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 id, 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));
1408 ___mark_reg_known(reg, imm);
1411 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1413 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1414 reg->s32_min_value = (s32)imm;
1415 reg->s32_max_value = (s32)imm;
1416 reg->u32_min_value = (u32)imm;
1417 reg->u32_max_value = (u32)imm;
1420 /* Mark the 'variable offset' part of a register as zero. This should be
1421 * used only on registers holding a pointer type.
1423 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1425 __mark_reg_known(reg, 0);
1428 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1430 __mark_reg_known(reg, 0);
1431 reg->type = SCALAR_VALUE;
1434 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1435 struct bpf_reg_state *regs, u32 regno)
1437 if (WARN_ON(regno >= MAX_BPF_REG)) {
1438 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1439 /* Something bad happened, let's kill all regs */
1440 for (regno = 0; regno < MAX_BPF_REG; regno++)
1441 __mark_reg_not_init(env, regs + regno);
1444 __mark_reg_known_zero(regs + regno);
1447 static void __mark_dynptr_reg(struct bpf_reg_state *reg, enum bpf_dynptr_type type,
1450 /* reg->type has no meaning for STACK_DYNPTR, but when we set reg for
1451 * callback arguments, it does need to be CONST_PTR_TO_DYNPTR, so simply
1452 * set it unconditionally as it is ignored for STACK_DYNPTR anyway.
1454 __mark_reg_known_zero(reg);
1455 reg->type = CONST_PTR_TO_DYNPTR;
1456 reg->dynptr.type = type;
1457 reg->dynptr.first_slot = first_slot;
1460 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
1462 if (base_type(reg->type) == PTR_TO_MAP_VALUE) {
1463 const struct bpf_map *map = reg->map_ptr;
1465 if (map->inner_map_meta) {
1466 reg->type = CONST_PTR_TO_MAP;
1467 reg->map_ptr = map->inner_map_meta;
1468 /* transfer reg's id which is unique for every map_lookup_elem
1469 * as UID of the inner map.
1471 if (btf_record_has_field(map->inner_map_meta->record, BPF_TIMER))
1472 reg->map_uid = reg->id;
1473 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
1474 reg->type = PTR_TO_XDP_SOCK;
1475 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
1476 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
1477 reg->type = PTR_TO_SOCKET;
1479 reg->type = PTR_TO_MAP_VALUE;
1484 reg->type &= ~PTR_MAYBE_NULL;
1487 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1489 return type_is_pkt_pointer(reg->type);
1492 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
1494 return reg_is_pkt_pointer(reg) ||
1495 reg->type == PTR_TO_PACKET_END;
1498 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
1499 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
1500 enum bpf_reg_type which)
1502 /* The register can already have a range from prior markings.
1503 * This is fine as long as it hasn't been advanced from its
1506 return reg->type == which &&
1509 tnum_equals_const(reg->var_off, 0);
1512 /* Reset the min/max bounds of a register */
1513 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1515 reg->smin_value = S64_MIN;
1516 reg->smax_value = S64_MAX;
1517 reg->umin_value = 0;
1518 reg->umax_value = U64_MAX;
1520 reg->s32_min_value = S32_MIN;
1521 reg->s32_max_value = S32_MAX;
1522 reg->u32_min_value = 0;
1523 reg->u32_max_value = U32_MAX;
1526 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
1528 reg->smin_value = S64_MIN;
1529 reg->smax_value = S64_MAX;
1530 reg->umin_value = 0;
1531 reg->umax_value = U64_MAX;
1534 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
1536 reg->s32_min_value = S32_MIN;
1537 reg->s32_max_value = S32_MAX;
1538 reg->u32_min_value = 0;
1539 reg->u32_max_value = U32_MAX;
1542 static void __update_reg32_bounds(struct bpf_reg_state *reg)
1544 struct tnum var32_off = tnum_subreg(reg->var_off);
1546 /* min signed is max(sign bit) | min(other bits) */
1547 reg->s32_min_value = max_t(s32, reg->s32_min_value,
1548 var32_off.value | (var32_off.mask & S32_MIN));
1549 /* max signed is min(sign bit) | max(other bits) */
1550 reg->s32_max_value = min_t(s32, reg->s32_max_value,
1551 var32_off.value | (var32_off.mask & S32_MAX));
1552 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
1553 reg->u32_max_value = min(reg->u32_max_value,
1554 (u32)(var32_off.value | var32_off.mask));
1557 static void __update_reg64_bounds(struct bpf_reg_state *reg)
1559 /* min signed is max(sign bit) | min(other bits) */
1560 reg->smin_value = max_t(s64, reg->smin_value,
1561 reg->var_off.value | (reg->var_off.mask & S64_MIN));
1562 /* max signed is min(sign bit) | max(other bits) */
1563 reg->smax_value = min_t(s64, reg->smax_value,
1564 reg->var_off.value | (reg->var_off.mask & S64_MAX));
1565 reg->umin_value = max(reg->umin_value, reg->var_off.value);
1566 reg->umax_value = min(reg->umax_value,
1567 reg->var_off.value | reg->var_off.mask);
1570 static void __update_reg_bounds(struct bpf_reg_state *reg)
1572 __update_reg32_bounds(reg);
1573 __update_reg64_bounds(reg);
1576 /* Uses signed min/max values to inform unsigned, and vice-versa */
1577 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
1579 /* Learn sign from signed bounds.
1580 * If we cannot cross the sign boundary, then signed and unsigned bounds
1581 * are the same, so combine. This works even in the negative case, e.g.
1582 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1584 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
1585 reg->s32_min_value = reg->u32_min_value =
1586 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1587 reg->s32_max_value = reg->u32_max_value =
1588 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1591 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1592 * boundary, so we must be careful.
1594 if ((s32)reg->u32_max_value >= 0) {
1595 /* Positive. We can't learn anything from the smin, but smax
1596 * is positive, hence safe.
1598 reg->s32_min_value = reg->u32_min_value;
1599 reg->s32_max_value = reg->u32_max_value =
1600 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1601 } else if ((s32)reg->u32_min_value < 0) {
1602 /* Negative. We can't learn anything from the smax, but smin
1603 * is negative, hence safe.
1605 reg->s32_min_value = reg->u32_min_value =
1606 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1607 reg->s32_max_value = reg->u32_max_value;
1611 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
1613 /* Learn sign from signed bounds.
1614 * If we cannot cross the sign boundary, then signed and unsigned bounds
1615 * are the same, so combine. This works even in the negative case, e.g.
1616 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1618 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1619 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1621 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1625 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1626 * boundary, so we must be careful.
1628 if ((s64)reg->umax_value >= 0) {
1629 /* Positive. We can't learn anything from the smin, but smax
1630 * is positive, hence safe.
1632 reg->smin_value = reg->umin_value;
1633 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1635 } else if ((s64)reg->umin_value < 0) {
1636 /* Negative. We can't learn anything from the smax, but smin
1637 * is negative, hence safe.
1639 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1641 reg->smax_value = reg->umax_value;
1645 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
1647 __reg32_deduce_bounds(reg);
1648 __reg64_deduce_bounds(reg);
1651 /* Attempts to improve var_off based on unsigned min/max information */
1652 static void __reg_bound_offset(struct bpf_reg_state *reg)
1654 struct tnum var64_off = tnum_intersect(reg->var_off,
1655 tnum_range(reg->umin_value,
1657 struct tnum var32_off = tnum_intersect(tnum_subreg(reg->var_off),
1658 tnum_range(reg->u32_min_value,
1659 reg->u32_max_value));
1661 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
1664 static void reg_bounds_sync(struct bpf_reg_state *reg)
1666 /* We might have learned new bounds from the var_off. */
1667 __update_reg_bounds(reg);
1668 /* We might have learned something about the sign bit. */
1669 __reg_deduce_bounds(reg);
1670 /* We might have learned some bits from the bounds. */
1671 __reg_bound_offset(reg);
1672 /* Intersecting with the old var_off might have improved our bounds
1673 * slightly, e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1674 * then new var_off is (0; 0x7f...fc) which improves our umax.
1676 __update_reg_bounds(reg);
1679 static bool __reg32_bound_s64(s32 a)
1681 return a >= 0 && a <= S32_MAX;
1684 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
1686 reg->umin_value = reg->u32_min_value;
1687 reg->umax_value = reg->u32_max_value;
1689 /* Attempt to pull 32-bit signed bounds into 64-bit bounds but must
1690 * be positive otherwise set to worse case bounds and refine later
1693 if (__reg32_bound_s64(reg->s32_min_value) &&
1694 __reg32_bound_s64(reg->s32_max_value)) {
1695 reg->smin_value = reg->s32_min_value;
1696 reg->smax_value = reg->s32_max_value;
1698 reg->smin_value = 0;
1699 reg->smax_value = U32_MAX;
1703 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
1705 /* special case when 64-bit register has upper 32-bit register
1706 * zeroed. Typically happens after zext or <<32, >>32 sequence
1707 * allowing us to use 32-bit bounds directly,
1709 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
1710 __reg_assign_32_into_64(reg);
1712 /* Otherwise the best we can do is push lower 32bit known and
1713 * unknown bits into register (var_off set from jmp logic)
1714 * then learn as much as possible from the 64-bit tnum
1715 * known and unknown bits. The previous smin/smax bounds are
1716 * invalid here because of jmp32 compare so mark them unknown
1717 * so they do not impact tnum bounds calculation.
1719 __mark_reg64_unbounded(reg);
1721 reg_bounds_sync(reg);
1724 static bool __reg64_bound_s32(s64 a)
1726 return a >= S32_MIN && a <= S32_MAX;
1729 static bool __reg64_bound_u32(u64 a)
1731 return a >= U32_MIN && a <= U32_MAX;
1734 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
1736 __mark_reg32_unbounded(reg);
1737 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
1738 reg->s32_min_value = (s32)reg->smin_value;
1739 reg->s32_max_value = (s32)reg->smax_value;
1741 if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
1742 reg->u32_min_value = (u32)reg->umin_value;
1743 reg->u32_max_value = (u32)reg->umax_value;
1745 reg_bounds_sync(reg);
1748 /* Mark a register as having a completely unknown (scalar) value. */
1749 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1750 struct bpf_reg_state *reg)
1753 * Clear type, id, off, and union(map_ptr, range) and
1754 * padding between 'type' and union
1756 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1757 reg->type = SCALAR_VALUE;
1758 reg->var_off = tnum_unknown;
1760 reg->precise = !env->bpf_capable;
1761 __mark_reg_unbounded(reg);
1764 static void mark_reg_unknown(struct bpf_verifier_env *env,
1765 struct bpf_reg_state *regs, u32 regno)
1767 if (WARN_ON(regno >= MAX_BPF_REG)) {
1768 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1769 /* Something bad happened, let's kill all regs except FP */
1770 for (regno = 0; regno < BPF_REG_FP; regno++)
1771 __mark_reg_not_init(env, regs + regno);
1774 __mark_reg_unknown(env, regs + regno);
1777 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1778 struct bpf_reg_state *reg)
1780 __mark_reg_unknown(env, reg);
1781 reg->type = NOT_INIT;
1784 static void mark_reg_not_init(struct bpf_verifier_env *env,
1785 struct bpf_reg_state *regs, u32 regno)
1787 if (WARN_ON(regno >= MAX_BPF_REG)) {
1788 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1789 /* Something bad happened, let's kill all regs except FP */
1790 for (regno = 0; regno < BPF_REG_FP; regno++)
1791 __mark_reg_not_init(env, regs + regno);
1794 __mark_reg_not_init(env, regs + regno);
1797 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
1798 struct bpf_reg_state *regs, u32 regno,
1799 enum bpf_reg_type reg_type,
1800 struct btf *btf, u32 btf_id,
1801 enum bpf_type_flag flag)
1803 if (reg_type == SCALAR_VALUE) {
1804 mark_reg_unknown(env, regs, regno);
1807 mark_reg_known_zero(env, regs, regno);
1808 regs[regno].type = PTR_TO_BTF_ID | flag;
1809 regs[regno].btf = btf;
1810 regs[regno].btf_id = btf_id;
1813 #define DEF_NOT_SUBREG (0)
1814 static void init_reg_state(struct bpf_verifier_env *env,
1815 struct bpf_func_state *state)
1817 struct bpf_reg_state *regs = state->regs;
1820 for (i = 0; i < MAX_BPF_REG; i++) {
1821 mark_reg_not_init(env, regs, i);
1822 regs[i].live = REG_LIVE_NONE;
1823 regs[i].parent = NULL;
1824 regs[i].subreg_def = DEF_NOT_SUBREG;
1828 regs[BPF_REG_FP].type = PTR_TO_STACK;
1829 mark_reg_known_zero(env, regs, BPF_REG_FP);
1830 regs[BPF_REG_FP].frameno = state->frameno;
1833 #define BPF_MAIN_FUNC (-1)
1834 static void init_func_state(struct bpf_verifier_env *env,
1835 struct bpf_func_state *state,
1836 int callsite, int frameno, int subprogno)
1838 state->callsite = callsite;
1839 state->frameno = frameno;
1840 state->subprogno = subprogno;
1841 state->callback_ret_range = tnum_range(0, 0);
1842 init_reg_state(env, state);
1843 mark_verifier_state_scratched(env);
1846 /* Similar to push_stack(), but for async callbacks */
1847 static struct bpf_verifier_state *push_async_cb(struct bpf_verifier_env *env,
1848 int insn_idx, int prev_insn_idx,
1851 struct bpf_verifier_stack_elem *elem;
1852 struct bpf_func_state *frame;
1854 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1858 elem->insn_idx = insn_idx;
1859 elem->prev_insn_idx = prev_insn_idx;
1860 elem->next = env->head;
1861 elem->log_pos = env->log.len_used;
1864 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1866 "The sequence of %d jumps is too complex for async cb.\n",
1870 /* Unlike push_stack() do not copy_verifier_state().
1871 * The caller state doesn't matter.
1872 * This is async callback. It starts in a fresh stack.
1873 * Initialize it similar to do_check_common().
1875 elem->st.branches = 1;
1876 frame = kzalloc(sizeof(*frame), GFP_KERNEL);
1879 init_func_state(env, frame,
1880 BPF_MAIN_FUNC /* callsite */,
1881 0 /* frameno within this callchain */,
1882 subprog /* subprog number within this prog */);
1883 elem->st.frame[0] = frame;
1886 free_verifier_state(env->cur_state, true);
1887 env->cur_state = NULL;
1888 /* pop all elements and return */
1889 while (!pop_stack(env, NULL, NULL, false));
1895 SRC_OP, /* register is used as source operand */
1896 DST_OP, /* register is used as destination operand */
1897 DST_OP_NO_MARK /* same as above, check only, don't mark */
1900 static int cmp_subprogs(const void *a, const void *b)
1902 return ((struct bpf_subprog_info *)a)->start -
1903 ((struct bpf_subprog_info *)b)->start;
1906 static int find_subprog(struct bpf_verifier_env *env, int off)
1908 struct bpf_subprog_info *p;
1910 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1911 sizeof(env->subprog_info[0]), cmp_subprogs);
1914 return p - env->subprog_info;
1918 static int add_subprog(struct bpf_verifier_env *env, int off)
1920 int insn_cnt = env->prog->len;
1923 if (off >= insn_cnt || off < 0) {
1924 verbose(env, "call to invalid destination\n");
1927 ret = find_subprog(env, off);
1930 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1931 verbose(env, "too many subprograms\n");
1934 /* determine subprog starts. The end is one before the next starts */
1935 env->subprog_info[env->subprog_cnt++].start = off;
1936 sort(env->subprog_info, env->subprog_cnt,
1937 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1938 return env->subprog_cnt - 1;
1941 #define MAX_KFUNC_DESCS 256
1942 #define MAX_KFUNC_BTFS 256
1944 struct bpf_kfunc_desc {
1945 struct btf_func_model func_model;
1951 struct bpf_kfunc_btf {
1953 struct module *module;
1957 struct bpf_kfunc_desc_tab {
1958 struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
1962 struct bpf_kfunc_btf_tab {
1963 struct bpf_kfunc_btf descs[MAX_KFUNC_BTFS];
1967 static int kfunc_desc_cmp_by_id_off(const void *a, const void *b)
1969 const struct bpf_kfunc_desc *d0 = a;
1970 const struct bpf_kfunc_desc *d1 = b;
1972 /* func_id is not greater than BTF_MAX_TYPE */
1973 return d0->func_id - d1->func_id ?: d0->offset - d1->offset;
1976 static int kfunc_btf_cmp_by_off(const void *a, const void *b)
1978 const struct bpf_kfunc_btf *d0 = a;
1979 const struct bpf_kfunc_btf *d1 = b;
1981 return d0->offset - d1->offset;
1984 static const struct bpf_kfunc_desc *
1985 find_kfunc_desc(const struct bpf_prog *prog, u32 func_id, u16 offset)
1987 struct bpf_kfunc_desc desc = {
1991 struct bpf_kfunc_desc_tab *tab;
1993 tab = prog->aux->kfunc_tab;
1994 return bsearch(&desc, tab->descs, tab->nr_descs,
1995 sizeof(tab->descs[0]), kfunc_desc_cmp_by_id_off);
1998 static struct btf *__find_kfunc_desc_btf(struct bpf_verifier_env *env,
2001 struct bpf_kfunc_btf kf_btf = { .offset = offset };
2002 struct bpf_kfunc_btf_tab *tab;
2003 struct bpf_kfunc_btf *b;
2008 tab = env->prog->aux->kfunc_btf_tab;
2009 b = bsearch(&kf_btf, tab->descs, tab->nr_descs,
2010 sizeof(tab->descs[0]), kfunc_btf_cmp_by_off);
2012 if (tab->nr_descs == MAX_KFUNC_BTFS) {
2013 verbose(env, "too many different module BTFs\n");
2014 return ERR_PTR(-E2BIG);
2017 if (bpfptr_is_null(env->fd_array)) {
2018 verbose(env, "kfunc offset > 0 without fd_array is invalid\n");
2019 return ERR_PTR(-EPROTO);
2022 if (copy_from_bpfptr_offset(&btf_fd, env->fd_array,
2023 offset * sizeof(btf_fd),
2025 return ERR_PTR(-EFAULT);
2027 btf = btf_get_by_fd(btf_fd);
2029 verbose(env, "invalid module BTF fd specified\n");
2033 if (!btf_is_module(btf)) {
2034 verbose(env, "BTF fd for kfunc is not a module BTF\n");
2036 return ERR_PTR(-EINVAL);
2039 mod = btf_try_get_module(btf);
2042 return ERR_PTR(-ENXIO);
2045 b = &tab->descs[tab->nr_descs++];
2050 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2051 kfunc_btf_cmp_by_off, NULL);
2056 void bpf_free_kfunc_btf_tab(struct bpf_kfunc_btf_tab *tab)
2061 while (tab->nr_descs--) {
2062 module_put(tab->descs[tab->nr_descs].module);
2063 btf_put(tab->descs[tab->nr_descs].btf);
2068 static struct btf *find_kfunc_desc_btf(struct bpf_verifier_env *env, s16 offset)
2072 /* In the future, this can be allowed to increase limit
2073 * of fd index into fd_array, interpreted as u16.
2075 verbose(env, "negative offset disallowed for kernel module function call\n");
2076 return ERR_PTR(-EINVAL);
2079 return __find_kfunc_desc_btf(env, offset);
2081 return btf_vmlinux ?: ERR_PTR(-ENOENT);
2084 static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id, s16 offset)
2086 const struct btf_type *func, *func_proto;
2087 struct bpf_kfunc_btf_tab *btf_tab;
2088 struct bpf_kfunc_desc_tab *tab;
2089 struct bpf_prog_aux *prog_aux;
2090 struct bpf_kfunc_desc *desc;
2091 const char *func_name;
2092 struct btf *desc_btf;
2093 unsigned long call_imm;
2097 prog_aux = env->prog->aux;
2098 tab = prog_aux->kfunc_tab;
2099 btf_tab = prog_aux->kfunc_btf_tab;
2102 verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
2106 if (!env->prog->jit_requested) {
2107 verbose(env, "JIT is required for calling kernel function\n");
2111 if (!bpf_jit_supports_kfunc_call()) {
2112 verbose(env, "JIT does not support calling kernel function\n");
2116 if (!env->prog->gpl_compatible) {
2117 verbose(env, "cannot call kernel function from non-GPL compatible program\n");
2121 tab = kzalloc(sizeof(*tab), GFP_KERNEL);
2124 prog_aux->kfunc_tab = tab;
2127 /* func_id == 0 is always invalid, but instead of returning an error, be
2128 * conservative and wait until the code elimination pass before returning
2129 * error, so that invalid calls that get pruned out can be in BPF programs
2130 * loaded from userspace. It is also required that offset be untouched
2133 if (!func_id && !offset)
2136 if (!btf_tab && offset) {
2137 btf_tab = kzalloc(sizeof(*btf_tab), GFP_KERNEL);
2140 prog_aux->kfunc_btf_tab = btf_tab;
2143 desc_btf = find_kfunc_desc_btf(env, offset);
2144 if (IS_ERR(desc_btf)) {
2145 verbose(env, "failed to find BTF for kernel function\n");
2146 return PTR_ERR(desc_btf);
2149 if (find_kfunc_desc(env->prog, func_id, offset))
2152 if (tab->nr_descs == MAX_KFUNC_DESCS) {
2153 verbose(env, "too many different kernel function calls\n");
2157 func = btf_type_by_id(desc_btf, func_id);
2158 if (!func || !btf_type_is_func(func)) {
2159 verbose(env, "kernel btf_id %u is not a function\n",
2163 func_proto = btf_type_by_id(desc_btf, func->type);
2164 if (!func_proto || !btf_type_is_func_proto(func_proto)) {
2165 verbose(env, "kernel function btf_id %u does not have a valid func_proto\n",
2170 func_name = btf_name_by_offset(desc_btf, func->name_off);
2171 addr = kallsyms_lookup_name(func_name);
2173 verbose(env, "cannot find address for kernel function %s\n",
2178 call_imm = BPF_CALL_IMM(addr);
2179 /* Check whether or not the relative offset overflows desc->imm */
2180 if ((unsigned long)(s32)call_imm != call_imm) {
2181 verbose(env, "address of kernel function %s is out of range\n",
2186 desc = &tab->descs[tab->nr_descs++];
2187 desc->func_id = func_id;
2188 desc->imm = call_imm;
2189 desc->offset = offset;
2190 err = btf_distill_func_proto(&env->log, desc_btf,
2191 func_proto, func_name,
2194 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2195 kfunc_desc_cmp_by_id_off, NULL);
2199 static int kfunc_desc_cmp_by_imm(const void *a, const void *b)
2201 const struct bpf_kfunc_desc *d0 = a;
2202 const struct bpf_kfunc_desc *d1 = b;
2204 if (d0->imm > d1->imm)
2206 else if (d0->imm < d1->imm)
2211 static void sort_kfunc_descs_by_imm(struct bpf_prog *prog)
2213 struct bpf_kfunc_desc_tab *tab;
2215 tab = prog->aux->kfunc_tab;
2219 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2220 kfunc_desc_cmp_by_imm, NULL);
2223 bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
2225 return !!prog->aux->kfunc_tab;
2228 const struct btf_func_model *
2229 bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
2230 const struct bpf_insn *insn)
2232 const struct bpf_kfunc_desc desc = {
2235 const struct bpf_kfunc_desc *res;
2236 struct bpf_kfunc_desc_tab *tab;
2238 tab = prog->aux->kfunc_tab;
2239 res = bsearch(&desc, tab->descs, tab->nr_descs,
2240 sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm);
2242 return res ? &res->func_model : NULL;
2245 static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
2247 struct bpf_subprog_info *subprog = env->subprog_info;
2248 struct bpf_insn *insn = env->prog->insnsi;
2249 int i, ret, insn_cnt = env->prog->len;
2251 /* Add entry function. */
2252 ret = add_subprog(env, 0);
2256 for (i = 0; i < insn_cnt; i++, insn++) {
2257 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
2258 !bpf_pseudo_kfunc_call(insn))
2261 if (!env->bpf_capable) {
2262 verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
2266 if (bpf_pseudo_func(insn) || bpf_pseudo_call(insn))
2267 ret = add_subprog(env, i + insn->imm + 1);
2269 ret = add_kfunc_call(env, insn->imm, insn->off);
2275 /* Add a fake 'exit' subprog which could simplify subprog iteration
2276 * logic. 'subprog_cnt' should not be increased.
2278 subprog[env->subprog_cnt].start = insn_cnt;
2280 if (env->log.level & BPF_LOG_LEVEL2)
2281 for (i = 0; i < env->subprog_cnt; i++)
2282 verbose(env, "func#%d @%d\n", i, subprog[i].start);
2287 static int check_subprogs(struct bpf_verifier_env *env)
2289 int i, subprog_start, subprog_end, off, cur_subprog = 0;
2290 struct bpf_subprog_info *subprog = env->subprog_info;
2291 struct bpf_insn *insn = env->prog->insnsi;
2292 int insn_cnt = env->prog->len;
2294 /* now check that all jumps are within the same subprog */
2295 subprog_start = subprog[cur_subprog].start;
2296 subprog_end = subprog[cur_subprog + 1].start;
2297 for (i = 0; i < insn_cnt; i++) {
2298 u8 code = insn[i].code;
2300 if (code == (BPF_JMP | BPF_CALL) &&
2301 insn[i].imm == BPF_FUNC_tail_call &&
2302 insn[i].src_reg != BPF_PSEUDO_CALL)
2303 subprog[cur_subprog].has_tail_call = true;
2304 if (BPF_CLASS(code) == BPF_LD &&
2305 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
2306 subprog[cur_subprog].has_ld_abs = true;
2307 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
2309 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
2311 off = i + insn[i].off + 1;
2312 if (off < subprog_start || off >= subprog_end) {
2313 verbose(env, "jump out of range from insn %d to %d\n", i, off);
2317 if (i == subprog_end - 1) {
2318 /* to avoid fall-through from one subprog into another
2319 * the last insn of the subprog should be either exit
2320 * or unconditional jump back
2322 if (code != (BPF_JMP | BPF_EXIT) &&
2323 code != (BPF_JMP | BPF_JA)) {
2324 verbose(env, "last insn is not an exit or jmp\n");
2327 subprog_start = subprog_end;
2329 if (cur_subprog < env->subprog_cnt)
2330 subprog_end = subprog[cur_subprog + 1].start;
2336 /* Parentage chain of this register (or stack slot) should take care of all
2337 * issues like callee-saved registers, stack slot allocation time, etc.
2339 static int mark_reg_read(struct bpf_verifier_env *env,
2340 const struct bpf_reg_state *state,
2341 struct bpf_reg_state *parent, u8 flag)
2343 bool writes = parent == state->parent; /* Observe write marks */
2347 /* if read wasn't screened by an earlier write ... */
2348 if (writes && state->live & REG_LIVE_WRITTEN)
2350 if (parent->live & REG_LIVE_DONE) {
2351 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
2352 reg_type_str(env, parent->type),
2353 parent->var_off.value, parent->off);
2356 /* The first condition is more likely to be true than the
2357 * second, checked it first.
2359 if ((parent->live & REG_LIVE_READ) == flag ||
2360 parent->live & REG_LIVE_READ64)
2361 /* The parentage chain never changes and
2362 * this parent was already marked as LIVE_READ.
2363 * There is no need to keep walking the chain again and
2364 * keep re-marking all parents as LIVE_READ.
2365 * This case happens when the same register is read
2366 * multiple times without writes into it in-between.
2367 * Also, if parent has the stronger REG_LIVE_READ64 set,
2368 * then no need to set the weak REG_LIVE_READ32.
2371 /* ... then we depend on parent's value */
2372 parent->live |= flag;
2373 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
2374 if (flag == REG_LIVE_READ64)
2375 parent->live &= ~REG_LIVE_READ32;
2377 parent = state->parent;
2382 if (env->longest_mark_read_walk < cnt)
2383 env->longest_mark_read_walk = cnt;
2387 /* This function is supposed to be used by the following 32-bit optimization
2388 * code only. It returns TRUE if the source or destination register operates
2389 * on 64-bit, otherwise return FALSE.
2391 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
2392 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
2397 class = BPF_CLASS(code);
2399 if (class == BPF_JMP) {
2400 /* BPF_EXIT for "main" will reach here. Return TRUE
2405 if (op == BPF_CALL) {
2406 /* BPF to BPF call will reach here because of marking
2407 * caller saved clobber with DST_OP_NO_MARK for which we
2408 * don't care the register def because they are anyway
2409 * marked as NOT_INIT already.
2411 if (insn->src_reg == BPF_PSEUDO_CALL)
2413 /* Helper call will reach here because of arg type
2414 * check, conservatively return TRUE.
2423 if (class == BPF_ALU64 || class == BPF_JMP ||
2424 /* BPF_END always use BPF_ALU class. */
2425 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
2428 if (class == BPF_ALU || class == BPF_JMP32)
2431 if (class == BPF_LDX) {
2433 return BPF_SIZE(code) == BPF_DW;
2434 /* LDX source must be ptr. */
2438 if (class == BPF_STX) {
2439 /* BPF_STX (including atomic variants) has multiple source
2440 * operands, one of which is a ptr. Check whether the caller is
2443 if (t == SRC_OP && reg->type != SCALAR_VALUE)
2445 return BPF_SIZE(code) == BPF_DW;
2448 if (class == BPF_LD) {
2449 u8 mode = BPF_MODE(code);
2452 if (mode == BPF_IMM)
2455 /* Both LD_IND and LD_ABS return 32-bit data. */
2459 /* Implicit ctx ptr. */
2460 if (regno == BPF_REG_6)
2463 /* Explicit source could be any width. */
2467 if (class == BPF_ST)
2468 /* The only source register for BPF_ST is a ptr. */
2471 /* Conservatively return true at default. */
2475 /* Return the regno defined by the insn, or -1. */
2476 static int insn_def_regno(const struct bpf_insn *insn)
2478 switch (BPF_CLASS(insn->code)) {
2484 if (BPF_MODE(insn->code) == BPF_ATOMIC &&
2485 (insn->imm & BPF_FETCH)) {
2486 if (insn->imm == BPF_CMPXCHG)
2489 return insn->src_reg;
2494 return insn->dst_reg;
2498 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
2499 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
2501 int dst_reg = insn_def_regno(insn);
2506 return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
2509 static void mark_insn_zext(struct bpf_verifier_env *env,
2510 struct bpf_reg_state *reg)
2512 s32 def_idx = reg->subreg_def;
2514 if (def_idx == DEF_NOT_SUBREG)
2517 env->insn_aux_data[def_idx - 1].zext_dst = true;
2518 /* The dst will be zero extended, so won't be sub-register anymore. */
2519 reg->subreg_def = DEF_NOT_SUBREG;
2522 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
2523 enum reg_arg_type t)
2525 struct bpf_verifier_state *vstate = env->cur_state;
2526 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2527 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
2528 struct bpf_reg_state *reg, *regs = state->regs;
2531 if (regno >= MAX_BPF_REG) {
2532 verbose(env, "R%d is invalid\n", regno);
2536 mark_reg_scratched(env, regno);
2539 rw64 = is_reg64(env, insn, regno, reg, t);
2541 /* check whether register used as source operand can be read */
2542 if (reg->type == NOT_INIT) {
2543 verbose(env, "R%d !read_ok\n", regno);
2546 /* We don't need to worry about FP liveness because it's read-only */
2547 if (regno == BPF_REG_FP)
2551 mark_insn_zext(env, reg);
2553 return mark_reg_read(env, reg, reg->parent,
2554 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
2556 /* check whether register used as dest operand can be written to */
2557 if (regno == BPF_REG_FP) {
2558 verbose(env, "frame pointer is read only\n");
2561 reg->live |= REG_LIVE_WRITTEN;
2562 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
2564 mark_reg_unknown(env, regs, regno);
2569 static void mark_jmp_point(struct bpf_verifier_env *env, int idx)
2571 env->insn_aux_data[idx].jmp_point = true;
2574 static bool is_jmp_point(struct bpf_verifier_env *env, int insn_idx)
2576 return env->insn_aux_data[insn_idx].jmp_point;
2579 /* for any branch, call, exit record the history of jmps in the given state */
2580 static int push_jmp_history(struct bpf_verifier_env *env,
2581 struct bpf_verifier_state *cur)
2583 u32 cnt = cur->jmp_history_cnt;
2584 struct bpf_idx_pair *p;
2587 if (!is_jmp_point(env, env->insn_idx))
2591 alloc_size = kmalloc_size_roundup(size_mul(cnt, sizeof(*p)));
2592 p = krealloc(cur->jmp_history, alloc_size, GFP_USER);
2595 p[cnt - 1].idx = env->insn_idx;
2596 p[cnt - 1].prev_idx = env->prev_insn_idx;
2597 cur->jmp_history = p;
2598 cur->jmp_history_cnt = cnt;
2602 /* Backtrack one insn at a time. If idx is not at the top of recorded
2603 * history then previous instruction came from straight line execution.
2605 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
2610 if (cnt && st->jmp_history[cnt - 1].idx == i) {
2611 i = st->jmp_history[cnt - 1].prev_idx;
2619 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
2621 const struct btf_type *func;
2622 struct btf *desc_btf;
2624 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
2627 desc_btf = find_kfunc_desc_btf(data, insn->off);
2628 if (IS_ERR(desc_btf))
2631 func = btf_type_by_id(desc_btf, insn->imm);
2632 return btf_name_by_offset(desc_btf, func->name_off);
2635 /* For given verifier state backtrack_insn() is called from the last insn to
2636 * the first insn. Its purpose is to compute a bitmask of registers and
2637 * stack slots that needs precision in the parent verifier state.
2639 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
2640 u32 *reg_mask, u64 *stack_mask)
2642 const struct bpf_insn_cbs cbs = {
2643 .cb_call = disasm_kfunc_name,
2644 .cb_print = verbose,
2645 .private_data = env,
2647 struct bpf_insn *insn = env->prog->insnsi + idx;
2648 u8 class = BPF_CLASS(insn->code);
2649 u8 opcode = BPF_OP(insn->code);
2650 u8 mode = BPF_MODE(insn->code);
2651 u32 dreg = 1u << insn->dst_reg;
2652 u32 sreg = 1u << insn->src_reg;
2655 if (insn->code == 0)
2657 if (env->log.level & BPF_LOG_LEVEL2) {
2658 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
2659 verbose(env, "%d: ", idx);
2660 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
2663 if (class == BPF_ALU || class == BPF_ALU64) {
2664 if (!(*reg_mask & dreg))
2666 if (opcode == BPF_MOV) {
2667 if (BPF_SRC(insn->code) == BPF_X) {
2669 * dreg needs precision after this insn
2670 * sreg needs precision before this insn
2676 * dreg needs precision after this insn.
2677 * Corresponding register is already marked
2678 * as precise=true in this verifier state.
2679 * No further markings in parent are necessary
2684 if (BPF_SRC(insn->code) == BPF_X) {
2686 * both dreg and sreg need precision
2691 * dreg still needs precision before this insn
2694 } else if (class == BPF_LDX) {
2695 if (!(*reg_mask & dreg))
2699 /* scalars can only be spilled into stack w/o losing precision.
2700 * Load from any other memory can be zero extended.
2701 * The desire to keep that precision is already indicated
2702 * by 'precise' mark in corresponding register of this state.
2703 * No further tracking necessary.
2705 if (insn->src_reg != BPF_REG_FP)
2708 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
2709 * that [fp - off] slot contains scalar that needs to be
2710 * tracked with precision
2712 spi = (-insn->off - 1) / BPF_REG_SIZE;
2714 verbose(env, "BUG spi %d\n", spi);
2715 WARN_ONCE(1, "verifier backtracking bug");
2718 *stack_mask |= 1ull << spi;
2719 } else if (class == BPF_STX || class == BPF_ST) {
2720 if (*reg_mask & dreg)
2721 /* stx & st shouldn't be using _scalar_ dst_reg
2722 * to access memory. It means backtracking
2723 * encountered a case of pointer subtraction.
2726 /* scalars can only be spilled into stack */
2727 if (insn->dst_reg != BPF_REG_FP)
2729 spi = (-insn->off - 1) / BPF_REG_SIZE;
2731 verbose(env, "BUG spi %d\n", spi);
2732 WARN_ONCE(1, "verifier backtracking bug");
2735 if (!(*stack_mask & (1ull << spi)))
2737 *stack_mask &= ~(1ull << spi);
2738 if (class == BPF_STX)
2740 } else if (class == BPF_JMP || class == BPF_JMP32) {
2741 if (opcode == BPF_CALL) {
2742 if (insn->src_reg == BPF_PSEUDO_CALL)
2744 /* BPF helpers that invoke callback subprogs are
2745 * equivalent to BPF_PSEUDO_CALL above
2747 if (insn->src_reg == 0 && is_callback_calling_function(insn->imm))
2749 /* regular helper call sets R0 */
2751 if (*reg_mask & 0x3f) {
2752 /* if backtracing was looking for registers R1-R5
2753 * they should have been found already.
2755 verbose(env, "BUG regs %x\n", *reg_mask);
2756 WARN_ONCE(1, "verifier backtracking bug");
2759 } else if (opcode == BPF_EXIT) {
2762 } else if (class == BPF_LD) {
2763 if (!(*reg_mask & dreg))
2766 /* It's ld_imm64 or ld_abs or ld_ind.
2767 * For ld_imm64 no further tracking of precision
2768 * into parent is necessary
2770 if (mode == BPF_IND || mode == BPF_ABS)
2771 /* to be analyzed */
2777 /* the scalar precision tracking algorithm:
2778 * . at the start all registers have precise=false.
2779 * . scalar ranges are tracked as normal through alu and jmp insns.
2780 * . once precise value of the scalar register is used in:
2781 * . ptr + scalar alu
2782 * . if (scalar cond K|scalar)
2783 * . helper_call(.., scalar, ...) where ARG_CONST is expected
2784 * backtrack through the verifier states and mark all registers and
2785 * stack slots with spilled constants that these scalar regisers
2786 * should be precise.
2787 * . during state pruning two registers (or spilled stack slots)
2788 * are equivalent if both are not precise.
2790 * Note the verifier cannot simply walk register parentage chain,
2791 * since many different registers and stack slots could have been
2792 * used to compute single precise scalar.
2794 * The approach of starting with precise=true for all registers and then
2795 * backtrack to mark a register as not precise when the verifier detects
2796 * that program doesn't care about specific value (e.g., when helper
2797 * takes register as ARG_ANYTHING parameter) is not safe.
2799 * It's ok to walk single parentage chain of the verifier states.
2800 * It's possible that this backtracking will go all the way till 1st insn.
2801 * All other branches will be explored for needing precision later.
2803 * The backtracking needs to deal with cases like:
2804 * 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)
2807 * if r5 > 0x79f goto pc+7
2808 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
2811 * call bpf_perf_event_output#25
2812 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
2816 * call foo // uses callee's r6 inside to compute r0
2820 * to track above reg_mask/stack_mask needs to be independent for each frame.
2822 * Also if parent's curframe > frame where backtracking started,
2823 * the verifier need to mark registers in both frames, otherwise callees
2824 * may incorrectly prune callers. This is similar to
2825 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
2827 * For now backtracking falls back into conservative marking.
2829 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
2830 struct bpf_verifier_state *st)
2832 struct bpf_func_state *func;
2833 struct bpf_reg_state *reg;
2836 /* big hammer: mark all scalars precise in this path.
2837 * pop_stack may still get !precise scalars.
2838 * We also skip current state and go straight to first parent state,
2839 * because precision markings in current non-checkpointed state are
2840 * not needed. See why in the comment in __mark_chain_precision below.
2842 for (st = st->parent; st; st = st->parent) {
2843 for (i = 0; i <= st->curframe; i++) {
2844 func = st->frame[i];
2845 for (j = 0; j < BPF_REG_FP; j++) {
2846 reg = &func->regs[j];
2847 if (reg->type != SCALAR_VALUE)
2849 reg->precise = true;
2851 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
2852 if (!is_spilled_reg(&func->stack[j]))
2854 reg = &func->stack[j].spilled_ptr;
2855 if (reg->type != SCALAR_VALUE)
2857 reg->precise = true;
2863 static void mark_all_scalars_imprecise(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
2865 struct bpf_func_state *func;
2866 struct bpf_reg_state *reg;
2869 for (i = 0; i <= st->curframe; i++) {
2870 func = st->frame[i];
2871 for (j = 0; j < BPF_REG_FP; j++) {
2872 reg = &func->regs[j];
2873 if (reg->type != SCALAR_VALUE)
2875 reg->precise = false;
2877 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
2878 if (!is_spilled_reg(&func->stack[j]))
2880 reg = &func->stack[j].spilled_ptr;
2881 if (reg->type != SCALAR_VALUE)
2883 reg->precise = false;
2889 * __mark_chain_precision() backtracks BPF program instruction sequence and
2890 * chain of verifier states making sure that register *regno* (if regno >= 0)
2891 * and/or stack slot *spi* (if spi >= 0) are marked as precisely tracked
2892 * SCALARS, as well as any other registers and slots that contribute to
2893 * a tracked state of given registers/stack slots, depending on specific BPF
2894 * assembly instructions (see backtrack_insns() for exact instruction handling
2895 * logic). This backtracking relies on recorded jmp_history and is able to
2896 * traverse entire chain of parent states. This process ends only when all the
2897 * necessary registers/slots and their transitive dependencies are marked as
2900 * One important and subtle aspect is that precise marks *do not matter* in
2901 * the currently verified state (current state). It is important to understand
2902 * why this is the case.
2904 * First, note that current state is the state that is not yet "checkpointed",
2905 * i.e., it is not yet put into env->explored_states, and it has no children
2906 * states as well. It's ephemeral, and can end up either a) being discarded if
2907 * compatible explored state is found at some point or BPF_EXIT instruction is
2908 * reached or b) checkpointed and put into env->explored_states, branching out
2909 * into one or more children states.
2911 * In the former case, precise markings in current state are completely
2912 * ignored by state comparison code (see regsafe() for details). Only
2913 * checkpointed ("old") state precise markings are important, and if old
2914 * state's register/slot is precise, regsafe() assumes current state's
2915 * register/slot as precise and checks value ranges exactly and precisely. If
2916 * states turn out to be compatible, current state's necessary precise
2917 * markings and any required parent states' precise markings are enforced
2918 * after the fact with propagate_precision() logic, after the fact. But it's
2919 * important to realize that in this case, even after marking current state
2920 * registers/slots as precise, we immediately discard current state. So what
2921 * actually matters is any of the precise markings propagated into current
2922 * state's parent states, which are always checkpointed (due to b) case above).
2923 * As such, for scenario a) it doesn't matter if current state has precise
2924 * markings set or not.
2926 * Now, for the scenario b), checkpointing and forking into child(ren)
2927 * state(s). Note that before current state gets to checkpointing step, any
2928 * processed instruction always assumes precise SCALAR register/slot
2929 * knowledge: if precise value or range is useful to prune jump branch, BPF
2930 * verifier takes this opportunity enthusiastically. Similarly, when
2931 * register's value is used to calculate offset or memory address, exact
2932 * knowledge of SCALAR range is assumed, checked, and enforced. So, similar to
2933 * what we mentioned above about state comparison ignoring precise markings
2934 * during state comparison, BPF verifier ignores and also assumes precise
2935 * markings *at will* during instruction verification process. But as verifier
2936 * assumes precision, it also propagates any precision dependencies across
2937 * parent states, which are not yet finalized, so can be further restricted
2938 * based on new knowledge gained from restrictions enforced by their children
2939 * states. This is so that once those parent states are finalized, i.e., when
2940 * they have no more active children state, state comparison logic in
2941 * is_state_visited() would enforce strict and precise SCALAR ranges, if
2942 * required for correctness.
2944 * To build a bit more intuition, note also that once a state is checkpointed,
2945 * the path we took to get to that state is not important. This is crucial
2946 * property for state pruning. When state is checkpointed and finalized at
2947 * some instruction index, it can be correctly and safely used to "short
2948 * circuit" any *compatible* state that reaches exactly the same instruction
2949 * index. I.e., if we jumped to that instruction from a completely different
2950 * code path than original finalized state was derived from, it doesn't
2951 * matter, current state can be discarded because from that instruction
2952 * forward having a compatible state will ensure we will safely reach the
2953 * exit. States describe preconditions for further exploration, but completely
2954 * forget the history of how we got here.
2956 * This also means that even if we needed precise SCALAR range to get to
2957 * finalized state, but from that point forward *that same* SCALAR register is
2958 * never used in a precise context (i.e., it's precise value is not needed for
2959 * correctness), it's correct and safe to mark such register as "imprecise"
2960 * (i.e., precise marking set to false). This is what we rely on when we do
2961 * not set precise marking in current state. If no child state requires
2962 * precision for any given SCALAR register, it's safe to dictate that it can
2963 * be imprecise. If any child state does require this register to be precise,
2964 * we'll mark it precise later retroactively during precise markings
2965 * propagation from child state to parent states.
2967 * Skipping precise marking setting in current state is a mild version of
2968 * relying on the above observation. But we can utilize this property even
2969 * more aggressively by proactively forgetting any precise marking in the
2970 * current state (which we inherited from the parent state), right before we
2971 * checkpoint it and branch off into new child state. This is done by
2972 * mark_all_scalars_imprecise() to hopefully get more permissive and generic
2973 * finalized states which help in short circuiting more future states.
2975 static int __mark_chain_precision(struct bpf_verifier_env *env, int frame, int regno,
2978 struct bpf_verifier_state *st = env->cur_state;
2979 int first_idx = st->first_insn_idx;
2980 int last_idx = env->insn_idx;
2981 struct bpf_func_state *func;
2982 struct bpf_reg_state *reg;
2983 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
2984 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
2985 bool skip_first = true;
2986 bool new_marks = false;
2989 if (!env->bpf_capable)
2992 /* Do sanity checks against current state of register and/or stack
2993 * slot, but don't set precise flag in current state, as precision
2994 * tracking in the current state is unnecessary.
2996 func = st->frame[frame];
2998 reg = &func->regs[regno];
2999 if (reg->type != SCALAR_VALUE) {
3000 WARN_ONCE(1, "backtracing misuse");
3007 if (!is_spilled_reg(&func->stack[spi])) {
3011 reg = &func->stack[spi].spilled_ptr;
3012 if (reg->type != SCALAR_VALUE) {
3022 if (!reg_mask && !stack_mask)
3026 DECLARE_BITMAP(mask, 64);
3027 u32 history = st->jmp_history_cnt;
3029 if (env->log.level & BPF_LOG_LEVEL2)
3030 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
3033 /* we are at the entry into subprog, which
3034 * is expected for global funcs, but only if
3035 * requested precise registers are R1-R5
3036 * (which are global func's input arguments)
3038 if (st->curframe == 0 &&
3039 st->frame[0]->subprogno > 0 &&
3040 st->frame[0]->callsite == BPF_MAIN_FUNC &&
3041 stack_mask == 0 && (reg_mask & ~0x3e) == 0) {
3042 bitmap_from_u64(mask, reg_mask);
3043 for_each_set_bit(i, mask, 32) {
3044 reg = &st->frame[0]->regs[i];
3045 if (reg->type != SCALAR_VALUE) {
3046 reg_mask &= ~(1u << i);
3049 reg->precise = true;
3054 verbose(env, "BUG backtracing func entry subprog %d reg_mask %x stack_mask %llx\n",
3055 st->frame[0]->subprogno, reg_mask, stack_mask);
3056 WARN_ONCE(1, "verifier backtracking bug");
3060 for (i = last_idx;;) {
3065 err = backtrack_insn(env, i, ®_mask, &stack_mask);
3067 if (err == -ENOTSUPP) {
3068 mark_all_scalars_precise(env, st);
3073 if (!reg_mask && !stack_mask)
3074 /* Found assignment(s) into tracked register in this state.
3075 * Since this state is already marked, just return.
3076 * Nothing to be tracked further in the parent state.
3081 i = get_prev_insn_idx(st, i, &history);
3082 if (i >= env->prog->len) {
3083 /* This can happen if backtracking reached insn 0
3084 * and there are still reg_mask or stack_mask
3086 * It means the backtracking missed the spot where
3087 * particular register was initialized with a constant.
3089 verbose(env, "BUG backtracking idx %d\n", i);
3090 WARN_ONCE(1, "verifier backtracking bug");
3099 func = st->frame[frame];
3100 bitmap_from_u64(mask, reg_mask);
3101 for_each_set_bit(i, mask, 32) {
3102 reg = &func->regs[i];
3103 if (reg->type != SCALAR_VALUE) {
3104 reg_mask &= ~(1u << i);
3109 reg->precise = true;
3112 bitmap_from_u64(mask, stack_mask);
3113 for_each_set_bit(i, mask, 64) {
3114 if (i >= func->allocated_stack / BPF_REG_SIZE) {
3115 /* the sequence of instructions:
3117 * 3: (7b) *(u64 *)(r3 -8) = r0
3118 * 4: (79) r4 = *(u64 *)(r10 -8)
3119 * doesn't contain jmps. It's backtracked
3120 * as a single block.
3121 * During backtracking insn 3 is not recognized as
3122 * stack access, so at the end of backtracking
3123 * stack slot fp-8 is still marked in stack_mask.
3124 * However the parent state may not have accessed
3125 * fp-8 and it's "unallocated" stack space.
3126 * In such case fallback to conservative.
3128 mark_all_scalars_precise(env, st);
3132 if (!is_spilled_reg(&func->stack[i])) {
3133 stack_mask &= ~(1ull << i);
3136 reg = &func->stack[i].spilled_ptr;
3137 if (reg->type != SCALAR_VALUE) {
3138 stack_mask &= ~(1ull << i);
3143 reg->precise = true;
3145 if (env->log.level & BPF_LOG_LEVEL2) {
3146 verbose(env, "parent %s regs=%x stack=%llx marks:",
3147 new_marks ? "didn't have" : "already had",
3148 reg_mask, stack_mask);
3149 print_verifier_state(env, func, true);
3152 if (!reg_mask && !stack_mask)
3157 last_idx = st->last_insn_idx;
3158 first_idx = st->first_insn_idx;
3163 int mark_chain_precision(struct bpf_verifier_env *env, int regno)
3165 return __mark_chain_precision(env, env->cur_state->curframe, regno, -1);
3168 static int mark_chain_precision_frame(struct bpf_verifier_env *env, int frame, int regno)
3170 return __mark_chain_precision(env, frame, regno, -1);
3173 static int mark_chain_precision_stack_frame(struct bpf_verifier_env *env, int frame, int spi)
3175 return __mark_chain_precision(env, frame, -1, spi);
3178 static bool is_spillable_regtype(enum bpf_reg_type type)
3180 switch (base_type(type)) {
3181 case PTR_TO_MAP_VALUE:
3185 case PTR_TO_PACKET_META:
3186 case PTR_TO_PACKET_END:
3187 case PTR_TO_FLOW_KEYS:
3188 case CONST_PTR_TO_MAP:
3190 case PTR_TO_SOCK_COMMON:
3191 case PTR_TO_TCP_SOCK:
3192 case PTR_TO_XDP_SOCK:
3197 case PTR_TO_MAP_KEY:
3204 /* Does this register contain a constant zero? */
3205 static bool register_is_null(struct bpf_reg_state *reg)
3207 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
3210 static bool register_is_const(struct bpf_reg_state *reg)
3212 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
3215 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
3217 return tnum_is_unknown(reg->var_off) &&
3218 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
3219 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
3220 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
3221 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
3224 static bool register_is_bounded(struct bpf_reg_state *reg)
3226 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
3229 static bool __is_pointer_value(bool allow_ptr_leaks,
3230 const struct bpf_reg_state *reg)
3232 if (allow_ptr_leaks)
3235 return reg->type != SCALAR_VALUE;
3238 static void save_register_state(struct bpf_func_state *state,
3239 int spi, struct bpf_reg_state *reg,
3244 state->stack[spi].spilled_ptr = *reg;
3245 if (size == BPF_REG_SIZE)
3246 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
3248 for (i = BPF_REG_SIZE; i > BPF_REG_SIZE - size; i--)
3249 state->stack[spi].slot_type[i - 1] = STACK_SPILL;
3251 /* size < 8 bytes spill */
3253 scrub_spilled_slot(&state->stack[spi].slot_type[i - 1]);
3256 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
3257 * stack boundary and alignment are checked in check_mem_access()
3259 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
3260 /* stack frame we're writing to */
3261 struct bpf_func_state *state,
3262 int off, int size, int value_regno,
3265 struct bpf_func_state *cur; /* state of the current function */
3266 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
3267 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
3268 struct bpf_reg_state *reg = NULL;
3270 err = grow_stack_state(state, round_up(slot + 1, BPF_REG_SIZE));
3273 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
3274 * so it's aligned access and [off, off + size) are within stack limits
3276 if (!env->allow_ptr_leaks &&
3277 state->stack[spi].slot_type[0] == STACK_SPILL &&
3278 size != BPF_REG_SIZE) {
3279 verbose(env, "attempt to corrupt spilled pointer on stack\n");
3283 cur = env->cur_state->frame[env->cur_state->curframe];
3284 if (value_regno >= 0)
3285 reg = &cur->regs[value_regno];
3286 if (!env->bypass_spec_v4) {
3287 bool sanitize = reg && is_spillable_regtype(reg->type);
3289 for (i = 0; i < size; i++) {
3290 if (state->stack[spi].slot_type[i] == STACK_INVALID) {
3297 env->insn_aux_data[insn_idx].sanitize_stack_spill = true;
3300 mark_stack_slot_scratched(env, spi);
3301 if (reg && !(off % BPF_REG_SIZE) && register_is_bounded(reg) &&
3302 !register_is_null(reg) && env->bpf_capable) {
3303 if (dst_reg != BPF_REG_FP) {
3304 /* The backtracking logic can only recognize explicit
3305 * stack slot address like [fp - 8]. Other spill of
3306 * scalar via different register has to be conservative.
