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 bpf_map_value_off_desc *kptr_off_desc;
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 reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
456 return reg->type == PTR_TO_MAP_VALUE &&
457 map_value_has_spin_lock(reg->map_ptr);
460 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
462 type = base_type(type);
463 return type == PTR_TO_SOCKET || type == PTR_TO_TCP_SOCK ||
464 type == PTR_TO_MEM || type == PTR_TO_BTF_ID;
467 static bool type_is_rdonly_mem(u32 type)
469 return type & MEM_RDONLY;
472 static bool type_may_be_null(u32 type)
474 return type & PTR_MAYBE_NULL;
477 static bool is_acquire_function(enum bpf_func_id func_id,
478 const struct bpf_map *map)
480 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
482 if (func_id == BPF_FUNC_sk_lookup_tcp ||
483 func_id == BPF_FUNC_sk_lookup_udp ||
484 func_id == BPF_FUNC_skc_lookup_tcp ||
485 func_id == BPF_FUNC_ringbuf_reserve ||
486 func_id == BPF_FUNC_kptr_xchg)
489 if (func_id == BPF_FUNC_map_lookup_elem &&
490 (map_type == BPF_MAP_TYPE_SOCKMAP ||
491 map_type == BPF_MAP_TYPE_SOCKHASH))
497 static bool is_ptr_cast_function(enum bpf_func_id func_id)
499 return func_id == BPF_FUNC_tcp_sock ||
500 func_id == BPF_FUNC_sk_fullsock ||
501 func_id == BPF_FUNC_skc_to_tcp_sock ||
502 func_id == BPF_FUNC_skc_to_tcp6_sock ||
503 func_id == BPF_FUNC_skc_to_udp6_sock ||
504 func_id == BPF_FUNC_skc_to_mptcp_sock ||
505 func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
506 func_id == BPF_FUNC_skc_to_tcp_request_sock;
509 static bool is_dynptr_ref_function(enum bpf_func_id func_id)
511 return func_id == BPF_FUNC_dynptr_data;
514 static bool helper_multiple_ref_obj_use(enum bpf_func_id func_id,
515 const struct bpf_map *map)
517 int ref_obj_uses = 0;
519 if (is_ptr_cast_function(func_id))
521 if (is_acquire_function(func_id, map))
523 if (is_dynptr_ref_function(func_id))
526 return ref_obj_uses > 1;
529 static bool is_cmpxchg_insn(const struct bpf_insn *insn)
531 return BPF_CLASS(insn->code) == BPF_STX &&
532 BPF_MODE(insn->code) == BPF_ATOMIC &&
533 insn->imm == BPF_CMPXCHG;
536 /* string representation of 'enum bpf_reg_type'
538 * Note that reg_type_str() can not appear more than once in a single verbose()
541 static const char *reg_type_str(struct bpf_verifier_env *env,
542 enum bpf_reg_type type)
544 char postfix[16] = {0}, prefix[32] = {0};
545 static const char * const str[] = {
547 [SCALAR_VALUE] = "scalar",
548 [PTR_TO_CTX] = "ctx",
549 [CONST_PTR_TO_MAP] = "map_ptr",
550 [PTR_TO_MAP_VALUE] = "map_value",
551 [PTR_TO_STACK] = "fp",
552 [PTR_TO_PACKET] = "pkt",
553 [PTR_TO_PACKET_META] = "pkt_meta",
554 [PTR_TO_PACKET_END] = "pkt_end",
555 [PTR_TO_FLOW_KEYS] = "flow_keys",
556 [PTR_TO_SOCKET] = "sock",
557 [PTR_TO_SOCK_COMMON] = "sock_common",
558 [PTR_TO_TCP_SOCK] = "tcp_sock",
559 [PTR_TO_TP_BUFFER] = "tp_buffer",
560 [PTR_TO_XDP_SOCK] = "xdp_sock",
561 [PTR_TO_BTF_ID] = "ptr_",
562 [PTR_TO_MEM] = "mem",
563 [PTR_TO_BUF] = "buf",
564 [PTR_TO_FUNC] = "func",
565 [PTR_TO_MAP_KEY] = "map_key",
566 [PTR_TO_DYNPTR] = "dynptr_ptr",
569 if (type & PTR_MAYBE_NULL) {
570 if (base_type(type) == PTR_TO_BTF_ID)
571 strncpy(postfix, "or_null_", 16);
573 strncpy(postfix, "_or_null", 16);
576 if (type & MEM_RDONLY)
577 strncpy(prefix, "rdonly_", 32);
578 if (type & MEM_ALLOC)
579 strncpy(prefix, "alloc_", 32);
581 strncpy(prefix, "user_", 32);
582 if (type & MEM_PERCPU)
583 strncpy(prefix, "percpu_", 32);
584 if (type & PTR_UNTRUSTED)
585 strncpy(prefix, "untrusted_", 32);
587 snprintf(env->type_str_buf, TYPE_STR_BUF_LEN, "%s%s%s",
588 prefix, str[base_type(type)], postfix);
589 return env->type_str_buf;
592 static char slot_type_char[] = {
593 [STACK_INVALID] = '?',
597 [STACK_DYNPTR] = 'd',
600 static void print_liveness(struct bpf_verifier_env *env,
601 enum bpf_reg_liveness live)
603 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
605 if (live & REG_LIVE_READ)
607 if (live & REG_LIVE_WRITTEN)
609 if (live & REG_LIVE_DONE)
613 static int get_spi(s32 off)
615 return (-off - 1) / BPF_REG_SIZE;
618 static bool is_spi_bounds_valid(struct bpf_func_state *state, int spi, int nr_slots)
620 int allocated_slots = state->allocated_stack / BPF_REG_SIZE;
622 /* We need to check that slots between [spi - nr_slots + 1, spi] are
623 * within [0, allocated_stack).
625 * Please note that the spi grows downwards. For example, a dynptr
626 * takes the size of two stack slots; the first slot will be at
627 * spi and the second slot will be at spi - 1.
629 return spi - nr_slots + 1 >= 0 && spi < allocated_slots;
632 static struct bpf_func_state *func(struct bpf_verifier_env *env,
633 const struct bpf_reg_state *reg)
635 struct bpf_verifier_state *cur = env->cur_state;
637 return cur->frame[reg->frameno];
640 static const char *kernel_type_name(const struct btf* btf, u32 id)
642 return btf_name_by_offset(btf, btf_type_by_id(btf, id)->name_off);
645 static void mark_reg_scratched(struct bpf_verifier_env *env, u32 regno)
647 env->scratched_regs |= 1U << regno;
650 static void mark_stack_slot_scratched(struct bpf_verifier_env *env, u32 spi)
652 env->scratched_stack_slots |= 1ULL << spi;
655 static bool reg_scratched(const struct bpf_verifier_env *env, u32 regno)
657 return (env->scratched_regs >> regno) & 1;
660 static bool stack_slot_scratched(const struct bpf_verifier_env *env, u64 regno)
662 return (env->scratched_stack_slots >> regno) & 1;
665 static bool verifier_state_scratched(const struct bpf_verifier_env *env)
667 return env->scratched_regs || env->scratched_stack_slots;
670 static void mark_verifier_state_clean(struct bpf_verifier_env *env)
672 env->scratched_regs = 0U;
673 env->scratched_stack_slots = 0ULL;
676 /* Used for printing the entire verifier state. */
677 static void mark_verifier_state_scratched(struct bpf_verifier_env *env)
679 env->scratched_regs = ~0U;
680 env->scratched_stack_slots = ~0ULL;
683 static enum bpf_dynptr_type arg_to_dynptr_type(enum bpf_arg_type arg_type)
685 switch (arg_type & DYNPTR_TYPE_FLAG_MASK) {
686 case DYNPTR_TYPE_LOCAL:
687 return BPF_DYNPTR_TYPE_LOCAL;
688 case DYNPTR_TYPE_RINGBUF:
689 return BPF_DYNPTR_TYPE_RINGBUF;
691 return BPF_DYNPTR_TYPE_INVALID;
695 static bool dynptr_type_refcounted(enum bpf_dynptr_type type)
697 return type == BPF_DYNPTR_TYPE_RINGBUF;
700 static int mark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
701 enum bpf_arg_type arg_type, int insn_idx)
703 struct bpf_func_state *state = func(env, reg);
704 enum bpf_dynptr_type type;
707 spi = get_spi(reg->off);
709 if (!is_spi_bounds_valid(state, spi, BPF_DYNPTR_NR_SLOTS))
712 for (i = 0; i < BPF_REG_SIZE; i++) {
713 state->stack[spi].slot_type[i] = STACK_DYNPTR;
714 state->stack[spi - 1].slot_type[i] = STACK_DYNPTR;
717 type = arg_to_dynptr_type(arg_type);
718 if (type == BPF_DYNPTR_TYPE_INVALID)
721 state->stack[spi].spilled_ptr.dynptr.first_slot = true;
722 state->stack[spi].spilled_ptr.dynptr.type = type;
723 state->stack[spi - 1].spilled_ptr.dynptr.type = type;
725 if (dynptr_type_refcounted(type)) {
726 /* The id is used to track proper releasing */
727 id = acquire_reference_state(env, insn_idx);
731 state->stack[spi].spilled_ptr.id = id;
732 state->stack[spi - 1].spilled_ptr.id = id;
738 static int unmark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
740 struct bpf_func_state *state = func(env, reg);
743 spi = get_spi(reg->off);
745 if (!is_spi_bounds_valid(state, spi, BPF_DYNPTR_NR_SLOTS))
748 for (i = 0; i < BPF_REG_SIZE; i++) {
749 state->stack[spi].slot_type[i] = STACK_INVALID;
750 state->stack[spi - 1].slot_type[i] = STACK_INVALID;
753 /* Invalidate any slices associated with this dynptr */
754 if (dynptr_type_refcounted(state->stack[spi].spilled_ptr.dynptr.type)) {
755 release_reference(env, state->stack[spi].spilled_ptr.id);
756 state->stack[spi].spilled_ptr.id = 0;
757 state->stack[spi - 1].spilled_ptr.id = 0;
760 state->stack[spi].spilled_ptr.dynptr.first_slot = false;
761 state->stack[spi].spilled_ptr.dynptr.type = 0;
762 state->stack[spi - 1].spilled_ptr.dynptr.type = 0;
767 static bool is_dynptr_reg_valid_uninit(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
769 struct bpf_func_state *state = func(env, reg);
770 int spi = get_spi(reg->off);
773 if (!is_spi_bounds_valid(state, spi, BPF_DYNPTR_NR_SLOTS))
776 for (i = 0; i < BPF_REG_SIZE; i++) {
777 if (state->stack[spi].slot_type[i] == STACK_DYNPTR ||
778 state->stack[spi - 1].slot_type[i] == STACK_DYNPTR)
785 bool is_dynptr_reg_valid_init(struct bpf_verifier_env *env,
786 struct bpf_reg_state *reg)
788 struct bpf_func_state *state = func(env, reg);
789 int spi = get_spi(reg->off);
792 if (!is_spi_bounds_valid(state, spi, BPF_DYNPTR_NR_SLOTS) ||
793 !state->stack[spi].spilled_ptr.dynptr.first_slot)
796 for (i = 0; i < BPF_REG_SIZE; i++) {
797 if (state->stack[spi].slot_type[i] != STACK_DYNPTR ||
798 state->stack[spi - 1].slot_type[i] != STACK_DYNPTR)
805 bool is_dynptr_type_expected(struct bpf_verifier_env *env,
806 struct bpf_reg_state *reg,
807 enum bpf_arg_type arg_type)
809 struct bpf_func_state *state = func(env, reg);
810 enum bpf_dynptr_type dynptr_type;
811 int spi = get_spi(reg->off);
813 /* ARG_PTR_TO_DYNPTR takes any type of dynptr */
814 if (arg_type == ARG_PTR_TO_DYNPTR)
817 dynptr_type = arg_to_dynptr_type(arg_type);
819 return state->stack[spi].spilled_ptr.dynptr.type == dynptr_type;
822 /* The reg state of a pointer or a bounded scalar was saved when
823 * it was spilled to the stack.
825 static bool is_spilled_reg(const struct bpf_stack_state *stack)
827 return stack->slot_type[BPF_REG_SIZE - 1] == STACK_SPILL;
830 static void scrub_spilled_slot(u8 *stype)
832 if (*stype != STACK_INVALID)
836 static void print_verifier_state(struct bpf_verifier_env *env,
837 const struct bpf_func_state *state,
840 const struct bpf_reg_state *reg;
845 verbose(env, " frame%d:", state->frameno);
846 for (i = 0; i < MAX_BPF_REG; i++) {
847 reg = &state->regs[i];
851 if (!print_all && !reg_scratched(env, i))
853 verbose(env, " R%d", i);
854 print_liveness(env, reg->live);
856 if (t == SCALAR_VALUE && reg->precise)
858 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
859 tnum_is_const(reg->var_off)) {
860 /* reg->off should be 0 for SCALAR_VALUE */
861 verbose(env, "%s", t == SCALAR_VALUE ? "" : reg_type_str(env, t));
862 verbose(env, "%lld", reg->var_off.value + reg->off);
864 const char *sep = "";
866 verbose(env, "%s", reg_type_str(env, t));
867 if (base_type(t) == PTR_TO_BTF_ID)
868 verbose(env, "%s", kernel_type_name(reg->btf, reg->btf_id));
871 * _a stands for append, was shortened to avoid multiline statements below.
872 * This macro is used to output a comma separated list of attributes.
874 #define verbose_a(fmt, ...) ({ verbose(env, "%s" fmt, sep, __VA_ARGS__); sep = ","; })
877 verbose_a("id=%d", reg->id);
878 if (reg_type_may_be_refcounted_or_null(t) && reg->ref_obj_id)
879 verbose_a("ref_obj_id=%d", reg->ref_obj_id);
880 if (t != SCALAR_VALUE)
881 verbose_a("off=%d", reg->off);
882 if (type_is_pkt_pointer(t))
883 verbose_a("r=%d", reg->range);
884 else if (base_type(t) == CONST_PTR_TO_MAP ||
885 base_type(t) == PTR_TO_MAP_KEY ||
886 base_type(t) == PTR_TO_MAP_VALUE)
887 verbose_a("ks=%d,vs=%d",
888 reg->map_ptr->key_size,
889 reg->map_ptr->value_size);
890 if (tnum_is_const(reg->var_off)) {
891 /* Typically an immediate SCALAR_VALUE, but
892 * could be a pointer whose offset is too big
895 verbose_a("imm=%llx", reg->var_off.value);
897 if (reg->smin_value != reg->umin_value &&
898 reg->smin_value != S64_MIN)
899 verbose_a("smin=%lld", (long long)reg->smin_value);
900 if (reg->smax_value != reg->umax_value &&
901 reg->smax_value != S64_MAX)
902 verbose_a("smax=%lld", (long long)reg->smax_value);
903 if (reg->umin_value != 0)
904 verbose_a("umin=%llu", (unsigned long long)reg->umin_value);
905 if (reg->umax_value != U64_MAX)
906 verbose_a("umax=%llu", (unsigned long long)reg->umax_value);
907 if (!tnum_is_unknown(reg->var_off)) {
910 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
911 verbose_a("var_off=%s", tn_buf);
913 if (reg->s32_min_value != reg->smin_value &&
914 reg->s32_min_value != S32_MIN)
915 verbose_a("s32_min=%d", (int)(reg->s32_min_value));
916 if (reg->s32_max_value != reg->smax_value &&
917 reg->s32_max_value != S32_MAX)
918 verbose_a("s32_max=%d", (int)(reg->s32_max_value));
919 if (reg->u32_min_value != reg->umin_value &&
920 reg->u32_min_value != U32_MIN)
921 verbose_a("u32_min=%d", (int)(reg->u32_min_value));
922 if (reg->u32_max_value != reg->umax_value &&
923 reg->u32_max_value != U32_MAX)
924 verbose_a("u32_max=%d", (int)(reg->u32_max_value));
931 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
932 char types_buf[BPF_REG_SIZE + 1];
936 for (j = 0; j < BPF_REG_SIZE; j++) {
937 if (state->stack[i].slot_type[j] != STACK_INVALID)
939 types_buf[j] = slot_type_char[
940 state->stack[i].slot_type[j]];
942 types_buf[BPF_REG_SIZE] = 0;
945 if (!print_all && !stack_slot_scratched(env, i))
947 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
948 print_liveness(env, state->stack[i].spilled_ptr.live);
949 if (is_spilled_reg(&state->stack[i])) {
950 reg = &state->stack[i].spilled_ptr;
952 verbose(env, "=%s", t == SCALAR_VALUE ? "" : reg_type_str(env, t));
953 if (t == SCALAR_VALUE && reg->precise)
955 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
956 verbose(env, "%lld", reg->var_off.value + reg->off);
958 verbose(env, "=%s", types_buf);
961 if (state->acquired_refs && state->refs[0].id) {
962 verbose(env, " refs=%d", state->refs[0].id);
963 for (i = 1; i < state->acquired_refs; i++)
964 if (state->refs[i].id)
965 verbose(env, ",%d", state->refs[i].id);
967 if (state->in_callback_fn)
969 if (state->in_async_callback_fn)
970 verbose(env, " async_cb");
972 mark_verifier_state_clean(env);
975 static inline u32 vlog_alignment(u32 pos)
977 return round_up(max(pos + BPF_LOG_MIN_ALIGNMENT / 2, BPF_LOG_ALIGNMENT),
978 BPF_LOG_MIN_ALIGNMENT) - pos - 1;
981 static void print_insn_state(struct bpf_verifier_env *env,
982 const struct bpf_func_state *state)
984 if (env->prev_log_len && env->prev_log_len == env->log.len_used) {
985 /* remove new line character */
986 bpf_vlog_reset(&env->log, env->prev_log_len - 1);
987 verbose(env, "%*c;", vlog_alignment(env->prev_insn_print_len), ' ');
989 verbose(env, "%d:", env->insn_idx);
991 print_verifier_state(env, state, false);
994 /* copy array src of length n * size bytes to dst. dst is reallocated if it's too
995 * small to hold src. This is different from krealloc since we don't want to preserve
996 * the contents of dst.
998 * Leaves dst untouched if src is NULL or length is zero. Returns NULL if memory could
1001 static void *copy_array(void *dst, const void *src, size_t n, size_t size, gfp_t flags)
1005 if (ZERO_OR_NULL_PTR(src))
1008 if (unlikely(check_mul_overflow(n, size, &bytes)))
1011 if (ksize(dst) < bytes) {
1013 dst = kmalloc_track_caller(bytes, flags);
1018 memcpy(dst, src, bytes);
1020 return dst ? dst : ZERO_SIZE_PTR;
1023 /* resize an array from old_n items to new_n items. the array is reallocated if it's too
1024 * small to hold new_n items. new items are zeroed out if the array grows.
1026 * Contrary to krealloc_array, does not free arr if new_n is zero.
1028 static void *realloc_array(void *arr, size_t old_n, size_t new_n, size_t size)
1032 if (!new_n || old_n == new_n)
1035 new_arr = krealloc_array(arr, new_n, size, GFP_KERNEL);
1043 memset(arr + old_n * size, 0, (new_n - old_n) * size);
1046 return arr ? arr : ZERO_SIZE_PTR;
1049 static int copy_reference_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
1051 dst->refs = copy_array(dst->refs, src->refs, src->acquired_refs,
1052 sizeof(struct bpf_reference_state), GFP_KERNEL);
1056 dst->acquired_refs = src->acquired_refs;
1060 static int copy_stack_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
1062 size_t n = src->allocated_stack / BPF_REG_SIZE;
1064 dst->stack = copy_array(dst->stack, src->stack, n, sizeof(struct bpf_stack_state),
1069 dst->allocated_stack = src->allocated_stack;
1073 static int resize_reference_state(struct bpf_func_state *state, size_t n)
1075 state->refs = realloc_array(state->refs, state->acquired_refs, n,
1076 sizeof(struct bpf_reference_state));
1080 state->acquired_refs = n;
1084 static int grow_stack_state(struct bpf_func_state *state, int size)
1086 size_t old_n = state->allocated_stack / BPF_REG_SIZE, n = size / BPF_REG_SIZE;
1091 state->stack = realloc_array(state->stack, old_n, n, sizeof(struct bpf_stack_state));
1095 state->allocated_stack = size;
1099 /* Acquire a pointer id from the env and update the state->refs to include
1100 * this new pointer reference.
1101 * On success, returns a valid pointer id to associate with the register
1102 * On failure, returns a negative errno.
1104 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
1106 struct bpf_func_state *state = cur_func(env);
1107 int new_ofs = state->acquired_refs;
1110 err = resize_reference_state(state, state->acquired_refs + 1);
1114 state->refs[new_ofs].id = id;
1115 state->refs[new_ofs].insn_idx = insn_idx;
1116 state->refs[new_ofs].callback_ref = state->in_callback_fn ? state->frameno : 0;
1121 /* release function corresponding to acquire_reference_state(). Idempotent. */
1122 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
1126 last_idx = state->acquired_refs - 1;
1127 for (i = 0; i < state->acquired_refs; i++) {
1128 if (state->refs[i].id == ptr_id) {
1129 /* Cannot release caller references in callbacks */
1130 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno)
1132 if (last_idx && i != last_idx)
1133 memcpy(&state->refs[i], &state->refs[last_idx],
1134 sizeof(*state->refs));
1135 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
1136 state->acquired_refs--;
1143 static void free_func_state(struct bpf_func_state *state)
1148 kfree(state->stack);
1152 static void clear_jmp_history(struct bpf_verifier_state *state)
1154 kfree(state->jmp_history);
1155 state->jmp_history = NULL;
1156 state->jmp_history_cnt = 0;
1159 static void free_verifier_state(struct bpf_verifier_state *state,
1164 for (i = 0; i <= state->curframe; i++) {
1165 free_func_state(state->frame[i]);
1166 state->frame[i] = NULL;
1168 clear_jmp_history(state);
1173 /* copy verifier state from src to dst growing dst stack space
1174 * when necessary to accommodate larger src stack
1176 static int copy_func_state(struct bpf_func_state *dst,
1177 const struct bpf_func_state *src)
1181 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
1182 err = copy_reference_state(dst, src);
1185 return copy_stack_state(dst, src);
1188 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
1189 const struct bpf_verifier_state *src)
1191 struct bpf_func_state *dst;
1194 dst_state->jmp_history = copy_array(dst_state->jmp_history, src->jmp_history,
1195 src->jmp_history_cnt, sizeof(struct bpf_idx_pair),
1197 if (!dst_state->jmp_history)
1199 dst_state->jmp_history_cnt = src->jmp_history_cnt;
1201 /* if dst has more stack frames then src frame, free them */
1202 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
1203 free_func_state(dst_state->frame[i]);
1204 dst_state->frame[i] = NULL;
1206 dst_state->speculative = src->speculative;
1207 dst_state->curframe = src->curframe;
1208 dst_state->active_spin_lock = src->active_spin_lock;
1209 dst_state->branches = src->branches;
1210 dst_state->parent = src->parent;
1211 dst_state->first_insn_idx = src->first_insn_idx;
1212 dst_state->last_insn_idx = src->last_insn_idx;
1213 for (i = 0; i <= src->curframe; i++) {
1214 dst = dst_state->frame[i];
1216 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
1219 dst_state->frame[i] = dst;
1221 err = copy_func_state(dst, src->frame[i]);
1228 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
1231 u32 br = --st->branches;
1233 /* WARN_ON(br > 1) technically makes sense here,
1234 * but see comment in push_stack(), hence:
1236 WARN_ONCE((int)br < 0,
1237 "BUG update_branch_counts:branches_to_explore=%d\n",
1245 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
1246 int *insn_idx, bool pop_log)
1248 struct bpf_verifier_state *cur = env->cur_state;
1249 struct bpf_verifier_stack_elem *elem, *head = env->head;
1252 if (env->head == NULL)
1256 err = copy_verifier_state(cur, &head->st);
1261 bpf_vlog_reset(&env->log, head->log_pos);
1263 *insn_idx = head->insn_idx;
1265 *prev_insn_idx = head->prev_insn_idx;
1267 free_verifier_state(&head->st, false);
1274 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
1275 int insn_idx, int prev_insn_idx,
1278 struct bpf_verifier_state *cur = env->cur_state;
1279 struct bpf_verifier_stack_elem *elem;
1282 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1286 elem->insn_idx = insn_idx;
1287 elem->prev_insn_idx = prev_insn_idx;
1288 elem->next = env->head;
1289 elem->log_pos = env->log.len_used;
1292 err = copy_verifier_state(&elem->st, cur);
1295 elem->st.speculative |= speculative;
1296 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1297 verbose(env, "The sequence of %d jumps is too complex.\n",
1301 if (elem->st.parent) {
1302 ++elem->st.parent->branches;
1303 /* WARN_ON(branches > 2) technically makes sense here,
1305 * 1. speculative states will bump 'branches' for non-branch
1307 * 2. is_state_visited() heuristics may decide not to create
1308 * a new state for a sequence of branches and all such current
1309 * and cloned states will be pointing to a single parent state
1310 * which might have large 'branches' count.
1315 free_verifier_state(env->cur_state, true);
1316 env->cur_state = NULL;
1317 /* pop all elements and return */
1318 while (!pop_stack(env, NULL, NULL, false));
1322 #define CALLER_SAVED_REGS 6
1323 static const int caller_saved[CALLER_SAVED_REGS] = {
1324 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
1327 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1328 struct bpf_reg_state *reg);
1330 /* This helper doesn't clear reg->id */
1331 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1333 reg->var_off = tnum_const(imm);
1334 reg->smin_value = (s64)imm;
1335 reg->smax_value = (s64)imm;
1336 reg->umin_value = imm;
1337 reg->umax_value = imm;
1339 reg->s32_min_value = (s32)imm;
1340 reg->s32_max_value = (s32)imm;
1341 reg->u32_min_value = (u32)imm;
1342 reg->u32_max_value = (u32)imm;
1345 /* Mark the unknown part of a register (variable offset or scalar value) as
1346 * known to have the value @imm.
1348 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1350 /* Clear id, off, and union(map_ptr, range) */
1351 memset(((u8 *)reg) + sizeof(reg->type), 0,
1352 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1353 ___mark_reg_known(reg, imm);
1356 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1358 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1359 reg->s32_min_value = (s32)imm;
1360 reg->s32_max_value = (s32)imm;
1361 reg->u32_min_value = (u32)imm;
1362 reg->u32_max_value = (u32)imm;
1365 /* Mark the 'variable offset' part of a register as zero. This should be
1366 * used only on registers holding a pointer type.
1368 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1370 __mark_reg_known(reg, 0);
1373 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1375 __mark_reg_known(reg, 0);
1376 reg->type = SCALAR_VALUE;
1379 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1380 struct bpf_reg_state *regs, u32 regno)
1382 if (WARN_ON(regno >= MAX_BPF_REG)) {
1383 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1384 /* Something bad happened, let's kill all regs */
1385 for (regno = 0; regno < MAX_BPF_REG; regno++)
1386 __mark_reg_not_init(env, regs + regno);
1389 __mark_reg_known_zero(regs + regno);
1392 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
1394 if (base_type(reg->type) == PTR_TO_MAP_VALUE) {
1395 const struct bpf_map *map = reg->map_ptr;
1397 if (map->inner_map_meta) {
1398 reg->type = CONST_PTR_TO_MAP;
1399 reg->map_ptr = map->inner_map_meta;
1400 /* transfer reg's id which is unique for every map_lookup_elem
1401 * as UID of the inner map.
1403 if (map_value_has_timer(map->inner_map_meta))
1404 reg->map_uid = reg->id;
1405 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
1406 reg->type = PTR_TO_XDP_SOCK;
1407 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
1408 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
1409 reg->type = PTR_TO_SOCKET;
1411 reg->type = PTR_TO_MAP_VALUE;
1416 reg->type &= ~PTR_MAYBE_NULL;
1419 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1421 return type_is_pkt_pointer(reg->type);
1424 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
1426 return reg_is_pkt_pointer(reg) ||
1427 reg->type == PTR_TO_PACKET_END;
1430 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
1431 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
1432 enum bpf_reg_type which)
1434 /* The register can already have a range from prior markings.
