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>
30 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
31 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
32 [_id] = & _name ## _verifier_ops,
33 #define BPF_MAP_TYPE(_id, _ops)
34 #define BPF_LINK_TYPE(_id, _name)
35 #include <linux/bpf_types.h>
41 /* bpf_check() is a static code analyzer that walks eBPF program
42 * instruction by instruction and updates register/stack state.
43 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
45 * The first pass is depth-first-search to check that the program is a DAG.
46 * It rejects the following programs:
47 * - larger than BPF_MAXINSNS insns
48 * - if loop is present (detected via back-edge)
49 * - unreachable insns exist (shouldn't be a forest. program = one function)
50 * - out of bounds or malformed jumps
51 * The second pass is all possible path descent from the 1st insn.
52 * Since it's analyzing all paths through the program, the length of the
53 * analysis is limited to 64k insn, which may be hit even if total number of
54 * insn is less then 4K, but there are too many branches that change stack/regs.
55 * Number of 'branches to be analyzed' is limited to 1k
57 * On entry to each instruction, each register has a type, and the instruction
58 * changes the types of the registers depending on instruction semantics.
59 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
62 * All registers are 64-bit.
63 * R0 - return register
64 * R1-R5 argument passing registers
65 * R6-R9 callee saved registers
66 * R10 - frame pointer read-only
68 * At the start of BPF program the register R1 contains a pointer to bpf_context
69 * and has type PTR_TO_CTX.
71 * Verifier tracks arithmetic operations on pointers in case:
72 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
73 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
74 * 1st insn copies R10 (which has FRAME_PTR) type into R1
75 * and 2nd arithmetic instruction is pattern matched to recognize
76 * that it wants to construct a pointer to some element within stack.
77 * So after 2nd insn, the register R1 has type PTR_TO_STACK
78 * (and -20 constant is saved for further stack bounds checking).
79 * Meaning that this reg is a pointer to stack plus known immediate constant.
81 * Most of the time the registers have SCALAR_VALUE type, which
82 * means the register has some value, but it's not a valid pointer.
83 * (like pointer plus pointer becomes SCALAR_VALUE type)
85 * When verifier sees load or store instructions the type of base register
86 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
87 * four pointer types recognized by check_mem_access() function.
89 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
90 * and the range of [ptr, ptr + map's value_size) is accessible.
92 * registers used to pass values to function calls are checked against
93 * function argument constraints.
95 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
96 * It means that the register type passed to this function must be
97 * PTR_TO_STACK and it will be used inside the function as
98 * 'pointer to map element key'
100 * For example the argument constraints for bpf_map_lookup_elem():
101 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
102 * .arg1_type = ARG_CONST_MAP_PTR,
103 * .arg2_type = ARG_PTR_TO_MAP_KEY,
105 * ret_type says that this function returns 'pointer to map elem value or null'
106 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
107 * 2nd argument should be a pointer to stack, which will be used inside
108 * the helper function as a pointer to map element key.
110 * On the kernel side the helper function looks like:
111 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
113 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
114 * void *key = (void *) (unsigned long) r2;
117 * here kernel can access 'key' and 'map' pointers safely, knowing that
118 * [key, key + map->key_size) bytes are valid and were initialized on
119 * the stack of eBPF program.
122 * Corresponding eBPF program may look like:
123 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
124 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
125 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
126 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
127 * here verifier looks at prototype of map_lookup_elem() and sees:
128 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
129 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
131 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
132 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
133 * and were initialized prior to this call.
134 * If it's ok, then verifier allows this BPF_CALL insn and looks at
135 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
136 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
137 * returns either pointer to map value or NULL.
139 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
140 * insn, the register holding that pointer in the true branch changes state to
141 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
142 * branch. See check_cond_jmp_op().
144 * After the call R0 is set to return type of the function and registers R1-R5
145 * are set to NOT_INIT to indicate that they are no longer readable.
147 * The following reference types represent a potential reference to a kernel
148 * resource which, after first being allocated, must be checked and freed by
150 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
152 * When the verifier sees a helper call return a reference type, it allocates a
153 * pointer id for the reference and stores it in the current function state.
154 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
155 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
156 * passes through a NULL-check conditional. For the branch wherein the state is
157 * changed to CONST_IMM, the verifier releases the reference.
159 * For each helper function that allocates a reference, such as
160 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
161 * bpf_sk_release(). When a reference type passes into the release function,
162 * the verifier also releases the reference. If any unchecked or unreleased
163 * reference remains at the end of the program, the verifier rejects it.
166 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
167 struct bpf_verifier_stack_elem {
168 /* verifer state is 'st'
169 * before processing instruction 'insn_idx'
170 * and after processing instruction 'prev_insn_idx'
172 struct bpf_verifier_state st;
175 struct bpf_verifier_stack_elem *next;
176 /* length of verifier log at the time this state was pushed on stack */
180 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
181 #define BPF_COMPLEXITY_LIMIT_STATES 64
183 #define BPF_MAP_KEY_POISON (1ULL << 63)
184 #define BPF_MAP_KEY_SEEN (1ULL << 62)
186 #define BPF_MAP_PTR_UNPRIV 1UL
187 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
188 POISON_POINTER_DELTA))
189 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
191 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx);
192 static int release_reference(struct bpf_verifier_env *env, int ref_obj_id);
194 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
196 return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON;
199 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
201 return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV;
204 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
205 const struct bpf_map *map, bool unpriv)
207 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
208 unpriv |= bpf_map_ptr_unpriv(aux);
209 aux->map_ptr_state = (unsigned long)map |
210 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
213 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
215 return aux->map_key_state & BPF_MAP_KEY_POISON;
218 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
220 return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
223 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
225 return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
228 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
230 bool poisoned = bpf_map_key_poisoned(aux);
232 aux->map_key_state = state | BPF_MAP_KEY_SEEN |
233 (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
236 static bool bpf_pseudo_call(const struct bpf_insn *insn)
238 return insn->code == (BPF_JMP | BPF_CALL) &&
239 insn->src_reg == BPF_PSEUDO_CALL;
242 static bool bpf_pseudo_kfunc_call(const struct bpf_insn *insn)
244 return insn->code == (BPF_JMP | BPF_CALL) &&
245 insn->src_reg == BPF_PSEUDO_KFUNC_CALL;
248 struct bpf_call_arg_meta {
249 struct bpf_map *map_ptr;
265 struct btf_field *kptr_field;
266 u8 uninit_dynptr_regno;
269 struct btf *btf_vmlinux;
271 static DEFINE_MUTEX(bpf_verifier_lock);
273 static const struct bpf_line_info *
274 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
276 const struct bpf_line_info *linfo;
277 const struct bpf_prog *prog;
281 nr_linfo = prog->aux->nr_linfo;
283 if (!nr_linfo || insn_off >= prog->len)
286 linfo = prog->aux->linfo;
287 for (i = 1; i < nr_linfo; i++)
288 if (insn_off < linfo[i].insn_off)
291 return &linfo[i - 1];
294 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
299 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
301 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
302 "verifier log line truncated - local buffer too short\n");
304 if (log->level == BPF_LOG_KERNEL) {
305 bool newline = n > 0 && log->kbuf[n - 1] == '\n';
307 pr_err("BPF: %s%s", log->kbuf, newline ? "" : "\n");
311 n = min(log->len_total - log->len_used - 1, n);
313 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
319 static void bpf_vlog_reset(struct bpf_verifier_log *log, u32 new_pos)
323 if (!bpf_verifier_log_needed(log))
326 log->len_used = new_pos;
327 if (put_user(zero, log->ubuf + new_pos))
331 /* log_level controls verbosity level of eBPF verifier.
332 * bpf_verifier_log_write() is used to dump the verification trace to the log,
333 * so the user can figure out what's wrong with the program
335 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
336 const char *fmt, ...)
340 if (!bpf_verifier_log_needed(&env->log))
344 bpf_verifier_vlog(&env->log, fmt, args);
347 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
349 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
351 struct bpf_verifier_env *env = private_data;
354 if (!bpf_verifier_log_needed(&env->log))
358 bpf_verifier_vlog(&env->log, fmt, args);
362 __printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
363 const char *fmt, ...)
367 if (!bpf_verifier_log_needed(log))
371 bpf_verifier_vlog(log, fmt, args);
374 EXPORT_SYMBOL_GPL(bpf_log);
376 static const char *ltrim(const char *s)
384 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
386 const char *prefix_fmt, ...)
388 const struct bpf_line_info *linfo;
390 if (!bpf_verifier_log_needed(&env->log))
393 linfo = find_linfo(env, insn_off);
394 if (!linfo || linfo == env->prev_linfo)
400 va_start(args, prefix_fmt);
401 bpf_verifier_vlog(&env->log, prefix_fmt, args);
406 ltrim(btf_name_by_offset(env->prog->aux->btf,
409 env->prev_linfo = linfo;
412 static void verbose_invalid_scalar(struct bpf_verifier_env *env,
413 struct bpf_reg_state *reg,
414 struct tnum *range, const char *ctx,
415 const char *reg_name)
419 verbose(env, "At %s the register %s ", ctx, reg_name);
420 if (!tnum_is_unknown(reg->var_off)) {
421 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
422 verbose(env, "has value %s", tn_buf);
424 verbose(env, "has unknown scalar value");
426 tnum_strn(tn_buf, sizeof(tn_buf), *range);
427 verbose(env, " should have been in %s\n", tn_buf);
430 static bool type_is_pkt_pointer(enum bpf_reg_type type)
432 type = base_type(type);
433 return type == PTR_TO_PACKET ||
434 type == PTR_TO_PACKET_META;
437 static bool type_is_sk_pointer(enum bpf_reg_type type)
439 return type == PTR_TO_SOCKET ||
440 type == PTR_TO_SOCK_COMMON ||
441 type == PTR_TO_TCP_SOCK ||
442 type == PTR_TO_XDP_SOCK;
445 static bool reg_type_not_null(enum bpf_reg_type type)
447 return type == PTR_TO_SOCKET ||
448 type == PTR_TO_TCP_SOCK ||
449 type == PTR_TO_MAP_VALUE ||
450 type == PTR_TO_MAP_KEY ||
451 type == PTR_TO_SOCK_COMMON;
454 static bool type_is_ptr_alloc_obj(u32 type)
456 return base_type(type) == PTR_TO_BTF_ID && type_flag(type) & MEM_ALLOC;
459 static struct btf_record *reg_btf_record(const struct bpf_reg_state *reg)
461 struct btf_record *rec = NULL;
462 struct btf_struct_meta *meta;
464 if (reg->type == PTR_TO_MAP_VALUE) {
465 rec = reg->map_ptr->record;
466 } else if (type_is_ptr_alloc_obj(reg->type)) {
467 meta = btf_find_struct_meta(reg->btf, reg->btf_id);
474 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
476 return btf_record_has_field(reg_btf_record(reg), BPF_SPIN_LOCK);
479 static bool type_is_rdonly_mem(u32 type)
481 return type & MEM_RDONLY;
484 static bool type_may_be_null(u32 type)
486 return type & PTR_MAYBE_NULL;
489 static bool is_acquire_function(enum bpf_func_id func_id,
490 const struct bpf_map *map)
492 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
494 if (func_id == BPF_FUNC_sk_lookup_tcp ||
495 func_id == BPF_FUNC_sk_lookup_udp ||
496 func_id == BPF_FUNC_skc_lookup_tcp ||
497 func_id == BPF_FUNC_ringbuf_reserve ||
498 func_id == BPF_FUNC_kptr_xchg)
501 if (func_id == BPF_FUNC_map_lookup_elem &&
502 (map_type == BPF_MAP_TYPE_SOCKMAP ||
503 map_type == BPF_MAP_TYPE_SOCKHASH))
509 static bool is_ptr_cast_function(enum bpf_func_id func_id)
511 return func_id == BPF_FUNC_tcp_sock ||
512 func_id == BPF_FUNC_sk_fullsock ||
513 func_id == BPF_FUNC_skc_to_tcp_sock ||
514 func_id == BPF_FUNC_skc_to_tcp6_sock ||
515 func_id == BPF_FUNC_skc_to_udp6_sock ||
516 func_id == BPF_FUNC_skc_to_mptcp_sock ||
517 func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
518 func_id == BPF_FUNC_skc_to_tcp_request_sock;
521 static bool is_dynptr_ref_function(enum bpf_func_id func_id)
523 return func_id == BPF_FUNC_dynptr_data;
526 static bool is_callback_calling_function(enum bpf_func_id func_id)
528 return func_id == BPF_FUNC_for_each_map_elem ||
529 func_id == BPF_FUNC_timer_set_callback ||
530 func_id == BPF_FUNC_find_vma ||
531 func_id == BPF_FUNC_loop ||
532 func_id == BPF_FUNC_user_ringbuf_drain;
535 static bool is_storage_get_function(enum bpf_func_id func_id)
537 return func_id == BPF_FUNC_sk_storage_get ||
538 func_id == BPF_FUNC_inode_storage_get ||
539 func_id == BPF_FUNC_task_storage_get ||
540 func_id == BPF_FUNC_cgrp_storage_get;
543 static bool helper_multiple_ref_obj_use(enum bpf_func_id func_id,
544 const struct bpf_map *map)
546 int ref_obj_uses = 0;
548 if (is_ptr_cast_function(func_id))
550 if (is_acquire_function(func_id, map))
552 if (is_dynptr_ref_function(func_id))
555 return ref_obj_uses > 1;
558 static bool is_cmpxchg_insn(const struct bpf_insn *insn)
560 return BPF_CLASS(insn->code) == BPF_STX &&
561 BPF_MODE(insn->code) == BPF_ATOMIC &&
562 insn->imm == BPF_CMPXCHG;
565 /* string representation of 'enum bpf_reg_type'
567 * Note that reg_type_str() can not appear more than once in a single verbose()
570 static const char *reg_type_str(struct bpf_verifier_env *env,
571 enum bpf_reg_type type)
573 char postfix[16] = {0}, prefix[64] = {0};
574 static const char * const str[] = {
576 [SCALAR_VALUE] = "scalar",
577 [PTR_TO_CTX] = "ctx",
578 [CONST_PTR_TO_MAP] = "map_ptr",
579 [PTR_TO_MAP_VALUE] = "map_value",
580 [PTR_TO_STACK] = "fp",
581 [PTR_TO_PACKET] = "pkt",
582 [PTR_TO_PACKET_META] = "pkt_meta",
583 [PTR_TO_PACKET_END] = "pkt_end",
584 [PTR_TO_FLOW_KEYS] = "flow_keys",
585 [PTR_TO_SOCKET] = "sock",
586 [PTR_TO_SOCK_COMMON] = "sock_common",
587 [PTR_TO_TCP_SOCK] = "tcp_sock",
588 [PTR_TO_TP_BUFFER] = "tp_buffer",
589 [PTR_TO_XDP_SOCK] = "xdp_sock",
590 [PTR_TO_BTF_ID] = "ptr_",
591 [PTR_TO_MEM] = "mem",
592 [PTR_TO_BUF] = "buf",
593 [PTR_TO_FUNC] = "func",
594 [PTR_TO_MAP_KEY] = "map_key",
595 [CONST_PTR_TO_DYNPTR] = "dynptr_ptr",
598 if (type & PTR_MAYBE_NULL) {
599 if (base_type(type) == PTR_TO_BTF_ID)
600 strncpy(postfix, "or_null_", 16);
602 strncpy(postfix, "_or_null", 16);
605 snprintf(prefix, sizeof(prefix), "%s%s%s%s%s%s%s",
606 type & MEM_RDONLY ? "rdonly_" : "",
607 type & MEM_RINGBUF ? "ringbuf_" : "",
608 type & MEM_USER ? "user_" : "",
609 type & MEM_PERCPU ? "percpu_" : "",
610 type & MEM_RCU ? "rcu_" : "",
611 type & PTR_UNTRUSTED ? "untrusted_" : "",
612 type & PTR_TRUSTED ? "trusted_" : ""
615 snprintf(env->type_str_buf, TYPE_STR_BUF_LEN, "%s%s%s",
616 prefix, str[base_type(type)], postfix);
617 return env->type_str_buf;
620 static char slot_type_char[] = {
621 [STACK_INVALID] = '?',
625 [STACK_DYNPTR] = 'd',
628 static void print_liveness(struct bpf_verifier_env *env,
629 enum bpf_reg_liveness live)
631 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
633 if (live & REG_LIVE_READ)
635 if (live & REG_LIVE_WRITTEN)
637 if (live & REG_LIVE_DONE)
641 static int get_spi(s32 off)
643 return (-off - 1) / BPF_REG_SIZE;
646 static bool is_spi_bounds_valid(struct bpf_func_state *state, int spi, int nr_slots)
648 int allocated_slots = state->allocated_stack / BPF_REG_SIZE;
650 /* We need to check that slots between [spi - nr_slots + 1, spi] are
651 * within [0, allocated_stack).
653 * Please note that the spi grows downwards. For example, a dynptr
654 * takes the size of two stack slots; the first slot will be at
655 * spi and the second slot will be at spi - 1.
657 return spi - nr_slots + 1 >= 0 && spi < allocated_slots;
660 static struct bpf_func_state *func(struct bpf_verifier_env *env,
661 const struct bpf_reg_state *reg)
663 struct bpf_verifier_state *cur = env->cur_state;
665 return cur->frame[reg->frameno];
668 static const char *kernel_type_name(const struct btf* btf, u32 id)
670 return btf_name_by_offset(btf, btf_type_by_id(btf, id)->name_off);
673 static void mark_reg_scratched(struct bpf_verifier_env *env, u32 regno)
675 env->scratched_regs |= 1U << regno;
678 static void mark_stack_slot_scratched(struct bpf_verifier_env *env, u32 spi)
680 env->scratched_stack_slots |= 1ULL << spi;
683 static bool reg_scratched(const struct bpf_verifier_env *env, u32 regno)
685 return (env->scratched_regs >> regno) & 1;
688 static bool stack_slot_scratched(const struct bpf_verifier_env *env, u64 regno)
690 return (env->scratched_stack_slots >> regno) & 1;
693 static bool verifier_state_scratched(const struct bpf_verifier_env *env)
695 return env->scratched_regs || env->scratched_stack_slots;
698 static void mark_verifier_state_clean(struct bpf_verifier_env *env)
700 env->scratched_regs = 0U;
701 env->scratched_stack_slots = 0ULL;
704 /* Used for printing the entire verifier state. */
705 static void mark_verifier_state_scratched(struct bpf_verifier_env *env)
707 env->scratched_regs = ~0U;
708 env->scratched_stack_slots = ~0ULL;
711 static enum bpf_dynptr_type arg_to_dynptr_type(enum bpf_arg_type arg_type)
713 switch (arg_type & DYNPTR_TYPE_FLAG_MASK) {
714 case DYNPTR_TYPE_LOCAL:
715 return BPF_DYNPTR_TYPE_LOCAL;
716 case DYNPTR_TYPE_RINGBUF:
717 return BPF_DYNPTR_TYPE_RINGBUF;
719 return BPF_DYNPTR_TYPE_INVALID;
723 static bool dynptr_type_refcounted(enum bpf_dynptr_type type)
725 return type == BPF_DYNPTR_TYPE_RINGBUF;
728 static void __mark_dynptr_reg(struct bpf_reg_state *reg,
729 enum bpf_dynptr_type type,
732 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
733 struct bpf_reg_state *reg);
735 static void mark_dynptr_stack_regs(struct bpf_reg_state *sreg1,
736 struct bpf_reg_state *sreg2,
737 enum bpf_dynptr_type type)
739 __mark_dynptr_reg(sreg1, type, true);
740 __mark_dynptr_reg(sreg2, type, false);
743 static void mark_dynptr_cb_reg(struct bpf_reg_state *reg,
744 enum bpf_dynptr_type type)
746 __mark_dynptr_reg(reg, type, true);
750 static int mark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
751 enum bpf_arg_type arg_type, int insn_idx)
753 struct bpf_func_state *state = func(env, reg);
754 enum bpf_dynptr_type type;
757 spi = get_spi(reg->off);
759 if (!is_spi_bounds_valid(state, spi, BPF_DYNPTR_NR_SLOTS))
762 for (i = 0; i < BPF_REG_SIZE; i++) {
763 state->stack[spi].slot_type[i] = STACK_DYNPTR;
764 state->stack[spi - 1].slot_type[i] = STACK_DYNPTR;
767 type = arg_to_dynptr_type(arg_type);
768 if (type == BPF_DYNPTR_TYPE_INVALID)
771 mark_dynptr_stack_regs(&state->stack[spi].spilled_ptr,
772 &state->stack[spi - 1].spilled_ptr, type);
774 if (dynptr_type_refcounted(type)) {
775 /* The id is used to track proper releasing */
776 id = acquire_reference_state(env, insn_idx);
780 state->stack[spi].spilled_ptr.ref_obj_id = id;
781 state->stack[spi - 1].spilled_ptr.ref_obj_id = id;
787 static int unmark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
789 struct bpf_func_state *state = func(env, reg);
792 spi = get_spi(reg->off);
794 if (!is_spi_bounds_valid(state, spi, BPF_DYNPTR_NR_SLOTS))
797 for (i = 0; i < BPF_REG_SIZE; i++) {
798 state->stack[spi].slot_type[i] = STACK_INVALID;
799 state->stack[spi - 1].slot_type[i] = STACK_INVALID;
802 /* Invalidate any slices associated with this dynptr */
803 if (dynptr_type_refcounted(state->stack[spi].spilled_ptr.dynptr.type))
804 WARN_ON_ONCE(release_reference(env, state->stack[spi].spilled_ptr.ref_obj_id));
806 __mark_reg_not_init(env, &state->stack[spi].spilled_ptr);
807 __mark_reg_not_init(env, &state->stack[spi - 1].spilled_ptr);
811 static bool is_dynptr_reg_valid_uninit(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
813 struct bpf_func_state *state = func(env, reg);
816 if (reg->type == CONST_PTR_TO_DYNPTR)
819 spi = get_spi(reg->off);
820 if (!is_spi_bounds_valid(state, spi, BPF_DYNPTR_NR_SLOTS))
823 for (i = 0; i < BPF_REG_SIZE; i++) {
824 if (state->stack[spi].slot_type[i] == STACK_DYNPTR ||
825 state->stack[spi - 1].slot_type[i] == STACK_DYNPTR)
832 static bool is_dynptr_reg_valid_init(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
834 struct bpf_func_state *state = func(env, reg);
838 /* This already represents first slot of initialized bpf_dynptr */
839 if (reg->type == CONST_PTR_TO_DYNPTR)
842 spi = get_spi(reg->off);
843 if (!is_spi_bounds_valid(state, spi, BPF_DYNPTR_NR_SLOTS) ||
844 !state->stack[spi].spilled_ptr.dynptr.first_slot)
847 for (i = 0; i < BPF_REG_SIZE; i++) {
848 if (state->stack[spi].slot_type[i] != STACK_DYNPTR ||
849 state->stack[spi - 1].slot_type[i] != STACK_DYNPTR)
856 static bool is_dynptr_type_expected(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
857 enum bpf_arg_type arg_type)
859 struct bpf_func_state *state = func(env, reg);
860 enum bpf_dynptr_type dynptr_type;
863 /* ARG_PTR_TO_DYNPTR takes any type of dynptr */
864 if (arg_type == ARG_PTR_TO_DYNPTR)
867 dynptr_type = arg_to_dynptr_type(arg_type);
868 if (reg->type == CONST_PTR_TO_DYNPTR) {
869 return reg->dynptr.type == dynptr_type;
871 spi = get_spi(reg->off);
872 return state->stack[spi].spilled_ptr.dynptr.type == dynptr_type;
876 /* The reg state of a pointer or a bounded scalar was saved when
877 * it was spilled to the stack.
879 static bool is_spilled_reg(const struct bpf_stack_state *stack)
881 return stack->slot_type[BPF_REG_SIZE - 1] == STACK_SPILL;
884 static void scrub_spilled_slot(u8 *stype)
886 if (*stype != STACK_INVALID)
890 static void print_verifier_state(struct bpf_verifier_env *env,
891 const struct bpf_func_state *state,
894 const struct bpf_reg_state *reg;
899 verbose(env, " frame%d:", state->frameno);
900 for (i = 0; i < MAX_BPF_REG; i++) {
901 reg = &state->regs[i];
905 if (!print_all && !reg_scratched(env, i))
907 verbose(env, " R%d", i);
908 print_liveness(env, reg->live);
910 if (t == SCALAR_VALUE && reg->precise)
912 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
913 tnum_is_const(reg->var_off)) {
914 /* reg->off should be 0 for SCALAR_VALUE */
915 verbose(env, "%s", t == SCALAR_VALUE ? "" : reg_type_str(env, t));
916 verbose(env, "%lld", reg->var_off.value + reg->off);
918 const char *sep = "";
920 verbose(env, "%s", reg_type_str(env, t));
921 if (base_type(t) == PTR_TO_BTF_ID)
922 verbose(env, "%s", kernel_type_name(reg->btf, reg->btf_id));
925 * _a stands for append, was shortened to avoid multiline statements below.
926 * This macro is used to output a comma separated list of attributes.
928 #define verbose_a(fmt, ...) ({ verbose(env, "%s" fmt, sep, __VA_ARGS__); sep = ","; })
931 verbose_a("id=%d", reg->id);
933 verbose_a("ref_obj_id=%d", reg->ref_obj_id);
934 if (t != SCALAR_VALUE)
935 verbose_a("off=%d", reg->off);
936 if (type_is_pkt_pointer(t))
937 verbose_a("r=%d", reg->range);
938 else if (base_type(t) == CONST_PTR_TO_MAP ||
939 base_type(t) == PTR_TO_MAP_KEY ||
940 base_type(t) == PTR_TO_MAP_VALUE)
941 verbose_a("ks=%d,vs=%d",
942 reg->map_ptr->key_size,
943 reg->map_ptr->value_size);
944 if (tnum_is_const(reg->var_off)) {
945 /* Typically an immediate SCALAR_VALUE, but
946 * could be a pointer whose offset is too big
949 verbose_a("imm=%llx", reg->var_off.value);
951 if (reg->smin_value != reg->umin_value &&
952 reg->smin_value != S64_MIN)
953 verbose_a("smin=%lld", (long long)reg->smin_value);
954 if (reg->smax_value != reg->umax_value &&
955 reg->smax_value != S64_MAX)
956 verbose_a("smax=%lld", (long long)reg->smax_value);
957 if (reg->umin_value != 0)
958 verbose_a("umin=%llu", (unsigned long long)reg->umin_value);
959 if (reg->umax_value != U64_MAX)
960 verbose_a("umax=%llu", (unsigned long long)reg->umax_value);
961 if (!tnum_is_unknown(reg->var_off)) {
964 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
965 verbose_a("var_off=%s", tn_buf);
967 if (reg->s32_min_value != reg->smin_value &&
968 reg->s32_min_value != S32_MIN)
969 verbose_a("s32_min=%d", (int)(reg->s32_min_value));
970 if (reg->s32_max_value != reg->smax_value &&
971 reg->s32_max_value != S32_MAX)
972 verbose_a("s32_max=%d", (int)(reg->s32_max_value));
973 if (reg->u32_min_value != reg->umin_value &&
974 reg->u32_min_value != U32_MIN)
975 verbose_a("u32_min=%d", (int)(reg->u32_min_value));
976 if (reg->u32_max_value != reg->umax_value &&
977 reg->u32_max_value != U32_MAX)
978 verbose_a("u32_max=%d", (int)(reg->u32_max_value));
985 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
986 char types_buf[BPF_REG_SIZE + 1];
990 for (j = 0; j < BPF_REG_SIZE; j++) {
991 if (state->stack[i].slot_type[j] != STACK_INVALID)
993 types_buf[j] = slot_type_char[
994 state->stack[i].slot_type[j]];
996 types_buf[BPF_REG_SIZE] = 0;
999 if (!print_all && !stack_slot_scratched(env, i))
1001 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
1002 print_liveness(env, state->stack[i].spilled_ptr.live);
1003 if (is_spilled_reg(&state->stack[i])) {
1004 reg = &state->stack[i].spilled_ptr;
1006 verbose(env, "=%s", t == SCALAR_VALUE ? "" : reg_type_str(env, t));
1007 if (t == SCALAR_VALUE && reg->precise)
1009 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
1010 verbose(env, "%lld", reg->var_off.value + reg->off);
1012 verbose(env, "=%s", types_buf);
1015 if (state->acquired_refs && state->refs[0].id) {
1016 verbose(env, " refs=%d", state->refs[0].id);
1017 for (i = 1; i < state->acquired_refs; i++)
1018 if (state->refs[i].id)
1019 verbose(env, ",%d", state->refs[i].id);
1021 if (state->in_callback_fn)
1022 verbose(env, " cb");
1023 if (state->in_async_callback_fn)
1024 verbose(env, " async_cb");
1026 mark_verifier_state_clean(env);
1029 static inline u32 vlog_alignment(u32 pos)
1031 return round_up(max(pos + BPF_LOG_MIN_ALIGNMENT / 2, BPF_LOG_ALIGNMENT),
1032 BPF_LOG_MIN_ALIGNMENT) - pos - 1;
1035 static void print_insn_state(struct bpf_verifier_env *env,
1036 const struct bpf_func_state *state)
1038 if (env->prev_log_len && env->prev_log_len == env->log.len_used) {
1039 /* remove new line character */
1040 bpf_vlog_reset(&env->log, env->prev_log_len - 1);
1041 verbose(env, "%*c;", vlog_alignment(env->prev_insn_print_len), ' ');
1043 verbose(env, "%d:", env->insn_idx);
1045 print_verifier_state(env, state, false);
1048 /* copy array src of length n * size bytes to dst. dst is reallocated if it's too
1049 * small to hold src. This is different from krealloc since we don't want to preserve
1050 * the contents of dst.
1052 * Leaves dst untouched if src is NULL or length is zero. Returns NULL if memory could
1055 static void *copy_array(void *dst, const void *src, size_t n, size_t size, gfp_t flags)
1059 if (ZERO_OR_NULL_PTR(src))
1062 if (unlikely(check_mul_overflow(n, size, &bytes)))
1065 if (ksize(dst) < ksize(src)) {
1067 dst = kmalloc_track_caller(kmalloc_size_roundup(bytes), flags);
1072 memcpy(dst, src, bytes);
1074 return dst ? dst : ZERO_SIZE_PTR;
1077 /* resize an array from old_n items to new_n items. the array is reallocated if it's too
1078 * small to hold new_n items. new items are zeroed out if the array grows.
1080 * Contrary to krealloc_array, does not free arr if new_n is zero.
1082 static void *realloc_array(void *arr, size_t old_n, size_t new_n, size_t size)
1087 if (!new_n || old_n == new_n)
1090 alloc_size = kmalloc_size_roundup(size_mul(new_n, size));
1091 new_arr = krealloc(arr, alloc_size, GFP_KERNEL);
1099 memset(arr + old_n * size, 0, (new_n - old_n) * size);
1102 return arr ? arr : ZERO_SIZE_PTR;
1105 static int copy_reference_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
1107 dst->refs = copy_array(dst->refs, src->refs, src->acquired_refs,
1108 sizeof(struct bpf_reference_state), GFP_KERNEL);
1112 dst->acquired_refs = src->acquired_refs;
1116 static int copy_stack_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
1118 size_t n = src->allocated_stack / BPF_REG_SIZE;
1120 dst->stack = copy_array(dst->stack, src->stack, n, sizeof(struct bpf_stack_state),
1125 dst->allocated_stack = src->allocated_stack;
1129 static int resize_reference_state(struct bpf_func_state *state, size_t n)
1131 state->refs = realloc_array(state->refs, state->acquired_refs, n,
1132 sizeof(struct bpf_reference_state));
1136 state->acquired_refs = n;
1140 static int grow_stack_state(struct bpf_func_state *state, int size)
1142 size_t old_n = state->allocated_stack / BPF_REG_SIZE, n = size / BPF_REG_SIZE;
1147 state->stack = realloc_array(state->stack, old_n, n, sizeof(struct bpf_stack_state));
1151 state->allocated_stack = size;
1155 /* Acquire a pointer id from the env and update the state->refs to include
1156 * this new pointer reference.
1157 * On success, returns a valid pointer id to associate with the register
1158 * On failure, returns a negative errno.
1160 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
1162 struct bpf_func_state *state = cur_func(env);
1163 int new_ofs = state->acquired_refs;
1166 err = resize_reference_state(state, state->acquired_refs + 1);
1170 state->refs[new_ofs].id = id;
1171 state->refs[new_ofs].insn_idx = insn_idx;
1172 state->refs[new_ofs].callback_ref = state->in_callback_fn ? state->frameno : 0;
1177 /* release function corresponding to acquire_reference_state(). Idempotent. */
1178 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
1182 last_idx = state->acquired_refs - 1;
1183 for (i = 0; i < state->acquired_refs; i++) {
1184 if (state->refs[i].id == ptr_id) {
1185 /* Cannot release caller references in callbacks */
1186 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno)
1188 if (last_idx && i != last_idx)
1189 memcpy(&state->refs[i], &state->refs[last_idx],
1190 sizeof(*state->refs));
1191 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
1192 state->acquired_refs--;
1199 static void free_func_state(struct bpf_func_state *state)
1204 kfree(state->stack);
1208 static void clear_jmp_history(struct bpf_verifier_state *state)
1210 kfree(state->jmp_history);
1211 state->jmp_history = NULL;
1212 state->jmp_history_cnt = 0;
1215 static void free_verifier_state(struct bpf_verifier_state *state,
1220 for (i = 0; i <= state->curframe; i++) {
1221 free_func_state(state->frame[i]);
1222 state->frame[i] = NULL;
1224 clear_jmp_history(state);
1229 /* copy verifier state from src to dst growing dst stack space
1230 * when necessary to accommodate larger src stack
1232 static int copy_func_state(struct bpf_func_state *dst,
1233 const struct bpf_func_state *src)
1237 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
1238 err = copy_reference_state(dst, src);
1241 return copy_stack_state(dst, src);
1244 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
1245 const struct bpf_verifier_state *src)
1247 struct bpf_func_state *dst;
1250 dst_state->jmp_history = copy_array(dst_state->jmp_history, src->jmp_history,
1251 src->jmp_history_cnt, sizeof(struct bpf_idx_pair),
1253 if (!dst_state->jmp_history)
1255 dst_state->jmp_history_cnt = src->jmp_history_cnt;
1257 /* if dst has more stack frames then src frame, free them */
1258 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
1259 free_func_state(dst_state->frame[i]);
1260 dst_state->frame[i] = NULL;
1262 dst_state->speculative = src->speculative;
1263 dst_state->active_rcu_lock = src->active_rcu_lock;
1264 dst_state->curframe = src->curframe;
1265 dst_state->active_lock.ptr = src->active_lock.ptr;
1266 dst_state->active_lock.id = src->active_lock.id;
1267 dst_state->branches = src->branches;
1268 dst_state->parent = src->parent;
1269 dst_state->first_insn_idx = src->first_insn_idx;
1270 dst_state->last_insn_idx = src->last_insn_idx;
1271 for (i = 0; i <= src->curframe; i++) {
1272 dst = dst_state->frame[i];
1274 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
1277 dst_state->frame[i] = dst;
1279 err = copy_func_state(dst, src->frame[i]);
1286 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
1289 u32 br = --st->branches;
1291 /* WARN_ON(br > 1) technically makes sense here,
1292 * but see comment in push_stack(), hence:
1294 WARN_ONCE((int)br < 0,
1295 "BUG update_branch_counts:branches_to_explore=%d\n",
1303 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
1304 int *insn_idx, bool pop_log)
1306 struct bpf_verifier_state *cur = env->cur_state;
1307 struct bpf_verifier_stack_elem *elem, *head = env->head;
1310 if (env->head == NULL)
1314 err = copy_verifier_state(cur, &head->st);
1319 bpf_vlog_reset(&env->log, head->log_pos);
1321 *insn_idx = head->insn_idx;
1323 *prev_insn_idx = head->prev_insn_idx;
1325 free_verifier_state(&head->st, false);
1332 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
1333 int insn_idx, int prev_insn_idx,
1336 struct bpf_verifier_state *cur = env->cur_state;
1337 struct bpf_verifier_stack_elem *elem;
1340 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1344 elem->insn_idx = insn_idx;
1345 elem->prev_insn_idx = prev_insn_idx;
1346 elem->next = env->head;
1347 elem->log_pos = env->log.len_used;
1350 err = copy_verifier_state(&elem->st, cur);
1353 elem->st.speculative |= speculative;
1354 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1355 verbose(env, "The sequence of %d jumps is too complex.\n",
1359 if (elem->st.parent) {
1360 ++elem->st.parent->branches;
1361 /* WARN_ON(branches > 2) technically makes sense here,
1363 * 1. speculative states will bump 'branches' for non-branch
1365 * 2. is_state_visited() heuristics may decide not to create
1366 * a new state for a sequence of branches and all such current
1367 * and cloned states will be pointing to a single parent state
1368 * which might have large 'branches' count.
1373 free_verifier_state(env->cur_state, true);
1374 env->cur_state = NULL;
1375 /* pop all elements and return */
1376 while (!pop_stack(env, NULL, NULL, false));
1380 #define CALLER_SAVED_REGS 6
1381 static const int caller_saved[CALLER_SAVED_REGS] = {
1382 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
1385 /* This helper doesn't clear reg->id */
1386 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1388 reg->var_off = tnum_const(imm);
1389 reg->smin_value = (s64)imm;
1390 reg->smax_value = (s64)imm;
1391 reg->umin_value = imm;
1392 reg->umax_value = imm;
1394 reg->s32_min_value = (s32)imm;
1395 reg->s32_max_value = (s32)imm;
1396 reg->u32_min_value = (u32)imm;
1397 reg->u32_max_value = (u32)imm;
1400 /* Mark the unknown part of a register (variable offset or scalar value) as
1401 * known to have the value @imm.
1403 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1405 /* Clear off and union(map_ptr, range) */
1406 memset(((u8 *)reg) + sizeof(reg->type), 0,
1407 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1409 reg->ref_obj_id = 0;
1410 ___mark_reg_known(reg, imm);
1413 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1415 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1416 reg->s32_min_value = (s32)imm;
1417 reg->s32_max_value = (s32)imm;
1418 reg->u32_min_value = (u32)imm;
1419 reg->u32_max_value = (u32)imm;
1422 /* Mark the 'variable offset' part of a register as zero. This should be
1423 * used only on registers holding a pointer type.
1425 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1427 __mark_reg_known(reg, 0);
1430 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1432 __mark_reg_known(reg, 0);
1433 reg->type = SCALAR_VALUE;
1436 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1437 struct bpf_reg_state *regs, u32 regno)
1439 if (WARN_ON(regno >= MAX_BPF_REG)) {
1440 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1441 /* Something bad happened, let's kill all regs */
1442 for (regno = 0; regno < MAX_BPF_REG; regno++)
1443 __mark_reg_not_init(env, regs + regno);
1446 __mark_reg_known_zero(regs + regno);
1449 static void __mark_dynptr_reg(struct bpf_reg_state *reg, enum bpf_dynptr_type type,
1452 /* reg->type has no meaning for STACK_DYNPTR, but when we set reg for
1453 * callback arguments, it does need to be CONST_PTR_TO_DYNPTR, so simply
1454 * set it unconditionally as it is ignored for STACK_DYNPTR anyway.
1456 __mark_reg_known_zero(reg);
1457 reg->type = CONST_PTR_TO_DYNPTR;
1458 reg->dynptr.type = type;
1459 reg->dynptr.first_slot = first_slot;
1462 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
1464 if (base_type(reg->type) == PTR_TO_MAP_VALUE) {
1465 const struct bpf_map *map = reg->map_ptr;
1467 if (map->inner_map_meta) {
1468 reg->type = CONST_PTR_TO_MAP;
1469 reg->map_ptr = map->inner_map_meta;
1470 /* transfer reg's id which is unique for every map_lookup_elem
1471 * as UID of the inner map.
1473 if (btf_record_has_field(map->inner_map_meta->record, BPF_TIMER))
1474 reg->map_uid = reg->id;
1475 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
1476 reg->type = PTR_TO_XDP_SOCK;
1477 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
1478 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
1479 reg->type = PTR_TO_SOCKET;
1481 reg->type = PTR_TO_MAP_VALUE;
1486 reg->type &= ~PTR_MAYBE_NULL;
1489 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1491 return type_is_pkt_pointer(reg->type);
1494 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
1496 return reg_is_pkt_pointer(reg) ||
1497 reg->type == PTR_TO_PACKET_END;
1500 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
1501 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
1502 enum bpf_reg_type which)
1504 /* The register can already have a range from prior markings.
1505 * This is fine as long as it hasn't been advanced from its
1508 return reg->type == which &&
1511 tnum_equals_const(reg->var_off, 0);
1514 /* Reset the min/max bounds of a register */
1515 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1517 reg->smin_value = S64_MIN;
1518 reg->smax_value = S64_MAX;
1519 reg->umin_value = 0;
1520 reg->umax_value = U64_MAX;
1522 reg->s32_min_value = S32_MIN;
1523 reg->s32_max_value = S32_MAX;
1524 reg->u32_min_value = 0;
1525 reg->u32_max_value = U32_MAX;
1528 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
1530 reg->smin_value = S64_MIN;
1531 reg->smax_value = S64_MAX;
1532 reg->umin_value = 0;
1533 reg->umax_value = U64_MAX;
1536 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
1538 reg->s32_min_value = S32_MIN;
1539 reg->s32_max_value = S32_MAX;
1540 reg->u32_min_value = 0;
1541 reg->u32_max_value = U32_MAX;
1544 static void __update_reg32_bounds(struct bpf_reg_state *reg)
1546 struct tnum var32_off = tnum_subreg(reg->var_off);
1548 /* min signed is max(sign bit) | min(other bits) */
1549 reg->s32_min_value = max_t(s32, reg->s32_min_value,
1550 var32_off.value | (var32_off.mask & S32_MIN));
1551 /* max signed is min(sign bit) | max(other bits) */
1552 reg->s32_max_value = min_t(s32, reg->s32_max_value,
1553 var32_off.value | (var32_off.mask & S32_MAX));
1554 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
1555 reg->u32_max_value = min(reg->u32_max_value,
1556 (u32)(var32_off.value | var32_off.mask));
1559 static void __update_reg64_bounds(struct bpf_reg_state *reg)
1561 /* min signed is max(sign bit) | min(other bits) */
1562 reg->smin_value = max_t(s64, reg->smin_value,
1563 reg->var_off.value | (reg->var_off.mask & S64_MIN));
1564 /* max signed is min(sign bit) | max(other bits) */
1565 reg->smax_value = min_t(s64, reg->smax_value,
1566 reg->var_off.value | (reg->var_off.mask & S64_MAX));
1567 reg->umin_value = max(reg->umin_value, reg->var_off.value);
1568 reg->umax_value = min(reg->umax_value,
1569 reg->var_off.value | reg->var_off.mask);
1572 static void __update_reg_bounds(struct bpf_reg_state *reg)
1574 __update_reg32_bounds(reg);
1575 __update_reg64_bounds(reg);
1578 /* Uses signed min/max values to inform unsigned, and vice-versa */
1579 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
1581 /* Learn sign from signed bounds.
1582 * If we cannot cross the sign boundary, then signed and unsigned bounds
1583 * are the same, so combine. This works even in the negative case, e.g.
1584 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1586 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
1587 reg->s32_min_value = reg->u32_min_value =
1588 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1589 reg->s32_max_value = reg->u32_max_value =
1590 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1593 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1594 * boundary, so we must be careful.
1596 if ((s32)reg->u32_max_value >= 0) {
1597 /* Positive. We can't learn anything from the smin, but smax
1598 * is positive, hence safe.
1600 reg->s32_min_value = reg->u32_min_value;
1601 reg->s32_max_value = reg->u32_max_value =
1602 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1603 } else if ((s32)reg->u32_min_value < 0) {
1604 /* Negative. We can't learn anything from the smax, but smin
1605 * is negative, hence safe.
1607 reg->s32_min_value = reg->u32_min_value =
1608 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1609 reg->s32_max_value = reg->u32_max_value;
1613 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
1615 /* Learn sign from signed bounds.
1616 * If we cannot cross the sign boundary, then signed and unsigned bounds
1617 * are the same, so combine. This works even in the negative case, e.g.
1618 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1620 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1621 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1623 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1627 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1628 * boundary, so we must be careful.
1630 if ((s64)reg->umax_value >= 0) {
1631 /* Positive. We can't learn anything from the smin, but smax
1632 * is positive, hence safe.
1634 reg->smin_value = reg->umin_value;
1635 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1637 } else if ((s64)reg->umin_value < 0) {
1638 /* Negative. We can't learn anything from the smax, but smin
1639 * is negative, hence safe.
1641 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1643 reg->smax_value = reg->umax_value;
1647 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
1649 __reg32_deduce_bounds(reg);
1650 __reg64_deduce_bounds(reg);
1653 /* Attempts to improve var_off based on unsigned min/max information */
1654 static void __reg_bound_offset(struct bpf_reg_state *reg)
1656 struct tnum var64_off = tnum_intersect(reg->var_off,
1657 tnum_range(reg->umin_value,
1659 struct tnum var32_off = tnum_intersect(tnum_subreg(reg->var_off),
1660 tnum_range(reg->u32_min_value,
1661 reg->u32_max_value));
1663 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
1666 static void reg_bounds_sync(struct bpf_reg_state *reg)
1668 /* We might have learned new bounds from the var_off. */
1669 __update_reg_bounds(reg);
1670 /* We might have learned something about the sign bit. */
1671 __reg_deduce_bounds(reg);
1672 /* We might have learned some bits from the bounds. */
1673 __reg_bound_offset(reg);
1674 /* Intersecting with the old var_off might have improved our bounds
1675 * slightly, e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1676 * then new var_off is (0; 0x7f...fc) which improves our umax.
1678 __update_reg_bounds(reg);
1681 static bool __reg32_bound_s64(s32 a)
1683 return a >= 0 && a <= S32_MAX;
1686 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
1688 reg->umin_value = reg->u32_min_value;
1689 reg->umax_value = reg->u32_max_value;
1691 /* Attempt to pull 32-bit signed bounds into 64-bit bounds but must
1692 * be positive otherwise set to worse case bounds and refine later
1695 if (__reg32_bound_s64(reg->s32_min_value) &&
1696 __reg32_bound_s64(reg->s32_max_value)) {
1697 reg->smin_value = reg->s32_min_value;
1698 reg->smax_value = reg->s32_max_value;
1700 reg->smin_value = 0;
1701 reg->smax_value = U32_MAX;
1705 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
1707 /* special case when 64-bit register has upper 32-bit register
1708 * zeroed. Typically happens after zext or <<32, >>32 sequence
1709 * allowing us to use 32-bit bounds directly,
1711 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
1712 __reg_assign_32_into_64(reg);
1714 /* Otherwise the best we can do is push lower 32bit known and
1715 * unknown bits into register (var_off set from jmp logic)
1716 * then learn as much as possible from the 64-bit tnum
1717 * known and unknown bits. The previous smin/smax bounds are
1718 * invalid here because of jmp32 compare so mark them unknown
1719 * so they do not impact tnum bounds calculation.
1721 __mark_reg64_unbounded(reg);
1723 reg_bounds_sync(reg);
1726 static bool __reg64_bound_s32(s64 a)
1728 return a >= S32_MIN && a <= S32_MAX;
1731 static bool __reg64_bound_u32(u64 a)
1733 return a >= U32_MIN && a <= U32_MAX;
1736 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
1738 __mark_reg32_unbounded(reg);
1739 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
1740 reg->s32_min_value = (s32)reg->smin_value;
1741 reg->s32_max_value = (s32)reg->smax_value;
1743 if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
1744 reg->u32_min_value = (u32)reg->umin_value;
1745 reg->u32_max_value = (u32)reg->umax_value;
1747 reg_bounds_sync(reg);
1750 /* Mark a register as having a completely unknown (scalar) value. */
1751 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1752 struct bpf_reg_state *reg)
1755 * Clear type, off, and union(map_ptr, range) and
1756 * padding between 'type' and union
1758 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1759 reg->type = SCALAR_VALUE;
1761 reg->ref_obj_id = 0;
1762 reg->var_off = tnum_unknown;
1764 reg->precise = !env->bpf_capable;
1765 __mark_reg_unbounded(reg);
1768 static void mark_reg_unknown(struct bpf_verifier_env *env,
1769 struct bpf_reg_state *regs, u32 regno)
1771 if (WARN_ON(regno >= MAX_BPF_REG)) {
1772 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1773 /* Something bad happened, let's kill all regs except FP */
1774 for (regno = 0; regno < BPF_REG_FP; regno++)
1775 __mark_reg_not_init(env, regs + regno);
1778 __mark_reg_unknown(env, regs + regno);
1781 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1782 struct bpf_reg_state *reg)
1784 __mark_reg_unknown(env, reg);
1785 reg->type = NOT_INIT;
1788 static void mark_reg_not_init(struct bpf_verifier_env *env,
1789 struct bpf_reg_state *regs, u32 regno)
1791 if (WARN_ON(regno >= MAX_BPF_REG)) {
1792 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1793 /* Something bad happened, let's kill all regs except FP */
1794 for (regno = 0; regno < BPF_REG_FP; regno++)
1795 __mark_reg_not_init(env, regs + regno);
1798 __mark_reg_not_init(env, regs + regno);
1801 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
1802 struct bpf_reg_state *regs, u32 regno,
1803 enum bpf_reg_type reg_type,
1804 struct btf *btf, u32 btf_id,
1805 enum bpf_type_flag flag)
1807 if (reg_type == SCALAR_VALUE) {
1808 mark_reg_unknown(env, regs, regno);
1811 mark_reg_known_zero(env, regs, regno);
1812 regs[regno].type = PTR_TO_BTF_ID | flag;
1813 regs[regno].btf = btf;
1814 regs[regno].btf_id = btf_id;
1817 #define DEF_NOT_SUBREG (0)
1818 static void init_reg_state(struct bpf_verifier_env *env,
1819 struct bpf_func_state *state)
1821 struct bpf_reg_state *regs = state->regs;
1824 for (i = 0; i < MAX_BPF_REG; i++) {
1825 mark_reg_not_init(env, regs, i);
1826 regs[i].live = REG_LIVE_NONE;
1827 regs[i].parent = NULL;
1828 regs[i].subreg_def = DEF_NOT_SUBREG;
1832 regs[BPF_REG_FP].type = PTR_TO_STACK;
1833 mark_reg_known_zero(env, regs, BPF_REG_FP);
1834 regs[BPF_REG_FP].frameno = state->frameno;
1837 #define BPF_MAIN_FUNC (-1)
1838 static void init_func_state(struct bpf_verifier_env *env,
1839 struct bpf_func_state *state,
1840 int callsite, int frameno, int subprogno)
1842 state->callsite = callsite;
1843 state->frameno = frameno;
1844 state->subprogno = subprogno;
1845 state->callback_ret_range = tnum_range(0, 0);
1846 init_reg_state(env, state);
1847 mark_verifier_state_scratched(env);
1850 /* Similar to push_stack(), but for async callbacks */
1851 static struct bpf_verifier_state *push_async_cb(struct bpf_verifier_env *env,
1852 int insn_idx, int prev_insn_idx,
1855 struct bpf_verifier_stack_elem *elem;
1856 struct bpf_func_state *frame;
1858 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1862 elem->insn_idx = insn_idx;
1863 elem->prev_insn_idx = prev_insn_idx;
1864 elem->next = env->head;
1865 elem->log_pos = env->log.len_used;
1868 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1870 "The sequence of %d jumps is too complex for async cb.\n",
1874 /* Unlike push_stack() do not copy_verifier_state().
1875 * The caller state doesn't matter.
1876 * This is async callback. It starts in a fresh stack.
1877 * Initialize it similar to do_check_common().
1879 elem->st.branches = 1;
1880 frame = kzalloc(sizeof(*frame), GFP_KERNEL);
1883 init_func_state(env, frame,
1884 BPF_MAIN_FUNC /* callsite */,
1885 0 /* frameno within this callchain */,
1886 subprog /* subprog number within this prog */);
1887 elem->st.frame[0] = frame;
1890 free_verifier_state(env->cur_state, true);
1891 env->cur_state = NULL;
1892 /* pop all elements and return */
1893 while (!pop_stack(env, NULL, NULL, false));
1899 SRC_OP, /* register is used as source operand */
1900 DST_OP, /* register is used as destination operand */
1901 DST_OP_NO_MARK /* same as above, check only, don't mark */
1904 static int cmp_subprogs(const void *a, const void *b)
1906 return ((struct bpf_subprog_info *)a)->start -
1907 ((struct bpf_subprog_info *)b)->start;
1910 static int find_subprog(struct bpf_verifier_env *env, int off)
1912 struct bpf_subprog_info *p;
1914 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1915 sizeof(env->subprog_info[0]), cmp_subprogs);
1918 return p - env->subprog_info;
1922 static int add_subprog(struct bpf_verifier_env *env, int off)
1924 int insn_cnt = env->prog->len;
1927 if (off >= insn_cnt || off < 0) {
1928 verbose(env, "call to invalid destination\n");
1931 ret = find_subprog(env, off);
1934 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1935 verbose(env, "too many subprograms\n");
1938 /* determine subprog starts. The end is one before the next starts */
1939 env->subprog_info[env->subprog_cnt++].start = off;
1940 sort(env->subprog_info, env->subprog_cnt,
1941 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1942 return env->subprog_cnt - 1;
1945 #define MAX_KFUNC_DESCS 256
1946 #define MAX_KFUNC_BTFS 256
1948 struct bpf_kfunc_desc {
1949 struct btf_func_model func_model;
1955 struct bpf_kfunc_btf {
1957 struct module *module;
1961 struct bpf_kfunc_desc_tab {
1962 struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
1966 struct bpf_kfunc_btf_tab {
1967 struct bpf_kfunc_btf descs[MAX_KFUNC_BTFS];
1971 static int kfunc_desc_cmp_by_id_off(const void *a, const void *b)
1973 const struct bpf_kfunc_desc *d0 = a;
1974 const struct bpf_kfunc_desc *d1 = b;
1976 /* func_id is not greater than BTF_MAX_TYPE */
1977 return d0->func_id - d1->func_id ?: d0->offset - d1->offset;
1980 static int kfunc_btf_cmp_by_off(const void *a, const void *b)
1982 const struct bpf_kfunc_btf *d0 = a;
1983 const struct bpf_kfunc_btf *d1 = b;
1985 return d0->offset - d1->offset;
1988 static const struct bpf_kfunc_desc *
1989 find_kfunc_desc(const struct bpf_prog *prog, u32 func_id, u16 offset)
1991 struct bpf_kfunc_desc desc = {
1995 struct bpf_kfunc_desc_tab *tab;
1997 tab = prog->aux->kfunc_tab;
1998 return bsearch(&desc, tab->descs, tab->nr_descs,
1999 sizeof(tab->descs[0]), kfunc_desc_cmp_by_id_off);
2002 static struct btf *__find_kfunc_desc_btf(struct bpf_verifier_env *env,
2005 struct bpf_kfunc_btf kf_btf = { .offset = offset };
2006 struct bpf_kfunc_btf_tab *tab;
2007 struct bpf_kfunc_btf *b;
2012 tab = env->prog->aux->kfunc_btf_tab;
2013 b = bsearch(&kf_btf, tab->descs, tab->nr_descs,
2014 sizeof(tab->descs[0]), kfunc_btf_cmp_by_off);
2016 if (tab->nr_descs == MAX_KFUNC_BTFS) {
2017 verbose(env, "too many different module BTFs\n");
2018 return ERR_PTR(-E2BIG);
2021 if (bpfptr_is_null(env->fd_array)) {
2022 verbose(env, "kfunc offset > 0 without fd_array is invalid\n");
2023 return ERR_PTR(-EPROTO);
2026 if (copy_from_bpfptr_offset(&btf_fd, env->fd_array,
2027 offset * sizeof(btf_fd),
2029 return ERR_PTR(-EFAULT);
2031 btf = btf_get_by_fd(btf_fd);
2033 verbose(env, "invalid module BTF fd specified\n");
2037 if (!btf_is_module(btf)) {
2038 verbose(env, "BTF fd for kfunc is not a module BTF\n");
2040 return ERR_PTR(-EINVAL);
2043 mod = btf_try_get_module(btf);
2046 return ERR_PTR(-ENXIO);
2049 b = &tab->descs[tab->nr_descs++];
2054 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2055 kfunc_btf_cmp_by_off, NULL);
2060 void bpf_free_kfunc_btf_tab(struct bpf_kfunc_btf_tab *tab)
2065 while (tab->nr_descs--) {
2066 module_put(tab->descs[tab->nr_descs].module);
2067 btf_put(tab->descs[tab->nr_descs].btf);
2072 static struct btf *find_kfunc_desc_btf(struct bpf_verifier_env *env, s16 offset)
2076 /* In the future, this can be allowed to increase limit
2077 * of fd index into fd_array, interpreted as u16.
