bt, env->cur_state->curframe);
3978 /* Do sanity checks against current state of register and/or stack
3979 * slot, but don't set precise flag in current state, as precision
3980 * tracking in the current state is unnecessary.
3982 func = st->frame[bt->frame];
3984 reg = &func->regs[regno];
3985 if (reg->type != SCALAR_VALUE) {
3986 WARN_ONCE(1, "backtracing misuse");
3989 bt_set_reg(bt, regno);
3996 DECLARE_BITMAP(mask, 64);
3997 u32 history = st->jmp_history_cnt;
3999 if (env->log.level & BPF_LOG_LEVEL2) {
4000 verbose(env, "mark_precise: frame%d: last_idx %d first_idx %d subseq_idx %d \n",
4001 bt->frame, last_idx, first_idx, subseq_idx);
4004 /* If some register with scalar ID is marked as precise,
4005 * make sure that all registers sharing this ID are also precise.
4006 * This is needed to estimate effect of find_equal_scalars().
4007 * Do this at the last instruction of each state,
4008 * bpf_reg_state::id fields are valid for these instructions.
4010 * Allows to track precision in situation like below:
4012 * r2 = unknown value
4016 * r1 = r2 // r1 and r2 now share the same ID
4018 * --- state #1 {r1.id = A, r2.id = A} ---
4020 * if (r2 > 10) goto exit; // find_equal_scalars() assigns range to r1
4022 * --- state #2 {r1.id = A, r2.id = A} ---
4024 * r3 += r1 // need to mark both r1 and r2
4026 if (mark_precise_scalar_ids(env, st))
4030 /* we are at the entry into subprog, which
4031 * is expected for global funcs, but only if
4032 * requested precise registers are R1-R5
4033 * (which are global func's input arguments)
4035 if (st->curframe == 0 &&
4036 st->frame[0]->subprogno > 0 &&
4037 st->frame[0]->callsite == BPF_MAIN_FUNC &&
4038 bt_stack_mask(bt) == 0 &&
4039 (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) == 0) {
4040 bitmap_from_u64(mask, bt_reg_mask(bt));
4041 for_each_set_bit(i, mask, 32) {
4042 reg = &st->frame[0]->regs[i];
4043 if (reg->type != SCALAR_VALUE) {
4044 bt_clear_reg(bt, i);
4047 reg->precise = true;
4052 verbose(env, "BUG backtracking func entry subprog %d reg_mask %x stack_mask %llx\n",
4053 st->frame[0]->subprogno, bt_reg_mask(bt), bt_stack_mask(bt));
4054 WARN_ONCE(1, "verifier backtracking bug");
4058 for (i = last_idx;;) {
4063 err = backtrack_insn(env, i, subseq_idx, bt);
4065 if (err == -ENOTSUPP) {
4066 mark_all_scalars_precise(env, env->cur_state);
4073 /* Found assignment(s) into tracked register in this state.
4074 * Since this state is already marked, just return.
4075 * Nothing to be tracked further in the parent state.
4081 i = get_prev_insn_idx(st, i, &history);
4082 if (i >= env->prog->len) {
4083 /* This can happen if backtracking reached insn 0
4084 * and there are still reg_mask or stack_mask
4086 * It means the backtracking missed the spot where
4087 * particular register was initialized with a constant.
4089 verbose(env, "BUG backtracking idx %d\n", i);
4090 WARN_ONCE(1, "verifier backtracking bug");
4098 for (fr = bt->frame; fr >= 0; fr--) {
4099 func = st->frame[fr];
4100 bitmap_from_u64(mask, bt_frame_reg_mask(bt, fr));
4101 for_each_set_bit(i, mask, 32) {
4102 reg = &func->regs[i];
4103 if (reg->type != SCALAR_VALUE) {
4104 bt_clear_frame_reg(bt, fr, i);
4108 bt_clear_frame_reg(bt, fr, i);
4110 reg->precise = true;
4113 bitmap_from_u64(mask, bt_frame_stack_mask(bt, fr));
4114 for_each_set_bit(i, mask, 64) {
4115 if (i >= func->allocated_stack / BPF_REG_SIZE) {
4116 /* the sequence of instructions:
4118 * 3: (7b) *(u64 *)(r3 -8) = r0
4119 * 4: (79) r4 = *(u64 *)(r10 -8)
4120 * doesn't contain jmps. It's backtracked
4121 * as a single block.
4122 * During backtracking insn 3 is not recognized as
4123 * stack access, so at the end of backtracking
4124 * stack slot fp-8 is still marked in stack_mask.
4125 * However the parent state may not have accessed
4126 * fp-8 and it's "unallocated" stack space.
4127 * In such case fallback to conservative.
4129 mark_all_scalars_precise(env, env->cur_state);
4134 if (!is_spilled_scalar_reg(&func->stack[i])) {
4135 bt_clear_frame_slot(bt, fr, i);
4138 reg = &func->stack[i].spilled_ptr;
4140 bt_clear_frame_slot(bt, fr, i);
4142 reg->precise = true;
4144 if (env->log.level & BPF_LOG_LEVEL2) {
4145 fmt_reg_mask(env->tmp_str_buf, TMP_STR_BUF_LEN,
4146 bt_frame_reg_mask(bt, fr));
4147 verbose(env, "mark_precise: frame%d: parent state regs=%s ",
4148 fr, env->tmp_str_buf);
4149 fmt_stack_mask(env->tmp_str_buf, TMP_STR_BUF_LEN,
4150 bt_frame_stack_mask(bt, fr));
4151 verbose(env, "stack=%s: ", env->tmp_str_buf);
4152 print_verifier_state(env, func, true);
4159 subseq_idx = first_idx;
4160 last_idx = st->last_insn_idx;
4161 first_idx = st->first_insn_idx;
4164 /* if we still have requested precise regs or slots, we missed
4165 * something (e.g., stack access through non-r10 register), so
4166 * fallback to marking all precise
4168 if (!bt_empty(bt)) {
4169 mark_all_scalars_precise(env, env->cur_state);
4176 int mark_chain_precision(struct bpf_verifier_env *env, int regno)
4178 return __mark_chain_precision(env, regno);
4181 /* mark_chain_precision_batch() assumes that env->bt is set in the caller to
4182 * desired reg and stack masks across all relevant frames
4184 static int mark_chain_precision_batch(struct bpf_verifier_env *env)
4186 return __mark_chain_precision(env, -1);
4189 static bool is_spillable_regtype(enum bpf_reg_type type)
4191 switch (base_type(type)) {
4192 case PTR_TO_MAP_VALUE:
4196 case PTR_TO_PACKET_META:
4197 case PTR_TO_PACKET_END:
4198 case PTR_TO_FLOW_KEYS:
4199 case CONST_PTR_TO_MAP:
4201 case PTR_TO_SOCK_COMMON:
4202 case PTR_TO_TCP_SOCK:
4203 case PTR_TO_XDP_SOCK:
4208 case PTR_TO_MAP_KEY:
4215 /* Does this register contain a constant zero? */
4216 static bool register_is_null(struct bpf_reg_state *reg)
4218 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
4221 static bool register_is_const(struct bpf_reg_state *reg)
4223 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
4226 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
4228 return tnum_is_unknown(reg->var_off) &&
4229 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
4230 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
4231 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
4232 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
4235 static bool register_is_bounded(struct bpf_reg_state *reg)
4237 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
4240 static bool __is_pointer_value(bool allow_ptr_leaks,
4241 const struct bpf_reg_state *reg)
4243 if (allow_ptr_leaks)
4246 return reg->type != SCALAR_VALUE;
4249 /* Copy src state preserving dst->parent and dst->live fields */
4250 static void copy_register_state(struct bpf_reg_state *dst, const struct bpf_reg_state *src)
4252 struct bpf_reg_state *parent = dst->parent;
4253 enum bpf_reg_liveness live = dst->live;
4256 dst->parent = parent;
4260 static void save_register_state(struct bpf_func_state *state,
4261 int spi, struct bpf_reg_state *reg,
4266 copy_register_state(&state->stack[spi].spilled_ptr, reg);
4267 if (size == BPF_REG_SIZE)
4268 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
4270 for (i = BPF_REG_SIZE; i > BPF_REG_SIZE - size; i--)
4271 state->stack[spi].slot_type[i - 1] = STACK_SPILL;
4273 /* size < 8 bytes spill */
4275 scrub_spilled_slot(&state->stack[spi].slot_type[i - 1]);
4278 static bool is_bpf_st_mem(struct bpf_insn *insn)
4280 return BPF_CLASS(insn->code) == BPF_ST && BPF_MODE(insn->code) == BPF_MEM;
4283 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
4284 * stack boundary and alignment are checked in check_mem_access()
4286 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
4287 /* stack frame we're writing to */
4288 struct bpf_func_state *state,
4289 int off, int size, int value_regno,
4292 struct bpf_func_state *cur; /* state of the current function */
4293 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
4294 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
4295 struct bpf_reg_state *reg = NULL;
4296 u32 dst_reg = insn->dst_reg;
4298 err = grow_stack_state(state, round_up(slot + 1, BPF_REG_SIZE));
4301 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
4302 * so it's aligned access and [off, off + size) are within stack limits
4304 if (!env->allow_ptr_leaks &&
4305 state->stack[spi].slot_type[0] == STACK_SPILL &&
4306 size != BPF_REG_SIZE) {
4307 verbose(env, "attempt to corrupt spilled pointer on stack\n");
4311 cur = env->cur_state->frame[env->cur_state->curframe];
4312 if (value_regno >= 0)
4313 reg = &cur->regs[value_regno];
4314 if (!env->bypass_spec_v4) {
4315 bool sanitize = reg && is_spillable_regtype(reg->type);
4317 for (i = 0; i < size; i++) {
4318 u8 type = state->stack[spi].slot_type[i];
4320 if (type != STACK_MISC && type != STACK_ZERO) {
4327 env->insn_aux_data[insn_idx].sanitize_stack_spill = true;
4330 err = destroy_if_dynptr_stack_slot(env, state, spi);
4334 mark_stack_slot_scratched(env, spi);
4335 if (reg && !(off % BPF_REG_SIZE) && register_is_bounded(reg) &&
4336 !register_is_null(reg) && env->bpf_capable) {
4337 if (dst_reg != BPF_REG_FP) {
4338 /* The backtracking logic can only recognize explicit
4339 * stack slot address like [fp - 8]. Other spill of
4340 * scalar via different register has to be conservative.
4341 * Backtrack from here and mark all registers as precise
4342 * that contributed into 'reg' being a constant.
4344 err = mark_chain_precision(env, value_regno);
4348 save_register_state(state, spi, reg, size);
4349 /* Break the relation on a narrowing spill. */
4350 if (fls64(reg->umax_value) > BITS_PER_BYTE * size)
4351 state->stack[spi].spilled_ptr.id = 0;
4352 } else if (!reg && !(off % BPF_REG_SIZE) && is_bpf_st_mem(insn) &&
4353 insn->imm != 0 && env->bpf_capable) {
4354 struct bpf_reg_state fake_reg = {};
4356 __mark_reg_known(&fake_reg, (u32)insn->imm);
4357 fake_reg.type = SCALAR_VALUE;
4358 save_register_state(state, spi, &fake_reg, size);
4359 } else if (reg && is_spillable_regtype(reg->type)) {
4360 /* register containing pointer is being spilled into stack */
4361 if (size != BPF_REG_SIZE) {
4362 verbose_linfo(env, insn_idx, "; ");
4363 verbose(env, "invalid size of register spill\n");
4366 if (state != cur && reg->type == PTR_TO_STACK) {
4367 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
4370 save_register_state(state, spi, reg, size);
4372 u8 type = STACK_MISC;
4374 /* regular write of data into stack destroys any spilled ptr */
4375 state->stack[spi].spilled_ptr.type = NOT_INIT;
4376 /* Mark slots as STACK_MISC if they belonged to spilled ptr/dynptr/iter. */
4377 if (is_stack_slot_special(&state->stack[spi]))
4378 for (i = 0; i < BPF_REG_SIZE; i++)
4379 scrub_spilled_slot(&state->stack[spi].slot_type[i]);
4381 /* only mark the slot as written if all 8 bytes were written
4382 * otherwise read propagation may incorrectly stop too soon
4383 * when stack slots are partially written.
4384 * This heuristic means that read propagation will be
4385 * conservative, since it will add reg_live_read marks
4386 * to stack slots all the way to first state when programs
4387 * writes+reads less than 8 bytes
4389 if (size == BPF_REG_SIZE)
4390 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
4392 /* when we zero initialize stack slots mark them as such */
4393 if ((reg && register_is_null(reg)) ||
4394 (!reg && is_bpf_st_mem(insn) && insn->imm == 0)) {
4395 /* backtracking doesn't work for STACK_ZERO yet. */
4396 err = mark_chain_precision(env, value_regno);
4402 /* Mark slots affected by this stack write. */
4403 for (i = 0; i < size; i++)
4404 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
4410 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
4411 * known to contain a variable offset.
4412 * This function checks whether the write is permitted and conservatively
4413 * tracks the effects of the write, considering that each stack slot in the
4414 * dynamic range is potentially written to.
4416 * 'off' includes 'regno->off'.
4417 * 'value_regno' can be -1, meaning that an unknown value is being written to
4420 * Spilled pointers in range are not marked as written because we don't know
4421 * what's going to be actually written. This means that read propagation for
4422 * future reads cannot be terminated by this write.
4424 * For privileged programs, uninitialized stack slots are considered
4425 * initialized by this write (even though we don't know exactly what offsets
4426 * are going to be written to). The idea is that we don't want the verifier to
4427 * reject future reads that access slots written to through variable offsets.
4429 static int check_stack_write_var_off(struct bpf_verifier_env *env,
4430 /* func where register points to */
4431 struct bpf_func_state *state,
4432 int ptr_regno, int off, int size,
4433 int value_regno, int insn_idx)
4435 struct bpf_func_state *cur; /* state of the current function */
4436 int min_off, max_off;
4438 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
4439 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
4440 bool writing_zero = false;
4441 /* set if the fact that we're writing a zero is used to let any
4442 * stack slots remain STACK_ZERO
4444 bool zero_used = false;
4446 cur = env->cur_state->frame[env->cur_state->curframe];
4447 ptr_reg = &cur->regs[ptr_regno];
4448 min_off = ptr_reg->smin_value + off;
4449 max_off = ptr_reg->smax_value + off + size;
4450 if (value_regno >= 0)
4451 value_reg = &cur->regs[value_regno];
4452 if ((value_reg && register_is_null(value_reg)) ||
4453 (!value_reg && is_bpf_st_mem(insn) && insn->imm == 0))
4454 writing_zero = true;
4456 err = grow_stack_state(state, round_up(-min_off, BPF_REG_SIZE));
4460 for (i = min_off; i < max_off; i++) {
4464 err = destroy_if_dynptr_stack_slot(env, state, spi);
4469 /* Variable offset writes destroy any spilled pointers in range. */
4470 for (i = min_off; i < max_off; i++) {
4471 u8 new_type, *stype;
4475 spi = slot / BPF_REG_SIZE;
4476 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
4477 mark_stack_slot_scratched(env, spi);
4479 if (!env->allow_ptr_leaks && *stype != STACK_MISC && *stype != STACK_ZERO) {
4480 /* Reject the write if range we may write to has not
4481 * been initialized beforehand. If we didn't reject
4482 * here, the ptr status would be erased below (even
4483 * though not all slots are actually overwritten),
4484 * possibly opening the door to leaks.
4486 * We do however catch STACK_INVALID case below, and
4487 * only allow reading possibly uninitialized memory
4488 * later for CAP_PERFMON, as the write may not happen to
4491 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
4496 /* Erase all spilled pointers. */
4497 state->stack[spi].spilled_ptr.type = NOT_INIT;
4499 /* Update the slot type. */
4500 new_type = STACK_MISC;
4501 if (writing_zero && *stype == STACK_ZERO) {
4502 new_type = STACK_ZERO;
4505 /* If the slot is STACK_INVALID, we check whether it's OK to
4506 * pretend that it will be initialized by this write. The slot
4507 * might not actually be written to, and so if we mark it as
4508 * initialized future reads might leak uninitialized memory.
4509 * For privileged programs, we will accept such reads to slots
4510 * that may or may not be written because, if we're reject
4511 * them, the error would be too confusing.
4513 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
4514 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
4521 /* backtracking doesn't work for STACK_ZERO yet. */
4522 err = mark_chain_precision(env, value_regno);
4529 /* When register 'dst_regno' is assigned some values from stack[min_off,
4530 * max_off), we set the register's type according to the types of the
4531 * respective stack slots. If all the stack values are known to be zeros, then
4532 * so is the destination reg. Otherwise, the register is considered to be
4533 * SCALAR. This function does not deal with register filling; the caller must
4534 * ensure that all spilled registers in the stack range have been marked as
4537 static void mark_reg_stack_read(struct bpf_verifier_env *env,
4538 /* func where src register points to */
4539 struct bpf_func_state *ptr_state,
4540 int min_off, int max_off, int dst_regno)
4542 struct bpf_verifier_state *vstate = env->cur_state;
4543 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4548 for (i = min_off; i < max_off; i++) {
4550 spi = slot / BPF_REG_SIZE;
4551 mark_stack_slot_scratched(env, spi);
4552 stype = ptr_state->stack[spi].slot_type;
4553 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
4557 if (zeros == max_off - min_off) {
4558 /* any access_size read into register is zero extended,
4559 * so the whole register == const_zero
4561 __mark_reg_const_zero(&state->regs[dst_regno]);
4562 /* backtracking doesn't support STACK_ZERO yet,
4563 * so mark it precise here, so that later
4564 * backtracking can stop here.
4565 * Backtracking may not need this if this register
4566 * doesn't participate in pointer adjustment.
4567 * Forward propagation of precise flag is not
4568 * necessary either. This mark is only to stop
4569 * backtracking. Any register that contributed
4570 * to const 0 was marked precise before spill.
4572 state->regs[dst_regno].precise = true;
4574 /* have read misc data from the stack */
4575 mark_reg_unknown(env, state->regs, dst_regno);
4577 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4580 /* Read the stack at 'off' and put the results into the register indicated by
4581 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
4584 * 'dst_regno' can be -1, meaning that the read value is not going to a
4587 * The access is assumed to be within the current stack bounds.
4589 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
4590 /* func where src register points to */
4591 struct bpf_func_state *reg_state,
4592 int off, int size, int dst_regno)
4594 struct bpf_verifier_state *vstate = env->cur_state;
4595 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4596 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
4597 struct bpf_reg_state *reg;
4600 stype = reg_state->stack[spi].slot_type;
4601 reg = ®_state->stack[spi].spilled_ptr;
4603 mark_stack_slot_scratched(env, spi);
4605 if (is_spilled_reg(®_state->stack[spi])) {
4608 for (i = BPF_REG_SIZE - 1; i > 0 && stype[i - 1] == STACK_SPILL; i--)
4611 if (size != BPF_REG_SIZE || spill_size != BPF_REG_SIZE) {
4612 if (reg->type != SCALAR_VALUE) {
4613 verbose_linfo(env, env->insn_idx, "; ");
4614 verbose(env, "invalid size of register fill\n");
4618 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4622 if (!(off % BPF_REG_SIZE) && size == spill_size) {
4623 /* The earlier check_reg_arg() has decided the
4624 * subreg_def for this insn. Save it first.
4626 s32 subreg_def = state->regs[dst_regno].subreg_def;
4628 copy_register_state(&state->regs[dst_regno], reg);
4629 state->regs[dst_regno].subreg_def = subreg_def;
4631 for (i = 0; i < size; i++) {
4632 type = stype[(slot - i) % BPF_REG_SIZE];
4633 if (type == STACK_SPILL)
4635 if (type == STACK_MISC)
4637 if (type == STACK_INVALID && env->allow_uninit_stack)
4639 verbose(env, "invalid read from stack off %d+%d size %d\n",
4643 mark_reg_unknown(env, state->regs, dst_regno);
4645 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4649 if (dst_regno >= 0) {
4650 /* restore register state from stack */
4651 copy_register_state(&state->regs[dst_regno], reg);
4652 /* mark reg as written since spilled pointer state likely
4653 * has its liveness marks cleared by is_state_visited()
4654 * which resets stack/reg liveness for state transitions
4656 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4657 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
4658 /* If dst_regno==-1, the caller is asking us whether
4659 * it is acceptable to use this value as a SCALAR_VALUE
4661 * We must not allow unprivileged callers to do that
4662 * with spilled pointers.
4664 verbose(env, "leaking pointer from stack off %d\n",
4668 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4670 for (i = 0; i < size; i++) {
4671 type = stype[(slot - i) % BPF_REG_SIZE];
4672 if (type == STACK_MISC)
4674 if (type == STACK_ZERO)
4676 if (type == STACK_INVALID && env->allow_uninit_stack)
4678 verbose(env, "invalid read from stack off %d+%d size %d\n",
4682 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4684 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
4689 enum bpf_access_src {
4690 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
4691 ACCESS_HELPER = 2, /* the access is performed by a helper */
4694 static int check_stack_range_initialized(struct bpf_verifier_env *env,
4695 int regno, int off, int access_size,
4696 bool zero_size_allowed,
4697 enum bpf_access_src type,
4698 struct bpf_call_arg_meta *meta);
4700 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
4702 return cur_regs(env) + regno;
4705 /* Read the stack at 'ptr_regno + off' and put the result into the register
4707 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
4708 * but not its variable offset.
4709 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
4711 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
4712 * filling registers (i.e. reads of spilled register cannot be detected when
4713 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
4714 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
4715 * offset; for a fixed offset check_stack_read_fixed_off should be used
4718 static int check_stack_read_var_off(struct bpf_verifier_env *env,
4719 int ptr_regno, int off, int size, int dst_regno)
4721 /* The state of the source register. */
4722 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
4723 struct bpf_func_state *ptr_state = func(env, reg);
4725 int min_off, max_off;
4727 /* Note that we pass a NULL meta, so raw access will not be permitted.
4729 err = check_stack_range_initialized(env, ptr_regno, off, size,
4730 false, ACCESS_DIRECT, NULL);
4734 min_off = reg->smin_value + off;
4735 max_off = reg->smax_value + off;
4736 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
4740 /* check_stack_read dispatches to check_stack_read_fixed_off or
4741 * check_stack_read_var_off.
4743 * The caller must ensure that the offset falls within the allocated stack
4746 * 'dst_regno' is a register which will receive the value from the stack. It
4747 * can be -1, meaning that the read value is not going to a register.
4749 static int check_stack_read(struct bpf_verifier_env *env,
4750 int ptr_regno, int off, int size,
4753 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
4754 struct bpf_func_state *state = func(env, reg);
4756 /* Some accesses are only permitted with a static offset. */
4757 bool var_off = !tnum_is_const(reg->var_off);
4759 /* The offset is required to be static when reads don't go to a
4760 * register, in order to not leak pointers (see
4761 * check_stack_read_fixed_off).
4763 if (dst_regno < 0 && var_off) {
4766 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4767 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
4771 /* Variable offset is prohibited for unprivileged mode for simplicity
4772 * since it requires corresponding support in Spectre masking for stack
4773 * ALU. See also retrieve_ptr_limit(). The check in
4774 * check_stack_access_for_ptr_arithmetic() called by
4775 * adjust_ptr_min_max_vals() prevents users from creating stack pointers
4776 * with variable offsets, therefore no check is required here. Further,
4777 * just checking it here would be insufficient as speculative stack
4778 * writes could still lead to unsafe speculative behaviour.
4781 off += reg->var_off.value;
4782 err = check_stack_read_fixed_off(env, state, off, size,
4785 /* Variable offset stack reads need more conservative handling
4786 * than fixed offset ones. Note that dst_regno >= 0 on this
4789 err = check_stack_read_var_off(env, ptr_regno, off, size,
4796 /* check_stack_write dispatches to check_stack_write_fixed_off or
4797 * check_stack_write_var_off.
4799 * 'ptr_regno' is the register used as a pointer into the stack.
4800 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
4801 * 'value_regno' is the register whose value we're writing to the stack. It can
4802 * be -1, meaning that we're not writing from a register.
4804 * The caller must ensure that the offset falls within the maximum stack size.
4806 static int check_stack_write(struct bpf_verifier_env *env,
4807 int ptr_regno, int off, int size,
4808 int value_regno, int insn_idx)
4810 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
4811 struct bpf_func_state *state = func(env, reg);
4814 if (tnum_is_const(reg->var_off)) {
4815 off += reg->var_off.value;
4816 err = check_stack_write_fixed_off(env, state, off, size,
4817 value_regno, insn_idx);
4819 /* Variable offset stack reads need more conservative handling
4820 * than fixed offset ones.
4822 err = check_stack_write_var_off(env, state,
4823 ptr_regno, off, size,
4824 value_regno, insn_idx);
4829 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
4830 int off, int size, enum bpf_access_type type)
4832 struct bpf_reg_state *regs = cur_regs(env);
4833 struct bpf_map *map = regs[regno].map_ptr;
4834 u32 cap = bpf_map_flags_to_cap(map);
4836 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
4837 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
4838 map->value_size, off, size);
4842 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
4843 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
4844 map->value_size, off, size);
4851 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
4852 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
4853 int off, int size, u32 mem_size,
4854 bool zero_size_allowed)
4856 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
4857 struct bpf_reg_state *reg;
4859 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
4862 reg = &cur_regs(env)[regno];
4863 switch (reg->type) {
4864 case PTR_TO_MAP_KEY:
4865 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
4866 mem_size, off, size);
4868 case PTR_TO_MAP_VALUE:
4869 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
4870 mem_size, off, size);
4873 case PTR_TO_PACKET_META:
4874 case PTR_TO_PACKET_END:
4875 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
4876 off, size, regno, reg->id, off, mem_size);
4880 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
4881 mem_size, off, size);
4887 /* check read/write into a memory region with possible variable offset */
4888 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
4889 int off, int size, u32 mem_size,
4890 bool zero_size_allowed)
4892 struct bpf_verifier_state *vstate = env->cur_state;
4893 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4894 struct bpf_reg_state *reg = &state->regs[regno];
4897 /* We may have adjusted the register pointing to memory region, so we
4898 * need to try adding each of min_value and max_value to off
4899 * to make sure our theoretical access will be safe.
4901 * The minimum value is only important with signed
4902 * comparisons where we can't assume the floor of a
4903 * value is 0. If we are using signed variables for our
4904 * index'es we need to make sure that whatever we use
4905 * will have a set floor within our range.
4907 if (reg->smin_value < 0 &&
4908 (reg->smin_value == S64_MIN ||
4909 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
4910 reg->smin_value + off < 0)) {
4911 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
4915 err = __check_mem_access(env, regno, reg->smin_value + off, size,
4916 mem_size, zero_size_allowed);
4918 verbose(env, "R%d min value is outside of the allowed memory range\n",
4923 /* If we haven't set a max value then we need to bail since we can't be
4924 * sure we won't do bad things.
4925 * If reg->umax_value + off could overflow, treat that as unbounded too.
4927 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
4928 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
4932 err = __check_mem_access(env, regno, reg->umax_value + off, size,
4933 mem_size, zero_size_allowed);
4935 verbose(env, "R%d max value is outside of the allowed memory range\n",
4943 static int __check_ptr_off_reg(struct bpf_verifier_env *env,
4944 const struct bpf_reg_state *reg, int regno,
4947 /* Access to this pointer-typed register or passing it to a helper
4948 * is only allowed in its original, unmodified form.
4952 verbose(env, "negative offset %s ptr R%d off=%d disallowed\n",
4953 reg_type_str(env, reg->type), regno, reg->off);
4957 if (!fixed_off_ok && reg->off) {
4958 verbose(env, "dereference of modified %s ptr R%d off=%d disallowed\n",
4959 reg_type_str(env, reg->type), regno, reg->off);
4963 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4966 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4967 verbose(env, "variable %s access var_off=%s disallowed\n",
4968 reg_type_str(env, reg->type), tn_buf);
4975 int check_ptr_off_reg(struct bpf_verifier_env *env,
4976 const struct bpf_reg_state *reg, int regno)
4978 return __check_ptr_off_reg(env, reg, regno, false);
4981 static int map_kptr_match_type(struct bpf_verifier_env *env,
4982 struct btf_field *kptr_field,
4983 struct bpf_reg_state *reg, u32 regno)
4985 const char *targ_name = btf_type_name(kptr_field->kptr.btf, kptr_field->kptr.btf_id);
4986 int perm_flags = PTR_MAYBE_NULL | PTR_TRUSTED | MEM_RCU;
4987 const char *reg_name = "";
4989 /* Only unreferenced case accepts untrusted pointers */
4990 if (kptr_field->type == BPF_KPTR_UNREF)
4991 perm_flags |= PTR_UNTRUSTED;
4993 if (base_type(reg->type) != PTR_TO_BTF_ID || (type_flag(reg->type) & ~perm_flags))
4996 if (!btf_is_kernel(reg->btf)) {
4997 verbose(env, "R%d must point to kernel BTF\n", regno);
5000 /* We need to verify reg->type and reg->btf, before accessing reg->btf */
5001 reg_name = btf_type_name(reg->btf, reg->btf_id);
5003 /* For ref_ptr case, release function check should ensure we get one
5004 * referenced PTR_TO_BTF_ID, and that its fixed offset is 0. For the
5005 * normal store of unreferenced kptr, we must ensure var_off is zero.
5006 * Since ref_ptr cannot be accessed directly by BPF insns, checks for
5007 * reg->off and reg->ref_obj_id are not needed here.
5009 if (__check_ptr_off_reg(env, reg, regno, true))
5012 /* A full type match is needed, as BTF can be vmlinux or module BTF, and
5013 * we also need to take into account the reg->off.
5015 * We want to support cases like:
5023 * v = func(); // PTR_TO_BTF_ID
5024 * val->foo = v; // reg->off is zero, btf and btf_id match type
5025 * val->bar = &v->br; // reg->off is still zero, but we need to retry with
5026 * // first member type of struct after comparison fails
5027 * val->baz = &v->bz; // reg->off is non-zero, so struct needs to be walked
5030 * In the kptr_ref case, check_func_arg_reg_off already ensures reg->off
5031 * is zero. We must also ensure that btf_struct_ids_match does not walk
5032 * the struct to match type against first member of struct, i.e. reject
5033 * second case from above. Hence, when type is BPF_KPTR_REF, we set
5034 * strict mode to true for type match.
5036 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
5037 kptr_field->kptr.btf, kptr_field->kptr.btf_id,
5038 kptr_field->type == BPF_KPTR_REF))
5042 verbose(env, "invalid kptr access, R%d type=%s%s ", regno,
5043 reg_type_str(env, reg->type), reg_name);
5044 verbose(env, "expected=%s%s", reg_type_str(env, PTR_TO_BTF_ID), targ_name);
5045 if (kptr_field->type == BPF_KPTR_UNREF)
5046 verbose(env, " or %s%s\n", reg_type_str(env, PTR_TO_BTF_ID | PTR_UNTRUSTED),
5053 /* The non-sleepable programs and sleepable programs with explicit bpf_rcu_read_lock()
5054 * can dereference RCU protected pointers and result is PTR_TRUSTED.
5056 static bool in_rcu_cs(struct bpf_verifier_env *env)
5058 return env->cur_state->active_rcu_lock || !env->prog->aux->sleepable;
5061 /* Once GCC supports btf_type_tag the following mechanism will be replaced with tag check */
5062 BTF_SET_START(rcu_protected_types)
5063 BTF_ID(struct, prog_test_ref_kfunc)
5064 BTF_ID(struct, cgroup)
5065 BTF_ID(struct, bpf_cpumask)
5066 BTF_ID(struct, task_struct)
5067 BTF_SET_END(rcu_protected_types)
5069 static bool rcu_protected_object(const struct btf *btf, u32 btf_id)
5071 if (!btf_is_kernel(btf))
5073 return btf_id_set_contains(&rcu_protected_types, btf_id);
5076 static bool rcu_safe_kptr(const struct btf_field *field)
5078 const struct btf_field_kptr *kptr = &field->kptr;
5080 return field->type == BPF_KPTR_REF && rcu_protected_object(kptr->btf, kptr->btf_id);
5083 static int check_map_kptr_access(struct bpf_verifier_env *env, u32 regno,
5084 int value_regno, int insn_idx,
5085 struct btf_field *kptr_field)
5087 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
5088 int class = BPF_CLASS(insn->code);
5089 struct bpf_reg_state *val_reg;
5091 /* Things we already checked for in check_map_access and caller:
5092 * - Reject cases where variable offset may touch kptr
5093 * - size of access (must be BPF_DW)
5094 * - tnum_is_const(reg->var_off)
5095 * - kptr_field->offset == off + reg->var_off.value
5097 /* Only BPF_[LDX,STX,ST] | BPF_MEM | BPF_DW is supported */
5098 if (BPF_MODE(insn->code) != BPF_MEM) {
5099 verbose(env, "kptr in map can only be accessed using BPF_MEM instruction mode\n");
5103 /* We only allow loading referenced kptr, since it will be marked as
5104 * untrusted, similar to unreferenced kptr.
5106 if (class != BPF_LDX && kptr_field->type == BPF_KPTR_REF) {
5107 verbose(env, "store to referenced kptr disallowed\n");
5111 if (class == BPF_LDX) {
5112 val_reg = reg_state(env, value_regno);
5113 /* We can simply mark the value_regno receiving the pointer
5114 * value from map as PTR_TO_BTF_ID, with the correct type.
5116 mark_btf_ld_reg(env, cur_regs(env), value_regno, PTR_TO_BTF_ID, kptr_field->kptr.btf,
5117 kptr_field->kptr.btf_id,
5118 rcu_safe_kptr(kptr_field) && in_rcu_cs(env) ?
5119 PTR_MAYBE_NULL | MEM_RCU :
5120 PTR_MAYBE_NULL | PTR_UNTRUSTED);
5121 /* For mark_ptr_or_null_reg */
5122 val_reg->id = ++env->id_gen;
5123 } else if (class == BPF_STX) {
5124 val_reg = reg_state(env, value_regno);
5125 if (!register_is_null(val_reg) &&
5126 map_kptr_match_type(env, kptr_field, val_reg, value_regno))
5128 } else if (class == BPF_ST) {
5130 verbose(env, "BPF_ST imm must be 0 when storing to kptr at off=%u\n",
5131 kptr_field->offset);
5135 verbose(env, "kptr in map can only be accessed using BPF_LDX/BPF_STX/BPF_ST\n");
5141 /* check read/write into a map element with possible variable offset */
5142 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
5143 int off, int size, bool zero_size_allowed,
5144 enum bpf_access_src src)
5146 struct bpf_verifier_state *vstate = env->cur_state;
5147 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5148 struct bpf_reg_state *reg = &state->regs[regno];
5149 struct bpf_map *map = reg->map_ptr;
5150 struct btf_record *rec;
5153 err = check_mem_region_access(env, regno, off, size, map->value_size,
5158 if (IS_ERR_OR_NULL(map->record))
5161 for (i = 0; i < rec->cnt; i++) {
5162 struct btf_field *field = &rec->fields[i];
5163 u32 p = field->offset;
5165 /* If any part of a field can be touched by load/store, reject
5166 * this program. To check that [x1, x2) overlaps with [y1, y2),
5167 * it is sufficient to check x1 < y2 && y1 < x2.
5169 if (reg->smin_value + off < p + btf_field_type_size(field->type) &&
5170 p < reg->umax_value + off + size) {
5171 switch (field->type) {
5172 case BPF_KPTR_UNREF:
5174 if (src != ACCESS_DIRECT) {
5175 verbose(env, "kptr cannot be accessed indirectly by helper\n");
5178 if (!tnum_is_const(reg->var_off)) {
5179 verbose(env, "kptr access cannot have variable offset\n");
5182 if (p != off + reg->var_off.value) {
5183 verbose(env, "kptr access misaligned expected=%u off=%llu\n",
5184 p, off + reg->var_off.value);
5187 if (size != bpf_size_to_bytes(BPF_DW)) {
5188 verbose(env, "kptr access size must be BPF_DW\n");
5193 verbose(env, "%s cannot be accessed directly by load/store\n",
5194 btf_field_type_name(field->type));
5202 #define MAX_PACKET_OFF 0xffff
5204 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
5205 const struct bpf_call_arg_meta *meta,
5206 enum bpf_access_type t)
5208 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
5210 switch (prog_type) {
5211 /* Program types only with direct read access go here! */
5212 case BPF_PROG_TYPE_LWT_IN:
5213 case BPF_PROG_TYPE_LWT_OUT:
5214 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
5215 case BPF_PROG_TYPE_SK_REUSEPORT:
5216 case BPF_PROG_TYPE_FLOW_DISSECTOR:
5217 case BPF_PROG_TYPE_CGROUP_SKB:
5222 /* Program types with direct read + write access go here! */
5223 case BPF_PROG_TYPE_SCHED_CLS:
5224 case BPF_PROG_TYPE_SCHED_ACT:
5225 case BPF_PROG_TYPE_XDP:
5226 case BPF_PROG_TYPE_LWT_XMIT:
5227 case BPF_PROG_TYPE_SK_SKB:
5228 case BPF_PROG_TYPE_SK_MSG:
5230 return meta->pkt_access;
5232 env->seen_direct_write = true;
5235 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
5237 env->seen_direct_write = true;
5246 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
5247 int size, bool zero_size_allowed)
5249 struct bpf_reg_state *regs = cur_regs(env);
5250 struct bpf_reg_state *reg = ®s[regno];
5253 /* We may have added a variable offset to the packet pointer; but any
5254 * reg->range we have comes after that. We are only checking the fixed
5258 /* We don't allow negative numbers, because we aren't tracking enough
5259 * detail to prove they're safe.
5261 if (reg->smin_value < 0) {
5262 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
5267 err = reg->range < 0 ? -EINVAL :
5268 __check_mem_access(env, regno, off, size, reg->range,
5271 verbose(env, "R%d offset is outside of the packet\n", regno);
5275 /* __check_mem_access has made sure "off + size - 1" is within u16.
5276 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
5277 * otherwise find_good_pkt_pointers would have refused to set range info
5278 * that __check_mem_access would have rejected this pkt access.
5279 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
5281 env->prog->aux->max_pkt_offset =
5282 max_t(u32, env->prog->aux->max_pkt_offset,
5283 off + reg->umax_value + size - 1);
5288 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
5289 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
5290 enum bpf_access_type t, enum bpf_reg_type *reg_type,
5291 struct btf **btf, u32 *btf_id)
5293 struct bpf_insn_access_aux info = {
5294 .reg_type = *reg_type,
5298 if (env->ops->is_valid_access &&
5299 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
5300 /* A non zero info.ctx_field_size indicates that this field is a
5301 * candidate for later verifier transformation to load the whole
5302 * field and then apply a mask when accessed with a narrower
5303 * access than actual ctx access size. A zero info.ctx_field_size
5304 * will only allow for whole field access and rejects any other
5305 * type of narrower access.
5307 *reg_type = info.reg_type;
5309 if (base_type(*reg_type) == PTR_TO_BTF_ID) {
5311 *btf_id = info.btf_id;
5313 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
5315 /* remember the offset of last byte accessed in ctx */
5316 if (env->prog->aux->max_ctx_offset < off + size)
5317 env->prog->aux->max_ctx_offset = off + size;
5321 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
5325 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
5328 if (size < 0 || off < 0 ||
5329 (u64)off + size > sizeof(struct bpf_flow_keys)) {
5330 verbose(env, "invalid access to flow keys off=%d size=%d\n",
5337 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
5338 u32 regno, int off, int size,
5339 enum bpf_access_type t)
5341 struct bpf_reg_state *regs = cur_regs(env);
5342 struct bpf_reg_state *reg = ®s[regno];
5343 struct bpf_insn_access_aux info = {};
5346 if (reg->smin_value < 0) {
5347 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
5352 switch (reg->type) {
5353 case PTR_TO_SOCK_COMMON:
5354 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
5357 valid = bpf_sock_is_valid_access(off, size, t, &info);
5359 case PTR_TO_TCP_SOCK:
5360 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
5362 case PTR_TO_XDP_SOCK:
5363 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
5371 env->insn_aux_data[insn_idx].ctx_field_size =
5372 info.ctx_field_size;
5376 verbose(env, "R%d invalid %s access off=%d size=%d\n",
5377 regno, reg_type_str(env, reg->type), off, size);
5382 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
5384 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
5387 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
5389 const struct bpf_reg_state *reg = reg_state(env, regno);
5391 return reg->type == PTR_TO_CTX;
5394 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
5396 const struct bpf_reg_state *reg = reg_state(env, regno);
5398 return type_is_sk_pointer(reg->type);
5401 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
5403 const struct bpf_reg_state *reg = reg_state(env, regno);
5405 return type_is_pkt_pointer(reg->type);
5408 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
5410 const struct bpf_reg_state *reg = reg_state(env, regno);
5412 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
5413 return reg->type == PTR_TO_FLOW_KEYS;
5416 static u32 *reg2btf_ids[__BPF_REG_TYPE_MAX] = {
5418 [PTR_TO_SOCKET] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK],
5419 [PTR_TO_SOCK_COMMON] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
5420 [PTR_TO_TCP_SOCK] = &btf_sock_ids[BTF_SOCK_TYPE_TCP],
5422 [CONST_PTR_TO_MAP] = btf_bpf_map_id,
5425 static bool is_trusted_reg(const struct bpf_reg_state *reg)
5427 /* A referenced register is always trusted. */
5428 if (reg->ref_obj_id)
5431 /* Types listed in the reg2btf_ids are always trusted */
5432 if (reg2btf_ids[base_type(reg->type)])
5435 /* If a register is not referenced, it is trusted if it has the
5436 * MEM_ALLOC or PTR_TRUSTED type modifiers, and no others. Some of the
5437 * other type modifiers may be safe, but we elect to take an opt-in
5438 * approach here as some (e.g. PTR_UNTRUSTED and PTR_MAYBE_NULL) are
5441 * Eventually, we should make PTR_TRUSTED the single source of truth
5442 * for whether a register is trusted.
5444 return type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS &&
5445 !bpf_type_has_unsafe_modifiers(reg->type);
5448 static bool is_rcu_reg(const struct bpf_reg_state *reg)
5450 return reg->type & MEM_RCU;
5453 static void clear_trusted_flags(enum bpf_type_flag *flag)
5455 *flag &= ~(BPF_REG_TRUSTED_MODIFIERS | MEM_RCU);
5458 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
5459 const struct bpf_reg_state *reg,
5460 int off, int size, bool strict)
5462 struct tnum reg_off;
5465 /* Byte size accesses are always allowed. */
5466 if (!strict || size == 1)
5469 /* For platforms that do not have a Kconfig enabling
5470 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
5471 * NET_IP_ALIGN is universally set to '2'. And on platforms
5472 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
5473 * to this code only in strict mode where we want to emulate
5474 * the NET_IP_ALIGN==2 checking. Therefore use an
5475 * unconditional IP align value of '2'.
5479 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
5480 if (!tnum_is_aligned(reg_off, size)) {
5483 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5485 "misaligned packet access off %d+%s+%d+%d size %d\n",
5486 ip_align, tn_buf, reg->off, off, size);
5493 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
5494 const struct bpf_reg_state *reg,
5495 const char *pointer_desc,
5496 int off, int size, bool strict)
5498 struct tnum reg_off;
5500 /* Byte size accesses are always allowed. */
5501 if (!strict || size == 1)
5504 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
5505 if (!tnum_is_aligned(reg_off, size)) {
5508 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5509 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
5510 pointer_desc, tn_buf, reg->off, off, size);
5517 static int check_ptr_alignment(struct bpf_verifier_env *env,
5518 const struct bpf_reg_state *reg, int off,
5519 int size, bool strict_alignment_once)
5521 bool strict = env->strict_alignment || strict_alignment_once;
5522 const char *pointer_desc = "";
5524 switch (reg->type) {
5526 case PTR_TO_PACKET_META:
5527 /* Special case, because of NET_IP_ALIGN. Given metadata sits
5528 * right in front, treat it the very same way.
5530 return check_pkt_ptr_alignment(env, reg, off, size, strict);
5531 case PTR_TO_FLOW_KEYS:
5532 pointer_desc = "flow keys ";
5534 case PTR_TO_MAP_KEY:
5535 pointer_desc = "key ";
5537 case PTR_TO_MAP_VALUE:
5538 pointer_desc = "value ";
5541 pointer_desc = "context ";
5544 pointer_desc = "stack ";
5545 /* The stack spill tracking logic in check_stack_write_fixed_off()
5546 * and check_stack_read_fixed_off() relies on stack accesses being
5552 pointer_desc = "sock ";
5554 case PTR_TO_SOCK_COMMON:
5555 pointer_desc = "sock_common ";
5557 case PTR_TO_TCP_SOCK:
5558 pointer_desc = "tcp_sock ";
5560 case PTR_TO_XDP_SOCK:
5561 pointer_desc = "xdp_sock ";
5566 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
5570 static int update_stack_depth(struct bpf_verifier_env *env,
5571 const struct bpf_func_state *func,
5574 u16 stack = env->subprog_info[func->subprogno].stack_depth;
5579 /* update known max for given subprogram */
5580 env->subprog_info[func->subprogno].stack_depth = -off;
5584 /* starting from main bpf function walk all instructions of the function
5585 * and recursively walk all callees that given function can call.
5586 * Ignore jump and exit insns.
5587 * Since recursion is prevented by check_cfg() this algorithm
5588 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
5590 static int check_max_stack_depth(struct bpf_verifier_env *env)
5592 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
5593 struct bpf_subprog_info *subprog = env->subprog_info;
5594 struct bpf_insn *insn = env->prog->insnsi;
5595 bool tail_call_reachable = false;
5596 int ret_insn[MAX_CALL_FRAMES];
5597 int ret_prog[MAX_CALL_FRAMES];
5601 /* protect against potential stack overflow that might happen when
5602 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
5603 * depth for such case down to 256 so that the worst case scenario
5604 * would result in 8k stack size (32 which is tailcall limit * 256 =
5607 * To get the idea what might happen, see an example:
5608 * func1 -> sub rsp, 128
5609 * subfunc1 -> sub rsp, 256
5610 * tailcall1 -> add rsp, 256
5611 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
5612 * subfunc2 -> sub rsp, 64
5613 * subfunc22 -> sub rsp, 128
5614 * tailcall2 -> add rsp, 128
5615 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
5617 * tailcall will unwind the current stack frame but it will not get rid
5618 * of caller's stack as shown on the example above.
5620 if (idx && subprog[idx].has_tail_call && depth >= 256) {
5622 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
5626 /* round up to 32-bytes, since this is granularity
5627 * of interpreter stack size
5629 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
5630 if (depth > MAX_BPF_STACK) {
5631 verbose(env, "combined stack size of %d calls is %d. Too large\n",
5636 subprog_end = subprog[idx + 1].start;
5637 for (; i < subprog_end; i++) {
5640 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
5642 /* remember insn and function to return to */
5643 ret_insn[frame] = i + 1;
5644 ret_prog[frame] = idx;
5646 /* find the callee */
5647 next_insn = i + insn[i].imm + 1;
5648 idx = find_subprog(env, next_insn);
5650 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5654 if (subprog[idx].is_async_cb) {
5655 if (subprog[idx].has_tail_call) {
5656 verbose(env, "verifier bug. subprog has tail_call and async cb\n");
5659 /* async callbacks don't increase bpf prog stack size unless called directly */
5660 if (!bpf_pseudo_call(insn + i))
5665 if (subprog[idx].has_tail_call)
5666 tail_call_reachable = true;
5669 if (frame >= MAX_CALL_FRAMES) {
5670 verbose(env, "the call stack of %d frames is too deep !\n",
5676 /* if tail call got detected across bpf2bpf calls then mark each of the
5677 * currently present subprog frames as tail call reachable subprogs;
5678 * this info will be utilized by JIT so that we will be preserving the
5679 * tail call counter throughout bpf2bpf calls combined with tailcalls
5681 if (tail_call_reachable)
5682 for (j = 0; j < frame; j++)
5683 subprog[ret_prog[j]].tail_call_reachable = true;
5684 if (subprog[0].tail_call_reachable)
5685 env->prog->aux->tail_call_reachable = true;
5687 /* end of for() loop means the last insn of the 'subprog'
5688 * was reached. Doesn't matter whether it was JA or EXIT
5692 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
5694 i = ret_insn[frame];
5695 idx = ret_prog[frame];
5699 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
5700 static int get_callee_stack_depth(struct bpf_verifier_env *env,
5701 const struct bpf_insn *insn, int idx)
5703 int start = idx + insn->imm + 1, subprog;
5705 subprog = find_subprog(env, start);
5707 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5711 return env->subprog_info[subprog].stack_depth;
5715 static int __check_buffer_access(struct bpf_verifier_env *env,
5716 const char *buf_info,
5717 const struct bpf_reg_state *reg,
5718 int regno, int off, int size)
5722 "R%d invalid %s buffer access: off=%d, size=%d\n",
5723 regno, buf_info, off, size);
5726 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
5729 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5731 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
5732 regno, off, tn_buf);
5739 static int check_tp_buffer_access(struct bpf_verifier_env *env,
5740 const struct bpf_reg_state *reg,
5741 int regno, int off, int size)
5745 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
5749 if (off + size > env->prog->aux->max_tp_access)
5750 env->prog->aux->max_tp_access = off + size;
5755 static int check_buffer_access(struct bpf_verifier_env *env,
5756 const struct bpf_reg_state *reg,
5757 int regno, int off, int size,
5758 bool zero_size_allowed,
5761 const char *buf_info = type_is_rdonly_mem(reg->type) ? "rdonly" : "rdwr";
5764 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
5768 if (off + size > *max_access)
5769 *max_access = off + size;
5774 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
5775 static void zext_32_to_64(struct bpf_reg_state *reg)
5777 reg->var_off = tnum_subreg(reg->var_off);
5778 __reg_assign_32_into_64(reg);
5781 /* truncate register to smaller size (in bytes)
5782 * must be called with size < BPF_REG_SIZE
5784 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
5788 /* clear high bits in bit representation */
5789 reg->var_off = tnum_cast(reg->var_off, size);
5791 /* fix arithmetic bounds */
5792 mask = ((u64)1 << (size * 8)) - 1;
5793 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
5794 reg->umin_value &= mask;
5795 reg->umax_value &= mask;
5797 reg->umin_value = 0;
5798 reg->umax_value = mask;
5800 reg->smin_value = reg->umin_value;
5801 reg->smax_value = reg->umax_value;
5803 /* If size is smaller than 32bit register the 32bit register
5804 * values are also truncated so we push 64-bit bounds into
5805 * 32-bit bounds. Above were truncated < 32-bits already.
5809 __reg_combine_64_into_32(reg);
5812 static bool bpf_map_is_rdonly(const struct bpf_map *map)
5814 /* A map is considered read-only if the following condition are true:
5816 * 1) BPF program side cannot change any of the map content. The
5817 * BPF_F_RDONLY_PROG flag is throughout the lifetime of a map
5818 * and was set at map creation time.
5819 * 2) The map value(s) have been initialized from user space by a
5820 * loader and then "frozen", such that no new map update/delete
5821 * operations from syscall side are possible for the rest of
5822 * the map's lifetime from that point onwards.
5823 * 3) Any parallel/pending map update/delete operations from syscall
5824 * side have been completed. Only after that point, it's safe to
5825 * assume that map value(s) are immutable.
5827 return (map->map_flags & BPF_F_RDONLY_PROG) &&
5828 READ_ONCE(map->frozen) &&
5829 !bpf_map_write_active(map);
5832 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
5838 err = map->ops->map_direct_value_addr(map, &addr, off);
5841 ptr = (void *)(long)addr + off;
5845 *val = (u64)*(u8 *)ptr;
5848 *val = (u64)*(u16 *)ptr;
5851 *val = (u64)*(u32 *)ptr;
5862 #define BTF_TYPE_SAFE_RCU(__type) __PASTE(__type, __safe_rcu)
5863 #define BTF_TYPE_SAFE_RCU_OR_NULL(__type) __PASTE(__type, __safe_rcu_or_null)
5864 #define BTF_TYPE_SAFE_TRUSTED(__type) __PASTE(__type, __safe_trusted)
5867 * Allow list few fields as RCU trusted or full trusted.
5868 * This logic doesn't allow mix tagging and will be removed once GCC supports
5872 /* RCU trusted: these fields are trusted in RCU CS and never NULL */
5873 BTF_TYPE_SAFE_RCU(struct task_struct) {
5874 const cpumask_t *cpus_ptr;
5875 struct css_set __rcu *cgroups;
5876 struct task_struct __rcu *real_parent;
5877 struct task_struct *group_leader;
5880 BTF_TYPE_SAFE_RCU(struct cgroup) {
5881 /* cgrp->kn is always accessible as documented in kernel/cgroup/cgroup.c */
5882 struct kernfs_node *kn;
5885 BTF_TYPE_SAFE_RCU(struct css_set) {
5886 struct cgroup *dfl_cgrp;
5889 /* RCU trusted: these fields are trusted in RCU CS and can be NULL */
5890 BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct) {
5891 struct file __rcu *exe_file;
5894 /* skb->sk, req->sk are not RCU protected, but we mark them as such
5895 * because bpf prog accessible sockets are SOCK_RCU_FREE.
5897 BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff) {
5901 BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock) {
5905 /* full trusted: these fields are trusted even outside of RCU CS and never NULL */
5906 BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta) {
5907 struct seq_file *seq;
5910 BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task) {
5911 struct bpf_iter_meta *meta;
5912 struct task_struct *task;
5915 BTF_TYPE_SAFE_TRUSTED(struct linux_binprm) {
5919 BTF_TYPE_SAFE_TRUSTED(struct file) {
5920 struct inode *f_inode;
5923 BTF_TYPE_SAFE_TRUSTED(struct dentry) {
5924 /* no negative dentry-s in places where bpf can see it */
5925 struct inode *d_inode;
5928 BTF_TYPE_SAFE_TRUSTED(struct socket) {
5932 static bool type_is_rcu(struct bpf_verifier_env *env,
5933 struct bpf_reg_state *reg,
5934 const char *field_name, u32 btf_id)
5936 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct task_struct));
5937 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct cgroup));
5938 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct css_set));
5940 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_rcu");
5943 static bool type_is_rcu_or_null(struct bpf_verifier_env *env,
5944 struct bpf_reg_state *reg,
5945 const char *field_name, u32 btf_id)
5947 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct));
5948 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff));
5949 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock));
5951 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_rcu_or_null");
5954 static bool type_is_trusted(struct bpf_verifier_env *env,
5955 struct bpf_reg_state *reg,
5956 const char *field_name, u32 btf_id)
5958 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta));
5959 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task));
5960 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct linux_binprm));
5961 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct file));
5962 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct dentry));
5963 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct socket));
5965 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_trusted");
5968 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
5969 struct bpf_reg_state *regs,
5970 int regno, int off, int size,
5971 enum bpf_access_type atype,
5974 struct bpf_reg_state *reg = regs + regno;
5975 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
5976 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
5977 const char *field_name = NULL;
5978 enum bpf_type_flag flag = 0;
5982 if (!env->allow_ptr_leaks) {
5984 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
5988 if (!env->prog->gpl_compatible && btf_is_kernel(reg->btf)) {
5990 "Cannot access kernel 'struct %s' from non-GPL compatible program\n",
5996 "R%d is ptr_%s invalid negative access: off=%d\n",
6000 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
6003 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6005 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
6006 regno, tname, off, tn_buf);
6010 if (reg->type & MEM_USER) {
6012 "R%d is ptr_%s access user memory: off=%d\n",
6017 if (reg->type & MEM_PERCPU) {
6019 "R%d is ptr_%s access percpu memory: off=%d\n",
6024 if (env->ops->btf_struct_access && !type_is_alloc(reg->type) && atype == BPF_WRITE) {
6025 if (!btf_is_kernel(reg->btf)) {
6026 verbose(env, "verifier internal error: reg->btf must be kernel btf\n");
6029 ret = env->ops->btf_struct_access(&env->log, reg, off, size);
6031 /* Writes are permitted with default btf_struct_access for
6032 * program allocated objects (which always have ref_obj_id > 0),
6033 * but not for untrusted PTR_TO_BTF_ID | MEM_ALLOC.