1435 * This is fine as long as it hasn't been advanced from its
1438 return reg->type == which &&
1441 tnum_equals_const(reg->var_off, 0);
1444 /* Reset the min/max bounds of a register */
1445 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1447 reg->smin_value = S64_MIN;
1448 reg->smax_value = S64_MAX;
1449 reg->umin_value = 0;
1450 reg->umax_value = U64_MAX;
1452 reg->s32_min_value = S32_MIN;
1453 reg->s32_max_value = S32_MAX;
1454 reg->u32_min_value = 0;
1455 reg->u32_max_value = U32_MAX;
1458 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
1460 reg->smin_value = S64_MIN;
1461 reg->smax_value = S64_MAX;
1462 reg->umin_value = 0;
1463 reg->umax_value = U64_MAX;
1466 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
1468 reg->s32_min_value = S32_MIN;
1469 reg->s32_max_value = S32_MAX;
1470 reg->u32_min_value = 0;
1471 reg->u32_max_value = U32_MAX;
1474 static void __update_reg32_bounds(struct bpf_reg_state *reg)
1476 struct tnum var32_off = tnum_subreg(reg->var_off);
1478 /* min signed is max(sign bit) | min(other bits) */
1479 reg->s32_min_value = max_t(s32, reg->s32_min_value,
1480 var32_off.value | (var32_off.mask & S32_MIN));
1481 /* max signed is min(sign bit) | max(other bits) */
1482 reg->s32_max_value = min_t(s32, reg->s32_max_value,
1483 var32_off.value | (var32_off.mask & S32_MAX));
1484 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
1485 reg->u32_max_value = min(reg->u32_max_value,
1486 (u32)(var32_off.value | var32_off.mask));
1489 static void __update_reg64_bounds(struct bpf_reg_state *reg)
1491 /* min signed is max(sign bit) | min(other bits) */
1492 reg->smin_value = max_t(s64, reg->smin_value,
1493 reg->var_off.value | (reg->var_off.mask & S64_MIN));
1494 /* max signed is min(sign bit) | max(other bits) */
1495 reg->smax_value = min_t(s64, reg->smax_value,
1496 reg->var_off.value | (reg->var_off.mask & S64_MAX));
1497 reg->umin_value = max(reg->umin_value, reg->var_off.value);
1498 reg->umax_value = min(reg->umax_value,
1499 reg->var_off.value | reg->var_off.mask);
1502 static void __update_reg_bounds(struct bpf_reg_state *reg)
1504 __update_reg32_bounds(reg);
1505 __update_reg64_bounds(reg);
1508 /* Uses signed min/max values to inform unsigned, and vice-versa */
1509 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
1511 /* Learn sign from signed bounds.
1512 * If we cannot cross the sign boundary, then signed and unsigned bounds
1513 * are the same, so combine. This works even in the negative case, e.g.
1514 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1516 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
1517 reg->s32_min_value = reg->u32_min_value =
1518 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1519 reg->s32_max_value = reg->u32_max_value =
1520 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1523 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1524 * boundary, so we must be careful.
1526 if ((s32)reg->u32_max_value >= 0) {
1527 /* Positive. We can't learn anything from the smin, but smax
1528 * is positive, hence safe.
1530 reg->s32_min_value = reg->u32_min_value;
1531 reg->s32_max_value = reg->u32_max_value =
1532 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1533 } else if ((s32)reg->u32_min_value < 0) {
1534 /* Negative. We can't learn anything from the smax, but smin
1535 * is negative, hence safe.
1537 reg->s32_min_value = reg->u32_min_value =
1538 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1539 reg->s32_max_value = reg->u32_max_value;
1543 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
1545 /* Learn sign from signed bounds.
1546 * If we cannot cross the sign boundary, then signed and unsigned bounds
1547 * are the same, so combine. This works even in the negative case, e.g.
1548 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1550 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1551 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1553 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1557 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1558 * boundary, so we must be careful.
1560 if ((s64)reg->umax_value >= 0) {
1561 /* Positive. We can't learn anything from the smin, but smax
1562 * is positive, hence safe.
1564 reg->smin_value = reg->umin_value;
1565 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1567 } else if ((s64)reg->umin_value < 0) {
1568 /* Negative. We can't learn anything from the smax, but smin
1569 * is negative, hence safe.
1571 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1573 reg->smax_value = reg->umax_value;
1577 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
1579 __reg32_deduce_bounds(reg);
1580 __reg64_deduce_bounds(reg);
1583 /* Attempts to improve var_off based on unsigned min/max information */
1584 static void __reg_bound_offset(struct bpf_reg_state *reg)
1586 struct tnum var64_off = tnum_intersect(reg->var_off,
1587 tnum_range(reg->umin_value,
1589 struct tnum var32_off = tnum_intersect(tnum_subreg(reg->var_off),
1590 tnum_range(reg->u32_min_value,
1591 reg->u32_max_value));
1593 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
1596 static void reg_bounds_sync(struct bpf_reg_state *reg)
1598 /* We might have learned new bounds from the var_off. */
1599 __update_reg_bounds(reg);
1600 /* We might have learned something about the sign bit. */
1601 __reg_deduce_bounds(reg);
1602 /* We might have learned some bits from the bounds. */
1603 __reg_bound_offset(reg);
1604 /* Intersecting with the old var_off might have improved our bounds
1605 * slightly, e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1606 * then new var_off is (0; 0x7f...fc) which improves our umax.
1608 __update_reg_bounds(reg);
1611 static bool __reg32_bound_s64(s32 a)
1613 return a >= 0 && a <= S32_MAX;
1616 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
1618 reg->umin_value = reg->u32_min_value;
1619 reg->umax_value = reg->u32_max_value;
1621 /* Attempt to pull 32-bit signed bounds into 64-bit bounds but must
1622 * be positive otherwise set to worse case bounds and refine later
1625 if (__reg32_bound_s64(reg->s32_min_value) &&
1626 __reg32_bound_s64(reg->s32_max_value)) {
1627 reg->smin_value = reg->s32_min_value;
1628 reg->smax_value = reg->s32_max_value;
1630 reg->smin_value = 0;
1631 reg->smax_value = U32_MAX;
1635 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
1637 /* special case when 64-bit register has upper 32-bit register
1638 * zeroed. Typically happens after zext or <<32, >>32 sequence
1639 * allowing us to use 32-bit bounds directly,
1641 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
1642 __reg_assign_32_into_64(reg);
1644 /* Otherwise the best we can do is push lower 32bit known and
1645 * unknown bits into register (var_off set from jmp logic)
1646 * then learn as much as possible from the 64-bit tnum
1647 * known and unknown bits. The previous smin/smax bounds are
1648 * invalid here because of jmp32 compare so mark them unknown
1649 * so they do not impact tnum bounds calculation.
1651 __mark_reg64_unbounded(reg);
1653 reg_bounds_sync(reg);
1656 static bool __reg64_bound_s32(s64 a)
1658 return a >= S32_MIN && a <= S32_MAX;
1661 static bool __reg64_bound_u32(u64 a)
1663 return a >= U32_MIN && a <= U32_MAX;
1666 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
1668 __mark_reg32_unbounded(reg);
1669 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
1670 reg->s32_min_value = (s32)reg->smin_value;
1671 reg->s32_max_value = (s32)reg->smax_value;
1673 if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
1674 reg->u32_min_value = (u32)reg->umin_value;
1675 reg->u32_max_value = (u32)reg->umax_value;
1677 reg_bounds_sync(reg);
1680 /* Mark a register as having a completely unknown (scalar) value. */
1681 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1682 struct bpf_reg_state *reg)
1685 * Clear type, id, off, and union(map_ptr, range) and
1686 * padding between 'type' and union
1688 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1689 reg->type = SCALAR_VALUE;
1690 reg->var_off = tnum_unknown;
1692 reg->precise = env->subprog_cnt > 1 || !env->bpf_capable;
1693 __mark_reg_unbounded(reg);
1696 static void mark_reg_unknown(struct bpf_verifier_env *env,
1697 struct bpf_reg_state *regs, u32 regno)
1699 if (WARN_ON(regno >= MAX_BPF_REG)) {
1700 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1701 /* Something bad happened, let's kill all regs except FP */
1702 for (regno = 0; regno < BPF_REG_FP; regno++)
1703 __mark_reg_not_init(env, regs + regno);
1706 __mark_reg_unknown(env, regs + regno);
1709 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1710 struct bpf_reg_state *reg)
1712 __mark_reg_unknown(env, reg);
1713 reg->type = NOT_INIT;
1716 static void mark_reg_not_init(struct bpf_verifier_env *env,
1717 struct bpf_reg_state *regs, u32 regno)
1719 if (WARN_ON(regno >= MAX_BPF_REG)) {
1720 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1721 /* Something bad happened, let's kill all regs except FP */
1722 for (regno = 0; regno < BPF_REG_FP; regno++)
1723 __mark_reg_not_init(env, regs + regno);
1726 __mark_reg_not_init(env, regs + regno);
1729 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
1730 struct bpf_reg_state *regs, u32 regno,
1731 enum bpf_reg_type reg_type,
1732 struct btf *btf, u32 btf_id,
1733 enum bpf_type_flag flag)
1735 if (reg_type == SCALAR_VALUE) {
1736 mark_reg_unknown(env, regs, regno);
1739 mark_reg_known_zero(env, regs, regno);
1740 regs[regno].type = PTR_TO_BTF_ID | flag;
1741 regs[regno].btf = btf;
1742 regs[regno].btf_id = btf_id;
1745 #define DEF_NOT_SUBREG (0)
1746 static void init_reg_state(struct bpf_verifier_env *env,
1747 struct bpf_func_state *state)
1749 struct bpf_reg_state *regs = state->regs;
1752 for (i = 0; i < MAX_BPF_REG; i++) {
1753 mark_reg_not_init(env, regs, i);
1754 regs[i].live = REG_LIVE_NONE;
1755 regs[i].parent = NULL;
1756 regs[i].subreg_def = DEF_NOT_SUBREG;
1760 regs[BPF_REG_FP].type = PTR_TO_STACK;
1761 mark_reg_known_zero(env, regs, BPF_REG_FP);
1762 regs[BPF_REG_FP].frameno = state->frameno;
1765 #define BPF_MAIN_FUNC (-1)
1766 static void init_func_state(struct bpf_verifier_env *env,
1767 struct bpf_func_state *state,
1768 int callsite, int frameno, int subprogno)
1770 state->callsite = callsite;
1771 state->frameno = frameno;
1772 state->subprogno = subprogno;
1773 state->callback_ret_range = tnum_range(0, 0);
1774 init_reg_state(env, state);
1775 mark_verifier_state_scratched(env);
1778 /* Similar to push_stack(), but for async callbacks */
1779 static struct bpf_verifier_state *push_async_cb(struct bpf_verifier_env *env,
1780 int insn_idx, int prev_insn_idx,
1783 struct bpf_verifier_stack_elem *elem;
1784 struct bpf_func_state *frame;
1786 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1790 elem->insn_idx = insn_idx;
1791 elem->prev_insn_idx = prev_insn_idx;
1792 elem->next = env->head;
1793 elem->log_pos = env->log.len_used;
1796 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1798 "The sequence of %d jumps is too complex for async cb.\n",
1802 /* Unlike push_stack() do not copy_verifier_state().
1803 * The caller state doesn't matter.
1804 * This is async callback. It starts in a fresh stack.
1805 * Initialize it similar to do_check_common().
1807 elem->st.branches = 1;
1808 frame = kzalloc(sizeof(*frame), GFP_KERNEL);
1811 init_func_state(env, frame,
1812 BPF_MAIN_FUNC /* callsite */,
1813 0 /* frameno within this callchain */,
1814 subprog /* subprog number within this prog */);
1815 elem->st.frame[0] = frame;
1818 free_verifier_state(env->cur_state, true);
1819 env->cur_state = NULL;
1820 /* pop all elements and return */
1821 while (!pop_stack(env, NULL, NULL, false));
1827 SRC_OP, /* register is used as source operand */
1828 DST_OP, /* register is used as destination operand */
1829 DST_OP_NO_MARK /* same as above, check only, don't mark */
1832 static int cmp_subprogs(const void *a, const void *b)
1834 return ((struct bpf_subprog_info *)a)->start -
1835 ((struct bpf_subprog_info *)b)->start;
1838 static int find_subprog(struct bpf_verifier_env *env, int off)
1840 struct bpf_subprog_info *p;
1842 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1843 sizeof(env->subprog_info[0]), cmp_subprogs);
1846 return p - env->subprog_info;
1850 static int add_subprog(struct bpf_verifier_env *env, int off)
1852 int insn_cnt = env->prog->len;
1855 if (off >= insn_cnt || off < 0) {
1856 verbose(env, "call to invalid destination\n");
1859 ret = find_subprog(env, off);
1862 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1863 verbose(env, "too many subprograms\n");
1866 /* determine subprog starts. The end is one before the next starts */
1867 env->subprog_info[env->subprog_cnt++].start = off;
1868 sort(env->subprog_info, env->subprog_cnt,
1869 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1870 return env->subprog_cnt - 1;
1873 #define MAX_KFUNC_DESCS 256
1874 #define MAX_KFUNC_BTFS 256
1876 struct bpf_kfunc_desc {
1877 struct btf_func_model func_model;
1883 struct bpf_kfunc_btf {
1885 struct module *module;
1889 struct bpf_kfunc_desc_tab {
1890 struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
1894 struct bpf_kfunc_btf_tab {
1895 struct bpf_kfunc_btf descs[MAX_KFUNC_BTFS];
1899 static int kfunc_desc_cmp_by_id_off(const void *a, const void *b)
1901 const struct bpf_kfunc_desc *d0 = a;
1902 const struct bpf_kfunc_desc *d1 = b;
1904 /* func_id is not greater than BTF_MAX_TYPE */
1905 return d0->func_id - d1->func_id ?: d0->offset - d1->offset;
1908 static int kfunc_btf_cmp_by_off(const void *a, const void *b)
1910 const struct bpf_kfunc_btf *d0 = a;
1911 const struct bpf_kfunc_btf *d1 = b;
1913 return d0->offset - d1->offset;
1916 static const struct bpf_kfunc_desc *
1917 find_kfunc_desc(const struct bpf_prog *prog, u32 func_id, u16 offset)
1919 struct bpf_kfunc_desc desc = {
1923 struct bpf_kfunc_desc_tab *tab;
1925 tab = prog->aux->kfunc_tab;
1926 return bsearch(&desc, tab->descs, tab->nr_descs,
1927 sizeof(tab->descs[0]), kfunc_desc_cmp_by_id_off);
1930 static struct btf *__find_kfunc_desc_btf(struct bpf_verifier_env *env,
1933 struct bpf_kfunc_btf kf_btf = { .offset = offset };
1934 struct bpf_kfunc_btf_tab *tab;
1935 struct bpf_kfunc_btf *b;
1940 tab = env->prog->aux->kfunc_btf_tab;
1941 b = bsearch(&kf_btf, tab->descs, tab->nr_descs,
1942 sizeof(tab->descs[0]), kfunc_btf_cmp_by_off);
1944 if (tab->nr_descs == MAX_KFUNC_BTFS) {
1945 verbose(env, "too many different module BTFs\n");
1946 return ERR_PTR(-E2BIG);
1949 if (bpfptr_is_null(env->fd_array)) {
1950 verbose(env, "kfunc offset > 0 without fd_array is invalid\n");
1951 return ERR_PTR(-EPROTO);
1954 if (copy_from_bpfptr_offset(&btf_fd, env->fd_array,
1955 offset * sizeof(btf_fd),
1957 return ERR_PTR(-EFAULT);
1959 btf = btf_get_by_fd(btf_fd);
1961 verbose(env, "invalid module BTF fd specified\n");
1965 if (!btf_is_module(btf)) {
1966 verbose(env, "BTF fd for kfunc is not a module BTF\n");
1968 return ERR_PTR(-EINVAL);
1971 mod = btf_try_get_module(btf);
1974 return ERR_PTR(-ENXIO);
1977 b = &tab->descs[tab->nr_descs++];
1982 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1983 kfunc_btf_cmp_by_off, NULL);
1988 void bpf_free_kfunc_btf_tab(struct bpf_kfunc_btf_tab *tab)
1993 while (tab->nr_descs--) {
1994 module_put(tab->descs[tab->nr_descs].module);
1995 btf_put(tab->descs[tab->nr_descs].btf);
2000 static struct btf *find_kfunc_desc_btf(struct bpf_verifier_env *env, s16 offset)
2004 /* In the future, this can be allowed to increase limit
2005 * of fd index into fd_array, interpreted as u16.
2007 verbose(env, "negative offset disallowed for kernel module function call\n");
2008 return ERR_PTR(-EINVAL);
2011 return __find_kfunc_desc_btf(env, offset);
2013 return btf_vmlinux ?: ERR_PTR(-ENOENT);
2016 static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id, s16 offset)
2018 const struct btf_type *func, *func_proto;
2019 struct bpf_kfunc_btf_tab *btf_tab;
2020 struct bpf_kfunc_desc_tab *tab;
2021 struct bpf_prog_aux *prog_aux;
2022 struct bpf_kfunc_desc *desc;
2023 const char *func_name;
2024 struct btf *desc_btf;
2025 unsigned long call_imm;
2029 prog_aux = env->prog->aux;
2030 tab = prog_aux->kfunc_tab;
2031 btf_tab = prog_aux->kfunc_btf_tab;
2034 verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
2038 if (!env->prog->jit_requested) {
2039 verbose(env, "JIT is required for calling kernel function\n");
2043 if (!bpf_jit_supports_kfunc_call()) {
2044 verbose(env, "JIT does not support calling kernel function\n");
2048 if (!env->prog->gpl_compatible) {
2049 verbose(env, "cannot call kernel function from non-GPL compatible program\n");
2053 tab = kzalloc(sizeof(*tab), GFP_KERNEL);
2056 prog_aux->kfunc_tab = tab;
2059 /* func_id == 0 is always invalid, but instead of returning an error, be
2060 * conservative and wait until the code elimination pass before returning
2061 * error, so that invalid calls that get pruned out can be in BPF programs
2062 * loaded from userspace. It is also required that offset be untouched
2065 if (!func_id && !offset)
2068 if (!btf_tab && offset) {
2069 btf_tab = kzalloc(sizeof(*btf_tab), GFP_KERNEL);
2072 prog_aux->kfunc_btf_tab = btf_tab;
2075 desc_btf = find_kfunc_desc_btf(env, offset);
2076 if (IS_ERR(desc_btf)) {
2077 verbose(env, "failed to find BTF for kernel function\n");
2078 return PTR_ERR(desc_btf);
2081 if (find_kfunc_desc(env->prog, func_id, offset))
2084 if (tab->nr_descs == MAX_KFUNC_DESCS) {
2085 verbose(env, "too many different kernel function calls\n");
2089 func = btf_type_by_id(desc_btf, func_id);
2090 if (!func || !btf_type_is_func(func)) {
2091 verbose(env, "kernel btf_id %u is not a function\n",
2095 func_proto = btf_type_by_id(desc_btf, func->type);
2096 if (!func_proto || !btf_type_is_func_proto(func_proto)) {
2097 verbose(env, "kernel function btf_id %u does not have a valid func_proto\n",
2102 func_name = btf_name_by_offset(desc_btf, func->name_off);
2103 addr = kallsyms_lookup_name(func_name);
2105 verbose(env, "cannot find address for kernel function %s\n",
2110 call_imm = BPF_CALL_IMM(addr);
2111 /* Check whether or not the relative offset overflows desc->imm */
2112 if ((unsigned long)(s32)call_imm != call_imm) {
2113 verbose(env, "address of kernel function %s is out of range\n",
2118 desc = &tab->descs[tab->nr_descs++];
2119 desc->func_id = func_id;
2120 desc->imm = call_imm;
2121 desc->offset = offset;
2122 err = btf_distill_func_proto(&env->log, desc_btf,
2123 func_proto, func_name,
2126 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2127 kfunc_desc_cmp_by_id_off, NULL);
2131 static int kfunc_desc_cmp_by_imm(const void *a, const void *b)
2133 const struct bpf_kfunc_desc *d0 = a;
2134 const struct bpf_kfunc_desc *d1 = b;
2136 if (d0->imm > d1->imm)
2138 else if (d0->imm < d1->imm)
2143 static void sort_kfunc_descs_by_imm(struct bpf_prog *prog)
2145 struct bpf_kfunc_desc_tab *tab;
2147 tab = prog->aux->kfunc_tab;
2151 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2152 kfunc_desc_cmp_by_imm, NULL);
2155 bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
2157 return !!prog->aux->kfunc_tab;
2160 const struct btf_func_model *
2161 bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
2162 const struct bpf_insn *insn)
2164 const struct bpf_kfunc_desc desc = {
2167 const struct bpf_kfunc_desc *res;
2168 struct bpf_kfunc_desc_tab *tab;
2170 tab = prog->aux->kfunc_tab;
2171 res = bsearch(&desc, tab->descs, tab->nr_descs,
2172 sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm);
2174 return res ? &res->func_model : NULL;
2177 static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
2179 struct bpf_subprog_info *subprog = env->subprog_info;
2180 struct bpf_insn *insn = env->prog->insnsi;
2181 int i, ret, insn_cnt = env->prog->len;
2183 /* Add entry function. */
2184 ret = add_subprog(env, 0);
2188 for (i = 0; i < insn_cnt; i++, insn++) {
2189 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
2190 !bpf_pseudo_kfunc_call(insn))
2193 if (!env->bpf_capable) {
2194 verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
2198 if (bpf_pseudo_func(insn) || bpf_pseudo_call(insn))
2199 ret = add_subprog(env, i + insn->imm + 1);
2201 ret = add_kfunc_call(env, insn->imm, insn->off);
2207 /* Add a fake 'exit' subprog which could simplify subprog iteration
2208 * logic. 'subprog_cnt' should not be increased.
2210 subprog[env->subprog_cnt].start = insn_cnt;
2212 if (env->log.level & BPF_LOG_LEVEL2)
2213 for (i = 0; i < env->subprog_cnt; i++)
2214 verbose(env, "func#%d @%d\n", i, subprog[i].start);
2219 static int check_subprogs(struct bpf_verifier_env *env)
2221 int i, subprog_start, subprog_end, off, cur_subprog = 0;
2222 struct bpf_subprog_info *subprog = env->subprog_info;
2223 struct bpf_insn *insn = env->prog->insnsi;
2224 int insn_cnt = env->prog->len;
2226 /* now check that all jumps are within the same subprog */
2227 subprog_start = subprog[cur_subprog].start;
2228 subprog_end = subprog[cur_subprog + 1].start;
2229 for (i = 0; i < insn_cnt; i++) {
2230 u8 code = insn[i].code;
2232 if (code == (BPF_JMP | BPF_CALL) &&
2233 insn[i].imm == BPF_FUNC_tail_call &&
2234 insn[i].src_reg != BPF_PSEUDO_CALL)
2235 subprog[cur_subprog].has_tail_call = true;
2236 if (BPF_CLASS(code) == BPF_LD &&
2237 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
2238 subprog[cur_subprog].has_ld_abs = true;
2239 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
2241 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
2243 off = i + insn[i].off + 1;
2244 if (off < subprog_start || off >= subprog_end) {
2245 verbose(env, "jump out of range from insn %d to %d\n", i, off);
2249 if (i == subprog_end - 1) {
2250 /* to avoid fall-through from one subprog into another
2251 * the last insn of the subprog should be either exit
2252 * or unconditional jump back
2254 if (code != (BPF_JMP | BPF_EXIT) &&
2255 code != (BPF_JMP | BPF_JA)) {
2256 verbose(env, "last insn is not an exit or jmp\n");
2259 subprog_start = subprog_end;
2261 if (cur_subprog < env->subprog_cnt)
2262 subprog_end = subprog[cur_subprog + 1].start;
2268 /* Parentage chain of this register (or stack slot) should take care of all
2269 * issues like callee-saved registers, stack slot allocation time, etc.
2271 static int mark_reg_read(struct bpf_verifier_env *env,
2272 const struct bpf_reg_state *state,
2273 struct bpf_reg_state *parent, u8 flag)
2275 bool writes = parent == state->parent; /* Observe write marks */
2279 /* if read wasn't screened by an earlier write ... */
2280 if (writes && state->live & REG_LIVE_WRITTEN)
2282 if (parent->live & REG_LIVE_DONE) {
2283 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
2284 reg_type_str(env, parent->type),
2285 parent->var_off.value, parent->off);
2288 /* The first condition is more likely to be true than the
2289 * second, checked it first.
2291 if ((parent->live & REG_LIVE_READ) == flag ||
2292 parent->live & REG_LIVE_READ64)
2293 /* The parentage chain never changes and
2294 * this parent was already marked as LIVE_READ.
2295 * There is no need to keep walking the chain again and
2296 * keep re-marking all parents as LIVE_READ.
2297 * This case happens when the same register is read
2298 * multiple times without writes into it in-between.
2299 * Also, if parent has the stronger REG_LIVE_READ64 set,
2300 * then no need to set the weak REG_LIVE_READ32.
2303 /* ... then we depend on parent's value */
2304 parent->live |= flag;
2305 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
2306 if (flag == REG_LIVE_READ64)
2307 parent->live &= ~REG_LIVE_READ32;
2309 parent = state->parent;
2314 if (env->longest_mark_read_walk < cnt)
2315 env->longest_mark_read_walk = cnt;
2319 /* This function is supposed to be used by the following 32-bit optimization
2320 * code only. It returns TRUE if the source or destination register operates
2321 * on 64-bit, otherwise return FALSE.
2323 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
2324 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
2329 class = BPF_CLASS(code);
2331 if (class == BPF_JMP) {
2332 /* BPF_EXIT for "main" will reach here. Return TRUE
2337 if (op == BPF_CALL) {
2338 /* BPF to BPF call will reach here because of marking
2339 * caller saved clobber with DST_OP_NO_MARK for which we
2340 * don't care the register def because they are anyway
2341 * marked as NOT_INIT already.
2343 if (insn->src_reg == BPF_PSEUDO_CALL)
2345 /* Helper call will reach here because of arg type
2346 * check, conservatively return TRUE.
2355 if (class == BPF_ALU64 || class == BPF_JMP ||
2356 /* BPF_END always use BPF_ALU class. */
2357 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
2360 if (class == BPF_ALU || class == BPF_JMP32)
2363 if (class == BPF_LDX) {
2365 return BPF_SIZE(code) == BPF_DW;
2366 /* LDX source must be ptr. */
2370 if (class == BPF_STX) {
2371 /* BPF_STX (including atomic variants) has multiple source
2372 * operands, one of which is a ptr. Check whether the caller is
2375 if (t == SRC_OP && reg->type != SCALAR_VALUE)
2377 return BPF_SIZE(code) == BPF_DW;
2380 if (class == BPF_LD) {
2381 u8 mode = BPF_MODE(code);
2384 if (mode == BPF_IMM)
2387 /* Both LD_IND and LD_ABS return 32-bit data. */
2391 /* Implicit ctx ptr. */
2392 if (regno == BPF_REG_6)
2395 /* Explicit source could be any width. */
2399 if (class == BPF_ST)
2400 /* The only source register for BPF_ST is a ptr. */
2403 /* Conservatively return true at default. */
2407 /* Return the regno defined by the insn, or -1. */
2408 static int insn_def_regno(const struct bpf_insn *insn)
2410 switch (BPF_CLASS(insn->code)) {
2416 if (BPF_MODE(insn->code) == BPF_ATOMIC &&
2417 (insn->imm & BPF_FETCH)) {
2418 if (insn->imm == BPF_CMPXCHG)
2421 return insn->src_reg;
2426 return insn->dst_reg;
2430 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
2431 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
2433 int dst_reg = insn_def_regno(insn);
2438 return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
2441 static void mark_insn_zext(struct bpf_verifier_env *env,
2442 struct bpf_reg_state *reg)
2444 s32 def_idx = reg->subreg_def;
2446 if (def_idx == DEF_NOT_SUBREG)
2449 env->insn_aux_data[def_idx - 1].zext_dst = true;
2450 /* The dst will be zero extended, so won't be sub-register anymore. */
2451 reg->subreg_def = DEF_NOT_SUBREG;
2454 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
2455 enum reg_arg_type t)
2457 struct bpf_verifier_state *vstate = env->cur_state;
2458 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2459 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
2460 struct bpf_reg_state *reg, *regs = state->regs;
2463 if (regno >= MAX_BPF_REG) {
2464 verbose(env, "R%d is invalid\n", regno);
2468 mark_reg_scratched(env, regno);
2471 rw64 = is_reg64(env, insn, regno, reg, t);
2473 /* check whether register used as source operand can be read */
2474 if (reg->type == NOT_INIT) {
2475 verbose(env, "R%d !read_ok\n", regno);
2478 /* We don't need to worry about FP liveness because it's read-only */
2479 if (regno == BPF_REG_FP)
2483 mark_insn_zext(env, reg);
2485 return mark_reg_read(env, reg, reg->parent,
2486 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
2488 /* check whether register used as dest operand can be written to */
2489 if (regno == BPF_REG_FP) {
2490 verbose(env, "frame pointer is read only\n");
2493 reg->live |= REG_LIVE_WRITTEN;
2494 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
2496 mark_reg_unknown(env, regs, regno);
2501 /* for any branch, call, exit record the history of jmps in the given state */
2502 static int push_jmp_history(struct bpf_verifier_env *env,
2503 struct bpf_verifier_state *cur)
2505 u32 cnt = cur->jmp_history_cnt;
2506 struct bpf_idx_pair *p;
2509 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
2512 p[cnt - 1].idx = env->insn_idx;
2513 p[cnt - 1].prev_idx = env->prev_insn_idx;
2514 cur->jmp_history = p;
2515 cur->jmp_history_cnt = cnt;
2519 /* Backtrack one insn at a time. If idx is not at the top of recorded
2520 * history then previous instruction came from straight line execution.