2079 verbose(env, "negative offset disallowed for kernel module function call\n");
2080 return ERR_PTR(-EINVAL);
2083 return __find_kfunc_desc_btf(env, offset);
2085 return btf_vmlinux ?: ERR_PTR(-ENOENT);
2088 static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id, s16 offset)
2090 const struct btf_type *func, *func_proto;
2091 struct bpf_kfunc_btf_tab *btf_tab;
2092 struct bpf_kfunc_desc_tab *tab;
2093 struct bpf_prog_aux *prog_aux;
2094 struct bpf_kfunc_desc *desc;
2095 const char *func_name;
2096 struct btf *desc_btf;
2097 unsigned long call_imm;
2101 prog_aux = env->prog->aux;
2102 tab = prog_aux->kfunc_tab;
2103 btf_tab = prog_aux->kfunc_btf_tab;
2106 verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
2110 if (!env->prog->jit_requested) {
2111 verbose(env, "JIT is required for calling kernel function\n");
2115 if (!bpf_jit_supports_kfunc_call()) {
2116 verbose(env, "JIT does not support calling kernel function\n");
2120 if (!env->prog->gpl_compatible) {
2121 verbose(env, "cannot call kernel function from non-GPL compatible program\n");
2125 tab = kzalloc(sizeof(*tab), GFP_KERNEL);
2128 prog_aux->kfunc_tab = tab;
2131 /* func_id == 0 is always invalid, but instead of returning an error, be
2132 * conservative and wait until the code elimination pass before returning
2133 * error, so that invalid calls that get pruned out can be in BPF programs
2134 * loaded from userspace. It is also required that offset be untouched
2137 if (!func_id && !offset)
2140 if (!btf_tab && offset) {
2141 btf_tab = kzalloc(sizeof(*btf_tab), GFP_KERNEL);
2144 prog_aux->kfunc_btf_tab = btf_tab;
2147 desc_btf = find_kfunc_desc_btf(env, offset);
2148 if (IS_ERR(desc_btf)) {
2149 verbose(env, "failed to find BTF for kernel function\n");
2150 return PTR_ERR(desc_btf);
2153 if (find_kfunc_desc(env->prog, func_id, offset))
2156 if (tab->nr_descs == MAX_KFUNC_DESCS) {
2157 verbose(env, "too many different kernel function calls\n");
2161 func = btf_type_by_id(desc_btf, func_id);
2162 if (!func || !btf_type_is_func(func)) {
2163 verbose(env, "kernel btf_id %u is not a function\n",
2167 func_proto = btf_type_by_id(desc_btf, func->type);
2168 if (!func_proto || !btf_type_is_func_proto(func_proto)) {
2169 verbose(env, "kernel function btf_id %u does not have a valid func_proto\n",
2174 func_name = btf_name_by_offset(desc_btf, func->name_off);
2175 addr = kallsyms_lookup_name(func_name);
2177 verbose(env, "cannot find address for kernel function %s\n",
2182 call_imm = BPF_CALL_IMM(addr);
2183 /* Check whether or not the relative offset overflows desc->imm */
2184 if ((unsigned long)(s32)call_imm != call_imm) {
2185 verbose(env, "address of kernel function %s is out of range\n",
2190 desc = &tab->descs[tab->nr_descs++];
2191 desc->func_id = func_id;
2192 desc->imm = call_imm;
2193 desc->offset = offset;
2194 err = btf_distill_func_proto(&env->log, desc_btf,
2195 func_proto, func_name,
2198 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2199 kfunc_desc_cmp_by_id_off, NULL);
2203 static int kfunc_desc_cmp_by_imm(const void *a, const void *b)
2205 const struct bpf_kfunc_desc *d0 = a;
2206 const struct bpf_kfunc_desc *d1 = b;
2208 if (d0->imm > d1->imm)
2210 else if (d0->imm < d1->imm)
2215 static void sort_kfunc_descs_by_imm(struct bpf_prog *prog)
2217 struct bpf_kfunc_desc_tab *tab;
2219 tab = prog->aux->kfunc_tab;
2223 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2224 kfunc_desc_cmp_by_imm, NULL);
2227 bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
2229 return !!prog->aux->kfunc_tab;
2232 const struct btf_func_model *
2233 bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
2234 const struct bpf_insn *insn)
2236 const struct bpf_kfunc_desc desc = {
2239 const struct bpf_kfunc_desc *res;
2240 struct bpf_kfunc_desc_tab *tab;
2242 tab = prog->aux->kfunc_tab;
2243 res = bsearch(&desc, tab->descs, tab->nr_descs,
2244 sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm);
2246 return res ? &res->func_model : NULL;
2249 static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
2251 struct bpf_subprog_info *subprog = env->subprog_info;
2252 struct bpf_insn *insn = env->prog->insnsi;
2253 int i, ret, insn_cnt = env->prog->len;
2255 /* Add entry function. */
2256 ret = add_subprog(env, 0);
2260 for (i = 0; i < insn_cnt; i++, insn++) {
2261 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
2262 !bpf_pseudo_kfunc_call(insn))
2265 if (!env->bpf_capable) {
2266 verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
2270 if (bpf_pseudo_func(insn) || bpf_pseudo_call(insn))
2271 ret = add_subprog(env, i + insn->imm + 1);
2273 ret = add_kfunc_call(env, insn->imm, insn->off);
2279 /* Add a fake 'exit' subprog which could simplify subprog iteration
2280 * logic. 'subprog_cnt' should not be increased.
2282 subprog[env->subprog_cnt].start = insn_cnt;
2284 if (env->log.level & BPF_LOG_LEVEL2)
2285 for (i = 0; i < env->subprog_cnt; i++)
2286 verbose(env, "func#%d @%d\n", i, subprog[i].start);
2291 static int check_subprogs(struct bpf_verifier_env *env)
2293 int i, subprog_start, subprog_end, off, cur_subprog = 0;
2294 struct bpf_subprog_info *subprog = env->subprog_info;
2295 struct bpf_insn *insn = env->prog->insnsi;
2296 int insn_cnt = env->prog->len;
2298 /* now check that all jumps are within the same subprog */
2299 subprog_start = subprog[cur_subprog].start;
2300 subprog_end = subprog[cur_subprog + 1].start;
2301 for (i = 0; i < insn_cnt; i++) {
2302 u8 code = insn[i].code;
2304 if (code == (BPF_JMP | BPF_CALL) &&
2305 insn[i].imm == BPF_FUNC_tail_call &&
2306 insn[i].src_reg != BPF_PSEUDO_CALL)
2307 subprog[cur_subprog].has_tail_call = true;
2308 if (BPF_CLASS(code) == BPF_LD &&
2309 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
2310 subprog[cur_subprog].has_ld_abs = true;
2311 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
2313 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
2315 off = i + insn[i].off + 1;
2316 if (off < subprog_start || off >= subprog_end) {
2317 verbose(env, "jump out of range from insn %d to %d\n", i, off);
2321 if (i == subprog_end - 1) {
2322 /* to avoid fall-through from one subprog into another
2323 * the last insn of the subprog should be either exit
2324 * or unconditional jump back
2326 if (code != (BPF_JMP | BPF_EXIT) &&
2327 code != (BPF_JMP | BPF_JA)) {
2328 verbose(env, "last insn is not an exit or jmp\n");
2331 subprog_start = subprog_end;
2333 if (cur_subprog < env->subprog_cnt)
2334 subprog_end = subprog[cur_subprog + 1].start;
2340 /* Parentage chain of this register (or stack slot) should take care of all
2341 * issues like callee-saved registers, stack slot allocation time, etc.
2343 static int mark_reg_read(struct bpf_verifier_env *env,
2344 const struct bpf_reg_state *state,
2345 struct bpf_reg_state *parent, u8 flag)
2347 bool writes = parent == state->parent; /* Observe write marks */
2351 /* if read wasn't screened by an earlier write ... */
2352 if (writes && state->live & REG_LIVE_WRITTEN)
2354 if (parent->live & REG_LIVE_DONE) {
2355 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
2356 reg_type_str(env, parent->type),
2357 parent->var_off.value, parent->off);
2360 /* The first condition is more likely to be true than the
2361 * second, checked it first.
2363 if ((parent->live & REG_LIVE_READ) == flag ||
2364 parent->live & REG_LIVE_READ64)
2365 /* The parentage chain never changes and
2366 * this parent was already marked as LIVE_READ.
2367 * There is no need to keep walking the chain again and
2368 * keep re-marking all parents as LIVE_READ.
2369 * This case happens when the same register is read
2370 * multiple times without writes into it in-between.
2371 * Also, if parent has the stronger REG_LIVE_READ64 set,
2372 * then no need to set the weak REG_LIVE_READ32.
2375 /* ... then we depend on parent's value */
2376 parent->live |= flag;
2377 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
2378 if (flag == REG_LIVE_READ64)
2379 parent->live &= ~REG_LIVE_READ32;
2381 parent = state->parent;
2386 if (env->longest_mark_read_walk < cnt)
2387 env->longest_mark_read_walk = cnt;
2391 /* This function is supposed to be used by the following 32-bit optimization
2392 * code only. It returns TRUE if the source or destination register operates
2393 * on 64-bit, otherwise return FALSE.
2395 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
2396 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
2401 class = BPF_CLASS(code);
2403 if (class == BPF_JMP) {
2404 /* BPF_EXIT for "main" will reach here. Return TRUE
2409 if (op == BPF_CALL) {
2410 /* BPF to BPF call will reach here because of marking
2411 * caller saved clobber with DST_OP_NO_MARK for which we
2412 * don't care the register def because they are anyway
2413 * marked as NOT_INIT already.
2415 if (insn->src_reg == BPF_PSEUDO_CALL)
2417 /* Helper call will reach here because of arg type
2418 * check, conservatively return TRUE.
2427 if (class == BPF_ALU64 || class == BPF_JMP ||
2428 /* BPF_END always use BPF_ALU class. */
2429 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
2432 if (class == BPF_ALU || class == BPF_JMP32)
2435 if (class == BPF_LDX) {
2437 return BPF_SIZE(code) == BPF_DW;
2438 /* LDX source must be ptr. */
2442 if (class == BPF_STX) {
2443 /* BPF_STX (including atomic variants) has multiple source
2444 * operands, one of which is a ptr. Check whether the caller is
2447 if (t == SRC_OP && reg->type != SCALAR_VALUE)
2449 return BPF_SIZE(code) == BPF_DW;
2452 if (class == BPF_LD) {
2453 u8 mode = BPF_MODE(code);
2456 if (mode == BPF_IMM)
2459 /* Both LD_IND and LD_ABS return 32-bit data. */
2463 /* Implicit ctx ptr. */
2464 if (regno == BPF_REG_6)
2467 /* Explicit source could be any width. */
2471 if (class == BPF_ST)
2472 /* The only source register for BPF_ST is a ptr. */
2475 /* Conservatively return true at default. */
2479 /* Return the regno defined by the insn, or -1. */
2480 static int insn_def_regno(const struct bpf_insn *insn)
2482 switch (BPF_CLASS(insn->code)) {
2488 if (BPF_MODE(insn->code) == BPF_ATOMIC &&
2489 (insn->imm & BPF_FETCH)) {
2490 if (insn->imm == BPF_CMPXCHG)
2493 return insn->src_reg;
2498 return insn->dst_reg;
2502 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
2503 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
2505 int dst_reg = insn_def_regno(insn);
2510 return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
2513 static void mark_insn_zext(struct bpf_verifier_env *env,
2514 struct bpf_reg_state *reg)
2516 s32 def_idx = reg->subreg_def;
2518 if (def_idx == DEF_NOT_SUBREG)
2521 env->insn_aux_data[def_idx - 1].zext_dst = true;
2522 /* The dst will be zero extended, so won't be sub-register anymore. */
2523 reg->subreg_def = DEF_NOT_SUBREG;
2526 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
2527 enum reg_arg_type t)
2529 struct bpf_verifier_state *vstate = env->cur_state;
2530 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2531 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
2532 struct bpf_reg_state *reg, *regs = state->regs;
2535 if (regno >= MAX_BPF_REG) {
2536 verbose(env, "R%d is invalid\n", regno);
2540 mark_reg_scratched(env, regno);
2543 rw64 = is_reg64(env, insn, regno, reg, t);
2545 /* check whether register used as source operand can be read */
2546 if (reg->type == NOT_INIT) {
2547 verbose(env, "R%d !read_ok\n", regno);
2550 /* We don't need to worry about FP liveness because it's read-only */
2551 if (regno == BPF_REG_FP)
2555 mark_insn_zext(env, reg);
2557 return mark_reg_read(env, reg, reg->parent,
2558 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
2560 /* check whether register used as dest operand can be written to */
2561 if (regno == BPF_REG_FP) {
2562 verbose(env, "frame pointer is read only\n");
2565 reg->live |= REG_LIVE_WRITTEN;
2566 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
2568 mark_reg_unknown(env, regs, regno);
2573 static void mark_jmp_point(struct bpf_verifier_env *env, int idx)
2575 env->insn_aux_data[idx].jmp_point = true;
2578 static bool is_jmp_point(struct bpf_verifier_env *env, int insn_idx)
2580 return env->insn_aux_data[insn_idx].jmp_point;
2583 /* for any branch, call, exit record the history of jmps in the given state */
2584 static int push_jmp_history(struct bpf_verifier_env *env,
2585 struct bpf_verifier_state *cur)
2587 u32 cnt = cur->jmp_history_cnt;
2588 struct bpf_idx_pair *p;
2591 if (!is_jmp_point(env, env->insn_idx))
2595 alloc_size = kmalloc_size_roundup(size_mul(cnt, sizeof(*p)));
2596 p = krealloc(cur->jmp_history, alloc_size, GFP_USER);
2599 p[cnt - 1].idx = env->insn_idx;
2600 p[cnt - 1].prev_idx = env->prev_insn_idx;
2601 cur->jmp_history = p;
2602 cur->jmp_history_cnt = cnt;
2606 /* Backtrack one insn at a time. If idx is not at the top of recorded
2607 * history then previous instruction came from straight line execution.
2609 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
2614 if (cnt && st->jmp_history[cnt - 1].idx == i) {
2615 i = st->jmp_history[cnt - 1].prev_idx;
2623 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
2625 const struct btf_type *func;
2626 struct btf *desc_btf;
2628 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
2631 desc_btf = find_kfunc_desc_btf(data, insn->off);
2632 if (IS_ERR(desc_btf))
2635 func = btf_type_by_id(desc_btf, insn->imm);
2636 return btf_name_by_offset(desc_btf, func->name_off);
2639 /* For given verifier state backtrack_insn() is called from the last insn to
2640 * the first insn. Its purpose is to compute a bitmask of registers and
2641 * stack slots that needs precision in the parent verifier state.
2643 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
2644 u32 *reg_mask, u64 *stack_mask)
2646 const struct bpf_insn_cbs cbs = {
2647 .cb_call = disasm_kfunc_name,
2648 .cb_print = verbose,
2649 .private_data = env,
2651 struct bpf_insn *insn = env->prog->insnsi + idx;
2652 u8 class = BPF_CLASS(insn->code);
2653 u8 opcode = BPF_OP(insn->code);
2654 u8 mode = BPF_MODE(insn->code);
2655 u32 dreg = 1u << insn->dst_reg;
2656 u32 sreg = 1u << insn->src_reg;
2659 if (insn->code == 0)
2661 if (env->log.level & BPF_LOG_LEVEL2) {
2662 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
2663 verbose(env, "%d: ", idx);
2664 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
2667 if (class == BPF_ALU || class == BPF_ALU64) {
2668 if (!(*reg_mask & dreg))
2670 if (opcode == BPF_MOV) {
2671 if (BPF_SRC(insn->code) == BPF_X) {
2673 * dreg needs precision after this insn
2674 * sreg needs precision before this insn
2680 * dreg needs precision after this insn.
2681 * Corresponding register is already marked
2682 * as precise=true in this verifier state.
2683 * No further markings in parent are necessary
2688 if (BPF_SRC(insn->code) == BPF_X) {
2690 * both dreg and sreg need precision
2695 * dreg still needs precision before this insn
2698 } else if (class == BPF_LDX) {
2699 if (!(*reg_mask & dreg))
2703 /* scalars can only be spilled into stack w/o losing precision.
2704 * Load from any other memory can be zero extended.
2705 * The desire to keep that precision is already indicated
2706 * by 'precise' mark in corresponding register of this state.
2707 * No further tracking necessary.
2709 if (insn->src_reg != BPF_REG_FP)
2712 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
2713 * that [fp - off] slot contains scalar that needs to be
2714 * tracked with precision
2716 spi = (-insn->off - 1) / BPF_REG_SIZE;
2718 verbose(env, "BUG spi %d\n", spi);
2719 WARN_ONCE(1, "verifier backtracking bug");
2722 *stack_mask |= 1ull << spi;
2723 } else if (class == BPF_STX || class == BPF_ST) {
2724 if (*reg_mask & dreg)
2725 /* stx & st shouldn't be using _scalar_ dst_reg
2726 * to access memory. It means backtracking
2727 * encountered a case of pointer subtraction.
2730 /* scalars can only be spilled into stack */
2731 if (insn->dst_reg != BPF_REG_FP)
2733 spi = (-insn->off - 1) / BPF_REG_SIZE;
2735 verbose(env, "BUG spi %d\n", spi);
2736 WARN_ONCE(1, "verifier backtracking bug");
2739 if (!(*stack_mask & (1ull << spi)))
2741 *stack_mask &= ~(1ull << spi);
2742 if (class == BPF_STX)
2744 } else if (class == BPF_JMP || class == BPF_JMP32) {
2745 if (opcode == BPF_CALL) {
2746 if (insn->src_reg == BPF_PSEUDO_CALL)
2748 /* BPF helpers that invoke callback subprogs are
2749 * equivalent to BPF_PSEUDO_CALL above
2751 if (insn->src_reg == 0 && is_callback_calling_function(insn->imm))
2753 /* regular helper call sets R0 */
2755 if (*reg_mask & 0x3f) {
2756 /* if backtracing was looking for registers R1-R5
2757 * they should have been found already.
2759 verbose(env, "BUG regs %x\n", *reg_mask);
2760 WARN_ONCE(1, "verifier backtracking bug");
2763 } else if (opcode == BPF_EXIT) {
2766 } else if (class == BPF_LD) {
2767 if (!(*reg_mask & dreg))
2770 /* It's ld_imm64 or ld_abs or ld_ind.
2771 * For ld_imm64 no further tracking of precision
2772 * into parent is necessary
2774 if (mode == BPF_IND || mode == BPF_ABS)
2775 /* to be analyzed */
2781 /* the scalar precision tracking algorithm:
2782 * . at the start all registers have precise=false.
2783 * . scalar ranges are tracked as normal through alu and jmp insns.
2784 * . once precise value of the scalar register is used in:
2785 * . ptr + scalar alu
2786 * . if (scalar cond K|scalar)
2787 * . helper_call(.., scalar, ...) where ARG_CONST is expected
2788 * backtrack through the verifier states and mark all registers and
2789 * stack slots with spilled constants that these scalar regisers
2790 * should be precise.
2791 * . during state pruning two registers (or spilled stack slots)
2792 * are equivalent if both are not precise.
2794 * Note the verifier cannot simply walk register parentage chain,
2795 * since many different registers and stack slots could have been
2796 * used to compute single precise scalar.
2798 * The approach of starting with precise=true for all registers and then
2799 * backtrack to mark a register as not precise when the verifier detects
2800 * that program doesn't care about specific value (e.g., when helper
2801 * takes register as ARG_ANYTHING parameter) is not safe.
2803 * It's ok to walk single parentage chain of the verifier states.
2804 * It's possible that this backtracking will go all the way till 1st insn.
2805 * All other branches will be explored for needing precision later.
2807 * The backtracking needs to deal with cases like:
2808 * 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)
2811 * if r5 > 0x79f goto pc+7
2812 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
2815 * call bpf_perf_event_output#25
2816 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
2820 * call foo // uses callee's r6 inside to compute r0
2824 * to track above reg_mask/stack_mask needs to be independent for each frame.
2826 * Also if parent's curframe > frame where backtracking started,
2827 * the verifier need to mark registers in both frames, otherwise callees
2828 * may incorrectly prune callers. This is similar to
2829 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
2831 * For now backtracking falls back into conservative marking.
2833 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
2834 struct bpf_verifier_state *st)
2836 struct bpf_func_state *func;
2837 struct bpf_reg_state *reg;
2840 /* big hammer: mark all scalars precise in this path.
2841 * pop_stack may still get !precise scalars.
2842 * We also skip current state and go straight to first parent state,
2843 * because precision markings in current non-checkpointed state are
2844 * not needed. See why in the comment in __mark_chain_precision below.
2846 for (st = st->parent; st; st = st->parent) {
2847 for (i = 0; i <= st->curframe; i++) {
2848 func = st->frame[i];
2849 for (j = 0; j < BPF_REG_FP; j++) {
2850 reg = &func->regs[j];
2851 if (reg->type != SCALAR_VALUE)
2853 reg->precise = true;
2855 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
2856 if (!is_spilled_reg(&func->stack[j]))
2858 reg = &func->stack[j].spilled_ptr;
2859 if (reg->type != SCALAR_VALUE)
2861 reg->precise = true;
2867 static void mark_all_scalars_imprecise(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
2869 struct bpf_func_state *func;
2870 struct bpf_reg_state *reg;
2873 for (i = 0; i <= st->curframe; i++) {
2874 func = st->frame[i];
2875 for (j = 0; j < BPF_REG_FP; j++) {
2876 reg = &func->regs[j];
2877 if (reg->type != SCALAR_VALUE)
2879 reg->precise = false;
2881 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
2882 if (!is_spilled_reg(&func->stack[j]))
2884 reg = &func->stack[j].spilled_ptr;
2885 if (reg->type != SCALAR_VALUE)
2887 reg->precise = false;
2893 * __mark_chain_precision() backtracks BPF program instruction sequence and
2894 * chain of verifier states making sure that register *regno* (if regno >= 0)
2895 * and/or stack slot *spi* (if spi >= 0) are marked as precisely tracked
2896 * SCALARS, as well as any other registers and slots that contribute to
2897 * a tracked state of given registers/stack slots, depending on specific BPF
2898 * assembly instructions (see backtrack_insns() for exact instruction handling
2899 * logic). This backtracking relies on recorded jmp_history and is able to
2900 * traverse entire chain of parent states. This process ends only when all the
2901 * necessary registers/slots and their transitive dependencies are marked as
2904 * One important and subtle aspect is that precise marks *do not matter* in
2905 * the currently verified state (current state). It is important to understand
2906 * why this is the case.
2908 * First, note that current state is the state that is not yet "checkpointed",
2909 * i.e., it is not yet put into env->explored_states, and it has no children
2910 * states as well. It's ephemeral, and can end up either a) being discarded if
2911 * compatible explored state is found at some point or BPF_EXIT instruction is
2912 * reached or b) checkpointed and put into env->explored_states, branching out
2913 * into one or more children states.
2915 * In the former case, precise markings in current state are completely
2916 * ignored by state comparison code (see regsafe() for details). Only
2917 * checkpointed ("old") state precise markings are important, and if old
2918 * state's register/slot is precise, regsafe() assumes current state's
2919 * register/slot as precise and checks value ranges exactly and precisely. If
2920 * states turn out to be compatible, current state's necessary precise
2921 * markings and any required parent states' precise markings are enforced
2922 * after the fact with propagate_precision() logic, after the fact. But it's
2923 * important to realize that in this case, even after marking current state
2924 * registers/slots as precise, we immediately discard current state. So what
2925 * actually matters is any of the precise markings propagated into current
2926 * state's parent states, which are always checkpointed (due to b) case above).
2927 * As such, for scenario a) it doesn't matter if current state has precise
2928 * markings set or not.
2930 * Now, for the scenario b), checkpointing and forking into child(ren)
2931 * state(s). Note that before current state gets to checkpointing step, any
2932 * processed instruction always assumes precise SCALAR register/slot
2933 * knowledge: if precise value or range is useful to prune jump branch, BPF
2934 * verifier takes this opportunity enthusiastically. Similarly, when
2935 * register's value is used to calculate offset or memory address, exact
2936 * knowledge of SCALAR range is assumed, checked, and enforced. So, similar to
2937 * what we mentioned above about state comparison ignoring precise markings
2938 * during state comparison, BPF verifier ignores and also assumes precise
2939 * markings *at will* during instruction verification process. But as verifier
2940 * assumes precision, it also propagates any precision dependencies across
2941 * parent states, which are not yet finalized, so can be further restricted
2942 * based on new knowledge gained from restrictions enforced by their children
2943 * states. This is so that once those parent states are finalized, i.e., when
2944 * they have no more active children state, state comparison logic in
2945 * is_state_visited() would enforce strict and precise SCALAR ranges, if
2946 * required for correctness.
2948 * To build a bit more intuition, note also that once a state is checkpointed,
2949 * the path we took to get to that state is not important. This is crucial
2950 * property for state pruning. When state is checkpointed and finalized at
2951 * some instruction index, it can be correctly and safely used to "short
2952 * circuit" any *compatible* state that reaches exactly the same instruction
2953 * index. I.e., if we jumped to that instruction from a completely different
2954 * code path than original finalized state was derived from, it doesn't
2955 * matter, current state can be discarded because from that instruction
2956 * forward having a compatible state will ensure we will safely reach the
2957 * exit. States describe preconditions for further exploration, but completely
2958 * forget the history of how we got here.
2960 * This also means that even if we needed precise SCALAR range to get to
2961 * finalized state, but from that point forward *that same* SCALAR register is
2962 * never used in a precise context (i.e., it's precise value is not needed for
2963 * correctness), it's correct and safe to mark such register as "imprecise"
2964 * (i.e., precise marking set to false). This is what we rely on when we do
2965 * not set precise marking in current state. If no child state requires
2966 * precision for any given SCALAR register, it's safe to dictate that it can
2967 * be imprecise. If any child state does require this register to be precise,
2968 * we'll mark it precise later retroactively during precise markings
2969 * propagation from child state to parent states.
2971 * Skipping precise marking setting in current state is a mild version of
2972 * relying on the above observation. But we can utilize this property even
2973 * more aggressively by proactively forgetting any precise marking in the
2974 * current state (which we inherited from the parent state), right before we
2975 * checkpoint it and branch off into new child state. This is done by
2976 * mark_all_scalars_imprecise() to hopefully get more permissive and generic
2977 * finalized states which help in short circuiting more future states.
2979 static int __mark_chain_precision(struct bpf_verifier_env *env, int frame, int regno,
2982 struct bpf_verifier_state *st = env->cur_state;
2983 int first_idx = st->first_insn_idx;
2984 int last_idx = env->insn_idx;
2985 struct bpf_func_state *func;
2986 struct bpf_reg_state *reg;
2987 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
2988 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
2989 bool skip_first = true;
2990 bool new_marks = false;
2993 if (!env->bpf_capable)
2996 /* Do sanity checks against current state of register and/or stack
2997 * slot, but don't set precise flag in current state, as precision
2998 * tracking in the current state is unnecessary.
3000 func = st->frame[frame];
3002 reg = &func->regs[regno];
3003 if (reg->type != SCALAR_VALUE) {
3004 WARN_ONCE(1, "backtracing misuse");
3011 if (!is_spilled_reg(&func->stack[spi])) {
3015 reg = &func->stack[spi].spilled_ptr;
3016 if (reg->type != SCALAR_VALUE) {
3026 if (!reg_mask && !stack_mask)
3030 DECLARE_BITMAP(mask, 64);
3031 u32 history = st->jmp_history_cnt;
3033 if (env->log.level & BPF_LOG_LEVEL2)
3034 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
3037 /* we are at the entry into subprog, which
3038 * is expected for global funcs, but only if
3039 * requested precise registers are R1-R5
3040 * (which are global func's input arguments)
3042 if (st->curframe == 0 &&
3043 st->frame[0]->subprogno > 0 &&
3044 st->frame[0]->callsite == BPF_MAIN_FUNC &&
3045 stack_mask == 0 && (reg_mask & ~0x3e) == 0) {
3046 bitmap_from_u64(mask, reg_mask);
3047 for_each_set_bit(i, mask, 32) {
3048 reg = &st->frame[0]->regs[i];
3049 if (reg->type != SCALAR_VALUE) {
3050 reg_mask &= ~(1u << i);
3053 reg->precise = true;
3058 verbose(env, "BUG backtracing func entry subprog %d reg_mask %x stack_mask %llx\n",
3059 st->frame[0]->subprogno, reg_mask, stack_mask);
3060 WARN_ONCE(1, "verifier backtracking bug");
3064 for (i = last_idx;;) {
3069 err = backtrack_insn(env, i, ®_mask, &stack_mask);
3071 if (err == -ENOTSUPP) {
3072 mark_all_scalars_precise(env, st);
3077 if (!reg_mask && !stack_mask)
3078 /* Found assignment(s) into tracked register in this state.
3079 * Since this state is already marked, just return.
3080 * Nothing to be tracked further in the parent state.
3085 i = get_prev_insn_idx(st, i, &history);
3086 if (i >= env->prog->len) {
3087 /* This can happen if backtracking reached insn 0
3088 * and there are still reg_mask or stack_mask
3090 * It means the backtracking missed the spot where
3091 * particular register was initialized with a constant.
3093 verbose(env, "BUG backtracking idx %d\n", i);
3094 WARN_ONCE(1, "verifier backtracking bug");
3103 func = st->frame[frame];
3104 bitmap_from_u64(mask, reg_mask);
3105 for_each_set_bit(i, mask, 32) {
3106 reg = &func->regs[i];
3107 if (reg->type != SCALAR_VALUE) {
3108 reg_mask &= ~(1u << i);
3113 reg->precise = true;
3116 bitmap_from_u64(mask, stack_mask);
3117 for_each_set_bit(i, mask, 64) {
3118 if (i >= func->allocated_stack / BPF_REG_SIZE) {
3119 /* the sequence of instructions:
3121 * 3: (7b) *(u64 *)(r3 -8) = r0
3122 * 4: (79) r4 = *(u64 *)(r10 -8)
3123 * doesn't contain jmps. It's backtracked
3124 * as a single block.
3125 * During backtracking insn 3 is not recognized as
3126 * stack access, so at the end of backtracking
3127 * stack slot fp-8 is still marked in stack_mask.
3128 * However the parent state may not have accessed
3129 * fp-8 and it's "unallocated" stack space.
3130 * In such case fallback to conservative.
3132 mark_all_scalars_precise(env, st);
3136 if (!is_spilled_reg(&func->stack[i])) {
3137 stack_mask &= ~(1ull << i);
3140 reg = &func->stack[i].spilled_ptr;
3141 if (reg->type != SCALAR_VALUE) {
3142 stack_mask &= ~(1ull << i);
3147 reg->precise = true;
3149 if (env->log.level & BPF_LOG_LEVEL2) {
3150 verbose(env, "parent %s regs=%x stack=%llx marks:",
3151 new_marks ? "didn't have" : "already had",
3152 reg_mask, stack_mask);
3153 print_verifier_state(env, func, true);
3156 if (!reg_mask && !stack_mask)
3161 last_idx = st->last_insn_idx;
3162 first_idx = st->first_insn_idx;
3167 int mark_chain_precision(struct bpf_verifier_env *env, int regno)
3169 return __mark_chain_precision(env, env->cur_state->curframe, regno, -1);
3172 static int mark_chain_precision_frame(struct bpf_verifier_env *env, int frame, int regno)
3174 return __mark_chain_precision(env, frame, regno, -1);
3177 static int mark_chain_precision_stack_frame(struct bpf_verifier_env *env, int frame, int spi)
3179 return __mark_chain_precision(env, frame, -1, spi);
3182 static bool is_spillable_regtype(enum bpf_reg_type type)
3184 switch (base_type(type)) {
3185 case PTR_TO_MAP_VALUE:
3189 case PTR_TO_PACKET_META:
3190 case PTR_TO_PACKET_END:
3191 case PTR_TO_FLOW_KEYS:
3192 case CONST_PTR_TO_MAP:
3194 case PTR_TO_SOCK_COMMON:
3195 case PTR_TO_TCP_SOCK:
3196 case PTR_TO_XDP_SOCK:
3201 case PTR_TO_MAP_KEY:
3208 /* Does this register contain a constant zero? */
3209 static bool register_is_null(struct bpf_reg_state *reg)
3211 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
3214 static bool register_is_const(struct bpf_reg_state *reg)
3216 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
3219 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
3221 return tnum_is_unknown(reg->var_off) &&
3222 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
3223 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
3224 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
3225 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
3228 static bool register_is_bounded(struct bpf_reg_state *reg)
3230 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
3233 static bool __is_pointer_value(bool allow_ptr_leaks,
3234 const struct bpf_reg_state *reg)
3236 if (allow_ptr_leaks)
3239 return reg->type != SCALAR_VALUE;
3242 static void save_register_state(struct bpf_func_state *state,
3243 int spi, struct bpf_reg_state *reg,
3248 state->stack[spi].spilled_ptr = *reg;
3249 if (size == BPF_REG_SIZE)
3250 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
3252 for (i = BPF_REG_SIZE; i > BPF_REG_SIZE - size; i--)
3253 state->stack[spi].slot_type[i - 1] = STACK_SPILL;
3255 /* size < 8 bytes spill */
3257 scrub_spilled_slot(&state->stack[spi].slot_type[i - 1]);
3260 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
3261 * stack boundary and alignment are checked in check_mem_access()
3263 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
3264 /* stack frame we're writing to */
3265 struct bpf_func_state *state,
3266 int off, int size, int value_regno,
3269 struct bpf_func_state *cur; /* state of the current function */
3270 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
3271 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
3272 struct bpf_reg_state *reg = NULL;
3274 err = grow_stack_state(state, round_up(slot + 1, BPF_REG_SIZE));
3277 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
3278 * so it's aligned access and [off, off + size) are within stack limits
3280 if (!env->allow_ptr_leaks &&
3281 state->stack[spi].slot_type[0] == STACK_SPILL &&
3282 size != BPF_REG_SIZE) {
3283 verbose(env, "attempt to corrupt spilled pointer on stack\n");
3287 cur = env->cur_state->frame[env->cur_state->curframe];
3288 if (value_regno >= 0)
3289 reg = &cur->regs[value_regno];
3290 if (!env->bypass_spec_v4) {
3291 bool sanitize = reg && is_spillable_regtype(reg->type);
3293 for (i = 0; i < size; i++) {
3294 if (state->stack[spi].slot_type[i] == STACK_INVALID) {
3301 env->insn_aux_data[insn_idx].sanitize_stack_spill = true;
3304 mark_stack_slot_scratched(env, spi);
3305 if (reg && !(off % BPF_REG_SIZE) && register_is_bounded(reg) &&
3306 !register_is_null(reg) && env->bpf_capable) {
3307 if (dst_reg != BPF_REG_FP) {
3308 /* The backtracking logic can only recognize explicit
3309 * stack slot address like [fp - 8]. Other spill of
3310 * scalar via different register has to be conservative.
3311 * Backtrack from here and mark all registers as precise
3312 * that contributed into 'reg' being a constant.
3314 err = mark_chain_precision(env, value_regno);
3318 save_register_state(state, spi, reg, size);
3319 } else if (reg && is_spillable_regtype(reg->type)) {
3320 /* register containing pointer is being spilled into stack */
3321 if (size != BPF_REG_SIZE) {
3322 verbose_linfo(env, insn_idx, "; ");
3323 verbose(env, "invalid size of register spill\n");
3326 if (state != cur && reg->type == PTR_TO_STACK) {
3327 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
3330 save_register_state(state, spi, reg, size);
3332 u8 type = STACK_MISC;
3334 /* regular write of data into stack destroys any spilled ptr */
3335 state->stack[spi].spilled_ptr.type = NOT_INIT;
3336 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
3337 if (is_spilled_reg(&state->stack[spi]))
3338 for (i = 0; i < BPF_REG_SIZE; i++)
3339 scrub_spilled_slot(&state->stack[spi].slot_type[i]);
3341 /* only mark the slot as written if all 8 bytes were written
3342 * otherwise read propagation may incorrectly stop too soon
3343 * when stack slots are partially written.
3344 * This heuristic means that read propagation will be
3345 * conservative, since it will add reg_live_read marks
3346 * to stack slots all the way to first state when programs
3347 * writes+reads less than 8 bytes
3349 if (size == BPF_REG_SIZE)
3350 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
3352 /* when we zero initialize stack slots mark them as such */
3353 if (reg && register_is_null(reg)) {
3354 /* backtracking doesn't work for STACK_ZERO yet. */
3355 err = mark_chain_precision(env, value_regno);
3361 /* Mark slots affected by this stack write. */
3362 for (i = 0; i < size; i++)
3363 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
3369 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
3370 * known to contain a variable offset.
3371 * This function checks whether the write is permitted and conservatively
3372 * tracks the effects of the write, considering that each stack slot in the
3373 * dynamic range is potentially written to.
3375 * 'off' includes 'regno->off'.
3376 * 'value_regno' can be -1, meaning that an unknown value is being written to
3379 * Spilled pointers in range are not marked as written because we don't know
3380 * what's going to be actually written. This means that read propagation for
3381 * future reads cannot be terminated by this write.
3383 * For privileged programs, uninitialized stack slots are considered
3384 * initialized by this write (even though we don't know exactly what offsets
3385 * are going to be written to). The idea is that we don't want the verifier to
3386 * reject future reads that access slots written to through variable offsets.
3388 static int check_stack_write_var_off(struct bpf_verifier_env *env,
3389 /* func where register points to */
3390 struct bpf_func_state *state,
3391 int ptr_regno, int off, int size,
3392 int value_regno, int insn_idx)
3394 struct bpf_func_state *cur; /* state of the current function */
3395 int min_off, max_off;
3397 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
3398 bool writing_zero = false;
3399 /* set if the fact that we're writing a zero is used to let any
3400 * stack slots remain STACK_ZERO
3402 bool zero_used = false;
3404 cur = env->cur_state->frame[env->cur_state->curframe];
3405 ptr_reg = &cur->regs[ptr_regno];
3406 min_off = ptr_reg->smin_value + off;
3407 max_off = ptr_reg->smax_value + off + size;
3408 if (value_regno >= 0)
3409 value_reg = &cur->regs[value_regno];
3410 if (value_reg && register_is_null(value_reg))
3411 writing_zero = true;
3413 err = grow_stack_state(state, round_up(-min_off, BPF_REG_SIZE));
3418 /* Variable offset writes destroy any spilled pointers in range. */
3419 for (i = min_off; i < max_off; i++) {
3420 u8 new_type, *stype;
3424 spi = slot / BPF_REG_SIZE;
3425 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
3426 mark_stack_slot_scratched(env, spi);
3428 if (!env->allow_ptr_leaks && *stype != STACK_MISC && *stype != STACK_ZERO) {
3429 /* Reject the write if range we may write to has not
3430 * been initialized beforehand. If we didn't reject
3431 * here, the ptr status would be erased below (even
3432 * though not all slots are actually overwritten),
3433 * possibly opening the door to leaks.
3435 * We do however catch STACK_INVALID case below, and
3436 * only allow reading possibly uninitialized memory
3437 * later for CAP_PERFMON, as the write may not happen to
3440 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
3445 /* Erase all spilled pointers. */
3446 state->stack[spi].spilled_ptr.type = NOT_INIT;
3448 /* Update the slot type. */
3449 new_type = STACK_MISC;
3450 if (writing_zero && *stype == STACK_ZERO) {
3451 new_type = STACK_ZERO;
3454 /* If the slot is STACK_INVALID, we check whether it's OK to
3455 * pretend that it will be initialized by this write. The slot
3456 * might not actually be written to, and so if we mark it as
3457 * initialized future reads might leak uninitialized memory.
3458 * For privileged programs, we will accept such reads to slots
3459 * that may or may not be written because, if we're reject
3460 * them, the error would be too confusing.
3462 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
3463 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
3470 /* backtracking doesn't work for STACK_ZERO yet. */
3471 err = mark_chain_precision(env, value_regno);
3478 /* When register 'dst_regno' is assigned some values from stack[min_off,
3479 * max_off), we set the register's type according to the types of the
3480 * respective stack slots. If all the stack values are known to be zeros, then
3481 * so is the destination reg. Otherwise, the register is considered to be
3482 * SCALAR. This function does not deal with register filling; the caller must
3483 * ensure that all spilled registers in the stack range have been marked as
3486 static void mark_reg_stack_read(struct bpf_verifier_env *env,
3487 /* func where src register points to */
3488 struct bpf_func_state *ptr_state,
3489 int min_off, int max_off, int dst_regno)
3491 struct bpf_verifier_state *vstate = env->cur_state;
3492 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3497 for (i = min_off; i < max_off; i++) {
3499 spi = slot / BPF_REG_SIZE;
3500 stype = ptr_state->stack[spi].slot_type;
3501 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
3505 if (zeros == max_off - min_off) {
3506 /* any access_size read into register is zero extended,
3507 * so the whole register == const_zero
3509 __mark_reg_const_zero(&state->regs[dst_regno]);
3510 /* backtracking doesn't support STACK_ZERO yet,
3511 * so mark it precise here, so that later
3512 * backtracking can stop here.
3513 * Backtracking may not need this if this register
3514 * doesn't participate in pointer adjustment.
3515 * Forward propagation of precise flag is not
3516 * necessary either. This mark is only to stop
3517 * backtracking. Any register that contributed
3518 * to const 0 was marked precise before spill.
3520 state->regs[dst_regno].precise = true;
3522 /* have read misc data from the stack */
3523 mark_reg_unknown(env, state->regs, dst_regno);
3525 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
3528 /* Read the stack at 'off' and put the results into the register indicated by
3529 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
3532 * 'dst_regno' can be -1, meaning that the read value is not going to a
3535 * The access is assumed to be within the current stack bounds.
3537 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
3538 /* func where src register points to */
3539 struct bpf_func_state *reg_state,
3540 int off, int size, int dst_regno)
3542 struct bpf_verifier_state *vstate = env->cur_state;
3543 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3544 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
3545 struct bpf_reg_state *reg;
3548 stype = reg_state->stack[spi].slot_type;
3549 reg = ®_state->stack[spi].spilled_ptr;
3551 if (is_spilled_reg(®_state->stack[spi])) {
3554 for (i = BPF_REG_SIZE - 1; i > 0 && stype[i - 1] == STACK_SPILL; i--)
3557 if (size != BPF_REG_SIZE || spill_size != BPF_REG_SIZE) {
3558 if (reg->type != SCALAR_VALUE) {
3559 verbose_linfo(env, env->insn_idx, "; ");
3560 verbose(env, "invalid size of register fill\n");
3564 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
3568 if (!(off % BPF_REG_SIZE) && size == spill_size) {
3569 /* The earlier check_reg_arg() has decided the
3570 * subreg_def for this insn. Save it first.
3572 s32 subreg_def = state->regs[dst_regno].subreg_def;
3574 state->regs[dst_regno] = *reg;
3575 state->regs[dst_regno].subreg_def = subreg_def;
3577 for (i = 0; i < size; i++) {
3578 type = stype[(slot - i) % BPF_REG_SIZE];
3579 if (type == STACK_SPILL)
3581 if (type == STACK_MISC)
3583 verbose(env, "invalid read from stack off %d+%d size %d\n",
3587 mark_reg_unknown(env, state->regs, dst_regno);
3589 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
3593 if (dst_regno >= 0) {
3594 /* restore register state from stack */
3595 state->regs[dst_regno] = *reg;
3596 /* mark reg as written since spilled pointer state likely
3597 * has its liveness marks cleared by is_state_visited()
3598 * which resets stack/reg liveness for state transitions
3600 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
3601 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
3602 /* If dst_regno==-1, the caller is asking us whether
3603 * it is acceptable to use this value as a SCALAR_VALUE
3605 * We must not allow unprivileged callers to do that
3606 * with spilled pointers.
3608 verbose(env, "leaking pointer from stack off %d\n",
3612 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
3614 for (i = 0; i < size; i++) {
3615 type = stype[(slot - i) % BPF_REG_SIZE];
3616 if (type == STACK_MISC)
3618 if (type == STACK_ZERO)
3620 verbose(env, "invalid read from stack off %d+%d size %d\n",
3624 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
3626 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
3631 enum bpf_access_src {
3632 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
3633 ACCESS_HELPER = 2, /* the access is performed by a helper */
3636 static int check_stack_range_initialized(struct bpf_verifier_env *env,
3637 int regno, int off, int access_size,
3638 bool zero_size_allowed,
3639 enum bpf_access_src type,
3640 struct bpf_call_arg_meta *meta);
3642 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
3644 return cur_regs(env) + regno;
3647 /* Read the stack at 'ptr_regno + off' and put the result into the register
3649 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
3650 * but not its variable offset.
3651 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
3653 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
3654 * filling registers (i.e. reads of spilled register cannot be detected when
3655 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
3656 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
3657 * offset; for a fixed offset check_stack_read_fixed_off should be used
3660 static int check_stack_read_var_off(struct bpf_verifier_env *env,
3661 int ptr_regno, int off, int size, int dst_regno)
3663 /* The state of the source register. */
3664 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3665 struct bpf_func_state *ptr_state = func(env, reg);
3667 int min_off, max_off;
3669 /* Note that we pass a NULL meta, so raw access will not be permitted.
3671 err = check_stack_range_initialized(env, ptr_regno, off, size,
3672 false, ACCESS_DIRECT, NULL);
3676 min_off = reg->smin_value + off;
3677 max_off = reg->smax_value + off;
3678 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
3682 /* check_stack_read dispatches to check_stack_read_fixed_off or
3683 * check_stack_read_var_off.
3685 * The caller must ensure that the offset falls within the allocated stack
3688 * 'dst_regno' is a register which will receive the value from the stack. It
3689 * can be -1, meaning that the read value is not going to a register.
3691 static int check_stack_read(struct bpf_verifier_env *env,
3692 int ptr_regno, int off, int size,
3695 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3696 struct bpf_func_state *state = func(env, reg);
3698 /* Some accesses are only permitted with a static offset. */
3699 bool var_off = !tnum_is_const(reg->var_off);
3701 /* The offset is required to be static when reads don't go to a
3702 * register, in order to not leak pointers (see
3703 * check_stack_read_fixed_off).
3705 if (dst_regno < 0 && var_off) {
3708 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3709 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
3713 /* Variable offset is prohibited for unprivileged mode for simplicity
3714 * since it requires corresponding support in Spectre masking for stack
3715 * ALU. See also retrieve_ptr_limit().
3717 if (!env->bypass_spec_v1 && var_off) {
3720 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3721 verbose(env, "R%d variable offset stack access prohibited for !root, var_off=%s\n",
3727 off += reg->var_off.value;
3728 err = check_stack_read_fixed_off(env, state, off, size,
3731 /* Variable offset stack reads need more conservative handling
3732 * than fixed offset ones. Note that dst_regno >= 0 on this
3735 err = check_stack_read_var_off(env, ptr_regno, off, size,
3742 /* check_stack_write dispatches to check_stack_write_fixed_off or
3743 * check_stack_write_var_off.
3745 * 'ptr_regno' is the register used as a pointer into the stack.
3746 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
3747 * 'value_regno' is the register whose value we're writing to the stack. It can
3748 * be -1, meaning that we're not writing from a register.
3750 * The caller must ensure that the offset falls within the maximum stack size.
3752 static int check_stack_write(struct bpf_verifier_env *env,
3753 int ptr_regno, int off, int size,
3754 int value_regno, int insn_idx)
3756 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3757 struct bpf_func_state *state = func(env, reg);
3760 if (tnum_is_const(reg->var_off)) {
3761 off += reg->var_off.value;
3762 err = check_stack_write_fixed_off(env, state, off, size,
3763 value_regno, insn_idx);
3765 /* Variable offset stack reads need more conservative handling
3766 * than fixed offset ones.
3768 err = check_stack_write_var_off(env, state,
3769 ptr_regno, off, size,
3770 value_regno, insn_idx);
3775 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
3776 int off, int size, enum bpf_access_type type)
3778 struct bpf_reg_state *regs = cur_regs(env);
3779 struct bpf_map *map = regs[regno].map_ptr;
3780 u32 cap = bpf_map_flags_to_cap(map);
3782 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
3783 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
3784 map->value_size, off, size);
3788 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
3789 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
3790 map->value_size, off, size);
3797 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
3798 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
3799 int off, int size, u32 mem_size,
3800 bool zero_size_allowed)
3802 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
3803 struct bpf_reg_state *reg;
3805 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
3808 reg = &cur_regs(env)[regno];
3809 switch (reg->type) {
3810 case PTR_TO_MAP_KEY:
3811 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
3812 mem_size, off, size);
3814 case PTR_TO_MAP_VALUE:
3815 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
3816 mem_size, off, size);
3819 case PTR_TO_PACKET_META:
3820 case PTR_TO_PACKET_END:
3821 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
3822 off, size, regno, reg->id, off, mem_size);
3826 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
3827 mem_size, off, size);
3833 /* check read/write into a memory region with possible variable offset */
3834 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
3835 int off, int size, u32 mem_size,
3836 bool zero_size_allowed)
3838 struct bpf_verifier_state *vstate = env->cur_state;
3839 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3840 struct bpf_reg_state *reg = &state->regs[regno];
3843 /* We may have adjusted the register pointing to memory region, so we
3844 * need to try adding each of min_value and max_value to off
3845 * to make sure our theoretical access will be safe.
3847 * The minimum value is only important with signed
3848 * comparisons where we can't assume the floor of a
3849 * value is 0. If we are using signed variables for our
3850 * index'es we need to make sure that whatever we use
3851 * will have a set floor within our range.
3853 if (reg->smin_value < 0 &&
3854 (reg->smin_value == S64_MIN ||
3855 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
3856 reg->smin_value + off < 0)) {
3857 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3861 err = __check_mem_access(env, regno, reg->smin_value + off, size,
3862 mem_size, zero_size_allowed);
3864 verbose(env, "R%d min value is outside of the allowed memory range\n",
3869 /* If we haven't set a max value then we need to bail since we can't be
3870 * sure we won't do bad things.
3871 * If reg->umax_value + off could overflow, treat that as unbounded too.
3873 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
3874 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
3878 err = __check_mem_access(env, regno, reg->umax_value + off, size,
3879 mem_size, zero_size_allowed);
3881 verbose(env, "R%d max value is outside of the allowed memory range\n",
3889 static int __check_ptr_off_reg(struct bpf_verifier_env *env,
3890 const struct bpf_reg_state *reg, int regno,
3893 /* Access to this pointer-typed register or passing it to a helper
3894 * is only allowed in its original, unmodified form.
3898 verbose(env, "negative offset %s ptr R%d off=%d disallowed\n",
3899 reg_type_str(env, reg->type), regno, reg->off);
3903 if (!fixed_off_ok && reg->off) {
3904 verbose(env, "dereference of modified %s ptr R%d off=%d disallowed\n",
3905 reg_type_str(env, reg->type), regno, reg->off);
3909 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3912 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3913 verbose(env, "variable %s access var_off=%s disallowed\n",
3914 reg_type_str(env, reg->type), tn_buf);
3921 int check_ptr_off_reg(struct bpf_verifier_env *env,
3922 const struct bpf_reg_state *reg, int regno)
3924 return __check_ptr_off_reg(env, reg, regno, false);
3927 static int map_kptr_match_type(struct bpf_verifier_env *env,
3928 struct btf_field *kptr_field,
3929 struct bpf_reg_state *reg, u32 regno)
3931 const char *targ_name = kernel_type_name(kptr_field->kptr.btf, kptr_field->kptr.btf_id);
3932 int perm_flags = PTR_MAYBE_NULL | PTR_TRUSTED;
3933 const char *reg_name = "";
3935 /* Only unreferenced case accepts untrusted pointers */
3936 if (kptr_field->type == BPF_KPTR_UNREF)
3937 perm_flags |= PTR_UNTRUSTED;
3939 if (base_type(reg->type) != PTR_TO_BTF_ID || (type_flag(reg->type) & ~perm_flags))
3942 if (!btf_is_kernel(reg->btf)) {
3943 verbose(env, "R%d must point to kernel BTF\n", regno);
3946 /* We need to verify reg->type and reg->btf, before accessing reg->btf */
3947 reg_name = kernel_type_name(reg->btf, reg->btf_id);
3949 /* For ref_ptr case, release function check should ensure we get one
3950 * referenced PTR_TO_BTF_ID, and that its fixed offset is 0. For the
3951 * normal store of unreferenced kptr, we must ensure var_off is zero.
3952 * Since ref_ptr cannot be accessed directly by BPF insns, checks for
3953 * reg->off and reg->ref_obj_id are not needed here.
3955 if (__check_ptr_off_reg(env, reg, regno, true))
3958 /* A full type match is needed, as BTF can be vmlinux or module BTF, and
3959 * we also need to take into account the reg->off.
3961 * We want to support cases like:
3969 * v = func(); // PTR_TO_BTF_ID
3970 * val->foo = v; // reg->off is zero, btf and btf_id match type
3971 * val->bar = &v->br; // reg->off is still zero, but we need to retry with
3972 * // first member type of struct after comparison fails
3973 * val->baz = &v->bz; // reg->off is non-zero, so struct needs to be walked
3976 * In the kptr_ref case, check_func_arg_reg_off already ensures reg->off
3977 * is zero. We must also ensure that btf_struct_ids_match does not walk
3978 * the struct to match type against first member of struct, i.e. reject
3979 * second case from above. Hence, when type is BPF_KPTR_REF, we set
3980 * strict mode to true for type match.
3982 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
3983 kptr_field->kptr.btf, kptr_field->kptr.btf_id,
3984 kptr_field->type == BPF_KPTR_REF))
3988 verbose(env, "invalid kptr access, R%d type=%s%s ", regno,
3989 reg_type_str(env, reg->type), reg_name);
3990 verbose(env, "expected=%s%s", reg_type_str(env, PTR_TO_BTF_ID), targ_name);
3991 if (kptr_field->type == BPF_KPTR_UNREF)
3992 verbose(env, " or %s%s\n", reg_type_str(env, PTR_TO_BTF_ID | PTR_UNTRUSTED),
3999 static int check_map_kptr_access(struct bpf_verifier_env *env, u32 regno,
4000 int value_regno, int insn_idx,
4001 struct btf_field *kptr_field)
4003 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
4004 int class = BPF_CLASS(insn->code);
4005 struct bpf_reg_state *val_reg;
4007 /* Things we already checked for in check_map_access and caller:
4008 * - Reject cases where variable offset may touch kptr
4009 * - size of access (must be BPF_DW)
4010 * - tnum_is_const(reg->var_off)
4011 * - kptr_field->offset == off + reg->var_off.value
4013 /* Only BPF_[LDX,STX,ST] | BPF_MEM | BPF_DW is supported */
4014 if (BPF_MODE(insn->code) != BPF_MEM) {
4015 verbose(env, "kptr in map can only be accessed using BPF_MEM instruction mode\n");
4019 /* We only allow loading referenced kptr, since it will be marked as
4020 * untrusted, similar to unreferenced kptr.
4022 if (class != BPF_LDX && kptr_field->type == BPF_KPTR_REF) {
4023 verbose(env, "store to referenced kptr disallowed\n");
4027 if (class == BPF_LDX) {
4028 val_reg = reg_state(env, value_regno);
4029 /* We can simply mark the value_regno receiving the pointer
4030 * value from map as PTR_TO_BTF_ID, with the correct type.
4032 mark_btf_ld_reg(env, cur_regs(env), value_regno, PTR_TO_BTF_ID, kptr_field->kptr.btf,
4033 kptr_field->kptr.btf_id, PTR_MAYBE_NULL | PTR_UNTRUSTED);
4034 /* For mark_ptr_or_null_reg */
4035 val_reg->id = ++env->id_gen;
4036 } else if (class == BPF_STX) {
4037 val_reg = reg_state(env, value_regno);
4038 if (!register_is_null(val_reg) &&
4039 map_kptr_match_type(env, kptr_field, val_reg, value_regno))
4041 } else if (class == BPF_ST) {
4043 verbose(env, "BPF_ST imm must be 0 when storing to kptr at off=%u\n",
4044 kptr_field->offset);
4048 verbose(env, "kptr in map can only be accessed using BPF_LDX/BPF_STX/BPF_ST\n");
4054 /* check read/write into a map element with possible variable offset */
4055 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
4056 int off, int size, bool zero_size_allowed,
4057 enum bpf_access_src src)
4059 struct bpf_verifier_state *vstate = env->cur_state;
4060 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4061 struct bpf_reg_state *reg = &state->regs[regno];
4062 struct bpf_map *map = reg->map_ptr;
4063 struct btf_record *rec;
4066 err = check_mem_region_access(env, regno, off, size, map->value_size,
4071 if (IS_ERR_OR_NULL(map->record))
4074 for (i = 0; i < rec->cnt; i++) {
4075 struct btf_field *field = &rec->fields[i];
4076 u32 p = field->offset;
4078 /* If any part of a field can be touched by load/store, reject
4079 * this program. To check that [x1, x2) overlaps with [y1, y2),
4080 * it is sufficient to check x1 < y2 && y1 < x2.
4082 if (reg->smin_value + off < p + btf_field_type_size(field->type) &&
4083 p < reg->umax_value + off + size) {
4084 switch (field->type) {
4085 case BPF_KPTR_UNREF:
4087 if (src != ACCESS_DIRECT) {
4088 verbose(env, "kptr cannot be accessed indirectly by helper\n");
4091 if (!tnum_is_const(reg->var_off)) {
4092 verbose(env, "kptr access cannot have variable offset\n");
4095 if (p != off + reg->var_off.value) {
4096 verbose(env, "kptr access misaligned expected=%u off=%llu\n",
4097 p, off + reg->var_off.value);
4100 if (size != bpf_size_to_bytes(BPF_DW)) {
4101 verbose(env, "kptr access size must be BPF_DW\n");
4106 verbose(env, "%s cannot be accessed directly by load/store\n",
4107 btf_field_type_name(field->type));
4115 #define MAX_PACKET_OFF 0xffff
4117 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
4118 const struct bpf_call_arg_meta *meta,
4119 enum bpf_access_type t)
4121 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
4123 switch (prog_type) {
4124 /* Program types only with direct read access go here! */
4125 case BPF_PROG_TYPE_LWT_IN:
4126 case BPF_PROG_TYPE_LWT_OUT:
4127 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
4128 case BPF_PROG_TYPE_SK_REUSEPORT:
4129 case BPF_PROG_TYPE_FLOW_DISSECTOR:
4130 case BPF_PROG_TYPE_CGROUP_SKB:
4135 /* Program types with direct read + write access go here! */
4136 case BPF_PROG_TYPE_SCHED_CLS:
4137 case BPF_PROG_TYPE_SCHED_ACT:
4138 case BPF_PROG_TYPE_XDP:
4139 case BPF_PROG_TYPE_LWT_XMIT:
4140 case BPF_PROG_TYPE_SK_SKB:
4141 case BPF_PROG_TYPE_SK_MSG:
4143 return meta->pkt_access;
4145 env->seen_direct_write = true;
4148 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
4150 env->seen_direct_write = true;
4159 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
4160 int size, bool zero_size_allowed)
4162 struct bpf_reg_state *regs = cur_regs(env);
4163 struct bpf_reg_state *reg = ®s[regno];
4166 /* We may have added a variable offset to the packet pointer; but any
4167 * reg->range we have comes after that. We are only checking the fixed
4171 /* We don't allow negative numbers, because we aren't tracking enough
4172 * detail to prove they're safe.
4174 if (reg->smin_value < 0) {
4175 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
4180 err = reg->range < 0 ? -EINVAL :
4181 __check_mem_access(env, regno, off, size, reg->range,
4184 verbose(env, "R%d offset is outside of the packet\n", regno);
4188 /* __check_mem_access has made sure "off + size - 1" is within u16.
4189 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
4190 * otherwise find_good_pkt_pointers would have refused to set range info
4191 * that __check_mem_access would have rejected this pkt access.