6035 if (atype != BPF_READ && !type_is_ptr_alloc_obj(reg->type)) {
6036 verbose(env, "only read is supported\n");
6040 if (type_is_alloc(reg->type) && !type_is_non_owning_ref(reg->type) &&
6042 verbose(env, "verifier internal error: ref_obj_id for allocated object must be non-zero\n");
6046 ret = btf_struct_access(&env->log, reg, off, size, atype, &btf_id, &flag, &field_name);
6052 if (ret != PTR_TO_BTF_ID) {
6055 } else if (type_flag(reg->type) & PTR_UNTRUSTED) {
6056 /* If this is an untrusted pointer, all pointers formed by walking it
6057 * also inherit the untrusted flag.
6059 flag = PTR_UNTRUSTED;
6061 } else if (is_trusted_reg(reg) || is_rcu_reg(reg)) {
6062 /* By default any pointer obtained from walking a trusted pointer is no
6063 * longer trusted, unless the field being accessed has explicitly been
6064 * marked as inheriting its parent's state of trust (either full or RCU).
6066 * 'cgroups' pointer is untrusted if task->cgroups dereference
6067 * happened in a sleepable program outside of bpf_rcu_read_lock()
6068 * section. In a non-sleepable program it's trusted while in RCU CS (aka MEM_RCU).
6069 * Note bpf_rcu_read_unlock() converts MEM_RCU pointers to PTR_UNTRUSTED.
6071 * A regular RCU-protected pointer with __rcu tag can also be deemed
6072 * trusted if we are in an RCU CS. Such pointer can be NULL.
6074 if (type_is_trusted(env, reg, field_name, btf_id)) {
6075 flag |= PTR_TRUSTED;
6076 } else if (in_rcu_cs(env) && !type_may_be_null(reg->type)) {
6077 if (type_is_rcu(env, reg, field_name, btf_id)) {
6078 /* ignore __rcu tag and mark it MEM_RCU */
6080 } else if (flag & MEM_RCU ||
6081 type_is_rcu_or_null(env, reg, field_name, btf_id)) {
6082 /* __rcu tagged pointers can be NULL */
6083 flag |= MEM_RCU | PTR_MAYBE_NULL;
6085 /* We always trust them */
6086 if (type_is_rcu_or_null(env, reg, field_name, btf_id) &&
6087 flag & PTR_UNTRUSTED)
6088 flag &= ~PTR_UNTRUSTED;
6089 } else if (flag & (MEM_PERCPU | MEM_USER)) {
6092 /* walking unknown pointers yields old deprecated PTR_TO_BTF_ID */
6093 clear_trusted_flags(&flag);
6097 * If not in RCU CS or MEM_RCU pointer can be NULL then
6098 * aggressively mark as untrusted otherwise such
6099 * pointers will be plain PTR_TO_BTF_ID without flags
6100 * and will be allowed to be passed into helpers for
6103 flag = PTR_UNTRUSTED;
6106 /* Old compat. Deprecated */
6107 clear_trusted_flags(&flag);
6110 if (atype == BPF_READ && value_regno >= 0)
6111 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id, flag);
6116 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
6117 struct bpf_reg_state *regs,
6118 int regno, int off, int size,
6119 enum bpf_access_type atype,
6122 struct bpf_reg_state *reg = regs + regno;
6123 struct bpf_map *map = reg->map_ptr;
6124 struct bpf_reg_state map_reg;
6125 enum bpf_type_flag flag = 0;
6126 const struct btf_type *t;
6132 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
6136 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
6137 verbose(env, "map_ptr access not supported for map type %d\n",
6142 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
6143 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
6145 if (!env->allow_ptr_leaks) {
6147 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
6153 verbose(env, "R%d is %s invalid negative access: off=%d\n",
6158 if (atype != BPF_READ) {
6159 verbose(env, "only read from %s is supported\n", tname);
6163 /* Simulate access to a PTR_TO_BTF_ID */
6164 memset(&map_reg, 0, sizeof(map_reg));
6165 mark_btf_ld_reg(env, &map_reg, 0, PTR_TO_BTF_ID, btf_vmlinux, *map->ops->map_btf_id, 0);
6166 ret = btf_struct_access(&env->log, &map_reg, off, size, atype, &btf_id, &flag, NULL);
6170 if (value_regno >= 0)
6171 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id, flag);
6176 /* Check that the stack access at the given offset is within bounds. The
6177 * maximum valid offset is -1.
6179 * The minimum valid offset is -MAX_BPF_STACK for writes, and
6180 * -state->allocated_stack for reads.
6182 static int check_stack_slot_within_bounds(int off,
6183 struct bpf_func_state *state,
6184 enum bpf_access_type t)
6189 min_valid_off = -MAX_BPF_STACK;
6191 min_valid_off = -state->allocated_stack;
6193 if (off < min_valid_off || off > -1)
6198 /* Check that the stack access at 'regno + off' falls within the maximum stack
6201 * 'off' includes `regno->offset`, but not its dynamic part (if any).
6203 static int check_stack_access_within_bounds(
6204 struct bpf_verifier_env *env,
6205 int regno, int off, int access_size,
6206 enum bpf_access_src src, enum bpf_access_type type)
6208 struct bpf_reg_state *regs = cur_regs(env);
6209 struct bpf_reg_state *reg = regs + regno;
6210 struct bpf_func_state *state = func(env, reg);
6211 int min_off, max_off;
6215 if (src == ACCESS_HELPER)
6216 /* We don't know if helpers are reading or writing (or both). */
6217 err_extra = " indirect access to";
6218 else if (type == BPF_READ)
6219 err_extra = " read from";
6221 err_extra = " write to";
6223 if (tnum_is_const(reg->var_off)) {
6224 min_off = reg->var_off.value + off;
6225 if (access_size > 0)
6226 max_off = min_off + access_size - 1;
6230 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
6231 reg->smin_value <= -BPF_MAX_VAR_OFF) {
6232 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
6236 min_off = reg->smin_value + off;
6237 if (access_size > 0)
6238 max_off = reg->smax_value + off + access_size - 1;
6243 err = check_stack_slot_within_bounds(min_off, state, type);
6245 err = check_stack_slot_within_bounds(max_off, state, type);
6248 if (tnum_is_const(reg->var_off)) {
6249 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
6250 err_extra, regno, off, access_size);
6254 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6255 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
6256 err_extra, regno, tn_buf, access_size);
6262 /* check whether memory at (regno + off) is accessible for t = (read | write)
6263 * if t==write, value_regno is a register which value is stored into memory
6264 * if t==read, value_regno is a register which will receive the value from memory
6265 * if t==write && value_regno==-1, some unknown value is stored into memory
6266 * if t==read && value_regno==-1, don't care what we read from memory
6268 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
6269 int off, int bpf_size, enum bpf_access_type t,
6270 int value_regno, bool strict_alignment_once)
6272 struct bpf_reg_state *regs = cur_regs(env);
6273 struct bpf_reg_state *reg = regs + regno;
6274 struct bpf_func_state *state;
6277 size = bpf_size_to_bytes(bpf_size);
6281 /* alignment checks will add in reg->off themselves */
6282 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
6286 /* for access checks, reg->off is just part of off */
6289 if (reg->type == PTR_TO_MAP_KEY) {
6290 if (t == BPF_WRITE) {
6291 verbose(env, "write to change key R%d not allowed\n", regno);
6295 err = check_mem_region_access(env, regno, off, size,
6296 reg->map_ptr->key_size, false);
6299 if (value_regno >= 0)
6300 mark_reg_unknown(env, regs, value_regno);
6301 } else if (reg->type == PTR_TO_MAP_VALUE) {
6302 struct btf_field *kptr_field = NULL;
6304 if (t == BPF_WRITE && value_regno >= 0 &&
6305 is_pointer_value(env, value_regno)) {
6306 verbose(env, "R%d leaks addr into map\n", value_regno);
6309 err = check_map_access_type(env, regno, off, size, t);
6312 err = check_map_access(env, regno, off, size, false, ACCESS_DIRECT);
6315 if (tnum_is_const(reg->var_off))
6316 kptr_field = btf_record_find(reg->map_ptr->record,
6317 off + reg->var_off.value, BPF_KPTR);
6319 err = check_map_kptr_access(env, regno, value_regno, insn_idx, kptr_field);
6320 } else if (t == BPF_READ && value_regno >= 0) {
6321 struct bpf_map *map = reg->map_ptr;
6323 /* if map is read-only, track its contents as scalars */
6324 if (tnum_is_const(reg->var_off) &&
6325 bpf_map_is_rdonly(map) &&
6326 map->ops->map_direct_value_addr) {
6327 int map_off = off + reg->var_off.value;
6330 err = bpf_map_direct_read(map, map_off, size,
6335 regs[value_regno].type = SCALAR_VALUE;
6336 __mark_reg_known(®s[value_regno], val);
6338 mark_reg_unknown(env, regs, value_regno);
6341 } else if (base_type(reg->type) == PTR_TO_MEM) {
6342 bool rdonly_mem = type_is_rdonly_mem(reg->type);
6344 if (type_may_be_null(reg->type)) {
6345 verbose(env, "R%d invalid mem access '%s'\n", regno,
6346 reg_type_str(env, reg->type));
6350 if (t == BPF_WRITE && rdonly_mem) {
6351 verbose(env, "R%d cannot write into %s\n",
6352 regno, reg_type_str(env, reg->type));
6356 if (t == BPF_WRITE && value_regno >= 0 &&
6357 is_pointer_value(env, value_regno)) {
6358 verbose(env, "R%d leaks addr into mem\n", value_regno);
6362 err = check_mem_region_access(env, regno, off, size,
6363 reg->mem_size, false);
6364 if (!err && value_regno >= 0 && (t == BPF_READ || rdonly_mem))
6365 mark_reg_unknown(env, regs, value_regno);
6366 } else if (reg->type == PTR_TO_CTX) {
6367 enum bpf_reg_type reg_type = SCALAR_VALUE;
6368 struct btf *btf = NULL;
6371 if (t == BPF_WRITE && value_regno >= 0 &&
6372 is_pointer_value(env, value_regno)) {
6373 verbose(env, "R%d leaks addr into ctx\n", value_regno);
6377 err = check_ptr_off_reg(env, reg, regno);
6381 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf,
6384 verbose_linfo(env, insn_idx, "; ");
6385 if (!err && t == BPF_READ && value_regno >= 0) {
6386 /* ctx access returns either a scalar, or a
6387 * PTR_TO_PACKET[_META,_END]. In the latter
6388 * case, we know the offset is zero.
6390 if (reg_type == SCALAR_VALUE) {
6391 mark_reg_unknown(env, regs, value_regno);
6393 mark_reg_known_zero(env, regs,
6395 if (type_may_be_null(reg_type))
6396 regs[value_regno].id = ++env->id_gen;
6397 /* A load of ctx field could have different
6398 * actual load size with the one encoded in the
6399 * insn. When the dst is PTR, it is for sure not
6402 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
6403 if (base_type(reg_type) == PTR_TO_BTF_ID) {
6404 regs[value_regno].btf = btf;
6405 regs[value_regno].btf_id = btf_id;
6408 regs[value_regno].type = reg_type;
6411 } else if (reg->type == PTR_TO_STACK) {
6412 /* Basic bounds checks. */
6413 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
6417 state = func(env, reg);
6418 err = update_stack_depth(env, state, off);
6423 err = check_stack_read(env, regno, off, size,
6426 err = check_stack_write(env, regno, off, size,
6427 value_regno, insn_idx);
6428 } else if (reg_is_pkt_pointer(reg)) {
6429 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
6430 verbose(env, "cannot write into packet\n");
6433 if (t == BPF_WRITE && value_regno >= 0 &&
6434 is_pointer_value(env, value_regno)) {
6435 verbose(env, "R%d leaks addr into packet\n",
6439 err = check_packet_access(env, regno, off, size, false);
6440 if (!err && t == BPF_READ && value_regno >= 0)
6441 mark_reg_unknown(env, regs, value_regno);
6442 } else if (reg->type == PTR_TO_FLOW_KEYS) {
6443 if (t == BPF_WRITE && value_regno >= 0 &&
6444 is_pointer_value(env, value_regno)) {
6445 verbose(env, "R%d leaks addr into flow keys\n",
6450 err = check_flow_keys_access(env, off, size);
6451 if (!err && t == BPF_READ && value_regno >= 0)
6452 mark_reg_unknown(env, regs, value_regno);
6453 } else if (type_is_sk_pointer(reg->type)) {
6454 if (t == BPF_WRITE) {
6455 verbose(env, "R%d cannot write into %s\n",
6456 regno, reg_type_str(env, reg->type));
6459 err = check_sock_access(env, insn_idx, regno, off, size, t);
6460 if (!err && value_regno >= 0)
6461 mark_reg_unknown(env, regs, value_regno);
6462 } else if (reg->type == PTR_TO_TP_BUFFER) {
6463 err = check_tp_buffer_access(env, reg, regno, off, size);
6464 if (!err && t == BPF_READ && value_regno >= 0)
6465 mark_reg_unknown(env, regs, value_regno);
6466 } else if (base_type(reg->type) == PTR_TO_BTF_ID &&
6467 !type_may_be_null(reg->type)) {
6468 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
6470 } else if (reg->type == CONST_PTR_TO_MAP) {
6471 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
6473 } else if (base_type(reg->type) == PTR_TO_BUF) {
6474 bool rdonly_mem = type_is_rdonly_mem(reg->type);
6478 if (t == BPF_WRITE) {
6479 verbose(env, "R%d cannot write into %s\n",
6480 regno, reg_type_str(env, reg->type));
6483 max_access = &env->prog->aux->max_rdonly_access;
6485 max_access = &env->prog->aux->max_rdwr_access;
6488 err = check_buffer_access(env, reg, regno, off, size, false,
6491 if (!err && value_regno >= 0 && (rdonly_mem || t == BPF_READ))
6492 mark_reg_unknown(env, regs, value_regno);
6494 verbose(env, "R%d invalid mem access '%s'\n", regno,
6495 reg_type_str(env, reg->type));
6499 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
6500 regs[value_regno].type == SCALAR_VALUE) {
6501 /* b/h/w load zero-extends, mark upper bits as known 0 */
6502 coerce_reg_to_size(®s[value_regno], size);
6507 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
6512 switch (insn->imm) {
6514 case BPF_ADD | BPF_FETCH:
6516 case BPF_AND | BPF_FETCH:
6518 case BPF_OR | BPF_FETCH:
6520 case BPF_XOR | BPF_FETCH:
6525 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
6529 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
6530 verbose(env, "invalid atomic operand size\n");
6534 /* check src1 operand */
6535 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6539 /* check src2 operand */
6540 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6544 if (insn->imm == BPF_CMPXCHG) {
6545 /* Check comparison of R0 with memory location */
6546 const u32 aux_reg = BPF_REG_0;
6548 err = check_reg_arg(env, aux_reg, SRC_OP);
6552 if (is_pointer_value(env, aux_reg)) {
6553 verbose(env, "R%d leaks addr into mem\n", aux_reg);
6558 if (is_pointer_value(env, insn->src_reg)) {
6559 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
6563 if (is_ctx_reg(env, insn->dst_reg) ||
6564 is_pkt_reg(env, insn->dst_reg) ||
6565 is_flow_key_reg(env, insn->dst_reg) ||
6566 is_sk_reg(env, insn->dst_reg)) {
6567 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
6569 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
6573 if (insn->imm & BPF_FETCH) {
6574 if (insn->imm == BPF_CMPXCHG)
6575 load_reg = BPF_REG_0;
6577 load_reg = insn->src_reg;
6579 /* check and record load of old value */
6580 err = check_reg_arg(env, load_reg, DST_OP);
6584 /* This instruction accesses a memory location but doesn't
6585 * actually load it into a register.
6590 /* Check whether we can read the memory, with second call for fetch
6591 * case to simulate the register fill.
6593 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
6594 BPF_SIZE(insn->code), BPF_READ, -1, true);
6595 if (!err && load_reg >= 0)
6596 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
6597 BPF_SIZE(insn->code), BPF_READ, load_reg,
6602 /* Check whether we can write into the same memory. */
6603 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
6604 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
6611 /* When register 'regno' is used to read the stack (either directly or through
6612 * a helper function) make sure that it's within stack boundary and, depending
6613 * on the access type, that all elements of the stack are initialized.
6615 * 'off' includes 'regno->off', but not its dynamic part (if any).
6617 * All registers that have been spilled on the stack in the slots within the
6618 * read offsets are marked as read.
6620 static int check_stack_range_initialized(
6621 struct bpf_verifier_env *env, int regno, int off,
6622 int access_size, bool zero_size_allowed,
6623 enum bpf_access_src type, struct bpf_call_arg_meta *meta)
6625 struct bpf_reg_state *reg = reg_state(env, regno);
6626 struct bpf_func_state *state = func(env, reg);
6627 int err, min_off, max_off, i, j, slot, spi;
6628 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
6629 enum bpf_access_type bounds_check_type;
6630 /* Some accesses can write anything into the stack, others are
6633 bool clobber = false;
6635 if (access_size == 0 && !zero_size_allowed) {
6636 verbose(env, "invalid zero-sized read\n");
6640 if (type == ACCESS_HELPER) {
6641 /* The bounds checks for writes are more permissive than for
6642 * reads. However, if raw_mode is not set, we'll do extra
6645 bounds_check_type = BPF_WRITE;
6648 bounds_check_type = BPF_READ;
6650 err = check_stack_access_within_bounds(env, regno, off, access_size,
6651 type, bounds_check_type);
6656 if (tnum_is_const(reg->var_off)) {
6657 min_off = max_off = reg->var_off.value + off;
6659 /* Variable offset is prohibited for unprivileged mode for
6660 * simplicity since it requires corresponding support in
6661 * Spectre masking for stack ALU.
6662 * See also retrieve_ptr_limit().
6664 if (!env->bypass_spec_v1) {
6667 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6668 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
6669 regno, err_extra, tn_buf);
6672 /* Only initialized buffer on stack is allowed to be accessed
6673 * with variable offset. With uninitialized buffer it's hard to
6674 * guarantee that whole memory is marked as initialized on
6675 * helper return since specific bounds are unknown what may
6676 * cause uninitialized stack leaking.
6678 if (meta && meta->raw_mode)
6681 min_off = reg->smin_value + off;
6682 max_off = reg->smax_value + off;
6685 if (meta && meta->raw_mode) {
6686 /* Ensure we won't be overwriting dynptrs when simulating byte
6687 * by byte access in check_helper_call using meta.access_size.
6688 * This would be a problem if we have a helper in the future
6691 * helper(uninit_mem, len, dynptr)
6693 * Now, uninint_mem may overlap with dynptr pointer. Hence, it
6694 * may end up writing to dynptr itself when touching memory from
6695 * arg 1. This can be relaxed on a case by case basis for known
6696 * safe cases, but reject due to the possibilitiy of aliasing by
6699 for (i = min_off; i < max_off + access_size; i++) {
6700 int stack_off = -i - 1;
6703 /* raw_mode may write past allocated_stack */
6704 if (state->allocated_stack <= stack_off)
6706 if (state->stack[spi].slot_type[stack_off % BPF_REG_SIZE] == STACK_DYNPTR) {
6707 verbose(env, "potential write to dynptr at off=%d disallowed\n", i);
6711 meta->access_size = access_size;
6712 meta->regno = regno;
6716 for (i = min_off; i < max_off + access_size; i++) {
6720 spi = slot / BPF_REG_SIZE;
6721 if (state->allocated_stack <= slot)
6723 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
6724 if (*stype == STACK_MISC)
6726 if ((*stype == STACK_ZERO) ||
6727 (*stype == STACK_INVALID && env->allow_uninit_stack)) {
6729 /* helper can write anything into the stack */
6730 *stype = STACK_MISC;
6735 if (is_spilled_reg(&state->stack[spi]) &&
6736 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
6737 env->allow_ptr_leaks)) {
6739 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
6740 for (j = 0; j < BPF_REG_SIZE; j++)
6741 scrub_spilled_slot(&state->stack[spi].slot_type[j]);
6747 if (tnum_is_const(reg->var_off)) {
6748 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
6749 err_extra, regno, min_off, i - min_off, access_size);
6753 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6754 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
6755 err_extra, regno, tn_buf, i - min_off, access_size);
6759 /* reading any byte out of 8-byte 'spill_slot' will cause
6760 * the whole slot to be marked as 'read'
6762 mark_reg_read(env, &state->stack[spi].spilled_ptr,
6763 state->stack[spi].spilled_ptr.parent,
6765 /* We do not set REG_LIVE_WRITTEN for stack slot, as we can not
6766 * be sure that whether stack slot is written to or not. Hence,
6767 * we must still conservatively propagate reads upwards even if
6768 * helper may write to the entire memory range.
6771 return update_stack_depth(env, state, min_off);
6774 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
6775 int access_size, bool zero_size_allowed,
6776 struct bpf_call_arg_meta *meta)
6778 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
6781 switch (base_type(reg->type)) {
6783 case PTR_TO_PACKET_META:
6784 return check_packet_access(env, regno, reg->off, access_size,
6786 case PTR_TO_MAP_KEY:
6787 if (meta && meta->raw_mode) {
6788 verbose(env, "R%d cannot write into %s\n", regno,
6789 reg_type_str(env, reg->type));
6792 return check_mem_region_access(env, regno, reg->off, access_size,
6793 reg->map_ptr->key_size, false);
6794 case PTR_TO_MAP_VALUE:
6795 if (check_map_access_type(env, regno, reg->off, access_size,
6796 meta && meta->raw_mode ? BPF_WRITE :
6799 return check_map_access(env, regno, reg->off, access_size,
6800 zero_size_allowed, ACCESS_HELPER);
6802 if (type_is_rdonly_mem(reg->type)) {
6803 if (meta && meta->raw_mode) {
6804 verbose(env, "R%d cannot write into %s\n", regno,
6805 reg_type_str(env, reg->type));
6809 return check_mem_region_access(env, regno, reg->off,
6810 access_size, reg->mem_size,
6813 if (type_is_rdonly_mem(reg->type)) {
6814 if (meta && meta->raw_mode) {
6815 verbose(env, "R%d cannot write into %s\n", regno,
6816 reg_type_str(env, reg->type));
6820 max_access = &env->prog->aux->max_rdonly_access;
6822 max_access = &env->prog->aux->max_rdwr_access;
6824 return check_buffer_access(env, reg, regno, reg->off,
6825 access_size, zero_size_allowed,
6828 return check_stack_range_initialized(
6830 regno, reg->off, access_size,
6831 zero_size_allowed, ACCESS_HELPER, meta);
6833 return check_ptr_to_btf_access(env, regs, regno, reg->off,
6834 access_size, BPF_READ, -1);
6836 /* in case the function doesn't know how to access the context,
6837 * (because we are in a program of type SYSCALL for example), we
6838 * can not statically check its size.
6839 * Dynamically check it now.
6841 if (!env->ops->convert_ctx_access) {
6842 enum bpf_access_type atype = meta && meta->raw_mode ? BPF_WRITE : BPF_READ;
6843 int offset = access_size - 1;
6845 /* Allow zero-byte read from PTR_TO_CTX */
6846 if (access_size == 0)
6847 return zero_size_allowed ? 0 : -EACCES;
6849 return check_mem_access(env, env->insn_idx, regno, offset, BPF_B,
6854 default: /* scalar_value or invalid ptr */
6855 /* Allow zero-byte read from NULL, regardless of pointer type */
6856 if (zero_size_allowed && access_size == 0 &&
6857 register_is_null(reg))
6860 verbose(env, "R%d type=%s ", regno,
6861 reg_type_str(env, reg->type));
6862 verbose(env, "expected=%s\n", reg_type_str(env, PTR_TO_STACK));
6867 static int check_mem_size_reg(struct bpf_verifier_env *env,
6868 struct bpf_reg_state *reg, u32 regno,
6869 bool zero_size_allowed,
6870 struct bpf_call_arg_meta *meta)
6874 /* This is used to refine r0 return value bounds for helpers
6875 * that enforce this value as an upper bound on return values.
6876 * See do_refine_retval_range() for helpers that can refine
6877 * the return value. C type of helper is u32 so we pull register
6878 * bound from umax_value however, if negative verifier errors
6879 * out. Only upper bounds can be learned because retval is an
6880 * int type and negative retvals are allowed.
6882 meta->msize_max_value = reg->umax_value;
6884 /* The register is SCALAR_VALUE; the access check
6885 * happens using its boundaries.
6887 if (!tnum_is_const(reg->var_off))
6888 /* For unprivileged variable accesses, disable raw
6889 * mode so that the program is required to
6890 * initialize all the memory that the helper could
6891 * just partially fill up.
6895 if (reg->smin_value < 0) {
6896 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
6901 if (reg->umin_value == 0) {
6902 err = check_helper_mem_access(env, regno - 1, 0,
6909 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
6910 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
6914 err = check_helper_mem_access(env, regno - 1,
6916 zero_size_allowed, meta);
6918 err = mark_chain_precision(env, regno);
6922 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
6923 u32 regno, u32 mem_size)
6925 bool may_be_null = type_may_be_null(reg->type);
6926 struct bpf_reg_state saved_reg;
6927 struct bpf_call_arg_meta meta;
6930 if (register_is_null(reg))
6933 memset(&meta, 0, sizeof(meta));
6934 /* Assuming that the register contains a value check if the memory
6935 * access is safe. Temporarily save and restore the register's state as
6936 * the conversion shouldn't be visible to a caller.
6940 mark_ptr_not_null_reg(reg);
6943 err = check_helper_mem_access(env, regno, mem_size, true, &meta);
6944 /* Check access for BPF_WRITE */
6945 meta.raw_mode = true;
6946 err = err ?: check_helper_mem_access(env, regno, mem_size, true, &meta);
6954 static int check_kfunc_mem_size_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
6957 struct bpf_reg_state *mem_reg = &cur_regs(env)[regno - 1];
6958 bool may_be_null = type_may_be_null(mem_reg->type);
6959 struct bpf_reg_state saved_reg;
6960 struct bpf_call_arg_meta meta;
6963 WARN_ON_ONCE(regno < BPF_REG_2 || regno > BPF_REG_5);
6965 memset(&meta, 0, sizeof(meta));
6968 saved_reg = *mem_reg;
6969 mark_ptr_not_null_reg(mem_reg);
6972 err = check_mem_size_reg(env, reg, regno, true, &meta);
6973 /* Check access for BPF_WRITE */
6974 meta.raw_mode = true;
6975 err = err ?: check_mem_size_reg(env, reg, regno, true, &meta);
6978 *mem_reg = saved_reg;
6982 /* Implementation details:
6983 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL.
6984 * bpf_obj_new returns PTR_TO_BTF_ID | MEM_ALLOC | PTR_MAYBE_NULL.
6985 * Two bpf_map_lookups (even with the same key) will have different reg->id.
6986 * Two separate bpf_obj_new will also have different reg->id.
6987 * For traditional PTR_TO_MAP_VALUE or PTR_TO_BTF_ID | MEM_ALLOC, the verifier
6988 * clears reg->id after value_or_null->value transition, since the verifier only
6989 * cares about the range of access to valid map value pointer and doesn't care
6990 * about actual address of the map element.
6991 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
6992 * reg->id > 0 after value_or_null->value transition. By doing so
6993 * two bpf_map_lookups will be considered two different pointers that
6994 * point to different bpf_spin_locks. Likewise for pointers to allocated objects
6995 * returned from bpf_obj_new.
6996 * The verifier allows taking only one bpf_spin_lock at a time to avoid
6998 * Since only one bpf_spin_lock is allowed the checks are simpler than
6999 * reg_is_refcounted() logic. The verifier needs to remember only
7000 * one spin_lock instead of array of acquired_refs.
7001 * cur_state->active_lock remembers which map value element or allocated
7002 * object got locked and clears it after bpf_spin_unlock.
7004 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
7007 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7008 struct bpf_verifier_state *cur = env->cur_state;
7009 bool is_const = tnum_is_const(reg->var_off);
7010 u64 val = reg->var_off.value;
7011 struct bpf_map *map = NULL;
7012 struct btf *btf = NULL;
7013 struct btf_record *rec;
7017 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
7021 if (reg->type == PTR_TO_MAP_VALUE) {
7025 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
7033 rec = reg_btf_record(reg);
7034 if (!btf_record_has_field(rec, BPF_SPIN_LOCK)) {
7035 verbose(env, "%s '%s' has no valid bpf_spin_lock\n", map ? "map" : "local",
7036 map ? map->name : "kptr");
7039 if (rec->spin_lock_off != val + reg->off) {
7040 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock' that is at %d\n",
7041 val + reg->off, rec->spin_lock_off);
7045 if (cur->active_lock.ptr) {
7047 "Locking two bpf_spin_locks are not allowed\n");
7051 cur->active_lock.ptr = map;
7053 cur->active_lock.ptr = btf;
7054 cur->active_lock.id = reg->id;
7063 if (!cur->active_lock.ptr) {
7064 verbose(env, "bpf_spin_unlock without taking a lock\n");
7067 if (cur->active_lock.ptr != ptr ||
7068 cur->active_lock.id != reg->id) {
7069 verbose(env, "bpf_spin_unlock of different lock\n");
7073 invalidate_non_owning_refs(env);
7075 cur->active_lock.ptr = NULL;
7076 cur->active_lock.id = 0;
7081 static int process_timer_func(struct bpf_verifier_env *env, int regno,
7082 struct bpf_call_arg_meta *meta)
7084 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7085 bool is_const = tnum_is_const(reg->var_off);
7086 struct bpf_map *map = reg->map_ptr;
7087 u64 val = reg->var_off.value;
7091 "R%d doesn't have constant offset. bpf_timer has to be at the constant offset\n",
7096 verbose(env, "map '%s' has to have BTF in order to use bpf_timer\n",
7100 if (!btf_record_has_field(map->record, BPF_TIMER)) {
7101 verbose(env, "map '%s' has no valid bpf_timer\n", map->name);
7104 if (map->record->timer_off != val + reg->off) {
7105 verbose(env, "off %lld doesn't point to 'struct bpf_timer' that is at %d\n",
7106 val + reg->off, map->record->timer_off);
7109 if (meta->map_ptr) {
7110 verbose(env, "verifier bug. Two map pointers in a timer helper\n");
7113 meta->map_uid = reg->map_uid;
7114 meta->map_ptr = map;
7118 static int process_kptr_func(struct bpf_verifier_env *env, int regno,
7119 struct bpf_call_arg_meta *meta)
7121 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7122 struct bpf_map *map_ptr = reg->map_ptr;
7123 struct btf_field *kptr_field;
7126 if (!tnum_is_const(reg->var_off)) {
7128 "R%d doesn't have constant offset. kptr has to be at the constant offset\n",
7132 if (!map_ptr->btf) {
7133 verbose(env, "map '%s' has to have BTF in order to use bpf_kptr_xchg\n",
7137 if (!btf_record_has_field(map_ptr->record, BPF_KPTR)) {
7138 verbose(env, "map '%s' has no valid kptr\n", map_ptr->name);
7142 meta->map_ptr = map_ptr;
7143 kptr_off = reg->off + reg->var_off.value;
7144 kptr_field = btf_record_find(map_ptr->record, kptr_off, BPF_KPTR);
7146 verbose(env, "off=%d doesn't point to kptr\n", kptr_off);
7149 if (kptr_field->type != BPF_KPTR_REF) {
7150 verbose(env, "off=%d kptr isn't referenced kptr\n", kptr_off);
7153 meta->kptr_field = kptr_field;
7157 /* There are two register types representing a bpf_dynptr, one is PTR_TO_STACK
7158 * which points to a stack slot, and the other is CONST_PTR_TO_DYNPTR.
7160 * In both cases we deal with the first 8 bytes, but need to mark the next 8
7161 * bytes as STACK_DYNPTR in case of PTR_TO_STACK. In case of
7162 * CONST_PTR_TO_DYNPTR, we are guaranteed to get the beginning of the object.
7164 * Mutability of bpf_dynptr is at two levels, one is at the level of struct
7165 * bpf_dynptr itself, i.e. whether the helper is receiving a pointer to struct
7166 * bpf_dynptr or pointer to const struct bpf_dynptr. In the former case, it can
7167 * mutate the view of the dynptr and also possibly destroy it. In the latter
7168 * case, it cannot mutate the bpf_dynptr itself but it can still mutate the
7169 * memory that dynptr points to.
7171 * The verifier will keep track both levels of mutation (bpf_dynptr's in
7172 * reg->type and the memory's in reg->dynptr.type), but there is no support for
7173 * readonly dynptr view yet, hence only the first case is tracked and checked.
7175 * This is consistent with how C applies the const modifier to a struct object,
7176 * where the pointer itself inside bpf_dynptr becomes const but not what it
7179 * Helpers which do not mutate the bpf_dynptr set MEM_RDONLY in their argument
7180 * type, and declare it as 'const struct bpf_dynptr *' in their prototype.
7182 static int process_dynptr_func(struct bpf_verifier_env *env, int regno, int insn_idx,
7183 enum bpf_arg_type arg_type, int clone_ref_obj_id)
7185 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7188 /* MEM_UNINIT and MEM_RDONLY are exclusive, when applied to an
7189 * ARG_PTR_TO_DYNPTR (or ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_*):
7191 if ((arg_type & (MEM_UNINIT | MEM_RDONLY)) == (MEM_UNINIT | MEM_RDONLY)) {
7192 verbose(env, "verifier internal error: misconfigured dynptr helper type flags\n");
7196 /* MEM_UNINIT - Points to memory that is an appropriate candidate for
7197 * constructing a mutable bpf_dynptr object.
7199 * Currently, this is only possible with PTR_TO_STACK
7200 * pointing to a region of at least 16 bytes which doesn't
7201 * contain an existing bpf_dynptr.
7203 * MEM_RDONLY - Points to a initialized bpf_dynptr that will not be
7204 * mutated or destroyed. However, the memory it points to
7207 * None - Points to a initialized dynptr that can be mutated and
7208 * destroyed, including mutation of the memory it points
7211 if (arg_type & MEM_UNINIT) {
7214 if (!is_dynptr_reg_valid_uninit(env, reg)) {
7215 verbose(env, "Dynptr has to be an uninitialized dynptr\n");
7219 /* we write BPF_DW bits (8 bytes) at a time */
7220 for (i = 0; i < BPF_DYNPTR_SIZE; i += 8) {
7221 err = check_mem_access(env, insn_idx, regno,
7222 i, BPF_DW, BPF_WRITE, -1, false);
7227 err = mark_stack_slots_dynptr(env, reg, arg_type, insn_idx, clone_ref_obj_id);
7228 } else /* MEM_RDONLY and None case from above */ {
7229 /* For the reg->type == PTR_TO_STACK case, bpf_dynptr is never const */
7230 if (reg->type == CONST_PTR_TO_DYNPTR && !(arg_type & MEM_RDONLY)) {
7231 verbose(env, "cannot pass pointer to const bpf_dynptr, the helper mutates it\n");
7235 if (!is_dynptr_reg_valid_init(env, reg)) {
7237 "Expected an initialized dynptr as arg #%d\n",
7242 /* Fold modifiers (in this case, MEM_RDONLY) when checking expected type */
7243 if (!is_dynptr_type_expected(env, reg, arg_type & ~MEM_RDONLY)) {
7245 "Expected a dynptr of type %s as arg #%d\n",
7246 dynptr_type_str(arg_to_dynptr_type(arg_type)), regno);
7250 err = mark_dynptr_read(env, reg);
7255 static u32 iter_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg, int spi)
7257 struct bpf_func_state *state = func(env, reg);
7259 return state->stack[spi].spilled_ptr.ref_obj_id;
7262 static bool is_iter_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7264 return meta->kfunc_flags & (KF_ITER_NEW | KF_ITER_NEXT | KF_ITER_DESTROY);
7267 static bool is_iter_new_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7269 return meta->kfunc_flags & KF_ITER_NEW;
7272 static bool is_iter_next_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7274 return meta->kfunc_flags & KF_ITER_NEXT;
7277 static bool is_iter_destroy_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7279 return meta->kfunc_flags & KF_ITER_DESTROY;
7282 static bool is_kfunc_arg_iter(struct bpf_kfunc_call_arg_meta *meta, int arg)
7284 /* btf_check_iter_kfuncs() guarantees that first argument of any iter
7285 * kfunc is iter state pointer
7287 return arg == 0 && is_iter_kfunc(meta);
7290 static int process_iter_arg(struct bpf_verifier_env *env, int regno, int insn_idx,
7291 struct bpf_kfunc_call_arg_meta *meta)
7293 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7294 const struct btf_type *t;
7295 const struct btf_param *arg;
7296 int spi, err, i, nr_slots;
7299 /* btf_check_iter_kfuncs() ensures we don't need to validate anything here */
7300 arg = &btf_params(meta->func_proto)[0];
7301 t = btf_type_skip_modifiers(meta->btf, arg->type, NULL); /* PTR */
7302 t = btf_type_skip_modifiers(meta->btf, t->type, &btf_id); /* STRUCT */
7303 nr_slots = t->size / BPF_REG_SIZE;
7305 if (is_iter_new_kfunc(meta)) {
7306 /* bpf_iter_<type>_new() expects pointer to uninit iter state */
7307 if (!is_iter_reg_valid_uninit(env, reg, nr_slots)) {
7308 verbose(env, "expected uninitialized iter_%s as arg #%d\n",
7309 iter_type_str(meta->btf, btf_id), regno);
7313 for (i = 0; i < nr_slots * 8; i += BPF_REG_SIZE) {
7314 err = check_mem_access(env, insn_idx, regno,
7315 i, BPF_DW, BPF_WRITE, -1, false);
7320 err = mark_stack_slots_iter(env, reg, insn_idx, meta->btf, btf_id, nr_slots);
7324 /* iter_next() or iter_destroy() expect initialized iter state*/
7325 if (!is_iter_reg_valid_init(env, reg, meta->btf, btf_id, nr_slots)) {
7326 verbose(env, "expected an initialized iter_%s as arg #%d\n",
7327 iter_type_str(meta->btf, btf_id), regno);
7331 spi = iter_get_spi(env, reg, nr_slots);
7335 err = mark_iter_read(env, reg, spi, nr_slots);
7339 /* remember meta->iter info for process_iter_next_call() */
7340 meta->iter.spi = spi;
7341 meta->iter.frameno = reg->frameno;
7342 meta->ref_obj_id = iter_ref_obj_id(env, reg, spi);
7344 if (is_iter_destroy_kfunc(meta)) {
7345 err = unmark_stack_slots_iter(env, reg, nr_slots);
7354 /* process_iter_next_call() is called when verifier gets to iterator's next
7355 * "method" (e.g., bpf_iter_num_next() for numbers iterator) call. We'll refer
7356 * to it as just "iter_next()" in comments below.
7358 * BPF verifier relies on a crucial contract for any iter_next()
7359 * implementation: it should *eventually* return NULL, and once that happens
7360 * it should keep returning NULL. That is, once iterator exhausts elements to
7361 * iterate, it should never reset or spuriously return new elements.
7363 * With the assumption of such contract, process_iter_next_call() simulates
7364 * a fork in the verifier state to validate loop logic correctness and safety
7365 * without having to simulate infinite amount of iterations.
7367 * In current state, we first assume that iter_next() returned NULL and
7368 * iterator state is set to DRAINED (BPF_ITER_STATE_DRAINED). In such
7369 * conditions we should not form an infinite loop and should eventually reach
7372 * Besides that, we also fork current state and enqueue it for later
7373 * verification. In a forked state we keep iterator state as ACTIVE
7374 * (BPF_ITER_STATE_ACTIVE) and assume non-NULL return from iter_next(). We
7375 * also bump iteration depth to prevent erroneous infinite loop detection
7376 * later on (see iter_active_depths_differ() comment for details). In this
7377 * state we assume that we'll eventually loop back to another iter_next()
7378 * calls (it could be in exactly same location or in some other instruction,
7379 * it doesn't matter, we don't make any unnecessary assumptions about this,
7380 * everything revolves around iterator state in a stack slot, not which
7381 * instruction is calling iter_next()). When that happens, we either will come
7382 * to iter_next() with equivalent state and can conclude that next iteration
7383 * will proceed in exactly the same way as we just verified, so it's safe to
7384 * assume that loop converges. If not, we'll go on another iteration
7385 * simulation with a different input state, until all possible starting states
7386 * are validated or we reach maximum number of instructions limit.
7388 * This way, we will either exhaustively discover all possible input states
7389 * that iterator loop can start with and eventually will converge, or we'll
7390 * effectively regress into bounded loop simulation logic and either reach
7391 * maximum number of instructions if loop is not provably convergent, or there
7392 * is some statically known limit on number of iterations (e.g., if there is
7393 * an explicit `if n > 100 then break;` statement somewhere in the loop).
7395 * One very subtle but very important aspect is that we *always* simulate NULL
7396 * condition first (as the current state) before we simulate non-NULL case.
7397 * This has to do with intricacies of scalar precision tracking. By simulating
7398 * "exit condition" of iter_next() returning NULL first, we make sure all the
7399 * relevant precision marks *that will be set **after** we exit iterator loop*
7400 * are propagated backwards to common parent state of NULL and non-NULL
7401 * branches. Thanks to that, state equivalence checks done later in forked
7402 * state, when reaching iter_next() for ACTIVE iterator, can assume that
7403 * precision marks are finalized and won't change. Because simulating another
7404 * ACTIVE iterator iteration won't change them (because given same input
7405 * states we'll end up with exactly same output states which we are currently
7406 * comparing; and verification after the loop already propagated back what
7407 * needs to be **additionally** tracked as precise). It's subtle, grok
7408 * precision tracking for more intuitive understanding.
7410 static int process_iter_next_call(struct bpf_verifier_env *env, int insn_idx,
7411 struct bpf_kfunc_call_arg_meta *meta)
7413 struct bpf_verifier_state *cur_st = env->cur_state, *queued_st;
7414 struct bpf_func_state *cur_fr = cur_st->frame[cur_st->curframe], *queued_fr;
7415 struct bpf_reg_state *cur_iter, *queued_iter;
7416 int iter_frameno = meta->iter.frameno;
7417 int iter_spi = meta->iter.spi;
7419 BTF_TYPE_EMIT(struct bpf_iter);
7421 cur_iter = &env->cur_state->frame[iter_frameno]->stack[iter_spi].spilled_ptr;
7423 if (cur_iter->iter.state != BPF_ITER_STATE_ACTIVE &&
7424 cur_iter->iter.state != BPF_ITER_STATE_DRAINED) {
7425 verbose(env, "verifier internal error: unexpected iterator state %d (%s)\n",
7426 cur_iter->iter.state, iter_state_str(cur_iter->iter.state));
7430 if (cur_iter->iter.state == BPF_ITER_STATE_ACTIVE) {
7431 /* branch out active iter state */
7432 queued_st = push_stack(env, insn_idx + 1, insn_idx, false);
7436 queued_iter = &queued_st->frame[iter_frameno]->stack[iter_spi].spilled_ptr;
7437 queued_iter->iter.state = BPF_ITER_STATE_ACTIVE;
7438 queued_iter->iter.depth++;
7440 queued_fr = queued_st->frame[queued_st->curframe];
7441 mark_ptr_not_null_reg(&queued_fr->regs[BPF_REG_0]);
7444 /* switch to DRAINED state, but keep the depth unchanged */
7445 /* mark current iter state as drained and assume returned NULL */
7446 cur_iter->iter.state = BPF_ITER_STATE_DRAINED;
7447 __mark_reg_const_zero(&cur_fr->regs[BPF_REG_0]);
7452 static bool arg_type_is_mem_size(enum bpf_arg_type type)
7454 return type == ARG_CONST_SIZE ||
7455 type == ARG_CONST_SIZE_OR_ZERO;
7458 static bool arg_type_is_release(enum bpf_arg_type type)
7460 return type & OBJ_RELEASE;
7463 static bool arg_type_is_dynptr(enum bpf_arg_type type)
7465 return base_type(type) == ARG_PTR_TO_DYNPTR;
7468 static int int_ptr_type_to_size(enum bpf_arg_type type)
7470 if (type == ARG_PTR_TO_INT)
7472 else if (type == ARG_PTR_TO_LONG)
7478 static int resolve_map_arg_type(struct bpf_verifier_env *env,
7479 const struct bpf_call_arg_meta *meta,
7480 enum bpf_arg_type *arg_type)
7482 if (!meta->map_ptr) {
7483 /* kernel subsystem misconfigured verifier */
7484 verbose(env, "invalid map_ptr to access map->type\n");
7488 switch (meta->map_ptr->map_type) {
7489 case BPF_MAP_TYPE_SOCKMAP:
7490 case BPF_MAP_TYPE_SOCKHASH:
7491 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
7492 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
7494 verbose(env, "invalid arg_type for sockmap/sockhash\n");
7498 case BPF_MAP_TYPE_BLOOM_FILTER:
7499 if (meta->func_id == BPF_FUNC_map_peek_elem)
7500 *arg_type = ARG_PTR_TO_MAP_VALUE;
7508 struct bpf_reg_types {
7509 const enum bpf_reg_type types[10];
7513 static const struct bpf_reg_types sock_types = {
7523 static const struct bpf_reg_types btf_id_sock_common_types = {
7530 PTR_TO_BTF_ID | PTR_TRUSTED,
7532 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
7536 static const struct bpf_reg_types mem_types = {
7544 PTR_TO_MEM | MEM_RINGBUF,
7546 PTR_TO_BTF_ID | PTR_TRUSTED,
7550 static const struct bpf_reg_types int_ptr_types = {
7560 static const struct bpf_reg_types spin_lock_types = {
7563 PTR_TO_BTF_ID | MEM_ALLOC,
7567 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
7568 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
7569 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
7570 static const struct bpf_reg_types ringbuf_mem_types = { .types = { PTR_TO_MEM | MEM_RINGBUF } };
7571 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
7572 static const struct bpf_reg_types btf_ptr_types = {
7575 PTR_TO_BTF_ID | PTR_TRUSTED,
7576 PTR_TO_BTF_ID | MEM_RCU,
7579 static const struct bpf_reg_types percpu_btf_ptr_types = {
7581 PTR_TO_BTF_ID | MEM_PERCPU,
7582 PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED,
7585 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
7586 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
7587 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
7588 static const struct bpf_reg_types timer_types = { .types = { PTR_TO_MAP_VALUE } };
7589 static const struct bpf_reg_types kptr_types = { .types = { PTR_TO_MAP_VALUE } };
7590 static const struct bpf_reg_types dynptr_types = {
7593 CONST_PTR_TO_DYNPTR,
7597 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
7598 [ARG_PTR_TO_MAP_KEY] = &mem_types,
7599 [ARG_PTR_TO_MAP_VALUE] = &mem_types,
7600 [ARG_CONST_SIZE] = &scalar_types,
7601 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
7602 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
7603 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
7604 [ARG_PTR_TO_CTX] = &context_types,
7605 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
7607 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
7609 [ARG_PTR_TO_SOCKET] = &fullsock_types,
7610 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
7611 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
7612 [ARG_PTR_TO_MEM] = &mem_types,
7613 [ARG_PTR_TO_RINGBUF_MEM] = &ringbuf_mem_types,
7614 [ARG_PTR_TO_INT] = &int_ptr_types,
7615 [ARG_PTR_TO_LONG] = &int_ptr_types,
7616 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
7617 [ARG_PTR_TO_FUNC] = &func_ptr_types,
7618 [ARG_PTR_TO_STACK] = &stack_ptr_types,
7619 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
7620 [ARG_PTR_TO_TIMER] = &timer_types,
7621 [ARG_PTR_TO_KPTR] = &kptr_types,
7622 [ARG_PTR_TO_DYNPTR] = &dynptr_types,
7625 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
7626 enum bpf_arg_type arg_type,
7627 const u32 *arg_btf_id,
7628 struct bpf_call_arg_meta *meta)
7630 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7631 enum bpf_reg_type expected, type = reg->type;
7632 const struct bpf_reg_types *compatible;
7635 compatible = compatible_reg_types[base_type(arg_type)];
7637 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
7641 /* ARG_PTR_TO_MEM + RDONLY is compatible with PTR_TO_MEM and PTR_TO_MEM + RDONLY,
7642 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM and NOT with PTR_TO_MEM + RDONLY
7644 * Same for MAYBE_NULL:
7646 * ARG_PTR_TO_MEM + MAYBE_NULL is compatible with PTR_TO_MEM and PTR_TO_MEM + MAYBE_NULL,
7647 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM but NOT with PTR_TO_MEM + MAYBE_NULL
7649 * ARG_PTR_TO_MEM is compatible with PTR_TO_MEM that is tagged with a dynptr type.
7651 * Therefore we fold these flags depending on the arg_type before comparison.
7653 if (arg_type & MEM_RDONLY)
7654 type &= ~MEM_RDONLY;
7655 if (arg_type & PTR_MAYBE_NULL)
7656 type &= ~PTR_MAYBE_NULL;
7657 if (base_type(arg_type) == ARG_PTR_TO_MEM)
7658 type &= ~DYNPTR_TYPE_FLAG_MASK;
7660 if (meta->func_id == BPF_FUNC_kptr_xchg && type_is_alloc(type))
7663 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
7664 expected = compatible->types[i];
7665 if (expected == NOT_INIT)
7668 if (type == expected)
7672 verbose(env, "R%d type=%s expected=", regno, reg_type_str(env, reg->type));
7673 for (j = 0; j + 1 < i; j++)
7674 verbose(env, "%s, ", reg_type_str(env, compatible->types[j]));
7675 verbose(env, "%s\n", reg_type_str(env, compatible->types[j]));
7679 if (base_type(reg->type) != PTR_TO_BTF_ID)
7682 if (compatible == &mem_types) {
7683 if (!(arg_type & MEM_RDONLY)) {
7685 "%s() may write into memory pointed by R%d type=%s\n",
7686 func_id_name(meta->func_id),
7687 regno, reg_type_str(env, reg->type));
7693 switch ((int)reg->type) {
7695 case PTR_TO_BTF_ID | PTR_TRUSTED:
7696 case PTR_TO_BTF_ID | MEM_RCU:
7697 case PTR_TO_BTF_ID | PTR_MAYBE_NULL:
7698 case PTR_TO_BTF_ID | PTR_MAYBE_NULL | MEM_RCU:
7700 /* For bpf_sk_release, it needs to match against first member
7701 * 'struct sock_common', hence make an exception for it. This
7702 * allows bpf_sk_release to work for multiple socket types.
7704 bool strict_type_match = arg_type_is_release(arg_type) &&
7705 meta->func_id != BPF_FUNC_sk_release;
7707 if (type_may_be_null(reg->type) &&
7708 (!type_may_be_null(arg_type) || arg_type_is_release(arg_type))) {
7709 verbose(env, "Possibly NULL pointer passed to helper arg%d\n", regno);
7714 if (!compatible->btf_id) {
7715 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
7718 arg_btf_id = compatible->btf_id;
7721 if (meta->func_id == BPF_FUNC_kptr_xchg) {
7722 if (map_kptr_match_type(env, meta->kptr_field, reg, regno))
7725 if (arg_btf_id == BPF_PTR_POISON) {
7726 verbose(env, "verifier internal error:");
7727 verbose(env, "R%d has non-overwritten BPF_PTR_POISON type\n",
7732 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
7733 btf_vmlinux, *arg_btf_id,
7734 strict_type_match)) {
7735 verbose(env, "R%d is of type %s but %s is expected\n",
7736 regno, btf_type_name(reg->btf, reg->btf_id),
7737 btf_type_name(btf_vmlinux, *arg_btf_id));
7743 case PTR_TO_BTF_ID | MEM_ALLOC:
7744 if (meta->func_id != BPF_FUNC_spin_lock && meta->func_id != BPF_FUNC_spin_unlock &&
7745 meta->func_id != BPF_FUNC_kptr_xchg) {
7746 verbose(env, "verifier internal error: unimplemented handling of MEM_ALLOC\n");
7749 /* Handled by helper specific checks */
7751 case PTR_TO_BTF_ID | MEM_PERCPU:
7752 case PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED:
7753 /* Handled by helper specific checks */
7756 verbose(env, "verifier internal error: invalid PTR_TO_BTF_ID register for type match\n");
7762 static struct btf_field *
7763 reg_find_field_offset(const struct bpf_reg_state *reg, s32 off, u32 fields)
7765 struct btf_field *field;
7766 struct btf_record *rec;
7768 rec = reg_btf_record(reg);
7772 field = btf_record_find(rec, off, fields);
7779 int check_func_arg_reg_off(struct bpf_verifier_env *env,
7780 const struct bpf_reg_state *reg, int regno,
7781 enum bpf_arg_type arg_type)
7783 u32 type = reg->type;
7785 /* When referenced register is passed to release function, its fixed
7788 * We will check arg_type_is_release reg has ref_obj_id when storing
7789 * meta->release_regno.
7791 if (arg_type_is_release(arg_type)) {
7792 /* ARG_PTR_TO_DYNPTR with OBJ_RELEASE is a bit special, as it
7793 * may not directly point to the object being released, but to
7794 * dynptr pointing to such object, which might be at some offset
7795 * on the stack. In that case, we simply to fallback to the
7798 if (arg_type_is_dynptr(arg_type) && type == PTR_TO_STACK)
7801 if ((type_is_ptr_alloc_obj(type) || type_is_non_owning_ref(type)) && reg->off) {
7802 if (reg_find_field_offset(reg, reg->off, BPF_GRAPH_NODE_OR_ROOT))
7803 return __check_ptr_off_reg(env, reg, regno, true);
7805 verbose(env, "R%d must have zero offset when passed to release func\n",
7807 verbose(env, "No graph node or root found at R%d type:%s off:%d\n", regno,
7808 btf_type_name(reg->btf, reg->btf_id), reg->off);
7812 /* Doing check_ptr_off_reg check for the offset will catch this
7813 * because fixed_off_ok is false, but checking here allows us
7814 * to give the user a better error message.
7817 verbose(env, "R%d must have zero offset when passed to release func or trusted arg to kfunc\n",
7821 return __check_ptr_off_reg(env, reg, regno, false);
7825 /* Pointer types where both fixed and variable offset is explicitly allowed: */
7828 case PTR_TO_PACKET_META:
7829 case PTR_TO_MAP_KEY:
7830 case PTR_TO_MAP_VALUE:
7832 case PTR_TO_MEM | MEM_RDONLY:
7833 case PTR_TO_MEM | MEM_RINGBUF:
7835 case PTR_TO_BUF | MEM_RDONLY:
7838 /* All the rest must be rejected, except PTR_TO_BTF_ID which allows
7842 case PTR_TO_BTF_ID | MEM_ALLOC:
7843 case PTR_TO_BTF_ID | PTR_TRUSTED:
7844 case PTR_TO_BTF_ID | MEM_RCU:
7845 case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF:
7846 /* When referenced PTR_TO_BTF_ID is passed to release function,
7847 * its fixed offset must be 0. In the other cases, fixed offset
7848 * can be non-zero. This was already checked above. So pass
7849 * fixed_off_ok as true to allow fixed offset for all other
7850 * cases. var_off always must be 0 for PTR_TO_BTF_ID, hence we
7851 * still need to do checks instead of returning.