2522 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
2527 if (cnt && st->jmp_history[cnt - 1].idx == i) {
2528 i = st->jmp_history[cnt - 1].prev_idx;
2536 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
2538 const struct btf_type *func;
2539 struct btf *desc_btf;
2541 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
2544 desc_btf = find_kfunc_desc_btf(data, insn->off);
2545 if (IS_ERR(desc_btf))
2548 func = btf_type_by_id(desc_btf, insn->imm);
2549 return btf_name_by_offset(desc_btf, func->name_off);
2552 /* For given verifier state backtrack_insn() is called from the last insn to
2553 * the first insn. Its purpose is to compute a bitmask of registers and
2554 * stack slots that needs precision in the parent verifier state.
2556 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
2557 u32 *reg_mask, u64 *stack_mask)
2559 const struct bpf_insn_cbs cbs = {
2560 .cb_call = disasm_kfunc_name,
2561 .cb_print = verbose,
2562 .private_data = env,
2564 struct bpf_insn *insn = env->prog->insnsi + idx;
2565 u8 class = BPF_CLASS(insn->code);
2566 u8 opcode = BPF_OP(insn->code);
2567 u8 mode = BPF_MODE(insn->code);
2568 u32 dreg = 1u << insn->dst_reg;
2569 u32 sreg = 1u << insn->src_reg;
2572 if (insn->code == 0)
2574 if (env->log.level & BPF_LOG_LEVEL2) {
2575 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
2576 verbose(env, "%d: ", idx);
2577 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
2580 if (class == BPF_ALU || class == BPF_ALU64) {
2581 if (!(*reg_mask & dreg))
2583 if (opcode == BPF_MOV) {
2584 if (BPF_SRC(insn->code) == BPF_X) {
2586 * dreg needs precision after this insn
2587 * sreg needs precision before this insn
2593 * dreg needs precision after this insn.
2594 * Corresponding register is already marked
2595 * as precise=true in this verifier state.
2596 * No further markings in parent are necessary
2601 if (BPF_SRC(insn->code) == BPF_X) {
2603 * both dreg and sreg need precision
2608 * dreg still needs precision before this insn
2611 } else if (class == BPF_LDX) {
2612 if (!(*reg_mask & dreg))
2616 /* scalars can only be spilled into stack w/o losing precision.
2617 * Load from any other memory can be zero extended.
2618 * The desire to keep that precision is already indicated
2619 * by 'precise' mark in corresponding register of this state.
2620 * No further tracking necessary.
2622 if (insn->src_reg != BPF_REG_FP)
2625 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
2626 * that [fp - off] slot contains scalar that needs to be
2627 * tracked with precision
2629 spi = (-insn->off - 1) / BPF_REG_SIZE;
2631 verbose(env, "BUG spi %d\n", spi);
2632 WARN_ONCE(1, "verifier backtracking bug");
2635 *stack_mask |= 1ull << spi;
2636 } else if (class == BPF_STX || class == BPF_ST) {
2637 if (*reg_mask & dreg)
2638 /* stx & st shouldn't be using _scalar_ dst_reg
2639 * to access memory. It means backtracking
2640 * encountered a case of pointer subtraction.
2643 /* scalars can only be spilled into stack */
2644 if (insn->dst_reg != BPF_REG_FP)
2646 spi = (-insn->off - 1) / BPF_REG_SIZE;
2648 verbose(env, "BUG spi %d\n", spi);
2649 WARN_ONCE(1, "verifier backtracking bug");
2652 if (!(*stack_mask & (1ull << spi)))
2654 *stack_mask &= ~(1ull << spi);
2655 if (class == BPF_STX)
2657 } else if (class == BPF_JMP || class == BPF_JMP32) {
2658 if (opcode == BPF_CALL) {
2659 if (insn->src_reg == BPF_PSEUDO_CALL)
2661 /* regular helper call sets R0 */
2663 if (*reg_mask & 0x3f) {
2664 /* if backtracing was looking for registers R1-R5
2665 * they should have been found already.
2667 verbose(env, "BUG regs %x\n", *reg_mask);
2668 WARN_ONCE(1, "verifier backtracking bug");
2671 } else if (opcode == BPF_EXIT) {
2674 } else if (class == BPF_LD) {
2675 if (!(*reg_mask & dreg))
2678 /* It's ld_imm64 or ld_abs or ld_ind.
2679 * For ld_imm64 no further tracking of precision
2680 * into parent is necessary
2682 if (mode == BPF_IND || mode == BPF_ABS)
2683 /* to be analyzed */
2689 /* the scalar precision tracking algorithm:
2690 * . at the start all registers have precise=false.
2691 * . scalar ranges are tracked as normal through alu and jmp insns.
2692 * . once precise value of the scalar register is used in:
2693 * . ptr + scalar alu
2694 * . if (scalar cond K|scalar)
2695 * . helper_call(.., scalar, ...) where ARG_CONST is expected
2696 * backtrack through the verifier states and mark all registers and
2697 * stack slots with spilled constants that these scalar regisers
2698 * should be precise.
2699 * . during state pruning two registers (or spilled stack slots)
2700 * are equivalent if both are not precise.
2702 * Note the verifier cannot simply walk register parentage chain,
2703 * since many different registers and stack slots could have been
2704 * used to compute single precise scalar.
2706 * The approach of starting with precise=true for all registers and then
2707 * backtrack to mark a register as not precise when the verifier detects
2708 * that program doesn't care about specific value (e.g., when helper
2709 * takes register as ARG_ANYTHING parameter) is not safe.
2711 * It's ok to walk single parentage chain of the verifier states.
2712 * It's possible that this backtracking will go all the way till 1st insn.
2713 * All other branches will be explored for needing precision later.
2715 * The backtracking needs to deal with cases like:
2716 * 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)
2719 * if r5 > 0x79f goto pc+7
2720 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
2723 * call bpf_perf_event_output#25
2724 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
2728 * call foo // uses callee's r6 inside to compute r0
2732 * to track above reg_mask/stack_mask needs to be independent for each frame.
2734 * Also if parent's curframe > frame where backtracking started,
2735 * the verifier need to mark registers in both frames, otherwise callees
2736 * may incorrectly prune callers. This is similar to
2737 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
2739 * For now backtracking falls back into conservative marking.
2741 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
2742 struct bpf_verifier_state *st)
2744 struct bpf_func_state *func;
2745 struct bpf_reg_state *reg;
2748 /* big hammer: mark all scalars precise in this path.
2749 * pop_stack may still get !precise scalars.
2751 for (; st; st = st->parent)
2752 for (i = 0; i <= st->curframe; i++) {
2753 func = st->frame[i];
2754 for (j = 0; j < BPF_REG_FP; j++) {
2755 reg = &func->regs[j];
2756 if (reg->type != SCALAR_VALUE)
2758 reg->precise = true;
2760 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
2761 if (!is_spilled_reg(&func->stack[j]))
2763 reg = &func->stack[j].spilled_ptr;
2764 if (reg->type != SCALAR_VALUE)
2766 reg->precise = true;
2771 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
2774 struct bpf_verifier_state *st = env->cur_state;
2775 int first_idx = st->first_insn_idx;
2776 int last_idx = env->insn_idx;
2777 struct bpf_func_state *func;
2778 struct bpf_reg_state *reg;
2779 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
2780 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
2781 bool skip_first = true;
2782 bool new_marks = false;
2785 if (!env->bpf_capable)
2788 func = st->frame[st->curframe];
2790 reg = &func->regs[regno];
2791 if (reg->type != SCALAR_VALUE) {
2792 WARN_ONCE(1, "backtracing misuse");
2799 reg->precise = true;
2803 if (!is_spilled_reg(&func->stack[spi])) {
2807 reg = &func->stack[spi].spilled_ptr;
2808 if (reg->type != SCALAR_VALUE) {
2816 reg->precise = true;
2822 if (!reg_mask && !stack_mask)
2825 DECLARE_BITMAP(mask, 64);
2826 u32 history = st->jmp_history_cnt;
2828 if (env->log.level & BPF_LOG_LEVEL2)
2829 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
2830 for (i = last_idx;;) {
2835 err = backtrack_insn(env, i, ®_mask, &stack_mask);
2837 if (err == -ENOTSUPP) {
2838 mark_all_scalars_precise(env, st);
2843 if (!reg_mask && !stack_mask)
2844 /* Found assignment(s) into tracked register in this state.
2845 * Since this state is already marked, just return.
2846 * Nothing to be tracked further in the parent state.
2851 i = get_prev_insn_idx(st, i, &history);
2852 if (i >= env->prog->len) {
2853 /* This can happen if backtracking reached insn 0
2854 * and there are still reg_mask or stack_mask
2856 * It means the backtracking missed the spot where
2857 * particular register was initialized with a constant.
2859 verbose(env, "BUG backtracking idx %d\n", i);
2860 WARN_ONCE(1, "verifier backtracking bug");
2869 func = st->frame[st->curframe];
2870 bitmap_from_u64(mask, reg_mask);
2871 for_each_set_bit(i, mask, 32) {
2872 reg = &func->regs[i];
2873 if (reg->type != SCALAR_VALUE) {
2874 reg_mask &= ~(1u << i);
2879 reg->precise = true;
2882 bitmap_from_u64(mask, stack_mask);
2883 for_each_set_bit(i, mask, 64) {
2884 if (i >= func->allocated_stack / BPF_REG_SIZE) {
2885 /* the sequence of instructions:
2887 * 3: (7b) *(u64 *)(r3 -8) = r0
2888 * 4: (79) r4 = *(u64 *)(r10 -8)
2889 * doesn't contain jmps. It's backtracked
2890 * as a single block.
2891 * During backtracking insn 3 is not recognized as
2892 * stack access, so at the end of backtracking
2893 * stack slot fp-8 is still marked in stack_mask.
2894 * However the parent state may not have accessed
2895 * fp-8 and it's "unallocated" stack space.
2896 * In such case fallback to conservative.
2898 mark_all_scalars_precise(env, st);
2902 if (!is_spilled_reg(&func->stack[i])) {
2903 stack_mask &= ~(1ull << i);
2906 reg = &func->stack[i].spilled_ptr;
2907 if (reg->type != SCALAR_VALUE) {
2908 stack_mask &= ~(1ull << i);
2913 reg->precise = true;
2915 if (env->log.level & BPF_LOG_LEVEL2) {
2916 verbose(env, "parent %s regs=%x stack=%llx marks:",
2917 new_marks ? "didn't have" : "already had",
2918 reg_mask, stack_mask);
2919 print_verifier_state(env, func, true);
2922 if (!reg_mask && !stack_mask)
2927 last_idx = st->last_insn_idx;
2928 first_idx = st->first_insn_idx;
2933 int mark_chain_precision(struct bpf_verifier_env *env, int regno)
2935 return __mark_chain_precision(env, regno, -1);
2938 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
2940 return __mark_chain_precision(env, -1, spi);
2943 static bool is_spillable_regtype(enum bpf_reg_type type)
2945 switch (base_type(type)) {
2946 case PTR_TO_MAP_VALUE:
2950 case PTR_TO_PACKET_META:
2951 case PTR_TO_PACKET_END:
2952 case PTR_TO_FLOW_KEYS:
2953 case CONST_PTR_TO_MAP:
2955 case PTR_TO_SOCK_COMMON:
2956 case PTR_TO_TCP_SOCK:
2957 case PTR_TO_XDP_SOCK:
2962 case PTR_TO_MAP_KEY:
2969 /* Does this register contain a constant zero? */
2970 static bool register_is_null(struct bpf_reg_state *reg)
2972 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
2975 static bool register_is_const(struct bpf_reg_state *reg)
2977 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
2980 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
2982 return tnum_is_unknown(reg->var_off) &&
2983 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
2984 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
2985 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
2986 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
2989 static bool register_is_bounded(struct bpf_reg_state *reg)
2991 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
2994 static bool __is_pointer_value(bool allow_ptr_leaks,
2995 const struct bpf_reg_state *reg)
2997 if (allow_ptr_leaks)
3000 return reg->type != SCALAR_VALUE;
3003 static void save_register_state(struct bpf_func_state *state,
3004 int spi, struct bpf_reg_state *reg,
3009 state->stack[spi].spilled_ptr = *reg;
3010 if (size == BPF_REG_SIZE)
3011 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
3013 for (i = BPF_REG_SIZE; i > BPF_REG_SIZE - size; i--)
3014 state->stack[spi].slot_type[i - 1] = STACK_SPILL;
3016 /* size < 8 bytes spill */
3018 scrub_spilled_slot(&state->stack[spi].slot_type[i - 1]);
3021 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
3022 * stack boundary and alignment are checked in check_mem_access()
3024 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
3025 /* stack frame we're writing to */
3026 struct bpf_func_state *state,
3027 int off, int size, int value_regno,
3030 struct bpf_func_state *cur; /* state of the current function */
3031 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
3032 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
3033 struct bpf_reg_state *reg = NULL;
3035 err = grow_stack_state(state, round_up(slot + 1, BPF_REG_SIZE));
3038 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
3039 * so it's aligned access and [off, off + size) are within stack limits
3041 if (!env->allow_ptr_leaks &&
3042 state->stack[spi].slot_type[0] == STACK_SPILL &&
3043 size != BPF_REG_SIZE) {
3044 verbose(env, "attempt to corrupt spilled pointer on stack\n");
3048 cur = env->cur_state->frame[env->cur_state->curframe];
3049 if (value_regno >= 0)
3050 reg = &cur->regs[value_regno];
3051 if (!env->bypass_spec_v4) {
3052 bool sanitize = reg && is_spillable_regtype(reg->type);
3054 for (i = 0; i < size; i++) {
3055 if (state->stack[spi].slot_type[i] == STACK_INVALID) {
3062 env->insn_aux_data[insn_idx].sanitize_stack_spill = true;
3065 mark_stack_slot_scratched(env, spi);
3066 if (reg && !(off % BPF_REG_SIZE) && register_is_bounded(reg) &&
3067 !register_is_null(reg) && env->bpf_capable) {
3068 if (dst_reg != BPF_REG_FP) {
3069 /* The backtracking logic can only recognize explicit
3070 * stack slot address like [fp - 8]. Other spill of
3071 * scalar via different register has to be conservative.
3072 * Backtrack from here and mark all registers as precise
3073 * that contributed into 'reg' being a constant.
3075 err = mark_chain_precision(env, value_regno);
3079 save_register_state(state, spi, reg, size);
3080 } else if (reg && is_spillable_regtype(reg->type)) {
3081 /* register containing pointer is being spilled into stack */
3082 if (size != BPF_REG_SIZE) {
3083 verbose_linfo(env, insn_idx, "; ");
3084 verbose(env, "invalid size of register spill\n");
3087 if (state != cur && reg->type == PTR_TO_STACK) {
3088 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
3091 save_register_state(state, spi, reg, size);
3093 u8 type = STACK_MISC;
3095 /* regular write of data into stack destroys any spilled ptr */
3096 state->stack[spi].spilled_ptr.type = NOT_INIT;
3097 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
3098 if (is_spilled_reg(&state->stack[spi]))
3099 for (i = 0; i < BPF_REG_SIZE; i++)
3100 scrub_spilled_slot(&state->stack[spi].slot_type[i]);
3102 /* only mark the slot as written if all 8 bytes were written
3103 * otherwise read propagation may incorrectly stop too soon
3104 * when stack slots are partially written.
3105 * This heuristic means that read propagation will be
3106 * conservative, since it will add reg_live_read marks
3107 * to stack slots all the way to first state when programs
3108 * writes+reads less than 8 bytes
3110 if (size == BPF_REG_SIZE)
3111 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
3113 /* when we zero initialize stack slots mark them as such */
3114 if (reg && register_is_null(reg)) {
3115 /* backtracking doesn't work for STACK_ZERO yet. */
3116 err = mark_chain_precision(env, value_regno);
3122 /* Mark slots affected by this stack write. */
3123 for (i = 0; i < size; i++)
3124 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
3130 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
3131 * known to contain a variable offset.
3132 * This function checks whether the write is permitted and conservatively
3133 * tracks the effects of the write, considering that each stack slot in the
3134 * dynamic range is potentially written to.
3136 * 'off' includes 'regno->off'.
3137 * 'value_regno' can be -1, meaning that an unknown value is being written to
3140 * Spilled pointers in range are not marked as written because we don't know
3141 * what's going to be actually written. This means that read propagation for
3142 * future reads cannot be terminated by this write.
3144 * For privileged programs, uninitialized stack slots are considered
3145 * initialized by this write (even though we don't know exactly what offsets
3146 * are going to be written to). The idea is that we don't want the verifier to
3147 * reject future reads that access slots written to through variable offsets.
3149 static int check_stack_write_var_off(struct bpf_verifier_env *env,
3150 /* func where register points to */
3151 struct bpf_func_state *state,
3152 int ptr_regno, int off, int size,
3153 int value_regno, int insn_idx)
3155 struct bpf_func_state *cur; /* state of the current function */
3156 int min_off, max_off;
3158 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
3159 bool writing_zero = false;
3160 /* set if the fact that we're writing a zero is used to let any
3161 * stack slots remain STACK_ZERO
3163 bool zero_used = false;
3165 cur = env->cur_state->frame[env->cur_state->curframe];
3166 ptr_reg = &cur->regs[ptr_regno];
3167 min_off = ptr_reg->smin_value + off;
3168 max_off = ptr_reg->smax_value + off + size;
3169 if (value_regno >= 0)
3170 value_reg = &cur->regs[value_regno];
3171 if (value_reg && register_is_null(value_reg))
3172 writing_zero = true;
3174 err = grow_stack_state(state, round_up(-min_off, BPF_REG_SIZE));
3179 /* Variable offset writes destroy any spilled pointers in range. */
3180 for (i = min_off; i < max_off; i++) {
3181 u8 new_type, *stype;
3185 spi = slot / BPF_REG_SIZE;
3186 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
3187 mark_stack_slot_scratched(env, spi);
3189 if (!env->allow_ptr_leaks
3190 && *stype != NOT_INIT
3191 && *stype != SCALAR_VALUE) {
3192 /* Reject the write if there's are spilled pointers in
3193 * range. If we didn't reject here, the ptr status
3194 * would be erased below (even though not all slots are
3195 * actually overwritten), possibly opening the door to
3198 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
3203 /* Erase all spilled pointers. */
3204 state->stack[spi].spilled_ptr.type = NOT_INIT;
3206 /* Update the slot type. */
3207 new_type = STACK_MISC;
3208 if (writing_zero && *stype == STACK_ZERO) {
3209 new_type = STACK_ZERO;
3212 /* If the slot is STACK_INVALID, we check whether it's OK to
3213 * pretend that it will be initialized by this write. The slot
3214 * might not actually be written to, and so if we mark it as
3215 * initialized future reads might leak uninitialized memory.
3216 * For privileged programs, we will accept such reads to slots
3217 * that may or may not be written because, if we're reject
3218 * them, the error would be too confusing.
3220 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
3221 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
3228 /* backtracking doesn't work for STACK_ZERO yet. */
3229 err = mark_chain_precision(env, value_regno);
3236 /* When register 'dst_regno' is assigned some values from stack[min_off,
3237 * max_off), we set the register's type according to the types of the
3238 * respective stack slots. If all the stack values are known to be zeros, then
3239 * so is the destination reg. Otherwise, the register is considered to be
3240 * SCALAR. This function does not deal with register filling; the caller must
3241 * ensure that all spilled registers in the stack range have been marked as
3244 static void mark_reg_stack_read(struct bpf_verifier_env *env,
3245 /* func where src register points to */
3246 struct bpf_func_state *ptr_state,
3247 int min_off, int max_off, int dst_regno)
3249 struct bpf_verifier_state *vstate = env->cur_state;
3250 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3255 for (i = min_off; i < max_off; i++) {
3257 spi = slot / BPF_REG_SIZE;
3258 stype = ptr_state->stack[spi].slot_type;
3259 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
3263 if (zeros == max_off - min_off) {
3264 /* any access_size read into register is zero extended,
3265 * so the whole register == const_zero
3267 __mark_reg_const_zero(&state->regs[dst_regno]);
3268 /* backtracking doesn't support STACK_ZERO yet,
3269 * so mark it precise here, so that later
3270 * backtracking can stop here.
3271 * Backtracking may not need this if this register
3272 * doesn't participate in pointer adjustment.
3273 * Forward propagation of precise flag is not
3274 * necessary either. This mark is only to stop
3275 * backtracking. Any register that contributed
3276 * to const 0 was marked precise before spill.
3278 state->regs[dst_regno].precise = true;
3280 /* have read misc data from the stack */
3281 mark_reg_unknown(env, state->regs, dst_regno);
3283 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
3286 /* Read the stack at 'off' and put the results into the register indicated by
3287 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
3290 * 'dst_regno' can be -1, meaning that the read value is not going to a
3293 * The access is assumed to be within the current stack bounds.
3295 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
3296 /* func where src register points to */
3297 struct bpf_func_state *reg_state,
3298 int off, int size, int dst_regno)
3300 struct bpf_verifier_state *vstate = env->cur_state;
3301 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3302 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
3303 struct bpf_reg_state *reg;
3306 stype = reg_state->stack[spi].slot_type;
3307 reg = ®_state->stack[spi].spilled_ptr;
3309 if (is_spilled_reg(®_state->stack[spi])) {
3312 for (i = BPF_REG_SIZE - 1; i > 0 && stype[i - 1] == STACK_SPILL; i--)
3315 if (size != BPF_REG_SIZE || spill_size != BPF_REG_SIZE) {
3316 if (reg->type != SCALAR_VALUE) {
3317 verbose_linfo(env, env->insn_idx, "; ");
3318 verbose(env, "invalid size of register fill\n");
3322 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
3326 if (!(off % BPF_REG_SIZE) && size == spill_size) {
3327 /* The earlier check_reg_arg() has decided the
3328 * subreg_def for this insn. Save it first.
3330 s32 subreg_def = state->regs[dst_regno].subreg_def;
3332 state->regs[dst_regno] = *reg;
3333 state->regs[dst_regno].subreg_def = subreg_def;
3335 for (i = 0; i < size; i++) {
3336 type = stype[(slot - i) % BPF_REG_SIZE];
3337 if (type == STACK_SPILL)
3339 if (type == STACK_MISC)
3341 verbose(env, "invalid read from stack off %d+%d size %d\n",
3345 mark_reg_unknown(env, state->regs, dst_regno);
3347 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
3351 if (dst_regno >= 0) {
3352 /* restore register state from stack */
3353 state->regs[dst_regno] = *reg;
3354 /* mark reg as written since spilled pointer state likely
3355 * has its liveness marks cleared by is_state_visited()
3356 * which resets stack/reg liveness for state transitions
3358 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
3359 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
3360 /* If dst_regno==-1, the caller is asking us whether
3361 * it is acceptable to use this value as a SCALAR_VALUE
3363 * We must not allow unprivileged callers to do that
3364 * with spilled pointers.
3366 verbose(env, "leaking pointer from stack off %d\n",
3370 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
3372 for (i = 0; i < size; i++) {
3373 type = stype[(slot - i) % BPF_REG_SIZE];
3374 if (type == STACK_MISC)
3376 if (type == STACK_ZERO)
3378 verbose(env, "invalid read from stack off %d+%d size %d\n",
3382 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
3384 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
3389 enum bpf_access_src {
3390 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
3391 ACCESS_HELPER = 2, /* the access is performed by a helper */
3394 static int check_stack_range_initialized(struct bpf_verifier_env *env,
3395 int regno, int off, int access_size,
3396 bool zero_size_allowed,
3397 enum bpf_access_src type,
3398 struct bpf_call_arg_meta *meta);
3400 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
3402 return cur_regs(env) + regno;
3405 /* Read the stack at 'ptr_regno + off' and put the result into the register
3407 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
3408 * but not its variable offset.
3409 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
3411 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
3412 * filling registers (i.e. reads of spilled register cannot be detected when
3413 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
3414 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
3415 * offset; for a fixed offset check_stack_read_fixed_off should be used
3418 static int check_stack_read_var_off(struct bpf_verifier_env *env,
3419 int ptr_regno, int off, int size, int dst_regno)
3421 /* The state of the source register. */
3422 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3423 struct bpf_func_state *ptr_state = func(env, reg);
3425 int min_off, max_off;
3427 /* Note that we pass a NULL meta, so raw access will not be permitted.
3429 err = check_stack_range_initialized(env, ptr_regno, off, size,
3430 false, ACCESS_DIRECT, NULL);
3434 min_off = reg->smin_value + off;
3435 max_off = reg->smax_value + off;
3436 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
3440 /* check_stack_read dispatches to check_stack_read_fixed_off or
3441 * check_stack_read_var_off.
3443 * The caller must ensure that the offset falls within the allocated stack
3446 * 'dst_regno' is a register which will receive the value from the stack. It
3447 * can be -1, meaning that the read value is not going to a register.
3449 static int check_stack_read(struct bpf_verifier_env *env,
3450 int ptr_regno, int off, int size,
3453 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3454 struct bpf_func_state *state = func(env, reg);
3456 /* Some accesses are only permitted with a static offset. */
3457 bool var_off = !tnum_is_const(reg->var_off);
3459 /* The offset is required to be static when reads don't go to a
3460 * register, in order to not leak pointers (see
3461 * check_stack_read_fixed_off).
3463 if (dst_regno < 0 && var_off) {
3466 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3467 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
3471 /* Variable offset is prohibited for unprivileged mode for simplicity
3472 * since it requires corresponding support in Spectre masking for stack
3473 * ALU. See also retrieve_ptr_limit().
3475 if (!env->bypass_spec_v1 && var_off) {
3478 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3479 verbose(env, "R%d variable offset stack access prohibited for !root, var_off=%s\n",
3485 off += reg->var_off.value;
3486 err = check_stack_read_fixed_off(env, state, off, size,
3489 /* Variable offset stack reads need more conservative handling
3490 * than fixed offset ones. Note that dst_regno >= 0 on this
3493 err = check_stack_read_var_off(env, ptr_regno, off, size,
3500 /* check_stack_write dispatches to check_stack_write_fixed_off or
3501 * check_stack_write_var_off.
3503 * 'ptr_regno' is the register used as a pointer into the stack.
3504 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
3505 * 'value_regno' is the register whose value we're writing to the stack. It can
3506 * be -1, meaning that we're not writing from a register.