4192 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
4194 env->prog->aux->max_pkt_offset =
4195 max_t(u32, env->prog->aux->max_pkt_offset,
4196 off + reg->umax_value + size - 1);
4201 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
4202 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
4203 enum bpf_access_type t, enum bpf_reg_type *reg_type,
4204 struct btf **btf, u32 *btf_id)
4206 struct bpf_insn_access_aux info = {
4207 .reg_type = *reg_type,
4211 if (env->ops->is_valid_access &&
4212 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
4213 /* A non zero info.ctx_field_size indicates that this field is a
4214 * candidate for later verifier transformation to load the whole
4215 * field and then apply a mask when accessed with a narrower
4216 * access than actual ctx access size. A zero info.ctx_field_size
4217 * will only allow for whole field access and rejects any other
4218 * type of narrower access.
4220 *reg_type = info.reg_type;
4222 if (base_type(*reg_type) == PTR_TO_BTF_ID) {
4224 *btf_id = info.btf_id;
4226 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
4228 /* remember the offset of last byte accessed in ctx */
4229 if (env->prog->aux->max_ctx_offset < off + size)
4230 env->prog->aux->max_ctx_offset = off + size;
4234 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
4238 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
4241 if (size < 0 || off < 0 ||
4242 (u64)off + size > sizeof(struct bpf_flow_keys)) {
4243 verbose(env, "invalid access to flow keys off=%d size=%d\n",
4250 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
4251 u32 regno, int off, int size,
4252 enum bpf_access_type t)
4254 struct bpf_reg_state *regs = cur_regs(env);
4255 struct bpf_reg_state *reg = ®s[regno];
4256 struct bpf_insn_access_aux info = {};
4259 if (reg->smin_value < 0) {
4260 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
4265 switch (reg->type) {
4266 case PTR_TO_SOCK_COMMON:
4267 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
4270 valid = bpf_sock_is_valid_access(off, size, t, &info);
4272 case PTR_TO_TCP_SOCK:
4273 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
4275 case PTR_TO_XDP_SOCK:
4276 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
4284 env->insn_aux_data[insn_idx].ctx_field_size =
4285 info.ctx_field_size;
4289 verbose(env, "R%d invalid %s access off=%d size=%d\n",
4290 regno, reg_type_str(env, reg->type), off, size);
4295 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
4297 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
4300 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
4302 const struct bpf_reg_state *reg = reg_state(env, regno);
4304 return reg->type == PTR_TO_CTX;
4307 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
4309 const struct bpf_reg_state *reg = reg_state(env, regno);
4311 return type_is_sk_pointer(reg->type);
4314 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
4316 const struct bpf_reg_state *reg = reg_state(env, regno);
4318 return type_is_pkt_pointer(reg->type);
4321 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
4323 const struct bpf_reg_state *reg = reg_state(env, regno);
4325 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
4326 return reg->type == PTR_TO_FLOW_KEYS;
4329 static bool is_trusted_reg(const struct bpf_reg_state *reg)
4331 /* A referenced register is always trusted. */
4332 if (reg->ref_obj_id)
4335 /* If a register is not referenced, it is trusted if it has the
4336 * MEM_ALLOC or PTR_TRUSTED type modifiers, and no others. Some of the
4337 * other type modifiers may be safe, but we elect to take an opt-in
4338 * approach here as some (e.g. PTR_UNTRUSTED and PTR_MAYBE_NULL) are
4341 * Eventually, we should make PTR_TRUSTED the single source of truth
4342 * for whether a register is trusted.
4344 return type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS &&
4345 !bpf_type_has_unsafe_modifiers(reg->type);
4348 static bool is_rcu_reg(const struct bpf_reg_state *reg)
4350 return reg->type & MEM_RCU;
4353 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
4354 const struct bpf_reg_state *reg,
4355 int off, int size, bool strict)
4357 struct tnum reg_off;
4360 /* Byte size accesses are always allowed. */
4361 if (!strict || size == 1)
4364 /* For platforms that do not have a Kconfig enabling
4365 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
4366 * NET_IP_ALIGN is universally set to '2'. And on platforms
4367 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
4368 * to this code only in strict mode where we want to emulate
4369 * the NET_IP_ALIGN==2 checking. Therefore use an
4370 * unconditional IP align value of '2'.
4374 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
4375 if (!tnum_is_aligned(reg_off, size)) {
4378 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4380 "misaligned packet access off %d+%s+%d+%d size %d\n",
4381 ip_align, tn_buf, reg->off, off, size);
4388 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
4389 const struct bpf_reg_state *reg,
4390 const char *pointer_desc,
4391 int off, int size, bool strict)
4393 struct tnum reg_off;
4395 /* Byte size accesses are always allowed. */
4396 if (!strict || size == 1)
4399 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
4400 if (!tnum_is_aligned(reg_off, size)) {
4403 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4404 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
4405 pointer_desc, tn_buf, reg->off, off, size);
4412 static int check_ptr_alignment(struct bpf_verifier_env *env,
4413 const struct bpf_reg_state *reg, int off,
4414 int size, bool strict_alignment_once)
4416 bool strict = env->strict_alignment || strict_alignment_once;
4417 const char *pointer_desc = "";
4419 switch (reg->type) {
4421 case PTR_TO_PACKET_META:
4422 /* Special case, because of NET_IP_ALIGN. Given metadata sits
4423 * right in front, treat it the very same way.
4425 return check_pkt_ptr_alignment(env, reg, off, size, strict);
4426 case PTR_TO_FLOW_KEYS:
4427 pointer_desc = "flow keys ";
4429 case PTR_TO_MAP_KEY:
4430 pointer_desc = "key ";
4432 case PTR_TO_MAP_VALUE:
4433 pointer_desc = "value ";
4436 pointer_desc = "context ";
4439 pointer_desc = "stack ";
4440 /* The stack spill tracking logic in check_stack_write_fixed_off()
4441 * and check_stack_read_fixed_off() relies on stack accesses being
4447 pointer_desc = "sock ";
4449 case PTR_TO_SOCK_COMMON:
4450 pointer_desc = "sock_common ";
4452 case PTR_TO_TCP_SOCK:
4453 pointer_desc = "tcp_sock ";
4455 case PTR_TO_XDP_SOCK:
4456 pointer_desc = "xdp_sock ";
4461 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
4465 static int update_stack_depth(struct bpf_verifier_env *env,
4466 const struct bpf_func_state *func,
4469 u16 stack = env->subprog_info[func->subprogno].stack_depth;
4474 /* update known max for given subprogram */
4475 env->subprog_info[func->subprogno].stack_depth = -off;
4479 /* starting from main bpf function walk all instructions of the function
4480 * and recursively walk all callees that given function can call.
4481 * Ignore jump and exit insns.
4482 * Since recursion is prevented by check_cfg() this algorithm
4483 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
4485 static int check_max_stack_depth(struct bpf_verifier_env *env)
4487 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
4488 struct bpf_subprog_info *subprog = env->subprog_info;
4489 struct bpf_insn *insn = env->prog->insnsi;
4490 bool tail_call_reachable = false;
4491 int ret_insn[MAX_CALL_FRAMES];
4492 int ret_prog[MAX_CALL_FRAMES];
4496 /* protect against potential stack overflow that might happen when
4497 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
4498 * depth for such case down to 256 so that the worst case scenario
4499 * would result in 8k stack size (32 which is tailcall limit * 256 =
4502 * To get the idea what might happen, see an example:
4503 * func1 -> sub rsp, 128
4504 * subfunc1 -> sub rsp, 256
4505 * tailcall1 -> add rsp, 256
4506 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
4507 * subfunc2 -> sub rsp, 64
4508 * subfunc22 -> sub rsp, 128
4509 * tailcall2 -> add rsp, 128
4510 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
4512 * tailcall will unwind the current stack frame but it will not get rid
4513 * of caller's stack as shown on the example above.
4515 if (idx && subprog[idx].has_tail_call && depth >= 256) {
4517 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
4521 /* round up to 32-bytes, since this is granularity
4522 * of interpreter stack size
4524 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
4525 if (depth > MAX_BPF_STACK) {
4526 verbose(env, "combined stack size of %d calls is %d. Too large\n",
4531 subprog_end = subprog[idx + 1].start;
4532 for (; i < subprog_end; i++) {
4535 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
4537 /* remember insn and function to return to */
4538 ret_insn[frame] = i + 1;
4539 ret_prog[frame] = idx;
4541 /* find the callee */
4542 next_insn = i + insn[i].imm + 1;
4543 idx = find_subprog(env, next_insn);
4545 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
4549 if (subprog[idx].is_async_cb) {
4550 if (subprog[idx].has_tail_call) {
4551 verbose(env, "verifier bug. subprog has tail_call and async cb\n");
4554 /* async callbacks don't increase bpf prog stack size */
4559 if (subprog[idx].has_tail_call)
4560 tail_call_reachable = true;
4563 if (frame >= MAX_CALL_FRAMES) {
4564 verbose(env, "the call stack of %d frames is too deep !\n",
4570 /* if tail call got detected across bpf2bpf calls then mark each of the
4571 * currently present subprog frames as tail call reachable subprogs;
4572 * this info will be utilized by JIT so that we will be preserving the
4573 * tail call counter throughout bpf2bpf calls combined with tailcalls
4575 if (tail_call_reachable)
4576 for (j = 0; j < frame; j++)
4577 subprog[ret_prog[j]].tail_call_reachable = true;
4578 if (subprog[0].tail_call_reachable)
4579 env->prog->aux->tail_call_reachable = true;
4581 /* end of for() loop means the last insn of the 'subprog'
4582 * was reached. Doesn't matter whether it was JA or EXIT
4586 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
4588 i = ret_insn[frame];
4589 idx = ret_prog[frame];
4593 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
4594 static int get_callee_stack_depth(struct bpf_verifier_env *env,
4595 const struct bpf_insn *insn, int idx)
4597 int start = idx + insn->imm + 1, subprog;
4599 subprog = find_subprog(env, start);
4601 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
4605 return env->subprog_info[subprog].stack_depth;
4609 static int __check_buffer_access(struct bpf_verifier_env *env,
4610 const char *buf_info,
4611 const struct bpf_reg_state *reg,
4612 int regno, int off, int size)
4616 "R%d invalid %s buffer access: off=%d, size=%d\n",
4617 regno, buf_info, off, size);
4620 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4623 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4625 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
4626 regno, off, tn_buf);
4633 static int check_tp_buffer_access(struct bpf_verifier_env *env,
4634 const struct bpf_reg_state *reg,
4635 int regno, int off, int size)
4639 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
4643 if (off + size > env->prog->aux->max_tp_access)
4644 env->prog->aux->max_tp_access = off + size;
4649 static int check_buffer_access(struct bpf_verifier_env *env,
4650 const struct bpf_reg_state *reg,
4651 int regno, int off, int size,
4652 bool zero_size_allowed,
4655 const char *buf_info = type_is_rdonly_mem(reg->type) ? "rdonly" : "rdwr";
4658 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
4662 if (off + size > *max_access)
4663 *max_access = off + size;
4668 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
4669 static void zext_32_to_64(struct bpf_reg_state *reg)
4671 reg->var_off = tnum_subreg(reg->var_off);
4672 __reg_assign_32_into_64(reg);
4675 /* truncate register to smaller size (in bytes)
4676 * must be called with size < BPF_REG_SIZE
4678 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
4682 /* clear high bits in bit representation */
4683 reg->var_off = tnum_cast(reg->var_off, size);
4685 /* fix arithmetic bounds */
4686 mask = ((u64)1 << (size * 8)) - 1;
4687 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
4688 reg->umin_value &= mask;
4689 reg->umax_value &= mask;
4691 reg->umin_value = 0;
4692 reg->umax_value = mask;
4694 reg->smin_value = reg->umin_value;
4695 reg->smax_value = reg->umax_value;
4697 /* If size is smaller than 32bit register the 32bit register
4698 * values are also truncated so we push 64-bit bounds into
4699 * 32-bit bounds. Above were truncated < 32-bits already.
4703 __reg_combine_64_into_32(reg);
4706 static bool bpf_map_is_rdonly(const struct bpf_map *map)
4708 /* A map is considered read-only if the following condition are true:
4710 * 1) BPF program side cannot change any of the map content. The
4711 * BPF_F_RDONLY_PROG flag is throughout the lifetime of a map
4712 * and was set at map creation time.
4713 * 2) The map value(s) have been initialized from user space by a
4714 * loader and then "frozen", such that no new map update/delete
4715 * operations from syscall side are possible for the rest of
4716 * the map's lifetime from that point onwards.
4717 * 3) Any parallel/pending map update/delete operations from syscall
4718 * side have been completed. Only after that point, it's safe to
4719 * assume that map value(s) are immutable.
4721 return (map->map_flags & BPF_F_RDONLY_PROG) &&
4722 READ_ONCE(map->frozen) &&
4723 !bpf_map_write_active(map);
4726 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
4732 err = map->ops->map_direct_value_addr(map, &addr, off);
4735 ptr = (void *)(long)addr + off;
4739 *val = (u64)*(u8 *)ptr;
4742 *val = (u64)*(u16 *)ptr;
4745 *val = (u64)*(u32 *)ptr;
4756 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
4757 struct bpf_reg_state *regs,
4758 int regno, int off, int size,
4759 enum bpf_access_type atype,
4762 struct bpf_reg_state *reg = regs + regno;
4763 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
4764 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
4765 enum bpf_type_flag flag = 0;
4769 if (!env->allow_ptr_leaks) {
4771 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
4775 if (!env->prog->gpl_compatible && btf_is_kernel(reg->btf)) {
4777 "Cannot access kernel 'struct %s' from non-GPL compatible program\n",
4783 "R%d is ptr_%s invalid negative access: off=%d\n",
4787 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4790 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4792 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
4793 regno, tname, off, tn_buf);
4797 if (reg->type & MEM_USER) {
4799 "R%d is ptr_%s access user memory: off=%d\n",
4804 if (reg->type & MEM_PERCPU) {
4806 "R%d is ptr_%s access percpu memory: off=%d\n",
4811 if (env->ops->btf_struct_access && !type_is_alloc(reg->type)) {
4812 if (!btf_is_kernel(reg->btf)) {
4813 verbose(env, "verifier internal error: reg->btf must be kernel btf\n");
4816 ret = env->ops->btf_struct_access(&env->log, reg, off, size, atype, &btf_id, &flag);
4818 /* Writes are permitted with default btf_struct_access for
4819 * program allocated objects (which always have ref_obj_id > 0),
4820 * but not for untrusted PTR_TO_BTF_ID | MEM_ALLOC.
4822 if (atype != BPF_READ && reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
4823 verbose(env, "only read is supported\n");
4827 if (type_is_alloc(reg->type) && !reg->ref_obj_id) {
4828 verbose(env, "verifier internal error: ref_obj_id for allocated object must be non-zero\n");
4832 ret = btf_struct_access(&env->log, reg, off, size, atype, &btf_id, &flag);
4838 /* If this is an untrusted pointer, all pointers formed by walking it
4839 * also inherit the untrusted flag.
4841 if (type_flag(reg->type) & PTR_UNTRUSTED)
4842 flag |= PTR_UNTRUSTED;
4844 /* By default any pointer obtained from walking a trusted pointer is
4845 * no longer trusted except the rcu case below.
4847 flag &= ~PTR_TRUSTED;
4849 if (flag & MEM_RCU) {
4850 /* Mark value register as MEM_RCU only if it is protected by
4851 * bpf_rcu_read_lock() and the ptr reg is rcu or trusted. MEM_RCU
4852 * itself can already indicate trustedness inside the rcu
4853 * read lock region. Also mark rcu pointer as PTR_MAYBE_NULL since
4854 * it could be null in some cases.
4856 if (!env->cur_state->active_rcu_lock ||
4857 !(is_trusted_reg(reg) || is_rcu_reg(reg)))
4860 flag |= PTR_MAYBE_NULL;
4861 } else if (reg->type & MEM_RCU) {
4862 /* ptr (reg) is marked as MEM_RCU, but the struct field is not tagged
4863 * with __rcu. Mark the flag as PTR_UNTRUSTED conservatively.
4865 flag |= PTR_UNTRUSTED;
4868 if (atype == BPF_READ && value_regno >= 0)
4869 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id, flag);
4874 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
4875 struct bpf_reg_state *regs,
4876 int regno, int off, int size,
4877 enum bpf_access_type atype,
4880 struct bpf_reg_state *reg = regs + regno;
4881 struct bpf_map *map = reg->map_ptr;
4882 struct bpf_reg_state map_reg;
4883 enum bpf_type_flag flag = 0;
4884 const struct btf_type *t;
4890 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
4894 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
4895 verbose(env, "map_ptr access not supported for map type %d\n",
4900 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
4901 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
4903 if (!env->allow_ptr_leaks) {
4905 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
4911 verbose(env, "R%d is %s invalid negative access: off=%d\n",
4916 if (atype != BPF_READ) {
4917 verbose(env, "only read from %s is supported\n", tname);
4921 /* Simulate access to a PTR_TO_BTF_ID */
4922 memset(&map_reg, 0, sizeof(map_reg));
4923 mark_btf_ld_reg(env, &map_reg, 0, PTR_TO_BTF_ID, btf_vmlinux, *map->ops->map_btf_id, 0);
4924 ret = btf_struct_access(&env->log, &map_reg, off, size, atype, &btf_id, &flag);
4928 if (value_regno >= 0)
4929 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id, flag);
4934 /* Check that the stack access at the given offset is within bounds. The
4935 * maximum valid offset is -1.
4937 * The minimum valid offset is -MAX_BPF_STACK for writes, and
4938 * -state->allocated_stack for reads.
4940 static int check_stack_slot_within_bounds(int off,
4941 struct bpf_func_state *state,
4942 enum bpf_access_type t)
4947 min_valid_off = -MAX_BPF_STACK;
4949 min_valid_off = -state->allocated_stack;
4951 if (off < min_valid_off || off > -1)
4956 /* Check that the stack access at 'regno + off' falls within the maximum stack
4959 * 'off' includes `regno->offset`, but not its dynamic part (if any).
4961 static int check_stack_access_within_bounds(
4962 struct bpf_verifier_env *env,
4963 int regno, int off, int access_size,
4964 enum bpf_access_src src, enum bpf_access_type type)
4966 struct bpf_reg_state *regs = cur_regs(env);
4967 struct bpf_reg_state *reg = regs + regno;
4968 struct bpf_func_state *state = func(env, reg);
4969 int min_off, max_off;
4973 if (src == ACCESS_HELPER)
4974 /* We don't know if helpers are reading or writing (or both). */
4975 err_extra = " indirect access to";
4976 else if (type == BPF_READ)
4977 err_extra = " read from";
4979 err_extra = " write to";
4981 if (tnum_is_const(reg->var_off)) {
4982 min_off = reg->var_off.value + off;
4983 if (access_size > 0)
4984 max_off = min_off + access_size - 1;
4988 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
4989 reg->smin_value <= -BPF_MAX_VAR_OFF) {
4990 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
4994 min_off = reg->smin_value + off;
4995 if (access_size > 0)
4996 max_off = reg->smax_value + off + access_size - 1;
5001 err = check_stack_slot_within_bounds(min_off, state, type);
5003 err = check_stack_slot_within_bounds(max_off, state, type);
5006 if (tnum_is_const(reg->var_off)) {
5007 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
5008 err_extra, regno, off, access_size);
5012 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5013 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
5014 err_extra, regno, tn_buf, access_size);
5020 /* check whether memory at (regno + off) is accessible for t = (read | write)
5021 * if t==write, value_regno is a register which value is stored into memory
5022 * if t==read, value_regno is a register which will receive the value from memory
5023 * if t==write && value_regno==-1, some unknown value is stored into memory
5024 * if t==read && value_regno==-1, don't care what we read from memory
5026 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
5027 int off, int bpf_size, enum bpf_access_type t,
5028 int value_regno, bool strict_alignment_once)
5030 struct bpf_reg_state *regs = cur_regs(env);
5031 struct bpf_reg_state *reg = regs + regno;
5032 struct bpf_func_state *state;
5035 size = bpf_size_to_bytes(bpf_size);
5039 /* alignment checks will add in reg->off themselves */
5040 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
5044 /* for access checks, reg->off is just part of off */
5047 if (reg->type == PTR_TO_MAP_KEY) {
5048 if (t == BPF_WRITE) {
5049 verbose(env, "write to change key R%d not allowed\n", regno);
5053 err = check_mem_region_access(env, regno, off, size,
5054 reg->map_ptr->key_size, false);
5057 if (value_regno >= 0)
5058 mark_reg_unknown(env, regs, value_regno);
5059 } else if (reg->type == PTR_TO_MAP_VALUE) {
5060 struct btf_field *kptr_field = NULL;
5062 if (t == BPF_WRITE && value_regno >= 0 &&
5063 is_pointer_value(env, value_regno)) {
5064 verbose(env, "R%d leaks addr into map\n", value_regno);
5067 err = check_map_access_type(env, regno, off, size, t);
5070 err = check_map_access(env, regno, off, size, false, ACCESS_DIRECT);
5073 if (tnum_is_const(reg->var_off))
5074 kptr_field = btf_record_find(reg->map_ptr->record,
5075 off + reg->var_off.value, BPF_KPTR);
5077 err = check_map_kptr_access(env, regno, value_regno, insn_idx, kptr_field);
5078 } else if (t == BPF_READ && value_regno >= 0) {
5079 struct bpf_map *map = reg->map_ptr;
5081 /* if map is read-only, track its contents as scalars */
5082 if (tnum_is_const(reg->var_off) &&
5083 bpf_map_is_rdonly(map) &&
5084 map->ops->map_direct_value_addr) {
5085 int map_off = off + reg->var_off.value;
5088 err = bpf_map_direct_read(map, map_off, size,
5093 regs[value_regno].type = SCALAR_VALUE;
5094 __mark_reg_known(®s[value_regno], val);
5096 mark_reg_unknown(env, regs, value_regno);
5099 } else if (base_type(reg->type) == PTR_TO_MEM) {
5100 bool rdonly_mem = type_is_rdonly_mem(reg->type);
5102 if (type_may_be_null(reg->type)) {
5103 verbose(env, "R%d invalid mem access '%s'\n", regno,
5104 reg_type_str(env, reg->type));
5108 if (t == BPF_WRITE && rdonly_mem) {
5109 verbose(env, "R%d cannot write into %s\n",
5110 regno, reg_type_str(env, reg->type));
5114 if (t == BPF_WRITE && value_regno >= 0 &&
5115 is_pointer_value(env, value_regno)) {
5116 verbose(env, "R%d leaks addr into mem\n", value_regno);
5120 err = check_mem_region_access(env, regno, off, size,
5121 reg->mem_size, false);
5122 if (!err && value_regno >= 0 && (t == BPF_READ || rdonly_mem))
5123 mark_reg_unknown(env, regs, value_regno);
5124 } else if (reg->type == PTR_TO_CTX) {
5125 enum bpf_reg_type reg_type = SCALAR_VALUE;
5126 struct btf *btf = NULL;
5129 if (t == BPF_WRITE && value_regno >= 0 &&
5130 is_pointer_value(env, value_regno)) {
5131 verbose(env, "R%d leaks addr into ctx\n", value_regno);
5135 err = check_ptr_off_reg(env, reg, regno);
5139 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf,
5142 verbose_linfo(env, insn_idx, "; ");
5143 if (!err && t == BPF_READ && value_regno >= 0) {
5144 /* ctx access returns either a scalar, or a
5145 * PTR_TO_PACKET[_META,_END]. In the latter
5146 * case, we know the offset is zero.
5148 if (reg_type == SCALAR_VALUE) {
5149 mark_reg_unknown(env, regs, value_regno);
5151 mark_reg_known_zero(env, regs,
5153 if (type_may_be_null(reg_type))
5154 regs[value_regno].id = ++env->id_gen;
5155 /* A load of ctx field could have different
5156 * actual load size with the one encoded in the
5157 * insn. When the dst is PTR, it is for sure not
5160 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
5161 if (base_type(reg_type) == PTR_TO_BTF_ID) {
5162 regs[value_regno].btf = btf;
5163 regs[value_regno].btf_id = btf_id;
5166 regs[value_regno].type = reg_type;
5169 } else if (reg->type == PTR_TO_STACK) {
5170 /* Basic bounds checks. */
5171 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
5175 state = func(env, reg);
5176 err = update_stack_depth(env, state, off);
5181 err = check_stack_read(env, regno, off, size,
5184 err = check_stack_write(env, regno, off, size,
5185 value_regno, insn_idx);
5186 } else if (reg_is_pkt_pointer(reg)) {
5187 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
5188 verbose(env, "cannot write into packet\n");
5191 if (t == BPF_WRITE && value_regno >= 0 &&
5192 is_pointer_value(env, value_regno)) {
5193 verbose(env, "R%d leaks addr into packet\n",
5197 err = check_packet_access(env, regno, off, size, false);
5198 if (!err && t == BPF_READ && value_regno >= 0)
5199 mark_reg_unknown(env, regs, value_regno);
5200 } else if (reg->type == PTR_TO_FLOW_KEYS) {
5201 if (t == BPF_WRITE && value_regno >= 0 &&
5202 is_pointer_value(env, value_regno)) {
5203 verbose(env, "R%d leaks addr into flow keys\n",
5208 err = check_flow_keys_access(env, off, size);
5209 if (!err && t == BPF_READ && value_regno >= 0)
5210 mark_reg_unknown(env, regs, value_regno);
5211 } else if (type_is_sk_pointer(reg->type)) {
5212 if (t == BPF_WRITE) {
5213 verbose(env, "R%d cannot write into %s\n",
5214 regno, reg_type_str(env, reg->type));
5217 err = check_sock_access(env, insn_idx, regno, off, size, t);
5218 if (!err && value_regno >= 0)
5219 mark_reg_unknown(env, regs, value_regno);
5220 } else if (reg->type == PTR_TO_TP_BUFFER) {
5221 err = check_tp_buffer_access(env, reg, regno, off, size);
5222 if (!err && t == BPF_READ && value_regno >= 0)
5223 mark_reg_unknown(env, regs, value_regno);
5224 } else if (base_type(reg->type) == PTR_TO_BTF_ID &&
5225 !type_may_be_null(reg->type)) {
5226 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
5228 } else if (reg->type == CONST_PTR_TO_MAP) {
5229 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
5231 } else if (base_type(reg->type) == PTR_TO_BUF) {
5232 bool rdonly_mem = type_is_rdonly_mem(reg->type);
5236 if (t == BPF_WRITE) {
5237 verbose(env, "R%d cannot write into %s\n",
5238 regno, reg_type_str(env, reg->type));
5241 max_access = &env->prog->aux->max_rdonly_access;
5243 max_access = &env->prog->aux->max_rdwr_access;
5246 err = check_buffer_access(env, reg, regno, off, size, false,
5249 if (!err && value_regno >= 0 && (rdonly_mem || t == BPF_READ))
5250 mark_reg_unknown(env, regs, value_regno);
5252 verbose(env, "R%d invalid mem access '%s'\n", regno,
5253 reg_type_str(env, reg->type));
5257 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
5258 regs[value_regno].type == SCALAR_VALUE) {
5259 /* b/h/w load zero-extends, mark upper bits as known 0 */
5260 coerce_reg_to_size(®s[value_regno], size);
5265 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
5270 switch (insn->imm) {
5272 case BPF_ADD | BPF_FETCH:
5274 case BPF_AND | BPF_FETCH:
5276 case BPF_OR | BPF_FETCH:
5278 case BPF_XOR | BPF_FETCH:
5283 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
5287 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
5288 verbose(env, "invalid atomic operand size\n");
5292 /* check src1 operand */
5293 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5297 /* check src2 operand */
5298 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5302 if (insn->imm == BPF_CMPXCHG) {
5303 /* Check comparison of R0 with memory location */
5304 const u32 aux_reg = BPF_REG_0;
5306 err = check_reg_arg(env, aux_reg, SRC_OP);
5310 if (is_pointer_value(env, aux_reg)) {
5311 verbose(env, "R%d leaks addr into mem\n", aux_reg);
5316 if (is_pointer_value(env, insn->src_reg)) {
5317 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
5321 if (is_ctx_reg(env, insn->dst_reg) ||
5322 is_pkt_reg(env, insn->dst_reg) ||
5323 is_flow_key_reg(env, insn->dst_reg) ||
5324 is_sk_reg(env, insn->dst_reg)) {
5325 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
5327 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
5331 if (insn->imm & BPF_FETCH) {
5332 if (insn->imm == BPF_CMPXCHG)
5333 load_reg = BPF_REG_0;
5335 load_reg = insn->src_reg;
5337 /* check and record load of old value */
5338 err = check_reg_arg(env, load_reg, DST_OP);
5342 /* This instruction accesses a memory location but doesn't
5343 * actually load it into a register.
5348 /* Check whether we can read the memory, with second call for fetch
5349 * case to simulate the register fill.
5351 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
5352 BPF_SIZE(insn->code), BPF_READ, -1, true);
5353 if (!err && load_reg >= 0)
5354 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
5355 BPF_SIZE(insn->code), BPF_READ, load_reg,
5360 /* Check whether we can write into the same memory. */
5361 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
5362 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
5369 /* When register 'regno' is used to read the stack (either directly or through
5370 * a helper function) make sure that it's within stack boundary and, depending
5371 * on the access type, that all elements of the stack are initialized.
5373 * 'off' includes 'regno->off', but not its dynamic part (if any).
5375 * All registers that have been spilled on the stack in the slots within the
5376 * read offsets are marked as read.
5378 static int check_stack_range_initialized(
5379 struct bpf_verifier_env *env, int regno, int off,
5380 int access_size, bool zero_size_allowed,
5381 enum bpf_access_src type, struct bpf_call_arg_meta *meta)
5383 struct bpf_reg_state *reg = reg_state(env, regno);
5384 struct bpf_func_state *state = func(env, reg);
5385 int err, min_off, max_off, i, j, slot, spi;
5386 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
5387 enum bpf_access_type bounds_check_type;
5388 /* Some accesses can write anything into the stack, others are
5391 bool clobber = false;
5393 if (access_size == 0 && !zero_size_allowed) {
5394 verbose(env, "invalid zero-sized read\n");
5398 if (type == ACCESS_HELPER) {
5399 /* The bounds checks for writes are more permissive than for
5400 * reads. However, if raw_mode is not set, we'll do extra
5403 bounds_check_type = BPF_WRITE;
5406 bounds_check_type = BPF_READ;
5408 err = check_stack_access_within_bounds(env, regno, off, access_size,
5409 type, bounds_check_type);
5414 if (tnum_is_const(reg->var_off)) {
5415 min_off = max_off = reg->var_off.value + off;
5417 /* Variable offset is prohibited for unprivileged mode for
5418 * simplicity since it requires corresponding support in
5419 * Spectre masking for stack ALU.
5420 * See also retrieve_ptr_limit().
5422 if (!env->bypass_spec_v1) {
5425 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5426 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
5427 regno, err_extra, tn_buf);
5430 /* Only initialized buffer on stack is allowed to be accessed
5431 * with variable offset. With uninitialized buffer it's hard to
5432 * guarantee that whole memory is marked as initialized on
5433 * helper return since specific bounds are unknown what may
5434 * cause uninitialized stack leaking.
5436 if (meta && meta->raw_mode)
5439 min_off = reg->smin_value + off;
5440 max_off = reg->smax_value + off;
5443 if (meta && meta->raw_mode) {
5444 meta->access_size = access_size;
5445 meta->regno = regno;
5449 for (i = min_off; i < max_off + access_size; i++) {
5453 spi = slot / BPF_REG_SIZE;
5454 if (state->allocated_stack <= slot)
5456 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
5457 if (*stype == STACK_MISC)
5459 if (*stype == STACK_ZERO) {
5461 /* helper can write anything into the stack */
5462 *stype = STACK_MISC;
5467 if (is_spilled_reg(&state->stack[spi]) &&
5468 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
5469 env->allow_ptr_leaks)) {
5471 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
5472 for (j = 0; j < BPF_REG_SIZE; j++)
5473 scrub_spilled_slot(&state->stack[spi].slot_type[j]);
5479 if (tnum_is_const(reg->var_off)) {
5480 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
5481 err_extra, regno, min_off, i - min_off, access_size);
5485 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5486 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
5487 err_extra, regno, tn_buf, i - min_off, access_size);
5491 /* reading any byte out of 8-byte 'spill_slot' will cause
5492 * the whole slot to be marked as 'read'
5494 mark_reg_read(env, &state->stack[spi].spilled_ptr,
5495 state->stack[spi].spilled_ptr.parent,
5497 /* We do not set REG_LIVE_WRITTEN for stack slot, as we can not
5498 * be sure that whether stack slot is written to or not. Hence,
5499 * we must still conservatively propagate reads upwards even if
5500 * helper may write to the entire memory range.
5503 return update_stack_depth(env, state, min_off);
5506 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
5507 int access_size, bool zero_size_allowed,
5508 struct bpf_call_arg_meta *meta)
5510 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5513 switch (base_type(reg->type)) {
5515 case PTR_TO_PACKET_META:
5516 return check_packet_access(env, regno, reg->off, access_size,
5518 case PTR_TO_MAP_KEY:
5519 if (meta && meta->raw_mode) {
5520 verbose(env, "R%d cannot write into %s\n", regno,
5521 reg_type_str(env, reg->type));
5524 return check_mem_region_access(env, regno, reg->off, access_size,
5525 reg->map_ptr->key_size, false);
5526 case PTR_TO_MAP_VALUE:
5527 if (check_map_access_type(env, regno, reg->off, access_size,
5528 meta && meta->raw_mode ? BPF_WRITE :
5531 return check_map_access(env, regno, reg->off, access_size,
5532 zero_size_allowed, ACCESS_HELPER);
5534 if (type_is_rdonly_mem(reg->type)) {
5535 if (meta && meta->raw_mode) {
5536 verbose(env, "R%d cannot write into %s\n", regno,
5537 reg_type_str(env, reg->type));
5541 return check_mem_region_access(env, regno, reg->off,
5542 access_size, reg->mem_size,
5545 if (type_is_rdonly_mem(reg->type)) {
5546 if (meta && meta->raw_mode) {
5547 verbose(env, "R%d cannot write into %s\n", regno,
5548 reg_type_str(env, reg->type));
5552 max_access = &env->prog->aux->max_rdonly_access;
5554 max_access = &env->prog->aux->max_rdwr_access;
5556 return check_buffer_access(env, reg, regno, reg->off,
5557 access_size, zero_size_allowed,
5560 return check_stack_range_initialized(
5562 regno, reg->off, access_size,
5563 zero_size_allowed, ACCESS_HELPER, meta);
5565 /* in case the function doesn't know how to access the context,
5566 * (because we are in a program of type SYSCALL for example), we
5567 * can not statically check its size.
5568 * Dynamically check it now.
5570 if (!env->ops->convert_ctx_access) {
5571 enum bpf_access_type atype = meta && meta->raw_mode ? BPF_WRITE : BPF_READ;
5572 int offset = access_size - 1;
5574 /* Allow zero-byte read from PTR_TO_CTX */
5575 if (access_size == 0)
5576 return zero_size_allowed ? 0 : -EACCES;
5578 return check_mem_access(env, env->insn_idx, regno, offset, BPF_B,
5583 default: /* scalar_value or invalid ptr */
5584 /* Allow zero-byte read from NULL, regardless of pointer type */
5585 if (zero_size_allowed && access_size == 0 &&
5586 register_is_null(reg))
5589 verbose(env, "R%d type=%s ", regno,
5590 reg_type_str(env, reg->type));
5591 verbose(env, "expected=%s\n", reg_type_str(env, PTR_TO_STACK));
5596 static int check_mem_size_reg(struct bpf_verifier_env *env,
5597 struct bpf_reg_state *reg, u32 regno,
5598 bool zero_size_allowed,
5599 struct bpf_call_arg_meta *meta)
5603 /* This is used to refine r0 return value bounds for helpers
5604 * that enforce this value as an upper bound on return values.
5605 * See do_refine_retval_range() for helpers that can refine
5606 * the return value. C type of helper is u32 so we pull register
5607 * bound from umax_value however, if negative verifier errors
5608 * out. Only upper bounds can be learned because retval is an
5609 * int type and negative retvals are allowed.
5611 meta->msize_max_value = reg->umax_value;
5613 /* The register is SCALAR_VALUE; the access check
5614 * happens using its boundaries.
5616 if (!tnum_is_const(reg->var_off))
5617 /* For unprivileged variable accesses, disable raw
5618 * mode so that the program is required to
5619 * initialize all the memory that the helper could
5620 * just partially fill up.
5624 if (reg->smin_value < 0) {
5625 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
5630 if (reg->umin_value == 0) {
5631 err = check_helper_mem_access(env, regno - 1, 0,
5638 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
5639 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
5643 err = check_helper_mem_access(env, regno - 1,
5645 zero_size_allowed, meta);
5647 err = mark_chain_precision(env, regno);
5651 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
5652 u32 regno, u32 mem_size)
5654 bool may_be_null = type_may_be_null(reg->type);
5655 struct bpf_reg_state saved_reg;
5656 struct bpf_call_arg_meta meta;
5659 if (register_is_null(reg))
5662 memset(&meta, 0, sizeof(meta));
5663 /* Assuming that the register contains a value check if the memory
5664 * access is safe. Temporarily save and restore the register's state as
5665 * the conversion shouldn't be visible to a caller.
5669 mark_ptr_not_null_reg(reg);
5672 err = check_helper_mem_access(env, regno, mem_size, true, &meta);
5673 /* Check access for BPF_WRITE */
5674 meta.raw_mode = true;
5675 err = err ?: check_helper_mem_access(env, regno, mem_size, true, &meta);
5683 static int check_kfunc_mem_size_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
5686 struct bpf_reg_state *mem_reg = &cur_regs(env)[regno - 1];
5687 bool may_be_null = type_may_be_null(mem_reg->type);
5688 struct bpf_reg_state saved_reg;
5689 struct bpf_call_arg_meta meta;
5692 WARN_ON_ONCE(regno < BPF_REG_2 || regno > BPF_REG_5);
5694 memset(&meta, 0, sizeof(meta));
5697 saved_reg = *mem_reg;
5698 mark_ptr_not_null_reg(mem_reg);
5701 err = check_mem_size_reg(env, reg, regno, true, &meta);
5702 /* Check access for BPF_WRITE */
5703 meta.raw_mode = true;
5704 err = err ?: check_mem_size_reg(env, reg, regno, true, &meta);
5707 *mem_reg = saved_reg;
5711 /* Implementation details:
5712 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL.
5713 * bpf_obj_new returns PTR_TO_BTF_ID | MEM_ALLOC | PTR_MAYBE_NULL.
5714 * Two bpf_map_lookups (even with the same key) will have different reg->id.
5715 * Two separate bpf_obj_new will also have different reg->id.
5716 * For traditional PTR_TO_MAP_VALUE or PTR_TO_BTF_ID | MEM_ALLOC, the verifier
5717 * clears reg->id after value_or_null->value transition, since the verifier only
5718 * cares about the range of access to valid map value pointer and doesn't care
5719 * about actual address of the map element.
5720 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
5721 * reg->id > 0 after value_or_null->value transition. By doing so
5722 * two bpf_map_lookups will be considered two different pointers that
5723 * point to different bpf_spin_locks. Likewise for pointers to allocated objects
5724 * returned from bpf_obj_new.
5725 * The verifier allows taking only one bpf_spin_lock at a time to avoid
5727 * Since only one bpf_spin_lock is allowed the checks are simpler than
5728 * reg_is_refcounted() logic. The verifier needs to remember only
5729 * one spin_lock instead of array of acquired_refs.
5730 * cur_state->active_lock remembers which map value element or allocated
5731 * object got locked and clears it after bpf_spin_unlock.
5733 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
5736 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5737 struct bpf_verifier_state *cur = env->cur_state;
5738 bool is_const = tnum_is_const(reg->var_off);
5739 u64 val = reg->var_off.value;
5740 struct bpf_map *map = NULL;
5741 struct btf *btf = NULL;
5742 struct btf_record *rec;
5746 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
5750 if (reg->type == PTR_TO_MAP_VALUE) {
5754 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
5762 rec = reg_btf_record(reg);
5763 if (!btf_record_has_field(rec, BPF_SPIN_LOCK)) {
5764 verbose(env, "%s '%s' has no valid bpf_spin_lock\n", map ? "map" : "local",
5765 map ? map->name : "kptr");
5768 if (rec->spin_lock_off != val + reg->off) {
5769 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock' that is at %d\n",
5770 val + reg->off, rec->spin_lock_off);
5774 if (cur->active_lock.ptr) {
5776 "Locking two bpf_spin_locks are not allowed\n");
5780 cur->active_lock.ptr = map;
5782 cur->active_lock.ptr = btf;
5783 cur->active_lock.id = reg->id;
5785 struct bpf_func_state *fstate = cur_func(env);
5794 if (!cur->active_lock.ptr) {
5795 verbose(env, "bpf_spin_unlock without taking a lock\n");
5798 if (cur->active_lock.ptr != ptr ||
5799 cur->active_lock.id != reg->id) {
5800 verbose(env, "bpf_spin_unlock of different lock\n");
5803 cur->active_lock.ptr = NULL;
5804 cur->active_lock.id = 0;
5806 for (i = fstate->acquired_refs - 1; i >= 0; i--) {
5809 /* Complain on error because this reference state cannot
5810 * be freed before this point, as bpf_spin_lock critical
5811 * section does not allow functions that release the
5812 * allocated object immediately.
5814 if (!fstate->refs[i].release_on_unlock)
5816 err = release_reference(env, fstate->refs[i].id);
5818 verbose(env, "failed to release release_on_unlock reference");
5826 static int process_timer_func(struct bpf_verifier_env *env, int regno,
5827 struct bpf_call_arg_meta *meta)
5829 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5830 bool is_const = tnum_is_const(reg->var_off);
5831 struct bpf_map *map = reg->map_ptr;
5832 u64 val = reg->var_off.value;
5836 "R%d doesn't have constant offset. bpf_timer has to be at the constant offset\n",
5841 verbose(env, "map '%s' has to have BTF in order to use bpf_timer\n",
5845 if (!btf_record_has_field(map->record, BPF_TIMER)) {
5846 verbose(env, "map '%s' has no valid bpf_timer\n", map->name);
5849 if (map->record->timer_off != val + reg->off) {
5850 verbose(env, "off %lld doesn't point to 'struct bpf_timer' that is at %d\n",
5851 val + reg->off, map->record->timer_off);
5854 if (meta->map_ptr) {
5855 verbose(env, "verifier bug. Two map pointers in a timer helper\n");
5858 meta->map_uid = reg->map_uid;
5859 meta->map_ptr = map;
5863 static int process_kptr_func(struct bpf_verifier_env *env, int regno,
5864 struct bpf_call_arg_meta *meta)
5866 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5867 struct bpf_map *map_ptr = reg->map_ptr;
5868 struct btf_field *kptr_field;
5871 if (!tnum_is_const(reg->var_off)) {
5873 "R%d doesn't have constant offset. kptr has to be at the constant offset\n",
5877 if (!map_ptr->btf) {
5878 verbose(env, "map '%s' has to have BTF in order to use bpf_kptr_xchg\n",
5882 if (!btf_record_has_field(map_ptr->record, BPF_KPTR)) {
5883 verbose(env, "map '%s' has no valid kptr\n", map_ptr->name);
5887 meta->map_ptr = map_ptr;
5888 kptr_off = reg->off + reg->var_off.value;
5889 kptr_field = btf_record_find(map_ptr->record, kptr_off, BPF_KPTR);
5891 verbose(env, "off=%d doesn't point to kptr\n", kptr_off);
5894 if (kptr_field->type != BPF_KPTR_REF) {
5895 verbose(env, "off=%d kptr isn't referenced kptr\n", kptr_off);
5898 meta->kptr_field = kptr_field;
5902 /* There are two register types representing a bpf_dynptr, one is PTR_TO_STACK
5903 * which points to a stack slot, and the other is CONST_PTR_TO_DYNPTR.
5905 * In both cases we deal with the first 8 bytes, but need to mark the next 8
5906 * bytes as STACK_DYNPTR in case of PTR_TO_STACK. In case of
5907 * CONST_PTR_TO_DYNPTR, we are guaranteed to get the beginning of the object.
5909 * Mutability of bpf_dynptr is at two levels, one is at the level of struct
5910 * bpf_dynptr itself, i.e. whether the helper is receiving a pointer to struct
5911 * bpf_dynptr or pointer to const struct bpf_dynptr. In the former case, it can
5912 * mutate the view of the dynptr and also possibly destroy it. In the latter
5913 * case, it cannot mutate the bpf_dynptr itself but it can still mutate the
5914 * memory that dynptr points to.
5916 * The verifier will keep track both levels of mutation (bpf_dynptr's in
5917 * reg->type and the memory's in reg->dynptr.type), but there is no support for
5918 * readonly dynptr view yet, hence only the first case is tracked and checked.
5920 * This is consistent with how C applies the const modifier to a struct object,
5921 * where the pointer itself inside bpf_dynptr becomes const but not what it
5924 * Helpers which do not mutate the bpf_dynptr set MEM_RDONLY in their argument
5925 * type, and declare it as 'const struct bpf_dynptr *' in their prototype.
5927 int process_dynptr_func(struct bpf_verifier_env *env, int regno,
5928 enum bpf_arg_type arg_type, struct bpf_call_arg_meta *meta)
5930 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5932 /* MEM_UNINIT and MEM_RDONLY are exclusive, when applied to an
5933 * ARG_PTR_TO_DYNPTR (or ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_*):
5935 if ((arg_type & (MEM_UNINIT | MEM_RDONLY)) == (MEM_UNINIT | MEM_RDONLY)) {
5936 verbose(env, "verifier internal error: misconfigured dynptr helper type flags\n");
5939 /* CONST_PTR_TO_DYNPTR already has fixed and var_off as 0 due to
5940 * check_func_arg_reg_off's logic. We only need to check offset
5941 * alignment for PTR_TO_STACK.
5943 if (reg->type == PTR_TO_STACK && (reg->off % BPF_REG_SIZE)) {
5944 verbose(env, "cannot pass in dynptr at an offset=%d\n", reg->off);
5947 /* MEM_UNINIT - Points to memory that is an appropriate candidate for
5948 * constructing a mutable bpf_dynptr object.
5950 * Currently, this is only possible with PTR_TO_STACK
5951 * pointing to a region of at least 16 bytes which doesn't
5952 * contain an existing bpf_dynptr.
5954 * MEM_RDONLY - Points to a initialized bpf_dynptr that will not be
5955 * mutated or destroyed. However, the memory it points to
5958 * None - Points to a initialized dynptr that can be mutated and
5959 * destroyed, including mutation of the memory it points
5962 if (arg_type & MEM_UNINIT) {
5963 if (!is_dynptr_reg_valid_uninit(env, reg)) {
5964 verbose(env, "Dynptr has to be an uninitialized dynptr\n");
5968 /* We only support one dynptr being uninitialized at the moment,
5969 * which is sufficient for the helper functions we have right now.
5971 if (meta->uninit_dynptr_regno) {
5972 verbose(env, "verifier internal error: multiple uninitialized dynptr args\n");
5976 meta->uninit_dynptr_regno = regno;
5977 } else /* MEM_RDONLY and None case from above */ {
5978 /* For the reg->type == PTR_TO_STACK case, bpf_dynptr is never const */
5979 if (reg->type == CONST_PTR_TO_DYNPTR && !(arg_type & MEM_RDONLY)) {
5980 verbose(env, "cannot pass pointer to const bpf_dynptr, the helper mutates it\n");
5984 if (!is_dynptr_reg_valid_init(env, reg)) {
5986 "Expected an initialized dynptr as arg #%d\n",
5991 /* Fold modifiers (in this case, MEM_RDONLY) when checking expected type */
5992 if (!is_dynptr_type_expected(env, reg, arg_type & ~MEM_RDONLY)) {
5993 const char *err_extra = "";
5995 switch (arg_type & DYNPTR_TYPE_FLAG_MASK) {
5996 case DYNPTR_TYPE_LOCAL:
5997 err_extra = "local";
5999 case DYNPTR_TYPE_RINGBUF:
6000 err_extra = "ringbuf";
6003 err_extra = "<unknown>";
6007 "Expected a dynptr of type %s as arg #%d\n",
6015 static bool arg_type_is_mem_size(enum bpf_arg_type type)
6017 return type == ARG_CONST_SIZE ||
6018 type == ARG_CONST_SIZE_OR_ZERO;
6021 static bool arg_type_is_release(enum bpf_arg_type type)
6023 return type & OBJ_RELEASE;
6026 static bool arg_type_is_dynptr(enum bpf_arg_type type)
6028 return base_type(type) == ARG_PTR_TO_DYNPTR;
6031 static int int_ptr_type_to_size(enum bpf_arg_type type)
6033 if (type == ARG_PTR_TO_INT)
6035 else if (type == ARG_PTR_TO_LONG)
6041 static int resolve_map_arg_type(struct bpf_verifier_env *env,
6042 const struct bpf_call_arg_meta *meta,
6043 enum bpf_arg_type *arg_type)
6045 if (!meta->map_ptr) {
6046 /* kernel subsystem misconfigured verifier */
6047 verbose(env, "invalid map_ptr to access map->type\n");
6051 switch (meta->map_ptr->map_type) {
6052 case BPF_MAP_TYPE_SOCKMAP:
6053 case BPF_MAP_TYPE_SOCKHASH:
6054 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
6055 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
6057 verbose(env, "invalid arg_type for sockmap/sockhash\n");
6061 case BPF_MAP_TYPE_BLOOM_FILTER:
6062 if (meta->func_id == BPF_FUNC_map_peek_elem)
6063 *arg_type = ARG_PTR_TO_MAP_VALUE;
6071 struct bpf_reg_types {
6072 const enum bpf_reg_type types[10];
6076 static const struct bpf_reg_types sock_types = {
6086 static const struct bpf_reg_types btf_id_sock_common_types = {
6093 PTR_TO_BTF_ID | PTR_TRUSTED,
6095 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
6099 static const struct bpf_reg_types mem_types = {
6107 PTR_TO_MEM | MEM_RINGBUF,
6112 static const struct bpf_reg_types int_ptr_types = {
6122 static const struct bpf_reg_types spin_lock_types = {
6125 PTR_TO_BTF_ID | MEM_ALLOC,
6129 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
6130 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
6131 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
6132 static const struct bpf_reg_types ringbuf_mem_types = { .types = { PTR_TO_MEM | MEM_RINGBUF } };
6133 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
6134 static const struct bpf_reg_types btf_ptr_types = {
6137 PTR_TO_BTF_ID | PTR_TRUSTED,
6138 PTR_TO_BTF_ID | MEM_RCU,
6141 static const struct bpf_reg_types percpu_btf_ptr_types = {
6143 PTR_TO_BTF_ID | MEM_PERCPU,
6144 PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED,
6147 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
6148 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
6149 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
6150 static const struct bpf_reg_types timer_types = { .types = { PTR_TO_MAP_VALUE } };
6151 static const struct bpf_reg_types kptr_types = { .types = { PTR_TO_MAP_VALUE } };
6152 static const struct bpf_reg_types dynptr_types = {
6155 CONST_PTR_TO_DYNPTR,
6159 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
6160 [ARG_PTR_TO_MAP_KEY] = &mem_types,
6161 [ARG_PTR_TO_MAP_VALUE] = &mem_types,
6162 [ARG_CONST_SIZE] = &scalar_types,
6163 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
6164 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
6165 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
6166 [ARG_PTR_TO_CTX] = &context_types,
6167 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
6169 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
6171 [ARG_PTR_TO_SOCKET] = &fullsock_types,
6172 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
6173 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
6174 [ARG_PTR_TO_MEM] = &mem_types,
6175 [ARG_PTR_TO_RINGBUF_MEM] = &ringbuf_mem_types,
6176 [ARG_PTR_TO_INT] = &int_ptr_types,
6177 [ARG_PTR_TO_LONG] = &int_ptr_types,
6178 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
6179 [ARG_PTR_TO_FUNC] = &func_ptr_types,
6180 [ARG_PTR_TO_STACK] = &stack_ptr_types,
6181 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
6182 [ARG_PTR_TO_TIMER] = &timer_types,
6183 [ARG_PTR_TO_KPTR] = &kptr_types,
6184 [ARG_PTR_TO_DYNPTR] = &dynptr_types,
6187 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
6188 enum bpf_arg_type arg_type,
6189 const u32 *arg_btf_id,
6190 struct bpf_call_arg_meta *meta)
6192 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
6193 enum bpf_reg_type expected, type = reg->type;
6194 const struct bpf_reg_types *compatible;
6197 compatible = compatible_reg_types[base_type(arg_type)];
6199 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
6203 /* ARG_PTR_TO_MEM + RDONLY is compatible with PTR_TO_MEM and PTR_TO_MEM + RDONLY,
6204 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM and NOT with PTR_TO_MEM + RDONLY
6206 * Same for MAYBE_NULL:
6208 * ARG_PTR_TO_MEM + MAYBE_NULL is compatible with PTR_TO_MEM and PTR_TO_MEM + MAYBE_NULL,
6209 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM but NOT with PTR_TO_MEM + MAYBE_NULL
6211 * Therefore we fold these flags depending on the arg_type before comparison.
6213 if (arg_type & MEM_RDONLY)
6214 type &= ~MEM_RDONLY;
6215 if (arg_type & PTR_MAYBE_NULL)
6216 type &= ~PTR_MAYBE_NULL;
6218 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
6219 expected = compatible->types[i];
6220 if (expected == NOT_INIT)
6223 if (type == expected)
6227 verbose(env, "R%d type=%s expected=", regno, reg_type_str(env, reg->type));
6228 for (j = 0; j + 1 < i; j++)
6229 verbose(env, "%s, ", reg_type_str(env, compatible->types[j]));
6230 verbose(env, "%s\n", reg_type_str(env, compatible->types[j]));
6234 if (reg->type == PTR_TO_BTF_ID || reg->type & PTR_TRUSTED) {
6235 /* For bpf_sk_release, it needs to match against first member
6236 * 'struct sock_common', hence make an exception for it. This
6237 * allows bpf_sk_release to work for multiple socket types.
6239 bool strict_type_match = arg_type_is_release(arg_type) &&
6240 meta->func_id != BPF_FUNC_sk_release;
6243 if (!compatible->btf_id) {
6244 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
6247 arg_btf_id = compatible->btf_id;
6250 if (meta->func_id == BPF_FUNC_kptr_xchg) {
6251 if (map_kptr_match_type(env, meta->kptr_field, reg, regno))
6254 if (arg_btf_id == BPF_PTR_POISON) {
6255 verbose(env, "verifier internal error:");
6256 verbose(env, "R%d has non-overwritten BPF_PTR_POISON type\n",
6261 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
6262 btf_vmlinux, *arg_btf_id,
6263 strict_type_match)) {
6264 verbose(env, "R%d is of type %s but %s is expected\n",
6265 regno, kernel_type_name(reg->btf, reg->btf_id),
6266 kernel_type_name(btf_vmlinux, *arg_btf_id));
6270 } else if (type_is_alloc(reg->type)) {
6271 if (meta->func_id != BPF_FUNC_spin_lock && meta->func_id != BPF_FUNC_spin_unlock) {
6272 verbose(env, "verifier internal error: unimplemented handling of MEM_ALLOC\n");
6280 int check_func_arg_reg_off(struct bpf_verifier_env *env,
6281 const struct bpf_reg_state *reg, int regno,
6282 enum bpf_arg_type arg_type)
6284 u32 type = reg->type;
6286 /* When referenced register is passed to release function, its fixed
6289 * We will check arg_type_is_release reg has ref_obj_id when storing
6290 * meta->release_regno.
6292 if (arg_type_is_release(arg_type)) {
6293 /* ARG_PTR_TO_DYNPTR with OBJ_RELEASE is a bit special, as it
6294 * may not directly point to the object being released, but to
6295 * dynptr pointing to such object, which might be at some offset
6296 * on the stack. In that case, we simply to fallback to the
6299 if (arg_type_is_dynptr(arg_type) && type == PTR_TO_STACK)
6301 /* Doing check_ptr_off_reg check for the offset will catch this
6302 * because fixed_off_ok is false, but checking here allows us
6303 * to give the user a better error message.
6306 verbose(env, "R%d must have zero offset when passed to release func or trusted arg to kfunc\n",
6310 return __check_ptr_off_reg(env, reg, regno, false);
6314 /* Pointer types where both fixed and variable offset is explicitly allowed: */
6317 case PTR_TO_PACKET_META:
6318 case PTR_TO_MAP_KEY:
6319 case PTR_TO_MAP_VALUE:
6321 case PTR_TO_MEM | MEM_RDONLY:
6322 case PTR_TO_MEM | MEM_RINGBUF:
6324 case PTR_TO_BUF | MEM_RDONLY:
6327 /* All the rest must be rejected, except PTR_TO_BTF_ID which allows
6331 case PTR_TO_BTF_ID | MEM_ALLOC:
6332 case PTR_TO_BTF_ID | PTR_TRUSTED:
6333 case PTR_TO_BTF_ID | MEM_RCU:
6334 case PTR_TO_BTF_ID | MEM_ALLOC | PTR_TRUSTED:
6335 /* When referenced PTR_TO_BTF_ID is passed to release function,
6336 * its fixed offset must be 0. In the other cases, fixed offset
6337 * can be non-zero. This was already checked above. So pass
6338 * fixed_off_ok as true to allow fixed offset for all other
6339 * cases. var_off always must be 0 for PTR_TO_BTF_ID, hence we
6340 * still need to do checks instead of returning.