7853 return __check_ptr_off_reg(env, reg, regno, true);
7855 return __check_ptr_off_reg(env, reg, regno, false);
7859 static struct bpf_reg_state *get_dynptr_arg_reg(struct bpf_verifier_env *env,
7860 const struct bpf_func_proto *fn,
7861 struct bpf_reg_state *regs)
7863 struct bpf_reg_state *state = NULL;
7866 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++)
7867 if (arg_type_is_dynptr(fn->arg_type[i])) {
7869 verbose(env, "verifier internal error: multiple dynptr args\n");
7872 state = ®s[BPF_REG_1 + i];
7876 verbose(env, "verifier internal error: no dynptr arg found\n");
7881 static int dynptr_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
7883 struct bpf_func_state *state = func(env, reg);
7886 if (reg->type == CONST_PTR_TO_DYNPTR)
7888 spi = dynptr_get_spi(env, reg);
7891 return state->stack[spi].spilled_ptr.id;
7894 static int dynptr_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
7896 struct bpf_func_state *state = func(env, reg);
7899 if (reg->type == CONST_PTR_TO_DYNPTR)
7900 return reg->ref_obj_id;
7901 spi = dynptr_get_spi(env, reg);
7904 return state->stack[spi].spilled_ptr.ref_obj_id;
7907 static enum bpf_dynptr_type dynptr_get_type(struct bpf_verifier_env *env,
7908 struct bpf_reg_state *reg)
7910 struct bpf_func_state *state = func(env, reg);
7913 if (reg->type == CONST_PTR_TO_DYNPTR)
7914 return reg->dynptr.type;
7916 spi = __get_spi(reg->off);
7918 verbose(env, "verifier internal error: invalid spi when querying dynptr type\n");
7919 return BPF_DYNPTR_TYPE_INVALID;
7922 return state->stack[spi].spilled_ptr.dynptr.type;
7925 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
7926 struct bpf_call_arg_meta *meta,
7927 const struct bpf_func_proto *fn,
7930 u32 regno = BPF_REG_1 + arg;
7931 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7932 enum bpf_arg_type arg_type = fn->arg_type[arg];
7933 enum bpf_reg_type type = reg->type;
7934 u32 *arg_btf_id = NULL;
7937 if (arg_type == ARG_DONTCARE)
7940 err = check_reg_arg(env, regno, SRC_OP);
7944 if (arg_type == ARG_ANYTHING) {
7945 if (is_pointer_value(env, regno)) {
7946 verbose(env, "R%d leaks addr into helper function\n",
7953 if (type_is_pkt_pointer(type) &&
7954 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
7955 verbose(env, "helper access to the packet is not allowed\n");
7959 if (base_type(arg_type) == ARG_PTR_TO_MAP_VALUE) {
7960 err = resolve_map_arg_type(env, meta, &arg_type);
7965 if (register_is_null(reg) && type_may_be_null(arg_type))
7966 /* A NULL register has a SCALAR_VALUE type, so skip
7969 goto skip_type_check;
7971 /* arg_btf_id and arg_size are in a union. */
7972 if (base_type(arg_type) == ARG_PTR_TO_BTF_ID ||
7973 base_type(arg_type) == ARG_PTR_TO_SPIN_LOCK)
7974 arg_btf_id = fn->arg_btf_id[arg];
7976 err = check_reg_type(env, regno, arg_type, arg_btf_id, meta);
7980 err = check_func_arg_reg_off(env, reg, regno, arg_type);
7985 if (arg_type_is_release(arg_type)) {
7986 if (arg_type_is_dynptr(arg_type)) {
7987 struct bpf_func_state *state = func(env, reg);
7990 /* Only dynptr created on stack can be released, thus
7991 * the get_spi and stack state checks for spilled_ptr
7992 * should only be done before process_dynptr_func for
7995 if (reg->type == PTR_TO_STACK) {
7996 spi = dynptr_get_spi(env, reg);
7997 if (spi < 0 || !state->stack[spi].spilled_ptr.ref_obj_id) {
7998 verbose(env, "arg %d is an unacquired reference\n", regno);
8002 verbose(env, "cannot release unowned const bpf_dynptr\n");
8005 } else if (!reg->ref_obj_id && !register_is_null(reg)) {
8006 verbose(env, "R%d must be referenced when passed to release function\n",
8010 if (meta->release_regno) {
8011 verbose(env, "verifier internal error: more than one release argument\n");
8014 meta->release_regno = regno;
8017 if (reg->ref_obj_id) {
8018 if (meta->ref_obj_id) {
8019 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
8020 regno, reg->ref_obj_id,
8024 meta->ref_obj_id = reg->ref_obj_id;
8027 switch (base_type(arg_type)) {
8028 case ARG_CONST_MAP_PTR:
8029 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
8030 if (meta->map_ptr) {
8031 /* Use map_uid (which is unique id of inner map) to reject:
8032 * inner_map1 = bpf_map_lookup_elem(outer_map, key1)
8033 * inner_map2 = bpf_map_lookup_elem(outer_map, key2)
8034 * if (inner_map1 && inner_map2) {
8035 * timer = bpf_map_lookup_elem(inner_map1);
8037 * // mismatch would have been allowed
8038 * bpf_timer_init(timer, inner_map2);
8041 * Comparing map_ptr is enough to distinguish normal and outer maps.
8043 if (meta->map_ptr != reg->map_ptr ||
8044 meta->map_uid != reg->map_uid) {
8046 "timer pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
8047 meta->map_uid, reg->map_uid);
8051 meta->map_ptr = reg->map_ptr;
8052 meta->map_uid = reg->map_uid;
8054 case ARG_PTR_TO_MAP_KEY:
8055 /* bpf_map_xxx(..., map_ptr, ..., key) call:
8056 * check that [key, key + map->key_size) are within
8057 * stack limits and initialized
8059 if (!meta->map_ptr) {
8060 /* in function declaration map_ptr must come before
8061 * map_key, so that it's verified and known before
8062 * we have to check map_key here. Otherwise it means
8063 * that kernel subsystem misconfigured verifier
8065 verbose(env, "invalid map_ptr to access map->key\n");
8068 err = check_helper_mem_access(env, regno,
8069 meta->map_ptr->key_size, false,
8072 case ARG_PTR_TO_MAP_VALUE:
8073 if (type_may_be_null(arg_type) && register_is_null(reg))
8076 /* bpf_map_xxx(..., map_ptr, ..., value) call:
8077 * check [value, value + map->value_size) validity
8079 if (!meta->map_ptr) {
8080 /* kernel subsystem misconfigured verifier */
8081 verbose(env, "invalid map_ptr to access map->value\n");
8084 meta->raw_mode = arg_type & MEM_UNINIT;
8085 err = check_helper_mem_access(env, regno,
8086 meta->map_ptr->value_size, false,
8089 case ARG_PTR_TO_PERCPU_BTF_ID:
8091 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
8094 meta->ret_btf = reg->btf;
8095 meta->ret_btf_id = reg->btf_id;
8097 case ARG_PTR_TO_SPIN_LOCK:
8098 if (in_rbtree_lock_required_cb(env)) {
8099 verbose(env, "can't spin_{lock,unlock} in rbtree cb\n");
8102 if (meta->func_id == BPF_FUNC_spin_lock) {
8103 err = process_spin_lock(env, regno, true);
8106 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
8107 err = process_spin_lock(env, regno, false);
8111 verbose(env, "verifier internal error\n");
8115 case ARG_PTR_TO_TIMER:
8116 err = process_timer_func(env, regno, meta);
8120 case ARG_PTR_TO_FUNC:
8121 meta->subprogno = reg->subprogno;
8123 case ARG_PTR_TO_MEM:
8124 /* The access to this pointer is only checked when we hit the
8125 * next is_mem_size argument below.
8127 meta->raw_mode = arg_type & MEM_UNINIT;
8128 if (arg_type & MEM_FIXED_SIZE) {
8129 err = check_helper_mem_access(env, regno,
8130 fn->arg_size[arg], false,
8134 case ARG_CONST_SIZE:
8135 err = check_mem_size_reg(env, reg, regno, false, meta);
8137 case ARG_CONST_SIZE_OR_ZERO:
8138 err = check_mem_size_reg(env, reg, regno, true, meta);
8140 case ARG_PTR_TO_DYNPTR:
8141 err = process_dynptr_func(env, regno, insn_idx, arg_type, 0);
8145 case ARG_CONST_ALLOC_SIZE_OR_ZERO:
8146 if (!tnum_is_const(reg->var_off)) {
8147 verbose(env, "R%d is not a known constant'\n",
8151 meta->mem_size = reg->var_off.value;
8152 err = mark_chain_precision(env, regno);
8156 case ARG_PTR_TO_INT:
8157 case ARG_PTR_TO_LONG:
8159 int size = int_ptr_type_to_size(arg_type);
8161 err = check_helper_mem_access(env, regno, size, false, meta);
8164 err = check_ptr_alignment(env, reg, 0, size, true);
8167 case ARG_PTR_TO_CONST_STR:
8169 struct bpf_map *map = reg->map_ptr;
8174 if (!bpf_map_is_rdonly(map)) {
8175 verbose(env, "R%d does not point to a readonly map'\n", regno);
8179 if (!tnum_is_const(reg->var_off)) {
8180 verbose(env, "R%d is not a constant address'\n", regno);
8184 if (!map->ops->map_direct_value_addr) {
8185 verbose(env, "no direct value access support for this map type\n");
8189 err = check_map_access(env, regno, reg->off,
8190 map->value_size - reg->off, false,
8195 map_off = reg->off + reg->var_off.value;
8196 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
8198 verbose(env, "direct value access on string failed\n");
8202 str_ptr = (char *)(long)(map_addr);
8203 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
8204 verbose(env, "string is not zero-terminated\n");
8209 case ARG_PTR_TO_KPTR:
8210 err = process_kptr_func(env, regno, meta);
8219 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
8221 enum bpf_attach_type eatype = env->prog->expected_attach_type;
8222 enum bpf_prog_type type = resolve_prog_type(env->prog);
8224 if (func_id != BPF_FUNC_map_update_elem)
8227 /* It's not possible to get access to a locked struct sock in these
8228 * contexts, so updating is safe.
8231 case BPF_PROG_TYPE_TRACING:
8232 if (eatype == BPF_TRACE_ITER)
8235 case BPF_PROG_TYPE_SOCKET_FILTER:
8236 case BPF_PROG_TYPE_SCHED_CLS:
8237 case BPF_PROG_TYPE_SCHED_ACT:
8238 case BPF_PROG_TYPE_XDP:
8239 case BPF_PROG_TYPE_SK_REUSEPORT:
8240 case BPF_PROG_TYPE_FLOW_DISSECTOR:
8241 case BPF_PROG_TYPE_SK_LOOKUP:
8247 verbose(env, "cannot update sockmap in this context\n");
8251 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
8253 return env->prog->jit_requested &&
8254 bpf_jit_supports_subprog_tailcalls();
8257 static int check_map_func_compatibility(struct bpf_verifier_env *env,
8258 struct bpf_map *map, int func_id)
8263 /* We need a two way check, first is from map perspective ... */
8264 switch (map->map_type) {
8265 case BPF_MAP_TYPE_PROG_ARRAY:
8266 if (func_id != BPF_FUNC_tail_call)
8269 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
8270 if (func_id != BPF_FUNC_perf_event_read &&
8271 func_id != BPF_FUNC_perf_event_output &&
8272 func_id != BPF_FUNC_skb_output &&
8273 func_id != BPF_FUNC_perf_event_read_value &&
8274 func_id != BPF_FUNC_xdp_output)
8277 case BPF_MAP_TYPE_RINGBUF:
8278 if (func_id != BPF_FUNC_ringbuf_output &&
8279 func_id != BPF_FUNC_ringbuf_reserve &&
8280 func_id != BPF_FUNC_ringbuf_query &&
8281 func_id != BPF_FUNC_ringbuf_reserve_dynptr &&
8282 func_id != BPF_FUNC_ringbuf_submit_dynptr &&
8283 func_id != BPF_FUNC_ringbuf_discard_dynptr)
8286 case BPF_MAP_TYPE_USER_RINGBUF:
8287 if (func_id != BPF_FUNC_user_ringbuf_drain)
8290 case BPF_MAP_TYPE_STACK_TRACE:
8291 if (func_id != BPF_FUNC_get_stackid)
8294 case BPF_MAP_TYPE_CGROUP_ARRAY:
8295 if (func_id != BPF_FUNC_skb_under_cgroup &&
8296 func_id != BPF_FUNC_current_task_under_cgroup)
8299 case BPF_MAP_TYPE_CGROUP_STORAGE:
8300 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
8301 if (func_id != BPF_FUNC_get_local_storage)
8304 case BPF_MAP_TYPE_DEVMAP:
8305 case BPF_MAP_TYPE_DEVMAP_HASH:
8306 if (func_id != BPF_FUNC_redirect_map &&
8307 func_id != BPF_FUNC_map_lookup_elem)
8310 /* Restrict bpf side of cpumap and xskmap, open when use-cases
8313 case BPF_MAP_TYPE_CPUMAP:
8314 if (func_id != BPF_FUNC_redirect_map)
8317 case BPF_MAP_TYPE_XSKMAP:
8318 if (func_id != BPF_FUNC_redirect_map &&
8319 func_id != BPF_FUNC_map_lookup_elem)
8322 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
8323 case BPF_MAP_TYPE_HASH_OF_MAPS:
8324 if (func_id != BPF_FUNC_map_lookup_elem)
8327 case BPF_MAP_TYPE_SOCKMAP:
8328 if (func_id != BPF_FUNC_sk_redirect_map &&
8329 func_id != BPF_FUNC_sock_map_update &&
8330 func_id != BPF_FUNC_map_delete_elem &&
8331 func_id != BPF_FUNC_msg_redirect_map &&
8332 func_id != BPF_FUNC_sk_select_reuseport &&
8333 func_id != BPF_FUNC_map_lookup_elem &&
8334 !may_update_sockmap(env, func_id))
8337 case BPF_MAP_TYPE_SOCKHASH:
8338 if (func_id != BPF_FUNC_sk_redirect_hash &&
8339 func_id != BPF_FUNC_sock_hash_update &&
8340 func_id != BPF_FUNC_map_delete_elem &&
8341 func_id != BPF_FUNC_msg_redirect_hash &&
8342 func_id != BPF_FUNC_sk_select_reuseport &&
8343 func_id != BPF_FUNC_map_lookup_elem &&
8344 !may_update_sockmap(env, func_id))
8347 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
8348 if (func_id != BPF_FUNC_sk_select_reuseport)
8351 case BPF_MAP_TYPE_QUEUE:
8352 case BPF_MAP_TYPE_STACK:
8353 if (func_id != BPF_FUNC_map_peek_elem &&
8354 func_id != BPF_FUNC_map_pop_elem &&
8355 func_id != BPF_FUNC_map_push_elem)
8358 case BPF_MAP_TYPE_SK_STORAGE:
8359 if (func_id != BPF_FUNC_sk_storage_get &&
8360 func_id != BPF_FUNC_sk_storage_delete &&
8361 func_id != BPF_FUNC_kptr_xchg)
8364 case BPF_MAP_TYPE_INODE_STORAGE:
8365 if (func_id != BPF_FUNC_inode_storage_get &&
8366 func_id != BPF_FUNC_inode_storage_delete &&
8367 func_id != BPF_FUNC_kptr_xchg)
8370 case BPF_MAP_TYPE_TASK_STORAGE:
8371 if (func_id != BPF_FUNC_task_storage_get &&
8372 func_id != BPF_FUNC_task_storage_delete &&
8373 func_id != BPF_FUNC_kptr_xchg)
8376 case BPF_MAP_TYPE_CGRP_STORAGE:
8377 if (func_id != BPF_FUNC_cgrp_storage_get &&
8378 func_id != BPF_FUNC_cgrp_storage_delete &&
8379 func_id != BPF_FUNC_kptr_xchg)
8382 case BPF_MAP_TYPE_BLOOM_FILTER:
8383 if (func_id != BPF_FUNC_map_peek_elem &&
8384 func_id != BPF_FUNC_map_push_elem)
8391 /* ... and second from the function itself. */
8393 case BPF_FUNC_tail_call:
8394 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
8396 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
8397 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
8401 case BPF_FUNC_perf_event_read:
8402 case BPF_FUNC_perf_event_output:
8403 case BPF_FUNC_perf_event_read_value:
8404 case BPF_FUNC_skb_output:
8405 case BPF_FUNC_xdp_output:
8406 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
8409 case BPF_FUNC_ringbuf_output:
8410 case BPF_FUNC_ringbuf_reserve:
8411 case BPF_FUNC_ringbuf_query:
8412 case BPF_FUNC_ringbuf_reserve_dynptr:
8413 case BPF_FUNC_ringbuf_submit_dynptr:
8414 case BPF_FUNC_ringbuf_discard_dynptr:
8415 if (map->map_type != BPF_MAP_TYPE_RINGBUF)
8418 case BPF_FUNC_user_ringbuf_drain:
8419 if (map->map_type != BPF_MAP_TYPE_USER_RINGBUF)
8422 case BPF_FUNC_get_stackid:
8423 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
8426 case BPF_FUNC_current_task_under_cgroup:
8427 case BPF_FUNC_skb_under_cgroup:
8428 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
8431 case BPF_FUNC_redirect_map:
8432 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
8433 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
8434 map->map_type != BPF_MAP_TYPE_CPUMAP &&
8435 map->map_type != BPF_MAP_TYPE_XSKMAP)
8438 case BPF_FUNC_sk_redirect_map:
8439 case BPF_FUNC_msg_redirect_map:
8440 case BPF_FUNC_sock_map_update:
8441 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
8444 case BPF_FUNC_sk_redirect_hash:
8445 case BPF_FUNC_msg_redirect_hash:
8446 case BPF_FUNC_sock_hash_update:
8447 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
8450 case BPF_FUNC_get_local_storage:
8451 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
8452 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
8455 case BPF_FUNC_sk_select_reuseport:
8456 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
8457 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
8458 map->map_type != BPF_MAP_TYPE_SOCKHASH)
8461 case BPF_FUNC_map_pop_elem:
8462 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
8463 map->map_type != BPF_MAP_TYPE_STACK)
8466 case BPF_FUNC_map_peek_elem:
8467 case BPF_FUNC_map_push_elem:
8468 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
8469 map->map_type != BPF_MAP_TYPE_STACK &&
8470 map->map_type != BPF_MAP_TYPE_BLOOM_FILTER)
8473 case BPF_FUNC_map_lookup_percpu_elem:
8474 if (map->map_type != BPF_MAP_TYPE_PERCPU_ARRAY &&
8475 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
8476 map->map_type != BPF_MAP_TYPE_LRU_PERCPU_HASH)
8479 case BPF_FUNC_sk_storage_get:
8480 case BPF_FUNC_sk_storage_delete:
8481 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
8484 case BPF_FUNC_inode_storage_get:
8485 case BPF_FUNC_inode_storage_delete:
8486 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
8489 case BPF_FUNC_task_storage_get:
8490 case BPF_FUNC_task_storage_delete:
8491 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
8494 case BPF_FUNC_cgrp_storage_get:
8495 case BPF_FUNC_cgrp_storage_delete:
8496 if (map->map_type != BPF_MAP_TYPE_CGRP_STORAGE)
8505 verbose(env, "cannot pass map_type %d into func %s#%d\n",
8506 map->map_type, func_id_name(func_id), func_id);
8510 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
8514 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
8516 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
8518 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
8520 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
8522 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
8525 /* We only support one arg being in raw mode at the moment,
8526 * which is sufficient for the helper functions we have
8532 static bool check_args_pair_invalid(const struct bpf_func_proto *fn, int arg)
8534 bool is_fixed = fn->arg_type[arg] & MEM_FIXED_SIZE;
8535 bool has_size = fn->arg_size[arg] != 0;
8536 bool is_next_size = false;
8538 if (arg + 1 < ARRAY_SIZE(fn->arg_type))
8539 is_next_size = arg_type_is_mem_size(fn->arg_type[arg + 1]);
8541 if (base_type(fn->arg_type[arg]) != ARG_PTR_TO_MEM)
8542 return is_next_size;
8544 return has_size == is_next_size || is_next_size == is_fixed;
8547 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
8549 /* bpf_xxx(..., buf, len) call will access 'len'
8550 * bytes from memory 'buf'. Both arg types need
8551 * to be paired, so make sure there's no buggy
8552 * helper function specification.
8554 if (arg_type_is_mem_size(fn->arg1_type) ||
8555 check_args_pair_invalid(fn, 0) ||
8556 check_args_pair_invalid(fn, 1) ||
8557 check_args_pair_invalid(fn, 2) ||
8558 check_args_pair_invalid(fn, 3) ||
8559 check_args_pair_invalid(fn, 4))
8565 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
8569 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
8570 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_BTF_ID)
8571 return !!fn->arg_btf_id[i];
8572 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_SPIN_LOCK)
8573 return fn->arg_btf_id[i] == BPF_PTR_POISON;
8574 if (base_type(fn->arg_type[i]) != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i] &&
8575 /* arg_btf_id and arg_size are in a union. */
8576 (base_type(fn->arg_type[i]) != ARG_PTR_TO_MEM ||
8577 !(fn->arg_type[i] & MEM_FIXED_SIZE)))
8584 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
8586 return check_raw_mode_ok(fn) &&
8587 check_arg_pair_ok(fn) &&
8588 check_btf_id_ok(fn) ? 0 : -EINVAL;
8591 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
8592 * are now invalid, so turn them into unknown SCALAR_VALUE.
8594 * This also applies to dynptr slices belonging to skb and xdp dynptrs,
8595 * since these slices point to packet data.
8597 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
8599 struct bpf_func_state *state;
8600 struct bpf_reg_state *reg;
8602 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
8603 if (reg_is_pkt_pointer_any(reg) || reg_is_dynptr_slice_pkt(reg))
8604 mark_reg_invalid(env, reg);
8610 BEYOND_PKT_END = -2,
8613 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
8615 struct bpf_func_state *state = vstate->frame[vstate->curframe];
8616 struct bpf_reg_state *reg = &state->regs[regn];
8618 if (reg->type != PTR_TO_PACKET)
8619 /* PTR_TO_PACKET_META is not supported yet */
8622 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
8623 * How far beyond pkt_end it goes is unknown.
8624 * if (!range_open) it's the case of pkt >= pkt_end
8625 * if (range_open) it's the case of pkt > pkt_end
8626 * hence this pointer is at least 1 byte bigger than pkt_end
8629 reg->range = BEYOND_PKT_END;
8631 reg->range = AT_PKT_END;
8634 /* The pointer with the specified id has released its reference to kernel
8635 * resources. Identify all copies of the same pointer and clear the reference.
8637 static int release_reference(struct bpf_verifier_env *env,
8640 struct bpf_func_state *state;
8641 struct bpf_reg_state *reg;
8644 err = release_reference_state(cur_func(env), ref_obj_id);
8648 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
8649 if (reg->ref_obj_id == ref_obj_id)
8650 mark_reg_invalid(env, reg);
8656 static void invalidate_non_owning_refs(struct bpf_verifier_env *env)
8658 struct bpf_func_state *unused;
8659 struct bpf_reg_state *reg;
8661 bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
8662 if (type_is_non_owning_ref(reg->type))
8663 mark_reg_invalid(env, reg);
8667 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
8668 struct bpf_reg_state *regs)
8672 /* after the call registers r0 - r5 were scratched */
8673 for (i = 0; i < CALLER_SAVED_REGS; i++) {
8674 mark_reg_not_init(env, regs, caller_saved[i]);
8675 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
8679 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
8680 struct bpf_func_state *caller,
8681 struct bpf_func_state *callee,
8684 static int set_callee_state(struct bpf_verifier_env *env,
8685 struct bpf_func_state *caller,
8686 struct bpf_func_state *callee, int insn_idx);
8688 static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
8689 int *insn_idx, int subprog,
8690 set_callee_state_fn set_callee_state_cb)
8692 struct bpf_verifier_state *state = env->cur_state;
8693 struct bpf_func_state *caller, *callee;
8696 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
8697 verbose(env, "the call stack of %d frames is too deep\n",
8698 state->curframe + 2);
8702 caller = state->frame[state->curframe];
8703 if (state->frame[state->curframe + 1]) {
8704 verbose(env, "verifier bug. Frame %d already allocated\n",
8705 state->curframe + 1);
8709 err = btf_check_subprog_call(env, subprog, caller->regs);
8712 if (subprog_is_global(env, subprog)) {
8714 verbose(env, "Caller passes invalid args into func#%d\n",
8718 if (env->log.level & BPF_LOG_LEVEL)
8720 "Func#%d is global and valid. Skipping.\n",
8722 clear_caller_saved_regs(env, caller->regs);
8724 /* All global functions return a 64-bit SCALAR_VALUE */
8725 mark_reg_unknown(env, caller->regs, BPF_REG_0);
8726 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
8728 /* continue with next insn after call */
8733 /* set_callee_state is used for direct subprog calls, but we are
8734 * interested in validating only BPF helpers that can call subprogs as
8737 if (set_callee_state_cb != set_callee_state) {
8738 if (bpf_pseudo_kfunc_call(insn) &&
8739 !is_callback_calling_kfunc(insn->imm)) {
8740 verbose(env, "verifier bug: kfunc %s#%d not marked as callback-calling\n",
8741 func_id_name(insn->imm), insn->imm);
8743 } else if (!bpf_pseudo_kfunc_call(insn) &&
8744 !is_callback_calling_function(insn->imm)) { /* helper */
8745 verbose(env, "verifier bug: helper %s#%d not marked as callback-calling\n",
8746 func_id_name(insn->imm), insn->imm);
8751 if (insn->code == (BPF_JMP | BPF_CALL) &&
8752 insn->src_reg == 0 &&
8753 insn->imm == BPF_FUNC_timer_set_callback) {
8754 struct bpf_verifier_state *async_cb;
8756 /* there is no real recursion here. timer callbacks are async */
8757 env->subprog_info[subprog].is_async_cb = true;
8758 async_cb = push_async_cb(env, env->subprog_info[subprog].start,
8759 *insn_idx, subprog);
8762 callee = async_cb->frame[0];
8763 callee->async_entry_cnt = caller->async_entry_cnt + 1;
8765 /* Convert bpf_timer_set_callback() args into timer callback args */
8766 err = set_callee_state_cb(env, caller, callee, *insn_idx);
8770 clear_caller_saved_regs(env, caller->regs);
8771 mark_reg_unknown(env, caller->regs, BPF_REG_0);
8772 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
8773 /* continue with next insn after call */
8777 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
8780 state->frame[state->curframe + 1] = callee;
8782 /* callee cannot access r0, r6 - r9 for reading and has to write
8783 * into its own stack before reading from it.
8784 * callee can read/write into caller's stack
8786 init_func_state(env, callee,
8787 /* remember the callsite, it will be used by bpf_exit */
8788 *insn_idx /* callsite */,
8789 state->curframe + 1 /* frameno within this callchain */,
8790 subprog /* subprog number within this prog */);
8792 /* Transfer references to the callee */
8793 err = copy_reference_state(callee, caller);
8797 err = set_callee_state_cb(env, caller, callee, *insn_idx);
8801 clear_caller_saved_regs(env, caller->regs);
8803 /* only increment it after check_reg_arg() finished */
8806 /* and go analyze first insn of the callee */
8807 *insn_idx = env->subprog_info[subprog].start - 1;
8809 if (env->log.level & BPF_LOG_LEVEL) {
8810 verbose(env, "caller:\n");
8811 print_verifier_state(env, caller, true);
8812 verbose(env, "callee:\n");
8813 print_verifier_state(env, callee, true);
8818 free_func_state(callee);
8819 state->frame[state->curframe + 1] = NULL;
8823 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
8824 struct bpf_func_state *caller,
8825 struct bpf_func_state *callee)
8827 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
8828 * void *callback_ctx, u64 flags);
8829 * callback_fn(struct bpf_map *map, void *key, void *value,
8830 * void *callback_ctx);
8832 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
8834 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
8835 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
8836 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
8838 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
8839 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
8840 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
8842 /* pointer to stack or null */
8843 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
8846 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
8850 static int set_callee_state(struct bpf_verifier_env *env,
8851 struct bpf_func_state *caller,
8852 struct bpf_func_state *callee, int insn_idx)
8856 /* copy r1 - r5 args that callee can access. The copy includes parent
8857 * pointers, which connects us up to the liveness chain
8859 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
8860 callee->regs[i] = caller->regs[i];
8864 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
8867 int subprog, target_insn;
8869 target_insn = *insn_idx + insn->imm + 1;
8870 subprog = find_subprog(env, target_insn);
8872 verbose(env, "verifier bug. No program starts at insn %d\n",
8877 return __check_func_call(env, insn, insn_idx, subprog, set_callee_state);
8880 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
8881 struct bpf_func_state *caller,
8882 struct bpf_func_state *callee,
8885 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
8886 struct bpf_map *map;
8889 if (bpf_map_ptr_poisoned(insn_aux)) {
8890 verbose(env, "tail_call abusing map_ptr\n");
8894 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
8895 if (!map->ops->map_set_for_each_callback_args ||
8896 !map->ops->map_for_each_callback) {
8897 verbose(env, "callback function not allowed for map\n");
8901 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
8905 callee->in_callback_fn = true;
8906 callee->callback_ret_range = tnum_range(0, 1);
8910 static int set_loop_callback_state(struct bpf_verifier_env *env,
8911 struct bpf_func_state *caller,
8912 struct bpf_func_state *callee,
8915 /* bpf_loop(u32 nr_loops, void *callback_fn, void *callback_ctx,
8917 * callback_fn(u32 index, void *callback_ctx);
8919 callee->regs[BPF_REG_1].type = SCALAR_VALUE;
8920 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
8923 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
8924 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
8925 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
8927 callee->in_callback_fn = true;
8928 callee->callback_ret_range = tnum_range(0, 1);
8932 static int set_timer_callback_state(struct bpf_verifier_env *env,
8933 struct bpf_func_state *caller,
8934 struct bpf_func_state *callee,
8937 struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
8939 /* bpf_timer_set_callback(struct bpf_timer *timer, void *callback_fn);
8940 * callback_fn(struct bpf_map *map, void *key, void *value);
8942 callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP;
8943 __mark_reg_known_zero(&callee->regs[BPF_REG_1]);
8944 callee->regs[BPF_REG_1].map_ptr = map_ptr;
8946 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
8947 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
8948 callee->regs[BPF_REG_2].map_ptr = map_ptr;
8950 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
8951 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
8952 callee->regs[BPF_REG_3].map_ptr = map_ptr;
8955 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
8956 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
8957 callee->in_async_callback_fn = true;
8958 callee->callback_ret_range = tnum_range(0, 1);
8962 static int set_find_vma_callback_state(struct bpf_verifier_env *env,
8963 struct bpf_func_state *caller,
8964 struct bpf_func_state *callee,
8967 /* bpf_find_vma(struct task_struct *task, u64 addr,
8968 * void *callback_fn, void *callback_ctx, u64 flags)
8969 * (callback_fn)(struct task_struct *task,
8970 * struct vm_area_struct *vma, void *callback_ctx);
8972 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
8974 callee->regs[BPF_REG_2].type = PTR_TO_BTF_ID;
8975 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
8976 callee->regs[BPF_REG_2].btf = btf_vmlinux;
8977 callee->regs[BPF_REG_2].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_VMA],
8979 /* pointer to stack or null */
8980 callee->regs[BPF_REG_3] = caller->regs[BPF_REG_4];
8983 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
8984 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
8985 callee->in_callback_fn = true;
8986 callee->callback_ret_range = tnum_range(0, 1);
8990 static int set_user_ringbuf_callback_state(struct bpf_verifier_env *env,
8991 struct bpf_func_state *caller,
8992 struct bpf_func_state *callee,
8995 /* bpf_user_ringbuf_drain(struct bpf_map *map, void *callback_fn, void
8996 * callback_ctx, u64 flags);
8997 * callback_fn(const struct bpf_dynptr_t* dynptr, void *callback_ctx);
8999 __mark_reg_not_init(env, &callee->regs[BPF_REG_0]);
9000 mark_dynptr_cb_reg(env, &callee->regs[BPF_REG_1], BPF_DYNPTR_TYPE_LOCAL);
9001 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
9004 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
9005 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9006 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9008 callee->in_callback_fn = true;
9009 callee->callback_ret_range = tnum_range(0, 1);
9013 static int set_rbtree_add_callback_state(struct bpf_verifier_env *env,
9014 struct bpf_func_state *caller,
9015 struct bpf_func_state *callee,
9018 /* void bpf_rbtree_add_impl(struct bpf_rb_root *root, struct bpf_rb_node *node,
9019 * bool (less)(struct bpf_rb_node *a, const struct bpf_rb_node *b));
9021 * 'struct bpf_rb_node *node' arg to bpf_rbtree_add_impl is the same PTR_TO_BTF_ID w/ offset
9022 * that 'less' callback args will be receiving. However, 'node' arg was release_reference'd
9023 * by this point, so look at 'root'
9025 struct btf_field *field;
9027 field = reg_find_field_offset(&caller->regs[BPF_REG_1], caller->regs[BPF_REG_1].off,
9029 if (!field || !field->graph_root.value_btf_id)
9032 mark_reg_graph_node(callee->regs, BPF_REG_1, &field->graph_root);
9033 ref_set_non_owning(env, &callee->regs[BPF_REG_1]);
9034 mark_reg_graph_node(callee->regs, BPF_REG_2, &field->graph_root);
9035 ref_set_non_owning(env, &callee->regs[BPF_REG_2]);
9037 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
9038 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9039 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9040 callee->in_callback_fn = true;
9041 callee->callback_ret_range = tnum_range(0, 1);
9045 static bool is_rbtree_lock_required_kfunc(u32 btf_id);
9047 /* Are we currently verifying the callback for a rbtree helper that must
9048 * be called with lock held? If so, no need to complain about unreleased
9051 static bool in_rbtree_lock_required_cb(struct bpf_verifier_env *env)
9053 struct bpf_verifier_state *state = env->cur_state;
9054 struct bpf_insn *insn = env->prog->insnsi;
9055 struct bpf_func_state *callee;
9058 if (!state->curframe)
9061 callee = state->frame[state->curframe];
9063 if (!callee->in_callback_fn)
9066 kfunc_btf_id = insn[callee->callsite].imm;
9067 return is_rbtree_lock_required_kfunc(kfunc_btf_id);
9070 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
9072 struct bpf_verifier_state *state = env->cur_state;
9073 struct bpf_func_state *caller, *callee;
9074 struct bpf_reg_state *r0;
9077 callee = state->frame[state->curframe];
9078 r0 = &callee->regs[BPF_REG_0];
9079 if (r0->type == PTR_TO_STACK) {
9080 /* technically it's ok to return caller's stack pointer
9081 * (or caller's caller's pointer) back to the caller,
9082 * since these pointers are valid. Only current stack
9083 * pointer will be invalid as soon as function exits,
9084 * but let's be conservative
9086 verbose(env, "cannot return stack pointer to the caller\n");
9090 caller = state->frame[state->curframe - 1];
9091 if (callee->in_callback_fn) {
9092 /* enforce R0 return value range [0, 1]. */
9093 struct tnum range = callee->callback_ret_range;
9095 if (r0->type != SCALAR_VALUE) {
9096 verbose(env, "R0 not a scalar value\n");
9099 if (!tnum_in(range, r0->var_off)) {
9100 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
9104 /* return to the caller whatever r0 had in the callee */
9105 caller->regs[BPF_REG_0] = *r0;
9108 /* callback_fn frame should have released its own additions to parent's
9109 * reference state at this point, or check_reference_leak would
9110 * complain, hence it must be the same as the caller. There is no need
9113 if (!callee->in_callback_fn) {
9114 /* Transfer references to the caller */
9115 err = copy_reference_state(caller, callee);
9120 *insn_idx = callee->callsite + 1;
9121 if (env->log.level & BPF_LOG_LEVEL) {
9122 verbose(env, "returning from callee:\n");
9123 print_verifier_state(env, callee, true);
9124 verbose(env, "to caller at %d:\n", *insn_idx);
9125 print_verifier_state(env, caller, true);
9127 /* clear everything in the callee */
9128 free_func_state(callee);
9129 state->frame[state->curframe--] = NULL;
9133 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
9135 struct bpf_call_arg_meta *meta)
9137 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
9139 if (ret_type != RET_INTEGER)
9143 case BPF_FUNC_get_stack:
9144 case BPF_FUNC_get_task_stack:
9145 case BPF_FUNC_probe_read_str:
9146 case BPF_FUNC_probe_read_kernel_str:
9147 case BPF_FUNC_probe_read_user_str:
9148 ret_reg->smax_value = meta->msize_max_value;
9149 ret_reg->s32_max_value = meta->msize_max_value;
9150 ret_reg->smin_value = -MAX_ERRNO;
9151 ret_reg->s32_min_value = -MAX_ERRNO;
9152 reg_bounds_sync(ret_reg);
9154 case BPF_FUNC_get_smp_processor_id:
9155 ret_reg->umax_value = nr_cpu_ids - 1;
9156 ret_reg->u32_max_value = nr_cpu_ids - 1;
9157 ret_reg->smax_value = nr_cpu_ids - 1;
9158 ret_reg->s32_max_value = nr_cpu_ids - 1;
9159 ret_reg->umin_value = 0;
9160 ret_reg->u32_min_value = 0;
9161 ret_reg->smin_value = 0;
9162 ret_reg->s32_min_value = 0;
9163 reg_bounds_sync(ret_reg);
9169 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
9170 int func_id, int insn_idx)
9172 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
9173 struct bpf_map *map = meta->map_ptr;
9175 if (func_id != BPF_FUNC_tail_call &&
9176 func_id != BPF_FUNC_map_lookup_elem &&
9177 func_id != BPF_FUNC_map_update_elem &&
9178 func_id != BPF_FUNC_map_delete_elem &&
9179 func_id != BPF_FUNC_map_push_elem &&
9180 func_id != BPF_FUNC_map_pop_elem &&
9181 func_id != BPF_FUNC_map_peek_elem &&
9182 func_id != BPF_FUNC_for_each_map_elem &&
9183 func_id != BPF_FUNC_redirect_map &&
9184 func_id != BPF_FUNC_map_lookup_percpu_elem)
9188 verbose(env, "kernel subsystem misconfigured verifier\n");
9192 /* In case of read-only, some additional restrictions
9193 * need to be applied in order to prevent altering the
9194 * state of the map from program side.
9196 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
9197 (func_id == BPF_FUNC_map_delete_elem ||
9198 func_id == BPF_FUNC_map_update_elem ||
9199 func_id == BPF_FUNC_map_push_elem ||
9200 func_id == BPF_FUNC_map_pop_elem)) {
9201 verbose(env, "write into map forbidden\n");
9205 if (!BPF_MAP_PTR(aux->map_ptr_state))
9206 bpf_map_ptr_store(aux, meta->map_ptr,
9207 !meta->map_ptr->bypass_spec_v1);
9208 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
9209 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
9210 !meta->map_ptr->bypass_spec_v1);
9215 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
9216 int func_id, int insn_idx)
9218 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
9219 struct bpf_reg_state *regs = cur_regs(env), *reg;
9220 struct bpf_map *map = meta->map_ptr;
9224 if (func_id != BPF_FUNC_tail_call)
9226 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
9227 verbose(env, "kernel subsystem misconfigured verifier\n");
9231 reg = ®s[BPF_REG_3];
9232 val = reg->var_off.value;
9233 max = map->max_entries;
9235 if (!(register_is_const(reg) && val < max)) {
9236 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
9240 err = mark_chain_precision(env, BPF_REG_3);
9243 if (bpf_map_key_unseen(aux))
9244 bpf_map_key_store(aux, val);
9245 else if (!bpf_map_key_poisoned(aux) &&
9246 bpf_map_key_immediate(aux) != val)
9247 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
9251 static int check_reference_leak(struct bpf_verifier_env *env)
9253 struct bpf_func_state *state = cur_func(env);
9254 bool refs_lingering = false;
9257 if (state->frameno && !state->in_callback_fn)
9260 for (i = 0; i < state->acquired_refs; i++) {
9261 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno)
9263 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
9264 state->refs[i].id, state->refs[i].insn_idx);
9265 refs_lingering = true;
9267 return refs_lingering ? -EINVAL : 0;
9270 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
9271 struct bpf_reg_state *regs)
9273 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
9274 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
9275 struct bpf_map *fmt_map = fmt_reg->map_ptr;
9276 struct bpf_bprintf_data data = {};
9277 int err, fmt_map_off, num_args;
9281 /* data must be an array of u64 */
9282 if (data_len_reg->var_off.value % 8)
9284 num_args = data_len_reg->var_off.value / 8;
9286 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
9287 * and map_direct_value_addr is set.
9289 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
9290 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
9293 verbose(env, "verifier bug\n");
9296 fmt = (char *)(long)fmt_addr + fmt_map_off;
9298 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
9299 * can focus on validating the format specifiers.
9301 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, num_args, &data);
9303 verbose(env, "Invalid format string\n");
9308 static int check_get_func_ip(struct bpf_verifier_env *env)
9310 enum bpf_prog_type type = resolve_prog_type(env->prog);
9311 int func_id = BPF_FUNC_get_func_ip;
9313 if (type == BPF_PROG_TYPE_TRACING) {
9314 if (!bpf_prog_has_trampoline(env->prog)) {
9315 verbose(env, "func %s#%d supported only for fentry/fexit/fmod_ret programs\n",
9316 func_id_name(func_id), func_id);
9320 } else if (type == BPF_PROG_TYPE_KPROBE) {
9324 verbose(env, "func %s#%d not supported for program type %d\n",
9325 func_id_name(func_id), func_id, type);
9329 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
9331 return &env->insn_aux_data[env->insn_idx];
9334 static bool loop_flag_is_zero(struct bpf_verifier_env *env)
9336 struct bpf_reg_state *regs = cur_regs(env);
9337 struct bpf_reg_state *reg = ®s[BPF_REG_4];
9338 bool reg_is_null = register_is_null(reg);
9341 mark_chain_precision(env, BPF_REG_4);
9346 static void update_loop_inline_state(struct bpf_verifier_env *env, u32 subprogno)
9348 struct bpf_loop_inline_state *state = &cur_aux(env)->loop_inline_state;
9350 if (!state->initialized) {
9351 state->initialized = 1;
9352 state->fit_for_inline = loop_flag_is_zero(env);
9353 state->callback_subprogno = subprogno;
9357 if (!state->fit_for_inline)
9360 state->fit_for_inline = (loop_flag_is_zero(env) &&
9361 state->callback_subprogno == subprogno);
9364 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
9367 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
9368 const struct bpf_func_proto *fn = NULL;
9369 enum bpf_return_type ret_type;
9370 enum bpf_type_flag ret_flag;
9371 struct bpf_reg_state *regs;
9372 struct bpf_call_arg_meta meta;
9373 int insn_idx = *insn_idx_p;
9375 int i, err, func_id;
9377 /* find function prototype */
9378 func_id = insn->imm;
9379 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
9380 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
9385 if (env->ops->get_func_proto)
9386 fn = env->ops->get_func_proto(func_id, env->prog);
9388 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
9393 /* eBPF programs must be GPL compatible to use GPL-ed functions */
9394 if (!env->prog->gpl_compatible && fn->gpl_only) {
9395 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
9399 if (fn->allowed && !fn->allowed(env->prog)) {
9400 verbose(env, "helper call is not allowed in probe\n");
9404 if (!env->prog->aux->sleepable && fn->might_sleep) {
9405 verbose(env, "helper call might sleep in a non-sleepable prog\n");
9409 /* With LD_ABS/IND some JITs save/restore skb from r1. */
9410 changes_data = bpf_helper_changes_pkt_data(fn->func);
9411 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
9412 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
9413 func_id_name(func_id), func_id);
9417 memset(&meta, 0, sizeof(meta));
9418 meta.pkt_access = fn->pkt_access;
9420 err = check_func_proto(fn, func_id);
9422 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
9423 func_id_name(func_id), func_id);
9427 if (env->cur_state->active_rcu_lock) {
9428 if (fn->might_sleep) {
9429 verbose(env, "sleepable helper %s#%d in rcu_read_lock region\n",
9430 func_id_name(func_id), func_id);
9434 if (env->prog->aux->sleepable && is_storage_get_function(func_id))
9435 env->insn_aux_data[insn_idx].storage_get_func_atomic = true;
9438 meta.func_id = func_id;
9440 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
9441 err = check_func_arg(env, i, &meta, fn, insn_idx);
9446 err = record_func_map(env, &meta, func_id, insn_idx);
9450 err = record_func_key(env, &meta, func_id, insn_idx);
9454 /* Mark slots with STACK_MISC in case of raw mode, stack offset
9455 * is inferred from register state.
9457 for (i = 0; i < meta.access_size; i++) {
9458 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
9459 BPF_WRITE, -1, false);
9464 regs = cur_regs(env);
9466 if (meta.release_regno) {
9468 /* This can only be set for PTR_TO_STACK, as CONST_PTR_TO_DYNPTR cannot
9469 * be released by any dynptr helper. Hence, unmark_stack_slots_dynptr
9470 * is safe to do directly.
9472 if (arg_type_is_dynptr(fn->arg_type[meta.release_regno - BPF_REG_1])) {
9473 if (regs[meta.release_regno].type == CONST_PTR_TO_DYNPTR) {
9474 verbose(env, "verifier internal error: CONST_PTR_TO_DYNPTR cannot be released\n");
9477 err = unmark_stack_slots_dynptr(env, ®s[meta.release_regno]);
9478 } else if (meta.ref_obj_id) {
9479 err = release_reference(env, meta.ref_obj_id);
9480 } else if (register_is_null(®s[meta.release_regno])) {
9481 /* meta.ref_obj_id can only be 0 if register that is meant to be
9482 * released is NULL, which must be > R0.
9487 verbose(env, "func %s#%d reference has not been acquired before\n",
9488 func_id_name(func_id), func_id);
9494 case BPF_FUNC_tail_call:
9495 err = check_reference_leak(env);
9497 verbose(env, "tail_call would lead to reference leak\n");
9501 case BPF_FUNC_get_local_storage:
9502 /* check that flags argument in get_local_storage(map, flags) is 0,
9503 * this is required because get_local_storage() can't return an error.
9505 if (!register_is_null(®s[BPF_REG_2])) {
9506 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
9510 case BPF_FUNC_for_each_map_elem:
9511 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9512 set_map_elem_callback_state);
9514 case BPF_FUNC_timer_set_callback:
9515 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9516 set_timer_callback_state);
9518 case BPF_FUNC_find_vma:
9519 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9520 set_find_vma_callback_state);
9522 case BPF_FUNC_snprintf:
9523 err = check_bpf_snprintf_call(env, regs);
9526 update_loop_inline_state(env, meta.subprogno);
9527 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9528 set_loop_callback_state);
9530 case BPF_FUNC_dynptr_from_mem:
9531 if (regs[BPF_REG_1].type != PTR_TO_MAP_VALUE) {
9532 verbose(env, "Unsupported reg type %s for bpf_dynptr_from_mem data\n",
9533 reg_type_str(env, regs[BPF_REG_1].type));
9537 case BPF_FUNC_set_retval:
9538 if (prog_type == BPF_PROG_TYPE_LSM &&
9539 env->prog->expected_attach_type == BPF_LSM_CGROUP) {
9540 if (!env->prog->aux->attach_func_proto->type) {
9541 /* Make sure programs that attach to void
9542 * hooks don't try to modify return value.
9544 verbose(env, "BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
9549 case BPF_FUNC_dynptr_data:
9551 struct bpf_reg_state *reg;
9554 reg = get_dynptr_arg_reg(env, fn, regs);
9559 if (meta.dynptr_id) {
9560 verbose(env, "verifier internal error: meta.dynptr_id already set\n");
9563 if (meta.ref_obj_id) {
9564 verbose(env, "verifier internal error: meta.ref_obj_id already set\n");
9568 id = dynptr_id(env, reg);
9570 verbose(env, "verifier internal error: failed to obtain dynptr id\n");
9574 ref_obj_id = dynptr_ref_obj_id(env, reg);
9575 if (ref_obj_id < 0) {
9576 verbose(env, "verifier internal error: failed to obtain dynptr ref_obj_id\n");
9580 meta.dynptr_id = id;
9581 meta.ref_obj_id = ref_obj_id;
9585 case BPF_FUNC_dynptr_write:
9587 enum bpf_dynptr_type dynptr_type;
9588 struct bpf_reg_state *reg;
9590 reg = get_dynptr_arg_reg(env, fn, regs);
9594 dynptr_type = dynptr_get_type(env, reg);
9595 if (dynptr_type == BPF_DYNPTR_TYPE_INVALID)
9598 if (dynptr_type == BPF_DYNPTR_TYPE_SKB)
9599 /* this will trigger clear_all_pkt_pointers(), which will
9600 * invalidate all dynptr slices associated with the skb
9602 changes_data = true;
9606 case BPF_FUNC_user_ringbuf_drain:
9607 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
9608 set_user_ringbuf_callback_state);
9615 /* reset caller saved regs */
9616 for (i = 0; i < CALLER_SAVED_REGS; i++) {
9617 mark_reg_not_init(env, regs, caller_saved[i]);
9618 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
9621 /* helper call returns 64-bit value. */
9622 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
9624 /* update return register (already marked as written above) */
9625 ret_type = fn->ret_type;
9626 ret_flag = type_flag(ret_type);
9628 switch (base_type(ret_type)) {
9630 /* sets type to SCALAR_VALUE */
9631 mark_reg_unknown(env, regs, BPF_REG_0);
9634 regs[BPF_REG_0].type = NOT_INIT;
9636 case RET_PTR_TO_MAP_VALUE:
9637 /* There is no offset yet applied, variable or fixed */
9638 mark_reg_known_zero(env, regs, BPF_REG_0);
9639 /* remember map_ptr, so that check_map_access()
9640 * can check 'value_size' boundary of memory access
9641 * to map element returned from bpf_map_lookup_elem()
9643 if (meta.map_ptr == NULL) {
9645 "kernel subsystem misconfigured verifier\n");
9648 regs[BPF_REG_0].map_ptr = meta.map_ptr;
9649 regs[BPF_REG_0].map_uid = meta.map_uid;
9650 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE | ret_flag;
9651 if (!type_may_be_null(ret_type) &&
9652 btf_record_has_field(meta.map_ptr->record, BPF_SPIN_LOCK)) {
9653 regs[BPF_REG_0].id = ++env->id_gen;
9656 case RET_PTR_TO_SOCKET:
9657 mark_reg_known_zero(env, regs, BPF_REG_0);
9658 regs[BPF_REG_0].type = PTR_TO_SOCKET | ret_flag;
9660 case RET_PTR_TO_SOCK_COMMON:
9661 mark_reg_known_zero(env, regs, BPF_REG_0);
9662 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON | ret_flag;
9664 case RET_PTR_TO_TCP_SOCK:
9665 mark_reg_known_zero(env, regs, BPF_REG_0);
9666 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK | ret_flag;
9668 case RET_PTR_TO_MEM:
9669 mark_reg_known_zero(env, regs, BPF_REG_0);
9670 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
9671 regs[BPF_REG_0].mem_size = meta.mem_size;
9673 case RET_PTR_TO_MEM_OR_BTF_ID:
9675 const struct btf_type *t;
9677 mark_reg_known_zero(env, regs, BPF_REG_0);
9678 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
9679 if (!btf_type_is_struct(t)) {
9681 const struct btf_type *ret;
9684 /* resolve the type size of ksym. */
9685 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
9687 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
9688 verbose(env, "unable to resolve the size of type '%s': %ld\n",
9689 tname, PTR_ERR(ret));
9692 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
9693 regs[BPF_REG_0].mem_size = tsize;
9695 /* MEM_RDONLY may be carried from ret_flag, but it
9696 * doesn't apply on PTR_TO_BTF_ID. Fold it, otherwise
9697 * it will confuse the check of PTR_TO_BTF_ID in
9698 * check_mem_access().
9700 ret_flag &= ~MEM_RDONLY;
9702 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
9703 regs[BPF_REG_0].btf = meta.ret_btf;
9704 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
9708 case RET_PTR_TO_BTF_ID:
9710 struct btf *ret_btf;
9713 mark_reg_known_zero(env, regs, BPF_REG_0);
9714 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
9715 if (func_id == BPF_FUNC_kptr_xchg) {
9716 ret_btf = meta.kptr_field->kptr.btf;
9717 ret_btf_id = meta.kptr_field->kptr.btf_id;
9718 if (!btf_is_kernel(ret_btf))
9719 regs[BPF_REG_0].type |= MEM_ALLOC;
9721 if (fn->ret_btf_id == BPF_PTR_POISON) {
9722 verbose(env, "verifier internal error:");
9723 verbose(env, "func %s has non-overwritten BPF_PTR_POISON return type\n",
9724 func_id_name(func_id));
9727 ret_btf = btf_vmlinux;
9728 ret_btf_id = *fn->ret_btf_id;
9730 if (ret_btf_id == 0) {
9731 verbose(env, "invalid return type %u of func %s#%d\n",
9732 base_type(ret_type), func_id_name(func_id),
9736 regs[BPF_REG_0].btf = ret_btf;
9737 regs[BPF_REG_0].btf_id = ret_btf_id;
9741 verbose(env, "unknown return type %u of func %s#%d\n",
9742 base_type(ret_type), func_id_name(func_id), func_id);
9746 if (type_may_be_null(regs[BPF_REG_0].type))
9747 regs[BPF_REG_0].id = ++env->id_gen;
9749 if (helper_multiple_ref_obj_use(func_id, meta.map_ptr)) {
9750 verbose(env, "verifier internal error: func %s#%d sets ref_obj_id more than once\n",
9751 func_id_name(func_id), func_id);
9755 if (is_dynptr_ref_function(func_id))
9756 regs[BPF_REG_0].dynptr_id = meta.dynptr_id;
9758 if (is_ptr_cast_function(func_id) || is_dynptr_ref_function(func_id)) {
9759 /* For release_reference() */
9760 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
9761 } else if (is_acquire_function(func_id, meta.map_ptr)) {
9762 int id = acquire_reference_state(env, insn_idx);
9766 /* For mark_ptr_or_null_reg() */
9767 regs[BPF_REG_0].id = id;
9768 /* For release_reference() */
9769 regs[BPF_REG_0].ref_obj_id = id;
9772 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
9774 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
9778 if ((func_id == BPF_FUNC_get_stack ||
9779 func_id == BPF_FUNC_get_task_stack) &&
9780 !env->prog->has_callchain_buf) {
9781 const char *err_str;
9783 #ifdef CONFIG_PERF_EVENTS
9784 err = get_callchain_buffers(sysctl_perf_event_max_stack);
9785 err_str = "cannot get callchain buffer for func %s#%d\n";
9788 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
9791 verbose(env, err_str, func_id_name(func_id), func_id);
9795 env->prog->has_callchain_buf = true;
9798 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
9799 env->prog->call_get_stack = true;
9801 if (func_id == BPF_FUNC_get_func_ip) {
9802 if (check_get_func_ip(env))
9804 env->prog->call_get_func_ip = true;
9808 clear_all_pkt_pointers(env);
9812 /* mark_btf_func_reg_size() is used when the reg size is determined by
9813 * the BTF func_proto's return value size and argument.