3508 * The caller must ensure that the offset falls within the maximum stack size.
3510 static int check_stack_write(struct bpf_verifier_env *env,
3511 int ptr_regno, int off, int size,
3512 int value_regno, int insn_idx)
3514 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3515 struct bpf_func_state *state = func(env, reg);
3518 if (tnum_is_const(reg->var_off)) {
3519 off += reg->var_off.value;
3520 err = check_stack_write_fixed_off(env, state, off, size,
3521 value_regno, insn_idx);
3523 /* Variable offset stack reads need more conservative handling
3524 * than fixed offset ones.
3526 err = check_stack_write_var_off(env, state,
3527 ptr_regno, off, size,
3528 value_regno, insn_idx);
3533 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
3534 int off, int size, enum bpf_access_type type)
3536 struct bpf_reg_state *regs = cur_regs(env);
3537 struct bpf_map *map = regs[regno].map_ptr;
3538 u32 cap = bpf_map_flags_to_cap(map);
3540 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
3541 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
3542 map->value_size, off, size);
3546 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
3547 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
3548 map->value_size, off, size);
3555 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
3556 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
3557 int off, int size, u32 mem_size,
3558 bool zero_size_allowed)
3560 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
3561 struct bpf_reg_state *reg;
3563 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
3566 reg = &cur_regs(env)[regno];
3567 switch (reg->type) {
3568 case PTR_TO_MAP_KEY:
3569 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
3570 mem_size, off, size);
3572 case PTR_TO_MAP_VALUE:
3573 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
3574 mem_size, off, size);
3577 case PTR_TO_PACKET_META:
3578 case PTR_TO_PACKET_END:
3579 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
3580 off, size, regno, reg->id, off, mem_size);
3584 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
3585 mem_size, off, size);
3591 /* check read/write into a memory region with possible variable offset */
3592 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
3593 int off, int size, u32 mem_size,
3594 bool zero_size_allowed)
3596 struct bpf_verifier_state *vstate = env->cur_state;
3597 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3598 struct bpf_reg_state *reg = &state->regs[regno];
3601 /* We may have adjusted the register pointing to memory region, so we
3602 * need to try adding each of min_value and max_value to off
3603 * to make sure our theoretical access will be safe.
3605 * The minimum value is only important with signed
3606 * comparisons where we can't assume the floor of a
3607 * value is 0. If we are using signed variables for our
3608 * index'es we need to make sure that whatever we use
3609 * will have a set floor within our range.
3611 if (reg->smin_value < 0 &&
3612 (reg->smin_value == S64_MIN ||
3613 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
3614 reg->smin_value + off < 0)) {
3615 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3619 err = __check_mem_access(env, regno, reg->smin_value + off, size,
3620 mem_size, zero_size_allowed);
3622 verbose(env, "R%d min value is outside of the allowed memory range\n",
3627 /* If we haven't set a max value then we need to bail since we can't be
3628 * sure we won't do bad things.
3629 * If reg->umax_value + off could overflow, treat that as unbounded too.
3631 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
3632 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
3636 err = __check_mem_access(env, regno, reg->umax_value + off, size,
3637 mem_size, zero_size_allowed);
3639 verbose(env, "R%d max value is outside of the allowed memory range\n",
3647 static int __check_ptr_off_reg(struct bpf_verifier_env *env,
3648 const struct bpf_reg_state *reg, int regno,
3651 /* Access to this pointer-typed register or passing it to a helper
3652 * is only allowed in its original, unmodified form.
3656 verbose(env, "negative offset %s ptr R%d off=%d disallowed\n",
3657 reg_type_str(env, reg->type), regno, reg->off);
3661 if (!fixed_off_ok && reg->off) {
3662 verbose(env, "dereference of modified %s ptr R%d off=%d disallowed\n",
3663 reg_type_str(env, reg->type), regno, reg->off);
3667 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3670 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3671 verbose(env, "variable %s access var_off=%s disallowed\n",
3672 reg_type_str(env, reg->type), tn_buf);
3679 int check_ptr_off_reg(struct bpf_verifier_env *env,
3680 const struct bpf_reg_state *reg, int regno)
3682 return __check_ptr_off_reg(env, reg, regno, false);
3685 static int map_kptr_match_type(struct bpf_verifier_env *env,
3686 struct bpf_map_value_off_desc *off_desc,
3687 struct bpf_reg_state *reg, u32 regno)
3689 const char *targ_name = kernel_type_name(off_desc->kptr.btf, off_desc->kptr.btf_id);
3690 int perm_flags = PTR_MAYBE_NULL;
3691 const char *reg_name = "";
3693 /* Only unreferenced case accepts untrusted pointers */
3694 if (off_desc->type == BPF_KPTR_UNREF)
3695 perm_flags |= PTR_UNTRUSTED;
3697 if (base_type(reg->type) != PTR_TO_BTF_ID || (type_flag(reg->type) & ~perm_flags))
3700 if (!btf_is_kernel(reg->btf)) {
3701 verbose(env, "R%d must point to kernel BTF\n", regno);
3704 /* We need to verify reg->type and reg->btf, before accessing reg->btf */
3705 reg_name = kernel_type_name(reg->btf, reg->btf_id);
3707 /* For ref_ptr case, release function check should ensure we get one
3708 * referenced PTR_TO_BTF_ID, and that its fixed offset is 0. For the
3709 * normal store of unreferenced kptr, we must ensure var_off is zero.
3710 * Since ref_ptr cannot be accessed directly by BPF insns, checks for
3711 * reg->off and reg->ref_obj_id are not needed here.
3713 if (__check_ptr_off_reg(env, reg, regno, true))
3716 /* A full type match is needed, as BTF can be vmlinux or module BTF, and
3717 * we also need to take into account the reg->off.
3719 * We want to support cases like:
3727 * v = func(); // PTR_TO_BTF_ID
3728 * val->foo = v; // reg->off is zero, btf and btf_id match type
3729 * val->bar = &v->br; // reg->off is still zero, but we need to retry with
3730 * // first member type of struct after comparison fails
3731 * val->baz = &v->bz; // reg->off is non-zero, so struct needs to be walked
3734 * In the kptr_ref case, check_func_arg_reg_off already ensures reg->off
3735 * is zero. We must also ensure that btf_struct_ids_match does not walk
3736 * the struct to match type against first member of struct, i.e. reject
3737 * second case from above. Hence, when type is BPF_KPTR_REF, we set
3738 * strict mode to true for type match.
3740 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
3741 off_desc->kptr.btf, off_desc->kptr.btf_id,
3742 off_desc->type == BPF_KPTR_REF))
3746 verbose(env, "invalid kptr access, R%d type=%s%s ", regno,
3747 reg_type_str(env, reg->type), reg_name);
3748 verbose(env, "expected=%s%s", reg_type_str(env, PTR_TO_BTF_ID), targ_name);
3749 if (off_desc->type == BPF_KPTR_UNREF)
3750 verbose(env, " or %s%s\n", reg_type_str(env, PTR_TO_BTF_ID | PTR_UNTRUSTED),
3757 static int check_map_kptr_access(struct bpf_verifier_env *env, u32 regno,
3758 int value_regno, int insn_idx,
3759 struct bpf_map_value_off_desc *off_desc)
3761 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
3762 int class = BPF_CLASS(insn->code);
3763 struct bpf_reg_state *val_reg;
3765 /* Things we already checked for in check_map_access and caller:
3766 * - Reject cases where variable offset may touch kptr
3767 * - size of access (must be BPF_DW)
3768 * - tnum_is_const(reg->var_off)
3769 * - off_desc->offset == off + reg->var_off.value
3771 /* Only BPF_[LDX,STX,ST] | BPF_MEM | BPF_DW is supported */
3772 if (BPF_MODE(insn->code) != BPF_MEM) {
3773 verbose(env, "kptr in map can only be accessed using BPF_MEM instruction mode\n");
3777 /* We only allow loading referenced kptr, since it will be marked as
3778 * untrusted, similar to unreferenced kptr.
3780 if (class != BPF_LDX && off_desc->type == BPF_KPTR_REF) {
3781 verbose(env, "store to referenced kptr disallowed\n");
3785 if (class == BPF_LDX) {
3786 val_reg = reg_state(env, value_regno);
3787 /* We can simply mark the value_regno receiving the pointer
3788 * value from map as PTR_TO_BTF_ID, with the correct type.
3790 mark_btf_ld_reg(env, cur_regs(env), value_regno, PTR_TO_BTF_ID, off_desc->kptr.btf,
3791 off_desc->kptr.btf_id, PTR_MAYBE_NULL | PTR_UNTRUSTED);
3792 /* For mark_ptr_or_null_reg */
3793 val_reg->id = ++env->id_gen;
3794 } else if (class == BPF_STX) {
3795 val_reg = reg_state(env, value_regno);
3796 if (!register_is_null(val_reg) &&
3797 map_kptr_match_type(env, off_desc, val_reg, value_regno))
3799 } else if (class == BPF_ST) {
3801 verbose(env, "BPF_ST imm must be 0 when storing to kptr at off=%u\n",
3806 verbose(env, "kptr in map can only be accessed using BPF_LDX/BPF_STX/BPF_ST\n");
3812 /* check read/write into a map element with possible variable offset */
3813 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
3814 int off, int size, bool zero_size_allowed,
3815 enum bpf_access_src src)
3817 struct bpf_verifier_state *vstate = env->cur_state;
3818 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3819 struct bpf_reg_state *reg = &state->regs[regno];
3820 struct bpf_map *map = reg->map_ptr;
3823 err = check_mem_region_access(env, regno, off, size, map->value_size,
3828 if (map_value_has_spin_lock(map)) {
3829 u32 lock = map->spin_lock_off;
3831 /* if any part of struct bpf_spin_lock can be touched by
3832 * load/store reject this program.
3833 * To check that [x1, x2) overlaps with [y1, y2)
3834 * it is sufficient to check x1 < y2 && y1 < x2.
3836 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
3837 lock < reg->umax_value + off + size) {
3838 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
3842 if (map_value_has_timer(map)) {
3843 u32 t = map->timer_off;
3845 if (reg->smin_value + off < t + sizeof(struct bpf_timer) &&
3846 t < reg->umax_value + off + size) {
3847 verbose(env, "bpf_timer cannot be accessed directly by load/store\n");
3851 if (map_value_has_kptrs(map)) {
3852 struct bpf_map_value_off *tab = map->kptr_off_tab;
3855 for (i = 0; i < tab->nr_off; i++) {
3856 u32 p = tab->off[i].offset;
3858 if (reg->smin_value + off < p + sizeof(u64) &&
3859 p < reg->umax_value + off + size) {
3860 if (src != ACCESS_DIRECT) {
3861 verbose(env, "kptr cannot be accessed indirectly by helper\n");
3864 if (!tnum_is_const(reg->var_off)) {
3865 verbose(env, "kptr access cannot have variable offset\n");
3868 if (p != off + reg->var_off.value) {
3869 verbose(env, "kptr access misaligned expected=%u off=%llu\n",
3870 p, off + reg->var_off.value);
3873 if (size != bpf_size_to_bytes(BPF_DW)) {
3874 verbose(env, "kptr access size must be BPF_DW\n");
3884 #define MAX_PACKET_OFF 0xffff
3886 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
3887 const struct bpf_call_arg_meta *meta,
3888 enum bpf_access_type t)
3890 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
3892 switch (prog_type) {
3893 /* Program types only with direct read access go here! */
3894 case BPF_PROG_TYPE_LWT_IN:
3895 case BPF_PROG_TYPE_LWT_OUT:
3896 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
3897 case BPF_PROG_TYPE_SK_REUSEPORT:
3898 case BPF_PROG_TYPE_FLOW_DISSECTOR:
3899 case BPF_PROG_TYPE_CGROUP_SKB:
3904 /* Program types with direct read + write access go here! */
3905 case BPF_PROG_TYPE_SCHED_CLS:
3906 case BPF_PROG_TYPE_SCHED_ACT:
3907 case BPF_PROG_TYPE_XDP:
3908 case BPF_PROG_TYPE_LWT_XMIT:
3909 case BPF_PROG_TYPE_SK_SKB:
3910 case BPF_PROG_TYPE_SK_MSG:
3912 return meta->pkt_access;
3914 env->seen_direct_write = true;
3917 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
3919 env->seen_direct_write = true;
3928 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
3929 int size, bool zero_size_allowed)
3931 struct bpf_reg_state *regs = cur_regs(env);
3932 struct bpf_reg_state *reg = ®s[regno];
3935 /* We may have added a variable offset to the packet pointer; but any
3936 * reg->range we have comes after that. We are only checking the fixed
3940 /* We don't allow negative numbers, because we aren't tracking enough
3941 * detail to prove they're safe.
3943 if (reg->smin_value < 0) {
3944 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3949 err = reg->range < 0 ? -EINVAL :
3950 __check_mem_access(env, regno, off, size, reg->range,
3953 verbose(env, "R%d offset is outside of the packet\n", regno);
3957 /* __check_mem_access has made sure "off + size - 1" is within u16.
3958 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
3959 * otherwise find_good_pkt_pointers would have refused to set range info
3960 * that __check_mem_access would have rejected this pkt access.
3961 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
3963 env->prog->aux->max_pkt_offset =
3964 max_t(u32, env->prog->aux->max_pkt_offset,
3965 off + reg->umax_value + size - 1);
3970 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
3971 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
3972 enum bpf_access_type t, enum bpf_reg_type *reg_type,
3973 struct btf **btf, u32 *btf_id)
3975 struct bpf_insn_access_aux info = {
3976 .reg_type = *reg_type,
3980 if (env->ops->is_valid_access &&
3981 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
3982 /* A non zero info.ctx_field_size indicates that this field is a
3983 * candidate for later verifier transformation to load the whole
3984 * field and then apply a mask when accessed with a narrower
3985 * access than actual ctx access size. A zero info.ctx_field_size
3986 * will only allow for whole field access and rejects any other
3987 * type of narrower access.
3989 *reg_type = info.reg_type;
3991 if (base_type(*reg_type) == PTR_TO_BTF_ID) {
3993 *btf_id = info.btf_id;
3995 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
3997 /* remember the offset of last byte accessed in ctx */
3998 if (env->prog->aux->max_ctx_offset < off + size)
3999 env->prog->aux->max_ctx_offset = off + size;
4003 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
4007 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
4010 if (size < 0 || off < 0 ||
4011 (u64)off + size > sizeof(struct bpf_flow_keys)) {
4012 verbose(env, "invalid access to flow keys off=%d size=%d\n",
4019 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
4020 u32 regno, int off, int size,
4021 enum bpf_access_type t)
4023 struct bpf_reg_state *regs = cur_regs(env);
4024 struct bpf_reg_state *reg = ®s[regno];
4025 struct bpf_insn_access_aux info = {};
4028 if (reg->smin_value < 0) {
4029 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
4034 switch (reg->type) {
4035 case PTR_TO_SOCK_COMMON:
4036 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
4039 valid = bpf_sock_is_valid_access(off, size, t, &info);
4041 case PTR_TO_TCP_SOCK:
4042 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
4044 case PTR_TO_XDP_SOCK:
4045 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
4053 env->insn_aux_data[insn_idx].ctx_field_size =
4054 info.ctx_field_size;
4058 verbose(env, "R%d invalid %s access off=%d size=%d\n",
4059 regno, reg_type_str(env, reg->type), off, size);
4064 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
4066 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
4069 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
4071 const struct bpf_reg_state *reg = reg_state(env, regno);
4073 return reg->type == PTR_TO_CTX;
4076 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
4078 const struct bpf_reg_state *reg = reg_state(env, regno);
4080 return type_is_sk_pointer(reg->type);
4083 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
4085 const struct bpf_reg_state *reg = reg_state(env, regno);
4087 return type_is_pkt_pointer(reg->type);
4090 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
4092 const struct bpf_reg_state *reg = reg_state(env, regno);
4094 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
4095 return reg->type == PTR_TO_FLOW_KEYS;
4098 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
4099 const struct bpf_reg_state *reg,
4100 int off, int size, bool strict)
4102 struct tnum reg_off;
4105 /* Byte size accesses are always allowed. */
4106 if (!strict || size == 1)
4109 /* For platforms that do not have a Kconfig enabling
4110 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
4111 * NET_IP_ALIGN is universally set to '2'. And on platforms
4112 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
4113 * to this code only in strict mode where we want to emulate
4114 * the NET_IP_ALIGN==2 checking. Therefore use an
4115 * unconditional IP align value of '2'.
4119 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
4120 if (!tnum_is_aligned(reg_off, size)) {
4123 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4125 "misaligned packet access off %d+%s+%d+%d size %d\n",
4126 ip_align, tn_buf, reg->off, off, size);
4133 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
4134 const struct bpf_reg_state *reg,
4135 const char *pointer_desc,
4136 int off, int size, bool strict)
4138 struct tnum reg_off;
4140 /* Byte size accesses are always allowed. */
4141 if (!strict || size == 1)
4144 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
4145 if (!tnum_is_aligned(reg_off, size)) {
4148 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4149 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
4150 pointer_desc, tn_buf, reg->off, off, size);
4157 static int check_ptr_alignment(struct bpf_verifier_env *env,
4158 const struct bpf_reg_state *reg, int off,
4159 int size, bool strict_alignment_once)
4161 bool strict = env->strict_alignment || strict_alignment_once;
4162 const char *pointer_desc = "";
4164 switch (reg->type) {
4166 case PTR_TO_PACKET_META:
4167 /* Special case, because of NET_IP_ALIGN. Given metadata sits
4168 * right in front, treat it the very same way.
4170 return check_pkt_ptr_alignment(env, reg, off, size, strict);
4171 case PTR_TO_FLOW_KEYS:
4172 pointer_desc = "flow keys ";
4174 case PTR_TO_MAP_KEY:
4175 pointer_desc = "key ";
4177 case PTR_TO_MAP_VALUE:
4178 pointer_desc = "value ";
4181 pointer_desc = "context ";
4184 pointer_desc = "stack ";
4185 /* The stack spill tracking logic in check_stack_write_fixed_off()
4186 * and check_stack_read_fixed_off() relies on stack accesses being
4192 pointer_desc = "sock ";
4194 case PTR_TO_SOCK_COMMON:
4195 pointer_desc = "sock_common ";
4197 case PTR_TO_TCP_SOCK:
4198 pointer_desc = "tcp_sock ";
4200 case PTR_TO_XDP_SOCK:
4201 pointer_desc = "xdp_sock ";
4206 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
4210 static int update_stack_depth(struct bpf_verifier_env *env,
4211 const struct bpf_func_state *func,
4214 u16 stack = env->subprog_info[func->subprogno].stack_depth;
4219 /* update known max for given subprogram */
4220 env->subprog_info[func->subprogno].stack_depth = -off;
4224 /* starting from main bpf function walk all instructions of the function
4225 * and recursively walk all callees that given function can call.
4226 * Ignore jump and exit insns.
4227 * Since recursion is prevented by check_cfg() this algorithm
4228 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
4230 static int check_max_stack_depth(struct bpf_verifier_env *env)
4232 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
4233 struct bpf_subprog_info *subprog = env->subprog_info;
4234 struct bpf_insn *insn = env->prog->insnsi;
4235 bool tail_call_reachable = false;
4236 int ret_insn[MAX_CALL_FRAMES];
4237 int ret_prog[MAX_CALL_FRAMES];
4241 /* protect against potential stack overflow that might happen when
4242 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
4243 * depth for such case down to 256 so that the worst case scenario
4244 * would result in 8k stack size (32 which is tailcall limit * 256 =
4247 * To get the idea what might happen, see an example:
4248 * func1 -> sub rsp, 128
4249 * subfunc1 -> sub rsp, 256
4250 * tailcall1 -> add rsp, 256
4251 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
4252 * subfunc2 -> sub rsp, 64
4253 * subfunc22 -> sub rsp, 128
4254 * tailcall2 -> add rsp, 128
4255 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
4257 * tailcall will unwind the current stack frame but it will not get rid
4258 * of caller's stack as shown on the example above.
4260 if (idx && subprog[idx].has_tail_call && depth >= 256) {
4262 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
4266 /* round up to 32-bytes, since this is granularity
4267 * of interpreter stack size
4269 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
4270 if (depth > MAX_BPF_STACK) {
4271 verbose(env, "combined stack size of %d calls is %d. Too large\n",
4276 subprog_end = subprog[idx + 1].start;
4277 for (; i < subprog_end; i++) {
4280 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
4282 /* remember insn and function to return to */
4283 ret_insn[frame] = i + 1;
4284 ret_prog[frame] = idx;
4286 /* find the callee */
4287 next_insn = i + insn[i].imm + 1;
4288 idx = find_subprog(env, next_insn);
4290 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
4294 if (subprog[idx].is_async_cb) {
4295 if (subprog[idx].has_tail_call) {
4296 verbose(env, "verifier bug. subprog has tail_call and async cb\n");
4299 /* async callbacks don't increase bpf prog stack size */
4304 if (subprog[idx].has_tail_call)
4305 tail_call_reachable = true;
4308 if (frame >= MAX_CALL_FRAMES) {
4309 verbose(env, "the call stack of %d frames is too deep !\n",
4315 /* if tail call got detected across bpf2bpf calls then mark each of the
4316 * currently present subprog frames as tail call reachable subprogs;
4317 * this info will be utilized by JIT so that we will be preserving the
4318 * tail call counter throughout bpf2bpf calls combined with tailcalls
4320 if (tail_call_reachable)
4321 for (j = 0; j < frame; j++)
4322 subprog[ret_prog[j]].tail_call_reachable = true;
4323 if (subprog[0].tail_call_reachable)
4324 env->prog->aux->tail_call_reachable = true;
4326 /* end of for() loop means the last insn of the 'subprog'
4327 * was reached. Doesn't matter whether it was JA or EXIT
4331 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
4333 i = ret_insn[frame];
4334 idx = ret_prog[frame];
4338 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
4339 static int get_callee_stack_depth(struct bpf_verifier_env *env,
4340 const struct bpf_insn *insn, int idx)
4342 int start = idx + insn->imm + 1, subprog;
4344 subprog = find_subprog(env, start);
4346 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
4350 return env->subprog_info[subprog].stack_depth;
4354 static int __check_buffer_access(struct bpf_verifier_env *env,
4355 const char *buf_info,
4356 const struct bpf_reg_state *reg,
4357 int regno, int off, int size)
4361 "R%d invalid %s buffer access: off=%d, size=%d\n",
4362 regno, buf_info, off, size);
4365 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4368 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4370 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
4371 regno, off, tn_buf);
4378 static int check_tp_buffer_access(struct bpf_verifier_env *env,
4379 const struct bpf_reg_state *reg,
4380 int regno, int off, int size)
4384 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
4388 if (off + size > env->prog->aux->max_tp_access)
4389 env->prog->aux->max_tp_access = off + size;
4394 static int check_buffer_access(struct bpf_verifier_env *env,
4395 const struct bpf_reg_state *reg,
4396 int regno, int off, int size,
4397 bool zero_size_allowed,
4400 const char *buf_info = type_is_rdonly_mem(reg->type) ? "rdonly" : "rdwr";
4403 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
4407 if (off + size > *max_access)
4408 *max_access = off + size;
4413 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
4414 static void zext_32_to_64(struct bpf_reg_state *reg)
4416 reg->var_off = tnum_subreg(reg->var_off);
4417 __reg_assign_32_into_64(reg);
4420 /* truncate register to smaller size (in bytes)
4421 * must be called with size < BPF_REG_SIZE
4423 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
4427 /* clear high bits in bit representation */
4428 reg->var_off = tnum_cast(reg->var_off, size);
4430 /* fix arithmetic bounds */
4431 mask = ((u64)1 << (size * 8)) - 1;
4432 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
4433 reg->umin_value &= mask;
4434 reg->umax_value &= mask;
4436 reg->umin_value = 0;
4437 reg->umax_value = mask;
4439 reg->smin_value = reg->umin_value;
4440 reg->smax_value = reg->umax_value;
4442 /* If size is smaller than 32bit register the 32bit register
4443 * values are also truncated so we push 64-bit bounds into
4444 * 32-bit bounds. Above were truncated < 32-bits already.
4448 __reg_combine_64_into_32(reg);
4451 static bool bpf_map_is_rdonly(const struct bpf_map *map)
4453 /* A map is considered read-only if the following condition are true:
4455 * 1) BPF program side cannot change any of the map content. The
4456 * BPF_F_RDONLY_PROG flag is throughout the lifetime of a map
4457 * and was set at map creation time.
4458 * 2) The map value(s) have been initialized from user space by a
4459 * loader and then "frozen", such that no new map update/delete
4460 * operations from syscall side are possible for the rest of
4461 * the map's lifetime from that point onwards.
4462 * 3) Any parallel/pending map update/delete operations from syscall
4463 * side have been completed. Only after that point, it's safe to
4464 * assume that map value(s) are immutable.
4466 return (map->map_flags & BPF_F_RDONLY_PROG) &&
4467 READ_ONCE(map->frozen) &&
4468 !bpf_map_write_active(map);
4471 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
4477 err = map->ops->map_direct_value_addr(map, &addr, off);
4480 ptr = (void *)(long)addr + off;
4484 *val = (u64)*(u8 *)ptr;
4487 *val = (u64)*(u16 *)ptr;
4490 *val = (u64)*(u32 *)ptr;
4501 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
4502 struct bpf_reg_state *regs,
4503 int regno, int off, int size,
4504 enum bpf_access_type atype,
4507 struct bpf_reg_state *reg = regs + regno;
4508 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
4509 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
4510 enum bpf_type_flag flag = 0;
4516 "R%d is ptr_%s invalid negative access: off=%d\n",
4520 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4523 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4525 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
4526 regno, tname, off, tn_buf);
4530 if (reg->type & MEM_USER) {
4532 "R%d is ptr_%s access user memory: off=%d\n",
4537 if (reg->type & MEM_PERCPU) {
4539 "R%d is ptr_%s access percpu memory: off=%d\n",
4544 if (env->ops->btf_struct_access) {
4545 ret = env->ops->btf_struct_access(&env->log, reg->btf, t,
4546 off, size, atype, &btf_id, &flag);
4548 if (atype != BPF_READ) {
4549 verbose(env, "only read is supported\n");
4553 ret = btf_struct_access(&env->log, reg->btf, t, off, size,
4554 atype, &btf_id, &flag);
4560 /* If this is an untrusted pointer, all pointers formed by walking it
4561 * also inherit the untrusted flag.
4563 if (type_flag(reg->type) & PTR_UNTRUSTED)
4564 flag |= PTR_UNTRUSTED;
4566 if (atype == BPF_READ && value_regno >= 0)
4567 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id, flag);
4572 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
4573 struct bpf_reg_state *regs,
4574 int regno, int off, int size,
4575 enum bpf_access_type atype,
4578 struct bpf_reg_state *reg = regs + regno;
4579 struct bpf_map *map = reg->map_ptr;
4580 enum bpf_type_flag flag = 0;
4581 const struct btf_type *t;
4587 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
4591 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
4592 verbose(env, "map_ptr access not supported for map type %d\n",
4597 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
4598 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
4600 if (!env->allow_ptr_to_map_access) {
4602 "%s access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
4608 verbose(env, "R%d is %s invalid negative access: off=%d\n",
4613 if (atype != BPF_READ) {
4614 verbose(env, "only read from %s is supported\n", tname);
4618 ret = btf_struct_access(&env->log, btf_vmlinux, t, off, size, atype, &btf_id, &flag);
4622 if (value_regno >= 0)
4623 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id, flag);
4628 /* Check that the stack access at the given offset is within bounds. The
4629 * maximum valid offset is -1.
4631 * The minimum valid offset is -MAX_BPF_STACK for writes, and
4632 * -state->allocated_stack for reads.
4634 static int check_stack_slot_within_bounds(int off,
4635 struct bpf_func_state *state,
4636 enum bpf_access_type t)
4641 min_valid_off = -MAX_BPF_STACK;
4643 min_valid_off = -state->allocated_stack;
4645 if (off < min_valid_off || off > -1)
4650 /* Check that the stack access at 'regno + off' falls within the maximum stack
4653 * 'off' includes `regno->offset`, but not its dynamic part (if any).