6342 return __check_ptr_off_reg(env, reg, regno, true);
6344 return __check_ptr_off_reg(env, reg, regno, false);
6348 static u32 dynptr_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
6350 struct bpf_func_state *state = func(env, reg);
6353 if (reg->type == CONST_PTR_TO_DYNPTR)
6354 return reg->ref_obj_id;
6356 spi = get_spi(reg->off);
6357 return state->stack[spi].spilled_ptr.ref_obj_id;
6360 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
6361 struct bpf_call_arg_meta *meta,
6362 const struct bpf_func_proto *fn)
6364 u32 regno = BPF_REG_1 + arg;
6365 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
6366 enum bpf_arg_type arg_type = fn->arg_type[arg];
6367 enum bpf_reg_type type = reg->type;
6368 u32 *arg_btf_id = NULL;
6371 if (arg_type == ARG_DONTCARE)
6374 err = check_reg_arg(env, regno, SRC_OP);
6378 if (arg_type == ARG_ANYTHING) {
6379 if (is_pointer_value(env, regno)) {
6380 verbose(env, "R%d leaks addr into helper function\n",
6387 if (type_is_pkt_pointer(type) &&
6388 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
6389 verbose(env, "helper access to the packet is not allowed\n");
6393 if (base_type(arg_type) == ARG_PTR_TO_MAP_VALUE) {
6394 err = resolve_map_arg_type(env, meta, &arg_type);
6399 if (register_is_null(reg) && type_may_be_null(arg_type))
6400 /* A NULL register has a SCALAR_VALUE type, so skip
6403 goto skip_type_check;
6405 /* arg_btf_id and arg_size are in a union. */
6406 if (base_type(arg_type) == ARG_PTR_TO_BTF_ID ||
6407 base_type(arg_type) == ARG_PTR_TO_SPIN_LOCK)
6408 arg_btf_id = fn->arg_btf_id[arg];
6410 err = check_reg_type(env, regno, arg_type, arg_btf_id, meta);
6414 err = check_func_arg_reg_off(env, reg, regno, arg_type);
6419 if (arg_type_is_release(arg_type)) {
6420 if (arg_type_is_dynptr(arg_type)) {
6421 struct bpf_func_state *state = func(env, reg);
6424 /* Only dynptr created on stack can be released, thus
6425 * the get_spi and stack state checks for spilled_ptr
6426 * should only be done before process_dynptr_func for
6429 if (reg->type == PTR_TO_STACK) {
6430 spi = get_spi(reg->off);
6431 if (!is_spi_bounds_valid(state, spi, BPF_DYNPTR_NR_SLOTS) ||
6432 !state->stack[spi].spilled_ptr.ref_obj_id) {
6433 verbose(env, "arg %d is an unacquired reference\n", regno);
6437 verbose(env, "cannot release unowned const bpf_dynptr\n");
6440 } else if (!reg->ref_obj_id && !register_is_null(reg)) {
6441 verbose(env, "R%d must be referenced when passed to release function\n",
6445 if (meta->release_regno) {
6446 verbose(env, "verifier internal error: more than one release argument\n");
6449 meta->release_regno = regno;
6452 if (reg->ref_obj_id) {
6453 if (meta->ref_obj_id) {
6454 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
6455 regno, reg->ref_obj_id,
6459 meta->ref_obj_id = reg->ref_obj_id;
6462 switch (base_type(arg_type)) {
6463 case ARG_CONST_MAP_PTR:
6464 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
6465 if (meta->map_ptr) {
6466 /* Use map_uid (which is unique id of inner map) to reject:
6467 * inner_map1 = bpf_map_lookup_elem(outer_map, key1)
6468 * inner_map2 = bpf_map_lookup_elem(outer_map, key2)
6469 * if (inner_map1 && inner_map2) {
6470 * timer = bpf_map_lookup_elem(inner_map1);
6472 * // mismatch would have been allowed
6473 * bpf_timer_init(timer, inner_map2);
6476 * Comparing map_ptr is enough to distinguish normal and outer maps.
6478 if (meta->map_ptr != reg->map_ptr ||
6479 meta->map_uid != reg->map_uid) {
6481 "timer pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
6482 meta->map_uid, reg->map_uid);
6486 meta->map_ptr = reg->map_ptr;
6487 meta->map_uid = reg->map_uid;
6489 case ARG_PTR_TO_MAP_KEY:
6490 /* bpf_map_xxx(..., map_ptr, ..., key) call:
6491 * check that [key, key + map->key_size) are within
6492 * stack limits and initialized
6494 if (!meta->map_ptr) {
6495 /* in function declaration map_ptr must come before
6496 * map_key, so that it's verified and known before
6497 * we have to check map_key here. Otherwise it means
6498 * that kernel subsystem misconfigured verifier
6500 verbose(env, "invalid map_ptr to access map->key\n");
6503 err = check_helper_mem_access(env, regno,
6504 meta->map_ptr->key_size, false,
6507 case ARG_PTR_TO_MAP_VALUE:
6508 if (type_may_be_null(arg_type) && register_is_null(reg))
6511 /* bpf_map_xxx(..., map_ptr, ..., value) call:
6512 * check [value, value + map->value_size) validity
6514 if (!meta->map_ptr) {
6515 /* kernel subsystem misconfigured verifier */
6516 verbose(env, "invalid map_ptr to access map->value\n");
6519 meta->raw_mode = arg_type & MEM_UNINIT;
6520 err = check_helper_mem_access(env, regno,
6521 meta->map_ptr->value_size, false,
6524 case ARG_PTR_TO_PERCPU_BTF_ID:
6526 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
6529 meta->ret_btf = reg->btf;
6530 meta->ret_btf_id = reg->btf_id;
6532 case ARG_PTR_TO_SPIN_LOCK:
6533 if (meta->func_id == BPF_FUNC_spin_lock) {
6534 err = process_spin_lock(env, regno, true);
6537 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
6538 err = process_spin_lock(env, regno, false);
6542 verbose(env, "verifier internal error\n");
6546 case ARG_PTR_TO_TIMER:
6547 err = process_timer_func(env, regno, meta);
6551 case ARG_PTR_TO_FUNC:
6552 meta->subprogno = reg->subprogno;
6554 case ARG_PTR_TO_MEM:
6555 /* The access to this pointer is only checked when we hit the
6556 * next is_mem_size argument below.
6558 meta->raw_mode = arg_type & MEM_UNINIT;
6559 if (arg_type & MEM_FIXED_SIZE) {
6560 err = check_helper_mem_access(env, regno,
6561 fn->arg_size[arg], false,
6565 case ARG_CONST_SIZE:
6566 err = check_mem_size_reg(env, reg, regno, false, meta);
6568 case ARG_CONST_SIZE_OR_ZERO:
6569 err = check_mem_size_reg(env, reg, regno, true, meta);
6571 case ARG_PTR_TO_DYNPTR:
6572 err = process_dynptr_func(env, regno, arg_type, meta);
6576 case ARG_CONST_ALLOC_SIZE_OR_ZERO:
6577 if (!tnum_is_const(reg->var_off)) {
6578 verbose(env, "R%d is not a known constant'\n",
6582 meta->mem_size = reg->var_off.value;
6583 err = mark_chain_precision(env, regno);
6587 case ARG_PTR_TO_INT:
6588 case ARG_PTR_TO_LONG:
6590 int size = int_ptr_type_to_size(arg_type);
6592 err = check_helper_mem_access(env, regno, size, false, meta);
6595 err = check_ptr_alignment(env, reg, 0, size, true);
6598 case ARG_PTR_TO_CONST_STR:
6600 struct bpf_map *map = reg->map_ptr;
6605 if (!bpf_map_is_rdonly(map)) {
6606 verbose(env, "R%d does not point to a readonly map'\n", regno);
6610 if (!tnum_is_const(reg->var_off)) {
6611 verbose(env, "R%d is not a constant address'\n", regno);
6615 if (!map->ops->map_direct_value_addr) {
6616 verbose(env, "no direct value access support for this map type\n");
6620 err = check_map_access(env, regno, reg->off,
6621 map->value_size - reg->off, false,
6626 map_off = reg->off + reg->var_off.value;
6627 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
6629 verbose(env, "direct value access on string failed\n");
6633 str_ptr = (char *)(long)(map_addr);
6634 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
6635 verbose(env, "string is not zero-terminated\n");
6640 case ARG_PTR_TO_KPTR:
6641 err = process_kptr_func(env, regno, meta);
6650 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
6652 enum bpf_attach_type eatype = env->prog->expected_attach_type;
6653 enum bpf_prog_type type = resolve_prog_type(env->prog);
6655 if (func_id != BPF_FUNC_map_update_elem)
6658 /* It's not possible to get access to a locked struct sock in these
6659 * contexts, so updating is safe.
6662 case BPF_PROG_TYPE_TRACING:
6663 if (eatype == BPF_TRACE_ITER)
6666 case BPF_PROG_TYPE_SOCKET_FILTER:
6667 case BPF_PROG_TYPE_SCHED_CLS:
6668 case BPF_PROG_TYPE_SCHED_ACT:
6669 case BPF_PROG_TYPE_XDP:
6670 case BPF_PROG_TYPE_SK_REUSEPORT:
6671 case BPF_PROG_TYPE_FLOW_DISSECTOR:
6672 case BPF_PROG_TYPE_SK_LOOKUP:
6678 verbose(env, "cannot update sockmap in this context\n");
6682 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
6684 return env->prog->jit_requested &&
6685 bpf_jit_supports_subprog_tailcalls();
6688 static int check_map_func_compatibility(struct bpf_verifier_env *env,
6689 struct bpf_map *map, int func_id)
6694 /* We need a two way check, first is from map perspective ... */
6695 switch (map->map_type) {
6696 case BPF_MAP_TYPE_PROG_ARRAY:
6697 if (func_id != BPF_FUNC_tail_call)
6700 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
6701 if (func_id != BPF_FUNC_perf_event_read &&
6702 func_id != BPF_FUNC_perf_event_output &&
6703 func_id != BPF_FUNC_skb_output &&
6704 func_id != BPF_FUNC_perf_event_read_value &&
6705 func_id != BPF_FUNC_xdp_output)
6708 case BPF_MAP_TYPE_RINGBUF:
6709 if (func_id != BPF_FUNC_ringbuf_output &&
6710 func_id != BPF_FUNC_ringbuf_reserve &&
6711 func_id != BPF_FUNC_ringbuf_query &&
6712 func_id != BPF_FUNC_ringbuf_reserve_dynptr &&
6713 func_id != BPF_FUNC_ringbuf_submit_dynptr &&
6714 func_id != BPF_FUNC_ringbuf_discard_dynptr)
6717 case BPF_MAP_TYPE_USER_RINGBUF:
6718 if (func_id != BPF_FUNC_user_ringbuf_drain)
6721 case BPF_MAP_TYPE_STACK_TRACE:
6722 if (func_id != BPF_FUNC_get_stackid)
6725 case BPF_MAP_TYPE_CGROUP_ARRAY:
6726 if (func_id != BPF_FUNC_skb_under_cgroup &&
6727 func_id != BPF_FUNC_current_task_under_cgroup)
6730 case BPF_MAP_TYPE_CGROUP_STORAGE:
6731 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
6732 if (func_id != BPF_FUNC_get_local_storage)
6735 case BPF_MAP_TYPE_DEVMAP:
6736 case BPF_MAP_TYPE_DEVMAP_HASH:
6737 if (func_id != BPF_FUNC_redirect_map &&
6738 func_id != BPF_FUNC_map_lookup_elem)
6741 /* Restrict bpf side of cpumap and xskmap, open when use-cases
6744 case BPF_MAP_TYPE_CPUMAP:
6745 if (func_id != BPF_FUNC_redirect_map)
6748 case BPF_MAP_TYPE_XSKMAP:
6749 if (func_id != BPF_FUNC_redirect_map &&
6750 func_id != BPF_FUNC_map_lookup_elem)
6753 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
6754 case BPF_MAP_TYPE_HASH_OF_MAPS:
6755 if (func_id != BPF_FUNC_map_lookup_elem)
6758 case BPF_MAP_TYPE_SOCKMAP:
6759 if (func_id != BPF_FUNC_sk_redirect_map &&
6760 func_id != BPF_FUNC_sock_map_update &&
6761 func_id != BPF_FUNC_map_delete_elem &&
6762 func_id != BPF_FUNC_msg_redirect_map &&
6763 func_id != BPF_FUNC_sk_select_reuseport &&
6764 func_id != BPF_FUNC_map_lookup_elem &&
6765 !may_update_sockmap(env, func_id))
6768 case BPF_MAP_TYPE_SOCKHASH:
6769 if (func_id != BPF_FUNC_sk_redirect_hash &&
6770 func_id != BPF_FUNC_sock_hash_update &&
6771 func_id != BPF_FUNC_map_delete_elem &&
6772 func_id != BPF_FUNC_msg_redirect_hash &&
6773 func_id != BPF_FUNC_sk_select_reuseport &&
6774 func_id != BPF_FUNC_map_lookup_elem &&
6775 !may_update_sockmap(env, func_id))
6778 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
6779 if (func_id != BPF_FUNC_sk_select_reuseport)
6782 case BPF_MAP_TYPE_QUEUE:
6783 case BPF_MAP_TYPE_STACK:
6784 if (func_id != BPF_FUNC_map_peek_elem &&
6785 func_id != BPF_FUNC_map_pop_elem &&
6786 func_id != BPF_FUNC_map_push_elem)
6789 case BPF_MAP_TYPE_SK_STORAGE:
6790 if (func_id != BPF_FUNC_sk_storage_get &&
6791 func_id != BPF_FUNC_sk_storage_delete)
6794 case BPF_MAP_TYPE_INODE_STORAGE:
6795 if (func_id != BPF_FUNC_inode_storage_get &&
6796 func_id != BPF_FUNC_inode_storage_delete)
6799 case BPF_MAP_TYPE_TASK_STORAGE:
6800 if (func_id != BPF_FUNC_task_storage_get &&
6801 func_id != BPF_FUNC_task_storage_delete)
6804 case BPF_MAP_TYPE_CGRP_STORAGE:
6805 if (func_id != BPF_FUNC_cgrp_storage_get &&
6806 func_id != BPF_FUNC_cgrp_storage_delete)
6809 case BPF_MAP_TYPE_BLOOM_FILTER:
6810 if (func_id != BPF_FUNC_map_peek_elem &&
6811 func_id != BPF_FUNC_map_push_elem)
6818 /* ... and second from the function itself. */
6820 case BPF_FUNC_tail_call:
6821 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
6823 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
6824 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
6828 case BPF_FUNC_perf_event_read:
6829 case BPF_FUNC_perf_event_output:
6830 case BPF_FUNC_perf_event_read_value:
6831 case BPF_FUNC_skb_output:
6832 case BPF_FUNC_xdp_output:
6833 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
6836 case BPF_FUNC_ringbuf_output:
6837 case BPF_FUNC_ringbuf_reserve:
6838 case BPF_FUNC_ringbuf_query:
6839 case BPF_FUNC_ringbuf_reserve_dynptr:
6840 case BPF_FUNC_ringbuf_submit_dynptr:
6841 case BPF_FUNC_ringbuf_discard_dynptr:
6842 if (map->map_type != BPF_MAP_TYPE_RINGBUF)
6845 case BPF_FUNC_user_ringbuf_drain:
6846 if (map->map_type != BPF_MAP_TYPE_USER_RINGBUF)
6849 case BPF_FUNC_get_stackid:
6850 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
6853 case BPF_FUNC_current_task_under_cgroup:
6854 case BPF_FUNC_skb_under_cgroup:
6855 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
6858 case BPF_FUNC_redirect_map:
6859 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
6860 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
6861 map->map_type != BPF_MAP_TYPE_CPUMAP &&
6862 map->map_type != BPF_MAP_TYPE_XSKMAP)
6865 case BPF_FUNC_sk_redirect_map:
6866 case BPF_FUNC_msg_redirect_map:
6867 case BPF_FUNC_sock_map_update:
6868 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
6871 case BPF_FUNC_sk_redirect_hash:
6872 case BPF_FUNC_msg_redirect_hash:
6873 case BPF_FUNC_sock_hash_update:
6874 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
6877 case BPF_FUNC_get_local_storage:
6878 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
6879 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
6882 case BPF_FUNC_sk_select_reuseport:
6883 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
6884 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
6885 map->map_type != BPF_MAP_TYPE_SOCKHASH)
6888 case BPF_FUNC_map_pop_elem:
6889 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
6890 map->map_type != BPF_MAP_TYPE_STACK)
6893 case BPF_FUNC_map_peek_elem:
6894 case BPF_FUNC_map_push_elem:
6895 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
6896 map->map_type != BPF_MAP_TYPE_STACK &&
6897 map->map_type != BPF_MAP_TYPE_BLOOM_FILTER)
6900 case BPF_FUNC_map_lookup_percpu_elem:
6901 if (map->map_type != BPF_MAP_TYPE_PERCPU_ARRAY &&
6902 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
6903 map->map_type != BPF_MAP_TYPE_LRU_PERCPU_HASH)
6906 case BPF_FUNC_sk_storage_get:
6907 case BPF_FUNC_sk_storage_delete:
6908 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
6911 case BPF_FUNC_inode_storage_get:
6912 case BPF_FUNC_inode_storage_delete:
6913 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
6916 case BPF_FUNC_task_storage_get:
6917 case BPF_FUNC_task_storage_delete:
6918 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
6921 case BPF_FUNC_cgrp_storage_get:
6922 case BPF_FUNC_cgrp_storage_delete:
6923 if (map->map_type != BPF_MAP_TYPE_CGRP_STORAGE)
6932 verbose(env, "cannot pass map_type %d into func %s#%d\n",
6933 map->map_type, func_id_name(func_id), func_id);
6937 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
6941 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
6943 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
6945 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
6947 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
6949 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
6952 /* We only support one arg being in raw mode at the moment,
6953 * which is sufficient for the helper functions we have
6959 static bool check_args_pair_invalid(const struct bpf_func_proto *fn, int arg)
6961 bool is_fixed = fn->arg_type[arg] & MEM_FIXED_SIZE;
6962 bool has_size = fn->arg_size[arg] != 0;
6963 bool is_next_size = false;
6965 if (arg + 1 < ARRAY_SIZE(fn->arg_type))
6966 is_next_size = arg_type_is_mem_size(fn->arg_type[arg + 1]);
6968 if (base_type(fn->arg_type[arg]) != ARG_PTR_TO_MEM)
6969 return is_next_size;
6971 return has_size == is_next_size || is_next_size == is_fixed;
6974 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
6976 /* bpf_xxx(..., buf, len) call will access 'len'
6977 * bytes from memory 'buf'. Both arg types need
6978 * to be paired, so make sure there's no buggy
6979 * helper function specification.
6981 if (arg_type_is_mem_size(fn->arg1_type) ||
6982 check_args_pair_invalid(fn, 0) ||
6983 check_args_pair_invalid(fn, 1) ||
6984 check_args_pair_invalid(fn, 2) ||
6985 check_args_pair_invalid(fn, 3) ||
6986 check_args_pair_invalid(fn, 4))
6992 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
6996 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
6997 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_BTF_ID)
6998 return !!fn->arg_btf_id[i];
6999 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_SPIN_LOCK)
7000 return fn->arg_btf_id[i] == BPF_PTR_POISON;
7001 if (base_type(fn->arg_type[i]) != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i] &&
7002 /* arg_btf_id and arg_size are in a union. */
7003 (base_type(fn->arg_type[i]) != ARG_PTR_TO_MEM ||
7004 !(fn->arg_type[i] & MEM_FIXED_SIZE)))
7011 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
7013 return check_raw_mode_ok(fn) &&
7014 check_arg_pair_ok(fn) &&
7015 check_btf_id_ok(fn) ? 0 : -EINVAL;
7018 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
7019 * are now invalid, so turn them into unknown SCALAR_VALUE.
7021 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
7023 struct bpf_func_state *state;
7024 struct bpf_reg_state *reg;
7026 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
7027 if (reg_is_pkt_pointer_any(reg))
7028 __mark_reg_unknown(env, reg);
7034 BEYOND_PKT_END = -2,
7037 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
7039 struct bpf_func_state *state = vstate->frame[vstate->curframe];
7040 struct bpf_reg_state *reg = &state->regs[regn];
7042 if (reg->type != PTR_TO_PACKET)
7043 /* PTR_TO_PACKET_META is not supported yet */
7046 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
7047 * How far beyond pkt_end it goes is unknown.
7048 * if (!range_open) it's the case of pkt >= pkt_end
7049 * if (range_open) it's the case of pkt > pkt_end
7050 * hence this pointer is at least 1 byte bigger than pkt_end
7053 reg->range = BEYOND_PKT_END;
7055 reg->range = AT_PKT_END;
7058 /* The pointer with the specified id has released its reference to kernel
7059 * resources. Identify all copies of the same pointer and clear the reference.
7061 static int release_reference(struct bpf_verifier_env *env,
7064 struct bpf_func_state *state;
7065 struct bpf_reg_state *reg;
7068 err = release_reference_state(cur_func(env), ref_obj_id);
7072 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
7073 if (reg->ref_obj_id == ref_obj_id) {
7074 if (!env->allow_ptr_leaks)
7075 __mark_reg_not_init(env, reg);
7077 __mark_reg_unknown(env, reg);
7084 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
7085 struct bpf_reg_state *regs)
7089 /* after the call registers r0 - r5 were scratched */
7090 for (i = 0; i < CALLER_SAVED_REGS; i++) {
7091 mark_reg_not_init(env, regs, caller_saved[i]);
7092 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
7096 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
7097 struct bpf_func_state *caller,
7098 struct bpf_func_state *callee,
7101 static int set_callee_state(struct bpf_verifier_env *env,
7102 struct bpf_func_state *caller,
7103 struct bpf_func_state *callee, int insn_idx);
7105 static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
7106 int *insn_idx, int subprog,
7107 set_callee_state_fn set_callee_state_cb)
7109 struct bpf_verifier_state *state = env->cur_state;
7110 struct bpf_func_info_aux *func_info_aux;
7111 struct bpf_func_state *caller, *callee;
7113 bool is_global = false;
7115 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
7116 verbose(env, "the call stack of %d frames is too deep\n",
7117 state->curframe + 2);
7121 caller = state->frame[state->curframe];
7122 if (state->frame[state->curframe + 1]) {
7123 verbose(env, "verifier bug. Frame %d already allocated\n",
7124 state->curframe + 1);
7128 func_info_aux = env->prog->aux->func_info_aux;
7130 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
7131 err = btf_check_subprog_call(env, subprog, caller->regs);
7136 verbose(env, "Caller passes invalid args into func#%d\n",
7140 if (env->log.level & BPF_LOG_LEVEL)
7142 "Func#%d is global and valid. Skipping.\n",
7144 clear_caller_saved_regs(env, caller->regs);
7146 /* All global functions return a 64-bit SCALAR_VALUE */
7147 mark_reg_unknown(env, caller->regs, BPF_REG_0);
7148 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
7150 /* continue with next insn after call */
7155 /* set_callee_state is used for direct subprog calls, but we are
7156 * interested in validating only BPF helpers that can call subprogs as
7159 if (set_callee_state_cb != set_callee_state && !is_callback_calling_function(insn->imm)) {
7160 verbose(env, "verifier bug: helper %s#%d is not marked as callback-calling\n",
7161 func_id_name(insn->imm), insn->imm);
7165 if (insn->code == (BPF_JMP | BPF_CALL) &&
7166 insn->src_reg == 0 &&
7167 insn->imm == BPF_FUNC_timer_set_callback) {
7168 struct bpf_verifier_state *async_cb;
7170 /* there is no real recursion here. timer callbacks are async */
7171 env->subprog_info[subprog].is_async_cb = true;
7172 async_cb = push_async_cb(env, env->subprog_info[subprog].start,
7173 *insn_idx, subprog);
7176 callee = async_cb->frame[0];
7177 callee->async_entry_cnt = caller->async_entry_cnt + 1;
7179 /* Convert bpf_timer_set_callback() args into timer callback args */
7180 err = set_callee_state_cb(env, caller, callee, *insn_idx);
7184 clear_caller_saved_regs(env, caller->regs);
7185 mark_reg_unknown(env, caller->regs, BPF_REG_0);
7186 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
7187 /* continue with next insn after call */
7191 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
7194 state->frame[state->curframe + 1] = callee;
7196 /* callee cannot access r0, r6 - r9 for reading and has to write
7197 * into its own stack before reading from it.
7198 * callee can read/write into caller's stack
7200 init_func_state(env, callee,
7201 /* remember the callsite, it will be used by bpf_exit */
7202 *insn_idx /* callsite */,
7203 state->curframe + 1 /* frameno within this callchain */,
7204 subprog /* subprog number within this prog */);
7206 /* Transfer references to the callee */
7207 err = copy_reference_state(callee, caller);
7211 err = set_callee_state_cb(env, caller, callee, *insn_idx);
7215 clear_caller_saved_regs(env, caller->regs);
7217 /* only increment it after check_reg_arg() finished */
7220 /* and go analyze first insn of the callee */
7221 *insn_idx = env->subprog_info[subprog].start - 1;
7223 if (env->log.level & BPF_LOG_LEVEL) {
7224 verbose(env, "caller:\n");
7225 print_verifier_state(env, caller, true);
7226 verbose(env, "callee:\n");
7227 print_verifier_state(env, callee, true);
7232 free_func_state(callee);
7233 state->frame[state->curframe + 1] = NULL;
7237 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
7238 struct bpf_func_state *caller,
7239 struct bpf_func_state *callee)
7241 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
7242 * void *callback_ctx, u64 flags);
7243 * callback_fn(struct bpf_map *map, void *key, void *value,
7244 * void *callback_ctx);
7246 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
7248 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
7249 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
7250 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
7252 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
7253 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
7254 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
7256 /* pointer to stack or null */
7257 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
7260 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
7264 static int set_callee_state(struct bpf_verifier_env *env,
7265 struct bpf_func_state *caller,
7266 struct bpf_func_state *callee, int insn_idx)
7270 /* copy r1 - r5 args that callee can access. The copy includes parent
7271 * pointers, which connects us up to the liveness chain
7273 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
7274 callee->regs[i] = caller->regs[i];
7278 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
7281 int subprog, target_insn;
7283 target_insn = *insn_idx + insn->imm + 1;
7284 subprog = find_subprog(env, target_insn);
7286 verbose(env, "verifier bug. No program starts at insn %d\n",
7291 return __check_func_call(env, insn, insn_idx, subprog, set_callee_state);
7294 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
7295 struct bpf_func_state *caller,
7296 struct bpf_func_state *callee,
7299 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
7300 struct bpf_map *map;
7303 if (bpf_map_ptr_poisoned(insn_aux)) {
7304 verbose(env, "tail_call abusing map_ptr\n");
7308 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
7309 if (!map->ops->map_set_for_each_callback_args ||
7310 !map->ops->map_for_each_callback) {
7311 verbose(env, "callback function not allowed for map\n");
7315 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
7319 callee->in_callback_fn = true;
7320 callee->callback_ret_range = tnum_range(0, 1);
7324 static int set_loop_callback_state(struct bpf_verifier_env *env,
7325 struct bpf_func_state *caller,
7326 struct bpf_func_state *callee,
7329 /* bpf_loop(u32 nr_loops, void *callback_fn, void *callback_ctx,
7331 * callback_fn(u32 index, void *callback_ctx);
7333 callee->regs[BPF_REG_1].type = SCALAR_VALUE;
7334 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
7337 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
7338 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
7339 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
7341 callee->in_callback_fn = true;
7342 callee->callback_ret_range = tnum_range(0, 1);
7346 static int set_timer_callback_state(struct bpf_verifier_env *env,
7347 struct bpf_func_state *caller,
7348 struct bpf_func_state *callee,
7351 struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
7353 /* bpf_timer_set_callback(struct bpf_timer *timer, void *callback_fn);
7354 * callback_fn(struct bpf_map *map, void *key, void *value);
7356 callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP;
7357 __mark_reg_known_zero(&callee->regs[BPF_REG_1]);
7358 callee->regs[BPF_REG_1].map_ptr = map_ptr;
7360 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
7361 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
7362 callee->regs[BPF_REG_2].map_ptr = map_ptr;
7364 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
7365 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
7366 callee->regs[BPF_REG_3].map_ptr = map_ptr;
7369 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
7370 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
7371 callee->in_async_callback_fn = true;
7372 callee->callback_ret_range = tnum_range(0, 1);
7376 static int set_find_vma_callback_state(struct bpf_verifier_env *env,
7377 struct bpf_func_state *caller,
7378 struct bpf_func_state *callee,
7381 /* bpf_find_vma(struct task_struct *task, u64 addr,
7382 * void *callback_fn, void *callback_ctx, u64 flags)
7383 * (callback_fn)(struct task_struct *task,
7384 * struct vm_area_struct *vma, void *callback_ctx);
7386 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
7388 callee->regs[BPF_REG_2].type = PTR_TO_BTF_ID;
7389 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
7390 callee->regs[BPF_REG_2].btf = btf_vmlinux;
7391 callee->regs[BPF_REG_2].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_VMA],
7393 /* pointer to stack or null */
7394 callee->regs[BPF_REG_3] = caller->regs[BPF_REG_4];
7397 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
7398 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
7399 callee->in_callback_fn = true;
7400 callee->callback_ret_range = tnum_range(0, 1);
7404 static int set_user_ringbuf_callback_state(struct bpf_verifier_env *env,
7405 struct bpf_func_state *caller,
7406 struct bpf_func_state *callee,
7409 /* bpf_user_ringbuf_drain(struct bpf_map *map, void *callback_fn, void
7410 * callback_ctx, u64 flags);
7411 * callback_fn(const struct bpf_dynptr_t* dynptr, void *callback_ctx);
7413 __mark_reg_not_init(env, &callee->regs[BPF_REG_0]);
7414 mark_dynptr_cb_reg(&callee->regs[BPF_REG_1], BPF_DYNPTR_TYPE_LOCAL);
7415 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
7418 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
7419 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
7420 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
7422 callee->in_callback_fn = true;
7423 callee->callback_ret_range = tnum_range(0, 1);
7427 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
7429 struct bpf_verifier_state *state = env->cur_state;
7430 struct bpf_func_state *caller, *callee;
7431 struct bpf_reg_state *r0;
7434 callee = state->frame[state->curframe];
7435 r0 = &callee->regs[BPF_REG_0];
7436 if (r0->type == PTR_TO_STACK) {
7437 /* technically it's ok to return caller's stack pointer
7438 * (or caller's caller's pointer) back to the caller,
7439 * since these pointers are valid. Only current stack
7440 * pointer will be invalid as soon as function exits,
7441 * but let's be conservative
7443 verbose(env, "cannot return stack pointer to the caller\n");
7447 caller = state->frame[state->curframe - 1];
7448 if (callee->in_callback_fn) {
7449 /* enforce R0 return value range [0, 1]. */
7450 struct tnum range = callee->callback_ret_range;
7452 if (r0->type != SCALAR_VALUE) {
7453 verbose(env, "R0 not a scalar value\n");
7456 if (!tnum_in(range, r0->var_off)) {
7457 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
7461 /* return to the caller whatever r0 had in the callee */
7462 caller->regs[BPF_REG_0] = *r0;
7465 /* callback_fn frame should have released its own additions to parent's
7466 * reference state at this point, or check_reference_leak would
7467 * complain, hence it must be the same as the caller. There is no need
7470 if (!callee->in_callback_fn) {
7471 /* Transfer references to the caller */
7472 err = copy_reference_state(caller, callee);
7477 *insn_idx = callee->callsite + 1;
7478 if (env->log.level & BPF_LOG_LEVEL) {
7479 verbose(env, "returning from callee:\n");
7480 print_verifier_state(env, callee, true);
7481 verbose(env, "to caller at %d:\n", *insn_idx);
7482 print_verifier_state(env, caller, true);
7484 /* clear everything in the callee */
7485 free_func_state(callee);
7486 state->frame[state->curframe--] = NULL;
7490 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
7492 struct bpf_call_arg_meta *meta)
7494 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
7496 if (ret_type != RET_INTEGER ||
7497 (func_id != BPF_FUNC_get_stack &&
7498 func_id != BPF_FUNC_get_task_stack &&
7499 func_id != BPF_FUNC_probe_read_str &&
7500 func_id != BPF_FUNC_probe_read_kernel_str &&
7501 func_id != BPF_FUNC_probe_read_user_str))
7504 ret_reg->smax_value = meta->msize_max_value;
7505 ret_reg->s32_max_value = meta->msize_max_value;
7506 ret_reg->smin_value = -MAX_ERRNO;
7507 ret_reg->s32_min_value = -MAX_ERRNO;
7508 reg_bounds_sync(ret_reg);
7512 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
7513 int func_id, int insn_idx)
7515 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
7516 struct bpf_map *map = meta->map_ptr;
7518 if (func_id != BPF_FUNC_tail_call &&
7519 func_id != BPF_FUNC_map_lookup_elem &&
7520 func_id != BPF_FUNC_map_update_elem &&
7521 func_id != BPF_FUNC_map_delete_elem &&
7522 func_id != BPF_FUNC_map_push_elem &&
7523 func_id != BPF_FUNC_map_pop_elem &&
7524 func_id != BPF_FUNC_map_peek_elem &&
7525 func_id != BPF_FUNC_for_each_map_elem &&
7526 func_id != BPF_FUNC_redirect_map &&
7527 func_id != BPF_FUNC_map_lookup_percpu_elem)
7531 verbose(env, "kernel subsystem misconfigured verifier\n");
7535 /* In case of read-only, some additional restrictions
7536 * need to be applied in order to prevent altering the
7537 * state of the map from program side.
7539 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
7540 (func_id == BPF_FUNC_map_delete_elem ||
7541 func_id == BPF_FUNC_map_update_elem ||
7542 func_id == BPF_FUNC_map_push_elem ||
7543 func_id == BPF_FUNC_map_pop_elem)) {
7544 verbose(env, "write into map forbidden\n");
7548 if (!BPF_MAP_PTR(aux->map_ptr_state))
7549 bpf_map_ptr_store(aux, meta->map_ptr,
7550 !meta->map_ptr->bypass_spec_v1);
7551 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
7552 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
7553 !meta->map_ptr->bypass_spec_v1);
7558 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
7559 int func_id, int insn_idx)
7561 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
7562 struct bpf_reg_state *regs = cur_regs(env), *reg;
7563 struct bpf_map *map = meta->map_ptr;
7567 if (func_id != BPF_FUNC_tail_call)
7569 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
7570 verbose(env, "kernel subsystem misconfigured verifier\n");
7574 reg = ®s[BPF_REG_3];
7575 val = reg->var_off.value;
7576 max = map->max_entries;
7578 if (!(register_is_const(reg) && val < max)) {
7579 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
7583 err = mark_chain_precision(env, BPF_REG_3);
7586 if (bpf_map_key_unseen(aux))
7587 bpf_map_key_store(aux, val);
7588 else if (!bpf_map_key_poisoned(aux) &&
7589 bpf_map_key_immediate(aux) != val)
7590 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
7594 static int check_reference_leak(struct bpf_verifier_env *env)
7596 struct bpf_func_state *state = cur_func(env);
7597 bool refs_lingering = false;
7600 if (state->frameno && !state->in_callback_fn)
7603 for (i = 0; i < state->acquired_refs; i++) {
7604 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno)
7606 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
7607 state->refs[i].id, state->refs[i].insn_idx);
7608 refs_lingering = true;
7610 return refs_lingering ? -EINVAL : 0;
7613 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
7614 struct bpf_reg_state *regs)
7616 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
7617 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
7618 struct bpf_map *fmt_map = fmt_reg->map_ptr;
7619 struct bpf_bprintf_data data = {};
7620 int err, fmt_map_off, num_args;
7624 /* data must be an array of u64 */
7625 if (data_len_reg->var_off.value % 8)
7627 num_args = data_len_reg->var_off.value / 8;
7629 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
7630 * and map_direct_value_addr is set.
7632 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
7633 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
7636 verbose(env, "verifier bug\n");
7639 fmt = (char *)(long)fmt_addr + fmt_map_off;
7641 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
7642 * can focus on validating the format specifiers.
7644 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, num_args, &data);
7646 verbose(env, "Invalid format string\n");
7651 static int check_get_func_ip(struct bpf_verifier_env *env)
7653 enum bpf_prog_type type = resolve_prog_type(env->prog);
7654 int func_id = BPF_FUNC_get_func_ip;
7656 if (type == BPF_PROG_TYPE_TRACING) {
7657 if (!bpf_prog_has_trampoline(env->prog)) {
7658 verbose(env, "func %s#%d supported only for fentry/fexit/fmod_ret programs\n",
7659 func_id_name(func_id), func_id);
7663 } else if (type == BPF_PROG_TYPE_KPROBE) {
7667 verbose(env, "func %s#%d not supported for program type %d\n",
7668 func_id_name(func_id), func_id, type);
7672 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
7674 return &env->insn_aux_data[env->insn_idx];
7677 static bool loop_flag_is_zero(struct bpf_verifier_env *env)
7679 struct bpf_reg_state *regs = cur_regs(env);
7680 struct bpf_reg_state *reg = ®s[BPF_REG_4];
7681 bool reg_is_null = register_is_null(reg);
7684 mark_chain_precision(env, BPF_REG_4);
7689 static void update_loop_inline_state(struct bpf_verifier_env *env, u32 subprogno)
7691 struct bpf_loop_inline_state *state = &cur_aux(env)->loop_inline_state;
7693 if (!state->initialized) {
7694 state->initialized = 1;
7695 state->fit_for_inline = loop_flag_is_zero(env);
7696 state->callback_subprogno = subprogno;
7700 if (!state->fit_for_inline)
7703 state->fit_for_inline = (loop_flag_is_zero(env) &&
7704 state->callback_subprogno == subprogno);
7707 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
7710 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
7711 const struct bpf_func_proto *fn = NULL;
7712 enum bpf_return_type ret_type;
7713 enum bpf_type_flag ret_flag;
7714 struct bpf_reg_state *regs;
7715 struct bpf_call_arg_meta meta;
7716 int insn_idx = *insn_idx_p;
7718 int i, err, func_id;
7720 /* find function prototype */
7721 func_id = insn->imm;
7722 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
7723 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
7728 if (env->ops->get_func_proto)
7729 fn = env->ops->get_func_proto(func_id, env->prog);
7731 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
7736 /* eBPF programs must be GPL compatible to use GPL-ed functions */
7737 if (!env->prog->gpl_compatible && fn->gpl_only) {
7738 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
7742 if (fn->allowed && !fn->allowed(env->prog)) {
7743 verbose(env, "helper call is not allowed in probe\n");
7747 if (!env->prog->aux->sleepable && fn->might_sleep) {
7748 verbose(env, "helper call might sleep in a non-sleepable prog\n");
7752 /* With LD_ABS/IND some JITs save/restore skb from r1. */
7753 changes_data = bpf_helper_changes_pkt_data(fn->func);
7754 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
7755 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
7756 func_id_name(func_id), func_id);
7760 memset(&meta, 0, sizeof(meta));
7761 meta.pkt_access = fn->pkt_access;
7763 err = check_func_proto(fn, func_id);
7765 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
7766 func_id_name(func_id), func_id);
7770 if (env->cur_state->active_rcu_lock) {
7771 if (fn->might_sleep) {
7772 verbose(env, "sleepable helper %s#%d in rcu_read_lock region\n",
7773 func_id_name(func_id), func_id);
7777 if (env->prog->aux->sleepable && is_storage_get_function(func_id))
7778 env->insn_aux_data[insn_idx].storage_get_func_atomic = true;
7781 meta.func_id = func_id;
7783 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
7784 err = check_func_arg(env, i, &meta, fn);
7789 err = record_func_map(env, &meta, func_id, insn_idx);
7793 err = record_func_key(env, &meta, func_id, insn_idx);
7797 /* Mark slots with STACK_MISC in case of raw mode, stack offset
7798 * is inferred from register state.
7800 for (i = 0; i < meta.access_size; i++) {
7801 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
7802 BPF_WRITE, -1, false);
7807 regs = cur_regs(env);
7809 /* This can only be set for PTR_TO_STACK, as CONST_PTR_TO_DYNPTR cannot
7810 * be reinitialized by any dynptr helper. Hence, mark_stack_slots_dynptr
7811 * is safe to do directly.
7813 if (meta.uninit_dynptr_regno) {
7814 if (regs[meta.uninit_dynptr_regno].type == CONST_PTR_TO_DYNPTR) {
7815 verbose(env, "verifier internal error: CONST_PTR_TO_DYNPTR cannot be initialized\n");
7818 /* we write BPF_DW bits (8 bytes) at a time */
7819 for (i = 0; i < BPF_DYNPTR_SIZE; i += 8) {
7820 err = check_mem_access(env, insn_idx, meta.uninit_dynptr_regno,
7821 i, BPF_DW, BPF_WRITE, -1, false);
7826 err = mark_stack_slots_dynptr(env, ®s[meta.uninit_dynptr_regno],
7827 fn->arg_type[meta.uninit_dynptr_regno - BPF_REG_1],
7833 if (meta.release_regno) {
7835 /* This can only be set for PTR_TO_STACK, as CONST_PTR_TO_DYNPTR cannot
7836 * be released by any dynptr helper. Hence, unmark_stack_slots_dynptr
7837 * is safe to do directly.
7839 if (arg_type_is_dynptr(fn->arg_type[meta.release_regno - BPF_REG_1])) {
7840 if (regs[meta.release_regno].type == CONST_PTR_TO_DYNPTR) {
7841 verbose(env, "verifier internal error: CONST_PTR_TO_DYNPTR cannot be released\n");
7844 err = unmark_stack_slots_dynptr(env, ®s[meta.release_regno]);
7845 } else if (meta.ref_obj_id) {
7846 err = release_reference(env, meta.ref_obj_id);
7847 } else if (register_is_null(®s[meta.release_regno])) {
7848 /* meta.ref_obj_id can only be 0 if register that is meant to be
7849 * released is NULL, which must be > R0.
7854 verbose(env, "func %s#%d reference has not been acquired before\n",
7855 func_id_name(func_id), func_id);
7861 case BPF_FUNC_tail_call:
7862 err = check_reference_leak(env);
7864 verbose(env, "tail_call would lead to reference leak\n");
7868 case BPF_FUNC_get_local_storage:
7869 /* check that flags argument in get_local_storage(map, flags) is 0,
7870 * this is required because get_local_storage() can't return an error.
7872 if (!register_is_null(®s[BPF_REG_2])) {
7873 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
7877 case BPF_FUNC_for_each_map_elem:
7878 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
7879 set_map_elem_callback_state);
7881 case BPF_FUNC_timer_set_callback:
7882 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
7883 set_timer_callback_state);
7885 case BPF_FUNC_find_vma:
7886 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
7887 set_find_vma_callback_state);
7889 case BPF_FUNC_snprintf:
7890 err = check_bpf_snprintf_call(env, regs);
7893 update_loop_inline_state(env, meta.subprogno);
7894 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
7895 set_loop_callback_state);
7897 case BPF_FUNC_dynptr_from_mem:
7898 if (regs[BPF_REG_1].type != PTR_TO_MAP_VALUE) {
7899 verbose(env, "Unsupported reg type %s for bpf_dynptr_from_mem data\n",
7900 reg_type_str(env, regs[BPF_REG_1].type));
7904 case BPF_FUNC_set_retval:
7905 if (prog_type == BPF_PROG_TYPE_LSM &&
7906 env->prog->expected_attach_type == BPF_LSM_CGROUP) {
7907 if (!env->prog->aux->attach_func_proto->type) {
7908 /* Make sure programs that attach to void
7909 * hooks don't try to modify return value.
7911 verbose(env, "BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
7916 case BPF_FUNC_dynptr_data:
7917 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
7918 if (arg_type_is_dynptr(fn->arg_type[i])) {
7919 struct bpf_reg_state *reg = ®s[BPF_REG_1 + i];
7921 if (meta.ref_obj_id) {
7922 verbose(env, "verifier internal error: meta.ref_obj_id already set\n");
7926 meta.ref_obj_id = dynptr_ref_obj_id(env, reg);
7930 if (i == MAX_BPF_FUNC_REG_ARGS) {
7931 verbose(env, "verifier internal error: no dynptr in bpf_dynptr_data()\n");
7935 case BPF_FUNC_user_ringbuf_drain:
7936 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
7937 set_user_ringbuf_callback_state);
7944 /* reset caller saved regs */
7945 for (i = 0; i < CALLER_SAVED_REGS; i++) {
7946 mark_reg_not_init(env, regs, caller_saved[i]);
7947 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
7950 /* helper call returns 64-bit value. */
7951 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
7953 /* update return register (already marked as written above) */
7954 ret_type = fn->ret_type;
7955 ret_flag = type_flag(ret_type);
7957 switch (base_type(ret_type)) {
7959 /* sets type to SCALAR_VALUE */
7960 mark_reg_unknown(env, regs, BPF_REG_0);
7963 regs[BPF_REG_0].type = NOT_INIT;
7965 case RET_PTR_TO_MAP_VALUE:
7966 /* There is no offset yet applied, variable or fixed */
7967 mark_reg_known_zero(env, regs, BPF_REG_0);
7968 /* remember map_ptr, so that check_map_access()
7969 * can check 'value_size' boundary of memory access
7970 * to map element returned from bpf_map_lookup_elem()
7972 if (meta.map_ptr == NULL) {
7974 "kernel subsystem misconfigured verifier\n");
7977 regs[BPF_REG_0].map_ptr = meta.map_ptr;
7978 regs[BPF_REG_0].map_uid = meta.map_uid;
7979 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE | ret_flag;
7980 if (!type_may_be_null(ret_type) &&
7981 btf_record_has_field(meta.map_ptr->record, BPF_SPIN_LOCK)) {
7982 regs[BPF_REG_0].id = ++env->id_gen;
7985 case RET_PTR_TO_SOCKET:
7986 mark_reg_known_zero(env, regs, BPF_REG_0);
7987 regs[BPF_REG_0].type = PTR_TO_SOCKET | ret_flag;
7989 case RET_PTR_TO_SOCK_COMMON:
7990 mark_reg_known_zero(env, regs, BPF_REG_0);
7991 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON | ret_flag;
7993 case RET_PTR_TO_TCP_SOCK:
7994 mark_reg_known_zero(env, regs, BPF_REG_0);
7995 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK | ret_flag;
7997 case RET_PTR_TO_MEM:
7998 mark_reg_known_zero(env, regs, BPF_REG_0);
7999 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
8000 regs[BPF_REG_0].mem_size = meta.mem_size;
8002 case RET_PTR_TO_MEM_OR_BTF_ID:
8004 const struct btf_type *t;
8006 mark_reg_known_zero(env, regs, BPF_REG_0);
8007 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
8008 if (!btf_type_is_struct(t)) {
8010 const struct btf_type *ret;
8013 /* resolve the type size of ksym. */
8014 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
8016 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
8017 verbose(env, "unable to resolve the size of type '%s': %ld\n",
8018 tname, PTR_ERR(ret));
8021 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
8022 regs[BPF_REG_0].mem_size = tsize;
8024 /* MEM_RDONLY may be carried from ret_flag, but it
8025 * doesn't apply on PTR_TO_BTF_ID. Fold it, otherwise
8026 * it will confuse the check of PTR_TO_BTF_ID in
8027 * check_mem_access().
8029 ret_flag &= ~MEM_RDONLY;
8031 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
8032 regs[BPF_REG_0].btf = meta.ret_btf;
8033 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
8037 case RET_PTR_TO_BTF_ID:
8039 struct btf *ret_btf;
8042 mark_reg_known_zero(env, regs, BPF_REG_0);
8043 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
8044 if (func_id == BPF_FUNC_kptr_xchg) {
8045 ret_btf = meta.kptr_field->kptr.btf;
8046 ret_btf_id = meta.kptr_field->kptr.btf_id;
8048 if (fn->ret_btf_id == BPF_PTR_POISON) {
8049 verbose(env, "verifier internal error:");
8050 verbose(env, "func %s has non-overwritten BPF_PTR_POISON return type\n",
8051 func_id_name(func_id));
8054 ret_btf = btf_vmlinux;
8055 ret_btf_id = *fn->ret_btf_id;
8057 if (ret_btf_id == 0) {
8058 verbose(env, "invalid return type %u of func %s#%d\n",
8059 base_type(ret_type), func_id_name(func_id),
8063 regs[BPF_REG_0].btf = ret_btf;
8064 regs[BPF_REG_0].btf_id = ret_btf_id;
8068 verbose(env, "unknown return type %u of func %s#%d\n",
8069 base_type(ret_type), func_id_name(func_id), func_id);
8073 if (type_may_be_null(regs[BPF_REG_0].type))
8074 regs[BPF_REG_0].id = ++env->id_gen;
8076 if (helper_multiple_ref_obj_use(func_id, meta.map_ptr)) {
8077 verbose(env, "verifier internal error: func %s#%d sets ref_obj_id more than once\n",
8078 func_id_name(func_id), func_id);
8082 if (is_ptr_cast_function(func_id) || is_dynptr_ref_function(func_id)) {
8083 /* For release_reference() */
8084 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
8085 } else if (is_acquire_function(func_id, meta.map_ptr)) {
8086 int id = acquire_reference_state(env, insn_idx);
8090 /* For mark_ptr_or_null_reg() */
8091 regs[BPF_REG_0].id = id;
8092 /* For release_reference() */
8093 regs[BPF_REG_0].ref_obj_id = id;
8096 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
8098 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
8102 if ((func_id == BPF_FUNC_get_stack ||
8103 func_id == BPF_FUNC_get_task_stack) &&
8104 !env->prog->has_callchain_buf) {
8105 const char *err_str;
8107 #ifdef CONFIG_PERF_EVENTS
8108 err = get_callchain_buffers(sysctl_perf_event_max_stack);
8109 err_str = "cannot get callchain buffer for func %s#%d\n";
8112 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
8115 verbose(env, err_str, func_id_name(func_id), func_id);
8119 env->prog->has_callchain_buf = true;
8122 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
8123 env->prog->call_get_stack = true;
8125 if (func_id == BPF_FUNC_get_func_ip) {
8126 if (check_get_func_ip(env))
8128 env->prog->call_get_func_ip = true;
8132 clear_all_pkt_pointers(env);
8136 /* mark_btf_func_reg_size() is used when the reg size is determined by
8137 * the BTF func_proto's return value size and argument.
8139 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
8142 struct bpf_reg_state *reg = &cur_regs(env)[regno];
8144 if (regno == BPF_REG_0) {
8145 /* Function return value */
8146 reg->live |= REG_LIVE_WRITTEN;
8147 reg->subreg_def = reg_size == sizeof(u64) ?
8148 DEF_NOT_SUBREG : env->insn_idx + 1;
8150 /* Function argument */
8151 if (reg_size == sizeof(u64)) {
8152 mark_insn_zext(env, reg);
8153 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
8155 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
8160 struct bpf_kfunc_call_arg_meta {
8165 const struct btf_type *func_proto;
8166 const char *func_name;
8167 /* Out parameters */
8182 struct btf_field *field;
8186 static bool is_kfunc_acquire(struct bpf_kfunc_call_arg_meta *meta)
8188 return meta->kfunc_flags & KF_ACQUIRE;
8191 static bool is_kfunc_ret_null(struct bpf_kfunc_call_arg_meta *meta)
8193 return meta->kfunc_flags & KF_RET_NULL;
8196 static bool is_kfunc_release(struct bpf_kfunc_call_arg_meta *meta)
8198 return meta->kfunc_flags & KF_RELEASE;
8201 static bool is_kfunc_trusted_args(struct bpf_kfunc_call_arg_meta *meta)
8203 return meta->kfunc_flags & KF_TRUSTED_ARGS;
8206 static bool is_kfunc_sleepable(struct bpf_kfunc_call_arg_meta *meta)
8208 return meta->kfunc_flags & KF_SLEEPABLE;
8211 static bool is_kfunc_destructive(struct bpf_kfunc_call_arg_meta *meta)
8213 return meta->kfunc_flags & KF_DESTRUCTIVE;
8216 static bool is_kfunc_rcu(struct bpf_kfunc_call_arg_meta *meta)
8218 return meta->kfunc_flags & KF_RCU;
8221 static bool is_kfunc_arg_kptr_get(struct bpf_kfunc_call_arg_meta *meta, int arg)
8223 return arg == 0 && (meta->kfunc_flags & KF_KPTR_GET);
8226 static bool __kfunc_param_match_suffix(const struct btf *btf,
8227 const struct btf_param *arg,
8230 int suffix_len = strlen(suffix), len;
8231 const char *param_name;
8233 /* In the future, this can be ported to use BTF tagging */
8234 param_name = btf_name_by_offset(btf, arg->name_off);
8235 if (str_is_empty(param_name))
8237 len = strlen(param_name);
8238 if (len < suffix_len)
8240 param_name += len - suffix_len;
8241 return !strncmp(param_name, suffix, suffix_len);
8244 static bool is_kfunc_arg_mem_size(const struct btf *btf,
8245 const struct btf_param *arg,
8246 const struct bpf_reg_state *reg)
8248 const struct btf_type *t;
8250 t = btf_type_skip_modifiers(btf, arg->type, NULL);
8251 if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
8254 return __kfunc_param_match_suffix(btf, arg, "__sz");
8257 static bool is_kfunc_arg_constant(const struct btf *btf, const struct btf_param *arg)
8259 return __kfunc_param_match_suffix(btf, arg, "__k");
8262 static bool is_kfunc_arg_ignore(const struct btf *btf, const struct btf_param *arg)
8264 return __kfunc_param_match_suffix(btf, arg, "__ign");
8267 static bool is_kfunc_arg_alloc_obj(const struct btf *btf, const struct btf_param *arg)
8269 return __kfunc_param_match_suffix(btf, arg, "__alloc");
8272 static bool is_kfunc_arg_scalar_with_name(const struct btf *btf,
8273 const struct btf_param *arg,
8276 int len, target_len = strlen(name);
8277 const char *param_name;
8279 param_name = btf_name_by_offset(btf, arg->name_off);
8280 if (str_is_empty(param_name))
8282 len = strlen(param_name);
8283 if (len != target_len)
8285 if (strcmp(param_name, name))
8293 KF_ARG_LIST_HEAD_ID,
8294 KF_ARG_LIST_NODE_ID,
8297 BTF_ID_LIST(kf_arg_btf_ids)
8298 BTF_ID(struct, bpf_dynptr_kern)
8299 BTF_ID(struct, bpf_list_head)
8300 BTF_ID(struct, bpf_list_node)
8302 static bool __is_kfunc_ptr_arg_type(const struct btf *btf,
8303 const struct btf_param *arg, int type)
8305 const struct btf_type *t;
8308 t = btf_type_skip_modifiers(btf, arg->type, NULL);
8311 if (!btf_type_is_ptr(t))
8313 t = btf_type_skip_modifiers(btf, t->type, &res_id);
8316 return btf_types_are_same(btf, res_id, btf_vmlinux, kf_arg_btf_ids[type]);
8319 static bool is_kfunc_arg_dynptr(const struct btf *btf, const struct btf_param *arg)
8321 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_DYNPTR_ID);
8324 static bool is_kfunc_arg_list_head(const struct btf *btf, const struct btf_param *arg)
8326 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_HEAD_ID);
8329 static bool is_kfunc_arg_list_node(const struct btf *btf, const struct btf_param *arg)
8331 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_NODE_ID);
8334 /* Returns true if struct is composed of scalars, 4 levels of nesting allowed */
8335 static bool __btf_type_is_scalar_struct(struct bpf_verifier_env *env,
8336 const struct btf *btf,
8337 const struct btf_type *t, int rec)
8339 const struct btf_type *member_type;
8340 const struct btf_member *member;
8343 if (!btf_type_is_struct(t))
8346 for_each_member(i, t, member) {
8347 const struct btf_array *array;
8349 member_type = btf_type_skip_modifiers(btf, member->type, NULL);
8350 if (btf_type_is_struct(member_type)) {
8352 verbose(env, "max struct nesting depth exceeded\n");
8355 if (!__btf_type_is_scalar_struct(env, btf, member_type, rec + 1))
8359 if (btf_type_is_array(member_type)) {
8360 array = btf_array(member_type);
8363 member_type = btf_type_skip_modifiers(btf, array->type, NULL);
8364 if (!btf_type_is_scalar(member_type))
8368 if (!btf_type_is_scalar(member_type))
8375 static u32 *reg2btf_ids[__BPF_REG_TYPE_MAX] = {
8377 [PTR_TO_SOCKET] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK],
8378 [PTR_TO_SOCK_COMMON] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
8379 [PTR_TO_TCP_SOCK] = &btf_sock_ids[BTF_SOCK_TYPE_TCP],
8383 enum kfunc_ptr_arg_type {
8385 KF_ARG_PTR_TO_ALLOC_BTF_ID, /* Allocated object */
8386 KF_ARG_PTR_TO_KPTR, /* PTR_TO_KPTR but type specific */
8387 KF_ARG_PTR_TO_DYNPTR,
8388 KF_ARG_PTR_TO_LIST_HEAD,
8389 KF_ARG_PTR_TO_LIST_NODE,
8390 KF_ARG_PTR_TO_BTF_ID, /* Also covers reg2btf_ids conversions */
8392 KF_ARG_PTR_TO_MEM_SIZE, /* Size derived from next argument, skip it */
8395 enum special_kfunc_type {
8396 KF_bpf_obj_new_impl,
8397 KF_bpf_obj_drop_impl,
8398 KF_bpf_list_push_front,
8399 KF_bpf_list_push_back,
8400 KF_bpf_list_pop_front,
8401 KF_bpf_list_pop_back,
8402 KF_bpf_cast_to_kern_ctx,
8404 KF_bpf_rcu_read_lock,
8405 KF_bpf_rcu_read_unlock,
8408 BTF_SET_START(special_kfunc_set)
8409 BTF_ID(func, bpf_obj_new_impl)
8410 BTF_ID(func, bpf_obj_drop_impl)
8411 BTF_ID(func, bpf_list_push_front)
8412 BTF_ID(func, bpf_list_push_back)
8413 BTF_ID(func, bpf_list_pop_front)
8414 BTF_ID(func, bpf_list_pop_back)
8415 BTF_ID(func, bpf_cast_to_kern_ctx)
8416 BTF_ID(func, bpf_rdonly_cast)
8417 BTF_SET_END(special_kfunc_set)
8419 BTF_ID_LIST(special_kfunc_list)
8420 BTF_ID(func, bpf_obj_new_impl)
8421 BTF_ID(func, bpf_obj_drop_impl)
8422 BTF_ID(func, bpf_list_push_front)
8423 BTF_ID(func, bpf_list_push_back)
8424 BTF_ID(func, bpf_list_pop_front)
8425 BTF_ID(func, bpf_list_pop_back)
8426 BTF_ID(func, bpf_cast_to_kern_ctx)
8427 BTF_ID(func, bpf_rdonly_cast)
8428 BTF_ID(func, bpf_rcu_read_lock)
8429 BTF_ID(func, bpf_rcu_read_unlock)
8431 static bool is_kfunc_bpf_rcu_read_lock(struct bpf_kfunc_call_arg_meta *meta)
8433 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_lock];
8436 static bool is_kfunc_bpf_rcu_read_unlock(struct bpf_kfunc_call_arg_meta *meta)
8438 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_unlock];
8441 static enum kfunc_ptr_arg_type
8442 get_kfunc_ptr_arg_type(struct bpf_verifier_env *env,
8443 struct bpf_kfunc_call_arg_meta *meta,
8444 const struct btf_type *t, const struct btf_type *ref_t,
8445 const char *ref_tname, const struct btf_param *args,
8446 int argno, int nargs)
8448 u32 regno = argno + 1;
8449 struct bpf_reg_state *regs = cur_regs(env);
8450 struct bpf_reg_state *reg = ®s[regno];
8451 bool arg_mem_size = false;
8453 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx])
8454 return KF_ARG_PTR_TO_CTX;
8456 /* In this function, we verify the kfunc's BTF as per the argument type,
8457 * leaving the rest of the verification with respect to the register
8458 * type to our caller. When a set of conditions hold in the BTF type of
8459 * arguments, we resolve it to a known kfunc_ptr_arg_type.