3307 * Backtrack from here and mark all registers as precise
3308 * that contributed into 'reg' being a constant.
3310 err = mark_chain_precision(env, value_regno);
3314 save_register_state(state, spi, reg, size);
3315 } else if (reg && is_spillable_regtype(reg->type)) {
3316 /* register containing pointer is being spilled into stack */
3317 if (size != BPF_REG_SIZE) {
3318 verbose_linfo(env, insn_idx, "; ");
3319 verbose(env, "invalid size of register spill\n");
3322 if (state != cur && reg->type == PTR_TO_STACK) {
3323 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
3326 save_register_state(state, spi, reg, size);
3328 u8 type = STACK_MISC;
3330 /* regular write of data into stack destroys any spilled ptr */
3331 state->stack[spi].spilled_ptr.type = NOT_INIT;
3332 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
3333 if (is_spilled_reg(&state->stack[spi]))
3334 for (i = 0; i < BPF_REG_SIZE; i++)
3335 scrub_spilled_slot(&state->stack[spi].slot_type[i]);
3337 /* only mark the slot as written if all 8 bytes were written
3338 * otherwise read propagation may incorrectly stop too soon
3339 * when stack slots are partially written.
3340 * This heuristic means that read propagation will be
3341 * conservative, since it will add reg_live_read marks
3342 * to stack slots all the way to first state when programs
3343 * writes+reads less than 8 bytes
3345 if (size == BPF_REG_SIZE)
3346 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
3348 /* when we zero initialize stack slots mark them as such */
3349 if (reg && register_is_null(reg)) {
3350 /* backtracking doesn't work for STACK_ZERO yet. */
3351 err = mark_chain_precision(env, value_regno);
3357 /* Mark slots affected by this stack write. */
3358 for (i = 0; i < size; i++)
3359 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
3365 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
3366 * known to contain a variable offset.
3367 * This function checks whether the write is permitted and conservatively
3368 * tracks the effects of the write, considering that each stack slot in the
3369 * dynamic range is potentially written to.
3371 * 'off' includes 'regno->off'.
3372 * 'value_regno' can be -1, meaning that an unknown value is being written to
3375 * Spilled pointers in range are not marked as written because we don't know
3376 * what's going to be actually written. This means that read propagation for
3377 * future reads cannot be terminated by this write.
3379 * For privileged programs, uninitialized stack slots are considered
3380 * initialized by this write (even though we don't know exactly what offsets
3381 * are going to be written to). The idea is that we don't want the verifier to
3382 * reject future reads that access slots written to through variable offsets.
3384 static int check_stack_write_var_off(struct bpf_verifier_env *env,
3385 /* func where register points to */
3386 struct bpf_func_state *state,
3387 int ptr_regno, int off, int size,
3388 int value_regno, int insn_idx)
3390 struct bpf_func_state *cur; /* state of the current function */
3391 int min_off, max_off;
3393 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
3394 bool writing_zero = false;
3395 /* set if the fact that we're writing a zero is used to let any
3396 * stack slots remain STACK_ZERO
3398 bool zero_used = false;
3400 cur = env->cur_state->frame[env->cur_state->curframe];
3401 ptr_reg = &cur->regs[ptr_regno];
3402 min_off = ptr_reg->smin_value + off;
3403 max_off = ptr_reg->smax_value + off + size;
3404 if (value_regno >= 0)
3405 value_reg = &cur->regs[value_regno];
3406 if (value_reg && register_is_null(value_reg))
3407 writing_zero = true;
3409 err = grow_stack_state(state, round_up(-min_off, BPF_REG_SIZE));
3414 /* Variable offset writes destroy any spilled pointers in range. */
3415 for (i = min_off; i < max_off; i++) {
3416 u8 new_type, *stype;
3420 spi = slot / BPF_REG_SIZE;
3421 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
3422 mark_stack_slot_scratched(env, spi);
3424 if (!env->allow_ptr_leaks && *stype != STACK_MISC && *stype != STACK_ZERO) {
3425 /* Reject the write if range we may write to has not
3426 * been initialized beforehand. If we didn't reject
3427 * here, the ptr status would be erased below (even
3428 * though not all slots are actually overwritten),
3429 * possibly opening the door to leaks.
3431 * We do however catch STACK_INVALID case below, and
3432 * only allow reading possibly uninitialized memory
3433 * later for CAP_PERFMON, as the write may not happen to
3436 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
3441 /* Erase all spilled pointers. */
3442 state->stack[spi].spilled_ptr.type = NOT_INIT;
3444 /* Update the slot type. */
3445 new_type = STACK_MISC;
3446 if (writing_zero && *stype == STACK_ZERO) {
3447 new_type = STACK_ZERO;
3450 /* If the slot is STACK_INVALID, we check whether it's OK to
3451 * pretend that it will be initialized by this write. The slot
3452 * might not actually be written to, and so if we mark it as
3453 * initialized future reads might leak uninitialized memory.
3454 * For privileged programs, we will accept such reads to slots
3455 * that may or may not be written because, if we're reject
3456 * them, the error would be too confusing.
3458 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
3459 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
3466 /* backtracking doesn't work for STACK_ZERO yet. */
3467 err = mark_chain_precision(env, value_regno);
3474 /* When register 'dst_regno' is assigned some values from stack[min_off,
3475 * max_off), we set the register's type according to the types of the
3476 * respective stack slots. If all the stack values are known to be zeros, then
3477 * so is the destination reg. Otherwise, the register is considered to be
3478 * SCALAR. This function does not deal with register filling; the caller must
3479 * ensure that all spilled registers in the stack range have been marked as
3482 static void mark_reg_stack_read(struct bpf_verifier_env *env,
3483 /* func where src register points to */
3484 struct bpf_func_state *ptr_state,
3485 int min_off, int max_off, int dst_regno)
3487 struct bpf_verifier_state *vstate = env->cur_state;
3488 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3493 for (i = min_off; i < max_off; i++) {
3495 spi = slot / BPF_REG_SIZE;
3496 stype = ptr_state->stack[spi].slot_type;
3497 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
3501 if (zeros == max_off - min_off) {
3502 /* any access_size read into register is zero extended,
3503 * so the whole register == const_zero
3505 __mark_reg_const_zero(&state->regs[dst_regno]);
3506 /* backtracking doesn't support STACK_ZERO yet,
3507 * so mark it precise here, so that later
3508 * backtracking can stop here.
3509 * Backtracking may not need this if this register
3510 * doesn't participate in pointer adjustment.
3511 * Forward propagation of precise flag is not
3512 * necessary either. This mark is only to stop
3513 * backtracking. Any register that contributed
3514 * to const 0 was marked precise before spill.
3516 state->regs[dst_regno].precise = true;
3518 /* have read misc data from the stack */
3519 mark_reg_unknown(env, state->regs, dst_regno);
3521 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
3524 /* Read the stack at 'off' and put the results into the register indicated by
3525 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
3528 * 'dst_regno' can be -1, meaning that the read value is not going to a
3531 * The access is assumed to be within the current stack bounds.
3533 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
3534 /* func where src register points to */
3535 struct bpf_func_state *reg_state,
3536 int off, int size, int dst_regno)
3538 struct bpf_verifier_state *vstate = env->cur_state;
3539 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3540 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
3541 struct bpf_reg_state *reg;
3544 stype = reg_state->stack[spi].slot_type;
3545 reg = ®_state->stack[spi].spilled_ptr;
3547 if (is_spilled_reg(®_state->stack[spi])) {
3550 for (i = BPF_REG_SIZE - 1; i > 0 && stype[i - 1] == STACK_SPILL; i--)
3553 if (size != BPF_REG_SIZE || spill_size != BPF_REG_SIZE) {
3554 if (reg->type != SCALAR_VALUE) {
3555 verbose_linfo(env, env->insn_idx, "; ");
3556 verbose(env, "invalid size of register fill\n");
3560 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
3564 if (!(off % BPF_REG_SIZE) && size == spill_size) {
3565 /* The earlier check_reg_arg() has decided the
3566 * subreg_def for this insn. Save it first.
3568 s32 subreg_def = state->regs[dst_regno].subreg_def;
3570 state->regs[dst_regno] = *reg;
3571 state->regs[dst_regno].subreg_def = subreg_def;
3573 for (i = 0; i < size; i++) {
3574 type = stype[(slot - i) % BPF_REG_SIZE];
3575 if (type == STACK_SPILL)
3577 if (type == STACK_MISC)
3579 verbose(env, "invalid read from stack off %d+%d size %d\n",
3583 mark_reg_unknown(env, state->regs, dst_regno);
3585 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
3589 if (dst_regno >= 0) {
3590 /* restore register state from stack */
3591 state->regs[dst_regno] = *reg;
3592 /* mark reg as written since spilled pointer state likely
3593 * has its liveness marks cleared by is_state_visited()
3594 * which resets stack/reg liveness for state transitions
3596 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
3597 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
3598 /* If dst_regno==-1, the caller is asking us whether
3599 * it is acceptable to use this value as a SCALAR_VALUE
3601 * We must not allow unprivileged callers to do that
3602 * with spilled pointers.
3604 verbose(env, "leaking pointer from stack off %d\n",
3608 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
3610 for (i = 0; i < size; i++) {
3611 type = stype[(slot - i) % BPF_REG_SIZE];
3612 if (type == STACK_MISC)
3614 if (type == STACK_ZERO)
3616 verbose(env, "invalid read from stack off %d+%d size %d\n",
3620 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
3622 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
3627 enum bpf_access_src {
3628 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
3629 ACCESS_HELPER = 2, /* the access is performed by a helper */
3632 static int check_stack_range_initialized(struct bpf_verifier_env *env,
3633 int regno, int off, int access_size,
3634 bool zero_size_allowed,
3635 enum bpf_access_src type,
3636 struct bpf_call_arg_meta *meta);
3638 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
3640 return cur_regs(env) + regno;
3643 /* Read the stack at 'ptr_regno + off' and put the result into the register
3645 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
3646 * but not its variable offset.
3647 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
3649 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
3650 * filling registers (i.e. reads of spilled register cannot be detected when
3651 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
3652 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
3653 * offset; for a fixed offset check_stack_read_fixed_off should be used
3656 static int check_stack_read_var_off(struct bpf_verifier_env *env,
3657 int ptr_regno, int off, int size, int dst_regno)
3659 /* The state of the source register. */
3660 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3661 struct bpf_func_state *ptr_state = func(env, reg);
3663 int min_off, max_off;
3665 /* Note that we pass a NULL meta, so raw access will not be permitted.
3667 err = check_stack_range_initialized(env, ptr_regno, off, size,
3668 false, ACCESS_DIRECT, NULL);
3672 min_off = reg->smin_value + off;
3673 max_off = reg->smax_value + off;
3674 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
3678 /* check_stack_read dispatches to check_stack_read_fixed_off or
3679 * check_stack_read_var_off.
3681 * The caller must ensure that the offset falls within the allocated stack
3684 * 'dst_regno' is a register which will receive the value from the stack. It
3685 * can be -1, meaning that the read value is not going to a register.
3687 static int check_stack_read(struct bpf_verifier_env *env,
3688 int ptr_regno, int off, int size,
3691 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3692 struct bpf_func_state *state = func(env, reg);
3694 /* Some accesses are only permitted with a static offset. */
3695 bool var_off = !tnum_is_const(reg->var_off);
3697 /* The offset is required to be static when reads don't go to a
3698 * register, in order to not leak pointers (see
3699 * check_stack_read_fixed_off).
3701 if (dst_regno < 0 && var_off) {
3704 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3705 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
3709 /* Variable offset is prohibited for unprivileged mode for simplicity
3710 * since it requires corresponding support in Spectre masking for stack
3711 * ALU. See also retrieve_ptr_limit().
3713 if (!env->bypass_spec_v1 && var_off) {
3716 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3717 verbose(env, "R%d variable offset stack access prohibited for !root, var_off=%s\n",
3723 off += reg->var_off.value;
3724 err = check_stack_read_fixed_off(env, state, off, size,
3727 /* Variable offset stack reads need more conservative handling
3728 * than fixed offset ones. Note that dst_regno >= 0 on this
3731 err = check_stack_read_var_off(env, ptr_regno, off, size,
3738 /* check_stack_write dispatches to check_stack_write_fixed_off or
3739 * check_stack_write_var_off.
3741 * 'ptr_regno' is the register used as a pointer into the stack.
3742 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
3743 * 'value_regno' is the register whose value we're writing to the stack. It can
3744 * be -1, meaning that we're not writing from a register.
3746 * The caller must ensure that the offset falls within the maximum stack size.
3748 static int check_stack_write(struct bpf_verifier_env *env,
3749 int ptr_regno, int off, int size,
3750 int value_regno, int insn_idx)
3752 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3753 struct bpf_func_state *state = func(env, reg);
3756 if (tnum_is_const(reg->var_off)) {
3757 off += reg->var_off.value;
3758 err = check_stack_write_fixed_off(env, state, off, size,
3759 value_regno, insn_idx);
3761 /* Variable offset stack reads need more conservative handling
3762 * than fixed offset ones.
3764 err = check_stack_write_var_off(env, state,
3765 ptr_regno, off, size,
3766 value_regno, insn_idx);
3771 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
3772 int off, int size, enum bpf_access_type type)
3774 struct bpf_reg_state *regs = cur_regs(env);
3775 struct bpf_map *map = regs[regno].map_ptr;
3776 u32 cap = bpf_map_flags_to_cap(map);
3778 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
3779 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
3780 map->value_size, off, size);
3784 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
3785 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
3786 map->value_size, off, size);
3793 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
3794 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
3795 int off, int size, u32 mem_size,
3796 bool zero_size_allowed)
3798 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
3799 struct bpf_reg_state *reg;
3801 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
3804 reg = &cur_regs(env)[regno];
3805 switch (reg->type) {
3806 case PTR_TO_MAP_KEY:
3807 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
3808 mem_size, off, size);
3810 case PTR_TO_MAP_VALUE:
3811 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
3812 mem_size, off, size);
3815 case PTR_TO_PACKET_META:
3816 case PTR_TO_PACKET_END:
3817 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
3818 off, size, regno, reg->id, off, mem_size);
3822 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
3823 mem_size, off, size);
3829 /* check read/write into a memory region with possible variable offset */
3830 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
3831 int off, int size, u32 mem_size,
3832 bool zero_size_allowed)
3834 struct bpf_verifier_state *vstate = env->cur_state;
3835 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3836 struct bpf_reg_state *reg = &state->regs[regno];
3839 /* We may have adjusted the register pointing to memory region, so we
3840 * need to try adding each of min_value and max_value to off
3841 * to make sure our theoretical access will be safe.
3843 * The minimum value is only important with signed
3844 * comparisons where we can't assume the floor of a
3845 * value is 0. If we are using signed variables for our
3846 * index'es we need to make sure that whatever we use
3847 * will have a set floor within our range.
3849 if (reg->smin_value < 0 &&
3850 (reg->smin_value == S64_MIN ||
3851 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
3852 reg->smin_value + off < 0)) {
3853 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3857 err = __check_mem_access(env, regno, reg->smin_value + off, size,
3858 mem_size, zero_size_allowed);
3860 verbose(env, "R%d min value is outside of the allowed memory range\n",
3865 /* If we haven't set a max value then we need to bail since we can't be
3866 * sure we won't do bad things.
3867 * If reg->umax_value + off could overflow, treat that as unbounded too.
3869 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
3870 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
3874 err = __check_mem_access(env, regno, reg->umax_value + off, size,
3875 mem_size, zero_size_allowed);
3877 verbose(env, "R%d max value is outside of the allowed memory range\n",
3885 static int __check_ptr_off_reg(struct bpf_verifier_env *env,
3886 const struct bpf_reg_state *reg, int regno,
3889 /* Access to this pointer-typed register or passing it to a helper
3890 * is only allowed in its original, unmodified form.
3894 verbose(env, "negative offset %s ptr R%d off=%d disallowed\n",
3895 reg_type_str(env, reg->type), regno, reg->off);
3899 if (!fixed_off_ok && reg->off) {
3900 verbose(env, "dereference of modified %s ptr R%d off=%d disallowed\n",
3901 reg_type_str(env, reg->type), regno, reg->off);
3905 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3908 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3909 verbose(env, "variable %s access var_off=%s disallowed\n",
3910 reg_type_str(env, reg->type), tn_buf);
3917 int check_ptr_off_reg(struct bpf_verifier_env *env,
3918 const struct bpf_reg_state *reg, int regno)
3920 return __check_ptr_off_reg(env, reg, regno, false);
3923 static int map_kptr_match_type(struct bpf_verifier_env *env,
3924 struct btf_field *kptr_field,
3925 struct bpf_reg_state *reg, u32 regno)
3927 const char *targ_name = kernel_type_name(kptr_field->kptr.btf, kptr_field->kptr.btf_id);
3928 int perm_flags = PTR_MAYBE_NULL | PTR_TRUSTED;
3929 const char *reg_name = "";
3931 /* Only unreferenced case accepts untrusted pointers */
3932 if (kptr_field->type == BPF_KPTR_UNREF)
3933 perm_flags |= PTR_UNTRUSTED;
3935 if (base_type(reg->type) != PTR_TO_BTF_ID || (type_flag(reg->type) & ~perm_flags))
3938 if (!btf_is_kernel(reg->btf)) {
3939 verbose(env, "R%d must point to kernel BTF\n", regno);
3942 /* We need to verify reg->type and reg->btf, before accessing reg->btf */
3943 reg_name = kernel_type_name(reg->btf, reg->btf_id);
3945 /* For ref_ptr case, release function check should ensure we get one
3946 * referenced PTR_TO_BTF_ID, and that its fixed offset is 0. For the
3947 * normal store of unreferenced kptr, we must ensure var_off is zero.
3948 * Since ref_ptr cannot be accessed directly by BPF insns, checks for
3949 * reg->off and reg->ref_obj_id are not needed here.
3951 if (__check_ptr_off_reg(env, reg, regno, true))
3954 /* A full type match is needed, as BTF can be vmlinux or module BTF, and
3955 * we also need to take into account the reg->off.
3957 * We want to support cases like:
3965 * v = func(); // PTR_TO_BTF_ID
3966 * val->foo = v; // reg->off is zero, btf and btf_id match type
3967 * val->bar = &v->br; // reg->off is still zero, but we need to retry with
3968 * // first member type of struct after comparison fails
3969 * val->baz = &v->bz; // reg->off is non-zero, so struct needs to be walked
3972 * In the kptr_ref case, check_func_arg_reg_off already ensures reg->off
3973 * is zero. We must also ensure that btf_struct_ids_match does not walk
3974 * the struct to match type against first member of struct, i.e. reject
3975 * second case from above. Hence, when type is BPF_KPTR_REF, we set
3976 * strict mode to true for type match.
3978 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
3979 kptr_field->kptr.btf, kptr_field->kptr.btf_id,
3980 kptr_field->type == BPF_KPTR_REF))
3984 verbose(env, "invalid kptr access, R%d type=%s%s ", regno,
3985 reg_type_str(env, reg->type), reg_name);
3986 verbose(env, "expected=%s%s", reg_type_str(env, PTR_TO_BTF_ID), targ_name);
3987 if (kptr_field->type == BPF_KPTR_UNREF)
3988 verbose(env, " or %s%s\n", reg_type_str(env, PTR_TO_BTF_ID | PTR_UNTRUSTED),
3995 static int check_map_kptr_access(struct bpf_verifier_env *env, u32 regno,
3996 int value_regno, int insn_idx,
3997 struct btf_field *kptr_field)
3999 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
4000 int class = BPF_CLASS(insn->code);
4001 struct bpf_reg_state *val_reg;
4003 /* Things we already checked for in check_map_access and caller:
4004 * - Reject cases where variable offset may touch kptr
4005 * - size of access (must be BPF_DW)
4006 * - tnum_is_const(reg->var_off)
4007 * - kptr_field->offset == off + reg->var_off.value
4009 /* Only BPF_[LDX,STX,ST] | BPF_MEM | BPF_DW is supported */
4010 if (BPF_MODE(insn->code) != BPF_MEM) {
4011 verbose(env, "kptr in map can only be accessed using BPF_MEM instruction mode\n");
4015 /* We only allow loading referenced kptr, since it will be marked as
4016 * untrusted, similar to unreferenced kptr.
4018 if (class != BPF_LDX && kptr_field->type == BPF_KPTR_REF) {
4019 verbose(env, "store to referenced kptr disallowed\n");
4023 if (class == BPF_LDX) {
4024 val_reg = reg_state(env, value_regno);
4025 /* We can simply mark the value_regno receiving the pointer
4026 * value from map as PTR_TO_BTF_ID, with the correct type.
4028 mark_btf_ld_reg(env, cur_regs(env), value_regno, PTR_TO_BTF_ID, kptr_field->kptr.btf,
4029 kptr_field->kptr.btf_id, PTR_MAYBE_NULL | PTR_UNTRUSTED);
4030 /* For mark_ptr_or_null_reg */
4031 val_reg->id = ++env->id_gen;
4032 } else if (class == BPF_STX) {
4033 val_reg = reg_state(env, value_regno);
4034 if (!register_is_null(val_reg) &&
4035 map_kptr_match_type(env, kptr_field, val_reg, value_regno))
4037 } else if (class == BPF_ST) {
4039 verbose(env, "BPF_ST imm must be 0 when storing to kptr at off=%u\n",
4040 kptr_field->offset);
4044 verbose(env, "kptr in map can only be accessed using BPF_LDX/BPF_STX/BPF_ST\n");
4050 /* check read/write into a map element with possible variable offset */
4051 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
4052 int off, int size, bool zero_size_allowed,
4053 enum bpf_access_src src)
4055 struct bpf_verifier_state *vstate = env->cur_state;
4056 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4057 struct bpf_reg_state *reg = &state->regs[regno];
4058 struct bpf_map *map = reg->map_ptr;
4059 struct btf_record *rec;
4062 err = check_mem_region_access(env, regno, off, size, map->value_size,
4067 if (IS_ERR_OR_NULL(map->record))
4070 for (i = 0; i < rec->cnt; i++) {
4071 struct btf_field *field = &rec->fields[i];
4072 u32 p = field->offset;
4074 /* If any part of a field can be touched by load/store, reject
4075 * this program. To check that [x1, x2) overlaps with [y1, y2),
4076 * it is sufficient to check x1 < y2 && y1 < x2.
4078 if (reg->smin_value + off < p + btf_field_type_size(field->type) &&
4079 p < reg->umax_value + off + size) {
4080 switch (field->type) {
4081 case BPF_KPTR_UNREF:
4083 if (src != ACCESS_DIRECT) {
4084 verbose(env, "kptr cannot be accessed indirectly by helper\n");
4087 if (!tnum_is_const(reg->var_off)) {
4088 verbose(env, "kptr access cannot have variable offset\n");
4091 if (p != off + reg->var_off.value) {
4092 verbose(env, "kptr access misaligned expected=%u off=%llu\n",
4093 p, off + reg->var_off.value);
4096 if (size != bpf_size_to_bytes(BPF_DW)) {
4097 verbose(env, "kptr access size must be BPF_DW\n");
4102 verbose(env, "%s cannot be accessed directly by load/store\n",
4103 btf_field_type_name(field->type));
4111 #define MAX_PACKET_OFF 0xffff
4113 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
4114 const struct bpf_call_arg_meta *meta,
4115 enum bpf_access_type t)
4117 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
4119 switch (prog_type) {
4120 /* Program types only with direct read access go here! */
4121 case BPF_PROG_TYPE_LWT_IN:
4122 case BPF_PROG_TYPE_LWT_OUT:
4123 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
4124 case BPF_PROG_TYPE_SK_REUSEPORT:
4125 case BPF_PROG_TYPE_FLOW_DISSECTOR:
4126 case BPF_PROG_TYPE_CGROUP_SKB:
4131 /* Program types with direct read + write access go here! */
4132 case BPF_PROG_TYPE_SCHED_CLS:
4133 case BPF_PROG_TYPE_SCHED_ACT:
4134 case BPF_PROG_TYPE_XDP:
4135 case BPF_PROG_TYPE_LWT_XMIT:
4136 case BPF_PROG_TYPE_SK_SKB:
4137 case BPF_PROG_TYPE_SK_MSG:
4139 return meta->pkt_access;
4141 env->seen_direct_write = true;
4144 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
4146 env->seen_direct_write = true;
4155 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
4156 int size, bool zero_size_allowed)
4158 struct bpf_reg_state *regs = cur_regs(env);
4159 struct bpf_reg_state *reg = ®s[regno];
4162 /* We may have added a variable offset to the packet pointer; but any
4163 * reg->range we have comes after that. We are only checking the fixed
4167 /* We don't allow negative numbers, because we aren't tracking enough
4168 * detail to prove they're safe.
4170 if (reg->smin_value < 0) {
4171 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
4176 err = reg->range < 0 ? -EINVAL :
4177 __check_mem_access(env, regno, off, size, reg->range,
4180 verbose(env, "R%d offset is outside of the packet\n", regno);
4184 /* __check_mem_access has made sure "off + size - 1" is within u16.
4185 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
4186 * otherwise find_good_pkt_pointers would have refused to set range info
4187 * that __check_mem_access would have rejected this pkt access.
4188 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
4190 env->prog->aux->max_pkt_offset =
4191 max_t(u32, env->prog->aux->max_pkt_offset,
4192 off + reg->umax_value + size - 1);
4197 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
4198 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
4199 enum bpf_access_type t, enum bpf_reg_type *reg_type,
4200 struct btf **btf, u32 *btf_id)
4202 struct bpf_insn_access_aux info = {
4203 .reg_type = *reg_type,
4207 if (env->ops->is_valid_access &&
4208 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
4209 /* A non zero info.ctx_field_size indicates that this field is a
4210 * candidate for later verifier transformation to load the whole
4211 * field and then apply a mask when accessed with a narrower
4212 * access than actual ctx access size. A zero info.ctx_field_size
4213 * will only allow for whole field access and rejects any other
4214 * type of narrower access.
4216 *reg_type = info.reg_type;
4218 if (base_type(*reg_type) == PTR_TO_BTF_ID) {
4220 *btf_id = info.btf_id;
4222 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
4224 /* remember the offset of last byte accessed in ctx */
4225 if (env->prog->aux->max_ctx_offset < off + size)
4226 env->prog->aux->max_ctx_offset = off + size;
4230 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
4234 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
4237 if (size < 0 || off < 0 ||
4238 (u64)off + size > sizeof(struct bpf_flow_keys)) {
4239 verbose(env, "invalid access to flow keys off=%d size=%d\n",
4246 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
4247 u32 regno, int off, int size,
4248 enum bpf_access_type t)
4250 struct bpf_reg_state *regs = cur_regs(env);
4251 struct bpf_reg_state *reg = ®s[regno];
4252 struct bpf_insn_access_aux info = {};
4255 if (reg->smin_value < 0) {
4256 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
4261 switch (reg->type) {
4262 case PTR_TO_SOCK_COMMON:
4263 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
4266 valid = bpf_sock_is_valid_access(off, size, t, &info);
4268 case PTR_TO_TCP_SOCK:
4269 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
4271 case PTR_TO_XDP_SOCK:
4272 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
4280 env->insn_aux_data[insn_idx].ctx_field_size =
4281 info.ctx_field_size;
4285 verbose(env, "R%d invalid %s access off=%d size=%d\n",
4286 regno, reg_type_str(env, reg->type), off, size);
4291 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
4293 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
4296 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
4298 const struct bpf_reg_state *reg = reg_state(env, regno);
4300 return reg->type == PTR_TO_CTX;
4303 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
4305 const struct bpf_reg_state *reg = reg_state(env, regno);
4307 return type_is_sk_pointer(reg->type);
4310 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
4312 const struct bpf_reg_state *reg = reg_state(env, regno);
4314 return type_is_pkt_pointer(reg->type);
4317 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
4319 const struct bpf_reg_state *reg = reg_state(env, regno);
4321 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
4322 return reg->type == PTR_TO_FLOW_KEYS;
4325 static bool is_trusted_reg(const struct bpf_reg_state *reg)
4327 /* A referenced register is always trusted. */
4328 if (reg->ref_obj_id)
4331 /* If a register is not referenced, it is trusted if it has the
4332 * MEM_ALLOC or PTR_TRUSTED type modifiers, and no others. Some of the
4333 * other type modifiers may be safe, but we elect to take an opt-in
4334 * approach here as some (e.g. PTR_UNTRUSTED and PTR_MAYBE_NULL) are
4337 * Eventually, we should make PTR_TRUSTED the single source of truth
4338 * for whether a register is trusted.
4340 return type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS &&
4341 !bpf_type_has_unsafe_modifiers(reg->type);
4344 static bool is_rcu_reg(const struct bpf_reg_state *reg)
4346 return reg->type & MEM_RCU;
4349 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
4350 const struct bpf_reg_state *reg,
4351 int off, int size, bool strict)
4353 struct tnum reg_off;
4356 /* Byte size accesses are always allowed. */
4357 if (!strict || size == 1)
4360 /* For platforms that do not have a Kconfig enabling
4361 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
4362 * NET_IP_ALIGN is universally set to '2'. And on platforms
4363 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
4364 * to this code only in strict mode where we want to emulate
4365 * the NET_IP_ALIGN==2 checking. Therefore use an
4366 * unconditional IP align value of '2'.
4370 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
4371 if (!tnum_is_aligned(reg_off, size)) {
4374 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4376 "misaligned packet access off %d+%s+%d+%d size %d\n",
4377 ip_align, tn_buf, reg->off, off, size);
4384 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
4385 const struct bpf_reg_state *reg,
4386 const char *pointer_desc,
4387 int off, int size, bool strict)
4389 struct tnum reg_off;
4391 /* Byte size accesses are always allowed. */
4392 if (!strict || size == 1)
4395 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
4396 if (!tnum_is_aligned(reg_off, size)) {
4399 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4400 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
4401 pointer_desc, tn_buf, reg->off, off, size);
4408 static int check_ptr_alignment(struct bpf_verifier_env *env,
4409 const struct bpf_reg_state *reg, int off,
4410 int size, bool strict_alignment_once)
4412 bool strict = env->strict_alignment || strict_alignment_once;
4413 const char *pointer_desc = "";
4415 switch (reg->type) {
4417 case PTR_TO_PACKET_META:
4418 /* Special case, because of NET_IP_ALIGN. Given metadata sits
4419 * right in front, treat it the very same way.
4421 return check_pkt_ptr_alignment(env, reg, off, size, strict);
4422 case PTR_TO_FLOW_KEYS:
4423 pointer_desc = "flow keys ";
4425 case PTR_TO_MAP_KEY:
4426 pointer_desc = "key ";
4428 case PTR_TO_MAP_VALUE:
4429 pointer_desc = "value ";
4432 pointer_desc = "context ";
4435 pointer_desc = "stack ";
4436 /* The stack spill tracking logic in check_stack_write_fixed_off()
4437 * and check_stack_read_fixed_off() relies on stack accesses being
4443 pointer_desc = "sock ";
4445 case PTR_TO_SOCK_COMMON:
4446 pointer_desc = "sock_common ";
4448 case PTR_TO_TCP_SOCK:
4449 pointer_desc = "tcp_sock ";
4451 case PTR_TO_XDP_SOCK:
4452 pointer_desc = "xdp_sock ";
4457 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
4461 static int update_stack_depth(struct bpf_verifier_env *env,
4462 const struct bpf_func_state *func,
4465 u16 stack = env->subprog_info[func->subprogno].stack_depth;
4470 /* update known max for given subprogram */
4471 env->subprog_info[func->subprogno].stack_depth = -off;
4475 /* starting from main bpf function walk all instructions of the function
4476 * and recursively walk all callees that given function can call.
4477 * Ignore jump and exit insns.
4478 * Since recursion is prevented by check_cfg() this algorithm
4479 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
4481 static int check_max_stack_depth(struct bpf_verifier_env *env)
4483 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
4484 struct bpf_subprog_info *subprog = env->subprog_info;
4485 struct bpf_insn *insn = env->prog->insnsi;
4486 bool tail_call_reachable = false;
4487 int ret_insn[MAX_CALL_FRAMES];
4488 int ret_prog[MAX_CALL_FRAMES];
4492 /* protect against potential stack overflow that might happen when
4493 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
4494 * depth for such case down to 256 so that the worst case scenario
4495 * would result in 8k stack size (32 which is tailcall limit * 256 =
4498 * To get the idea what might happen, see an example:
4499 * func1 -> sub rsp, 128
4500 * subfunc1 -> sub rsp, 256
4501 * tailcall1 -> add rsp, 256
4502 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
4503 * subfunc2 -> sub rsp, 64
4504 * subfunc22 -> sub rsp, 128
4505 * tailcall2 -> add rsp, 128
4506 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
4508 * tailcall will unwind the current stack frame but it will not get rid
4509 * of caller's stack as shown on the example above.
4511 if (idx && subprog[idx].has_tail_call && depth >= 256) {
4513 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
4517 /* round up to 32-bytes, since this is granularity
4518 * of interpreter stack size
4520 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
4521 if (depth > MAX_BPF_STACK) {
4522 verbose(env, "combined stack size of %d calls is %d. Too large\n",
4527 subprog_end = subprog[idx + 1].start;
4528 for (; i < subprog_end; i++) {
4531 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
4533 /* remember insn and function to return to */
4534 ret_insn[frame] = i + 1;
4535 ret_prog[frame] = idx;
4537 /* find the callee */
4538 next_insn = i + insn[i].imm + 1;
4539 idx = find_subprog(env, next_insn);
4541 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
4545 if (subprog[idx].is_async_cb) {
4546 if (subprog[idx].has_tail_call) {
4547 verbose(env, "verifier bug. subprog has tail_call and async cb\n");
4550 /* async callbacks don't increase bpf prog stack size */
4555 if (subprog[idx].has_tail_call)
4556 tail_call_reachable = true;
4559 if (frame >= MAX_CALL_FRAMES) {
4560 verbose(env, "the call stack of %d frames is too deep !\n",
4566 /* if tail call got detected across bpf2bpf calls then mark each of the
4567 * currently present subprog frames as tail call reachable subprogs;
4568 * this info will be utilized by JIT so that we will be preserving the
4569 * tail call counter throughout bpf2bpf calls combined with tailcalls
4571 if (tail_call_reachable)
4572 for (j = 0; j < frame; j++)
4573 subprog[ret_prog[j]].tail_call_reachable = true;
4574 if (subprog[0].tail_call_reachable)
4575 env->prog->aux->tail_call_reachable = true;
4577 /* end of for() loop means the last insn of the 'subprog'
4578 * was reached. Doesn't matter whether it was JA or EXIT
4582 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
4584 i = ret_insn[frame];
4585 idx = ret_prog[frame];
4589 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
4590 static int get_callee_stack_depth(struct bpf_verifier_env *env,
4591 const struct bpf_insn *insn, int idx)
4593 int start = idx + insn->imm + 1, subprog;
4595 subprog = find_subprog(env, start);
4597 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
4601 return env->subprog_info[subprog].stack_depth;
4605 static int __check_buffer_access(struct bpf_verifier_env *env,
4606 const char *buf_info,
4607 const struct bpf_reg_state *reg,
4608 int regno, int off, int size)
4612 "R%d invalid %s buffer access: off=%d, size=%d\n",
4613 regno, buf_info, off, size);
4616 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4619 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4621 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
4622 regno, off, tn_buf);
4629 static int check_tp_buffer_access(struct bpf_verifier_env *env,
4630 const struct bpf_reg_state *reg,
4631 int regno, int off, int size)
4635 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
4639 if (off + size > env->prog->aux->max_tp_access)
4640 env->prog->aux->max_tp_access = off + size;
4645 static int check_buffer_access(struct bpf_verifier_env *env,
4646 const struct bpf_reg_state *reg,
4647 int regno, int off, int size,
4648 bool zero_size_allowed,
4651 const char *buf_info = type_is_rdonly_mem(reg->type) ? "rdonly" : "rdwr";
4654 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
4658 if (off + size > *max_access)
4659 *max_access = off + size;
4664 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
4665 static void zext_32_to_64(struct bpf_reg_state *reg)
4667 reg->var_off = tnum_subreg(reg->var_off);
4668 __reg_assign_32_into_64(reg);
4671 /* truncate register to smaller size (in bytes)
4672 * must be called with size < BPF_REG_SIZE
4674 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
4678 /* clear high bits in bit representation */
4679 reg->var_off = tnum_cast(reg->var_off, size);
4681 /* fix arithmetic bounds */
4682 mask = ((u64)1 << (size * 8)) - 1;
4683 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
4684 reg->umin_value &= mask;
4685 reg->umax_value &= mask;
4687 reg->umin_value = 0;
4688 reg->umax_value = mask;
4690 reg->smin_value = reg->umin_value;
4691 reg->smax_value = reg->umax_value;
4693 /* If size is smaller than 32bit register the 32bit register
4694 * values are also truncated so we push 64-bit bounds into
4695 * 32-bit bounds. Above were truncated < 32-bits already.
4699 __reg_combine_64_into_32(reg);
4702 static bool bpf_map_is_rdonly(const struct bpf_map *map)
4704 /* A map is considered read-only if the following condition are true:
4706 * 1) BPF program side cannot change any of the map content. The
4707 * BPF_F_RDONLY_PROG flag is throughout the lifetime of a map
4708 * and was set at map creation time.
4709 * 2) The map value(s) have been initialized from user space by a
4710 * loader and then "frozen", such that no new map update/delete
4711 * operations from syscall side are possible for the rest of
4712 * the map's lifetime from that point onwards.
4713 * 3) Any parallel/pending map update/delete operations from syscall
4714 * side have been completed. Only after that point, it's safe to
4715 * assume that map value(s) are immutable.
4717 return (map->map_flags & BPF_F_RDONLY_PROG) &&
4718 READ_ONCE(map->frozen) &&
4719 !bpf_map_write_active(map);
4722 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
4728 err = map->ops->map_direct_value_addr(map, &addr, off);
4731 ptr = (void *)(long)addr + off;
4735 *val = (u64)*(u8 *)ptr;
4738 *val = (u64)*(u16 *)ptr;
4741 *val = (u64)*(u32 *)ptr;
4752 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
4753 struct bpf_reg_state *regs,
4754 int regno, int off, int size,
4755 enum bpf_access_type atype,
4758 struct bpf_reg_state *reg = regs + regno;
4759 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
4760 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
4761 enum bpf_type_flag flag = 0;
4765 if (!env->allow_ptr_leaks) {
4767 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
4771 if (!env->prog->gpl_compatible && btf_is_kernel(reg->btf)) {
4773 "Cannot access kernel 'struct %s' from non-GPL compatible program\n",
4779 "R%d is ptr_%s invalid negative access: off=%d\n",
4783 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4786 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4788 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
4789 regno, tname, off, tn_buf);
4793 if (reg->type & MEM_USER) {
4795 "R%d is ptr_%s access user memory: off=%d\n",
4800 if (reg->type & MEM_PERCPU) {
4802 "R%d is ptr_%s access percpu memory: off=%d\n",
4807 if (env->ops->btf_struct_access && !type_is_alloc(reg->type)) {
4808 if (!btf_is_kernel(reg->btf)) {
4809 verbose(env, "verifier internal error: reg->btf must be kernel btf\n");
4812 ret = env->ops->btf_struct_access(&env->log, reg, off, size, atype, &btf_id, &flag);
4814 /* Writes are permitted with default btf_struct_access for
4815 * program allocated objects (which always have ref_obj_id > 0),
4816 * but not for untrusted PTR_TO_BTF_ID | MEM_ALLOC.
4818 if (atype != BPF_READ && reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
4819 verbose(env, "only read is supported\n");
4823 if (type_is_alloc(reg->type) && !reg->ref_obj_id) {
4824 verbose(env, "verifier internal error: ref_obj_id for allocated object must be non-zero\n");
4828 ret = btf_struct_access(&env->log, reg, off, size, atype, &btf_id, &flag);
4834 /* If this is an untrusted pointer, all pointers formed by walking it
4835 * also inherit the untrusted flag.
4837 if (type_flag(reg->type) & PTR_UNTRUSTED)
4838 flag |= PTR_UNTRUSTED;
4840 /* By default any pointer obtained from walking a trusted pointer is
4841 * no longer trusted except the rcu case below.
4843 flag &= ~PTR_TRUSTED;
4845 if (flag & MEM_RCU) {
4846 /* Mark value register as MEM_RCU only if it is protected by
4847 * bpf_rcu_read_lock() and the ptr reg is rcu or trusted. MEM_RCU
4848 * itself can already indicate trustedness inside the rcu
4849 * read lock region. Also mark rcu pointer as PTR_MAYBE_NULL since
4850 * it could be null in some cases.
4852 if (!env->cur_state->active_rcu_lock ||
4853 !(is_trusted_reg(reg) || is_rcu_reg(reg)))
4856 flag |= PTR_MAYBE_NULL;
4857 } else if (reg->type & MEM_RCU) {
4858 /* ptr (reg) is marked as MEM_RCU, but the struct field is not tagged
4859 * with __rcu. Mark the flag as PTR_UNTRUSTED conservatively.
4861 flag |= PTR_UNTRUSTED;
4864 if (atype == BPF_READ && value_regno >= 0)
4865 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id, flag);
4870 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
4871 struct bpf_reg_state *regs,
4872 int regno, int off, int size,
4873 enum bpf_access_type atype,
4876 struct bpf_reg_state *reg = regs + regno;
4877 struct bpf_map *map = reg->map_ptr;
4878 struct bpf_reg_state map_reg;
4879 enum bpf_type_flag flag = 0;
4880 const struct btf_type *t;
4886 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
4890 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
4891 verbose(env, "map_ptr access not supported for map type %d\n",
4896 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
4897 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
4899 if (!env->allow_ptr_leaks) {
4901 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
4907 verbose(env, "R%d is %s invalid negative access: off=%d\n",
4912 if (atype != BPF_READ) {
4913 verbose(env, "only read from %s is supported\n", tname);
4917 /* Simulate access to a PTR_TO_BTF_ID */
4918 memset(&map_reg, 0, sizeof(map_reg));
4919 mark_btf_ld_reg(env, &map_reg, 0, PTR_TO_BTF_ID, btf_vmlinux, *map->ops->map_btf_id, 0);
4920 ret = btf_struct_access(&env->log, &map_reg, off, size, atype, &btf_id, &flag);
4924 if (value_regno >= 0)
4925 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id, flag);
4930 /* Check that the stack access at the given offset is within bounds. The
4931 * maximum valid offset is -1.
4933 * The minimum valid offset is -MAX_BPF_STACK for writes, and
4934 * -state->allocated_stack for reads.
4936 static int check_stack_slot_within_bounds(int off,
4937 struct bpf_func_state *state,
4938 enum bpf_access_type t)
4943 min_valid_off = -MAX_BPF_STACK;
4945 min_valid_off = -state->allocated_stack;
4947 if (off < min_valid_off || off > -1)
4952 /* Check that the stack access at 'regno + off' falls within the maximum stack
4955 * 'off' includes `regno->offset`, but not its dynamic part (if any).
4957 static int check_stack_access_within_bounds(
4958 struct bpf_verifier_env *env,
4959 int regno, int off, int access_size,
4960 enum bpf_access_src src, enum bpf_access_type type)
4962 struct bpf_reg_state *regs = cur_regs(env);
4963 struct bpf_reg_state *reg = regs + regno;
4964 struct bpf_func_state *state = func(env, reg);
4965 int min_off, max_off;
4969 if (src == ACCESS_HELPER)
4970 /* We don't know if helpers are reading or writing (or both). */
4971 err_extra = " indirect access to";
4972 else if (type == BPF_READ)
4973 err_extra = " read from";
4975 err_extra = " write to";
4977 if (tnum_is_const(reg->var_off)) {
4978 min_off = reg->var_off.value + off;
4979 if (access_size > 0)
4980 max_off = min_off + access_size - 1;
4984 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
4985 reg->smin_value <= -BPF_MAX_VAR_OFF) {
4986 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
4990 min_off = reg->smin_value + off;
4991 if (access_size > 0)
4992 max_off = reg->smax_value + off + access_size - 1;
4997 err = check_stack_slot_within_bounds(min_off, state, type);
4999 err = check_stack_slot_within_bounds(max_off, state, type);
5002 if (tnum_is_const(reg->var_off)) {
5003 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
5004 err_extra, regno, off, access_size);
5008 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5009 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
5010 err_extra, regno, tn_buf, access_size);
5016 /* check whether memory at (regno + off) is accessible for t = (read | write)
5017 * if t==write, value_regno is a register which value is stored into memory
5018 * if t==read, value_regno is a register which will receive the value from memory
5019 * if t==write && value_regno==-1, some unknown value is stored into memory
5020 * if t==read && value_regno==-1, don't care what we read from memory
5022 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
5023 int off, int bpf_size, enum bpf_access_type t,
5024 int value_regno, bool strict_alignment_once)
5026 struct bpf_reg_state *regs = cur_regs(env);
5027 struct bpf_reg_state *reg = regs + regno;
5028 struct bpf_func_state *state;
5031 size = bpf_size_to_bytes(bpf_size);
5035 /* alignment checks will add in reg->off themselves */
5036 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
5040 /* for access checks, reg->off is just part of off */
5043 if (reg->type == PTR_TO_MAP_KEY) {
5044 if (t == BPF_WRITE) {
5045 verbose(env, "write to change key R%d not allowed\n", regno);
5049 err = check_mem_region_access(env, regno, off, size,
5050 reg->map_ptr->key_size, false);
5053 if (value_regno >= 0)
5054 mark_reg_unknown(env, regs, value_regno);
5055 } else if (reg->type == PTR_TO_MAP_VALUE) {
5056 struct btf_field *kptr_field = NULL;
5058 if (t == BPF_WRITE && value_regno >= 0 &&
5059 is_pointer_value(env, value_regno)) {
5060 verbose(env, "R%d leaks addr into map\n", value_regno);
5063 err = check_map_access_type(env, regno, off, size, t);
5066 err = check_map_access(env, regno, off, size, false, ACCESS_DIRECT);
5069 if (tnum_is_const(reg->var_off))
5070 kptr_field = btf_record_find(reg->map_ptr->record,
5071 off + reg->var_off.value, BPF_KPTR);
5073 err = check_map_kptr_access(env, regno, value_regno, insn_idx, kptr_field);
5074 } else if (t == BPF_READ && value_regno >= 0) {
5075 struct bpf_map *map = reg->map_ptr;
5077 /* if map is read-only, track its contents as scalars */
5078 if (tnum_is_const(reg->var_off) &&
5079 bpf_map_is_rdonly(map) &&
5080 map->ops->map_direct_value_addr) {
5081 int map_off = off + reg->var_off.value;
5084 err = bpf_map_direct_read(map, map_off, size,
5089 regs[value_regno].type = SCALAR_VALUE;
5090 __mark_reg_known(®s[value_regno], val);
5092 mark_reg_unknown(env, regs, value_regno);
5095 } else if (base_type(reg->type) == PTR_TO_MEM) {
5096 bool rdonly_mem = type_is_rdonly_mem(reg->type);
5098 if (type_may_be_null(reg->type)) {
5099 verbose(env, "R%d invalid mem access '%s'\n", regno,
5100 reg_type_str(env, reg->type));
5104 if (t == BPF_WRITE && rdonly_mem) {
5105 verbose(env, "R%d cannot write into %s\n",
5106 regno, reg_type_str(env, reg->type));
5110 if (t == BPF_WRITE && value_regno >= 0 &&
5111 is_pointer_value(env, value_regno)) {
5112 verbose(env, "R%d leaks addr into mem\n", value_regno);
5116 err = check_mem_region_access(env, regno, off, size,
5117 reg->mem_size, false);
5118 if (!err && value_regno >= 0 && (t == BPF_READ || rdonly_mem))
5119 mark_reg_unknown(env, regs, value_regno);
5120 } else if (reg->type == PTR_TO_CTX) {
5121 enum bpf_reg_type reg_type = SCALAR_VALUE;
5122 struct btf *btf = NULL;
5125 if (t == BPF_WRITE && value_regno >= 0 &&
5126 is_pointer_value(env, value_regno)) {
5127 verbose(env, "R%d leaks addr into ctx\n", value_regno);
5131 err = check_ptr_off_reg(env, reg, regno);
5135 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf,
5138 verbose_linfo(env, insn_idx, "; ");
5139 if (!err && t == BPF_READ && value_regno >= 0) {
5140 /* ctx access returns either a scalar, or a
5141 * PTR_TO_PACKET[_META,_END]. In the latter
5142 * case, we know the offset is zero.
5144 if (reg_type == SCALAR_VALUE) {
5145 mark_reg_unknown(env, regs, value_regno);
5147 mark_reg_known_zero(env, regs,
5149 if (type_may_be_null(reg_type))
5150 regs[value_regno].id = ++env->id_gen;
5151 /* A load of ctx field could have different
5152 * actual load size with the one encoded in the
5153 * insn. When the dst is PTR, it is for sure not
5156 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
5157 if (base_type(reg_type) == PTR_TO_BTF_ID) {
5158 regs[value_regno].btf = btf;
5159 regs[value_regno].btf_id = btf_id;
5162 regs[value_regno].type = reg_type;
5165 } else if (reg->type == PTR_TO_STACK) {
5166 /* Basic bounds checks. */
5167 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
5171 state = func(env, reg);
5172 err = update_stack_depth(env, state, off);
5177 err = check_stack_read(env, regno, off, size,
5180 err = check_stack_write(env, regno, off, size,
5181 value_regno, insn_idx);
5182 } else if (reg_is_pkt_pointer(reg)) {
5183 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
5184 verbose(env, "cannot write into packet\n");
5187 if (t == BPF_WRITE && value_regno >= 0 &&
5188 is_pointer_value(env, value_regno)) {
5189 verbose(env, "R%d leaks addr into packet\n",
5193 err = check_packet_access(env, regno, off, size, false);
5194 if (!err && t == BPF_READ && value_regno >= 0)
5195 mark_reg_unknown(env, regs, value_regno);
5196 } else if (reg->type == PTR_TO_FLOW_KEYS) {
5197 if (t == BPF_WRITE && value_regno >= 0 &&
5198 is_pointer_value(env, value_regno)) {
5199 verbose(env, "R%d leaks addr into flow keys\n",
5204 err = check_flow_keys_access(env, off, size);
5205 if (!err && t == BPF_READ && value_regno >= 0)
5206 mark_reg_unknown(env, regs, value_regno);
5207 } else if (type_is_sk_pointer(reg->type)) {
5208 if (t == BPF_WRITE) {
5209 verbose(env, "R%d cannot write into %s\n",
5210 regno, reg_type_str(env, reg->type));
5213 err = check_sock_access(env, insn_idx, regno, off, size, t);
5214 if (!err && value_regno >= 0)
5215 mark_reg_unknown(env, regs, value_regno);
5216 } else if (reg->type == PTR_TO_TP_BUFFER) {
5217 err = check_tp_buffer_access(env, reg, regno, off, size);
5218 if (!err && t == BPF_READ && value_regno >= 0)
5219 mark_reg_unknown(env, regs, value_regno);
5220 } else if (base_type(reg->type) == PTR_TO_BTF_ID &&
5221 !type_may_be_null(reg->type)) {
5222 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
5224 } else if (reg->type == CONST_PTR_TO_MAP) {
5225 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
5227 } else if (base_type(reg->type) == PTR_TO_BUF) {
5228 bool rdonly_mem = type_is_rdonly_mem(reg->type);
5232 if (t == BPF_WRITE) {
5233 verbose(env, "R%d cannot write into %s\n",
5234 regno, reg_type_str(env, reg->type));
5237 max_access = &env->prog->aux->max_rdonly_access;
5239 max_access = &env->prog->aux->max_rdwr_access;
5242 err = check_buffer_access(env, reg, regno, off, size, false,
5245 if (!err && value_regno >= 0 && (rdonly_mem || t == BPF_READ))
5246 mark_reg_unknown(env, regs, value_regno);
5248 verbose(env, "R%d invalid mem access '%s'\n", regno,
5249 reg_type_str(env, reg->type));
5253 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
5254 regs[value_regno].type == SCALAR_VALUE) {
5255 /* b/h/w load zero-extends, mark upper bits as known 0 */
5256 coerce_reg_to_size(®s[value_regno], size);
5261 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
5266 switch (insn->imm) {
5268 case BPF_ADD | BPF_FETCH:
5270 case BPF_AND | BPF_FETCH:
5272 case BPF_OR | BPF_FETCH:
5274 case BPF_XOR | BPF_FETCH:
5279 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
5283 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
5284 verbose(env, "invalid atomic operand size\n");
5288 /* check src1 operand */
5289 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5293 /* check src2 operand */
5294 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5298 if (insn->imm == BPF_CMPXCHG) {
5299 /* Check comparison of R0 with memory location */
5300 const u32 aux_reg = BPF_REG_0;
5302 err = check_reg_arg(env, aux_reg, SRC_OP);
5306 if (is_pointer_value(env, aux_reg)) {
5307 verbose(env, "R%d leaks addr into mem\n", aux_reg);
5312 if (is_pointer_value(env, insn->src_reg)) {
5313 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
5317 if (is_ctx_reg(env, insn->dst_reg) ||
5318 is_pkt_reg(env, insn->dst_reg) ||
5319 is_flow_key_reg(env, insn->dst_reg) ||
5320 is_sk_reg(env, insn->dst_reg)) {
5321 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
5323 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
5327 if (insn->imm & BPF_FETCH) {
5328 if (insn->imm == BPF_CMPXCHG)
5329 load_reg = BPF_REG_0;
5331 load_reg = insn->src_reg;
5333 /* check and record load of old value */
5334 err = check_reg_arg(env, load_reg, DST_OP);
5338 /* This instruction accesses a memory location but doesn't
5339 * actually load it into a register.
5344 /* Check whether we can read the memory, with second call for fetch
5345 * case to simulate the register fill.