4655 static int check_stack_access_within_bounds(
4656 struct bpf_verifier_env *env,
4657 int regno, int off, int access_size,
4658 enum bpf_access_src src, enum bpf_access_type type)
4660 struct bpf_reg_state *regs = cur_regs(env);
4661 struct bpf_reg_state *reg = regs + regno;
4662 struct bpf_func_state *state = func(env, reg);
4663 int min_off, max_off;
4667 if (src == ACCESS_HELPER)
4668 /* We don't know if helpers are reading or writing (or both). */
4669 err_extra = " indirect access to";
4670 else if (type == BPF_READ)
4671 err_extra = " read from";
4673 err_extra = " write to";
4675 if (tnum_is_const(reg->var_off)) {
4676 min_off = reg->var_off.value + off;
4677 if (access_size > 0)
4678 max_off = min_off + access_size - 1;
4682 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
4683 reg->smin_value <= -BPF_MAX_VAR_OFF) {
4684 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
4688 min_off = reg->smin_value + off;
4689 if (access_size > 0)
4690 max_off = reg->smax_value + off + access_size - 1;
4695 err = check_stack_slot_within_bounds(min_off, state, type);
4697 err = check_stack_slot_within_bounds(max_off, state, type);
4700 if (tnum_is_const(reg->var_off)) {
4701 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
4702 err_extra, regno, off, access_size);
4706 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4707 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
4708 err_extra, regno, tn_buf, access_size);
4714 /* check whether memory at (regno + off) is accessible for t = (read | write)
4715 * if t==write, value_regno is a register which value is stored into memory
4716 * if t==read, value_regno is a register which will receive the value from memory
4717 * if t==write && value_regno==-1, some unknown value is stored into memory
4718 * if t==read && value_regno==-1, don't care what we read from memory
4720 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
4721 int off, int bpf_size, enum bpf_access_type t,
4722 int value_regno, bool strict_alignment_once)
4724 struct bpf_reg_state *regs = cur_regs(env);
4725 struct bpf_reg_state *reg = regs + regno;
4726 struct bpf_func_state *state;
4729 size = bpf_size_to_bytes(bpf_size);
4733 /* alignment checks will add in reg->off themselves */
4734 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
4738 /* for access checks, reg->off is just part of off */
4741 if (reg->type == PTR_TO_MAP_KEY) {
4742 if (t == BPF_WRITE) {
4743 verbose(env, "write to change key R%d not allowed\n", regno);
4747 err = check_mem_region_access(env, regno, off, size,
4748 reg->map_ptr->key_size, false);
4751 if (value_regno >= 0)
4752 mark_reg_unknown(env, regs, value_regno);
4753 } else if (reg->type == PTR_TO_MAP_VALUE) {
4754 struct bpf_map_value_off_desc *kptr_off_desc = NULL;
4756 if (t == BPF_WRITE && value_regno >= 0 &&
4757 is_pointer_value(env, value_regno)) {
4758 verbose(env, "R%d leaks addr into map\n", value_regno);
4761 err = check_map_access_type(env, regno, off, size, t);
4764 err = check_map_access(env, regno, off, size, false, ACCESS_DIRECT);
4767 if (tnum_is_const(reg->var_off))
4768 kptr_off_desc = bpf_map_kptr_off_contains(reg->map_ptr,
4769 off + reg->var_off.value);
4770 if (kptr_off_desc) {
4771 err = check_map_kptr_access(env, regno, value_regno, insn_idx,
4773 } else if (t == BPF_READ && value_regno >= 0) {
4774 struct bpf_map *map = reg->map_ptr;
4776 /* if map is read-only, track its contents as scalars */
4777 if (tnum_is_const(reg->var_off) &&
4778 bpf_map_is_rdonly(map) &&
4779 map->ops->map_direct_value_addr) {
4780 int map_off = off + reg->var_off.value;
4783 err = bpf_map_direct_read(map, map_off, size,
4788 regs[value_regno].type = SCALAR_VALUE;
4789 __mark_reg_known(®s[value_regno], val);
4791 mark_reg_unknown(env, regs, value_regno);
4794 } else if (base_type(reg->type) == PTR_TO_MEM) {
4795 bool rdonly_mem = type_is_rdonly_mem(reg->type);
4797 if (type_may_be_null(reg->type)) {
4798 verbose(env, "R%d invalid mem access '%s'\n", regno,
4799 reg_type_str(env, reg->type));
4803 if (t == BPF_WRITE && rdonly_mem) {
4804 verbose(env, "R%d cannot write into %s\n",
4805 regno, reg_type_str(env, reg->type));
4809 if (t == BPF_WRITE && value_regno >= 0 &&
4810 is_pointer_value(env, value_regno)) {
4811 verbose(env, "R%d leaks addr into mem\n", value_regno);
4815 err = check_mem_region_access(env, regno, off, size,
4816 reg->mem_size, false);
4817 if (!err && value_regno >= 0 && (t == BPF_READ || rdonly_mem))
4818 mark_reg_unknown(env, regs, value_regno);
4819 } else if (reg->type == PTR_TO_CTX) {
4820 enum bpf_reg_type reg_type = SCALAR_VALUE;
4821 struct btf *btf = NULL;
4824 if (t == BPF_WRITE && value_regno >= 0 &&
4825 is_pointer_value(env, value_regno)) {
4826 verbose(env, "R%d leaks addr into ctx\n", value_regno);
4830 err = check_ptr_off_reg(env, reg, regno);
4834 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf,
4837 verbose_linfo(env, insn_idx, "; ");
4838 if (!err && t == BPF_READ && value_regno >= 0) {
4839 /* ctx access returns either a scalar, or a
4840 * PTR_TO_PACKET[_META,_END]. In the latter
4841 * case, we know the offset is zero.
4843 if (reg_type == SCALAR_VALUE) {
4844 mark_reg_unknown(env, regs, value_regno);
4846 mark_reg_known_zero(env, regs,
4848 if (type_may_be_null(reg_type))
4849 regs[value_regno].id = ++env->id_gen;
4850 /* A load of ctx field could have different
4851 * actual load size with the one encoded in the
4852 * insn. When the dst is PTR, it is for sure not
4855 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
4856 if (base_type(reg_type) == PTR_TO_BTF_ID) {
4857 regs[value_regno].btf = btf;
4858 regs[value_regno].btf_id = btf_id;
4861 regs[value_regno].type = reg_type;
4864 } else if (reg->type == PTR_TO_STACK) {
4865 /* Basic bounds checks. */
4866 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
4870 state = func(env, reg);
4871 err = update_stack_depth(env, state, off);
4876 err = check_stack_read(env, regno, off, size,
4879 err = check_stack_write(env, regno, off, size,
4880 value_regno, insn_idx);
4881 } else if (reg_is_pkt_pointer(reg)) {
4882 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
4883 verbose(env, "cannot write into packet\n");
4886 if (t == BPF_WRITE && value_regno >= 0 &&
4887 is_pointer_value(env, value_regno)) {
4888 verbose(env, "R%d leaks addr into packet\n",
4892 err = check_packet_access(env, regno, off, size, false);
4893 if (!err && t == BPF_READ && value_regno >= 0)
4894 mark_reg_unknown(env, regs, value_regno);
4895 } else if (reg->type == PTR_TO_FLOW_KEYS) {
4896 if (t == BPF_WRITE && value_regno >= 0 &&
4897 is_pointer_value(env, value_regno)) {
4898 verbose(env, "R%d leaks addr into flow keys\n",
4903 err = check_flow_keys_access(env, off, size);
4904 if (!err && t == BPF_READ && value_regno >= 0)
4905 mark_reg_unknown(env, regs, value_regno);
4906 } else if (type_is_sk_pointer(reg->type)) {
4907 if (t == BPF_WRITE) {
4908 verbose(env, "R%d cannot write into %s\n",
4909 regno, reg_type_str(env, reg->type));
4912 err = check_sock_access(env, insn_idx, regno, off, size, t);
4913 if (!err && value_regno >= 0)
4914 mark_reg_unknown(env, regs, value_regno);
4915 } else if (reg->type == PTR_TO_TP_BUFFER) {
4916 err = check_tp_buffer_access(env, reg, regno, off, size);
4917 if (!err && t == BPF_READ && value_regno >= 0)
4918 mark_reg_unknown(env, regs, value_regno);
4919 } else if (base_type(reg->type) == PTR_TO_BTF_ID &&
4920 !type_may_be_null(reg->type)) {
4921 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
4923 } else if (reg->type == CONST_PTR_TO_MAP) {
4924 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
4926 } else if (base_type(reg->type) == PTR_TO_BUF) {
4927 bool rdonly_mem = type_is_rdonly_mem(reg->type);
4931 if (t == BPF_WRITE) {
4932 verbose(env, "R%d cannot write into %s\n",
4933 regno, reg_type_str(env, reg->type));
4936 max_access = &env->prog->aux->max_rdonly_access;
4938 max_access = &env->prog->aux->max_rdwr_access;
4941 err = check_buffer_access(env, reg, regno, off, size, false,
4944 if (!err && value_regno >= 0 && (rdonly_mem || t == BPF_READ))
4945 mark_reg_unknown(env, regs, value_regno);
4947 verbose(env, "R%d invalid mem access '%s'\n", regno,
4948 reg_type_str(env, reg->type));
4952 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
4953 regs[value_regno].type == SCALAR_VALUE) {
4954 /* b/h/w load zero-extends, mark upper bits as known 0 */
4955 coerce_reg_to_size(®s[value_regno], size);
4960 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
4965 switch (insn->imm) {
4967 case BPF_ADD | BPF_FETCH:
4969 case BPF_AND | BPF_FETCH:
4971 case BPF_OR | BPF_FETCH:
4973 case BPF_XOR | BPF_FETCH:
4978 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
4982 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
4983 verbose(env, "invalid atomic operand size\n");
4987 /* check src1 operand */
4988 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4992 /* check src2 operand */
4993 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4997 if (insn->imm == BPF_CMPXCHG) {
4998 /* Check comparison of R0 with memory location */
4999 const u32 aux_reg = BPF_REG_0;
5001 err = check_reg_arg(env, aux_reg, SRC_OP);
5005 if (is_pointer_value(env, aux_reg)) {
5006 verbose(env, "R%d leaks addr into mem\n", aux_reg);
5011 if (is_pointer_value(env, insn->src_reg)) {
5012 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
5016 if (is_ctx_reg(env, insn->dst_reg) ||
5017 is_pkt_reg(env, insn->dst_reg) ||
5018 is_flow_key_reg(env, insn->dst_reg) ||
5019 is_sk_reg(env, insn->dst_reg)) {
5020 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
5022 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
5026 if (insn->imm & BPF_FETCH) {
5027 if (insn->imm == BPF_CMPXCHG)
5028 load_reg = BPF_REG_0;
5030 load_reg = insn->src_reg;
5032 /* check and record load of old value */
5033 err = check_reg_arg(env, load_reg, DST_OP);
5037 /* This instruction accesses a memory location but doesn't
5038 * actually load it into a register.
5043 /* Check whether we can read the memory, with second call for fetch
5044 * case to simulate the register fill.
5046 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
5047 BPF_SIZE(insn->code), BPF_READ, -1, true);
5048 if (!err && load_reg >= 0)
5049 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
5050 BPF_SIZE(insn->code), BPF_READ, load_reg,
5055 /* Check whether we can write into the same memory. */
5056 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
5057 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
5064 /* When register 'regno' is used to read the stack (either directly or through
5065 * a helper function) make sure that it's within stack boundary and, depending
5066 * on the access type, that all elements of the stack are initialized.
5068 * 'off' includes 'regno->off', but not its dynamic part (if any).
5070 * All registers that have been spilled on the stack in the slots within the
5071 * read offsets are marked as read.
5073 static int check_stack_range_initialized(
5074 struct bpf_verifier_env *env, int regno, int off,
5075 int access_size, bool zero_size_allowed,
5076 enum bpf_access_src type, struct bpf_call_arg_meta *meta)
5078 struct bpf_reg_state *reg = reg_state(env, regno);
5079 struct bpf_func_state *state = func(env, reg);
5080 int err, min_off, max_off, i, j, slot, spi;
5081 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
5082 enum bpf_access_type bounds_check_type;
5083 /* Some accesses can write anything into the stack, others are
5086 bool clobber = false;
5088 if (access_size == 0 && !zero_size_allowed) {
5089 verbose(env, "invalid zero-sized read\n");
5093 if (type == ACCESS_HELPER) {
5094 /* The bounds checks for writes are more permissive than for
5095 * reads. However, if raw_mode is not set, we'll do extra
5098 bounds_check_type = BPF_WRITE;
5101 bounds_check_type = BPF_READ;
5103 err = check_stack_access_within_bounds(env, regno, off, access_size,
5104 type, bounds_check_type);
5109 if (tnum_is_const(reg->var_off)) {
5110 min_off = max_off = reg->var_off.value + off;
5112 /* Variable offset is prohibited for unprivileged mode for
5113 * simplicity since it requires corresponding support in
5114 * Spectre masking for stack ALU.
5115 * See also retrieve_ptr_limit().
5117 if (!env->bypass_spec_v1) {
5120 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5121 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
5122 regno, err_extra, tn_buf);
5125 /* Only initialized buffer on stack is allowed to be accessed
5126 * with variable offset. With uninitialized buffer it's hard to
5127 * guarantee that whole memory is marked as initialized on
5128 * helper return since specific bounds are unknown what may
5129 * cause uninitialized stack leaking.
5131 if (meta && meta->raw_mode)
5134 min_off = reg->smin_value + off;
5135 max_off = reg->smax_value + off;
5138 if (meta && meta->raw_mode) {
5139 meta->access_size = access_size;
5140 meta->regno = regno;
5144 for (i = min_off; i < max_off + access_size; i++) {
5148 spi = slot / BPF_REG_SIZE;
5149 if (state->allocated_stack <= slot)
5151 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
5152 if (*stype == STACK_MISC)
5154 if (*stype == STACK_ZERO) {
5156 /* helper can write anything into the stack */
5157 *stype = STACK_MISC;
5162 if (is_spilled_reg(&state->stack[spi]) &&
5163 base_type(state->stack[spi].spilled_ptr.type) == PTR_TO_BTF_ID)
5166 if (is_spilled_reg(&state->stack[spi]) &&
5167 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
5168 env->allow_ptr_leaks)) {
5170 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
5171 for (j = 0; j < BPF_REG_SIZE; j++)
5172 scrub_spilled_slot(&state->stack[spi].slot_type[j]);
5178 if (tnum_is_const(reg->var_off)) {
5179 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
5180 err_extra, regno, min_off, i - min_off, access_size);
5184 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5185 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
5186 err_extra, regno, tn_buf, i - min_off, access_size);
5190 /* reading any byte out of 8-byte 'spill_slot' will cause
5191 * the whole slot to be marked as 'read'
5193 mark_reg_read(env, &state->stack[spi].spilled_ptr,
5194 state->stack[spi].spilled_ptr.parent,
5197 return update_stack_depth(env, state, min_off);
5200 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
5201 int access_size, bool zero_size_allowed,
5202 struct bpf_call_arg_meta *meta)
5204 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5207 switch (base_type(reg->type)) {
5209 case PTR_TO_PACKET_META:
5210 return check_packet_access(env, regno, reg->off, access_size,
5212 case PTR_TO_MAP_KEY:
5213 if (meta && meta->raw_mode) {
5214 verbose(env, "R%d cannot write into %s\n", regno,
5215 reg_type_str(env, reg->type));
5218 return check_mem_region_access(env, regno, reg->off, access_size,
5219 reg->map_ptr->key_size, false);
5220 case PTR_TO_MAP_VALUE:
5221 if (check_map_access_type(env, regno, reg->off, access_size,
5222 meta && meta->raw_mode ? BPF_WRITE :
5225 return check_map_access(env, regno, reg->off, access_size,
5226 zero_size_allowed, ACCESS_HELPER);
5228 if (type_is_rdonly_mem(reg->type)) {
5229 if (meta && meta->raw_mode) {
5230 verbose(env, "R%d cannot write into %s\n", regno,
5231 reg_type_str(env, reg->type));
5235 return check_mem_region_access(env, regno, reg->off,
5236 access_size, reg->mem_size,
5239 if (type_is_rdonly_mem(reg->type)) {
5240 if (meta && meta->raw_mode) {
5241 verbose(env, "R%d cannot write into %s\n", regno,
5242 reg_type_str(env, reg->type));
5246 max_access = &env->prog->aux->max_rdonly_access;
5248 max_access = &env->prog->aux->max_rdwr_access;
5250 return check_buffer_access(env, reg, regno, reg->off,
5251 access_size, zero_size_allowed,
5254 return check_stack_range_initialized(
5256 regno, reg->off, access_size,
5257 zero_size_allowed, ACCESS_HELPER, meta);
5259 /* in case the function doesn't know how to access the context,
5260 * (because we are in a program of type SYSCALL for example), we
5261 * can not statically check its size.
5262 * Dynamically check it now.
5264 if (!env->ops->convert_ctx_access) {
5265 enum bpf_access_type atype = meta && meta->raw_mode ? BPF_WRITE : BPF_READ;
5266 int offset = access_size - 1;
5268 /* Allow zero-byte read from PTR_TO_CTX */
5269 if (access_size == 0)
5270 return zero_size_allowed ? 0 : -EACCES;
5272 return check_mem_access(env, env->insn_idx, regno, offset, BPF_B,
5277 default: /* scalar_value or invalid ptr */
5278 /* Allow zero-byte read from NULL, regardless of pointer type */
5279 if (zero_size_allowed && access_size == 0 &&
5280 register_is_null(reg))
5283 verbose(env, "R%d type=%s ", regno,
5284 reg_type_str(env, reg->type));
5285 verbose(env, "expected=%s\n", reg_type_str(env, PTR_TO_STACK));
5290 static int check_mem_size_reg(struct bpf_verifier_env *env,
5291 struct bpf_reg_state *reg, u32 regno,
5292 bool zero_size_allowed,
5293 struct bpf_call_arg_meta *meta)
5297 /* This is used to refine r0 return value bounds for helpers
5298 * that enforce this value as an upper bound on return values.
5299 * See do_refine_retval_range() for helpers that can refine
5300 * the return value. C type of helper is u32 so we pull register
5301 * bound from umax_value however, if negative verifier errors
5302 * out. Only upper bounds can be learned because retval is an
5303 * int type and negative retvals are allowed.
5305 meta->msize_max_value = reg->umax_value;
5307 /* The register is SCALAR_VALUE; the access check
5308 * happens using its boundaries.
5310 if (!tnum_is_const(reg->var_off))
5311 /* For unprivileged variable accesses, disable raw
5312 * mode so that the program is required to
5313 * initialize all the memory that the helper could
5314 * just partially fill up.
5318 if (reg->smin_value < 0) {
5319 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
5324 if (reg->umin_value == 0) {
5325 err = check_helper_mem_access(env, regno - 1, 0,
5332 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
5333 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
5337 err = check_helper_mem_access(env, regno - 1,
5339 zero_size_allowed, meta);
5341 err = mark_chain_precision(env, regno);
5345 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
5346 u32 regno, u32 mem_size)
5348 bool may_be_null = type_may_be_null(reg->type);
5349 struct bpf_reg_state saved_reg;
5350 struct bpf_call_arg_meta meta;
5353 if (register_is_null(reg))
5356 memset(&meta, 0, sizeof(meta));
5357 /* Assuming that the register contains a value check if the memory
5358 * access is safe. Temporarily save and restore the register's state as
5359 * the conversion shouldn't be visible to a caller.
5363 mark_ptr_not_null_reg(reg);
5366 err = check_helper_mem_access(env, regno, mem_size, true, &meta);
5367 /* Check access for BPF_WRITE */
5368 meta.raw_mode = true;
5369 err = err ?: check_helper_mem_access(env, regno, mem_size, true, &meta);
5377 int check_kfunc_mem_size_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
5380 struct bpf_reg_state *mem_reg = &cur_regs(env)[regno - 1];
5381 bool may_be_null = type_may_be_null(mem_reg->type);
5382 struct bpf_reg_state saved_reg;
5383 struct bpf_call_arg_meta meta;
5386 WARN_ON_ONCE(regno < BPF_REG_2 || regno > BPF_REG_5);
5388 memset(&meta, 0, sizeof(meta));
5391 saved_reg = *mem_reg;
5392 mark_ptr_not_null_reg(mem_reg);
5395 err = check_mem_size_reg(env, reg, regno, true, &meta);
5396 /* Check access for BPF_WRITE */
5397 meta.raw_mode = true;
5398 err = err ?: check_mem_size_reg(env, reg, regno, true, &meta);
5401 *mem_reg = saved_reg;
5405 /* Implementation details:
5406 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
5407 * Two bpf_map_lookups (even with the same key) will have different reg->id.
5408 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
5409 * value_or_null->value transition, since the verifier only cares about
5410 * the range of access to valid map value pointer and doesn't care about actual
5411 * address of the map element.
5412 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
5413 * reg->id > 0 after value_or_null->value transition. By doing so
5414 * two bpf_map_lookups will be considered two different pointers that
5415 * point to different bpf_spin_locks.
5416 * The verifier allows taking only one bpf_spin_lock at a time to avoid
5418 * Since only one bpf_spin_lock is allowed the checks are simpler than
5419 * reg_is_refcounted() logic. The verifier needs to remember only
5420 * one spin_lock instead of array of acquired_refs.
5421 * cur_state->active_spin_lock remembers which map value element got locked
5422 * and clears it after bpf_spin_unlock.
5424 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
5427 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5428 struct bpf_verifier_state *cur = env->cur_state;
5429 bool is_const = tnum_is_const(reg->var_off);
5430 struct bpf_map *map = reg->map_ptr;
5431 u64 val = reg->var_off.value;
5435 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
5441 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
5445 if (!map_value_has_spin_lock(map)) {
5446 if (map->spin_lock_off == -E2BIG)
5448 "map '%s' has more than one 'struct bpf_spin_lock'\n",
5450 else if (map->spin_lock_off == -ENOENT)
5452 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
5456 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
5460 if (map->spin_lock_off != val + reg->off) {
5461 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
5466 if (cur->active_spin_lock) {
5468 "Locking two bpf_spin_locks are not allowed\n");
5471 cur->active_spin_lock = reg->id;
5473 if (!cur->active_spin_lock) {
5474 verbose(env, "bpf_spin_unlock without taking a lock\n");
5477 if (cur->active_spin_lock != reg->id) {
5478 verbose(env, "bpf_spin_unlock of different lock\n");
5481 cur->active_spin_lock = 0;
5486 static int process_timer_func(struct bpf_verifier_env *env, int regno,
5487 struct bpf_call_arg_meta *meta)
5489 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5490 bool is_const = tnum_is_const(reg->var_off);
5491 struct bpf_map *map = reg->map_ptr;
5492 u64 val = reg->var_off.value;
5496 "R%d doesn't have constant offset. bpf_timer has to be at the constant offset\n",
5501 verbose(env, "map '%s' has to have BTF in order to use bpf_timer\n",
5505 if (!map_value_has_timer(map)) {
5506 if (map->timer_off == -E2BIG)
5508 "map '%s' has more than one 'struct bpf_timer'\n",
5510 else if (map->timer_off == -ENOENT)
5512 "map '%s' doesn't have 'struct bpf_timer'\n",
5516 "map '%s' is not a struct type or bpf_timer is mangled\n",
5520 if (map->timer_off != val + reg->off) {
5521 verbose(env, "off %lld doesn't point to 'struct bpf_timer' that is at %d\n",
5522 val + reg->off, map->timer_off);
5525 if (meta->map_ptr) {
5526 verbose(env, "verifier bug. Two map pointers in a timer helper\n");
5529 meta->map_uid = reg->map_uid;
5530 meta->map_ptr = map;
5534 static int process_kptr_func(struct bpf_verifier_env *env, int regno,
5535 struct bpf_call_arg_meta *meta)
5537 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5538 struct bpf_map_value_off_desc *off_desc;
5539 struct bpf_map *map_ptr = reg->map_ptr;
5543 if (!tnum_is_const(reg->var_off)) {
5545 "R%d doesn't have constant offset. kptr has to be at the constant offset\n",
5549 if (!map_ptr->btf) {
5550 verbose(env, "map '%s' has to have BTF in order to use bpf_kptr_xchg\n",
5554 if (!map_value_has_kptrs(map_ptr)) {
5555 ret = PTR_ERR_OR_ZERO(map_ptr->kptr_off_tab);
5557 verbose(env, "map '%s' has more than %d kptr\n", map_ptr->name,
5558 BPF_MAP_VALUE_OFF_MAX);
5559 else if (ret == -EEXIST)
5560 verbose(env, "map '%s' has repeating kptr BTF tags\n", map_ptr->name);
5562 verbose(env, "map '%s' has no valid kptr\n", map_ptr->name);
5566 meta->map_ptr = map_ptr;
5567 kptr_off = reg->off + reg->var_off.value;
5568 off_desc = bpf_map_kptr_off_contains(map_ptr, kptr_off);
5570 verbose(env, "off=%d doesn't point to kptr\n", kptr_off);
5573 if (off_desc->type != BPF_KPTR_REF) {
5574 verbose(env, "off=%d kptr isn't referenced kptr\n", kptr_off);
5577 meta->kptr_off_desc = off_desc;
5581 static bool arg_type_is_mem_size(enum bpf_arg_type type)
5583 return type == ARG_CONST_SIZE ||
5584 type == ARG_CONST_SIZE_OR_ZERO;
5587 static bool arg_type_is_release(enum bpf_arg_type type)
5589 return type & OBJ_RELEASE;
5592 static bool arg_type_is_dynptr(enum bpf_arg_type type)
5594 return base_type(type) == ARG_PTR_TO_DYNPTR;
5597 static int int_ptr_type_to_size(enum bpf_arg_type type)
5599 if (type == ARG_PTR_TO_INT)
5601 else if (type == ARG_PTR_TO_LONG)
5607 static int resolve_map_arg_type(struct bpf_verifier_env *env,
5608 const struct bpf_call_arg_meta *meta,
5609 enum bpf_arg_type *arg_type)
5611 if (!meta->map_ptr) {
5612 /* kernel subsystem misconfigured verifier */
5613 verbose(env, "invalid map_ptr to access map->type\n");
5617 switch (meta->map_ptr->map_type) {
5618 case BPF_MAP_TYPE_SOCKMAP:
5619 case BPF_MAP_TYPE_SOCKHASH:
5620 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
5621 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
5623 verbose(env, "invalid arg_type for sockmap/sockhash\n");
5627 case BPF_MAP_TYPE_BLOOM_FILTER:
5628 if (meta->func_id == BPF_FUNC_map_peek_elem)
5629 *arg_type = ARG_PTR_TO_MAP_VALUE;
5637 struct bpf_reg_types {
5638 const enum bpf_reg_type types[10];
5642 static const struct bpf_reg_types map_key_value_types = {
5652 static const struct bpf_reg_types sock_types = {
5662 static const struct bpf_reg_types btf_id_sock_common_types = {
5670 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
5674 static const struct bpf_reg_types mem_types = {
5682 PTR_TO_MEM | MEM_ALLOC,
5687 static const struct bpf_reg_types int_ptr_types = {
5697 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
5698 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
5699 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
5700 static const struct bpf_reg_types alloc_mem_types = { .types = { PTR_TO_MEM | MEM_ALLOC } };
5701 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
5702 static const struct bpf_reg_types btf_ptr_types = { .types = { PTR_TO_BTF_ID } };
5703 static const struct bpf_reg_types spin_lock_types = { .types = { PTR_TO_MAP_VALUE } };
5704 static const struct bpf_reg_types percpu_btf_ptr_types = { .types = { PTR_TO_BTF_ID | MEM_PERCPU } };
5705 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
5706 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
5707 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
5708 static const struct bpf_reg_types timer_types = { .types = { PTR_TO_MAP_VALUE } };
5709 static const struct bpf_reg_types kptr_types = { .types = { PTR_TO_MAP_VALUE } };
5710 static const struct bpf_reg_types dynptr_types = {
5713 PTR_TO_DYNPTR | DYNPTR_TYPE_LOCAL,
5717 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
5718 [ARG_PTR_TO_MAP_KEY] = &map_key_value_types,
5719 [ARG_PTR_TO_MAP_VALUE] = &map_key_value_types,
5720 [ARG_CONST_SIZE] = &scalar_types,
5721 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
5722 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
5723 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
5724 [ARG_PTR_TO_CTX] = &context_types,
5725 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
5727 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
5729 [ARG_PTR_TO_SOCKET] = &fullsock_types,
5730 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
5731 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
5732 [ARG_PTR_TO_MEM] = &mem_types,
5733 [ARG_PTR_TO_ALLOC_MEM] = &alloc_mem_types,
5734 [ARG_PTR_TO_INT] = &int_ptr_types,
5735 [ARG_PTR_TO_LONG] = &int_ptr_types,
5736 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
5737 [ARG_PTR_TO_FUNC] = &func_ptr_types,
5738 [ARG_PTR_TO_STACK] = &stack_ptr_types,
5739 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
5740 [ARG_PTR_TO_TIMER] = &timer_types,
5741 [ARG_PTR_TO_KPTR] = &kptr_types,
5742 [ARG_PTR_TO_DYNPTR] = &dynptr_types,
5745 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
5746 enum bpf_arg_type arg_type,
5747 const u32 *arg_btf_id,
5748 struct bpf_call_arg_meta *meta)
5750 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5751 enum bpf_reg_type expected, type = reg->type;
5752 const struct bpf_reg_types *compatible;
5755 compatible = compatible_reg_types[base_type(arg_type)];
5757 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
5761 /* ARG_PTR_TO_MEM + RDONLY is compatible with PTR_TO_MEM and PTR_TO_MEM + RDONLY,
5762 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM and NOT with PTR_TO_MEM + RDONLY
5764 * Same for MAYBE_NULL:
5766 * ARG_PTR_TO_MEM + MAYBE_NULL is compatible with PTR_TO_MEM and PTR_TO_MEM + MAYBE_NULL,
5767 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM but NOT with PTR_TO_MEM + MAYBE_NULL
5769 * Therefore we fold these flags depending on the arg_type before comparison.