8461 if (btf_get_prog_ctx_type(&env->log, meta->btf, t, resolve_prog_type(env->prog), argno))
8462 return KF_ARG_PTR_TO_CTX;
8464 if (is_kfunc_arg_alloc_obj(meta->btf, &args[argno]))
8465 return KF_ARG_PTR_TO_ALLOC_BTF_ID;
8467 if (is_kfunc_arg_kptr_get(meta, argno)) {
8468 if (!btf_type_is_ptr(ref_t)) {
8469 verbose(env, "arg#0 BTF type must be a double pointer for kptr_get kfunc\n");
8472 ref_t = btf_type_by_id(meta->btf, ref_t->type);
8473 ref_tname = btf_name_by_offset(meta->btf, ref_t->name_off);
8474 if (!btf_type_is_struct(ref_t)) {
8475 verbose(env, "kernel function %s args#0 pointer type %s %s is not supported\n",
8476 meta->func_name, btf_type_str(ref_t), ref_tname);
8479 return KF_ARG_PTR_TO_KPTR;
8482 if (is_kfunc_arg_dynptr(meta->btf, &args[argno]))
8483 return KF_ARG_PTR_TO_DYNPTR;
8485 if (is_kfunc_arg_list_head(meta->btf, &args[argno]))
8486 return KF_ARG_PTR_TO_LIST_HEAD;
8488 if (is_kfunc_arg_list_node(meta->btf, &args[argno]))
8489 return KF_ARG_PTR_TO_LIST_NODE;
8491 if ((base_type(reg->type) == PTR_TO_BTF_ID || reg2btf_ids[base_type(reg->type)])) {
8492 if (!btf_type_is_struct(ref_t)) {
8493 verbose(env, "kernel function %s args#%d pointer type %s %s is not supported\n",
8494 meta->func_name, argno, btf_type_str(ref_t), ref_tname);
8497 return KF_ARG_PTR_TO_BTF_ID;
8500 if (argno + 1 < nargs && is_kfunc_arg_mem_size(meta->btf, &args[argno + 1], ®s[regno + 1]))
8501 arg_mem_size = true;
8503 /* This is the catch all argument type of register types supported by
8504 * check_helper_mem_access. However, we only allow when argument type is
8505 * pointer to scalar, or struct composed (recursively) of scalars. When
8506 * arg_mem_size is true, the pointer can be void *.
8508 if (!btf_type_is_scalar(ref_t) && !__btf_type_is_scalar_struct(env, meta->btf, ref_t, 0) &&
8509 (arg_mem_size ? !btf_type_is_void(ref_t) : 1)) {
8510 verbose(env, "arg#%d pointer type %s %s must point to %sscalar, or struct with scalar\n",
8511 argno, btf_type_str(ref_t), ref_tname, arg_mem_size ? "void, " : "");
8514 return arg_mem_size ? KF_ARG_PTR_TO_MEM_SIZE : KF_ARG_PTR_TO_MEM;
8517 static int process_kf_arg_ptr_to_btf_id(struct bpf_verifier_env *env,
8518 struct bpf_reg_state *reg,
8519 const struct btf_type *ref_t,
8520 const char *ref_tname, u32 ref_id,
8521 struct bpf_kfunc_call_arg_meta *meta,
8524 const struct btf_type *reg_ref_t;
8525 bool strict_type_match = false;
8526 const struct btf *reg_btf;
8527 const char *reg_ref_tname;
8530 if (base_type(reg->type) == PTR_TO_BTF_ID) {
8532 reg_ref_id = reg->btf_id;
8534 reg_btf = btf_vmlinux;
8535 reg_ref_id = *reg2btf_ids[base_type(reg->type)];
8538 if (is_kfunc_trusted_args(meta) || (is_kfunc_release(meta) && reg->ref_obj_id))
8539 strict_type_match = true;
8541 reg_ref_t = btf_type_skip_modifiers(reg_btf, reg_ref_id, ®_ref_id);
8542 reg_ref_tname = btf_name_by_offset(reg_btf, reg_ref_t->name_off);
8543 if (!btf_struct_ids_match(&env->log, reg_btf, reg_ref_id, reg->off, meta->btf, ref_id, strict_type_match)) {
8544 verbose(env, "kernel function %s args#%d expected pointer to %s %s but R%d has a pointer to %s %s\n",
8545 meta->func_name, argno, btf_type_str(ref_t), ref_tname, argno + 1,
8546 btf_type_str(reg_ref_t), reg_ref_tname);
8552 static int process_kf_arg_ptr_to_kptr(struct bpf_verifier_env *env,
8553 struct bpf_reg_state *reg,
8554 const struct btf_type *ref_t,
8555 const char *ref_tname,
8556 struct bpf_kfunc_call_arg_meta *meta,
8559 struct btf_field *kptr_field;
8561 /* check_func_arg_reg_off allows var_off for
8562 * PTR_TO_MAP_VALUE, but we need fixed offset to find
8565 if (!tnum_is_const(reg->var_off)) {
8566 verbose(env, "arg#0 must have constant offset\n");
8570 kptr_field = btf_record_find(reg->map_ptr->record, reg->off + reg->var_off.value, BPF_KPTR);
8571 if (!kptr_field || kptr_field->type != BPF_KPTR_REF) {
8572 verbose(env, "arg#0 no referenced kptr at map value offset=%llu\n",
8573 reg->off + reg->var_off.value);
8577 if (!btf_struct_ids_match(&env->log, meta->btf, ref_t->type, 0, kptr_field->kptr.btf,
8578 kptr_field->kptr.btf_id, true)) {
8579 verbose(env, "kernel function %s args#%d expected pointer to %s %s\n",
8580 meta->func_name, argno, btf_type_str(ref_t), ref_tname);
8586 static int ref_set_release_on_unlock(struct bpf_verifier_env *env, u32 ref_obj_id)
8588 struct bpf_func_state *state = cur_func(env);
8589 struct bpf_reg_state *reg;
8592 /* bpf_spin_lock only allows calling list_push and list_pop, no BPF
8593 * subprogs, no global functions. This means that the references would
8594 * not be released inside the critical section but they may be added to
8595 * the reference state, and the acquired_refs are never copied out for a
8596 * different frame as BPF to BPF calls don't work in bpf_spin_lock
8597 * critical sections.
8600 verbose(env, "verifier internal error: ref_obj_id is zero for release_on_unlock\n");
8603 for (i = 0; i < state->acquired_refs; i++) {
8604 if (state->refs[i].id == ref_obj_id) {
8605 if (state->refs[i].release_on_unlock) {
8606 verbose(env, "verifier internal error: expected false release_on_unlock");
8609 state->refs[i].release_on_unlock = true;
8610 /* Now mark everyone sharing same ref_obj_id as untrusted */
8611 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
8612 if (reg->ref_obj_id == ref_obj_id)
8613 reg->type |= PTR_UNTRUSTED;
8618 verbose(env, "verifier internal error: ref state missing for ref_obj_id\n");
8622 /* Implementation details:
8624 * Each register points to some region of memory, which we define as an
8625 * allocation. Each allocation may embed a bpf_spin_lock which protects any
8626 * special BPF objects (bpf_list_head, bpf_rb_root, etc.) part of the same
8627 * allocation. The lock and the data it protects are colocated in the same
8630 * Hence, everytime a register holds a pointer value pointing to such
8631 * allocation, the verifier preserves a unique reg->id for it.
8633 * The verifier remembers the lock 'ptr' and the lock 'id' whenever
8634 * bpf_spin_lock is called.
8636 * To enable this, lock state in the verifier captures two values:
8637 * active_lock.ptr = Register's type specific pointer
8638 * active_lock.id = A unique ID for each register pointer value
8640 * Currently, PTR_TO_MAP_VALUE and PTR_TO_BTF_ID | MEM_ALLOC are the two
8641 * supported register types.
8643 * The active_lock.ptr in case of map values is the reg->map_ptr, and in case of
8644 * allocated objects is the reg->btf pointer.
8646 * The active_lock.id is non-unique for maps supporting direct_value_addr, as we
8647 * can establish the provenance of the map value statically for each distinct
8648 * lookup into such maps. They always contain a single map value hence unique
8649 * IDs for each pseudo load pessimizes the algorithm and rejects valid programs.
8651 * So, in case of global variables, they use array maps with max_entries = 1,
8652 * hence their active_lock.ptr becomes map_ptr and id = 0 (since they all point
8653 * into the same map value as max_entries is 1, as described above).
8655 * In case of inner map lookups, the inner map pointer has same map_ptr as the
8656 * outer map pointer (in verifier context), but each lookup into an inner map
8657 * assigns a fresh reg->id to the lookup, so while lookups into distinct inner
8658 * maps from the same outer map share the same map_ptr as active_lock.ptr, they
8659 * will get different reg->id assigned to each lookup, hence different
8662 * In case of allocated objects, active_lock.ptr is the reg->btf, and the
8663 * reg->id is a unique ID preserved after the NULL pointer check on the pointer
8664 * returned from bpf_obj_new. Each allocation receives a new reg->id.
8666 static int check_reg_allocation_locked(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
8671 switch ((int)reg->type) {
8672 case PTR_TO_MAP_VALUE:
8675 case PTR_TO_BTF_ID | MEM_ALLOC:
8676 case PTR_TO_BTF_ID | MEM_ALLOC | PTR_TRUSTED:
8680 verbose(env, "verifier internal error: unknown reg type for lock check\n");
8685 if (!env->cur_state->active_lock.ptr)
8687 if (env->cur_state->active_lock.ptr != ptr ||
8688 env->cur_state->active_lock.id != id) {
8689 verbose(env, "held lock and object are not in the same allocation\n");
8695 static bool is_bpf_list_api_kfunc(u32 btf_id)
8697 return btf_id == special_kfunc_list[KF_bpf_list_push_front] ||
8698 btf_id == special_kfunc_list[KF_bpf_list_push_back] ||
8699 btf_id == special_kfunc_list[KF_bpf_list_pop_front] ||
8700 btf_id == special_kfunc_list[KF_bpf_list_pop_back];
8703 static int process_kf_arg_ptr_to_list_head(struct bpf_verifier_env *env,
8704 struct bpf_reg_state *reg, u32 regno,
8705 struct bpf_kfunc_call_arg_meta *meta)
8707 struct btf_field *field;
8708 struct btf_record *rec;
8711 if (meta->btf != btf_vmlinux || !is_bpf_list_api_kfunc(meta->func_id)) {
8712 verbose(env, "verifier internal error: bpf_list_head argument for unknown kfunc\n");
8716 if (!tnum_is_const(reg->var_off)) {
8718 "R%d doesn't have constant offset. bpf_list_head has to be at the constant offset\n",
8723 rec = reg_btf_record(reg);
8724 list_head_off = reg->off + reg->var_off.value;
8725 field = btf_record_find(rec, list_head_off, BPF_LIST_HEAD);
8727 verbose(env, "bpf_list_head not found at offset=%u\n", list_head_off);
8731 /* All functions require bpf_list_head to be protected using a bpf_spin_lock */
8732 if (check_reg_allocation_locked(env, reg)) {
8733 verbose(env, "bpf_spin_lock at off=%d must be held for bpf_list_head\n",
8734 rec->spin_lock_off);
8738 if (meta->arg_list_head.field) {
8739 verbose(env, "verifier internal error: repeating bpf_list_head arg\n");
8742 meta->arg_list_head.field = field;
8746 static int process_kf_arg_ptr_to_list_node(struct bpf_verifier_env *env,
8747 struct bpf_reg_state *reg, u32 regno,
8748 struct bpf_kfunc_call_arg_meta *meta)
8750 const struct btf_type *et, *t;
8751 struct btf_field *field;
8752 struct btf_record *rec;
8755 if (meta->btf != btf_vmlinux ||
8756 (meta->func_id != special_kfunc_list[KF_bpf_list_push_front] &&
8757 meta->func_id != special_kfunc_list[KF_bpf_list_push_back])) {
8758 verbose(env, "verifier internal error: bpf_list_node argument for unknown kfunc\n");
8762 if (!tnum_is_const(reg->var_off)) {
8764 "R%d doesn't have constant offset. bpf_list_node has to be at the constant offset\n",
8769 rec = reg_btf_record(reg);
8770 list_node_off = reg->off + reg->var_off.value;
8771 field = btf_record_find(rec, list_node_off, BPF_LIST_NODE);
8772 if (!field || field->offset != list_node_off) {
8773 verbose(env, "bpf_list_node not found at offset=%u\n", list_node_off);
8777 field = meta->arg_list_head.field;
8779 et = btf_type_by_id(field->list_head.btf, field->list_head.value_btf_id);
8780 t = btf_type_by_id(reg->btf, reg->btf_id);
8781 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, 0, field->list_head.btf,
8782 field->list_head.value_btf_id, true)) {
8783 verbose(env, "operation on bpf_list_head expects arg#1 bpf_list_node at offset=%d "
8784 "in struct %s, but arg is at offset=%d in struct %s\n",
8785 field->list_head.node_offset, btf_name_by_offset(field->list_head.btf, et->name_off),
8786 list_node_off, btf_name_by_offset(reg->btf, t->name_off));
8790 if (list_node_off != field->list_head.node_offset) {
8791 verbose(env, "arg#1 offset=%d, but expected bpf_list_node at offset=%d in struct %s\n",
8792 list_node_off, field->list_head.node_offset,
8793 btf_name_by_offset(field->list_head.btf, et->name_off));
8796 /* Set arg#1 for expiration after unlock */
8797 return ref_set_release_on_unlock(env, reg->ref_obj_id);
8800 static int check_kfunc_args(struct bpf_verifier_env *env, struct bpf_kfunc_call_arg_meta *meta)
8802 const char *func_name = meta->func_name, *ref_tname;
8803 const struct btf *btf = meta->btf;
8804 const struct btf_param *args;
8808 args = (const struct btf_param *)(meta->func_proto + 1);
8809 nargs = btf_type_vlen(meta->func_proto);
8810 if (nargs > MAX_BPF_FUNC_REG_ARGS) {
8811 verbose(env, "Function %s has %d > %d args\n", func_name, nargs,
8812 MAX_BPF_FUNC_REG_ARGS);
8816 /* Check that BTF function arguments match actual types that the
8819 for (i = 0; i < nargs; i++) {
8820 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[i + 1];
8821 const struct btf_type *t, *ref_t, *resolve_ret;
8822 enum bpf_arg_type arg_type = ARG_DONTCARE;
8823 u32 regno = i + 1, ref_id, type_size;
8824 bool is_ret_buf_sz = false;
8827 t = btf_type_skip_modifiers(btf, args[i].type, NULL);
8829 if (is_kfunc_arg_ignore(btf, &args[i]))
8832 if (btf_type_is_scalar(t)) {
8833 if (reg->type != SCALAR_VALUE) {
8834 verbose(env, "R%d is not a scalar\n", regno);
8838 if (is_kfunc_arg_constant(meta->btf, &args[i])) {
8839 if (meta->arg_constant.found) {
8840 verbose(env, "verifier internal error: only one constant argument permitted\n");
8843 if (!tnum_is_const(reg->var_off)) {
8844 verbose(env, "R%d must be a known constant\n", regno);
8847 ret = mark_chain_precision(env, regno);
8850 meta->arg_constant.found = true;
8851 meta->arg_constant.value = reg->var_off.value;
8852 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdonly_buf_size")) {
8853 meta->r0_rdonly = true;
8854 is_ret_buf_sz = true;
8855 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdwr_buf_size")) {
8856 is_ret_buf_sz = true;
8859 if (is_ret_buf_sz) {
8860 if (meta->r0_size) {
8861 verbose(env, "2 or more rdonly/rdwr_buf_size parameters for kfunc");
8865 if (!tnum_is_const(reg->var_off)) {
8866 verbose(env, "R%d is not a const\n", regno);
8870 meta->r0_size = reg->var_off.value;
8871 ret = mark_chain_precision(env, regno);
8878 if (!btf_type_is_ptr(t)) {
8879 verbose(env, "Unrecognized arg#%d type %s\n", i, btf_type_str(t));
8883 if (reg->ref_obj_id) {
8884 if (is_kfunc_release(meta) && meta->ref_obj_id) {
8885 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
8886 regno, reg->ref_obj_id,
8890 meta->ref_obj_id = reg->ref_obj_id;
8891 if (is_kfunc_release(meta))
8892 meta->release_regno = regno;
8895 ref_t = btf_type_skip_modifiers(btf, t->type, &ref_id);
8896 ref_tname = btf_name_by_offset(btf, ref_t->name_off);
8898 kf_arg_type = get_kfunc_ptr_arg_type(env, meta, t, ref_t, ref_tname, args, i, nargs);
8899 if (kf_arg_type < 0)
8902 switch (kf_arg_type) {
8903 case KF_ARG_PTR_TO_ALLOC_BTF_ID:
8904 case KF_ARG_PTR_TO_BTF_ID:
8905 if (!is_kfunc_trusted_args(meta) && !is_kfunc_rcu(meta))
8908 if (!is_trusted_reg(reg)) {
8909 if (!is_kfunc_rcu(meta)) {
8910 verbose(env, "R%d must be referenced or trusted\n", regno);
8913 if (!is_rcu_reg(reg)) {
8914 verbose(env, "R%d must be a rcu pointer\n", regno);
8920 case KF_ARG_PTR_TO_CTX:
8921 /* Trusted arguments have the same offset checks as release arguments */
8922 arg_type |= OBJ_RELEASE;
8924 case KF_ARG_PTR_TO_KPTR:
8925 case KF_ARG_PTR_TO_DYNPTR:
8926 case KF_ARG_PTR_TO_LIST_HEAD:
8927 case KF_ARG_PTR_TO_LIST_NODE:
8928 case KF_ARG_PTR_TO_MEM:
8929 case KF_ARG_PTR_TO_MEM_SIZE:
8930 /* Trusted by default */
8937 if (is_kfunc_release(meta) && reg->ref_obj_id)
8938 arg_type |= OBJ_RELEASE;
8939 ret = check_func_arg_reg_off(env, reg, regno, arg_type);
8943 switch (kf_arg_type) {
8944 case KF_ARG_PTR_TO_CTX:
8945 if (reg->type != PTR_TO_CTX) {
8946 verbose(env, "arg#%d expected pointer to ctx, but got %s\n", i, btf_type_str(t));
8950 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
8951 ret = get_kern_ctx_btf_id(&env->log, resolve_prog_type(env->prog));
8954 meta->ret_btf_id = ret;
8957 case KF_ARG_PTR_TO_ALLOC_BTF_ID:
8958 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
8959 verbose(env, "arg#%d expected pointer to allocated object\n", i);
8962 if (!reg->ref_obj_id) {
8963 verbose(env, "allocated object must be referenced\n");
8966 if (meta->btf == btf_vmlinux &&
8967 meta->func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
8968 meta->arg_obj_drop.btf = reg->btf;
8969 meta->arg_obj_drop.btf_id = reg->btf_id;
8972 case KF_ARG_PTR_TO_KPTR:
8973 if (reg->type != PTR_TO_MAP_VALUE) {
8974 verbose(env, "arg#0 expected pointer to map value\n");
8977 ret = process_kf_arg_ptr_to_kptr(env, reg, ref_t, ref_tname, meta, i);
8981 case KF_ARG_PTR_TO_DYNPTR:
8982 if (reg->type != PTR_TO_STACK &&
8983 reg->type != CONST_PTR_TO_DYNPTR) {
8984 verbose(env, "arg#%d expected pointer to stack or dynptr_ptr\n", i);
8988 ret = process_dynptr_func(env, regno, ARG_PTR_TO_DYNPTR | MEM_RDONLY, NULL);
8992 case KF_ARG_PTR_TO_LIST_HEAD:
8993 if (reg->type != PTR_TO_MAP_VALUE &&
8994 reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
8995 verbose(env, "arg#%d expected pointer to map value or allocated object\n", i);
8998 if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
8999 verbose(env, "allocated object must be referenced\n");
9002 ret = process_kf_arg_ptr_to_list_head(env, reg, regno, meta);
9006 case KF_ARG_PTR_TO_LIST_NODE:
9007 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
9008 verbose(env, "arg#%d expected pointer to allocated object\n", i);
9011 if (!reg->ref_obj_id) {
9012 verbose(env, "allocated object must be referenced\n");
9015 ret = process_kf_arg_ptr_to_list_node(env, reg, regno, meta);
9019 case KF_ARG_PTR_TO_BTF_ID:
9020 /* Only base_type is checked, further checks are done here */
9021 if ((base_type(reg->type) != PTR_TO_BTF_ID ||
9022 (bpf_type_has_unsafe_modifiers(reg->type) && !is_rcu_reg(reg))) &&
9023 !reg2btf_ids[base_type(reg->type)]) {
9024 verbose(env, "arg#%d is %s ", i, reg_type_str(env, reg->type));
9025 verbose(env, "expected %s or socket\n",
9026 reg_type_str(env, base_type(reg->type) |
9027 (type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS)));
9030 ret = process_kf_arg_ptr_to_btf_id(env, reg, ref_t, ref_tname, ref_id, meta, i);
9034 case KF_ARG_PTR_TO_MEM:
9035 resolve_ret = btf_resolve_size(btf, ref_t, &type_size);
9036 if (IS_ERR(resolve_ret)) {
9037 verbose(env, "arg#%d reference type('%s %s') size cannot be determined: %ld\n",
9038 i, btf_type_str(ref_t), ref_tname, PTR_ERR(resolve_ret));
9041 ret = check_mem_reg(env, reg, regno, type_size);
9045 case KF_ARG_PTR_TO_MEM_SIZE:
9046 ret = check_kfunc_mem_size_reg(env, ®s[regno + 1], regno + 1);
9048 verbose(env, "arg#%d arg#%d memory, len pair leads to invalid memory access\n", i, i + 1);
9051 /* Skip next '__sz' argument */
9057 if (is_kfunc_release(meta) && !meta->release_regno) {
9058 verbose(env, "release kernel function %s expects refcounted PTR_TO_BTF_ID\n",
9066 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
9069 const struct btf_type *t, *func, *func_proto, *ptr_type;
9070 struct bpf_reg_state *regs = cur_regs(env);
9071 const char *func_name, *ptr_type_name;
9072 bool sleepable, rcu_lock, rcu_unlock;
9073 struct bpf_kfunc_call_arg_meta meta;
9074 u32 i, nargs, func_id, ptr_type_id;
9075 int err, insn_idx = *insn_idx_p;
9076 const struct btf_param *args;
9077 const struct btf_type *ret_t;
9078 struct btf *desc_btf;
9081 /* skip for now, but return error when we find this in fixup_kfunc_call */
9085 desc_btf = find_kfunc_desc_btf(env, insn->off);
9086 if (IS_ERR(desc_btf))
9087 return PTR_ERR(desc_btf);
9089 func_id = insn->imm;
9090 func = btf_type_by_id(desc_btf, func_id);
9091 func_name = btf_name_by_offset(desc_btf, func->name_off);
9092 func_proto = btf_type_by_id(desc_btf, func->type);
9094 kfunc_flags = btf_kfunc_id_set_contains(desc_btf, resolve_prog_type(env->prog), func_id);
9096 verbose(env, "calling kernel function %s is not allowed\n",
9101 /* Prepare kfunc call metadata */
9102 memset(&meta, 0, sizeof(meta));
9103 meta.btf = desc_btf;
9104 meta.func_id = func_id;
9105 meta.kfunc_flags = *kfunc_flags;
9106 meta.func_proto = func_proto;
9107 meta.func_name = func_name;
9109 if (is_kfunc_destructive(&meta) && !capable(CAP_SYS_BOOT)) {
9110 verbose(env, "destructive kfunc calls require CAP_SYS_BOOT capability\n");
9114 sleepable = is_kfunc_sleepable(&meta);
9115 if (sleepable && !env->prog->aux->sleepable) {
9116 verbose(env, "program must be sleepable to call sleepable kfunc %s\n", func_name);
9120 rcu_lock = is_kfunc_bpf_rcu_read_lock(&meta);
9121 rcu_unlock = is_kfunc_bpf_rcu_read_unlock(&meta);
9122 if ((rcu_lock || rcu_unlock) && !env->rcu_tag_supported) {
9123 verbose(env, "no vmlinux btf rcu tag support for kfunc %s\n", func_name);
9127 if (env->cur_state->active_rcu_lock) {
9128 struct bpf_func_state *state;
9129 struct bpf_reg_state *reg;
9132 verbose(env, "nested rcu read lock (kernel function %s)\n", func_name);
9134 } else if (rcu_unlock) {
9135 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
9136 if (reg->type & MEM_RCU) {
9137 reg->type &= ~(MEM_RCU | PTR_MAYBE_NULL);
9138 reg->type |= PTR_UNTRUSTED;
9141 env->cur_state->active_rcu_lock = false;
9142 } else if (sleepable) {
9143 verbose(env, "kernel func %s is sleepable within rcu_read_lock region\n", func_name);
9146 } else if (rcu_lock) {
9147 env->cur_state->active_rcu_lock = true;
9148 } else if (rcu_unlock) {
9149 verbose(env, "unmatched rcu read unlock (kernel function %s)\n", func_name);
9153 /* Check the arguments */
9154 err = check_kfunc_args(env, &meta);
9157 /* In case of release function, we get register number of refcounted
9158 * PTR_TO_BTF_ID in bpf_kfunc_arg_meta, do the release now.
9160 if (meta.release_regno) {
9161 err = release_reference(env, regs[meta.release_regno].ref_obj_id);
9163 verbose(env, "kfunc %s#%d reference has not been acquired before\n",
9164 func_name, func_id);
9169 for (i = 0; i < CALLER_SAVED_REGS; i++)
9170 mark_reg_not_init(env, regs, caller_saved[i]);
9172 /* Check return type */
9173 t = btf_type_skip_modifiers(desc_btf, func_proto->type, NULL);
9175 if (is_kfunc_acquire(&meta) && !btf_type_is_struct_ptr(meta.btf, t)) {
9176 /* Only exception is bpf_obj_new_impl */
9177 if (meta.btf != btf_vmlinux || meta.func_id != special_kfunc_list[KF_bpf_obj_new_impl]) {
9178 verbose(env, "acquire kernel function does not return PTR_TO_BTF_ID\n");
9183 if (btf_type_is_scalar(t)) {
9184 mark_reg_unknown(env, regs, BPF_REG_0);
9185 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
9186 } else if (btf_type_is_ptr(t)) {
9187 ptr_type = btf_type_skip_modifiers(desc_btf, t->type, &ptr_type_id);
9189 if (meta.btf == btf_vmlinux && btf_id_set_contains(&special_kfunc_set, meta.func_id)) {
9190 if (meta.func_id == special_kfunc_list[KF_bpf_obj_new_impl]) {
9191 struct btf *ret_btf;
9194 if (unlikely(!bpf_global_ma_set))
9197 if (((u64)(u32)meta.arg_constant.value) != meta.arg_constant.value) {
9198 verbose(env, "local type ID argument must be in range [0, U32_MAX]\n");
9202 ret_btf = env->prog->aux->btf;
9203 ret_btf_id = meta.arg_constant.value;
9205 /* This may be NULL due to user not supplying a BTF */
9207 verbose(env, "bpf_obj_new requires prog BTF\n");
9211 ret_t = btf_type_by_id(ret_btf, ret_btf_id);
9212 if (!ret_t || !__btf_type_is_struct(ret_t)) {
9213 verbose(env, "bpf_obj_new type ID argument must be of a struct\n");
9217 mark_reg_known_zero(env, regs, BPF_REG_0);
9218 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
9219 regs[BPF_REG_0].btf = ret_btf;
9220 regs[BPF_REG_0].btf_id = ret_btf_id;
9222 env->insn_aux_data[insn_idx].obj_new_size = ret_t->size;
9223 env->insn_aux_data[insn_idx].kptr_struct_meta =
9224 btf_find_struct_meta(ret_btf, ret_btf_id);
9225 } else if (meta.func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
9226 env->insn_aux_data[insn_idx].kptr_struct_meta =
9227 btf_find_struct_meta(meta.arg_obj_drop.btf,
9228 meta.arg_obj_drop.btf_id);
9229 } else if (meta.func_id == special_kfunc_list[KF_bpf_list_pop_front] ||
9230 meta.func_id == special_kfunc_list[KF_bpf_list_pop_back]) {
9231 struct btf_field *field = meta.arg_list_head.field;
9233 mark_reg_known_zero(env, regs, BPF_REG_0);
9234 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
9235 regs[BPF_REG_0].btf = field->list_head.btf;
9236 regs[BPF_REG_0].btf_id = field->list_head.value_btf_id;
9237 regs[BPF_REG_0].off = field->list_head.node_offset;
9238 } else if (meta.func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
9239 mark_reg_known_zero(env, regs, BPF_REG_0);
9240 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_TRUSTED;
9241 regs[BPF_REG_0].btf = desc_btf;
9242 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
9243 } else if (meta.func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
9244 ret_t = btf_type_by_id(desc_btf, meta.arg_constant.value);
9245 if (!ret_t || !btf_type_is_struct(ret_t)) {
9247 "kfunc bpf_rdonly_cast type ID argument must be of a struct\n");
9251 mark_reg_known_zero(env, regs, BPF_REG_0);
9252 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
9253 regs[BPF_REG_0].btf = desc_btf;
9254 regs[BPF_REG_0].btf_id = meta.arg_constant.value;
9256 verbose(env, "kernel function %s unhandled dynamic return type\n",
9260 } else if (!__btf_type_is_struct(ptr_type)) {
9261 if (!meta.r0_size) {
9262 ptr_type_name = btf_name_by_offset(desc_btf,
9263 ptr_type->name_off);
9265 "kernel function %s returns pointer type %s %s is not supported\n",
9267 btf_type_str(ptr_type),
9272 mark_reg_known_zero(env, regs, BPF_REG_0);
9273 regs[BPF_REG_0].type = PTR_TO_MEM;
9274 regs[BPF_REG_0].mem_size = meta.r0_size;
9277 regs[BPF_REG_0].type |= MEM_RDONLY;
9279 /* Ensures we don't access the memory after a release_reference() */
9280 if (meta.ref_obj_id)
9281 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
9283 mark_reg_known_zero(env, regs, BPF_REG_0);
9284 regs[BPF_REG_0].btf = desc_btf;
9285 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
9286 regs[BPF_REG_0].btf_id = ptr_type_id;
9289 if (is_kfunc_ret_null(&meta)) {
9290 regs[BPF_REG_0].type |= PTR_MAYBE_NULL;
9291 /* For mark_ptr_or_null_reg, see 93c230e3f5bd6 */
9292 regs[BPF_REG_0].id = ++env->id_gen;
9294 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
9295 if (is_kfunc_acquire(&meta)) {
9296 int id = acquire_reference_state(env, insn_idx);
9300 if (is_kfunc_ret_null(&meta))
9301 regs[BPF_REG_0].id = id;
9302 regs[BPF_REG_0].ref_obj_id = id;
9304 if (reg_may_point_to_spin_lock(®s[BPF_REG_0]) && !regs[BPF_REG_0].id)
9305 regs[BPF_REG_0].id = ++env->id_gen;
9306 } /* else { add_kfunc_call() ensures it is btf_type_is_void(t) } */
9308 nargs = btf_type_vlen(func_proto);
9309 args = (const struct btf_param *)(func_proto + 1);
9310 for (i = 0; i < nargs; i++) {
9313 t = btf_type_skip_modifiers(desc_btf, args[i].type, NULL);
9314 if (btf_type_is_ptr(t))
9315 mark_btf_func_reg_size(env, regno, sizeof(void *));
9317 /* scalar. ensured by btf_check_kfunc_arg_match() */
9318 mark_btf_func_reg_size(env, regno, t->size);
9324 static bool signed_add_overflows(s64 a, s64 b)
9326 /* Do the add in u64, where overflow is well-defined */
9327 s64 res = (s64)((u64)a + (u64)b);
9334 static bool signed_add32_overflows(s32 a, s32 b)
9336 /* Do the add in u32, where overflow is well-defined */
9337 s32 res = (s32)((u32)a + (u32)b);
9344 static bool signed_sub_overflows(s64 a, s64 b)
9346 /* Do the sub in u64, where overflow is well-defined */
9347 s64 res = (s64)((u64)a - (u64)b);
9354 static bool signed_sub32_overflows(s32 a, s32 b)
9356 /* Do the sub in u32, where overflow is well-defined */
9357 s32 res = (s32)((u32)a - (u32)b);
9364 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
9365 const struct bpf_reg_state *reg,
9366 enum bpf_reg_type type)
9368 bool known = tnum_is_const(reg->var_off);
9369 s64 val = reg->var_off.value;
9370 s64 smin = reg->smin_value;
9372 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
9373 verbose(env, "math between %s pointer and %lld is not allowed\n",
9374 reg_type_str(env, type), val);
9378 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
9379 verbose(env, "%s pointer offset %d is not allowed\n",
9380 reg_type_str(env, type), reg->off);
9384 if (smin == S64_MIN) {
9385 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
9386 reg_type_str(env, type));
9390 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
9391 verbose(env, "value %lld makes %s pointer be out of bounds\n",
9392 smin, reg_type_str(env, type));
9407 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
9408 u32 *alu_limit, bool mask_to_left)
9410 u32 max = 0, ptr_limit = 0;
9412 switch (ptr_reg->type) {
9414 /* Offset 0 is out-of-bounds, but acceptable start for the
9415 * left direction, see BPF_REG_FP. Also, unknown scalar
9416 * offset where we would need to deal with min/max bounds is
9417 * currently prohibited for unprivileged.
9419 max = MAX_BPF_STACK + mask_to_left;
9420 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
9422 case PTR_TO_MAP_VALUE:
9423 max = ptr_reg->map_ptr->value_size;
9424 ptr_limit = (mask_to_left ?
9425 ptr_reg->smin_value :
9426 ptr_reg->umax_value) + ptr_reg->off;
9432 if (ptr_limit >= max)
9433 return REASON_LIMIT;
9434 *alu_limit = ptr_limit;
9438 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
9439 const struct bpf_insn *insn)
9441 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
9444 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
9445 u32 alu_state, u32 alu_limit)
9447 /* If we arrived here from different branches with different
9448 * state or limits to sanitize, then this won't work.
9450 if (aux->alu_state &&
9451 (aux->alu_state != alu_state ||
9452 aux->alu_limit != alu_limit))
9453 return REASON_PATHS;
9455 /* Corresponding fixup done in do_misc_fixups(). */
9456 aux->alu_state = alu_state;
9457 aux->alu_limit = alu_limit;
9461 static int sanitize_val_alu(struct bpf_verifier_env *env,
9462 struct bpf_insn *insn)
9464 struct bpf_insn_aux_data *aux = cur_aux(env);
9466 if (can_skip_alu_sanitation(env, insn))
9469 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
9472 static bool sanitize_needed(u8 opcode)
9474 return opcode == BPF_ADD || opcode == BPF_SUB;
9477 struct bpf_sanitize_info {
9478 struct bpf_insn_aux_data aux;
9482 static struct bpf_verifier_state *
9483 sanitize_speculative_path(struct bpf_verifier_env *env,
9484 const struct bpf_insn *insn,
9485 u32 next_idx, u32 curr_idx)
9487 struct bpf_verifier_state *branch;
9488 struct bpf_reg_state *regs;
9490 branch = push_stack(env, next_idx, curr_idx, true);
9491 if (branch && insn) {
9492 regs = branch->frame[branch->curframe]->regs;
9493 if (BPF_SRC(insn->code) == BPF_K) {
9494 mark_reg_unknown(env, regs, insn->dst_reg);
9495 } else if (BPF_SRC(insn->code) == BPF_X) {
9496 mark_reg_unknown(env, regs, insn->dst_reg);
9497 mark_reg_unknown(env, regs, insn->src_reg);
9503 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
9504 struct bpf_insn *insn,
9505 const struct bpf_reg_state *ptr_reg,
9506 const struct bpf_reg_state *off_reg,
9507 struct bpf_reg_state *dst_reg,
9508 struct bpf_sanitize_info *info,
9509 const bool commit_window)
9511 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
9512 struct bpf_verifier_state *vstate = env->cur_state;
9513 bool off_is_imm = tnum_is_const(off_reg->var_off);
9514 bool off_is_neg = off_reg->smin_value < 0;
9515 bool ptr_is_dst_reg = ptr_reg == dst_reg;
9516 u8 opcode = BPF_OP(insn->code);
9517 u32 alu_state, alu_limit;
9518 struct bpf_reg_state tmp;
9522 if (can_skip_alu_sanitation(env, insn))
9525 /* We already marked aux for masking from non-speculative
9526 * paths, thus we got here in the first place. We only care
9527 * to explore bad access from here.
9529 if (vstate->speculative)
9532 if (!commit_window) {
9533 if (!tnum_is_const(off_reg->var_off) &&
9534 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
9535 return REASON_BOUNDS;
9537 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
9538 (opcode == BPF_SUB && !off_is_neg);
9541 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
9545 if (commit_window) {
9546 /* In commit phase we narrow the masking window based on
9547 * the observed pointer move after the simulated operation.
9549 alu_state = info->aux.alu_state;
9550 alu_limit = abs(info->aux.alu_limit - alu_limit);
9552 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
9553 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
9554 alu_state |= ptr_is_dst_reg ?
9555 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
9557 /* Limit pruning on unknown scalars to enable deep search for
9558 * potential masking differences from other program paths.
9561 env->explore_alu_limits = true;
9564 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
9568 /* If we're in commit phase, we're done here given we already
9569 * pushed the truncated dst_reg into the speculative verification
9572 * Also, when register is a known constant, we rewrite register-based
9573 * operation to immediate-based, and thus do not need masking (and as
9574 * a consequence, do not need to simulate the zero-truncation either).
9576 if (commit_window || off_is_imm)
9579 /* Simulate and find potential out-of-bounds access under
9580 * speculative execution from truncation as a result of
9581 * masking when off was not within expected range. If off
9582 * sits in dst, then we temporarily need to move ptr there
9583 * to simulate dst (== 0) +/-= ptr. Needed, for example,
9584 * for cases where we use K-based arithmetic in one direction
9585 * and truncated reg-based in the other in order to explore
9588 if (!ptr_is_dst_reg) {
9590 *dst_reg = *ptr_reg;
9592 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
9594 if (!ptr_is_dst_reg && ret)
9596 return !ret ? REASON_STACK : 0;
9599 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
9601 struct bpf_verifier_state *vstate = env->cur_state;
9603 /* If we simulate paths under speculation, we don't update the
9604 * insn as 'seen' such that when we verify unreachable paths in
9605 * the non-speculative domain, sanitize_dead_code() can still
9606 * rewrite/sanitize them.
9608 if (!vstate->speculative)
9609 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
9612 static int sanitize_err(struct bpf_verifier_env *env,
9613 const struct bpf_insn *insn, int reason,
9614 const struct bpf_reg_state *off_reg,
9615 const struct bpf_reg_state *dst_reg)
9617 static const char *err = "pointer arithmetic with it prohibited for !root";
9618 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
9619 u32 dst = insn->dst_reg, src = insn->src_reg;
9623 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
9624 off_reg == dst_reg ? dst : src, err);
9627 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
9628 off_reg == dst_reg ? src : dst, err);
9631 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
9635 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
9639 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
9643 verbose(env, "verifier internal error: unknown reason (%d)\n",
9651 /* check that stack access falls within stack limits and that 'reg' doesn't
9652 * have a variable offset.
9654 * Variable offset is prohibited for unprivileged mode for simplicity since it
9655 * requires corresponding support in Spectre masking for stack ALU. See also
9656 * retrieve_ptr_limit().
9659 * 'off' includes 'reg->off'.
9661 static int check_stack_access_for_ptr_arithmetic(
9662 struct bpf_verifier_env *env,
9664 const struct bpf_reg_state *reg,
9667 if (!tnum_is_const(reg->var_off)) {
9670 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
9671 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
9672 regno, tn_buf, off);
9676 if (off >= 0 || off < -MAX_BPF_STACK) {
9677 verbose(env, "R%d stack pointer arithmetic goes out of range, "
9678 "prohibited for !root; off=%d\n", regno, off);
9685 static int sanitize_check_bounds(struct bpf_verifier_env *env,
9686 const struct bpf_insn *insn,
9687 const struct bpf_reg_state *dst_reg)
9689 u32 dst = insn->dst_reg;
9691 /* For unprivileged we require that resulting offset must be in bounds
9692 * in order to be able to sanitize access later on.
9694 if (env->bypass_spec_v1)
9697 switch (dst_reg->type) {
9699 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
9700 dst_reg->off + dst_reg->var_off.value))
9703 case PTR_TO_MAP_VALUE:
9704 if (check_map_access(env, dst, dst_reg->off, 1, false, ACCESS_HELPER)) {
9705 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
9706 "prohibited for !root\n", dst);
9717 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
9718 * Caller should also handle BPF_MOV case separately.
9719 * If we return -EACCES, caller may want to try again treating pointer as a
9720 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
9722 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
9723 struct bpf_insn *insn,
9724 const struct bpf_reg_state *ptr_reg,
9725 const struct bpf_reg_state *off_reg)
9727 struct bpf_verifier_state *vstate = env->cur_state;
9728 struct bpf_func_state *state = vstate->frame[vstate->curframe];
9729 struct bpf_reg_state *regs = state->regs, *dst_reg;
9730 bool known = tnum_is_const(off_reg->var_off);
9731 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
9732 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
9733 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
9734 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
9735 struct bpf_sanitize_info info = {};
9736 u8 opcode = BPF_OP(insn->code);
9737 u32 dst = insn->dst_reg;
9740 dst_reg = ®s[dst];
9742 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
9743 smin_val > smax_val || umin_val > umax_val) {
9744 /* Taint dst register if offset had invalid bounds derived from
9745 * e.g. dead branches.
9747 __mark_reg_unknown(env, dst_reg);
9751 if (BPF_CLASS(insn->code) != BPF_ALU64) {
9752 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
9753 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
9754 __mark_reg_unknown(env, dst_reg);
9759 "R%d 32-bit pointer arithmetic prohibited\n",
9764 if (ptr_reg->type & PTR_MAYBE_NULL) {
9765 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
9766 dst, reg_type_str(env, ptr_reg->type));
9770 switch (base_type(ptr_reg->type)) {
9771 case CONST_PTR_TO_MAP:
9772 /* smin_val represents the known value */
9773 if (known && smin_val == 0 && opcode == BPF_ADD)
9776 case PTR_TO_PACKET_END:
9778 case PTR_TO_SOCK_COMMON:
9779 case PTR_TO_TCP_SOCK:
9780 case PTR_TO_XDP_SOCK:
9781 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
9782 dst, reg_type_str(env, ptr_reg->type));
9788 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
9789 * The id may be overwritten later if we create a new variable offset.
9791 dst_reg->type = ptr_reg->type;
9792 dst_reg->id = ptr_reg->id;
9794 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
9795 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
9798 /* pointer types do not carry 32-bit bounds at the moment. */
9799 __mark_reg32_unbounded(dst_reg);
9801 if (sanitize_needed(opcode)) {
9802 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
9805 return sanitize_err(env, insn, ret, off_reg, dst_reg);
9810 /* We can take a fixed offset as long as it doesn't overflow
9811 * the s32 'off' field
9813 if (known && (ptr_reg->off + smin_val ==
9814 (s64)(s32)(ptr_reg->off + smin_val))) {
9815 /* pointer += K. Accumulate it into fixed offset */
9816 dst_reg->smin_value = smin_ptr;
9817 dst_reg->smax_value = smax_ptr;
9818 dst_reg->umin_value = umin_ptr;
9819 dst_reg->umax_value = umax_ptr;
9820 dst_reg->var_off = ptr_reg->var_off;
9821 dst_reg->off = ptr_reg->off + smin_val;
9822 dst_reg->raw = ptr_reg->raw;
9825 /* A new variable offset is created. Note that off_reg->off
9826 * == 0, since it's a scalar.
9827 * dst_reg gets the pointer type and since some positive
9828 * integer value was added to the pointer, give it a new 'id'
9829 * if it's a PTR_TO_PACKET.
9830 * this creates a new 'base' pointer, off_reg (variable) gets
9831 * added into the variable offset, and we copy the fixed offset
9834 if (signed_add_overflows(smin_ptr, smin_val) ||
9835 signed_add_overflows(smax_ptr, smax_val)) {
9836 dst_reg->smin_value = S64_MIN;
9837 dst_reg->smax_value = S64_MAX;
9839 dst_reg->smin_value = smin_ptr + smin_val;
9840 dst_reg->smax_value = smax_ptr + smax_val;
9842 if (umin_ptr + umin_val < umin_ptr ||
9843 umax_ptr + umax_val < umax_ptr) {
9844 dst_reg->umin_value = 0;
9845 dst_reg->umax_value = U64_MAX;
9847 dst_reg->umin_value = umin_ptr + umin_val;
9848 dst_reg->umax_value = umax_ptr + umax_val;
9850 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
9851 dst_reg->off = ptr_reg->off;
9852 dst_reg->raw = ptr_reg->raw;
9853 if (reg_is_pkt_pointer(ptr_reg)) {
9854 dst_reg->id = ++env->id_gen;
9855 /* something was added to pkt_ptr, set range to zero */
9856 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
9860 if (dst_reg == off_reg) {
9861 /* scalar -= pointer. Creates an unknown scalar */
9862 verbose(env, "R%d tried to subtract pointer from scalar\n",
9866 /* We don't allow subtraction from FP, because (according to
9867 * test_verifier.c test "invalid fp arithmetic", JITs might not
9868 * be able to deal with it.
9870 if (ptr_reg->type == PTR_TO_STACK) {
9871 verbose(env, "R%d subtraction from stack pointer prohibited\n",
9875 if (known && (ptr_reg->off - smin_val ==
9876 (s64)(s32)(ptr_reg->off - smin_val))) {
9877 /* pointer -= K. Subtract it from fixed offset */
9878 dst_reg->smin_value = smin_ptr;
9879 dst_reg->smax_value = smax_ptr;
9880 dst_reg->umin_value = umin_ptr;
9881 dst_reg->umax_value = umax_ptr;
9882 dst_reg->var_off = ptr_reg->var_off;
9883 dst_reg->id = ptr_reg->id;
9884 dst_reg->off = ptr_reg->off - smin_val;
9885 dst_reg->raw = ptr_reg->raw;
9888 /* A new variable offset is created. If the subtrahend is known
9889 * nonnegative, then any reg->range we had before is still good.
9891 if (signed_sub_overflows(smin_ptr, smax_val) ||
9892 signed_sub_overflows(smax_ptr, smin_val)) {
9893 /* Overflow possible, we know nothing */
9894 dst_reg->smin_value = S64_MIN;
9895 dst_reg->smax_value = S64_MAX;
9897 dst_reg->smin_value = smin_ptr - smax_val;
9898 dst_reg->smax_value = smax_ptr - smin_val;
9900 if (umin_ptr < umax_val) {
9901 /* Overflow possible, we know nothing */
9902 dst_reg->umin_value = 0;
9903 dst_reg->umax_value = U64_MAX;
9905 /* Cannot overflow (as long as bounds are consistent) */
9906 dst_reg->umin_value = umin_ptr - umax_val;
9907 dst_reg->umax_value = umax_ptr - umin_val;
9909 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
9910 dst_reg->off = ptr_reg->off;
9911 dst_reg->raw = ptr_reg->raw;
9912 if (reg_is_pkt_pointer(ptr_reg)) {
9913 dst_reg->id = ++env->id_gen;
9914 /* something was added to pkt_ptr, set range to zero */
9916 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
9922 /* bitwise ops on pointers are troublesome, prohibit. */
9923 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
9924 dst, bpf_alu_string[opcode >> 4]);
9927 /* other operators (e.g. MUL,LSH) produce non-pointer results */
9928 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
9929 dst, bpf_alu_string[opcode >> 4]);
9933 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
9935 reg_bounds_sync(dst_reg);
9936 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
9938 if (sanitize_needed(opcode)) {
9939 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
9942 return sanitize_err(env, insn, ret, off_reg, dst_reg);
9948 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
9949 struct bpf_reg_state *src_reg)
9951 s32 smin_val = src_reg->s32_min_value;
9952 s32 smax_val = src_reg->s32_max_value;
9953 u32 umin_val = src_reg->u32_min_value;
9954 u32 umax_val = src_reg->u32_max_value;
9956 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
9957 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
9958 dst_reg->s32_min_value = S32_MIN;
9959 dst_reg->s32_max_value = S32_MAX;
9961 dst_reg->s32_min_value += smin_val;
9962 dst_reg->s32_max_value += smax_val;
9964 if (dst_reg->u32_min_value + umin_val < umin_val ||
9965 dst_reg->u32_max_value + umax_val < umax_val) {
9966 dst_reg->u32_min_value = 0;
9967 dst_reg->u32_max_value = U32_MAX;
9969 dst_reg->u32_min_value += umin_val;
9970 dst_reg->u32_max_value += umax_val;
9974 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
9975 struct bpf_reg_state *src_reg)
9977 s64 smin_val = src_reg->smin_value;
9978 s64 smax_val = src_reg->smax_value;
9979 u64 umin_val = src_reg->umin_value;
9980 u64 umax_val = src_reg->umax_value;
9982 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
9983 signed_add_overflows(dst_reg->smax_value, smax_val)) {
9984 dst_reg->smin_value = S64_MIN;
9985 dst_reg->smax_value = S64_MAX;
9987 dst_reg->smin_value += smin_val;
9988 dst_reg->smax_value += smax_val;
9990 if (dst_reg->umin_value + umin_val < umin_val ||
9991 dst_reg->umax_value + umax_val < umax_val) {
9992 dst_reg->umin_value = 0;
9993 dst_reg->umax_value = U64_MAX;
9995 dst_reg->umin_value += umin_val;
9996 dst_reg->umax_value += umax_val;
10000 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
10001 struct bpf_reg_state *src_reg)
10003 s32 smin_val = src_reg->s32_min_value;
10004 s32 smax_val = src_reg->s32_max_value;
10005 u32 umin_val = src_reg->u32_min_value;
10006 u32 umax_val = src_reg->u32_max_value;
10008 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
10009 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
10010 /* Overflow possible, we know nothing */
10011 dst_reg->s32_min_value = S32_MIN;
10012 dst_reg->s32_max_value = S32_MAX;
10014 dst_reg->s32_min_value -= smax_val;
10015 dst_reg->s32_max_value -= smin_val;
10017 if (dst_reg->u32_min_value < umax_val) {
10018 /* Overflow possible, we know nothing */
10019 dst_reg->u32_min_value = 0;
10020 dst_reg->u32_max_value = U32_MAX;
10022 /* Cannot overflow (as long as bounds are consistent) */
10023 dst_reg->u32_min_value -= umax_val;
10024 dst_reg->u32_max_value -= umin_val;
10028 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
10029 struct bpf_reg_state *src_reg)
10031 s64 smin_val = src_reg->smin_value;
10032 s64 smax_val = src_reg->smax_value;
10033 u64 umin_val = src_reg->umin_value;
10034 u64 umax_val = src_reg->umax_value;
10036 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
10037 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
10038 /* Overflow possible, we know nothing */
10039 dst_reg->smin_value = S64_MIN;
10040 dst_reg->smax_value = S64_MAX;
10042 dst_reg->smin_value -= smax_val;
10043 dst_reg->smax_value -= smin_val;
10045 if (dst_reg->umin_value < umax_val) {
10046 /* Overflow possible, we know nothing */
10047 dst_reg->umin_value = 0;
10048 dst_reg->umax_value = U64_MAX;
10050 /* Cannot overflow (as long as bounds are consistent) */
10051 dst_reg->umin_value -= umax_val;
10052 dst_reg->umax_value -= umin_val;
10056 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
10057 struct bpf_reg_state *src_reg)
10059 s32 smin_val = src_reg->s32_min_value;
10060 u32 umin_val = src_reg->u32_min_value;
10061 u32 umax_val = src_reg->u32_max_value;
10063 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
10064 /* Ain't nobody got time to multiply that sign */
10065 __mark_reg32_unbounded(dst_reg);
10068 /* Both values are positive, so we can work with unsigned and
10069 * copy the result to signed (unless it exceeds S32_MAX).
10071 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
10072 /* Potential overflow, we know nothing */
10073 __mark_reg32_unbounded(dst_reg);
10076 dst_reg->u32_min_value *= umin_val;
10077 dst_reg->u32_max_value *= umax_val;
10078 if (dst_reg->u32_max_value > S32_MAX) {
10079 /* Overflow possible, we know nothing */
10080 dst_reg->s32_min_value = S32_MIN;
10081 dst_reg->s32_max_value = S32_MAX;
10083 dst_reg->s32_min_value = dst_reg->u32_min_value;
10084 dst_reg->s32_max_value = dst_reg->u32_max_value;
10088 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
10089 struct bpf_reg_state *src_reg)
10091 s64 smin_val = src_reg->smin_value;
10092 u64 umin_val = src_reg->umin_value;
10093 u64 umax_val = src_reg->umax_value;
10095 if (smin_val < 0 || dst_reg->smin_value < 0) {
10096 /* Ain't nobody got time to multiply that sign */
10097 __mark_reg64_unbounded(dst_reg);
10100 /* Both values are positive, so we can work with unsigned and
10101 * copy the result to signed (unless it exceeds S64_MAX).
10103 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
10104 /* Potential overflow, we know nothing */
10105 __mark_reg64_unbounded(dst_reg);
10108 dst_reg->umin_value *= umin_val;
10109 dst_reg->umax_value *= umax_val;
10110 if (dst_reg->umax_value > S64_MAX) {
10111 /* Overflow possible, we know nothing */
10112 dst_reg->smin_value = S64_MIN;
10113 dst_reg->smax_value = S64_MAX;
10115 dst_reg->smin_value = dst_reg->umin_value;
10116 dst_reg->smax_value = dst_reg->umax_value;
10120 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
10121 struct bpf_reg_state *src_reg)
10123 bool src_known = tnum_subreg_is_const(src_reg->var_off);
10124 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
10125 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
10126 s32 smin_val = src_reg->s32_min_value;
10127 u32 umax_val = src_reg->u32_max_value;
10129 if (src_known && dst_known) {
10130 __mark_reg32_known(dst_reg, var32_off.value);
10134 /* We get our minimum from the var_off, since that's inherently
10135 * bitwise. Our maximum is the minimum of the operands' maxima.
10137 dst_reg->u32_min_value = var32_off.value;
10138 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
10139 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
10140 /* Lose signed bounds when ANDing negative numbers,
10141 * ain't nobody got time for that.
10143 dst_reg->s32_min_value = S32_MIN;
10144 dst_reg->s32_max_value = S32_MAX;
10146 /* ANDing two positives gives a positive, so safe to
10147 * cast result into s64.
10149 dst_reg->s32_min_value = dst_reg->u32_min_value;
10150 dst_reg->s32_max_value = dst_reg->u32_max_value;
10154 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
10155 struct bpf_reg_state *src_reg)
10157 bool src_known = tnum_is_const(src_reg->var_off);
10158 bool dst_known = tnum_is_const(dst_reg->var_off);
10159 s64 smin_val = src_reg->smin_value;
10160 u64 umax_val = src_reg->umax_value;
10162 if (src_known && dst_known) {
10163 __mark_reg_known(dst_reg, dst_reg->var_off.value);
10167 /* We get our minimum from the var_off, since that's inherently
10168 * bitwise. Our maximum is the minimum of the operands' maxima.
10170 dst_reg->umin_value = dst_reg->var_off.value;
10171 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
10172 if (dst_reg->smin_value < 0 || smin_val < 0) {
10173 /* Lose signed bounds when ANDing negative numbers,
10174 * ain't nobody got time for that.
10176 dst_reg->smin_value = S64_MIN;
10177 dst_reg->smax_value = S64_MAX;
10179 /* ANDing two positives gives a positive, so safe to
10180 * cast result into s64.
10182 dst_reg->smin_value = dst_reg->umin_value;
10183 dst_reg->smax_value = dst_reg->umax_value;
10185 /* We may learn something more from the var_off */
10186 __update_reg_bounds(dst_reg);
10189 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
10190 struct bpf_reg_state *src_reg)
10192 bool src_known = tnum_subreg_is_const(src_reg->var_off);
10193 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
10194 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
10195 s32 smin_val = src_reg->s32_min_value;
10196 u32 umin_val = src_reg->u32_min_value;
10198 if (src_known && dst_known) {
10199 __mark_reg32_known(dst_reg, var32_off.value);
10203 /* We get our maximum from the var_off, and our minimum is the
10204 * maximum of the operands' minima
10206 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
10207 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
10208 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
10209 /* Lose signed bounds when ORing negative numbers,
10210 * ain't nobody got time for that.
10212 dst_reg->s32_min_value = S32_MIN;
10213 dst_reg->s32_max_value = S32_MAX;
10215 /* ORing two positives gives a positive, so safe to
10216 * cast result into s64.
10218 dst_reg->s32_min_value = dst_reg->u32_min_value;
10219 dst_reg->s32_max_value = dst_reg->u32_max_value;
10223 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
10224 struct bpf_reg_state *src_reg)
10226 bool src_known = tnum_is_const(src_reg->var_off);
10227 bool dst_known = tnum_is_const(dst_reg->var_off);
10228 s64 smin_val = src_reg->smin_value;
10229 u64 umin_val = src_reg->umin_value;
10231 if (src_known && dst_known) {
10232 __mark_reg_known(dst_reg, dst_reg->var_off.value);
10236 /* We get our maximum from the var_off, and our minimum is the
10237 * maximum of the operands' minima
10239 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
10240 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
10241 if (dst_reg->smin_value < 0 || smin_val < 0) {
10242 /* Lose signed bounds when ORing negative numbers,
10243 * ain't nobody got time for that.
10245 dst_reg->smin_value = S64_MIN;
10246 dst_reg->smax_value = S64_MAX;
10248 /* ORing two positives gives a positive, so safe to
10249 * cast result into s64.
10251 dst_reg->smin_value = dst_reg->umin_value;
10252 dst_reg->smax_value = dst_reg->umax_value;
10254 /* We may learn something more from the var_off */
10255 __update_reg_bounds(dst_reg);
10258 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
10259 struct bpf_reg_state *src_reg)
10261 bool src_known = tnum_subreg_is_const(src_reg->var_off);
10262 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
10263 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
10264 s32 smin_val = src_reg->s32_min_value;
10266 if (src_known && dst_known) {
10267 __mark_reg32_known(dst_reg, var32_off.value);
10271 /* We get both minimum and maximum from the var32_off. */
10272 dst_reg->u32_min_value = var32_off.value;
10273 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
10275 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
10276 /* XORing two positive sign numbers gives a positive,
10277 * so safe to cast u32 result into s32.