5347 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
5348 BPF_SIZE(insn->code), BPF_READ, -1, true);
5349 if (!err && load_reg >= 0)
5350 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
5351 BPF_SIZE(insn->code), BPF_READ, load_reg,
5356 /* Check whether we can write into the same memory. */
5357 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
5358 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
5365 /* When register 'regno' is used to read the stack (either directly or through
5366 * a helper function) make sure that it's within stack boundary and, depending
5367 * on the access type, that all elements of the stack are initialized.
5369 * 'off' includes 'regno->off', but not its dynamic part (if any).
5371 * All registers that have been spilled on the stack in the slots within the
5372 * read offsets are marked as read.
5374 static int check_stack_range_initialized(
5375 struct bpf_verifier_env *env, int regno, int off,
5376 int access_size, bool zero_size_allowed,
5377 enum bpf_access_src type, struct bpf_call_arg_meta *meta)
5379 struct bpf_reg_state *reg = reg_state(env, regno);
5380 struct bpf_func_state *state = func(env, reg);
5381 int err, min_off, max_off, i, j, slot, spi;
5382 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
5383 enum bpf_access_type bounds_check_type;
5384 /* Some accesses can write anything into the stack, others are
5387 bool clobber = false;
5389 if (access_size == 0 && !zero_size_allowed) {
5390 verbose(env, "invalid zero-sized read\n");
5394 if (type == ACCESS_HELPER) {
5395 /* The bounds checks for writes are more permissive than for
5396 * reads. However, if raw_mode is not set, we'll do extra
5399 bounds_check_type = BPF_WRITE;
5402 bounds_check_type = BPF_READ;
5404 err = check_stack_access_within_bounds(env, regno, off, access_size,
5405 type, bounds_check_type);
5410 if (tnum_is_const(reg->var_off)) {
5411 min_off = max_off = reg->var_off.value + off;
5413 /* Variable offset is prohibited for unprivileged mode for
5414 * simplicity since it requires corresponding support in
5415 * Spectre masking for stack ALU.
5416 * See also retrieve_ptr_limit().
5418 if (!env->bypass_spec_v1) {
5421 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5422 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
5423 regno, err_extra, tn_buf);
5426 /* Only initialized buffer on stack is allowed to be accessed
5427 * with variable offset. With uninitialized buffer it's hard to
5428 * guarantee that whole memory is marked as initialized on
5429 * helper return since specific bounds are unknown what may
5430 * cause uninitialized stack leaking.
5432 if (meta && meta->raw_mode)
5435 min_off = reg->smin_value + off;
5436 max_off = reg->smax_value + off;
5439 if (meta && meta->raw_mode) {
5440 meta->access_size = access_size;
5441 meta->regno = regno;
5445 for (i = min_off; i < max_off + access_size; i++) {
5449 spi = slot / BPF_REG_SIZE;
5450 if (state->allocated_stack <= slot)
5452 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
5453 if (*stype == STACK_MISC)
5455 if (*stype == STACK_ZERO) {
5457 /* helper can write anything into the stack */
5458 *stype = STACK_MISC;
5463 if (is_spilled_reg(&state->stack[spi]) &&
5464 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
5465 env->allow_ptr_leaks)) {
5467 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
5468 for (j = 0; j < BPF_REG_SIZE; j++)
5469 scrub_spilled_slot(&state->stack[spi].slot_type[j]);
5475 if (tnum_is_const(reg->var_off)) {
5476 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
5477 err_extra, regno, min_off, i - min_off, access_size);
5481 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5482 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
5483 err_extra, regno, tn_buf, i - min_off, access_size);
5487 /* reading any byte out of 8-byte 'spill_slot' will cause
5488 * the whole slot to be marked as 'read'
5490 mark_reg_read(env, &state->stack[spi].spilled_ptr,
5491 state->stack[spi].spilled_ptr.parent,
5493 /* We do not set REG_LIVE_WRITTEN for stack slot, as we can not
5494 * be sure that whether stack slot is written to or not. Hence,
5495 * we must still conservatively propagate reads upwards even if
5496 * helper may write to the entire memory range.
5499 return update_stack_depth(env, state, min_off);
5502 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
5503 int access_size, bool zero_size_allowed,
5504 struct bpf_call_arg_meta *meta)
5506 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5509 switch (base_type(reg->type)) {
5511 case PTR_TO_PACKET_META:
5512 return check_packet_access(env, regno, reg->off, access_size,
5514 case PTR_TO_MAP_KEY:
5515 if (meta && meta->raw_mode) {
5516 verbose(env, "R%d cannot write into %s\n", regno,
5517 reg_type_str(env, reg->type));
5520 return check_mem_region_access(env, regno, reg->off, access_size,
5521 reg->map_ptr->key_size, false);
5522 case PTR_TO_MAP_VALUE:
5523 if (check_map_access_type(env, regno, reg->off, access_size,
5524 meta && meta->raw_mode ? BPF_WRITE :
5527 return check_map_access(env, regno, reg->off, access_size,
5528 zero_size_allowed, ACCESS_HELPER);
5530 if (type_is_rdonly_mem(reg->type)) {
5531 if (meta && meta->raw_mode) {
5532 verbose(env, "R%d cannot write into %s\n", regno,
5533 reg_type_str(env, reg->type));
5537 return check_mem_region_access(env, regno, reg->off,
5538 access_size, reg->mem_size,
5541 if (type_is_rdonly_mem(reg->type)) {
5542 if (meta && meta->raw_mode) {
5543 verbose(env, "R%d cannot write into %s\n", regno,
5544 reg_type_str(env, reg->type));
5548 max_access = &env->prog->aux->max_rdonly_access;
5550 max_access = &env->prog->aux->max_rdwr_access;
5552 return check_buffer_access(env, reg, regno, reg->off,
5553 access_size, zero_size_allowed,
5556 return check_stack_range_initialized(
5558 regno, reg->off, access_size,
5559 zero_size_allowed, ACCESS_HELPER, meta);
5561 /* in case the function doesn't know how to access the context,
5562 * (because we are in a program of type SYSCALL for example), we
5563 * can not statically check its size.
5564 * Dynamically check it now.
5566 if (!env->ops->convert_ctx_access) {
5567 enum bpf_access_type atype = meta && meta->raw_mode ? BPF_WRITE : BPF_READ;
5568 int offset = access_size - 1;
5570 /* Allow zero-byte read from PTR_TO_CTX */
5571 if (access_size == 0)
5572 return zero_size_allowed ? 0 : -EACCES;
5574 return check_mem_access(env, env->insn_idx, regno, offset, BPF_B,
5579 default: /* scalar_value or invalid ptr */
5580 /* Allow zero-byte read from NULL, regardless of pointer type */
5581 if (zero_size_allowed && access_size == 0 &&
5582 register_is_null(reg))
5585 verbose(env, "R%d type=%s ", regno,
5586 reg_type_str(env, reg->type));
5587 verbose(env, "expected=%s\n", reg_type_str(env, PTR_TO_STACK));
5592 static int check_mem_size_reg(struct bpf_verifier_env *env,
5593 struct bpf_reg_state *reg, u32 regno,
5594 bool zero_size_allowed,
5595 struct bpf_call_arg_meta *meta)
5599 /* This is used to refine r0 return value bounds for helpers
5600 * that enforce this value as an upper bound on return values.
5601 * See do_refine_retval_range() for helpers that can refine
5602 * the return value. C type of helper is u32 so we pull register
5603 * bound from umax_value however, if negative verifier errors
5604 * out. Only upper bounds can be learned because retval is an
5605 * int type and negative retvals are allowed.
5607 meta->msize_max_value = reg->umax_value;
5609 /* The register is SCALAR_VALUE; the access check
5610 * happens using its boundaries.
5612 if (!tnum_is_const(reg->var_off))
5613 /* For unprivileged variable accesses, disable raw
5614 * mode so that the program is required to
5615 * initialize all the memory that the helper could
5616 * just partially fill up.
5620 if (reg->smin_value < 0) {
5621 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
5626 if (reg->umin_value == 0) {
5627 err = check_helper_mem_access(env, regno - 1, 0,
5634 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
5635 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
5639 err = check_helper_mem_access(env, regno - 1,
5641 zero_size_allowed, meta);
5643 err = mark_chain_precision(env, regno);
5647 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
5648 u32 regno, u32 mem_size)
5650 bool may_be_null = type_may_be_null(reg->type);
5651 struct bpf_reg_state saved_reg;
5652 struct bpf_call_arg_meta meta;
5655 if (register_is_null(reg))
5658 memset(&meta, 0, sizeof(meta));
5659 /* Assuming that the register contains a value check if the memory
5660 * access is safe. Temporarily save and restore the register's state as
5661 * the conversion shouldn't be visible to a caller.
5665 mark_ptr_not_null_reg(reg);
5668 err = check_helper_mem_access(env, regno, mem_size, true, &meta);
5669 /* Check access for BPF_WRITE */
5670 meta.raw_mode = true;
5671 err = err ?: check_helper_mem_access(env, regno, mem_size, true, &meta);
5679 static int check_kfunc_mem_size_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
5682 struct bpf_reg_state *mem_reg = &cur_regs(env)[regno - 1];
5683 bool may_be_null = type_may_be_null(mem_reg->type);
5684 struct bpf_reg_state saved_reg;
5685 struct bpf_call_arg_meta meta;
5688 WARN_ON_ONCE(regno < BPF_REG_2 || regno > BPF_REG_5);
5690 memset(&meta, 0, sizeof(meta));
5693 saved_reg = *mem_reg;
5694 mark_ptr_not_null_reg(mem_reg);
5697 err = check_mem_size_reg(env, reg, regno, true, &meta);
5698 /* Check access for BPF_WRITE */
5699 meta.raw_mode = true;
5700 err = err ?: check_mem_size_reg(env, reg, regno, true, &meta);
5703 *mem_reg = saved_reg;
5707 /* Implementation details:
5708 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL.
5709 * bpf_obj_new returns PTR_TO_BTF_ID | MEM_ALLOC | PTR_MAYBE_NULL.
5710 * Two bpf_map_lookups (even with the same key) will have different reg->id.
5711 * Two separate bpf_obj_new will also have different reg->id.
5712 * For traditional PTR_TO_MAP_VALUE or PTR_TO_BTF_ID | MEM_ALLOC, the verifier
5713 * clears reg->id after value_or_null->value transition, since the verifier only
5714 * cares about the range of access to valid map value pointer and doesn't care
5715 * about actual address of the map element.
5716 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
5717 * reg->id > 0 after value_or_null->value transition. By doing so
5718 * two bpf_map_lookups will be considered two different pointers that
5719 * point to different bpf_spin_locks. Likewise for pointers to allocated objects
5720 * returned from bpf_obj_new.
5721 * The verifier allows taking only one bpf_spin_lock at a time to avoid
5723 * Since only one bpf_spin_lock is allowed the checks are simpler than
5724 * reg_is_refcounted() logic. The verifier needs to remember only
5725 * one spin_lock instead of array of acquired_refs.
5726 * cur_state->active_lock remembers which map value element or allocated
5727 * object got locked and clears it after bpf_spin_unlock.
5729 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
5732 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5733 struct bpf_verifier_state *cur = env->cur_state;
5734 bool is_const = tnum_is_const(reg->var_off);
5735 u64 val = reg->var_off.value;
5736 struct bpf_map *map = NULL;
5737 struct btf *btf = NULL;
5738 struct btf_record *rec;
5742 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
5746 if (reg->type == PTR_TO_MAP_VALUE) {
5750 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
5758 rec = reg_btf_record(reg);
5759 if (!btf_record_has_field(rec, BPF_SPIN_LOCK)) {
5760 verbose(env, "%s '%s' has no valid bpf_spin_lock\n", map ? "map" : "local",
5761 map ? map->name : "kptr");
5764 if (rec->spin_lock_off != val + reg->off) {
5765 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock' that is at %d\n",
5766 val + reg->off, rec->spin_lock_off);
5770 if (cur->active_lock.ptr) {
5772 "Locking two bpf_spin_locks are not allowed\n");
5776 cur->active_lock.ptr = map;
5778 cur->active_lock.ptr = btf;
5779 cur->active_lock.id = reg->id;
5781 struct bpf_func_state *fstate = cur_func(env);
5790 if (!cur->active_lock.ptr) {
5791 verbose(env, "bpf_spin_unlock without taking a lock\n");
5794 if (cur->active_lock.ptr != ptr ||
5795 cur->active_lock.id != reg->id) {
5796 verbose(env, "bpf_spin_unlock of different lock\n");
5799 cur->active_lock.ptr = NULL;
5800 cur->active_lock.id = 0;
5802 for (i = fstate->acquired_refs - 1; i >= 0; i--) {
5805 /* Complain on error because this reference state cannot
5806 * be freed before this point, as bpf_spin_lock critical
5807 * section does not allow functions that release the
5808 * allocated object immediately.
5810 if (!fstate->refs[i].release_on_unlock)
5812 err = release_reference(env, fstate->refs[i].id);
5814 verbose(env, "failed to release release_on_unlock reference");
5822 static int process_timer_func(struct bpf_verifier_env *env, int regno,
5823 struct bpf_call_arg_meta *meta)
5825 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5826 bool is_const = tnum_is_const(reg->var_off);
5827 struct bpf_map *map = reg->map_ptr;
5828 u64 val = reg->var_off.value;
5832 "R%d doesn't have constant offset. bpf_timer has to be at the constant offset\n",
5837 verbose(env, "map '%s' has to have BTF in order to use bpf_timer\n",
5841 if (!btf_record_has_field(map->record, BPF_TIMER)) {
5842 verbose(env, "map '%s' has no valid bpf_timer\n", map->name);
5845 if (map->record->timer_off != val + reg->off) {
5846 verbose(env, "off %lld doesn't point to 'struct bpf_timer' that is at %d\n",
5847 val + reg->off, map->record->timer_off);
5850 if (meta->map_ptr) {
5851 verbose(env, "verifier bug. Two map pointers in a timer helper\n");
5854 meta->map_uid = reg->map_uid;
5855 meta->map_ptr = map;
5859 static int process_kptr_func(struct bpf_verifier_env *env, int regno,
5860 struct bpf_call_arg_meta *meta)
5862 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5863 struct bpf_map *map_ptr = reg->map_ptr;
5864 struct btf_field *kptr_field;
5867 if (!tnum_is_const(reg->var_off)) {
5869 "R%d doesn't have constant offset. kptr has to be at the constant offset\n",
5873 if (!map_ptr->btf) {
5874 verbose(env, "map '%s' has to have BTF in order to use bpf_kptr_xchg\n",
5878 if (!btf_record_has_field(map_ptr->record, BPF_KPTR)) {
5879 verbose(env, "map '%s' has no valid kptr\n", map_ptr->name);
5883 meta->map_ptr = map_ptr;
5884 kptr_off = reg->off + reg->var_off.value;
5885 kptr_field = btf_record_find(map_ptr->record, kptr_off, BPF_KPTR);
5887 verbose(env, "off=%d doesn't point to kptr\n", kptr_off);
5890 if (kptr_field->type != BPF_KPTR_REF) {
5891 verbose(env, "off=%d kptr isn't referenced kptr\n", kptr_off);
5894 meta->kptr_field = kptr_field;
5898 /* There are two register types representing a bpf_dynptr, one is PTR_TO_STACK
5899 * which points to a stack slot, and the other is CONST_PTR_TO_DYNPTR.
5901 * In both cases we deal with the first 8 bytes, but need to mark the next 8
5902 * bytes as STACK_DYNPTR in case of PTR_TO_STACK. In case of
5903 * CONST_PTR_TO_DYNPTR, we are guaranteed to get the beginning of the object.
5905 * Mutability of bpf_dynptr is at two levels, one is at the level of struct
5906 * bpf_dynptr itself, i.e. whether the helper is receiving a pointer to struct
5907 * bpf_dynptr or pointer to const struct bpf_dynptr. In the former case, it can
5908 * mutate the view of the dynptr and also possibly destroy it. In the latter
5909 * case, it cannot mutate the bpf_dynptr itself but it can still mutate the
5910 * memory that dynptr points to.
5912 * The verifier will keep track both levels of mutation (bpf_dynptr's in
5913 * reg->type and the memory's in reg->dynptr.type), but there is no support for
5914 * readonly dynptr view yet, hence only the first case is tracked and checked.
5916 * This is consistent with how C applies the const modifier to a struct object,
5917 * where the pointer itself inside bpf_dynptr becomes const but not what it
5920 * Helpers which do not mutate the bpf_dynptr set MEM_RDONLY in their argument
5921 * type, and declare it as 'const struct bpf_dynptr *' in their prototype.
5923 int process_dynptr_func(struct bpf_verifier_env *env, int regno,
5924 enum bpf_arg_type arg_type, struct bpf_call_arg_meta *meta)
5926 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5928 /* MEM_UNINIT and MEM_RDONLY are exclusive, when applied to an
5929 * ARG_PTR_TO_DYNPTR (or ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_*):
5931 if ((arg_type & (MEM_UNINIT | MEM_RDONLY)) == (MEM_UNINIT | MEM_RDONLY)) {
5932 verbose(env, "verifier internal error: misconfigured dynptr helper type flags\n");
5935 /* CONST_PTR_TO_DYNPTR already has fixed and var_off as 0 due to
5936 * check_func_arg_reg_off's logic. We only need to check offset
5937 * alignment for PTR_TO_STACK.
5939 if (reg->type == PTR_TO_STACK && (reg->off % BPF_REG_SIZE)) {
5940 verbose(env, "cannot pass in dynptr at an offset=%d\n", reg->off);
5943 /* MEM_UNINIT - Points to memory that is an appropriate candidate for
5944 * constructing a mutable bpf_dynptr object.
5946 * Currently, this is only possible with PTR_TO_STACK
5947 * pointing to a region of at least 16 bytes which doesn't
5948 * contain an existing bpf_dynptr.
5950 * MEM_RDONLY - Points to a initialized bpf_dynptr that will not be
5951 * mutated or destroyed. However, the memory it points to
5954 * None - Points to a initialized dynptr that can be mutated and
5955 * destroyed, including mutation of the memory it points
5958 if (arg_type & MEM_UNINIT) {
5959 if (!is_dynptr_reg_valid_uninit(env, reg)) {
5960 verbose(env, "Dynptr has to be an uninitialized dynptr\n");
5964 /* We only support one dynptr being uninitialized at the moment,
5965 * which is sufficient for the helper functions we have right now.
5967 if (meta->uninit_dynptr_regno) {
5968 verbose(env, "verifier internal error: multiple uninitialized dynptr args\n");
5972 meta->uninit_dynptr_regno = regno;
5973 } else /* MEM_RDONLY and None case from above */ {
5974 /* For the reg->type == PTR_TO_STACK case, bpf_dynptr is never const */
5975 if (reg->type == CONST_PTR_TO_DYNPTR && !(arg_type & MEM_RDONLY)) {
5976 verbose(env, "cannot pass pointer to const bpf_dynptr, the helper mutates it\n");
5980 if (!is_dynptr_reg_valid_init(env, reg)) {
5982 "Expected an initialized dynptr as arg #%d\n",
5987 /* Fold modifiers (in this case, MEM_RDONLY) when checking expected type */
5988 if (!is_dynptr_type_expected(env, reg, arg_type & ~MEM_RDONLY)) {
5989 const char *err_extra = "";
5991 switch (arg_type & DYNPTR_TYPE_FLAG_MASK) {
5992 case DYNPTR_TYPE_LOCAL:
5993 err_extra = "local";
5995 case DYNPTR_TYPE_RINGBUF:
5996 err_extra = "ringbuf";
5999 err_extra = "<unknown>";
6003 "Expected a dynptr of type %s as arg #%d\n",
6011 static bool arg_type_is_mem_size(enum bpf_arg_type type)
6013 return type == ARG_CONST_SIZE ||
6014 type == ARG_CONST_SIZE_OR_ZERO;
6017 static bool arg_type_is_release(enum bpf_arg_type type)
6019 return type & OBJ_RELEASE;
6022 static bool arg_type_is_dynptr(enum bpf_arg_type type)
6024 return base_type(type) == ARG_PTR_TO_DYNPTR;
6027 static int int_ptr_type_to_size(enum bpf_arg_type type)
6029 if (type == ARG_PTR_TO_INT)
6031 else if (type == ARG_PTR_TO_LONG)
6037 static int resolve_map_arg_type(struct bpf_verifier_env *env,
6038 const struct bpf_call_arg_meta *meta,
6039 enum bpf_arg_type *arg_type)
6041 if (!meta->map_ptr) {
6042 /* kernel subsystem misconfigured verifier */
6043 verbose(env, "invalid map_ptr to access map->type\n");
6047 switch (meta->map_ptr->map_type) {
6048 case BPF_MAP_TYPE_SOCKMAP:
6049 case BPF_MAP_TYPE_SOCKHASH:
6050 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
6051 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
6053 verbose(env, "invalid arg_type for sockmap/sockhash\n");
6057 case BPF_MAP_TYPE_BLOOM_FILTER:
6058 if (meta->func_id == BPF_FUNC_map_peek_elem)
6059 *arg_type = ARG_PTR_TO_MAP_VALUE;
6067 struct bpf_reg_types {
6068 const enum bpf_reg_type types[10];
6072 static const struct bpf_reg_types sock_types = {
6082 static const struct bpf_reg_types btf_id_sock_common_types = {
6089 PTR_TO_BTF_ID | PTR_TRUSTED,
6091 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
6095 static const struct bpf_reg_types mem_types = {
6103 PTR_TO_MEM | MEM_RINGBUF,
6108 static const struct bpf_reg_types int_ptr_types = {
6118 static const struct bpf_reg_types spin_lock_types = {
6121 PTR_TO_BTF_ID | MEM_ALLOC,
6125 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
6126 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
6127 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
6128 static const struct bpf_reg_types ringbuf_mem_types = { .types = { PTR_TO_MEM | MEM_RINGBUF } };
6129 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
6130 static const struct bpf_reg_types btf_ptr_types = {
6133 PTR_TO_BTF_ID | PTR_TRUSTED,
6134 PTR_TO_BTF_ID | MEM_RCU,
6137 static const struct bpf_reg_types percpu_btf_ptr_types = {
6139 PTR_TO_BTF_ID | MEM_PERCPU,
6140 PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED,
6143 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
6144 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
6145 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
6146 static const struct bpf_reg_types timer_types = { .types = { PTR_TO_MAP_VALUE } };
6147 static const struct bpf_reg_types kptr_types = { .types = { PTR_TO_MAP_VALUE } };
6148 static const struct bpf_reg_types dynptr_types = {
6151 CONST_PTR_TO_DYNPTR,
6155 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
6156 [ARG_PTR_TO_MAP_KEY] = &mem_types,
6157 [ARG_PTR_TO_MAP_VALUE] = &mem_types,
6158 [ARG_CONST_SIZE] = &scalar_types,
6159 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
6160 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
6161 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
6162 [ARG_PTR_TO_CTX] = &context_types,
6163 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
6165 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
6167 [ARG_PTR_TO_SOCKET] = &fullsock_types,
6168 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
6169 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
6170 [ARG_PTR_TO_MEM] = &mem_types,
6171 [ARG_PTR_TO_RINGBUF_MEM] = &ringbuf_mem_types,
6172 [ARG_PTR_TO_INT] = &int_ptr_types,
6173 [ARG_PTR_TO_LONG] = &int_ptr_types,
6174 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
6175 [ARG_PTR_TO_FUNC] = &func_ptr_types,
6176 [ARG_PTR_TO_STACK] = &stack_ptr_types,
6177 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
6178 [ARG_PTR_TO_TIMER] = &timer_types,
6179 [ARG_PTR_TO_KPTR] = &kptr_types,
6180 [ARG_PTR_TO_DYNPTR] = &dynptr_types,
6183 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
6184 enum bpf_arg_type arg_type,
6185 const u32 *arg_btf_id,
6186 struct bpf_call_arg_meta *meta)
6188 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
6189 enum bpf_reg_type expected, type = reg->type;
6190 const struct bpf_reg_types *compatible;
6193 compatible = compatible_reg_types[base_type(arg_type)];
6195 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
6199 /* ARG_PTR_TO_MEM + RDONLY is compatible with PTR_TO_MEM and PTR_TO_MEM + RDONLY,
6200 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM and NOT with PTR_TO_MEM + RDONLY
6202 * Same for MAYBE_NULL:
6204 * ARG_PTR_TO_MEM + MAYBE_NULL is compatible with PTR_TO_MEM and PTR_TO_MEM + MAYBE_NULL,
6205 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM but NOT with PTR_TO_MEM + MAYBE_NULL
6207 * Therefore we fold these flags depending on the arg_type before comparison.
6209 if (arg_type & MEM_RDONLY)
6210 type &= ~MEM_RDONLY;
6211 if (arg_type & PTR_MAYBE_NULL)
6212 type &= ~PTR_MAYBE_NULL;
6214 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
6215 expected = compatible->types[i];
6216 if (expected == NOT_INIT)
6219 if (type == expected)
6223 verbose(env, "R%d type=%s expected=", regno, reg_type_str(env, reg->type));
6224 for (j = 0; j + 1 < i; j++)
6225 verbose(env, "%s, ", reg_type_str(env, compatible->types[j]));
6226 verbose(env, "%s\n", reg_type_str(env, compatible->types[j]));
6230 if (reg->type == PTR_TO_BTF_ID || reg->type & PTR_TRUSTED) {
6231 /* For bpf_sk_release, it needs to match against first member
6232 * 'struct sock_common', hence make an exception for it. This
6233 * allows bpf_sk_release to work for multiple socket types.
6235 bool strict_type_match = arg_type_is_release(arg_type) &&
6236 meta->func_id != BPF_FUNC_sk_release;
6239 if (!compatible->btf_id) {
6240 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
6243 arg_btf_id = compatible->btf_id;
6246 if (meta->func_id == BPF_FUNC_kptr_xchg) {
6247 if (map_kptr_match_type(env, meta->kptr_field, reg, regno))
6250 if (arg_btf_id == BPF_PTR_POISON) {
6251 verbose(env, "verifier internal error:");
6252 verbose(env, "R%d has non-overwritten BPF_PTR_POISON type\n",
6257 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
6258 btf_vmlinux, *arg_btf_id,
6259 strict_type_match)) {
6260 verbose(env, "R%d is of type %s but %s is expected\n",
6261 regno, kernel_type_name(reg->btf, reg->btf_id),
6262 kernel_type_name(btf_vmlinux, *arg_btf_id));
6266 } else if (type_is_alloc(reg->type)) {
6267 if (meta->func_id != BPF_FUNC_spin_lock && meta->func_id != BPF_FUNC_spin_unlock) {
6268 verbose(env, "verifier internal error: unimplemented handling of MEM_ALLOC\n");
6276 int check_func_arg_reg_off(struct bpf_verifier_env *env,
6277 const struct bpf_reg_state *reg, int regno,
6278 enum bpf_arg_type arg_type)
6280 u32 type = reg->type;
6282 /* When referenced register is passed to release function, its fixed
6285 * We will check arg_type_is_release reg has ref_obj_id when storing
6286 * meta->release_regno.
6288 if (arg_type_is_release(arg_type)) {
6289 /* ARG_PTR_TO_DYNPTR with OBJ_RELEASE is a bit special, as it
6290 * may not directly point to the object being released, but to
6291 * dynptr pointing to such object, which might be at some offset
6292 * on the stack. In that case, we simply to fallback to the
6295 if (arg_type_is_dynptr(arg_type) && type == PTR_TO_STACK)
6297 /* Doing check_ptr_off_reg check for the offset will catch this
6298 * because fixed_off_ok is false, but checking here allows us
6299 * to give the user a better error message.
6302 verbose(env, "R%d must have zero offset when passed to release func or trusted arg to kfunc\n",
6306 return __check_ptr_off_reg(env, reg, regno, false);
6310 /* Pointer types where both fixed and variable offset is explicitly allowed: */
6313 case PTR_TO_PACKET_META:
6314 case PTR_TO_MAP_KEY:
6315 case PTR_TO_MAP_VALUE:
6317 case PTR_TO_MEM | MEM_RDONLY:
6318 case PTR_TO_MEM | MEM_RINGBUF:
6320 case PTR_TO_BUF | MEM_RDONLY:
6323 /* All the rest must be rejected, except PTR_TO_BTF_ID which allows
6327 case PTR_TO_BTF_ID | MEM_ALLOC:
6328 case PTR_TO_BTF_ID | PTR_TRUSTED:
6329 case PTR_TO_BTF_ID | MEM_RCU:
6330 case PTR_TO_BTF_ID | MEM_ALLOC | PTR_TRUSTED:
6331 /* When referenced PTR_TO_BTF_ID is passed to release function,
6332 * its fixed offset must be 0. In the other cases, fixed offset
6333 * can be non-zero. This was already checked above. So pass
6334 * fixed_off_ok as true to allow fixed offset for all other
6335 * cases. var_off always must be 0 for PTR_TO_BTF_ID, hence we
6336 * still need to do checks instead of returning.
6338 return __check_ptr_off_reg(env, reg, regno, true);
6340 return __check_ptr_off_reg(env, reg, regno, false);
6344 static u32 dynptr_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
6346 struct bpf_func_state *state = func(env, reg);
6349 if (reg->type == CONST_PTR_TO_DYNPTR)
6350 return reg->ref_obj_id;
6352 spi = get_spi(reg->off);
6353 return state->stack[spi].spilled_ptr.ref_obj_id;
6356 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
6357 struct bpf_call_arg_meta *meta,
6358 const struct bpf_func_proto *fn)
6360 u32 regno = BPF_REG_1 + arg;
6361 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
6362 enum bpf_arg_type arg_type = fn->arg_type[arg];
6363 enum bpf_reg_type type = reg->type;
6364 u32 *arg_btf_id = NULL;
6367 if (arg_type == ARG_DONTCARE)
6370 err = check_reg_arg(env, regno, SRC_OP);
6374 if (arg_type == ARG_ANYTHING) {
6375 if (is_pointer_value(env, regno)) {
6376 verbose(env, "R%d leaks addr into helper function\n",
6383 if (type_is_pkt_pointer(type) &&
6384 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
6385 verbose(env, "helper access to the packet is not allowed\n");
6389 if (base_type(arg_type) == ARG_PTR_TO_MAP_VALUE) {
6390 err = resolve_map_arg_type(env, meta, &arg_type);
6395 if (register_is_null(reg) && type_may_be_null(arg_type))
6396 /* A NULL register has a SCALAR_VALUE type, so skip
6399 goto skip_type_check;
6401 /* arg_btf_id and arg_size are in a union. */
6402 if (base_type(arg_type) == ARG_PTR_TO_BTF_ID ||
6403 base_type(arg_type) == ARG_PTR_TO_SPIN_LOCK)
6404 arg_btf_id = fn->arg_btf_id[arg];
6406 err = check_reg_type(env, regno, arg_type, arg_btf_id, meta);
6410 err = check_func_arg_reg_off(env, reg, regno, arg_type);
6415 if (arg_type_is_release(arg_type)) {
6416 if (arg_type_is_dynptr(arg_type)) {
6417 struct bpf_func_state *state = func(env, reg);
6420 /* Only dynptr created on stack can be released, thus
6421 * the get_spi and stack state checks for spilled_ptr
6422 * should only be done before process_dynptr_func for
6425 if (reg->type == PTR_TO_STACK) {
6426 spi = get_spi(reg->off);
6427 if (!is_spi_bounds_valid(state, spi, BPF_DYNPTR_NR_SLOTS) ||
6428 !state->stack[spi].spilled_ptr.ref_obj_id) {
6429 verbose(env, "arg %d is an unacquired reference\n", regno);
6433 verbose(env, "cannot release unowned const bpf_dynptr\n");
6436 } else if (!reg->ref_obj_id && !register_is_null(reg)) {
6437 verbose(env, "R%d must be referenced when passed to release function\n",
6441 if (meta->release_regno) {
6442 verbose(env, "verifier internal error: more than one release argument\n");
6445 meta->release_regno = regno;
6448 if (reg->ref_obj_id) {
6449 if (meta->ref_obj_id) {
6450 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
6451 regno, reg->ref_obj_id,
6455 meta->ref_obj_id = reg->ref_obj_id;
6458 switch (base_type(arg_type)) {
6459 case ARG_CONST_MAP_PTR:
6460 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
6461 if (meta->map_ptr) {
6462 /* Use map_uid (which is unique id of inner map) to reject:
6463 * inner_map1 = bpf_map_lookup_elem(outer_map, key1)
6464 * inner_map2 = bpf_map_lookup_elem(outer_map, key2)
6465 * if (inner_map1 && inner_map2) {
6466 * timer = bpf_map_lookup_elem(inner_map1);
6468 * // mismatch would have been allowed
6469 * bpf_timer_init(timer, inner_map2);
6472 * Comparing map_ptr is enough to distinguish normal and outer maps.
6474 if (meta->map_ptr != reg->map_ptr ||
6475 meta->map_uid != reg->map_uid) {
6477 "timer pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
6478 meta->map_uid, reg->map_uid);
6482 meta->map_ptr = reg->map_ptr;
6483 meta->map_uid = reg->map_uid;
6485 case ARG_PTR_TO_MAP_KEY:
6486 /* bpf_map_xxx(..., map_ptr, ..., key) call:
6487 * check that [key, key + map->key_size) are within
6488 * stack limits and initialized
6490 if (!meta->map_ptr) {
6491 /* in function declaration map_ptr must come before
6492 * map_key, so that it's verified and known before
6493 * we have to check map_key here. Otherwise it means
6494 * that kernel subsystem misconfigured verifier
6496 verbose(env, "invalid map_ptr to access map->key\n");
6499 err = check_helper_mem_access(env, regno,
6500 meta->map_ptr->key_size, false,
6503 case ARG_PTR_TO_MAP_VALUE:
6504 if (type_may_be_null(arg_type) && register_is_null(reg))
6507 /* bpf_map_xxx(..., map_ptr, ..., value) call:
6508 * check [value, value + map->value_size) validity
6510 if (!meta->map_ptr) {
6511 /* kernel subsystem misconfigured verifier */
6512 verbose(env, "invalid map_ptr to access map->value\n");
6515 meta->raw_mode = arg_type & MEM_UNINIT;
6516 err = check_helper_mem_access(env, regno,
6517 meta->map_ptr->value_size, false,
6520 case ARG_PTR_TO_PERCPU_BTF_ID:
6522 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
6525 meta->ret_btf = reg->btf;
6526 meta->ret_btf_id = reg->btf_id;
6528 case ARG_PTR_TO_SPIN_LOCK:
6529 if (meta->func_id == BPF_FUNC_spin_lock) {
6530 err = process_spin_lock(env, regno, true);
6533 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
6534 err = process_spin_lock(env, regno, false);
6538 verbose(env, "verifier internal error\n");
6542 case ARG_PTR_TO_TIMER:
6543 err = process_timer_func(env, regno, meta);
6547 case ARG_PTR_TO_FUNC:
6548 meta->subprogno = reg->subprogno;
6550 case ARG_PTR_TO_MEM:
6551 /* The access to this pointer is only checked when we hit the
6552 * next is_mem_size argument below.
6554 meta->raw_mode = arg_type & MEM_UNINIT;
6555 if (arg_type & MEM_FIXED_SIZE) {
6556 err = check_helper_mem_access(env, regno,
6557 fn->arg_size[arg], false,
6561 case ARG_CONST_SIZE:
6562 err = check_mem_size_reg(env, reg, regno, false, meta);
6564 case ARG_CONST_SIZE_OR_ZERO:
6565 err = check_mem_size_reg(env, reg, regno, true, meta);
6567 case ARG_PTR_TO_DYNPTR:
6568 err = process_dynptr_func(env, regno, arg_type, meta);
6572 case ARG_CONST_ALLOC_SIZE_OR_ZERO:
6573 if (!tnum_is_const(reg->var_off)) {
6574 verbose(env, "R%d is not a known constant'\n",
6578 meta->mem_size = reg->var_off.value;
6579 err = mark_chain_precision(env, regno);
6583 case ARG_PTR_TO_INT:
6584 case ARG_PTR_TO_LONG:
6586 int size = int_ptr_type_to_size(arg_type);
6588 err = check_helper_mem_access(env, regno, size, false, meta);
6591 err = check_ptr_alignment(env, reg, 0, size, true);
6594 case ARG_PTR_TO_CONST_STR:
6596 struct bpf_map *map = reg->map_ptr;
6601 if (!bpf_map_is_rdonly(map)) {
6602 verbose(env, "R%d does not point to a readonly map'\n", regno);
6606 if (!tnum_is_const(reg->var_off)) {
6607 verbose(env, "R%d is not a constant address'\n", regno);
6611 if (!map->ops->map_direct_value_addr) {
6612 verbose(env, "no direct value access support for this map type\n");
6616 err = check_map_access(env, regno, reg->off,
6617 map->value_size - reg->off, false,
6622 map_off = reg->off + reg->var_off.value;
6623 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
6625 verbose(env, "direct value access on string failed\n");
6629 str_ptr = (char *)(long)(map_addr);
6630 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
6631 verbose(env, "string is not zero-terminated\n");
6636 case ARG_PTR_TO_KPTR:
6637 err = process_kptr_func(env, regno, meta);
6646 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
6648 enum bpf_attach_type eatype = env->prog->expected_attach_type;
6649 enum bpf_prog_type type = resolve_prog_type(env->prog);
6651 if (func_id != BPF_FUNC_map_update_elem)
6654 /* It's not possible to get access to a locked struct sock in these
6655 * contexts, so updating is safe.
6658 case BPF_PROG_TYPE_TRACING:
6659 if (eatype == BPF_TRACE_ITER)
6662 case BPF_PROG_TYPE_SOCKET_FILTER:
6663 case BPF_PROG_TYPE_SCHED_CLS:
6664 case BPF_PROG_TYPE_SCHED_ACT:
6665 case BPF_PROG_TYPE_XDP:
6666 case BPF_PROG_TYPE_SK_REUSEPORT:
6667 case BPF_PROG_TYPE_FLOW_DISSECTOR:
6668 case BPF_PROG_TYPE_SK_LOOKUP:
6674 verbose(env, "cannot update sockmap in this context\n");
6678 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
6680 return env->prog->jit_requested &&
6681 bpf_jit_supports_subprog_tailcalls();
6684 static int check_map_func_compatibility(struct bpf_verifier_env *env,
6685 struct bpf_map *map, int func_id)
6690 /* We need a two way check, first is from map perspective ... */
6691 switch (map->map_type) {
6692 case BPF_MAP_TYPE_PROG_ARRAY:
6693 if (func_id != BPF_FUNC_tail_call)
6696 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
6697 if (func_id != BPF_FUNC_perf_event_read &&
6698 func_id != BPF_FUNC_perf_event_output &&
6699 func_id != BPF_FUNC_skb_output &&
6700 func_id != BPF_FUNC_perf_event_read_value &&
6701 func_id != BPF_FUNC_xdp_output)
6704 case BPF_MAP_TYPE_RINGBUF:
6705 if (func_id != BPF_FUNC_ringbuf_output &&
6706 func_id != BPF_FUNC_ringbuf_reserve &&
6707 func_id != BPF_FUNC_ringbuf_query &&
6708 func_id != BPF_FUNC_ringbuf_reserve_dynptr &&
6709 func_id != BPF_FUNC_ringbuf_submit_dynptr &&
6710 func_id != BPF_FUNC_ringbuf_discard_dynptr)
6713 case BPF_MAP_TYPE_USER_RINGBUF:
6714 if (func_id != BPF_FUNC_user_ringbuf_drain)
6717 case BPF_MAP_TYPE_STACK_TRACE:
6718 if (func_id != BPF_FUNC_get_stackid)
6721 case BPF_MAP_TYPE_CGROUP_ARRAY:
6722 if (func_id != BPF_FUNC_skb_under_cgroup &&
6723 func_id != BPF_FUNC_current_task_under_cgroup)
6726 case BPF_MAP_TYPE_CGROUP_STORAGE:
6727 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
6728 if (func_id != BPF_FUNC_get_local_storage)
6731 case BPF_MAP_TYPE_DEVMAP:
6732 case BPF_MAP_TYPE_DEVMAP_HASH:
6733 if (func_id != BPF_FUNC_redirect_map &&
6734 func_id != BPF_FUNC_map_lookup_elem)
6737 /* Restrict bpf side of cpumap and xskmap, open when use-cases
6740 case BPF_MAP_TYPE_CPUMAP:
6741 if (func_id != BPF_FUNC_redirect_map)
6744 case BPF_MAP_TYPE_XSKMAP:
6745 if (func_id != BPF_FUNC_redirect_map &&
6746 func_id != BPF_FUNC_map_lookup_elem)
6749 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
6750 case BPF_MAP_TYPE_HASH_OF_MAPS:
6751 if (func_id != BPF_FUNC_map_lookup_elem)
6754 case BPF_MAP_TYPE_SOCKMAP:
6755 if (func_id != BPF_FUNC_sk_redirect_map &&
6756 func_id != BPF_FUNC_sock_map_update &&
6757 func_id != BPF_FUNC_map_delete_elem &&
6758 func_id != BPF_FUNC_msg_redirect_map &&
6759 func_id != BPF_FUNC_sk_select_reuseport &&
6760 func_id != BPF_FUNC_map_lookup_elem &&
6761 !may_update_sockmap(env, func_id))
6764 case BPF_MAP_TYPE_SOCKHASH:
6765 if (func_id != BPF_FUNC_sk_redirect_hash &&
6766 func_id != BPF_FUNC_sock_hash_update &&
6767 func_id != BPF_FUNC_map_delete_elem &&
6768 func_id != BPF_FUNC_msg_redirect_hash &&
6769 func_id != BPF_FUNC_sk_select_reuseport &&
6770 func_id != BPF_FUNC_map_lookup_elem &&
6771 !may_update_sockmap(env, func_id))
6774 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
6775 if (func_id != BPF_FUNC_sk_select_reuseport)
6778 case BPF_MAP_TYPE_QUEUE:
6779 case BPF_MAP_TYPE_STACK:
6780 if (func_id != BPF_FUNC_map_peek_elem &&
6781 func_id != BPF_FUNC_map_pop_elem &&
6782 func_id != BPF_FUNC_map_push_elem)
6785 case BPF_MAP_TYPE_SK_STORAGE:
6786 if (func_id != BPF_FUNC_sk_storage_get &&
6787 func_id != BPF_FUNC_sk_storage_delete)
6790 case BPF_MAP_TYPE_INODE_STORAGE:
6791 if (func_id != BPF_FUNC_inode_storage_get &&
6792 func_id != BPF_FUNC_inode_storage_delete)
6795 case BPF_MAP_TYPE_TASK_STORAGE:
6796 if (func_id != BPF_FUNC_task_storage_get &&
6797 func_id != BPF_FUNC_task_storage_delete)
6800 case BPF_MAP_TYPE_CGRP_STORAGE:
6801 if (func_id != BPF_FUNC_cgrp_storage_get &&
6802 func_id != BPF_FUNC_cgrp_storage_delete)
6805 case BPF_MAP_TYPE_BLOOM_FILTER:
6806 if (func_id != BPF_FUNC_map_peek_elem &&
6807 func_id != BPF_FUNC_map_push_elem)
6814 /* ... and second from the function itself. */
6816 case BPF_FUNC_tail_call:
6817 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
6819 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
6820 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
6824 case BPF_FUNC_perf_event_read:
6825 case BPF_FUNC_perf_event_output:
6826 case BPF_FUNC_perf_event_read_value:
6827 case BPF_FUNC_skb_output:
6828 case BPF_FUNC_xdp_output:
6829 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
6832 case BPF_FUNC_ringbuf_output:
6833 case BPF_FUNC_ringbuf_reserve:
6834 case BPF_FUNC_ringbuf_query:
6835 case BPF_FUNC_ringbuf_reserve_dynptr:
6836 case BPF_FUNC_ringbuf_submit_dynptr:
6837 case BPF_FUNC_ringbuf_discard_dynptr:
6838 if (map->map_type != BPF_MAP_TYPE_RINGBUF)
6841 case BPF_FUNC_user_ringbuf_drain:
6842 if (map->map_type != BPF_MAP_TYPE_USER_RINGBUF)
6845 case BPF_FUNC_get_stackid:
6846 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
6849 case BPF_FUNC_current_task_under_cgroup:
6850 case BPF_FUNC_skb_under_cgroup:
6851 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
6854 case BPF_FUNC_redirect_map:
6855 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
6856 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
6857 map->map_type != BPF_MAP_TYPE_CPUMAP &&
6858 map->map_type != BPF_MAP_TYPE_XSKMAP)
6861 case BPF_FUNC_sk_redirect_map:
6862 case BPF_FUNC_msg_redirect_map:
6863 case BPF_FUNC_sock_map_update:
6864 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
6867 case BPF_FUNC_sk_redirect_hash:
6868 case BPF_FUNC_msg_redirect_hash:
6869 case BPF_FUNC_sock_hash_update:
6870 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
6873 case BPF_FUNC_get_local_storage:
6874 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
6875 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
6878 case BPF_FUNC_sk_select_reuseport:
6879 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
6880 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
6881 map->map_type != BPF_MAP_TYPE_SOCKHASH)
6884 case BPF_FUNC_map_pop_elem:
6885 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
6886 map->map_type != BPF_MAP_TYPE_STACK)
6889 case BPF_FUNC_map_peek_elem:
6890 case BPF_FUNC_map_push_elem:
6891 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
6892 map->map_type != BPF_MAP_TYPE_STACK &&
6893 map->map_type != BPF_MAP_TYPE_BLOOM_FILTER)
6896 case BPF_FUNC_map_lookup_percpu_elem:
6897 if (map->map_type != BPF_MAP_TYPE_PERCPU_ARRAY &&
6898 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
6899 map->map_type != BPF_MAP_TYPE_LRU_PERCPU_HASH)
6902 case BPF_FUNC_sk_storage_get:
6903 case BPF_FUNC_sk_storage_delete:
6904 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
6907 case BPF_FUNC_inode_storage_get:
6908 case BPF_FUNC_inode_storage_delete:
6909 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
6912 case BPF_FUNC_task_storage_get:
6913 case BPF_FUNC_task_storage_delete:
6914 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
6917 case BPF_FUNC_cgrp_storage_get:
6918 case BPF_FUNC_cgrp_storage_delete:
6919 if (map->map_type != BPF_MAP_TYPE_CGRP_STORAGE)
6928 verbose(env, "cannot pass map_type %d into func %s#%d\n",
6929 map->map_type, func_id_name(func_id), func_id);
6933 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
6937 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
6939 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
6941 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
6943 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
6945 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
6948 /* We only support one arg being in raw mode at the moment,
6949 * which is sufficient for the helper functions we have
6955 static bool check_args_pair_invalid(const struct bpf_func_proto *fn, int arg)
6957 bool is_fixed = fn->arg_type[arg] & MEM_FIXED_SIZE;
6958 bool has_size = fn->arg_size[arg] != 0;
6959 bool is_next_size = false;
6961 if (arg + 1 < ARRAY_SIZE(fn->arg_type))
6962 is_next_size = arg_type_is_mem_size(fn->arg_type[arg + 1]);
6964 if (base_type(fn->arg_type[arg]) != ARG_PTR_TO_MEM)
6965 return is_next_size;
6967 return has_size == is_next_size || is_next_size == is_fixed;
6970 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
6972 /* bpf_xxx(..., buf, len) call will access 'len'
6973 * bytes from memory 'buf'. Both arg types need
6974 * to be paired, so make sure there's no buggy
6975 * helper function specification.
6977 if (arg_type_is_mem_size(fn->arg1_type) ||
6978 check_args_pair_invalid(fn, 0) ||
6979 check_args_pair_invalid(fn, 1) ||
6980 check_args_pair_invalid(fn, 2) ||
6981 check_args_pair_invalid(fn, 3) ||
6982 check_args_pair_invalid(fn, 4))
6988 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
6992 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
6993 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_BTF_ID)
6994 return !!fn->arg_btf_id[i];
6995 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_SPIN_LOCK)
6996 return fn->arg_btf_id[i] == BPF_PTR_POISON;
6997 if (base_type(fn->arg_type[i]) != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i] &&
6998 /* arg_btf_id and arg_size are in a union. */
6999 (base_type(fn->arg_type[i]) != ARG_PTR_TO_MEM ||
7000 !(fn->arg_type[i] & MEM_FIXED_SIZE)))
7007 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
7009 return check_raw_mode_ok(fn) &&
7010 check_arg_pair_ok(fn) &&
7011 check_btf_id_ok(fn) ? 0 : -EINVAL;
7014 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
7015 * are now invalid, so turn them into unknown SCALAR_VALUE.
7017 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
7019 struct bpf_func_state *state;
7020 struct bpf_reg_state *reg;
7022 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
7023 if (reg_is_pkt_pointer_any(reg))
7024 __mark_reg_unknown(env, reg);
7030 BEYOND_PKT_END = -2,
7033 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
7035 struct bpf_func_state *state = vstate->frame[vstate->curframe];
7036 struct bpf_reg_state *reg = &state->regs[regn];
7038 if (reg->type != PTR_TO_PACKET)
7039 /* PTR_TO_PACKET_META is not supported yet */
7042 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
7043 * How far beyond pkt_end it goes is unknown.
7044 * if (!range_open) it's the case of pkt >= pkt_end
7045 * if (range_open) it's the case of pkt > pkt_end
7046 * hence this pointer is at least 1 byte bigger than pkt_end
7049 reg->range = BEYOND_PKT_END;
7051 reg->range = AT_PKT_END;
7054 /* The pointer with the specified id has released its reference to kernel
7055 * resources. Identify all copies of the same pointer and clear the reference.
7057 static int release_reference(struct bpf_verifier_env *env,
7060 struct bpf_func_state *state;
7061 struct bpf_reg_state *reg;
7064 err = release_reference_state(cur_func(env), ref_obj_id);
7068 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
7069 if (reg->ref_obj_id == ref_obj_id) {
7070 if (!env->allow_ptr_leaks)
7071 __mark_reg_not_init(env, reg);
7073 __mark_reg_unknown(env, reg);
7080 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
7081 struct bpf_reg_state *regs)
7085 /* after the call registers r0 - r5 were scratched */
7086 for (i = 0; i < CALLER_SAVED_REGS; i++) {
7087 mark_reg_not_init(env, regs, caller_saved[i]);
7088 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
7092 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
7093 struct bpf_func_state *caller,
7094 struct bpf_func_state *callee,
7097 static int set_callee_state(struct bpf_verifier_env *env,
7098 struct bpf_func_state *caller,
7099 struct bpf_func_state *callee, int insn_idx);
7101 static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
7102 int *insn_idx, int subprog,
7103 set_callee_state_fn set_callee_state_cb)
7105 struct bpf_verifier_state *state = env->cur_state;
7106 struct bpf_func_info_aux *func_info_aux;
7107 struct bpf_func_state *caller, *callee;
7109 bool is_global = false;
7111 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
7112 verbose(env, "the call stack of %d frames is too deep\n",
7113 state->curframe + 2);
7117 caller = state->frame[state->curframe];
7118 if (state->frame[state->curframe + 1]) {
7119 verbose(env, "verifier bug. Frame %d already allocated\n",
7120 state->curframe + 1);
7124 func_info_aux = env->prog->aux->func_info_aux;
7126 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
7127 err = btf_check_subprog_call(env, subprog, caller->regs);
7132 verbose(env, "Caller passes invalid args into func#%d\n",
7136 if (env->log.level & BPF_LOG_LEVEL)
7138 "Func#%d is global and valid. Skipping.\n",
7140 clear_caller_saved_regs(env, caller->regs);
7142 /* All global functions return a 64-bit SCALAR_VALUE */
7143 mark_reg_unknown(env, caller->regs, BPF_REG_0);
7144 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
7146 /* continue with next insn after call */
7151 /* set_callee_state is used for direct subprog calls, but we are
7152 * interested in validating only BPF helpers that can call subprogs as
7155 if (set_callee_state_cb != set_callee_state && !is_callback_calling_function(insn->imm)) {
7156 verbose(env, "verifier bug: helper %s#%d is not marked as callback-calling\n",
7157 func_id_name(insn->imm), insn->imm);
7161 if (insn->code == (BPF_JMP | BPF_CALL) &&
7162 insn->src_reg == 0 &&
7163 insn->imm == BPF_FUNC_timer_set_callback) {
7164 struct bpf_verifier_state *async_cb;
7166 /* there is no real recursion here. timer callbacks are async */
7167 env->subprog_info[subprog].is_async_cb = true;
7168 async_cb = push_async_cb(env, env->subprog_info[subprog].start,
7169 *insn_idx, subprog);
7172 callee = async_cb->frame[0];
7173 callee->async_entry_cnt = caller->async_entry_cnt + 1;
7175 /* Convert bpf_timer_set_callback() args into timer callback args */
7176 err = set_callee_state_cb(env, caller, callee, *insn_idx);
7180 clear_caller_saved_regs(env, caller->regs);
7181 mark_reg_unknown(env, caller->regs, BPF_REG_0);
7182 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
7183 /* continue with next insn after call */
7187 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
7190 state->frame[state->curframe + 1] = callee;
7192 /* callee cannot access r0, r6 - r9 for reading and has to write
7193 * into its own stack before reading from it.