5771 if (arg_type & MEM_RDONLY)
5772 type &= ~MEM_RDONLY;
5773 if (arg_type & PTR_MAYBE_NULL)
5774 type &= ~PTR_MAYBE_NULL;
5776 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
5777 expected = compatible->types[i];
5778 if (expected == NOT_INIT)
5781 if (type == expected)
5785 verbose(env, "R%d type=%s expected=", regno, reg_type_str(env, reg->type));
5786 for (j = 0; j + 1 < i; j++)
5787 verbose(env, "%s, ", reg_type_str(env, compatible->types[j]));
5788 verbose(env, "%s\n", reg_type_str(env, compatible->types[j]));
5792 if (reg->type == PTR_TO_BTF_ID) {
5793 /* For bpf_sk_release, it needs to match against first member
5794 * 'struct sock_common', hence make an exception for it. This
5795 * allows bpf_sk_release to work for multiple socket types.
5797 bool strict_type_match = arg_type_is_release(arg_type) &&
5798 meta->func_id != BPF_FUNC_sk_release;
5801 if (!compatible->btf_id) {
5802 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
5805 arg_btf_id = compatible->btf_id;
5808 if (meta->func_id == BPF_FUNC_kptr_xchg) {
5809 if (map_kptr_match_type(env, meta->kptr_off_desc, reg, regno))
5812 if (arg_btf_id == BPF_PTR_POISON) {
5813 verbose(env, "verifier internal error:");
5814 verbose(env, "R%d has non-overwritten BPF_PTR_POISON type\n",
5819 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
5820 btf_vmlinux, *arg_btf_id,
5821 strict_type_match)) {
5822 verbose(env, "R%d is of type %s but %s is expected\n",
5823 regno, kernel_type_name(reg->btf, reg->btf_id),
5824 kernel_type_name(btf_vmlinux, *arg_btf_id));
5833 int check_func_arg_reg_off(struct bpf_verifier_env *env,
5834 const struct bpf_reg_state *reg, int regno,
5835 enum bpf_arg_type arg_type)
5837 enum bpf_reg_type type = reg->type;
5838 bool fixed_off_ok = false;
5840 switch ((u32)type) {
5841 /* Pointer types where reg offset is explicitly allowed: */
5843 if (arg_type_is_dynptr(arg_type) && reg->off % BPF_REG_SIZE) {
5844 verbose(env, "cannot pass in dynptr at an offset\n");
5849 case PTR_TO_PACKET_META:
5850 case PTR_TO_MAP_KEY:
5851 case PTR_TO_MAP_VALUE:
5853 case PTR_TO_MEM | MEM_RDONLY:
5854 case PTR_TO_MEM | MEM_ALLOC:
5856 case PTR_TO_BUF | MEM_RDONLY:
5858 /* Some of the argument types nevertheless require a
5859 * zero register offset.
5861 if (base_type(arg_type) != ARG_PTR_TO_ALLOC_MEM)
5864 /* All the rest must be rejected, except PTR_TO_BTF_ID which allows
5868 /* When referenced PTR_TO_BTF_ID is passed to release function,
5869 * it's fixed offset must be 0. In the other cases, fixed offset
5872 if (arg_type_is_release(arg_type) && reg->off) {
5873 verbose(env, "R%d must have zero offset when passed to release func\n",
5877 /* For arg is release pointer, fixed_off_ok must be false, but
5878 * we already checked and rejected reg->off != 0 above, so set
5879 * to true to allow fixed offset for all other cases.
5881 fixed_off_ok = true;
5886 return __check_ptr_off_reg(env, reg, regno, fixed_off_ok);
5889 static u32 stack_slot_get_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
5891 struct bpf_func_state *state = func(env, reg);
5892 int spi = get_spi(reg->off);
5894 return state->stack[spi].spilled_ptr.id;
5897 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
5898 struct bpf_call_arg_meta *meta,
5899 const struct bpf_func_proto *fn)
5901 u32 regno = BPF_REG_1 + arg;
5902 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5903 enum bpf_arg_type arg_type = fn->arg_type[arg];
5904 enum bpf_reg_type type = reg->type;
5905 u32 *arg_btf_id = NULL;
5908 if (arg_type == ARG_DONTCARE)
5911 err = check_reg_arg(env, regno, SRC_OP);
5915 if (arg_type == ARG_ANYTHING) {
5916 if (is_pointer_value(env, regno)) {
5917 verbose(env, "R%d leaks addr into helper function\n",
5924 if (type_is_pkt_pointer(type) &&
5925 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
5926 verbose(env, "helper access to the packet is not allowed\n");
5930 if (base_type(arg_type) == ARG_PTR_TO_MAP_VALUE) {
5931 err = resolve_map_arg_type(env, meta, &arg_type);
5936 if (register_is_null(reg) && type_may_be_null(arg_type))
5937 /* A NULL register has a SCALAR_VALUE type, so skip
5940 goto skip_type_check;
5942 /* arg_btf_id and arg_size are in a union. */
5943 if (base_type(arg_type) == ARG_PTR_TO_BTF_ID)
5944 arg_btf_id = fn->arg_btf_id[arg];
5946 err = check_reg_type(env, regno, arg_type, arg_btf_id, meta);
5950 err = check_func_arg_reg_off(env, reg, regno, arg_type);
5955 if (arg_type_is_release(arg_type)) {
5956 if (arg_type_is_dynptr(arg_type)) {
5957 struct bpf_func_state *state = func(env, reg);
5958 int spi = get_spi(reg->off);
5960 if (!is_spi_bounds_valid(state, spi, BPF_DYNPTR_NR_SLOTS) ||
5961 !state->stack[spi].spilled_ptr.id) {
5962 verbose(env, "arg %d is an unacquired reference\n", regno);
5965 } else if (!reg->ref_obj_id && !register_is_null(reg)) {
5966 verbose(env, "R%d must be referenced when passed to release function\n",
5970 if (meta->release_regno) {
5971 verbose(env, "verifier internal error: more than one release argument\n");
5974 meta->release_regno = regno;
5977 if (reg->ref_obj_id) {
5978 if (meta->ref_obj_id) {
5979 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
5980 regno, reg->ref_obj_id,
5984 meta->ref_obj_id = reg->ref_obj_id;
5987 switch (base_type(arg_type)) {
5988 case ARG_CONST_MAP_PTR:
5989 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
5990 if (meta->map_ptr) {
5991 /* Use map_uid (which is unique id of inner map) to reject:
5992 * inner_map1 = bpf_map_lookup_elem(outer_map, key1)
5993 * inner_map2 = bpf_map_lookup_elem(outer_map, key2)
5994 * if (inner_map1 && inner_map2) {
5995 * timer = bpf_map_lookup_elem(inner_map1);
5997 * // mismatch would have been allowed
5998 * bpf_timer_init(timer, inner_map2);
6001 * Comparing map_ptr is enough to distinguish normal and outer maps.
6003 if (meta->map_ptr != reg->map_ptr ||
6004 meta->map_uid != reg->map_uid) {
6006 "timer pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
6007 meta->map_uid, reg->map_uid);
6011 meta->map_ptr = reg->map_ptr;
6012 meta->map_uid = reg->map_uid;
6014 case ARG_PTR_TO_MAP_KEY:
6015 /* bpf_map_xxx(..., map_ptr, ..., key) call:
6016 * check that [key, key + map->key_size) are within
6017 * stack limits and initialized
6019 if (!meta->map_ptr) {
6020 /* in function declaration map_ptr must come before
6021 * map_key, so that it's verified and known before
6022 * we have to check map_key here. Otherwise it means
6023 * that kernel subsystem misconfigured verifier
6025 verbose(env, "invalid map_ptr to access map->key\n");
6028 err = check_helper_mem_access(env, regno,
6029 meta->map_ptr->key_size, false,
6032 case ARG_PTR_TO_MAP_VALUE:
6033 if (type_may_be_null(arg_type) && register_is_null(reg))
6036 /* bpf_map_xxx(..., map_ptr, ..., value) call:
6037 * check [value, value + map->value_size) validity
6039 if (!meta->map_ptr) {
6040 /* kernel subsystem misconfigured verifier */
6041 verbose(env, "invalid map_ptr to access map->value\n");
6044 meta->raw_mode = arg_type & MEM_UNINIT;
6045 err = check_helper_mem_access(env, regno,
6046 meta->map_ptr->value_size, false,
6049 case ARG_PTR_TO_PERCPU_BTF_ID:
6051 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
6054 meta->ret_btf = reg->btf;
6055 meta->ret_btf_id = reg->btf_id;
6057 case ARG_PTR_TO_SPIN_LOCK:
6058 if (meta->func_id == BPF_FUNC_spin_lock) {
6059 if (process_spin_lock(env, regno, true))
6061 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
6062 if (process_spin_lock(env, regno, false))
6065 verbose(env, "verifier internal error\n");
6069 case ARG_PTR_TO_TIMER:
6070 if (process_timer_func(env, regno, meta))
6073 case ARG_PTR_TO_FUNC:
6074 meta->subprogno = reg->subprogno;
6076 case ARG_PTR_TO_MEM:
6077 /* The access to this pointer is only checked when we hit the
6078 * next is_mem_size argument below.
6080 meta->raw_mode = arg_type & MEM_UNINIT;
6081 if (arg_type & MEM_FIXED_SIZE) {
6082 err = check_helper_mem_access(env, regno,
6083 fn->arg_size[arg], false,
6087 case ARG_CONST_SIZE:
6088 err = check_mem_size_reg(env, reg, regno, false, meta);
6090 case ARG_CONST_SIZE_OR_ZERO:
6091 err = check_mem_size_reg(env, reg, regno, true, meta);
6093 case ARG_PTR_TO_DYNPTR:
6094 /* We only need to check for initialized / uninitialized helper
6095 * dynptr args if the dynptr is not PTR_TO_DYNPTR, as the
6096 * assumption is that if it is, that a helper function
6097 * initialized the dynptr on behalf of the BPF program.
6099 if (base_type(reg->type) == PTR_TO_DYNPTR)
6101 if (arg_type & MEM_UNINIT) {
6102 if (!is_dynptr_reg_valid_uninit(env, reg)) {
6103 verbose(env, "Dynptr has to be an uninitialized dynptr\n");
6107 /* We only support one dynptr being uninitialized at the moment,
6108 * which is sufficient for the helper functions we have right now.
6110 if (meta->uninit_dynptr_regno) {
6111 verbose(env, "verifier internal error: multiple uninitialized dynptr args\n");
6115 meta->uninit_dynptr_regno = regno;
6116 } else if (!is_dynptr_reg_valid_init(env, reg)) {
6118 "Expected an initialized dynptr as arg #%d\n",
6121 } else if (!is_dynptr_type_expected(env, reg, arg_type)) {
6122 const char *err_extra = "";
6124 switch (arg_type & DYNPTR_TYPE_FLAG_MASK) {
6125 case DYNPTR_TYPE_LOCAL:
6126 err_extra = "local";
6128 case DYNPTR_TYPE_RINGBUF:
6129 err_extra = "ringbuf";
6132 err_extra = "<unknown>";
6136 "Expected a dynptr of type %s as arg #%d\n",
6137 err_extra, arg + 1);
6141 case ARG_CONST_ALLOC_SIZE_OR_ZERO:
6142 if (!tnum_is_const(reg->var_off)) {
6143 verbose(env, "R%d is not a known constant'\n",
6147 meta->mem_size = reg->var_off.value;
6148 err = mark_chain_precision(env, regno);
6152 case ARG_PTR_TO_INT:
6153 case ARG_PTR_TO_LONG:
6155 int size = int_ptr_type_to_size(arg_type);
6157 err = check_helper_mem_access(env, regno, size, false, meta);
6160 err = check_ptr_alignment(env, reg, 0, size, true);
6163 case ARG_PTR_TO_CONST_STR:
6165 struct bpf_map *map = reg->map_ptr;
6170 if (!bpf_map_is_rdonly(map)) {
6171 verbose(env, "R%d does not point to a readonly map'\n", regno);
6175 if (!tnum_is_const(reg->var_off)) {
6176 verbose(env, "R%d is not a constant address'\n", regno);
6180 if (!map->ops->map_direct_value_addr) {
6181 verbose(env, "no direct value access support for this map type\n");
6185 err = check_map_access(env, regno, reg->off,
6186 map->value_size - reg->off, false,
6191 map_off = reg->off + reg->var_off.value;
6192 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
6194 verbose(env, "direct value access on string failed\n");
6198 str_ptr = (char *)(long)(map_addr);
6199 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
6200 verbose(env, "string is not zero-terminated\n");
6205 case ARG_PTR_TO_KPTR:
6206 if (process_kptr_func(env, regno, meta))
6214 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
6216 enum bpf_attach_type eatype = env->prog->expected_attach_type;
6217 enum bpf_prog_type type = resolve_prog_type(env->prog);
6219 if (func_id != BPF_FUNC_map_update_elem)
6222 /* It's not possible to get access to a locked struct sock in these
6223 * contexts, so updating is safe.
6226 case BPF_PROG_TYPE_TRACING:
6227 if (eatype == BPF_TRACE_ITER)
6230 case BPF_PROG_TYPE_SOCKET_FILTER:
6231 case BPF_PROG_TYPE_SCHED_CLS:
6232 case BPF_PROG_TYPE_SCHED_ACT:
6233 case BPF_PROG_TYPE_XDP:
6234 case BPF_PROG_TYPE_SK_REUSEPORT:
6235 case BPF_PROG_TYPE_FLOW_DISSECTOR:
6236 case BPF_PROG_TYPE_SK_LOOKUP:
6242 verbose(env, "cannot update sockmap in this context\n");
6246 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
6248 return env->prog->jit_requested &&
6249 bpf_jit_supports_subprog_tailcalls();
6252 static int check_map_func_compatibility(struct bpf_verifier_env *env,
6253 struct bpf_map *map, int func_id)
6258 /* We need a two way check, first is from map perspective ... */
6259 switch (map->map_type) {
6260 case BPF_MAP_TYPE_PROG_ARRAY:
6261 if (func_id != BPF_FUNC_tail_call)
6264 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
6265 if (func_id != BPF_FUNC_perf_event_read &&
6266 func_id != BPF_FUNC_perf_event_output &&
6267 func_id != BPF_FUNC_skb_output &&
6268 func_id != BPF_FUNC_perf_event_read_value &&
6269 func_id != BPF_FUNC_xdp_output)
6272 case BPF_MAP_TYPE_RINGBUF:
6273 if (func_id != BPF_FUNC_ringbuf_output &&
6274 func_id != BPF_FUNC_ringbuf_reserve &&
6275 func_id != BPF_FUNC_ringbuf_query &&
6276 func_id != BPF_FUNC_ringbuf_reserve_dynptr &&
6277 func_id != BPF_FUNC_ringbuf_submit_dynptr &&
6278 func_id != BPF_FUNC_ringbuf_discard_dynptr)
6281 case BPF_MAP_TYPE_USER_RINGBUF:
6282 if (func_id != BPF_FUNC_user_ringbuf_drain)
6285 case BPF_MAP_TYPE_STACK_TRACE:
6286 if (func_id != BPF_FUNC_get_stackid)
6289 case BPF_MAP_TYPE_CGROUP_ARRAY:
6290 if (func_id != BPF_FUNC_skb_under_cgroup &&
6291 func_id != BPF_FUNC_current_task_under_cgroup)
6294 case BPF_MAP_TYPE_CGROUP_STORAGE:
6295 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
6296 if (func_id != BPF_FUNC_get_local_storage)
6299 case BPF_MAP_TYPE_DEVMAP:
6300 case BPF_MAP_TYPE_DEVMAP_HASH:
6301 if (func_id != BPF_FUNC_redirect_map &&
6302 func_id != BPF_FUNC_map_lookup_elem)
6305 /* Restrict bpf side of cpumap and xskmap, open when use-cases
6308 case BPF_MAP_TYPE_CPUMAP:
6309 if (func_id != BPF_FUNC_redirect_map)
6312 case BPF_MAP_TYPE_XSKMAP:
6313 if (func_id != BPF_FUNC_redirect_map &&
6314 func_id != BPF_FUNC_map_lookup_elem)
6317 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
6318 case BPF_MAP_TYPE_HASH_OF_MAPS:
6319 if (func_id != BPF_FUNC_map_lookup_elem)
6322 case BPF_MAP_TYPE_SOCKMAP:
6323 if (func_id != BPF_FUNC_sk_redirect_map &&
6324 func_id != BPF_FUNC_sock_map_update &&
6325 func_id != BPF_FUNC_map_delete_elem &&
6326 func_id != BPF_FUNC_msg_redirect_map &&
6327 func_id != BPF_FUNC_sk_select_reuseport &&
6328 func_id != BPF_FUNC_map_lookup_elem &&
6329 !may_update_sockmap(env, func_id))
6332 case BPF_MAP_TYPE_SOCKHASH:
6333 if (func_id != BPF_FUNC_sk_redirect_hash &&
6334 func_id != BPF_FUNC_sock_hash_update &&
6335 func_id != BPF_FUNC_map_delete_elem &&
6336 func_id != BPF_FUNC_msg_redirect_hash &&
6337 func_id != BPF_FUNC_sk_select_reuseport &&
6338 func_id != BPF_FUNC_map_lookup_elem &&
6339 !may_update_sockmap(env, func_id))
6342 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
6343 if (func_id != BPF_FUNC_sk_select_reuseport)
6346 case BPF_MAP_TYPE_QUEUE:
6347 case BPF_MAP_TYPE_STACK:
6348 if (func_id != BPF_FUNC_map_peek_elem &&
6349 func_id != BPF_FUNC_map_pop_elem &&
6350 func_id != BPF_FUNC_map_push_elem)
6353 case BPF_MAP_TYPE_SK_STORAGE:
6354 if (func_id != BPF_FUNC_sk_storage_get &&
6355 func_id != BPF_FUNC_sk_storage_delete)
6358 case BPF_MAP_TYPE_INODE_STORAGE:
6359 if (func_id != BPF_FUNC_inode_storage_get &&
6360 func_id != BPF_FUNC_inode_storage_delete)
6363 case BPF_MAP_TYPE_TASK_STORAGE:
6364 if (func_id != BPF_FUNC_task_storage_get &&
6365 func_id != BPF_FUNC_task_storage_delete)
6368 case BPF_MAP_TYPE_BLOOM_FILTER:
6369 if (func_id != BPF_FUNC_map_peek_elem &&
6370 func_id != BPF_FUNC_map_push_elem)
6377 /* ... and second from the function itself. */
6379 case BPF_FUNC_tail_call:
6380 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
6382 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
6383 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
6387 case BPF_FUNC_perf_event_read:
6388 case BPF_FUNC_perf_event_output:
6389 case BPF_FUNC_perf_event_read_value:
6390 case BPF_FUNC_skb_output:
6391 case BPF_FUNC_xdp_output:
6392 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
6395 case BPF_FUNC_ringbuf_output:
6396 case BPF_FUNC_ringbuf_reserve:
6397 case BPF_FUNC_ringbuf_query:
6398 case BPF_FUNC_ringbuf_reserve_dynptr:
6399 case BPF_FUNC_ringbuf_submit_dynptr:
6400 case BPF_FUNC_ringbuf_discard_dynptr:
6401 if (map->map_type != BPF_MAP_TYPE_RINGBUF)
6404 case BPF_FUNC_user_ringbuf_drain:
6405 if (map->map_type != BPF_MAP_TYPE_USER_RINGBUF)
6408 case BPF_FUNC_get_stackid:
6409 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
6412 case BPF_FUNC_current_task_under_cgroup:
6413 case BPF_FUNC_skb_under_cgroup:
6414 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
6417 case BPF_FUNC_redirect_map:
6418 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
6419 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
6420 map->map_type != BPF_MAP_TYPE_CPUMAP &&
6421 map->map_type != BPF_MAP_TYPE_XSKMAP)
6424 case BPF_FUNC_sk_redirect_map:
6425 case BPF_FUNC_msg_redirect_map:
6426 case BPF_FUNC_sock_map_update:
6427 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
6430 case BPF_FUNC_sk_redirect_hash:
6431 case BPF_FUNC_msg_redirect_hash:
6432 case BPF_FUNC_sock_hash_update:
6433 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
6436 case BPF_FUNC_get_local_storage:
6437 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
6438 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
6441 case BPF_FUNC_sk_select_reuseport:
6442 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
6443 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
6444 map->map_type != BPF_MAP_TYPE_SOCKHASH)
6447 case BPF_FUNC_map_pop_elem:
6448 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
6449 map->map_type != BPF_MAP_TYPE_STACK)
6452 case BPF_FUNC_map_peek_elem:
6453 case BPF_FUNC_map_push_elem:
6454 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
6455 map->map_type != BPF_MAP_TYPE_STACK &&
6456 map->map_type != BPF_MAP_TYPE_BLOOM_FILTER)
6459 case BPF_FUNC_map_lookup_percpu_elem:
6460 if (map->map_type != BPF_MAP_TYPE_PERCPU_ARRAY &&
6461 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
6462 map->map_type != BPF_MAP_TYPE_LRU_PERCPU_HASH)
6465 case BPF_FUNC_sk_storage_get:
6466 case BPF_FUNC_sk_storage_delete:
6467 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
6470 case BPF_FUNC_inode_storage_get:
6471 case BPF_FUNC_inode_storage_delete:
6472 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
6475 case BPF_FUNC_task_storage_get:
6476 case BPF_FUNC_task_storage_delete:
6477 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
6486 verbose(env, "cannot pass map_type %d into func %s#%d\n",
6487 map->map_type, func_id_name(func_id), func_id);
6491 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
6495 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
6497 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
6499 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
6501 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
6503 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
6506 /* We only support one arg being in raw mode at the moment,
6507 * which is sufficient for the helper functions we have
6513 static bool check_args_pair_invalid(const struct bpf_func_proto *fn, int arg)
6515 bool is_fixed = fn->arg_type[arg] & MEM_FIXED_SIZE;
6516 bool has_size = fn->arg_size[arg] != 0;
6517 bool is_next_size = false;
6519 if (arg + 1 < ARRAY_SIZE(fn->arg_type))
6520 is_next_size = arg_type_is_mem_size(fn->arg_type[arg + 1]);
6522 if (base_type(fn->arg_type[arg]) != ARG_PTR_TO_MEM)
6523 return is_next_size;
6525 return has_size == is_next_size || is_next_size == is_fixed;
6528 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
6530 /* bpf_xxx(..., buf, len) call will access 'len'
6531 * bytes from memory 'buf'. Both arg types need
6532 * to be paired, so make sure there's no buggy
6533 * helper function specification.
6535 if (arg_type_is_mem_size(fn->arg1_type) ||
6536 check_args_pair_invalid(fn, 0) ||
6537 check_args_pair_invalid(fn, 1) ||
6538 check_args_pair_invalid(fn, 2) ||
6539 check_args_pair_invalid(fn, 3) ||
6540 check_args_pair_invalid(fn, 4))
6546 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
6550 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
6551 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_BTF_ID && !fn->arg_btf_id[i])
6554 if (base_type(fn->arg_type[i]) != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i] &&
6555 /* arg_btf_id and arg_size are in a union. */
6556 (base_type(fn->arg_type[i]) != ARG_PTR_TO_MEM ||
6557 !(fn->arg_type[i] & MEM_FIXED_SIZE)))
6564 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
6566 return check_raw_mode_ok(fn) &&
6567 check_arg_pair_ok(fn) &&
6568 check_btf_id_ok(fn) ? 0 : -EINVAL;
6571 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
6572 * are now invalid, so turn them into unknown SCALAR_VALUE.
6574 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
6576 struct bpf_func_state *state;
6577 struct bpf_reg_state *reg;
6579 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
6580 if (reg_is_pkt_pointer_any(reg))
6581 __mark_reg_unknown(env, reg);
6587 BEYOND_PKT_END = -2,
6590 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
6592 struct bpf_func_state *state = vstate->frame[vstate->curframe];
6593 struct bpf_reg_state *reg = &state->regs[regn];
6595 if (reg->type != PTR_TO_PACKET)
6596 /* PTR_TO_PACKET_META is not supported yet */
6599 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
6600 * How far beyond pkt_end it goes is unknown.
6601 * if (!range_open) it's the case of pkt >= pkt_end
6602 * if (range_open) it's the case of pkt > pkt_end
6603 * hence this pointer is at least 1 byte bigger than pkt_end
6606 reg->range = BEYOND_PKT_END;
6608 reg->range = AT_PKT_END;
6611 /* The pointer with the specified id has released its reference to kernel
6612 * resources. Identify all copies of the same pointer and clear the reference.
6614 static int release_reference(struct bpf_verifier_env *env,
6617 struct bpf_func_state *state;
6618 struct bpf_reg_state *reg;
6621 err = release_reference_state(cur_func(env), ref_obj_id);
6625 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
6626 if (reg->ref_obj_id == ref_obj_id)
6627 __mark_reg_unknown(env, reg);
6633 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
6634 struct bpf_reg_state *regs)
6638 /* after the call registers r0 - r5 were scratched */
6639 for (i = 0; i < CALLER_SAVED_REGS; i++) {
6640 mark_reg_not_init(env, regs, caller_saved[i]);
6641 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
6645 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
6646 struct bpf_func_state *caller,
6647 struct bpf_func_state *callee,
6650 static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
6651 int *insn_idx, int subprog,
6652 set_callee_state_fn set_callee_state_cb)
6654 struct bpf_verifier_state *state = env->cur_state;
6655 struct bpf_func_info_aux *func_info_aux;
6656 struct bpf_func_state *caller, *callee;
6658 bool is_global = false;
6660 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
6661 verbose(env, "the call stack of %d frames is too deep\n",
6662 state->curframe + 2);
6666 caller = state->frame[state->curframe];
6667 if (state->frame[state->curframe + 1]) {
6668 verbose(env, "verifier bug. Frame %d already allocated\n",
6669 state->curframe + 1);
6673 func_info_aux = env->prog->aux->func_info_aux;
6675 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
6676 err = btf_check_subprog_call(env, subprog, caller->regs);
6681 verbose(env, "Caller passes invalid args into func#%d\n",
6685 if (env->log.level & BPF_LOG_LEVEL)
6687 "Func#%d is global and valid. Skipping.\n",
6689 clear_caller_saved_regs(env, caller->regs);
6691 /* All global functions return a 64-bit SCALAR_VALUE */
6692 mark_reg_unknown(env, caller->regs, BPF_REG_0);
6693 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
6695 /* continue with next insn after call */
6700 if (insn->code == (BPF_JMP | BPF_CALL) &&
6701 insn->src_reg == 0 &&
6702 insn->imm == BPF_FUNC_timer_set_callback) {
6703 struct bpf_verifier_state *async_cb;
6705 /* there is no real recursion here. timer callbacks are async */
6706 env->subprog_info[subprog].is_async_cb = true;
6707 async_cb = push_async_cb(env, env->subprog_info[subprog].start,
6708 *insn_idx, subprog);
6711 callee = async_cb->frame[0];
6712 callee->async_entry_cnt = caller->async_entry_cnt + 1;
6714 /* Convert bpf_timer_set_callback() args into timer callback args */
6715 err = set_callee_state_cb(env, caller, callee, *insn_idx);
6719 clear_caller_saved_regs(env, caller->regs);
6720 mark_reg_unknown(env, caller->regs, BPF_REG_0);
6721 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
6722 /* continue with next insn after call */
6726 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
6729 state->frame[state->curframe + 1] = callee;
6731 /* callee cannot access r0, r6 - r9 for reading and has to write
6732 * into its own stack before reading from it.