10279 dst_reg->s32_min_value = dst_reg->u32_min_value;
10280 dst_reg->s32_max_value = dst_reg->u32_max_value;
10282 dst_reg->s32_min_value = S32_MIN;
10283 dst_reg->s32_max_value = S32_MAX;
10287 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
10288 struct bpf_reg_state *src_reg)
10290 bool src_known = tnum_is_const(src_reg->var_off);
10291 bool dst_known = tnum_is_const(dst_reg->var_off);
10292 s64 smin_val = src_reg->smin_value;
10294 if (src_known && dst_known) {
10295 /* dst_reg->var_off.value has been updated earlier */
10296 __mark_reg_known(dst_reg, dst_reg->var_off.value);
10300 /* We get both minimum and maximum from the var_off. */
10301 dst_reg->umin_value = dst_reg->var_off.value;
10302 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
10304 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
10305 /* XORing two positive sign numbers gives a positive,
10306 * so safe to cast u64 result into s64.
10308 dst_reg->smin_value = dst_reg->umin_value;
10309 dst_reg->smax_value = dst_reg->umax_value;
10311 dst_reg->smin_value = S64_MIN;
10312 dst_reg->smax_value = S64_MAX;
10315 __update_reg_bounds(dst_reg);
10318 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
10319 u64 umin_val, u64 umax_val)
10321 /* We lose all sign bit information (except what we can pick
10324 dst_reg->s32_min_value = S32_MIN;
10325 dst_reg->s32_max_value = S32_MAX;
10326 /* If we might shift our top bit out, then we know nothing */
10327 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
10328 dst_reg->u32_min_value = 0;
10329 dst_reg->u32_max_value = U32_MAX;
10331 dst_reg->u32_min_value <<= umin_val;
10332 dst_reg->u32_max_value <<= umax_val;
10336 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
10337 struct bpf_reg_state *src_reg)
10339 u32 umax_val = src_reg->u32_max_value;
10340 u32 umin_val = src_reg->u32_min_value;
10341 /* u32 alu operation will zext upper bits */
10342 struct tnum subreg = tnum_subreg(dst_reg->var_off);
10344 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
10345 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
10346 /* Not required but being careful mark reg64 bounds as unknown so
10347 * that we are forced to pick them up from tnum and zext later and
10348 * if some path skips this step we are still safe.
10350 __mark_reg64_unbounded(dst_reg);
10351 __update_reg32_bounds(dst_reg);
10354 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
10355 u64 umin_val, u64 umax_val)
10357 /* Special case <<32 because it is a common compiler pattern to sign
10358 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
10359 * positive we know this shift will also be positive so we can track
10360 * bounds correctly. Otherwise we lose all sign bit information except
10361 * what we can pick up from var_off. Perhaps we can generalize this
10362 * later to shifts of any length.
10364 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
10365 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
10367 dst_reg->smax_value = S64_MAX;
10369 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
10370 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
10372 dst_reg->smin_value = S64_MIN;
10374 /* If we might shift our top bit out, then we know nothing */
10375 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
10376 dst_reg->umin_value = 0;
10377 dst_reg->umax_value = U64_MAX;
10379 dst_reg->umin_value <<= umin_val;
10380 dst_reg->umax_value <<= umax_val;
10384 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
10385 struct bpf_reg_state *src_reg)
10387 u64 umax_val = src_reg->umax_value;
10388 u64 umin_val = src_reg->umin_value;
10390 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
10391 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
10392 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
10394 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
10395 /* We may learn something more from the var_off */
10396 __update_reg_bounds(dst_reg);
10399 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
10400 struct bpf_reg_state *src_reg)
10402 struct tnum subreg = tnum_subreg(dst_reg->var_off);
10403 u32 umax_val = src_reg->u32_max_value;
10404 u32 umin_val = src_reg->u32_min_value;
10406 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
10407 * be negative, then either:
10408 * 1) src_reg might be zero, so the sign bit of the result is
10409 * unknown, so we lose our signed bounds
10410 * 2) it's known negative, thus the unsigned bounds capture the
10412 * 3) the signed bounds cross zero, so they tell us nothing
10414 * If the value in dst_reg is known nonnegative, then again the
10415 * unsigned bounds capture the signed bounds.
10416 * Thus, in all cases it suffices to blow away our signed bounds
10417 * and rely on inferring new ones from the unsigned bounds and
10418 * var_off of the result.
10420 dst_reg->s32_min_value = S32_MIN;
10421 dst_reg->s32_max_value = S32_MAX;
10423 dst_reg->var_off = tnum_rshift(subreg, umin_val);
10424 dst_reg->u32_min_value >>= umax_val;
10425 dst_reg->u32_max_value >>= umin_val;
10427 __mark_reg64_unbounded(dst_reg);
10428 __update_reg32_bounds(dst_reg);
10431 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
10432 struct bpf_reg_state *src_reg)
10434 u64 umax_val = src_reg->umax_value;
10435 u64 umin_val = src_reg->umin_value;
10437 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
10438 * be negative, then either:
10439 * 1) src_reg might be zero, so the sign bit of the result is
10440 * unknown, so we lose our signed bounds
10441 * 2) it's known negative, thus the unsigned bounds capture the
10443 * 3) the signed bounds cross zero, so they tell us nothing
10445 * If the value in dst_reg is known nonnegative, then again the
10446 * unsigned bounds capture the signed bounds.
10447 * Thus, in all cases it suffices to blow away our signed bounds
10448 * and rely on inferring new ones from the unsigned bounds and
10449 * var_off of the result.
10451 dst_reg->smin_value = S64_MIN;
10452 dst_reg->smax_value = S64_MAX;
10453 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
10454 dst_reg->umin_value >>= umax_val;
10455 dst_reg->umax_value >>= umin_val;
10457 /* Its not easy to operate on alu32 bounds here because it depends
10458 * on bits being shifted in. Take easy way out and mark unbounded
10459 * so we can recalculate later from tnum.
10461 __mark_reg32_unbounded(dst_reg);
10462 __update_reg_bounds(dst_reg);
10465 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
10466 struct bpf_reg_state *src_reg)
10468 u64 umin_val = src_reg->u32_min_value;
10470 /* Upon reaching here, src_known is true and
10471 * umax_val is equal to umin_val.
10473 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
10474 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
10476 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
10478 /* blow away the dst_reg umin_value/umax_value and rely on
10479 * dst_reg var_off to refine the result.
10481 dst_reg->u32_min_value = 0;
10482 dst_reg->u32_max_value = U32_MAX;
10484 __mark_reg64_unbounded(dst_reg);
10485 __update_reg32_bounds(dst_reg);
10488 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
10489 struct bpf_reg_state *src_reg)
10491 u64 umin_val = src_reg->umin_value;
10493 /* Upon reaching here, src_known is true and umax_val is equal
10496 dst_reg->smin_value >>= umin_val;
10497 dst_reg->smax_value >>= umin_val;
10499 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
10501 /* blow away the dst_reg umin_value/umax_value and rely on
10502 * dst_reg var_off to refine the result.
10504 dst_reg->umin_value = 0;
10505 dst_reg->umax_value = U64_MAX;
10507 /* Its not easy to operate on alu32 bounds here because it depends
10508 * on bits being shifted in from upper 32-bits. Take easy way out
10509 * and mark unbounded so we can recalculate later from tnum.
10511 __mark_reg32_unbounded(dst_reg);
10512 __update_reg_bounds(dst_reg);
10515 /* WARNING: This function does calculations on 64-bit values, but the actual
10516 * execution may occur on 32-bit values. Therefore, things like bitshifts
10517 * need extra checks in the 32-bit case.
10519 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
10520 struct bpf_insn *insn,
10521 struct bpf_reg_state *dst_reg,
10522 struct bpf_reg_state src_reg)
10524 struct bpf_reg_state *regs = cur_regs(env);
10525 u8 opcode = BPF_OP(insn->code);
10527 s64 smin_val, smax_val;
10528 u64 umin_val, umax_val;
10529 s32 s32_min_val, s32_max_val;
10530 u32 u32_min_val, u32_max_val;
10531 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
10532 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
10535 smin_val = src_reg.smin_value;
10536 smax_val = src_reg.smax_value;
10537 umin_val = src_reg.umin_value;
10538 umax_val = src_reg.umax_value;
10540 s32_min_val = src_reg.s32_min_value;
10541 s32_max_val = src_reg.s32_max_value;
10542 u32_min_val = src_reg.u32_min_value;
10543 u32_max_val = src_reg.u32_max_value;
10546 src_known = tnum_subreg_is_const(src_reg.var_off);
10548 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
10549 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
10550 /* Taint dst register if offset had invalid bounds
10551 * derived from e.g. dead branches.
10553 __mark_reg_unknown(env, dst_reg);
10557 src_known = tnum_is_const(src_reg.var_off);
10559 (smin_val != smax_val || umin_val != umax_val)) ||
10560 smin_val > smax_val || umin_val > umax_val) {
10561 /* Taint dst register if offset had invalid bounds
10562 * derived from e.g. dead branches.
10564 __mark_reg_unknown(env, dst_reg);
10570 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
10571 __mark_reg_unknown(env, dst_reg);
10575 if (sanitize_needed(opcode)) {
10576 ret = sanitize_val_alu(env, insn);
10578 return sanitize_err(env, insn, ret, NULL, NULL);
10581 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
10582 * There are two classes of instructions: The first class we track both
10583 * alu32 and alu64 sign/unsigned bounds independently this provides the
10584 * greatest amount of precision when alu operations are mixed with jmp32
10585 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
10586 * and BPF_OR. This is possible because these ops have fairly easy to
10587 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
10588 * See alu32 verifier tests for examples. The second class of
10589 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
10590 * with regards to tracking sign/unsigned bounds because the bits may
10591 * cross subreg boundaries in the alu64 case. When this happens we mark
10592 * the reg unbounded in the subreg bound space and use the resulting
10593 * tnum to calculate an approximation of the sign/unsigned bounds.
10597 scalar32_min_max_add(dst_reg, &src_reg);
10598 scalar_min_max_add(dst_reg, &src_reg);
10599 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
10602 scalar32_min_max_sub(dst_reg, &src_reg);
10603 scalar_min_max_sub(dst_reg, &src_reg);
10604 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
10607 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
10608 scalar32_min_max_mul(dst_reg, &src_reg);
10609 scalar_min_max_mul(dst_reg, &src_reg);
10612 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
10613 scalar32_min_max_and(dst_reg, &src_reg);
10614 scalar_min_max_and(dst_reg, &src_reg);
10617 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
10618 scalar32_min_max_or(dst_reg, &src_reg);
10619 scalar_min_max_or(dst_reg, &src_reg);
10622 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
10623 scalar32_min_max_xor(dst_reg, &src_reg);
10624 scalar_min_max_xor(dst_reg, &src_reg);
10627 if (umax_val >= insn_bitness) {
10628 /* Shifts greater than 31 or 63 are undefined.
10629 * This includes shifts by a negative number.
10631 mark_reg_unknown(env, regs, insn->dst_reg);
10635 scalar32_min_max_lsh(dst_reg, &src_reg);
10637 scalar_min_max_lsh(dst_reg, &src_reg);
10640 if (umax_val >= insn_bitness) {
10641 /* Shifts greater than 31 or 63 are undefined.
10642 * This includes shifts by a negative number.
10644 mark_reg_unknown(env, regs, insn->dst_reg);
10648 scalar32_min_max_rsh(dst_reg, &src_reg);
10650 scalar_min_max_rsh(dst_reg, &src_reg);
10653 if (umax_val >= insn_bitness) {
10654 /* Shifts greater than 31 or 63 are undefined.
10655 * This includes shifts by a negative number.
10657 mark_reg_unknown(env, regs, insn->dst_reg);
10661 scalar32_min_max_arsh(dst_reg, &src_reg);
10663 scalar_min_max_arsh(dst_reg, &src_reg);
10666 mark_reg_unknown(env, regs, insn->dst_reg);
10670 /* ALU32 ops are zero extended into 64bit register */
10672 zext_32_to_64(dst_reg);
10673 reg_bounds_sync(dst_reg);
10677 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
10680 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
10681 struct bpf_insn *insn)
10683 struct bpf_verifier_state *vstate = env->cur_state;
10684 struct bpf_func_state *state = vstate->frame[vstate->curframe];
10685 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
10686 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
10687 u8 opcode = BPF_OP(insn->code);
10690 dst_reg = ®s[insn->dst_reg];
10692 if (dst_reg->type != SCALAR_VALUE)
10695 /* Make sure ID is cleared otherwise dst_reg min/max could be
10696 * incorrectly propagated into other registers by find_equal_scalars()
10699 if (BPF_SRC(insn->code) == BPF_X) {
10700 src_reg = ®s[insn->src_reg];
10701 if (src_reg->type != SCALAR_VALUE) {
10702 if (dst_reg->type != SCALAR_VALUE) {
10703 /* Combining two pointers by any ALU op yields
10704 * an arbitrary scalar. Disallow all math except
10705 * pointer subtraction
10707 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
10708 mark_reg_unknown(env, regs, insn->dst_reg);
10711 verbose(env, "R%d pointer %s pointer prohibited\n",
10713 bpf_alu_string[opcode >> 4]);
10716 /* scalar += pointer
10717 * This is legal, but we have to reverse our
10718 * src/dest handling in computing the range
10720 err = mark_chain_precision(env, insn->dst_reg);
10723 return adjust_ptr_min_max_vals(env, insn,
10726 } else if (ptr_reg) {
10727 /* pointer += scalar */
10728 err = mark_chain_precision(env, insn->src_reg);
10731 return adjust_ptr_min_max_vals(env, insn,
10733 } else if (dst_reg->precise) {
10734 /* if dst_reg is precise, src_reg should be precise as well */
10735 err = mark_chain_precision(env, insn->src_reg);
10740 /* Pretend the src is a reg with a known value, since we only
10741 * need to be able to read from this state.
10743 off_reg.type = SCALAR_VALUE;
10744 __mark_reg_known(&off_reg, insn->imm);
10745 src_reg = &off_reg;
10746 if (ptr_reg) /* pointer += K */
10747 return adjust_ptr_min_max_vals(env, insn,
10751 /* Got here implies adding two SCALAR_VALUEs */
10752 if (WARN_ON_ONCE(ptr_reg)) {
10753 print_verifier_state(env, state, true);
10754 verbose(env, "verifier internal error: unexpected ptr_reg\n");
10757 if (WARN_ON(!src_reg)) {
10758 print_verifier_state(env, state, true);
10759 verbose(env, "verifier internal error: no src_reg\n");
10762 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
10765 /* check validity of 32-bit and 64-bit arithmetic operations */
10766 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
10768 struct bpf_reg_state *regs = cur_regs(env);
10769 u8 opcode = BPF_OP(insn->code);
10772 if (opcode == BPF_END || opcode == BPF_NEG) {
10773 if (opcode == BPF_NEG) {
10774 if (BPF_SRC(insn->code) != BPF_K ||
10775 insn->src_reg != BPF_REG_0 ||
10776 insn->off != 0 || insn->imm != 0) {
10777 verbose(env, "BPF_NEG uses reserved fields\n");
10781 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
10782 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
10783 BPF_CLASS(insn->code) == BPF_ALU64) {
10784 verbose(env, "BPF_END uses reserved fields\n");
10789 /* check src operand */
10790 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
10794 if (is_pointer_value(env, insn->dst_reg)) {
10795 verbose(env, "R%d pointer arithmetic prohibited\n",
10800 /* check dest operand */
10801 err = check_reg_arg(env, insn->dst_reg, DST_OP);
10805 } else if (opcode == BPF_MOV) {
10807 if (BPF_SRC(insn->code) == BPF_X) {
10808 if (insn->imm != 0 || insn->off != 0) {
10809 verbose(env, "BPF_MOV uses reserved fields\n");
10813 /* check src operand */
10814 err = check_reg_arg(env, insn->src_reg, SRC_OP);
10818 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
10819 verbose(env, "BPF_MOV uses reserved fields\n");
10824 /* check dest operand, mark as required later */
10825 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
10829 if (BPF_SRC(insn->code) == BPF_X) {
10830 struct bpf_reg_state *src_reg = regs + insn->src_reg;
10831 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
10833 if (BPF_CLASS(insn->code) == BPF_ALU64) {
10835 * copy register state to dest reg
10837 if (src_reg->type == SCALAR_VALUE && !src_reg->id)
10838 /* Assign src and dst registers the same ID
10839 * that will be used by find_equal_scalars()
10840 * to propagate min/max range.
10842 src_reg->id = ++env->id_gen;
10843 *dst_reg = *src_reg;
10844 dst_reg->live |= REG_LIVE_WRITTEN;
10845 dst_reg->subreg_def = DEF_NOT_SUBREG;
10847 /* R1 = (u32) R2 */
10848 if (is_pointer_value(env, insn->src_reg)) {
10850 "R%d partial copy of pointer\n",
10853 } else if (src_reg->type == SCALAR_VALUE) {
10854 *dst_reg = *src_reg;
10855 /* Make sure ID is cleared otherwise
10856 * dst_reg min/max could be incorrectly
10857 * propagated into src_reg by find_equal_scalars()
10860 dst_reg->live |= REG_LIVE_WRITTEN;
10861 dst_reg->subreg_def = env->insn_idx + 1;
10863 mark_reg_unknown(env, regs,
10866 zext_32_to_64(dst_reg);
10867 reg_bounds_sync(dst_reg);
10871 * remember the value we stored into this reg
10873 /* clear any state __mark_reg_known doesn't set */
10874 mark_reg_unknown(env, regs, insn->dst_reg);
10875 regs[insn->dst_reg].type = SCALAR_VALUE;
10876 if (BPF_CLASS(insn->code) == BPF_ALU64) {
10877 __mark_reg_known(regs + insn->dst_reg,
10880 __mark_reg_known(regs + insn->dst_reg,
10885 } else if (opcode > BPF_END) {
10886 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
10889 } else { /* all other ALU ops: and, sub, xor, add, ... */
10891 if (BPF_SRC(insn->code) == BPF_X) {
10892 if (insn->imm != 0 || insn->off != 0) {
10893 verbose(env, "BPF_ALU uses reserved fields\n");
10896 /* check src1 operand */
10897 err = check_reg_arg(env, insn->src_reg, SRC_OP);
10901 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
10902 verbose(env, "BPF_ALU uses reserved fields\n");
10907 /* check src2 operand */
10908 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
10912 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
10913 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
10914 verbose(env, "div by zero\n");
10918 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
10919 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
10920 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
10922 if (insn->imm < 0 || insn->imm >= size) {
10923 verbose(env, "invalid shift %d\n", insn->imm);
10928 /* check dest operand */
10929 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
10933 return adjust_reg_min_max_vals(env, insn);
10939 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
10940 struct bpf_reg_state *dst_reg,
10941 enum bpf_reg_type type,
10942 bool range_right_open)
10944 struct bpf_func_state *state;
10945 struct bpf_reg_state *reg;
10948 if (dst_reg->off < 0 ||
10949 (dst_reg->off == 0 && range_right_open))
10950 /* This doesn't give us any range */
10953 if (dst_reg->umax_value > MAX_PACKET_OFF ||
10954 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
10955 /* Risk of overflow. For instance, ptr + (1<<63) may be less
10956 * than pkt_end, but that's because it's also less than pkt.
10960 new_range = dst_reg->off;
10961 if (range_right_open)
10964 /* Examples for register markings:
10966 * pkt_data in dst register:
10970 * if (r2 > pkt_end) goto <handle exception>
10975 * if (r2 < pkt_end) goto <access okay>
10976 * <handle exception>
10979 * r2 == dst_reg, pkt_end == src_reg
10980 * r2=pkt(id=n,off=8,r=0)
10981 * r3=pkt(id=n,off=0,r=0)
10983 * pkt_data in src register:
10987 * if (pkt_end >= r2) goto <access okay>
10988 * <handle exception>
10992 * if (pkt_end <= r2) goto <handle exception>
10996 * pkt_end == dst_reg, r2 == src_reg
10997 * r2=pkt(id=n,off=8,r=0)
10998 * r3=pkt(id=n,off=0,r=0)
11000 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
11001 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
11002 * and [r3, r3 + 8-1) respectively is safe to access depending on
11006 /* If our ids match, then we must have the same max_value. And we
11007 * don't care about the other reg's fixed offset, since if it's too big
11008 * the range won't allow anything.
11009 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
11011 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
11012 if (reg->type == type && reg->id == dst_reg->id)
11013 /* keep the maximum range already checked */
11014 reg->range = max(reg->range, new_range);
11018 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
11020 struct tnum subreg = tnum_subreg(reg->var_off);
11021 s32 sval = (s32)val;
11025 if (tnum_is_const(subreg))
11026 return !!tnum_equals_const(subreg, val);
11029 if (tnum_is_const(subreg))
11030 return !tnum_equals_const(subreg, val);
11033 if ((~subreg.mask & subreg.value) & val)
11035 if (!((subreg.mask | subreg.value) & val))
11039 if (reg->u32_min_value > val)
11041 else if (reg->u32_max_value <= val)
11045 if (reg->s32_min_value > sval)
11047 else if (reg->s32_max_value <= sval)
11051 if (reg->u32_max_value < val)
11053 else if (reg->u32_min_value >= val)
11057 if (reg->s32_max_value < sval)
11059 else if (reg->s32_min_value >= sval)
11063 if (reg->u32_min_value >= val)
11065 else if (reg->u32_max_value < val)
11069 if (reg->s32_min_value >= sval)
11071 else if (reg->s32_max_value < sval)
11075 if (reg->u32_max_value <= val)
11077 else if (reg->u32_min_value > val)
11081 if (reg->s32_max_value <= sval)
11083 else if (reg->s32_min_value > sval)
11092 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
11094 s64 sval = (s64)val;
11098 if (tnum_is_const(reg->var_off))
11099 return !!tnum_equals_const(reg->var_off, val);
11102 if (tnum_is_const(reg->var_off))
11103 return !tnum_equals_const(reg->var_off, val);
11106 if ((~reg->var_off.mask & reg->var_off.value) & val)
11108 if (!((reg->var_off.mask | reg->var_off.value) & val))
11112 if (reg->umin_value > val)
11114 else if (reg->umax_value <= val)
11118 if (reg->smin_value > sval)
11120 else if (reg->smax_value <= sval)
11124 if (reg->umax_value < val)
11126 else if (reg->umin_value >= val)
11130 if (reg->smax_value < sval)
11132 else if (reg->smin_value >= sval)
11136 if (reg->umin_value >= val)
11138 else if (reg->umax_value < val)
11142 if (reg->smin_value >= sval)
11144 else if (reg->smax_value < sval)
11148 if (reg->umax_value <= val)
11150 else if (reg->umin_value > val)
11154 if (reg->smax_value <= sval)
11156 else if (reg->smin_value > sval)
11164 /* compute branch direction of the expression "if (reg opcode val) goto target;"
11166 * 1 - branch will be taken and "goto target" will be executed
11167 * 0 - branch will not be taken and fall-through to next insn
11168 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
11171 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
11174 if (__is_pointer_value(false, reg)) {
11175 if (!reg_type_not_null(reg->type))
11178 /* If pointer is valid tests against zero will fail so we can
11179 * use this to direct branch taken.
11195 return is_branch32_taken(reg, val, opcode);
11196 return is_branch64_taken(reg, val, opcode);
11199 static int flip_opcode(u32 opcode)
11201 /* How can we transform "a <op> b" into "b <op> a"? */
11202 static const u8 opcode_flip[16] = {
11203 /* these stay the same */
11204 [BPF_JEQ >> 4] = BPF_JEQ,
11205 [BPF_JNE >> 4] = BPF_JNE,
11206 [BPF_JSET >> 4] = BPF_JSET,
11207 /* these swap "lesser" and "greater" (L and G in the opcodes) */
11208 [BPF_JGE >> 4] = BPF_JLE,
11209 [BPF_JGT >> 4] = BPF_JLT,
11210 [BPF_JLE >> 4] = BPF_JGE,
11211 [BPF_JLT >> 4] = BPF_JGT,
11212 [BPF_JSGE >> 4] = BPF_JSLE,
11213 [BPF_JSGT >> 4] = BPF_JSLT,
11214 [BPF_JSLE >> 4] = BPF_JSGE,
11215 [BPF_JSLT >> 4] = BPF_JSGT
11217 return opcode_flip[opcode >> 4];
11220 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
11221 struct bpf_reg_state *src_reg,
11224 struct bpf_reg_state *pkt;
11226 if (src_reg->type == PTR_TO_PACKET_END) {
11228 } else if (dst_reg->type == PTR_TO_PACKET_END) {
11230 opcode = flip_opcode(opcode);
11235 if (pkt->range >= 0)
11240 /* pkt <= pkt_end */
11243 /* pkt > pkt_end */
11244 if (pkt->range == BEYOND_PKT_END)
11245 /* pkt has at last one extra byte beyond pkt_end */
11246 return opcode == BPF_JGT;
11249 /* pkt < pkt_end */
11252 /* pkt >= pkt_end */
11253 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
11254 return opcode == BPF_JGE;
11260 /* Adjusts the register min/max values in the case that the dst_reg is the
11261 * variable register that we are working on, and src_reg is a constant or we're
11262 * simply doing a BPF_K check.
11263 * In JEQ/JNE cases we also adjust the var_off values.
11265 static void reg_set_min_max(struct bpf_reg_state *true_reg,
11266 struct bpf_reg_state *false_reg,
11267 u64 val, u32 val32,
11268 u8 opcode, bool is_jmp32)
11270 struct tnum false_32off = tnum_subreg(false_reg->var_off);
11271 struct tnum false_64off = false_reg->var_off;
11272 struct tnum true_32off = tnum_subreg(true_reg->var_off);
11273 struct tnum true_64off = true_reg->var_off;
11274 s64 sval = (s64)val;
11275 s32 sval32 = (s32)val32;
11277 /* If the dst_reg is a pointer, we can't learn anything about its
11278 * variable offset from the compare (unless src_reg were a pointer into
11279 * the same object, but we don't bother with that.
11280 * Since false_reg and true_reg have the same type by construction, we
11281 * only need to check one of them for pointerness.
11283 if (__is_pointer_value(false, false_reg))
11287 /* JEQ/JNE comparison doesn't change the register equivalence.
11290 * if (r1 == 42) goto label;
11292 * label: // here both r1 and r2 are known to be 42.
11294 * Hence when marking register as known preserve it's ID.
11298 __mark_reg32_known(true_reg, val32);
11299 true_32off = tnum_subreg(true_reg->var_off);
11301 ___mark_reg_known(true_reg, val);
11302 true_64off = true_reg->var_off;
11307 __mark_reg32_known(false_reg, val32);
11308 false_32off = tnum_subreg(false_reg->var_off);
11310 ___mark_reg_known(false_reg, val);
11311 false_64off = false_reg->var_off;
11316 false_32off = tnum_and(false_32off, tnum_const(~val32));
11317 if (is_power_of_2(val32))
11318 true_32off = tnum_or(true_32off,
11319 tnum_const(val32));
11321 false_64off = tnum_and(false_64off, tnum_const(~val));
11322 if (is_power_of_2(val))
11323 true_64off = tnum_or(true_64off,
11331 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
11332 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
11334 false_reg->u32_max_value = min(false_reg->u32_max_value,
11336 true_reg->u32_min_value = max(true_reg->u32_min_value,
11339 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
11340 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
11342 false_reg->umax_value = min(false_reg->umax_value, false_umax);
11343 true_reg->umin_value = max(true_reg->umin_value, true_umin);
11351 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
11352 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
11354 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
11355 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
11357 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
11358 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
11360 false_reg->smax_value = min(false_reg->smax_value, false_smax);
11361 true_reg->smin_value = max(true_reg->smin_value, true_smin);
11369 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
11370 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
11372 false_reg->u32_min_value = max(false_reg->u32_min_value,
11374 true_reg->u32_max_value = min(true_reg->u32_max_value,
11377 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
11378 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
11380 false_reg->umin_value = max(false_reg->umin_value, false_umin);
11381 true_reg->umax_value = min(true_reg->umax_value, true_umax);
11389 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
11390 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
11392 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
11393 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
11395 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
11396 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
11398 false_reg->smin_value = max(false_reg->smin_value, false_smin);
11399 true_reg->smax_value = min(true_reg->smax_value, true_smax);
11408 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
11409 tnum_subreg(false_32off));
11410 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
11411 tnum_subreg(true_32off));
11412 __reg_combine_32_into_64(false_reg);
11413 __reg_combine_32_into_64(true_reg);
11415 false_reg->var_off = false_64off;
11416 true_reg->var_off = true_64off;
11417 __reg_combine_64_into_32(false_reg);
11418 __reg_combine_64_into_32(true_reg);
11422 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
11423 * the variable reg.
11425 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
11426 struct bpf_reg_state *false_reg,
11427 u64 val, u32 val32,
11428 u8 opcode, bool is_jmp32)
11430 opcode = flip_opcode(opcode);
11431 /* This uses zero as "not present in table"; luckily the zero opcode,
11432 * BPF_JA, can't get here.
11435 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
11438 /* Regs are known to be equal, so intersect their min/max/var_off */
11439 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
11440 struct bpf_reg_state *dst_reg)
11442 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
11443 dst_reg->umin_value);
11444 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
11445 dst_reg->umax_value);
11446 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
11447 dst_reg->smin_value);
11448 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
11449 dst_reg->smax_value);
11450 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
11452 reg_bounds_sync(src_reg);
11453 reg_bounds_sync(dst_reg);
11456 static void reg_combine_min_max(struct bpf_reg_state *true_src,
11457 struct bpf_reg_state *true_dst,
11458 struct bpf_reg_state *false_src,
11459 struct bpf_reg_state *false_dst,
11464 __reg_combine_min_max(true_src, true_dst);
11467 __reg_combine_min_max(false_src, false_dst);
11472 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
11473 struct bpf_reg_state *reg, u32 id,
11476 if (type_may_be_null(reg->type) && reg->id == id &&
11477 (is_rcu_reg(reg) || !WARN_ON_ONCE(!reg->id))) {
11478 /* Old offset (both fixed and variable parts) should have been
11479 * known-zero, because we don't allow pointer arithmetic on
11480 * pointers that might be NULL. If we see this happening, don't
11481 * convert the register.
11483 * But in some cases, some helpers that return local kptrs
11484 * advance offset for the returned pointer. In those cases, it
11485 * is fine to expect to see reg->off.
11487 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value || !tnum_equals_const(reg->var_off, 0)))
11489 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC | PTR_MAYBE_NULL) && WARN_ON_ONCE(reg->off))
11492 reg->type = SCALAR_VALUE;
11493 /* We don't need id and ref_obj_id from this point
11494 * onwards anymore, thus we should better reset it,
11495 * so that state pruning has chances to take effect.
11498 reg->ref_obj_id = 0;
11503 mark_ptr_not_null_reg(reg);
11505 if (!reg_may_point_to_spin_lock(reg)) {
11506 /* For not-NULL ptr, reg->ref_obj_id will be reset
11507 * in release_reference().
11509 * reg->id is still used by spin_lock ptr. Other
11510 * than spin_lock ptr type, reg->id can be reset.
11517 /* The logic is similar to find_good_pkt_pointers(), both could eventually
11518 * be folded together at some point.
11520 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
11523 struct bpf_func_state *state = vstate->frame[vstate->curframe];
11524 struct bpf_reg_state *regs = state->regs, *reg;
11525 u32 ref_obj_id = regs[regno].ref_obj_id;
11526 u32 id = regs[regno].id;
11528 if (ref_obj_id && ref_obj_id == id && is_null)
11529 /* regs[regno] is in the " == NULL" branch.
11530 * No one could have freed the reference state before
11531 * doing the NULL check.
11533 WARN_ON_ONCE(release_reference_state(state, id));
11535 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
11536 mark_ptr_or_null_reg(state, reg, id, is_null);
11540 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
11541 struct bpf_reg_state *dst_reg,
11542 struct bpf_reg_state *src_reg,
11543 struct bpf_verifier_state *this_branch,
11544 struct bpf_verifier_state *other_branch)
11546 if (BPF_SRC(insn->code) != BPF_X)
11549 /* Pointers are always 64-bit. */
11550 if (BPF_CLASS(insn->code) == BPF_JMP32)
11553 switch (BPF_OP(insn->code)) {
11555 if ((dst_reg->type == PTR_TO_PACKET &&
11556 src_reg->type == PTR_TO_PACKET_END) ||
11557 (dst_reg->type == PTR_TO_PACKET_META &&
11558 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
11559 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
11560 find_good_pkt_pointers(this_branch, dst_reg,
11561 dst_reg->type, false);
11562 mark_pkt_end(other_branch, insn->dst_reg, true);
11563 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
11564 src_reg->type == PTR_TO_PACKET) ||
11565 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
11566 src_reg->type == PTR_TO_PACKET_META)) {
11567 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
11568 find_good_pkt_pointers(other_branch, src_reg,
11569 src_reg->type, true);
11570 mark_pkt_end(this_branch, insn->src_reg, false);
11576 if ((dst_reg->type == PTR_TO_PACKET &&
11577 src_reg->type == PTR_TO_PACKET_END) ||
11578 (dst_reg->type == PTR_TO_PACKET_META &&
11579 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
11580 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
11581 find_good_pkt_pointers(other_branch, dst_reg,
11582 dst_reg->type, true);
11583 mark_pkt_end(this_branch, insn->dst_reg, false);
11584 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
11585 src_reg->type == PTR_TO_PACKET) ||
11586 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
11587 src_reg->type == PTR_TO_PACKET_META)) {
11588 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
11589 find_good_pkt_pointers(this_branch, src_reg,
11590 src_reg->type, false);
11591 mark_pkt_end(other_branch, insn->src_reg, true);
11597 if ((dst_reg->type == PTR_TO_PACKET &&
11598 src_reg->type == PTR_TO_PACKET_END) ||
11599 (dst_reg->type == PTR_TO_PACKET_META &&
11600 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
11601 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
11602 find_good_pkt_pointers(this_branch, dst_reg,
11603 dst_reg->type, true);
11604 mark_pkt_end(other_branch, insn->dst_reg, false);
11605 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
11606 src_reg->type == PTR_TO_PACKET) ||
11607 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
11608 src_reg->type == PTR_TO_PACKET_META)) {
11609 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
11610 find_good_pkt_pointers(other_branch, src_reg,
11611 src_reg->type, false);
11612 mark_pkt_end(this_branch, insn->src_reg, true);
11618 if ((dst_reg->type == PTR_TO_PACKET &&
11619 src_reg->type == PTR_TO_PACKET_END) ||
11620 (dst_reg->type == PTR_TO_PACKET_META &&
11621 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
11622 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
11623 find_good_pkt_pointers(other_branch, dst_reg,
11624 dst_reg->type, false);
11625 mark_pkt_end(this_branch, insn->dst_reg, true);
11626 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
11627 src_reg->type == PTR_TO_PACKET) ||
11628 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
11629 src_reg->type == PTR_TO_PACKET_META)) {
11630 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
11631 find_good_pkt_pointers(this_branch, src_reg,
11632 src_reg->type, true);
11633 mark_pkt_end(other_branch, insn->src_reg, false);
11645 static void find_equal_scalars(struct bpf_verifier_state *vstate,
11646 struct bpf_reg_state *known_reg)
11648 struct bpf_func_state *state;
11649 struct bpf_reg_state *reg;
11651 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
11652 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
11657 static int check_cond_jmp_op(struct bpf_verifier_env *env,
11658 struct bpf_insn *insn, int *insn_idx)
11660 struct bpf_verifier_state *this_branch = env->cur_state;
11661 struct bpf_verifier_state *other_branch;
11662 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
11663 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
11664 struct bpf_reg_state *eq_branch_regs;
11665 u8 opcode = BPF_OP(insn->code);
11670 /* Only conditional jumps are expected to reach here. */
11671 if (opcode == BPF_JA || opcode > BPF_JSLE) {
11672 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
11676 if (BPF_SRC(insn->code) == BPF_X) {
11677 if (insn->imm != 0) {
11678 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
11682 /* check src1 operand */
11683 err = check_reg_arg(env, insn->src_reg, SRC_OP);
11687 if (is_pointer_value(env, insn->src_reg)) {
11688 verbose(env, "R%d pointer comparison prohibited\n",
11692 src_reg = ®s[insn->src_reg];
11694 if (insn->src_reg != BPF_REG_0) {
11695 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
11700 /* check src2 operand */
11701 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
11705 dst_reg = ®s[insn->dst_reg];
11706 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
11708 if (BPF_SRC(insn->code) == BPF_K) {
11709 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
11710 } else if (src_reg->type == SCALAR_VALUE &&
11711 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
11712 pred = is_branch_taken(dst_reg,
11713 tnum_subreg(src_reg->var_off).value,
11716 } else if (src_reg->type == SCALAR_VALUE &&
11717 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
11718 pred = is_branch_taken(dst_reg,
11719 src_reg->var_off.value,
11722 } else if (reg_is_pkt_pointer_any(dst_reg) &&
11723 reg_is_pkt_pointer_any(src_reg) &&
11725 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
11729 /* If we get here with a dst_reg pointer type it is because
11730 * above is_branch_taken() special cased the 0 comparison.
11732 if (!__is_pointer_value(false, dst_reg))
11733 err = mark_chain_precision(env, insn->dst_reg);
11734 if (BPF_SRC(insn->code) == BPF_X && !err &&
11735 !__is_pointer_value(false, src_reg))
11736 err = mark_chain_precision(env, insn->src_reg);
11742 /* Only follow the goto, ignore fall-through. If needed, push
11743 * the fall-through branch for simulation under speculative
11746 if (!env->bypass_spec_v1 &&
11747 !sanitize_speculative_path(env, insn, *insn_idx + 1,
11750 *insn_idx += insn->off;
11752 } else if (pred == 0) {
11753 /* Only follow the fall-through branch, since that's where the
11754 * program will go. If needed, push the goto branch for
11755 * simulation under speculative execution.
11757 if (!env->bypass_spec_v1 &&
11758 !sanitize_speculative_path(env, insn,
11759 *insn_idx + insn->off + 1,
11765 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
11769 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
11771 /* detect if we are comparing against a constant value so we can adjust
11772 * our min/max values for our dst register.
11773 * this is only legit if both are scalars (or pointers to the same
11774 * object, I suppose, see the PTR_MAYBE_NULL related if block below),
11775 * because otherwise the different base pointers mean the offsets aren't
11778 if (BPF_SRC(insn->code) == BPF_X) {
11779 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
11781 if (dst_reg->type == SCALAR_VALUE &&
11782 src_reg->type == SCALAR_VALUE) {
11783 if (tnum_is_const(src_reg->var_off) ||
11785 tnum_is_const(tnum_subreg(src_reg->var_off))))
11786 reg_set_min_max(&other_branch_regs[insn->dst_reg],
11788 src_reg->var_off.value,
11789 tnum_subreg(src_reg->var_off).value,
11791 else if (tnum_is_const(dst_reg->var_off) ||
11793 tnum_is_const(tnum_subreg(dst_reg->var_off))))
11794 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
11796 dst_reg->var_off.value,
11797 tnum_subreg(dst_reg->var_off).value,
11799 else if (!is_jmp32 &&
11800 (opcode == BPF_JEQ || opcode == BPF_JNE))
11801 /* Comparing for equality, we can combine knowledge */
11802 reg_combine_min_max(&other_branch_regs[insn->src_reg],
11803 &other_branch_regs[insn->dst_reg],
11804 src_reg, dst_reg, opcode);
11806 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
11807 find_equal_scalars(this_branch, src_reg);
11808 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
11812 } else if (dst_reg->type == SCALAR_VALUE) {
11813 reg_set_min_max(&other_branch_regs[insn->dst_reg],
11814 dst_reg, insn->imm, (u32)insn->imm,
11818 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
11819 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
11820 find_equal_scalars(this_branch, dst_reg);
11821 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
11824 /* if one pointer register is compared to another pointer
11825 * register check if PTR_MAYBE_NULL could be lifted.
11826 * E.g. register A - maybe null
11827 * register B - not null
11828 * for JNE A, B, ... - A is not null in the false branch;
11829 * for JEQ A, B, ... - A is not null in the true branch.
11831 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_X &&
11832 __is_pointer_value(false, src_reg) && __is_pointer_value(false, dst_reg) &&
11833 type_may_be_null(src_reg->type) != type_may_be_null(dst_reg->type)) {
11834 eq_branch_regs = NULL;
11837 eq_branch_regs = other_branch_regs;
11840 eq_branch_regs = regs;
11846 if (eq_branch_regs) {
11847 if (type_may_be_null(src_reg->type))
11848 mark_ptr_not_null_reg(&eq_branch_regs[insn->src_reg]);
11850 mark_ptr_not_null_reg(&eq_branch_regs[insn->dst_reg]);
11854 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
11855 * NOTE: these optimizations below are related with pointer comparison
11856 * which will never be JMP32.
11858 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
11859 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
11860 type_may_be_null(dst_reg->type)) {
11861 /* Mark all identical registers in each branch as either
11862 * safe or unknown depending R == 0 or R != 0 conditional.
11864 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
11865 opcode == BPF_JNE);
11866 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
11867 opcode == BPF_JEQ);
11868 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
11869 this_branch, other_branch) &&
11870 is_pointer_value(env, insn->dst_reg)) {
11871 verbose(env, "R%d pointer comparison prohibited\n",
11875 if (env->log.level & BPF_LOG_LEVEL)
11876 print_insn_state(env, this_branch->frame[this_branch->curframe]);
11880 /* verify BPF_LD_IMM64 instruction */
11881 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
11883 struct bpf_insn_aux_data *aux = cur_aux(env);
11884 struct bpf_reg_state *regs = cur_regs(env);
11885 struct bpf_reg_state *dst_reg;
11886 struct bpf_map *map;
11889 if (BPF_SIZE(insn->code) != BPF_DW) {
11890 verbose(env, "invalid BPF_LD_IMM insn\n");
11893 if (insn->off != 0) {
11894 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
11898 err = check_reg_arg(env, insn->dst_reg, DST_OP);
11902 dst_reg = ®s[insn->dst_reg];
11903 if (insn->src_reg == 0) {
11904 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
11906 dst_reg->type = SCALAR_VALUE;
11907 __mark_reg_known(®s[insn->dst_reg], imm);
11911 /* All special src_reg cases are listed below. From this point onwards
11912 * we either succeed and assign a corresponding dst_reg->type after
11913 * zeroing the offset, or fail and reject the program.
11915 mark_reg_known_zero(env, regs, insn->dst_reg);
11917 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
11918 dst_reg->type = aux->btf_var.reg_type;
11919 switch (base_type(dst_reg->type)) {
11921 dst_reg->mem_size = aux->btf_var.mem_size;
11923 case PTR_TO_BTF_ID:
11924 dst_reg->btf = aux->btf_var.btf;
11925 dst_reg->btf_id = aux->btf_var.btf_id;
11928 verbose(env, "bpf verifier is misconfigured\n");
11934 if (insn->src_reg == BPF_PSEUDO_FUNC) {
11935 struct bpf_prog_aux *aux = env->prog->aux;
11936 u32 subprogno = find_subprog(env,
11937 env->insn_idx + insn->imm + 1);
11939 if (!aux->func_info) {
11940 verbose(env, "missing btf func_info\n");
11943 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
11944 verbose(env, "callback function not static\n");
11948 dst_reg->type = PTR_TO_FUNC;
11949 dst_reg->subprogno = subprogno;
11953 map = env->used_maps[aux->map_index];
11954 dst_reg->map_ptr = map;
11956 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
11957 insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
11958 dst_reg->type = PTR_TO_MAP_VALUE;
11959 dst_reg->off = aux->map_off;
11960 WARN_ON_ONCE(map->max_entries != 1);
11961 /* We want reg->id to be same (0) as map_value is not distinct */
11962 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
11963 insn->src_reg == BPF_PSEUDO_MAP_IDX) {
11964 dst_reg->type = CONST_PTR_TO_MAP;
11966 verbose(env, "bpf verifier is misconfigured\n");
11973 static bool may_access_skb(enum bpf_prog_type type)
11976 case BPF_PROG_TYPE_SOCKET_FILTER:
11977 case BPF_PROG_TYPE_SCHED_CLS:
11978 case BPF_PROG_TYPE_SCHED_ACT:
11985 /* verify safety of LD_ABS|LD_IND instructions:
11986 * - they can only appear in the programs where ctx == skb
11987 * - since they are wrappers of function calls, they scratch R1-R5 registers,
11988 * preserve R6-R9, and store return value into R0
11991 * ctx == skb == R6 == CTX
11994 * SRC == any register
11995 * IMM == 32-bit immediate
11998 * R0 - 8/16/32-bit skb data converted to cpu endianness
12000 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
12002 struct bpf_reg_state *regs = cur_regs(env);
12003 static const int ctx_reg = BPF_REG_6;
12004 u8 mode = BPF_MODE(insn->code);
12007 if (!may_access_skb(resolve_prog_type(env->prog))) {
12008 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
12012 if (!env->ops->gen_ld_abs) {
12013 verbose(env, "bpf verifier is misconfigured\n");
12017 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
12018 BPF_SIZE(insn->code) == BPF_DW ||
12019 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
12020 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
12024 /* check whether implicit source operand (register R6) is readable */
12025 err = check_reg_arg(env, ctx_reg, SRC_OP);
12029 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
12030 * gen_ld_abs() may terminate the program at runtime, leading to
12033 err = check_reference_leak(env);
12035 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
12039 if (env->cur_state->active_lock.ptr) {
12040 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
12044 if (env->cur_state->active_rcu_lock) {
12045 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_rcu_read_lock-ed region\n");
12049 if (regs[ctx_reg].type != PTR_TO_CTX) {
12051 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
12055 if (mode == BPF_IND) {
12056 /* check explicit source operand */
12057 err = check_reg_arg(env, insn->src_reg, SRC_OP);
12062 err = check_ptr_off_reg(env, ®s[ctx_reg], ctx_reg);
12066 /* reset caller saved regs to unreadable */
12067 for (i = 0; i < CALLER_SAVED_REGS; i++) {
12068 mark_reg_not_init(env, regs, caller_saved[i]);
12069 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
12072 /* mark destination R0 register as readable, since it contains
12073 * the value fetched from the packet.
12074 * Already marked as written above.
12076 mark_reg_unknown(env, regs, BPF_REG_0);
12077 /* ld_abs load up to 32-bit skb data. */
12078 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
12082 static int check_return_code(struct bpf_verifier_env *env)
12084 struct tnum enforce_attach_type_range = tnum_unknown;
12085 const struct bpf_prog *prog = env->prog;
12086 struct bpf_reg_state *reg;
12087 struct tnum range = tnum_range(0, 1);
12088 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
12090 struct bpf_func_state *frame = env->cur_state->frame[0];
12091 const bool is_subprog = frame->subprogno;
12093 /* LSM and struct_ops func-ptr's return type could be "void" */
12095 switch (prog_type) {
12096 case BPF_PROG_TYPE_LSM:
12097 if (prog->expected_attach_type == BPF_LSM_CGROUP)
12098 /* See below, can be 0 or 0-1 depending on hook. */
12101 case BPF_PROG_TYPE_STRUCT_OPS:
12102 if (!prog->aux->attach_func_proto->type)
12110 /* eBPF calling convention is such that R0 is used
12111 * to return the value from eBPF program.
12112 * Make sure that it's readable at this time
12113 * of bpf_exit, which means that program wrote
12114 * something into it earlier
12116 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
12120 if (is_pointer_value(env, BPF_REG_0)) {
12121 verbose(env, "R0 leaks addr as return value\n");
12125 reg = cur_regs(env) + BPF_REG_0;
12127 if (frame->in_async_callback_fn) {
12128 /* enforce return zero from async callbacks like timer */
12129 if (reg->type != SCALAR_VALUE) {
12130 verbose(env, "In async callback the register R0 is not a known value (%s)\n",
12131 reg_type_str(env, reg->type));
12135 if (!tnum_in(tnum_const(0), reg->var_off)) {
12136 verbose_invalid_scalar(env, reg, &range, "async callback", "R0");
12143 if (reg->type != SCALAR_VALUE) {
12144 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
12145 reg_type_str(env, reg->type));
12151 switch (prog_type) {
12152 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
12153 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
12154 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
12155 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
12156 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
12157 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
12158 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
12159 range = tnum_range(1, 1);
12160 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
12161 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
12162 range = tnum_range(0, 3);
12164 case BPF_PROG_TYPE_CGROUP_SKB:
12165 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
12166 range = tnum_range(0, 3);
12167 enforce_attach_type_range = tnum_range(2, 3);
12170 case BPF_PROG_TYPE_CGROUP_SOCK:
12171 case BPF_PROG_TYPE_SOCK_OPS:
12172 case BPF_PROG_TYPE_CGROUP_DEVICE:
12173 case BPF_PROG_TYPE_CGROUP_SYSCTL:
12174 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
12176 case BPF_PROG_TYPE_RAW_TRACEPOINT:
12177 if (!env->prog->aux->attach_btf_id)
12179 range = tnum_const(0);
12181 case BPF_PROG_TYPE_TRACING:
12182 switch (env->prog->expected_attach_type) {
12183 case BPF_TRACE_FENTRY:
12184 case BPF_TRACE_FEXIT:
12185 range = tnum_const(0);
12187 case BPF_TRACE_RAW_TP:
12188 case BPF_MODIFY_RETURN:
12190 case BPF_TRACE_ITER:
12196 case BPF_PROG_TYPE_SK_LOOKUP:
12197 range = tnum_range(SK_DROP, SK_PASS);
12200 case BPF_PROG_TYPE_LSM:
12201 if (env->prog->expected_attach_type != BPF_LSM_CGROUP) {
12202 /* Regular BPF_PROG_TYPE_LSM programs can return
12207 if (!env->prog->aux->attach_func_proto->type) {
12208 /* Make sure programs that attach to void
12209 * hooks don't try to modify return value.
12211 range = tnum_range(1, 1);
12215 case BPF_PROG_TYPE_EXT:
12216 /* freplace program can return anything as its return value
12217 * depends on the to-be-replaced kernel func or bpf program.
12223 if (reg->type != SCALAR_VALUE) {
12224 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
12225 reg_type_str(env, reg->type));
12229 if (!tnum_in(range, reg->var_off)) {
12230 verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
12231 if (prog->expected_attach_type == BPF_LSM_CGROUP &&
12232 prog_type == BPF_PROG_TYPE_LSM &&
12233 !prog->aux->attach_func_proto->type)
12234 verbose(env, "Note, BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
12238 if (!tnum_is_unknown(enforce_attach_type_range) &&
12239 tnum_in(enforce_attach_type_range, reg->var_off))
12240 env->prog->enforce_expected_attach_type = 1;
12244 /* non-recursive DFS pseudo code
12245 * 1 procedure DFS-iterative(G,v):
12246 * 2 label v as discovered
12247 * 3 let S be a stack
12249 * 5 while S is not empty
12251 * 7 if t is what we're looking for:
12253 * 9 for all edges e in G.adjacentEdges(t) do
12254 * 10 if edge e is already labelled
12255 * 11 continue with the next edge
12256 * 12 w <- G.adjacentVertex(t,e)
12257 * 13 if vertex w is not discovered and not explored
12258 * 14 label e as tree-edge
12259 * 15 label w as discovered
12262 * 18 else if vertex w is discovered
12263 * 19 label e as back-edge
12265 * 21 // vertex w is explored
12266 * 22 label e as forward- or cross-edge
12267 * 23 label t as explored
12271 * 0x10 - discovered
12272 * 0x11 - discovered and fall-through edge labelled
12273 * 0x12 - discovered and fall-through and branch edges labelled
12284 static u32 state_htab_size(struct bpf_verifier_env *env)
12286 return env->prog->len;
12289 static struct bpf_verifier_state_list **explored_state(
12290 struct bpf_verifier_env *env,
12293 struct bpf_verifier_state *cur = env->cur_state;
12294 struct bpf_func_state *state = cur->frame[cur->curframe];
12296 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
12299 static void mark_prune_point(struct bpf_verifier_env *env, int idx)
12301 env->insn_aux_data[idx].prune_point = true;
12304 static bool is_prune_point(struct bpf_verifier_env *env, int insn_idx)
12306 return env->insn_aux_data[insn_idx].prune_point;
12310 DONE_EXPLORING = 0,
12311 KEEP_EXPLORING = 1,
12314 /* t, w, e - match pseudo-code above:
12315 * t - index of current instruction
12316 * w - next instruction
12319 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
12322 int *insn_stack = env->cfg.insn_stack;
12323 int *insn_state = env->cfg.insn_state;
12325 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
12326 return DONE_EXPLORING;
12328 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
12329 return DONE_EXPLORING;
12331 if (w < 0 || w >= env->prog->len) {
12332 verbose_linfo(env, t, "%d: ", t);
12333 verbose(env, "jump out of range from insn %d to %d\n", t, w);
12338 /* mark branch target for state pruning */
12339 mark_prune_point(env, w);
12340 mark_jmp_point(env, w);
12343 if (insn_state[w] == 0) {
12345 insn_state[t] = DISCOVERED | e;
12346 insn_state[w] = DISCOVERED;
12347 if (env->cfg.cur_stack >= env->prog->len)
12349 insn_stack[env->cfg.cur_stack++] = w;
12350 return KEEP_EXPLORING;
12351 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
12352 if (loop_ok && env->bpf_capable)
12353 return DONE_EXPLORING;
12354 verbose_linfo(env, t, "%d: ", t);
12355 verbose_linfo(env, w, "%d: ", w);
12356 verbose(env, "back-edge from insn %d to %d\n", t, w);
12358 } else if (insn_state[w] == EXPLORED) {
12359 /* forward- or cross-edge */
12360 insn_state[t] = DISCOVERED | e;
12362 verbose(env, "insn state internal bug\n");
12365 return DONE_EXPLORING;
12368 static int visit_func_call_insn(int t, struct bpf_insn *insns,
12369 struct bpf_verifier_env *env,
12374 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
12378 mark_prune_point(env, t + 1);
12379 /* when we exit from subprog, we need to record non-linear history */
12380 mark_jmp_point(env, t + 1);
12382 if (visit_callee) {
12383 mark_prune_point(env, t);
12384 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env,
12385 /* It's ok to allow recursion from CFG point of
12386 * view. __check_func_call() will do the actual
12389 bpf_pseudo_func(insns + t));
12394 /* Visits the instruction at index t and returns one of the following:
12395 * < 0 - an error occurred
12396 * DONE_EXPLORING - the instruction was fully explored
12397 * KEEP_EXPLORING - there is still work to be done before it is fully explored
12399 static int visit_insn(int t, struct bpf_verifier_env *env)
12401 struct bpf_insn *insns = env->prog->insnsi;
12404 if (bpf_pseudo_func(insns + t))
12405 return visit_func_call_insn(t, insns, env, true);
12407 /* All non-branch instructions have a single fall-through edge. */
12408 if (BPF_CLASS(insns[t].code) != BPF_JMP &&
12409 BPF_CLASS(insns[t].code) != BPF_JMP32)
12410 return push_insn(t, t + 1, FALLTHROUGH, env, false);
12412 switch (BPF_OP(insns[t].code)) {
12414 return DONE_EXPLORING;
12417 if (insns[t].imm == BPF_FUNC_timer_set_callback)
12418 /* Mark this call insn as a prune point to trigger
12419 * is_state_visited() check before call itself is
12420 * processed by __check_func_call(). Otherwise new
12421 * async state will be pushed for further exploration.
12423 mark_prune_point(env, t);
12424 return visit_func_call_insn(t, insns, env,
12425 insns[t].src_reg == BPF_PSEUDO_CALL);
12428 if (BPF_SRC(insns[t].code) != BPF_K)
12431 /* unconditional jump with single edge */
12432 ret = push_insn(t, t + insns[t].off + 1, FALLTHROUGH, env,
12437 mark_prune_point(env, t + insns[t].off + 1);
12438 mark_jmp_point(env, t + insns[t].off + 1);
12443 /* conditional jump with two edges */
12444 mark_prune_point(env, t);
12446 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
12450 return push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
12454 /* non-recursive depth-first-search to detect loops in BPF program
12455 * loop == back-edge in directed graph
12457 static int check_cfg(struct bpf_verifier_env *env)
12459 int insn_cnt = env->prog->len;
12460 int *insn_stack, *insn_state;
12464 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
12468 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
12470 kvfree(insn_state);
12474 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
12475 insn_stack[0] = 0; /* 0 is the first instruction */
12476 env->cfg.cur_stack = 1;
12478 while (env->cfg.cur_stack > 0) {
12479 int t = insn_stack[env->cfg.cur_stack - 1];
12481 ret = visit_insn(t, env);
12483 case DONE_EXPLORING:
12484 insn_state[t] = EXPLORED;
12485 env->cfg.cur_stack--;
12487 case KEEP_EXPLORING:
12491 verbose(env, "visit_insn internal bug\n");
12498 if (env->cfg.cur_stack < 0) {
12499 verbose(env, "pop stack internal bug\n");
12504 for (i = 0; i < insn_cnt; i++) {
12505 if (insn_state[i] != EXPLORED) {
12506 verbose(env, "unreachable insn %d\n", i);
12511 ret = 0; /* cfg looks good */
12514 kvfree(insn_state);
12515 kvfree(insn_stack);
12516 env->cfg.insn_state = env->cfg.insn_stack = NULL;
12520 static int check_abnormal_return(struct bpf_verifier_env *env)
12524 for (i = 1; i < env->subprog_cnt; i++) {
12525 if (env->subprog_info[i].has_ld_abs) {
12526 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
12529 if (env->subprog_info[i].has_tail_call) {
12530 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
12537 /* The minimum supported BTF func info size */
12538 #define MIN_BPF_FUNCINFO_SIZE 8
12539 #define MAX_FUNCINFO_REC_SIZE 252
12541 static int check_btf_func(struct bpf_verifier_env *env,
12542 const union bpf_attr *attr,
12545 const struct btf_type *type, *func_proto, *ret_type;
12546 u32 i, nfuncs, urec_size, min_size;
12547 u32 krec_size = sizeof(struct bpf_func_info);
12548 struct bpf_func_info *krecord;
12549 struct bpf_func_info_aux *info_aux = NULL;
12550 struct bpf_prog *prog;
12551 const struct btf *btf;
12553 u32 prev_offset = 0;
12554 bool scalar_return;
12557 nfuncs = attr->func_info_cnt;
12559 if (check_abnormal_return(env))
12564 if (nfuncs != env->subprog_cnt) {
12565 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
12569 urec_size = attr->func_info_rec_size;
12570 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
12571 urec_size > MAX_FUNCINFO_REC_SIZE ||
12572 urec_size % sizeof(u32)) {
12573 verbose(env, "invalid func info rec size %u\n", urec_size);
12578 btf = prog->aux->btf;
12580 urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
12581 min_size = min_t(u32, krec_size, urec_size);
12583 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
12586 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
12590 for (i = 0; i < nfuncs; i++) {
12591 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
12593 if (ret == -E2BIG) {
12594 verbose(env, "nonzero tailing record in func info");
12595 /* set the size kernel expects so loader can zero
12596 * out the rest of the record.