7194 * callee can read/write into caller's stack
7196 init_func_state(env, callee,
7197 /* remember the callsite, it will be used by bpf_exit */
7198 *insn_idx /* callsite */,
7199 state->curframe + 1 /* frameno within this callchain */,
7200 subprog /* subprog number within this prog */);
7202 /* Transfer references to the callee */
7203 err = copy_reference_state(callee, caller);
7207 err = set_callee_state_cb(env, caller, callee, *insn_idx);
7211 clear_caller_saved_regs(env, caller->regs);
7213 /* only increment it after check_reg_arg() finished */
7216 /* and go analyze first insn of the callee */
7217 *insn_idx = env->subprog_info[subprog].start - 1;
7219 if (env->log.level & BPF_LOG_LEVEL) {
7220 verbose(env, "caller:\n");
7221 print_verifier_state(env, caller, true);
7222 verbose(env, "callee:\n");
7223 print_verifier_state(env, callee, true);
7228 free_func_state(callee);
7229 state->frame[state->curframe + 1] = NULL;
7233 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
7234 struct bpf_func_state *caller,
7235 struct bpf_func_state *callee)
7237 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
7238 * void *callback_ctx, u64 flags);
7239 * callback_fn(struct bpf_map *map, void *key, void *value,
7240 * void *callback_ctx);
7242 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
7244 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
7245 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
7246 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
7248 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
7249 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
7250 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
7252 /* pointer to stack or null */
7253 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
7256 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
7260 static int set_callee_state(struct bpf_verifier_env *env,
7261 struct bpf_func_state *caller,
7262 struct bpf_func_state *callee, int insn_idx)
7266 /* copy r1 - r5 args that callee can access. The copy includes parent
7267 * pointers, which connects us up to the liveness chain
7269 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
7270 callee->regs[i] = caller->regs[i];
7274 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
7277 int subprog, target_insn;
7279 target_insn = *insn_idx + insn->imm + 1;
7280 subprog = find_subprog(env, target_insn);
7282 verbose(env, "verifier bug. No program starts at insn %d\n",
7287 return __check_func_call(env, insn, insn_idx, subprog, set_callee_state);
7290 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
7291 struct bpf_func_state *caller,
7292 struct bpf_func_state *callee,
7295 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
7296 struct bpf_map *map;
7299 if (bpf_map_ptr_poisoned(insn_aux)) {
7300 verbose(env, "tail_call abusing map_ptr\n");
7304 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
7305 if (!map->ops->map_set_for_each_callback_args ||
7306 !map->ops->map_for_each_callback) {
7307 verbose(env, "callback function not allowed for map\n");
7311 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
7315 callee->in_callback_fn = true;
7316 callee->callback_ret_range = tnum_range(0, 1);
7320 static int set_loop_callback_state(struct bpf_verifier_env *env,
7321 struct bpf_func_state *caller,
7322 struct bpf_func_state *callee,
7325 /* bpf_loop(u32 nr_loops, void *callback_fn, void *callback_ctx,
7327 * callback_fn(u32 index, void *callback_ctx);
7329 callee->regs[BPF_REG_1].type = SCALAR_VALUE;
7330 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
7333 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
7334 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
7335 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
7337 callee->in_callback_fn = true;
7338 callee->callback_ret_range = tnum_range(0, 1);
7342 static int set_timer_callback_state(struct bpf_verifier_env *env,
7343 struct bpf_func_state *caller,
7344 struct bpf_func_state *callee,
7347 struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
7349 /* bpf_timer_set_callback(struct bpf_timer *timer, void *callback_fn);
7350 * callback_fn(struct bpf_map *map, void *key, void *value);
7352 callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP;
7353 __mark_reg_known_zero(&callee->regs[BPF_REG_1]);
7354 callee->regs[BPF_REG_1].map_ptr = map_ptr;
7356 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
7357 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
7358 callee->regs[BPF_REG_2].map_ptr = map_ptr;
7360 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
7361 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
7362 callee->regs[BPF_REG_3].map_ptr = map_ptr;
7365 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
7366 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
7367 callee->in_async_callback_fn = true;
7368 callee->callback_ret_range = tnum_range(0, 1);
7372 static int set_find_vma_callback_state(struct bpf_verifier_env *env,
7373 struct bpf_func_state *caller,
7374 struct bpf_func_state *callee,
7377 /* bpf_find_vma(struct task_struct *task, u64 addr,
7378 * void *callback_fn, void *callback_ctx, u64 flags)
7379 * (callback_fn)(struct task_struct *task,
7380 * struct vm_area_struct *vma, void *callback_ctx);
7382 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
7384 callee->regs[BPF_REG_2].type = PTR_TO_BTF_ID;
7385 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
7386 callee->regs[BPF_REG_2].btf = btf_vmlinux;
7387 callee->regs[BPF_REG_2].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_VMA],
7389 /* pointer to stack or null */
7390 callee->regs[BPF_REG_3] = caller->regs[BPF_REG_4];
7393 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
7394 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
7395 callee->in_callback_fn = true;
7396 callee->callback_ret_range = tnum_range(0, 1);
7400 static int set_user_ringbuf_callback_state(struct bpf_verifier_env *env,
7401 struct bpf_func_state *caller,
7402 struct bpf_func_state *callee,
7405 /* bpf_user_ringbuf_drain(struct bpf_map *map, void *callback_fn, void
7406 * callback_ctx, u64 flags);
7407 * callback_fn(const struct bpf_dynptr_t* dynptr, void *callback_ctx);
7409 __mark_reg_not_init(env, &callee->regs[BPF_REG_0]);
7410 mark_dynptr_cb_reg(&callee->regs[BPF_REG_1], BPF_DYNPTR_TYPE_LOCAL);
7411 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
7414 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
7415 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
7416 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
7418 callee->in_callback_fn = true;
7419 callee->callback_ret_range = tnum_range(0, 1);
7423 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
7425 struct bpf_verifier_state *state = env->cur_state;
7426 struct bpf_func_state *caller, *callee;
7427 struct bpf_reg_state *r0;
7430 callee = state->frame[state->curframe];
7431 r0 = &callee->regs[BPF_REG_0];
7432 if (r0->type == PTR_TO_STACK) {
7433 /* technically it's ok to return caller's stack pointer
7434 * (or caller's caller's pointer) back to the caller,
7435 * since these pointers are valid. Only current stack
7436 * pointer will be invalid as soon as function exits,
7437 * but let's be conservative
7439 verbose(env, "cannot return stack pointer to the caller\n");
7443 caller = state->frame[state->curframe - 1];
7444 if (callee->in_callback_fn) {
7445 /* enforce R0 return value range [0, 1]. */
7446 struct tnum range = callee->callback_ret_range;
7448 if (r0->type != SCALAR_VALUE) {
7449 verbose(env, "R0 not a scalar value\n");
7452 if (!tnum_in(range, r0->var_off)) {
7453 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
7457 /* return to the caller whatever r0 had in the callee */
7458 caller->regs[BPF_REG_0] = *r0;
7461 /* callback_fn frame should have released its own additions to parent's
7462 * reference state at this point, or check_reference_leak would
7463 * complain, hence it must be the same as the caller. There is no need
7466 if (!callee->in_callback_fn) {
7467 /* Transfer references to the caller */
7468 err = copy_reference_state(caller, callee);
7473 *insn_idx = callee->callsite + 1;
7474 if (env->log.level & BPF_LOG_LEVEL) {
7475 verbose(env, "returning from callee:\n");
7476 print_verifier_state(env, callee, true);
7477 verbose(env, "to caller at %d:\n", *insn_idx);
7478 print_verifier_state(env, caller, true);
7480 /* clear everything in the callee */
7481 free_func_state(callee);
7482 state->frame[state->curframe--] = NULL;
7486 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
7488 struct bpf_call_arg_meta *meta)
7490 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
7492 if (ret_type != RET_INTEGER ||
7493 (func_id != BPF_FUNC_get_stack &&
7494 func_id != BPF_FUNC_get_task_stack &&
7495 func_id != BPF_FUNC_probe_read_str &&
7496 func_id != BPF_FUNC_probe_read_kernel_str &&
7497 func_id != BPF_FUNC_probe_read_user_str))
7500 ret_reg->smax_value = meta->msize_max_value;
7501 ret_reg->s32_max_value = meta->msize_max_value;
7502 ret_reg->smin_value = -MAX_ERRNO;
7503 ret_reg->s32_min_value = -MAX_ERRNO;
7504 reg_bounds_sync(ret_reg);
7508 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
7509 int func_id, int insn_idx)
7511 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
7512 struct bpf_map *map = meta->map_ptr;
7514 if (func_id != BPF_FUNC_tail_call &&
7515 func_id != BPF_FUNC_map_lookup_elem &&
7516 func_id != BPF_FUNC_map_update_elem &&
7517 func_id != BPF_FUNC_map_delete_elem &&
7518 func_id != BPF_FUNC_map_push_elem &&
7519 func_id != BPF_FUNC_map_pop_elem &&
7520 func_id != BPF_FUNC_map_peek_elem &&
7521 func_id != BPF_FUNC_for_each_map_elem &&
7522 func_id != BPF_FUNC_redirect_map &&
7523 func_id != BPF_FUNC_map_lookup_percpu_elem)
7527 verbose(env, "kernel subsystem misconfigured verifier\n");
7531 /* In case of read-only, some additional restrictions
7532 * need to be applied in order to prevent altering the
7533 * state of the map from program side.
7535 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
7536 (func_id == BPF_FUNC_map_delete_elem ||
7537 func_id == BPF_FUNC_map_update_elem ||
7538 func_id == BPF_FUNC_map_push_elem ||
7539 func_id == BPF_FUNC_map_pop_elem)) {
7540 verbose(env, "write into map forbidden\n");
7544 if (!BPF_MAP_PTR(aux->map_ptr_state))
7545 bpf_map_ptr_store(aux, meta->map_ptr,
7546 !meta->map_ptr->bypass_spec_v1);
7547 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
7548 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
7549 !meta->map_ptr->bypass_spec_v1);
7554 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
7555 int func_id, int insn_idx)
7557 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
7558 struct bpf_reg_state *regs = cur_regs(env), *reg;
7559 struct bpf_map *map = meta->map_ptr;
7563 if (func_id != BPF_FUNC_tail_call)
7565 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
7566 verbose(env, "kernel subsystem misconfigured verifier\n");
7570 reg = ®s[BPF_REG_3];
7571 val = reg->var_off.value;
7572 max = map->max_entries;
7574 if (!(register_is_const(reg) && val < max)) {
7575 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
7579 err = mark_chain_precision(env, BPF_REG_3);
7582 if (bpf_map_key_unseen(aux))
7583 bpf_map_key_store(aux, val);
7584 else if (!bpf_map_key_poisoned(aux) &&
7585 bpf_map_key_immediate(aux) != val)
7586 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
7590 static int check_reference_leak(struct bpf_verifier_env *env)
7592 struct bpf_func_state *state = cur_func(env);
7593 bool refs_lingering = false;
7596 if (state->frameno && !state->in_callback_fn)
7599 for (i = 0; i < state->acquired_refs; i++) {
7600 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno)
7602 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
7603 state->refs[i].id, state->refs[i].insn_idx);
7604 refs_lingering = true;
7606 return refs_lingering ? -EINVAL : 0;
7609 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
7610 struct bpf_reg_state *regs)
7612 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
7613 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
7614 struct bpf_map *fmt_map = fmt_reg->map_ptr;
7615 struct bpf_bprintf_data data = {};
7616 int err, fmt_map_off, num_args;
7620 /* data must be an array of u64 */
7621 if (data_len_reg->var_off.value % 8)
7623 num_args = data_len_reg->var_off.value / 8;
7625 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
7626 * and map_direct_value_addr is set.
7628 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
7629 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
7632 verbose(env, "verifier bug\n");
7635 fmt = (char *)(long)fmt_addr + fmt_map_off;
7637 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
7638 * can focus on validating the format specifiers.
7640 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, num_args, &data);
7642 verbose(env, "Invalid format string\n");
7647 static int check_get_func_ip(struct bpf_verifier_env *env)
7649 enum bpf_prog_type type = resolve_prog_type(env->prog);
7650 int func_id = BPF_FUNC_get_func_ip;
7652 if (type == BPF_PROG_TYPE_TRACING) {
7653 if (!bpf_prog_has_trampoline(env->prog)) {
7654 verbose(env, "func %s#%d supported only for fentry/fexit/fmod_ret programs\n",
7655 func_id_name(func_id), func_id);
7659 } else if (type == BPF_PROG_TYPE_KPROBE) {
7663 verbose(env, "func %s#%d not supported for program type %d\n",
7664 func_id_name(func_id), func_id, type);
7668 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
7670 return &env->insn_aux_data[env->insn_idx];
7673 static bool loop_flag_is_zero(struct bpf_verifier_env *env)
7675 struct bpf_reg_state *regs = cur_regs(env);
7676 struct bpf_reg_state *reg = ®s[BPF_REG_4];
7677 bool reg_is_null = register_is_null(reg);
7680 mark_chain_precision(env, BPF_REG_4);
7685 static void update_loop_inline_state(struct bpf_verifier_env *env, u32 subprogno)
7687 struct bpf_loop_inline_state *state = &cur_aux(env)->loop_inline_state;
7689 if (!state->initialized) {
7690 state->initialized = 1;
7691 state->fit_for_inline = loop_flag_is_zero(env);
7692 state->callback_subprogno = subprogno;
7696 if (!state->fit_for_inline)
7699 state->fit_for_inline = (loop_flag_is_zero(env) &&
7700 state->callback_subprogno == subprogno);
7703 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
7706 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
7707 const struct bpf_func_proto *fn = NULL;
7708 enum bpf_return_type ret_type;
7709 enum bpf_type_flag ret_flag;
7710 struct bpf_reg_state *regs;
7711 struct bpf_call_arg_meta meta;
7712 int insn_idx = *insn_idx_p;
7714 int i, err, func_id;
7716 /* find function prototype */
7717 func_id = insn->imm;
7718 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
7719 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
7724 if (env->ops->get_func_proto)
7725 fn = env->ops->get_func_proto(func_id, env->prog);
7727 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
7732 /* eBPF programs must be GPL compatible to use GPL-ed functions */
7733 if (!env->prog->gpl_compatible && fn->gpl_only) {
7734 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
7738 if (fn->allowed && !fn->allowed(env->prog)) {
7739 verbose(env, "helper call is not allowed in probe\n");
7743 if (!env->prog->aux->sleepable && fn->might_sleep) {
7744 verbose(env, "helper call might sleep in a non-sleepable prog\n");
7748 /* With LD_ABS/IND some JITs save/restore skb from r1. */
7749 changes_data = bpf_helper_changes_pkt_data(fn->func);
7750 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
7751 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
7752 func_id_name(func_id), func_id);
7756 memset(&meta, 0, sizeof(meta));
7757 meta.pkt_access = fn->pkt_access;
7759 err = check_func_proto(fn, func_id);
7761 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
7762 func_id_name(func_id), func_id);
7766 if (env->cur_state->active_rcu_lock) {
7767 if (fn->might_sleep) {
7768 verbose(env, "sleepable helper %s#%d in rcu_read_lock region\n",
7769 func_id_name(func_id), func_id);
7773 if (env->prog->aux->sleepable && is_storage_get_function(func_id))
7774 env->insn_aux_data[insn_idx].storage_get_func_atomic = true;
7777 meta.func_id = func_id;
7779 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
7780 err = check_func_arg(env, i, &meta, fn);
7785 err = record_func_map(env, &meta, func_id, insn_idx);
7789 err = record_func_key(env, &meta, func_id, insn_idx);
7793 /* Mark slots with STACK_MISC in case of raw mode, stack offset
7794 * is inferred from register state.
7796 for (i = 0; i < meta.access_size; i++) {
7797 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
7798 BPF_WRITE, -1, false);
7803 regs = cur_regs(env);
7805 /* This can only be set for PTR_TO_STACK, as CONST_PTR_TO_DYNPTR cannot
7806 * be reinitialized by any dynptr helper. Hence, mark_stack_slots_dynptr
7807 * is safe to do directly.
7809 if (meta.uninit_dynptr_regno) {
7810 if (regs[meta.uninit_dynptr_regno].type == CONST_PTR_TO_DYNPTR) {
7811 verbose(env, "verifier internal error: CONST_PTR_TO_DYNPTR cannot be initialized\n");
7814 /* we write BPF_DW bits (8 bytes) at a time */
7815 for (i = 0; i < BPF_DYNPTR_SIZE; i += 8) {
7816 err = check_mem_access(env, insn_idx, meta.uninit_dynptr_regno,
7817 i, BPF_DW, BPF_WRITE, -1, false);
7822 err = mark_stack_slots_dynptr(env, ®s[meta.uninit_dynptr_regno],
7823 fn->arg_type[meta.uninit_dynptr_regno - BPF_REG_1],
7829 if (meta.release_regno) {
7831 /* This can only be set for PTR_TO_STACK, as CONST_PTR_TO_DYNPTR cannot
7832 * be released by any dynptr helper. Hence, unmark_stack_slots_dynptr
7833 * is safe to do directly.
7835 if (arg_type_is_dynptr(fn->arg_type[meta.release_regno - BPF_REG_1])) {
7836 if (regs[meta.release_regno].type == CONST_PTR_TO_DYNPTR) {
7837 verbose(env, "verifier internal error: CONST_PTR_TO_DYNPTR cannot be released\n");
7840 err = unmark_stack_slots_dynptr(env, ®s[meta.release_regno]);
7841 } else if (meta.ref_obj_id) {
7842 err = release_reference(env, meta.ref_obj_id);
7843 } else if (register_is_null(®s[meta.release_regno])) {
7844 /* meta.ref_obj_id can only be 0 if register that is meant to be
7845 * released is NULL, which must be > R0.
7850 verbose(env, "func %s#%d reference has not been acquired before\n",
7851 func_id_name(func_id), func_id);
7857 case BPF_FUNC_tail_call:
7858 err = check_reference_leak(env);
7860 verbose(env, "tail_call would lead to reference leak\n");
7864 case BPF_FUNC_get_local_storage:
7865 /* check that flags argument in get_local_storage(map, flags) is 0,
7866 * this is required because get_local_storage() can't return an error.
7868 if (!register_is_null(®s[BPF_REG_2])) {
7869 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
7873 case BPF_FUNC_for_each_map_elem:
7874 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
7875 set_map_elem_callback_state);
7877 case BPF_FUNC_timer_set_callback:
7878 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
7879 set_timer_callback_state);
7881 case BPF_FUNC_find_vma:
7882 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
7883 set_find_vma_callback_state);
7885 case BPF_FUNC_snprintf:
7886 err = check_bpf_snprintf_call(env, regs);
7889 update_loop_inline_state(env, meta.subprogno);
7890 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
7891 set_loop_callback_state);
7893 case BPF_FUNC_dynptr_from_mem:
7894 if (regs[BPF_REG_1].type != PTR_TO_MAP_VALUE) {
7895 verbose(env, "Unsupported reg type %s for bpf_dynptr_from_mem data\n",
7896 reg_type_str(env, regs[BPF_REG_1].type));
7900 case BPF_FUNC_set_retval:
7901 if (prog_type == BPF_PROG_TYPE_LSM &&
7902 env->prog->expected_attach_type == BPF_LSM_CGROUP) {
7903 if (!env->prog->aux->attach_func_proto->type) {
7904 /* Make sure programs that attach to void
7905 * hooks don't try to modify return value.
7907 verbose(env, "BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
7912 case BPF_FUNC_dynptr_data:
7913 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
7914 if (arg_type_is_dynptr(fn->arg_type[i])) {
7915 struct bpf_reg_state *reg = ®s[BPF_REG_1 + i];
7917 if (meta.ref_obj_id) {
7918 verbose(env, "verifier internal error: meta.ref_obj_id already set\n");
7922 meta.ref_obj_id = dynptr_ref_obj_id(env, reg);
7926 if (i == MAX_BPF_FUNC_REG_ARGS) {
7927 verbose(env, "verifier internal error: no dynptr in bpf_dynptr_data()\n");
7931 case BPF_FUNC_user_ringbuf_drain:
7932 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
7933 set_user_ringbuf_callback_state);
7940 /* reset caller saved regs */
7941 for (i = 0; i < CALLER_SAVED_REGS; i++) {
7942 mark_reg_not_init(env, regs, caller_saved[i]);
7943 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
7946 /* helper call returns 64-bit value. */
7947 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
7949 /* update return register (already marked as written above) */
7950 ret_type = fn->ret_type;
7951 ret_flag = type_flag(ret_type);
7953 switch (base_type(ret_type)) {
7955 /* sets type to SCALAR_VALUE */
7956 mark_reg_unknown(env, regs, BPF_REG_0);
7959 regs[BPF_REG_0].type = NOT_INIT;
7961 case RET_PTR_TO_MAP_VALUE:
7962 /* There is no offset yet applied, variable or fixed */
7963 mark_reg_known_zero(env, regs, BPF_REG_0);
7964 /* remember map_ptr, so that check_map_access()
7965 * can check 'value_size' boundary of memory access
7966 * to map element returned from bpf_map_lookup_elem()
7968 if (meta.map_ptr == NULL) {
7970 "kernel subsystem misconfigured verifier\n");
7973 regs[BPF_REG_0].map_ptr = meta.map_ptr;
7974 regs[BPF_REG_0].map_uid = meta.map_uid;
7975 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE | ret_flag;
7976 if (!type_may_be_null(ret_type) &&
7977 btf_record_has_field(meta.map_ptr->record, BPF_SPIN_LOCK)) {
7978 regs[BPF_REG_0].id = ++env->id_gen;
7981 case RET_PTR_TO_SOCKET:
7982 mark_reg_known_zero(env, regs, BPF_REG_0);
7983 regs[BPF_REG_0].type = PTR_TO_SOCKET | ret_flag;
7985 case RET_PTR_TO_SOCK_COMMON:
7986 mark_reg_known_zero(env, regs, BPF_REG_0);
7987 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON | ret_flag;
7989 case RET_PTR_TO_TCP_SOCK:
7990 mark_reg_known_zero(env, regs, BPF_REG_0);
7991 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK | ret_flag;
7993 case RET_PTR_TO_MEM:
7994 mark_reg_known_zero(env, regs, BPF_REG_0);
7995 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
7996 regs[BPF_REG_0].mem_size = meta.mem_size;
7998 case RET_PTR_TO_MEM_OR_BTF_ID:
8000 const struct btf_type *t;
8002 mark_reg_known_zero(env, regs, BPF_REG_0);
8003 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
8004 if (!btf_type_is_struct(t)) {
8006 const struct btf_type *ret;
8009 /* resolve the type size of ksym. */
8010 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
8012 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
8013 verbose(env, "unable to resolve the size of type '%s': %ld\n",
8014 tname, PTR_ERR(ret));
8017 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
8018 regs[BPF_REG_0].mem_size = tsize;
8020 /* MEM_RDONLY may be carried from ret_flag, but it
8021 * doesn't apply on PTR_TO_BTF_ID. Fold it, otherwise
8022 * it will confuse the check of PTR_TO_BTF_ID in
8023 * check_mem_access().
8025 ret_flag &= ~MEM_RDONLY;
8027 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
8028 regs[BPF_REG_0].btf = meta.ret_btf;
8029 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
8033 case RET_PTR_TO_BTF_ID:
8035 struct btf *ret_btf;
8038 mark_reg_known_zero(env, regs, BPF_REG_0);
8039 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
8040 if (func_id == BPF_FUNC_kptr_xchg) {
8041 ret_btf = meta.kptr_field->kptr.btf;
8042 ret_btf_id = meta.kptr_field->kptr.btf_id;
8044 if (fn->ret_btf_id == BPF_PTR_POISON) {
8045 verbose(env, "verifier internal error:");
8046 verbose(env, "func %s has non-overwritten BPF_PTR_POISON return type\n",
8047 func_id_name(func_id));
8050 ret_btf = btf_vmlinux;
8051 ret_btf_id = *fn->ret_btf_id;
8053 if (ret_btf_id == 0) {
8054 verbose(env, "invalid return type %u of func %s#%d\n",
8055 base_type(ret_type), func_id_name(func_id),
8059 regs[BPF_REG_0].btf = ret_btf;
8060 regs[BPF_REG_0].btf_id = ret_btf_id;
8064 verbose(env, "unknown return type %u of func %s#%d\n",
8065 base_type(ret_type), func_id_name(func_id), func_id);
8069 if (type_may_be_null(regs[BPF_REG_0].type))
8070 regs[BPF_REG_0].id = ++env->id_gen;
8072 if (helper_multiple_ref_obj_use(func_id, meta.map_ptr)) {
8073 verbose(env, "verifier internal error: func %s#%d sets ref_obj_id more than once\n",
8074 func_id_name(func_id), func_id);
8078 if (is_ptr_cast_function(func_id) || is_dynptr_ref_function(func_id)) {
8079 /* For release_reference() */
8080 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
8081 } else if (is_acquire_function(func_id, meta.map_ptr)) {
8082 int id = acquire_reference_state(env, insn_idx);
8086 /* For mark_ptr_or_null_reg() */
8087 regs[BPF_REG_0].id = id;
8088 /* For release_reference() */
8089 regs[BPF_REG_0].ref_obj_id = id;
8092 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
8094 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
8098 if ((func_id == BPF_FUNC_get_stack ||
8099 func_id == BPF_FUNC_get_task_stack) &&
8100 !env->prog->has_callchain_buf) {
8101 const char *err_str;
8103 #ifdef CONFIG_PERF_EVENTS
8104 err = get_callchain_buffers(sysctl_perf_event_max_stack);
8105 err_str = "cannot get callchain buffer for func %s#%d\n";
8108 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
8111 verbose(env, err_str, func_id_name(func_id), func_id);
8115 env->prog->has_callchain_buf = true;
8118 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
8119 env->prog->call_get_stack = true;
8121 if (func_id == BPF_FUNC_get_func_ip) {
8122 if (check_get_func_ip(env))
8124 env->prog->call_get_func_ip = true;
8128 clear_all_pkt_pointers(env);
8132 /* mark_btf_func_reg_size() is used when the reg size is determined by
8133 * the BTF func_proto's return value size and argument.
8135 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
8138 struct bpf_reg_state *reg = &cur_regs(env)[regno];
8140 if (regno == BPF_REG_0) {
8141 /* Function return value */
8142 reg->live |= REG_LIVE_WRITTEN;
8143 reg->subreg_def = reg_size == sizeof(u64) ?
8144 DEF_NOT_SUBREG : env->insn_idx + 1;
8146 /* Function argument */
8147 if (reg_size == sizeof(u64)) {
8148 mark_insn_zext(env, reg);
8149 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
8151 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
8156 struct bpf_kfunc_call_arg_meta {
8161 const struct btf_type *func_proto;
8162 const char *func_name;
8163 /* Out parameters */
8178 struct btf_field *field;
8182 static bool is_kfunc_acquire(struct bpf_kfunc_call_arg_meta *meta)
8184 return meta->kfunc_flags & KF_ACQUIRE;
8187 static bool is_kfunc_ret_null(struct bpf_kfunc_call_arg_meta *meta)
8189 return meta->kfunc_flags & KF_RET_NULL;
8192 static bool is_kfunc_release(struct bpf_kfunc_call_arg_meta *meta)
8194 return meta->kfunc_flags & KF_RELEASE;
8197 static bool is_kfunc_trusted_args(struct bpf_kfunc_call_arg_meta *meta)
8199 return meta->kfunc_flags & KF_TRUSTED_ARGS;
8202 static bool is_kfunc_sleepable(struct bpf_kfunc_call_arg_meta *meta)
8204 return meta->kfunc_flags & KF_SLEEPABLE;
8207 static bool is_kfunc_destructive(struct bpf_kfunc_call_arg_meta *meta)
8209 return meta->kfunc_flags & KF_DESTRUCTIVE;
8212 static bool is_kfunc_rcu(struct bpf_kfunc_call_arg_meta *meta)
8214 return meta->kfunc_flags & KF_RCU;
8217 static bool is_kfunc_arg_kptr_get(struct bpf_kfunc_call_arg_meta *meta, int arg)
8219 return arg == 0 && (meta->kfunc_flags & KF_KPTR_GET);
8222 static bool __kfunc_param_match_suffix(const struct btf *btf,
8223 const struct btf_param *arg,
8226 int suffix_len = strlen(suffix), len;
8227 const char *param_name;
8229 /* In the future, this can be ported to use BTF tagging */
8230 param_name = btf_name_by_offset(btf, arg->name_off);
8231 if (str_is_empty(param_name))
8233 len = strlen(param_name);
8234 if (len < suffix_len)
8236 param_name += len - suffix_len;
8237 return !strncmp(param_name, suffix, suffix_len);
8240 static bool is_kfunc_arg_mem_size(const struct btf *btf,
8241 const struct btf_param *arg,
8242 const struct bpf_reg_state *reg)
8244 const struct btf_type *t;
8246 t = btf_type_skip_modifiers(btf, arg->type, NULL);
8247 if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
8250 return __kfunc_param_match_suffix(btf, arg, "__sz");
8253 static bool is_kfunc_arg_constant(const struct btf *btf, const struct btf_param *arg)
8255 return __kfunc_param_match_suffix(btf, arg, "__k");
8258 static bool is_kfunc_arg_ignore(const struct btf *btf, const struct btf_param *arg)
8260 return __kfunc_param_match_suffix(btf, arg, "__ign");
8263 static bool is_kfunc_arg_alloc_obj(const struct btf *btf, const struct btf_param *arg)
8265 return __kfunc_param_match_suffix(btf, arg, "__alloc");
8268 static bool is_kfunc_arg_scalar_with_name(const struct btf *btf,
8269 const struct btf_param *arg,
8272 int len, target_len = strlen(name);
8273 const char *param_name;
8275 param_name = btf_name_by_offset(btf, arg->name_off);
8276 if (str_is_empty(param_name))
8278 len = strlen(param_name);
8279 if (len != target_len)
8281 if (strcmp(param_name, name))
8289 KF_ARG_LIST_HEAD_ID,
8290 KF_ARG_LIST_NODE_ID,
8293 BTF_ID_LIST(kf_arg_btf_ids)
8294 BTF_ID(struct, bpf_dynptr_kern)
8295 BTF_ID(struct, bpf_list_head)
8296 BTF_ID(struct, bpf_list_node)
8298 static bool __is_kfunc_ptr_arg_type(const struct btf *btf,
8299 const struct btf_param *arg, int type)
8301 const struct btf_type *t;
8304 t = btf_type_skip_modifiers(btf, arg->type, NULL);
8307 if (!btf_type_is_ptr(t))
8309 t = btf_type_skip_modifiers(btf, t->type, &res_id);
8312 return btf_types_are_same(btf, res_id, btf_vmlinux, kf_arg_btf_ids[type]);
8315 static bool is_kfunc_arg_dynptr(const struct btf *btf, const struct btf_param *arg)
8317 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_DYNPTR_ID);
8320 static bool is_kfunc_arg_list_head(const struct btf *btf, const struct btf_param *arg)
8322 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_HEAD_ID);
8325 static bool is_kfunc_arg_list_node(const struct btf *btf, const struct btf_param *arg)
8327 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_NODE_ID);
8330 /* Returns true if struct is composed of scalars, 4 levels of nesting allowed */
8331 static bool __btf_type_is_scalar_struct(struct bpf_verifier_env *env,
8332 const struct btf *btf,
8333 const struct btf_type *t, int rec)
8335 const struct btf_type *member_type;
8336 const struct btf_member *member;
8339 if (!btf_type_is_struct(t))
8342 for_each_member(i, t, member) {
8343 const struct btf_array *array;
8345 member_type = btf_type_skip_modifiers(btf, member->type, NULL);
8346 if (btf_type_is_struct(member_type)) {
8348 verbose(env, "max struct nesting depth exceeded\n");
8351 if (!__btf_type_is_scalar_struct(env, btf, member_type, rec + 1))
8355 if (btf_type_is_array(member_type)) {
8356 array = btf_array(member_type);
8359 member_type = btf_type_skip_modifiers(btf, array->type, NULL);
8360 if (!btf_type_is_scalar(member_type))
8364 if (!btf_type_is_scalar(member_type))
8371 static u32 *reg2btf_ids[__BPF_REG_TYPE_MAX] = {
8373 [PTR_TO_SOCKET] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK],
8374 [PTR_TO_SOCK_COMMON] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
8375 [PTR_TO_TCP_SOCK] = &btf_sock_ids[BTF_SOCK_TYPE_TCP],
8379 enum kfunc_ptr_arg_type {
8381 KF_ARG_PTR_TO_ALLOC_BTF_ID, /* Allocated object */
8382 KF_ARG_PTR_TO_KPTR, /* PTR_TO_KPTR but type specific */
8383 KF_ARG_PTR_TO_DYNPTR,
8384 KF_ARG_PTR_TO_LIST_HEAD,
8385 KF_ARG_PTR_TO_LIST_NODE,
8386 KF_ARG_PTR_TO_BTF_ID, /* Also covers reg2btf_ids conversions */
8388 KF_ARG_PTR_TO_MEM_SIZE, /* Size derived from next argument, skip it */
8391 enum special_kfunc_type {
8392 KF_bpf_obj_new_impl,
8393 KF_bpf_obj_drop_impl,
8394 KF_bpf_list_push_front,
8395 KF_bpf_list_push_back,
8396 KF_bpf_list_pop_front,
8397 KF_bpf_list_pop_back,
8398 KF_bpf_cast_to_kern_ctx,
8400 KF_bpf_rcu_read_lock,
8401 KF_bpf_rcu_read_unlock,
8404 BTF_SET_START(special_kfunc_set)
8405 BTF_ID(func, bpf_obj_new_impl)
8406 BTF_ID(func, bpf_obj_drop_impl)
8407 BTF_ID(func, bpf_list_push_front)
8408 BTF_ID(func, bpf_list_push_back)
8409 BTF_ID(func, bpf_list_pop_front)
8410 BTF_ID(func, bpf_list_pop_back)
8411 BTF_ID(func, bpf_cast_to_kern_ctx)
8412 BTF_ID(func, bpf_rdonly_cast)
8413 BTF_SET_END(special_kfunc_set)
8415 BTF_ID_LIST(special_kfunc_list)
8416 BTF_ID(func, bpf_obj_new_impl)
8417 BTF_ID(func, bpf_obj_drop_impl)
8418 BTF_ID(func, bpf_list_push_front)
8419 BTF_ID(func, bpf_list_push_back)
8420 BTF_ID(func, bpf_list_pop_front)
8421 BTF_ID(func, bpf_list_pop_back)
8422 BTF_ID(func, bpf_cast_to_kern_ctx)
8423 BTF_ID(func, bpf_rdonly_cast)
8424 BTF_ID(func, bpf_rcu_read_lock)
8425 BTF_ID(func, bpf_rcu_read_unlock)
8427 static bool is_kfunc_bpf_rcu_read_lock(struct bpf_kfunc_call_arg_meta *meta)
8429 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_lock];
8432 static bool is_kfunc_bpf_rcu_read_unlock(struct bpf_kfunc_call_arg_meta *meta)
8434 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_unlock];
8437 static enum kfunc_ptr_arg_type
8438 get_kfunc_ptr_arg_type(struct bpf_verifier_env *env,
8439 struct bpf_kfunc_call_arg_meta *meta,
8440 const struct btf_type *t, const struct btf_type *ref_t,
8441 const char *ref_tname, const struct btf_param *args,
8442 int argno, int nargs)
8444 u32 regno = argno + 1;
8445 struct bpf_reg_state *regs = cur_regs(env);
8446 struct bpf_reg_state *reg = ®s[regno];
8447 bool arg_mem_size = false;
8449 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx])
8450 return KF_ARG_PTR_TO_CTX;
8452 /* In this function, we verify the kfunc's BTF as per the argument type,
8453 * leaving the rest of the verification with respect to the register
8454 * type to our caller. When a set of conditions hold in the BTF type of
8455 * arguments, we resolve it to a known kfunc_ptr_arg_type.
8457 if (btf_get_prog_ctx_type(&env->log, meta->btf, t, resolve_prog_type(env->prog), argno))
8458 return KF_ARG_PTR_TO_CTX;
8460 if (is_kfunc_arg_alloc_obj(meta->btf, &args[argno]))
8461 return KF_ARG_PTR_TO_ALLOC_BTF_ID;
8463 if (is_kfunc_arg_kptr_get(meta, argno)) {
8464 if (!btf_type_is_ptr(ref_t)) {
8465 verbose(env, "arg#0 BTF type must be a double pointer for kptr_get kfunc\n");
8468 ref_t = btf_type_by_id(meta->btf, ref_t->type);
8469 ref_tname = btf_name_by_offset(meta->btf, ref_t->name_off);
8470 if (!btf_type_is_struct(ref_t)) {
8471 verbose(env, "kernel function %s args#0 pointer type %s %s is not supported\n",
8472 meta->func_name, btf_type_str(ref_t), ref_tname);
8475 return KF_ARG_PTR_TO_KPTR;
8478 if (is_kfunc_arg_dynptr(meta->btf, &args[argno]))
8479 return KF_ARG_PTR_TO_DYNPTR;
8481 if (is_kfunc_arg_list_head(meta->btf, &args[argno]))
8482 return KF_ARG_PTR_TO_LIST_HEAD;
8484 if (is_kfunc_arg_list_node(meta->btf, &args[argno]))
8485 return KF_ARG_PTR_TO_LIST_NODE;
8487 if ((base_type(reg->type) == PTR_TO_BTF_ID || reg2btf_ids[base_type(reg->type)])) {
8488 if (!btf_type_is_struct(ref_t)) {
8489 verbose(env, "kernel function %s args#%d pointer type %s %s is not supported\n",
8490 meta->func_name, argno, btf_type_str(ref_t), ref_tname);
8493 return KF_ARG_PTR_TO_BTF_ID;
8496 if (argno + 1 < nargs && is_kfunc_arg_mem_size(meta->btf, &args[argno + 1], ®s[regno + 1]))
8497 arg_mem_size = true;
8499 /* This is the catch all argument type of register types supported by
8500 * check_helper_mem_access. However, we only allow when argument type is
8501 * pointer to scalar, or struct composed (recursively) of scalars. When
8502 * arg_mem_size is true, the pointer can be void *.
8504 if (!btf_type_is_scalar(ref_t) && !__btf_type_is_scalar_struct(env, meta->btf, ref_t, 0) &&
8505 (arg_mem_size ? !btf_type_is_void(ref_t) : 1)) {
8506 verbose(env, "arg#%d pointer type %s %s must point to %sscalar, or struct with scalar\n",
8507 argno, btf_type_str(ref_t), ref_tname, arg_mem_size ? "void, " : "");
8510 return arg_mem_size ? KF_ARG_PTR_TO_MEM_SIZE : KF_ARG_PTR_TO_MEM;
8513 static int process_kf_arg_ptr_to_btf_id(struct bpf_verifier_env *env,
8514 struct bpf_reg_state *reg,
8515 const struct btf_type *ref_t,
8516 const char *ref_tname, u32 ref_id,
8517 struct bpf_kfunc_call_arg_meta *meta,
8520 const struct btf_type *reg_ref_t;
8521 bool strict_type_match = false;
8522 const struct btf *reg_btf;
8523 const char *reg_ref_tname;
8526 if (base_type(reg->type) == PTR_TO_BTF_ID) {
8528 reg_ref_id = reg->btf_id;
8530 reg_btf = btf_vmlinux;
8531 reg_ref_id = *reg2btf_ids[base_type(reg->type)];
8534 if (is_kfunc_trusted_args(meta) || (is_kfunc_release(meta) && reg->ref_obj_id))
8535 strict_type_match = true;
8537 reg_ref_t = btf_type_skip_modifiers(reg_btf, reg_ref_id, ®_ref_id);
8538 reg_ref_tname = btf_name_by_offset(reg_btf, reg_ref_t->name_off);
8539 if (!btf_struct_ids_match(&env->log, reg_btf, reg_ref_id, reg->off, meta->btf, ref_id, strict_type_match)) {
8540 verbose(env, "kernel function %s args#%d expected pointer to %s %s but R%d has a pointer to %s %s\n",
8541 meta->func_name, argno, btf_type_str(ref_t), ref_tname, argno + 1,
8542 btf_type_str(reg_ref_t), reg_ref_tname);
8548 static int process_kf_arg_ptr_to_kptr(struct bpf_verifier_env *env,
8549 struct bpf_reg_state *reg,
8550 const struct btf_type *ref_t,
8551 const char *ref_tname,
8552 struct bpf_kfunc_call_arg_meta *meta,
8555 struct btf_field *kptr_field;
8557 /* check_func_arg_reg_off allows var_off for
8558 * PTR_TO_MAP_VALUE, but we need fixed offset to find
8561 if (!tnum_is_const(reg->var_off)) {
8562 verbose(env, "arg#0 must have constant offset\n");
8566 kptr_field = btf_record_find(reg->map_ptr->record, reg->off + reg->var_off.value, BPF_KPTR);
8567 if (!kptr_field || kptr_field->type != BPF_KPTR_REF) {
8568 verbose(env, "arg#0 no referenced kptr at map value offset=%llu\n",
8569 reg->off + reg->var_off.value);
8573 if (!btf_struct_ids_match(&env->log, meta->btf, ref_t->type, 0, kptr_field->kptr.btf,
8574 kptr_field->kptr.btf_id, true)) {
8575 verbose(env, "kernel function %s args#%d expected pointer to %s %s\n",
8576 meta->func_name, argno, btf_type_str(ref_t), ref_tname);
8582 static int ref_set_release_on_unlock(struct bpf_verifier_env *env, u32 ref_obj_id)
8584 struct bpf_func_state *state = cur_func(env);
8585 struct bpf_reg_state *reg;
8588 /* bpf_spin_lock only allows calling list_push and list_pop, no BPF
8589 * subprogs, no global functions. This means that the references would
8590 * not be released inside the critical section but they may be added to
8591 * the reference state, and the acquired_refs are never copied out for a
8592 * different frame as BPF to BPF calls don't work in bpf_spin_lock
8593 * critical sections.
8596 verbose(env, "verifier internal error: ref_obj_id is zero for release_on_unlock\n");
8599 for (i = 0; i < state->acquired_refs; i++) {
8600 if (state->refs[i].id == ref_obj_id) {
8601 if (state->refs[i].release_on_unlock) {
8602 verbose(env, "verifier internal error: expected false release_on_unlock");
8605 state->refs[i].release_on_unlock = true;
8606 /* Now mark everyone sharing same ref_obj_id as untrusted */
8607 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
8608 if (reg->ref_obj_id == ref_obj_id)
8609 reg->type |= PTR_UNTRUSTED;
8614 verbose(env, "verifier internal error: ref state missing for ref_obj_id\n");
8618 /* Implementation details:
8620 * Each register points to some region of memory, which we define as an
8621 * allocation. Each allocation may embed a bpf_spin_lock which protects any
8622 * special BPF objects (bpf_list_head, bpf_rb_root, etc.) part of the same
8623 * allocation. The lock and the data it protects are colocated in the same
8626 * Hence, everytime a register holds a pointer value pointing to such
8627 * allocation, the verifier preserves a unique reg->id for it.
8629 * The verifier remembers the lock 'ptr' and the lock 'id' whenever
8630 * bpf_spin_lock is called.
8632 * To enable this, lock state in the verifier captures two values:
8633 * active_lock.ptr = Register's type specific pointer
8634 * active_lock.id = A unique ID for each register pointer value
8636 * Currently, PTR_TO_MAP_VALUE and PTR_TO_BTF_ID | MEM_ALLOC are the two
8637 * supported register types.
8639 * The active_lock.ptr in case of map values is the reg->map_ptr, and in case of
8640 * allocated objects is the reg->btf pointer.
8642 * The active_lock.id is non-unique for maps supporting direct_value_addr, as we
8643 * can establish the provenance of the map value statically for each distinct
8644 * lookup into such maps. They always contain a single map value hence unique
8645 * IDs for each pseudo load pessimizes the algorithm and rejects valid programs.
8647 * So, in case of global variables, they use array maps with max_entries = 1,
8648 * hence their active_lock.ptr becomes map_ptr and id = 0 (since they all point
8649 * into the same map value as max_entries is 1, as described above).
8651 * In case of inner map lookups, the inner map pointer has same map_ptr as the
8652 * outer map pointer (in verifier context), but each lookup into an inner map
8653 * assigns a fresh reg->id to the lookup, so while lookups into distinct inner
8654 * maps from the same outer map share the same map_ptr as active_lock.ptr, they
8655 * will get different reg->id assigned to each lookup, hence different
8658 * In case of allocated objects, active_lock.ptr is the reg->btf, and the
8659 * reg->id is a unique ID preserved after the NULL pointer check on the pointer
8660 * returned from bpf_obj_new. Each allocation receives a new reg->id.
8662 static int check_reg_allocation_locked(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
8667 switch ((int)reg->type) {
8668 case PTR_TO_MAP_VALUE:
8671 case PTR_TO_BTF_ID | MEM_ALLOC:
8672 case PTR_TO_BTF_ID | MEM_ALLOC | PTR_TRUSTED:
8676 verbose(env, "verifier internal error: unknown reg type for lock check\n");
8681 if (!env->cur_state->active_lock.ptr)
8683 if (env->cur_state->active_lock.ptr != ptr ||
8684 env->cur_state->active_lock.id != id) {
8685 verbose(env, "held lock and object are not in the same allocation\n");
8691 static bool is_bpf_list_api_kfunc(u32 btf_id)
8693 return btf_id == special_kfunc_list[KF_bpf_list_push_front] ||
8694 btf_id == special_kfunc_list[KF_bpf_list_push_back] ||
8695 btf_id == special_kfunc_list[KF_bpf_list_pop_front] ||
8696 btf_id == special_kfunc_list[KF_bpf_list_pop_back];
8699 static int process_kf_arg_ptr_to_list_head(struct bpf_verifier_env *env,
8700 struct bpf_reg_state *reg, u32 regno,
8701 struct bpf_kfunc_call_arg_meta *meta)
8703 struct btf_field *field;
8704 struct btf_record *rec;
8707 if (meta->btf != btf_vmlinux || !is_bpf_list_api_kfunc(meta->func_id)) {
8708 verbose(env, "verifier internal error: bpf_list_head argument for unknown kfunc\n");
8712 if (!tnum_is_const(reg->var_off)) {
8714 "R%d doesn't have constant offset. bpf_list_head has to be at the constant offset\n",
8719 rec = reg_btf_record(reg);
8720 list_head_off = reg->off + reg->var_off.value;
8721 field = btf_record_find(rec, list_head_off, BPF_LIST_HEAD);
8723 verbose(env, "bpf_list_head not found at offset=%u\n", list_head_off);
8727 /* All functions require bpf_list_head to be protected using a bpf_spin_lock */
8728 if (check_reg_allocation_locked(env, reg)) {
8729 verbose(env, "bpf_spin_lock at off=%d must be held for bpf_list_head\n",
8730 rec->spin_lock_off);
8734 if (meta->arg_list_head.field) {
8735 verbose(env, "verifier internal error: repeating bpf_list_head arg\n");
8738 meta->arg_list_head.field = field;
8742 static int process_kf_arg_ptr_to_list_node(struct bpf_verifier_env *env,
8743 struct bpf_reg_state *reg, u32 regno,
8744 struct bpf_kfunc_call_arg_meta *meta)
8746 const struct btf_type *et, *t;
8747 struct btf_field *field;
8748 struct btf_record *rec;
8751 if (meta->btf != btf_vmlinux ||
8752 (meta->func_id != special_kfunc_list[KF_bpf_list_push_front] &&
8753 meta->func_id != special_kfunc_list[KF_bpf_list_push_back])) {
8754 verbose(env, "verifier internal error: bpf_list_node argument for unknown kfunc\n");
8758 if (!tnum_is_const(reg->var_off)) {
8760 "R%d doesn't have constant offset. bpf_list_node has to be at the constant offset\n",
8765 rec = reg_btf_record(reg);
8766 list_node_off = reg->off + reg->var_off.value;
8767 field = btf_record_find(rec, list_node_off, BPF_LIST_NODE);
8768 if (!field || field->offset != list_node_off) {
8769 verbose(env, "bpf_list_node not found at offset=%u\n", list_node_off);
8773 field = meta->arg_list_head.field;
8775 et = btf_type_by_id(field->list_head.btf, field->list_head.value_btf_id);
8776 t = btf_type_by_id(reg->btf, reg->btf_id);
8777 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, 0, field->list_head.btf,
8778 field->list_head.value_btf_id, true)) {
8779 verbose(env, "operation on bpf_list_head expects arg#1 bpf_list_node at offset=%d "
8780 "in struct %s, but arg is at offset=%d in struct %s\n",
8781 field->list_head.node_offset, btf_name_by_offset(field->list_head.btf, et->name_off),
8782 list_node_off, btf_name_by_offset(reg->btf, t->name_off));
8786 if (list_node_off != field->list_head.node_offset) {
8787 verbose(env, "arg#1 offset=%d, but expected bpf_list_node at offset=%d in struct %s\n",
8788 list_node_off, field->list_head.node_offset,
8789 btf_name_by_offset(field->list_head.btf, et->name_off));
8792 /* Set arg#1 for expiration after unlock */
8793 return ref_set_release_on_unlock(env, reg->ref_obj_id);
8796 static int check_kfunc_args(struct bpf_verifier_env *env, struct bpf_kfunc_call_arg_meta *meta)
8798 const char *func_name = meta->func_name, *ref_tname;
8799 const struct btf *btf = meta->btf;
8800 const struct btf_param *args;
8804 args = (const struct btf_param *)(meta->func_proto + 1);
8805 nargs = btf_type_vlen(meta->func_proto);
8806 if (nargs > MAX_BPF_FUNC_REG_ARGS) {
8807 verbose(env, "Function %s has %d > %d args\n", func_name, nargs,
8808 MAX_BPF_FUNC_REG_ARGS);
8812 /* Check that BTF function arguments match actual types that the
8815 for (i = 0; i < nargs; i++) {
8816 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[i + 1];
8817 const struct btf_type *t, *ref_t, *resolve_ret;
8818 enum bpf_arg_type arg_type = ARG_DONTCARE;
8819 u32 regno = i + 1, ref_id, type_size;
8820 bool is_ret_buf_sz = false;
8823 t = btf_type_skip_modifiers(btf, args[i].type, NULL);
8825 if (is_kfunc_arg_ignore(btf, &args[i]))
8828 if (btf_type_is_scalar(t)) {
8829 if (reg->type != SCALAR_VALUE) {
8830 verbose(env, "R%d is not a scalar\n", regno);
8834 if (is_kfunc_arg_constant(meta->btf, &args[i])) {
8835 if (meta->arg_constant.found) {
8836 verbose(env, "verifier internal error: only one constant argument permitted\n");
8839 if (!tnum_is_const(reg->var_off)) {
8840 verbose(env, "R%d must be a known constant\n", regno);
8843 ret = mark_chain_precision(env, regno);
8846 meta->arg_constant.found = true;
8847 meta->arg_constant.value = reg->var_off.value;
8848 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdonly_buf_size")) {
8849 meta->r0_rdonly = true;
8850 is_ret_buf_sz = true;
8851 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdwr_buf_size")) {
8852 is_ret_buf_sz = true;
8855 if (is_ret_buf_sz) {
8856 if (meta->r0_size) {
8857 verbose(env, "2 or more rdonly/rdwr_buf_size parameters for kfunc");
8861 if (!tnum_is_const(reg->var_off)) {
8862 verbose(env, "R%d is not a const\n", regno);
8866 meta->r0_size = reg->var_off.value;
8867 ret = mark_chain_precision(env, regno);
8874 if (!btf_type_is_ptr(t)) {
8875 verbose(env, "Unrecognized arg#%d type %s\n", i, btf_type_str(t));
8879 if (reg->ref_obj_id) {
8880 if (is_kfunc_release(meta) && meta->ref_obj_id) {
8881 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
8882 regno, reg->ref_obj_id,
8886 meta->ref_obj_id = reg->ref_obj_id;
8887 if (is_kfunc_release(meta))
8888 meta->release_regno = regno;
8891 ref_t = btf_type_skip_modifiers(btf, t->type, &ref_id);
8892 ref_tname = btf_name_by_offset(btf, ref_t->name_off);
8894 kf_arg_type = get_kfunc_ptr_arg_type(env, meta, t, ref_t, ref_tname, args, i, nargs);
8895 if (kf_arg_type < 0)
8898 switch (kf_arg_type) {
8899 case KF_ARG_PTR_TO_ALLOC_BTF_ID:
8900 case KF_ARG_PTR_TO_BTF_ID:
8901 if (!is_kfunc_trusted_args(meta) && !is_kfunc_rcu(meta))
8904 if (!is_trusted_reg(reg)) {
8905 if (!is_kfunc_rcu(meta)) {
8906 verbose(env, "R%d must be referenced or trusted\n", regno);
8909 if (!is_rcu_reg(reg)) {
8910 verbose(env, "R%d must be a rcu pointer\n", regno);
8916 case KF_ARG_PTR_TO_CTX:
8917 /* Trusted arguments have the same offset checks as release arguments */
8918 arg_type |= OBJ_RELEASE;
8920 case KF_ARG_PTR_TO_KPTR:
8921 case KF_ARG_PTR_TO_DYNPTR:
8922 case KF_ARG_PTR_TO_LIST_HEAD:
8923 case KF_ARG_PTR_TO_LIST_NODE:
8924 case KF_ARG_PTR_TO_MEM:
8925 case KF_ARG_PTR_TO_MEM_SIZE:
8926 /* Trusted by default */
8933 if (is_kfunc_release(meta) && reg->ref_obj_id)
8934 arg_type |= OBJ_RELEASE;
8935 ret = check_func_arg_reg_off(env, reg, regno, arg_type);
8939 switch (kf_arg_type) {
8940 case KF_ARG_PTR_TO_CTX:
8941 if (reg->type != PTR_TO_CTX) {
8942 verbose(env, "arg#%d expected pointer to ctx, but got %s\n", i, btf_type_str(t));
8946 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
8947 ret = get_kern_ctx_btf_id(&env->log, resolve_prog_type(env->prog));
8950 meta->ret_btf_id = ret;
8953 case KF_ARG_PTR_TO_ALLOC_BTF_ID:
8954 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
8955 verbose(env, "arg#%d expected pointer to allocated object\n", i);
8958 if (!reg->ref_obj_id) {
8959 verbose(env, "allocated object must be referenced\n");
8962 if (meta->btf == btf_vmlinux &&
8963 meta->func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
8964 meta->arg_obj_drop.btf = reg->btf;
8965 meta->arg_obj_drop.btf_id = reg->btf_id;
8968 case KF_ARG_PTR_TO_KPTR:
8969 if (reg->type != PTR_TO_MAP_VALUE) {
8970 verbose(env, "arg#0 expected pointer to map value\n");
8973 ret = process_kf_arg_ptr_to_kptr(env, reg, ref_t, ref_tname, meta, i);
8977 case KF_ARG_PTR_TO_DYNPTR:
8978 if (reg->type != PTR_TO_STACK &&
8979 reg->type != CONST_PTR_TO_DYNPTR) {
8980 verbose(env, "arg#%d expected pointer to stack or dynptr_ptr\n", i);
8984 ret = process_dynptr_func(env, regno, ARG_PTR_TO_DYNPTR | MEM_RDONLY, NULL);
8988 case KF_ARG_PTR_TO_LIST_HEAD:
8989 if (reg->type != PTR_TO_MAP_VALUE &&
8990 reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
8991 verbose(env, "arg#%d expected pointer to map value or allocated object\n", i);
8994 if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
8995 verbose(env, "allocated object must be referenced\n");
8998 ret = process_kf_arg_ptr_to_list_head(env, reg, regno, meta);
9002 case KF_ARG_PTR_TO_LIST_NODE:
9003 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
9004 verbose(env, "arg#%d expected pointer to allocated object\n", i);
9007 if (!reg->ref_obj_id) {
9008 verbose(env, "allocated object must be referenced\n");
9011 ret = process_kf_arg_ptr_to_list_node(env, reg, regno, meta);
9015 case KF_ARG_PTR_TO_BTF_ID:
9016 /* Only base_type is checked, further checks are done here */
9017 if ((base_type(reg->type) != PTR_TO_BTF_ID ||
9018 (bpf_type_has_unsafe_modifiers(reg->type) && !is_rcu_reg(reg))) &&
9019 !reg2btf_ids[base_type(reg->type)]) {
9020 verbose(env, "arg#%d is %s ", i, reg_type_str(env, reg->type));
9021 verbose(env, "expected %s or socket\n",
9022 reg_type_str(env, base_type(reg->type) |
9023 (type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS)));
9026 ret = process_kf_arg_ptr_to_btf_id(env, reg, ref_t, ref_tname, ref_id, meta, i);
9030 case KF_ARG_PTR_TO_MEM:
9031 resolve_ret = btf_resolve_size(btf, ref_t, &type_size);
9032 if (IS_ERR(resolve_ret)) {
9033 verbose(env, "arg#%d reference type('%s %s') size cannot be determined: %ld\n",
9034 i, btf_type_str(ref_t), ref_tname, PTR_ERR(resolve_ret));
9037 ret = check_mem_reg(env, reg, regno, type_size);
9041 case KF_ARG_PTR_TO_MEM_SIZE:
9042 ret = check_kfunc_mem_size_reg(env, ®s[regno + 1], regno + 1);
9044 verbose(env, "arg#%d arg#%d memory, len pair leads to invalid memory access\n", i, i + 1);
9047 /* Skip next '__sz' argument */
9053 if (is_kfunc_release(meta) && !meta->release_regno) {
9054 verbose(env, "release kernel function %s expects refcounted PTR_TO_BTF_ID\n",
9062 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
9065 const struct btf_type *t, *func, *func_proto, *ptr_type;
9066 struct bpf_reg_state *regs = cur_regs(env);
9067 const char *func_name, *ptr_type_name;
9068 bool sleepable, rcu_lock, rcu_unlock;
9069 struct bpf_kfunc_call_arg_meta meta;
9070 u32 i, nargs, func_id, ptr_type_id;
9071 int err, insn_idx = *insn_idx_p;
9072 const struct btf_param *args;
9073 const struct btf_type *ret_t;
9074 struct btf *desc_btf;
9077 /* skip for now, but return error when we find this in fixup_kfunc_call */
9081 desc_btf = find_kfunc_desc_btf(env, insn->off);
9082 if (IS_ERR(desc_btf))
9083 return PTR_ERR(desc_btf);
9085 func_id = insn->imm;
9086 func = btf_type_by_id(desc_btf, func_id);
9087 func_name = btf_name_by_offset(desc_btf, func->name_off);
9088 func_proto = btf_type_by_id(desc_btf, func->type);
9090 kfunc_flags = btf_kfunc_id_set_contains(desc_btf, resolve_prog_type(env->prog), func_id);
9092 verbose(env, "calling kernel function %s is not allowed\n",
9097 /* Prepare kfunc call metadata */
9098 memset(&meta, 0, sizeof(meta));
9099 meta.btf = desc_btf;
9100 meta.func_id = func_id;
9101 meta.kfunc_flags = *kfunc_flags;
9102 meta.func_proto = func_proto;
9103 meta.func_name = func_name;
9105 if (is_kfunc_destructive(&meta) && !capable(CAP_SYS_BOOT)) {
9106 verbose(env, "destructive kfunc calls require CAP_SYS_BOOT capability\n");
9110 sleepable = is_kfunc_sleepable(&meta);
9111 if (sleepable && !env->prog->aux->sleepable) {
9112 verbose(env, "program must be sleepable to call sleepable kfunc %s\n", func_name);
9116 rcu_lock = is_kfunc_bpf_rcu_read_lock(&meta);
9117 rcu_unlock = is_kfunc_bpf_rcu_read_unlock(&meta);
9118 if ((rcu_lock || rcu_unlock) && !env->rcu_tag_supported) {
9119 verbose(env, "no vmlinux btf rcu tag support for kfunc %s\n", func_name);
9123 if (env->cur_state->active_rcu_lock) {
9124 struct bpf_func_state *state;
9125 struct bpf_reg_state *reg;
9128 verbose(env, "nested rcu read lock (kernel function %s)\n", func_name);
9130 } else if (rcu_unlock) {
9131 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
9132 if (reg->type & MEM_RCU) {
9133 reg->type &= ~(MEM_RCU | PTR_MAYBE_NULL);
9134 reg->type |= PTR_UNTRUSTED;
9137 env->cur_state->active_rcu_lock = false;
9138 } else if (sleepable) {
9139 verbose(env, "kernel func %s is sleepable within rcu_read_lock region\n", func_name);
9142 } else if (rcu_lock) {
9143 env->cur_state->active_rcu_lock = true;
9144 } else if (rcu_unlock) {
9145 verbose(env, "unmatched rcu read unlock (kernel function %s)\n", func_name);
9149 /* Check the arguments */
9150 err = check_kfunc_args(env, &meta);
9153 /* In case of release function, we get register number of refcounted
9154 * PTR_TO_BTF_ID in bpf_kfunc_arg_meta, do the release now.