9815 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
9818 struct bpf_reg_state *reg = &cur_regs(env)[regno];
9820 if (regno == BPF_REG_0) {
9821 /* Function return value */
9822 reg->live |= REG_LIVE_WRITTEN;
9823 reg->subreg_def = reg_size == sizeof(u64) ?
9824 DEF_NOT_SUBREG : env->insn_idx + 1;
9826 /* Function argument */
9827 if (reg_size == sizeof(u64)) {
9828 mark_insn_zext(env, reg);
9829 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
9831 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
9836 static bool is_kfunc_acquire(struct bpf_kfunc_call_arg_meta *meta)
9838 return meta->kfunc_flags & KF_ACQUIRE;
9841 static bool is_kfunc_release(struct bpf_kfunc_call_arg_meta *meta)
9843 return meta->kfunc_flags & KF_RELEASE;
9846 static bool is_kfunc_trusted_args(struct bpf_kfunc_call_arg_meta *meta)
9848 return (meta->kfunc_flags & KF_TRUSTED_ARGS) || is_kfunc_release(meta);
9851 static bool is_kfunc_sleepable(struct bpf_kfunc_call_arg_meta *meta)
9853 return meta->kfunc_flags & KF_SLEEPABLE;
9856 static bool is_kfunc_destructive(struct bpf_kfunc_call_arg_meta *meta)
9858 return meta->kfunc_flags & KF_DESTRUCTIVE;
9861 static bool is_kfunc_rcu(struct bpf_kfunc_call_arg_meta *meta)
9863 return meta->kfunc_flags & KF_RCU;
9866 static bool __kfunc_param_match_suffix(const struct btf *btf,
9867 const struct btf_param *arg,
9870 int suffix_len = strlen(suffix), len;
9871 const char *param_name;
9873 /* In the future, this can be ported to use BTF tagging */
9874 param_name = btf_name_by_offset(btf, arg->name_off);
9875 if (str_is_empty(param_name))
9877 len = strlen(param_name);
9878 if (len < suffix_len)
9880 param_name += len - suffix_len;
9881 return !strncmp(param_name, suffix, suffix_len);
9884 static bool is_kfunc_arg_mem_size(const struct btf *btf,
9885 const struct btf_param *arg,
9886 const struct bpf_reg_state *reg)
9888 const struct btf_type *t;
9890 t = btf_type_skip_modifiers(btf, arg->type, NULL);
9891 if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
9894 return __kfunc_param_match_suffix(btf, arg, "__sz");
9897 static bool is_kfunc_arg_const_mem_size(const struct btf *btf,
9898 const struct btf_param *arg,
9899 const struct bpf_reg_state *reg)
9901 const struct btf_type *t;
9903 t = btf_type_skip_modifiers(btf, arg->type, NULL);
9904 if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
9907 return __kfunc_param_match_suffix(btf, arg, "__szk");
9910 static bool is_kfunc_arg_optional(const struct btf *btf, const struct btf_param *arg)
9912 return __kfunc_param_match_suffix(btf, arg, "__opt");
9915 static bool is_kfunc_arg_constant(const struct btf *btf, const struct btf_param *arg)
9917 return __kfunc_param_match_suffix(btf, arg, "__k");
9920 static bool is_kfunc_arg_ignore(const struct btf *btf, const struct btf_param *arg)
9922 return __kfunc_param_match_suffix(btf, arg, "__ign");
9925 static bool is_kfunc_arg_alloc_obj(const struct btf *btf, const struct btf_param *arg)
9927 return __kfunc_param_match_suffix(btf, arg, "__alloc");
9930 static bool is_kfunc_arg_uninit(const struct btf *btf, const struct btf_param *arg)
9932 return __kfunc_param_match_suffix(btf, arg, "__uninit");
9935 static bool is_kfunc_arg_refcounted_kptr(const struct btf *btf, const struct btf_param *arg)
9937 return __kfunc_param_match_suffix(btf, arg, "__refcounted_kptr");
9940 static bool is_kfunc_arg_scalar_with_name(const struct btf *btf,
9941 const struct btf_param *arg,
9944 int len, target_len = strlen(name);
9945 const char *param_name;
9947 param_name = btf_name_by_offset(btf, arg->name_off);
9948 if (str_is_empty(param_name))
9950 len = strlen(param_name);
9951 if (len != target_len)
9953 if (strcmp(param_name, name))
9961 KF_ARG_LIST_HEAD_ID,
9962 KF_ARG_LIST_NODE_ID,
9967 BTF_ID_LIST(kf_arg_btf_ids)
9968 BTF_ID(struct, bpf_dynptr_kern)
9969 BTF_ID(struct, bpf_list_head)
9970 BTF_ID(struct, bpf_list_node)
9971 BTF_ID(struct, bpf_rb_root)
9972 BTF_ID(struct, bpf_rb_node)
9974 static bool __is_kfunc_ptr_arg_type(const struct btf *btf,
9975 const struct btf_param *arg, int type)
9977 const struct btf_type *t;
9980 t = btf_type_skip_modifiers(btf, arg->type, NULL);
9983 if (!btf_type_is_ptr(t))
9985 t = btf_type_skip_modifiers(btf, t->type, &res_id);
9988 return btf_types_are_same(btf, res_id, btf_vmlinux, kf_arg_btf_ids[type]);
9991 static bool is_kfunc_arg_dynptr(const struct btf *btf, const struct btf_param *arg)
9993 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_DYNPTR_ID);
9996 static bool is_kfunc_arg_list_head(const struct btf *btf, const struct btf_param *arg)
9998 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_HEAD_ID);
10001 static bool is_kfunc_arg_list_node(const struct btf *btf, const struct btf_param *arg)
10003 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_NODE_ID);
10006 static bool is_kfunc_arg_rbtree_root(const struct btf *btf, const struct btf_param *arg)
10008 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RB_ROOT_ID);
10011 static bool is_kfunc_arg_rbtree_node(const struct btf *btf, const struct btf_param *arg)
10013 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RB_NODE_ID);
10016 static bool is_kfunc_arg_callback(struct bpf_verifier_env *env, const struct btf *btf,
10017 const struct btf_param *arg)
10019 const struct btf_type *t;
10021 t = btf_type_resolve_func_ptr(btf, arg->type, NULL);
10028 /* Returns true if struct is composed of scalars, 4 levels of nesting allowed */
10029 static bool __btf_type_is_scalar_struct(struct bpf_verifier_env *env,
10030 const struct btf *btf,
10031 const struct btf_type *t, int rec)
10033 const struct btf_type *member_type;
10034 const struct btf_member *member;
10037 if (!btf_type_is_struct(t))
10040 for_each_member(i, t, member) {
10041 const struct btf_array *array;
10043 member_type = btf_type_skip_modifiers(btf, member->type, NULL);
10044 if (btf_type_is_struct(member_type)) {
10046 verbose(env, "max struct nesting depth exceeded\n");
10049 if (!__btf_type_is_scalar_struct(env, btf, member_type, rec + 1))
10053 if (btf_type_is_array(member_type)) {
10054 array = btf_array(member_type);
10055 if (!array->nelems)
10057 member_type = btf_type_skip_modifiers(btf, array->type, NULL);
10058 if (!btf_type_is_scalar(member_type))
10062 if (!btf_type_is_scalar(member_type))
10068 enum kfunc_ptr_arg_type {
10070 KF_ARG_PTR_TO_ALLOC_BTF_ID, /* Allocated object */
10071 KF_ARG_PTR_TO_REFCOUNTED_KPTR, /* Refcounted local kptr */
10072 KF_ARG_PTR_TO_DYNPTR,
10073 KF_ARG_PTR_TO_ITER,
10074 KF_ARG_PTR_TO_LIST_HEAD,
10075 KF_ARG_PTR_TO_LIST_NODE,
10076 KF_ARG_PTR_TO_BTF_ID, /* Also covers reg2btf_ids conversions */
10078 KF_ARG_PTR_TO_MEM_SIZE, /* Size derived from next argument, skip it */
10079 KF_ARG_PTR_TO_CALLBACK,
10080 KF_ARG_PTR_TO_RB_ROOT,
10081 KF_ARG_PTR_TO_RB_NODE,
10084 enum special_kfunc_type {
10085 KF_bpf_obj_new_impl,
10086 KF_bpf_obj_drop_impl,
10087 KF_bpf_refcount_acquire_impl,
10088 KF_bpf_list_push_front_impl,
10089 KF_bpf_list_push_back_impl,
10090 KF_bpf_list_pop_front,
10091 KF_bpf_list_pop_back,
10092 KF_bpf_cast_to_kern_ctx,
10093 KF_bpf_rdonly_cast,
10094 KF_bpf_rcu_read_lock,
10095 KF_bpf_rcu_read_unlock,
10096 KF_bpf_rbtree_remove,
10097 KF_bpf_rbtree_add_impl,
10098 KF_bpf_rbtree_first,
10099 KF_bpf_dynptr_from_skb,
10100 KF_bpf_dynptr_from_xdp,
10101 KF_bpf_dynptr_slice,
10102 KF_bpf_dynptr_slice_rdwr,
10103 KF_bpf_dynptr_clone,
10106 BTF_SET_START(special_kfunc_set)
10107 BTF_ID(func, bpf_obj_new_impl)
10108 BTF_ID(func, bpf_obj_drop_impl)
10109 BTF_ID(func, bpf_refcount_acquire_impl)
10110 BTF_ID(func, bpf_list_push_front_impl)
10111 BTF_ID(func, bpf_list_push_back_impl)
10112 BTF_ID(func, bpf_list_pop_front)
10113 BTF_ID(func, bpf_list_pop_back)
10114 BTF_ID(func, bpf_cast_to_kern_ctx)
10115 BTF_ID(func, bpf_rdonly_cast)
10116 BTF_ID(func, bpf_rbtree_remove)
10117 BTF_ID(func, bpf_rbtree_add_impl)
10118 BTF_ID(func, bpf_rbtree_first)
10119 BTF_ID(func, bpf_dynptr_from_skb)
10120 BTF_ID(func, bpf_dynptr_from_xdp)
10121 BTF_ID(func, bpf_dynptr_slice)
10122 BTF_ID(func, bpf_dynptr_slice_rdwr)
10123 BTF_ID(func, bpf_dynptr_clone)
10124 BTF_SET_END(special_kfunc_set)
10126 BTF_ID_LIST(special_kfunc_list)
10127 BTF_ID(func, bpf_obj_new_impl)
10128 BTF_ID(func, bpf_obj_drop_impl)
10129 BTF_ID(func, bpf_refcount_acquire_impl)
10130 BTF_ID(func, bpf_list_push_front_impl)
10131 BTF_ID(func, bpf_list_push_back_impl)
10132 BTF_ID(func, bpf_list_pop_front)
10133 BTF_ID(func, bpf_list_pop_back)
10134 BTF_ID(func, bpf_cast_to_kern_ctx)
10135 BTF_ID(func, bpf_rdonly_cast)
10136 BTF_ID(func, bpf_rcu_read_lock)
10137 BTF_ID(func, bpf_rcu_read_unlock)
10138 BTF_ID(func, bpf_rbtree_remove)
10139 BTF_ID(func, bpf_rbtree_add_impl)
10140 BTF_ID(func, bpf_rbtree_first)
10141 BTF_ID(func, bpf_dynptr_from_skb)
10142 BTF_ID(func, bpf_dynptr_from_xdp)
10143 BTF_ID(func, bpf_dynptr_slice)
10144 BTF_ID(func, bpf_dynptr_slice_rdwr)
10145 BTF_ID(func, bpf_dynptr_clone)
10147 static bool is_kfunc_ret_null(struct bpf_kfunc_call_arg_meta *meta)
10149 if (meta->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl] &&
10150 meta->arg_owning_ref) {
10154 return meta->kfunc_flags & KF_RET_NULL;
10157 static bool is_kfunc_bpf_rcu_read_lock(struct bpf_kfunc_call_arg_meta *meta)
10159 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_lock];
10162 static bool is_kfunc_bpf_rcu_read_unlock(struct bpf_kfunc_call_arg_meta *meta)
10164 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_unlock];
10167 static enum kfunc_ptr_arg_type
10168 get_kfunc_ptr_arg_type(struct bpf_verifier_env *env,
10169 struct bpf_kfunc_call_arg_meta *meta,
10170 const struct btf_type *t, const struct btf_type *ref_t,
10171 const char *ref_tname, const struct btf_param *args,
10172 int argno, int nargs)
10174 u32 regno = argno + 1;
10175 struct bpf_reg_state *regs = cur_regs(env);
10176 struct bpf_reg_state *reg = ®s[regno];
10177 bool arg_mem_size = false;
10179 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx])
10180 return KF_ARG_PTR_TO_CTX;
10182 /* In this function, we verify the kfunc's BTF as per the argument type,
10183 * leaving the rest of the verification with respect to the register
10184 * type to our caller. When a set of conditions hold in the BTF type of
10185 * arguments, we resolve it to a known kfunc_ptr_arg_type.
10187 if (btf_get_prog_ctx_type(&env->log, meta->btf, t, resolve_prog_type(env->prog), argno))
10188 return KF_ARG_PTR_TO_CTX;
10190 if (is_kfunc_arg_alloc_obj(meta->btf, &args[argno]))
10191 return KF_ARG_PTR_TO_ALLOC_BTF_ID;
10193 if (is_kfunc_arg_refcounted_kptr(meta->btf, &args[argno]))
10194 return KF_ARG_PTR_TO_REFCOUNTED_KPTR;
10196 if (is_kfunc_arg_dynptr(meta->btf, &args[argno]))
10197 return KF_ARG_PTR_TO_DYNPTR;
10199 if (is_kfunc_arg_iter(meta, argno))
10200 return KF_ARG_PTR_TO_ITER;
10202 if (is_kfunc_arg_list_head(meta->btf, &args[argno]))
10203 return KF_ARG_PTR_TO_LIST_HEAD;
10205 if (is_kfunc_arg_list_node(meta->btf, &args[argno]))
10206 return KF_ARG_PTR_TO_LIST_NODE;
10208 if (is_kfunc_arg_rbtree_root(meta->btf, &args[argno]))
10209 return KF_ARG_PTR_TO_RB_ROOT;
10211 if (is_kfunc_arg_rbtree_node(meta->btf, &args[argno]))
10212 return KF_ARG_PTR_TO_RB_NODE;
10214 if ((base_type(reg->type) == PTR_TO_BTF_ID || reg2btf_ids[base_type(reg->type)])) {
10215 if (!btf_type_is_struct(ref_t)) {
10216 verbose(env, "kernel function %s args#%d pointer type %s %s is not supported\n",
10217 meta->func_name, argno, btf_type_str(ref_t), ref_tname);
10220 return KF_ARG_PTR_TO_BTF_ID;
10223 if (is_kfunc_arg_callback(env, meta->btf, &args[argno]))
10224 return KF_ARG_PTR_TO_CALLBACK;
10227 if (argno + 1 < nargs &&
10228 (is_kfunc_arg_mem_size(meta->btf, &args[argno + 1], ®s[regno + 1]) ||
10229 is_kfunc_arg_const_mem_size(meta->btf, &args[argno + 1], ®s[regno + 1])))
10230 arg_mem_size = true;
10232 /* This is the catch all argument type of register types supported by
10233 * check_helper_mem_access. However, we only allow when argument type is
10234 * pointer to scalar, or struct composed (recursively) of scalars. When
10235 * arg_mem_size is true, the pointer can be void *.
10237 if (!btf_type_is_scalar(ref_t) && !__btf_type_is_scalar_struct(env, meta->btf, ref_t, 0) &&
10238 (arg_mem_size ? !btf_type_is_void(ref_t) : 1)) {
10239 verbose(env, "arg#%d pointer type %s %s must point to %sscalar, or struct with scalar\n",
10240 argno, btf_type_str(ref_t), ref_tname, arg_mem_size ? "void, " : "");
10243 return arg_mem_size ? KF_ARG_PTR_TO_MEM_SIZE : KF_ARG_PTR_TO_MEM;
10246 static int process_kf_arg_ptr_to_btf_id(struct bpf_verifier_env *env,
10247 struct bpf_reg_state *reg,
10248 const struct btf_type *ref_t,
10249 const char *ref_tname, u32 ref_id,
10250 struct bpf_kfunc_call_arg_meta *meta,
10253 const struct btf_type *reg_ref_t;
10254 bool strict_type_match = false;
10255 const struct btf *reg_btf;
10256 const char *reg_ref_tname;
10259 if (base_type(reg->type) == PTR_TO_BTF_ID) {
10260 reg_btf = reg->btf;
10261 reg_ref_id = reg->btf_id;
10263 reg_btf = btf_vmlinux;
10264 reg_ref_id = *reg2btf_ids[base_type(reg->type)];
10267 /* Enforce strict type matching for calls to kfuncs that are acquiring
10268 * or releasing a reference, or are no-cast aliases. We do _not_
10269 * enforce strict matching for plain KF_TRUSTED_ARGS kfuncs by default,
10270 * as we want to enable BPF programs to pass types that are bitwise
10271 * equivalent without forcing them to explicitly cast with something
10272 * like bpf_cast_to_kern_ctx().
10274 * For example, say we had a type like the following:
10276 * struct bpf_cpumask {
10277 * cpumask_t cpumask;
10278 * refcount_t usage;
10281 * Note that as specified in <linux/cpumask.h>, cpumask_t is typedef'ed
10282 * to a struct cpumask, so it would be safe to pass a struct
10283 * bpf_cpumask * to a kfunc expecting a struct cpumask *.
10285 * The philosophy here is similar to how we allow scalars of different
10286 * types to be passed to kfuncs as long as the size is the same. The
10287 * only difference here is that we're simply allowing
10288 * btf_struct_ids_match() to walk the struct at the 0th offset, and
10291 if (is_kfunc_acquire(meta) ||
10292 (is_kfunc_release(meta) && reg->ref_obj_id) ||
10293 btf_type_ids_nocast_alias(&env->log, reg_btf, reg_ref_id, meta->btf, ref_id))
10294 strict_type_match = true;
10296 WARN_ON_ONCE(is_kfunc_trusted_args(meta) && reg->off);
10298 reg_ref_t = btf_type_skip_modifiers(reg_btf, reg_ref_id, ®_ref_id);
10299 reg_ref_tname = btf_name_by_offset(reg_btf, reg_ref_t->name_off);
10300 if (!btf_struct_ids_match(&env->log, reg_btf, reg_ref_id, reg->off, meta->btf, ref_id, strict_type_match)) {
10301 verbose(env, "kernel function %s args#%d expected pointer to %s %s but R%d has a pointer to %s %s\n",
10302 meta->func_name, argno, btf_type_str(ref_t), ref_tname, argno + 1,
10303 btf_type_str(reg_ref_t), reg_ref_tname);
10309 static int ref_set_non_owning(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
10311 struct bpf_verifier_state *state = env->cur_state;
10313 if (!state->active_lock.ptr) {
10314 verbose(env, "verifier internal error: ref_set_non_owning w/o active lock\n");
10318 if (type_flag(reg->type) & NON_OWN_REF) {
10319 verbose(env, "verifier internal error: NON_OWN_REF already set\n");
10323 reg->type |= NON_OWN_REF;
10327 static int ref_convert_owning_non_owning(struct bpf_verifier_env *env, u32 ref_obj_id)
10329 struct bpf_func_state *state, *unused;
10330 struct bpf_reg_state *reg;
10333 state = cur_func(env);
10336 verbose(env, "verifier internal error: ref_obj_id is zero for "
10337 "owning -> non-owning conversion\n");
10341 for (i = 0; i < state->acquired_refs; i++) {
10342 if (state->refs[i].id != ref_obj_id)
10345 /* Clear ref_obj_id here so release_reference doesn't clobber
10348 bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
10349 if (reg->ref_obj_id == ref_obj_id) {
10350 reg->ref_obj_id = 0;
10351 ref_set_non_owning(env, reg);
10357 verbose(env, "verifier internal error: ref state missing for ref_obj_id\n");
10361 /* Implementation details:
10363 * Each register points to some region of memory, which we define as an
10364 * allocation. Each allocation may embed a bpf_spin_lock which protects any
10365 * special BPF objects (bpf_list_head, bpf_rb_root, etc.) part of the same
10366 * allocation. The lock and the data it protects are colocated in the same
10369 * Hence, everytime a register holds a pointer value pointing to such
10370 * allocation, the verifier preserves a unique reg->id for it.
10372 * The verifier remembers the lock 'ptr' and the lock 'id' whenever
10373 * bpf_spin_lock is called.
10375 * To enable this, lock state in the verifier captures two values:
10376 * active_lock.ptr = Register's type specific pointer
10377 * active_lock.id = A unique ID for each register pointer value
10379 * Currently, PTR_TO_MAP_VALUE and PTR_TO_BTF_ID | MEM_ALLOC are the two
10380 * supported register types.
10382 * The active_lock.ptr in case of map values is the reg->map_ptr, and in case of
10383 * allocated objects is the reg->btf pointer.
10385 * The active_lock.id is non-unique for maps supporting direct_value_addr, as we
10386 * can establish the provenance of the map value statically for each distinct
10387 * lookup into such maps. They always contain a single map value hence unique
10388 * IDs for each pseudo load pessimizes the algorithm and rejects valid programs.
10390 * So, in case of global variables, they use array maps with max_entries = 1,
10391 * hence their active_lock.ptr becomes map_ptr and id = 0 (since they all point
10392 * into the same map value as max_entries is 1, as described above).
10394 * In case of inner map lookups, the inner map pointer has same map_ptr as the
10395 * outer map pointer (in verifier context), but each lookup into an inner map
10396 * assigns a fresh reg->id to the lookup, so while lookups into distinct inner
10397 * maps from the same outer map share the same map_ptr as active_lock.ptr, they
10398 * will get different reg->id assigned to each lookup, hence different
10401 * In case of allocated objects, active_lock.ptr is the reg->btf, and the
10402 * reg->id is a unique ID preserved after the NULL pointer check on the pointer
10403 * returned from bpf_obj_new. Each allocation receives a new reg->id.
10405 static int check_reg_allocation_locked(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
10410 switch ((int)reg->type) {
10411 case PTR_TO_MAP_VALUE:
10412 ptr = reg->map_ptr;
10414 case PTR_TO_BTF_ID | MEM_ALLOC:
10418 verbose(env, "verifier internal error: unknown reg type for lock check\n");
10423 if (!env->cur_state->active_lock.ptr)
10425 if (env->cur_state->active_lock.ptr != ptr ||
10426 env->cur_state->active_lock.id != id) {
10427 verbose(env, "held lock and object are not in the same allocation\n");
10433 static bool is_bpf_list_api_kfunc(u32 btf_id)
10435 return btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
10436 btf_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
10437 btf_id == special_kfunc_list[KF_bpf_list_pop_front] ||
10438 btf_id == special_kfunc_list[KF_bpf_list_pop_back];
10441 static bool is_bpf_rbtree_api_kfunc(u32 btf_id)
10443 return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl] ||
10444 btf_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
10445 btf_id == special_kfunc_list[KF_bpf_rbtree_first];
10448 static bool is_bpf_graph_api_kfunc(u32 btf_id)
10450 return is_bpf_list_api_kfunc(btf_id) || is_bpf_rbtree_api_kfunc(btf_id) ||
10451 btf_id == special_kfunc_list[KF_bpf_refcount_acquire_impl];
10454 static bool is_callback_calling_kfunc(u32 btf_id)
10456 return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl];
10459 static bool is_rbtree_lock_required_kfunc(u32 btf_id)
10461 return is_bpf_rbtree_api_kfunc(btf_id);
10464 static bool check_kfunc_is_graph_root_api(struct bpf_verifier_env *env,
10465 enum btf_field_type head_field_type,
10470 switch (head_field_type) {
10471 case BPF_LIST_HEAD:
10472 ret = is_bpf_list_api_kfunc(kfunc_btf_id);
10475 ret = is_bpf_rbtree_api_kfunc(kfunc_btf_id);
10478 verbose(env, "verifier internal error: unexpected graph root argument type %s\n",
10479 btf_field_type_name(head_field_type));
10484 verbose(env, "verifier internal error: %s head arg for unknown kfunc\n",
10485 btf_field_type_name(head_field_type));
10489 static bool check_kfunc_is_graph_node_api(struct bpf_verifier_env *env,
10490 enum btf_field_type node_field_type,
10495 switch (node_field_type) {
10496 case BPF_LIST_NODE:
10497 ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
10498 kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_back_impl]);
10501 ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
10502 kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl]);
10505 verbose(env, "verifier internal error: unexpected graph node argument type %s\n",
10506 btf_field_type_name(node_field_type));
10511 verbose(env, "verifier internal error: %s node arg for unknown kfunc\n",
10512 btf_field_type_name(node_field_type));
10517 __process_kf_arg_ptr_to_graph_root(struct bpf_verifier_env *env,
10518 struct bpf_reg_state *reg, u32 regno,
10519 struct bpf_kfunc_call_arg_meta *meta,
10520 enum btf_field_type head_field_type,
10521 struct btf_field **head_field)
10523 const char *head_type_name;
10524 struct btf_field *field;
10525 struct btf_record *rec;
10528 if (meta->btf != btf_vmlinux) {
10529 verbose(env, "verifier internal error: unexpected btf mismatch in kfunc call\n");
10533 if (!check_kfunc_is_graph_root_api(env, head_field_type, meta->func_id))
10536 head_type_name = btf_field_type_name(head_field_type);
10537 if (!tnum_is_const(reg->var_off)) {
10539 "R%d doesn't have constant offset. %s has to be at the constant offset\n",
10540 regno, head_type_name);
10544 rec = reg_btf_record(reg);
10545 head_off = reg->off + reg->var_off.value;
10546 field = btf_record_find(rec, head_off, head_field_type);
10548 verbose(env, "%s not found at offset=%u\n", head_type_name, head_off);
10552 /* All functions require bpf_list_head to be protected using a bpf_spin_lock */
10553 if (check_reg_allocation_locked(env, reg)) {
10554 verbose(env, "bpf_spin_lock at off=%d must be held for %s\n",
10555 rec->spin_lock_off, head_type_name);
10560 verbose(env, "verifier internal error: repeating %s arg\n", head_type_name);
10563 *head_field = field;
10567 static int process_kf_arg_ptr_to_list_head(struct bpf_verifier_env *env,
10568 struct bpf_reg_state *reg, u32 regno,
10569 struct bpf_kfunc_call_arg_meta *meta)
10571 return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, BPF_LIST_HEAD,
10572 &meta->arg_list_head.field);
10575 static int process_kf_arg_ptr_to_rbtree_root(struct bpf_verifier_env *env,
10576 struct bpf_reg_state *reg, u32 regno,
10577 struct bpf_kfunc_call_arg_meta *meta)
10579 return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, BPF_RB_ROOT,
10580 &meta->arg_rbtree_root.field);
10584 __process_kf_arg_ptr_to_graph_node(struct bpf_verifier_env *env,
10585 struct bpf_reg_state *reg, u32 regno,
10586 struct bpf_kfunc_call_arg_meta *meta,
10587 enum btf_field_type head_field_type,
10588 enum btf_field_type node_field_type,
10589 struct btf_field **node_field)
10591 const char *node_type_name;
10592 const struct btf_type *et, *t;
10593 struct btf_field *field;
10596 if (meta->btf != btf_vmlinux) {
10597 verbose(env, "verifier internal error: unexpected btf mismatch in kfunc call\n");
10601 if (!check_kfunc_is_graph_node_api(env, node_field_type, meta->func_id))
10604 node_type_name = btf_field_type_name(node_field_type);
10605 if (!tnum_is_const(reg->var_off)) {
10607 "R%d doesn't have constant offset. %s has to be at the constant offset\n",
10608 regno, node_type_name);
10612 node_off = reg->off + reg->var_off.value;
10613 field = reg_find_field_offset(reg, node_off, node_field_type);
10614 if (!field || field->offset != node_off) {
10615 verbose(env, "%s not found at offset=%u\n", node_type_name, node_off);
10619 field = *node_field;
10621 et = btf_type_by_id(field->graph_root.btf, field->graph_root.value_btf_id);
10622 t = btf_type_by_id(reg->btf, reg->btf_id);
10623 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, 0, field->graph_root.btf,
10624 field->graph_root.value_btf_id, true)) {
10625 verbose(env, "operation on %s expects arg#1 %s at offset=%d "
10626 "in struct %s, but arg is at offset=%d in struct %s\n",
10627 btf_field_type_name(head_field_type),
10628 btf_field_type_name(node_field_type),
10629 field->graph_root.node_offset,
10630 btf_name_by_offset(field->graph_root.btf, et->name_off),
10631 node_off, btf_name_by_offset(reg->btf, t->name_off));
10634 meta->arg_btf = reg->btf;
10635 meta->arg_btf_id = reg->btf_id;
10637 if (node_off != field->graph_root.node_offset) {
10638 verbose(env, "arg#1 offset=%d, but expected %s at offset=%d in struct %s\n",
10639 node_off, btf_field_type_name(node_field_type),
10640 field->graph_root.node_offset,
10641 btf_name_by_offset(field->graph_root.btf, et->name_off));
10648 static int process_kf_arg_ptr_to_list_node(struct bpf_verifier_env *env,
10649 struct bpf_reg_state *reg, u32 regno,
10650 struct bpf_kfunc_call_arg_meta *meta)
10652 return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta,
10653 BPF_LIST_HEAD, BPF_LIST_NODE,
10654 &meta->arg_list_head.field);
10657 static int process_kf_arg_ptr_to_rbtree_node(struct bpf_verifier_env *env,
10658 struct bpf_reg_state *reg, u32 regno,
10659 struct bpf_kfunc_call_arg_meta *meta)
10661 return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta,
10662 BPF_RB_ROOT, BPF_RB_NODE,
10663 &meta->arg_rbtree_root.field);
10666 static int check_kfunc_args(struct bpf_verifier_env *env, struct bpf_kfunc_call_arg_meta *meta,
10669 const char *func_name = meta->func_name, *ref_tname;
10670 const struct btf *btf = meta->btf;
10671 const struct btf_param *args;
10672 struct btf_record *rec;
10676 args = (const struct btf_param *)(meta->func_proto + 1);
10677 nargs = btf_type_vlen(meta->func_proto);
10678 if (nargs > MAX_BPF_FUNC_REG_ARGS) {
10679 verbose(env, "Function %s has %d > %d args\n", func_name, nargs,
10680 MAX_BPF_FUNC_REG_ARGS);
10684 /* Check that BTF function arguments match actual types that the
10687 for (i = 0; i < nargs; i++) {
10688 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[i + 1];
10689 const struct btf_type *t, *ref_t, *resolve_ret;
10690 enum bpf_arg_type arg_type = ARG_DONTCARE;
10691 u32 regno = i + 1, ref_id, type_size;
10692 bool is_ret_buf_sz = false;
10695 t = btf_type_skip_modifiers(btf, args[i].type, NULL);
10697 if (is_kfunc_arg_ignore(btf, &args[i]))
10700 if (btf_type_is_scalar(t)) {
10701 if (reg->type != SCALAR_VALUE) {
10702 verbose(env, "R%d is not a scalar\n", regno);
10706 if (is_kfunc_arg_constant(meta->btf, &args[i])) {
10707 if (meta->arg_constant.found) {
10708 verbose(env, "verifier internal error: only one constant argument permitted\n");
10711 if (!tnum_is_const(reg->var_off)) {
10712 verbose(env, "R%d must be a known constant\n", regno);
10715 ret = mark_chain_precision(env, regno);
10718 meta->arg_constant.found = true;
10719 meta->arg_constant.value = reg->var_off.value;
10720 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdonly_buf_size")) {
10721 meta->r0_rdonly = true;
10722 is_ret_buf_sz = true;
10723 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdwr_buf_size")) {
10724 is_ret_buf_sz = true;
10727 if (is_ret_buf_sz) {
10728 if (meta->r0_size) {
10729 verbose(env, "2 or more rdonly/rdwr_buf_size parameters for kfunc");
10733 if (!tnum_is_const(reg->var_off)) {
10734 verbose(env, "R%d is not a const\n", regno);
10738 meta->r0_size = reg->var_off.value;
10739 ret = mark_chain_precision(env, regno);
10746 if (!btf_type_is_ptr(t)) {
10747 verbose(env, "Unrecognized arg#%d type %s\n", i, btf_type_str(t));
10751 if ((is_kfunc_trusted_args(meta) || is_kfunc_rcu(meta)) &&
10752 (register_is_null(reg) || type_may_be_null(reg->type))) {
10753 verbose(env, "Possibly NULL pointer passed to trusted arg%d\n", i);
10757 if (reg->ref_obj_id) {
10758 if (is_kfunc_release(meta) && meta->ref_obj_id) {
10759 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
10760 regno, reg->ref_obj_id,
10764 meta->ref_obj_id = reg->ref_obj_id;
10765 if (is_kfunc_release(meta))
10766 meta->release_regno = regno;
10769 ref_t = btf_type_skip_modifiers(btf, t->type, &ref_id);
10770 ref_tname = btf_name_by_offset(btf, ref_t->name_off);
10772 kf_arg_type = get_kfunc_ptr_arg_type(env, meta, t, ref_t, ref_tname, args, i, nargs);
10773 if (kf_arg_type < 0)
10774 return kf_arg_type;
10776 switch (kf_arg_type) {
10777 case KF_ARG_PTR_TO_ALLOC_BTF_ID:
10778 case KF_ARG_PTR_TO_BTF_ID:
10779 if (!is_kfunc_trusted_args(meta) && !is_kfunc_rcu(meta))
10782 if (!is_trusted_reg(reg)) {
10783 if (!is_kfunc_rcu(meta)) {
10784 verbose(env, "R%d must be referenced or trusted\n", regno);
10787 if (!is_rcu_reg(reg)) {
10788 verbose(env, "R%d must be a rcu pointer\n", regno);
10794 case KF_ARG_PTR_TO_CTX:
10795 /* Trusted arguments have the same offset checks as release arguments */
10796 arg_type |= OBJ_RELEASE;
10798 case KF_ARG_PTR_TO_DYNPTR:
10799 case KF_ARG_PTR_TO_ITER:
10800 case KF_ARG_PTR_TO_LIST_HEAD:
10801 case KF_ARG_PTR_TO_LIST_NODE:
10802 case KF_ARG_PTR_TO_RB_ROOT:
10803 case KF_ARG_PTR_TO_RB_NODE:
10804 case KF_ARG_PTR_TO_MEM:
10805 case KF_ARG_PTR_TO_MEM_SIZE:
10806 case KF_ARG_PTR_TO_CALLBACK:
10807 case KF_ARG_PTR_TO_REFCOUNTED_KPTR:
10808 /* Trusted by default */
10815 if (is_kfunc_release(meta) && reg->ref_obj_id)
10816 arg_type |= OBJ_RELEASE;
10817 ret = check_func_arg_reg_off(env, reg, regno, arg_type);
10821 switch (kf_arg_type) {
10822 case KF_ARG_PTR_TO_CTX:
10823 if (reg->type != PTR_TO_CTX) {
10824 verbose(env, "arg#%d expected pointer to ctx, but got %s\n", i, btf_type_str(t));
10828 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
10829 ret = get_kern_ctx_btf_id(&env->log, resolve_prog_type(env->prog));
10832 meta->ret_btf_id = ret;
10835 case KF_ARG_PTR_TO_ALLOC_BTF_ID:
10836 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
10837 verbose(env, "arg#%d expected pointer to allocated object\n", i);
10840 if (!reg->ref_obj_id) {
10841 verbose(env, "allocated object must be referenced\n");
10844 if (meta->btf == btf_vmlinux &&
10845 meta->func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
10846 meta->arg_btf = reg->btf;
10847 meta->arg_btf_id = reg->btf_id;
10850 case KF_ARG_PTR_TO_DYNPTR:
10852 enum bpf_arg_type dynptr_arg_type = ARG_PTR_TO_DYNPTR;
10853 int clone_ref_obj_id = 0;
10855 if (reg->type != PTR_TO_STACK &&
10856 reg->type != CONST_PTR_TO_DYNPTR) {
10857 verbose(env, "arg#%d expected pointer to stack or dynptr_ptr\n", i);
10861 if (reg->type == CONST_PTR_TO_DYNPTR)
10862 dynptr_arg_type |= MEM_RDONLY;
10864 if (is_kfunc_arg_uninit(btf, &args[i]))
10865 dynptr_arg_type |= MEM_UNINIT;
10867 if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) {
10868 dynptr_arg_type |= DYNPTR_TYPE_SKB;
10869 } else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_xdp]) {
10870 dynptr_arg_type |= DYNPTR_TYPE_XDP;
10871 } else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_clone] &&
10872 (dynptr_arg_type & MEM_UNINIT)) {
10873 enum bpf_dynptr_type parent_type = meta->initialized_dynptr.type;
10875 if (parent_type == BPF_DYNPTR_TYPE_INVALID) {
10876 verbose(env, "verifier internal error: no dynptr type for parent of clone\n");
10880 dynptr_arg_type |= (unsigned int)get_dynptr_type_flag(parent_type);
10881 clone_ref_obj_id = meta->initialized_dynptr.ref_obj_id;
10882 if (dynptr_type_refcounted(parent_type) && !clone_ref_obj_id) {
10883 verbose(env, "verifier internal error: missing ref obj id for parent of clone\n");
10888 ret = process_dynptr_func(env, regno, insn_idx, dynptr_arg_type, clone_ref_obj_id);
10892 if (!(dynptr_arg_type & MEM_UNINIT)) {
10893 int id = dynptr_id(env, reg);
10896 verbose(env, "verifier internal error: failed to obtain dynptr id\n");
10899 meta->initialized_dynptr.id = id;
10900 meta->initialized_dynptr.type = dynptr_get_type(env, reg);
10901 meta->initialized_dynptr.ref_obj_id = dynptr_ref_obj_id(env, reg);
10906 case KF_ARG_PTR_TO_ITER:
10907 ret = process_iter_arg(env, regno, insn_idx, meta);
10911 case KF_ARG_PTR_TO_LIST_HEAD:
10912 if (reg->type != PTR_TO_MAP_VALUE &&
10913 reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
10914 verbose(env, "arg#%d expected pointer to map value or allocated object\n", i);
10917 if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
10918 verbose(env, "allocated object must be referenced\n");
10921 ret = process_kf_arg_ptr_to_list_head(env, reg, regno, meta);
10925 case KF_ARG_PTR_TO_RB_ROOT:
10926 if (reg->type != PTR_TO_MAP_VALUE &&
10927 reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
10928 verbose(env, "arg#%d expected pointer to map value or allocated object\n", i);
10931 if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
10932 verbose(env, "allocated object must be referenced\n");
10935 ret = process_kf_arg_ptr_to_rbtree_root(env, reg, regno, meta);
10939 case KF_ARG_PTR_TO_LIST_NODE:
10940 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
10941 verbose(env, "arg#%d expected pointer to allocated object\n", i);
10944 if (!reg->ref_obj_id) {
10945 verbose(env, "allocated object must be referenced\n");
10948 ret = process_kf_arg_ptr_to_list_node(env, reg, regno, meta);
10952 case KF_ARG_PTR_TO_RB_NODE:
10953 if (meta->func_id == special_kfunc_list[KF_bpf_rbtree_remove]) {
10954 if (!type_is_non_owning_ref(reg->type) || reg->ref_obj_id) {
10955 verbose(env, "rbtree_remove node input must be non-owning ref\n");
10958 if (in_rbtree_lock_required_cb(env)) {
10959 verbose(env, "rbtree_remove not allowed in rbtree cb\n");
10963 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
10964 verbose(env, "arg#%d expected pointer to allocated object\n", i);
10967 if (!reg->ref_obj_id) {
10968 verbose(env, "allocated object must be referenced\n");
10973 ret = process_kf_arg_ptr_to_rbtree_node(env, reg, regno, meta);
10977 case KF_ARG_PTR_TO_BTF_ID:
10978 /* Only base_type is checked, further checks are done here */
10979 if ((base_type(reg->type) != PTR_TO_BTF_ID ||
10980 (bpf_type_has_unsafe_modifiers(reg->type) && !is_rcu_reg(reg))) &&
10981 !reg2btf_ids[base_type(reg->type)]) {
10982 verbose(env, "arg#%d is %s ", i, reg_type_str(env, reg->type));
10983 verbose(env, "expected %s or socket\n",
10984 reg_type_str(env, base_type(reg->type) |
10985 (type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS)));
10988 ret = process_kf_arg_ptr_to_btf_id(env, reg, ref_t, ref_tname, ref_id, meta, i);
10992 case KF_ARG_PTR_TO_MEM:
10993 resolve_ret = btf_resolve_size(btf, ref_t, &type_size);
10994 if (IS_ERR(resolve_ret)) {
10995 verbose(env, "arg#%d reference type('%s %s') size cannot be determined: %ld\n",
10996 i, btf_type_str(ref_t), ref_tname, PTR_ERR(resolve_ret));
10999 ret = check_mem_reg(env, reg, regno, type_size);
11003 case KF_ARG_PTR_TO_MEM_SIZE:
11005 struct bpf_reg_state *buff_reg = ®s[regno];
11006 const struct btf_param *buff_arg = &args[i];
11007 struct bpf_reg_state *size_reg = ®s[regno + 1];
11008 const struct btf_param *size_arg = &args[i + 1];
11010 if (!register_is_null(buff_reg) || !is_kfunc_arg_optional(meta->btf, buff_arg)) {
11011 ret = check_kfunc_mem_size_reg(env, size_reg, regno + 1);
11013 verbose(env, "arg#%d arg#%d memory, len pair leads to invalid memory access\n", i, i + 1);
11018 if (is_kfunc_arg_const_mem_size(meta->btf, size_arg, size_reg)) {
11019 if (meta->arg_constant.found) {
11020 verbose(env, "verifier internal error: only one constant argument permitted\n");
11023 if (!tnum_is_const(size_reg->var_off)) {
11024 verbose(env, "R%d must be a known constant\n", regno + 1);
11027 meta->arg_constant.found = true;
11028 meta->arg_constant.value = size_reg->var_off.value;
11031 /* Skip next '__sz' or '__szk' argument */
11035 case KF_ARG_PTR_TO_CALLBACK:
11036 meta->subprogno = reg->subprogno;
11038 case KF_ARG_PTR_TO_REFCOUNTED_KPTR:
11039 if (!type_is_ptr_alloc_obj(reg->type)) {
11040 verbose(env, "arg#%d is neither owning or non-owning ref\n", i);
11043 if (!type_is_non_owning_ref(reg->type))
11044 meta->arg_owning_ref = true;
11046 rec = reg_btf_record(reg);
11048 verbose(env, "verifier internal error: Couldn't find btf_record\n");
11052 if (rec->refcount_off < 0) {
11053 verbose(env, "arg#%d doesn't point to a type with bpf_refcount field\n", i);
11056 if (rec->refcount_off >= 0) {
11057 verbose(env, "bpf_refcount_acquire calls are disabled for now\n");
11060 meta->arg_btf = reg->btf;
11061 meta->arg_btf_id = reg->btf_id;
11066 if (is_kfunc_release(meta) && !meta->release_regno) {
11067 verbose(env, "release kernel function %s expects refcounted PTR_TO_BTF_ID\n",
11075 static int fetch_kfunc_meta(struct bpf_verifier_env *env,
11076 struct bpf_insn *insn,
11077 struct bpf_kfunc_call_arg_meta *meta,
11078 const char **kfunc_name)
11080 const struct btf_type *func, *func_proto;
11081 u32 func_id, *kfunc_flags;
11082 const char *func_name;
11083 struct btf *desc_btf;
11086 *kfunc_name = NULL;
11091 desc_btf = find_kfunc_desc_btf(env, insn->off);
11092 if (IS_ERR(desc_btf))
11093 return PTR_ERR(desc_btf);
11095 func_id = insn->imm;
11096 func = btf_type_by_id(desc_btf, func_id);
11097 func_name = btf_name_by_offset(desc_btf, func->name_off);
11099 *kfunc_name = func_name;
11100 func_proto = btf_type_by_id(desc_btf, func->type);
11102 kfunc_flags = btf_kfunc_id_set_contains(desc_btf, func_id, env->prog);
11103 if (!kfunc_flags) {
11107 memset(meta, 0, sizeof(*meta));
11108 meta->btf = desc_btf;
11109 meta->func_id = func_id;
11110 meta->kfunc_flags = *kfunc_flags;
11111 meta->func_proto = func_proto;
11112 meta->func_name = func_name;
11117 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
11120 const struct btf_type *t, *ptr_type;
11121 u32 i, nargs, ptr_type_id, release_ref_obj_id;
11122 struct bpf_reg_state *regs = cur_regs(env);
11123 const char *func_name, *ptr_type_name;
11124 bool sleepable, rcu_lock, rcu_unlock;
11125 struct bpf_kfunc_call_arg_meta meta;
11126 struct bpf_insn_aux_data *insn_aux;
11127 int err, insn_idx = *insn_idx_p;
11128 const struct btf_param *args;
11129 const struct btf_type *ret_t;
11130 struct btf *desc_btf;
11132 /* skip for now, but return error when we find this in fixup_kfunc_call */
11136 err = fetch_kfunc_meta(env, insn, &meta, &func_name);
11137 if (err == -EACCES && func_name)
11138 verbose(env, "calling kernel function %s is not allowed\n", func_name);
11141 desc_btf = meta.btf;
11142 insn_aux = &env->insn_aux_data[insn_idx];
11144 insn_aux->is_iter_next = is_iter_next_kfunc(&meta);
11146 if (is_kfunc_destructive(&meta) && !capable(CAP_SYS_BOOT)) {
11147 verbose(env, "destructive kfunc calls require CAP_SYS_BOOT capability\n");
11151 sleepable = is_kfunc_sleepable(&meta);
11152 if (sleepable && !env->prog->aux->sleepable) {
11153 verbose(env, "program must be sleepable to call sleepable kfunc %s\n", func_name);
11157 rcu_lock = is_kfunc_bpf_rcu_read_lock(&meta);
11158 rcu_unlock = is_kfunc_bpf_rcu_read_unlock(&meta);
11160 if (env->cur_state->active_rcu_lock) {
11161 struct bpf_func_state *state;
11162 struct bpf_reg_state *reg;
11165 verbose(env, "nested rcu read lock (kernel function %s)\n", func_name);
11167 } else if (rcu_unlock) {
11168 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
11169 if (reg->type & MEM_RCU) {
11170 reg->type &= ~(MEM_RCU | PTR_MAYBE_NULL);
11171 reg->type |= PTR_UNTRUSTED;
11174 env->cur_state->active_rcu_lock = false;
11175 } else if (sleepable) {
11176 verbose(env, "kernel func %s is sleepable within rcu_read_lock region\n", func_name);
11179 } else if (rcu_lock) {
11180 env->cur_state->active_rcu_lock = true;
11181 } else if (rcu_unlock) {
11182 verbose(env, "unmatched rcu read unlock (kernel function %s)\n", func_name);
11186 /* Check the arguments */
11187 err = check_kfunc_args(env, &meta, insn_idx);
11190 /* In case of release function, we get register number of refcounted
11191 * PTR_TO_BTF_ID in bpf_kfunc_arg_meta, do the release now.
11193 if (meta.release_regno) {
11194 err = release_reference(env, regs[meta.release_regno].ref_obj_id);
11196 verbose(env, "kfunc %s#%d reference has not been acquired before\n",
11197 func_name, meta.func_id);
11202 if (meta.func_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
11203 meta.func_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
11204 meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
11205 release_ref_obj_id = regs[BPF_REG_2].ref_obj_id;
11206 insn_aux->insert_off = regs[BPF_REG_2].off;
11207 insn_aux->kptr_struct_meta = btf_find_struct_meta(meta.arg_btf, meta.arg_btf_id);
11208 err = ref_convert_owning_non_owning(env, release_ref_obj_id);
11210 verbose(env, "kfunc %s#%d conversion of owning ref to non-owning failed\n",
11211 func_name, meta.func_id);
11215 err = release_reference(env, release_ref_obj_id);
11217 verbose(env, "kfunc %s#%d reference has not been acquired before\n",
11218 func_name, meta.func_id);
11223 if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
11224 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
11225 set_rbtree_add_callback_state);
11227 verbose(env, "kfunc %s#%d failed callback verification\n",
11228 func_name, meta.func_id);
11233 for (i = 0; i < CALLER_SAVED_REGS; i++)
11234 mark_reg_not_init(env, regs, caller_saved[i]);
11236 /* Check return type */
11237 t = btf_type_skip_modifiers(desc_btf, meta.func_proto->type, NULL);
11239 if (is_kfunc_acquire(&meta) && !btf_type_is_struct_ptr(meta.btf, t)) {
11240 /* Only exception is bpf_obj_new_impl */
11241 if (meta.btf != btf_vmlinux ||
11242 (meta.func_id != special_kfunc_list[KF_bpf_obj_new_impl] &&
11243 meta.func_id != special_kfunc_list[KF_bpf_refcount_acquire_impl])) {
11244 verbose(env, "acquire kernel function does not return PTR_TO_BTF_ID\n");
11249 if (btf_type_is_scalar(t)) {
11250 mark_reg_unknown(env, regs, BPF_REG_0);
11251 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
11252 } else if (btf_type_is_ptr(t)) {
11253 ptr_type = btf_type_skip_modifiers(desc_btf, t->type, &ptr_type_id);
11255 if (meta.btf == btf_vmlinux && btf_id_set_contains(&special_kfunc_set, meta.func_id)) {
11256 if (meta.func_id == special_kfunc_list[KF_bpf_obj_new_impl]) {
11257 struct btf *ret_btf;
11260 if (unlikely(!bpf_global_ma_set))
11263 if (((u64)(u32)meta.arg_constant.value) != meta.arg_constant.value) {
11264 verbose(env, "local type ID argument must be in range [0, U32_MAX]\n");
11268 ret_btf = env->prog->aux->btf;
11269 ret_btf_id = meta.arg_constant.value;
11271 /* This may be NULL due to user not supplying a BTF */
11273 verbose(env, "bpf_obj_new requires prog BTF\n");
11277 ret_t = btf_type_by_id(ret_btf, ret_btf_id);
11278 if (!ret_t || !__btf_type_is_struct(ret_t)) {
11279 verbose(env, "bpf_obj_new type ID argument must be of a struct\n");
11283 mark_reg_known_zero(env, regs, BPF_REG_0);
11284 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
11285 regs[BPF_REG_0].btf = ret_btf;
11286 regs[BPF_REG_0].btf_id = ret_btf_id;
11288 insn_aux->obj_new_size = ret_t->size;
11289 insn_aux->kptr_struct_meta =
11290 btf_find_struct_meta(ret_btf, ret_btf_id);
11291 } else if (meta.func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) {
11292 mark_reg_known_zero(env, regs, BPF_REG_0);
11293 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
11294 regs[BPF_REG_0].btf = meta.arg_btf;
11295 regs[BPF_REG_0].btf_id = meta.arg_btf_id;
11297 insn_aux->kptr_struct_meta =
11298 btf_find_struct_meta(meta.arg_btf,
11300 } else if (meta.func_id == special_kfunc_list[KF_bpf_list_pop_front] ||
11301 meta.func_id == special_kfunc_list[KF_bpf_list_pop_back]) {
11302 struct btf_field *field = meta.arg_list_head.field;
11304 mark_reg_graph_node(regs, BPF_REG_0, &field->graph_root);
11305 } else if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
11306 meta.func_id == special_kfunc_list[KF_bpf_rbtree_first]) {
11307 struct btf_field *field = meta.arg_rbtree_root.field;
11309 mark_reg_graph_node(regs, BPF_REG_0, &field->graph_root);
11310 } else if (meta.func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
11311 mark_reg_known_zero(env, regs, BPF_REG_0);
11312 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_TRUSTED;
11313 regs[BPF_REG_0].btf = desc_btf;
11314 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
11315 } else if (meta.func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
11316 ret_t = btf_type_by_id(desc_btf, meta.arg_constant.value);
11317 if (!ret_t || !btf_type_is_struct(ret_t)) {
11319 "kfunc bpf_rdonly_cast type ID argument must be of a struct\n");
11323 mark_reg_known_zero(env, regs, BPF_REG_0);
11324 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
11325 regs[BPF_REG_0].btf = desc_btf;
11326 regs[BPF_REG_0].btf_id = meta.arg_constant.value;
11327 } else if (meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice] ||
11328 meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice_rdwr]) {
11329 enum bpf_type_flag type_flag = get_dynptr_type_flag(meta.initialized_dynptr.type);
11331 mark_reg_known_zero(env, regs, BPF_REG_0);
11333 if (!meta.arg_constant.found) {
11334 verbose(env, "verifier internal error: bpf_dynptr_slice(_rdwr) no constant size\n");
11338 regs[BPF_REG_0].mem_size = meta.arg_constant.value;
11340 /* PTR_MAYBE_NULL will be added when is_kfunc_ret_null is checked */
11341 regs[BPF_REG_0].type = PTR_TO_MEM | type_flag;
11343 if (meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice]) {
11344 regs[BPF_REG_0].type |= MEM_RDONLY;
11346 /* this will set env->seen_direct_write to true */
11347 if (!may_access_direct_pkt_data(env, NULL, BPF_WRITE)) {
11348 verbose(env, "the prog does not allow writes to packet data\n");
11353 if (!meta.initialized_dynptr.id) {
11354 verbose(env, "verifier internal error: no dynptr id\n");
11357 regs[BPF_REG_0].dynptr_id = meta.initialized_dynptr.id;
11359 /* we don't need to set BPF_REG_0's ref obj id
11360 * because packet slices are not refcounted (see
11361 * dynptr_type_refcounted)
11364 verbose(env, "kernel function %s unhandled dynamic return type\n",
11368 } else if (!__btf_type_is_struct(ptr_type)) {
11369 if (!meta.r0_size) {
11372 if (!IS_ERR(btf_resolve_size(desc_btf, ptr_type, &sz))) {
11374 meta.r0_rdonly = true;
11377 if (!meta.r0_size) {
11378 ptr_type_name = btf_name_by_offset(desc_btf,
11379 ptr_type->name_off);
11381 "kernel function %s returns pointer type %s %s is not supported\n",
11383 btf_type_str(ptr_type),
11388 mark_reg_known_zero(env, regs, BPF_REG_0);
11389 regs[BPF_REG_0].type = PTR_TO_MEM;
11390 regs[BPF_REG_0].mem_size = meta.r0_size;
11392 if (meta.r0_rdonly)
11393 regs[BPF_REG_0].type |= MEM_RDONLY;
11395 /* Ensures we don't access the memory after a release_reference() */
11396 if (meta.ref_obj_id)
11397 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
11399 mark_reg_known_zero(env, regs, BPF_REG_0);
11400 regs[BPF_REG_0].btf = desc_btf;
11401 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
11402 regs[BPF_REG_0].btf_id = ptr_type_id;
11405 if (is_kfunc_ret_null(&meta)) {
11406 regs[BPF_REG_0].type |= PTR_MAYBE_NULL;
11407 /* For mark_ptr_or_null_reg, see 93c230e3f5bd6 */
11408 regs[BPF_REG_0].id = ++env->id_gen;
11410 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
11411 if (is_kfunc_acquire(&meta)) {
11412 int id = acquire_reference_state(env, insn_idx);
11416 if (is_kfunc_ret_null(&meta))
11417 regs[BPF_REG_0].id = id;
11418 regs[BPF_REG_0].ref_obj_id = id;
11419 } else if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_first]) {
11420 ref_set_non_owning(env, ®s[BPF_REG_0]);
11423 if (reg_may_point_to_spin_lock(®s[BPF_REG_0]) && !regs[BPF_REG_0].id)
11424 regs[BPF_REG_0].id = ++env->id_gen;
11425 } else if (btf_type_is_void(t)) {
11426 if (meta.btf == btf_vmlinux && btf_id_set_contains(&special_kfunc_set, meta.func_id)) {
11427 if (meta.func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
11428 insn_aux->kptr_struct_meta =
11429 btf_find_struct_meta(meta.arg_btf,
11435 nargs = btf_type_vlen(meta.func_proto);
11436 args = (const struct btf_param *)(meta.func_proto + 1);
11437 for (i = 0; i < nargs; i++) {
11440 t = btf_type_skip_modifiers(desc_btf, args[i].type, NULL);
11441 if (btf_type_is_ptr(t))
11442 mark_btf_func_reg_size(env, regno, sizeof(void *));
11444 /* scalar. ensured by btf_check_kfunc_arg_match() */
11445 mark_btf_func_reg_size(env, regno, t->size);
11448 if (is_iter_next_kfunc(&meta)) {
11449 err = process_iter_next_call(env, insn_idx, &meta);
11457 static bool signed_add_overflows(s64 a, s64 b)
11459 /* Do the add in u64, where overflow is well-defined */
11460 s64 res = (s64)((u64)a + (u64)b);
11467 static bool signed_add32_overflows(s32 a, s32 b)
11469 /* Do the add in u32, where overflow is well-defined */
11470 s32 res = (s32)((u32)a + (u32)b);
11477 static bool signed_sub_overflows(s64 a, s64 b)
11479 /* Do the sub in u64, where overflow is well-defined */
11480 s64 res = (s64)((u64)a - (u64)b);
11487 static bool signed_sub32_overflows(s32 a, s32 b)
11489 /* Do the sub in u32, where overflow is well-defined */
11490 s32 res = (s32)((u32)a - (u32)b);
11497 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
11498 const struct bpf_reg_state *reg,
11499 enum bpf_reg_type type)
11501 bool known = tnum_is_const(reg->var_off);
11502 s64 val = reg->var_off.value;
11503 s64 smin = reg->smin_value;
11505 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
11506 verbose(env, "math between %s pointer and %lld is not allowed\n",
11507 reg_type_str(env, type), val);
11511 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
11512 verbose(env, "%s pointer offset %d is not allowed\n",
11513 reg_type_str(env, type), reg->off);
11517 if (smin == S64_MIN) {
11518 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
11519 reg_type_str(env, type));
11523 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
11524 verbose(env, "value %lld makes %s pointer be out of bounds\n",
11525 smin, reg_type_str(env, type));
11533 REASON_BOUNDS = -1,
11540 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
11541 u32 *alu_limit, bool mask_to_left)
11543 u32 max = 0, ptr_limit = 0;
11545 switch (ptr_reg->type) {
11547 /* Offset 0 is out-of-bounds, but acceptable start for the
11548 * left direction, see BPF_REG_FP. Also, unknown scalar
11549 * offset where we would need to deal with min/max bounds is
11550 * currently prohibited for unprivileged.