12598 if (copy_to_bpfptr_offset(uattr,
12599 offsetof(union bpf_attr, func_info_rec_size),
12600 &min_size, sizeof(min_size)))
12606 if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
12611 /* check insn_off */
12614 if (krecord[i].insn_off) {
12616 "nonzero insn_off %u for the first func info record",
12617 krecord[i].insn_off);
12620 } else if (krecord[i].insn_off <= prev_offset) {
12622 "same or smaller insn offset (%u) than previous func info record (%u)",
12623 krecord[i].insn_off, prev_offset);
12627 if (env->subprog_info[i].start != krecord[i].insn_off) {
12628 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
12632 /* check type_id */
12633 type = btf_type_by_id(btf, krecord[i].type_id);
12634 if (!type || !btf_type_is_func(type)) {
12635 verbose(env, "invalid type id %d in func info",
12636 krecord[i].type_id);
12639 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
12641 func_proto = btf_type_by_id(btf, type->type);
12642 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
12643 /* btf_func_check() already verified it during BTF load */
12645 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
12647 btf_type_is_small_int(ret_type) || btf_is_any_enum(ret_type);
12648 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
12649 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
12652 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
12653 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
12657 prev_offset = krecord[i].insn_off;
12658 bpfptr_add(&urecord, urec_size);
12661 prog->aux->func_info = krecord;
12662 prog->aux->func_info_cnt = nfuncs;
12663 prog->aux->func_info_aux = info_aux;
12672 static void adjust_btf_func(struct bpf_verifier_env *env)
12674 struct bpf_prog_aux *aux = env->prog->aux;
12677 if (!aux->func_info)
12680 for (i = 0; i < env->subprog_cnt; i++)
12681 aux->func_info[i].insn_off = env->subprog_info[i].start;
12684 #define MIN_BPF_LINEINFO_SIZE offsetofend(struct bpf_line_info, line_col)
12685 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
12687 static int check_btf_line(struct bpf_verifier_env *env,
12688 const union bpf_attr *attr,
12691 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
12692 struct bpf_subprog_info *sub;
12693 struct bpf_line_info *linfo;
12694 struct bpf_prog *prog;
12695 const struct btf *btf;
12699 nr_linfo = attr->line_info_cnt;
12702 if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info))
12705 rec_size = attr->line_info_rec_size;
12706 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
12707 rec_size > MAX_LINEINFO_REC_SIZE ||
12708 rec_size & (sizeof(u32) - 1))
12711 /* Need to zero it in case the userspace may
12712 * pass in a smaller bpf_line_info object.
12714 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
12715 GFP_KERNEL | __GFP_NOWARN);
12720 btf = prog->aux->btf;
12723 sub = env->subprog_info;
12724 ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
12725 expected_size = sizeof(struct bpf_line_info);
12726 ncopy = min_t(u32, expected_size, rec_size);
12727 for (i = 0; i < nr_linfo; i++) {
12728 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
12730 if (err == -E2BIG) {
12731 verbose(env, "nonzero tailing record in line_info");
12732 if (copy_to_bpfptr_offset(uattr,
12733 offsetof(union bpf_attr, line_info_rec_size),
12734 &expected_size, sizeof(expected_size)))
12740 if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
12746 * Check insn_off to ensure
12747 * 1) strictly increasing AND
12748 * 2) bounded by prog->len
12750 * The linfo[0].insn_off == 0 check logically falls into
12751 * the later "missing bpf_line_info for func..." case
12752 * because the first linfo[0].insn_off must be the
12753 * first sub also and the first sub must have
12754 * subprog_info[0].start == 0.
12756 if ((i && linfo[i].insn_off <= prev_offset) ||
12757 linfo[i].insn_off >= prog->len) {
12758 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
12759 i, linfo[i].insn_off, prev_offset,
12765 if (!prog->insnsi[linfo[i].insn_off].code) {
12767 "Invalid insn code at line_info[%u].insn_off\n",
12773 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
12774 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
12775 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
12780 if (s != env->subprog_cnt) {
12781 if (linfo[i].insn_off == sub[s].start) {
12782 sub[s].linfo_idx = i;
12784 } else if (sub[s].start < linfo[i].insn_off) {
12785 verbose(env, "missing bpf_line_info for func#%u\n", s);
12791 prev_offset = linfo[i].insn_off;
12792 bpfptr_add(&ulinfo, rec_size);
12795 if (s != env->subprog_cnt) {
12796 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
12797 env->subprog_cnt - s, s);
12802 prog->aux->linfo = linfo;
12803 prog->aux->nr_linfo = nr_linfo;
12812 #define MIN_CORE_RELO_SIZE sizeof(struct bpf_core_relo)
12813 #define MAX_CORE_RELO_SIZE MAX_FUNCINFO_REC_SIZE
12815 static int check_core_relo(struct bpf_verifier_env *env,
12816 const union bpf_attr *attr,
12819 u32 i, nr_core_relo, ncopy, expected_size, rec_size;
12820 struct bpf_core_relo core_relo = {};
12821 struct bpf_prog *prog = env->prog;
12822 const struct btf *btf = prog->aux->btf;
12823 struct bpf_core_ctx ctx = {
12827 bpfptr_t u_core_relo;
12830 nr_core_relo = attr->core_relo_cnt;
12833 if (nr_core_relo > INT_MAX / sizeof(struct bpf_core_relo))
12836 rec_size = attr->core_relo_rec_size;
12837 if (rec_size < MIN_CORE_RELO_SIZE ||
12838 rec_size > MAX_CORE_RELO_SIZE ||
12839 rec_size % sizeof(u32))
12842 u_core_relo = make_bpfptr(attr->core_relos, uattr.is_kernel);
12843 expected_size = sizeof(struct bpf_core_relo);
12844 ncopy = min_t(u32, expected_size, rec_size);
12846 /* Unlike func_info and line_info, copy and apply each CO-RE
12847 * relocation record one at a time.
12849 for (i = 0; i < nr_core_relo; i++) {
12850 /* future proofing when sizeof(bpf_core_relo) changes */
12851 err = bpf_check_uarg_tail_zero(u_core_relo, expected_size, rec_size);
12853 if (err == -E2BIG) {
12854 verbose(env, "nonzero tailing record in core_relo");
12855 if (copy_to_bpfptr_offset(uattr,
12856 offsetof(union bpf_attr, core_relo_rec_size),
12857 &expected_size, sizeof(expected_size)))
12863 if (copy_from_bpfptr(&core_relo, u_core_relo, ncopy)) {
12868 if (core_relo.insn_off % 8 || core_relo.insn_off / 8 >= prog->len) {
12869 verbose(env, "Invalid core_relo[%u].insn_off:%u prog->len:%u\n",
12870 i, core_relo.insn_off, prog->len);
12875 err = bpf_core_apply(&ctx, &core_relo, i,
12876 &prog->insnsi[core_relo.insn_off / 8]);
12879 bpfptr_add(&u_core_relo, rec_size);
12884 static int check_btf_info(struct bpf_verifier_env *env,
12885 const union bpf_attr *attr,
12891 if (!attr->func_info_cnt && !attr->line_info_cnt) {
12892 if (check_abnormal_return(env))
12897 btf = btf_get_by_fd(attr->prog_btf_fd);
12899 return PTR_ERR(btf);
12900 if (btf_is_kernel(btf)) {
12904 env->prog->aux->btf = btf;
12906 err = check_btf_func(env, attr, uattr);
12910 err = check_btf_line(env, attr, uattr);
12914 err = check_core_relo(env, attr, uattr);
12921 /* check %cur's range satisfies %old's */
12922 static bool range_within(struct bpf_reg_state *old,
12923 struct bpf_reg_state *cur)
12925 return old->umin_value <= cur->umin_value &&
12926 old->umax_value >= cur->umax_value &&
12927 old->smin_value <= cur->smin_value &&
12928 old->smax_value >= cur->smax_value &&
12929 old->u32_min_value <= cur->u32_min_value &&
12930 old->u32_max_value >= cur->u32_max_value &&
12931 old->s32_min_value <= cur->s32_min_value &&
12932 old->s32_max_value >= cur->s32_max_value;
12935 /* If in the old state two registers had the same id, then they need to have
12936 * the same id in the new state as well. But that id could be different from
12937 * the old state, so we need to track the mapping from old to new ids.
12938 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
12939 * regs with old id 5 must also have new id 9 for the new state to be safe. But
12940 * regs with a different old id could still have new id 9, we don't care about
12942 * So we look through our idmap to see if this old id has been seen before. If
12943 * so, we require the new id to match; otherwise, we add the id pair to the map.
12945 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_id_pair *idmap)
12949 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
12950 if (!idmap[i].old) {
12951 /* Reached an empty slot; haven't seen this id before */
12952 idmap[i].old = old_id;
12953 idmap[i].cur = cur_id;
12956 if (idmap[i].old == old_id)
12957 return idmap[i].cur == cur_id;
12959 /* We ran out of idmap slots, which should be impossible */
12964 static void clean_func_state(struct bpf_verifier_env *env,
12965 struct bpf_func_state *st)
12967 enum bpf_reg_liveness live;
12970 for (i = 0; i < BPF_REG_FP; i++) {
12971 live = st->regs[i].live;
12972 /* liveness must not touch this register anymore */
12973 st->regs[i].live |= REG_LIVE_DONE;
12974 if (!(live & REG_LIVE_READ))
12975 /* since the register is unused, clear its state
12976 * to make further comparison simpler
12978 __mark_reg_not_init(env, &st->regs[i]);
12981 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
12982 live = st->stack[i].spilled_ptr.live;
12983 /* liveness must not touch this stack slot anymore */
12984 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
12985 if (!(live & REG_LIVE_READ)) {
12986 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
12987 for (j = 0; j < BPF_REG_SIZE; j++)
12988 st->stack[i].slot_type[j] = STACK_INVALID;
12993 static void clean_verifier_state(struct bpf_verifier_env *env,
12994 struct bpf_verifier_state *st)
12998 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
12999 /* all regs in this state in all frames were already marked */
13002 for (i = 0; i <= st->curframe; i++)
13003 clean_func_state(env, st->frame[i]);
13006 /* the parentage chains form a tree.
13007 * the verifier states are added to state lists at given insn and
13008 * pushed into state stack for future exploration.
13009 * when the verifier reaches bpf_exit insn some of the verifer states
13010 * stored in the state lists have their final liveness state already,
13011 * but a lot of states will get revised from liveness point of view when
13012 * the verifier explores other branches.
13015 * 2: if r1 == 100 goto pc+1
13018 * when the verifier reaches exit insn the register r0 in the state list of
13019 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
13020 * of insn 2 and goes exploring further. At the insn 4 it will walk the
13021 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
13023 * Since the verifier pushes the branch states as it sees them while exploring
13024 * the program the condition of walking the branch instruction for the second
13025 * time means that all states below this branch were already explored and
13026 * their final liveness marks are already propagated.
13027 * Hence when the verifier completes the search of state list in is_state_visited()
13028 * we can call this clean_live_states() function to mark all liveness states
13029 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
13030 * will not be used.
13031 * This function also clears the registers and stack for states that !READ
13032 * to simplify state merging.
13034 * Important note here that walking the same branch instruction in the callee
13035 * doesn't meant that the states are DONE. The verifier has to compare
13038 static void clean_live_states(struct bpf_verifier_env *env, int insn,
13039 struct bpf_verifier_state *cur)
13041 struct bpf_verifier_state_list *sl;
13044 sl = *explored_state(env, insn);
13046 if (sl->state.branches)
13048 if (sl->state.insn_idx != insn ||
13049 sl->state.curframe != cur->curframe)
13051 for (i = 0; i <= cur->curframe; i++)
13052 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
13054 clean_verifier_state(env, &sl->state);
13060 /* Returns true if (rold safe implies rcur safe) */
13061 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
13062 struct bpf_reg_state *rcur, struct bpf_id_pair *idmap)
13066 if (!(rold->live & REG_LIVE_READ))
13067 /* explored state didn't use this */
13070 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
13072 if (rold->type == NOT_INIT)
13073 /* explored state can't have used this */
13075 if (rcur->type == NOT_INIT)
13078 /* Enforce that register types have to match exactly, including their
13079 * modifiers (like PTR_MAYBE_NULL, MEM_RDONLY, etc), as a general
13082 * One can make a point that using a pointer register as unbounded
13083 * SCALAR would be technically acceptable, but this could lead to
13084 * pointer leaks because scalars are allowed to leak while pointers
13085 * are not. We could make this safe in special cases if root is
13086 * calling us, but it's probably not worth the hassle.
13088 * Also, register types that are *not* MAYBE_NULL could technically be
13089 * safe to use as their MAYBE_NULL variants (e.g., PTR_TO_MAP_VALUE
13090 * is safe to be used as PTR_TO_MAP_VALUE_OR_NULL, provided both point
13091 * to the same map).
13092 * However, if the old MAYBE_NULL register then got NULL checked,
13093 * doing so could have affected others with the same id, and we can't
13094 * check for that because we lost the id when we converted to
13095 * a non-MAYBE_NULL variant.
13096 * So, as a general rule we don't allow mixing MAYBE_NULL and
13097 * non-MAYBE_NULL registers as well.
13099 if (rold->type != rcur->type)
13102 switch (base_type(rold->type)) {
13106 if (env->explore_alu_limits)
13108 if (!rold->precise)
13110 /* new val must satisfy old val knowledge */
13111 return range_within(rold, rcur) &&
13112 tnum_in(rold->var_off, rcur->var_off);
13113 case PTR_TO_MAP_KEY:
13114 case PTR_TO_MAP_VALUE:
13115 /* If the new min/max/var_off satisfy the old ones and
13116 * everything else matches, we are OK.
13118 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, var_off)) == 0 &&
13119 range_within(rold, rcur) &&
13120 tnum_in(rold->var_off, rcur->var_off) &&
13121 check_ids(rold->id, rcur->id, idmap);
13122 case PTR_TO_PACKET_META:
13123 case PTR_TO_PACKET:
13124 /* We must have at least as much range as the old ptr
13125 * did, so that any accesses which were safe before are
13126 * still safe. This is true even if old range < old off,
13127 * since someone could have accessed through (ptr - k), or
13128 * even done ptr -= k in a register, to get a safe access.
13130 if (rold->range > rcur->range)
13132 /* If the offsets don't match, we can't trust our alignment;
13133 * nor can we be sure that we won't fall out of range.
13135 if (rold->off != rcur->off)
13137 /* id relations must be preserved */
13138 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
13140 /* new val must satisfy old val knowledge */
13141 return range_within(rold, rcur) &&
13142 tnum_in(rold->var_off, rcur->var_off);
13144 /* two stack pointers are equal only if they're pointing to
13145 * the same stack frame, since fp-8 in foo != fp-8 in bar
13147 return equal && rold->frameno == rcur->frameno;
13149 /* Only valid matches are exact, which memcmp() */
13153 /* Shouldn't get here; if we do, say it's not safe */
13158 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
13159 struct bpf_func_state *cur, struct bpf_id_pair *idmap)
13163 /* walk slots of the explored stack and ignore any additional
13164 * slots in the current stack, since explored(safe) state
13167 for (i = 0; i < old->allocated_stack; i++) {
13168 spi = i / BPF_REG_SIZE;
13170 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
13171 i += BPF_REG_SIZE - 1;
13172 /* explored state didn't use this */
13176 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
13179 /* explored stack has more populated slots than current stack
13180 * and these slots were used
13182 if (i >= cur->allocated_stack)
13185 /* if old state was safe with misc data in the stack
13186 * it will be safe with zero-initialized stack.
13187 * The opposite is not true
13189 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
13190 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
13192 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
13193 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
13194 /* Ex: old explored (safe) state has STACK_SPILL in
13195 * this stack slot, but current has STACK_MISC ->
13196 * this verifier states are not equivalent,
13197 * return false to continue verification of this path
13200 if (i % BPF_REG_SIZE != BPF_REG_SIZE - 1)
13202 if (!is_spilled_reg(&old->stack[spi]))
13204 if (!regsafe(env, &old->stack[spi].spilled_ptr,
13205 &cur->stack[spi].spilled_ptr, idmap))
13206 /* when explored and current stack slot are both storing
13207 * spilled registers, check that stored pointers types
13208 * are the same as well.
13209 * Ex: explored safe path could have stored
13210 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
13211 * but current path has stored:
13212 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
13213 * such verifier states are not equivalent.
13214 * return false to continue verification of this path
13221 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur,
13222 struct bpf_id_pair *idmap)
13226 if (old->acquired_refs != cur->acquired_refs)
13229 for (i = 0; i < old->acquired_refs; i++) {
13230 if (!check_ids(old->refs[i].id, cur->refs[i].id, idmap))
13237 /* compare two verifier states
13239 * all states stored in state_list are known to be valid, since
13240 * verifier reached 'bpf_exit' instruction through them
13242 * this function is called when verifier exploring different branches of
13243 * execution popped from the state stack. If it sees an old state that has
13244 * more strict register state and more strict stack state then this execution
13245 * branch doesn't need to be explored further, since verifier already
13246 * concluded that more strict state leads to valid finish.
13248 * Therefore two states are equivalent if register state is more conservative
13249 * and explored stack state is more conservative than the current one.
13252 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
13253 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
13255 * In other words if current stack state (one being explored) has more
13256 * valid slots than old one that already passed validation, it means
13257 * the verifier can stop exploring and conclude that current state is valid too
13259 * Similarly with registers. If explored state has register type as invalid
13260 * whereas register type in current state is meaningful, it means that
13261 * the current state will reach 'bpf_exit' instruction safely
13263 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
13264 struct bpf_func_state *cur)
13268 for (i = 0; i < MAX_BPF_REG; i++)
13269 if (!regsafe(env, &old->regs[i], &cur->regs[i],
13270 env->idmap_scratch))
13273 if (!stacksafe(env, old, cur, env->idmap_scratch))
13276 if (!refsafe(old, cur, env->idmap_scratch))
13282 static bool states_equal(struct bpf_verifier_env *env,
13283 struct bpf_verifier_state *old,
13284 struct bpf_verifier_state *cur)
13288 if (old->curframe != cur->curframe)
13291 memset(env->idmap_scratch, 0, sizeof(env->idmap_scratch));
13293 /* Verification state from speculative execution simulation
13294 * must never prune a non-speculative execution one.
13296 if (old->speculative && !cur->speculative)
13299 if (old->active_lock.ptr != cur->active_lock.ptr)
13302 /* Old and cur active_lock's have to be either both present
13305 if (!!old->active_lock.id != !!cur->active_lock.id)
13308 if (old->active_lock.id &&
13309 !check_ids(old->active_lock.id, cur->active_lock.id, env->idmap_scratch))
13312 if (old->active_rcu_lock != cur->active_rcu_lock)
13315 /* for states to be equal callsites have to be the same
13316 * and all frame states need to be equivalent
13318 for (i = 0; i <= old->curframe; i++) {
13319 if (old->frame[i]->callsite != cur->frame[i]->callsite)
13321 if (!func_states_equal(env, old->frame[i], cur->frame[i]))
13327 /* Return 0 if no propagation happened. Return negative error code if error
13328 * happened. Otherwise, return the propagated bit.
13330 static int propagate_liveness_reg(struct bpf_verifier_env *env,
13331 struct bpf_reg_state *reg,
13332 struct bpf_reg_state *parent_reg)
13334 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
13335 u8 flag = reg->live & REG_LIVE_READ;
13338 /* When comes here, read flags of PARENT_REG or REG could be any of
13339 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
13340 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
13342 if (parent_flag == REG_LIVE_READ64 ||
13343 /* Or if there is no read flag from REG. */
13345 /* Or if the read flag from REG is the same as PARENT_REG. */
13346 parent_flag == flag)
13349 err = mark_reg_read(env, reg, parent_reg, flag);
13356 /* A write screens off any subsequent reads; but write marks come from the
13357 * straight-line code between a state and its parent. When we arrive at an
13358 * equivalent state (jump target or such) we didn't arrive by the straight-line
13359 * code, so read marks in the state must propagate to the parent regardless
13360 * of the state's write marks. That's what 'parent == state->parent' comparison
13361 * in mark_reg_read() is for.
13363 static int propagate_liveness(struct bpf_verifier_env *env,
13364 const struct bpf_verifier_state *vstate,
13365 struct bpf_verifier_state *vparent)
13367 struct bpf_reg_state *state_reg, *parent_reg;
13368 struct bpf_func_state *state, *parent;
13369 int i, frame, err = 0;
13371 if (vparent->curframe != vstate->curframe) {
13372 WARN(1, "propagate_live: parent frame %d current frame %d\n",
13373 vparent->curframe, vstate->curframe);
13376 /* Propagate read liveness of registers... */
13377 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
13378 for (frame = 0; frame <= vstate->curframe; frame++) {
13379 parent = vparent->frame[frame];
13380 state = vstate->frame[frame];
13381 parent_reg = parent->regs;
13382 state_reg = state->regs;
13383 /* We don't need to worry about FP liveness, it's read-only */
13384 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
13385 err = propagate_liveness_reg(env, &state_reg[i],
13389 if (err == REG_LIVE_READ64)
13390 mark_insn_zext(env, &parent_reg[i]);
13393 /* Propagate stack slots. */
13394 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
13395 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
13396 parent_reg = &parent->stack[i].spilled_ptr;
13397 state_reg = &state->stack[i].spilled_ptr;
13398 err = propagate_liveness_reg(env, state_reg,
13407 /* find precise scalars in the previous equivalent state and
13408 * propagate them into the current state
13410 static int propagate_precision(struct bpf_verifier_env *env,
13411 const struct bpf_verifier_state *old)
13413 struct bpf_reg_state *state_reg;
13414 struct bpf_func_state *state;
13415 int i, err = 0, fr;
13417 for (fr = old->curframe; fr >= 0; fr--) {
13418 state = old->frame[fr];
13419 state_reg = state->regs;
13420 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
13421 if (state_reg->type != SCALAR_VALUE ||
13422 !state_reg->precise)
13424 if (env->log.level & BPF_LOG_LEVEL2)
13425 verbose(env, "frame %d: propagating r%d\n", i, fr);
13426 err = mark_chain_precision_frame(env, fr, i);
13431 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
13432 if (!is_spilled_reg(&state->stack[i]))
13434 state_reg = &state->stack[i].spilled_ptr;
13435 if (state_reg->type != SCALAR_VALUE ||
13436 !state_reg->precise)
13438 if (env->log.level & BPF_LOG_LEVEL2)
13439 verbose(env, "frame %d: propagating fp%d\n",
13440 (-i - 1) * BPF_REG_SIZE, fr);
13441 err = mark_chain_precision_stack_frame(env, fr, i);
13449 static bool states_maybe_looping(struct bpf_verifier_state *old,
13450 struct bpf_verifier_state *cur)
13452 struct bpf_func_state *fold, *fcur;
13453 int i, fr = cur->curframe;
13455 if (old->curframe != fr)
13458 fold = old->frame[fr];
13459 fcur = cur->frame[fr];
13460 for (i = 0; i < MAX_BPF_REG; i++)
13461 if (memcmp(&fold->regs[i], &fcur->regs[i],
13462 offsetof(struct bpf_reg_state, parent)))
13468 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
13470 struct bpf_verifier_state_list *new_sl;
13471 struct bpf_verifier_state_list *sl, **pprev;
13472 struct bpf_verifier_state *cur = env->cur_state, *new;
13473 int i, j, err, states_cnt = 0;
13474 bool add_new_state = env->test_state_freq ? true : false;
13476 /* bpf progs typically have pruning point every 4 instructions
13477 * http://vger.kernel.org/bpfconf2019.html#session-1
13478 * Do not add new state for future pruning if the verifier hasn't seen
13479 * at least 2 jumps and at least 8 instructions.
13480 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
13481 * In tests that amounts to up to 50% reduction into total verifier
13482 * memory consumption and 20% verifier time speedup.
13484 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
13485 env->insn_processed - env->prev_insn_processed >= 8)
13486 add_new_state = true;
13488 pprev = explored_state(env, insn_idx);
13491 clean_live_states(env, insn_idx, cur);
13495 if (sl->state.insn_idx != insn_idx)
13498 if (sl->state.branches) {
13499 struct bpf_func_state *frame = sl->state.frame[sl->state.curframe];
13501 if (frame->in_async_callback_fn &&
13502 frame->async_entry_cnt != cur->frame[cur->curframe]->async_entry_cnt) {
13503 /* Different async_entry_cnt means that the verifier is
13504 * processing another entry into async callback.
13505 * Seeing the same state is not an indication of infinite
13506 * loop or infinite recursion.
13507 * But finding the same state doesn't mean that it's safe
13508 * to stop processing the current state. The previous state
13509 * hasn't yet reached bpf_exit, since state.branches > 0.
13510 * Checking in_async_callback_fn alone is not enough either.
13511 * Since the verifier still needs to catch infinite loops
13512 * inside async callbacks.
13514 } else if (states_maybe_looping(&sl->state, cur) &&
13515 states_equal(env, &sl->state, cur)) {
13516 verbose_linfo(env, insn_idx, "; ");
13517 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
13520 /* if the verifier is processing a loop, avoid adding new state
13521 * too often, since different loop iterations have distinct
13522 * states and may not help future pruning.
13523 * This threshold shouldn't be too low to make sure that
13524 * a loop with large bound will be rejected quickly.
13525 * The most abusive loop will be:
13527 * if r1 < 1000000 goto pc-2
13528 * 1M insn_procssed limit / 100 == 10k peak states.
13529 * This threshold shouldn't be too high either, since states
13530 * at the end of the loop are likely to be useful in pruning.
13532 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
13533 env->insn_processed - env->prev_insn_processed < 100)
13534 add_new_state = false;
13537 if (states_equal(env, &sl->state, cur)) {
13539 /* reached equivalent register/stack state,
13540 * prune the search.
13541 * Registers read by the continuation are read by us.
13542 * If we have any write marks in env->cur_state, they
13543 * will prevent corresponding reads in the continuation
13544 * from reaching our parent (an explored_state). Our
13545 * own state will get the read marks recorded, but
13546 * they'll be immediately forgotten as we're pruning
13547 * this state and will pop a new one.
13549 err = propagate_liveness(env, &sl->state, cur);
13551 /* if previous state reached the exit with precision and
13552 * current state is equivalent to it (except precsion marks)
13553 * the precision needs to be propagated back in
13554 * the current state.
13556 err = err ? : push_jmp_history(env, cur);
13557 err = err ? : propagate_precision(env, &sl->state);
13563 /* when new state is not going to be added do not increase miss count.
13564 * Otherwise several loop iterations will remove the state
13565 * recorded earlier. The goal of these heuristics is to have
13566 * states from some iterations of the loop (some in the beginning
13567 * and some at the end) to help pruning.
13571 /* heuristic to determine whether this state is beneficial
13572 * to keep checking from state equivalence point of view.
13573 * Higher numbers increase max_states_per_insn and verification time,
13574 * but do not meaningfully decrease insn_processed.
13576 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
13577 /* the state is unlikely to be useful. Remove it to
13578 * speed up verification
13581 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
13582 u32 br = sl->state.branches;
13585 "BUG live_done but branches_to_explore %d\n",
13587 free_verifier_state(&sl->state, false);
13589 env->peak_states--;
13591 /* cannot free this state, since parentage chain may
13592 * walk it later. Add it for free_list instead to
13593 * be freed at the end of verification
13595 sl->next = env->free_list;
13596 env->free_list = sl;
13606 if (env->max_states_per_insn < states_cnt)
13607 env->max_states_per_insn = states_cnt;
13609 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
13612 if (!add_new_state)
13615 /* There were no equivalent states, remember the current one.
13616 * Technically the current state is not proven to be safe yet,
13617 * but it will either reach outer most bpf_exit (which means it's safe)
13618 * or it will be rejected. When there are no loops the verifier won't be
13619 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
13620 * again on the way to bpf_exit.
13621 * When looping the sl->state.branches will be > 0 and this state
13622 * will not be considered for equivalence until branches == 0.
13624 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
13627 env->total_states++;
13628 env->peak_states++;
13629 env->prev_jmps_processed = env->jmps_processed;
13630 env->prev_insn_processed = env->insn_processed;
13632 /* forget precise markings we inherited, see __mark_chain_precision */
13633 if (env->bpf_capable)
13634 mark_all_scalars_imprecise(env, cur);
13636 /* add new state to the head of linked list */
13637 new = &new_sl->state;
13638 err = copy_verifier_state(new, cur);
13640 free_verifier_state(new, false);
13644 new->insn_idx = insn_idx;
13645 WARN_ONCE(new->branches != 1,
13646 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
13649 cur->first_insn_idx = insn_idx;
13650 clear_jmp_history(cur);
13651 new_sl->next = *explored_state(env, insn_idx);
13652 *explored_state(env, insn_idx) = new_sl;
13653 /* connect new state to parentage chain. Current frame needs all
13654 * registers connected. Only r6 - r9 of the callers are alive (pushed
13655 * to the stack implicitly by JITs) so in callers' frames connect just
13656 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
13657 * the state of the call instruction (with WRITTEN set), and r0 comes
13658 * from callee with its full parentage chain, anyway.
13660 /* clear write marks in current state: the writes we did are not writes
13661 * our child did, so they don't screen off its reads from us.
13662 * (There are no read marks in current state, because reads always mark
13663 * their parent and current state never has children yet. Only
13664 * explored_states can get read marks.)
13666 for (j = 0; j <= cur->curframe; j++) {
13667 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
13668 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
13669 for (i = 0; i < BPF_REG_FP; i++)
13670 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
13673 /* all stack frames are accessible from callee, clear them all */
13674 for (j = 0; j <= cur->curframe; j++) {
13675 struct bpf_func_state *frame = cur->frame[j];
13676 struct bpf_func_state *newframe = new->frame[j];
13678 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
13679 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
13680 frame->stack[i].spilled_ptr.parent =
13681 &newframe->stack[i].spilled_ptr;
13687 /* Return true if it's OK to have the same insn return a different type. */
13688 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
13690 switch (base_type(type)) {
13692 case PTR_TO_SOCKET:
13693 case PTR_TO_SOCK_COMMON:
13694 case PTR_TO_TCP_SOCK:
13695 case PTR_TO_XDP_SOCK:
13696 case PTR_TO_BTF_ID:
13703 /* If an instruction was previously used with particular pointer types, then we
13704 * need to be careful to avoid cases such as the below, where it may be ok
13705 * for one branch accessing the pointer, but not ok for the other branch:
13710 * R1 = some_other_valid_ptr;
13713 * R2 = *(u32 *)(R1 + 0);
13715 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
13717 return src != prev && (!reg_type_mismatch_ok(src) ||
13718 !reg_type_mismatch_ok(prev));
13721 static int do_check(struct bpf_verifier_env *env)
13723 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
13724 struct bpf_verifier_state *state = env->cur_state;
13725 struct bpf_insn *insns = env->prog->insnsi;
13726 struct bpf_reg_state *regs;
13727 int insn_cnt = env->prog->len;
13728 bool do_print_state = false;
13729 int prev_insn_idx = -1;
13732 struct bpf_insn *insn;
13736 env->prev_insn_idx = prev_insn_idx;
13737 if (env->insn_idx >= insn_cnt) {
13738 verbose(env, "invalid insn idx %d insn_cnt %d\n",
13739 env->insn_idx, insn_cnt);
13743 insn = &insns[env->insn_idx];
13744 class = BPF_CLASS(insn->code);
13746 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
13748 "BPF program is too large. Processed %d insn\n",
13749 env->insn_processed);
13753 state->last_insn_idx = env->prev_insn_idx;
13755 if (is_prune_point(env, env->insn_idx)) {
13756 err = is_state_visited(env, env->insn_idx);
13760 /* found equivalent state, can prune the search */
13761 if (env->log.level & BPF_LOG_LEVEL) {
13762 if (do_print_state)
13763 verbose(env, "\nfrom %d to %d%s: safe\n",
13764 env->prev_insn_idx, env->insn_idx,
13765 env->cur_state->speculative ?
13766 " (speculative execution)" : "");
13768 verbose(env, "%d: safe\n", env->insn_idx);
13770 goto process_bpf_exit;
13774 if (is_jmp_point(env, env->insn_idx)) {
13775 err = push_jmp_history(env, state);
13780 if (signal_pending(current))
13783 if (need_resched())
13786 if (env->log.level & BPF_LOG_LEVEL2 && do_print_state) {
13787 verbose(env, "\nfrom %d to %d%s:",
13788 env->prev_insn_idx, env->insn_idx,
13789 env->cur_state->speculative ?
13790 " (speculative execution)" : "");
13791 print_verifier_state(env, state->frame[state->curframe], true);
13792 do_print_state = false;
13795 if (env->log.level & BPF_LOG_LEVEL) {
13796 const struct bpf_insn_cbs cbs = {
13797 .cb_call = disasm_kfunc_name,
13798 .cb_print = verbose,
13799 .private_data = env,
13802 if (verifier_state_scratched(env))
13803 print_insn_state(env, state->frame[state->curframe]);
13805 verbose_linfo(env, env->insn_idx, "; ");
13806 env->prev_log_len = env->log.len_used;
13807 verbose(env, "%d: ", env->insn_idx);
13808 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
13809 env->prev_insn_print_len = env->log.len_used - env->prev_log_len;
13810 env->prev_log_len = env->log.len_used;
13813 if (bpf_prog_is_dev_bound(env->prog->aux)) {
13814 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
13815 env->prev_insn_idx);
13820 regs = cur_regs(env);
13821 sanitize_mark_insn_seen(env);
13822 prev_insn_idx = env->insn_idx;
13824 if (class == BPF_ALU || class == BPF_ALU64) {
13825 err = check_alu_op(env, insn);
13829 } else if (class == BPF_LDX) {
13830 enum bpf_reg_type *prev_src_type, src_reg_type;
13832 /* check for reserved fields is already done */
13834 /* check src operand */
13835 err = check_reg_arg(env, insn->src_reg, SRC_OP);
13839 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
13843 src_reg_type = regs[insn->src_reg].type;
13845 /* check that memory (src_reg + off) is readable,
13846 * the state of dst_reg will be updated by this func
13848 err = check_mem_access(env, env->insn_idx, insn->src_reg,
13849 insn->off, BPF_SIZE(insn->code),
13850 BPF_READ, insn->dst_reg, false);
13854 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
13856 if (*prev_src_type == NOT_INIT) {
13857 /* saw a valid insn
13858 * dst_reg = *(u32 *)(src_reg + off)
13859 * save type to validate intersecting paths
13861 *prev_src_type = src_reg_type;
13863 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
13864 /* ABuser program is trying to use the same insn
13865 * dst_reg = *(u32*) (src_reg + off)
13866 * with different pointer types:
13867 * src_reg == ctx in one branch and
13868 * src_reg == stack|map in some other branch.
13871 verbose(env, "same insn cannot be used with different pointers\n");
13875 } else if (class == BPF_STX) {
13876 enum bpf_reg_type *prev_dst_type, dst_reg_type;
13878 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
13879 err = check_atomic(env, env->insn_idx, insn);
13886 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
13887 verbose(env, "BPF_STX uses reserved fields\n");
13891 /* check src1 operand */
13892 err = check_reg_arg(env, insn->src_reg, SRC_OP);
13895 /* check src2 operand */
13896 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
13900 dst_reg_type = regs[insn->dst_reg].type;
13902 /* check that memory (dst_reg + off) is writeable */
13903 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
13904 insn->off, BPF_SIZE(insn->code),
13905 BPF_WRITE, insn->src_reg, false);
13909 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
13911 if (*prev_dst_type == NOT_INIT) {
13912 *prev_dst_type = dst_reg_type;
13913 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
13914 verbose(env, "same insn cannot be used with different pointers\n");
13918 } else if (class == BPF_ST) {
13919 if (BPF_MODE(insn->code) != BPF_MEM ||
13920 insn->src_reg != BPF_REG_0) {
13921 verbose(env, "BPF_ST uses reserved fields\n");
13924 /* check src operand */
13925 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
13929 if (is_ctx_reg(env, insn->dst_reg)) {
13930 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
13932 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
13936 /* check that memory (dst_reg + off) is writeable */
13937 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
13938 insn->off, BPF_SIZE(insn->code),
13939 BPF_WRITE, -1, false);
13943 } else if (class == BPF_JMP || class == BPF_JMP32) {
13944 u8 opcode = BPF_OP(insn->code);
13946 env->jmps_processed++;
13947 if (opcode == BPF_CALL) {
13948 if (BPF_SRC(insn->code) != BPF_K ||
13949 (insn->src_reg != BPF_PSEUDO_KFUNC_CALL
13950 && insn->off != 0) ||
13951 (insn->src_reg != BPF_REG_0 &&
13952 insn->src_reg != BPF_PSEUDO_CALL &&
13953 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
13954 insn->dst_reg != BPF_REG_0 ||
13955 class == BPF_JMP32) {
13956 verbose(env, "BPF_CALL uses reserved fields\n");
13960 if (env->cur_state->active_lock.ptr) {
13961 if ((insn->src_reg == BPF_REG_0 && insn->imm != BPF_FUNC_spin_unlock) ||
13962 (insn->src_reg == BPF_PSEUDO_CALL) ||
13963 (insn->src_reg == BPF_PSEUDO_KFUNC_CALL &&
13964 (insn->off != 0 || !is_bpf_list_api_kfunc(insn->imm)))) {
13965 verbose(env, "function calls are not allowed while holding a lock\n");
13969 if (insn->src_reg == BPF_PSEUDO_CALL)
13970 err = check_func_call(env, insn, &env->insn_idx);
13971 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
13972 err = check_kfunc_call(env, insn, &env->insn_idx);
13974 err = check_helper_call(env, insn, &env->insn_idx);
13977 } else if (opcode == BPF_JA) {
13978 if (BPF_SRC(insn->code) != BPF_K ||
13980 insn->src_reg != BPF_REG_0 ||
13981 insn->dst_reg != BPF_REG_0 ||
13982 class == BPF_JMP32) {
13983 verbose(env, "BPF_JA uses reserved fields\n");
13987 env->insn_idx += insn->off + 1;
13990 } else if (opcode == BPF_EXIT) {
13991 if (BPF_SRC(insn->code) != BPF_K ||
13993 insn->src_reg != BPF_REG_0 ||
13994 insn->dst_reg != BPF_REG_0 ||
13995 class == BPF_JMP32) {
13996 verbose(env, "BPF_EXIT uses reserved fields\n");
14000 if (env->cur_state->active_lock.ptr) {
14001 verbose(env, "bpf_spin_unlock is missing\n");
14005 if (env->cur_state->active_rcu_lock) {
14006 verbose(env, "bpf_rcu_read_unlock is missing\n");
14010 /* We must do check_reference_leak here before
14011 * prepare_func_exit to handle the case when
14012 * state->curframe > 0, it may be a callback
14013 * function, for which reference_state must
14014 * match caller reference state when it exits.
14016 err = check_reference_leak(env);
14020 if (state->curframe) {
14021 /* exit from nested function */
14022 err = prepare_func_exit(env, &env->insn_idx);
14025 do_print_state = true;
14029 err = check_return_code(env);
14033 mark_verifier_state_scratched(env);
14034 update_branch_counts(env, env->cur_state);
14035 err = pop_stack(env, &prev_insn_idx,
14036 &env->insn_idx, pop_log);
14038 if (err != -ENOENT)
14042 do_print_state = true;
14046 err = check_cond_jmp_op(env, insn, &env->insn_idx);
14050 } else if (class == BPF_LD) {
14051 u8 mode = BPF_MODE(insn->code);
14053 if (mode == BPF_ABS || mode == BPF_IND) {
14054 err = check_ld_abs(env, insn);
14058 } else if (mode == BPF_IMM) {
14059 err = check_ld_imm(env, insn);
14064 sanitize_mark_insn_seen(env);
14066 verbose(env, "invalid BPF_LD mode\n");
14070 verbose(env, "unknown insn class %d\n", class);
14080 static int find_btf_percpu_datasec(struct btf *btf)
14082 const struct btf_type *t;
14087 * Both vmlinux and module each have their own ".data..percpu"
14088 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
14089 * types to look at only module's own BTF types.
14091 n = btf_nr_types(btf);
14092 if (btf_is_module(btf))
14093 i = btf_nr_types(btf_vmlinux);
14097 for(; i < n; i++) {
14098 t = btf_type_by_id(btf, i);
14099 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
14102 tname = btf_name_by_offset(btf, t->name_off);
14103 if (!strcmp(tname, ".data..percpu"))
14110 /* replace pseudo btf_id with kernel symbol address */
14111 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
14112 struct bpf_insn *insn,
14113 struct bpf_insn_aux_data *aux)
14115 const struct btf_var_secinfo *vsi;
14116 const struct btf_type *datasec;
14117 struct btf_mod_pair *btf_mod;
14118 const struct btf_type *t;
14119 const char *sym_name;
14120 bool percpu = false;
14121 u32 type, id = insn->imm;
14125 int i, btf_fd, err;
14127 btf_fd = insn[1].imm;
14129 btf = btf_get_by_fd(btf_fd);
14131 verbose(env, "invalid module BTF object FD specified.\n");
14135 if (!btf_vmlinux) {
14136 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
14143 t = btf_type_by_id(btf, id);
14145 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
14150 if (!btf_type_is_var(t)) {
14151 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR.\n", id);
14156 sym_name = btf_name_by_offset(btf, t->name_off);
14157 addr = kallsyms_lookup_name(sym_name);
14159 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
14165 datasec_id = find_btf_percpu_datasec(btf);
14166 if (datasec_id > 0) {
14167 datasec = btf_type_by_id(btf, datasec_id);
14168 for_each_vsi(i, datasec, vsi) {
14169 if (vsi->type == id) {
14176 insn[0].imm = (u32)addr;
14177 insn[1].imm = addr >> 32;
14180 t = btf_type_skip_modifiers(btf, type, NULL);
14182 aux->btf_var.reg_type = PTR_TO_BTF_ID | MEM_PERCPU;
14183 aux->btf_var.btf = btf;
14184 aux->btf_var.btf_id = type;
14185 } else if (!btf_type_is_struct(t)) {
14186 const struct btf_type *ret;
14190 /* resolve the type size of ksym. */
14191 ret = btf_resolve_size(btf, t, &tsize);
14193 tname = btf_name_by_offset(btf, t->name_off);
14194 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
14195 tname, PTR_ERR(ret));
14199 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
14200 aux->btf_var.mem_size = tsize;
14202 aux->btf_var.reg_type = PTR_TO_BTF_ID;
14203 aux->btf_var.btf = btf;
14204 aux->btf_var.btf_id = type;
14207 /* check whether we recorded this BTF (and maybe module) already */
14208 for (i = 0; i < env->used_btf_cnt; i++) {
14209 if (env->used_btfs[i].btf == btf) {
14215 if (env->used_btf_cnt >= MAX_USED_BTFS) {
14220 btf_mod = &env->used_btfs[env->used_btf_cnt];
14221 btf_mod->btf = btf;
14222 btf_mod->module = NULL;
14224 /* if we reference variables from kernel module, bump its refcount */
14225 if (btf_is_module(btf)) {
14226 btf_mod->module = btf_try_get_module(btf);
14227 if (!btf_mod->module) {
14233 env->used_btf_cnt++;
14241 static bool is_tracing_prog_type(enum bpf_prog_type type)
14244 case BPF_PROG_TYPE_KPROBE:
14245 case BPF_PROG_TYPE_TRACEPOINT:
14246 case BPF_PROG_TYPE_PERF_EVENT:
14247 case BPF_PROG_TYPE_RAW_TRACEPOINT:
14248 case BPF_PROG_TYPE_RAW_TRACEPOINT_WRITABLE:
14255 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
14256 struct bpf_map *map,
14257 struct bpf_prog *prog)
14260 enum bpf_prog_type prog_type = resolve_prog_type(prog);
14262 if (btf_record_has_field(map->record, BPF_LIST_HEAD)) {
14263 if (is_tracing_prog_type(prog_type)) {
14264 verbose(env, "tracing progs cannot use bpf_list_head yet\n");
14269 if (btf_record_has_field(map->record, BPF_SPIN_LOCK)) {
14270 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
14271 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
14275 if (is_tracing_prog_type(prog_type)) {
14276 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
14280 if (prog->aux->sleepable) {
14281 verbose(env, "sleepable progs cannot use bpf_spin_lock yet\n");
14286 if (btf_record_has_field(map->record, BPF_TIMER)) {
14287 if (is_tracing_prog_type(prog_type)) {
14288 verbose(env, "tracing progs cannot use bpf_timer yet\n");
14293 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
14294 !bpf_offload_prog_map_match(prog, map)) {
14295 verbose(env, "offload device mismatch between prog and map\n");
14299 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
14300 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
14304 if (prog->aux->sleepable)
14305 switch (map->map_type) {
14306 case BPF_MAP_TYPE_HASH:
14307 case BPF_MAP_TYPE_LRU_HASH:
14308 case BPF_MAP_TYPE_ARRAY:
14309 case BPF_MAP_TYPE_PERCPU_HASH:
14310 case BPF_MAP_TYPE_PERCPU_ARRAY:
14311 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
14312 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
14313 case BPF_MAP_TYPE_HASH_OF_MAPS:
14314 case BPF_MAP_TYPE_RINGBUF:
14315 case BPF_MAP_TYPE_USER_RINGBUF:
14316 case BPF_MAP_TYPE_INODE_STORAGE:
14317 case BPF_MAP_TYPE_SK_STORAGE:
14318 case BPF_MAP_TYPE_TASK_STORAGE:
14319 case BPF_MAP_TYPE_CGRP_STORAGE:
14323 "Sleepable programs can only use array, hash, ringbuf and local storage maps\n");
14330 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
14332 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
14333 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
14336 /* find and rewrite pseudo imm in ld_imm64 instructions:
14338 * 1. if it accesses map FD, replace it with actual map pointer.
14339 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
14341 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
14343 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
14345 struct bpf_insn *insn = env->prog->insnsi;
14346 int insn_cnt = env->prog->len;
14349 err = bpf_prog_calc_tag(env->prog);
14353 for (i = 0; i < insn_cnt; i++, insn++) {
14354 if (BPF_CLASS(insn->code) == BPF_LDX &&
14355 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
14356 verbose(env, "BPF_LDX uses reserved fields\n");
14360 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
14361 struct bpf_insn_aux_data *aux;
14362 struct bpf_map *map;
14367 if (i == insn_cnt - 1 || insn[1].code != 0 ||
14368 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
14369 insn[1].off != 0) {
14370 verbose(env, "invalid bpf_ld_imm64 insn\n");
14374 if (insn[0].src_reg == 0)
14375 /* valid generic load 64-bit imm */
14378 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
14379 aux = &env->insn_aux_data[i];
14380 err = check_pseudo_btf_id(env, insn, aux);
14386 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
14387 aux = &env->insn_aux_data[i];
14388 aux->ptr_type = PTR_TO_FUNC;
14392 /* In final convert_pseudo_ld_imm64() step, this is
14393 * converted into regular 64-bit imm load insn.
14395 switch (insn[0].src_reg) {
14396 case BPF_PSEUDO_MAP_VALUE:
14397 case BPF_PSEUDO_MAP_IDX_VALUE:
14399 case BPF_PSEUDO_MAP_FD:
14400 case BPF_PSEUDO_MAP_IDX:
14401 if (insn[1].imm == 0)
14405 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
14409 switch (insn[0].src_reg) {
14410 case BPF_PSEUDO_MAP_IDX_VALUE:
14411 case BPF_PSEUDO_MAP_IDX:
14412 if (bpfptr_is_null(env->fd_array)) {
14413 verbose(env, "fd_idx without fd_array is invalid\n");
14416 if (copy_from_bpfptr_offset(&fd, env->fd_array,
14417 insn[0].imm * sizeof(fd),
14427 map = __bpf_map_get(f);
14429 verbose(env, "fd %d is not pointing to valid bpf_map\n",
14431 return PTR_ERR(map);
14434 err = check_map_prog_compatibility(env, map, env->prog);
14440 aux = &env->insn_aux_data[i];
14441 if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
14442 insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
14443 addr = (unsigned long)map;
14445 u32 off = insn[1].imm;
14447 if (off >= BPF_MAX_VAR_OFF) {
14448 verbose(env, "direct value offset of %u is not allowed\n", off);
14453 if (!map->ops->map_direct_value_addr) {
14454 verbose(env, "no direct value access support for this map type\n");
14459 err = map->ops->map_direct_value_addr(map, &addr, off);
14461 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
14462 map->value_size, off);
14467 aux->map_off = off;
14471 insn[0].imm = (u32)addr;
14472 insn[1].imm = addr >> 32;
14474 /* check whether we recorded this map already */
14475 for (j = 0; j < env->used_map_cnt; j++) {
14476 if (env->used_maps[j] == map) {
14477 aux->map_index = j;
14483 if (env->used_map_cnt >= MAX_USED_MAPS) {
14488 /* hold the map. If the program is rejected by verifier,
14489 * the map will be released by release_maps() or it
14490 * will be used by the valid program until it's unloaded
14491 * and all maps are released in free_used_maps()
14495 aux->map_index = env->used_map_cnt;
14496 env->used_maps[env->used_map_cnt++] = map;
14498 if (bpf_map_is_cgroup_storage(map) &&
14499 bpf_cgroup_storage_assign(env->prog->aux, map)) {
14500 verbose(env, "only one cgroup storage of each type is allowed\n");
14512 /* Basic sanity check before we invest more work here. */
14513 if (!bpf_opcode_in_insntable(insn->code)) {
14514 verbose(env, "unknown opcode %02x\n", insn->code);
14519 /* now all pseudo BPF_LD_IMM64 instructions load valid
14520 * 'struct bpf_map *' into a register instead of user map_fd.
14521 * These pointers will be used later by verifier to validate map access.
14526 /* drop refcnt of maps used by the rejected program */
14527 static void release_maps(struct bpf_verifier_env *env)
14529 __bpf_free_used_maps(env->prog->aux, env->used_maps,
14530 env->used_map_cnt);
14533 /* drop refcnt of maps used by the rejected program */
14534 static void release_btfs(struct bpf_verifier_env *env)
14536 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
14537 env->used_btf_cnt);
14540 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
14541 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
14543 struct bpf_insn *insn = env->prog->insnsi;
14544 int insn_cnt = env->prog->len;
14547 for (i = 0; i < insn_cnt; i++, insn++) {
14548 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
14550 if (insn->src_reg == BPF_PSEUDO_FUNC)
14556 /* single env->prog->insni[off] instruction was replaced with the range
14557 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
14558 * [0, off) and [off, end) to new locations, so the patched range stays zero
14560 static void adjust_insn_aux_data(struct bpf_verifier_env *env,
14561 struct bpf_insn_aux_data *new_data,
14562 struct bpf_prog *new_prog, u32 off, u32 cnt)
14564 struct bpf_insn_aux_data *old_data = env->insn_aux_data;
14565 struct bpf_insn *insn = new_prog->insnsi;
14566 u32 old_seen = old_data[off].seen;
14570 /* aux info at OFF always needs adjustment, no matter fast path
14571 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
14572 * original insn at old prog.
14574 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
14578 prog_len = new_prog->len;
14580 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
14581 memcpy(new_data + off + cnt - 1, old_data + off,
14582 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
14583 for (i = off; i < off + cnt - 1; i++) {
14584 /* Expand insni[off]'s seen count to the patched range. */
14585 new_data[i].seen = old_seen;
14586 new_data[i].zext_dst = insn_has_def32(env, insn + i);
14588 env->insn_aux_data = new_data;
14592 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
14598 /* NOTE: fake 'exit' subprog should be updated as well. */
14599 for (i = 0; i <= env->subprog_cnt; i++) {
14600 if (env->subprog_info[i].start <= off)
14602 env->subprog_info[i].start += len - 1;
14606 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
14608 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
14609 int i, sz = prog->aux->size_poke_tab;
14610 struct bpf_jit_poke_descriptor *desc;
14612 for (i = 0; i < sz; i++) {
14614 if (desc->insn_idx <= off)
14616 desc->insn_idx += len - 1;
14620 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
14621 const struct bpf_insn *patch, u32 len)
14623 struct bpf_prog *new_prog;
14624 struct bpf_insn_aux_data *new_data = NULL;
14627 new_data = vzalloc(array_size(env->prog->len + len - 1,
14628 sizeof(struct bpf_insn_aux_data)));
14633 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
14634 if (IS_ERR(new_prog)) {
14635 if (PTR_ERR(new_prog) == -ERANGE)
14637 "insn %d cannot be patched due to 16-bit range\n",
14638 env->insn_aux_data[off].orig_idx);
14642 adjust_insn_aux_data(env, new_data, new_prog, off, len);
14643 adjust_subprog_starts(env, off, len);
14644 adjust_poke_descs(new_prog, off, len);
14648 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
14653 /* find first prog starting at or after off (first to remove) */
14654 for (i = 0; i < env->subprog_cnt; i++)
14655 if (env->subprog_info[i].start >= off)
14657 /* find first prog starting at or after off + cnt (first to stay) */
14658 for (j = i; j < env->subprog_cnt; j++)
14659 if (env->subprog_info[j].start >= off + cnt)
14661 /* if j doesn't start exactly at off + cnt, we are just removing
14662 * the front of previous prog
14664 if (env->subprog_info[j].start != off + cnt)
14668 struct bpf_prog_aux *aux = env->prog->aux;
14671 /* move fake 'exit' subprog as well */
14672 move = env->subprog_cnt + 1 - j;
14674 memmove(env->subprog_info + i,
14675 env->subprog_info + j,
14676 sizeof(*env->subprog_info) * move);
14677 env->subprog_cnt -= j - i;
14679 /* remove func_info */
14680 if (aux->func_info) {
14681 move = aux->func_info_cnt - j;
14683 memmove(aux->func_info + i,
14684 aux->func_info + j,
14685 sizeof(*aux->func_info) * move);
14686 aux->func_info_cnt -= j - i;
14687 /* func_info->insn_off is set after all code rewrites,
14688 * in adjust_btf_func() - no need to adjust
14692 /* convert i from "first prog to remove" to "first to adjust" */
14693 if (env->subprog_info[i].start == off)
14697 /* update fake 'exit' subprog as well */
14698 for (; i <= env->subprog_cnt; i++)
14699 env->subprog_info[i].start -= cnt;
14704 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
14707 struct bpf_prog *prog = env->prog;
14708 u32 i, l_off, l_cnt, nr_linfo;
14709 struct bpf_line_info *linfo;
14711 nr_linfo = prog->aux->nr_linfo;
14715 linfo = prog->aux->linfo;
14717 /* find first line info to remove, count lines to be removed */
14718 for (i = 0; i < nr_linfo; i++)
14719 if (linfo[i].insn_off >= off)
14724 for (; i < nr_linfo; i++)
14725 if (linfo[i].insn_off < off + cnt)
14730 /* First live insn doesn't match first live linfo, it needs to "inherit"
14731 * last removed linfo. prog is already modified, so prog->len == off
14732 * means no live instructions after (tail of the program was removed).
14734 if (prog->len != off && l_cnt &&
14735 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
14737 linfo[--i].insn_off = off + cnt;
14740 /* remove the line info which refer to the removed instructions */
14742 memmove(linfo + l_off, linfo + i,
14743 sizeof(*linfo) * (nr_linfo - i));
14745 prog->aux->nr_linfo -= l_cnt;
14746 nr_linfo = prog->aux->nr_linfo;
14749 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
14750 for (i = l_off; i < nr_linfo; i++)
14751 linfo[i].insn_off -= cnt;
14753 /* fix up all subprogs (incl. 'exit') which start >= off */
14754 for (i = 0; i <= env->subprog_cnt; i++)
14755 if (env->subprog_info[i].linfo_idx > l_off) {
14756 /* program may have started in the removed region but
14757 * may not be fully removed
14759 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
14760 env->subprog_info[i].linfo_idx -= l_cnt;
14762 env->subprog_info[i].linfo_idx = l_off;
14768 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
14770 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
14771 unsigned int orig_prog_len = env->prog->len;
14774 if (bpf_prog_is_dev_bound(env->prog->aux))
14775 bpf_prog_offload_remove_insns(env, off, cnt);
14777 err = bpf_remove_insns(env->prog, off, cnt);
14781 err = adjust_subprog_starts_after_remove(env, off, cnt);
14785 err = bpf_adj_linfo_after_remove(env, off, cnt);
14789 memmove(aux_data + off, aux_data + off + cnt,
14790 sizeof(*aux_data) * (orig_prog_len - off - cnt));
14795 /* The verifier does more data flow analysis than llvm and will not
14796 * explore branches that are dead at run time. Malicious programs can
14797 * have dead code too. Therefore replace all dead at-run-time code
14800 * Just nops are not optimal, e.g. if they would sit at the end of the
14801 * program and through another bug we would manage to jump there, then
14802 * we'd execute beyond program memory otherwise. Returning exception
14803 * code also wouldn't work since we can have subprogs where the dead
14804 * code could be located.