9156 if (meta.release_regno) {
9157 err = release_reference(env, regs[meta.release_regno].ref_obj_id);
9159 verbose(env, "kfunc %s#%d reference has not been acquired before\n",
9160 func_name, func_id);
9165 for (i = 0; i < CALLER_SAVED_REGS; i++)
9166 mark_reg_not_init(env, regs, caller_saved[i]);
9168 /* Check return type */
9169 t = btf_type_skip_modifiers(desc_btf, func_proto->type, NULL);
9171 if (is_kfunc_acquire(&meta) && !btf_type_is_struct_ptr(meta.btf, t)) {
9172 /* Only exception is bpf_obj_new_impl */
9173 if (meta.btf != btf_vmlinux || meta.func_id != special_kfunc_list[KF_bpf_obj_new_impl]) {
9174 verbose(env, "acquire kernel function does not return PTR_TO_BTF_ID\n");
9179 if (btf_type_is_scalar(t)) {
9180 mark_reg_unknown(env, regs, BPF_REG_0);
9181 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
9182 } else if (btf_type_is_ptr(t)) {
9183 ptr_type = btf_type_skip_modifiers(desc_btf, t->type, &ptr_type_id);
9185 if (meta.btf == btf_vmlinux && btf_id_set_contains(&special_kfunc_set, meta.func_id)) {
9186 if (meta.func_id == special_kfunc_list[KF_bpf_obj_new_impl]) {
9187 struct btf *ret_btf;
9190 if (unlikely(!bpf_global_ma_set))
9193 if (((u64)(u32)meta.arg_constant.value) != meta.arg_constant.value) {
9194 verbose(env, "local type ID argument must be in range [0, U32_MAX]\n");
9198 ret_btf = env->prog->aux->btf;
9199 ret_btf_id = meta.arg_constant.value;
9201 /* This may be NULL due to user not supplying a BTF */
9203 verbose(env, "bpf_obj_new requires prog BTF\n");
9207 ret_t = btf_type_by_id(ret_btf, ret_btf_id);
9208 if (!ret_t || !__btf_type_is_struct(ret_t)) {
9209 verbose(env, "bpf_obj_new type ID argument must be of a struct\n");
9213 mark_reg_known_zero(env, regs, BPF_REG_0);
9214 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
9215 regs[BPF_REG_0].btf = ret_btf;
9216 regs[BPF_REG_0].btf_id = ret_btf_id;
9218 env->insn_aux_data[insn_idx].obj_new_size = ret_t->size;
9219 env->insn_aux_data[insn_idx].kptr_struct_meta =
9220 btf_find_struct_meta(ret_btf, ret_btf_id);
9221 } else if (meta.func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
9222 env->insn_aux_data[insn_idx].kptr_struct_meta =
9223 btf_find_struct_meta(meta.arg_obj_drop.btf,
9224 meta.arg_obj_drop.btf_id);
9225 } else if (meta.func_id == special_kfunc_list[KF_bpf_list_pop_front] ||
9226 meta.func_id == special_kfunc_list[KF_bpf_list_pop_back]) {
9227 struct btf_field *field = meta.arg_list_head.field;
9229 mark_reg_known_zero(env, regs, BPF_REG_0);
9230 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
9231 regs[BPF_REG_0].btf = field->list_head.btf;
9232 regs[BPF_REG_0].btf_id = field->list_head.value_btf_id;
9233 regs[BPF_REG_0].off = field->list_head.node_offset;
9234 } else if (meta.func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
9235 mark_reg_known_zero(env, regs, BPF_REG_0);
9236 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_TRUSTED;
9237 regs[BPF_REG_0].btf = desc_btf;
9238 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
9239 } else if (meta.func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
9240 ret_t = btf_type_by_id(desc_btf, meta.arg_constant.value);
9241 if (!ret_t || !btf_type_is_struct(ret_t)) {
9243 "kfunc bpf_rdonly_cast type ID argument must be of a struct\n");
9247 mark_reg_known_zero(env, regs, BPF_REG_0);
9248 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
9249 regs[BPF_REG_0].btf = desc_btf;
9250 regs[BPF_REG_0].btf_id = meta.arg_constant.value;
9252 verbose(env, "kernel function %s unhandled dynamic return type\n",
9256 } else if (!__btf_type_is_struct(ptr_type)) {
9257 if (!meta.r0_size) {
9258 ptr_type_name = btf_name_by_offset(desc_btf,
9259 ptr_type->name_off);
9261 "kernel function %s returns pointer type %s %s is not supported\n",
9263 btf_type_str(ptr_type),
9268 mark_reg_known_zero(env, regs, BPF_REG_0);
9269 regs[BPF_REG_0].type = PTR_TO_MEM;
9270 regs[BPF_REG_0].mem_size = meta.r0_size;
9273 regs[BPF_REG_0].type |= MEM_RDONLY;
9275 /* Ensures we don't access the memory after a release_reference() */
9276 if (meta.ref_obj_id)
9277 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
9279 mark_reg_known_zero(env, regs, BPF_REG_0);
9280 regs[BPF_REG_0].btf = desc_btf;
9281 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
9282 regs[BPF_REG_0].btf_id = ptr_type_id;
9285 if (is_kfunc_ret_null(&meta)) {
9286 regs[BPF_REG_0].type |= PTR_MAYBE_NULL;
9287 /* For mark_ptr_or_null_reg, see 93c230e3f5bd6 */
9288 regs[BPF_REG_0].id = ++env->id_gen;
9290 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
9291 if (is_kfunc_acquire(&meta)) {
9292 int id = acquire_reference_state(env, insn_idx);
9296 if (is_kfunc_ret_null(&meta))
9297 regs[BPF_REG_0].id = id;
9298 regs[BPF_REG_0].ref_obj_id = id;
9300 if (reg_may_point_to_spin_lock(®s[BPF_REG_0]) && !regs[BPF_REG_0].id)
9301 regs[BPF_REG_0].id = ++env->id_gen;
9302 } /* else { add_kfunc_call() ensures it is btf_type_is_void(t) } */
9304 nargs = btf_type_vlen(func_proto);
9305 args = (const struct btf_param *)(func_proto + 1);
9306 for (i = 0; i < nargs; i++) {
9309 t = btf_type_skip_modifiers(desc_btf, args[i].type, NULL);
9310 if (btf_type_is_ptr(t))
9311 mark_btf_func_reg_size(env, regno, sizeof(void *));
9313 /* scalar. ensured by btf_check_kfunc_arg_match() */
9314 mark_btf_func_reg_size(env, regno, t->size);
9320 static bool signed_add_overflows(s64 a, s64 b)
9322 /* Do the add in u64, where overflow is well-defined */
9323 s64 res = (s64)((u64)a + (u64)b);
9330 static bool signed_add32_overflows(s32 a, s32 b)
9332 /* Do the add in u32, where overflow is well-defined */
9333 s32 res = (s32)((u32)a + (u32)b);
9340 static bool signed_sub_overflows(s64 a, s64 b)
9342 /* Do the sub in u64, where overflow is well-defined */
9343 s64 res = (s64)((u64)a - (u64)b);
9350 static bool signed_sub32_overflows(s32 a, s32 b)
9352 /* Do the sub in u32, where overflow is well-defined */
9353 s32 res = (s32)((u32)a - (u32)b);
9360 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
9361 const struct bpf_reg_state *reg,
9362 enum bpf_reg_type type)
9364 bool known = tnum_is_const(reg->var_off);
9365 s64 val = reg->var_off.value;
9366 s64 smin = reg->smin_value;
9368 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
9369 verbose(env, "math between %s pointer and %lld is not allowed\n",
9370 reg_type_str(env, type), val);
9374 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
9375 verbose(env, "%s pointer offset %d is not allowed\n",
9376 reg_type_str(env, type), reg->off);
9380 if (smin == S64_MIN) {
9381 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
9382 reg_type_str(env, type));
9386 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
9387 verbose(env, "value %lld makes %s pointer be out of bounds\n",
9388 smin, reg_type_str(env, type));
9403 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
9404 u32 *alu_limit, bool mask_to_left)
9406 u32 max = 0, ptr_limit = 0;
9408 switch (ptr_reg->type) {
9410 /* Offset 0 is out-of-bounds, but acceptable start for the
9411 * left direction, see BPF_REG_FP. Also, unknown scalar
9412 * offset where we would need to deal with min/max bounds is
9413 * currently prohibited for unprivileged.
9415 max = MAX_BPF_STACK + mask_to_left;
9416 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
9418 case PTR_TO_MAP_VALUE:
9419 max = ptr_reg->map_ptr->value_size;
9420 ptr_limit = (mask_to_left ?
9421 ptr_reg->smin_value :
9422 ptr_reg->umax_value) + ptr_reg->off;
9428 if (ptr_limit >= max)
9429 return REASON_LIMIT;
9430 *alu_limit = ptr_limit;
9434 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
9435 const struct bpf_insn *insn)
9437 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
9440 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
9441 u32 alu_state, u32 alu_limit)
9443 /* If we arrived here from different branches with different
9444 * state or limits to sanitize, then this won't work.
9446 if (aux->alu_state &&
9447 (aux->alu_state != alu_state ||
9448 aux->alu_limit != alu_limit))
9449 return REASON_PATHS;
9451 /* Corresponding fixup done in do_misc_fixups(). */
9452 aux->alu_state = alu_state;
9453 aux->alu_limit = alu_limit;
9457 static int sanitize_val_alu(struct bpf_verifier_env *env,
9458 struct bpf_insn *insn)
9460 struct bpf_insn_aux_data *aux = cur_aux(env);
9462 if (can_skip_alu_sanitation(env, insn))
9465 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
9468 static bool sanitize_needed(u8 opcode)
9470 return opcode == BPF_ADD || opcode == BPF_SUB;
9473 struct bpf_sanitize_info {
9474 struct bpf_insn_aux_data aux;
9478 static struct bpf_verifier_state *
9479 sanitize_speculative_path(struct bpf_verifier_env *env,
9480 const struct bpf_insn *insn,
9481 u32 next_idx, u32 curr_idx)
9483 struct bpf_verifier_state *branch;
9484 struct bpf_reg_state *regs;
9486 branch = push_stack(env, next_idx, curr_idx, true);
9487 if (branch && insn) {
9488 regs = branch->frame[branch->curframe]->regs;
9489 if (BPF_SRC(insn->code) == BPF_K) {
9490 mark_reg_unknown(env, regs, insn->dst_reg);
9491 } else if (BPF_SRC(insn->code) == BPF_X) {
9492 mark_reg_unknown(env, regs, insn->dst_reg);
9493 mark_reg_unknown(env, regs, insn->src_reg);
9499 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
9500 struct bpf_insn *insn,
9501 const struct bpf_reg_state *ptr_reg,
9502 const struct bpf_reg_state *off_reg,
9503 struct bpf_reg_state *dst_reg,
9504 struct bpf_sanitize_info *info,
9505 const bool commit_window)
9507 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
9508 struct bpf_verifier_state *vstate = env->cur_state;
9509 bool off_is_imm = tnum_is_const(off_reg->var_off);
9510 bool off_is_neg = off_reg->smin_value < 0;
9511 bool ptr_is_dst_reg = ptr_reg == dst_reg;
9512 u8 opcode = BPF_OP(insn->code);
9513 u32 alu_state, alu_limit;
9514 struct bpf_reg_state tmp;
9518 if (can_skip_alu_sanitation(env, insn))
9521 /* We already marked aux for masking from non-speculative
9522 * paths, thus we got here in the first place. We only care
9523 * to explore bad access from here.
9525 if (vstate->speculative)
9528 if (!commit_window) {
9529 if (!tnum_is_const(off_reg->var_off) &&
9530 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
9531 return REASON_BOUNDS;
9533 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
9534 (opcode == BPF_SUB && !off_is_neg);
9537 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
9541 if (commit_window) {
9542 /* In commit phase we narrow the masking window based on
9543 * the observed pointer move after the simulated operation.
9545 alu_state = info->aux.alu_state;
9546 alu_limit = abs(info->aux.alu_limit - alu_limit);
9548 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
9549 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
9550 alu_state |= ptr_is_dst_reg ?
9551 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
9553 /* Limit pruning on unknown scalars to enable deep search for
9554 * potential masking differences from other program paths.
9557 env->explore_alu_limits = true;
9560 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
9564 /* If we're in commit phase, we're done here given we already
9565 * pushed the truncated dst_reg into the speculative verification
9568 * Also, when register is a known constant, we rewrite register-based
9569 * operation to immediate-based, and thus do not need masking (and as
9570 * a consequence, do not need to simulate the zero-truncation either).
9572 if (commit_window || off_is_imm)
9575 /* Simulate and find potential out-of-bounds access under
9576 * speculative execution from truncation as a result of
9577 * masking when off was not within expected range. If off
9578 * sits in dst, then we temporarily need to move ptr there
9579 * to simulate dst (== 0) +/-= ptr. Needed, for example,
9580 * for cases where we use K-based arithmetic in one direction
9581 * and truncated reg-based in the other in order to explore
9584 if (!ptr_is_dst_reg) {
9586 *dst_reg = *ptr_reg;
9588 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
9590 if (!ptr_is_dst_reg && ret)
9592 return !ret ? REASON_STACK : 0;
9595 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
9597 struct bpf_verifier_state *vstate = env->cur_state;
9599 /* If we simulate paths under speculation, we don't update the
9600 * insn as 'seen' such that when we verify unreachable paths in
9601 * the non-speculative domain, sanitize_dead_code() can still
9602 * rewrite/sanitize them.
9604 if (!vstate->speculative)
9605 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
9608 static int sanitize_err(struct bpf_verifier_env *env,
9609 const struct bpf_insn *insn, int reason,
9610 const struct bpf_reg_state *off_reg,
9611 const struct bpf_reg_state *dst_reg)
9613 static const char *err = "pointer arithmetic with it prohibited for !root";
9614 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
9615 u32 dst = insn->dst_reg, src = insn->src_reg;
9619 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
9620 off_reg == dst_reg ? dst : src, err);
9623 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
9624 off_reg == dst_reg ? src : dst, err);
9627 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
9631 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
9635 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
9639 verbose(env, "verifier internal error: unknown reason (%d)\n",
9647 /* check that stack access falls within stack limits and that 'reg' doesn't
9648 * have a variable offset.
9650 * Variable offset is prohibited for unprivileged mode for simplicity since it
9651 * requires corresponding support in Spectre masking for stack ALU. See also
9652 * retrieve_ptr_limit().
9655 * 'off' includes 'reg->off'.
9657 static int check_stack_access_for_ptr_arithmetic(
9658 struct bpf_verifier_env *env,
9660 const struct bpf_reg_state *reg,
9663 if (!tnum_is_const(reg->var_off)) {
9666 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
9667 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
9668 regno, tn_buf, off);
9672 if (off >= 0 || off < -MAX_BPF_STACK) {
9673 verbose(env, "R%d stack pointer arithmetic goes out of range, "
9674 "prohibited for !root; off=%d\n", regno, off);
9681 static int sanitize_check_bounds(struct bpf_verifier_env *env,
9682 const struct bpf_insn *insn,
9683 const struct bpf_reg_state *dst_reg)
9685 u32 dst = insn->dst_reg;
9687 /* For unprivileged we require that resulting offset must be in bounds
9688 * in order to be able to sanitize access later on.
9690 if (env->bypass_spec_v1)
9693 switch (dst_reg->type) {
9695 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
9696 dst_reg->off + dst_reg->var_off.value))
9699 case PTR_TO_MAP_VALUE:
9700 if (check_map_access(env, dst, dst_reg->off, 1, false, ACCESS_HELPER)) {
9701 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
9702 "prohibited for !root\n", dst);
9713 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
9714 * Caller should also handle BPF_MOV case separately.
9715 * If we return -EACCES, caller may want to try again treating pointer as a
9716 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
9718 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
9719 struct bpf_insn *insn,
9720 const struct bpf_reg_state *ptr_reg,
9721 const struct bpf_reg_state *off_reg)
9723 struct bpf_verifier_state *vstate = env->cur_state;
9724 struct bpf_func_state *state = vstate->frame[vstate->curframe];
9725 struct bpf_reg_state *regs = state->regs, *dst_reg;
9726 bool known = tnum_is_const(off_reg->var_off);
9727 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
9728 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
9729 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
9730 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
9731 struct bpf_sanitize_info info = {};
9732 u8 opcode = BPF_OP(insn->code);
9733 u32 dst = insn->dst_reg;
9736 dst_reg = ®s[dst];
9738 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
9739 smin_val > smax_val || umin_val > umax_val) {
9740 /* Taint dst register if offset had invalid bounds derived from
9741 * e.g. dead branches.
9743 __mark_reg_unknown(env, dst_reg);
9747 if (BPF_CLASS(insn->code) != BPF_ALU64) {
9748 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
9749 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
9750 __mark_reg_unknown(env, dst_reg);
9755 "R%d 32-bit pointer arithmetic prohibited\n",
9760 if (ptr_reg->type & PTR_MAYBE_NULL) {
9761 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
9762 dst, reg_type_str(env, ptr_reg->type));
9766 switch (base_type(ptr_reg->type)) {
9767 case CONST_PTR_TO_MAP:
9768 /* smin_val represents the known value */
9769 if (known && smin_val == 0 && opcode == BPF_ADD)
9772 case PTR_TO_PACKET_END:
9774 case PTR_TO_SOCK_COMMON:
9775 case PTR_TO_TCP_SOCK:
9776 case PTR_TO_XDP_SOCK:
9777 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
9778 dst, reg_type_str(env, ptr_reg->type));
9784 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
9785 * The id may be overwritten later if we create a new variable offset.
9787 dst_reg->type = ptr_reg->type;
9788 dst_reg->id = ptr_reg->id;
9790 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
9791 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
9794 /* pointer types do not carry 32-bit bounds at the moment. */
9795 __mark_reg32_unbounded(dst_reg);
9797 if (sanitize_needed(opcode)) {
9798 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
9801 return sanitize_err(env, insn, ret, off_reg, dst_reg);
9806 /* We can take a fixed offset as long as it doesn't overflow
9807 * the s32 'off' field
9809 if (known && (ptr_reg->off + smin_val ==
9810 (s64)(s32)(ptr_reg->off + smin_val))) {
9811 /* pointer += K. Accumulate it into fixed offset */
9812 dst_reg->smin_value = smin_ptr;
9813 dst_reg->smax_value = smax_ptr;
9814 dst_reg->umin_value = umin_ptr;
9815 dst_reg->umax_value = umax_ptr;
9816 dst_reg->var_off = ptr_reg->var_off;
9817 dst_reg->off = ptr_reg->off + smin_val;
9818 dst_reg->raw = ptr_reg->raw;
9821 /* A new variable offset is created. Note that off_reg->off
9822 * == 0, since it's a scalar.
9823 * dst_reg gets the pointer type and since some positive
9824 * integer value was added to the pointer, give it a new 'id'
9825 * if it's a PTR_TO_PACKET.
9826 * this creates a new 'base' pointer, off_reg (variable) gets
9827 * added into the variable offset, and we copy the fixed offset
9830 if (signed_add_overflows(smin_ptr, smin_val) ||
9831 signed_add_overflows(smax_ptr, smax_val)) {
9832 dst_reg->smin_value = S64_MIN;
9833 dst_reg->smax_value = S64_MAX;
9835 dst_reg->smin_value = smin_ptr + smin_val;
9836 dst_reg->smax_value = smax_ptr + smax_val;
9838 if (umin_ptr + umin_val < umin_ptr ||
9839 umax_ptr + umax_val < umax_ptr) {
9840 dst_reg->umin_value = 0;
9841 dst_reg->umax_value = U64_MAX;
9843 dst_reg->umin_value = umin_ptr + umin_val;
9844 dst_reg->umax_value = umax_ptr + umax_val;
9846 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
9847 dst_reg->off = ptr_reg->off;
9848 dst_reg->raw = ptr_reg->raw;
9849 if (reg_is_pkt_pointer(ptr_reg)) {
9850 dst_reg->id = ++env->id_gen;
9851 /* something was added to pkt_ptr, set range to zero */
9852 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
9856 if (dst_reg == off_reg) {
9857 /* scalar -= pointer. Creates an unknown scalar */
9858 verbose(env, "R%d tried to subtract pointer from scalar\n",
9862 /* We don't allow subtraction from FP, because (according to
9863 * test_verifier.c test "invalid fp arithmetic", JITs might not
9864 * be able to deal with it.
9866 if (ptr_reg->type == PTR_TO_STACK) {
9867 verbose(env, "R%d subtraction from stack pointer prohibited\n",
9871 if (known && (ptr_reg->off - smin_val ==
9872 (s64)(s32)(ptr_reg->off - smin_val))) {
9873 /* pointer -= K. Subtract it from fixed offset */
9874 dst_reg->smin_value = smin_ptr;
9875 dst_reg->smax_value = smax_ptr;
9876 dst_reg->umin_value = umin_ptr;
9877 dst_reg->umax_value = umax_ptr;
9878 dst_reg->var_off = ptr_reg->var_off;
9879 dst_reg->id = ptr_reg->id;
9880 dst_reg->off = ptr_reg->off - smin_val;
9881 dst_reg->raw = ptr_reg->raw;
9884 /* A new variable offset is created. If the subtrahend is known
9885 * nonnegative, then any reg->range we had before is still good.
9887 if (signed_sub_overflows(smin_ptr, smax_val) ||
9888 signed_sub_overflows(smax_ptr, smin_val)) {
9889 /* Overflow possible, we know nothing */
9890 dst_reg->smin_value = S64_MIN;
9891 dst_reg->smax_value = S64_MAX;
9893 dst_reg->smin_value = smin_ptr - smax_val;
9894 dst_reg->smax_value = smax_ptr - smin_val;
9896 if (umin_ptr < umax_val) {
9897 /* Overflow possible, we know nothing */
9898 dst_reg->umin_value = 0;
9899 dst_reg->umax_value = U64_MAX;
9901 /* Cannot overflow (as long as bounds are consistent) */
9902 dst_reg->umin_value = umin_ptr - umax_val;
9903 dst_reg->umax_value = umax_ptr - umin_val;
9905 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
9906 dst_reg->off = ptr_reg->off;
9907 dst_reg->raw = ptr_reg->raw;
9908 if (reg_is_pkt_pointer(ptr_reg)) {
9909 dst_reg->id = ++env->id_gen;
9910 /* something was added to pkt_ptr, set range to zero */
9912 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
9918 /* bitwise ops on pointers are troublesome, prohibit. */
9919 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
9920 dst, bpf_alu_string[opcode >> 4]);
9923 /* other operators (e.g. MUL,LSH) produce non-pointer results */
9924 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
9925 dst, bpf_alu_string[opcode >> 4]);
9929 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
9931 reg_bounds_sync(dst_reg);
9932 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
9934 if (sanitize_needed(opcode)) {
9935 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
9938 return sanitize_err(env, insn, ret, off_reg, dst_reg);
9944 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
9945 struct bpf_reg_state *src_reg)
9947 s32 smin_val = src_reg->s32_min_value;
9948 s32 smax_val = src_reg->s32_max_value;
9949 u32 umin_val = src_reg->u32_min_value;
9950 u32 umax_val = src_reg->u32_max_value;
9952 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
9953 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
9954 dst_reg->s32_min_value = S32_MIN;
9955 dst_reg->s32_max_value = S32_MAX;
9957 dst_reg->s32_min_value += smin_val;
9958 dst_reg->s32_max_value += smax_val;
9960 if (dst_reg->u32_min_value + umin_val < umin_val ||
9961 dst_reg->u32_max_value + umax_val < umax_val) {
9962 dst_reg->u32_min_value = 0;
9963 dst_reg->u32_max_value = U32_MAX;
9965 dst_reg->u32_min_value += umin_val;
9966 dst_reg->u32_max_value += umax_val;
9970 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
9971 struct bpf_reg_state *src_reg)
9973 s64 smin_val = src_reg->smin_value;
9974 s64 smax_val = src_reg->smax_value;
9975 u64 umin_val = src_reg->umin_value;
9976 u64 umax_val = src_reg->umax_value;
9978 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
9979 signed_add_overflows(dst_reg->smax_value, smax_val)) {
9980 dst_reg->smin_value = S64_MIN;
9981 dst_reg->smax_value = S64_MAX;
9983 dst_reg->smin_value += smin_val;
9984 dst_reg->smax_value += smax_val;
9986 if (dst_reg->umin_value + umin_val < umin_val ||
9987 dst_reg->umax_value + umax_val < umax_val) {
9988 dst_reg->umin_value = 0;
9989 dst_reg->umax_value = U64_MAX;
9991 dst_reg->umin_value += umin_val;
9992 dst_reg->umax_value += umax_val;
9996 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
9997 struct bpf_reg_state *src_reg)
9999 s32 smin_val = src_reg->s32_min_value;
10000 s32 smax_val = src_reg->s32_max_value;
10001 u32 umin_val = src_reg->u32_min_value;
10002 u32 umax_val = src_reg->u32_max_value;
10004 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
10005 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
10006 /* Overflow possible, we know nothing */
10007 dst_reg->s32_min_value = S32_MIN;
10008 dst_reg->s32_max_value = S32_MAX;
10010 dst_reg->s32_min_value -= smax_val;
10011 dst_reg->s32_max_value -= smin_val;
10013 if (dst_reg->u32_min_value < umax_val) {
10014 /* Overflow possible, we know nothing */
10015 dst_reg->u32_min_value = 0;
10016 dst_reg->u32_max_value = U32_MAX;
10018 /* Cannot overflow (as long as bounds are consistent) */
10019 dst_reg->u32_min_value -= umax_val;
10020 dst_reg->u32_max_value -= umin_val;
10024 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
10025 struct bpf_reg_state *src_reg)
10027 s64 smin_val = src_reg->smin_value;
10028 s64 smax_val = src_reg->smax_value;
10029 u64 umin_val = src_reg->umin_value;
10030 u64 umax_val = src_reg->umax_value;
10032 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
10033 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
10034 /* Overflow possible, we know nothing */
10035 dst_reg->smin_value = S64_MIN;
10036 dst_reg->smax_value = S64_MAX;
10038 dst_reg->smin_value -= smax_val;
10039 dst_reg->smax_value -= smin_val;
10041 if (dst_reg->umin_value < umax_val) {
10042 /* Overflow possible, we know nothing */
10043 dst_reg->umin_value = 0;
10044 dst_reg->umax_value = U64_MAX;
10046 /* Cannot overflow (as long as bounds are consistent) */
10047 dst_reg->umin_value -= umax_val;
10048 dst_reg->umax_value -= umin_val;
10052 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
10053 struct bpf_reg_state *src_reg)
10055 s32 smin_val = src_reg->s32_min_value;
10056 u32 umin_val = src_reg->u32_min_value;
10057 u32 umax_val = src_reg->u32_max_value;
10059 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
10060 /* Ain't nobody got time to multiply that sign */
10061 __mark_reg32_unbounded(dst_reg);
10064 /* Both values are positive, so we can work with unsigned and
10065 * copy the result to signed (unless it exceeds S32_MAX).
10067 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
10068 /* Potential overflow, we know nothing */
10069 __mark_reg32_unbounded(dst_reg);
10072 dst_reg->u32_min_value *= umin_val;
10073 dst_reg->u32_max_value *= umax_val;
10074 if (dst_reg->u32_max_value > S32_MAX) {
10075 /* Overflow possible, we know nothing */
10076 dst_reg->s32_min_value = S32_MIN;
10077 dst_reg->s32_max_value = S32_MAX;
10079 dst_reg->s32_min_value = dst_reg->u32_min_value;
10080 dst_reg->s32_max_value = dst_reg->u32_max_value;
10084 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
10085 struct bpf_reg_state *src_reg)
10087 s64 smin_val = src_reg->smin_value;
10088 u64 umin_val = src_reg->umin_value;
10089 u64 umax_val = src_reg->umax_value;
10091 if (smin_val < 0 || dst_reg->smin_value < 0) {
10092 /* Ain't nobody got time to multiply that sign */
10093 __mark_reg64_unbounded(dst_reg);
10096 /* Both values are positive, so we can work with unsigned and
10097 * copy the result to signed (unless it exceeds S64_MAX).
10099 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
10100 /* Potential overflow, we know nothing */
10101 __mark_reg64_unbounded(dst_reg);
10104 dst_reg->umin_value *= umin_val;
10105 dst_reg->umax_value *= umax_val;
10106 if (dst_reg->umax_value > S64_MAX) {
10107 /* Overflow possible, we know nothing */
10108 dst_reg->smin_value = S64_MIN;
10109 dst_reg->smax_value = S64_MAX;
10111 dst_reg->smin_value = dst_reg->umin_value;
10112 dst_reg->smax_value = dst_reg->umax_value;
10116 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
10117 struct bpf_reg_state *src_reg)
10119 bool src_known = tnum_subreg_is_const(src_reg->var_off);
10120 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
10121 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
10122 s32 smin_val = src_reg->s32_min_value;
10123 u32 umax_val = src_reg->u32_max_value;
10125 if (src_known && dst_known) {
10126 __mark_reg32_known(dst_reg, var32_off.value);
10130 /* We get our minimum from the var_off, since that's inherently
10131 * bitwise. Our maximum is the minimum of the operands' maxima.
10133 dst_reg->u32_min_value = var32_off.value;
10134 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
10135 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
10136 /* Lose signed bounds when ANDing negative numbers,
10137 * ain't nobody got time for that.
10139 dst_reg->s32_min_value = S32_MIN;
10140 dst_reg->s32_max_value = S32_MAX;
10142 /* ANDing two positives gives a positive, so safe to
10143 * cast result into s64.
10145 dst_reg->s32_min_value = dst_reg->u32_min_value;
10146 dst_reg->s32_max_value = dst_reg->u32_max_value;
10150 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
10151 struct bpf_reg_state *src_reg)
10153 bool src_known = tnum_is_const(src_reg->var_off);
10154 bool dst_known = tnum_is_const(dst_reg->var_off);
10155 s64 smin_val = src_reg->smin_value;
10156 u64 umax_val = src_reg->umax_value;
10158 if (src_known && dst_known) {
10159 __mark_reg_known(dst_reg, dst_reg->var_off.value);
10163 /* We get our minimum from the var_off, since that's inherently
10164 * bitwise. Our maximum is the minimum of the operands' maxima.
10166 dst_reg->umin_value = dst_reg->var_off.value;
10167 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
10168 if (dst_reg->smin_value < 0 || smin_val < 0) {
10169 /* Lose signed bounds when ANDing negative numbers,
10170 * ain't nobody got time for that.
10172 dst_reg->smin_value = S64_MIN;
10173 dst_reg->smax_value = S64_MAX;
10175 /* ANDing two positives gives a positive, so safe to
10176 * cast result into s64.
10178 dst_reg->smin_value = dst_reg->umin_value;
10179 dst_reg->smax_value = dst_reg->umax_value;
10181 /* We may learn something more from the var_off */
10182 __update_reg_bounds(dst_reg);
10185 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
10186 struct bpf_reg_state *src_reg)
10188 bool src_known = tnum_subreg_is_const(src_reg->var_off);
10189 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
10190 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
10191 s32 smin_val = src_reg->s32_min_value;
10192 u32 umin_val = src_reg->u32_min_value;
10194 if (src_known && dst_known) {
10195 __mark_reg32_known(dst_reg, var32_off.value);
10199 /* We get our maximum from the var_off, and our minimum is the
10200 * maximum of the operands' minima
10202 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
10203 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
10204 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
10205 /* Lose signed bounds when ORing negative numbers,
10206 * ain't nobody got time for that.
10208 dst_reg->s32_min_value = S32_MIN;
10209 dst_reg->s32_max_value = S32_MAX;
10211 /* ORing two positives gives a positive, so safe to
10212 * cast result into s64.
10214 dst_reg->s32_min_value = dst_reg->u32_min_value;
10215 dst_reg->s32_max_value = dst_reg->u32_max_value;
10219 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
10220 struct bpf_reg_state *src_reg)
10222 bool src_known = tnum_is_const(src_reg->var_off);
10223 bool dst_known = tnum_is_const(dst_reg->var_off);
10224 s64 smin_val = src_reg->smin_value;
10225 u64 umin_val = src_reg->umin_value;
10227 if (src_known && dst_known) {
10228 __mark_reg_known(dst_reg, dst_reg->var_off.value);
10232 /* We get our maximum from the var_off, and our minimum is the
10233 * maximum of the operands' minima
10235 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
10236 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
10237 if (dst_reg->smin_value < 0 || smin_val < 0) {
10238 /* Lose signed bounds when ORing negative numbers,
10239 * ain't nobody got time for that.
10241 dst_reg->smin_value = S64_MIN;
10242 dst_reg->smax_value = S64_MAX;
10244 /* ORing two positives gives a positive, so safe to
10245 * cast result into s64.
10247 dst_reg->smin_value = dst_reg->umin_value;
10248 dst_reg->smax_value = dst_reg->umax_value;
10250 /* We may learn something more from the var_off */
10251 __update_reg_bounds(dst_reg);
10254 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
10255 struct bpf_reg_state *src_reg)
10257 bool src_known = tnum_subreg_is_const(src_reg->var_off);
10258 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
10259 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
10260 s32 smin_val = src_reg->s32_min_value;
10262 if (src_known && dst_known) {
10263 __mark_reg32_known(dst_reg, var32_off.value);
10267 /* We get both minimum and maximum from the var32_off. */
10268 dst_reg->u32_min_value = var32_off.value;
10269 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
10271 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
10272 /* XORing two positive sign numbers gives a positive,
10273 * so safe to cast u32 result into s32.
10275 dst_reg->s32_min_value = dst_reg->u32_min_value;
10276 dst_reg->s32_max_value = dst_reg->u32_max_value;
10278 dst_reg->s32_min_value = S32_MIN;
10279 dst_reg->s32_max_value = S32_MAX;
10283 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
10284 struct bpf_reg_state *src_reg)
10286 bool src_known = tnum_is_const(src_reg->var_off);
10287 bool dst_known = tnum_is_const(dst_reg->var_off);
10288 s64 smin_val = src_reg->smin_value;
10290 if (src_known && dst_known) {
10291 /* dst_reg->var_off.value has been updated earlier */
10292 __mark_reg_known(dst_reg, dst_reg->var_off.value);
10296 /* We get both minimum and maximum from the var_off. */
10297 dst_reg->umin_value = dst_reg->var_off.value;
10298 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
10300 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
10301 /* XORing two positive sign numbers gives a positive,
10302 * so safe to cast u64 result into s64.
10304 dst_reg->smin_value = dst_reg->umin_value;
10305 dst_reg->smax_value = dst_reg->umax_value;
10307 dst_reg->smin_value = S64_MIN;
10308 dst_reg->smax_value = S64_MAX;
10311 __update_reg_bounds(dst_reg);
10314 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
10315 u64 umin_val, u64 umax_val)
10317 /* We lose all sign bit information (except what we can pick
10320 dst_reg->s32_min_value = S32_MIN;
10321 dst_reg->s32_max_value = S32_MAX;
10322 /* If we might shift our top bit out, then we know nothing */
10323 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
10324 dst_reg->u32_min_value = 0;
10325 dst_reg->u32_max_value = U32_MAX;
10327 dst_reg->u32_min_value <<= umin_val;
10328 dst_reg->u32_max_value <<= umax_val;
10332 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
10333 struct bpf_reg_state *src_reg)
10335 u32 umax_val = src_reg->u32_max_value;
10336 u32 umin_val = src_reg->u32_min_value;
10337 /* u32 alu operation will zext upper bits */
10338 struct tnum subreg = tnum_subreg(dst_reg->var_off);
10340 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
10341 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
10342 /* Not required but being careful mark reg64 bounds as unknown so
10343 * that we are forced to pick them up from tnum and zext later and
10344 * if some path skips this step we are still safe.
10346 __mark_reg64_unbounded(dst_reg);
10347 __update_reg32_bounds(dst_reg);
10350 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
10351 u64 umin_val, u64 umax_val)
10353 /* Special case <<32 because it is a common compiler pattern to sign
10354 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
10355 * positive we know this shift will also be positive so we can track
10356 * bounds correctly. Otherwise we lose all sign bit information except
10357 * what we can pick up from var_off. Perhaps we can generalize this
10358 * later to shifts of any length.
10360 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
10361 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
10363 dst_reg->smax_value = S64_MAX;
10365 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
10366 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
10368 dst_reg->smin_value = S64_MIN;
10370 /* If we might shift our top bit out, then we know nothing */
10371 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
10372 dst_reg->umin_value = 0;
10373 dst_reg->umax_value = U64_MAX;
10375 dst_reg->umin_value <<= umin_val;
10376 dst_reg->umax_value <<= umax_val;
10380 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
10381 struct bpf_reg_state *src_reg)
10383 u64 umax_val = src_reg->umax_value;
10384 u64 umin_val = src_reg->umin_value;
10386 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
10387 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
10388 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
10390 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
10391 /* We may learn something more from the var_off */
10392 __update_reg_bounds(dst_reg);
10395 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
10396 struct bpf_reg_state *src_reg)
10398 struct tnum subreg = tnum_subreg(dst_reg->var_off);
10399 u32 umax_val = src_reg->u32_max_value;
10400 u32 umin_val = src_reg->u32_min_value;
10402 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
10403 * be negative, then either:
10404 * 1) src_reg might be zero, so the sign bit of the result is
10405 * unknown, so we lose our signed bounds
10406 * 2) it's known negative, thus the unsigned bounds capture the
10408 * 3) the signed bounds cross zero, so they tell us nothing
10410 * If the value in dst_reg is known nonnegative, then again the
10411 * unsigned bounds capture the signed bounds.
10412 * Thus, in all cases it suffices to blow away our signed bounds
10413 * and rely on inferring new ones from the unsigned bounds and
10414 * var_off of the result.
10416 dst_reg->s32_min_value = S32_MIN;
10417 dst_reg->s32_max_value = S32_MAX;
10419 dst_reg->var_off = tnum_rshift(subreg, umin_val);
10420 dst_reg->u32_min_value >>= umax_val;
10421 dst_reg->u32_max_value >>= umin_val;
10423 __mark_reg64_unbounded(dst_reg);
10424 __update_reg32_bounds(dst_reg);
10427 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
10428 struct bpf_reg_state *src_reg)
10430 u64 umax_val = src_reg->umax_value;
10431 u64 umin_val = src_reg->umin_value;
10433 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
10434 * be negative, then either:
10435 * 1) src_reg might be zero, so the sign bit of the result is
10436 * unknown, so we lose our signed bounds
10437 * 2) it's known negative, thus the unsigned bounds capture the
10439 * 3) the signed bounds cross zero, so they tell us nothing
10441 * If the value in dst_reg is known nonnegative, then again the
10442 * unsigned bounds capture the signed bounds.
10443 * Thus, in all cases it suffices to blow away our signed bounds
10444 * and rely on inferring new ones from the unsigned bounds and
10445 * var_off of the result.
10447 dst_reg->smin_value = S64_MIN;
10448 dst_reg->smax_value = S64_MAX;
10449 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
10450 dst_reg->umin_value >>= umax_val;
10451 dst_reg->umax_value >>= umin_val;
10453 /* Its not easy to operate on alu32 bounds here because it depends
10454 * on bits being shifted in. Take easy way out and mark unbounded
10455 * so we can recalculate later from tnum.
10457 __mark_reg32_unbounded(dst_reg);
10458 __update_reg_bounds(dst_reg);
10461 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
10462 struct bpf_reg_state *src_reg)
10464 u64 umin_val = src_reg->u32_min_value;
10466 /* Upon reaching here, src_known is true and
10467 * umax_val is equal to umin_val.
10469 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
10470 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
10472 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
10474 /* blow away the dst_reg umin_value/umax_value and rely on
10475 * dst_reg var_off to refine the result.
10477 dst_reg->u32_min_value = 0;
10478 dst_reg->u32_max_value = U32_MAX;
10480 __mark_reg64_unbounded(dst_reg);
10481 __update_reg32_bounds(dst_reg);
10484 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
10485 struct bpf_reg_state *src_reg)
10487 u64 umin_val = src_reg->umin_value;
10489 /* Upon reaching here, src_known is true and umax_val is equal
10492 dst_reg->smin_value >>= umin_val;
10493 dst_reg->smax_value >>= umin_val;
10495 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
10497 /* blow away the dst_reg umin_value/umax_value and rely on
10498 * dst_reg var_off to refine the result.
10500 dst_reg->umin_value = 0;
10501 dst_reg->umax_value = U64_MAX;
10503 /* Its not easy to operate on alu32 bounds here because it depends
10504 * on bits being shifted in from upper 32-bits. Take easy way out
10505 * and mark unbounded so we can recalculate later from tnum.
10507 __mark_reg32_unbounded(dst_reg);
10508 __update_reg_bounds(dst_reg);
10511 /* WARNING: This function does calculations on 64-bit values, but the actual
10512 * execution may occur on 32-bit values. Therefore, things like bitshifts
10513 * need extra checks in the 32-bit case.
10515 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
10516 struct bpf_insn *insn,
10517 struct bpf_reg_state *dst_reg,
10518 struct bpf_reg_state src_reg)
10520 struct bpf_reg_state *regs = cur_regs(env);
10521 u8 opcode = BPF_OP(insn->code);
10523 s64 smin_val, smax_val;
10524 u64 umin_val, umax_val;
10525 s32 s32_min_val, s32_max_val;
10526 u32 u32_min_val, u32_max_val;
10527 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
10528 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
10531 smin_val = src_reg.smin_value;
10532 smax_val = src_reg.smax_value;
10533 umin_val = src_reg.umin_value;
10534 umax_val = src_reg.umax_value;
10536 s32_min_val = src_reg.s32_min_value;
10537 s32_max_val = src_reg.s32_max_value;
10538 u32_min_val = src_reg.u32_min_value;
10539 u32_max_val = src_reg.u32_max_value;
10542 src_known = tnum_subreg_is_const(src_reg.var_off);
10544 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
10545 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
10546 /* Taint dst register if offset had invalid bounds
10547 * derived from e.g. dead branches.
10549 __mark_reg_unknown(env, dst_reg);
10553 src_known = tnum_is_const(src_reg.var_off);
10555 (smin_val != smax_val || umin_val != umax_val)) ||
10556 smin_val > smax_val || umin_val > umax_val) {
10557 /* Taint dst register if offset had invalid bounds
10558 * derived from e.g. dead branches.
10560 __mark_reg_unknown(env, dst_reg);
10566 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
10567 __mark_reg_unknown(env, dst_reg);
10571 if (sanitize_needed(opcode)) {
10572 ret = sanitize_val_alu(env, insn);
10574 return sanitize_err(env, insn, ret, NULL, NULL);
10577 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
10578 * There are two classes of instructions: The first class we track both
10579 * alu32 and alu64 sign/unsigned bounds independently this provides the
10580 * greatest amount of precision when alu operations are mixed with jmp32
10581 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
10582 * and BPF_OR. This is possible because these ops have fairly easy to
10583 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
10584 * See alu32 verifier tests for examples. The second class of
10585 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
10586 * with regards to tracking sign/unsigned bounds because the bits may
10587 * cross subreg boundaries in the alu64 case. When this happens we mark
10588 * the reg unbounded in the subreg bound space and use the resulting
10589 * tnum to calculate an approximation of the sign/unsigned bounds.
10593 scalar32_min_max_add(dst_reg, &src_reg);
10594 scalar_min_max_add(dst_reg, &src_reg);
10595 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
10598 scalar32_min_max_sub(dst_reg, &src_reg);
10599 scalar_min_max_sub(dst_reg, &src_reg);
10600 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
10603 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
10604 scalar32_min_max_mul(dst_reg, &src_reg);
10605 scalar_min_max_mul(dst_reg, &src_reg);
10608 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
10609 scalar32_min_max_and(dst_reg, &src_reg);
10610 scalar_min_max_and(dst_reg, &src_reg);
10613 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
10614 scalar32_min_max_or(dst_reg, &src_reg);
10615 scalar_min_max_or(dst_reg, &src_reg);
10618 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
10619 scalar32_min_max_xor(dst_reg, &src_reg);
10620 scalar_min_max_xor(dst_reg, &src_reg);
10623 if (umax_val >= insn_bitness) {
10624 /* Shifts greater than 31 or 63 are undefined.
10625 * This includes shifts by a negative number.
10627 mark_reg_unknown(env, regs, insn->dst_reg);
10631 scalar32_min_max_lsh(dst_reg, &src_reg);
10633 scalar_min_max_lsh(dst_reg, &src_reg);
10636 if (umax_val >= insn_bitness) {
10637 /* Shifts greater than 31 or 63 are undefined.
10638 * This includes shifts by a negative number.
10640 mark_reg_unknown(env, regs, insn->dst_reg);
10644 scalar32_min_max_rsh(dst_reg, &src_reg);
10646 scalar_min_max_rsh(dst_reg, &src_reg);
10649 if (umax_val >= insn_bitness) {
10650 /* Shifts greater than 31 or 63 are undefined.
10651 * This includes shifts by a negative number.
10653 mark_reg_unknown(env, regs, insn->dst_reg);
10657 scalar32_min_max_arsh(dst_reg, &src_reg);
10659 scalar_min_max_arsh(dst_reg, &src_reg);
10662 mark_reg_unknown(env, regs, insn->dst_reg);
10666 /* ALU32 ops are zero extended into 64bit register */
10668 zext_32_to_64(dst_reg);
10669 reg_bounds_sync(dst_reg);
10673 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
10676 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
10677 struct bpf_insn *insn)
10679 struct bpf_verifier_state *vstate = env->cur_state;
10680 struct bpf_func_state *state = vstate->frame[vstate->curframe];
10681 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
10682 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
10683 u8 opcode = BPF_OP(insn->code);
10686 dst_reg = ®s[insn->dst_reg];
10688 if (dst_reg->type != SCALAR_VALUE)
10691 /* Make sure ID is cleared otherwise dst_reg min/max could be
10692 * incorrectly propagated into other registers by find_equal_scalars()
10695 if (BPF_SRC(insn->code) == BPF_X) {
10696 src_reg = ®s[insn->src_reg];
10697 if (src_reg->type != SCALAR_VALUE) {
10698 if (dst_reg->type != SCALAR_VALUE) {
10699 /* Combining two pointers by any ALU op yields
10700 * an arbitrary scalar. Disallow all math except
10701 * pointer subtraction
10703 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
10704 mark_reg_unknown(env, regs, insn->dst_reg);
10707 verbose(env, "R%d pointer %s pointer prohibited\n",
10709 bpf_alu_string[opcode >> 4]);
10712 /* scalar += pointer
10713 * This is legal, but we have to reverse our
10714 * src/dest handling in computing the range
10716 err = mark_chain_precision(env, insn->dst_reg);
10719 return adjust_ptr_min_max_vals(env, insn,
10722 } else if (ptr_reg) {
10723 /* pointer += scalar */
10724 err = mark_chain_precision(env, insn->src_reg);
10727 return adjust_ptr_min_max_vals(env, insn,
10729 } else if (dst_reg->precise) {
10730 /* if dst_reg is precise, src_reg should be precise as well */
10731 err = mark_chain_precision(env, insn->src_reg);
10736 /* Pretend the src is a reg with a known value, since we only
10737 * need to be able to read from this state.