6733 * callee can read/write into caller's stack
6735 init_func_state(env, callee,
6736 /* remember the callsite, it will be used by bpf_exit */
6737 *insn_idx /* callsite */,
6738 state->curframe + 1 /* frameno within this callchain */,
6739 subprog /* subprog number within this prog */);
6741 /* Transfer references to the callee */
6742 err = copy_reference_state(callee, caller);
6746 err = set_callee_state_cb(env, caller, callee, *insn_idx);
6750 clear_caller_saved_regs(env, caller->regs);
6752 /* only increment it after check_reg_arg() finished */
6755 /* and go analyze first insn of the callee */
6756 *insn_idx = env->subprog_info[subprog].start - 1;
6758 if (env->log.level & BPF_LOG_LEVEL) {
6759 verbose(env, "caller:\n");
6760 print_verifier_state(env, caller, true);
6761 verbose(env, "callee:\n");
6762 print_verifier_state(env, callee, true);
6767 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
6768 struct bpf_func_state *caller,
6769 struct bpf_func_state *callee)
6771 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
6772 * void *callback_ctx, u64 flags);
6773 * callback_fn(struct bpf_map *map, void *key, void *value,
6774 * void *callback_ctx);
6776 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
6778 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
6779 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
6780 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
6782 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
6783 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
6784 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
6786 /* pointer to stack or null */
6787 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
6790 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
6794 static int set_callee_state(struct bpf_verifier_env *env,
6795 struct bpf_func_state *caller,
6796 struct bpf_func_state *callee, int insn_idx)
6800 /* copy r1 - r5 args that callee can access. The copy includes parent
6801 * pointers, which connects us up to the liveness chain
6803 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
6804 callee->regs[i] = caller->regs[i];
6808 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
6811 int subprog, target_insn;
6813 target_insn = *insn_idx + insn->imm + 1;
6814 subprog = find_subprog(env, target_insn);
6816 verbose(env, "verifier bug. No program starts at insn %d\n",
6821 return __check_func_call(env, insn, insn_idx, subprog, set_callee_state);
6824 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
6825 struct bpf_func_state *caller,
6826 struct bpf_func_state *callee,
6829 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
6830 struct bpf_map *map;
6833 if (bpf_map_ptr_poisoned(insn_aux)) {
6834 verbose(env, "tail_call abusing map_ptr\n");
6838 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
6839 if (!map->ops->map_set_for_each_callback_args ||
6840 !map->ops->map_for_each_callback) {
6841 verbose(env, "callback function not allowed for map\n");
6845 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
6849 callee->in_callback_fn = true;
6850 callee->callback_ret_range = tnum_range(0, 1);
6854 static int set_loop_callback_state(struct bpf_verifier_env *env,
6855 struct bpf_func_state *caller,
6856 struct bpf_func_state *callee,
6859 /* bpf_loop(u32 nr_loops, void *callback_fn, void *callback_ctx,
6861 * callback_fn(u32 index, void *callback_ctx);
6863 callee->regs[BPF_REG_1].type = SCALAR_VALUE;
6864 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
6867 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
6868 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
6869 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
6871 callee->in_callback_fn = true;
6872 callee->callback_ret_range = tnum_range(0, 1);
6876 static int set_timer_callback_state(struct bpf_verifier_env *env,
6877 struct bpf_func_state *caller,
6878 struct bpf_func_state *callee,
6881 struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
6883 /* bpf_timer_set_callback(struct bpf_timer *timer, void *callback_fn);
6884 * callback_fn(struct bpf_map *map, void *key, void *value);
6886 callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP;
6887 __mark_reg_known_zero(&callee->regs[BPF_REG_1]);
6888 callee->regs[BPF_REG_1].map_ptr = map_ptr;
6890 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
6891 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
6892 callee->regs[BPF_REG_2].map_ptr = map_ptr;
6894 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
6895 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
6896 callee->regs[BPF_REG_3].map_ptr = map_ptr;
6899 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
6900 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
6901 callee->in_async_callback_fn = true;
6902 callee->callback_ret_range = tnum_range(0, 1);
6906 static int set_find_vma_callback_state(struct bpf_verifier_env *env,
6907 struct bpf_func_state *caller,
6908 struct bpf_func_state *callee,
6911 /* bpf_find_vma(struct task_struct *task, u64 addr,
6912 * void *callback_fn, void *callback_ctx, u64 flags)
6913 * (callback_fn)(struct task_struct *task,
6914 * struct vm_area_struct *vma, void *callback_ctx);
6916 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
6918 callee->regs[BPF_REG_2].type = PTR_TO_BTF_ID;
6919 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
6920 callee->regs[BPF_REG_2].btf = btf_vmlinux;
6921 callee->regs[BPF_REG_2].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_VMA],
6923 /* pointer to stack or null */
6924 callee->regs[BPF_REG_3] = caller->regs[BPF_REG_4];
6927 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
6928 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
6929 callee->in_callback_fn = true;
6930 callee->callback_ret_range = tnum_range(0, 1);
6934 static int set_user_ringbuf_callback_state(struct bpf_verifier_env *env,
6935 struct bpf_func_state *caller,
6936 struct bpf_func_state *callee,
6939 /* bpf_user_ringbuf_drain(struct bpf_map *map, void *callback_fn, void
6940 * callback_ctx, u64 flags);
6941 * callback_fn(struct bpf_dynptr_t* dynptr, void *callback_ctx);
6943 __mark_reg_not_init(env, &callee->regs[BPF_REG_0]);
6944 callee->regs[BPF_REG_1].type = PTR_TO_DYNPTR | DYNPTR_TYPE_LOCAL;
6945 __mark_reg_known_zero(&callee->regs[BPF_REG_1]);
6946 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
6949 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
6950 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
6951 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
6953 callee->in_callback_fn = true;
6954 callee->callback_ret_range = tnum_range(0, 1);
6958 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
6960 struct bpf_verifier_state *state = env->cur_state;
6961 struct bpf_func_state *caller, *callee;
6962 struct bpf_reg_state *r0;
6965 callee = state->frame[state->curframe];
6966 r0 = &callee->regs[BPF_REG_0];
6967 if (r0->type == PTR_TO_STACK) {
6968 /* technically it's ok to return caller's stack pointer
6969 * (or caller's caller's pointer) back to the caller,
6970 * since these pointers are valid. Only current stack
6971 * pointer will be invalid as soon as function exits,
6972 * but let's be conservative
6974 verbose(env, "cannot return stack pointer to the caller\n");
6979 caller = state->frame[state->curframe];
6980 if (callee->in_callback_fn) {
6981 /* enforce R0 return value range [0, 1]. */
6982 struct tnum range = callee->callback_ret_range;
6984 if (r0->type != SCALAR_VALUE) {
6985 verbose(env, "R0 not a scalar value\n");
6988 if (!tnum_in(range, r0->var_off)) {
6989 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
6993 /* return to the caller whatever r0 had in the callee */
6994 caller->regs[BPF_REG_0] = *r0;
6997 /* callback_fn frame should have released its own additions to parent's
6998 * reference state at this point, or check_reference_leak would
6999 * complain, hence it must be the same as the caller. There is no need
7002 if (!callee->in_callback_fn) {
7003 /* Transfer references to the caller */
7004 err = copy_reference_state(caller, callee);
7009 *insn_idx = callee->callsite + 1;
7010 if (env->log.level & BPF_LOG_LEVEL) {
7011 verbose(env, "returning from callee:\n");
7012 print_verifier_state(env, callee, true);
7013 verbose(env, "to caller at %d:\n", *insn_idx);
7014 print_verifier_state(env, caller, true);
7016 /* clear everything in the callee */
7017 free_func_state(callee);
7018 state->frame[state->curframe + 1] = NULL;
7022 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
7024 struct bpf_call_arg_meta *meta)
7026 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
7028 if (ret_type != RET_INTEGER ||
7029 (func_id != BPF_FUNC_get_stack &&
7030 func_id != BPF_FUNC_get_task_stack &&
7031 func_id != BPF_FUNC_probe_read_str &&
7032 func_id != BPF_FUNC_probe_read_kernel_str &&
7033 func_id != BPF_FUNC_probe_read_user_str))
7036 ret_reg->smax_value = meta->msize_max_value;
7037 ret_reg->s32_max_value = meta->msize_max_value;
7038 ret_reg->smin_value = -MAX_ERRNO;
7039 ret_reg->s32_min_value = -MAX_ERRNO;
7040 reg_bounds_sync(ret_reg);
7044 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
7045 int func_id, int insn_idx)
7047 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
7048 struct bpf_map *map = meta->map_ptr;
7050 if (func_id != BPF_FUNC_tail_call &&
7051 func_id != BPF_FUNC_map_lookup_elem &&
7052 func_id != BPF_FUNC_map_update_elem &&
7053 func_id != BPF_FUNC_map_delete_elem &&
7054 func_id != BPF_FUNC_map_push_elem &&
7055 func_id != BPF_FUNC_map_pop_elem &&
7056 func_id != BPF_FUNC_map_peek_elem &&
7057 func_id != BPF_FUNC_for_each_map_elem &&
7058 func_id != BPF_FUNC_redirect_map &&
7059 func_id != BPF_FUNC_map_lookup_percpu_elem)
7063 verbose(env, "kernel subsystem misconfigured verifier\n");
7067 /* In case of read-only, some additional restrictions
7068 * need to be applied in order to prevent altering the
7069 * state of the map from program side.
7071 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
7072 (func_id == BPF_FUNC_map_delete_elem ||
7073 func_id == BPF_FUNC_map_update_elem ||
7074 func_id == BPF_FUNC_map_push_elem ||
7075 func_id == BPF_FUNC_map_pop_elem)) {
7076 verbose(env, "write into map forbidden\n");
7080 if (!BPF_MAP_PTR(aux->map_ptr_state))
7081 bpf_map_ptr_store(aux, meta->map_ptr,
7082 !meta->map_ptr->bypass_spec_v1);
7083 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
7084 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
7085 !meta->map_ptr->bypass_spec_v1);
7090 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
7091 int func_id, int insn_idx)
7093 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
7094 struct bpf_reg_state *regs = cur_regs(env), *reg;
7095 struct bpf_map *map = meta->map_ptr;
7099 if (func_id != BPF_FUNC_tail_call)
7101 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
7102 verbose(env, "kernel subsystem misconfigured verifier\n");
7106 reg = ®s[BPF_REG_3];
7107 val = reg->var_off.value;
7108 max = map->max_entries;
7110 if (!(register_is_const(reg) && val < max)) {
7111 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
7115 err = mark_chain_precision(env, BPF_REG_3);
7118 if (bpf_map_key_unseen(aux))
7119 bpf_map_key_store(aux, val);
7120 else if (!bpf_map_key_poisoned(aux) &&
7121 bpf_map_key_immediate(aux) != val)
7122 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
7126 static int check_reference_leak(struct bpf_verifier_env *env)
7128 struct bpf_func_state *state = cur_func(env);
7129 bool refs_lingering = false;
7132 if (state->frameno && !state->in_callback_fn)
7135 for (i = 0; i < state->acquired_refs; i++) {
7136 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno)
7138 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
7139 state->refs[i].id, state->refs[i].insn_idx);
7140 refs_lingering = true;
7142 return refs_lingering ? -EINVAL : 0;
7145 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
7146 struct bpf_reg_state *regs)
7148 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
7149 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
7150 struct bpf_map *fmt_map = fmt_reg->map_ptr;
7151 int err, fmt_map_off, num_args;
7155 /* data must be an array of u64 */
7156 if (data_len_reg->var_off.value % 8)
7158 num_args = data_len_reg->var_off.value / 8;
7160 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
7161 * and map_direct_value_addr is set.
7163 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
7164 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
7167 verbose(env, "verifier bug\n");
7170 fmt = (char *)(long)fmt_addr + fmt_map_off;
7172 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
7173 * can focus on validating the format specifiers.
7175 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, NULL, num_args);
7177 verbose(env, "Invalid format string\n");
7182 static int check_get_func_ip(struct bpf_verifier_env *env)
7184 enum bpf_prog_type type = resolve_prog_type(env->prog);
7185 int func_id = BPF_FUNC_get_func_ip;
7187 if (type == BPF_PROG_TYPE_TRACING) {
7188 if (!bpf_prog_has_trampoline(env->prog)) {
7189 verbose(env, "func %s#%d supported only for fentry/fexit/fmod_ret programs\n",
7190 func_id_name(func_id), func_id);
7194 } else if (type == BPF_PROG_TYPE_KPROBE) {
7198 verbose(env, "func %s#%d not supported for program type %d\n",
7199 func_id_name(func_id), func_id, type);
7203 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
7205 return &env->insn_aux_data[env->insn_idx];
7208 static bool loop_flag_is_zero(struct bpf_verifier_env *env)
7210 struct bpf_reg_state *regs = cur_regs(env);
7211 struct bpf_reg_state *reg = ®s[BPF_REG_4];
7212 bool reg_is_null = register_is_null(reg);
7215 mark_chain_precision(env, BPF_REG_4);
7220 static void update_loop_inline_state(struct bpf_verifier_env *env, u32 subprogno)
7222 struct bpf_loop_inline_state *state = &cur_aux(env)->loop_inline_state;
7224 if (!state->initialized) {
7225 state->initialized = 1;
7226 state->fit_for_inline = loop_flag_is_zero(env);
7227 state->callback_subprogno = subprogno;
7231 if (!state->fit_for_inline)
7234 state->fit_for_inline = (loop_flag_is_zero(env) &&
7235 state->callback_subprogno == subprogno);
7238 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
7241 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
7242 const struct bpf_func_proto *fn = NULL;
7243 enum bpf_return_type ret_type;
7244 enum bpf_type_flag ret_flag;
7245 struct bpf_reg_state *regs;
7246 struct bpf_call_arg_meta meta;
7247 int insn_idx = *insn_idx_p;
7249 int i, err, func_id;
7251 /* find function prototype */
7252 func_id = insn->imm;
7253 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
7254 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
7259 if (env->ops->get_func_proto)
7260 fn = env->ops->get_func_proto(func_id, env->prog);
7262 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
7267 /* eBPF programs must be GPL compatible to use GPL-ed functions */
7268 if (!env->prog->gpl_compatible && fn->gpl_only) {
7269 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
7273 if (fn->allowed && !fn->allowed(env->prog)) {
7274 verbose(env, "helper call is not allowed in probe\n");
7278 /* With LD_ABS/IND some JITs save/restore skb from r1. */
7279 changes_data = bpf_helper_changes_pkt_data(fn->func);
7280 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
7281 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
7282 func_id_name(func_id), func_id);
7286 memset(&meta, 0, sizeof(meta));
7287 meta.pkt_access = fn->pkt_access;
7289 err = check_func_proto(fn, func_id);
7291 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
7292 func_id_name(func_id), func_id);
7296 meta.func_id = func_id;
7298 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
7299 err = check_func_arg(env, i, &meta, fn);
7304 err = record_func_map(env, &meta, func_id, insn_idx);
7308 err = record_func_key(env, &meta, func_id, insn_idx);
7312 /* Mark slots with STACK_MISC in case of raw mode, stack offset
7313 * is inferred from register state.
7315 for (i = 0; i < meta.access_size; i++) {
7316 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
7317 BPF_WRITE, -1, false);
7322 regs = cur_regs(env);
7324 if (meta.uninit_dynptr_regno) {
7325 /* we write BPF_DW bits (8 bytes) at a time */
7326 for (i = 0; i < BPF_DYNPTR_SIZE; i += 8) {
7327 err = check_mem_access(env, insn_idx, meta.uninit_dynptr_regno,
7328 i, BPF_DW, BPF_WRITE, -1, false);
7333 err = mark_stack_slots_dynptr(env, ®s[meta.uninit_dynptr_regno],
7334 fn->arg_type[meta.uninit_dynptr_regno - BPF_REG_1],
7340 if (meta.release_regno) {
7342 if (arg_type_is_dynptr(fn->arg_type[meta.release_regno - BPF_REG_1]))
7343 err = unmark_stack_slots_dynptr(env, ®s[meta.release_regno]);
7344 else if (meta.ref_obj_id)
7345 err = release_reference(env, meta.ref_obj_id);
7346 /* meta.ref_obj_id can only be 0 if register that is meant to be
7347 * released is NULL, which must be > R0.
7349 else if (register_is_null(®s[meta.release_regno]))
7352 verbose(env, "func %s#%d reference has not been acquired before\n",
7353 func_id_name(func_id), func_id);
7359 case BPF_FUNC_tail_call:
7360 err = check_reference_leak(env);
7362 verbose(env, "tail_call would lead to reference leak\n");
7366 case BPF_FUNC_get_local_storage:
7367 /* check that flags argument in get_local_storage(map, flags) is 0,
7368 * this is required because get_local_storage() can't return an error.
7370 if (!register_is_null(®s[BPF_REG_2])) {
7371 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
7375 case BPF_FUNC_for_each_map_elem:
7376 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
7377 set_map_elem_callback_state);
7379 case BPF_FUNC_timer_set_callback:
7380 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
7381 set_timer_callback_state);
7383 case BPF_FUNC_find_vma:
7384 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
7385 set_find_vma_callback_state);
7387 case BPF_FUNC_snprintf:
7388 err = check_bpf_snprintf_call(env, regs);
7391 update_loop_inline_state(env, meta.subprogno);
7392 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
7393 set_loop_callback_state);
7395 case BPF_FUNC_dynptr_from_mem:
7396 if (regs[BPF_REG_1].type != PTR_TO_MAP_VALUE) {
7397 verbose(env, "Unsupported reg type %s for bpf_dynptr_from_mem data\n",
7398 reg_type_str(env, regs[BPF_REG_1].type));
7402 case BPF_FUNC_set_retval:
7403 if (prog_type == BPF_PROG_TYPE_LSM &&
7404 env->prog->expected_attach_type == BPF_LSM_CGROUP) {
7405 if (!env->prog->aux->attach_func_proto->type) {
7406 /* Make sure programs that attach to void
7407 * hooks don't try to modify return value.
7409 verbose(env, "BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
7414 case BPF_FUNC_dynptr_data:
7415 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
7416 if (arg_type_is_dynptr(fn->arg_type[i])) {
7417 struct bpf_reg_state *reg = ®s[BPF_REG_1 + i];
7419 if (meta.ref_obj_id) {
7420 verbose(env, "verifier internal error: meta.ref_obj_id already set\n");
7424 if (base_type(reg->type) != PTR_TO_DYNPTR)
7425 /* Find the id of the dynptr we're
7426 * tracking the reference of
7428 meta.ref_obj_id = stack_slot_get_id(env, reg);
7432 if (i == MAX_BPF_FUNC_REG_ARGS) {
7433 verbose(env, "verifier internal error: no dynptr in bpf_dynptr_data()\n");
7437 case BPF_FUNC_user_ringbuf_drain:
7438 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
7439 set_user_ringbuf_callback_state);
7446 /* reset caller saved regs */
7447 for (i = 0; i < CALLER_SAVED_REGS; i++) {
7448 mark_reg_not_init(env, regs, caller_saved[i]);
7449 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
7452 /* helper call returns 64-bit value. */
7453 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
7455 /* update return register (already marked as written above) */
7456 ret_type = fn->ret_type;
7457 ret_flag = type_flag(ret_type);
7459 switch (base_type(ret_type)) {
7461 /* sets type to SCALAR_VALUE */
7462 mark_reg_unknown(env, regs, BPF_REG_0);
7465 regs[BPF_REG_0].type = NOT_INIT;
7467 case RET_PTR_TO_MAP_VALUE:
7468 /* There is no offset yet applied, variable or fixed */
7469 mark_reg_known_zero(env, regs, BPF_REG_0);
7470 /* remember map_ptr, so that check_map_access()
7471 * can check 'value_size' boundary of memory access
7472 * to map element returned from bpf_map_lookup_elem()
7474 if (meta.map_ptr == NULL) {
7476 "kernel subsystem misconfigured verifier\n");
7479 regs[BPF_REG_0].map_ptr = meta.map_ptr;
7480 regs[BPF_REG_0].map_uid = meta.map_uid;
7481 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE | ret_flag;
7482 if (!type_may_be_null(ret_type) &&
7483 map_value_has_spin_lock(meta.map_ptr)) {
7484 regs[BPF_REG_0].id = ++env->id_gen;
7487 case RET_PTR_TO_SOCKET:
7488 mark_reg_known_zero(env, regs, BPF_REG_0);
7489 regs[BPF_REG_0].type = PTR_TO_SOCKET | ret_flag;
7491 case RET_PTR_TO_SOCK_COMMON:
7492 mark_reg_known_zero(env, regs, BPF_REG_0);
7493 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON | ret_flag;
7495 case RET_PTR_TO_TCP_SOCK:
7496 mark_reg_known_zero(env, regs, BPF_REG_0);
7497 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK | ret_flag;
7499 case RET_PTR_TO_ALLOC_MEM:
7500 mark_reg_known_zero(env, regs, BPF_REG_0);
7501 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
7502 regs[BPF_REG_0].mem_size = meta.mem_size;
7504 case RET_PTR_TO_MEM_OR_BTF_ID:
7506 const struct btf_type *t;
7508 mark_reg_known_zero(env, regs, BPF_REG_0);
7509 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
7510 if (!btf_type_is_struct(t)) {
7512 const struct btf_type *ret;
7515 /* resolve the type size of ksym. */
7516 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
7518 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
7519 verbose(env, "unable to resolve the size of type '%s': %ld\n",
7520 tname, PTR_ERR(ret));
7523 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
7524 regs[BPF_REG_0].mem_size = tsize;
7526 /* MEM_RDONLY may be carried from ret_flag, but it
7527 * doesn't apply on PTR_TO_BTF_ID. Fold it, otherwise
7528 * it will confuse the check of PTR_TO_BTF_ID in
7529 * check_mem_access().
7531 ret_flag &= ~MEM_RDONLY;
7533 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
7534 regs[BPF_REG_0].btf = meta.ret_btf;
7535 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
7539 case RET_PTR_TO_BTF_ID:
7541 struct btf *ret_btf;
7544 mark_reg_known_zero(env, regs, BPF_REG_0);
7545 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
7546 if (func_id == BPF_FUNC_kptr_xchg) {
7547 ret_btf = meta.kptr_off_desc->kptr.btf;
7548 ret_btf_id = meta.kptr_off_desc->kptr.btf_id;
7550 if (fn->ret_btf_id == BPF_PTR_POISON) {
7551 verbose(env, "verifier internal error:");
7552 verbose(env, "func %s has non-overwritten BPF_PTR_POISON return type\n",
7553 func_id_name(func_id));
7556 ret_btf = btf_vmlinux;
7557 ret_btf_id = *fn->ret_btf_id;
7559 if (ret_btf_id == 0) {
7560 verbose(env, "invalid return type %u of func %s#%d\n",
7561 base_type(ret_type), func_id_name(func_id),
7565 regs[BPF_REG_0].btf = ret_btf;
7566 regs[BPF_REG_0].btf_id = ret_btf_id;
7570 verbose(env, "unknown return type %u of func %s#%d\n",
7571 base_type(ret_type), func_id_name(func_id), func_id);
7575 if (type_may_be_null(regs[BPF_REG_0].type))
7576 regs[BPF_REG_0].id = ++env->id_gen;
7578 if (helper_multiple_ref_obj_use(func_id, meta.map_ptr)) {
7579 verbose(env, "verifier internal error: func %s#%d sets ref_obj_id more than once\n",
7580 func_id_name(func_id), func_id);
7584 if (is_ptr_cast_function(func_id) || is_dynptr_ref_function(func_id)) {
7585 /* For release_reference() */
7586 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
7587 } else if (is_acquire_function(func_id, meta.map_ptr)) {
7588 int id = acquire_reference_state(env, insn_idx);
7592 /* For mark_ptr_or_null_reg() */
7593 regs[BPF_REG_0].id = id;
7594 /* For release_reference() */
7595 regs[BPF_REG_0].ref_obj_id = id;
7598 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
7600 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
7604 if ((func_id == BPF_FUNC_get_stack ||
7605 func_id == BPF_FUNC_get_task_stack) &&
7606 !env->prog->has_callchain_buf) {
7607 const char *err_str;
7609 #ifdef CONFIG_PERF_EVENTS
7610 err = get_callchain_buffers(sysctl_perf_event_max_stack);
7611 err_str = "cannot get callchain buffer for func %s#%d\n";
7614 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
7617 verbose(env, err_str, func_id_name(func_id), func_id);
7621 env->prog->has_callchain_buf = true;
7624 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
7625 env->prog->call_get_stack = true;
7627 if (func_id == BPF_FUNC_get_func_ip) {
7628 if (check_get_func_ip(env))
7630 env->prog->call_get_func_ip = true;
7634 clear_all_pkt_pointers(env);
7638 /* mark_btf_func_reg_size() is used when the reg size is determined by
7639 * the BTF func_proto's return value size and argument.
7641 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
7644 struct bpf_reg_state *reg = &cur_regs(env)[regno];
7646 if (regno == BPF_REG_0) {
7647 /* Function return value */
7648 reg->live |= REG_LIVE_WRITTEN;
7649 reg->subreg_def = reg_size == sizeof(u64) ?