14806 static void sanitize_dead_code(struct bpf_verifier_env *env)
14808 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
14809 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
14810 struct bpf_insn *insn = env->prog->insnsi;
14811 const int insn_cnt = env->prog->len;
14814 for (i = 0; i < insn_cnt; i++) {
14815 if (aux_data[i].seen)
14817 memcpy(insn + i, &trap, sizeof(trap));
14818 aux_data[i].zext_dst = false;
14822 static bool insn_is_cond_jump(u8 code)
14826 if (BPF_CLASS(code) == BPF_JMP32)
14829 if (BPF_CLASS(code) != BPF_JMP)
14833 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
14836 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
14838 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
14839 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
14840 struct bpf_insn *insn = env->prog->insnsi;
14841 const int insn_cnt = env->prog->len;
14844 for (i = 0; i < insn_cnt; i++, insn++) {
14845 if (!insn_is_cond_jump(insn->code))
14848 if (!aux_data[i + 1].seen)
14849 ja.off = insn->off;
14850 else if (!aux_data[i + 1 + insn->off].seen)
14855 if (bpf_prog_is_dev_bound(env->prog->aux))
14856 bpf_prog_offload_replace_insn(env, i, &ja);
14858 memcpy(insn, &ja, sizeof(ja));
14862 static int opt_remove_dead_code(struct bpf_verifier_env *env)
14864 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
14865 int insn_cnt = env->prog->len;
14868 for (i = 0; i < insn_cnt; i++) {
14872 while (i + j < insn_cnt && !aux_data[i + j].seen)
14877 err = verifier_remove_insns(env, i, j);
14880 insn_cnt = env->prog->len;
14886 static int opt_remove_nops(struct bpf_verifier_env *env)
14888 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
14889 struct bpf_insn *insn = env->prog->insnsi;
14890 int insn_cnt = env->prog->len;
14893 for (i = 0; i < insn_cnt; i++) {
14894 if (memcmp(&insn[i], &ja, sizeof(ja)))
14897 err = verifier_remove_insns(env, i, 1);
14907 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
14908 const union bpf_attr *attr)
14910 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
14911 struct bpf_insn_aux_data *aux = env->insn_aux_data;
14912 int i, patch_len, delta = 0, len = env->prog->len;
14913 struct bpf_insn *insns = env->prog->insnsi;
14914 struct bpf_prog *new_prog;
14917 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
14918 zext_patch[1] = BPF_ZEXT_REG(0);
14919 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
14920 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
14921 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
14922 for (i = 0; i < len; i++) {
14923 int adj_idx = i + delta;
14924 struct bpf_insn insn;
14927 insn = insns[adj_idx];
14928 load_reg = insn_def_regno(&insn);
14929 if (!aux[adj_idx].zext_dst) {
14937 class = BPF_CLASS(code);
14938 if (load_reg == -1)
14941 /* NOTE: arg "reg" (the fourth one) is only used for
14942 * BPF_STX + SRC_OP, so it is safe to pass NULL
14945 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
14946 if (class == BPF_LD &&
14947 BPF_MODE(code) == BPF_IMM)
14952 /* ctx load could be transformed into wider load. */
14953 if (class == BPF_LDX &&
14954 aux[adj_idx].ptr_type == PTR_TO_CTX)
14957 imm_rnd = get_random_u32();
14958 rnd_hi32_patch[0] = insn;
14959 rnd_hi32_patch[1].imm = imm_rnd;
14960 rnd_hi32_patch[3].dst_reg = load_reg;
14961 patch = rnd_hi32_patch;
14963 goto apply_patch_buffer;
14966 /* Add in an zero-extend instruction if a) the JIT has requested
14967 * it or b) it's a CMPXCHG.
14969 * The latter is because: BPF_CMPXCHG always loads a value into
14970 * R0, therefore always zero-extends. However some archs'
14971 * equivalent instruction only does this load when the
14972 * comparison is successful. This detail of CMPXCHG is
14973 * orthogonal to the general zero-extension behaviour of the
14974 * CPU, so it's treated independently of bpf_jit_needs_zext.
14976 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
14979 /* Zero-extension is done by the caller. */
14980 if (bpf_pseudo_kfunc_call(&insn))
14983 if (WARN_ON(load_reg == -1)) {
14984 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
14988 zext_patch[0] = insn;
14989 zext_patch[1].dst_reg = load_reg;
14990 zext_patch[1].src_reg = load_reg;
14991 patch = zext_patch;
14993 apply_patch_buffer:
14994 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
14997 env->prog = new_prog;
14998 insns = new_prog->insnsi;
14999 aux = env->insn_aux_data;
15000 delta += patch_len - 1;
15006 /* convert load instructions that access fields of a context type into a
15007 * sequence of instructions that access fields of the underlying structure:
15008 * struct __sk_buff -> struct sk_buff
15009 * struct bpf_sock_ops -> struct sock
15011 static int convert_ctx_accesses(struct bpf_verifier_env *env)
15013 const struct bpf_verifier_ops *ops = env->ops;
15014 int i, cnt, size, ctx_field_size, delta = 0;
15015 const int insn_cnt = env->prog->len;
15016 struct bpf_insn insn_buf[16], *insn;
15017 u32 target_size, size_default, off;
15018 struct bpf_prog *new_prog;
15019 enum bpf_access_type type;
15020 bool is_narrower_load;
15022 if (ops->gen_prologue || env->seen_direct_write) {
15023 if (!ops->gen_prologue) {
15024 verbose(env, "bpf verifier is misconfigured\n");
15027 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
15029 if (cnt >= ARRAY_SIZE(insn_buf)) {
15030 verbose(env, "bpf verifier is misconfigured\n");
15033 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
15037 env->prog = new_prog;
15042 if (bpf_prog_is_dev_bound(env->prog->aux))
15045 insn = env->prog->insnsi + delta;
15047 for (i = 0; i < insn_cnt; i++, insn++) {
15048 bpf_convert_ctx_access_t convert_ctx_access;
15051 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
15052 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
15053 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
15054 insn->code == (BPF_LDX | BPF_MEM | BPF_DW)) {
15057 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
15058 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
15059 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
15060 insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
15061 insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
15062 insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
15063 insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
15064 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
15066 ctx_access = BPF_CLASS(insn->code) == BPF_STX;
15071 if (type == BPF_WRITE &&
15072 env->insn_aux_data[i + delta].sanitize_stack_spill) {
15073 struct bpf_insn patch[] = {
15078 cnt = ARRAY_SIZE(patch);
15079 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
15084 env->prog = new_prog;
15085 insn = new_prog->insnsi + i + delta;
15092 switch ((int)env->insn_aux_data[i + delta].ptr_type) {
15094 if (!ops->convert_ctx_access)
15096 convert_ctx_access = ops->convert_ctx_access;
15098 case PTR_TO_SOCKET:
15099 case PTR_TO_SOCK_COMMON:
15100 convert_ctx_access = bpf_sock_convert_ctx_access;
15102 case PTR_TO_TCP_SOCK:
15103 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
15105 case PTR_TO_XDP_SOCK:
15106 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
15108 case PTR_TO_BTF_ID:
15109 case PTR_TO_BTF_ID | PTR_UNTRUSTED:
15110 /* PTR_TO_BTF_ID | MEM_ALLOC always has a valid lifetime, unlike
15111 * PTR_TO_BTF_ID, and an active ref_obj_id, but the same cannot
15112 * be said once it is marked PTR_UNTRUSTED, hence we must handle
15113 * any faults for loads into such types. BPF_WRITE is disallowed
15116 case PTR_TO_BTF_ID | MEM_ALLOC | PTR_UNTRUSTED:
15117 if (type == BPF_READ) {
15118 insn->code = BPF_LDX | BPF_PROBE_MEM |
15119 BPF_SIZE((insn)->code);
15120 env->prog->aux->num_exentries++;
15127 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
15128 size = BPF_LDST_BYTES(insn);
15130 /* If the read access is a narrower load of the field,
15131 * convert to a 4/8-byte load, to minimum program type specific
15132 * convert_ctx_access changes. If conversion is successful,
15133 * we will apply proper mask to the result.
15135 is_narrower_load = size < ctx_field_size;
15136 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
15138 if (is_narrower_load) {
15141 if (type == BPF_WRITE) {
15142 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
15147 if (ctx_field_size == 4)
15149 else if (ctx_field_size == 8)
15150 size_code = BPF_DW;
15152 insn->off = off & ~(size_default - 1);
15153 insn->code = BPF_LDX | BPF_MEM | size_code;
15157 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
15159 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
15160 (ctx_field_size && !target_size)) {
15161 verbose(env, "bpf verifier is misconfigured\n");
15165 if (is_narrower_load && size < target_size) {
15166 u8 shift = bpf_ctx_narrow_access_offset(
15167 off, size, size_default) * 8;
15168 if (shift && cnt + 1 >= ARRAY_SIZE(insn_buf)) {
15169 verbose(env, "bpf verifier narrow ctx load misconfigured\n");
15172 if (ctx_field_size <= 4) {
15174 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
15177 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
15178 (1 << size * 8) - 1);
15181 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
15184 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
15185 (1ULL << size * 8) - 1);
15189 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
15195 /* keep walking new program and skip insns we just inserted */
15196 env->prog = new_prog;
15197 insn = new_prog->insnsi + i + delta;
15203 static int jit_subprogs(struct bpf_verifier_env *env)
15205 struct bpf_prog *prog = env->prog, **func, *tmp;
15206 int i, j, subprog_start, subprog_end = 0, len, subprog;
15207 struct bpf_map *map_ptr;
15208 struct bpf_insn *insn;
15209 void *old_bpf_func;
15210 int err, num_exentries;
15212 if (env->subprog_cnt <= 1)
15215 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
15216 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn))
15219 /* Upon error here we cannot fall back to interpreter but
15220 * need a hard reject of the program. Thus -EFAULT is
15221 * propagated in any case.
15223 subprog = find_subprog(env, i + insn->imm + 1);
15225 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
15226 i + insn->imm + 1);
15229 /* temporarily remember subprog id inside insn instead of
15230 * aux_data, since next loop will split up all insns into funcs
15232 insn->off = subprog;
15233 /* remember original imm in case JIT fails and fallback
15234 * to interpreter will be needed
15236 env->insn_aux_data[i].call_imm = insn->imm;
15237 /* point imm to __bpf_call_base+1 from JITs point of view */
15239 if (bpf_pseudo_func(insn))
15240 /* jit (e.g. x86_64) may emit fewer instructions
15241 * if it learns a u32 imm is the same as a u64 imm.
15242 * Force a non zero here.
15247 err = bpf_prog_alloc_jited_linfo(prog);
15249 goto out_undo_insn;
15252 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
15254 goto out_undo_insn;
15256 for (i = 0; i < env->subprog_cnt; i++) {
15257 subprog_start = subprog_end;
15258 subprog_end = env->subprog_info[i + 1].start;
15260 len = subprog_end - subprog_start;
15261 /* bpf_prog_run() doesn't call subprogs directly,
15262 * hence main prog stats include the runtime of subprogs.
15263 * subprogs don't have IDs and not reachable via prog_get_next_id
15264 * func[i]->stats will never be accessed and stays NULL
15266 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
15269 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
15270 len * sizeof(struct bpf_insn));
15271 func[i]->type = prog->type;
15272 func[i]->len = len;
15273 if (bpf_prog_calc_tag(func[i]))
15275 func[i]->is_func = 1;
15276 func[i]->aux->func_idx = i;
15277 /* Below members will be freed only at prog->aux */
15278 func[i]->aux->btf = prog->aux->btf;
15279 func[i]->aux->func_info = prog->aux->func_info;
15280 func[i]->aux->func_info_cnt = prog->aux->func_info_cnt;
15281 func[i]->aux->poke_tab = prog->aux->poke_tab;
15282 func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
15284 for (j = 0; j < prog->aux->size_poke_tab; j++) {
15285 struct bpf_jit_poke_descriptor *poke;
15287 poke = &prog->aux->poke_tab[j];
15288 if (poke->insn_idx < subprog_end &&
15289 poke->insn_idx >= subprog_start)
15290 poke->aux = func[i]->aux;
15293 func[i]->aux->name[0] = 'F';
15294 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
15295 func[i]->jit_requested = 1;
15296 func[i]->blinding_requested = prog->blinding_requested;
15297 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
15298 func[i]->aux->kfunc_btf_tab = prog->aux->kfunc_btf_tab;
15299 func[i]->aux->linfo = prog->aux->linfo;
15300 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
15301 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
15302 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
15304 insn = func[i]->insnsi;
15305 for (j = 0; j < func[i]->len; j++, insn++) {
15306 if (BPF_CLASS(insn->code) == BPF_LDX &&
15307 BPF_MODE(insn->code) == BPF_PROBE_MEM)
15310 func[i]->aux->num_exentries = num_exentries;
15311 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
15312 func[i] = bpf_int_jit_compile(func[i]);
15313 if (!func[i]->jited) {
15320 /* at this point all bpf functions were successfully JITed
15321 * now populate all bpf_calls with correct addresses and
15322 * run last pass of JIT
15324 for (i = 0; i < env->subprog_cnt; i++) {
15325 insn = func[i]->insnsi;
15326 for (j = 0; j < func[i]->len; j++, insn++) {
15327 if (bpf_pseudo_func(insn)) {
15328 subprog = insn->off;
15329 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
15330 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
15333 if (!bpf_pseudo_call(insn))
15335 subprog = insn->off;
15336 insn->imm = BPF_CALL_IMM(func[subprog]->bpf_func);
15339 /* we use the aux data to keep a list of the start addresses
15340 * of the JITed images for each function in the program
15342 * for some architectures, such as powerpc64, the imm field
15343 * might not be large enough to hold the offset of the start
15344 * address of the callee's JITed image from __bpf_call_base
15346 * in such cases, we can lookup the start address of a callee
15347 * by using its subprog id, available from the off field of
15348 * the call instruction, as an index for this list
15350 func[i]->aux->func = func;
15351 func[i]->aux->func_cnt = env->subprog_cnt;
15353 for (i = 0; i < env->subprog_cnt; i++) {
15354 old_bpf_func = func[i]->bpf_func;
15355 tmp = bpf_int_jit_compile(func[i]);
15356 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
15357 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
15364 /* finally lock prog and jit images for all functions and
15365 * populate kallsysm
15367 for (i = 0; i < env->subprog_cnt; i++) {
15368 bpf_prog_lock_ro(func[i]);
15369 bpf_prog_kallsyms_add(func[i]);
15372 /* Last step: make now unused interpreter insns from main
15373 * prog consistent for later dump requests, so they can
15374 * later look the same as if they were interpreted only.
15376 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
15377 if (bpf_pseudo_func(insn)) {
15378 insn[0].imm = env->insn_aux_data[i].call_imm;
15379 insn[1].imm = insn->off;
15383 if (!bpf_pseudo_call(insn))
15385 insn->off = env->insn_aux_data[i].call_imm;
15386 subprog = find_subprog(env, i + insn->off + 1);
15387 insn->imm = subprog;
15391 prog->bpf_func = func[0]->bpf_func;
15392 prog->jited_len = func[0]->jited_len;
15393 prog->aux->func = func;
15394 prog->aux->func_cnt = env->subprog_cnt;
15395 bpf_prog_jit_attempt_done(prog);
15398 /* We failed JIT'ing, so at this point we need to unregister poke
15399 * descriptors from subprogs, so that kernel is not attempting to
15400 * patch it anymore as we're freeing the subprog JIT memory.
15402 for (i = 0; i < prog->aux->size_poke_tab; i++) {
15403 map_ptr = prog->aux->poke_tab[i].tail_call.map;
15404 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
15406 /* At this point we're guaranteed that poke descriptors are not
15407 * live anymore. We can just unlink its descriptor table as it's
15408 * released with the main prog.
15410 for (i = 0; i < env->subprog_cnt; i++) {
15413 func[i]->aux->poke_tab = NULL;
15414 bpf_jit_free(func[i]);
15418 /* cleanup main prog to be interpreted */
15419 prog->jit_requested = 0;
15420 prog->blinding_requested = 0;
15421 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
15422 if (!bpf_pseudo_call(insn))
15425 insn->imm = env->insn_aux_data[i].call_imm;
15427 bpf_prog_jit_attempt_done(prog);
15431 static int fixup_call_args(struct bpf_verifier_env *env)
15433 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
15434 struct bpf_prog *prog = env->prog;
15435 struct bpf_insn *insn = prog->insnsi;
15436 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
15441 if (env->prog->jit_requested &&
15442 !bpf_prog_is_dev_bound(env->prog->aux)) {
15443 err = jit_subprogs(env);
15446 if (err == -EFAULT)
15449 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
15450 if (has_kfunc_call) {
15451 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
15454 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
15455 /* When JIT fails the progs with bpf2bpf calls and tail_calls
15456 * have to be rejected, since interpreter doesn't support them yet.
15458 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
15461 for (i = 0; i < prog->len; i++, insn++) {
15462 if (bpf_pseudo_func(insn)) {
15463 /* When JIT fails the progs with callback calls
15464 * have to be rejected, since interpreter doesn't support them yet.
15466 verbose(env, "callbacks are not allowed in non-JITed programs\n");
15470 if (!bpf_pseudo_call(insn))
15472 depth = get_callee_stack_depth(env, insn, i);
15475 bpf_patch_call_args(insn, depth);
15482 static int fixup_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
15483 struct bpf_insn *insn_buf, int insn_idx, int *cnt)
15485 const struct bpf_kfunc_desc *desc;
15488 verbose(env, "invalid kernel function call not eliminated in verifier pass\n");
15492 /* insn->imm has the btf func_id. Replace it with
15493 * an address (relative to __bpf_call_base).
15495 desc = find_kfunc_desc(env->prog, insn->imm, insn->off);
15497 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
15503 insn->imm = desc->imm;
15506 if (desc->func_id == special_kfunc_list[KF_bpf_obj_new_impl]) {
15507 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
15508 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
15509 u64 obj_new_size = env->insn_aux_data[insn_idx].obj_new_size;
15511 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_1, obj_new_size);
15512 insn_buf[1] = addr[0];
15513 insn_buf[2] = addr[1];
15514 insn_buf[3] = *insn;
15516 } else if (desc->func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
15517 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
15518 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
15520 insn_buf[0] = addr[0];
15521 insn_buf[1] = addr[1];
15522 insn_buf[2] = *insn;
15524 } else if (desc->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx] ||
15525 desc->func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
15526 insn_buf[0] = BPF_MOV64_REG(BPF_REG_0, BPF_REG_1);
15532 /* Do various post-verification rewrites in a single program pass.
15533 * These rewrites simplify JIT and interpreter implementations.
15535 static int do_misc_fixups(struct bpf_verifier_env *env)
15537 struct bpf_prog *prog = env->prog;
15538 enum bpf_attach_type eatype = prog->expected_attach_type;
15539 enum bpf_prog_type prog_type = resolve_prog_type(prog);
15540 struct bpf_insn *insn = prog->insnsi;
15541 const struct bpf_func_proto *fn;
15542 const int insn_cnt = prog->len;
15543 const struct bpf_map_ops *ops;
15544 struct bpf_insn_aux_data *aux;
15545 struct bpf_insn insn_buf[16];
15546 struct bpf_prog *new_prog;
15547 struct bpf_map *map_ptr;
15548 int i, ret, cnt, delta = 0;
15550 for (i = 0; i < insn_cnt; i++, insn++) {
15551 /* Make divide-by-zero exceptions impossible. */
15552 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
15553 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
15554 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
15555 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
15556 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
15557 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
15558 struct bpf_insn *patchlet;
15559 struct bpf_insn chk_and_div[] = {
15560 /* [R,W]x div 0 -> 0 */
15561 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
15562 BPF_JNE | BPF_K, insn->src_reg,
15564 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
15565 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
15568 struct bpf_insn chk_and_mod[] = {
15569 /* [R,W]x mod 0 -> [R,W]x */
15570 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
15571 BPF_JEQ | BPF_K, insn->src_reg,
15572 0, 1 + (is64 ? 0 : 1), 0),
15574 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
15575 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
15578 patchlet = isdiv ? chk_and_div : chk_and_mod;
15579 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
15580 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
15582 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
15587 env->prog = prog = new_prog;
15588 insn = new_prog->insnsi + i + delta;
15592 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
15593 if (BPF_CLASS(insn->code) == BPF_LD &&
15594 (BPF_MODE(insn->code) == BPF_ABS ||
15595 BPF_MODE(insn->code) == BPF_IND)) {
15596 cnt = env->ops->gen_ld_abs(insn, insn_buf);
15597 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
15598 verbose(env, "bpf verifier is misconfigured\n");
15602 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
15607 env->prog = prog = new_prog;
15608 insn = new_prog->insnsi + i + delta;
15612 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
15613 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
15614 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
15615 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
15616 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
15617 struct bpf_insn *patch = &insn_buf[0];
15618 bool issrc, isneg, isimm;
15621 aux = &env->insn_aux_data[i + delta];
15622 if (!aux->alu_state ||
15623 aux->alu_state == BPF_ALU_NON_POINTER)
15626 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
15627 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
15628 BPF_ALU_SANITIZE_SRC;
15629 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
15631 off_reg = issrc ? insn->src_reg : insn->dst_reg;
15633 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
15636 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
15637 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
15638 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
15639 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
15640 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
15641 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
15642 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
15645 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
15646 insn->src_reg = BPF_REG_AX;
15648 insn->code = insn->code == code_add ?
15649 code_sub : code_add;
15651 if (issrc && isneg && !isimm)
15652 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
15653 cnt = patch - insn_buf;
15655 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
15660 env->prog = prog = new_prog;
15661 insn = new_prog->insnsi + i + delta;
15665 if (insn->code != (BPF_JMP | BPF_CALL))
15667 if (insn->src_reg == BPF_PSEUDO_CALL)
15669 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
15670 ret = fixup_kfunc_call(env, insn, insn_buf, i + delta, &cnt);
15676 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
15681 env->prog = prog = new_prog;
15682 insn = new_prog->insnsi + i + delta;
15686 if (insn->imm == BPF_FUNC_get_route_realm)
15687 prog->dst_needed = 1;
15688 if (insn->imm == BPF_FUNC_get_prandom_u32)
15689 bpf_user_rnd_init_once();
15690 if (insn->imm == BPF_FUNC_override_return)
15691 prog->kprobe_override = 1;
15692 if (insn->imm == BPF_FUNC_tail_call) {
15693 /* If we tail call into other programs, we
15694 * cannot make any assumptions since they can
15695 * be replaced dynamically during runtime in
15696 * the program array.
15698 prog->cb_access = 1;
15699 if (!allow_tail_call_in_subprogs(env))
15700 prog->aux->stack_depth = MAX_BPF_STACK;
15701 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
15703 /* mark bpf_tail_call as different opcode to avoid
15704 * conditional branch in the interpreter for every normal
15705 * call and to prevent accidental JITing by JIT compiler
15706 * that doesn't support bpf_tail_call yet
15709 insn->code = BPF_JMP | BPF_TAIL_CALL;
15711 aux = &env->insn_aux_data[i + delta];
15712 if (env->bpf_capable && !prog->blinding_requested &&
15713 prog->jit_requested &&
15714 !bpf_map_key_poisoned(aux) &&
15715 !bpf_map_ptr_poisoned(aux) &&
15716 !bpf_map_ptr_unpriv(aux)) {
15717 struct bpf_jit_poke_descriptor desc = {
15718 .reason = BPF_POKE_REASON_TAIL_CALL,
15719 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
15720 .tail_call.key = bpf_map_key_immediate(aux),
15721 .insn_idx = i + delta,
15724 ret = bpf_jit_add_poke_descriptor(prog, &desc);
15726 verbose(env, "adding tail call poke descriptor failed\n");
15730 insn->imm = ret + 1;
15734 if (!bpf_map_ptr_unpriv(aux))
15737 /* instead of changing every JIT dealing with tail_call
15738 * emit two extra insns:
15739 * if (index >= max_entries) goto out;
15740 * index &= array->index_mask;
15741 * to avoid out-of-bounds cpu speculation
15743 if (bpf_map_ptr_poisoned(aux)) {
15744 verbose(env, "tail_call abusing map_ptr\n");
15748 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
15749 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
15750 map_ptr->max_entries, 2);
15751 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
15752 container_of(map_ptr,
15755 insn_buf[2] = *insn;
15757 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
15762 env->prog = prog = new_prog;
15763 insn = new_prog->insnsi + i + delta;
15767 if (insn->imm == BPF_FUNC_timer_set_callback) {
15768 /* The verifier will process callback_fn as many times as necessary
15769 * with different maps and the register states prepared by
15770 * set_timer_callback_state will be accurate.
15772 * The following use case is valid:
15773 * map1 is shared by prog1, prog2, prog3.
15774 * prog1 calls bpf_timer_init for some map1 elements
15775 * prog2 calls bpf_timer_set_callback for some map1 elements.
15776 * Those that were not bpf_timer_init-ed will return -EINVAL.
15777 * prog3 calls bpf_timer_start for some map1 elements.
15778 * Those that were not both bpf_timer_init-ed and
15779 * bpf_timer_set_callback-ed will return -EINVAL.
15781 struct bpf_insn ld_addrs[2] = {
15782 BPF_LD_IMM64(BPF_REG_3, (long)prog->aux),
15785 insn_buf[0] = ld_addrs[0];
15786 insn_buf[1] = ld_addrs[1];
15787 insn_buf[2] = *insn;
15790 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
15795 env->prog = prog = new_prog;
15796 insn = new_prog->insnsi + i + delta;
15797 goto patch_call_imm;
15800 if (is_storage_get_function(insn->imm)) {
15801 if (!env->prog->aux->sleepable ||
15802 env->insn_aux_data[i + delta].storage_get_func_atomic)
15803 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_ATOMIC);
15805 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_KERNEL);
15806 insn_buf[1] = *insn;
15809 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
15814 env->prog = prog = new_prog;
15815 insn = new_prog->insnsi + i + delta;
15816 goto patch_call_imm;
15819 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
15820 * and other inlining handlers are currently limited to 64 bit
15823 if (prog->jit_requested && BITS_PER_LONG == 64 &&
15824 (insn->imm == BPF_FUNC_map_lookup_elem ||
15825 insn->imm == BPF_FUNC_map_update_elem ||
15826 insn->imm == BPF_FUNC_map_delete_elem ||
15827 insn->imm == BPF_FUNC_map_push_elem ||
15828 insn->imm == BPF_FUNC_map_pop_elem ||
15829 insn->imm == BPF_FUNC_map_peek_elem ||
15830 insn->imm == BPF_FUNC_redirect_map ||
15831 insn->imm == BPF_FUNC_for_each_map_elem ||
15832 insn->imm == BPF_FUNC_map_lookup_percpu_elem)) {
15833 aux = &env->insn_aux_data[i + delta];
15834 if (bpf_map_ptr_poisoned(aux))
15835 goto patch_call_imm;
15837 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
15838 ops = map_ptr->ops;
15839 if (insn->imm == BPF_FUNC_map_lookup_elem &&
15840 ops->map_gen_lookup) {
15841 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
15842 if (cnt == -EOPNOTSUPP)
15843 goto patch_map_ops_generic;
15844 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
15845 verbose(env, "bpf verifier is misconfigured\n");
15849 new_prog = bpf_patch_insn_data(env, i + delta,
15855 env->prog = prog = new_prog;
15856 insn = new_prog->insnsi + i + delta;
15860 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
15861 (void *(*)(struct bpf_map *map, void *key))NULL));
15862 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
15863 (int (*)(struct bpf_map *map, void *key))NULL));
15864 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
15865 (int (*)(struct bpf_map *map, void *key, void *value,
15867 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
15868 (int (*)(struct bpf_map *map, void *value,
15870 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
15871 (int (*)(struct bpf_map *map, void *value))NULL));
15872 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
15873 (int (*)(struct bpf_map *map, void *value))NULL));
15874 BUILD_BUG_ON(!__same_type(ops->map_redirect,
15875 (int (*)(struct bpf_map *map, u64 index, u64 flags))NULL));
15876 BUILD_BUG_ON(!__same_type(ops->map_for_each_callback,
15877 (int (*)(struct bpf_map *map,
15878 bpf_callback_t callback_fn,
15879 void *callback_ctx,
15881 BUILD_BUG_ON(!__same_type(ops->map_lookup_percpu_elem,
15882 (void *(*)(struct bpf_map *map, void *key, u32 cpu))NULL));
15884 patch_map_ops_generic:
15885 switch (insn->imm) {
15886 case BPF_FUNC_map_lookup_elem:
15887 insn->imm = BPF_CALL_IMM(ops->map_lookup_elem);
15889 case BPF_FUNC_map_update_elem:
15890 insn->imm = BPF_CALL_IMM(ops->map_update_elem);
15892 case BPF_FUNC_map_delete_elem:
15893 insn->imm = BPF_CALL_IMM(ops->map_delete_elem);
15895 case BPF_FUNC_map_push_elem:
15896 insn->imm = BPF_CALL_IMM(ops->map_push_elem);
15898 case BPF_FUNC_map_pop_elem:
15899 insn->imm = BPF_CALL_IMM(ops->map_pop_elem);
15901 case BPF_FUNC_map_peek_elem:
15902 insn->imm = BPF_CALL_IMM(ops->map_peek_elem);
15904 case BPF_FUNC_redirect_map:
15905 insn->imm = BPF_CALL_IMM(ops->map_redirect);
15907 case BPF_FUNC_for_each_map_elem:
15908 insn->imm = BPF_CALL_IMM(ops->map_for_each_callback);
15910 case BPF_FUNC_map_lookup_percpu_elem:
15911 insn->imm = BPF_CALL_IMM(ops->map_lookup_percpu_elem);
15915 goto patch_call_imm;
15918 /* Implement bpf_jiffies64 inline. */
15919 if (prog->jit_requested && BITS_PER_LONG == 64 &&
15920 insn->imm == BPF_FUNC_jiffies64) {
15921 struct bpf_insn ld_jiffies_addr[2] = {
15922 BPF_LD_IMM64(BPF_REG_0,
15923 (unsigned long)&jiffies),
15926 insn_buf[0] = ld_jiffies_addr[0];
15927 insn_buf[1] = ld_jiffies_addr[1];
15928 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
15932 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
15938 env->prog = prog = new_prog;
15939 insn = new_prog->insnsi + i + delta;
15943 /* Implement bpf_get_func_arg inline. */
15944 if (prog_type == BPF_PROG_TYPE_TRACING &&
15945 insn->imm == BPF_FUNC_get_func_arg) {
15946 /* Load nr_args from ctx - 8 */
15947 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
15948 insn_buf[1] = BPF_JMP32_REG(BPF_JGE, BPF_REG_2, BPF_REG_0, 6);
15949 insn_buf[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_2, 3);
15950 insn_buf[3] = BPF_ALU64_REG(BPF_ADD, BPF_REG_2, BPF_REG_1);
15951 insn_buf[4] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_2, 0);
15952 insn_buf[5] = BPF_STX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
15953 insn_buf[6] = BPF_MOV64_IMM(BPF_REG_0, 0);
15954 insn_buf[7] = BPF_JMP_A(1);
15955 insn_buf[8] = BPF_MOV64_IMM(BPF_REG_0, -EINVAL);
15958 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
15963 env->prog = prog = new_prog;
15964 insn = new_prog->insnsi + i + delta;
15968 /* Implement bpf_get_func_ret inline. */
15969 if (prog_type == BPF_PROG_TYPE_TRACING &&
15970 insn->imm == BPF_FUNC_get_func_ret) {
15971 if (eatype == BPF_TRACE_FEXIT ||
15972 eatype == BPF_MODIFY_RETURN) {
15973 /* Load nr_args from ctx - 8 */
15974 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
15975 insn_buf[1] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_0, 3);
15976 insn_buf[2] = BPF_ALU64_REG(BPF_ADD, BPF_REG_0, BPF_REG_1);
15977 insn_buf[3] = BPF_LDX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
15978 insn_buf[4] = BPF_STX_MEM(BPF_DW, BPF_REG_2, BPF_REG_3, 0);
15979 insn_buf[5] = BPF_MOV64_IMM(BPF_REG_0, 0);
15982 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, -EOPNOTSUPP);
15986 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
15991 env->prog = prog = new_prog;
15992 insn = new_prog->insnsi + i + delta;
15996 /* Implement get_func_arg_cnt inline. */
15997 if (prog_type == BPF_PROG_TYPE_TRACING &&
15998 insn->imm == BPF_FUNC_get_func_arg_cnt) {
15999 /* Load nr_args from ctx - 8 */
16000 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
16002 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
16006 env->prog = prog = new_prog;
16007 insn = new_prog->insnsi + i + delta;
16011 /* Implement bpf_get_func_ip inline. */
16012 if (prog_type == BPF_PROG_TYPE_TRACING &&
16013 insn->imm == BPF_FUNC_get_func_ip) {
16014 /* Load IP address from ctx - 16 */
16015 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -16);
16017 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
16021 env->prog = prog = new_prog;
16022 insn = new_prog->insnsi + i + delta;
16027 fn = env->ops->get_func_proto(insn->imm, env->prog);
16028 /* all functions that have prototype and verifier allowed
16029 * programs to call them, must be real in-kernel functions
16033 "kernel subsystem misconfigured func %s#%d\n",
16034 func_id_name(insn->imm), insn->imm);
16037 insn->imm = fn->func - __bpf_call_base;
16040 /* Since poke tab is now finalized, publish aux to tracker. */
16041 for (i = 0; i < prog->aux->size_poke_tab; i++) {
16042 map_ptr = prog->aux->poke_tab[i].tail_call.map;
16043 if (!map_ptr->ops->map_poke_track ||
16044 !map_ptr->ops->map_poke_untrack ||
16045 !map_ptr->ops->map_poke_run) {
16046 verbose(env, "bpf verifier is misconfigured\n");
16050 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
16052 verbose(env, "tracking tail call prog failed\n");
16057 sort_kfunc_descs_by_imm(env->prog);
16062 static struct bpf_prog *inline_bpf_loop(struct bpf_verifier_env *env,
16065 u32 callback_subprogno,
16068 s32 r6_offset = stack_base + 0 * BPF_REG_SIZE;
16069 s32 r7_offset = stack_base + 1 * BPF_REG_SIZE;
16070 s32 r8_offset = stack_base + 2 * BPF_REG_SIZE;
16071 int reg_loop_max = BPF_REG_6;
16072 int reg_loop_cnt = BPF_REG_7;
16073 int reg_loop_ctx = BPF_REG_8;
16075 struct bpf_prog *new_prog;
16076 u32 callback_start;
16077 u32 call_insn_offset;
16078 s32 callback_offset;
16080 /* This represents an inlined version of bpf_iter.c:bpf_loop,
16081 * be careful to modify this code in sync.
16083 struct bpf_insn insn_buf[] = {
16084 /* Return error and jump to the end of the patch if
16085 * expected number of iterations is too big.
16087 BPF_JMP_IMM(BPF_JLE, BPF_REG_1, BPF_MAX_LOOPS, 2),
16088 BPF_MOV32_IMM(BPF_REG_0, -E2BIG),
16089 BPF_JMP_IMM(BPF_JA, 0, 0, 16),
16090 /* spill R6, R7, R8 to use these as loop vars */
16091 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_6, r6_offset),
16092 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_7, r7_offset),
16093 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_8, r8_offset),
16094 /* initialize loop vars */
16095 BPF_MOV64_REG(reg_loop_max, BPF_REG_1),
16096 BPF_MOV32_IMM(reg_loop_cnt, 0),
16097 BPF_MOV64_REG(reg_loop_ctx, BPF_REG_3),
16099 * if reg_loop_cnt >= reg_loop_max skip the loop body
16101 BPF_JMP_REG(BPF_JGE, reg_loop_cnt, reg_loop_max, 5),
16103 * correct callback offset would be set after patching
16105 BPF_MOV64_REG(BPF_REG_1, reg_loop_cnt),
16106 BPF_MOV64_REG(BPF_REG_2, reg_loop_ctx),
16108 /* increment loop counter */
16109 BPF_ALU64_IMM(BPF_ADD, reg_loop_cnt, 1),
16110 /* jump to loop header if callback returned 0 */
16111 BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, -6),
16112 /* return value of bpf_loop,
16113 * set R0 to the number of iterations
16115 BPF_MOV64_REG(BPF_REG_0, reg_loop_cnt),
16116 /* restore original values of R6, R7, R8 */
16117 BPF_LDX_MEM(BPF_DW, BPF_REG_6, BPF_REG_10, r6_offset),
16118 BPF_LDX_MEM(BPF_DW, BPF_REG_7, BPF_REG_10, r7_offset),
16119 BPF_LDX_MEM(BPF_DW, BPF_REG_8, BPF_REG_10, r8_offset),
16122 *cnt = ARRAY_SIZE(insn_buf);
16123 new_prog = bpf_patch_insn_data(env, position, insn_buf, *cnt);
16127 /* callback start is known only after patching */
16128 callback_start = env->subprog_info[callback_subprogno].start;
16129 /* Note: insn_buf[12] is an offset of BPF_CALL_REL instruction */
16130 call_insn_offset = position + 12;
16131 callback_offset = callback_start - call_insn_offset - 1;
16132 new_prog->insnsi[call_insn_offset].imm = callback_offset;
16137 static bool is_bpf_loop_call(struct bpf_insn *insn)
16139 return insn->code == (BPF_JMP | BPF_CALL) &&
16140 insn->src_reg == 0 &&
16141 insn->imm == BPF_FUNC_loop;
16144 /* For all sub-programs in the program (including main) check
16145 * insn_aux_data to see if there are bpf_loop calls that require
16146 * inlining. If such calls are found the calls are replaced with a
16147 * sequence of instructions produced by `inline_bpf_loop` function and
16148 * subprog stack_depth is increased by the size of 3 registers.
16149 * This stack space is used to spill values of the R6, R7, R8. These
16150 * registers are used to store the loop bound, counter and context
16153 static int optimize_bpf_loop(struct bpf_verifier_env *env)
16155 struct bpf_subprog_info *subprogs = env->subprog_info;
16156 int i, cur_subprog = 0, cnt, delta = 0;
16157 struct bpf_insn *insn = env->prog->insnsi;
16158 int insn_cnt = env->prog->len;
16159 u16 stack_depth = subprogs[cur_subprog].stack_depth;
16160 u16 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
16161 u16 stack_depth_extra = 0;
16163 for (i = 0; i < insn_cnt; i++, insn++) {
16164 struct bpf_loop_inline_state *inline_state =
16165 &env->insn_aux_data[i + delta].loop_inline_state;
16167 if (is_bpf_loop_call(insn) && inline_state->fit_for_inline) {
16168 struct bpf_prog *new_prog;
16170 stack_depth_extra = BPF_REG_SIZE * 3 + stack_depth_roundup;
16171 new_prog = inline_bpf_loop(env,
16173 -(stack_depth + stack_depth_extra),
16174 inline_state->callback_subprogno,
16180 env->prog = new_prog;
16181 insn = new_prog->insnsi + i + delta;
16184 if (subprogs[cur_subprog + 1].start == i + delta + 1) {
16185 subprogs[cur_subprog].stack_depth += stack_depth_extra;
16187 stack_depth = subprogs[cur_subprog].stack_depth;
16188 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
16189 stack_depth_extra = 0;
16193 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
16198 static void free_states(struct bpf_verifier_env *env)
16200 struct bpf_verifier_state_list *sl, *sln;
16203 sl = env->free_list;
16206 free_verifier_state(&sl->state, false);
16210 env->free_list = NULL;
16212 if (!env->explored_states)
16215 for (i = 0; i < state_htab_size(env); i++) {
16216 sl = env->explored_states[i];
16220 free_verifier_state(&sl->state, false);
16224 env->explored_states[i] = NULL;
16228 static int do_check_common(struct bpf_verifier_env *env, int subprog)
16230 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
16231 struct bpf_verifier_state *state;
16232 struct bpf_reg_state *regs;
16235 env->prev_linfo = NULL;
16238 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
16241 state->curframe = 0;
16242 state->speculative = false;
16243 state->branches = 1;
16244 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
16245 if (!state->frame[0]) {
16249 env->cur_state = state;
16250 init_func_state(env, state->frame[0],
16251 BPF_MAIN_FUNC /* callsite */,
16254 state->first_insn_idx = env->subprog_info[subprog].start;
16255 state->last_insn_idx = -1;
16257 regs = state->frame[state->curframe]->regs;
16258 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
16259 ret = btf_prepare_func_args(env, subprog, regs);
16262 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
16263 if (regs[i].type == PTR_TO_CTX)
16264 mark_reg_known_zero(env, regs, i);
16265 else if (regs[i].type == SCALAR_VALUE)
16266 mark_reg_unknown(env, regs, i);
16267 else if (base_type(regs[i].type) == PTR_TO_MEM) {
16268 const u32 mem_size = regs[i].mem_size;
16270 mark_reg_known_zero(env, regs, i);
16271 regs[i].mem_size = mem_size;
16272 regs[i].id = ++env->id_gen;
16276 /* 1st arg to a function */
16277 regs[BPF_REG_1].type = PTR_TO_CTX;
16278 mark_reg_known_zero(env, regs, BPF_REG_1);
16279 ret = btf_check_subprog_arg_match(env, subprog, regs);
16280 if (ret == -EFAULT)
16281 /* unlikely verifier bug. abort.
16282 * ret == 0 and ret < 0 are sadly acceptable for
16283 * main() function due to backward compatibility.
16284 * Like socket filter program may be written as:
16285 * int bpf_prog(struct pt_regs *ctx)
16286 * and never dereference that ctx in the program.
16287 * 'struct pt_regs' is a type mismatch for socket
16288 * filter that should be using 'struct __sk_buff'.
16293 ret = do_check(env);
16295 /* check for NULL is necessary, since cur_state can be freed inside
16296 * do_check() under memory pressure.
16298 if (env->cur_state) {
16299 free_verifier_state(env->cur_state, true);
16300 env->cur_state = NULL;
16302 while (!pop_stack(env, NULL, NULL, false));
16303 if (!ret && pop_log)
16304 bpf_vlog_reset(&env->log, 0);
16309 /* Verify all global functions in a BPF program one by one based on their BTF.
16310 * All global functions must pass verification. Otherwise the whole program is rejected.
16321 * foo() will be verified first for R1=any_scalar_value. During verification it
16322 * will be assumed that bar() already verified successfully and call to bar()
16323 * from foo() will be checked for type match only. Later bar() will be verified
16324 * independently to check that it's safe for R1=any_scalar_value.
16326 static int do_check_subprogs(struct bpf_verifier_env *env)
16328 struct bpf_prog_aux *aux = env->prog->aux;
16331 if (!aux->func_info)
16334 for (i = 1; i < env->subprog_cnt; i++) {
16335 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
16337 env->insn_idx = env->subprog_info[i].start;
16338 WARN_ON_ONCE(env->insn_idx == 0);
16339 ret = do_check_common(env, i);
16342 } else if (env->log.level & BPF_LOG_LEVEL) {
16344 "Func#%d is safe for any args that match its prototype\n",
16351 static int do_check_main(struct bpf_verifier_env *env)
16356 ret = do_check_common(env, 0);
16358 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
16363 static void print_verification_stats(struct bpf_verifier_env *env)
16367 if (env->log.level & BPF_LOG_STATS) {
16368 verbose(env, "verification time %lld usec\n",
16369 div_u64(env->verification_time, 1000));
16370 verbose(env, "stack depth ");
16371 for (i = 0; i < env->subprog_cnt; i++) {
16372 u32 depth = env->subprog_info[i].stack_depth;
16374 verbose(env, "%d", depth);
16375 if (i + 1 < env->subprog_cnt)
16378 verbose(env, "\n");
16380 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
16381 "total_states %d peak_states %d mark_read %d\n",
16382 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
16383 env->max_states_per_insn, env->total_states,
16384 env->peak_states, env->longest_mark_read_walk);
16387 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
16389 const struct btf_type *t, *func_proto;
16390 const struct bpf_struct_ops *st_ops;
16391 const struct btf_member *member;
16392 struct bpf_prog *prog = env->prog;
16393 u32 btf_id, member_idx;
16396 if (!prog->gpl_compatible) {
16397 verbose(env, "struct ops programs must have a GPL compatible license\n");
16401 btf_id = prog->aux->attach_btf_id;
16402 st_ops = bpf_struct_ops_find(btf_id);
16404 verbose(env, "attach_btf_id %u is not a supported struct\n",
16410 member_idx = prog->expected_attach_type;
16411 if (member_idx >= btf_type_vlen(t)) {
16412 verbose(env, "attach to invalid member idx %u of struct %s\n",
16413 member_idx, st_ops->name);
16417 member = &btf_type_member(t)[member_idx];
16418 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
16419 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
16422 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
16423 mname, member_idx, st_ops->name);
16427 if (st_ops->check_member) {
16428 int err = st_ops->check_member(t, member);
16431 verbose(env, "attach to unsupported member %s of struct %s\n",
16432 mname, st_ops->name);
16437 prog->aux->attach_func_proto = func_proto;
16438 prog->aux->attach_func_name = mname;
16439 env->ops = st_ops->verifier_ops;
16443 #define SECURITY_PREFIX "security_"
16445 static int check_attach_modify_return(unsigned long addr, const char *func_name)
16447 if (within_error_injection_list(addr) ||
16448 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
16454 /* list of non-sleepable functions that are otherwise on
16455 * ALLOW_ERROR_INJECTION list
16457 BTF_SET_START(btf_non_sleepable_error_inject)
16458 /* Three functions below can be called from sleepable and non-sleepable context.
16459 * Assume non-sleepable from bpf safety point of view.
16461 BTF_ID(func, __filemap_add_folio)
16462 BTF_ID(func, should_fail_alloc_page)
16463 BTF_ID(func, should_failslab)
16464 BTF_SET_END(btf_non_sleepable_error_inject)
16466 static int check_non_sleepable_error_inject(u32 btf_id)
16468 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
16471 int bpf_check_attach_target(struct bpf_verifier_log *log,
16472 const struct bpf_prog *prog,
16473 const struct bpf_prog *tgt_prog,
16475 struct bpf_attach_target_info *tgt_info)
16477 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
16478 const char prefix[] = "btf_trace_";
16479 int ret = 0, subprog = -1, i;
16480 const struct btf_type *t;
16481 bool conservative = true;
16487 bpf_log(log, "Tracing programs must provide btf_id\n");
16490 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
16493 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
16496 t = btf_type_by_id(btf, btf_id);
16498 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
16501 tname = btf_name_by_offset(btf, t->name_off);
16503 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
16507 struct bpf_prog_aux *aux = tgt_prog->aux;
16509 for (i = 0; i < aux->func_info_cnt; i++)
16510 if (aux->func_info[i].type_id == btf_id) {
16514 if (subprog == -1) {
16515 bpf_log(log, "Subprog %s doesn't exist\n", tname);
16518 conservative = aux->func_info_aux[subprog].unreliable;
16519 if (prog_extension) {
16520 if (conservative) {
16522 "Cannot replace static functions\n");
16525 if (!prog->jit_requested) {
16527 "Extension programs should be JITed\n");
16531 if (!tgt_prog->jited) {
16532 bpf_log(log, "Can attach to only JITed progs\n");
16535 if (tgt_prog->type == prog->type) {
16536 /* Cannot fentry/fexit another fentry/fexit program.
16537 * Cannot attach program extension to another extension.
16538 * It's ok to attach fentry/fexit to extension program.
16540 bpf_log(log, "Cannot recursively attach\n");
16543 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
16545 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
16546 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
16547 /* Program extensions can extend all program types
16548 * except fentry/fexit. The reason is the following.
16549 * The fentry/fexit programs are used for performance
16550 * analysis, stats and can be attached to any program
16551 * type except themselves. When extension program is
16552 * replacing XDP function it is necessary to allow
16553 * performance analysis of all functions. Both original
16554 * XDP program and its program extension. Hence
16555 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
16556 * allowed. If extending of fentry/fexit was allowed it
16557 * would be possible to create long call chain
16558 * fentry->extension->fentry->extension beyond
16559 * reasonable stack size. Hence extending fentry is not
16562 bpf_log(log, "Cannot extend fentry/fexit\n");
16566 if (prog_extension) {
16567 bpf_log(log, "Cannot replace kernel functions\n");
16572 switch (prog->expected_attach_type) {
16573 case BPF_TRACE_RAW_TP:
16576 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
16579 if (!btf_type_is_typedef(t)) {
16580 bpf_log(log, "attach_btf_id %u is not a typedef\n",
16584 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
16585 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
16589 tname += sizeof(prefix) - 1;
16590 t = btf_type_by_id(btf, t->type);
16591 if (!btf_type_is_ptr(t))
16592 /* should never happen in valid vmlinux build */
16594 t = btf_type_by_id(btf, t->type);
16595 if (!btf_type_is_func_proto(t))
16596 /* should never happen in valid vmlinux build */
16600 case BPF_TRACE_ITER:
16601 if (!btf_type_is_func(t)) {
16602 bpf_log(log, "attach_btf_id %u is not a function\n",
16606 t = btf_type_by_id(btf, t->type);
16607 if (!btf_type_is_func_proto(t))
16609 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
16614 if (!prog_extension)
16617 case BPF_MODIFY_RETURN:
16619 case BPF_LSM_CGROUP:
16620 case BPF_TRACE_FENTRY:
16621 case BPF_TRACE_FEXIT:
16622 if (!btf_type_is_func(t)) {
16623 bpf_log(log, "attach_btf_id %u is not a function\n",
16627 if (prog_extension &&
16628 btf_check_type_match(log, prog, btf, t))
16630 t = btf_type_by_id(btf, t->type);
16631 if (!btf_type_is_func_proto(t))
16634 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
16635 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
16636 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
16639 if (tgt_prog && conservative)
16642 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
16648 addr = (long) tgt_prog->bpf_func;
16650 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
16652 addr = kallsyms_lookup_name(tname);
16655 "The address of function %s cannot be found\n",
16661 if (prog->aux->sleepable) {
16663 switch (prog->type) {
16664 case BPF_PROG_TYPE_TRACING:
16666 /* fentry/fexit/fmod_ret progs can be sleepable if they are
16667 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
16669 if (!check_non_sleepable_error_inject(btf_id) &&
16670 within_error_injection_list(addr))
16672 /* fentry/fexit/fmod_ret progs can also be sleepable if they are
16673 * in the fmodret id set with the KF_SLEEPABLE flag.
16676 u32 *flags = btf_kfunc_is_modify_return(btf, btf_id);
16678 if (flags && (*flags & KF_SLEEPABLE))
16682 case BPF_PROG_TYPE_LSM:
16683 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
16684 * Only some of them are sleepable.
16686 if (bpf_lsm_is_sleepable_hook(btf_id))
16693 bpf_log(log, "%s is not sleepable\n", tname);
16696 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
16698 bpf_log(log, "can't modify return codes of BPF programs\n");
16702 if (btf_kfunc_is_modify_return(btf, btf_id) ||
16703 !check_attach_modify_return(addr, tname))
16706 bpf_log(log, "%s() is not modifiable\n", tname);
16713 tgt_info->tgt_addr = addr;
16714 tgt_info->tgt_name = tname;
16715 tgt_info->tgt_type = t;
16719 BTF_SET_START(btf_id_deny)
16722 BTF_ID(func, migrate_disable)
16723 BTF_ID(func, migrate_enable)
16725 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
16726 BTF_ID(func, rcu_read_unlock_strict)
16728 BTF_SET_END(btf_id_deny)
16730 static int check_attach_btf_id(struct bpf_verifier_env *env)
16732 struct bpf_prog *prog = env->prog;
16733 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
16734 struct bpf_attach_target_info tgt_info = {};
16735 u32 btf_id = prog->aux->attach_btf_id;
16736 struct bpf_trampoline *tr;
16740 if (prog->type == BPF_PROG_TYPE_SYSCALL) {
16741 if (prog->aux->sleepable)
16742 /* attach_btf_id checked to be zero already */
16744 verbose(env, "Syscall programs can only be sleepable\n");
16748 if (prog->aux->sleepable && prog->type != BPF_PROG_TYPE_TRACING &&
16749 prog->type != BPF_PROG_TYPE_LSM && prog->type != BPF_PROG_TYPE_KPROBE) {
16750 verbose(env, "Only fentry/fexit/fmod_ret, lsm, and kprobe/uprobe programs can be sleepable\n");
16754 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
16755 return check_struct_ops_btf_id(env);
16757 if (prog->type != BPF_PROG_TYPE_TRACING &&
16758 prog->type != BPF_PROG_TYPE_LSM &&
16759 prog->type != BPF_PROG_TYPE_EXT)
16762 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
16766 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
16767 /* to make freplace equivalent to their targets, they need to
16768 * inherit env->ops and expected_attach_type for the rest of the
16771 env->ops = bpf_verifier_ops[tgt_prog->type];
16772 prog->expected_attach_type = tgt_prog->expected_attach_type;
16775 /* store info about the attachment target that will be used later */
16776 prog->aux->attach_func_proto = tgt_info.tgt_type;
16777 prog->aux->attach_func_name = tgt_info.tgt_name;
16780 prog->aux->saved_dst_prog_type = tgt_prog->type;
16781 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
16784 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
16785 prog->aux->attach_btf_trace = true;
16787 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
16788 if (!bpf_iter_prog_supported(prog))
16793 if (prog->type == BPF_PROG_TYPE_LSM) {
16794 ret = bpf_lsm_verify_prog(&env->log, prog);
16797 } else if (prog->type == BPF_PROG_TYPE_TRACING &&
16798 btf_id_set_contains(&btf_id_deny, btf_id)) {
16802 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
16803 tr = bpf_trampoline_get(key, &tgt_info);
16807 prog->aux->dst_trampoline = tr;
16811 struct btf *bpf_get_btf_vmlinux(void)
16813 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
16814 mutex_lock(&bpf_verifier_lock);
16816 btf_vmlinux = btf_parse_vmlinux();
16817 mutex_unlock(&bpf_verifier_lock);
16819 return btf_vmlinux;
16822 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr)
16824 u64 start_time = ktime_get_ns();
16825 struct bpf_verifier_env *env;
16826 struct bpf_verifier_log *log;
16827 int i, len, ret = -EINVAL;
16830 /* no program is valid */
16831 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
16834 /* 'struct bpf_verifier_env' can be global, but since it's not small,
16835 * allocate/free it every time bpf_check() is called
16837 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
16842 len = (*prog)->len;
16843 env->insn_aux_data =
16844 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
16846 if (!env->insn_aux_data)
16848 for (i = 0; i < len; i++)
16849 env->insn_aux_data[i].orig_idx = i;
16851 env->ops = bpf_verifier_ops[env->prog->type];
16852 env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
16853 is_priv = bpf_capable();
16855 bpf_get_btf_vmlinux();
16857 /* grab the mutex to protect few globals used by verifier */
16859 mutex_lock(&bpf_verifier_lock);
16861 if (attr->log_level || attr->log_buf || attr->log_size) {
16862 /* user requested verbose verifier output
16863 * and supplied buffer to store the verification trace
16865 log->level = attr->log_level;
16866 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
16867 log->len_total = attr->log_size;
16869 /* log attributes have to be sane */
16870 if (!bpf_verifier_log_attr_valid(log)) {
16876 mark_verifier_state_clean(env);
16878 if (IS_ERR(btf_vmlinux)) {
16879 /* Either gcc or pahole or kernel are broken. */
16880 verbose(env, "in-kernel BTF is malformed\n");
16881 ret = PTR_ERR(btf_vmlinux);
16882 goto skip_full_check;
16885 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
16886 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
16887 env->strict_alignment = true;
16888 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
16889 env->strict_alignment = false;
16891 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
16892 env->allow_uninit_stack = bpf_allow_uninit_stack();
16893 env->bypass_spec_v1 = bpf_bypass_spec_v1();
16894 env->bypass_spec_v4 = bpf_bypass_spec_v4();
16895 env->bpf_capable = bpf_capable();
16896 env->rcu_tag_supported = btf_vmlinux &&
16897 btf_find_by_name_kind(btf_vmlinux, "rcu", BTF_KIND_TYPE_TAG) > 0;
16900 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
16902 env->explored_states = kvcalloc(state_htab_size(env),
16903 sizeof(struct bpf_verifier_state_list *),
16906 if (!env->explored_states)
16907 goto skip_full_check;
16909 ret = add_subprog_and_kfunc(env);
16911 goto skip_full_check;
16913 ret = check_subprogs(env);
16915 goto skip_full_check;
16917 ret = check_btf_info(env, attr, uattr);
16919 goto skip_full_check;
16921 ret = check_attach_btf_id(env);
16923 goto skip_full_check;
16925 ret = resolve_pseudo_ldimm64(env);
16927 goto skip_full_check;
16929 if (bpf_prog_is_dev_bound(env->prog->aux)) {
16930 ret = bpf_prog_offload_verifier_prep(env->prog);
16932 goto skip_full_check;
16935 ret = check_cfg(env);
16937 goto skip_full_check;
16939 ret = do_check_subprogs(env);
16940 ret = ret ?: do_check_main(env);
16942 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
16943 ret = bpf_prog_offload_finalize(env);
16946 kvfree(env->explored_states);
16949 ret = check_max_stack_depth(env);
16951 /* instruction rewrites happen after this point */
16953 ret = optimize_bpf_loop(env);
16957 opt_hard_wire_dead_code_branches(env);
16959 ret = opt_remove_dead_code(env);
16961 ret = opt_remove_nops(env);
16964 sanitize_dead_code(env);
16968 /* program is valid, convert *(u32*)(ctx + off) accesses */
16969 ret = convert_ctx_accesses(env);
16972 ret = do_misc_fixups(env);
16974 /* do 32-bit optimization after insn patching has done so those patched
16975 * insns could be handled correctly.
16977 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
16978 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
16979 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
16984 ret = fixup_call_args(env);
16986 env->verification_time = ktime_get_ns() - start_time;
16987 print_verification_stats(env);
16988 env->prog->aux->verified_insns = env->insn_processed;
16990 if (log->level && bpf_verifier_log_full(log))
16992 if (log->level && !log->ubuf) {
16994 goto err_release_maps;
16998 goto err_release_maps;
17000 if (env->used_map_cnt) {
17001 /* if program passed verifier, update used_maps in bpf_prog_info */
17002 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
17003 sizeof(env->used_maps[0]),
17006 if (!env->prog->aux->used_maps) {
17008 goto err_release_maps;
17011 memcpy(env->prog->aux->used_maps, env->used_maps,
17012 sizeof(env->used_maps[0]) * env->used_map_cnt);
17013 env->prog->aux->used_map_cnt = env->used_map_cnt;
17015 if (env->used_btf_cnt) {
17016 /* if program passed verifier, update used_btfs in bpf_prog_aux */
17017 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
17018 sizeof(env->used_btfs[0]),
17020 if (!env->prog->aux->used_btfs) {
17022 goto err_release_maps;
17025 memcpy(env->prog->aux->used_btfs, env->used_btfs,
17026 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
17027 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
17029 if (env->used_map_cnt || env->used_btf_cnt) {
17030 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
17031 * bpf_ld_imm64 instructions
17033 convert_pseudo_ld_imm64(env);
17036 adjust_btf_func(env);
17039 if (!env->prog->aux->used_maps)
17040 /* if we didn't copy map pointers into bpf_prog_info, release
17041 * them now. Otherwise free_used_maps() will release them.
17044 if (!env->prog->aux->used_btfs)
17047 /* extension progs temporarily inherit the attach_type of their targets
17048 for verification purposes, so set it back to zero before returning
17050 if (env->prog->type == BPF_PROG_TYPE_EXT)
17051 env->prog->expected_attach_type = 0;
17056 mutex_unlock(&bpf_verifier_lock);
17057 vfree(env->insn_aux_data);