10739 off_reg.type = SCALAR_VALUE;
10740 __mark_reg_known(&off_reg, insn->imm);
10741 src_reg = &off_reg;
10742 if (ptr_reg) /* pointer += K */
10743 return adjust_ptr_min_max_vals(env, insn,
10747 /* Got here implies adding two SCALAR_VALUEs */
10748 if (WARN_ON_ONCE(ptr_reg)) {
10749 print_verifier_state(env, state, true);
10750 verbose(env, "verifier internal error: unexpected ptr_reg\n");
10753 if (WARN_ON(!src_reg)) {
10754 print_verifier_state(env, state, true);
10755 verbose(env, "verifier internal error: no src_reg\n");
10758 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
10761 /* check validity of 32-bit and 64-bit arithmetic operations */
10762 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
10764 struct bpf_reg_state *regs = cur_regs(env);
10765 u8 opcode = BPF_OP(insn->code);
10768 if (opcode == BPF_END || opcode == BPF_NEG) {
10769 if (opcode == BPF_NEG) {
10770 if (BPF_SRC(insn->code) != BPF_K ||
10771 insn->src_reg != BPF_REG_0 ||
10772 insn->off != 0 || insn->imm != 0) {
10773 verbose(env, "BPF_NEG uses reserved fields\n");
10777 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
10778 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
10779 BPF_CLASS(insn->code) == BPF_ALU64) {
10780 verbose(env, "BPF_END uses reserved fields\n");
10785 /* check src operand */
10786 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
10790 if (is_pointer_value(env, insn->dst_reg)) {
10791 verbose(env, "R%d pointer arithmetic prohibited\n",
10796 /* check dest operand */
10797 err = check_reg_arg(env, insn->dst_reg, DST_OP);
10801 } else if (opcode == BPF_MOV) {
10803 if (BPF_SRC(insn->code) == BPF_X) {
10804 if (insn->imm != 0 || insn->off != 0) {
10805 verbose(env, "BPF_MOV uses reserved fields\n");
10809 /* check src operand */
10810 err = check_reg_arg(env, insn->src_reg, SRC_OP);
10814 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
10815 verbose(env, "BPF_MOV uses reserved fields\n");
10820 /* check dest operand, mark as required later */
10821 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
10825 if (BPF_SRC(insn->code) == BPF_X) {
10826 struct bpf_reg_state *src_reg = regs + insn->src_reg;
10827 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
10829 if (BPF_CLASS(insn->code) == BPF_ALU64) {
10831 * copy register state to dest reg
10833 if (src_reg->type == SCALAR_VALUE && !src_reg->id)
10834 /* Assign src and dst registers the same ID
10835 * that will be used by find_equal_scalars()
10836 * to propagate min/max range.
10838 src_reg->id = ++env->id_gen;
10839 *dst_reg = *src_reg;
10840 dst_reg->live |= REG_LIVE_WRITTEN;
10841 dst_reg->subreg_def = DEF_NOT_SUBREG;
10843 /* R1 = (u32) R2 */
10844 if (is_pointer_value(env, insn->src_reg)) {
10846 "R%d partial copy of pointer\n",
10849 } else if (src_reg->type == SCALAR_VALUE) {
10850 *dst_reg = *src_reg;
10851 /* Make sure ID is cleared otherwise
10852 * dst_reg min/max could be incorrectly
10853 * propagated into src_reg by find_equal_scalars()
10856 dst_reg->live |= REG_LIVE_WRITTEN;
10857 dst_reg->subreg_def = env->insn_idx + 1;
10859 mark_reg_unknown(env, regs,
10862 zext_32_to_64(dst_reg);
10863 reg_bounds_sync(dst_reg);
10867 * remember the value we stored into this reg
10869 /* clear any state __mark_reg_known doesn't set */
10870 mark_reg_unknown(env, regs, insn->dst_reg);
10871 regs[insn->dst_reg].type = SCALAR_VALUE;
10872 if (BPF_CLASS(insn->code) == BPF_ALU64) {
10873 __mark_reg_known(regs + insn->dst_reg,
10876 __mark_reg_known(regs + insn->dst_reg,
10881 } else if (opcode > BPF_END) {
10882 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
10885 } else { /* all other ALU ops: and, sub, xor, add, ... */
10887 if (BPF_SRC(insn->code) == BPF_X) {
10888 if (insn->imm != 0 || insn->off != 0) {
10889 verbose(env, "BPF_ALU uses reserved fields\n");
10892 /* check src1 operand */
10893 err = check_reg_arg(env, insn->src_reg, SRC_OP);
10897 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
10898 verbose(env, "BPF_ALU uses reserved fields\n");
10903 /* check src2 operand */
10904 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
10908 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
10909 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
10910 verbose(env, "div by zero\n");
10914 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
10915 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
10916 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
10918 if (insn->imm < 0 || insn->imm >= size) {
10919 verbose(env, "invalid shift %d\n", insn->imm);
10924 /* check dest operand */
10925 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
10929 return adjust_reg_min_max_vals(env, insn);
10935 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
10936 struct bpf_reg_state *dst_reg,
10937 enum bpf_reg_type type,
10938 bool range_right_open)
10940 struct bpf_func_state *state;
10941 struct bpf_reg_state *reg;
10944 if (dst_reg->off < 0 ||
10945 (dst_reg->off == 0 && range_right_open))
10946 /* This doesn't give us any range */
10949 if (dst_reg->umax_value > MAX_PACKET_OFF ||
10950 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
10951 /* Risk of overflow. For instance, ptr + (1<<63) may be less
10952 * than pkt_end, but that's because it's also less than pkt.
10956 new_range = dst_reg->off;
10957 if (range_right_open)
10960 /* Examples for register markings:
10962 * pkt_data in dst register:
10966 * if (r2 > pkt_end) goto <handle exception>
10971 * if (r2 < pkt_end) goto <access okay>
10972 * <handle exception>
10975 * r2 == dst_reg, pkt_end == src_reg
10976 * r2=pkt(id=n,off=8,r=0)
10977 * r3=pkt(id=n,off=0,r=0)
10979 * pkt_data in src register:
10983 * if (pkt_end >= r2) goto <access okay>
10984 * <handle exception>
10988 * if (pkt_end <= r2) goto <handle exception>
10992 * pkt_end == dst_reg, r2 == src_reg
10993 * r2=pkt(id=n,off=8,r=0)
10994 * r3=pkt(id=n,off=0,r=0)
10996 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
10997 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
10998 * and [r3, r3 + 8-1) respectively is safe to access depending on
11002 /* If our ids match, then we must have the same max_value. And we
11003 * don't care about the other reg's fixed offset, since if it's too big
11004 * the range won't allow anything.
11005 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
11007 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
11008 if (reg->type == type && reg->id == dst_reg->id)
11009 /* keep the maximum range already checked */
11010 reg->range = max(reg->range, new_range);
11014 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
11016 struct tnum subreg = tnum_subreg(reg->var_off);
11017 s32 sval = (s32)val;
11021 if (tnum_is_const(subreg))
11022 return !!tnum_equals_const(subreg, val);
11025 if (tnum_is_const(subreg))
11026 return !tnum_equals_const(subreg, val);
11029 if ((~subreg.mask & subreg.value) & val)
11031 if (!((subreg.mask | subreg.value) & val))
11035 if (reg->u32_min_value > val)
11037 else if (reg->u32_max_value <= val)
11041 if (reg->s32_min_value > sval)
11043 else if (reg->s32_max_value <= sval)
11047 if (reg->u32_max_value < val)
11049 else if (reg->u32_min_value >= val)
11053 if (reg->s32_max_value < sval)
11055 else if (reg->s32_min_value >= sval)
11059 if (reg->u32_min_value >= val)
11061 else if (reg->u32_max_value < val)
11065 if (reg->s32_min_value >= sval)
11067 else if (reg->s32_max_value < sval)
11071 if (reg->u32_max_value <= val)
11073 else if (reg->u32_min_value > val)
11077 if (reg->s32_max_value <= sval)
11079 else if (reg->s32_min_value > sval)
11088 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
11090 s64 sval = (s64)val;
11094 if (tnum_is_const(reg->var_off))
11095 return !!tnum_equals_const(reg->var_off, val);
11098 if (tnum_is_const(reg->var_off))
11099 return !tnum_equals_const(reg->var_off, val);
11102 if ((~reg->var_off.mask & reg->var_off.value) & val)
11104 if (!((reg->var_off.mask | reg->var_off.value) & val))
11108 if (reg->umin_value > val)
11110 else if (reg->umax_value <= val)
11114 if (reg->smin_value > sval)
11116 else if (reg->smax_value <= sval)
11120 if (reg->umax_value < val)
11122 else if (reg->umin_value >= val)
11126 if (reg->smax_value < sval)
11128 else if (reg->smin_value >= sval)
11132 if (reg->umin_value >= val)
11134 else if (reg->umax_value < val)
11138 if (reg->smin_value >= sval)
11140 else if (reg->smax_value < sval)
11144 if (reg->umax_value <= val)
11146 else if (reg->umin_value > val)
11150 if (reg->smax_value <= sval)
11152 else if (reg->smin_value > sval)
11160 /* compute branch direction of the expression "if (reg opcode val) goto target;"
11162 * 1 - branch will be taken and "goto target" will be executed
11163 * 0 - branch will not be taken and fall-through to next insn
11164 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
11167 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
11170 if (__is_pointer_value(false, reg)) {
11171 if (!reg_type_not_null(reg->type))
11174 /* If pointer is valid tests against zero will fail so we can
11175 * use this to direct branch taken.
11191 return is_branch32_taken(reg, val, opcode);
11192 return is_branch64_taken(reg, val, opcode);
11195 static int flip_opcode(u32 opcode)
11197 /* How can we transform "a <op> b" into "b <op> a"? */
11198 static const u8 opcode_flip[16] = {
11199 /* these stay the same */
11200 [BPF_JEQ >> 4] = BPF_JEQ,
11201 [BPF_JNE >> 4] = BPF_JNE,
11202 [BPF_JSET >> 4] = BPF_JSET,
11203 /* these swap "lesser" and "greater" (L and G in the opcodes) */
11204 [BPF_JGE >> 4] = BPF_JLE,
11205 [BPF_JGT >> 4] = BPF_JLT,
11206 [BPF_JLE >> 4] = BPF_JGE,
11207 [BPF_JLT >> 4] = BPF_JGT,
11208 [BPF_JSGE >> 4] = BPF_JSLE,
11209 [BPF_JSGT >> 4] = BPF_JSLT,
11210 [BPF_JSLE >> 4] = BPF_JSGE,
11211 [BPF_JSLT >> 4] = BPF_JSGT
11213 return opcode_flip[opcode >> 4];
11216 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
11217 struct bpf_reg_state *src_reg,
11220 struct bpf_reg_state *pkt;
11222 if (src_reg->type == PTR_TO_PACKET_END) {
11224 } else if (dst_reg->type == PTR_TO_PACKET_END) {
11226 opcode = flip_opcode(opcode);
11231 if (pkt->range >= 0)
11236 /* pkt <= pkt_end */
11239 /* pkt > pkt_end */
11240 if (pkt->range == BEYOND_PKT_END)
11241 /* pkt has at last one extra byte beyond pkt_end */
11242 return opcode == BPF_JGT;
11245 /* pkt < pkt_end */
11248 /* pkt >= pkt_end */
11249 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
11250 return opcode == BPF_JGE;
11256 /* Adjusts the register min/max values in the case that the dst_reg is the
11257 * variable register that we are working on, and src_reg is a constant or we're
11258 * simply doing a BPF_K check.
11259 * In JEQ/JNE cases we also adjust the var_off values.
11261 static void reg_set_min_max(struct bpf_reg_state *true_reg,
11262 struct bpf_reg_state *false_reg,
11263 u64 val, u32 val32,
11264 u8 opcode, bool is_jmp32)
11266 struct tnum false_32off = tnum_subreg(false_reg->var_off);
11267 struct tnum false_64off = false_reg->var_off;
11268 struct tnum true_32off = tnum_subreg(true_reg->var_off);
11269 struct tnum true_64off = true_reg->var_off;
11270 s64 sval = (s64)val;
11271 s32 sval32 = (s32)val32;
11273 /* If the dst_reg is a pointer, we can't learn anything about its
11274 * variable offset from the compare (unless src_reg were a pointer into
11275 * the same object, but we don't bother with that.
11276 * Since false_reg and true_reg have the same type by construction, we
11277 * only need to check one of them for pointerness.
11279 if (__is_pointer_value(false, false_reg))
11283 /* JEQ/JNE comparison doesn't change the register equivalence.
11286 * if (r1 == 42) goto label;
11288 * label: // here both r1 and r2 are known to be 42.
11290 * Hence when marking register as known preserve it's ID.
11294 __mark_reg32_known(true_reg, val32);
11295 true_32off = tnum_subreg(true_reg->var_off);
11297 ___mark_reg_known(true_reg, val);
11298 true_64off = true_reg->var_off;
11303 __mark_reg32_known(false_reg, val32);
11304 false_32off = tnum_subreg(false_reg->var_off);
11306 ___mark_reg_known(false_reg, val);
11307 false_64off = false_reg->var_off;
11312 false_32off = tnum_and(false_32off, tnum_const(~val32));
11313 if (is_power_of_2(val32))
11314 true_32off = tnum_or(true_32off,
11315 tnum_const(val32));
11317 false_64off = tnum_and(false_64off, tnum_const(~val));
11318 if (is_power_of_2(val))
11319 true_64off = tnum_or(true_64off,
11327 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
11328 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
11330 false_reg->u32_max_value = min(false_reg->u32_max_value,
11332 true_reg->u32_min_value = max(true_reg->u32_min_value,
11335 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
11336 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
11338 false_reg->umax_value = min(false_reg->umax_value, false_umax);
11339 true_reg->umin_value = max(true_reg->umin_value, true_umin);
11347 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
11348 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
11350 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
11351 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
11353 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
11354 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
11356 false_reg->smax_value = min(false_reg->smax_value, false_smax);
11357 true_reg->smin_value = max(true_reg->smin_value, true_smin);
11365 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
11366 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
11368 false_reg->u32_min_value = max(false_reg->u32_min_value,
11370 true_reg->u32_max_value = min(true_reg->u32_max_value,
11373 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
11374 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
11376 false_reg->umin_value = max(false_reg->umin_value, false_umin);
11377 true_reg->umax_value = min(true_reg->umax_value, true_umax);
11385 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
11386 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
11388 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
11389 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
11391 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
11392 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
11394 false_reg->smin_value = max(false_reg->smin_value, false_smin);
11395 true_reg->smax_value = min(true_reg->smax_value, true_smax);
11404 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
11405 tnum_subreg(false_32off));
11406 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
11407 tnum_subreg(true_32off));
11408 __reg_combine_32_into_64(false_reg);
11409 __reg_combine_32_into_64(true_reg);
11411 false_reg->var_off = false_64off;
11412 true_reg->var_off = true_64off;
11413 __reg_combine_64_into_32(false_reg);
11414 __reg_combine_64_into_32(true_reg);
11418 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
11419 * the variable reg.
11421 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
11422 struct bpf_reg_state *false_reg,
11423 u64 val, u32 val32,
11424 u8 opcode, bool is_jmp32)
11426 opcode = flip_opcode(opcode);
11427 /* This uses zero as "not present in table"; luckily the zero opcode,
11428 * BPF_JA, can't get here.
11431 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
11434 /* Regs are known to be equal, so intersect their min/max/var_off */
11435 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
11436 struct bpf_reg_state *dst_reg)
11438 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
11439 dst_reg->umin_value);
11440 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
11441 dst_reg->umax_value);
11442 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
11443 dst_reg->smin_value);
11444 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
11445 dst_reg->smax_value);
11446 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
11448 reg_bounds_sync(src_reg);
11449 reg_bounds_sync(dst_reg);
11452 static void reg_combine_min_max(struct bpf_reg_state *true_src,
11453 struct bpf_reg_state *true_dst,
11454 struct bpf_reg_state *false_src,
11455 struct bpf_reg_state *false_dst,
11460 __reg_combine_min_max(true_src, true_dst);
11463 __reg_combine_min_max(false_src, false_dst);
11468 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
11469 struct bpf_reg_state *reg, u32 id,
11472 if (type_may_be_null(reg->type) && reg->id == id &&
11473 (is_rcu_reg(reg) || !WARN_ON_ONCE(!reg->id))) {
11474 /* Old offset (both fixed and variable parts) should have been
11475 * known-zero, because we don't allow pointer arithmetic on
11476 * pointers that might be NULL. If we see this happening, don't
11477 * convert the register.
11479 * But in some cases, some helpers that return local kptrs
11480 * advance offset for the returned pointer. In those cases, it
11481 * is fine to expect to see reg->off.
11483 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value || !tnum_equals_const(reg->var_off, 0)))
11485 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC | PTR_MAYBE_NULL) && WARN_ON_ONCE(reg->off))
11488 reg->type = SCALAR_VALUE;
11489 /* We don't need id and ref_obj_id from this point
11490 * onwards anymore, thus we should better reset it,
11491 * so that state pruning has chances to take effect.
11494 reg->ref_obj_id = 0;
11499 mark_ptr_not_null_reg(reg);
11501 if (!reg_may_point_to_spin_lock(reg)) {
11502 /* For not-NULL ptr, reg->ref_obj_id will be reset
11503 * in release_reference().
11505 * reg->id is still used by spin_lock ptr. Other
11506 * than spin_lock ptr type, reg->id can be reset.
11513 /* The logic is similar to find_good_pkt_pointers(), both could eventually
11514 * be folded together at some point.
11516 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
11519 struct bpf_func_state *state = vstate->frame[vstate->curframe];
11520 struct bpf_reg_state *regs = state->regs, *reg;
11521 u32 ref_obj_id = regs[regno].ref_obj_id;
11522 u32 id = regs[regno].id;
11524 if (ref_obj_id && ref_obj_id == id && is_null)
11525 /* regs[regno] is in the " == NULL" branch.
11526 * No one could have freed the reference state before
11527 * doing the NULL check.
11529 WARN_ON_ONCE(release_reference_state(state, id));
11531 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
11532 mark_ptr_or_null_reg(state, reg, id, is_null);
11536 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
11537 struct bpf_reg_state *dst_reg,
11538 struct bpf_reg_state *src_reg,
11539 struct bpf_verifier_state *this_branch,
11540 struct bpf_verifier_state *other_branch)
11542 if (BPF_SRC(insn->code) != BPF_X)
11545 /* Pointers are always 64-bit. */
11546 if (BPF_CLASS(insn->code) == BPF_JMP32)
11549 switch (BPF_OP(insn->code)) {
11551 if ((dst_reg->type == PTR_TO_PACKET &&
11552 src_reg->type == PTR_TO_PACKET_END) ||
11553 (dst_reg->type == PTR_TO_PACKET_META &&
11554 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
11555 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
11556 find_good_pkt_pointers(this_branch, dst_reg,
11557 dst_reg->type, false);
11558 mark_pkt_end(other_branch, insn->dst_reg, true);
11559 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
11560 src_reg->type == PTR_TO_PACKET) ||
11561 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
11562 src_reg->type == PTR_TO_PACKET_META)) {
11563 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
11564 find_good_pkt_pointers(other_branch, src_reg,
11565 src_reg->type, true);
11566 mark_pkt_end(this_branch, insn->src_reg, false);
11572 if ((dst_reg->type == PTR_TO_PACKET &&
11573 src_reg->type == PTR_TO_PACKET_END) ||
11574 (dst_reg->type == PTR_TO_PACKET_META &&
11575 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
11576 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
11577 find_good_pkt_pointers(other_branch, dst_reg,
11578 dst_reg->type, true);
11579 mark_pkt_end(this_branch, insn->dst_reg, false);
11580 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
11581 src_reg->type == PTR_TO_PACKET) ||
11582 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
11583 src_reg->type == PTR_TO_PACKET_META)) {
11584 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
11585 find_good_pkt_pointers(this_branch, src_reg,
11586 src_reg->type, false);
11587 mark_pkt_end(other_branch, insn->src_reg, true);
11593 if ((dst_reg->type == PTR_TO_PACKET &&
11594 src_reg->type == PTR_TO_PACKET_END) ||
11595 (dst_reg->type == PTR_TO_PACKET_META &&
11596 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
11597 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
11598 find_good_pkt_pointers(this_branch, dst_reg,
11599 dst_reg->type, true);
11600 mark_pkt_end(other_branch, insn->dst_reg, false);
11601 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
11602 src_reg->type == PTR_TO_PACKET) ||
11603 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
11604 src_reg->type == PTR_TO_PACKET_META)) {
11605 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
11606 find_good_pkt_pointers(other_branch, src_reg,
11607 src_reg->type, false);
11608 mark_pkt_end(this_branch, insn->src_reg, true);
11614 if ((dst_reg->type == PTR_TO_PACKET &&
11615 src_reg->type == PTR_TO_PACKET_END) ||
11616 (dst_reg->type == PTR_TO_PACKET_META &&
11617 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
11618 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
11619 find_good_pkt_pointers(other_branch, dst_reg,
11620 dst_reg->type, false);
11621 mark_pkt_end(this_branch, insn->dst_reg, true);
11622 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
11623 src_reg->type == PTR_TO_PACKET) ||
11624 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
11625 src_reg->type == PTR_TO_PACKET_META)) {
11626 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
11627 find_good_pkt_pointers(this_branch, src_reg,
11628 src_reg->type, true);
11629 mark_pkt_end(other_branch, insn->src_reg, false);
11641 static void find_equal_scalars(struct bpf_verifier_state *vstate,
11642 struct bpf_reg_state *known_reg)
11644 struct bpf_func_state *state;
11645 struct bpf_reg_state *reg;
11647 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
11648 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
11653 static int check_cond_jmp_op(struct bpf_verifier_env *env,
11654 struct bpf_insn *insn, int *insn_idx)
11656 struct bpf_verifier_state *this_branch = env->cur_state;
11657 struct bpf_verifier_state *other_branch;
11658 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
11659 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
11660 struct bpf_reg_state *eq_branch_regs;
11661 u8 opcode = BPF_OP(insn->code);
11666 /* Only conditional jumps are expected to reach here. */
11667 if (opcode == BPF_JA || opcode > BPF_JSLE) {
11668 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
11672 if (BPF_SRC(insn->code) == BPF_X) {
11673 if (insn->imm != 0) {
11674 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
11678 /* check src1 operand */
11679 err = check_reg_arg(env, insn->src_reg, SRC_OP);
11683 if (is_pointer_value(env, insn->src_reg)) {
11684 verbose(env, "R%d pointer comparison prohibited\n",
11688 src_reg = ®s[insn->src_reg];
11690 if (insn->src_reg != BPF_REG_0) {
11691 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
11696 /* check src2 operand */
11697 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
11701 dst_reg = ®s[insn->dst_reg];
11702 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
11704 if (BPF_SRC(insn->code) == BPF_K) {
11705 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
11706 } else if (src_reg->type == SCALAR_VALUE &&
11707 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
11708 pred = is_branch_taken(dst_reg,
11709 tnum_subreg(src_reg->var_off).value,
11712 } else if (src_reg->type == SCALAR_VALUE &&
11713 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
11714 pred = is_branch_taken(dst_reg,
11715 src_reg->var_off.value,
11718 } else if (reg_is_pkt_pointer_any(dst_reg) &&
11719 reg_is_pkt_pointer_any(src_reg) &&
11721 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
11725 /* If we get here with a dst_reg pointer type it is because
11726 * above is_branch_taken() special cased the 0 comparison.
11728 if (!__is_pointer_value(false, dst_reg))
11729 err = mark_chain_precision(env, insn->dst_reg);
11730 if (BPF_SRC(insn->code) == BPF_X && !err &&
11731 !__is_pointer_value(false, src_reg))
11732 err = mark_chain_precision(env, insn->src_reg);
11738 /* Only follow the goto, ignore fall-through. If needed, push
11739 * the fall-through branch for simulation under speculative
11742 if (!env->bypass_spec_v1 &&
11743 !sanitize_speculative_path(env, insn, *insn_idx + 1,
11746 *insn_idx += insn->off;
11748 } else if (pred == 0) {
11749 /* Only follow the fall-through branch, since that's where the
11750 * program will go. If needed, push the goto branch for
11751 * simulation under speculative execution.
11753 if (!env->bypass_spec_v1 &&
11754 !sanitize_speculative_path(env, insn,
11755 *insn_idx + insn->off + 1,
11761 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
11765 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
11767 /* detect if we are comparing against a constant value so we can adjust
11768 * our min/max values for our dst register.
11769 * this is only legit if both are scalars (or pointers to the same
11770 * object, I suppose, see the PTR_MAYBE_NULL related if block below),
11771 * because otherwise the different base pointers mean the offsets aren't
11774 if (BPF_SRC(insn->code) == BPF_X) {
11775 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
11777 if (dst_reg->type == SCALAR_VALUE &&
11778 src_reg->type == SCALAR_VALUE) {
11779 if (tnum_is_const(src_reg->var_off) ||
11781 tnum_is_const(tnum_subreg(src_reg->var_off))))
11782 reg_set_min_max(&other_branch_regs[insn->dst_reg],
11784 src_reg->var_off.value,
11785 tnum_subreg(src_reg->var_off).value,
11787 else if (tnum_is_const(dst_reg->var_off) ||
11789 tnum_is_const(tnum_subreg(dst_reg->var_off))))
11790 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
11792 dst_reg->var_off.value,
11793 tnum_subreg(dst_reg->var_off).value,
11795 else if (!is_jmp32 &&
11796 (opcode == BPF_JEQ || opcode == BPF_JNE))
11797 /* Comparing for equality, we can combine knowledge */
11798 reg_combine_min_max(&other_branch_regs[insn->src_reg],
11799 &other_branch_regs[insn->dst_reg],
11800 src_reg, dst_reg, opcode);
11802 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
11803 find_equal_scalars(this_branch, src_reg);
11804 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
11808 } else if (dst_reg->type == SCALAR_VALUE) {
11809 reg_set_min_max(&other_branch_regs[insn->dst_reg],
11810 dst_reg, insn->imm, (u32)insn->imm,
11814 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
11815 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
11816 find_equal_scalars(this_branch, dst_reg);
11817 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
11820 /* if one pointer register is compared to another pointer
11821 * register check if PTR_MAYBE_NULL could be lifted.
11822 * E.g. register A - maybe null
11823 * register B - not null
11824 * for JNE A, B, ... - A is not null in the false branch;
11825 * for JEQ A, B, ... - A is not null in the true branch.
11827 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_X &&
11828 __is_pointer_value(false, src_reg) && __is_pointer_value(false, dst_reg) &&
11829 type_may_be_null(src_reg->type) != type_may_be_null(dst_reg->type)) {
11830 eq_branch_regs = NULL;
11833 eq_branch_regs = other_branch_regs;
11836 eq_branch_regs = regs;
11842 if (eq_branch_regs) {
11843 if (type_may_be_null(src_reg->type))
11844 mark_ptr_not_null_reg(&eq_branch_regs[insn->src_reg]);
11846 mark_ptr_not_null_reg(&eq_branch_regs[insn->dst_reg]);
11850 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
11851 * NOTE: these optimizations below are related with pointer comparison
11852 * which will never be JMP32.
11854 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
11855 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
11856 type_may_be_null(dst_reg->type)) {
11857 /* Mark all identical registers in each branch as either
11858 * safe or unknown depending R == 0 or R != 0 conditional.
11860 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
11861 opcode == BPF_JNE);
11862 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
11863 opcode == BPF_JEQ);
11864 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
11865 this_branch, other_branch) &&
11866 is_pointer_value(env, insn->dst_reg)) {
11867 verbose(env, "R%d pointer comparison prohibited\n",
11871 if (env->log.level & BPF_LOG_LEVEL)
11872 print_insn_state(env, this_branch->frame[this_branch->curframe]);
11876 /* verify BPF_LD_IMM64 instruction */
11877 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
11879 struct bpf_insn_aux_data *aux = cur_aux(env);
11880 struct bpf_reg_state *regs = cur_regs(env);
11881 struct bpf_reg_state *dst_reg;
11882 struct bpf_map *map;
11885 if (BPF_SIZE(insn->code) != BPF_DW) {
11886 verbose(env, "invalid BPF_LD_IMM insn\n");
11889 if (insn->off != 0) {
11890 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
11894 err = check_reg_arg(env, insn->dst_reg, DST_OP);
11898 dst_reg = ®s[insn->dst_reg];
11899 if (insn->src_reg == 0) {
11900 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
11902 dst_reg->type = SCALAR_VALUE;
11903 __mark_reg_known(®s[insn->dst_reg], imm);
11907 /* All special src_reg cases are listed below. From this point onwards
11908 * we either succeed and assign a corresponding dst_reg->type after
11909 * zeroing the offset, or fail and reject the program.
11911 mark_reg_known_zero(env, regs, insn->dst_reg);
11913 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
11914 dst_reg->type = aux->btf_var.reg_type;
11915 switch (base_type(dst_reg->type)) {
11917 dst_reg->mem_size = aux->btf_var.mem_size;
11919 case PTR_TO_BTF_ID:
11920 dst_reg->btf = aux->btf_var.btf;
11921 dst_reg->btf_id = aux->btf_var.btf_id;
11924 verbose(env, "bpf verifier is misconfigured\n");
11930 if (insn->src_reg == BPF_PSEUDO_FUNC) {
11931 struct bpf_prog_aux *aux = env->prog->aux;
11932 u32 subprogno = find_subprog(env,
11933 env->insn_idx + insn->imm + 1);
11935 if (!aux->func_info) {
11936 verbose(env, "missing btf func_info\n");
11939 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
11940 verbose(env, "callback function not static\n");
11944 dst_reg->type = PTR_TO_FUNC;
11945 dst_reg->subprogno = subprogno;
11949 map = env->used_maps[aux->map_index];
11950 dst_reg->map_ptr = map;
11952 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
11953 insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
11954 dst_reg->type = PTR_TO_MAP_VALUE;
11955 dst_reg->off = aux->map_off;
11956 WARN_ON_ONCE(map->max_entries != 1);
11957 /* We want reg->id to be same (0) as map_value is not distinct */
11958 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
11959 insn->src_reg == BPF_PSEUDO_MAP_IDX) {
11960 dst_reg->type = CONST_PTR_TO_MAP;
11962 verbose(env, "bpf verifier is misconfigured\n");
11969 static bool may_access_skb(enum bpf_prog_type type)
11972 case BPF_PROG_TYPE_SOCKET_FILTER:
11973 case BPF_PROG_TYPE_SCHED_CLS:
11974 case BPF_PROG_TYPE_SCHED_ACT:
11981 /* verify safety of LD_ABS|LD_IND instructions:
11982 * - they can only appear in the programs where ctx == skb
11983 * - since they are wrappers of function calls, they scratch R1-R5 registers,
11984 * preserve R6-R9, and store return value into R0
11987 * ctx == skb == R6 == CTX
11990 * SRC == any register
11991 * IMM == 32-bit immediate
11994 * R0 - 8/16/32-bit skb data converted to cpu endianness
11996 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
11998 struct bpf_reg_state *regs = cur_regs(env);
11999 static const int ctx_reg = BPF_REG_6;
12000 u8 mode = BPF_MODE(insn->code);
12003 if (!may_access_skb(resolve_prog_type(env->prog))) {
12004 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
12008 if (!env->ops->gen_ld_abs) {
12009 verbose(env, "bpf verifier is misconfigured\n");
12013 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
12014 BPF_SIZE(insn->code) == BPF_DW ||
12015 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
12016 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
12020 /* check whether implicit source operand (register R6) is readable */
12021 err = check_reg_arg(env, ctx_reg, SRC_OP);
12025 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
12026 * gen_ld_abs() may terminate the program at runtime, leading to
12029 err = check_reference_leak(env);
12031 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
12035 if (env->cur_state->active_lock.ptr) {
12036 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
12040 if (env->cur_state->active_rcu_lock) {
12041 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_rcu_read_lock-ed region\n");
12045 if (regs[ctx_reg].type != PTR_TO_CTX) {
12047 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
12051 if (mode == BPF_IND) {
12052 /* check explicit source operand */
12053 err = check_reg_arg(env, insn->src_reg, SRC_OP);
12058 err = check_ptr_off_reg(env, ®s[ctx_reg], ctx_reg);
12062 /* reset caller saved regs to unreadable */
12063 for (i = 0; i < CALLER_SAVED_REGS; i++) {
12064 mark_reg_not_init(env, regs, caller_saved[i]);
12065 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
12068 /* mark destination R0 register as readable, since it contains
12069 * the value fetched from the packet.
12070 * Already marked as written above.
12072 mark_reg_unknown(env, regs, BPF_REG_0);
12073 /* ld_abs load up to 32-bit skb data. */
12074 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
12078 static int check_return_code(struct bpf_verifier_env *env)
12080 struct tnum enforce_attach_type_range = tnum_unknown;
12081 const struct bpf_prog *prog = env->prog;
12082 struct bpf_reg_state *reg;
12083 struct tnum range = tnum_range(0, 1);
12084 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
12086 struct bpf_func_state *frame = env->cur_state->frame[0];
12087 const bool is_subprog = frame->subprogno;
12089 /* LSM and struct_ops func-ptr's return type could be "void" */
12091 switch (prog_type) {
12092 case BPF_PROG_TYPE_LSM:
12093 if (prog->expected_attach_type == BPF_LSM_CGROUP)
12094 /* See below, can be 0 or 0-1 depending on hook. */
12097 case BPF_PROG_TYPE_STRUCT_OPS:
12098 if (!prog->aux->attach_func_proto->type)
12106 /* eBPF calling convention is such that R0 is used
12107 * to return the value from eBPF program.
12108 * Make sure that it's readable at this time
12109 * of bpf_exit, which means that program wrote
12110 * something into it earlier
12112 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
12116 if (is_pointer_value(env, BPF_REG_0)) {
12117 verbose(env, "R0 leaks addr as return value\n");
12121 reg = cur_regs(env) + BPF_REG_0;
12123 if (frame->in_async_callback_fn) {
12124 /* enforce return zero from async callbacks like timer */
12125 if (reg->type != SCALAR_VALUE) {
12126 verbose(env, "In async callback the register R0 is not a known value (%s)\n",
12127 reg_type_str(env, reg->type));
12131 if (!tnum_in(tnum_const(0), reg->var_off)) {
12132 verbose_invalid_scalar(env, reg, &range, "async callback", "R0");
12139 if (reg->type != SCALAR_VALUE) {
12140 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
12141 reg_type_str(env, reg->type));
12147 switch (prog_type) {
12148 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
12149 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
12150 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
12151 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
12152 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
12153 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
12154 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
12155 range = tnum_range(1, 1);
12156 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
12157 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
12158 range = tnum_range(0, 3);
12160 case BPF_PROG_TYPE_CGROUP_SKB:
12161 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
12162 range = tnum_range(0, 3);
12163 enforce_attach_type_range = tnum_range(2, 3);
12166 case BPF_PROG_TYPE_CGROUP_SOCK:
12167 case BPF_PROG_TYPE_SOCK_OPS:
12168 case BPF_PROG_TYPE_CGROUP_DEVICE:
12169 case BPF_PROG_TYPE_CGROUP_SYSCTL:
12170 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
12172 case BPF_PROG_TYPE_RAW_TRACEPOINT:
12173 if (!env->prog->aux->attach_btf_id)
12175 range = tnum_const(0);
12177 case BPF_PROG_TYPE_TRACING:
12178 switch (env->prog->expected_attach_type) {
12179 case BPF_TRACE_FENTRY:
12180 case BPF_TRACE_FEXIT:
12181 range = tnum_const(0);
12183 case BPF_TRACE_RAW_TP:
12184 case BPF_MODIFY_RETURN:
12186 case BPF_TRACE_ITER:
12192 case BPF_PROG_TYPE_SK_LOOKUP:
12193 range = tnum_range(SK_DROP, SK_PASS);
12196 case BPF_PROG_TYPE_LSM:
12197 if (env->prog->expected_attach_type != BPF_LSM_CGROUP) {
12198 /* Regular BPF_PROG_TYPE_LSM programs can return
12203 if (!env->prog->aux->attach_func_proto->type) {
12204 /* Make sure programs that attach to void
12205 * hooks don't try to modify return value.
12207 range = tnum_range(1, 1);
12211 case BPF_PROG_TYPE_EXT:
12212 /* freplace program can return anything as its return value
12213 * depends on the to-be-replaced kernel func or bpf program.
12219 if (reg->type != SCALAR_VALUE) {
12220 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
12221 reg_type_str(env, reg->type));
12225 if (!tnum_in(range, reg->var_off)) {
12226 verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
12227 if (prog->expected_attach_type == BPF_LSM_CGROUP &&
12228 prog_type == BPF_PROG_TYPE_LSM &&
12229 !prog->aux->attach_func_proto->type)
12230 verbose(env, "Note, BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
12234 if (!tnum_is_unknown(enforce_attach_type_range) &&
12235 tnum_in(enforce_attach_type_range, reg->var_off))
12236 env->prog->enforce_expected_attach_type = 1;
12240 /* non-recursive DFS pseudo code
12241 * 1 procedure DFS-iterative(G,v):
12242 * 2 label v as discovered
12243 * 3 let S be a stack
12245 * 5 while S is not empty
12247 * 7 if t is what we're looking for:
12249 * 9 for all edges e in G.adjacentEdges(t) do
12250 * 10 if edge e is already labelled
12251 * 11 continue with the next edge
12252 * 12 w <- G.adjacentVertex(t,e)
12253 * 13 if vertex w is not discovered and not explored
12254 * 14 label e as tree-edge
12255 * 15 label w as discovered
12258 * 18 else if vertex w is discovered
12259 * 19 label e as back-edge
12261 * 21 // vertex w is explored
12262 * 22 label e as forward- or cross-edge
12263 * 23 label t as explored
12267 * 0x10 - discovered
12268 * 0x11 - discovered and fall-through edge labelled
12269 * 0x12 - discovered and fall-through and branch edges labelled
12280 static u32 state_htab_size(struct bpf_verifier_env *env)
12282 return env->prog->len;
12285 static struct bpf_verifier_state_list **explored_state(
12286 struct bpf_verifier_env *env,
12289 struct bpf_verifier_state *cur = env->cur_state;
12290 struct bpf_func_state *state = cur->frame[cur->curframe];
12292 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
12295 static void mark_prune_point(struct bpf_verifier_env *env, int idx)
12297 env->insn_aux_data[idx].prune_point = true;
12300 static bool is_prune_point(struct bpf_verifier_env *env, int insn_idx)
12302 return env->insn_aux_data[insn_idx].prune_point;
12306 DONE_EXPLORING = 0,
12307 KEEP_EXPLORING = 1,
12310 /* t, w, e - match pseudo-code above:
12311 * t - index of current instruction
12312 * w - next instruction
12315 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
12318 int *insn_stack = env->cfg.insn_stack;
12319 int *insn_state = env->cfg.insn_state;
12321 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
12322 return DONE_EXPLORING;
12324 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
12325 return DONE_EXPLORING;
12327 if (w < 0 || w >= env->prog->len) {
12328 verbose_linfo(env, t, "%d: ", t);
12329 verbose(env, "jump out of range from insn %d to %d\n", t, w);
12334 /* mark branch target for state pruning */
12335 mark_prune_point(env, w);
12336 mark_jmp_point(env, w);
12339 if (insn_state[w] == 0) {
12341 insn_state[t] = DISCOVERED | e;
12342 insn_state[w] = DISCOVERED;
12343 if (env->cfg.cur_stack >= env->prog->len)
12345 insn_stack[env->cfg.cur_stack++] = w;
12346 return KEEP_EXPLORING;
12347 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
12348 if (loop_ok && env->bpf_capable)
12349 return DONE_EXPLORING;
12350 verbose_linfo(env, t, "%d: ", t);
12351 verbose_linfo(env, w, "%d: ", w);
12352 verbose(env, "back-edge from insn %d to %d\n", t, w);
12354 } else if (insn_state[w] == EXPLORED) {
12355 /* forward- or cross-edge */
12356 insn_state[t] = DISCOVERED | e;
12358 verbose(env, "insn state internal bug\n");
12361 return DONE_EXPLORING;
12364 static int visit_func_call_insn(int t, struct bpf_insn *insns,
12365 struct bpf_verifier_env *env,
12370 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
12374 mark_prune_point(env, t + 1);
12375 /* when we exit from subprog, we need to record non-linear history */
12376 mark_jmp_point(env, t + 1);
12378 if (visit_callee) {
12379 mark_prune_point(env, t);
12380 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env,
12381 /* It's ok to allow recursion from CFG point of
12382 * view. __check_func_call() will do the actual
12385 bpf_pseudo_func(insns + t));
12390 /* Visits the instruction at index t and returns one of the following:
12391 * < 0 - an error occurred
12392 * DONE_EXPLORING - the instruction was fully explored
12393 * KEEP_EXPLORING - there is still work to be done before it is fully explored
12395 static int visit_insn(int t, struct bpf_verifier_env *env)
12397 struct bpf_insn *insns = env->prog->insnsi;
12400 if (bpf_pseudo_func(insns + t))
12401 return visit_func_call_insn(t, insns, env, true);
12403 /* All non-branch instructions have a single fall-through edge. */
12404 if (BPF_CLASS(insns[t].code) != BPF_JMP &&
12405 BPF_CLASS(insns[t].code) != BPF_JMP32)
12406 return push_insn(t, t + 1, FALLTHROUGH, env, false);
12408 switch (BPF_OP(insns[t].code)) {
12410 return DONE_EXPLORING;
12413 if (insns[t].imm == BPF_FUNC_timer_set_callback)
12414 /* Mark this call insn as a prune point to trigger
12415 * is_state_visited() check before call itself is
12416 * processed by __check_func_call(). Otherwise new
12417 * async state will be pushed for further exploration.
12419 mark_prune_point(env, t);
12420 return visit_func_call_insn(t, insns, env,
12421 insns[t].src_reg == BPF_PSEUDO_CALL);
12424 if (BPF_SRC(insns[t].code) != BPF_K)
12427 /* unconditional jump with single edge */
12428 ret = push_insn(t, t + insns[t].off + 1, FALLTHROUGH, env,
12433 mark_prune_point(env, t + insns[t].off + 1);
12434 mark_jmp_point(env, t + insns[t].off + 1);
12439 /* conditional jump with two edges */
12440 mark_prune_point(env, t);
12442 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
12446 return push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
12450 /* non-recursive depth-first-search to detect loops in BPF program
12451 * loop == back-edge in directed graph
12453 static int check_cfg(struct bpf_verifier_env *env)
12455 int insn_cnt = env->prog->len;
12456 int *insn_stack, *insn_state;
12460 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
12464 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
12466 kvfree(insn_state);
12470 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
12471 insn_stack[0] = 0; /* 0 is the first instruction */
12472 env->cfg.cur_stack = 1;
12474 while (env->cfg.cur_stack > 0) {
12475 int t = insn_stack[env->cfg.cur_stack - 1];
12477 ret = visit_insn(t, env);
12479 case DONE_EXPLORING:
12480 insn_state[t] = EXPLORED;
12481 env->cfg.cur_stack--;
12483 case KEEP_EXPLORING:
12487 verbose(env, "visit_insn internal bug\n");
12494 if (env->cfg.cur_stack < 0) {
12495 verbose(env, "pop stack internal bug\n");
12500 for (i = 0; i < insn_cnt; i++) {
12501 if (insn_state[i] != EXPLORED) {
12502 verbose(env, "unreachable insn %d\n", i);
12507 ret = 0; /* cfg looks good */
12510 kvfree(insn_state);
12511 kvfree(insn_stack);
12512 env->cfg.insn_state = env->cfg.insn_stack = NULL;
12516 static int check_abnormal_return(struct bpf_verifier_env *env)
12520 for (i = 1; i < env->subprog_cnt; i++) {
12521 if (env->subprog_info[i].has_ld_abs) {
12522 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
12525 if (env->subprog_info[i].has_tail_call) {
12526 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
12533 /* The minimum supported BTF func info size */
12534 #define MIN_BPF_FUNCINFO_SIZE 8
12535 #define MAX_FUNCINFO_REC_SIZE 252
12537 static int check_btf_func(struct bpf_verifier_env *env,
12538 const union bpf_attr *attr,
12541 const struct btf_type *type, *func_proto, *ret_type;
12542 u32 i, nfuncs, urec_size, min_size;
12543 u32 krec_size = sizeof(struct bpf_func_info);
12544 struct bpf_func_info *krecord;
12545 struct bpf_func_info_aux *info_aux = NULL;
12546 struct bpf_prog *prog;
12547 const struct btf *btf;
12549 u32 prev_offset = 0;
12550 bool scalar_return;
12553 nfuncs = attr->func_info_cnt;
12555 if (check_abnormal_return(env))
12560 if (nfuncs != env->subprog_cnt) {
12561 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
12565 urec_size = attr->func_info_rec_size;
12566 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
12567 urec_size > MAX_FUNCINFO_REC_SIZE ||
12568 urec_size % sizeof(u32)) {
12569 verbose(env, "invalid func info rec size %u\n", urec_size);
12574 btf = prog->aux->btf;
12576 urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
12577 min_size = min_t(u32, krec_size, urec_size);
12579 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
12582 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
12586 for (i = 0; i < nfuncs; i++) {
12587 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
12589 if (ret == -E2BIG) {
12590 verbose(env, "nonzero tailing record in func info");
12591 /* set the size kernel expects so loader can zero
12592 * out the rest of the record.
12594 if (copy_to_bpfptr_offset(uattr,
12595 offsetof(union bpf_attr, func_info_rec_size),
12596 &min_size, sizeof(min_size)))
12602 if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
12607 /* check insn_off */
12610 if (krecord[i].insn_off) {
12612 "nonzero insn_off %u for the first func info record",
12613 krecord[i].insn_off);
12616 } else if (krecord[i].insn_off <= prev_offset) {
12618 "same or smaller insn offset (%u) than previous func info record (%u)",
12619 krecord[i].insn_off, prev_offset);
12623 if (env->subprog_info[i].start != krecord[i].insn_off) {
12624 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
12628 /* check type_id */
12629 type = btf_type_by_id(btf, krecord[i].type_id);
12630 if (!type || !btf_type_is_func(type)) {
12631 verbose(env, "invalid type id %d in func info",
12632 krecord[i].type_id);
12635 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
12637 func_proto = btf_type_by_id(btf, type->type);
12638 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
12639 /* btf_func_check() already verified it during BTF load */
12641 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
12643 btf_type_is_small_int(ret_type) || btf_is_any_enum(ret_type);
12644 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
12645 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
12648 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
12649 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
12653 prev_offset = krecord[i].insn_off;
12654 bpfptr_add(&urecord, urec_size);
12657 prog->aux->func_info = krecord;
12658 prog->aux->func_info_cnt = nfuncs;
12659 prog->aux->func_info_aux = info_aux;
12668 static void adjust_btf_func(struct bpf_verifier_env *env)
12670 struct bpf_prog_aux *aux = env->prog->aux;
12673 if (!aux->func_info)
12676 for (i = 0; i < env->subprog_cnt; i++)
12677 aux->func_info[i].insn_off = env->subprog_info[i].start;
12680 #define MIN_BPF_LINEINFO_SIZE offsetofend(struct bpf_line_info, line_col)
12681 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
12683 static int check_btf_line(struct bpf_verifier_env *env,
12684 const union bpf_attr *attr,
12687 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
12688 struct bpf_subprog_info *sub;
12689 struct bpf_line_info *linfo;
12690 struct bpf_prog *prog;
12691 const struct btf *btf;
12695 nr_linfo = attr->line_info_cnt;
12698 if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info))
12701 rec_size = attr->line_info_rec_size;
12702 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
12703 rec_size > MAX_LINEINFO_REC_SIZE ||
12704 rec_size & (sizeof(u32) - 1))
12707 /* Need to zero it in case the userspace may
12708 * pass in a smaller bpf_line_info object.
12710 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
12711 GFP_KERNEL | __GFP_NOWARN);
12716 btf = prog->aux->btf;
12719 sub = env->subprog_info;
12720 ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
12721 expected_size = sizeof(struct bpf_line_info);
12722 ncopy = min_t(u32, expected_size, rec_size);
12723 for (i = 0; i < nr_linfo; i++) {
12724 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
12726 if (err == -E2BIG) {
12727 verbose(env, "nonzero tailing record in line_info");
12728 if (copy_to_bpfptr_offset(uattr,
12729 offsetof(union bpf_attr, line_info_rec_size),
12730 &expected_size, sizeof(expected_size)))
12736 if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
12742 * Check insn_off to ensure
12743 * 1) strictly increasing AND
12744 * 2) bounded by prog->len
12746 * The linfo[0].insn_off == 0 check logically falls into
12747 * the later "missing bpf_line_info for func..." case
12748 * because the first linfo[0].insn_off must be the
12749 * first sub also and the first sub must have
12750 * subprog_info[0].start == 0.
12752 if ((i && linfo[i].insn_off <= prev_offset) ||
12753 linfo[i].insn_off >= prog->len) {
12754 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
12755 i, linfo[i].insn_off, prev_offset,
12761 if (!prog->insnsi[linfo[i].insn_off].code) {
12763 "Invalid insn code at line_info[%u].insn_off\n",
12769 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
12770 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
12771 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
12776 if (s != env->subprog_cnt) {
12777 if (linfo[i].insn_off == sub[s].start) {
12778 sub[s].linfo_idx = i;
12780 } else if (sub[s].start < linfo[i].insn_off) {
12781 verbose(env, "missing bpf_line_info for func#%u\n", s);
12787 prev_offset = linfo[i].insn_off;
12788 bpfptr_add(&ulinfo, rec_size);
12791 if (s != env->subprog_cnt) {
12792 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
12793 env->subprog_cnt - s, s);
12798 prog->aux->linfo = linfo;
12799 prog->aux->nr_linfo = nr_linfo;
12808 #define MIN_CORE_RELO_SIZE sizeof(struct bpf_core_relo)
12809 #define MAX_CORE_RELO_SIZE MAX_FUNCINFO_REC_SIZE
12811 static int check_core_relo(struct bpf_verifier_env *env,
12812 const union bpf_attr *attr,
12815 u32 i, nr_core_relo, ncopy, expected_size, rec_size;
12816 struct bpf_core_relo core_relo = {};
12817 struct bpf_prog *prog = env->prog;
12818 const struct btf *btf = prog->aux->btf;
12819 struct bpf_core_ctx ctx = {
12823 bpfptr_t u_core_relo;
12826 nr_core_relo = attr->core_relo_cnt;
12829 if (nr_core_relo > INT_MAX / sizeof(struct bpf_core_relo))
12832 rec_size = attr->core_relo_rec_size;
12833 if (rec_size < MIN_CORE_RELO_SIZE ||
12834 rec_size > MAX_CORE_RELO_SIZE ||
12835 rec_size % sizeof(u32))
12838 u_core_relo = make_bpfptr(attr->core_relos, uattr.is_kernel);
12839 expected_size = sizeof(struct bpf_core_relo);
12840 ncopy = min_t(u32, expected_size, rec_size);
12842 /* Unlike func_info and line_info, copy and apply each CO-RE
12843 * relocation record one at a time.
12845 for (i = 0; i < nr_core_relo; i++) {
12846 /* future proofing when sizeof(bpf_core_relo) changes */
12847 err = bpf_check_uarg_tail_zero(u_core_relo, expected_size, rec_size);
12849 if (err == -E2BIG) {
12850 verbose(env, "nonzero tailing record in core_relo");
12851 if (copy_to_bpfptr_offset(uattr,
12852 offsetof(union bpf_attr, core_relo_rec_size),
12853 &expected_size, sizeof(expected_size)))
12859 if (copy_from_bpfptr(&core_relo, u_core_relo, ncopy)) {
12864 if (core_relo.insn_off % 8 || core_relo.insn_off / 8 >= prog->len) {
12865 verbose(env, "Invalid core_relo[%u].insn_off:%u prog->len:%u\n",
12866 i, core_relo.insn_off, prog->len);
12871 err = bpf_core_apply(&ctx, &core_relo, i,
12872 &prog->insnsi[core_relo.insn_off / 8]);
12875 bpfptr_add(&u_core_relo, rec_size);
12880 static int check_btf_info(struct bpf_verifier_env *env,
12881 const union bpf_attr *attr,
12887 if (!attr->func_info_cnt && !attr->line_info_cnt) {
12888 if (check_abnormal_return(env))
12893 btf = btf_get_by_fd(attr->prog_btf_fd);
12895 return PTR_ERR(btf);
12896 if (btf_is_kernel(btf)) {
12900 env->prog->aux->btf = btf;
12902 err = check_btf_func(env, attr, uattr);
12906 err = check_btf_line(env, attr, uattr);
12910 err = check_core_relo(env, attr, uattr);
12917 /* check %cur's range satisfies %old's */
12918 static bool range_within(struct bpf_reg_state *old,
12919 struct bpf_reg_state *cur)
12921 return old->umin_value <= cur->umin_value &&
12922 old->umax_value >= cur->umax_value &&
12923 old->smin_value <= cur->smin_value &&
12924 old->smax_value >= cur->smax_value &&
12925 old->u32_min_value <= cur->u32_min_value &&
12926 old->u32_max_value >= cur->u32_max_value &&
12927 old->s32_min_value <= cur->s32_min_value &&
12928 old->s32_max_value >= cur->s32_max_value;
12931 /* If in the old state two registers had the same id, then they need to have
12932 * the same id in the new state as well. But that id could be different from
12933 * the old state, so we need to track the mapping from old to new ids.
12934 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
12935 * regs with old id 5 must also have new id 9 for the new state to be safe. But
12936 * regs with a different old id could still have new id 9, we don't care about
12938 * So we look through our idmap to see if this old id has been seen before. If
12939 * so, we require the new id to match; otherwise, we add the id pair to the map.