11552 max = MAX_BPF_STACK + mask_to_left;
11553 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
11555 case PTR_TO_MAP_VALUE:
11556 max = ptr_reg->map_ptr->value_size;
11557 ptr_limit = (mask_to_left ?
11558 ptr_reg->smin_value :
11559 ptr_reg->umax_value) + ptr_reg->off;
11562 return REASON_TYPE;
11565 if (ptr_limit >= max)
11566 return REASON_LIMIT;
11567 *alu_limit = ptr_limit;
11571 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
11572 const struct bpf_insn *insn)
11574 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
11577 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
11578 u32 alu_state, u32 alu_limit)
11580 /* If we arrived here from different branches with different
11581 * state or limits to sanitize, then this won't work.
11583 if (aux->alu_state &&
11584 (aux->alu_state != alu_state ||
11585 aux->alu_limit != alu_limit))
11586 return REASON_PATHS;
11588 /* Corresponding fixup done in do_misc_fixups(). */
11589 aux->alu_state = alu_state;
11590 aux->alu_limit = alu_limit;
11594 static int sanitize_val_alu(struct bpf_verifier_env *env,
11595 struct bpf_insn *insn)
11597 struct bpf_insn_aux_data *aux = cur_aux(env);
11599 if (can_skip_alu_sanitation(env, insn))
11602 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
11605 static bool sanitize_needed(u8 opcode)
11607 return opcode == BPF_ADD || opcode == BPF_SUB;
11610 struct bpf_sanitize_info {
11611 struct bpf_insn_aux_data aux;
11615 static struct bpf_verifier_state *
11616 sanitize_speculative_path(struct bpf_verifier_env *env,
11617 const struct bpf_insn *insn,
11618 u32 next_idx, u32 curr_idx)
11620 struct bpf_verifier_state *branch;
11621 struct bpf_reg_state *regs;
11623 branch = push_stack(env, next_idx, curr_idx, true);
11624 if (branch && insn) {
11625 regs = branch->frame[branch->curframe]->regs;
11626 if (BPF_SRC(insn->code) == BPF_K) {
11627 mark_reg_unknown(env, regs, insn->dst_reg);
11628 } else if (BPF_SRC(insn->code) == BPF_X) {
11629 mark_reg_unknown(env, regs, insn->dst_reg);
11630 mark_reg_unknown(env, regs, insn->src_reg);
11636 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
11637 struct bpf_insn *insn,
11638 const struct bpf_reg_state *ptr_reg,
11639 const struct bpf_reg_state *off_reg,
11640 struct bpf_reg_state *dst_reg,
11641 struct bpf_sanitize_info *info,
11642 const bool commit_window)
11644 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
11645 struct bpf_verifier_state *vstate = env->cur_state;
11646 bool off_is_imm = tnum_is_const(off_reg->var_off);
11647 bool off_is_neg = off_reg->smin_value < 0;
11648 bool ptr_is_dst_reg = ptr_reg == dst_reg;
11649 u8 opcode = BPF_OP(insn->code);
11650 u32 alu_state, alu_limit;
11651 struct bpf_reg_state tmp;
11655 if (can_skip_alu_sanitation(env, insn))
11658 /* We already marked aux for masking from non-speculative
11659 * paths, thus we got here in the first place. We only care
11660 * to explore bad access from here.
11662 if (vstate->speculative)
11665 if (!commit_window) {
11666 if (!tnum_is_const(off_reg->var_off) &&
11667 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
11668 return REASON_BOUNDS;
11670 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
11671 (opcode == BPF_SUB && !off_is_neg);
11674 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
11678 if (commit_window) {
11679 /* In commit phase we narrow the masking window based on
11680 * the observed pointer move after the simulated operation.
11682 alu_state = info->aux.alu_state;
11683 alu_limit = abs(info->aux.alu_limit - alu_limit);
11685 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
11686 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
11687 alu_state |= ptr_is_dst_reg ?
11688 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
11690 /* Limit pruning on unknown scalars to enable deep search for
11691 * potential masking differences from other program paths.
11694 env->explore_alu_limits = true;
11697 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
11701 /* If we're in commit phase, we're done here given we already
11702 * pushed the truncated dst_reg into the speculative verification
11705 * Also, when register is a known constant, we rewrite register-based
11706 * operation to immediate-based, and thus do not need masking (and as
11707 * a consequence, do not need to simulate the zero-truncation either).
11709 if (commit_window || off_is_imm)
11712 /* Simulate and find potential out-of-bounds access under
11713 * speculative execution from truncation as a result of
11714 * masking when off was not within expected range. If off
11715 * sits in dst, then we temporarily need to move ptr there
11716 * to simulate dst (== 0) +/-= ptr. Needed, for example,
11717 * for cases where we use K-based arithmetic in one direction
11718 * and truncated reg-based in the other in order to explore
11721 if (!ptr_is_dst_reg) {
11723 copy_register_state(dst_reg, ptr_reg);
11725 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
11727 if (!ptr_is_dst_reg && ret)
11729 return !ret ? REASON_STACK : 0;
11732 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
11734 struct bpf_verifier_state *vstate = env->cur_state;
11736 /* If we simulate paths under speculation, we don't update the
11737 * insn as 'seen' such that when we verify unreachable paths in
11738 * the non-speculative domain, sanitize_dead_code() can still
11739 * rewrite/sanitize them.
11741 if (!vstate->speculative)
11742 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
11745 static int sanitize_err(struct bpf_verifier_env *env,
11746 const struct bpf_insn *insn, int reason,
11747 const struct bpf_reg_state *off_reg,
11748 const struct bpf_reg_state *dst_reg)
11750 static const char *err = "pointer arithmetic with it prohibited for !root";
11751 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
11752 u32 dst = insn->dst_reg, src = insn->src_reg;
11755 case REASON_BOUNDS:
11756 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
11757 off_reg == dst_reg ? dst : src, err);
11760 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
11761 off_reg == dst_reg ? src : dst, err);
11764 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
11768 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
11772 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
11776 verbose(env, "verifier internal error: unknown reason (%d)\n",
11784 /* check that stack access falls within stack limits and that 'reg' doesn't
11785 * have a variable offset.
11787 * Variable offset is prohibited for unprivileged mode for simplicity since it
11788 * requires corresponding support in Spectre masking for stack ALU. See also
11789 * retrieve_ptr_limit().
11792 * 'off' includes 'reg->off'.
11794 static int check_stack_access_for_ptr_arithmetic(
11795 struct bpf_verifier_env *env,
11797 const struct bpf_reg_state *reg,
11800 if (!tnum_is_const(reg->var_off)) {
11803 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
11804 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
11805 regno, tn_buf, off);
11809 if (off >= 0 || off < -MAX_BPF_STACK) {
11810 verbose(env, "R%d stack pointer arithmetic goes out of range, "
11811 "prohibited for !root; off=%d\n", regno, off);
11818 static int sanitize_check_bounds(struct bpf_verifier_env *env,
11819 const struct bpf_insn *insn,
11820 const struct bpf_reg_state *dst_reg)
11822 u32 dst = insn->dst_reg;
11824 /* For unprivileged we require that resulting offset must be in bounds
11825 * in order to be able to sanitize access later on.
11827 if (env->bypass_spec_v1)
11830 switch (dst_reg->type) {
11832 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
11833 dst_reg->off + dst_reg->var_off.value))
11836 case PTR_TO_MAP_VALUE:
11837 if (check_map_access(env, dst, dst_reg->off, 1, false, ACCESS_HELPER)) {
11838 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
11839 "prohibited for !root\n", dst);
11850 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
11851 * Caller should also handle BPF_MOV case separately.
11852 * If we return -EACCES, caller may want to try again treating pointer as a
11853 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
11855 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
11856 struct bpf_insn *insn,
11857 const struct bpf_reg_state *ptr_reg,
11858 const struct bpf_reg_state *off_reg)
11860 struct bpf_verifier_state *vstate = env->cur_state;
11861 struct bpf_func_state *state = vstate->frame[vstate->curframe];
11862 struct bpf_reg_state *regs = state->regs, *dst_reg;
11863 bool known = tnum_is_const(off_reg->var_off);
11864 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
11865 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
11866 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
11867 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
11868 struct bpf_sanitize_info info = {};
11869 u8 opcode = BPF_OP(insn->code);
11870 u32 dst = insn->dst_reg;
11873 dst_reg = ®s[dst];
11875 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
11876 smin_val > smax_val || umin_val > umax_val) {
11877 /* Taint dst register if offset had invalid bounds derived from
11878 * e.g. dead branches.
11880 __mark_reg_unknown(env, dst_reg);
11884 if (BPF_CLASS(insn->code) != BPF_ALU64) {
11885 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
11886 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
11887 __mark_reg_unknown(env, dst_reg);
11892 "R%d 32-bit pointer arithmetic prohibited\n",
11897 if (ptr_reg->type & PTR_MAYBE_NULL) {
11898 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
11899 dst, reg_type_str(env, ptr_reg->type));
11903 switch (base_type(ptr_reg->type)) {
11904 case CONST_PTR_TO_MAP:
11905 /* smin_val represents the known value */
11906 if (known && smin_val == 0 && opcode == BPF_ADD)
11909 case PTR_TO_PACKET_END:
11910 case PTR_TO_SOCKET:
11911 case PTR_TO_SOCK_COMMON:
11912 case PTR_TO_TCP_SOCK:
11913 case PTR_TO_XDP_SOCK:
11914 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
11915 dst, reg_type_str(env, ptr_reg->type));
11921 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
11922 * The id may be overwritten later if we create a new variable offset.
11924 dst_reg->type = ptr_reg->type;
11925 dst_reg->id = ptr_reg->id;
11927 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
11928 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
11931 /* pointer types do not carry 32-bit bounds at the moment. */
11932 __mark_reg32_unbounded(dst_reg);
11934 if (sanitize_needed(opcode)) {
11935 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
11938 return sanitize_err(env, insn, ret, off_reg, dst_reg);
11943 /* We can take a fixed offset as long as it doesn't overflow
11944 * the s32 'off' field
11946 if (known && (ptr_reg->off + smin_val ==
11947 (s64)(s32)(ptr_reg->off + smin_val))) {
11948 /* pointer += K. Accumulate it into fixed offset */
11949 dst_reg->smin_value = smin_ptr;
11950 dst_reg->smax_value = smax_ptr;
11951 dst_reg->umin_value = umin_ptr;
11952 dst_reg->umax_value = umax_ptr;
11953 dst_reg->var_off = ptr_reg->var_off;
11954 dst_reg->off = ptr_reg->off + smin_val;
11955 dst_reg->raw = ptr_reg->raw;
11958 /* A new variable offset is created. Note that off_reg->off
11959 * == 0, since it's a scalar.
11960 * dst_reg gets the pointer type and since some positive
11961 * integer value was added to the pointer, give it a new 'id'
11962 * if it's a PTR_TO_PACKET.
11963 * this creates a new 'base' pointer, off_reg (variable) gets
11964 * added into the variable offset, and we copy the fixed offset
11967 if (signed_add_overflows(smin_ptr, smin_val) ||
11968 signed_add_overflows(smax_ptr, smax_val)) {
11969 dst_reg->smin_value = S64_MIN;
11970 dst_reg->smax_value = S64_MAX;
11972 dst_reg->smin_value = smin_ptr + smin_val;
11973 dst_reg->smax_value = smax_ptr + smax_val;
11975 if (umin_ptr + umin_val < umin_ptr ||
11976 umax_ptr + umax_val < umax_ptr) {
11977 dst_reg->umin_value = 0;
11978 dst_reg->umax_value = U64_MAX;
11980 dst_reg->umin_value = umin_ptr + umin_val;
11981 dst_reg->umax_value = umax_ptr + umax_val;
11983 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
11984 dst_reg->off = ptr_reg->off;
11985 dst_reg->raw = ptr_reg->raw;
11986 if (reg_is_pkt_pointer(ptr_reg)) {
11987 dst_reg->id = ++env->id_gen;
11988 /* something was added to pkt_ptr, set range to zero */
11989 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
11993 if (dst_reg == off_reg) {
11994 /* scalar -= pointer. Creates an unknown scalar */
11995 verbose(env, "R%d tried to subtract pointer from scalar\n",
11999 /* We don't allow subtraction from FP, because (according to
12000 * test_verifier.c test "invalid fp arithmetic", JITs might not
12001 * be able to deal with it.
12003 if (ptr_reg->type == PTR_TO_STACK) {
12004 verbose(env, "R%d subtraction from stack pointer prohibited\n",
12008 if (known && (ptr_reg->off - smin_val ==
12009 (s64)(s32)(ptr_reg->off - smin_val))) {
12010 /* pointer -= K. Subtract it from fixed offset */
12011 dst_reg->smin_value = smin_ptr;
12012 dst_reg->smax_value = smax_ptr;
12013 dst_reg->umin_value = umin_ptr;
12014 dst_reg->umax_value = umax_ptr;
12015 dst_reg->var_off = ptr_reg->var_off;
12016 dst_reg->id = ptr_reg->id;
12017 dst_reg->off = ptr_reg->off - smin_val;
12018 dst_reg->raw = ptr_reg->raw;
12021 /* A new variable offset is created. If the subtrahend is known
12022 * nonnegative, then any reg->range we had before is still good.
12024 if (signed_sub_overflows(smin_ptr, smax_val) ||
12025 signed_sub_overflows(smax_ptr, smin_val)) {
12026 /* Overflow possible, we know nothing */
12027 dst_reg->smin_value = S64_MIN;
12028 dst_reg->smax_value = S64_MAX;
12030 dst_reg->smin_value = smin_ptr - smax_val;
12031 dst_reg->smax_value = smax_ptr - smin_val;
12033 if (umin_ptr < umax_val) {
12034 /* Overflow possible, we know nothing */
12035 dst_reg->umin_value = 0;
12036 dst_reg->umax_value = U64_MAX;
12038 /* Cannot overflow (as long as bounds are consistent) */
12039 dst_reg->umin_value = umin_ptr - umax_val;
12040 dst_reg->umax_value = umax_ptr - umin_val;
12042 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
12043 dst_reg->off = ptr_reg->off;
12044 dst_reg->raw = ptr_reg->raw;
12045 if (reg_is_pkt_pointer(ptr_reg)) {
12046 dst_reg->id = ++env->id_gen;
12047 /* something was added to pkt_ptr, set range to zero */
12049 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
12055 /* bitwise ops on pointers are troublesome, prohibit. */
12056 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
12057 dst, bpf_alu_string[opcode >> 4]);
12060 /* other operators (e.g. MUL,LSH) produce non-pointer results */
12061 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
12062 dst, bpf_alu_string[opcode >> 4]);
12066 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
12068 reg_bounds_sync(dst_reg);
12069 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
12071 if (sanitize_needed(opcode)) {
12072 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
12075 return sanitize_err(env, insn, ret, off_reg, dst_reg);
12081 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
12082 struct bpf_reg_state *src_reg)
12084 s32 smin_val = src_reg->s32_min_value;
12085 s32 smax_val = src_reg->s32_max_value;
12086 u32 umin_val = src_reg->u32_min_value;
12087 u32 umax_val = src_reg->u32_max_value;
12089 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
12090 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
12091 dst_reg->s32_min_value = S32_MIN;
12092 dst_reg->s32_max_value = S32_MAX;
12094 dst_reg->s32_min_value += smin_val;
12095 dst_reg->s32_max_value += smax_val;
12097 if (dst_reg->u32_min_value + umin_val < umin_val ||
12098 dst_reg->u32_max_value + umax_val < umax_val) {
12099 dst_reg->u32_min_value = 0;
12100 dst_reg->u32_max_value = U32_MAX;
12102 dst_reg->u32_min_value += umin_val;
12103 dst_reg->u32_max_value += umax_val;
12107 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
12108 struct bpf_reg_state *src_reg)
12110 s64 smin_val = src_reg->smin_value;
12111 s64 smax_val = src_reg->smax_value;
12112 u64 umin_val = src_reg->umin_value;
12113 u64 umax_val = src_reg->umax_value;
12115 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
12116 signed_add_overflows(dst_reg->smax_value, smax_val)) {
12117 dst_reg->smin_value = S64_MIN;
12118 dst_reg->smax_value = S64_MAX;
12120 dst_reg->smin_value += smin_val;
12121 dst_reg->smax_value += smax_val;
12123 if (dst_reg->umin_value + umin_val < umin_val ||
12124 dst_reg->umax_value + umax_val < umax_val) {
12125 dst_reg->umin_value = 0;
12126 dst_reg->umax_value = U64_MAX;
12128 dst_reg->umin_value += umin_val;
12129 dst_reg->umax_value += umax_val;
12133 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
12134 struct bpf_reg_state *src_reg)
12136 s32 smin_val = src_reg->s32_min_value;
12137 s32 smax_val = src_reg->s32_max_value;
12138 u32 umin_val = src_reg->u32_min_value;
12139 u32 umax_val = src_reg->u32_max_value;
12141 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
12142 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
12143 /* Overflow possible, we know nothing */
12144 dst_reg->s32_min_value = S32_MIN;
12145 dst_reg->s32_max_value = S32_MAX;
12147 dst_reg->s32_min_value -= smax_val;
12148 dst_reg->s32_max_value -= smin_val;
12150 if (dst_reg->u32_min_value < umax_val) {
12151 /* Overflow possible, we know nothing */
12152 dst_reg->u32_min_value = 0;
12153 dst_reg->u32_max_value = U32_MAX;
12155 /* Cannot overflow (as long as bounds are consistent) */
12156 dst_reg->u32_min_value -= umax_val;
12157 dst_reg->u32_max_value -= umin_val;
12161 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
12162 struct bpf_reg_state *src_reg)
12164 s64 smin_val = src_reg->smin_value;
12165 s64 smax_val = src_reg->smax_value;
12166 u64 umin_val = src_reg->umin_value;
12167 u64 umax_val = src_reg->umax_value;
12169 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
12170 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
12171 /* Overflow possible, we know nothing */
12172 dst_reg->smin_value = S64_MIN;
12173 dst_reg->smax_value = S64_MAX;
12175 dst_reg->smin_value -= smax_val;
12176 dst_reg->smax_value -= smin_val;
12178 if (dst_reg->umin_value < umax_val) {
12179 /* Overflow possible, we know nothing */
12180 dst_reg->umin_value = 0;
12181 dst_reg->umax_value = U64_MAX;
12183 /* Cannot overflow (as long as bounds are consistent) */
12184 dst_reg->umin_value -= umax_val;
12185 dst_reg->umax_value -= umin_val;
12189 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
12190 struct bpf_reg_state *src_reg)
12192 s32 smin_val = src_reg->s32_min_value;
12193 u32 umin_val = src_reg->u32_min_value;
12194 u32 umax_val = src_reg->u32_max_value;
12196 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
12197 /* Ain't nobody got time to multiply that sign */
12198 __mark_reg32_unbounded(dst_reg);
12201 /* Both values are positive, so we can work with unsigned and
12202 * copy the result to signed (unless it exceeds S32_MAX).
12204 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
12205 /* Potential overflow, we know nothing */
12206 __mark_reg32_unbounded(dst_reg);
12209 dst_reg->u32_min_value *= umin_val;
12210 dst_reg->u32_max_value *= umax_val;
12211 if (dst_reg->u32_max_value > S32_MAX) {
12212 /* Overflow possible, we know nothing */
12213 dst_reg->s32_min_value = S32_MIN;
12214 dst_reg->s32_max_value = S32_MAX;
12216 dst_reg->s32_min_value = dst_reg->u32_min_value;
12217 dst_reg->s32_max_value = dst_reg->u32_max_value;
12221 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
12222 struct bpf_reg_state *src_reg)
12224 s64 smin_val = src_reg->smin_value;
12225 u64 umin_val = src_reg->umin_value;
12226 u64 umax_val = src_reg->umax_value;
12228 if (smin_val < 0 || dst_reg->smin_value < 0) {
12229 /* Ain't nobody got time to multiply that sign */
12230 __mark_reg64_unbounded(dst_reg);
12233 /* Both values are positive, so we can work with unsigned and
12234 * copy the result to signed (unless it exceeds S64_MAX).
12236 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
12237 /* Potential overflow, we know nothing */
12238 __mark_reg64_unbounded(dst_reg);
12241 dst_reg->umin_value *= umin_val;
12242 dst_reg->umax_value *= umax_val;
12243 if (dst_reg->umax_value > S64_MAX) {
12244 /* Overflow possible, we know nothing */
12245 dst_reg->smin_value = S64_MIN;
12246 dst_reg->smax_value = S64_MAX;
12248 dst_reg->smin_value = dst_reg->umin_value;
12249 dst_reg->smax_value = dst_reg->umax_value;
12253 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
12254 struct bpf_reg_state *src_reg)
12256 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12257 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12258 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12259 s32 smin_val = src_reg->s32_min_value;
12260 u32 umax_val = src_reg->u32_max_value;
12262 if (src_known && dst_known) {
12263 __mark_reg32_known(dst_reg, var32_off.value);
12267 /* We get our minimum from the var_off, since that's inherently
12268 * bitwise. Our maximum is the minimum of the operands' maxima.
12270 dst_reg->u32_min_value = var32_off.value;
12271 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
12272 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
12273 /* Lose signed bounds when ANDing negative numbers,
12274 * ain't nobody got time for that.
12276 dst_reg->s32_min_value = S32_MIN;
12277 dst_reg->s32_max_value = S32_MAX;
12279 /* ANDing two positives gives a positive, so safe to
12280 * cast result into s64.
12282 dst_reg->s32_min_value = dst_reg->u32_min_value;
12283 dst_reg->s32_max_value = dst_reg->u32_max_value;
12287 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
12288 struct bpf_reg_state *src_reg)
12290 bool src_known = tnum_is_const(src_reg->var_off);
12291 bool dst_known = tnum_is_const(dst_reg->var_off);
12292 s64 smin_val = src_reg->smin_value;
12293 u64 umax_val = src_reg->umax_value;
12295 if (src_known && dst_known) {
12296 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12300 /* We get our minimum from the var_off, since that's inherently
12301 * bitwise. Our maximum is the minimum of the operands' maxima.
12303 dst_reg->umin_value = dst_reg->var_off.value;
12304 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
12305 if (dst_reg->smin_value < 0 || smin_val < 0) {
12306 /* Lose signed bounds when ANDing negative numbers,
12307 * ain't nobody got time for that.
12309 dst_reg->smin_value = S64_MIN;
12310 dst_reg->smax_value = S64_MAX;
12312 /* ANDing two positives gives a positive, so safe to
12313 * cast result into s64.
12315 dst_reg->smin_value = dst_reg->umin_value;
12316 dst_reg->smax_value = dst_reg->umax_value;
12318 /* We may learn something more from the var_off */
12319 __update_reg_bounds(dst_reg);
12322 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
12323 struct bpf_reg_state *src_reg)
12325 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12326 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12327 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12328 s32 smin_val = src_reg->s32_min_value;
12329 u32 umin_val = src_reg->u32_min_value;
12331 if (src_known && dst_known) {
12332 __mark_reg32_known(dst_reg, var32_off.value);
12336 /* We get our maximum from the var_off, and our minimum is the
12337 * maximum of the operands' minima
12339 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
12340 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
12341 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
12342 /* Lose signed bounds when ORing negative numbers,
12343 * ain't nobody got time for that.
12345 dst_reg->s32_min_value = S32_MIN;
12346 dst_reg->s32_max_value = S32_MAX;
12348 /* ORing two positives gives a positive, so safe to
12349 * cast result into s64.
12351 dst_reg->s32_min_value = dst_reg->u32_min_value;
12352 dst_reg->s32_max_value = dst_reg->u32_max_value;
12356 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
12357 struct bpf_reg_state *src_reg)
12359 bool src_known = tnum_is_const(src_reg->var_off);
12360 bool dst_known = tnum_is_const(dst_reg->var_off);
12361 s64 smin_val = src_reg->smin_value;
12362 u64 umin_val = src_reg->umin_value;
12364 if (src_known && dst_known) {
12365 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12369 /* We get our maximum from the var_off, and our minimum is the
12370 * maximum of the operands' minima
12372 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
12373 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
12374 if (dst_reg->smin_value < 0 || smin_val < 0) {
12375 /* Lose signed bounds when ORing negative numbers,
12376 * ain't nobody got time for that.
12378 dst_reg->smin_value = S64_MIN;
12379 dst_reg->smax_value = S64_MAX;
12381 /* ORing two positives gives a positive, so safe to
12382 * cast result into s64.
12384 dst_reg->smin_value = dst_reg->umin_value;
12385 dst_reg->smax_value = dst_reg->umax_value;
12387 /* We may learn something more from the var_off */
12388 __update_reg_bounds(dst_reg);
12391 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
12392 struct bpf_reg_state *src_reg)
12394 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12395 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12396 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12397 s32 smin_val = src_reg->s32_min_value;
12399 if (src_known && dst_known) {
12400 __mark_reg32_known(dst_reg, var32_off.value);
12404 /* We get both minimum and maximum from the var32_off. */
12405 dst_reg->u32_min_value = var32_off.value;
12406 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
12408 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
12409 /* XORing two positive sign numbers gives a positive,
12410 * so safe to cast u32 result into s32.
12412 dst_reg->s32_min_value = dst_reg->u32_min_value;
12413 dst_reg->s32_max_value = dst_reg->u32_max_value;
12415 dst_reg->s32_min_value = S32_MIN;
12416 dst_reg->s32_max_value = S32_MAX;
12420 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
12421 struct bpf_reg_state *src_reg)
12423 bool src_known = tnum_is_const(src_reg->var_off);
12424 bool dst_known = tnum_is_const(dst_reg->var_off);
12425 s64 smin_val = src_reg->smin_value;
12427 if (src_known && dst_known) {
12428 /* dst_reg->var_off.value has been updated earlier */
12429 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12433 /* We get both minimum and maximum from the var_off. */
12434 dst_reg->umin_value = dst_reg->var_off.value;
12435 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
12437 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
12438 /* XORing two positive sign numbers gives a positive,
12439 * so safe to cast u64 result into s64.
12441 dst_reg->smin_value = dst_reg->umin_value;
12442 dst_reg->smax_value = dst_reg->umax_value;
12444 dst_reg->smin_value = S64_MIN;
12445 dst_reg->smax_value = S64_MAX;
12448 __update_reg_bounds(dst_reg);
12451 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
12452 u64 umin_val, u64 umax_val)
12454 /* We lose all sign bit information (except what we can pick
12457 dst_reg->s32_min_value = S32_MIN;
12458 dst_reg->s32_max_value = S32_MAX;
12459 /* If we might shift our top bit out, then we know nothing */
12460 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
12461 dst_reg->u32_min_value = 0;
12462 dst_reg->u32_max_value = U32_MAX;
12464 dst_reg->u32_min_value <<= umin_val;
12465 dst_reg->u32_max_value <<= umax_val;
12469 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
12470 struct bpf_reg_state *src_reg)
12472 u32 umax_val = src_reg->u32_max_value;
12473 u32 umin_val = src_reg->u32_min_value;
12474 /* u32 alu operation will zext upper bits */
12475 struct tnum subreg = tnum_subreg(dst_reg->var_off);
12477 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
12478 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
12479 /* Not required but being careful mark reg64 bounds as unknown so
12480 * that we are forced to pick them up from tnum and zext later and
12481 * if some path skips this step we are still safe.
12483 __mark_reg64_unbounded(dst_reg);
12484 __update_reg32_bounds(dst_reg);
12487 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
12488 u64 umin_val, u64 umax_val)
12490 /* Special case <<32 because it is a common compiler pattern to sign
12491 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
12492 * positive we know this shift will also be positive so we can track
12493 * bounds correctly. Otherwise we lose all sign bit information except
12494 * what we can pick up from var_off. Perhaps we can generalize this
12495 * later to shifts of any length.
12497 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
12498 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
12500 dst_reg->smax_value = S64_MAX;
12502 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
12503 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
12505 dst_reg->smin_value = S64_MIN;
12507 /* If we might shift our top bit out, then we know nothing */
12508 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
12509 dst_reg->umin_value = 0;
12510 dst_reg->umax_value = U64_MAX;
12512 dst_reg->umin_value <<= umin_val;
12513 dst_reg->umax_value <<= umax_val;
12517 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
12518 struct bpf_reg_state *src_reg)
12520 u64 umax_val = src_reg->umax_value;
12521 u64 umin_val = src_reg->umin_value;
12523 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
12524 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
12525 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
12527 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
12528 /* We may learn something more from the var_off */
12529 __update_reg_bounds(dst_reg);
12532 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
12533 struct bpf_reg_state *src_reg)
12535 struct tnum subreg = tnum_subreg(dst_reg->var_off);
12536 u32 umax_val = src_reg->u32_max_value;
12537 u32 umin_val = src_reg->u32_min_value;
12539 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
12540 * be negative, then either:
12541 * 1) src_reg might be zero, so the sign bit of the result is
12542 * unknown, so we lose our signed bounds
12543 * 2) it's known negative, thus the unsigned bounds capture the
12545 * 3) the signed bounds cross zero, so they tell us nothing
12547 * If the value in dst_reg is known nonnegative, then again the
12548 * unsigned bounds capture the signed bounds.
12549 * Thus, in all cases it suffices to blow away our signed bounds
12550 * and rely on inferring new ones from the unsigned bounds and
12551 * var_off of the result.
12553 dst_reg->s32_min_value = S32_MIN;
12554 dst_reg->s32_max_value = S32_MAX;
12556 dst_reg->var_off = tnum_rshift(subreg, umin_val);
12557 dst_reg->u32_min_value >>= umax_val;
12558 dst_reg->u32_max_value >>= umin_val;
12560 __mark_reg64_unbounded(dst_reg);
12561 __update_reg32_bounds(dst_reg);
12564 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
12565 struct bpf_reg_state *src_reg)
12567 u64 umax_val = src_reg->umax_value;
12568 u64 umin_val = src_reg->umin_value;
12570 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
12571 * be negative, then either:
12572 * 1) src_reg might be zero, so the sign bit of the result is
12573 * unknown, so we lose our signed bounds
12574 * 2) it's known negative, thus the unsigned bounds capture the
12576 * 3) the signed bounds cross zero, so they tell us nothing
12578 * If the value in dst_reg is known nonnegative, then again the
12579 * unsigned bounds capture the signed bounds.
12580 * Thus, in all cases it suffices to blow away our signed bounds
12581 * and rely on inferring new ones from the unsigned bounds and
12582 * var_off of the result.
12584 dst_reg->smin_value = S64_MIN;
12585 dst_reg->smax_value = S64_MAX;
12586 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
12587 dst_reg->umin_value >>= umax_val;
12588 dst_reg->umax_value >>= umin_val;
12590 /* Its not easy to operate on alu32 bounds here because it depends
12591 * on bits being shifted in. Take easy way out and mark unbounded
12592 * so we can recalculate later from tnum.
12594 __mark_reg32_unbounded(dst_reg);
12595 __update_reg_bounds(dst_reg);
12598 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
12599 struct bpf_reg_state *src_reg)
12601 u64 umin_val = src_reg->u32_min_value;
12603 /* Upon reaching here, src_known is true and
12604 * umax_val is equal to umin_val.
12606 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
12607 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
12609 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
12611 /* blow away the dst_reg umin_value/umax_value and rely on
12612 * dst_reg var_off to refine the result.
12614 dst_reg->u32_min_value = 0;
12615 dst_reg->u32_max_value = U32_MAX;
12617 __mark_reg64_unbounded(dst_reg);
12618 __update_reg32_bounds(dst_reg);
12621 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
12622 struct bpf_reg_state *src_reg)
12624 u64 umin_val = src_reg->umin_value;
12626 /* Upon reaching here, src_known is true and umax_val is equal
12629 dst_reg->smin_value >>= umin_val;
12630 dst_reg->smax_value >>= umin_val;
12632 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
12634 /* blow away the dst_reg umin_value/umax_value and rely on
12635 * dst_reg var_off to refine the result.
12637 dst_reg->umin_value = 0;
12638 dst_reg->umax_value = U64_MAX;
12640 /* Its not easy to operate on alu32 bounds here because it depends
12641 * on bits being shifted in from upper 32-bits. Take easy way out
12642 * and mark unbounded so we can recalculate later from tnum.
12644 __mark_reg32_unbounded(dst_reg);
12645 __update_reg_bounds(dst_reg);
12648 /* WARNING: This function does calculations on 64-bit values, but the actual
12649 * execution may occur on 32-bit values. Therefore, things like bitshifts
12650 * need extra checks in the 32-bit case.
12652 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
12653 struct bpf_insn *insn,
12654 struct bpf_reg_state *dst_reg,
12655 struct bpf_reg_state src_reg)
12657 struct bpf_reg_state *regs = cur_regs(env);
12658 u8 opcode = BPF_OP(insn->code);
12660 s64 smin_val, smax_val;
12661 u64 umin_val, umax_val;
12662 s32 s32_min_val, s32_max_val;
12663 u32 u32_min_val, u32_max_val;
12664 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
12665 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
12668 smin_val = src_reg.smin_value;
12669 smax_val = src_reg.smax_value;
12670 umin_val = src_reg.umin_value;
12671 umax_val = src_reg.umax_value;
12673 s32_min_val = src_reg.s32_min_value;
12674 s32_max_val = src_reg.s32_max_value;
12675 u32_min_val = src_reg.u32_min_value;
12676 u32_max_val = src_reg.u32_max_value;
12679 src_known = tnum_subreg_is_const(src_reg.var_off);
12681 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
12682 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
12683 /* Taint dst register if offset had invalid bounds
12684 * derived from e.g. dead branches.
12686 __mark_reg_unknown(env, dst_reg);
12690 src_known = tnum_is_const(src_reg.var_off);
12692 (smin_val != smax_val || umin_val != umax_val)) ||
12693 smin_val > smax_val || umin_val > umax_val) {
12694 /* Taint dst register if offset had invalid bounds
12695 * derived from e.g. dead branches.
12697 __mark_reg_unknown(env, dst_reg);
12703 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
12704 __mark_reg_unknown(env, dst_reg);
12708 if (sanitize_needed(opcode)) {
12709 ret = sanitize_val_alu(env, insn);
12711 return sanitize_err(env, insn, ret, NULL, NULL);
12714 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
12715 * There are two classes of instructions: The first class we track both
12716 * alu32 and alu64 sign/unsigned bounds independently this provides the
12717 * greatest amount of precision when alu operations are mixed with jmp32
12718 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
12719 * and BPF_OR. This is possible because these ops have fairly easy to
12720 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
12721 * See alu32 verifier tests for examples. The second class of
12722 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
12723 * with regards to tracking sign/unsigned bounds because the bits may
12724 * cross subreg boundaries in the alu64 case. When this happens we mark
12725 * the reg unbounded in the subreg bound space and use the resulting
12726 * tnum to calculate an approximation of the sign/unsigned bounds.
12730 scalar32_min_max_add(dst_reg, &src_reg);
12731 scalar_min_max_add(dst_reg, &src_reg);
12732 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
12735 scalar32_min_max_sub(dst_reg, &src_reg);
12736 scalar_min_max_sub(dst_reg, &src_reg);
12737 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
12740 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
12741 scalar32_min_max_mul(dst_reg, &src_reg);
12742 scalar_min_max_mul(dst_reg, &src_reg);
12745 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
12746 scalar32_min_max_and(dst_reg, &src_reg);
12747 scalar_min_max_and(dst_reg, &src_reg);
12750 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
12751 scalar32_min_max_or(dst_reg, &src_reg);
12752 scalar_min_max_or(dst_reg, &src_reg);
12755 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
12756 scalar32_min_max_xor(dst_reg, &src_reg);
12757 scalar_min_max_xor(dst_reg, &src_reg);
12760 if (umax_val >= insn_bitness) {
12761 /* Shifts greater than 31 or 63 are undefined.
12762 * This includes shifts by a negative number.
12764 mark_reg_unknown(env, regs, insn->dst_reg);
12768 scalar32_min_max_lsh(dst_reg, &src_reg);
12770 scalar_min_max_lsh(dst_reg, &src_reg);
12773 if (umax_val >= insn_bitness) {
12774 /* Shifts greater than 31 or 63 are undefined.
12775 * This includes shifts by a negative number.
12777 mark_reg_unknown(env, regs, insn->dst_reg);
12781 scalar32_min_max_rsh(dst_reg, &src_reg);
12783 scalar_min_max_rsh(dst_reg, &src_reg);
12786 if (umax_val >= insn_bitness) {
12787 /* Shifts greater than 31 or 63 are undefined.
12788 * This includes shifts by a negative number.
12790 mark_reg_unknown(env, regs, insn->dst_reg);
12794 scalar32_min_max_arsh(dst_reg, &src_reg);
12796 scalar_min_max_arsh(dst_reg, &src_reg);
12799 mark_reg_unknown(env, regs, insn->dst_reg);
12803 /* ALU32 ops are zero extended into 64bit register */
12805 zext_32_to_64(dst_reg);
12806 reg_bounds_sync(dst_reg);
12810 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
12813 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
12814 struct bpf_insn *insn)
12816 struct bpf_verifier_state *vstate = env->cur_state;
12817 struct bpf_func_state *state = vstate->frame[vstate->curframe];
12818 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
12819 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
12820 u8 opcode = BPF_OP(insn->code);
12823 dst_reg = ®s[insn->dst_reg];
12825 if (dst_reg->type != SCALAR_VALUE)
12828 /* Make sure ID is cleared otherwise dst_reg min/max could be
12829 * incorrectly propagated into other registers by find_equal_scalars()
12832 if (BPF_SRC(insn->code) == BPF_X) {
12833 src_reg = ®s[insn->src_reg];
12834 if (src_reg->type != SCALAR_VALUE) {
12835 if (dst_reg->type != SCALAR_VALUE) {
12836 /* Combining two pointers by any ALU op yields
12837 * an arbitrary scalar. Disallow all math except
12838 * pointer subtraction
12840 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
12841 mark_reg_unknown(env, regs, insn->dst_reg);
12844 verbose(env, "R%d pointer %s pointer prohibited\n",
12846 bpf_alu_string[opcode >> 4]);
12849 /* scalar += pointer
12850 * This is legal, but we have to reverse our
12851 * src/dest handling in computing the range
12853 err = mark_chain_precision(env, insn->dst_reg);
12856 return adjust_ptr_min_max_vals(env, insn,
12859 } else if (ptr_reg) {
12860 /* pointer += scalar */
12861 err = mark_chain_precision(env, insn->src_reg);
12864 return adjust_ptr_min_max_vals(env, insn,
12866 } else if (dst_reg->precise) {
12867 /* if dst_reg is precise, src_reg should be precise as well */
12868 err = mark_chain_precision(env, insn->src_reg);
12873 /* Pretend the src is a reg with a known value, since we only
12874 * need to be able to read from this state.
12876 off_reg.type = SCALAR_VALUE;
12877 __mark_reg_known(&off_reg, insn->imm);
12878 src_reg = &off_reg;
12879 if (ptr_reg) /* pointer += K */
12880 return adjust_ptr_min_max_vals(env, insn,
12884 /* Got here implies adding two SCALAR_VALUEs */
12885 if (WARN_ON_ONCE(ptr_reg)) {
12886 print_verifier_state(env, state, true);
12887 verbose(env, "verifier internal error: unexpected ptr_reg\n");
12890 if (WARN_ON(!src_reg)) {
12891 print_verifier_state(env, state, true);
12892 verbose(env, "verifier internal error: no src_reg\n");
12895 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
12898 /* check validity of 32-bit and 64-bit arithmetic operations */
12899 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
12901 struct bpf_reg_state *regs = cur_regs(env);
12902 u8 opcode = BPF_OP(insn->code);
12905 if (opcode == BPF_END || opcode == BPF_NEG) {
12906 if (opcode == BPF_NEG) {
12907 if (BPF_SRC(insn->code) != BPF_K ||
12908 insn->src_reg != BPF_REG_0 ||
12909 insn->off != 0 || insn->imm != 0) {
12910 verbose(env, "BPF_NEG uses reserved fields\n");
12914 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
12915 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
12916 BPF_CLASS(insn->code) == BPF_ALU64) {
12917 verbose(env, "BPF_END uses reserved fields\n");
12922 /* check src operand */
12923 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
12927 if (is_pointer_value(env, insn->dst_reg)) {
12928 verbose(env, "R%d pointer arithmetic prohibited\n",
12933 /* check dest operand */
12934 err = check_reg_arg(env, insn->dst_reg, DST_OP);
12938 } else if (opcode == BPF_MOV) {
12940 if (BPF_SRC(insn->code) == BPF_X) {
12941 if (insn->imm != 0 || insn->off != 0) {
12942 verbose(env, "BPF_MOV uses reserved fields\n");
12946 /* check src operand */
12947 err = check_reg_arg(env, insn->src_reg, SRC_OP);
12951 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
12952 verbose(env, "BPF_MOV uses reserved fields\n");
12957 /* check dest operand, mark as required later */
12958 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
12962 if (BPF_SRC(insn->code) == BPF_X) {
12963 struct bpf_reg_state *src_reg = regs + insn->src_reg;
12964 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
12965 bool need_id = src_reg->type == SCALAR_VALUE && !src_reg->id &&
12966 !tnum_is_const(src_reg->var_off);
12968 if (BPF_CLASS(insn->code) == BPF_ALU64) {
12970 * copy register state to dest reg
12973 /* Assign src and dst registers the same ID
12974 * that will be used by find_equal_scalars()
12975 * to propagate min/max range.
12977 src_reg->id = ++env->id_gen;
12978 copy_register_state(dst_reg, src_reg);
12979 dst_reg->live |= REG_LIVE_WRITTEN;
12980 dst_reg->subreg_def = DEF_NOT_SUBREG;
12982 /* R1 = (u32) R2 */
12983 if (is_pointer_value(env, insn->src_reg)) {
12985 "R%d partial copy of pointer\n",
12988 } else if (src_reg->type == SCALAR_VALUE) {
12989 bool is_src_reg_u32 = src_reg->umax_value <= U32_MAX;
12991 if (is_src_reg_u32 && need_id)
12992 src_reg->id = ++env->id_gen;
12993 copy_register_state(dst_reg, src_reg);
12994 /* Make sure ID is cleared if src_reg is not in u32 range otherwise
12995 * dst_reg min/max could be incorrectly
12996 * propagated into src_reg by find_equal_scalars()
12998 if (!is_src_reg_u32)
13000 dst_reg->live |= REG_LIVE_WRITTEN;
13001 dst_reg->subreg_def = env->insn_idx + 1;
13003 mark_reg_unknown(env, regs,
13006 zext_32_to_64(dst_reg);
13007 reg_bounds_sync(dst_reg);
13011 * remember the value we stored into this reg
13013 /* clear any state __mark_reg_known doesn't set */
13014 mark_reg_unknown(env, regs, insn->dst_reg);
13015 regs[insn->dst_reg].type = SCALAR_VALUE;
13016 if (BPF_CLASS(insn->code) == BPF_ALU64) {
13017 __mark_reg_known(regs + insn->dst_reg,
13020 __mark_reg_known(regs + insn->dst_reg,
13025 } else if (opcode > BPF_END) {
13026 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
13029 } else { /* all other ALU ops: and, sub, xor, add, ... */
13031 if (BPF_SRC(insn->code) == BPF_X) {
13032 if (insn->imm != 0 || insn->off != 0) {
13033 verbose(env, "BPF_ALU uses reserved fields\n");
13036 /* check src1 operand */
13037 err = check_reg_arg(env, insn->src_reg, SRC_OP);
13041 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
13042 verbose(env, "BPF_ALU uses reserved fields\n");
13047 /* check src2 operand */
13048 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
13052 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
13053 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
13054 verbose(env, "div by zero\n");
13058 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
13059 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
13060 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
13062 if (insn->imm < 0 || insn->imm >= size) {
13063 verbose(env, "invalid shift %d\n", insn->imm);
13068 /* check dest operand */
13069 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
13073 return adjust_reg_min_max_vals(env, insn);
13079 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
13080 struct bpf_reg_state *dst_reg,
13081 enum bpf_reg_type type,
13082 bool range_right_open)
13084 struct bpf_func_state *state;
13085 struct bpf_reg_state *reg;
13088 if (dst_reg->off < 0 ||
13089 (dst_reg->off == 0 && range_right_open))
13090 /* This doesn't give us any range */
13093 if (dst_reg->umax_value > MAX_PACKET_OFF ||
13094 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
13095 /* Risk of overflow. For instance, ptr + (1<<63) may be less
13096 * than pkt_end, but that's because it's also less than pkt.
13100 new_range = dst_reg->off;
13101 if (range_right_open)
13104 /* Examples for register markings:
13106 * pkt_data in dst register:
13110 * if (r2 > pkt_end) goto <handle exception>
13115 * if (r2 < pkt_end) goto <access okay>
13116 * <handle exception>
13119 * r2 == dst_reg, pkt_end == src_reg
13120 * r2=pkt(id=n,off=8,r=0)
13121 * r3=pkt(id=n,off=0,r=0)
13123 * pkt_data in src register:
13127 * if (pkt_end >= r2) goto <access okay>
13128 * <handle exception>
13132 * if (pkt_end <= r2) goto <handle exception>
13136 * pkt_end == dst_reg, r2 == src_reg
13137 * r2=pkt(id=n,off=8,r=0)
13138 * r3=pkt(id=n,off=0,r=0)
13140 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
13141 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
13142 * and [r3, r3 + 8-1) respectively is safe to access depending on
13146 /* If our ids match, then we must have the same max_value. And we
13147 * don't care about the other reg's fixed offset, since if it's too big
13148 * the range won't allow anything.
13149 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
13151 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
13152 if (reg->type == type && reg->id == dst_reg->id)
13153 /* keep the maximum range already checked */
13154 reg->range = max(reg->range, new_range);
13158 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
13160 struct tnum subreg = tnum_subreg(reg->var_off);
13161 s32 sval = (s32)val;
13165 if (tnum_is_const(subreg))
13166 return !!tnum_equals_const(subreg, val);
13167 else if (val < reg->u32_min_value || val > reg->u32_max_value)
13171 if (tnum_is_const(subreg))
13172 return !tnum_equals_const(subreg, val);
13173 else if (val < reg->u32_min_value || val > reg->u32_max_value)
13177 if ((~subreg.mask & subreg.value) & val)
13179 if (!((subreg.mask | subreg.value) & val))
13183 if (reg->u32_min_value > val)
13185 else if (reg->u32_max_value <= val)
13189 if (reg->s32_min_value > sval)
13191 else if (reg->s32_max_value <= sval)
13195 if (reg->u32_max_value < val)
13197 else if (reg->u32_min_value >= val)
13201 if (reg->s32_max_value < sval)
13203 else if (reg->s32_min_value >= sval)
13207 if (reg->u32_min_value >= val)
13209 else if (reg->u32_max_value < val)
13213 if (reg->s32_min_value >= sval)
13215 else if (reg->s32_max_value < sval)
13219 if (reg->u32_max_value <= val)
13221 else if (reg->u32_min_value > val)
13225 if (reg->s32_max_value <= sval)
13227 else if (reg->s32_min_value > sval)
13236 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
13238 s64 sval = (s64)val;
13242 if (tnum_is_const(reg->var_off))
13243 return !!tnum_equals_const(reg->var_off, val);
13244 else if (val < reg->umin_value || val > reg->umax_value)
13248 if (tnum_is_const(reg->var_off))
13249 return !tnum_equals_const(reg->var_off, val);
13250 else if (val < reg->umin_value || val > reg->umax_value)
13254 if ((~reg->var_off.mask & reg->var_off.value) & val)
13256 if (!((reg->var_off.mask | reg->var_off.value) & val))
13260 if (reg->umin_value > val)
13262 else if (reg->umax_value <= val)
13266 if (reg->smin_value > sval)
13268 else if (reg->smax_value <= sval)
13272 if (reg->umax_value < val)
13274 else if (reg->umin_value >= val)
13278 if (reg->smax_value < sval)
13280 else if (reg->smin_value >= sval)
13284 if (reg->umin_value >= val)
13286 else if (reg->umax_value < val)
13290 if (reg->smin_value >= sval)
13292 else if (reg->smax_value < sval)
13296 if (reg->umax_value <= val)
13298 else if (reg->umin_value > val)
13302 if (reg->smax_value <= sval)
13304 else if (reg->smin_value > sval)
13312 /* compute branch direction of the expression "if (reg opcode val) goto target;"
13314 * 1 - branch will be taken and "goto target" will be executed
13315 * 0 - branch will not be taken and fall-through to next insn
13316 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
13319 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
13322 if (__is_pointer_value(false, reg)) {
13323 if (!reg_not_null(reg))
13326 /* If pointer is valid tests against zero will fail so we can
13327 * use this to direct branch taken.
13343 return is_branch32_taken(reg, val, opcode);
13344 return is_branch64_taken(reg, val, opcode);
13347 static int flip_opcode(u32 opcode)
13349 /* How can we transform "a <op> b" into "b <op> a"? */
13350 static const u8 opcode_flip[16] = {
13351 /* these stay the same */
13352 [BPF_JEQ >> 4] = BPF_JEQ,
13353 [BPF_JNE >> 4] = BPF_JNE,
13354 [BPF_JSET >> 4] = BPF_JSET,
13355 /* these swap "lesser" and "greater" (L and G in the opcodes) */
13356 [BPF_JGE >> 4] = BPF_JLE,
13357 [BPF_JGT >> 4] = BPF_JLT,
13358 [BPF_JLE >> 4] = BPF_JGE,
13359 [BPF_JLT >> 4] = BPF_JGT,
13360 [BPF_JSGE >> 4] = BPF_JSLE,
13361 [BPF_JSGT >> 4] = BPF_JSLT,
13362 [BPF_JSLE >> 4] = BPF_JSGE,
13363 [BPF_JSLT >> 4] = BPF_JSGT
13365 return opcode_flip[opcode >> 4];
13368 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
13369 struct bpf_reg_state *src_reg,
13372 struct bpf_reg_state *pkt;
13374 if (src_reg->type == PTR_TO_PACKET_END) {
13376 } else if (dst_reg->type == PTR_TO_PACKET_END) {
13378 opcode = flip_opcode(opcode);
13383 if (pkt->range >= 0)
13388 /* pkt <= pkt_end */
13391 /* pkt > pkt_end */
13392 if (pkt->range == BEYOND_PKT_END)
13393 /* pkt has at last one extra byte beyond pkt_end */
13394 return opcode == BPF_JGT;
13397 /* pkt < pkt_end */
13400 /* pkt >= pkt_end */
13401 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
13402 return opcode == BPF_JGE;
13408 /* Adjusts the register min/max values in the case that the dst_reg is the
13409 * variable register that we are working on, and src_reg is a constant or we're
13410 * simply doing a BPF_K check.
13411 * In JEQ/JNE cases we also adjust the var_off values.
13413 static void reg_set_min_max(struct bpf_reg_state *true_reg,
13414 struct bpf_reg_state *false_reg,
13415 u64 val, u32 val32,
13416 u8 opcode, bool is_jmp32)
13418 struct tnum false_32off = tnum_subreg(false_reg->var_off);
13419 struct tnum false_64off = false_reg->var_off;
13420 struct tnum true_32off = tnum_subreg(true_reg->var_off);
13421 struct tnum true_64off = true_reg->var_off;
13422 s64 sval = (s64)val;
13423 s32 sval32 = (s32)val32;
13425 /* If the dst_reg is a pointer, we can't learn anything about its
13426 * variable offset from the compare (unless src_reg were a pointer into
13427 * the same object, but we don't bother with that.
13428 * Since false_reg and true_reg have the same type by construction, we
13429 * only need to check one of them for pointerness.
13431 if (__is_pointer_value(false, false_reg))
13435 /* JEQ/JNE comparison doesn't change the register equivalence.
13438 * if (r1 == 42) goto label;
13440 * label: // here both r1 and r2 are known to be 42.
13442 * Hence when marking register as known preserve it's ID.
13446 __mark_reg32_known(true_reg, val32);
13447 true_32off = tnum_subreg(true_reg->var_off);
13449 ___mark_reg_known(true_reg, val);
13450 true_64off = true_reg->var_off;
13455 __mark_reg32_known(false_reg, val32);
13456 false_32off = tnum_subreg(false_reg->var_off);
13458 ___mark_reg_known(false_reg, val);
13459 false_64off = false_reg->var_off;
13464 false_32off = tnum_and(false_32off, tnum_const(~val32));
13465 if (is_power_of_2(val32))
13466 true_32off = tnum_or(true_32off,
13467 tnum_const(val32));
13469 false_64off = tnum_and(false_64off, tnum_const(~val));
13470 if (is_power_of_2(val))
13471 true_64off = tnum_or(true_64off,
13479 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
13480 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
13482 false_reg->u32_max_value = min(false_reg->u32_max_value,
13484 true_reg->u32_min_value = max(true_reg->u32_min_value,
13487 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
13488 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
13490 false_reg->umax_value = min(false_reg->umax_value, false_umax);
13491 true_reg->umin_value = max(true_reg->umin_value, true_umin);
13499 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
13500 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
13502 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
13503 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
13505 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
13506 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
13508 false_reg->smax_value = min(false_reg->smax_value, false_smax);
13509 true_reg->smin_value = max(true_reg->smin_value, true_smin);
13517 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
13518 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
13520 false_reg->u32_min_value = max(false_reg->u32_min_value,
13522 true_reg->u32_max_value = min(true_reg->u32_max_value,
13525 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
13526 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
13528 false_reg->umin_value = max(false_reg->umin_value, false_umin);
13529 true_reg->umax_value = min(true_reg->umax_value, true_umax);
13537 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
13538 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
13540 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
13541 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
13543 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
13544 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
13546 false_reg->smin_value = max(false_reg->smin_value, false_smin);
13547 true_reg->smax_value = min(true_reg->smax_value, true_smax);
13556 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
13557 tnum_subreg(false_32off));
13558 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
13559 tnum_subreg(true_32off));
13560 __reg_combine_32_into_64(false_reg);
13561 __reg_combine_32_into_64(true_reg);
13563 false_reg->var_off = false_64off;
13564 true_reg->var_off = true_64off;
13565 __reg_combine_64_into_32(false_reg);
13566 __reg_combine_64_into_32(true_reg);
13570 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
13571 * the variable reg.
13573 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
13574 struct bpf_reg_state *false_reg,
13575 u64 val, u32 val32,
13576 u8 opcode, bool is_jmp32)
13578 opcode = flip_opcode(opcode);
13579 /* This uses zero as "not present in table"; luckily the zero opcode,
13580 * BPF_JA, can't get here.
13583 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
13586 /* Regs are known to be equal, so intersect their min/max/var_off */
13587 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
13588 struct bpf_reg_state *dst_reg)
13590 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
13591 dst_reg->umin_value);
13592 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
13593 dst_reg->umax_value);
13594 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
13595 dst_reg->smin_value);
13596 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
13597 dst_reg->smax_value);
13598 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
13600 reg_bounds_sync(src_reg);
13601 reg_bounds_sync(dst_reg);
13604 static void reg_combine_min_max(struct bpf_reg_state *true_src,
13605 struct bpf_reg_state *true_dst,
13606 struct bpf_reg_state *false_src,
13607 struct bpf_reg_state *false_dst,
13612 __reg_combine_min_max(true_src, true_dst);
13615 __reg_combine_min_max(false_src, false_dst);
13620 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
13621 struct bpf_reg_state *reg, u32 id,
13624 if (type_may_be_null(reg->type) && reg->id == id &&
13625 (is_rcu_reg(reg) || !WARN_ON_ONCE(!reg->id))) {
13626 /* Old offset (both fixed and variable parts) should have been
13627 * known-zero, because we don't allow pointer arithmetic on
13628 * pointers that might be NULL. If we see this happening, don't
13629 * convert the register.
13631 * But in some cases, some helpers that return local kptrs
13632 * advance offset for the returned pointer. In those cases, it
13633 * is fine to expect to see reg->off.
13635 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value || !tnum_equals_const(reg->var_off, 0)))
13637 if (!(type_is_ptr_alloc_obj(reg->type) || type_is_non_owning_ref(reg->type)) &&
13638 WARN_ON_ONCE(reg->off))
13642 reg->type = SCALAR_VALUE;
13643 /* We don't need id and ref_obj_id from this point
13644 * onwards anymore, thus we should better reset it,
13645 * so that state pruning has chances to take effect.
13648 reg->ref_obj_id = 0;
13653 mark_ptr_not_null_reg(reg);
13655 if (!reg_may_point_to_spin_lock(reg)) {
13656 /* For not-NULL ptr, reg->ref_obj_id will be reset
13657 * in release_reference().
13659 * reg->id is still used by spin_lock ptr. Other
13660 * than spin_lock ptr type, reg->id can be reset.
13667 /* The logic is similar to find_good_pkt_pointers(), both could eventually
13668 * be folded together at some point.
13670 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
13673 struct bpf_func_state *state = vstate->frame[vstate->curframe];
13674 struct bpf_reg_state *regs = state->regs, *reg;
13675 u32 ref_obj_id = regs[regno].ref_obj_id;
13676 u32 id = regs[regno].id;
13678 if (ref_obj_id && ref_obj_id == id && is_null)
13679 /* regs[regno] is in the " == NULL" branch.
13680 * No one could have freed the reference state before
13681 * doing the NULL check.
13683 WARN_ON_ONCE(release_reference_state(state, id));
13685 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
13686 mark_ptr_or_null_reg(state, reg, id, is_null);
13690 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
13691 struct bpf_reg_state *dst_reg,
13692 struct bpf_reg_state *src_reg,
13693 struct bpf_verifier_state *this_branch,
13694 struct bpf_verifier_state *other_branch)
13696 if (BPF_SRC(insn->code) != BPF_X)
13699 /* Pointers are always 64-bit. */
13700 if (BPF_CLASS(insn->code) == BPF_JMP32)
13703 switch (BPF_OP(insn->code)) {
13705 if ((dst_reg->type == PTR_TO_PACKET &&
13706 src_reg->type == PTR_TO_PACKET_END) ||
13707 (dst_reg->type == PTR_TO_PACKET_META &&
13708 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13709 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
13710 find_good_pkt_pointers(this_branch, dst_reg,
13711 dst_reg->type, false);
13712 mark_pkt_end(other_branch, insn->dst_reg, true);
13713 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
13714 src_reg->type == PTR_TO_PACKET) ||
13715 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
13716 src_reg->type == PTR_TO_PACKET_META)) {
13717 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
13718 find_good_pkt_pointers(other_branch, src_reg,
13719 src_reg->type, true);
13720 mark_pkt_end(this_branch, insn->src_reg, false);
13726 if ((dst_reg->type == PTR_TO_PACKET &&
13727 src_reg->type == PTR_TO_PACKET_END) ||
13728 (dst_reg->type == PTR_TO_PACKET_META &&
13729 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13730 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
13731 find_good_pkt_pointers(other_branch, dst_reg,
13732 dst_reg->type, true);
13733 mark_pkt_end(this_branch, insn->dst_reg, false);
13734 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
13735 src_reg->type == PTR_TO_PACKET) ||
13736 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
13737 src_reg->type == PTR_TO_PACKET_META)) {
13738 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
13739 find_good_pkt_pointers(this_branch, src_reg,
13740 src_reg->type, false);
13741 mark_pkt_end(other_branch, insn->src_reg, true);
13747 if ((dst_reg->type == PTR_TO_PACKET &&
13748 src_reg->type == PTR_TO_PACKET_END) ||
13749 (dst_reg->type == PTR_TO_PACKET_META &&
13750 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13751 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
13752 find_good_pkt_pointers(this_branch, dst_reg,
13753 dst_reg->type, true);
13754 mark_pkt_end(other_branch, insn->dst_reg, false);
13755 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
13756 src_reg->type == PTR_TO_PACKET) ||
13757 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
13758 src_reg->type == PTR_TO_PACKET_META)) {
13759 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
13760 find_good_pkt_pointers(other_branch, src_reg,
13761 src_reg->type, false);
13762 mark_pkt_end(this_branch, insn->src_reg, true);
13768 if ((dst_reg->type == PTR_TO_PACKET &&
13769 src_reg->type == PTR_TO_PACKET_END) ||
13770 (dst_reg->type == PTR_TO_PACKET_META &&
13771 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
13772 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
13773 find_good_pkt_pointers(other_branch, dst_reg,
13774 dst_reg->type, false);
13775 mark_pkt_end(this_branch, insn->dst_reg, true);
13776 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
13777 src_reg->type == PTR_TO_PACKET) ||
13778 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
13779 src_reg->type == PTR_TO_PACKET_META)) {
13780 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
13781 find_good_pkt_pointers(this_branch, src_reg,
13782 src_reg->type, true);
13783 mark_pkt_end(other_branch, insn->src_reg, false);
13795 static void find_equal_scalars(struct bpf_verifier_state *vstate,
13796 struct bpf_reg_state *known_reg)
13798 struct bpf_func_state *state;
13799 struct bpf_reg_state *reg;
13801 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
13802 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
13803 copy_register_state(reg, known_reg);
13807 static int check_cond_jmp_op(struct bpf_verifier_env *env,
13808 struct bpf_insn *insn, int *insn_idx)
13810 struct bpf_verifier_state *this_branch = env->cur_state;
13811 struct bpf_verifier_state *other_branch;
13812 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
13813 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
13814 struct bpf_reg_state *eq_branch_regs;
13815 u8 opcode = BPF_OP(insn->code);
13820 /* Only conditional jumps are expected to reach here. */
13821 if (opcode == BPF_JA || opcode > BPF_JSLE) {
13822 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
13826 if (BPF_SRC(insn->code) == BPF_X) {
13827 if (insn->imm != 0) {
13828 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
13832 /* check src1 operand */
13833 err = check_reg_arg(env, insn->src_reg, SRC_OP);
13837 if (is_pointer_value(env, insn->src_reg)) {
13838 verbose(env, "R%d pointer comparison prohibited\n",
13842 src_reg = ®s[insn->src_reg];
13844 if (insn->src_reg != BPF_REG_0) {
13845 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
13850 /* check src2 operand */
13851 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
13855 dst_reg = ®s[insn->dst_reg];
13856 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
13858 if (BPF_SRC(insn->code) == BPF_K) {
13859 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
13860 } else if (src_reg->type == SCALAR_VALUE &&
13861 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
13862 pred = is_branch_taken(dst_reg,
13863 tnum_subreg(src_reg->var_off).value,
13866 } else if (src_reg->type == SCALAR_VALUE &&
13867 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
13868 pred = is_branch_taken(dst_reg,
13869 src_reg->var_off.value,
13872 } else if (dst_reg->type == SCALAR_VALUE &&
13873 is_jmp32 && tnum_is_const(tnum_subreg(dst_reg->var_off))) {
13874 pred = is_branch_taken(src_reg,
13875 tnum_subreg(dst_reg->var_off).value,
13876 flip_opcode(opcode),
13878 } else if (dst_reg->type == SCALAR_VALUE &&
13879 !is_jmp32 && tnum_is_const(dst_reg->var_off)) {
13880 pred = is_branch_taken(src_reg,
13881 dst_reg->var_off.value,
13882 flip_opcode(opcode),
13884 } else if (reg_is_pkt_pointer_any(dst_reg) &&
13885 reg_is_pkt_pointer_any(src_reg) &&
13887 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
13891 /* If we get here with a dst_reg pointer type it is because
13892 * above is_branch_taken() special cased the 0 comparison.
13894 if (!__is_pointer_value(false, dst_reg))
13895 err = mark_chain_precision(env, insn->dst_reg);
13896 if (BPF_SRC(insn->code) == BPF_X && !err &&
13897 !__is_pointer_value(false, src_reg))
13898 err = mark_chain_precision(env, insn->src_reg);
13904 /* Only follow the goto, ignore fall-through. If needed, push
13905 * the fall-through branch for simulation under speculative
13908 if (!env->bypass_spec_v1 &&
13909 !sanitize_speculative_path(env, insn, *insn_idx + 1,
13912 *insn_idx += insn->off;
13914 } else if (pred == 0) {
13915 /* Only follow the fall-through branch, since that's where the
13916 * program will go. If needed, push the goto branch for
13917 * simulation under speculative execution.
13919 if (!env->bypass_spec_v1 &&
13920 !sanitize_speculative_path(env, insn,
13921 *insn_idx + insn->off + 1,
13927 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
13931 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
13933 /* detect if we are comparing against a constant value so we can adjust
13934 * our min/max values for our dst register.
13935 * this is only legit if both are scalars (or pointers to the same
13936 * object, I suppose, see the PTR_MAYBE_NULL related if block below),
13937 * because otherwise the different base pointers mean the offsets aren't
13940 if (BPF_SRC(insn->code) == BPF_X) {
13941 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
13943 if (dst_reg->type == SCALAR_VALUE &&
13944 src_reg->type == SCALAR_VALUE) {
13945 if (tnum_is_const(src_reg->var_off) ||
13947 tnum_is_const(tnum_subreg(src_reg->var_off))))
13948 reg_set_min_max(&other_branch_regs[insn->dst_reg],
13950 src_reg->var_off.value,
13951 tnum_subreg(src_reg->var_off).value,
13953 else if (tnum_is_const(dst_reg->var_off) ||
13955 tnum_is_const(tnum_subreg(dst_reg->var_off))))
13956 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
13958 dst_reg->var_off.value,
13959 tnum_subreg(dst_reg->var_off).value,
13961 else if (!is_jmp32 &&
13962 (opcode == BPF_JEQ || opcode == BPF_JNE))
13963 /* Comparing for equality, we can combine knowledge */
13964 reg_combine_min_max(&other_branch_regs[insn->src_reg],
13965 &other_branch_regs[insn->dst_reg],
13966 src_reg, dst_reg, opcode);
13968 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
13969 find_equal_scalars(this_branch, src_reg);
13970 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
13974 } else if (dst_reg->type == SCALAR_VALUE) {
13975 reg_set_min_max(&other_branch_regs[insn->dst_reg],
13976 dst_reg, insn->imm, (u32)insn->imm,
13980 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
13981 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
13982 find_equal_scalars(this_branch, dst_reg);
13983 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
13986 /* if one pointer register is compared to another pointer
13987 * register check if PTR_MAYBE_NULL could be lifted.
13988 * E.g. register A - maybe null
13989 * register B - not null
13990 * for JNE A, B, ... - A is not null in the false branch;
13991 * for JEQ A, B, ... - A is not null in the true branch.
13993 * Since PTR_TO_BTF_ID points to a kernel struct that does
13994 * not need to be null checked by the BPF program, i.e.,
13995 * could be null even without PTR_MAYBE_NULL marking, so
13996 * only propagate nullness when neither reg is that type.
13998 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_X &&
13999 __is_pointer_value(false, src_reg) && __is_pointer_value(false, dst_reg) &&
14000 type_may_be_null(src_reg->type) != type_may_be_null(dst_reg->type) &&
14001 base_type(src_reg->type) != PTR_TO_BTF_ID &&
14002 base_type(dst_reg->type) != PTR_TO_BTF_ID) {
14003 eq_branch_regs = NULL;
14006 eq_branch_regs = other_branch_regs;
14009 eq_branch_regs = regs;
14015 if (eq_branch_regs) {
14016 if (type_may_be_null(src_reg->type))
14017 mark_ptr_not_null_reg(&eq_branch_regs[insn->src_reg]);
14019 mark_ptr_not_null_reg(&eq_branch_regs[insn->dst_reg]);
14023 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
14024 * NOTE: these optimizations below are related with pointer comparison
14025 * which will never be JMP32.
14027 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
14028 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
14029 type_may_be_null(dst_reg->type)) {
14030 /* Mark all identical registers in each branch as either
14031 * safe or unknown depending R == 0 or R != 0 conditional.
14033 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
14034 opcode == BPF_JNE);
14035 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
14036 opcode == BPF_JEQ);
14037 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
14038 this_branch, other_branch) &&
14039 is_pointer_value(env, insn->dst_reg)) {
14040 verbose(env, "R%d pointer comparison prohibited\n",
14044 if (env->log.level & BPF_LOG_LEVEL)
14045 print_insn_state(env, this_branch->frame[this_branch->curframe]);
14049 /* verify BPF_LD_IMM64 instruction */
14050 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
14052 struct bpf_insn_aux_data *aux = cur_aux(env);
14053 struct bpf_reg_state *regs = cur_regs(env);
14054 struct bpf_reg_state *dst_reg;
14055 struct bpf_map *map;
14058 if (BPF_SIZE(insn->code) != BPF_DW) {
14059 verbose(env, "invalid BPF_LD_IMM insn\n");
14062 if (insn->off != 0) {
14063 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
14067 err = check_reg_arg(env, insn->dst_reg, DST_OP);
14071 dst_reg = ®s[insn->dst_reg];
14072 if (insn->src_reg == 0) {
14073 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
14075 dst_reg->type = SCALAR_VALUE;
14076 __mark_reg_known(®s[insn->dst_reg], imm);
14080 /* All special src_reg cases are listed below. From this point onwards
14081 * we either succeed and assign a corresponding dst_reg->type after
14082 * zeroing the offset, or fail and reject the program.
14084 mark_reg_known_zero(env, regs, insn->dst_reg);
14086 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
14087 dst_reg->type = aux->btf_var.reg_type;
14088 switch (base_type(dst_reg->type)) {
14090 dst_reg->mem_size = aux->btf_var.mem_size;
14092 case PTR_TO_BTF_ID:
14093 dst_reg->btf = aux->btf_var.btf;
14094 dst_reg->btf_id = aux->btf_var.btf_id;
14097 verbose(env, "bpf verifier is misconfigured\n");
14103 if (insn->src_reg == BPF_PSEUDO_FUNC) {
14104 struct bpf_prog_aux *aux = env->prog->aux;
14105 u32 subprogno = find_subprog(env,
14106 env->insn_idx + insn->imm + 1);
14108 if (!aux->func_info) {
14109 verbose(env, "missing btf func_info\n");
14112 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
14113 verbose(env, "callback function not static\n");
14117 dst_reg->type = PTR_TO_FUNC;
14118 dst_reg->subprogno = subprogno;
14122 map = env->used_maps[aux->map_index];
14123 dst_reg->map_ptr = map;
14125 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
14126 insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
14127 dst_reg->type = PTR_TO_MAP_VALUE;
14128 dst_reg->off = aux->map_off;
14129 WARN_ON_ONCE(map->max_entries != 1);
14130 /* We want reg->id to be same (0) as map_value is not distinct */
14131 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
14132 insn->src_reg == BPF_PSEUDO_MAP_IDX) {
14133 dst_reg->type = CONST_PTR_TO_MAP;
14135 verbose(env, "bpf verifier is misconfigured\n");
14142 static bool may_access_skb(enum bpf_prog_type type)
14145 case BPF_PROG_TYPE_SOCKET_FILTER:
14146 case BPF_PROG_TYPE_SCHED_CLS:
14147 case BPF_PROG_TYPE_SCHED_ACT:
14154 /* verify safety of LD_ABS|LD_IND instructions:
14155 * - they can only appear in the programs where ctx == skb
14156 * - since they are wrappers of function calls, they scratch R1-R5 registers,
14157 * preserve R6-R9, and store return value into R0
14160 * ctx == skb == R6 == CTX
14163 * SRC == any register
14164 * IMM == 32-bit immediate
14167 * R0 - 8/16/32-bit skb data converted to cpu endianness
14169 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
14171 struct bpf_reg_state *regs = cur_regs(env);
14172 static const int ctx_reg = BPF_REG_6;
14173 u8 mode = BPF_MODE(insn->code);
14176 if (!may_access_skb(resolve_prog_type(env->prog))) {
14177 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
14181 if (!env->ops->gen_ld_abs) {
14182 verbose(env, "bpf verifier is misconfigured\n");
14186 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
14187 BPF_SIZE(insn->code) == BPF_DW ||
14188 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
14189 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
14193 /* check whether implicit source operand (register R6) is readable */
14194 err = check_reg_arg(env, ctx_reg, SRC_OP);
14198 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
14199 * gen_ld_abs() may terminate the program at runtime, leading to
14202 err = check_reference_leak(env);
14204 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
14208 if (env->cur_state->active_lock.ptr) {
14209 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
14213 if (env->cur_state->active_rcu_lock) {
14214 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_rcu_read_lock-ed region\n");
14218 if (regs[ctx_reg].type != PTR_TO_CTX) {
14220 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
14224 if (mode == BPF_IND) {
14225 /* check explicit source operand */
14226 err = check_reg_arg(env, insn->src_reg, SRC_OP);
14231 err = check_ptr_off_reg(env, ®s[ctx_reg], ctx_reg);
14235 /* reset caller saved regs to unreadable */
14236 for (i = 0; i < CALLER_SAVED_REGS; i++) {
14237 mark_reg_not_init(env, regs, caller_saved[i]);
14238 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
14241 /* mark destination R0 register as readable, since it contains
14242 * the value fetched from the packet.
14243 * Already marked as written above.
14245 mark_reg_unknown(env, regs, BPF_REG_0);
14246 /* ld_abs load up to 32-bit skb data. */
14247 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
14251 static int check_return_code(struct bpf_verifier_env *env)
14253 struct tnum enforce_attach_type_range = tnum_unknown;
14254 const struct bpf_prog *prog = env->prog;
14255 struct bpf_reg_state *reg;
14256 struct tnum range = tnum_range(0, 1);
14257 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
14259 struct bpf_func_state *frame = env->cur_state->frame[0];
14260 const bool is_subprog = frame->subprogno;
14262 /* LSM and struct_ops func-ptr's return type could be "void" */
14264 switch (prog_type) {
14265 case BPF_PROG_TYPE_LSM:
14266 if (prog->expected_attach_type == BPF_LSM_CGROUP)
14267 /* See below, can be 0 or 0-1 depending on hook. */
14270 case BPF_PROG_TYPE_STRUCT_OPS:
14271 if (!prog->aux->attach_func_proto->type)
14279 /* eBPF calling convention is such that R0 is used
14280 * to return the value from eBPF program.
14281 * Make sure that it's readable at this time
14282 * of bpf_exit, which means that program wrote
14283 * something into it earlier
14285 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
14289 if (is_pointer_value(env, BPF_REG_0)) {
14290 verbose(env, "R0 leaks addr as return value\n");
14294 reg = cur_regs(env) + BPF_REG_0;
14296 if (frame->in_async_callback_fn) {
14297 /* enforce return zero from async callbacks like timer */
14298 if (reg->type != SCALAR_VALUE) {
14299 verbose(env, "In async callback the register R0 is not a known value (%s)\n",
14300 reg_type_str(env, reg->type));
14304 if (!tnum_in(tnum_const(0), reg->var_off)) {
14305 verbose_invalid_scalar(env, reg, &range, "async callback", "R0");
14312 if (reg->type != SCALAR_VALUE) {
14313 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
14314 reg_type_str(env, reg->type));
14320 switch (prog_type) {
14321 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
14322 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
14323 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
14324 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
14325 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
14326 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
14327 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
14328 range = tnum_range(1, 1);
14329 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
14330 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
14331 range = tnum_range(0, 3);
14333 case BPF_PROG_TYPE_CGROUP_SKB:
14334 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
14335 range = tnum_range(0, 3);
14336 enforce_attach_type_range = tnum_range(2, 3);
14339 case BPF_PROG_TYPE_CGROUP_SOCK:
14340 case BPF_PROG_TYPE_SOCK_OPS:
14341 case BPF_PROG_TYPE_CGROUP_DEVICE:
14342 case BPF_PROG_TYPE_CGROUP_SYSCTL:
14343 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
14345 case BPF_PROG_TYPE_RAW_TRACEPOINT:
14346 if (!env->prog->aux->attach_btf_id)
14348 range = tnum_const(0);
14350 case BPF_PROG_TYPE_TRACING:
14351 switch (env->prog->expected_attach_type) {
14352 case BPF_TRACE_FENTRY:
14353 case BPF_TRACE_FEXIT:
14354 range = tnum_const(0);
14356 case BPF_TRACE_RAW_TP:
14357 case BPF_MODIFY_RETURN:
14359 case BPF_TRACE_ITER:
14365 case BPF_PROG_TYPE_SK_LOOKUP:
14366 range = tnum_range(SK_DROP, SK_PASS);
14369 case BPF_PROG_TYPE_LSM:
14370 if (env->prog->expected_attach_type != BPF_LSM_CGROUP) {
14371 /* Regular BPF_PROG_TYPE_LSM programs can return
14376 if (!env->prog->aux->attach_func_proto->type) {
14377 /* Make sure programs that attach to void
14378 * hooks don't try to modify return value.
14380 range = tnum_range(1, 1);
14384 case BPF_PROG_TYPE_NETFILTER:
14385 range = tnum_range(NF_DROP, NF_ACCEPT);
14387 case BPF_PROG_TYPE_EXT:
14388 /* freplace program can return anything as its return value
14389 * depends on the to-be-replaced kernel func or bpf program.
14395 if (reg->type != SCALAR_VALUE) {
14396 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
14397 reg_type_str(env, reg->type));
14401 if (!tnum_in(range, reg->var_off)) {
14402 verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
14403 if (prog->expected_attach_type == BPF_LSM_CGROUP &&
14404 prog_type == BPF_PROG_TYPE_LSM &&
14405 !prog->aux->attach_func_proto->type)
14406 verbose(env, "Note, BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
14410 if (!tnum_is_unknown(enforce_attach_type_range) &&
14411 tnum_in(enforce_attach_type_range, reg->var_off))
14412 env->prog->enforce_expected_attach_type = 1;
14416 /* non-recursive DFS pseudo code
14417 * 1 procedure DFS-iterative(G,v):
14418 * 2 label v as discovered
14419 * 3 let S be a stack
14421 * 5 while S is not empty
14423 * 7 if t is what we're looking for:
14425 * 9 for all edges e in G.adjacentEdges(t) do
14426 * 10 if edge e is already labelled
14427 * 11 continue with the next edge
14428 * 12 w <- G.adjacentVertex(t,e)
14429 * 13 if vertex w is not discovered and not explored
14430 * 14 label e as tree-edge
14431 * 15 label w as discovered
14434 * 18 else if vertex w is discovered
14435 * 19 label e as back-edge
14437 * 21 // vertex w is explored
14438 * 22 label e as forward- or cross-edge
14439 * 23 label t as explored
14443 * 0x10 - discovered
14444 * 0x11 - discovered and fall-through edge labelled
14445 * 0x12 - discovered and fall-through and branch edges labelled
14456 static u32 state_htab_size(struct bpf_verifier_env *env)
14458 return env->prog->len;
14461 static struct bpf_verifier_state_list **explored_state(
14462 struct bpf_verifier_env *env,
14465 struct bpf_verifier_state *cur = env->cur_state;
14466 struct bpf_func_state *state = cur->frame[cur->curframe];
14468 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
14471 static void mark_prune_point(struct bpf_verifier_env *env, int idx)
14473 env->insn_aux_data[idx].prune_point = true;
14476 static bool is_prune_point(struct bpf_verifier_env *env, int insn_idx)
14478 return env->insn_aux_data[insn_idx].prune_point;
14481 static void mark_force_checkpoint(struct bpf_verifier_env *env, int idx)
14483 env->insn_aux_data[idx].force_checkpoint = true;
14486 static bool is_force_checkpoint(struct bpf_verifier_env *env, int insn_idx)
14488 return env->insn_aux_data[insn_idx].force_checkpoint;
14493 DONE_EXPLORING = 0,
14494 KEEP_EXPLORING = 1,
14497 /* t, w, e - match pseudo-code above:
14498 * t - index of current instruction
14499 * w - next instruction
14502 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
14505 int *insn_stack = env->cfg.insn_stack;
14506 int *insn_state = env->cfg.insn_state;
14508 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
14509 return DONE_EXPLORING;
14511 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
14512 return DONE_EXPLORING;
14514 if (w < 0 || w >= env->prog->len) {
14515 verbose_linfo(env, t, "%d: ", t);
14516 verbose(env, "jump out of range from insn %d to %d\n", t, w);
14521 /* mark branch target for state pruning */
14522 mark_prune_point(env, w);
14523 mark_jmp_point(env, w);
14526 if (insn_state[w] == 0) {
14528 insn_state[t] = DISCOVERED | e;
14529 insn_state[w] = DISCOVERED;
14530 if (env->cfg.cur_stack >= env->prog->len)
14532 insn_stack[env->cfg.cur_stack++] = w;
14533 return KEEP_EXPLORING;
14534 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
14535 if (loop_ok && env->bpf_capable)
14536 return DONE_EXPLORING;
14537 verbose_linfo(env, t, "%d: ", t);
14538 verbose_linfo(env, w, "%d: ", w);
14539 verbose(env, "back-edge from insn %d to %d\n", t, w);
14541 } else if (insn_state[w] == EXPLORED) {
14542 /* forward- or cross-edge */
14543 insn_state[t] = DISCOVERED | e;
14545 verbose(env, "insn state internal bug\n");
14548 return DONE_EXPLORING;
14551 static int visit_func_call_insn(int t, struct bpf_insn *insns,
14552 struct bpf_verifier_env *env,
14557 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
14561 mark_prune_point(env, t + 1);
14562 /* when we exit from subprog, we need to record non-linear history */
14563 mark_jmp_point(env, t + 1);
14565 if (visit_callee) {
14566 mark_prune_point(env, t);
14567 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env,
14568 /* It's ok to allow recursion from CFG point of
14569 * view. __check_func_call() will do the actual
14572 bpf_pseudo_func(insns + t));
14577 /* Visits the instruction at index t and returns one of the following:
14578 * < 0 - an error occurred
14579 * DONE_EXPLORING - the instruction was fully explored
14580 * KEEP_EXPLORING - there is still work to be done before it is fully explored
14582 static int visit_insn(int t, struct bpf_verifier_env *env)
14584 struct bpf_insn *insns = env->prog->insnsi, *insn = &insns[t];
14587 if (bpf_pseudo_func(insn))
14588 return visit_func_call_insn(t, insns, env, true);
14590 /* All non-branch instructions have a single fall-through edge. */
14591 if (BPF_CLASS(insn->code) != BPF_JMP &&
14592 BPF_CLASS(insn->code) != BPF_JMP32)
14593 return push_insn(t, t + 1, FALLTHROUGH, env, false);
14595 switch (BPF_OP(insn->code)) {
14597 return DONE_EXPLORING;
14600 if (insn->src_reg == 0 && insn->imm == BPF_FUNC_timer_set_callback)
14601 /* Mark this call insn as a prune point to trigger
14602 * is_state_visited() check before call itself is
14603 * processed by __check_func_call(). Otherwise new
14604 * async state will be pushed for further exploration.
14606 mark_prune_point(env, t);
14607 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
14608 struct bpf_kfunc_call_arg_meta meta;
14610 ret = fetch_kfunc_meta(env, insn, &meta, NULL);
14611 if (ret == 0 && is_iter_next_kfunc(&meta)) {
14612 mark_prune_point(env, t);
14613 /* Checking and saving state checkpoints at iter_next() call
14614 * is crucial for fast convergence of open-coded iterator loop
14615 * logic, so we need to force it. If we don't do that,
14616 * is_state_visited() might skip saving a checkpoint, causing
14617 * unnecessarily long sequence of not checkpointed
14618 * instructions and jumps, leading to exhaustion of jump
14619 * history buffer, and potentially other undesired outcomes.
14620 * It is expected that with correct open-coded iterators
14621 * convergence will happen quickly, so we don't run a risk of
14622 * exhausting memory.
14624 mark_force_checkpoint(env, t);
14627 return visit_func_call_insn(t, insns, env, insn->src_reg == BPF_PSEUDO_CALL);
14630 if (BPF_SRC(insn->code) != BPF_K)
14633 /* unconditional jump with single edge */
14634 ret = push_insn(t, t + insn->off + 1, FALLTHROUGH, env,
14639 mark_prune_point(env, t + insn->off + 1);
14640 mark_jmp_point(env, t + insn->off + 1);
14645 /* conditional jump with two edges */
14646 mark_prune_point(env, t);
14648 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
14652 return push_insn(t, t + insn->off + 1, BRANCH, env, true);
14656 /* non-recursive depth-first-search to detect loops in BPF program
14657 * loop == back-edge in directed graph
14659 static int check_cfg(struct bpf_verifier_env *env)
14661 int insn_cnt = env->prog->len;
14662 int *insn_stack, *insn_state;
14666 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
14670 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
14672 kvfree(insn_state);
14676 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
14677 insn_stack[0] = 0; /* 0 is the first instruction */
14678 env->cfg.cur_stack = 1;
14680 while (env->cfg.cur_stack > 0) {
14681 int t = insn_stack[env->cfg.cur_stack - 1];
14683 ret = visit_insn(t, env);
14685 case DONE_EXPLORING:
14686 insn_state[t] = EXPLORED;
14687 env->cfg.cur_stack--;
14689 case KEEP_EXPLORING:
14693 verbose(env, "visit_insn internal bug\n");
14700 if (env->cfg.cur_stack < 0) {
14701 verbose(env, "pop stack internal bug\n");
14706 for (i = 0; i < insn_cnt; i++) {
14707 if (insn_state[i] != EXPLORED) {
14708 verbose(env, "unreachable insn %d\n", i);
14713 ret = 0; /* cfg looks good */
14716 kvfree(insn_state);
14717 kvfree(insn_stack);
14718 env->cfg.insn_state = env->cfg.insn_stack = NULL;
14722 static int check_abnormal_return(struct bpf_verifier_env *env)
14726 for (i = 1; i < env->subprog_cnt; i++) {
14727 if (env->subprog_info[i].has_ld_abs) {
14728 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
14731 if (env->subprog_info[i].has_tail_call) {
14732 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
14739 /* The minimum supported BTF func info size */
14740 #define MIN_BPF_FUNCINFO_SIZE 8
14741 #define MAX_FUNCINFO_REC_SIZE 252
14743 static int check_btf_func(struct bpf_verifier_env *env,
14744 const union bpf_attr *attr,
14747 const struct btf_type *type, *func_proto, *ret_type;
14748 u32 i, nfuncs, urec_size, min_size;
14749 u32 krec_size = sizeof(struct bpf_func_info);
14750 struct bpf_func_info *krecord;
14751 struct bpf_func_info_aux *info_aux = NULL;
14752 struct bpf_prog *prog;
14753 const struct btf *btf;
14755 u32 prev_offset = 0;
14756 bool scalar_return;
14759 nfuncs = attr->func_info_cnt;
14761 if (check_abnormal_return(env))
14766 if (nfuncs != env->subprog_cnt) {
14767 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
14771 urec_size = attr->func_info_rec_size;
14772 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
14773 urec_size > MAX_FUNCINFO_REC_SIZE ||
14774 urec_size % sizeof(u32)) {
14775 verbose(env, "invalid func info rec size %u\n", urec_size);
14780 btf = prog->aux->btf;
14782 urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
14783 min_size = min_t(u32, krec_size, urec_size);
14785 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
14788 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
14792 for (i = 0; i < nfuncs; i++) {
14793 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
14795 if (ret == -E2BIG) {
14796 verbose(env, "nonzero tailing record in func info");
14797 /* set the size kernel expects so loader can zero
14798 * out the rest of the record.
14800 if (copy_to_bpfptr_offset(uattr,
14801 offsetof(union bpf_attr, func_info_rec_size),
14802 &min_size, sizeof(min_size)))
14808 if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
14813 /* check insn_off */
14816 if (krecord[i].insn_off) {
14818 "nonzero insn_off %u for the first func info record",
14819 krecord[i].insn_off);
14822 } else if (krecord[i].insn_off <= prev_offset) {
14824 "same or smaller insn offset (%u) than previous func info record (%u)",
14825 krecord[i].insn_off, prev_offset);
14829 if (env->subprog_info[i].start != krecord[i].insn_off) {
14830 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
14834 /* check type_id */
14835 type = btf_type_by_id(btf, krecord[i].type_id);
14836 if (!type || !btf_type_is_func(type)) {
14837 verbose(env, "invalid type id %d in func info",
14838 krecord[i].type_id);
14841 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
14843 func_proto = btf_type_by_id(btf, type->type);
14844 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
14845 /* btf_func_check() already verified it during BTF load */
14847 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
14849 btf_type_is_small_int(ret_type) || btf_is_any_enum(ret_type);
14850 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
14851 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
14854 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
14855 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
14859 prev_offset = krecord[i].insn_off;
14860 bpfptr_add(&urecord, urec_size);
14863 prog->aux->func_info = krecord;
14864 prog->aux->func_info_cnt = nfuncs;
14865 prog->aux->func_info_aux = info_aux;
14874 static void adjust_btf_func(struct bpf_verifier_env *env)
14876 struct bpf_prog_aux *aux = env->prog->aux;
14879 if (!aux->func_info)
14882 for (i = 0; i < env->subprog_cnt; i++)
14883 aux->func_info[i].insn_off = env->subprog_info[i].start;
14886 #define MIN_BPF_LINEINFO_SIZE offsetofend(struct bpf_line_info, line_col)
14887 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
14889 static int check_btf_line(struct bpf_verifier_env *env,
14890 const union bpf_attr *attr,
14893 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
14894 struct bpf_subprog_info *sub;
14895 struct bpf_line_info *linfo;
14896 struct bpf_prog *prog;
14897 const struct btf *btf;
14901 nr_linfo = attr->line_info_cnt;
14904 if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info))
14907 rec_size = attr->line_info_rec_size;
14908 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
14909 rec_size > MAX_LINEINFO_REC_SIZE ||
14910 rec_size & (sizeof(u32) - 1))
14913 /* Need to zero it in case the userspace may
14914 * pass in a smaller bpf_line_info object.
14916 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
14917 GFP_KERNEL | __GFP_NOWARN);
14922 btf = prog->aux->btf;
14925 sub = env->subprog_info;
14926 ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
14927 expected_size = sizeof(struct bpf_line_info);
14928 ncopy = min_t(u32, expected_size, rec_size);
14929 for (i = 0; i < nr_linfo; i++) {
14930 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
14932 if (err == -E2BIG) {
14933 verbose(env, "nonzero tailing record in line_info");
14934 if (copy_to_bpfptr_offset(uattr,
14935 offsetof(union bpf_attr, line_info_rec_size),
14936 &expected_size, sizeof(expected_size)))
14942 if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
14948 * Check insn_off to ensure
14949 * 1) strictly increasing AND
14950 * 2) bounded by prog->len
14952 * The linfo[0].insn_off == 0 check logically falls into
14953 * the later "missing bpf_line_info for func..." case
14954 * because the first linfo[0].insn_off must be the
14955 * first sub also and the first sub must have
14956 * subprog_info[0].start == 0.
14958 if ((i && linfo[i].insn_off <= prev_offset) ||
14959 linfo[i].insn_off >= prog->len) {
14960 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
14961 i, linfo[i].insn_off, prev_offset,
14967 if (!prog->insnsi[linfo[i].insn_off].code) {
14969 "Invalid insn code at line_info[%u].insn_off\n",
14975 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
14976 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
14977 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
14982 if (s != env->subprog_cnt) {
14983 if (linfo[i].insn_off == sub[s].start) {
14984 sub[s].linfo_idx = i;
14986 } else if (sub[s].start < linfo[i].insn_off) {
14987 verbose(env, "missing bpf_line_info for func#%u\n", s);
14993 prev_offset = linfo[i].insn_off;
14994 bpfptr_add(&ulinfo, rec_size);
14997 if (s != env->subprog_cnt) {
14998 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
14999 env->subprog_cnt - s, s);
15004 prog->aux->linfo = linfo;
15005 prog->aux->nr_linfo = nr_linfo;
15014 #define MIN_CORE_RELO_SIZE sizeof(struct bpf_core_relo)
15015 #define MAX_CORE_RELO_SIZE MAX_FUNCINFO_REC_SIZE
15017 static int check_core_relo(struct bpf_verifier_env *env,
15018 const union bpf_attr *attr,
15021 u32 i, nr_core_relo, ncopy, expected_size, rec_size;
15022 struct bpf_core_relo core_relo = {};
15023 struct bpf_prog *prog = env->prog;
15024 const struct btf *btf = prog->aux->btf;
15025 struct bpf_core_ctx ctx = {
15029 bpfptr_t u_core_relo;
15032 nr_core_relo = attr->core_relo_cnt;
15035 if (nr_core_relo > INT_MAX / sizeof(struct bpf_core_relo))
15038 rec_size = attr->core_relo_rec_size;
15039 if (rec_size < MIN_CORE_RELO_SIZE ||
15040 rec_size > MAX_CORE_RELO_SIZE ||
15041 rec_size % sizeof(u32))
15044 u_core_relo = make_bpfptr(attr->core_relos, uattr.is_kernel);
15045 expected_size = sizeof(struct bpf_core_relo);
15046 ncopy = min_t(u32, expected_size, rec_size);
15048 /* Unlike func_info and line_info, copy and apply each CO-RE
15049 * relocation record one at a time.
15051 for (i = 0; i < nr_core_relo; i++) {
15052 /* future proofing when sizeof(bpf_core_relo) changes */
15053 err = bpf_check_uarg_tail_zero(u_core_relo, expected_size, rec_size);
15055 if (err == -E2BIG) {
15056 verbose(env, "nonzero tailing record in core_relo");
15057 if (copy_to_bpfptr_offset(uattr,
15058 offsetof(union bpf_attr, core_relo_rec_size),
15059 &expected_size, sizeof(expected_size)))
15065 if (copy_from_bpfptr(&core_relo, u_core_relo, ncopy)) {
15070 if (core_relo.insn_off % 8 || core_relo.insn_off / 8 >= prog->len) {
15071 verbose(env, "Invalid core_relo[%u].insn_off:%u prog->len:%u\n",
15072 i, core_relo.insn_off, prog->len);
15077 err = bpf_core_apply(&ctx, &core_relo, i,
15078 &prog->insnsi[core_relo.insn_off / 8]);
15081 bpfptr_add(&u_core_relo, rec_size);
15086 static int check_btf_info(struct bpf_verifier_env *env,
15087 const union bpf_attr *attr,
15093 if (!attr->func_info_cnt && !attr->line_info_cnt) {
15094 if (check_abnormal_return(env))
15099 btf = btf_get_by_fd(attr->prog_btf_fd);
15101 return PTR_ERR(btf);
15102 if (btf_is_kernel(btf)) {
15106 env->prog->aux->btf = btf;
15108 err = check_btf_func(env, attr, uattr);
15112 err = check_btf_line(env, attr, uattr);
15116 err = check_core_relo(env, attr, uattr);
15123 /* check %cur's range satisfies %old's */
15124 static bool range_within(struct bpf_reg_state *old,
15125 struct bpf_reg_state *cur)
15127 return old->umin_value <= cur->umin_value &&
15128 old->umax_value >= cur->umax_value &&
15129 old->smin_value <= cur->smin_value &&
15130 old->smax_value >= cur->smax_value &&
15131 old->u32_min_value <= cur->u32_min_value &&
15132 old->u32_max_value >= cur->u32_max_value &&
15133 old->s32_min_value <= cur->s32_min_value &&
15134 old->s32_max_value >= cur->s32_max_value;
15137 /* If in the old state two registers had the same id, then they need to have
15138 * the same id in the new state as well. But that id could be different from
15139 * the old state, so we need to track the mapping from old to new ids.
15140 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
15141 * regs with old id 5 must also have new id 9 for the new state to be safe. But
15142 * regs with a different old id could still have new id 9, we don't care about
15144 * So we look through our idmap to see if this old id has been seen before. If
15145 * so, we require the new id to match; otherwise, we add the id pair to the map.
15147 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap)
15149 struct bpf_id_pair *map = idmap->map;
15152 /* either both IDs should be set or both should be zero */
15153 if (!!old_id != !!cur_id)
15156 if (old_id == 0) /* cur_id == 0 as well */
15159 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
15161 /* Reached an empty slot; haven't seen this id before */
15162 map[i].old = old_id;
15163 map[i].cur = cur_id;
15166 if (map[i].old == old_id)
15167 return map[i].cur == cur_id;
15168 if (map[i].cur == cur_id)
15171 /* We ran out of idmap slots, which should be impossible */
15176 /* Similar to check_ids(), but allocate a unique temporary ID
15177 * for 'old_id' or 'cur_id' of zero.
15178 * This makes pairs like '0 vs unique ID', 'unique ID vs 0' valid.
15180 static bool check_scalar_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap)
15182 old_id = old_id ? old_id : ++idmap->tmp_id_gen;
15183 cur_id = cur_id ? cur_id : ++idmap->tmp_id_gen;
15185 return check_ids(old_id, cur_id, idmap);
15188 static void clean_func_state(struct bpf_verifier_env *env,
15189 struct bpf_func_state *st)
15191 enum bpf_reg_liveness live;
15194 for (i = 0; i < BPF_REG_FP; i++) {
15195 live = st->regs[i].live;
15196 /* liveness must not touch this register anymore */
15197 st->regs[i].live |= REG_LIVE_DONE;
15198 if (!(live & REG_LIVE_READ))
15199 /* since the register is unused, clear its state
15200 * to make further comparison simpler
15202 __mark_reg_not_init(env, &st->regs[i]);
15205 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
15206 live = st->stack[i].spilled_ptr.live;
15207 /* liveness must not touch this stack slot anymore */
15208 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
15209 if (!(live & REG_LIVE_READ)) {
15210 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
15211 for (j = 0; j < BPF_REG_SIZE; j++)
15212 st->stack[i].slot_type[j] = STACK_INVALID;
15217 static void clean_verifier_state(struct bpf_verifier_env *env,
15218 struct bpf_verifier_state *st)
15222 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
15223 /* all regs in this state in all frames were already marked */
15226 for (i = 0; i <= st->curframe; i++)
15227 clean_func_state(env, st->frame[i]);
15230 /* the parentage chains form a tree.
15231 * the verifier states are added to state lists at given insn and
15232 * pushed into state stack for future exploration.
15233 * when the verifier reaches bpf_exit insn some of the verifer states
15234 * stored in the state lists have their final liveness state already,
15235 * but a lot of states will get revised from liveness point of view when
15236 * the verifier explores other branches.
15239 * 2: if r1 == 100 goto pc+1
15242 * when the verifier reaches exit insn the register r0 in the state list of
15243 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
15244 * of insn 2 and goes exploring further. At the insn 4 it will walk the
15245 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
15247 * Since the verifier pushes the branch states as it sees them while exploring
15248 * the program the condition of walking the branch instruction for the second
15249 * time means that all states below this branch were already explored and
15250 * their final liveness marks are already propagated.
15251 * Hence when the verifier completes the search of state list in is_state_visited()
15252 * we can call this clean_live_states() function to mark all liveness states
15253 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
15254 * will not be used.
15255 * This function also clears the registers and stack for states that !READ
15256 * to simplify state merging.
15258 * Important note here that walking the same branch instruction in the callee
15259 * doesn't meant that the states are DONE. The verifier has to compare
15262 static void clean_live_states(struct bpf_verifier_env *env, int insn,
15263 struct bpf_verifier_state *cur)
15265 struct bpf_verifier_state_list *sl;
15268 sl = *explored_state(env, insn);
15270 if (sl->state.branches)
15272 if (sl->state.insn_idx != insn ||
15273 sl->state.curframe != cur->curframe)
15275 for (i = 0; i <= cur->curframe; i++)
15276 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
15278 clean_verifier_state(env, &sl->state);
15284 static bool regs_exact(const struct bpf_reg_state *rold,
15285 const struct bpf_reg_state *rcur,
15286 struct bpf_idmap *idmap)
15288 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
15289 check_ids(rold->id, rcur->id, idmap) &&
15290 check_ids(rold->ref_obj_id, rcur->ref_obj_id, idmap);
15293 /* Returns true if (rold safe implies rcur safe) */
15294 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
15295 struct bpf_reg_state *rcur, struct bpf_idmap *idmap)
15297 if (!(rold->live & REG_LIVE_READ))
15298 /* explored state didn't use this */
15300 if (rold->type == NOT_INIT)
15301 /* explored state can't have used this */
15303 if (rcur->type == NOT_INIT)
15306 /* Enforce that register types have to match exactly, including their
15307 * modifiers (like PTR_MAYBE_NULL, MEM_RDONLY, etc), as a general
15310 * One can make a point that using a pointer register as unbounded
15311 * SCALAR would be technically acceptable, but this could lead to
15312 * pointer leaks because scalars are allowed to leak while pointers
15313 * are not. We could make this safe in special cases if root is
15314 * calling us, but it's probably not worth the hassle.
15316 * Also, register types that are *not* MAYBE_NULL could technically be
15317 * safe to use as their MAYBE_NULL variants (e.g., PTR_TO_MAP_VALUE
15318 * is safe to be used as PTR_TO_MAP_VALUE_OR_NULL, provided both point
15319 * to the same map).
15320 * However, if the old MAYBE_NULL register then got NULL checked,
15321 * doing so could have affected others with the same id, and we can't
15322 * check for that because we lost the id when we converted to
15323 * a non-MAYBE_NULL variant.
15324 * So, as a general rule we don't allow mixing MAYBE_NULL and
15325 * non-MAYBE_NULL registers as well.
15327 if (rold->type != rcur->type)
15330 switch (base_type(rold->type)) {
15332 if (env->explore_alu_limits) {
15333 /* explore_alu_limits disables tnum_in() and range_within()
15334 * logic and requires everything to be strict
15336 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
15337 check_scalar_ids(rold->id, rcur->id, idmap);
15339 if (!rold->precise)
15341 /* Why check_ids() for scalar registers?
15343 * Consider the following BPF code:
15344 * 1: r6 = ... unbound scalar, ID=a ...
15345 * 2: r7 = ... unbound scalar, ID=b ...
15346 * 3: if (r6 > r7) goto +1
15348 * 5: if (r6 > X) goto ...
15349 * 6: ... memory operation using r7 ...
15351 * First verification path is [1-6]:
15352 * - at (4) same bpf_reg_state::id (b) would be assigned to r6 and r7;
15353 * - at (5) r6 would be marked <= X, find_equal_scalars() would also mark
15354 * r7 <= X, because r6 and r7 share same id.
15355 * Next verification path is [1-4, 6].
15357 * Instruction (6) would be reached in two states:
15358 * I. r6{.id=b}, r7{.id=b} via path 1-6;
15359 * II. r6{.id=a}, r7{.id=b} via path 1-4, 6.
15361 * Use check_ids() to distinguish these states.
15363 * Also verify that new value satisfies old value range knowledge.
15365 return range_within(rold, rcur) &&
15366 tnum_in(rold->var_off, rcur->var_off) &&
15367 check_scalar_ids(rold->id, rcur->id, idmap);
15368 case PTR_TO_MAP_KEY:
15369 case PTR_TO_MAP_VALUE:
15372 case PTR_TO_TP_BUFFER:
15373 /* If the new min/max/var_off satisfy the old ones and
15374 * everything else matches, we are OK.
15376 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, var_off)) == 0 &&
15377 range_within(rold, rcur) &&
15378 tnum_in(rold->var_off, rcur->var_off) &&
15379 check_ids(rold->id, rcur->id, idmap) &&
15380 check_ids(rold->ref_obj_id, rcur->ref_obj_id, idmap);
15381 case PTR_TO_PACKET_META:
15382 case PTR_TO_PACKET:
15383 /* We must have at least as much range as the old ptr
15384 * did, so that any accesses which were safe before are
15385 * still safe. This is true even if old range < old off,
15386 * since someone could have accessed through (ptr - k), or
15387 * even done ptr -= k in a register, to get a safe access.
15389 if (rold->range > rcur->range)
15391 /* If the offsets don't match, we can't trust our alignment;
15392 * nor can we be sure that we won't fall out of range.
15394 if (rold->off != rcur->off)
15396 /* id relations must be preserved */
15397 if (!check_ids(rold->id, rcur->id, idmap))
15399 /* new val must satisfy old val knowledge */
15400 return range_within(rold, rcur) &&
15401 tnum_in(rold->var_off, rcur->var_off);
15403 /* two stack pointers are equal only if they're pointing to
15404 * the same stack frame, since fp-8 in foo != fp-8 in bar
15406 return regs_exact(rold, rcur, idmap) && rold->frameno == rcur->frameno;
15408 return regs_exact(rold, rcur, idmap);
15412 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
15413 struct bpf_func_state *cur, struct bpf_idmap *idmap)
15417 /* walk slots of the explored stack and ignore any additional
15418 * slots in the current stack, since explored(safe) state
15421 for (i = 0; i < old->allocated_stack; i++) {
15422 struct bpf_reg_state *old_reg, *cur_reg;
15424 spi = i / BPF_REG_SIZE;
15426 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
15427 i += BPF_REG_SIZE - 1;
15428 /* explored state didn't use this */
15432 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
15435 if (env->allow_uninit_stack &&
15436 old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC)
15439 /* explored stack has more populated slots than current stack
15440 * and these slots were used
15442 if (i >= cur->allocated_stack)
15445 /* if old state was safe with misc data in the stack
15446 * it will be safe with zero-initialized stack.
15447 * The opposite is not true
15449 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
15450 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
15452 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
15453 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
15454 /* Ex: old explored (safe) state has STACK_SPILL in
15455 * this stack slot, but current has STACK_MISC ->
15456 * this verifier states are not equivalent,
15457 * return false to continue verification of this path
15460 if (i % BPF_REG_SIZE != BPF_REG_SIZE - 1)
15462 /* Both old and cur are having same slot_type */
15463 switch (old->stack[spi].slot_type[BPF_REG_SIZE - 1]) {
15465 /* when explored and current stack slot are both storing
15466 * spilled registers, check that stored pointers types
15467 * are the same as well.
15468 * Ex: explored safe path could have stored
15469 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
15470 * but current path has stored:
15471 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
15472 * such verifier states are not equivalent.
15473 * return false to continue verification of this path
15475 if (!regsafe(env, &old->stack[spi].spilled_ptr,
15476 &cur->stack[spi].spilled_ptr, idmap))
15480 old_reg = &old->stack[spi].spilled_ptr;
15481 cur_reg = &cur->stack[spi].spilled_ptr;
15482 if (old_reg->dynptr.type != cur_reg->dynptr.type ||
15483 old_reg->dynptr.first_slot != cur_reg->dynptr.first_slot ||
15484 !check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap))
15488 old_reg = &old->stack[spi].spilled_ptr;
15489 cur_reg = &cur->stack[spi].spilled_ptr;
15490 /* iter.depth is not compared between states as it
15491 * doesn't matter for correctness and would otherwise
15492 * prevent convergence; we maintain it only to prevent
15493 * infinite loop check triggering, see
15494 * iter_active_depths_differ()
15496 if (old_reg->iter.btf != cur_reg->iter.btf ||
15497 old_reg->iter.btf_id != cur_reg->iter.btf_id ||
15498 old_reg->iter.state != cur_reg->iter.state ||
15499 /* ignore {old_reg,cur_reg}->iter.depth, see above */
15500 !check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap))
15505 case STACK_INVALID:
15507 /* Ensure that new unhandled slot types return false by default */
15515 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur,
15516 struct bpf_idmap *idmap)
15520 if (old->acquired_refs != cur->acquired_refs)
15523 for (i = 0; i < old->acquired_refs; i++) {
15524 if (!check_ids(old->refs[i].id, cur->refs[i].id, idmap))
15531 /* compare two verifier states
15533 * all states stored in state_list are known to be valid, since
15534 * verifier reached 'bpf_exit' instruction through them
15536 * this function is called when verifier exploring different branches of
15537 * execution popped from the state stack. If it sees an old state that has
15538 * more strict register state and more strict stack state then this execution
15539 * branch doesn't need to be explored further, since verifier already
15540 * concluded that more strict state leads to valid finish.