12941 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_id_pair *idmap)
12945 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
12946 if (!idmap[i].old) {
12947 /* Reached an empty slot; haven't seen this id before */
12948 idmap[i].old = old_id;
12949 idmap[i].cur = cur_id;
12952 if (idmap[i].old == old_id)
12953 return idmap[i].cur == cur_id;
12955 /* We ran out of idmap slots, which should be impossible */
12960 static void clean_func_state(struct bpf_verifier_env *env,
12961 struct bpf_func_state *st)
12963 enum bpf_reg_liveness live;
12966 for (i = 0; i < BPF_REG_FP; i++) {
12967 live = st->regs[i].live;
12968 /* liveness must not touch this register anymore */
12969 st->regs[i].live |= REG_LIVE_DONE;
12970 if (!(live & REG_LIVE_READ))
12971 /* since the register is unused, clear its state
12972 * to make further comparison simpler
12974 __mark_reg_not_init(env, &st->regs[i]);
12977 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
12978 live = st->stack[i].spilled_ptr.live;
12979 /* liveness must not touch this stack slot anymore */
12980 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
12981 if (!(live & REG_LIVE_READ)) {
12982 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
12983 for (j = 0; j < BPF_REG_SIZE; j++)
12984 st->stack[i].slot_type[j] = STACK_INVALID;
12989 static void clean_verifier_state(struct bpf_verifier_env *env,
12990 struct bpf_verifier_state *st)
12994 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
12995 /* all regs in this state in all frames were already marked */
12998 for (i = 0; i <= st->curframe; i++)
12999 clean_func_state(env, st->frame[i]);
13002 /* the parentage chains form a tree.
13003 * the verifier states are added to state lists at given insn and
13004 * pushed into state stack for future exploration.
13005 * when the verifier reaches bpf_exit insn some of the verifer states
13006 * stored in the state lists have their final liveness state already,
13007 * but a lot of states will get revised from liveness point of view when
13008 * the verifier explores other branches.
13011 * 2: if r1 == 100 goto pc+1
13014 * when the verifier reaches exit insn the register r0 in the state list of
13015 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
13016 * of insn 2 and goes exploring further. At the insn 4 it will walk the
13017 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
13019 * Since the verifier pushes the branch states as it sees them while exploring
13020 * the program the condition of walking the branch instruction for the second
13021 * time means that all states below this branch were already explored and
13022 * their final liveness marks are already propagated.
13023 * Hence when the verifier completes the search of state list in is_state_visited()
13024 * we can call this clean_live_states() function to mark all liveness states
13025 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
13026 * will not be used.
13027 * This function also clears the registers and stack for states that !READ
13028 * to simplify state merging.
13030 * Important note here that walking the same branch instruction in the callee
13031 * doesn't meant that the states are DONE. The verifier has to compare
13034 static void clean_live_states(struct bpf_verifier_env *env, int insn,
13035 struct bpf_verifier_state *cur)
13037 struct bpf_verifier_state_list *sl;
13040 sl = *explored_state(env, insn);
13042 if (sl->state.branches)
13044 if (sl->state.insn_idx != insn ||
13045 sl->state.curframe != cur->curframe)
13047 for (i = 0; i <= cur->curframe; i++)
13048 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
13050 clean_verifier_state(env, &sl->state);
13056 /* Returns true if (rold safe implies rcur safe) */
13057 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
13058 struct bpf_reg_state *rcur, struct bpf_id_pair *idmap)
13062 if (!(rold->live & REG_LIVE_READ))
13063 /* explored state didn't use this */
13066 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
13068 if (rold->type == NOT_INIT)
13069 /* explored state can't have used this */
13071 if (rcur->type == NOT_INIT)
13073 switch (base_type(rold->type)) {
13077 if (env->explore_alu_limits)
13079 if (rcur->type == SCALAR_VALUE) {
13080 if (!rold->precise)
13082 /* new val must satisfy old val knowledge */
13083 return range_within(rold, rcur) &&
13084 tnum_in(rold->var_off, rcur->var_off);
13086 /* We're trying to use a pointer in place of a scalar.
13087 * Even if the scalar was unbounded, this could lead to
13088 * pointer leaks because scalars are allowed to leak
13089 * while pointers are not. We could make this safe in
13090 * special cases if root is calling us, but it's
13091 * probably not worth the hassle.
13095 case PTR_TO_MAP_KEY:
13096 case PTR_TO_MAP_VALUE:
13097 /* a PTR_TO_MAP_VALUE could be safe to use as a
13098 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
13099 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
13100 * checked, doing so could have affected others with the same
13101 * id, and we can't check for that because we lost the id when
13102 * we converted to a PTR_TO_MAP_VALUE.
13104 if (type_may_be_null(rold->type)) {
13105 if (!type_may_be_null(rcur->type))
13107 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
13109 /* Check our ids match any regs they're supposed to */
13110 return check_ids(rold->id, rcur->id, idmap);
13113 /* If the new min/max/var_off satisfy the old ones and
13114 * everything else matches, we are OK.
13115 * 'id' is not compared, since it's only used for maps with
13116 * bpf_spin_lock inside map element and in such cases if
13117 * the rest of the prog is valid for one map element then
13118 * it's valid for all map elements regardless of the key
13119 * used in bpf_map_lookup()
13121 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
13122 range_within(rold, rcur) &&
13123 tnum_in(rold->var_off, rcur->var_off) &&
13124 check_ids(rold->id, rcur->id, idmap);
13125 case PTR_TO_PACKET_META:
13126 case PTR_TO_PACKET:
13127 if (rcur->type != rold->type)
13129 /* We must have at least as much range as the old ptr
13130 * did, so that any accesses which were safe before are
13131 * still safe. This is true even if old range < old off,
13132 * since someone could have accessed through (ptr - k), or
13133 * even done ptr -= k in a register, to get a safe access.
13135 if (rold->range > rcur->range)
13137 /* If the offsets don't match, we can't trust our alignment;
13138 * nor can we be sure that we won't fall out of range.
13140 if (rold->off != rcur->off)
13142 /* id relations must be preserved */
13143 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
13145 /* new val must satisfy old val knowledge */
13146 return range_within(rold, rcur) &&
13147 tnum_in(rold->var_off, rcur->var_off);
13149 /* two stack pointers are equal only if they're pointing to
13150 * the same stack frame, since fp-8 in foo != fp-8 in bar
13152 return equal && rold->frameno == rcur->frameno;
13154 /* Only valid matches are exact, which memcmp() */
13158 /* Shouldn't get here; if we do, say it's not safe */
13163 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
13164 struct bpf_func_state *cur, struct bpf_id_pair *idmap)
13168 /* walk slots of the explored stack and ignore any additional
13169 * slots in the current stack, since explored(safe) state
13172 for (i = 0; i < old->allocated_stack; i++) {
13173 spi = i / BPF_REG_SIZE;
13175 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
13176 i += BPF_REG_SIZE - 1;
13177 /* explored state didn't use this */
13181 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
13184 /* explored stack has more populated slots than current stack
13185 * and these slots were used
13187 if (i >= cur->allocated_stack)
13190 /* if old state was safe with misc data in the stack
13191 * it will be safe with zero-initialized stack.
13192 * The opposite is not true
13194 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
13195 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
13197 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
13198 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
13199 /* Ex: old explored (safe) state has STACK_SPILL in
13200 * this stack slot, but current has STACK_MISC ->
13201 * this verifier states are not equivalent,
13202 * return false to continue verification of this path
13205 if (i % BPF_REG_SIZE != BPF_REG_SIZE - 1)
13207 if (!is_spilled_reg(&old->stack[spi]))
13209 if (!regsafe(env, &old->stack[spi].spilled_ptr,
13210 &cur->stack[spi].spilled_ptr, idmap))
13211 /* when explored and current stack slot are both storing
13212 * spilled registers, check that stored pointers types
13213 * are the same as well.
13214 * Ex: explored safe path could have stored
13215 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
13216 * but current path has stored:
13217 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
13218 * such verifier states are not equivalent.
13219 * return false to continue verification of this path
13226 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
13228 if (old->acquired_refs != cur->acquired_refs)
13230 return !memcmp(old->refs, cur->refs,
13231 sizeof(*old->refs) * old->acquired_refs);
13234 /* compare two verifier states
13236 * all states stored in state_list are known to be valid, since
13237 * verifier reached 'bpf_exit' instruction through them
13239 * this function is called when verifier exploring different branches of
13240 * execution popped from the state stack. If it sees an old state that has
13241 * more strict register state and more strict stack state then this execution
13242 * branch doesn't need to be explored further, since verifier already
13243 * concluded that more strict state leads to valid finish.
13245 * Therefore two states are equivalent if register state is more conservative
13246 * and explored stack state is more conservative than the current one.
13249 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
13250 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
13252 * In other words if current stack state (one being explored) has more
13253 * valid slots than old one that already passed validation, it means
13254 * the verifier can stop exploring and conclude that current state is valid too
13256 * Similarly with registers. If explored state has register type as invalid
13257 * whereas register type in current state is meaningful, it means that
13258 * the current state will reach 'bpf_exit' instruction safely
13260 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
13261 struct bpf_func_state *cur)
13265 for (i = 0; i < MAX_BPF_REG; i++)
13266 if (!regsafe(env, &old->regs[i], &cur->regs[i],
13267 env->idmap_scratch))
13270 if (!stacksafe(env, old, cur, env->idmap_scratch))
13273 if (!refsafe(old, cur))
13279 static bool states_equal(struct bpf_verifier_env *env,
13280 struct bpf_verifier_state *old,
13281 struct bpf_verifier_state *cur)
13285 if (old->curframe != cur->curframe)
13288 memset(env->idmap_scratch, 0, sizeof(env->idmap_scratch));
13290 /* Verification state from speculative execution simulation
13291 * must never prune a non-speculative execution one.
13293 if (old->speculative && !cur->speculative)
13296 if (old->active_lock.ptr != cur->active_lock.ptr)
13299 /* Old and cur active_lock's have to be either both present
13302 if (!!old->active_lock.id != !!cur->active_lock.id)
13305 if (old->active_lock.id &&
13306 !check_ids(old->active_lock.id, cur->active_lock.id, env->idmap_scratch))
13309 if (old->active_rcu_lock != cur->active_rcu_lock)
13312 /* for states to be equal callsites have to be the same
13313 * and all frame states need to be equivalent
13315 for (i = 0; i <= old->curframe; i++) {
13316 if (old->frame[i]->callsite != cur->frame[i]->callsite)
13318 if (!func_states_equal(env, old->frame[i], cur->frame[i]))
13324 /* Return 0 if no propagation happened. Return negative error code if error
13325 * happened. Otherwise, return the propagated bit.
13327 static int propagate_liveness_reg(struct bpf_verifier_env *env,
13328 struct bpf_reg_state *reg,
13329 struct bpf_reg_state *parent_reg)
13331 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
13332 u8 flag = reg->live & REG_LIVE_READ;
13335 /* When comes here, read flags of PARENT_REG or REG could be any of
13336 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
13337 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
13339 if (parent_flag == REG_LIVE_READ64 ||
13340 /* Or if there is no read flag from REG. */
13342 /* Or if the read flag from REG is the same as PARENT_REG. */
13343 parent_flag == flag)
13346 err = mark_reg_read(env, reg, parent_reg, flag);
13353 /* A write screens off any subsequent reads; but write marks come from the
13354 * straight-line code between a state and its parent. When we arrive at an
13355 * equivalent state (jump target or such) we didn't arrive by the straight-line
13356 * code, so read marks in the state must propagate to the parent regardless
13357 * of the state's write marks. That's what 'parent == state->parent' comparison
13358 * in mark_reg_read() is for.
13360 static int propagate_liveness(struct bpf_verifier_env *env,
13361 const struct bpf_verifier_state *vstate,
13362 struct bpf_verifier_state *vparent)
13364 struct bpf_reg_state *state_reg, *parent_reg;
13365 struct bpf_func_state *state, *parent;
13366 int i, frame, err = 0;
13368 if (vparent->curframe != vstate->curframe) {
13369 WARN(1, "propagate_live: parent frame %d current frame %d\n",
13370 vparent->curframe, vstate->curframe);
13373 /* Propagate read liveness of registers... */
13374 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
13375 for (frame = 0; frame <= vstate->curframe; frame++) {
13376 parent = vparent->frame[frame];
13377 state = vstate->frame[frame];
13378 parent_reg = parent->regs;
13379 state_reg = state->regs;
13380 /* We don't need to worry about FP liveness, it's read-only */
13381 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
13382 err = propagate_liveness_reg(env, &state_reg[i],
13386 if (err == REG_LIVE_READ64)
13387 mark_insn_zext(env, &parent_reg[i]);
13390 /* Propagate stack slots. */
13391 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
13392 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
13393 parent_reg = &parent->stack[i].spilled_ptr;
13394 state_reg = &state->stack[i].spilled_ptr;
13395 err = propagate_liveness_reg(env, state_reg,
13404 /* find precise scalars in the previous equivalent state and
13405 * propagate them into the current state
13407 static int propagate_precision(struct bpf_verifier_env *env,
13408 const struct bpf_verifier_state *old)
13410 struct bpf_reg_state *state_reg;
13411 struct bpf_func_state *state;
13412 int i, err = 0, fr;
13414 for (fr = old->curframe; fr >= 0; fr--) {
13415 state = old->frame[fr];
13416 state_reg = state->regs;
13417 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
13418 if (state_reg->type != SCALAR_VALUE ||
13419 !state_reg->precise)
13421 if (env->log.level & BPF_LOG_LEVEL2)
13422 verbose(env, "frame %d: propagating r%d\n", i, fr);
13423 err = mark_chain_precision_frame(env, fr, i);
13428 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
13429 if (!is_spilled_reg(&state->stack[i]))
13431 state_reg = &state->stack[i].spilled_ptr;
13432 if (state_reg->type != SCALAR_VALUE ||
13433 !state_reg->precise)
13435 if (env->log.level & BPF_LOG_LEVEL2)
13436 verbose(env, "frame %d: propagating fp%d\n",
13437 (-i - 1) * BPF_REG_SIZE, fr);
13438 err = mark_chain_precision_stack_frame(env, fr, i);
13446 static bool states_maybe_looping(struct bpf_verifier_state *old,
13447 struct bpf_verifier_state *cur)
13449 struct bpf_func_state *fold, *fcur;
13450 int i, fr = cur->curframe;
13452 if (old->curframe != fr)
13455 fold = old->frame[fr];
13456 fcur = cur->frame[fr];
13457 for (i = 0; i < MAX_BPF_REG; i++)
13458 if (memcmp(&fold->regs[i], &fcur->regs[i],
13459 offsetof(struct bpf_reg_state, parent)))
13465 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
13467 struct bpf_verifier_state_list *new_sl;
13468 struct bpf_verifier_state_list *sl, **pprev;
13469 struct bpf_verifier_state *cur = env->cur_state, *new;
13470 int i, j, err, states_cnt = 0;
13471 bool add_new_state = env->test_state_freq ? true : false;
13473 /* bpf progs typically have pruning point every 4 instructions
13474 * http://vger.kernel.org/bpfconf2019.html#session-1
13475 * Do not add new state for future pruning if the verifier hasn't seen
13476 * at least 2 jumps and at least 8 instructions.
13477 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
13478 * In tests that amounts to up to 50% reduction into total verifier
13479 * memory consumption and 20% verifier time speedup.
13481 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
13482 env->insn_processed - env->prev_insn_processed >= 8)
13483 add_new_state = true;
13485 pprev = explored_state(env, insn_idx);
13488 clean_live_states(env, insn_idx, cur);
13492 if (sl->state.insn_idx != insn_idx)
13495 if (sl->state.branches) {
13496 struct bpf_func_state *frame = sl->state.frame[sl->state.curframe];
13498 if (frame->in_async_callback_fn &&
13499 frame->async_entry_cnt != cur->frame[cur->curframe]->async_entry_cnt) {
13500 /* Different async_entry_cnt means that the verifier is
13501 * processing another entry into async callback.
13502 * Seeing the same state is not an indication of infinite
13503 * loop or infinite recursion.
13504 * But finding the same state doesn't mean that it's safe
13505 * to stop processing the current state. The previous state
13506 * hasn't yet reached bpf_exit, since state.branches > 0.
13507 * Checking in_async_callback_fn alone is not enough either.
13508 * Since the verifier still needs to catch infinite loops
13509 * inside async callbacks.
13511 } else if (states_maybe_looping(&sl->state, cur) &&
13512 states_equal(env, &sl->state, cur)) {
13513 verbose_linfo(env, insn_idx, "; ");
13514 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
13517 /* if the verifier is processing a loop, avoid adding new state
13518 * too often, since different loop iterations have distinct
13519 * states and may not help future pruning.
13520 * This threshold shouldn't be too low to make sure that
13521 * a loop with large bound will be rejected quickly.
13522 * The most abusive loop will be:
13524 * if r1 < 1000000 goto pc-2
13525 * 1M insn_procssed limit / 100 == 10k peak states.
13526 * This threshold shouldn't be too high either, since states
13527 * at the end of the loop are likely to be useful in pruning.
13529 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
13530 env->insn_processed - env->prev_insn_processed < 100)
13531 add_new_state = false;
13534 if (states_equal(env, &sl->state, cur)) {
13536 /* reached equivalent register/stack state,
13537 * prune the search.
13538 * Registers read by the continuation are read by us.
13539 * If we have any write marks in env->cur_state, they
13540 * will prevent corresponding reads in the continuation
13541 * from reaching our parent (an explored_state). Our
13542 * own state will get the read marks recorded, but
13543 * they'll be immediately forgotten as we're pruning
13544 * this state and will pop a new one.
13546 err = propagate_liveness(env, &sl->state, cur);
13548 /* if previous state reached the exit with precision and
13549 * current state is equivalent to it (except precsion marks)
13550 * the precision needs to be propagated back in
13551 * the current state.
13553 err = err ? : push_jmp_history(env, cur);
13554 err = err ? : propagate_precision(env, &sl->state);
13560 /* when new state is not going to be added do not increase miss count.
13561 * Otherwise several loop iterations will remove the state
13562 * recorded earlier. The goal of these heuristics is to have
13563 * states from some iterations of the loop (some in the beginning
13564 * and some at the end) to help pruning.
13568 /* heuristic to determine whether this state is beneficial
13569 * to keep checking from state equivalence point of view.
13570 * Higher numbers increase max_states_per_insn and verification time,
13571 * but do not meaningfully decrease insn_processed.
13573 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
13574 /* the state is unlikely to be useful. Remove it to
13575 * speed up verification
13578 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
13579 u32 br = sl->state.branches;
13582 "BUG live_done but branches_to_explore %d\n",
13584 free_verifier_state(&sl->state, false);
13586 env->peak_states--;
13588 /* cannot free this state, since parentage chain may
13589 * walk it later. Add it for free_list instead to
13590 * be freed at the end of verification
13592 sl->next = env->free_list;
13593 env->free_list = sl;
13603 if (env->max_states_per_insn < states_cnt)
13604 env->max_states_per_insn = states_cnt;
13606 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
13609 if (!add_new_state)
13612 /* There were no equivalent states, remember the current one.
13613 * Technically the current state is not proven to be safe yet,
13614 * but it will either reach outer most bpf_exit (which means it's safe)
13615 * or it will be rejected. When there are no loops the verifier won't be
13616 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
13617 * again on the way to bpf_exit.
13618 * When looping the sl->state.branches will be > 0 and this state
13619 * will not be considered for equivalence until branches == 0.
13621 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
13624 env->total_states++;
13625 env->peak_states++;
13626 env->prev_jmps_processed = env->jmps_processed;
13627 env->prev_insn_processed = env->insn_processed;
13629 /* forget precise markings we inherited, see __mark_chain_precision */
13630 if (env->bpf_capable)
13631 mark_all_scalars_imprecise(env, cur);
13633 /* add new state to the head of linked list */
13634 new = &new_sl->state;
13635 err = copy_verifier_state(new, cur);
13637 free_verifier_state(new, false);
13641 new->insn_idx = insn_idx;
13642 WARN_ONCE(new->branches != 1,
13643 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
13646 cur->first_insn_idx = insn_idx;
13647 clear_jmp_history(cur);
13648 new_sl->next = *explored_state(env, insn_idx);
13649 *explored_state(env, insn_idx) = new_sl;
13650 /* connect new state to parentage chain. Current frame needs all
13651 * registers connected. Only r6 - r9 of the callers are alive (pushed
13652 * to the stack implicitly by JITs) so in callers' frames connect just
13653 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
13654 * the state of the call instruction (with WRITTEN set), and r0 comes
13655 * from callee with its full parentage chain, anyway.
13657 /* clear write marks in current state: the writes we did are not writes
13658 * our child did, so they don't screen off its reads from us.
13659 * (There are no read marks in current state, because reads always mark
13660 * their parent and current state never has children yet. Only
13661 * explored_states can get read marks.)
13663 for (j = 0; j <= cur->curframe; j++) {
13664 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
13665 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
13666 for (i = 0; i < BPF_REG_FP; i++)
13667 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
13670 /* all stack frames are accessible from callee, clear them all */
13671 for (j = 0; j <= cur->curframe; j++) {
13672 struct bpf_func_state *frame = cur->frame[j];
13673 struct bpf_func_state *newframe = new->frame[j];
13675 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
13676 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
13677 frame->stack[i].spilled_ptr.parent =
13678 &newframe->stack[i].spilled_ptr;
13684 /* Return true if it's OK to have the same insn return a different type. */
13685 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
13687 switch (base_type(type)) {
13689 case PTR_TO_SOCKET:
13690 case PTR_TO_SOCK_COMMON:
13691 case PTR_TO_TCP_SOCK:
13692 case PTR_TO_XDP_SOCK:
13693 case PTR_TO_BTF_ID:
13700 /* If an instruction was previously used with particular pointer types, then we
13701 * need to be careful to avoid cases such as the below, where it may be ok
13702 * for one branch accessing the pointer, but not ok for the other branch:
13707 * R1 = some_other_valid_ptr;
13710 * R2 = *(u32 *)(R1 + 0);
13712 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
13714 return src != prev && (!reg_type_mismatch_ok(src) ||
13715 !reg_type_mismatch_ok(prev));
13718 static int do_check(struct bpf_verifier_env *env)
13720 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
13721 struct bpf_verifier_state *state = env->cur_state;
13722 struct bpf_insn *insns = env->prog->insnsi;
13723 struct bpf_reg_state *regs;
13724 int insn_cnt = env->prog->len;
13725 bool do_print_state = false;
13726 int prev_insn_idx = -1;
13729 struct bpf_insn *insn;
13733 env->prev_insn_idx = prev_insn_idx;
13734 if (env->insn_idx >= insn_cnt) {
13735 verbose(env, "invalid insn idx %d insn_cnt %d\n",
13736 env->insn_idx, insn_cnt);
13740 insn = &insns[env->insn_idx];
13741 class = BPF_CLASS(insn->code);
13743 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
13745 "BPF program is too large. Processed %d insn\n",
13746 env->insn_processed);
13750 state->last_insn_idx = env->prev_insn_idx;
13752 if (is_prune_point(env, env->insn_idx)) {
13753 err = is_state_visited(env, env->insn_idx);
13757 /* found equivalent state, can prune the search */
13758 if (env->log.level & BPF_LOG_LEVEL) {
13759 if (do_print_state)
13760 verbose(env, "\nfrom %d to %d%s: safe\n",
13761 env->prev_insn_idx, env->insn_idx,
13762 env->cur_state->speculative ?
13763 " (speculative execution)" : "");
13765 verbose(env, "%d: safe\n", env->insn_idx);
13767 goto process_bpf_exit;
13771 if (is_jmp_point(env, env->insn_idx)) {
13772 err = push_jmp_history(env, state);
13777 if (signal_pending(current))
13780 if (need_resched())
13783 if (env->log.level & BPF_LOG_LEVEL2 && do_print_state) {
13784 verbose(env, "\nfrom %d to %d%s:",
13785 env->prev_insn_idx, env->insn_idx,
13786 env->cur_state->speculative ?
13787 " (speculative execution)" : "");
13788 print_verifier_state(env, state->frame[state->curframe], true);
13789 do_print_state = false;
13792 if (env->log.level & BPF_LOG_LEVEL) {
13793 const struct bpf_insn_cbs cbs = {
13794 .cb_call = disasm_kfunc_name,
13795 .cb_print = verbose,
13796 .private_data = env,
13799 if (verifier_state_scratched(env))
13800 print_insn_state(env, state->frame[state->curframe]);
13802 verbose_linfo(env, env->insn_idx, "; ");
13803 env->prev_log_len = env->log.len_used;
13804 verbose(env, "%d: ", env->insn_idx);
13805 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
13806 env->prev_insn_print_len = env->log.len_used - env->prev_log_len;
13807 env->prev_log_len = env->log.len_used;
13810 if (bpf_prog_is_dev_bound(env->prog->aux)) {
13811 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
13812 env->prev_insn_idx);
13817 regs = cur_regs(env);
13818 sanitize_mark_insn_seen(env);
13819 prev_insn_idx = env->insn_idx;
13821 if (class == BPF_ALU || class == BPF_ALU64) {
13822 err = check_alu_op(env, insn);
13826 } else if (class == BPF_LDX) {
13827 enum bpf_reg_type *prev_src_type, src_reg_type;
13829 /* check for reserved fields is already done */
13831 /* check src operand */
13832 err = check_reg_arg(env, insn->src_reg, SRC_OP);
13836 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
13840 src_reg_type = regs[insn->src_reg].type;
13842 /* check that memory (src_reg + off) is readable,
13843 * the state of dst_reg will be updated by this func
13845 err = check_mem_access(env, env->insn_idx, insn->src_reg,
13846 insn->off, BPF_SIZE(insn->code),
13847 BPF_READ, insn->dst_reg, false);
13851 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
13853 if (*prev_src_type == NOT_INIT) {
13854 /* saw a valid insn
13855 * dst_reg = *(u32 *)(src_reg + off)
13856 * save type to validate intersecting paths
13858 *prev_src_type = src_reg_type;
13860 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
13861 /* ABuser program is trying to use the same insn
13862 * dst_reg = *(u32*) (src_reg + off)
13863 * with different pointer types:
13864 * src_reg == ctx in one branch and
13865 * src_reg == stack|map in some other branch.
13868 verbose(env, "same insn cannot be used with different pointers\n");
13872 } else if (class == BPF_STX) {
13873 enum bpf_reg_type *prev_dst_type, dst_reg_type;
13875 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
13876 err = check_atomic(env, env->insn_idx, insn);
13883 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
13884 verbose(env, "BPF_STX uses reserved fields\n");
13888 /* check src1 operand */
13889 err = check_reg_arg(env, insn->src_reg, SRC_OP);
13892 /* check src2 operand */
13893 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
13897 dst_reg_type = regs[insn->dst_reg].type;
13899 /* check that memory (dst_reg + off) is writeable */
13900 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
13901 insn->off, BPF_SIZE(insn->code),
13902 BPF_WRITE, insn->src_reg, false);
13906 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
13908 if (*prev_dst_type == NOT_INIT) {
13909 *prev_dst_type = dst_reg_type;
13910 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
13911 verbose(env, "same insn cannot be used with different pointers\n");
13915 } else if (class == BPF_ST) {
13916 if (BPF_MODE(insn->code) != BPF_MEM ||
13917 insn->src_reg != BPF_REG_0) {
13918 verbose(env, "BPF_ST uses reserved fields\n");
13921 /* check src operand */
13922 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
13926 if (is_ctx_reg(env, insn->dst_reg)) {
13927 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
13929 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
13933 /* check that memory (dst_reg + off) is writeable */
13934 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
13935 insn->off, BPF_SIZE(insn->code),
13936 BPF_WRITE, -1, false);
13940 } else if (class == BPF_JMP || class == BPF_JMP32) {
13941 u8 opcode = BPF_OP(insn->code);
13943 env->jmps_processed++;
13944 if (opcode == BPF_CALL) {
13945 if (BPF_SRC(insn->code) != BPF_K ||
13946 (insn->src_reg != BPF_PSEUDO_KFUNC_CALL
13947 && insn->off != 0) ||
13948 (insn->src_reg != BPF_REG_0 &&
13949 insn->src_reg != BPF_PSEUDO_CALL &&
13950 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
13951 insn->dst_reg != BPF_REG_0 ||
13952 class == BPF_JMP32) {
13953 verbose(env, "BPF_CALL uses reserved fields\n");
13957 if (env->cur_state->active_lock.ptr) {
13958 if ((insn->src_reg == BPF_REG_0 && insn->imm != BPF_FUNC_spin_unlock) ||
13959 (insn->src_reg == BPF_PSEUDO_CALL) ||
13960 (insn->src_reg == BPF_PSEUDO_KFUNC_CALL &&
13961 (insn->off != 0 || !is_bpf_list_api_kfunc(insn->imm)))) {
13962 verbose(env, "function calls are not allowed while holding a lock\n");
13966 if (insn->src_reg == BPF_PSEUDO_CALL)
13967 err = check_func_call(env, insn, &env->insn_idx);
13968 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
13969 err = check_kfunc_call(env, insn, &env->insn_idx);
13971 err = check_helper_call(env, insn, &env->insn_idx);
13974 } else if (opcode == BPF_JA) {
13975 if (BPF_SRC(insn->code) != BPF_K ||
13977 insn->src_reg != BPF_REG_0 ||
13978 insn->dst_reg != BPF_REG_0 ||
13979 class == BPF_JMP32) {
13980 verbose(env, "BPF_JA uses reserved fields\n");
13984 env->insn_idx += insn->off + 1;
13987 } else if (opcode == BPF_EXIT) {
13988 if (BPF_SRC(insn->code) != BPF_K ||
13990 insn->src_reg != BPF_REG_0 ||
13991 insn->dst_reg != BPF_REG_0 ||
13992 class == BPF_JMP32) {
13993 verbose(env, "BPF_EXIT uses reserved fields\n");
13997 if (env->cur_state->active_lock.ptr) {
13998 verbose(env, "bpf_spin_unlock is missing\n");
14002 if (env->cur_state->active_rcu_lock) {
14003 verbose(env, "bpf_rcu_read_unlock is missing\n");
14007 /* We must do check_reference_leak here before
14008 * prepare_func_exit to handle the case when
14009 * state->curframe > 0, it may be a callback
14010 * function, for which reference_state must
14011 * match caller reference state when it exits.
14013 err = check_reference_leak(env);
14017 if (state->curframe) {
14018 /* exit from nested function */
14019 err = prepare_func_exit(env, &env->insn_idx);
14022 do_print_state = true;
14026 err = check_return_code(env);
14030 mark_verifier_state_scratched(env);
14031 update_branch_counts(env, env->cur_state);
14032 err = pop_stack(env, &prev_insn_idx,
14033 &env->insn_idx, pop_log);
14035 if (err != -ENOENT)
14039 do_print_state = true;
14043 err = check_cond_jmp_op(env, insn, &env->insn_idx);
14047 } else if (class == BPF_LD) {
14048 u8 mode = BPF_MODE(insn->code);
14050 if (mode == BPF_ABS || mode == BPF_IND) {
14051 err = check_ld_abs(env, insn);
14055 } else if (mode == BPF_IMM) {
14056 err = check_ld_imm(env, insn);
14061 sanitize_mark_insn_seen(env);
14063 verbose(env, "invalid BPF_LD mode\n");
14067 verbose(env, "unknown insn class %d\n", class);
14077 static int find_btf_percpu_datasec(struct btf *btf)
14079 const struct btf_type *t;
14084 * Both vmlinux and module each have their own ".data..percpu"
14085 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
14086 * types to look at only module's own BTF types.
14088 n = btf_nr_types(btf);
14089 if (btf_is_module(btf))
14090 i = btf_nr_types(btf_vmlinux);
14094 for(; i < n; i++) {
14095 t = btf_type_by_id(btf, i);
14096 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
14099 tname = btf_name_by_offset(btf, t->name_off);
14100 if (!strcmp(tname, ".data..percpu"))
14107 /* replace pseudo btf_id with kernel symbol address */
14108 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
14109 struct bpf_insn *insn,
14110 struct bpf_insn_aux_data *aux)
14112 const struct btf_var_secinfo *vsi;
14113 const struct btf_type *datasec;
14114 struct btf_mod_pair *btf_mod;
14115 const struct btf_type *t;
14116 const char *sym_name;
14117 bool percpu = false;
14118 u32 type, id = insn->imm;
14122 int i, btf_fd, err;
14124 btf_fd = insn[1].imm;
14126 btf = btf_get_by_fd(btf_fd);
14128 verbose(env, "invalid module BTF object FD specified.\n");
14132 if (!btf_vmlinux) {
14133 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
14140 t = btf_type_by_id(btf, id);
14142 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
14147 if (!btf_type_is_var(t)) {
14148 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR.\n", id);
14153 sym_name = btf_name_by_offset(btf, t->name_off);
14154 addr = kallsyms_lookup_name(sym_name);
14156 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
14162 datasec_id = find_btf_percpu_datasec(btf);
14163 if (datasec_id > 0) {
14164 datasec = btf_type_by_id(btf, datasec_id);
14165 for_each_vsi(i, datasec, vsi) {
14166 if (vsi->type == id) {
14173 insn[0].imm = (u32)addr;
14174 insn[1].imm = addr >> 32;
14177 t = btf_type_skip_modifiers(btf, type, NULL);
14179 aux->btf_var.reg_type = PTR_TO_BTF_ID | MEM_PERCPU;
14180 aux->btf_var.btf = btf;
14181 aux->btf_var.btf_id = type;
14182 } else if (!btf_type_is_struct(t)) {
14183 const struct btf_type *ret;
14187 /* resolve the type size of ksym. */
14188 ret = btf_resolve_size(btf, t, &tsize);
14190 tname = btf_name_by_offset(btf, t->name_off);
14191 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
14192 tname, PTR_ERR(ret));
14196 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
14197 aux->btf_var.mem_size = tsize;
14199 aux->btf_var.reg_type = PTR_TO_BTF_ID;
14200 aux->btf_var.btf = btf;
14201 aux->btf_var.btf_id = type;
14204 /* check whether we recorded this BTF (and maybe module) already */
14205 for (i = 0; i < env->used_btf_cnt; i++) {
14206 if (env->used_btfs[i].btf == btf) {
14212 if (env->used_btf_cnt >= MAX_USED_BTFS) {
14217 btf_mod = &env->used_btfs[env->used_btf_cnt];
14218 btf_mod->btf = btf;
14219 btf_mod->module = NULL;
14221 /* if we reference variables from kernel module, bump its refcount */
14222 if (btf_is_module(btf)) {
14223 btf_mod->module = btf_try_get_module(btf);
14224 if (!btf_mod->module) {
14230 env->used_btf_cnt++;
14238 static bool is_tracing_prog_type(enum bpf_prog_type type)
14241 case BPF_PROG_TYPE_KPROBE:
14242 case BPF_PROG_TYPE_TRACEPOINT:
14243 case BPF_PROG_TYPE_PERF_EVENT:
14244 case BPF_PROG_TYPE_RAW_TRACEPOINT:
14245 case BPF_PROG_TYPE_RAW_TRACEPOINT_WRITABLE:
14252 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
14253 struct bpf_map *map,
14254 struct bpf_prog *prog)
14257 enum bpf_prog_type prog_type = resolve_prog_type(prog);
14259 if (btf_record_has_field(map->record, BPF_LIST_HEAD)) {
14260 if (is_tracing_prog_type(prog_type)) {
14261 verbose(env, "tracing progs cannot use bpf_list_head yet\n");
14266 if (btf_record_has_field(map->record, BPF_SPIN_LOCK)) {
14267 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
14268 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
14272 if (is_tracing_prog_type(prog_type)) {
14273 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
14277 if (prog->aux->sleepable) {
14278 verbose(env, "sleepable progs cannot use bpf_spin_lock yet\n");
14283 if (btf_record_has_field(map->record, BPF_TIMER)) {
14284 if (is_tracing_prog_type(prog_type)) {
14285 verbose(env, "tracing progs cannot use bpf_timer yet\n");
14290 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
14291 !bpf_offload_prog_map_match(prog, map)) {
14292 verbose(env, "offload device mismatch between prog and map\n");
14296 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
14297 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
14301 if (prog->aux->sleepable)
14302 switch (map->map_type) {
14303 case BPF_MAP_TYPE_HASH:
14304 case BPF_MAP_TYPE_LRU_HASH:
14305 case BPF_MAP_TYPE_ARRAY:
14306 case BPF_MAP_TYPE_PERCPU_HASH:
14307 case BPF_MAP_TYPE_PERCPU_ARRAY:
14308 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
14309 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
14310 case BPF_MAP_TYPE_HASH_OF_MAPS:
14311 case BPF_MAP_TYPE_RINGBUF:
14312 case BPF_MAP_TYPE_USER_RINGBUF:
14313 case BPF_MAP_TYPE_INODE_STORAGE:
14314 case BPF_MAP_TYPE_SK_STORAGE:
14315 case BPF_MAP_TYPE_TASK_STORAGE:
14316 case BPF_MAP_TYPE_CGRP_STORAGE:
14320 "Sleepable programs can only use array, hash, ringbuf and local storage maps\n");
14327 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
14329 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
14330 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
14333 /* find and rewrite pseudo imm in ld_imm64 instructions:
14335 * 1. if it accesses map FD, replace it with actual map pointer.
14336 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
14338 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
14340 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
14342 struct bpf_insn *insn = env->prog->insnsi;
14343 int insn_cnt = env->prog->len;
14346 err = bpf_prog_calc_tag(env->prog);
14350 for (i = 0; i < insn_cnt; i++, insn++) {
14351 if (BPF_CLASS(insn->code) == BPF_LDX &&
14352 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
14353 verbose(env, "BPF_LDX uses reserved fields\n");
14357 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
14358 struct bpf_insn_aux_data *aux;
14359 struct bpf_map *map;
14364 if (i == insn_cnt - 1 || insn[1].code != 0 ||
14365 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
14366 insn[1].off != 0) {
14367 verbose(env, "invalid bpf_ld_imm64 insn\n");
14371 if (insn[0].src_reg == 0)
14372 /* valid generic load 64-bit imm */
14375 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
14376 aux = &env->insn_aux_data[i];
14377 err = check_pseudo_btf_id(env, insn, aux);
14383 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
14384 aux = &env->insn_aux_data[i];
14385 aux->ptr_type = PTR_TO_FUNC;
14389 /* In final convert_pseudo_ld_imm64() step, this is
14390 * converted into regular 64-bit imm load insn.
14392 switch (insn[0].src_reg) {
14393 case BPF_PSEUDO_MAP_VALUE:
14394 case BPF_PSEUDO_MAP_IDX_VALUE:
14396 case BPF_PSEUDO_MAP_FD:
14397 case BPF_PSEUDO_MAP_IDX:
14398 if (insn[1].imm == 0)
14402 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
14406 switch (insn[0].src_reg) {
14407 case BPF_PSEUDO_MAP_IDX_VALUE:
14408 case BPF_PSEUDO_MAP_IDX:
14409 if (bpfptr_is_null(env->fd_array)) {
14410 verbose(env, "fd_idx without fd_array is invalid\n");
14413 if (copy_from_bpfptr_offset(&fd, env->fd_array,
14414 insn[0].imm * sizeof(fd),
14424 map = __bpf_map_get(f);
14426 verbose(env, "fd %d is not pointing to valid bpf_map\n",
14428 return PTR_ERR(map);
14431 err = check_map_prog_compatibility(env, map, env->prog);
14437 aux = &env->insn_aux_data[i];
14438 if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
14439 insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
14440 addr = (unsigned long)map;
14442 u32 off = insn[1].imm;
14444 if (off >= BPF_MAX_VAR_OFF) {
14445 verbose(env, "direct value offset of %u is not allowed\n", off);
14450 if (!map->ops->map_direct_value_addr) {
14451 verbose(env, "no direct value access support for this map type\n");
14456 err = map->ops->map_direct_value_addr(map, &addr, off);
14458 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
14459 map->value_size, off);
14464 aux->map_off = off;
14468 insn[0].imm = (u32)addr;
14469 insn[1].imm = addr >> 32;
14471 /* check whether we recorded this map already */
14472 for (j = 0; j < env->used_map_cnt; j++) {
14473 if (env->used_maps[j] == map) {
14474 aux->map_index = j;
14480 if (env->used_map_cnt >= MAX_USED_MAPS) {
14485 /* hold the map. If the program is rejected by verifier,
14486 * the map will be released by release_maps() or it
14487 * will be used by the valid program until it's unloaded
14488 * and all maps are released in free_used_maps()
14492 aux->map_index = env->used_map_cnt;
14493 env->used_maps[env->used_map_cnt++] = map;
14495 if (bpf_map_is_cgroup_storage(map) &&
14496 bpf_cgroup_storage_assign(env->prog->aux, map)) {
14497 verbose(env, "only one cgroup storage of each type is allowed\n");
14509 /* Basic sanity check before we invest more work here. */
14510 if (!bpf_opcode_in_insntable(insn->code)) {
14511 verbose(env, "unknown opcode %02x\n", insn->code);
14516 /* now all pseudo BPF_LD_IMM64 instructions load valid
14517 * 'struct bpf_map *' into a register instead of user map_fd.
14518 * These pointers will be used later by verifier to validate map access.
14523 /* drop refcnt of maps used by the rejected program */
14524 static void release_maps(struct bpf_verifier_env *env)
14526 __bpf_free_used_maps(env->prog->aux, env->used_maps,
14527 env->used_map_cnt);
14530 /* drop refcnt of maps used by the rejected program */
14531 static void release_btfs(struct bpf_verifier_env *env)
14533 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
14534 env->used_btf_cnt);
14537 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
14538 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
14540 struct bpf_insn *insn = env->prog->insnsi;
14541 int insn_cnt = env->prog->len;
14544 for (i = 0; i < insn_cnt; i++, insn++) {
14545 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
14547 if (insn->src_reg == BPF_PSEUDO_FUNC)
14553 /* single env->prog->insni[off] instruction was replaced with the range
14554 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
14555 * [0, off) and [off, end) to new locations, so the patched range stays zero
14557 static void adjust_insn_aux_data(struct bpf_verifier_env *env,
14558 struct bpf_insn_aux_data *new_data,
14559 struct bpf_prog *new_prog, u32 off, u32 cnt)
14561 struct bpf_insn_aux_data *old_data = env->insn_aux_data;
14562 struct bpf_insn *insn = new_prog->insnsi;
14563 u32 old_seen = old_data[off].seen;
14567 /* aux info at OFF always needs adjustment, no matter fast path
14568 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
14569 * original insn at old prog.
14571 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
14575 prog_len = new_prog->len;
14577 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
14578 memcpy(new_data + off + cnt - 1, old_data + off,
14579 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
14580 for (i = off; i < off + cnt - 1; i++) {
14581 /* Expand insni[off]'s seen count to the patched range. */
14582 new_data[i].seen = old_seen;
14583 new_data[i].zext_dst = insn_has_def32(env, insn + i);
14585 env->insn_aux_data = new_data;
14589 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
14595 /* NOTE: fake 'exit' subprog should be updated as well. */
14596 for (i = 0; i <= env->subprog_cnt; i++) {
14597 if (env->subprog_info[i].start <= off)
14599 env->subprog_info[i].start += len - 1;
14603 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
14605 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
14606 int i, sz = prog->aux->size_poke_tab;
14607 struct bpf_jit_poke_descriptor *desc;
14609 for (i = 0; i < sz; i++) {
14611 if (desc->insn_idx <= off)
14613 desc->insn_idx += len - 1;
14617 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
14618 const struct bpf_insn *patch, u32 len)
14620 struct bpf_prog *new_prog;
14621 struct bpf_insn_aux_data *new_data = NULL;
14624 new_data = vzalloc(array_size(env->prog->len + len - 1,
14625 sizeof(struct bpf_insn_aux_data)));
14630 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
14631 if (IS_ERR(new_prog)) {
14632 if (PTR_ERR(new_prog) == -ERANGE)
14634 "insn %d cannot be patched due to 16-bit range\n",
14635 env->insn_aux_data[off].orig_idx);
14639 adjust_insn_aux_data(env, new_data, new_prog, off, len);
14640 adjust_subprog_starts(env, off, len);
14641 adjust_poke_descs(new_prog, off, len);
14645 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
14650 /* find first prog starting at or after off (first to remove) */
14651 for (i = 0; i < env->subprog_cnt; i++)
14652 if (env->subprog_info[i].start >= off)
14654 /* find first prog starting at or after off + cnt (first to stay) */
14655 for (j = i; j < env->subprog_cnt; j++)
14656 if (env->subprog_info[j].start >= off + cnt)
14658 /* if j doesn't start exactly at off + cnt, we are just removing
14659 * the front of previous prog
14661 if (env->subprog_info[j].start != off + cnt)
14665 struct bpf_prog_aux *aux = env->prog->aux;
14668 /* move fake 'exit' subprog as well */
14669 move = env->subprog_cnt + 1 - j;
14671 memmove(env->subprog_info + i,
14672 env->subprog_info + j,
14673 sizeof(*env->subprog_info) * move);
14674 env->subprog_cnt -= j - i;
14676 /* remove func_info */
14677 if (aux->func_info) {
14678 move = aux->func_info_cnt - j;
14680 memmove(aux->func_info + i,
14681 aux->func_info + j,
14682 sizeof(*aux->func_info) * move);
14683 aux->func_info_cnt -= j - i;
14684 /* func_info->insn_off is set after all code rewrites,
14685 * in adjust_btf_func() - no need to adjust
14689 /* convert i from "first prog to remove" to "first to adjust" */
14690 if (env->subprog_info[i].start == off)
14694 /* update fake 'exit' subprog as well */
14695 for (; i <= env->subprog_cnt; i++)
14696 env->subprog_info[i].start -= cnt;
14701 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
14704 struct bpf_prog *prog = env->prog;
14705 u32 i, l_off, l_cnt, nr_linfo;
14706 struct bpf_line_info *linfo;
14708 nr_linfo = prog->aux->nr_linfo;
14712 linfo = prog->aux->linfo;
14714 /* find first line info to remove, count lines to be removed */
14715 for (i = 0; i < nr_linfo; i++)
14716 if (linfo[i].insn_off >= off)
14721 for (; i < nr_linfo; i++)
14722 if (linfo[i].insn_off < off + cnt)
14727 /* First live insn doesn't match first live linfo, it needs to "inherit"
14728 * last removed linfo. prog is already modified, so prog->len == off
14729 * means no live instructions after (tail of the program was removed).
14731 if (prog->len != off && l_cnt &&
14732 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
14734 linfo[--i].insn_off = off + cnt;
14737 /* remove the line info which refer to the removed instructions */
14739 memmove(linfo + l_off, linfo + i,
14740 sizeof(*linfo) * (nr_linfo - i));
14742 prog->aux->nr_linfo -= l_cnt;
14743 nr_linfo = prog->aux->nr_linfo;
14746 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
14747 for (i = l_off; i < nr_linfo; i++)
14748 linfo[i].insn_off -= cnt;
14750 /* fix up all subprogs (incl. 'exit') which start >= off */
14751 for (i = 0; i <= env->subprog_cnt; i++)
14752 if (env->subprog_info[i].linfo_idx > l_off) {
14753 /* program may have started in the removed region but
14754 * may not be fully removed
14756 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
14757 env->subprog_info[i].linfo_idx -= l_cnt;
14759 env->subprog_info[i].linfo_idx = l_off;
14765 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
14767 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
14768 unsigned int orig_prog_len = env->prog->len;
14771 if (bpf_prog_is_dev_bound(env->prog->aux))
14772 bpf_prog_offload_remove_insns(env, off, cnt);
14774 err = bpf_remove_insns(env->prog, off, cnt);
14778 err = adjust_subprog_starts_after_remove(env, off, cnt);
14782 err = bpf_adj_linfo_after_remove(env, off, cnt);
14786 memmove(aux_data + off, aux_data + off + cnt,
14787 sizeof(*aux_data) * (orig_prog_len - off - cnt));
14792 /* The verifier does more data flow analysis than llvm and will not
14793 * explore branches that are dead at run time. Malicious programs can
14794 * have dead code too. Therefore replace all dead at-run-time code
14797 * Just nops are not optimal, e.g. if they would sit at the end of the
14798 * program and through another bug we would manage to jump there, then
14799 * we'd execute beyond program memory otherwise. Returning exception
14800 * code also wouldn't work since we can have subprogs where the dead
14801 * code could be located.
14803 static void sanitize_dead_code(struct bpf_verifier_env *env)
14805 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
14806 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
14807 struct bpf_insn *insn = env->prog->insnsi;
14808 const int insn_cnt = env->prog->len;
14811 for (i = 0; i < insn_cnt; i++) {
14812 if (aux_data[i].seen)
14814 memcpy(insn + i, &trap, sizeof(trap));
14815 aux_data[i].zext_dst = false;
14819 static bool insn_is_cond_jump(u8 code)
14823 if (BPF_CLASS(code) == BPF_JMP32)
14826 if (BPF_CLASS(code) != BPF_JMP)
14830 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
14833 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
14835 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
14836 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
14837 struct bpf_insn *insn = env->prog->insnsi;
14838 const int insn_cnt = env->prog->len;
14841 for (i = 0; i < insn_cnt; i++, insn++) {
14842 if (!insn_is_cond_jump(insn->code))
14845 if (!aux_data[i + 1].seen)
14846 ja.off = insn->off;
14847 else if (!aux_data[i + 1 + insn->off].seen)
14852 if (bpf_prog_is_dev_bound(env->prog->aux))
14853 bpf_prog_offload_replace_insn(env, i, &ja);
14855 memcpy(insn, &ja, sizeof(ja));
14859 static int opt_remove_dead_code(struct bpf_verifier_env *env)
14861 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
14862 int insn_cnt = env->prog->len;
14865 for (i = 0; i < insn_cnt; i++) {
14869 while (i + j < insn_cnt && !aux_data[i + j].seen)
14874 err = verifier_remove_insns(env, i, j);
14877 insn_cnt = env->prog->len;
14883 static int opt_remove_nops(struct bpf_verifier_env *env)
14885 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
14886 struct bpf_insn *insn = env->prog->insnsi;
14887 int insn_cnt = env->prog->len;
14890 for (i = 0; i < insn_cnt; i++) {
14891 if (memcmp(&insn[i], &ja, sizeof(ja)))
14894 err = verifier_remove_insns(env, i, 1);
14904 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
14905 const union bpf_attr *attr)
14907 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
14908 struct bpf_insn_aux_data *aux = env->insn_aux_data;
14909 int i, patch_len, delta = 0, len = env->prog->len;
14910 struct bpf_insn *insns = env->prog->insnsi;
14911 struct bpf_prog *new_prog;
14914 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
14915 zext_patch[1] = BPF_ZEXT_REG(0);
14916 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
14917 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
14918 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
14919 for (i = 0; i < len; i++) {
14920 int adj_idx = i + delta;
14921 struct bpf_insn insn;
14924 insn = insns[adj_idx];
14925 load_reg = insn_def_regno(&insn);
14926 if (!aux[adj_idx].zext_dst) {
14934 class = BPF_CLASS(code);
14935 if (load_reg == -1)
14938 /* NOTE: arg "reg" (the fourth one) is only used for
14939 * BPF_STX + SRC_OP, so it is safe to pass NULL
14942 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
14943 if (class == BPF_LD &&
14944 BPF_MODE(code) == BPF_IMM)
14949 /* ctx load could be transformed into wider load. */
14950 if (class == BPF_LDX &&
14951 aux[adj_idx].ptr_type == PTR_TO_CTX)
14954 imm_rnd = get_random_u32();
14955 rnd_hi32_patch[0] = insn;
14956 rnd_hi32_patch[1].imm = imm_rnd;
14957 rnd_hi32_patch[3].dst_reg = load_reg;
14958 patch = rnd_hi32_patch;
14960 goto apply_patch_buffer;
14963 /* Add in an zero-extend instruction if a) the JIT has requested
14964 * it or b) it's a CMPXCHG.
14966 * The latter is because: BPF_CMPXCHG always loads a value into
14967 * R0, therefore always zero-extends. However some archs'
14968 * equivalent instruction only does this load when the
14969 * comparison is successful. This detail of CMPXCHG is
14970 * orthogonal to the general zero-extension behaviour of the
14971 * CPU, so it's treated independently of bpf_jit_needs_zext.