15542 * Therefore two states are equivalent if register state is more conservative
15543 * and explored stack state is more conservative than the current one.
15546 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
15547 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
15549 * In other words if current stack state (one being explored) has more
15550 * valid slots than old one that already passed validation, it means
15551 * the verifier can stop exploring and conclude that current state is valid too
15553 * Similarly with registers. If explored state has register type as invalid
15554 * whereas register type in current state is meaningful, it means that
15555 * the current state will reach 'bpf_exit' instruction safely
15557 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
15558 struct bpf_func_state *cur)
15562 for (i = 0; i < MAX_BPF_REG; i++)
15563 if (!regsafe(env, &old->regs[i], &cur->regs[i],
15564 &env->idmap_scratch))
15567 if (!stacksafe(env, old, cur, &env->idmap_scratch))
15570 if (!refsafe(old, cur, &env->idmap_scratch))
15576 static bool states_equal(struct bpf_verifier_env *env,
15577 struct bpf_verifier_state *old,
15578 struct bpf_verifier_state *cur)
15582 if (old->curframe != cur->curframe)
15585 env->idmap_scratch.tmp_id_gen = env->id_gen;
15586 memset(&env->idmap_scratch.map, 0, sizeof(env->idmap_scratch.map));
15588 /* Verification state from speculative execution simulation
15589 * must never prune a non-speculative execution one.
15591 if (old->speculative && !cur->speculative)
15594 if (old->active_lock.ptr != cur->active_lock.ptr)
15597 /* Old and cur active_lock's have to be either both present
15600 if (!!old->active_lock.id != !!cur->active_lock.id)
15603 if (old->active_lock.id &&
15604 !check_ids(old->active_lock.id, cur->active_lock.id, &env->idmap_scratch))
15607 if (old->active_rcu_lock != cur->active_rcu_lock)
15610 /* for states to be equal callsites have to be the same
15611 * and all frame states need to be equivalent
15613 for (i = 0; i <= old->curframe; i++) {
15614 if (old->frame[i]->callsite != cur->frame[i]->callsite)
15616 if (!func_states_equal(env, old->frame[i], cur->frame[i]))
15622 /* Return 0 if no propagation happened. Return negative error code if error
15623 * happened. Otherwise, return the propagated bit.
15625 static int propagate_liveness_reg(struct bpf_verifier_env *env,
15626 struct bpf_reg_state *reg,
15627 struct bpf_reg_state *parent_reg)
15629 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
15630 u8 flag = reg->live & REG_LIVE_READ;
15633 /* When comes here, read flags of PARENT_REG or REG could be any of
15634 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
15635 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
15637 if (parent_flag == REG_LIVE_READ64 ||
15638 /* Or if there is no read flag from REG. */
15640 /* Or if the read flag from REG is the same as PARENT_REG. */
15641 parent_flag == flag)
15644 err = mark_reg_read(env, reg, parent_reg, flag);
15651 /* A write screens off any subsequent reads; but write marks come from the
15652 * straight-line code between a state and its parent. When we arrive at an
15653 * equivalent state (jump target or such) we didn't arrive by the straight-line
15654 * code, so read marks in the state must propagate to the parent regardless
15655 * of the state's write marks. That's what 'parent == state->parent' comparison
15656 * in mark_reg_read() is for.
15658 static int propagate_liveness(struct bpf_verifier_env *env,
15659 const struct bpf_verifier_state *vstate,
15660 struct bpf_verifier_state *vparent)
15662 struct bpf_reg_state *state_reg, *parent_reg;
15663 struct bpf_func_state *state, *parent;
15664 int i, frame, err = 0;
15666 if (vparent->curframe != vstate->curframe) {
15667 WARN(1, "propagate_live: parent frame %d current frame %d\n",
15668 vparent->curframe, vstate->curframe);
15671 /* Propagate read liveness of registers... */
15672 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
15673 for (frame = 0; frame <= vstate->curframe; frame++) {
15674 parent = vparent->frame[frame];
15675 state = vstate->frame[frame];
15676 parent_reg = parent->regs;
15677 state_reg = state->regs;
15678 /* We don't need to worry about FP liveness, it's read-only */
15679 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
15680 err = propagate_liveness_reg(env, &state_reg[i],
15684 if (err == REG_LIVE_READ64)
15685 mark_insn_zext(env, &parent_reg[i]);
15688 /* Propagate stack slots. */
15689 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
15690 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
15691 parent_reg = &parent->stack[i].spilled_ptr;
15692 state_reg = &state->stack[i].spilled_ptr;
15693 err = propagate_liveness_reg(env, state_reg,
15702 /* find precise scalars in the previous equivalent state and
15703 * propagate them into the current state
15705 static int propagate_precision(struct bpf_verifier_env *env,
15706 const struct bpf_verifier_state *old)
15708 struct bpf_reg_state *state_reg;
15709 struct bpf_func_state *state;
15710 int i, err = 0, fr;
15713 for (fr = old->curframe; fr >= 0; fr--) {
15714 state = old->frame[fr];
15715 state_reg = state->regs;
15717 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
15718 if (state_reg->type != SCALAR_VALUE ||
15719 !state_reg->precise ||
15720 !(state_reg->live & REG_LIVE_READ))
15722 if (env->log.level & BPF_LOG_LEVEL2) {
15724 verbose(env, "frame %d: propagating r%d", fr, i);
15726 verbose(env, ",r%d", i);
15728 bt_set_frame_reg(&env->bt, fr, i);
15732 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
15733 if (!is_spilled_reg(&state->stack[i]))
15735 state_reg = &state->stack[i].spilled_ptr;
15736 if (state_reg->type != SCALAR_VALUE ||
15737 !state_reg->precise ||
15738 !(state_reg->live & REG_LIVE_READ))
15740 if (env->log.level & BPF_LOG_LEVEL2) {
15742 verbose(env, "frame %d: propagating fp%d",
15743 fr, (-i - 1) * BPF_REG_SIZE);
15745 verbose(env, ",fp%d", (-i - 1) * BPF_REG_SIZE);
15747 bt_set_frame_slot(&env->bt, fr, i);
15751 verbose(env, "\n");
15754 err = mark_chain_precision_batch(env);
15761 static bool states_maybe_looping(struct bpf_verifier_state *old,
15762 struct bpf_verifier_state *cur)
15764 struct bpf_func_state *fold, *fcur;
15765 int i, fr = cur->curframe;
15767 if (old->curframe != fr)
15770 fold = old->frame[fr];
15771 fcur = cur->frame[fr];
15772 for (i = 0; i < MAX_BPF_REG; i++)
15773 if (memcmp(&fold->regs[i], &fcur->regs[i],
15774 offsetof(struct bpf_reg_state, parent)))
15779 static bool is_iter_next_insn(struct bpf_verifier_env *env, int insn_idx)
15781 return env->insn_aux_data[insn_idx].is_iter_next;
15784 /* is_state_visited() handles iter_next() (see process_iter_next_call() for
15785 * terminology) calls specially: as opposed to bounded BPF loops, it *expects*
15786 * states to match, which otherwise would look like an infinite loop. So while
15787 * iter_next() calls are taken care of, we still need to be careful and
15788 * prevent erroneous and too eager declaration of "ininite loop", when
15789 * iterators are involved.
15791 * Here's a situation in pseudo-BPF assembly form:
15793 * 0: again: ; set up iter_next() call args
15794 * 1: r1 = &it ; <CHECKPOINT HERE>
15795 * 2: call bpf_iter_num_next ; this is iter_next() call
15796 * 3: if r0 == 0 goto done
15797 * 4: ... something useful here ...
15798 * 5: goto again ; another iteration
15801 * 8: call bpf_iter_num_destroy ; clean up iter state
15804 * This is a typical loop. Let's assume that we have a prune point at 1:,
15805 * before we get to `call bpf_iter_num_next` (e.g., because of that `goto
15806 * again`, assuming other heuristics don't get in a way).
15808 * When we first time come to 1:, let's say we have some state X. We proceed
15809 * to 2:, fork states, enqueue ACTIVE, validate NULL case successfully, exit.
15810 * Now we come back to validate that forked ACTIVE state. We proceed through
15811 * 3-5, come to goto, jump to 1:. Let's assume our state didn't change, so we
15812 * are converging. But the problem is that we don't know that yet, as this
15813 * convergence has to happen at iter_next() call site only. So if nothing is
15814 * done, at 1: verifier will use bounded loop logic and declare infinite
15815 * looping (and would be *technically* correct, if not for iterator's
15816 * "eventual sticky NULL" contract, see process_iter_next_call()). But we
15817 * don't want that. So what we do in process_iter_next_call() when we go on
15818 * another ACTIVE iteration, we bump slot->iter.depth, to mark that it's
15819 * a different iteration. So when we suspect an infinite loop, we additionally
15820 * check if any of the *ACTIVE* iterator states depths differ. If yes, we
15821 * pretend we are not looping and wait for next iter_next() call.
15823 * This only applies to ACTIVE state. In DRAINED state we don't expect to
15824 * loop, because that would actually mean infinite loop, as DRAINED state is
15825 * "sticky", and so we'll keep returning into the same instruction with the
15826 * same state (at least in one of possible code paths).
15828 * This approach allows to keep infinite loop heuristic even in the face of
15829 * active iterator. E.g., C snippet below is and will be detected as
15830 * inifintely looping:
15832 * struct bpf_iter_num it;
15835 * bpf_iter_num_new(&it, 0, 10);
15836 * while ((p = bpf_iter_num_next(&t))) {
15838 * while (x--) {} // <<-- infinite loop here
15842 static bool iter_active_depths_differ(struct bpf_verifier_state *old, struct bpf_verifier_state *cur)
15844 struct bpf_reg_state *slot, *cur_slot;
15845 struct bpf_func_state *state;
15848 for (fr = old->curframe; fr >= 0; fr--) {
15849 state = old->frame[fr];
15850 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
15851 if (state->stack[i].slot_type[0] != STACK_ITER)
15854 slot = &state->stack[i].spilled_ptr;
15855 if (slot->iter.state != BPF_ITER_STATE_ACTIVE)
15858 cur_slot = &cur->frame[fr]->stack[i].spilled_ptr;
15859 if (cur_slot->iter.depth != slot->iter.depth)
15866 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
15868 struct bpf_verifier_state_list *new_sl;
15869 struct bpf_verifier_state_list *sl, **pprev;
15870 struct bpf_verifier_state *cur = env->cur_state, *new;
15871 int i, j, err, states_cnt = 0;
15872 bool force_new_state = env->test_state_freq || is_force_checkpoint(env, insn_idx);
15873 bool add_new_state = force_new_state;
15875 /* bpf progs typically have pruning point every 4 instructions
15876 * http://vger.kernel.org/bpfconf2019.html#session-1
15877 * Do not add new state for future pruning if the verifier hasn't seen
15878 * at least 2 jumps and at least 8 instructions.
15879 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
15880 * In tests that amounts to up to 50% reduction into total verifier
15881 * memory consumption and 20% verifier time speedup.
15883 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
15884 env->insn_processed - env->prev_insn_processed >= 8)
15885 add_new_state = true;
15887 pprev = explored_state(env, insn_idx);
15890 clean_live_states(env, insn_idx, cur);
15894 if (sl->state.insn_idx != insn_idx)
15897 if (sl->state.branches) {
15898 struct bpf_func_state *frame = sl->state.frame[sl->state.curframe];
15900 if (frame->in_async_callback_fn &&
15901 frame->async_entry_cnt != cur->frame[cur->curframe]->async_entry_cnt) {
15902 /* Different async_entry_cnt means that the verifier is
15903 * processing another entry into async callback.
15904 * Seeing the same state is not an indication of infinite
15905 * loop or infinite recursion.
15906 * But finding the same state doesn't mean that it's safe
15907 * to stop processing the current state. The previous state
15908 * hasn't yet reached bpf_exit, since state.branches > 0.
15909 * Checking in_async_callback_fn alone is not enough either.
15910 * Since the verifier still needs to catch infinite loops
15911 * inside async callbacks.
15913 goto skip_inf_loop_check;
15915 /* BPF open-coded iterators loop detection is special.
15916 * states_maybe_looping() logic is too simplistic in detecting
15917 * states that *might* be equivalent, because it doesn't know
15918 * about ID remapping, so don't even perform it.
15919 * See process_iter_next_call() and iter_active_depths_differ()
15920 * for overview of the logic. When current and one of parent
15921 * states are detected as equivalent, it's a good thing: we prove
15922 * convergence and can stop simulating further iterations.
15923 * It's safe to assume that iterator loop will finish, taking into
15924 * account iter_next() contract of eventually returning
15925 * sticky NULL result.
15927 if (is_iter_next_insn(env, insn_idx)) {
15928 if (states_equal(env, &sl->state, cur)) {
15929 struct bpf_func_state *cur_frame;
15930 struct bpf_reg_state *iter_state, *iter_reg;
15933 cur_frame = cur->frame[cur->curframe];
15934 /* btf_check_iter_kfuncs() enforces that
15935 * iter state pointer is always the first arg
15937 iter_reg = &cur_frame->regs[BPF_REG_1];
15938 /* current state is valid due to states_equal(),
15939 * so we can assume valid iter and reg state,
15940 * no need for extra (re-)validations
15942 spi = __get_spi(iter_reg->off + iter_reg->var_off.value);
15943 iter_state = &func(env, iter_reg)->stack[spi].spilled_ptr;
15944 if (iter_state->iter.state == BPF_ITER_STATE_ACTIVE)
15947 goto skip_inf_loop_check;
15949 /* attempt to detect infinite loop to avoid unnecessary doomed work */
15950 if (states_maybe_looping(&sl->state, cur) &&
15951 states_equal(env, &sl->state, cur) &&
15952 !iter_active_depths_differ(&sl->state, cur)) {
15953 verbose_linfo(env, insn_idx, "; ");
15954 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
15957 /* if the verifier is processing a loop, avoid adding new state
15958 * too often, since different loop iterations have distinct
15959 * states and may not help future pruning.
15960 * This threshold shouldn't be too low to make sure that
15961 * a loop with large bound will be rejected quickly.
15962 * The most abusive loop will be:
15964 * if r1 < 1000000 goto pc-2
15965 * 1M insn_procssed limit / 100 == 10k peak states.
15966 * This threshold shouldn't be too high either, since states
15967 * at the end of the loop are likely to be useful in pruning.
15969 skip_inf_loop_check:
15970 if (!force_new_state &&
15971 env->jmps_processed - env->prev_jmps_processed < 20 &&
15972 env->insn_processed - env->prev_insn_processed < 100)
15973 add_new_state = false;
15976 if (states_equal(env, &sl->state, cur)) {
15979 /* reached equivalent register/stack state,
15980 * prune the search.
15981 * Registers read by the continuation are read by us.
15982 * If we have any write marks in env->cur_state, they
15983 * will prevent corresponding reads in the continuation
15984 * from reaching our parent (an explored_state). Our
15985 * own state will get the read marks recorded, but
15986 * they'll be immediately forgotten as we're pruning
15987 * this state and will pop a new one.
15989 err = propagate_liveness(env, &sl->state, cur);
15991 /* if previous state reached the exit with precision and
15992 * current state is equivalent to it (except precsion marks)
15993 * the precision needs to be propagated back in
15994 * the current state.
15996 err = err ? : push_jmp_history(env, cur);
15997 err = err ? : propagate_precision(env, &sl->state);
16003 /* when new state is not going to be added do not increase miss count.
16004 * Otherwise several loop iterations will remove the state
16005 * recorded earlier. The goal of these heuristics is to have
16006 * states from some iterations of the loop (some in the beginning
16007 * and some at the end) to help pruning.
16011 /* heuristic to determine whether this state is beneficial
16012 * to keep checking from state equivalence point of view.
16013 * Higher numbers increase max_states_per_insn and verification time,
16014 * but do not meaningfully decrease insn_processed.
16016 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
16017 /* the state is unlikely to be useful. Remove it to
16018 * speed up verification
16021 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
16022 u32 br = sl->state.branches;
16025 "BUG live_done but branches_to_explore %d\n",
16027 free_verifier_state(&sl->state, false);
16029 env->peak_states--;
16031 /* cannot free this state, since parentage chain may
16032 * walk it later. Add it for free_list instead to
16033 * be freed at the end of verification
16035 sl->next = env->free_list;
16036 env->free_list = sl;
16046 if (env->max_states_per_insn < states_cnt)
16047 env->max_states_per_insn = states_cnt;
16049 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
16052 if (!add_new_state)
16055 /* There were no equivalent states, remember the current one.
16056 * Technically the current state is not proven to be safe yet,
16057 * but it will either reach outer most bpf_exit (which means it's safe)
16058 * or it will be rejected. When there are no loops the verifier won't be
16059 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
16060 * again on the way to bpf_exit.
16061 * When looping the sl->state.branches will be > 0 and this state
16062 * will not be considered for equivalence until branches == 0.
16064 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
16067 env->total_states++;
16068 env->peak_states++;
16069 env->prev_jmps_processed = env->jmps_processed;
16070 env->prev_insn_processed = env->insn_processed;
16072 /* forget precise markings we inherited, see __mark_chain_precision */
16073 if (env->bpf_capable)
16074 mark_all_scalars_imprecise(env, cur);
16076 /* add new state to the head of linked list */
16077 new = &new_sl->state;
16078 err = copy_verifier_state(new, cur);
16080 free_verifier_state(new, false);
16084 new->insn_idx = insn_idx;
16085 WARN_ONCE(new->branches != 1,
16086 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
16089 cur->first_insn_idx = insn_idx;
16090 clear_jmp_history(cur);
16091 new_sl->next = *explored_state(env, insn_idx);
16092 *explored_state(env, insn_idx) = new_sl;
16093 /* connect new state to parentage chain. Current frame needs all
16094 * registers connected. Only r6 - r9 of the callers are alive (pushed
16095 * to the stack implicitly by JITs) so in callers' frames connect just
16096 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
16097 * the state of the call instruction (with WRITTEN set), and r0 comes
16098 * from callee with its full parentage chain, anyway.
16100 /* clear write marks in current state: the writes we did are not writes
16101 * our child did, so they don't screen off its reads from us.
16102 * (There are no read marks in current state, because reads always mark
16103 * their parent and current state never has children yet. Only
16104 * explored_states can get read marks.)
16106 for (j = 0; j <= cur->curframe; j++) {
16107 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
16108 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
16109 for (i = 0; i < BPF_REG_FP; i++)
16110 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
16113 /* all stack frames are accessible from callee, clear them all */
16114 for (j = 0; j <= cur->curframe; j++) {
16115 struct bpf_func_state *frame = cur->frame[j];
16116 struct bpf_func_state *newframe = new->frame[j];
16118 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
16119 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
16120 frame->stack[i].spilled_ptr.parent =
16121 &newframe->stack[i].spilled_ptr;
16127 /* Return true if it's OK to have the same insn return a different type. */
16128 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
16130 switch (base_type(type)) {
16132 case PTR_TO_SOCKET:
16133 case PTR_TO_SOCK_COMMON:
16134 case PTR_TO_TCP_SOCK:
16135 case PTR_TO_XDP_SOCK:
16136 case PTR_TO_BTF_ID:
16143 /* If an instruction was previously used with particular pointer types, then we
16144 * need to be careful to avoid cases such as the below, where it may be ok
16145 * for one branch accessing the pointer, but not ok for the other branch:
16150 * R1 = some_other_valid_ptr;
16153 * R2 = *(u32 *)(R1 + 0);
16155 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
16157 return src != prev && (!reg_type_mismatch_ok(src) ||
16158 !reg_type_mismatch_ok(prev));
16161 static int save_aux_ptr_type(struct bpf_verifier_env *env, enum bpf_reg_type type,
16162 bool allow_trust_missmatch)
16164 enum bpf_reg_type *prev_type = &env->insn_aux_data[env->insn_idx].ptr_type;
16166 if (*prev_type == NOT_INIT) {
16167 /* Saw a valid insn
16168 * dst_reg = *(u32 *)(src_reg + off)
16169 * save type to validate intersecting paths
16172 } else if (reg_type_mismatch(type, *prev_type)) {
16173 /* Abuser program is trying to use the same insn
16174 * dst_reg = *(u32*) (src_reg + off)
16175 * with different pointer types:
16176 * src_reg == ctx in one branch and
16177 * src_reg == stack|map in some other branch.
16180 if (allow_trust_missmatch &&
16181 base_type(type) == PTR_TO_BTF_ID &&
16182 base_type(*prev_type) == PTR_TO_BTF_ID) {
16184 * Have to support a use case when one path through
16185 * the program yields TRUSTED pointer while another
16186 * is UNTRUSTED. Fallback to UNTRUSTED to generate
16189 *prev_type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
16191 verbose(env, "same insn cannot be used with different pointers\n");
16199 static int do_check(struct bpf_verifier_env *env)
16201 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
16202 struct bpf_verifier_state *state = env->cur_state;
16203 struct bpf_insn *insns = env->prog->insnsi;
16204 struct bpf_reg_state *regs;
16205 int insn_cnt = env->prog->len;
16206 bool do_print_state = false;
16207 int prev_insn_idx = -1;
16210 struct bpf_insn *insn;
16214 env->prev_insn_idx = prev_insn_idx;
16215 if (env->insn_idx >= insn_cnt) {
16216 verbose(env, "invalid insn idx %d insn_cnt %d\n",
16217 env->insn_idx, insn_cnt);
16221 insn = &insns[env->insn_idx];
16222 class = BPF_CLASS(insn->code);
16224 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
16226 "BPF program is too large. Processed %d insn\n",
16227 env->insn_processed);
16231 state->last_insn_idx = env->prev_insn_idx;
16233 if (is_prune_point(env, env->insn_idx)) {
16234 err = is_state_visited(env, env->insn_idx);
16238 /* found equivalent state, can prune the search */
16239 if (env->log.level & BPF_LOG_LEVEL) {
16240 if (do_print_state)
16241 verbose(env, "\nfrom %d to %d%s: safe\n",
16242 env->prev_insn_idx, env->insn_idx,
16243 env->cur_state->speculative ?
16244 " (speculative execution)" : "");
16246 verbose(env, "%d: safe\n", env->insn_idx);
16248 goto process_bpf_exit;
16252 if (is_jmp_point(env, env->insn_idx)) {
16253 err = push_jmp_history(env, state);
16258 if (signal_pending(current))
16261 if (need_resched())
16264 if (env->log.level & BPF_LOG_LEVEL2 && do_print_state) {
16265 verbose(env, "\nfrom %d to %d%s:",
16266 env->prev_insn_idx, env->insn_idx,
16267 env->cur_state->speculative ?
16268 " (speculative execution)" : "");
16269 print_verifier_state(env, state->frame[state->curframe], true);
16270 do_print_state = false;
16273 if (env->log.level & BPF_LOG_LEVEL) {
16274 const struct bpf_insn_cbs cbs = {
16275 .cb_call = disasm_kfunc_name,
16276 .cb_print = verbose,
16277 .private_data = env,
16280 if (verifier_state_scratched(env))
16281 print_insn_state(env, state->frame[state->curframe]);
16283 verbose_linfo(env, env->insn_idx, "; ");
16284 env->prev_log_pos = env->log.end_pos;
16285 verbose(env, "%d: ", env->insn_idx);
16286 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
16287 env->prev_insn_print_pos = env->log.end_pos - env->prev_log_pos;
16288 env->prev_log_pos = env->log.end_pos;
16291 if (bpf_prog_is_offloaded(env->prog->aux)) {
16292 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
16293 env->prev_insn_idx);
16298 regs = cur_regs(env);
16299 sanitize_mark_insn_seen(env);
16300 prev_insn_idx = env->insn_idx;
16302 if (class == BPF_ALU || class == BPF_ALU64) {
16303 err = check_alu_op(env, insn);
16307 } else if (class == BPF_LDX) {
16308 enum bpf_reg_type src_reg_type;
16310 /* check for reserved fields is already done */
16312 /* check src operand */
16313 err = check_reg_arg(env, insn->src_reg, SRC_OP);
16317 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
16321 src_reg_type = regs[insn->src_reg].type;
16323 /* check that memory (src_reg + off) is readable,
16324 * the state of dst_reg will be updated by this func
16326 err = check_mem_access(env, env->insn_idx, insn->src_reg,
16327 insn->off, BPF_SIZE(insn->code),
16328 BPF_READ, insn->dst_reg, false);
16332 err = save_aux_ptr_type(env, src_reg_type, true);
16335 } else if (class == BPF_STX) {
16336 enum bpf_reg_type dst_reg_type;
16338 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
16339 err = check_atomic(env, env->insn_idx, insn);
16346 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
16347 verbose(env, "BPF_STX uses reserved fields\n");
16351 /* check src1 operand */
16352 err = check_reg_arg(env, insn->src_reg, SRC_OP);
16355 /* check src2 operand */
16356 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
16360 dst_reg_type = regs[insn->dst_reg].type;
16362 /* check that memory (dst_reg + off) is writeable */
16363 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
16364 insn->off, BPF_SIZE(insn->code),
16365 BPF_WRITE, insn->src_reg, false);
16369 err = save_aux_ptr_type(env, dst_reg_type, false);
16372 } else if (class == BPF_ST) {
16373 enum bpf_reg_type dst_reg_type;
16375 if (BPF_MODE(insn->code) != BPF_MEM ||
16376 insn->src_reg != BPF_REG_0) {
16377 verbose(env, "BPF_ST uses reserved fields\n");
16380 /* check src operand */
16381 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
16385 dst_reg_type = regs[insn->dst_reg].type;
16387 /* check that memory (dst_reg + off) is writeable */
16388 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
16389 insn->off, BPF_SIZE(insn->code),
16390 BPF_WRITE, -1, false);
16394 err = save_aux_ptr_type(env, dst_reg_type, false);
16397 } else if (class == BPF_JMP || class == BPF_JMP32) {
16398 u8 opcode = BPF_OP(insn->code);
16400 env->jmps_processed++;
16401 if (opcode == BPF_CALL) {
16402 if (BPF_SRC(insn->code) != BPF_K ||
16403 (insn->src_reg != BPF_PSEUDO_KFUNC_CALL
16404 && insn->off != 0) ||
16405 (insn->src_reg != BPF_REG_0 &&
16406 insn->src_reg != BPF_PSEUDO_CALL &&
16407 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
16408 insn->dst_reg != BPF_REG_0 ||
16409 class == BPF_JMP32) {
16410 verbose(env, "BPF_CALL uses reserved fields\n");
16414 if (env->cur_state->active_lock.ptr) {
16415 if ((insn->src_reg == BPF_REG_0 && insn->imm != BPF_FUNC_spin_unlock) ||
16416 (insn->src_reg == BPF_PSEUDO_CALL) ||
16417 (insn->src_reg == BPF_PSEUDO_KFUNC_CALL &&
16418 (insn->off != 0 || !is_bpf_graph_api_kfunc(insn->imm)))) {
16419 verbose(env, "function calls are not allowed while holding a lock\n");
16423 if (insn->src_reg == BPF_PSEUDO_CALL)
16424 err = check_func_call(env, insn, &env->insn_idx);
16425 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
16426 err = check_kfunc_call(env, insn, &env->insn_idx);
16428 err = check_helper_call(env, insn, &env->insn_idx);
16432 mark_reg_scratched(env, BPF_REG_0);
16433 } else if (opcode == BPF_JA) {
16434 if (BPF_SRC(insn->code) != BPF_K ||
16436 insn->src_reg != BPF_REG_0 ||
16437 insn->dst_reg != BPF_REG_0 ||
16438 class == BPF_JMP32) {
16439 verbose(env, "BPF_JA uses reserved fields\n");
16443 env->insn_idx += insn->off + 1;
16446 } else if (opcode == BPF_EXIT) {
16447 if (BPF_SRC(insn->code) != BPF_K ||
16449 insn->src_reg != BPF_REG_0 ||
16450 insn->dst_reg != BPF_REG_0 ||
16451 class == BPF_JMP32) {
16452 verbose(env, "BPF_EXIT uses reserved fields\n");
16456 if (env->cur_state->active_lock.ptr &&
16457 !in_rbtree_lock_required_cb(env)) {
16458 verbose(env, "bpf_spin_unlock is missing\n");
16462 if (env->cur_state->active_rcu_lock) {
16463 verbose(env, "bpf_rcu_read_unlock is missing\n");
16467 /* We must do check_reference_leak here before
16468 * prepare_func_exit to handle the case when
16469 * state->curframe > 0, it may be a callback
16470 * function, for which reference_state must
16471 * match caller reference state when it exits.
16473 err = check_reference_leak(env);
16477 if (state->curframe) {
16478 /* exit from nested function */
16479 err = prepare_func_exit(env, &env->insn_idx);
16482 do_print_state = true;
16486 err = check_return_code(env);
16490 mark_verifier_state_scratched(env);
16491 update_branch_counts(env, env->cur_state);
16492 err = pop_stack(env, &prev_insn_idx,
16493 &env->insn_idx, pop_log);
16495 if (err != -ENOENT)
16499 do_print_state = true;
16503 err = check_cond_jmp_op(env, insn, &env->insn_idx);
16507 } else if (class == BPF_LD) {
16508 u8 mode = BPF_MODE(insn->code);
16510 if (mode == BPF_ABS || mode == BPF_IND) {
16511 err = check_ld_abs(env, insn);
16515 } else if (mode == BPF_IMM) {
16516 err = check_ld_imm(env, insn);
16521 sanitize_mark_insn_seen(env);
16523 verbose(env, "invalid BPF_LD mode\n");
16527 verbose(env, "unknown insn class %d\n", class);
16537 static int find_btf_percpu_datasec(struct btf *btf)
16539 const struct btf_type *t;
16544 * Both vmlinux and module each have their own ".data..percpu"
16545 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
16546 * types to look at only module's own BTF types.
16548 n = btf_nr_types(btf);
16549 if (btf_is_module(btf))
16550 i = btf_nr_types(btf_vmlinux);
16554 for(; i < n; i++) {
16555 t = btf_type_by_id(btf, i);
16556 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
16559 tname = btf_name_by_offset(btf, t->name_off);
16560 if (!strcmp(tname, ".data..percpu"))
16567 /* replace pseudo btf_id with kernel symbol address */
16568 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
16569 struct bpf_insn *insn,
16570 struct bpf_insn_aux_data *aux)
16572 const struct btf_var_secinfo *vsi;
16573 const struct btf_type *datasec;
16574 struct btf_mod_pair *btf_mod;
16575 const struct btf_type *t;
16576 const char *sym_name;
16577 bool percpu = false;
16578 u32 type, id = insn->imm;
16582 int i, btf_fd, err;
16584 btf_fd = insn[1].imm;
16586 btf = btf_get_by_fd(btf_fd);
16588 verbose(env, "invalid module BTF object FD specified.\n");
16592 if (!btf_vmlinux) {
16593 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
16600 t = btf_type_by_id(btf, id);
16602 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
16607 if (!btf_type_is_var(t) && !btf_type_is_func(t)) {
16608 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR or KIND_FUNC\n", id);
16613 sym_name = btf_name_by_offset(btf, t->name_off);
16614 addr = kallsyms_lookup_name(sym_name);
16616 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
16621 insn[0].imm = (u32)addr;
16622 insn[1].imm = addr >> 32;
16624 if (btf_type_is_func(t)) {
16625 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
16626 aux->btf_var.mem_size = 0;
16630 datasec_id = find_btf_percpu_datasec(btf);
16631 if (datasec_id > 0) {
16632 datasec = btf_type_by_id(btf, datasec_id);
16633 for_each_vsi(i, datasec, vsi) {
16634 if (vsi->type == id) {
16642 t = btf_type_skip_modifiers(btf, type, NULL);
16644 aux->btf_var.reg_type = PTR_TO_BTF_ID | MEM_PERCPU;
16645 aux->btf_var.btf = btf;
16646 aux->btf_var.btf_id = type;
16647 } else if (!btf_type_is_struct(t)) {
16648 const struct btf_type *ret;
16652 /* resolve the type size of ksym. */
16653 ret = btf_resolve_size(btf, t, &tsize);
16655 tname = btf_name_by_offset(btf, t->name_off);
16656 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
16657 tname, PTR_ERR(ret));
16661 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
16662 aux->btf_var.mem_size = tsize;
16664 aux->btf_var.reg_type = PTR_TO_BTF_ID;
16665 aux->btf_var.btf = btf;
16666 aux->btf_var.btf_id = type;
16669 /* check whether we recorded this BTF (and maybe module) already */
16670 for (i = 0; i < env->used_btf_cnt; i++) {
16671 if (env->used_btfs[i].btf == btf) {
16677 if (env->used_btf_cnt >= MAX_USED_BTFS) {
16682 btf_mod = &env->used_btfs[env->used_btf_cnt];
16683 btf_mod->btf = btf;
16684 btf_mod->module = NULL;
16686 /* if we reference variables from kernel module, bump its refcount */
16687 if (btf_is_module(btf)) {
16688 btf_mod->module = btf_try_get_module(btf);
16689 if (!btf_mod->module) {
16695 env->used_btf_cnt++;
16703 static bool is_tracing_prog_type(enum bpf_prog_type type)
16706 case BPF_PROG_TYPE_KPROBE:
16707 case BPF_PROG_TYPE_TRACEPOINT:
16708 case BPF_PROG_TYPE_PERF_EVENT:
16709 case BPF_PROG_TYPE_RAW_TRACEPOINT:
16710 case BPF_PROG_TYPE_RAW_TRACEPOINT_WRITABLE:
16717 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
16718 struct bpf_map *map,
16719 struct bpf_prog *prog)
16722 enum bpf_prog_type prog_type = resolve_prog_type(prog);
16724 if (btf_record_has_field(map->record, BPF_LIST_HEAD) ||
16725 btf_record_has_field(map->record, BPF_RB_ROOT)) {
16726 if (is_tracing_prog_type(prog_type)) {
16727 verbose(env, "tracing progs cannot use bpf_{list_head,rb_root} yet\n");
16732 if (btf_record_has_field(map->record, BPF_SPIN_LOCK)) {
16733 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
16734 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
16738 if (is_tracing_prog_type(prog_type)) {
16739 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
16743 if (prog->aux->sleepable) {
16744 verbose(env, "sleepable progs cannot use bpf_spin_lock yet\n");
16749 if (btf_record_has_field(map->record, BPF_TIMER)) {
16750 if (is_tracing_prog_type(prog_type)) {
16751 verbose(env, "tracing progs cannot use bpf_timer yet\n");
16756 if ((bpf_prog_is_offloaded(prog->aux) || bpf_map_is_offloaded(map)) &&
16757 !bpf_offload_prog_map_match(prog, map)) {
16758 verbose(env, "offload device mismatch between prog and map\n");
16762 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
16763 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
16767 if (prog->aux->sleepable)
16768 switch (map->map_type) {
16769 case BPF_MAP_TYPE_HASH:
16770 case BPF_MAP_TYPE_LRU_HASH:
16771 case BPF_MAP_TYPE_ARRAY:
16772 case BPF_MAP_TYPE_PERCPU_HASH:
16773 case BPF_MAP_TYPE_PERCPU_ARRAY:
16774 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
16775 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
16776 case BPF_MAP_TYPE_HASH_OF_MAPS:
16777 case BPF_MAP_TYPE_RINGBUF:
16778 case BPF_MAP_TYPE_USER_RINGBUF:
16779 case BPF_MAP_TYPE_INODE_STORAGE:
16780 case BPF_MAP_TYPE_SK_STORAGE:
16781 case BPF_MAP_TYPE_TASK_STORAGE:
16782 case BPF_MAP_TYPE_CGRP_STORAGE:
16786 "Sleepable programs can only use array, hash, ringbuf and local storage maps\n");
16793 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
16795 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
16796 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
16799 /* find and rewrite pseudo imm in ld_imm64 instructions:
16801 * 1. if it accesses map FD, replace it with actual map pointer.
16802 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
16804 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
16806 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
16808 struct bpf_insn *insn = env->prog->insnsi;
16809 int insn_cnt = env->prog->len;
16812 err = bpf_prog_calc_tag(env->prog);
16816 for (i = 0; i < insn_cnt; i++, insn++) {
16817 if (BPF_CLASS(insn->code) == BPF_LDX &&
16818 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
16819 verbose(env, "BPF_LDX uses reserved fields\n");
16823 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
16824 struct bpf_insn_aux_data *aux;
16825 struct bpf_map *map;
16830 if (i == insn_cnt - 1 || insn[1].code != 0 ||
16831 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
16832 insn[1].off != 0) {
16833 verbose(env, "invalid bpf_ld_imm64 insn\n");
16837 if (insn[0].src_reg == 0)
16838 /* valid generic load 64-bit imm */
16841 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
16842 aux = &env->insn_aux_data[i];
16843 err = check_pseudo_btf_id(env, insn, aux);
16849 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
16850 aux = &env->insn_aux_data[i];
16851 aux->ptr_type = PTR_TO_FUNC;
16855 /* In final convert_pseudo_ld_imm64() step, this is
16856 * converted into regular 64-bit imm load insn.
16858 switch (insn[0].src_reg) {
16859 case BPF_PSEUDO_MAP_VALUE:
16860 case BPF_PSEUDO_MAP_IDX_VALUE:
16862 case BPF_PSEUDO_MAP_FD:
16863 case BPF_PSEUDO_MAP_IDX:
16864 if (insn[1].imm == 0)
16868 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
16872 switch (insn[0].src_reg) {
16873 case BPF_PSEUDO_MAP_IDX_VALUE:
16874 case BPF_PSEUDO_MAP_IDX:
16875 if (bpfptr_is_null(env->fd_array)) {
16876 verbose(env, "fd_idx without fd_array is invalid\n");
16879 if (copy_from_bpfptr_offset(&fd, env->fd_array,
16880 insn[0].imm * sizeof(fd),
16890 map = __bpf_map_get(f);
16892 verbose(env, "fd %d is not pointing to valid bpf_map\n",
16894 return PTR_ERR(map);
16897 err = check_map_prog_compatibility(env, map, env->prog);
16903 aux = &env->insn_aux_data[i];
16904 if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
16905 insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
16906 addr = (unsigned long)map;
16908 u32 off = insn[1].imm;
16910 if (off >= BPF_MAX_VAR_OFF) {
16911 verbose(env, "direct value offset of %u is not allowed\n", off);
16916 if (!map->ops->map_direct_value_addr) {
16917 verbose(env, "no direct value access support for this map type\n");
16922 err = map->ops->map_direct_value_addr(map, &addr, off);
16924 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
16925 map->value_size, off);
16930 aux->map_off = off;
16934 insn[0].imm = (u32)addr;
16935 insn[1].imm = addr >> 32;
16937 /* check whether we recorded this map already */
16938 for (j = 0; j < env->used_map_cnt; j++) {
16939 if (env->used_maps[j] == map) {
16940 aux->map_index = j;
16946 if (env->used_map_cnt >= MAX_USED_MAPS) {
16951 /* hold the map. If the program is rejected by verifier,
16952 * the map will be released by release_maps() or it
16953 * will be used by the valid program until it's unloaded
16954 * and all maps are released in free_used_maps()
16958 aux->map_index = env->used_map_cnt;
16959 env->used_maps[env->used_map_cnt++] = map;
16961 if (bpf_map_is_cgroup_storage(map) &&
16962 bpf_cgroup_storage_assign(env->prog->aux, map)) {
16963 verbose(env, "only one cgroup storage of each type is allowed\n");
16975 /* Basic sanity check before we invest more work here. */
16976 if (!bpf_opcode_in_insntable(insn->code)) {
16977 verbose(env, "unknown opcode %02x\n", insn->code);
16982 /* now all pseudo BPF_LD_IMM64 instructions load valid
16983 * 'struct bpf_map *' into a register instead of user map_fd.
16984 * These pointers will be used later by verifier to validate map access.
16989 /* drop refcnt of maps used by the rejected program */
16990 static void release_maps(struct bpf_verifier_env *env)
16992 __bpf_free_used_maps(env->prog->aux, env->used_maps,
16993 env->used_map_cnt);
16996 /* drop refcnt of maps used by the rejected program */
16997 static void release_btfs(struct bpf_verifier_env *env)
16999 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
17000 env->used_btf_cnt);
17003 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
17004 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
17006 struct bpf_insn *insn = env->prog->insnsi;
17007 int insn_cnt = env->prog->len;
17010 for (i = 0; i < insn_cnt; i++, insn++) {
17011 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
17013 if (insn->src_reg == BPF_PSEUDO_FUNC)
17019 /* single env->prog->insni[off] instruction was replaced with the range
17020 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
17021 * [0, off) and [off, end) to new locations, so the patched range stays zero
17023 static void adjust_insn_aux_data(struct bpf_verifier_env *env,
17024 struct bpf_insn_aux_data *new_data,
17025 struct bpf_prog *new_prog, u32 off, u32 cnt)
17027 struct bpf_insn_aux_data *old_data = env->insn_aux_data;
17028 struct bpf_insn *insn = new_prog->insnsi;
17029 u32 old_seen = old_data[off].seen;
17033 /* aux info at OFF always needs adjustment, no matter fast path
17034 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
17035 * original insn at old prog.
17037 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
17041 prog_len = new_prog->len;
17043 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
17044 memcpy(new_data + off + cnt - 1, old_data + off,
17045 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
17046 for (i = off; i < off + cnt - 1; i++) {
17047 /* Expand insni[off]'s seen count to the patched range. */
17048 new_data[i].seen = old_seen;
17049 new_data[i].zext_dst = insn_has_def32(env, insn + i);
17051 env->insn_aux_data = new_data;
17055 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
17061 /* NOTE: fake 'exit' subprog should be updated as well. */
17062 for (i = 0; i <= env->subprog_cnt; i++) {
17063 if (env->subprog_info[i].start <= off)
17065 env->subprog_info[i].start += len - 1;
17069 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
17071 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
17072 int i, sz = prog->aux->size_poke_tab;
17073 struct bpf_jit_poke_descriptor *desc;
17075 for (i = 0; i < sz; i++) {
17077 if (desc->insn_idx <= off)
17079 desc->insn_idx += len - 1;
17083 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
17084 const struct bpf_insn *patch, u32 len)
17086 struct bpf_prog *new_prog;
17087 struct bpf_insn_aux_data *new_data = NULL;
17090 new_data = vzalloc(array_size(env->prog->len + len - 1,
17091 sizeof(struct bpf_insn_aux_data)));
17096 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
17097 if (IS_ERR(new_prog)) {
17098 if (PTR_ERR(new_prog) == -ERANGE)
17100 "insn %d cannot be patched due to 16-bit range\n",
17101 env->insn_aux_data[off].orig_idx);
17105 adjust_insn_aux_data(env, new_data, new_prog, off, len);
17106 adjust_subprog_starts(env, off, len);
17107 adjust_poke_descs(new_prog, off, len);
17111 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
17116 /* find first prog starting at or after off (first to remove) */
17117 for (i = 0; i < env->subprog_cnt; i++)
17118 if (env->subprog_info[i].start >= off)
17120 /* find first prog starting at or after off + cnt (first to stay) */
17121 for (j = i; j < env->subprog_cnt; j++)
17122 if (env->subprog_info[j].start >= off + cnt)
17124 /* if j doesn't start exactly at off + cnt, we are just removing
17125 * the front of previous prog
17127 if (env->subprog_info[j].start != off + cnt)
17131 struct bpf_prog_aux *aux = env->prog->aux;
17134 /* move fake 'exit' subprog as well */
17135 move = env->subprog_cnt + 1 - j;
17137 memmove(env->subprog_info + i,
17138 env->subprog_info + j,
17139 sizeof(*env->subprog_info) * move);
17140 env->subprog_cnt -= j - i;
17142 /* remove func_info */
17143 if (aux->func_info) {
17144 move = aux->func_info_cnt - j;
17146 memmove(aux->func_info + i,
17147 aux->func_info + j,
17148 sizeof(*aux->func_info) * move);
17149 aux->func_info_cnt -= j - i;
17150 /* func_info->insn_off is set after all code rewrites,
17151 * in adjust_btf_func() - no need to adjust
17155 /* convert i from "first prog to remove" to "first to adjust" */
17156 if (env->subprog_info[i].start == off)
17160 /* update fake 'exit' subprog as well */
17161 for (; i <= env->subprog_cnt; i++)
17162 env->subprog_info[i].start -= cnt;
17167 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
17170 struct bpf_prog *prog = env->prog;
17171 u32 i, l_off, l_cnt, nr_linfo;
17172 struct bpf_line_info *linfo;
17174 nr_linfo = prog->aux->nr_linfo;
17178 linfo = prog->aux->linfo;
17180 /* find first line info to remove, count lines to be removed */
17181 for (i = 0; i < nr_linfo; i++)
17182 if (linfo[i].insn_off >= off)
17187 for (; i < nr_linfo; i++)
17188 if (linfo[i].insn_off < off + cnt)
17193 /* First live insn doesn't match first live linfo, it needs to "inherit"
17194 * last removed linfo. prog is already modified, so prog->len == off
17195 * means no live instructions after (tail of the program was removed).
17197 if (prog->len != off && l_cnt &&
17198 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
17200 linfo[--i].insn_off = off + cnt;
17203 /* remove the line info which refer to the removed instructions */
17205 memmove(linfo + l_off, linfo + i,
17206 sizeof(*linfo) * (nr_linfo - i));
17208 prog->aux->nr_linfo -= l_cnt;
17209 nr_linfo = prog->aux->nr_linfo;
17212 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
17213 for (i = l_off; i < nr_linfo; i++)
17214 linfo[i].insn_off -= cnt;
17216 /* fix up all subprogs (incl. 'exit') which start >= off */
17217 for (i = 0; i <= env->subprog_cnt; i++)
17218 if (env->subprog_info[i].linfo_idx > l_off) {
17219 /* program may have started in the removed region but
17220 * may not be fully removed
17222 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
17223 env->subprog_info[i].linfo_idx -= l_cnt;
17225 env->subprog_info[i].linfo_idx = l_off;
17231 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
17233 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17234 unsigned int orig_prog_len = env->prog->len;
17237 if (bpf_prog_is_offloaded(env->prog->aux))
17238 bpf_prog_offload_remove_insns(env, off, cnt);
17240 err = bpf_remove_insns(env->prog, off, cnt);
17244 err = adjust_subprog_starts_after_remove(env, off, cnt);
17248 err = bpf_adj_linfo_after_remove(env, off, cnt);
17252 memmove(aux_data + off, aux_data + off + cnt,
17253 sizeof(*aux_data) * (orig_prog_len - off - cnt));
17258 /* The verifier does more data flow analysis than llvm and will not
17259 * explore branches that are dead at run time. Malicious programs can
17260 * have dead code too. Therefore replace all dead at-run-time code
17263 * Just nops are not optimal, e.g. if they would sit at the end of the
17264 * program and through another bug we would manage to jump there, then
17265 * we'd execute beyond program memory otherwise. Returning exception
17266 * code also wouldn't work since we can have subprogs where the dead
17267 * code could be located.
17269 static void sanitize_dead_code(struct bpf_verifier_env *env)
17271 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17272 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
17273 struct bpf_insn *insn = env->prog->insnsi;
17274 const int insn_cnt = env->prog->len;
17277 for (i = 0; i < insn_cnt; i++) {
17278 if (aux_data[i].seen)
17280 memcpy(insn + i, &trap, sizeof(trap));
17281 aux_data[i].zext_dst = false;
17285 static bool insn_is_cond_jump(u8 code)
17289 if (BPF_CLASS(code) == BPF_JMP32)
17292 if (BPF_CLASS(code) != BPF_JMP)
17296 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
17299 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
17301 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17302 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
17303 struct bpf_insn *insn = env->prog->insnsi;
17304 const int insn_cnt = env->prog->len;
17307 for (i = 0; i < insn_cnt; i++, insn++) {
17308 if (!insn_is_cond_jump(insn->code))
17311 if (!aux_data[i + 1].seen)
17312 ja.off = insn->off;
17313 else if (!aux_data[i + 1 + insn->off].seen)
17318 if (bpf_prog_is_offloaded(env->prog->aux))
17319 bpf_prog_offload_replace_insn(env, i, &ja);
17321 memcpy(insn, &ja, sizeof(ja));
17325 static int opt_remove_dead_code(struct bpf_verifier_env *env)
17327 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
17328 int insn_cnt = env->prog->len;
17331 for (i = 0; i < insn_cnt; i++) {
17335 while (i + j < insn_cnt && !aux_data[i + j].seen)
17340 err = verifier_remove_insns(env, i, j);
17343 insn_cnt = env->prog->len;
17349 static int opt_remove_nops(struct bpf_verifier_env *env)
17351 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
17352 struct bpf_insn *insn = env->prog->insnsi;
17353 int insn_cnt = env->prog->len;
17356 for (i = 0; i < insn_cnt; i++) {
17357 if (memcmp(&insn[i], &ja, sizeof(ja)))
17360 err = verifier_remove_insns(env, i, 1);
17370 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
17371 const union bpf_attr *attr)
17373 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
17374 struct bpf_insn_aux_data *aux = env->insn_aux_data;
17375 int i, patch_len, delta = 0, len = env->prog->len;
17376 struct bpf_insn *insns = env->prog->insnsi;
17377 struct bpf_prog *new_prog;
17380 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
17381 zext_patch[1] = BPF_ZEXT_REG(0);
17382 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
17383 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
17384 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
17385 for (i = 0; i < len; i++) {
17386 int adj_idx = i + delta;
17387 struct bpf_insn insn;
17390 insn = insns[adj_idx];
17391 load_reg = insn_def_regno(&insn);
17392 if (!aux[adj_idx].zext_dst) {
17400 class = BPF_CLASS(code);
17401 if (load_reg == -1)
17404 /* NOTE: arg "reg" (the fourth one) is only used for
17405 * BPF_STX + SRC_OP, so it is safe to pass NULL
17408 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
17409 if (class == BPF_LD &&
17410 BPF_MODE(code) == BPF_IMM)
17415 /* ctx load could be transformed into wider load. */
17416 if (class == BPF_LDX &&
17417 aux[adj_idx].ptr_type == PTR_TO_CTX)
17420 imm_rnd = get_random_u32();
17421 rnd_hi32_patch[0] = insn;
17422 rnd_hi32_patch[1].imm = imm_rnd;
17423 rnd_hi32_patch[3].dst_reg = load_reg;
17424 patch = rnd_hi32_patch;
17426 goto apply_patch_buffer;
17429 /* Add in an zero-extend instruction if a) the JIT has requested
17430 * it or b) it's a CMPXCHG.
17432 * The latter is because: BPF_CMPXCHG always loads a value into
17433 * R0, therefore always zero-extends. However some archs'
17434 * equivalent instruction only does this load when the
17435 * comparison is successful. This detail of CMPXCHG is
17436 * orthogonal to the general zero-extension behaviour of the
17437 * CPU, so it's treated independently of bpf_jit_needs_zext.
17439 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
17442 /* Zero-extension is done by the caller. */
17443 if (bpf_pseudo_kfunc_call(&insn))
17446 if (WARN_ON(load_reg == -1)) {
17447 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
17451 zext_patch[0] = insn;
17452 zext_patch[1].dst_reg = load_reg;
17453 zext_patch[1].src_reg = load_reg;
17454 patch = zext_patch;
17456 apply_patch_buffer:
17457 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
17460 env->prog = new_prog;
17461 insns = new_prog->insnsi;
17462 aux = env->insn_aux_data;
17463 delta += patch_len - 1;
17469 /* convert load instructions that access fields of a context type into a
17470 * sequence of instructions that access fields of the underlying structure:
17471 * struct __sk_buff -> struct sk_buff
17472 * struct bpf_sock_ops -> struct sock
17474 static int convert_ctx_accesses(struct bpf_verifier_env *env)
17476 const struct bpf_verifier_ops *ops = env->ops;
17477 int i, cnt, size, ctx_field_size, delta = 0;
17478 const int insn_cnt = env->prog->len;
17479 struct bpf_insn insn_buf[16], *insn;
17480 u32 target_size, size_default, off;
17481 struct bpf_prog *new_prog;
17482 enum bpf_access_type type;
17483 bool is_narrower_load;
17485 if (ops->gen_prologue || env->seen_direct_write) {
17486 if (!ops->gen_prologue) {
17487 verbose(env, "bpf verifier is misconfigured\n");
17490 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
17492 if (cnt >= ARRAY_SIZE(insn_buf)) {
17493 verbose(env, "bpf verifier is misconfigured\n");
17496 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
17500 env->prog = new_prog;
17505 if (bpf_prog_is_offloaded(env->prog->aux))
17508 insn = env->prog->insnsi + delta;
17510 for (i = 0; i < insn_cnt; i++, insn++) {
17511 bpf_convert_ctx_access_t convert_ctx_access;
17513 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
17514 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
17515 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
17516 insn->code == (BPF_LDX | BPF_MEM | BPF_DW)) {
17518 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
17519 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
17520 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
17521 insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
17522 insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
17523 insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
17524 insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
17525 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
17531 if (type == BPF_WRITE &&
17532 env->insn_aux_data[i + delta].sanitize_stack_spill) {
17533 struct bpf_insn patch[] = {
17538 cnt = ARRAY_SIZE(patch);
17539 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
17544 env->prog = new_prog;
17545 insn = new_prog->insnsi + i + delta;
17549 switch ((int)env->insn_aux_data[i + delta].ptr_type) {
17551 if (!ops->convert_ctx_access)
17553 convert_ctx_access = ops->convert_ctx_access;
17555 case PTR_TO_SOCKET:
17556 case PTR_TO_SOCK_COMMON:
17557 convert_ctx_access = bpf_sock_convert_ctx_access;
17559 case PTR_TO_TCP_SOCK:
17560 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
17562 case PTR_TO_XDP_SOCK:
17563 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
17565 case PTR_TO_BTF_ID:
17566 case PTR_TO_BTF_ID | PTR_UNTRUSTED:
17567 /* PTR_TO_BTF_ID | MEM_ALLOC always has a valid lifetime, unlike
17568 * PTR_TO_BTF_ID, and an active ref_obj_id, but the same cannot
17569 * be said once it is marked PTR_UNTRUSTED, hence we must handle
17570 * any faults for loads into such types. BPF_WRITE is disallowed
17573 case PTR_TO_BTF_ID | MEM_ALLOC | PTR_UNTRUSTED:
17574 if (type == BPF_READ) {
17575 insn->code = BPF_LDX | BPF_PROBE_MEM |
17576 BPF_SIZE((insn)->code);
17577 env->prog->aux->num_exentries++;
17584 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
17585 size = BPF_LDST_BYTES(insn);
17587 /* If the read access is a narrower load of the field,
17588 * convert to a 4/8-byte load, to minimum program type specific
17589 * convert_ctx_access changes. If conversion is successful,
17590 * we will apply proper mask to the result.
17592 is_narrower_load = size < ctx_field_size;
17593 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
17595 if (is_narrower_load) {
17598 if (type == BPF_WRITE) {
17599 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
17604 if (ctx_field_size == 4)
17606 else if (ctx_field_size == 8)
17607 size_code = BPF_DW;
17609 insn->off = off & ~(size_default - 1);
17610 insn->code = BPF_LDX | BPF_MEM | size_code;
17614 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
17616 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
17617 (ctx_field_size && !target_size)) {
17618 verbose(env, "bpf verifier is misconfigured\n");
17622 if (is_narrower_load && size < target_size) {
17623 u8 shift = bpf_ctx_narrow_access_offset(
17624 off, size, size_default) * 8;
17625 if (shift && cnt + 1 >= ARRAY_SIZE(insn_buf)) {
17626 verbose(env, "bpf verifier narrow ctx load misconfigured\n");
17629 if (ctx_field_size <= 4) {
17631 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
17634 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
17635 (1 << size * 8) - 1);
17638 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
17641 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
17642 (1ULL << size * 8) - 1);
17646 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
17652 /* keep walking new program and skip insns we just inserted */
17653 env->prog = new_prog;
17654 insn = new_prog->insnsi + i + delta;
17660 static int jit_subprogs(struct bpf_verifier_env *env)
17662 struct bpf_prog *prog = env->prog, **func, *tmp;
17663 int i, j, subprog_start, subprog_end = 0, len, subprog;
17664 struct bpf_map *map_ptr;
17665 struct bpf_insn *insn;
17666 void *old_bpf_func;
17667 int err, num_exentries;
17669 if (env->subprog_cnt <= 1)
17672 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
17673 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn))
17676 /* Upon error here we cannot fall back to interpreter but
17677 * need a hard reject of the program. Thus -EFAULT is
17678 * propagated in any case.
17680 subprog = find_subprog(env, i + insn->imm + 1);
17682 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
17683 i + insn->imm + 1);
17686 /* temporarily remember subprog id inside insn instead of
17687 * aux_data, since next loop will split up all insns into funcs
17689 insn->off = subprog;
17690 /* remember original imm in case JIT fails and fallback
17691 * to interpreter will be needed
17693 env->insn_aux_data[i].call_imm = insn->imm;
17694 /* point imm to __bpf_call_base+1 from JITs point of view */
17696 if (bpf_pseudo_func(insn))
17697 /* jit (e.g. x86_64) may emit fewer instructions
17698 * if it learns a u32 imm is the same as a u64 imm.
17699 * Force a non zero here.
17704 err = bpf_prog_alloc_jited_linfo(prog);
17706 goto out_undo_insn;
17709 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
17711 goto out_undo_insn;
17713 for (i = 0; i < env->subprog_cnt; i++) {
17714 subprog_start = subprog_end;
17715 subprog_end = env->subprog_info[i + 1].start;
17717 len = subprog_end - subprog_start;
17718 /* bpf_prog_run() doesn't call subprogs directly,
17719 * hence main prog stats include the runtime of subprogs.
17720 * subprogs don't have IDs and not reachable via prog_get_next_id
17721 * func[i]->stats will never be accessed and stays NULL
17723 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
17726 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
17727 len * sizeof(struct bpf_insn));
17728 func[i]->type = prog->type;
17729 func[i]->len = len;
17730 if (bpf_prog_calc_tag(func[i]))
17732 func[i]->is_func = 1;
17733 func[i]->aux->func_idx = i;
17734 /* Below members will be freed only at prog->aux */
17735 func[i]->aux->btf = prog->aux->btf;
17736 func[i]->aux->func_info = prog->aux->func_info;
17737 func[i]->aux->func_info_cnt = prog->aux->func_info_cnt;
17738 func[i]->aux->poke_tab = prog->aux->poke_tab;
17739 func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
17741 for (j = 0; j < prog->aux->size_poke_tab; j++) {
17742 struct bpf_jit_poke_descriptor *poke;
17744 poke = &prog->aux->poke_tab[j];
17745 if (poke->insn_idx < subprog_end &&
17746 poke->insn_idx >= subprog_start)
17747 poke->aux = func[i]->aux;
17750 func[i]->aux->name[0] = 'F';
17751 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
17752 func[i]->jit_requested = 1;
17753 func[i]->blinding_requested = prog->blinding_requested;
17754 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
17755 func[i]->aux->kfunc_btf_tab = prog->aux->kfunc_btf_tab;
17756 func[i]->aux->linfo = prog->aux->linfo;
17757 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
17758 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
17759 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
17761 insn = func[i]->insnsi;
17762 for (j = 0; j < func[i]->len; j++, insn++) {
17763 if (BPF_CLASS(insn->code) == BPF_LDX &&
17764 BPF_MODE(insn->code) == BPF_PROBE_MEM)
17767 func[i]->aux->num_exentries = num_exentries;
17768 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
17769 func[i] = bpf_int_jit_compile(func[i]);
17770 if (!func[i]->jited) {
17777 /* at this point all bpf functions were successfully JITed
17778 * now populate all bpf_calls with correct addresses and
17779 * run last pass of JIT
17781 for (i = 0; i < env->subprog_cnt; i++) {
17782 insn = func[i]->insnsi;
17783 for (j = 0; j < func[i]->len; j++, insn++) {
17784 if (bpf_pseudo_func(insn)) {
17785 subprog = insn->off;
17786 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
17787 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
17790 if (!bpf_pseudo_call(insn))
17792 subprog = insn->off;
17793 insn->imm = BPF_CALL_IMM(func[subprog]->bpf_func);
17796 /* we use the aux data to keep a list of the start addresses
17797 * of the JITed images for each function in the program
17799 * for some architectures, such as powerpc64, the imm field
17800 * might not be large enough to hold the offset of the start
17801 * address of the callee's JITed image from __bpf_call_base
17803 * in such cases, we can lookup the start address of a callee
17804 * by using its subprog id, available from the off field of
17805 * the call instruction, as an index for this list
17807 func[i]->aux->func = func;
17808 func[i]->aux->func_cnt = env->subprog_cnt;
17810 for (i = 0; i < env->subprog_cnt; i++) {
17811 old_bpf_func = func[i]->bpf_func;
17812 tmp = bpf_int_jit_compile(func[i]);
17813 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
17814 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
17821 /* finally lock prog and jit images for all functions and
17822 * populate kallsysm. Begin at the first subprogram, since
17823 * bpf_prog_load will add the kallsyms for the main program.
17825 for (i = 1; i < env->subprog_cnt; i++) {
17826 bpf_prog_lock_ro(func[i]);
17827 bpf_prog_kallsyms_add(func[i]);
17830 /* Last step: make now unused interpreter insns from main
17831 * prog consistent for later dump requests, so they can
17832 * later look the same as if they were interpreted only.
17834 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
17835 if (bpf_pseudo_func(insn)) {
17836 insn[0].imm = env->insn_aux_data[i].call_imm;
17837 insn[1].imm = insn->off;
17841 if (!bpf_pseudo_call(insn))
17843 insn->off = env->insn_aux_data[i].call_imm;
17844 subprog = find_subprog(env, i + insn->off + 1);
17845 insn->imm = subprog;
17849 prog->bpf_func = func[0]->bpf_func;
17850 prog->jited_len = func[0]->jited_len;
17851 prog->aux->extable = func[0]->aux->extable;
17852 prog->aux->num_exentries = func[0]->aux->num_exentries;
17853 prog->aux->func = func;
17854 prog->aux->func_cnt = env->subprog_cnt;
17855 bpf_prog_jit_attempt_done(prog);
17858 /* We failed JIT'ing, so at this point we need to unregister poke
17859 * descriptors from subprogs, so that kernel is not attempting to
17860 * patch it anymore as we're freeing the subprog JIT memory.
17862 for (i = 0; i < prog->aux->size_poke_tab; i++) {
17863 map_ptr = prog->aux->poke_tab[i].tail_call.map;
17864 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
17866 /* At this point we're guaranteed that poke descriptors are not
17867 * live anymore. We can just unlink its descriptor table as it's
17868 * released with the main prog.
17870 for (i = 0; i < env->subprog_cnt; i++) {
17873 func[i]->aux->poke_tab = NULL;
17874 bpf_jit_free(func[i]);
17878 /* cleanup main prog to be interpreted */
17879 prog->jit_requested = 0;
17880 prog->blinding_requested = 0;
17881 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
17882 if (!bpf_pseudo_call(insn))
17885 insn->imm = env->insn_aux_data[i].call_imm;
17887 bpf_prog_jit_attempt_done(prog);
17891 static int fixup_call_args(struct bpf_verifier_env *env)
17893 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
17894 struct bpf_prog *prog = env->prog;
17895 struct bpf_insn *insn = prog->insnsi;
17896 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
17901 if (env->prog->jit_requested &&
17902 !bpf_prog_is_offloaded(env->prog->aux)) {
17903 err = jit_subprogs(env);
17906 if (err == -EFAULT)
17909 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
17910 if (has_kfunc_call) {
17911 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
17914 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
17915 /* When JIT fails the progs with bpf2bpf calls and tail_calls
17916 * have to be rejected, since interpreter doesn't support them yet.
17918 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
17921 for (i = 0; i < prog->len; i++, insn++) {
17922 if (bpf_pseudo_func(insn)) {
17923 /* When JIT fails the progs with callback calls
17924 * have to be rejected, since interpreter doesn't support them yet.
17926 verbose(env, "callbacks are not allowed in non-JITed programs\n");
17930 if (!bpf_pseudo_call(insn))
17932 depth = get_callee_stack_depth(env, insn, i);
17935 bpf_patch_call_args(insn, depth);
17942 /* replace a generic kfunc with a specialized version if necessary */
17943 static void specialize_kfunc(struct bpf_verifier_env *env,
17944 u32 func_id, u16 offset, unsigned long *addr)
17946 struct bpf_prog *prog = env->prog;
17947 bool seen_direct_write;
17951 if (bpf_dev_bound_kfunc_id(func_id)) {
17952 xdp_kfunc = bpf_dev_bound_resolve_kfunc(prog, func_id);
17954 *addr = (unsigned long)xdp_kfunc;
17957 /* fallback to default kfunc when not supported by netdev */
17963 if (func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) {
17964 seen_direct_write = env->seen_direct_write;
17965 is_rdonly = !may_access_direct_pkt_data(env, NULL, BPF_WRITE);
17968 *addr = (unsigned long)bpf_dynptr_from_skb_rdonly;
17970 /* restore env->seen_direct_write to its original value, since
17971 * may_access_direct_pkt_data mutates it
17973 env->seen_direct_write = seen_direct_write;
17977 static void __fixup_collection_insert_kfunc(struct bpf_insn_aux_data *insn_aux,
17978 u16 struct_meta_reg,
17979 u16 node_offset_reg,
17980 struct bpf_insn *insn,
17981 struct bpf_insn *insn_buf,
17984 struct btf_struct_meta *kptr_struct_meta = insn_aux->kptr_struct_meta;
17985 struct bpf_insn addr[2] = { BPF_LD_IMM64(struct_meta_reg, (long)kptr_struct_meta) };
17987 insn_buf[0] = addr[0];
17988 insn_buf[1] = addr[1];
17989 insn_buf[2] = BPF_MOV64_IMM(node_offset_reg, insn_aux->insert_off);
17990 insn_buf[3] = *insn;
17994 static int fixup_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
17995 struct bpf_insn *insn_buf, int insn_idx, int *cnt)
17997 const struct bpf_kfunc_desc *desc;
18000 verbose(env, "invalid kernel function call not eliminated in verifier pass\n");
18006 /* insn->imm has the btf func_id. Replace it with an offset relative to
18007 * __bpf_call_base, unless the JIT needs to call functions that are
18008 * further than 32 bits away (bpf_jit_supports_far_kfunc_call()).
18010 desc = find_kfunc_desc(env->prog, insn->imm, insn->off);
18012 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
18017 if (!bpf_jit_supports_far_kfunc_call())
18018 insn->imm = BPF_CALL_IMM(desc->addr);
18021 if (desc->func_id == special_kfunc_list[KF_bpf_obj_new_impl]) {
18022 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18023 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
18024 u64 obj_new_size = env->insn_aux_data[insn_idx].obj_new_size;
18026 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_1, obj_new_size);
18027 insn_buf[1] = addr[0];
18028 insn_buf[2] = addr[1];
18029 insn_buf[3] = *insn;
18031 } else if (desc->func_id == special_kfunc_list[KF_bpf_obj_drop_impl] ||
18032 desc->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) {
18033 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18034 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
18036 insn_buf[0] = addr[0];
18037 insn_buf[1] = addr[1];
18038 insn_buf[2] = *insn;
18040 } else if (desc->func_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
18041 desc->func_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
18042 desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
18043 int struct_meta_reg = BPF_REG_3;
18044 int node_offset_reg = BPF_REG_4;
18046 /* rbtree_add has extra 'less' arg, so args-to-fixup are in diff regs */
18047 if (desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
18048 struct_meta_reg = BPF_REG_4;
18049 node_offset_reg = BPF_REG_5;
18052 __fixup_collection_insert_kfunc(&env->insn_aux_data[insn_idx], struct_meta_reg,
18053 node_offset_reg, insn, insn_buf, cnt);
18054 } else if (desc->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx] ||
18055 desc->func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
18056 insn_buf[0] = BPF_MOV64_REG(BPF_REG_0, BPF_REG_1);
18062 /* Do various post-verification rewrites in a single program pass.
18063 * These rewrites simplify JIT and interpreter implementations.
18065 static int do_misc_fixups(struct bpf_verifier_env *env)
18067 struct bpf_prog *prog = env->prog;
18068 enum bpf_attach_type eatype = prog->expected_attach_type;
18069 enum bpf_prog_type prog_type = resolve_prog_type(prog);
18070 struct bpf_insn *insn = prog->insnsi;
18071 const struct bpf_func_proto *fn;
18072 const int insn_cnt = prog->len;
18073 const struct bpf_map_ops *ops;
18074 struct bpf_insn_aux_data *aux;
18075 struct bpf_insn insn_buf[16];
18076 struct bpf_prog *new_prog;
18077 struct bpf_map *map_ptr;
18078 int i, ret, cnt, delta = 0;
18080 for (i = 0; i < insn_cnt; i++, insn++) {
18081 /* Make divide-by-zero exceptions impossible. */
18082 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
18083 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
18084 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
18085 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
18086 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
18087 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
18088 struct bpf_insn *patchlet;
18089 struct bpf_insn chk_and_div[] = {
18090 /* [R,W]x div 0 -> 0 */
18091 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
18092 BPF_JNE | BPF_K, insn->src_reg,
18094 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
18095 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
18098 struct bpf_insn chk_and_mod[] = {
18099 /* [R,W]x mod 0 -> [R,W]x */
18100 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
18101 BPF_JEQ | BPF_K, insn->src_reg,
18102 0, 1 + (is64 ? 0 : 1), 0),
18104 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
18105 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
18108 patchlet = isdiv ? chk_and_div : chk_and_mod;
18109 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
18110 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
18112 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
18117 env->prog = prog = new_prog;
18118 insn = new_prog->insnsi + i + delta;
18122 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
18123 if (BPF_CLASS(insn->code) == BPF_LD &&
18124 (BPF_MODE(insn->code) == BPF_ABS ||
18125 BPF_MODE(insn->code) == BPF_IND)) {
18126 cnt = env->ops->gen_ld_abs(insn, insn_buf);
18127 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
18128 verbose(env, "bpf verifier is misconfigured\n");
18132 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18137 env->prog = prog = new_prog;
18138 insn = new_prog->insnsi + i + delta;
18142 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
18143 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
18144 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
18145 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
18146 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
18147 struct bpf_insn *patch = &insn_buf[0];
18148 bool issrc, isneg, isimm;
18151 aux = &env->insn_aux_data[i + delta];
18152 if (!aux->alu_state ||
18153 aux->alu_state == BPF_ALU_NON_POINTER)
18156 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
18157 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
18158 BPF_ALU_SANITIZE_SRC;
18159 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
18161 off_reg = issrc ? insn->src_reg : insn->dst_reg;
18163 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
18166 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
18167 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
18168 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
18169 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
18170 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
18171 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
18172 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
18175 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
18176 insn->src_reg = BPF_REG_AX;
18178 insn->code = insn->code == code_add ?
18179 code_sub : code_add;
18181 if (issrc && isneg && !isimm)
18182 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
18183 cnt = patch - insn_buf;
18185 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18190 env->prog = prog = new_prog;
18191 insn = new_prog->insnsi + i + delta;
18195 if (insn->code != (BPF_JMP | BPF_CALL))
18197 if (insn->src_reg == BPF_PSEUDO_CALL)
18199 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
18200 ret = fixup_kfunc_call(env, insn, insn_buf, i + delta, &cnt);
18206 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18211 env->prog = prog = new_prog;
18212 insn = new_prog->insnsi + i + delta;
18216 if (insn->imm == BPF_FUNC_get_route_realm)
18217 prog->dst_needed = 1;
18218 if (insn->imm == BPF_FUNC_get_prandom_u32)
18219 bpf_user_rnd_init_once();
18220 if (insn->imm == BPF_FUNC_override_return)
18221 prog->kprobe_override = 1;
18222 if (insn->imm == BPF_FUNC_tail_call) {
18223 /* If we tail call into other programs, we
18224 * cannot make any assumptions since they can
18225 * be replaced dynamically during runtime in
18226 * the program array.
18228 prog->cb_access = 1;
18229 if (!allow_tail_call_in_subprogs(env))
18230 prog->aux->stack_depth = MAX_BPF_STACK;
18231 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
18233 /* mark bpf_tail_call as different opcode to avoid
18234 * conditional branch in the interpreter for every normal
18235 * call and to prevent accidental JITing by JIT compiler
18236 * that doesn't support bpf_tail_call yet
18239 insn->code = BPF_JMP | BPF_TAIL_CALL;
18241 aux = &env->insn_aux_data[i + delta];
18242 if (env->bpf_capable && !prog->blinding_requested &&
18243 prog->jit_requested &&
18244 !bpf_map_key_poisoned(aux) &&
18245 !bpf_map_ptr_poisoned(aux) &&
18246 !bpf_map_ptr_unpriv(aux)) {
18247 struct bpf_jit_poke_descriptor desc = {
18248 .reason = BPF_POKE_REASON_TAIL_CALL,
18249 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
18250 .tail_call.key = bpf_map_key_immediate(aux),
18251 .insn_idx = i + delta,
18254 ret = bpf_jit_add_poke_descriptor(prog, &desc);
18256 verbose(env, "adding tail call poke descriptor failed\n");
18260 insn->imm = ret + 1;
18264 if (!bpf_map_ptr_unpriv(aux))
18267 /* instead of changing every JIT dealing with tail_call
18268 * emit two extra insns:
18269 * if (index >= max_entries) goto out;
18270 * index &= array->index_mask;
18271 * to avoid out-of-bounds cpu speculation
18273 if (bpf_map_ptr_poisoned(aux)) {
18274 verbose(env, "tail_call abusing map_ptr\n");
18278 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
18279 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
18280 map_ptr->max_entries, 2);
18281 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
18282 container_of(map_ptr,
18285 insn_buf[2] = *insn;
18287 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18292 env->prog = prog = new_prog;
18293 insn = new_prog->insnsi + i + delta;
18297 if (insn->imm == BPF_FUNC_timer_set_callback) {
18298 /* The verifier will process callback_fn as many times as necessary
18299 * with different maps and the register states prepared by
18300 * set_timer_callback_state will be accurate.
18302 * The following use case is valid:
18303 * map1 is shared by prog1, prog2, prog3.
18304 * prog1 calls bpf_timer_init for some map1 elements
18305 * prog2 calls bpf_timer_set_callback for some map1 elements.
18306 * Those that were not bpf_timer_init-ed will return -EINVAL.
18307 * prog3 calls bpf_timer_start for some map1 elements.
18308 * Those that were not both bpf_timer_init-ed and
18309 * bpf_timer_set_callback-ed will return -EINVAL.
18311 struct bpf_insn ld_addrs[2] = {
18312 BPF_LD_IMM64(BPF_REG_3, (long)prog->aux),
18315 insn_buf[0] = ld_addrs[0];
18316 insn_buf[1] = ld_addrs[1];
18317 insn_buf[2] = *insn;
18320 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18325 env->prog = prog = new_prog;
18326 insn = new_prog->insnsi + i + delta;
18327 goto patch_call_imm;
18330 if (is_storage_get_function(insn->imm)) {
18331 if (!env->prog->aux->sleepable ||
18332 env->insn_aux_data[i + delta].storage_get_func_atomic)
18333 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_ATOMIC);
18335 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_KERNEL);
18336 insn_buf[1] = *insn;
18339 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18344 env->prog = prog = new_prog;
18345 insn = new_prog->insnsi + i + delta;
18346 goto patch_call_imm;
18349 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
18350 * and other inlining handlers are currently limited to 64 bit
18353 if (prog->jit_requested && BITS_PER_LONG == 64 &&
18354 (insn->imm == BPF_FUNC_map_lookup_elem ||
18355 insn->imm == BPF_FUNC_map_update_elem ||
18356 insn->imm == BPF_FUNC_map_delete_elem ||
18357 insn->imm == BPF_FUNC_map_push_elem ||
18358 insn->imm == BPF_FUNC_map_pop_elem ||
18359 insn->imm == BPF_FUNC_map_peek_elem ||
18360 insn->imm == BPF_FUNC_redirect_map ||
18361 insn->imm == BPF_FUNC_for_each_map_elem ||
18362 insn->imm == BPF_FUNC_map_lookup_percpu_elem)) {
18363 aux = &env->insn_aux_data[i + delta];
18364 if (bpf_map_ptr_poisoned(aux))
18365 goto patch_call_imm;
18367 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
18368 ops = map_ptr->ops;
18369 if (insn->imm == BPF_FUNC_map_lookup_elem &&
18370 ops->map_gen_lookup) {
18371 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
18372 if (cnt == -EOPNOTSUPP)
18373 goto patch_map_ops_generic;
18374 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
18375 verbose(env, "bpf verifier is misconfigured\n");
18379 new_prog = bpf_patch_insn_data(env, i + delta,
18385 env->prog = prog = new_prog;
18386 insn = new_prog->insnsi + i + delta;
18390 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
18391 (void *(*)(struct bpf_map *map, void *key))NULL));
18392 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
18393 (long (*)(struct bpf_map *map, void *key))NULL));
18394 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
18395 (long (*)(struct bpf_map *map, void *key, void *value,
18397 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
18398 (long (*)(struct bpf_map *map, void *value,
18400 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
18401 (long (*)(struct bpf_map *map, void *value))NULL));
18402 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
18403 (long (*)(struct bpf_map *map, void *value))NULL));
18404 BUILD_BUG_ON(!__same_type(ops->map_redirect,
18405 (long (*)(struct bpf_map *map, u64 index, u64 flags))NULL));
18406 BUILD_BUG_ON(!__same_type(ops->map_for_each_callback,
18407 (long (*)(struct bpf_map *map,
18408 bpf_callback_t callback_fn,
18409 void *callback_ctx,
18411 BUILD_BUG_ON(!__same_type(ops->map_lookup_percpu_elem,
18412 (void *(*)(struct bpf_map *map, void *key, u32 cpu))NULL));
18414 patch_map_ops_generic:
18415 switch (insn->imm) {
18416 case BPF_FUNC_map_lookup_elem:
18417 insn->imm = BPF_CALL_IMM(ops->map_lookup_elem);
18419 case BPF_FUNC_map_update_elem:
18420 insn->imm = BPF_CALL_IMM(ops->map_update_elem);
18422 case BPF_FUNC_map_delete_elem:
18423 insn->imm = BPF_CALL_IMM(ops->map_delete_elem);
18425 case BPF_FUNC_map_push_elem:
18426 insn->imm = BPF_CALL_IMM(ops->map_push_elem);
18428 case BPF_FUNC_map_pop_elem:
18429 insn->imm = BPF_CALL_IMM(ops->map_pop_elem);
18431 case BPF_FUNC_map_peek_elem:
18432 insn->imm = BPF_CALL_IMM(ops->map_peek_elem);
18434 case BPF_FUNC_redirect_map:
18435 insn->imm = BPF_CALL_IMM(ops->map_redirect);
18437 case BPF_FUNC_for_each_map_elem:
18438 insn->imm = BPF_CALL_IMM(ops->map_for_each_callback);
18440 case BPF_FUNC_map_lookup_percpu_elem:
18441 insn->imm = BPF_CALL_IMM(ops->map_lookup_percpu_elem);
18445 goto patch_call_imm;
18448 /* Implement bpf_jiffies64 inline. */
18449 if (prog->jit_requested && BITS_PER_LONG == 64 &&
18450 insn->imm == BPF_FUNC_jiffies64) {
18451 struct bpf_insn ld_jiffies_addr[2] = {
18452 BPF_LD_IMM64(BPF_REG_0,
18453 (unsigned long)&jiffies),
18456 insn_buf[0] = ld_jiffies_addr[0];
18457 insn_buf[1] = ld_jiffies_addr[1];
18458 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
18462 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
18468 env->prog = prog = new_prog;
18469 insn = new_prog->insnsi + i + delta;
18473 /* Implement bpf_get_func_arg inline. */
18474 if (prog_type == BPF_PROG_TYPE_TRACING &&
18475 insn->imm == BPF_FUNC_get_func_arg) {
18476 /* Load nr_args from ctx - 8 */
18477 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
18478 insn_buf[1] = BPF_JMP32_REG(BPF_JGE, BPF_REG_2, BPF_REG_0, 6);
18479 insn_buf[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_2, 3);
18480 insn_buf[3] = BPF_ALU64_REG(BPF_ADD, BPF_REG_2, BPF_REG_1);
18481 insn_buf[4] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_2, 0);
18482 insn_buf[5] = BPF_STX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
18483 insn_buf[6] = BPF_MOV64_IMM(BPF_REG_0, 0);
18484 insn_buf[7] = BPF_JMP_A(1);
18485 insn_buf[8] = BPF_MOV64_IMM(BPF_REG_0, -EINVAL);
18488 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18493 env->prog = prog = new_prog;
18494 insn = new_prog->insnsi + i + delta;
18498 /* Implement bpf_get_func_ret inline. */
18499 if (prog_type == BPF_PROG_TYPE_TRACING &&
18500 insn->imm == BPF_FUNC_get_func_ret) {
18501 if (eatype == BPF_TRACE_FEXIT ||
18502 eatype == BPF_MODIFY_RETURN) {
18503 /* Load nr_args from ctx - 8 */
18504 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
18505 insn_buf[1] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_0, 3);
18506 insn_buf[2] = BPF_ALU64_REG(BPF_ADD, BPF_REG_0, BPF_REG_1);
18507 insn_buf[3] = BPF_LDX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
18508 insn_buf[4] = BPF_STX_MEM(BPF_DW, BPF_REG_2, BPF_REG_3, 0);
18509 insn_buf[5] = BPF_MOV64_IMM(BPF_REG_0, 0);
18512 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, -EOPNOTSUPP);
18516 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18521 env->prog = prog = new_prog;
18522 insn = new_prog->insnsi + i + delta;
18526 /* Implement get_func_arg_cnt inline. */
18527 if (prog_type == BPF_PROG_TYPE_TRACING &&
18528 insn->imm == BPF_FUNC_get_func_arg_cnt) {
18529 /* Load nr_args from ctx - 8 */
18530 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
18532 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
18536 env->prog = prog = new_prog;
18537 insn = new_prog->insnsi + i + delta;
18541 /* Implement bpf_get_func_ip inline. */
18542 if (prog_type == BPF_PROG_TYPE_TRACING &&
18543 insn->imm == BPF_FUNC_get_func_ip) {
18544 /* Load IP address from ctx - 16 */
18545 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -16);
18547 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
18551 env->prog = prog = new_prog;
18552 insn = new_prog->insnsi + i + delta;
18557 fn = env->ops->get_func_proto(insn->imm, env->prog);
18558 /* all functions that have prototype and verifier allowed
18559 * programs to call them, must be real in-kernel functions
18563 "kernel subsystem misconfigured func %s#%d\n",
18564 func_id_name(insn->imm), insn->imm);
18567 insn->imm = fn->func - __bpf_call_base;
18570 /* Since poke tab is now finalized, publish aux to tracker. */
18571 for (i = 0; i < prog->aux->size_poke_tab; i++) {
18572 map_ptr = prog->aux->poke_tab[i].tail_call.map;
18573 if (!map_ptr->ops->map_poke_track ||
18574 !map_ptr->ops->map_poke_untrack ||
18575 !map_ptr->ops->map_poke_run) {
18576 verbose(env, "bpf verifier is misconfigured\n");
18580 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
18582 verbose(env, "tracking tail call prog failed\n");
18587 sort_kfunc_descs_by_imm_off(env->prog);
18592 static struct bpf_prog *inline_bpf_loop(struct bpf_verifier_env *env,
18595 u32 callback_subprogno,
18598 s32 r6_offset = stack_base + 0 * BPF_REG_SIZE;
18599 s32 r7_offset = stack_base + 1 * BPF_REG_SIZE;
18600 s32 r8_offset = stack_base + 2 * BPF_REG_SIZE;
18601 int reg_loop_max = BPF_REG_6;
18602 int reg_loop_cnt = BPF_REG_7;
18603 int reg_loop_ctx = BPF_REG_8;
18605 struct bpf_prog *new_prog;
18606 u32 callback_start;
18607 u32 call_insn_offset;
18608 s32 callback_offset;
18610 /* This represents an inlined version of bpf_iter.c:bpf_loop,
18611 * be careful to modify this code in sync.
18613 struct bpf_insn insn_buf[] = {
18614 /* Return error and jump to the end of the patch if
18615 * expected number of iterations is too big.
18617 BPF_JMP_IMM(BPF_JLE, BPF_REG_1, BPF_MAX_LOOPS, 2),
18618 BPF_MOV32_IMM(BPF_REG_0, -E2BIG),
18619 BPF_JMP_IMM(BPF_JA, 0, 0, 16),
18620 /* spill R6, R7, R8 to use these as loop vars */
18621 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_6, r6_offset),
18622 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_7, r7_offset),
18623 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_8, r8_offset),
18624 /* initialize loop vars */
18625 BPF_MOV64_REG(reg_loop_max, BPF_REG_1),
18626 BPF_MOV32_IMM(reg_loop_cnt, 0),
18627 BPF_MOV64_REG(reg_loop_ctx, BPF_REG_3),
18629 * if reg_loop_cnt >= reg_loop_max skip the loop body
18631 BPF_JMP_REG(BPF_JGE, reg_loop_cnt, reg_loop_max, 5),
18633 * correct callback offset would be set after patching
18635 BPF_MOV64_REG(BPF_REG_1, reg_loop_cnt),
18636 BPF_MOV64_REG(BPF_REG_2, reg_loop_ctx),
18638 /* increment loop counter */
18639 BPF_ALU64_IMM(BPF_ADD, reg_loop_cnt, 1),
18640 /* jump to loop header if callback returned 0 */
18641 BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, -6),
18642 /* return value of bpf_loop,
18643 * set R0 to the number of iterations
18645 BPF_MOV64_REG(BPF_REG_0, reg_loop_cnt),
18646 /* restore original values of R6, R7, R8 */
18647 BPF_LDX_MEM(BPF_DW, BPF_REG_6, BPF_REG_10, r6_offset),
18648 BPF_LDX_MEM(BPF_DW, BPF_REG_7, BPF_REG_10, r7_offset),
18649 BPF_LDX_MEM(BPF_DW, BPF_REG_8, BPF_REG_10, r8_offset),
18652 *cnt = ARRAY_SIZE(insn_buf);
18653 new_prog = bpf_patch_insn_data(env, position, insn_buf, *cnt);
18657 /* callback start is known only after patching */
18658 callback_start = env->subprog_info[callback_subprogno].start;
18659 /* Note: insn_buf[12] is an offset of BPF_CALL_REL instruction */
18660 call_insn_offset = position + 12;
18661 callback_offset = callback_start - call_insn_offset - 1;
18662 new_prog->insnsi[call_insn_offset].imm = callback_offset;
18667 static bool is_bpf_loop_call(struct bpf_insn *insn)
18669 return insn->code == (BPF_JMP | BPF_CALL) &&
18670 insn->src_reg == 0 &&
18671 insn->imm == BPF_FUNC_loop;
18674 /* For all sub-programs in the program (including main) check
18675 * insn_aux_data to see if there are bpf_loop calls that require
18676 * inlining. If such calls are found the calls are replaced with a
18677 * sequence of instructions produced by `inline_bpf_loop` function and
18678 * subprog stack_depth is increased by the size of 3 registers.
18679 * This stack space is used to spill values of the R6, R7, R8. These
18680 * registers are used to store the loop bound, counter and context
18683 static int optimize_bpf_loop(struct bpf_verifier_env *env)
18685 struct bpf_subprog_info *subprogs = env->subprog_info;
18686 int i, cur_subprog = 0, cnt, delta = 0;
18687 struct bpf_insn *insn = env->prog->insnsi;
18688 int insn_cnt = env->prog->len;
18689 u16 stack_depth = subprogs[cur_subprog].stack_depth;
18690 u16 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
18691 u16 stack_depth_extra = 0;
18693 for (i = 0; i < insn_cnt; i++, insn++) {
18694 struct bpf_loop_inline_state *inline_state =
18695 &env->insn_aux_data[i + delta].loop_inline_state;
18697 if (is_bpf_loop_call(insn) && inline_state->fit_for_inline) {
18698 struct bpf_prog *new_prog;
18700 stack_depth_extra = BPF_REG_SIZE * 3 + stack_depth_roundup;
18701 new_prog = inline_bpf_loop(env,
18703 -(stack_depth + stack_depth_extra),
18704 inline_state->callback_subprogno,
18710 env->prog = new_prog;
18711 insn = new_prog->insnsi + i + delta;
18714 if (subprogs[cur_subprog + 1].start == i + delta + 1) {
18715 subprogs[cur_subprog].stack_depth += stack_depth_extra;
18717 stack_depth = subprogs[cur_subprog].stack_depth;
18718 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
18719 stack_depth_extra = 0;
18723 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
18728 static void free_states(struct bpf_verifier_env *env)
18730 struct bpf_verifier_state_list *sl, *sln;
18733 sl = env->free_list;
18736 free_verifier_state(&sl->state, false);
18740 env->free_list = NULL;
18742 if (!env->explored_states)
18745 for (i = 0; i < state_htab_size(env); i++) {
18746 sl = env->explored_states[i];
18750 free_verifier_state(&sl->state, false);
18754 env->explored_states[i] = NULL;
18758 static int do_check_common(struct bpf_verifier_env *env, int subprog)
18760 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
18761 struct bpf_verifier_state *state;
18762 struct bpf_reg_state *regs;
18765 env->prev_linfo = NULL;
18768 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
18771 state->curframe = 0;
18772 state->speculative = false;
18773 state->branches = 1;
18774 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
18775 if (!state->frame[0]) {
18779 env->cur_state = state;
18780 init_func_state(env, state->frame[0],
18781 BPF_MAIN_FUNC /* callsite */,
18784 state->first_insn_idx = env->subprog_info[subprog].start;
18785 state->last_insn_idx = -1;
18787 regs = state->frame[state->curframe]->regs;
18788 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
18789 ret = btf_prepare_func_args(env, subprog, regs);
18792 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
18793 if (regs[i].type == PTR_TO_CTX)
18794 mark_reg_known_zero(env, regs, i);
18795 else if (regs[i].type == SCALAR_VALUE)
18796 mark_reg_unknown(env, regs, i);
18797 else if (base_type(regs[i].type) == PTR_TO_MEM) {
18798 const u32 mem_size = regs[i].mem_size;
18800 mark_reg_known_zero(env, regs, i);
18801 regs[i].mem_size = mem_size;
18802 regs[i].id = ++env->id_gen;
18806 /* 1st arg to a function */
18807 regs[BPF_REG_1].type = PTR_TO_CTX;
18808 mark_reg_known_zero(env, regs, BPF_REG_1);
18809 ret = btf_check_subprog_arg_match(env, subprog, regs);
18810 if (ret == -EFAULT)
18811 /* unlikely verifier bug. abort.
18812 * ret == 0 and ret < 0 are sadly acceptable for
18813 * main() function due to backward compatibility.
18814 * Like socket filter program may be written as:
18815 * int bpf_prog(struct pt_regs *ctx)
18816 * and never dereference that ctx in the program.
18817 * 'struct pt_regs' is a type mismatch for socket
18818 * filter that should be using 'struct __sk_buff'.
18823 ret = do_check(env);
18825 /* check for NULL is necessary, since cur_state can be freed inside
18826 * do_check() under memory pressure.
18828 if (env->cur_state) {
18829 free_verifier_state(env->cur_state, true);
18830 env->cur_state = NULL;
18832 while (!pop_stack(env, NULL, NULL, false));
18833 if (!ret && pop_log)
18834 bpf_vlog_reset(&env->log, 0);
18839 /* Verify all global functions in a BPF program one by one based on their BTF.
18840 * All global functions must pass verification. Otherwise the whole program is rejected.
18851 * foo() will be verified first for R1=any_scalar_value. During verification it
18852 * will be assumed that bar() already verified successfully and call to bar()
18853 * from foo() will be checked for type match only. Later bar() will be verified
18854 * independently to check that it's safe for R1=any_scalar_value.
18856 static int do_check_subprogs(struct bpf_verifier_env *env)
18858 struct bpf_prog_aux *aux = env->prog->aux;
18861 if (!aux->func_info)
18864 for (i = 1; i < env->subprog_cnt; i++) {
18865 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
18867 env->insn_idx = env->subprog_info[i].start;
18868 WARN_ON_ONCE(env->insn_idx == 0);
18869 ret = do_check_common(env, i);
18872 } else if (env->log.level & BPF_LOG_LEVEL) {
18874 "Func#%d is safe for any args that match its prototype\n",
18881 static int do_check_main(struct bpf_verifier_env *env)
18886 ret = do_check_common(env, 0);
18888 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
18893 static void print_verification_stats(struct bpf_verifier_env *env)
18897 if (env->log.level & BPF_LOG_STATS) {
18898 verbose(env, "verification time %lld usec\n",
18899 div_u64(env->verification_time, 1000));
18900 verbose(env, "stack depth ");
18901 for (i = 0; i < env->subprog_cnt; i++) {
18902 u32 depth = env->subprog_info[i].stack_depth;
18904 verbose(env, "%d", depth);
18905 if (i + 1 < env->subprog_cnt)
18908 verbose(env, "\n");
18910 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
18911 "total_states %d peak_states %d mark_read %d\n",
18912 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
18913 env->max_states_per_insn, env->total_states,
18914 env->peak_states, env->longest_mark_read_walk);
18917 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
18919 const struct btf_type *t, *func_proto;
18920 const struct bpf_struct_ops *st_ops;
18921 const struct btf_member *member;
18922 struct bpf_prog *prog = env->prog;
18923 u32 btf_id, member_idx;
18926 if (!prog->gpl_compatible) {
18927 verbose(env, "struct ops programs must have a GPL compatible license\n");
18931 btf_id = prog->aux->attach_btf_id;
18932 st_ops = bpf_struct_ops_find(btf_id);
18934 verbose(env, "attach_btf_id %u is not a supported struct\n",
18940 member_idx = prog->expected_attach_type;
18941 if (member_idx >= btf_type_vlen(t)) {
18942 verbose(env, "attach to invalid member idx %u of struct %s\n",
18943 member_idx, st_ops->name);
18947 member = &btf_type_member(t)[member_idx];
18948 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
18949 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
18952 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
18953 mname, member_idx, st_ops->name);
18957 if (st_ops->check_member) {
18958 int err = st_ops->check_member(t, member, prog);
18961 verbose(env, "attach to unsupported member %s of struct %s\n",
18962 mname, st_ops->name);
18967 prog->aux->attach_func_proto = func_proto;
18968 prog->aux->attach_func_name = mname;
18969 env->ops = st_ops->verifier_ops;
18973 #define SECURITY_PREFIX "security_"
18975 static int check_attach_modify_return(unsigned long addr, const char *func_name)
18977 if (within_error_injection_list(addr) ||
18978 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
18984 /* list of non-sleepable functions that are otherwise on
18985 * ALLOW_ERROR_INJECTION list
18987 BTF_SET_START(btf_non_sleepable_error_inject)
18988 /* Three functions below can be called from sleepable and non-sleepable context.
18989 * Assume non-sleepable from bpf safety point of view.
18991 BTF_ID(func, __filemap_add_folio)
18992 BTF_ID(func, should_fail_alloc_page)
18993 BTF_ID(func, should_failslab)
18994 BTF_SET_END(btf_non_sleepable_error_inject)
18996 static int check_non_sleepable_error_inject(u32 btf_id)
18998 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
19001 int bpf_check_attach_target(struct bpf_verifier_log *log,
19002 const struct bpf_prog *prog,
19003 const struct bpf_prog *tgt_prog,
19005 struct bpf_attach_target_info *tgt_info)
19007 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
19008 const char prefix[] = "btf_trace_";
19009 int ret = 0, subprog = -1, i;
19010 const struct btf_type *t;
19011 bool conservative = true;
19015 struct module *mod = NULL;
19018 bpf_log(log, "Tracing programs must provide btf_id\n");
19021 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
19024 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
19027 t = btf_type_by_id(btf, btf_id);
19029 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
19032 tname = btf_name_by_offset(btf, t->name_off);
19034 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
19038 struct bpf_prog_aux *aux = tgt_prog->aux;
19040 if (bpf_prog_is_dev_bound(prog->aux) &&
19041 !bpf_prog_dev_bound_match(prog, tgt_prog)) {
19042 bpf_log(log, "Target program bound device mismatch");
19046 for (i = 0; i < aux->func_info_cnt; i++)
19047 if (aux->func_info[i].type_id == btf_id) {
19051 if (subprog == -1) {
19052 bpf_log(log, "Subprog %s doesn't exist\n", tname);
19055 conservative = aux->func_info_aux[subprog].unreliable;
19056 if (prog_extension) {
19057 if (conservative) {
19059 "Cannot replace static functions\n");
19062 if (!prog->jit_requested) {
19064 "Extension programs should be JITed\n");
19068 if (!tgt_prog->jited) {
19069 bpf_log(log, "Can attach to only JITed progs\n");
19072 if (tgt_prog->type == prog->type) {
19073 /* Cannot fentry/fexit another fentry/fexit program.
19074 * Cannot attach program extension to another extension.
19075 * It's ok to attach fentry/fexit to extension program.
19077 bpf_log(log, "Cannot recursively attach\n");
19080 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
19082 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
19083 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
19084 /* Program extensions can extend all program types
19085 * except fentry/fexit. The reason is the following.
19086 * The fentry/fexit programs are used for performance
19087 * analysis, stats and can be attached to any program
19088 * type except themselves. When extension program is
19089 * replacing XDP function it is necessary to allow
19090 * performance analysis of all functions. Both original
19091 * XDP program and its program extension. Hence
19092 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
19093 * allowed. If extending of fentry/fexit was allowed it
19094 * would be possible to create long call chain
19095 * fentry->extension->fentry->extension beyond
19096 * reasonable stack size. Hence extending fentry is not
19099 bpf_log(log, "Cannot extend fentry/fexit\n");
19103 if (prog_extension) {
19104 bpf_log(log, "Cannot replace kernel functions\n");
19109 switch (prog->expected_attach_type) {
19110 case BPF_TRACE_RAW_TP:
19113 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
19116 if (!btf_type_is_typedef(t)) {
19117 bpf_log(log, "attach_btf_id %u is not a typedef\n",
19121 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
19122 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
19126 tname += sizeof(prefix) - 1;
19127 t = btf_type_by_id(btf, t->type);
19128 if (!btf_type_is_ptr(t))
19129 /* should never happen in valid vmlinux build */
19131 t = btf_type_by_id(btf, t->type);
19132 if (!btf_type_is_func_proto(t))
19133 /* should never happen in valid vmlinux build */
19137 case BPF_TRACE_ITER:
19138 if (!btf_type_is_func(t)) {
19139 bpf_log(log, "attach_btf_id %u is not a function\n",
19143 t = btf_type_by_id(btf, t->type);
19144 if (!btf_type_is_func_proto(t))
19146 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
19151 if (!prog_extension)
19154 case BPF_MODIFY_RETURN:
19156 case BPF_LSM_CGROUP:
19157 case BPF_TRACE_FENTRY:
19158 case BPF_TRACE_FEXIT:
19159 if (!btf_type_is_func(t)) {
19160 bpf_log(log, "attach_btf_id %u is not a function\n",
19164 if (prog_extension &&
19165 btf_check_type_match(log, prog, btf, t))
19167 t = btf_type_by_id(btf, t->type);
19168 if (!btf_type_is_func_proto(t))
19171 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
19172 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
19173 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
19176 if (tgt_prog && conservative)
19179 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
19185 addr = (long) tgt_prog->bpf_func;
19187 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
19189 if (btf_is_module(btf)) {
19190 mod = btf_try_get_module(btf);
19192 addr = find_kallsyms_symbol_value(mod, tname);
19196 addr = kallsyms_lookup_name(tname);
19201 "The address of function %s cannot be found\n",
19207 if (prog->aux->sleepable) {
19209 switch (prog->type) {
19210 case BPF_PROG_TYPE_TRACING:
19212 /* fentry/fexit/fmod_ret progs can be sleepable if they are
19213 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
19215 if (!check_non_sleepable_error_inject(btf_id) &&
19216 within_error_injection_list(addr))
19218 /* fentry/fexit/fmod_ret progs can also be sleepable if they are
19219 * in the fmodret id set with the KF_SLEEPABLE flag.
19222 u32 *flags = btf_kfunc_is_modify_return(btf, btf_id,
19225 if (flags && (*flags & KF_SLEEPABLE))
19229 case BPF_PROG_TYPE_LSM:
19230 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
19231 * Only some of them are sleepable.
19233 if (bpf_lsm_is_sleepable_hook(btf_id))
19241 bpf_log(log, "%s is not sleepable\n", tname);
19244 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
19247 bpf_log(log, "can't modify return codes of BPF programs\n");
19251 if (btf_kfunc_is_modify_return(btf, btf_id, prog) ||
19252 !check_attach_modify_return(addr, tname))
19256 bpf_log(log, "%s() is not modifiable\n", tname);
19263 tgt_info->tgt_addr = addr;
19264 tgt_info->tgt_name = tname;
19265 tgt_info->tgt_type = t;
19266 tgt_info->tgt_mod = mod;
19270 BTF_SET_START(btf_id_deny)
19273 BTF_ID(func, migrate_disable)
19274 BTF_ID(func, migrate_enable)
19276 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
19277 BTF_ID(func, rcu_read_unlock_strict)
19279 #if defined(CONFIG_DEBUG_PREEMPT) || defined(CONFIG_TRACE_PREEMPT_TOGGLE)
19280 BTF_ID(func, preempt_count_add)
19281 BTF_ID(func, preempt_count_sub)
19283 #ifdef CONFIG_PREEMPT_RCU
19284 BTF_ID(func, __rcu_read_lock)
19285 BTF_ID(func, __rcu_read_unlock)
19287 BTF_SET_END(btf_id_deny)
19289 static bool can_be_sleepable(struct bpf_prog *prog)
19291 if (prog->type == BPF_PROG_TYPE_TRACING) {
19292 switch (prog->expected_attach_type) {
19293 case BPF_TRACE_FENTRY:
19294 case BPF_TRACE_FEXIT:
19295 case BPF_MODIFY_RETURN:
19296 case BPF_TRACE_ITER:
19302 return prog->type == BPF_PROG_TYPE_LSM ||
19303 prog->type == BPF_PROG_TYPE_KPROBE /* only for uprobes */ ||
19304 prog->type == BPF_PROG_TYPE_STRUCT_OPS;
19307 static int check_attach_btf_id(struct bpf_verifier_env *env)
19309 struct bpf_prog *prog = env->prog;
19310 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
19311 struct bpf_attach_target_info tgt_info = {};
19312 u32 btf_id = prog->aux->attach_btf_id;
19313 struct bpf_trampoline *tr;
19317 if (prog->type == BPF_PROG_TYPE_SYSCALL) {
19318 if (prog->aux->sleepable)
19319 /* attach_btf_id checked to be zero already */
19321 verbose(env, "Syscall programs can only be sleepable\n");
19325 if (prog->aux->sleepable && !can_be_sleepable(prog)) {
19326 verbose(env, "Only fentry/fexit/fmod_ret, lsm, iter, uprobe, and struct_ops programs can be sleepable\n");
19330 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
19331 return check_struct_ops_btf_id(env);
19333 if (prog->type != BPF_PROG_TYPE_TRACING &&
19334 prog->type != BPF_PROG_TYPE_LSM &&
19335 prog->type != BPF_PROG_TYPE_EXT)
19338 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
19342 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
19343 /* to make freplace equivalent to their targets, they need to
19344 * inherit env->ops and expected_attach_type for the rest of the
19347 env->ops = bpf_verifier_ops[tgt_prog->type];
19348 prog->expected_attach_type = tgt_prog->expected_attach_type;
19351 /* store info about the attachment target that will be used later */
19352 prog->aux->attach_func_proto = tgt_info.tgt_type;
19353 prog->aux->attach_func_name = tgt_info.tgt_name;
19354 prog->aux->mod = tgt_info.tgt_mod;
19357 prog->aux->saved_dst_prog_type = tgt_prog->type;
19358 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
19361 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
19362 prog->aux->attach_btf_trace = true;
19364 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
19365 if (!bpf_iter_prog_supported(prog))
19370 if (prog->type == BPF_PROG_TYPE_LSM) {
19371 ret = bpf_lsm_verify_prog(&env->log, prog);
19374 } else if (prog->type == BPF_PROG_TYPE_TRACING &&
19375 btf_id_set_contains(&btf_id_deny, btf_id)) {
19379 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
19380 tr = bpf_trampoline_get(key, &tgt_info);
19384 prog->aux->dst_trampoline = tr;
19388 struct btf *bpf_get_btf_vmlinux(void)
19390 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
19391 mutex_lock(&bpf_verifier_lock);
19393 btf_vmlinux = btf_parse_vmlinux();
19394 mutex_unlock(&bpf_verifier_lock);
19396 return btf_vmlinux;
19399 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr, __u32 uattr_size)
19401 u64 start_time = ktime_get_ns();
19402 struct bpf_verifier_env *env;
19403 int i, len, ret = -EINVAL, err;
19407 /* no program is valid */
19408 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
19411 /* 'struct bpf_verifier_env' can be global, but since it's not small,
19412 * allocate/free it every time bpf_check() is called
19414 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
19420 len = (*prog)->len;
19421 env->insn_aux_data =
19422 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
19424 if (!env->insn_aux_data)
19426 for (i = 0; i < len; i++)
19427 env->insn_aux_data[i].orig_idx = i;
19429 env->ops = bpf_verifier_ops[env->prog->type];
19430 env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
19431 is_priv = bpf_capable();
19433 bpf_get_btf_vmlinux();
19435 /* grab the mutex to protect few globals used by verifier */
19437 mutex_lock(&bpf_verifier_lock);
19439 /* user could have requested verbose verifier output
19440 * and supplied buffer to store the verification trace
19442 ret = bpf_vlog_init(&env->log, attr->log_level,
19443 (char __user *) (unsigned long) attr->log_buf,
19448 mark_verifier_state_clean(env);
19450 if (IS_ERR(btf_vmlinux)) {
19451 /* Either gcc or pahole or kernel are broken. */
19452 verbose(env, "in-kernel BTF is malformed\n");
19453 ret = PTR_ERR(btf_vmlinux);
19454 goto skip_full_check;
19457 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
19458 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
19459 env->strict_alignment = true;
19460 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
19461 env->strict_alignment = false;
19463 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
19464 env->allow_uninit_stack = bpf_allow_uninit_stack();
19465 env->bypass_spec_v1 = bpf_bypass_spec_v1();
19466 env->bypass_spec_v4 = bpf_bypass_spec_v4();
19467 env->bpf_capable = bpf_capable();
19470 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
19472 env->explored_states = kvcalloc(state_htab_size(env),
19473 sizeof(struct bpf_verifier_state_list *),
19476 if (!env->explored_states)
19477 goto skip_full_check;
19479 ret = add_subprog_and_kfunc(env);
19481 goto skip_full_check;
19483 ret = check_subprogs(env);
19485 goto skip_full_check;
19487 ret = check_btf_info(env, attr, uattr);
19489 goto skip_full_check;
19491 ret = check_attach_btf_id(env);
19493 goto skip_full_check;
19495 ret = resolve_pseudo_ldimm64(env);
19497 goto skip_full_check;
19499 if (bpf_prog_is_offloaded(env->prog->aux)) {
19500 ret = bpf_prog_offload_verifier_prep(env->prog);
19502 goto skip_full_check;
19505 ret = check_cfg(env);
19507 goto skip_full_check;
19509 ret = do_check_subprogs(env);
19510 ret = ret ?: do_check_main(env);
19512 if (ret == 0 && bpf_prog_is_offloaded(env->prog->aux))
19513 ret = bpf_prog_offload_finalize(env);
19516 kvfree(env->explored_states);
19519 ret = check_max_stack_depth(env);
19521 /* instruction rewrites happen after this point */
19523 ret = optimize_bpf_loop(env);
19527 opt_hard_wire_dead_code_branches(env);
19529 ret = opt_remove_dead_code(env);
19531 ret = opt_remove_nops(env);
19534 sanitize_dead_code(env);
19538 /* program is valid, convert *(u32*)(ctx + off) accesses */
19539 ret = convert_ctx_accesses(env);
19542 ret = do_misc_fixups(env);
19544 /* do 32-bit optimization after insn patching has done so those patched
19545 * insns could be handled correctly.
19547 if (ret == 0 && !bpf_prog_is_offloaded(env->prog->aux)) {
19548 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
19549 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
19554 ret = fixup_call_args(env);
19556 env->verification_time = ktime_get_ns() - start_time;
19557 print_verification_stats(env);
19558 env->prog->aux->verified_insns = env->insn_processed;
19560 /* preserve original error even if log finalization is successful */
19561 err = bpf_vlog_finalize(&env->log, &log_true_size);
19565 if (uattr_size >= offsetofend(union bpf_attr, log_true_size) &&
19566 copy_to_bpfptr_offset(uattr, offsetof(union bpf_attr, log_true_size),
19567 &log_true_size, sizeof(log_true_size))) {
19569 goto err_release_maps;
19573 goto err_release_maps;
19575 if (env->used_map_cnt) {
19576 /* if program passed verifier, update used_maps in bpf_prog_info */
19577 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
19578 sizeof(env->used_maps[0]),
19581 if (!env->prog->aux->used_maps) {
19583 goto err_release_maps;
19586 memcpy(env->prog->aux->used_maps, env->used_maps,
19587 sizeof(env->used_maps[0]) * env->used_map_cnt);
19588 env->prog->aux->used_map_cnt = env->used_map_cnt;
19590 if (env->used_btf_cnt) {
19591 /* if program passed verifier, update used_btfs in bpf_prog_aux */
19592 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
19593 sizeof(env->used_btfs[0]),
19595 if (!env->prog->aux->used_btfs) {
19597 goto err_release_maps;
19600 memcpy(env->prog->aux->used_btfs, env->used_btfs,
19601 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
19602 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
19604 if (env->used_map_cnt || env->used_btf_cnt) {
19605 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
19606 * bpf_ld_imm64 instructions
19608 convert_pseudo_ld_imm64(env);
19611 adjust_btf_func(env);
19614 if (!env->prog->aux->used_maps)
19615 /* if we didn't copy map pointers into bpf_prog_info, release
19616 * them now. Otherwise free_used_maps() will release them.
19619 if (!env->prog->aux->used_btfs)
19622 /* extension progs temporarily inherit the attach_type of their targets
19623 for verification purposes, so set it back to zero before returning
19625 if (env->prog->type == BPF_PROG_TYPE_EXT)
19626 env->prog->expected_attach_type = 0;
19631 mutex_unlock(&bpf_verifier_lock);
19632 vfree(env->insn_aux_data);