14973 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
14976 /* Zero-extension is done by the caller. */
14977 if (bpf_pseudo_kfunc_call(&insn))
14980 if (WARN_ON(load_reg == -1)) {
14981 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
14985 zext_patch[0] = insn;
14986 zext_patch[1].dst_reg = load_reg;
14987 zext_patch[1].src_reg = load_reg;
14988 patch = zext_patch;
14990 apply_patch_buffer:
14991 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
14994 env->prog = new_prog;
14995 insns = new_prog->insnsi;
14996 aux = env->insn_aux_data;
14997 delta += patch_len - 1;
15003 /* convert load instructions that access fields of a context type into a
15004 * sequence of instructions that access fields of the underlying structure:
15005 * struct __sk_buff -> struct sk_buff
15006 * struct bpf_sock_ops -> struct sock
15008 static int convert_ctx_accesses(struct bpf_verifier_env *env)
15010 const struct bpf_verifier_ops *ops = env->ops;
15011 int i, cnt, size, ctx_field_size, delta = 0;
15012 const int insn_cnt = env->prog->len;
15013 struct bpf_insn insn_buf[16], *insn;
15014 u32 target_size, size_default, off;
15015 struct bpf_prog *new_prog;
15016 enum bpf_access_type type;
15017 bool is_narrower_load;
15019 if (ops->gen_prologue || env->seen_direct_write) {
15020 if (!ops->gen_prologue) {
15021 verbose(env, "bpf verifier is misconfigured\n");
15024 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
15026 if (cnt >= ARRAY_SIZE(insn_buf)) {
15027 verbose(env, "bpf verifier is misconfigured\n");
15030 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
15034 env->prog = new_prog;
15039 if (bpf_prog_is_dev_bound(env->prog->aux))
15042 insn = env->prog->insnsi + delta;
15044 for (i = 0; i < insn_cnt; i++, insn++) {
15045 bpf_convert_ctx_access_t convert_ctx_access;
15048 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
15049 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
15050 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
15051 insn->code == (BPF_LDX | BPF_MEM | BPF_DW)) {
15054 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
15055 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
15056 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
15057 insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
15058 insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
15059 insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
15060 insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
15061 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
15063 ctx_access = BPF_CLASS(insn->code) == BPF_STX;
15068 if (type == BPF_WRITE &&
15069 env->insn_aux_data[i + delta].sanitize_stack_spill) {
15070 struct bpf_insn patch[] = {
15075 cnt = ARRAY_SIZE(patch);
15076 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
15081 env->prog = new_prog;
15082 insn = new_prog->insnsi + i + delta;
15089 switch ((int)env->insn_aux_data[i + delta].ptr_type) {
15091 if (!ops->convert_ctx_access)
15093 convert_ctx_access = ops->convert_ctx_access;
15095 case PTR_TO_SOCKET:
15096 case PTR_TO_SOCK_COMMON:
15097 convert_ctx_access = bpf_sock_convert_ctx_access;
15099 case PTR_TO_TCP_SOCK:
15100 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
15102 case PTR_TO_XDP_SOCK:
15103 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
15105 case PTR_TO_BTF_ID:
15106 case PTR_TO_BTF_ID | PTR_UNTRUSTED:
15107 /* PTR_TO_BTF_ID | MEM_ALLOC always has a valid lifetime, unlike
15108 * PTR_TO_BTF_ID, and an active ref_obj_id, but the same cannot
15109 * be said once it is marked PTR_UNTRUSTED, hence we must handle
15110 * any faults for loads into such types. BPF_WRITE is disallowed
15113 case PTR_TO_BTF_ID | MEM_ALLOC | PTR_UNTRUSTED:
15114 if (type == BPF_READ) {
15115 insn->code = BPF_LDX | BPF_PROBE_MEM |
15116 BPF_SIZE((insn)->code);
15117 env->prog->aux->num_exentries++;
15124 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
15125 size = BPF_LDST_BYTES(insn);
15127 /* If the read access is a narrower load of the field,
15128 * convert to a 4/8-byte load, to minimum program type specific
15129 * convert_ctx_access changes. If conversion is successful,
15130 * we will apply proper mask to the result.
15132 is_narrower_load = size < ctx_field_size;
15133 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
15135 if (is_narrower_load) {
15138 if (type == BPF_WRITE) {
15139 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
15144 if (ctx_field_size == 4)
15146 else if (ctx_field_size == 8)
15147 size_code = BPF_DW;
15149 insn->off = off & ~(size_default - 1);
15150 insn->code = BPF_LDX | BPF_MEM | size_code;
15154 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
15156 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
15157 (ctx_field_size && !target_size)) {
15158 verbose(env, "bpf verifier is misconfigured\n");
15162 if (is_narrower_load && size < target_size) {
15163 u8 shift = bpf_ctx_narrow_access_offset(
15164 off, size, size_default) * 8;
15165 if (shift && cnt + 1 >= ARRAY_SIZE(insn_buf)) {
15166 verbose(env, "bpf verifier narrow ctx load misconfigured\n");
15169 if (ctx_field_size <= 4) {
15171 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
15174 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
15175 (1 << size * 8) - 1);
15178 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
15181 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
15182 (1ULL << size * 8) - 1);
15186 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
15192 /* keep walking new program and skip insns we just inserted */
15193 env->prog = new_prog;
15194 insn = new_prog->insnsi + i + delta;
15200 static int jit_subprogs(struct bpf_verifier_env *env)
15202 struct bpf_prog *prog = env->prog, **func, *tmp;
15203 int i, j, subprog_start, subprog_end = 0, len, subprog;
15204 struct bpf_map *map_ptr;
15205 struct bpf_insn *insn;
15206 void *old_bpf_func;
15207 int err, num_exentries;
15209 if (env->subprog_cnt <= 1)
15212 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
15213 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn))
15216 /* Upon error here we cannot fall back to interpreter but
15217 * need a hard reject of the program. Thus -EFAULT is
15218 * propagated in any case.
15220 subprog = find_subprog(env, i + insn->imm + 1);
15222 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
15223 i + insn->imm + 1);
15226 /* temporarily remember subprog id inside insn instead of
15227 * aux_data, since next loop will split up all insns into funcs
15229 insn->off = subprog;
15230 /* remember original imm in case JIT fails and fallback
15231 * to interpreter will be needed
15233 env->insn_aux_data[i].call_imm = insn->imm;
15234 /* point imm to __bpf_call_base+1 from JITs point of view */
15236 if (bpf_pseudo_func(insn))
15237 /* jit (e.g. x86_64) may emit fewer instructions
15238 * if it learns a u32 imm is the same as a u64 imm.
15239 * Force a non zero here.
15244 err = bpf_prog_alloc_jited_linfo(prog);
15246 goto out_undo_insn;
15249 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
15251 goto out_undo_insn;
15253 for (i = 0; i < env->subprog_cnt; i++) {
15254 subprog_start = subprog_end;
15255 subprog_end = env->subprog_info[i + 1].start;
15257 len = subprog_end - subprog_start;
15258 /* bpf_prog_run() doesn't call subprogs directly,
15259 * hence main prog stats include the runtime of subprogs.
15260 * subprogs don't have IDs and not reachable via prog_get_next_id
15261 * func[i]->stats will never be accessed and stays NULL
15263 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
15266 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
15267 len * sizeof(struct bpf_insn));
15268 func[i]->type = prog->type;
15269 func[i]->len = len;
15270 if (bpf_prog_calc_tag(func[i]))
15272 func[i]->is_func = 1;
15273 func[i]->aux->func_idx = i;
15274 /* Below members will be freed only at prog->aux */
15275 func[i]->aux->btf = prog->aux->btf;
15276 func[i]->aux->func_info = prog->aux->func_info;
15277 func[i]->aux->func_info_cnt = prog->aux->func_info_cnt;
15278 func[i]->aux->poke_tab = prog->aux->poke_tab;
15279 func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
15281 for (j = 0; j < prog->aux->size_poke_tab; j++) {
15282 struct bpf_jit_poke_descriptor *poke;
15284 poke = &prog->aux->poke_tab[j];
15285 if (poke->insn_idx < subprog_end &&
15286 poke->insn_idx >= subprog_start)
15287 poke->aux = func[i]->aux;
15290 func[i]->aux->name[0] = 'F';
15291 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
15292 func[i]->jit_requested = 1;
15293 func[i]->blinding_requested = prog->blinding_requested;
15294 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
15295 func[i]->aux->kfunc_btf_tab = prog->aux->kfunc_btf_tab;
15296 func[i]->aux->linfo = prog->aux->linfo;
15297 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
15298 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
15299 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
15301 insn = func[i]->insnsi;
15302 for (j = 0; j < func[i]->len; j++, insn++) {
15303 if (BPF_CLASS(insn->code) == BPF_LDX &&
15304 BPF_MODE(insn->code) == BPF_PROBE_MEM)
15307 func[i]->aux->num_exentries = num_exentries;
15308 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
15309 func[i] = bpf_int_jit_compile(func[i]);
15310 if (!func[i]->jited) {
15317 /* at this point all bpf functions were successfully JITed
15318 * now populate all bpf_calls with correct addresses and
15319 * run last pass of JIT
15321 for (i = 0; i < env->subprog_cnt; i++) {
15322 insn = func[i]->insnsi;
15323 for (j = 0; j < func[i]->len; j++, insn++) {
15324 if (bpf_pseudo_func(insn)) {
15325 subprog = insn->off;
15326 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
15327 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
15330 if (!bpf_pseudo_call(insn))
15332 subprog = insn->off;
15333 insn->imm = BPF_CALL_IMM(func[subprog]->bpf_func);
15336 /* we use the aux data to keep a list of the start addresses
15337 * of the JITed images for each function in the program
15339 * for some architectures, such as powerpc64, the imm field
15340 * might not be large enough to hold the offset of the start
15341 * address of the callee's JITed image from __bpf_call_base
15343 * in such cases, we can lookup the start address of a callee
15344 * by using its subprog id, available from the off field of
15345 * the call instruction, as an index for this list
15347 func[i]->aux->func = func;
15348 func[i]->aux->func_cnt = env->subprog_cnt;
15350 for (i = 0; i < env->subprog_cnt; i++) {
15351 old_bpf_func = func[i]->bpf_func;
15352 tmp = bpf_int_jit_compile(func[i]);
15353 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
15354 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
15361 /* finally lock prog and jit images for all functions and
15362 * populate kallsysm
15364 for (i = 0; i < env->subprog_cnt; i++) {
15365 bpf_prog_lock_ro(func[i]);
15366 bpf_prog_kallsyms_add(func[i]);
15369 /* Last step: make now unused interpreter insns from main
15370 * prog consistent for later dump requests, so they can
15371 * later look the same as if they were interpreted only.
15373 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
15374 if (bpf_pseudo_func(insn)) {
15375 insn[0].imm = env->insn_aux_data[i].call_imm;
15376 insn[1].imm = insn->off;
15380 if (!bpf_pseudo_call(insn))
15382 insn->off = env->insn_aux_data[i].call_imm;
15383 subprog = find_subprog(env, i + insn->off + 1);
15384 insn->imm = subprog;
15388 prog->bpf_func = func[0]->bpf_func;
15389 prog->jited_len = func[0]->jited_len;
15390 prog->aux->func = func;
15391 prog->aux->func_cnt = env->subprog_cnt;
15392 bpf_prog_jit_attempt_done(prog);
15395 /* We failed JIT'ing, so at this point we need to unregister poke
15396 * descriptors from subprogs, so that kernel is not attempting to
15397 * patch it anymore as we're freeing the subprog JIT memory.
15399 for (i = 0; i < prog->aux->size_poke_tab; i++) {
15400 map_ptr = prog->aux->poke_tab[i].tail_call.map;
15401 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
15403 /* At this point we're guaranteed that poke descriptors are not
15404 * live anymore. We can just unlink its descriptor table as it's
15405 * released with the main prog.
15407 for (i = 0; i < env->subprog_cnt; i++) {
15410 func[i]->aux->poke_tab = NULL;
15411 bpf_jit_free(func[i]);
15415 /* cleanup main prog to be interpreted */
15416 prog->jit_requested = 0;
15417 prog->blinding_requested = 0;
15418 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
15419 if (!bpf_pseudo_call(insn))
15422 insn->imm = env->insn_aux_data[i].call_imm;
15424 bpf_prog_jit_attempt_done(prog);
15428 static int fixup_call_args(struct bpf_verifier_env *env)
15430 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
15431 struct bpf_prog *prog = env->prog;
15432 struct bpf_insn *insn = prog->insnsi;
15433 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
15438 if (env->prog->jit_requested &&
15439 !bpf_prog_is_dev_bound(env->prog->aux)) {
15440 err = jit_subprogs(env);
15443 if (err == -EFAULT)
15446 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
15447 if (has_kfunc_call) {
15448 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
15451 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
15452 /* When JIT fails the progs with bpf2bpf calls and tail_calls
15453 * have to be rejected, since interpreter doesn't support them yet.
15455 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
15458 for (i = 0; i < prog->len; i++, insn++) {
15459 if (bpf_pseudo_func(insn)) {
15460 /* When JIT fails the progs with callback calls
15461 * have to be rejected, since interpreter doesn't support them yet.
15463 verbose(env, "callbacks are not allowed in non-JITed programs\n");
15467 if (!bpf_pseudo_call(insn))
15469 depth = get_callee_stack_depth(env, insn, i);
15472 bpf_patch_call_args(insn, depth);
15479 static int fixup_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
15480 struct bpf_insn *insn_buf, int insn_idx, int *cnt)
15482 const struct bpf_kfunc_desc *desc;
15485 verbose(env, "invalid kernel function call not eliminated in verifier pass\n");
15489 /* insn->imm has the btf func_id. Replace it with
15490 * an address (relative to __bpf_call_base).
15492 desc = find_kfunc_desc(env->prog, insn->imm, insn->off);
15494 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
15500 insn->imm = desc->imm;
15503 if (desc->func_id == special_kfunc_list[KF_bpf_obj_new_impl]) {
15504 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
15505 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
15506 u64 obj_new_size = env->insn_aux_data[insn_idx].obj_new_size;
15508 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_1, obj_new_size);
15509 insn_buf[1] = addr[0];
15510 insn_buf[2] = addr[1];
15511 insn_buf[3] = *insn;
15513 } else if (desc->func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
15514 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
15515 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
15517 insn_buf[0] = addr[0];
15518 insn_buf[1] = addr[1];
15519 insn_buf[2] = *insn;
15521 } else if (desc->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx] ||
15522 desc->func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
15523 insn_buf[0] = BPF_MOV64_REG(BPF_REG_0, BPF_REG_1);
15529 /* Do various post-verification rewrites in a single program pass.
15530 * These rewrites simplify JIT and interpreter implementations.
15532 static int do_misc_fixups(struct bpf_verifier_env *env)
15534 struct bpf_prog *prog = env->prog;
15535 enum bpf_attach_type eatype = prog->expected_attach_type;
15536 enum bpf_prog_type prog_type = resolve_prog_type(prog);
15537 struct bpf_insn *insn = prog->insnsi;
15538 const struct bpf_func_proto *fn;
15539 const int insn_cnt = prog->len;
15540 const struct bpf_map_ops *ops;
15541 struct bpf_insn_aux_data *aux;
15542 struct bpf_insn insn_buf[16];
15543 struct bpf_prog *new_prog;
15544 struct bpf_map *map_ptr;
15545 int i, ret, cnt, delta = 0;
15547 for (i = 0; i < insn_cnt; i++, insn++) {
15548 /* Make divide-by-zero exceptions impossible. */
15549 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
15550 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
15551 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
15552 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
15553 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
15554 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
15555 struct bpf_insn *patchlet;
15556 struct bpf_insn chk_and_div[] = {
15557 /* [R,W]x div 0 -> 0 */
15558 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
15559 BPF_JNE | BPF_K, insn->src_reg,
15561 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
15562 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
15565 struct bpf_insn chk_and_mod[] = {
15566 /* [R,W]x mod 0 -> [R,W]x */
15567 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
15568 BPF_JEQ | BPF_K, insn->src_reg,
15569 0, 1 + (is64 ? 0 : 1), 0),
15571 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
15572 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
15575 patchlet = isdiv ? chk_and_div : chk_and_mod;
15576 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
15577 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
15579 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
15584 env->prog = prog = new_prog;
15585 insn = new_prog->insnsi + i + delta;
15589 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
15590 if (BPF_CLASS(insn->code) == BPF_LD &&
15591 (BPF_MODE(insn->code) == BPF_ABS ||
15592 BPF_MODE(insn->code) == BPF_IND)) {
15593 cnt = env->ops->gen_ld_abs(insn, insn_buf);
15594 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
15595 verbose(env, "bpf verifier is misconfigured\n");
15599 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
15604 env->prog = prog = new_prog;
15605 insn = new_prog->insnsi + i + delta;
15609 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
15610 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
15611 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
15612 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
15613 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
15614 struct bpf_insn *patch = &insn_buf[0];
15615 bool issrc, isneg, isimm;
15618 aux = &env->insn_aux_data[i + delta];
15619 if (!aux->alu_state ||
15620 aux->alu_state == BPF_ALU_NON_POINTER)
15623 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
15624 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
15625 BPF_ALU_SANITIZE_SRC;
15626 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
15628 off_reg = issrc ? insn->src_reg : insn->dst_reg;
15630 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
15633 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
15634 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
15635 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
15636 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
15637 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
15638 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
15639 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
15642 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
15643 insn->src_reg = BPF_REG_AX;
15645 insn->code = insn->code == code_add ?
15646 code_sub : code_add;
15648 if (issrc && isneg && !isimm)
15649 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
15650 cnt = patch - insn_buf;
15652 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
15657 env->prog = prog = new_prog;
15658 insn = new_prog->insnsi + i + delta;
15662 if (insn->code != (BPF_JMP | BPF_CALL))
15664 if (insn->src_reg == BPF_PSEUDO_CALL)
15666 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
15667 ret = fixup_kfunc_call(env, insn, insn_buf, i + delta, &cnt);
15673 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
15678 env->prog = prog = new_prog;
15679 insn = new_prog->insnsi + i + delta;
15683 if (insn->imm == BPF_FUNC_get_route_realm)
15684 prog->dst_needed = 1;
15685 if (insn->imm == BPF_FUNC_get_prandom_u32)
15686 bpf_user_rnd_init_once();
15687 if (insn->imm == BPF_FUNC_override_return)
15688 prog->kprobe_override = 1;
15689 if (insn->imm == BPF_FUNC_tail_call) {
15690 /* If we tail call into other programs, we
15691 * cannot make any assumptions since they can
15692 * be replaced dynamically during runtime in
15693 * the program array.
15695 prog->cb_access = 1;
15696 if (!allow_tail_call_in_subprogs(env))
15697 prog->aux->stack_depth = MAX_BPF_STACK;
15698 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
15700 /* mark bpf_tail_call as different opcode to avoid
15701 * conditional branch in the interpreter for every normal
15702 * call and to prevent accidental JITing by JIT compiler
15703 * that doesn't support bpf_tail_call yet
15706 insn->code = BPF_JMP | BPF_TAIL_CALL;
15708 aux = &env->insn_aux_data[i + delta];
15709 if (env->bpf_capable && !prog->blinding_requested &&
15710 prog->jit_requested &&
15711 !bpf_map_key_poisoned(aux) &&
15712 !bpf_map_ptr_poisoned(aux) &&
15713 !bpf_map_ptr_unpriv(aux)) {
15714 struct bpf_jit_poke_descriptor desc = {
15715 .reason = BPF_POKE_REASON_TAIL_CALL,
15716 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
15717 .tail_call.key = bpf_map_key_immediate(aux),
15718 .insn_idx = i + delta,
15721 ret = bpf_jit_add_poke_descriptor(prog, &desc);
15723 verbose(env, "adding tail call poke descriptor failed\n");
15727 insn->imm = ret + 1;
15731 if (!bpf_map_ptr_unpriv(aux))
15734 /* instead of changing every JIT dealing with tail_call
15735 * emit two extra insns:
15736 * if (index >= max_entries) goto out;
15737 * index &= array->index_mask;
15738 * to avoid out-of-bounds cpu speculation
15740 if (bpf_map_ptr_poisoned(aux)) {
15741 verbose(env, "tail_call abusing map_ptr\n");
15745 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
15746 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
15747 map_ptr->max_entries, 2);
15748 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
15749 container_of(map_ptr,
15752 insn_buf[2] = *insn;
15754 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
15759 env->prog = prog = new_prog;
15760 insn = new_prog->insnsi + i + delta;
15764 if (insn->imm == BPF_FUNC_timer_set_callback) {
15765 /* The verifier will process callback_fn as many times as necessary
15766 * with different maps and the register states prepared by
15767 * set_timer_callback_state will be accurate.
15769 * The following use case is valid:
15770 * map1 is shared by prog1, prog2, prog3.
15771 * prog1 calls bpf_timer_init for some map1 elements
15772 * prog2 calls bpf_timer_set_callback for some map1 elements.
15773 * Those that were not bpf_timer_init-ed will return -EINVAL.
15774 * prog3 calls bpf_timer_start for some map1 elements.
15775 * Those that were not both bpf_timer_init-ed and
15776 * bpf_timer_set_callback-ed will return -EINVAL.
15778 struct bpf_insn ld_addrs[2] = {
15779 BPF_LD_IMM64(BPF_REG_3, (long)prog->aux),
15782 insn_buf[0] = ld_addrs[0];
15783 insn_buf[1] = ld_addrs[1];
15784 insn_buf[2] = *insn;
15787 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
15792 env->prog = prog = new_prog;
15793 insn = new_prog->insnsi + i + delta;
15794 goto patch_call_imm;
15797 if (is_storage_get_function(insn->imm)) {
15798 if (!env->prog->aux->sleepable ||
15799 env->insn_aux_data[i + delta].storage_get_func_atomic)
15800 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_ATOMIC);
15802 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_KERNEL);
15803 insn_buf[1] = *insn;
15806 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
15811 env->prog = prog = new_prog;
15812 insn = new_prog->insnsi + i + delta;
15813 goto patch_call_imm;
15816 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
15817 * and other inlining handlers are currently limited to 64 bit
15820 if (prog->jit_requested && BITS_PER_LONG == 64 &&
15821 (insn->imm == BPF_FUNC_map_lookup_elem ||
15822 insn->imm == BPF_FUNC_map_update_elem ||
15823 insn->imm == BPF_FUNC_map_delete_elem ||
15824 insn->imm == BPF_FUNC_map_push_elem ||
15825 insn->imm == BPF_FUNC_map_pop_elem ||
15826 insn->imm == BPF_FUNC_map_peek_elem ||
15827 insn->imm == BPF_FUNC_redirect_map ||
15828 insn->imm == BPF_FUNC_for_each_map_elem ||
15829 insn->imm == BPF_FUNC_map_lookup_percpu_elem)) {
15830 aux = &env->insn_aux_data[i + delta];
15831 if (bpf_map_ptr_poisoned(aux))
15832 goto patch_call_imm;
15834 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
15835 ops = map_ptr->ops;
15836 if (insn->imm == BPF_FUNC_map_lookup_elem &&
15837 ops->map_gen_lookup) {
15838 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
15839 if (cnt == -EOPNOTSUPP)
15840 goto patch_map_ops_generic;
15841 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
15842 verbose(env, "bpf verifier is misconfigured\n");
15846 new_prog = bpf_patch_insn_data(env, i + delta,
15852 env->prog = prog = new_prog;
15853 insn = new_prog->insnsi + i + delta;
15857 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
15858 (void *(*)(struct bpf_map *map, void *key))NULL));
15859 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
15860 (int (*)(struct bpf_map *map, void *key))NULL));
15861 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
15862 (int (*)(struct bpf_map *map, void *key, void *value,
15864 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
15865 (int (*)(struct bpf_map *map, void *value,
15867 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
15868 (int (*)(struct bpf_map *map, void *value))NULL));
15869 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
15870 (int (*)(struct bpf_map *map, void *value))NULL));
15871 BUILD_BUG_ON(!__same_type(ops->map_redirect,
15872 (int (*)(struct bpf_map *map, u64 index, u64 flags))NULL));
15873 BUILD_BUG_ON(!__same_type(ops->map_for_each_callback,
15874 (int (*)(struct bpf_map *map,
15875 bpf_callback_t callback_fn,
15876 void *callback_ctx,
15878 BUILD_BUG_ON(!__same_type(ops->map_lookup_percpu_elem,
15879 (void *(*)(struct bpf_map *map, void *key, u32 cpu))NULL));
15881 patch_map_ops_generic:
15882 switch (insn->imm) {
15883 case BPF_FUNC_map_lookup_elem:
15884 insn->imm = BPF_CALL_IMM(ops->map_lookup_elem);
15886 case BPF_FUNC_map_update_elem:
15887 insn->imm = BPF_CALL_IMM(ops->map_update_elem);
15889 case BPF_FUNC_map_delete_elem:
15890 insn->imm = BPF_CALL_IMM(ops->map_delete_elem);
15892 case BPF_FUNC_map_push_elem:
15893 insn->imm = BPF_CALL_IMM(ops->map_push_elem);
15895 case BPF_FUNC_map_pop_elem:
15896 insn->imm = BPF_CALL_IMM(ops->map_pop_elem);
15898 case BPF_FUNC_map_peek_elem:
15899 insn->imm = BPF_CALL_IMM(ops->map_peek_elem);
15901 case BPF_FUNC_redirect_map:
15902 insn->imm = BPF_CALL_IMM(ops->map_redirect);
15904 case BPF_FUNC_for_each_map_elem:
15905 insn->imm = BPF_CALL_IMM(ops->map_for_each_callback);
15907 case BPF_FUNC_map_lookup_percpu_elem:
15908 insn->imm = BPF_CALL_IMM(ops->map_lookup_percpu_elem);
15912 goto patch_call_imm;
15915 /* Implement bpf_jiffies64 inline. */
15916 if (prog->jit_requested && BITS_PER_LONG == 64 &&
15917 insn->imm == BPF_FUNC_jiffies64) {
15918 struct bpf_insn ld_jiffies_addr[2] = {
15919 BPF_LD_IMM64(BPF_REG_0,
15920 (unsigned long)&jiffies),
15923 insn_buf[0] = ld_jiffies_addr[0];
15924 insn_buf[1] = ld_jiffies_addr[1];
15925 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
15929 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
15935 env->prog = prog = new_prog;
15936 insn = new_prog->insnsi + i + delta;
15940 /* Implement bpf_get_func_arg inline. */
15941 if (prog_type == BPF_PROG_TYPE_TRACING &&
15942 insn->imm == BPF_FUNC_get_func_arg) {
15943 /* Load nr_args from ctx - 8 */
15944 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
15945 insn_buf[1] = BPF_JMP32_REG(BPF_JGE, BPF_REG_2, BPF_REG_0, 6);
15946 insn_buf[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_2, 3);
15947 insn_buf[3] = BPF_ALU64_REG(BPF_ADD, BPF_REG_2, BPF_REG_1);
15948 insn_buf[4] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_2, 0);
15949 insn_buf[5] = BPF_STX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
15950 insn_buf[6] = BPF_MOV64_IMM(BPF_REG_0, 0);
15951 insn_buf[7] = BPF_JMP_A(1);
15952 insn_buf[8] = BPF_MOV64_IMM(BPF_REG_0, -EINVAL);
15955 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
15960 env->prog = prog = new_prog;
15961 insn = new_prog->insnsi + i + delta;
15965 /* Implement bpf_get_func_ret inline. */
15966 if (prog_type == BPF_PROG_TYPE_TRACING &&
15967 insn->imm == BPF_FUNC_get_func_ret) {
15968 if (eatype == BPF_TRACE_FEXIT ||
15969 eatype == BPF_MODIFY_RETURN) {
15970 /* Load nr_args from ctx - 8 */
15971 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
15972 insn_buf[1] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_0, 3);
15973 insn_buf[2] = BPF_ALU64_REG(BPF_ADD, BPF_REG_0, BPF_REG_1);
15974 insn_buf[3] = BPF_LDX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
15975 insn_buf[4] = BPF_STX_MEM(BPF_DW, BPF_REG_2, BPF_REG_3, 0);
15976 insn_buf[5] = BPF_MOV64_IMM(BPF_REG_0, 0);
15979 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, -EOPNOTSUPP);
15983 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
15988 env->prog = prog = new_prog;
15989 insn = new_prog->insnsi + i + delta;
15993 /* Implement get_func_arg_cnt inline. */
15994 if (prog_type == BPF_PROG_TYPE_TRACING &&
15995 insn->imm == BPF_FUNC_get_func_arg_cnt) {
15996 /* Load nr_args from ctx - 8 */
15997 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
15999 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
16003 env->prog = prog = new_prog;
16004 insn = new_prog->insnsi + i + delta;
16008 /* Implement bpf_get_func_ip inline. */
16009 if (prog_type == BPF_PROG_TYPE_TRACING &&
16010 insn->imm == BPF_FUNC_get_func_ip) {
16011 /* Load IP address from ctx - 16 */
16012 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -16);
16014 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
16018 env->prog = prog = new_prog;
16019 insn = new_prog->insnsi + i + delta;
16024 fn = env->ops->get_func_proto(insn->imm, env->prog);
16025 /* all functions that have prototype and verifier allowed
16026 * programs to call them, must be real in-kernel functions
16030 "kernel subsystem misconfigured func %s#%d\n",
16031 func_id_name(insn->imm), insn->imm);
16034 insn->imm = fn->func - __bpf_call_base;
16037 /* Since poke tab is now finalized, publish aux to tracker. */
16038 for (i = 0; i < prog->aux->size_poke_tab; i++) {
16039 map_ptr = prog->aux->poke_tab[i].tail_call.map;
16040 if (!map_ptr->ops->map_poke_track ||
16041 !map_ptr->ops->map_poke_untrack ||
16042 !map_ptr->ops->map_poke_run) {
16043 verbose(env, "bpf verifier is misconfigured\n");
16047 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
16049 verbose(env, "tracking tail call prog failed\n");
16054 sort_kfunc_descs_by_imm(env->prog);
16059 static struct bpf_prog *inline_bpf_loop(struct bpf_verifier_env *env,
16062 u32 callback_subprogno,
16065 s32 r6_offset = stack_base + 0 * BPF_REG_SIZE;
16066 s32 r7_offset = stack_base + 1 * BPF_REG_SIZE;
16067 s32 r8_offset = stack_base + 2 * BPF_REG_SIZE;
16068 int reg_loop_max = BPF_REG_6;
16069 int reg_loop_cnt = BPF_REG_7;
16070 int reg_loop_ctx = BPF_REG_8;
16072 struct bpf_prog *new_prog;
16073 u32 callback_start;
16074 u32 call_insn_offset;
16075 s32 callback_offset;
16077 /* This represents an inlined version of bpf_iter.c:bpf_loop,
16078 * be careful to modify this code in sync.
16080 struct bpf_insn insn_buf[] = {
16081 /* Return error and jump to the end of the patch if
16082 * expected number of iterations is too big.
16084 BPF_JMP_IMM(BPF_JLE, BPF_REG_1, BPF_MAX_LOOPS, 2),
16085 BPF_MOV32_IMM(BPF_REG_0, -E2BIG),
16086 BPF_JMP_IMM(BPF_JA, 0, 0, 16),
16087 /* spill R6, R7, R8 to use these as loop vars */
16088 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_6, r6_offset),
16089 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_7, r7_offset),
16090 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_8, r8_offset),
16091 /* initialize loop vars */
16092 BPF_MOV64_REG(reg_loop_max, BPF_REG_1),
16093 BPF_MOV32_IMM(reg_loop_cnt, 0),
16094 BPF_MOV64_REG(reg_loop_ctx, BPF_REG_3),
16096 * if reg_loop_cnt >= reg_loop_max skip the loop body
16098 BPF_JMP_REG(BPF_JGE, reg_loop_cnt, reg_loop_max, 5),
16100 * correct callback offset would be set after patching
16102 BPF_MOV64_REG(BPF_REG_1, reg_loop_cnt),
16103 BPF_MOV64_REG(BPF_REG_2, reg_loop_ctx),
16105 /* increment loop counter */
16106 BPF_ALU64_IMM(BPF_ADD, reg_loop_cnt, 1),
16107 /* jump to loop header if callback returned 0 */
16108 BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, -6),
16109 /* return value of bpf_loop,
16110 * set R0 to the number of iterations
16112 BPF_MOV64_REG(BPF_REG_0, reg_loop_cnt),
16113 /* restore original values of R6, R7, R8 */
16114 BPF_LDX_MEM(BPF_DW, BPF_REG_6, BPF_REG_10, r6_offset),
16115 BPF_LDX_MEM(BPF_DW, BPF_REG_7, BPF_REG_10, r7_offset),
16116 BPF_LDX_MEM(BPF_DW, BPF_REG_8, BPF_REG_10, r8_offset),
16119 *cnt = ARRAY_SIZE(insn_buf);
16120 new_prog = bpf_patch_insn_data(env, position, insn_buf, *cnt);
16124 /* callback start is known only after patching */
16125 callback_start = env->subprog_info[callback_subprogno].start;
16126 /* Note: insn_buf[12] is an offset of BPF_CALL_REL instruction */
16127 call_insn_offset = position + 12;
16128 callback_offset = callback_start - call_insn_offset - 1;
16129 new_prog->insnsi[call_insn_offset].imm = callback_offset;
16134 static bool is_bpf_loop_call(struct bpf_insn *insn)
16136 return insn->code == (BPF_JMP | BPF_CALL) &&
16137 insn->src_reg == 0 &&
16138 insn->imm == BPF_FUNC_loop;
16141 /* For all sub-programs in the program (including main) check
16142 * insn_aux_data to see if there are bpf_loop calls that require
16143 * inlining. If such calls are found the calls are replaced with a
16144 * sequence of instructions produced by `inline_bpf_loop` function and
16145 * subprog stack_depth is increased by the size of 3 registers.
16146 * This stack space is used to spill values of the R6, R7, R8. These
16147 * registers are used to store the loop bound, counter and context
16150 static int optimize_bpf_loop(struct bpf_verifier_env *env)
16152 struct bpf_subprog_info *subprogs = env->subprog_info;
16153 int i, cur_subprog = 0, cnt, delta = 0;
16154 struct bpf_insn *insn = env->prog->insnsi;
16155 int insn_cnt = env->prog->len;
16156 u16 stack_depth = subprogs[cur_subprog].stack_depth;
16157 u16 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
16158 u16 stack_depth_extra = 0;
16160 for (i = 0; i < insn_cnt; i++, insn++) {
16161 struct bpf_loop_inline_state *inline_state =
16162 &env->insn_aux_data[i + delta].loop_inline_state;
16164 if (is_bpf_loop_call(insn) && inline_state->fit_for_inline) {
16165 struct bpf_prog *new_prog;
16167 stack_depth_extra = BPF_REG_SIZE * 3 + stack_depth_roundup;
16168 new_prog = inline_bpf_loop(env,
16170 -(stack_depth + stack_depth_extra),
16171 inline_state->callback_subprogno,
16177 env->prog = new_prog;
16178 insn = new_prog->insnsi + i + delta;
16181 if (subprogs[cur_subprog + 1].start == i + delta + 1) {
16182 subprogs[cur_subprog].stack_depth += stack_depth_extra;
16184 stack_depth = subprogs[cur_subprog].stack_depth;
16185 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
16186 stack_depth_extra = 0;
16190 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
16195 static void free_states(struct bpf_verifier_env *env)
16197 struct bpf_verifier_state_list *sl, *sln;
16200 sl = env->free_list;
16203 free_verifier_state(&sl->state, false);
16207 env->free_list = NULL;
16209 if (!env->explored_states)
16212 for (i = 0; i < state_htab_size(env); i++) {
16213 sl = env->explored_states[i];
16217 free_verifier_state(&sl->state, false);
16221 env->explored_states[i] = NULL;
16225 static int do_check_common(struct bpf_verifier_env *env, int subprog)
16227 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
16228 struct bpf_verifier_state *state;
16229 struct bpf_reg_state *regs;
16232 env->prev_linfo = NULL;
16235 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
16238 state->curframe = 0;
16239 state->speculative = false;
16240 state->branches = 1;
16241 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
16242 if (!state->frame[0]) {
16246 env->cur_state = state;
16247 init_func_state(env, state->frame[0],
16248 BPF_MAIN_FUNC /* callsite */,
16251 state->first_insn_idx = env->subprog_info[subprog].start;
16252 state->last_insn_idx = -1;
16254 regs = state->frame[state->curframe]->regs;
16255 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
16256 ret = btf_prepare_func_args(env, subprog, regs);
16259 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
16260 if (regs[i].type == PTR_TO_CTX)
16261 mark_reg_known_zero(env, regs, i);
16262 else if (regs[i].type == SCALAR_VALUE)
16263 mark_reg_unknown(env, regs, i);
16264 else if (base_type(regs[i].type) == PTR_TO_MEM) {
16265 const u32 mem_size = regs[i].mem_size;
16267 mark_reg_known_zero(env, regs, i);
16268 regs[i].mem_size = mem_size;
16269 regs[i].id = ++env->id_gen;
16273 /* 1st arg to a function */
16274 regs[BPF_REG_1].type = PTR_TO_CTX;
16275 mark_reg_known_zero(env, regs, BPF_REG_1);
16276 ret = btf_check_subprog_arg_match(env, subprog, regs);
16277 if (ret == -EFAULT)
16278 /* unlikely verifier bug. abort.
16279 * ret == 0 and ret < 0 are sadly acceptable for
16280 * main() function due to backward compatibility.
16281 * Like socket filter program may be written as:
16282 * int bpf_prog(struct pt_regs *ctx)
16283 * and never dereference that ctx in the program.
16284 * 'struct pt_regs' is a type mismatch for socket
16285 * filter that should be using 'struct __sk_buff'.
16290 ret = do_check(env);
16292 /* check for NULL is necessary, since cur_state can be freed inside
16293 * do_check() under memory pressure.
16295 if (env->cur_state) {
16296 free_verifier_state(env->cur_state, true);
16297 env->cur_state = NULL;
16299 while (!pop_stack(env, NULL, NULL, false));
16300 if (!ret && pop_log)
16301 bpf_vlog_reset(&env->log, 0);
16306 /* Verify all global functions in a BPF program one by one based on their BTF.
16307 * All global functions must pass verification. Otherwise the whole program is rejected.
16318 * foo() will be verified first for R1=any_scalar_value. During verification it
16319 * will be assumed that bar() already verified successfully and call to bar()
16320 * from foo() will be checked for type match only. Later bar() will be verified
16321 * independently to check that it's safe for R1=any_scalar_value.
16323 static int do_check_subprogs(struct bpf_verifier_env *env)
16325 struct bpf_prog_aux *aux = env->prog->aux;
16328 if (!aux->func_info)
16331 for (i = 1; i < env->subprog_cnt; i++) {
16332 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
16334 env->insn_idx = env->subprog_info[i].start;
16335 WARN_ON_ONCE(env->insn_idx == 0);
16336 ret = do_check_common(env, i);
16339 } else if (env->log.level & BPF_LOG_LEVEL) {
16341 "Func#%d is safe for any args that match its prototype\n",
16348 static int do_check_main(struct bpf_verifier_env *env)
16353 ret = do_check_common(env, 0);
16355 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
16360 static void print_verification_stats(struct bpf_verifier_env *env)
16364 if (env->log.level & BPF_LOG_STATS) {
16365 verbose(env, "verification time %lld usec\n",
16366 div_u64(env->verification_time, 1000));
16367 verbose(env, "stack depth ");
16368 for (i = 0; i < env->subprog_cnt; i++) {
16369 u32 depth = env->subprog_info[i].stack_depth;
16371 verbose(env, "%d", depth);
16372 if (i + 1 < env->subprog_cnt)
16375 verbose(env, "\n");
16377 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
16378 "total_states %d peak_states %d mark_read %d\n",
16379 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
16380 env->max_states_per_insn, env->total_states,
16381 env->peak_states, env->longest_mark_read_walk);
16384 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
16386 const struct btf_type *t, *func_proto;
16387 const struct bpf_struct_ops *st_ops;
16388 const struct btf_member *member;
16389 struct bpf_prog *prog = env->prog;
16390 u32 btf_id, member_idx;
16393 if (!prog->gpl_compatible) {
16394 verbose(env, "struct ops programs must have a GPL compatible license\n");
16398 btf_id = prog->aux->attach_btf_id;
16399 st_ops = bpf_struct_ops_find(btf_id);
16401 verbose(env, "attach_btf_id %u is not a supported struct\n",
16407 member_idx = prog->expected_attach_type;
16408 if (member_idx >= btf_type_vlen(t)) {
16409 verbose(env, "attach to invalid member idx %u of struct %s\n",
16410 member_idx, st_ops->name);
16414 member = &btf_type_member(t)[member_idx];
16415 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
16416 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
16419 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
16420 mname, member_idx, st_ops->name);
16424 if (st_ops->check_member) {
16425 int err = st_ops->check_member(t, member);
16428 verbose(env, "attach to unsupported member %s of struct %s\n",
16429 mname, st_ops->name);
16434 prog->aux->attach_func_proto = func_proto;
16435 prog->aux->attach_func_name = mname;
16436 env->ops = st_ops->verifier_ops;
16440 #define SECURITY_PREFIX "security_"
16442 static int check_attach_modify_return(unsigned long addr, const char *func_name)
16444 if (within_error_injection_list(addr) ||
16445 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
16451 /* list of non-sleepable functions that are otherwise on
16452 * ALLOW_ERROR_INJECTION list
16454 BTF_SET_START(btf_non_sleepable_error_inject)
16455 /* Three functions below can be called from sleepable and non-sleepable context.
16456 * Assume non-sleepable from bpf safety point of view.
16458 BTF_ID(func, __filemap_add_folio)
16459 BTF_ID(func, should_fail_alloc_page)
16460 BTF_ID(func, should_failslab)
16461 BTF_SET_END(btf_non_sleepable_error_inject)
16463 static int check_non_sleepable_error_inject(u32 btf_id)
16465 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
16468 int bpf_check_attach_target(struct bpf_verifier_log *log,
16469 const struct bpf_prog *prog,
16470 const struct bpf_prog *tgt_prog,
16472 struct bpf_attach_target_info *tgt_info)
16474 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
16475 const char prefix[] = "btf_trace_";
16476 int ret = 0, subprog = -1, i;
16477 const struct btf_type *t;
16478 bool conservative = true;
16484 bpf_log(log, "Tracing programs must provide btf_id\n");
16487 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
16490 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
16493 t = btf_type_by_id(btf, btf_id);
16495 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
16498 tname = btf_name_by_offset(btf, t->name_off);
16500 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
16504 struct bpf_prog_aux *aux = tgt_prog->aux;
16506 for (i = 0; i < aux->func_info_cnt; i++)
16507 if (aux->func_info[i].type_id == btf_id) {
16511 if (subprog == -1) {
16512 bpf_log(log, "Subprog %s doesn't exist\n", tname);
16515 conservative = aux->func_info_aux[subprog].unreliable;
16516 if (prog_extension) {
16517 if (conservative) {
16519 "Cannot replace static functions\n");
16522 if (!prog->jit_requested) {
16524 "Extension programs should be JITed\n");
16528 if (!tgt_prog->jited) {
16529 bpf_log(log, "Can attach to only JITed progs\n");
16532 if (tgt_prog->type == prog->type) {
16533 /* Cannot fentry/fexit another fentry/fexit program.
16534 * Cannot attach program extension to another extension.
16535 * It's ok to attach fentry/fexit to extension program.
16537 bpf_log(log, "Cannot recursively attach\n");
16540 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
16542 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
16543 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
16544 /* Program extensions can extend all program types
16545 * except fentry/fexit. The reason is the following.
16546 * The fentry/fexit programs are used for performance
16547 * analysis, stats and can be attached to any program
16548 * type except themselves. When extension program is
16549 * replacing XDP function it is necessary to allow
16550 * performance analysis of all functions. Both original
16551 * XDP program and its program extension. Hence
16552 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
16553 * allowed. If extending of fentry/fexit was allowed it
16554 * would be possible to create long call chain
16555 * fentry->extension->fentry->extension beyond
16556 * reasonable stack size. Hence extending fentry is not
16559 bpf_log(log, "Cannot extend fentry/fexit\n");
16563 if (prog_extension) {
16564 bpf_log(log, "Cannot replace kernel functions\n");
16569 switch (prog->expected_attach_type) {
16570 case BPF_TRACE_RAW_TP:
16573 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
16576 if (!btf_type_is_typedef(t)) {
16577 bpf_log(log, "attach_btf_id %u is not a typedef\n",
16581 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
16582 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
16586 tname += sizeof(prefix) - 1;
16587 t = btf_type_by_id(btf, t->type);
16588 if (!btf_type_is_ptr(t))
16589 /* should never happen in valid vmlinux build */
16591 t = btf_type_by_id(btf, t->type);
16592 if (!btf_type_is_func_proto(t))
16593 /* should never happen in valid vmlinux build */
16597 case BPF_TRACE_ITER:
16598 if (!btf_type_is_func(t)) {
16599 bpf_log(log, "attach_btf_id %u is not a function\n",
16603 t = btf_type_by_id(btf, t->type);
16604 if (!btf_type_is_func_proto(t))
16606 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
16611 if (!prog_extension)
16614 case BPF_MODIFY_RETURN:
16616 case BPF_LSM_CGROUP:
16617 case BPF_TRACE_FENTRY:
16618 case BPF_TRACE_FEXIT:
16619 if (!btf_type_is_func(t)) {
16620 bpf_log(log, "attach_btf_id %u is not a function\n",
16624 if (prog_extension &&
16625 btf_check_type_match(log, prog, btf, t))
16627 t = btf_type_by_id(btf, t->type);
16628 if (!btf_type_is_func_proto(t))
16631 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
16632 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
16633 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
16636 if (tgt_prog && conservative)
16639 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
16645 addr = (long) tgt_prog->bpf_func;
16647 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
16649 addr = kallsyms_lookup_name(tname);
16652 "The address of function %s cannot be found\n",
16658 if (prog->aux->sleepable) {
16660 switch (prog->type) {
16661 case BPF_PROG_TYPE_TRACING:
16663 /* fentry/fexit/fmod_ret progs can be sleepable if they are
16664 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
16666 if (!check_non_sleepable_error_inject(btf_id) &&
16667 within_error_injection_list(addr))
16669 /* fentry/fexit/fmod_ret progs can also be sleepable if they are
16670 * in the fmodret id set with the KF_SLEEPABLE flag.
16673 u32 *flags = btf_kfunc_is_modify_return(btf, btf_id);
16675 if (flags && (*flags & KF_SLEEPABLE))
16679 case BPF_PROG_TYPE_LSM:
16680 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
16681 * Only some of them are sleepable.
16683 if (bpf_lsm_is_sleepable_hook(btf_id))
16690 bpf_log(log, "%s is not sleepable\n", tname);
16693 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
16695 bpf_log(log, "can't modify return codes of BPF programs\n");
16699 if (btf_kfunc_is_modify_return(btf, btf_id) ||
16700 !check_attach_modify_return(addr, tname))
16703 bpf_log(log, "%s() is not modifiable\n", tname);
16710 tgt_info->tgt_addr = addr;
16711 tgt_info->tgt_name = tname;
16712 tgt_info->tgt_type = t;
16716 BTF_SET_START(btf_id_deny)
16719 BTF_ID(func, migrate_disable)
16720 BTF_ID(func, migrate_enable)
16722 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
16723 BTF_ID(func, rcu_read_unlock_strict)
16725 BTF_SET_END(btf_id_deny)
16727 static int check_attach_btf_id(struct bpf_verifier_env *env)
16729 struct bpf_prog *prog = env->prog;
16730 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
16731 struct bpf_attach_target_info tgt_info = {};
16732 u32 btf_id = prog->aux->attach_btf_id;
16733 struct bpf_trampoline *tr;
16737 if (prog->type == BPF_PROG_TYPE_SYSCALL) {
16738 if (prog->aux->sleepable)
16739 /* attach_btf_id checked to be zero already */
16741 verbose(env, "Syscall programs can only be sleepable\n");
16745 if (prog->aux->sleepable && prog->type != BPF_PROG_TYPE_TRACING &&
16746 prog->type != BPF_PROG_TYPE_LSM && prog->type != BPF_PROG_TYPE_KPROBE) {
16747 verbose(env, "Only fentry/fexit/fmod_ret, lsm, and kprobe/uprobe programs can be sleepable\n");
16751 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
16752 return check_struct_ops_btf_id(env);
16754 if (prog->type != BPF_PROG_TYPE_TRACING &&
16755 prog->type != BPF_PROG_TYPE_LSM &&
16756 prog->type != BPF_PROG_TYPE_EXT)
16759 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
16763 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
16764 /* to make freplace equivalent to their targets, they need to
16765 * inherit env->ops and expected_attach_type for the rest of the
16768 env->ops = bpf_verifier_ops[tgt_prog->type];
16769 prog->expected_attach_type = tgt_prog->expected_attach_type;
16772 /* store info about the attachment target that will be used later */
16773 prog->aux->attach_func_proto = tgt_info.tgt_type;
16774 prog->aux->attach_func_name = tgt_info.tgt_name;
16777 prog->aux->saved_dst_prog_type = tgt_prog->type;
16778 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
16781 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
16782 prog->aux->attach_btf_trace = true;
16784 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
16785 if (!bpf_iter_prog_supported(prog))
16790 if (prog->type == BPF_PROG_TYPE_LSM) {
16791 ret = bpf_lsm_verify_prog(&env->log, prog);
16794 } else if (prog->type == BPF_PROG_TYPE_TRACING &&
16795 btf_id_set_contains(&btf_id_deny, btf_id)) {
16799 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
16800 tr = bpf_trampoline_get(key, &tgt_info);
16804 prog->aux->dst_trampoline = tr;
16808 struct btf *bpf_get_btf_vmlinux(void)
16810 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
16811 mutex_lock(&bpf_verifier_lock);
16813 btf_vmlinux = btf_parse_vmlinux();
16814 mutex_unlock(&bpf_verifier_lock);
16816 return btf_vmlinux;
16819 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr)
16821 u64 start_time = ktime_get_ns();
16822 struct bpf_verifier_env *env;
16823 struct bpf_verifier_log *log;
16824 int i, len, ret = -EINVAL;
16827 /* no program is valid */
16828 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
16831 /* 'struct bpf_verifier_env' can be global, but since it's not small,
16832 * allocate/free it every time bpf_check() is called
16834 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
16839 len = (*prog)->len;
16840 env->insn_aux_data =
16841 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
16843 if (!env->insn_aux_data)
16845 for (i = 0; i < len; i++)
16846 env->insn_aux_data[i].orig_idx = i;
16848 env->ops = bpf_verifier_ops[env->prog->type];
16849 env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
16850 is_priv = bpf_capable();
16852 bpf_get_btf_vmlinux();
16854 /* grab the mutex to protect few globals used by verifier */
16856 mutex_lock(&bpf_verifier_lock);
16858 if (attr->log_level || attr->log_buf || attr->log_size) {
16859 /* user requested verbose verifier output
16860 * and supplied buffer to store the verification trace
16862 log->level = attr->log_level;
16863 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
16864 log->len_total = attr->log_size;
16866 /* log attributes have to be sane */
16867 if (!bpf_verifier_log_attr_valid(log)) {
16873 mark_verifier_state_clean(env);
16875 if (IS_ERR(btf_vmlinux)) {
16876 /* Either gcc or pahole or kernel are broken. */
16877 verbose(env, "in-kernel BTF is malformed\n");
16878 ret = PTR_ERR(btf_vmlinux);
16879 goto skip_full_check;
16882 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
16883 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
16884 env->strict_alignment = true;
16885 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
16886 env->strict_alignment = false;
16888 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
16889 env->allow_uninit_stack = bpf_allow_uninit_stack();
16890 env->bypass_spec_v1 = bpf_bypass_spec_v1();
16891 env->bypass_spec_v4 = bpf_bypass_spec_v4();
16892 env->bpf_capable = bpf_capable();
16893 env->rcu_tag_supported = btf_vmlinux &&
16894 btf_find_by_name_kind(btf_vmlinux, "rcu", BTF_KIND_TYPE_TAG) > 0;
16897 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
16899 env->explored_states = kvcalloc(state_htab_size(env),
16900 sizeof(struct bpf_verifier_state_list *),
16903 if (!env->explored_states)
16904 goto skip_full_check;
16906 ret = add_subprog_and_kfunc(env);
16908 goto skip_full_check;
16910 ret = check_subprogs(env);
16912 goto skip_full_check;
16914 ret = check_btf_info(env, attr, uattr);
16916 goto skip_full_check;
16918 ret = check_attach_btf_id(env);
16920 goto skip_full_check;
16922 ret = resolve_pseudo_ldimm64(env);
16924 goto skip_full_check;
16926 if (bpf_prog_is_dev_bound(env->prog->aux)) {
16927 ret = bpf_prog_offload_verifier_prep(env->prog);
16929 goto skip_full_check;
16932 ret = check_cfg(env);
16934 goto skip_full_check;
16936 ret = do_check_subprogs(env);
16937 ret = ret ?: do_check_main(env);
16939 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
16940 ret = bpf_prog_offload_finalize(env);
16943 kvfree(env->explored_states);
16946 ret = check_max_stack_depth(env);
16948 /* instruction rewrites happen after this point */
16950 ret = optimize_bpf_loop(env);
16954 opt_hard_wire_dead_code_branches(env);
16956 ret = opt_remove_dead_code(env);
16958 ret = opt_remove_nops(env);
16961 sanitize_dead_code(env);
16965 /* program is valid, convert *(u32*)(ctx + off) accesses */
16966 ret = convert_ctx_accesses(env);
16969 ret = do_misc_fixups(env);
16971 /* do 32-bit optimization after insn patching has done so those patched
16972 * insns could be handled correctly.
16974 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
16975 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
16976 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
16981 ret = fixup_call_args(env);
16983 env->verification_time = ktime_get_ns() - start_time;
16984 print_verification_stats(env);
16985 env->prog->aux->verified_insns = env->insn_processed;
16987 if (log->level && bpf_verifier_log_full(log))
16989 if (log->level && !log->ubuf) {
16991 goto err_release_maps;
16995 goto err_release_maps;
16997 if (env->used_map_cnt) {
16998 /* if program passed verifier, update used_maps in bpf_prog_info */
16999 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
17000 sizeof(env->used_maps[0]),
17003 if (!env->prog->aux->used_maps) {
17005 goto err_release_maps;
17008 memcpy(env->prog->aux->used_maps, env->used_maps,
17009 sizeof(env->used_maps[0]) * env->used_map_cnt);
17010 env->prog->aux->used_map_cnt = env->used_map_cnt;
17012 if (env->used_btf_cnt) {
17013 /* if program passed verifier, update used_btfs in bpf_prog_aux */
17014 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
17015 sizeof(env->used_btfs[0]),
17017 if (!env->prog->aux->used_btfs) {
17019 goto err_release_maps;
17022 memcpy(env->prog->aux->used_btfs, env->used_btfs,
17023 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
17024 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
17026 if (env->used_map_cnt || env->used_btf_cnt) {
17027 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
17028 * bpf_ld_imm64 instructions
17030 convert_pseudo_ld_imm64(env);
17033 adjust_btf_func(env);
17036 if (!env->prog->aux->used_maps)
17037 /* if we didn't copy map pointers into bpf_prog_info, release
17038 * them now. Otherwise free_used_maps() will release them.
17041 if (!env->prog->aux->used_btfs)
17044 /* extension progs temporarily inherit the attach_type of their targets
17045 for verification purposes, so set it back to zero before returning
17047 if (env->prog->type == BPF_PROG_TYPE_EXT)
17048 env->prog->expected_attach_type = 0;
17053 mutex_unlock(&bpf_verifier_lock);
17054 vfree(env->insn_aux_data);