7650 DEF_NOT_SUBREG : env->insn_idx + 1;
7652 /* Function argument */
7653 if (reg_size == sizeof(u64)) {
7654 mark_insn_zext(env, reg);
7655 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
7657 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
7662 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
7665 const struct btf_type *t, *func, *func_proto, *ptr_type;
7666 struct bpf_reg_state *regs = cur_regs(env);
7667 struct bpf_kfunc_arg_meta meta = { 0 };
7668 const char *func_name, *ptr_type_name;
7669 u32 i, nargs, func_id, ptr_type_id;
7670 int err, insn_idx = *insn_idx_p;
7671 const struct btf_param *args;
7672 struct btf *desc_btf;
7676 /* skip for now, but return error when we find this in fixup_kfunc_call */
7680 desc_btf = find_kfunc_desc_btf(env, insn->off);
7681 if (IS_ERR(desc_btf))
7682 return PTR_ERR(desc_btf);
7684 func_id = insn->imm;
7685 func = btf_type_by_id(desc_btf, func_id);
7686 func_name = btf_name_by_offset(desc_btf, func->name_off);
7687 func_proto = btf_type_by_id(desc_btf, func->type);
7689 kfunc_flags = btf_kfunc_id_set_contains(desc_btf, resolve_prog_type(env->prog), func_id);
7691 verbose(env, "calling kernel function %s is not allowed\n",
7695 if (*kfunc_flags & KF_DESTRUCTIVE && !capable(CAP_SYS_BOOT)) {
7696 verbose(env, "destructive kfunc calls require CAP_SYS_BOOT capabilities\n");
7700 acq = *kfunc_flags & KF_ACQUIRE;
7702 meta.flags = *kfunc_flags;
7704 /* Check the arguments */
7705 err = btf_check_kfunc_arg_match(env, desc_btf, func_id, regs, &meta);
7708 /* In case of release function, we get register number of refcounted
7709 * PTR_TO_BTF_ID back from btf_check_kfunc_arg_match, do the release now
7712 err = release_reference(env, regs[err].ref_obj_id);
7714 verbose(env, "kfunc %s#%d reference has not been acquired before\n",
7715 func_name, func_id);
7720 for (i = 0; i < CALLER_SAVED_REGS; i++)
7721 mark_reg_not_init(env, regs, caller_saved[i]);
7723 /* Check return type */
7724 t = btf_type_skip_modifiers(desc_btf, func_proto->type, NULL);
7726 if (acq && !btf_type_is_struct_ptr(desc_btf, t)) {
7727 verbose(env, "acquire kernel function does not return PTR_TO_BTF_ID\n");
7731 if (btf_type_is_scalar(t)) {
7732 mark_reg_unknown(env, regs, BPF_REG_0);
7733 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
7734 } else if (btf_type_is_ptr(t)) {
7735 ptr_type = btf_type_skip_modifiers(desc_btf, t->type,
7737 if (!btf_type_is_struct(ptr_type)) {
7738 if (!meta.r0_size) {
7739 ptr_type_name = btf_name_by_offset(desc_btf,
7740 ptr_type->name_off);
7742 "kernel function %s returns pointer type %s %s is not supported\n",
7744 btf_type_str(ptr_type),
7749 mark_reg_known_zero(env, regs, BPF_REG_0);
7750 regs[BPF_REG_0].type = PTR_TO_MEM;
7751 regs[BPF_REG_0].mem_size = meta.r0_size;
7754 regs[BPF_REG_0].type |= MEM_RDONLY;
7756 /* Ensures we don't access the memory after a release_reference() */
7757 if (meta.ref_obj_id)
7758 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
7760 mark_reg_known_zero(env, regs, BPF_REG_0);
7761 regs[BPF_REG_0].btf = desc_btf;
7762 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
7763 regs[BPF_REG_0].btf_id = ptr_type_id;
7765 if (*kfunc_flags & KF_RET_NULL) {
7766 regs[BPF_REG_0].type |= PTR_MAYBE_NULL;
7767 /* For mark_ptr_or_null_reg, see 93c230e3f5bd6 */
7768 regs[BPF_REG_0].id = ++env->id_gen;
7770 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
7772 int id = acquire_reference_state(env, insn_idx);
7776 regs[BPF_REG_0].id = id;
7777 regs[BPF_REG_0].ref_obj_id = id;
7779 } /* else { add_kfunc_call() ensures it is btf_type_is_void(t) } */
7781 nargs = btf_type_vlen(func_proto);
7782 args = (const struct btf_param *)(func_proto + 1);
7783 for (i = 0; i < nargs; i++) {
7786 t = btf_type_skip_modifiers(desc_btf, args[i].type, NULL);
7787 if (btf_type_is_ptr(t))
7788 mark_btf_func_reg_size(env, regno, sizeof(void *));
7790 /* scalar. ensured by btf_check_kfunc_arg_match() */
7791 mark_btf_func_reg_size(env, regno, t->size);
7797 static bool signed_add_overflows(s64 a, s64 b)
7799 /* Do the add in u64, where overflow is well-defined */
7800 s64 res = (s64)((u64)a + (u64)b);
7807 static bool signed_add32_overflows(s32 a, s32 b)
7809 /* Do the add in u32, where overflow is well-defined */
7810 s32 res = (s32)((u32)a + (u32)b);
7817 static bool signed_sub_overflows(s64 a, s64 b)
7819 /* Do the sub in u64, where overflow is well-defined */
7820 s64 res = (s64)((u64)a - (u64)b);
7827 static bool signed_sub32_overflows(s32 a, s32 b)
7829 /* Do the sub in u32, where overflow is well-defined */
7830 s32 res = (s32)((u32)a - (u32)b);
7837 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
7838 const struct bpf_reg_state *reg,
7839 enum bpf_reg_type type)
7841 bool known = tnum_is_const(reg->var_off);
7842 s64 val = reg->var_off.value;
7843 s64 smin = reg->smin_value;
7845 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
7846 verbose(env, "math between %s pointer and %lld is not allowed\n",
7847 reg_type_str(env, type), val);
7851 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
7852 verbose(env, "%s pointer offset %d is not allowed\n",
7853 reg_type_str(env, type), reg->off);
7857 if (smin == S64_MIN) {
7858 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
7859 reg_type_str(env, type));
7863 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
7864 verbose(env, "value %lld makes %s pointer be out of bounds\n",
7865 smin, reg_type_str(env, type));
7880 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
7881 u32 *alu_limit, bool mask_to_left)
7883 u32 max = 0, ptr_limit = 0;
7885 switch (ptr_reg->type) {
7887 /* Offset 0 is out-of-bounds, but acceptable start for the
7888 * left direction, see BPF_REG_FP. Also, unknown scalar
7889 * offset where we would need to deal with min/max bounds is
7890 * currently prohibited for unprivileged.
7892 max = MAX_BPF_STACK + mask_to_left;
7893 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
7895 case PTR_TO_MAP_VALUE:
7896 max = ptr_reg->map_ptr->value_size;
7897 ptr_limit = (mask_to_left ?
7898 ptr_reg->smin_value :
7899 ptr_reg->umax_value) + ptr_reg->off;
7905 if (ptr_limit >= max)
7906 return REASON_LIMIT;
7907 *alu_limit = ptr_limit;
7911 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
7912 const struct bpf_insn *insn)
7914 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
7917 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
7918 u32 alu_state, u32 alu_limit)
7920 /* If we arrived here from different branches with different
7921 * state or limits to sanitize, then this won't work.
7923 if (aux->alu_state &&
7924 (aux->alu_state != alu_state ||
7925 aux->alu_limit != alu_limit))
7926 return REASON_PATHS;
7928 /* Corresponding fixup done in do_misc_fixups(). */
7929 aux->alu_state = alu_state;
7930 aux->alu_limit = alu_limit;
7934 static int sanitize_val_alu(struct bpf_verifier_env *env,
7935 struct bpf_insn *insn)
7937 struct bpf_insn_aux_data *aux = cur_aux(env);
7939 if (can_skip_alu_sanitation(env, insn))
7942 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
7945 static bool sanitize_needed(u8 opcode)
7947 return opcode == BPF_ADD || opcode == BPF_SUB;
7950 struct bpf_sanitize_info {
7951 struct bpf_insn_aux_data aux;
7955 static struct bpf_verifier_state *
7956 sanitize_speculative_path(struct bpf_verifier_env *env,
7957 const struct bpf_insn *insn,
7958 u32 next_idx, u32 curr_idx)
7960 struct bpf_verifier_state *branch;
7961 struct bpf_reg_state *regs;
7963 branch = push_stack(env, next_idx, curr_idx, true);
7964 if (branch && insn) {
7965 regs = branch->frame[branch->curframe]->regs;
7966 if (BPF_SRC(insn->code) == BPF_K) {
7967 mark_reg_unknown(env, regs, insn->dst_reg);
7968 } else if (BPF_SRC(insn->code) == BPF_X) {
7969 mark_reg_unknown(env, regs, insn->dst_reg);
7970 mark_reg_unknown(env, regs, insn->src_reg);
7976 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
7977 struct bpf_insn *insn,
7978 const struct bpf_reg_state *ptr_reg,
7979 const struct bpf_reg_state *off_reg,
7980 struct bpf_reg_state *dst_reg,
7981 struct bpf_sanitize_info *info,
7982 const bool commit_window)
7984 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
7985 struct bpf_verifier_state *vstate = env->cur_state;
7986 bool off_is_imm = tnum_is_const(off_reg->var_off);
7987 bool off_is_neg = off_reg->smin_value < 0;
7988 bool ptr_is_dst_reg = ptr_reg == dst_reg;
7989 u8 opcode = BPF_OP(insn->code);
7990 u32 alu_state, alu_limit;
7991 struct bpf_reg_state tmp;
7995 if (can_skip_alu_sanitation(env, insn))
7998 /* We already marked aux for masking from non-speculative
7999 * paths, thus we got here in the first place. We only care
8000 * to explore bad access from here.
8002 if (vstate->speculative)
8005 if (!commit_window) {
8006 if (!tnum_is_const(off_reg->var_off) &&
8007 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
8008 return REASON_BOUNDS;
8010 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
8011 (opcode == BPF_SUB && !off_is_neg);
8014 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
8018 if (commit_window) {
8019 /* In commit phase we narrow the masking window based on
8020 * the observed pointer move after the simulated operation.
8022 alu_state = info->aux.alu_state;
8023 alu_limit = abs(info->aux.alu_limit - alu_limit);
8025 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
8026 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
8027 alu_state |= ptr_is_dst_reg ?
8028 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
8030 /* Limit pruning on unknown scalars to enable deep search for
8031 * potential masking differences from other program paths.
8034 env->explore_alu_limits = true;
8037 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
8041 /* If we're in commit phase, we're done here given we already
8042 * pushed the truncated dst_reg into the speculative verification
8045 * Also, when register is a known constant, we rewrite register-based
8046 * operation to immediate-based, and thus do not need masking (and as
8047 * a consequence, do not need to simulate the zero-truncation either).
8049 if (commit_window || off_is_imm)
8052 /* Simulate and find potential out-of-bounds access under
8053 * speculative execution from truncation as a result of
8054 * masking when off was not within expected range. If off
8055 * sits in dst, then we temporarily need to move ptr there
8056 * to simulate dst (== 0) +/-= ptr. Needed, for example,
8057 * for cases where we use K-based arithmetic in one direction
8058 * and truncated reg-based in the other in order to explore
8061 if (!ptr_is_dst_reg) {
8063 *dst_reg = *ptr_reg;
8065 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
8067 if (!ptr_is_dst_reg && ret)
8069 return !ret ? REASON_STACK : 0;
8072 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
8074 struct bpf_verifier_state *vstate = env->cur_state;
8076 /* If we simulate paths under speculation, we don't update the
8077 * insn as 'seen' such that when we verify unreachable paths in
8078 * the non-speculative domain, sanitize_dead_code() can still
8079 * rewrite/sanitize them.
8081 if (!vstate->speculative)
8082 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
8085 static int sanitize_err(struct bpf_verifier_env *env,
8086 const struct bpf_insn *insn, int reason,
8087 const struct bpf_reg_state *off_reg,
8088 const struct bpf_reg_state *dst_reg)
8090 static const char *err = "pointer arithmetic with it prohibited for !root";
8091 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
8092 u32 dst = insn->dst_reg, src = insn->src_reg;
8096 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
8097 off_reg == dst_reg ? dst : src, err);
8100 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
8101 off_reg == dst_reg ? src : dst, err);
8104 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
8108 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
8112 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
8116 verbose(env, "verifier internal error: unknown reason (%d)\n",
8124 /* check that stack access falls within stack limits and that 'reg' doesn't
8125 * have a variable offset.
8127 * Variable offset is prohibited for unprivileged mode for simplicity since it
8128 * requires corresponding support in Spectre masking for stack ALU. See also
8129 * retrieve_ptr_limit().
8132 * 'off' includes 'reg->off'.
8134 static int check_stack_access_for_ptr_arithmetic(
8135 struct bpf_verifier_env *env,
8137 const struct bpf_reg_state *reg,
8140 if (!tnum_is_const(reg->var_off)) {
8143 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
8144 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
8145 regno, tn_buf, off);
8149 if (off >= 0 || off < -MAX_BPF_STACK) {
8150 verbose(env, "R%d stack pointer arithmetic goes out of range, "
8151 "prohibited for !root; off=%d\n", regno, off);
8158 static int sanitize_check_bounds(struct bpf_verifier_env *env,
8159 const struct bpf_insn *insn,
8160 const struct bpf_reg_state *dst_reg)
8162 u32 dst = insn->dst_reg;
8164 /* For unprivileged we require that resulting offset must be in bounds
8165 * in order to be able to sanitize access later on.
8167 if (env->bypass_spec_v1)
8170 switch (dst_reg->type) {
8172 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
8173 dst_reg->off + dst_reg->var_off.value))
8176 case PTR_TO_MAP_VALUE:
8177 if (check_map_access(env, dst, dst_reg->off, 1, false, ACCESS_HELPER)) {
8178 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
8179 "prohibited for !root\n", dst);
8190 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
8191 * Caller should also handle BPF_MOV case separately.
8192 * If we return -EACCES, caller may want to try again treating pointer as a
8193 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
8195 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
8196 struct bpf_insn *insn,
8197 const struct bpf_reg_state *ptr_reg,
8198 const struct bpf_reg_state *off_reg)
8200 struct bpf_verifier_state *vstate = env->cur_state;
8201 struct bpf_func_state *state = vstate->frame[vstate->curframe];
8202 struct bpf_reg_state *regs = state->regs, *dst_reg;
8203 bool known = tnum_is_const(off_reg->var_off);
8204 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
8205 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
8206 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
8207 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
8208 struct bpf_sanitize_info info = {};
8209 u8 opcode = BPF_OP(insn->code);
8210 u32 dst = insn->dst_reg;
8213 dst_reg = ®s[dst];
8215 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
8216 smin_val > smax_val || umin_val > umax_val) {
8217 /* Taint dst register if offset had invalid bounds derived from
8218 * e.g. dead branches.
8220 __mark_reg_unknown(env, dst_reg);
8224 if (BPF_CLASS(insn->code) != BPF_ALU64) {
8225 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
8226 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
8227 __mark_reg_unknown(env, dst_reg);
8232 "R%d 32-bit pointer arithmetic prohibited\n",
8237 if (ptr_reg->type & PTR_MAYBE_NULL) {
8238 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
8239 dst, reg_type_str(env, ptr_reg->type));
8243 switch (base_type(ptr_reg->type)) {
8244 case CONST_PTR_TO_MAP:
8245 /* smin_val represents the known value */
8246 if (known && smin_val == 0 && opcode == BPF_ADD)
8249 case PTR_TO_PACKET_END:
8251 case PTR_TO_SOCK_COMMON:
8252 case PTR_TO_TCP_SOCK:
8253 case PTR_TO_XDP_SOCK:
8254 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
8255 dst, reg_type_str(env, ptr_reg->type));
8261 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
8262 * The id may be overwritten later if we create a new variable offset.
8264 dst_reg->type = ptr_reg->type;
8265 dst_reg->id = ptr_reg->id;
8267 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
8268 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
8271 /* pointer types do not carry 32-bit bounds at the moment. */
8272 __mark_reg32_unbounded(dst_reg);
8274 if (sanitize_needed(opcode)) {
8275 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
8278 return sanitize_err(env, insn, ret, off_reg, dst_reg);
8283 /* We can take a fixed offset as long as it doesn't overflow
8284 * the s32 'off' field
8286 if (known && (ptr_reg->off + smin_val ==
8287 (s64)(s32)(ptr_reg->off + smin_val))) {
8288 /* pointer += K. Accumulate it into fixed offset */
8289 dst_reg->smin_value = smin_ptr;
8290 dst_reg->smax_value = smax_ptr;
8291 dst_reg->umin_value = umin_ptr;
8292 dst_reg->umax_value = umax_ptr;
8293 dst_reg->var_off = ptr_reg->var_off;
8294 dst_reg->off = ptr_reg->off + smin_val;
8295 dst_reg->raw = ptr_reg->raw;
8298 /* A new variable offset is created. Note that off_reg->off
8299 * == 0, since it's a scalar.
8300 * dst_reg gets the pointer type and since some positive
8301 * integer value was added to the pointer, give it a new 'id'
8302 * if it's a PTR_TO_PACKET.
8303 * this creates a new 'base' pointer, off_reg (variable) gets
8304 * added into the variable offset, and we copy the fixed offset
8307 if (signed_add_overflows(smin_ptr, smin_val) ||
8308 signed_add_overflows(smax_ptr, smax_val)) {
8309 dst_reg->smin_value = S64_MIN;
8310 dst_reg->smax_value = S64_MAX;
8312 dst_reg->smin_value = smin_ptr + smin_val;
8313 dst_reg->smax_value = smax_ptr + smax_val;
8315 if (umin_ptr + umin_val < umin_ptr ||
8316 umax_ptr + umax_val < umax_ptr) {
8317 dst_reg->umin_value = 0;
8318 dst_reg->umax_value = U64_MAX;
8320 dst_reg->umin_value = umin_ptr + umin_val;
8321 dst_reg->umax_value = umax_ptr + umax_val;
8323 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
8324 dst_reg->off = ptr_reg->off;
8325 dst_reg->raw = ptr_reg->raw;
8326 if (reg_is_pkt_pointer(ptr_reg)) {
8327 dst_reg->id = ++env->id_gen;
8328 /* something was added to pkt_ptr, set range to zero */
8329 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
8333 if (dst_reg == off_reg) {
8334 /* scalar -= pointer. Creates an unknown scalar */
8335 verbose(env, "R%d tried to subtract pointer from scalar\n",
8339 /* We don't allow subtraction from FP, because (according to
8340 * test_verifier.c test "invalid fp arithmetic", JITs might not
8341 * be able to deal with it.
8343 if (ptr_reg->type == PTR_TO_STACK) {
8344 verbose(env, "R%d subtraction from stack pointer prohibited\n",
8348 if (known && (ptr_reg->off - smin_val ==
8349 (s64)(s32)(ptr_reg->off - smin_val))) {
8350 /* pointer -= K. Subtract it from fixed offset */
8351 dst_reg->smin_value = smin_ptr;
8352 dst_reg->smax_value = smax_ptr;
8353 dst_reg->umin_value = umin_ptr;
8354 dst_reg->umax_value = umax_ptr;
8355 dst_reg->var_off = ptr_reg->var_off;
8356 dst_reg->id = ptr_reg->id;
8357 dst_reg->off = ptr_reg->off - smin_val;
8358 dst_reg->raw = ptr_reg->raw;
8361 /* A new variable offset is created. If the subtrahend is known
8362 * nonnegative, then any reg->range we had before is still good.
8364 if (signed_sub_overflows(smin_ptr, smax_val) ||
8365 signed_sub_overflows(smax_ptr, smin_val)) {
8366 /* Overflow possible, we know nothing */
8367 dst_reg->smin_value = S64_MIN;
8368 dst_reg->smax_value = S64_MAX;
8370 dst_reg->smin_value = smin_ptr - smax_val;
8371 dst_reg->smax_value = smax_ptr - smin_val;
8373 if (umin_ptr < umax_val) {
8374 /* Overflow possible, we know nothing */
8375 dst_reg->umin_value = 0;
8376 dst_reg->umax_value = U64_MAX;
8378 /* Cannot overflow (as long as bounds are consistent) */
8379 dst_reg->umin_value = umin_ptr - umax_val;
8380 dst_reg->umax_value = umax_ptr - umin_val;
8382 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
8383 dst_reg->off = ptr_reg->off;
8384 dst_reg->raw = ptr_reg->raw;
8385 if (reg_is_pkt_pointer(ptr_reg)) {
8386 dst_reg->id = ++env->id_gen;
8387 /* something was added to pkt_ptr, set range to zero */
8389 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
8395 /* bitwise ops on pointers are troublesome, prohibit. */
8396 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
8397 dst, bpf_alu_string[opcode >> 4]);
8400 /* other operators (e.g. MUL,LSH) produce non-pointer results */
8401 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
8402 dst, bpf_alu_string[opcode >> 4]);
8406 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
8408 reg_bounds_sync(dst_reg);
8409 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
8411 if (sanitize_needed(opcode)) {
8412 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
8415 return sanitize_err(env, insn, ret, off_reg, dst_reg);
8421 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
8422 struct bpf_reg_state *src_reg)
8424 s32 smin_val = src_reg->s32_min_value;
8425 s32 smax_val = src_reg->s32_max_value;
8426 u32 umin_val = src_reg->u32_min_value;
8427 u32 umax_val = src_reg->u32_max_value;
8429 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
8430 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
8431 dst_reg->s32_min_value = S32_MIN;
8432 dst_reg->s32_max_value = S32_MAX;
8434 dst_reg->s32_min_value += smin_val;
8435 dst_reg->s32_max_value += smax_val;
8437 if (dst_reg->u32_min_value + umin_val < umin_val ||
8438 dst_reg->u32_max_value + umax_val < umax_val) {
8439 dst_reg->u32_min_value = 0;
8440 dst_reg->u32_max_value = U32_MAX;
8442 dst_reg->u32_min_value += umin_val;
8443 dst_reg->u32_max_value += umax_val;
8447 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
8448 struct bpf_reg_state *src_reg)
8450 s64 smin_val = src_reg->smin_value;
8451 s64 smax_val = src_reg->smax_value;
8452 u64 umin_val = src_reg->umin_value;
8453 u64 umax_val = src_reg->umax_value;
8455 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
8456 signed_add_overflows(dst_reg->smax_value, smax_val)) {
8457 dst_reg->smin_value = S64_MIN;
8458 dst_reg->smax_value = S64_MAX;
8460 dst_reg->smin_value += smin_val;
8461 dst_reg->smax_value += smax_val;
8463 if (dst_reg->umin_value + umin_val < umin_val ||
8464 dst_reg->umax_value + umax_val < umax_val) {
8465 dst_reg->umin_value = 0;
8466 dst_reg->umax_value = U64_MAX;
8468 dst_reg->umin_value += umin_val;
8469 dst_reg->umax_value += umax_val;
8473 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
8474 struct bpf_reg_state *src_reg)
8476 s32 smin_val = src_reg->s32_min_value;
8477 s32 smax_val = src_reg->s32_max_value;
8478 u32 umin_val = src_reg->u32_min_value;
8479 u32 umax_val = src_reg->u32_max_value;
8481 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
8482 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
8483 /* Overflow possible, we know nothing */
8484 dst_reg->s32_min_value = S32_MIN;
8485 dst_reg->s32_max_value = S32_MAX;
8487 dst_reg->s32_min_value -= smax_val;
8488 dst_reg->s32_max_value -= smin_val;
8490 if (dst_reg->u32_min_value < umax_val) {
8491 /* Overflow possible, we know nothing */
8492 dst_reg->u32_min_value = 0;
8493 dst_reg->u32_max_value = U32_MAX;
8495 /* Cannot overflow (as long as bounds are consistent) */
8496 dst_reg->u32_min_value -= umax_val;
8497 dst_reg->u32_max_value -= umin_val;
8501 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
8502 struct bpf_reg_state *src_reg)
8504 s64 smin_val = src_reg->smin_value;
8505 s64 smax_val = src_reg->smax_value;
8506 u64 umin_val = src_reg->umin_value;
8507 u64 umax_val = src_reg->umax_value;
8509 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
8510 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
8511 /* Overflow possible, we know nothing */
8512 dst_reg->smin_value = S64_MIN;
8513 dst_reg->smax_value = S64_MAX;
8515 dst_reg->smin_value -= smax_val;
8516 dst_reg->smax_value -= smin_val;
8518 if (dst_reg->umin_value < umax_val) {
8519 /* Overflow possible, we know nothing */
8520 dst_reg->umin_value = 0;
8521 dst_reg->umax_value = U64_MAX;
8523 /* Cannot overflow (as long as bounds are consistent) */
8524 dst_reg->umin_value -= umax_val;
8525 dst_reg->umax_value -= umin_val;
8529 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
8530 struct bpf_reg_state *src_reg)
8532 s32 smin_val = src_reg->s32_min_value;
8533 u32 umin_val = src_reg->u32_min_value;
8534 u32 umax_val = src_reg->u32_max_value;
8536 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
8537 /* Ain't nobody got time to multiply that sign */
8538 __mark_reg32_unbounded(dst_reg);
8541 /* Both values are positive, so we can work with unsigned and
8542 * copy the result to signed (unless it exceeds S32_MAX).
8544 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
8545 /* Potential overflow, we know nothing */
8546 __mark_reg32_unbounded(dst_reg);
8549 dst_reg->u32_min_value *= umin_val;
8550 dst_reg->u32_max_value *= umax_val;
8551 if (dst_reg->u32_max_value > S32_MAX) {
8552 /* Overflow possible, we know nothing */
8553 dst_reg->s32_min_value = S32_MIN;
8554 dst_reg->s32_max_value = S32_MAX;
8556 dst_reg->s32_min_value = dst_reg->u32_min_value;
8557 dst_reg->s32_max_value = dst_reg->u32_max_value;
8561 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
8562 struct bpf_reg_state *src_reg)
8564 s64 smin_val = src_reg->smin_value;
8565 u64 umin_val = src_reg->umin_value;
8566 u64 umax_val = src_reg->umax_value;
8568 if (smin_val < 0 || dst_reg->smin_value < 0) {
8569 /* Ain't nobody got time to multiply that sign */
8570 __mark_reg64_unbounded(dst_reg);
8573 /* Both values are positive, so we can work with unsigned and
8574 * copy the result to signed (unless it exceeds S64_MAX).
8576 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
8577 /* Potential overflow, we know nothing */
8578 __mark_reg64_unbounded(dst_reg);
8581 dst_reg->umin_value *= umin_val;
8582 dst_reg->umax_value *= umax_val;
8583 if (dst_reg->umax_value > S64_MAX) {
8584 /* Overflow possible, we know nothing */
8585 dst_reg->smin_value = S64_MIN;
8586 dst_reg->smax_value = S64_MAX;
8588 dst_reg->smin_value = dst_reg->umin_value;
8589 dst_reg->smax_value = dst_reg->umax_value;
8593 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
8594 struct bpf_reg_state *src_reg)
8596 bool src_known = tnum_subreg_is_const(src_reg->var_off);
8597 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
8598 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
8599 s32 smin_val = src_reg->s32_min_value;
8600 u32 umax_val = src_reg->u32_max_value;
8602 if (src_known && dst_known) {
8603 __mark_reg32_known(dst_reg, var32_off.value);
8607 /* We get our minimum from the var_off, since that's inherently
8608 * bitwise. Our maximum is the minimum of the operands' maxima.
8610 dst_reg->u32_min_value = var32_off.value;
8611 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
8612 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
8613 /* Lose signed bounds when ANDing negative numbers,
8614 * ain't nobody got time for that.
8616 dst_reg->s32_min_value = S32_MIN;
8617 dst_reg->s32_max_value = S32_MAX;
8619 /* ANDing two positives gives a positive, so safe to
8620 * cast result into s64.
8622 dst_reg->s32_min_value = dst_reg->u32_min_value;
8623 dst_reg->s32_max_value = dst_reg->u32_max_value;
8627 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
8628 struct bpf_reg_state *src_reg)
8630 bool src_known = tnum_is_const(src_reg->var_off);
8631 bool dst_known = tnum_is_const(dst_reg->var_off);
8632 s64 smin_val = src_reg->smin_value;
8633 u64 umax_val = src_reg->umax_value;
8635 if (src_known && dst_known) {
8636 __mark_reg_known(dst_reg, dst_reg->var_off.value);
8640 /* We get our minimum from the var_off, since that's inherently
8641 * bitwise. Our maximum is the minimum of the operands' maxima.
8643 dst_reg->umin_value = dst_reg->var_off.value;
8644 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
8645 if (dst_reg->smin_value < 0 || smin_val < 0) {
8646 /* Lose signed bounds when ANDing negative numbers,
8647 * ain't nobody got time for that.
8649 dst_reg->smin_value = S64_MIN;
8650 dst_reg->smax_value = S64_MAX;
8652 /* ANDing two positives gives a positive, so safe to
8653 * cast result into s64.
8655 dst_reg->smin_value = dst_reg->umin_value;
8656 dst_reg->smax_value = dst_reg->umax_value;
8658 /* We may learn something more from the var_off */
8659 __update_reg_bounds(dst_reg);
8662 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
8663 struct bpf_reg_state *src_reg)
8665 bool src_known = tnum_subreg_is_const(src_reg->var_off);
8666 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
8667 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
8668 s32 smin_val = src_reg->s32_min_value;
8669 u32 umin_val = src_reg->u32_min_value;
8671 if (src_known && dst_known) {
8672 __mark_reg32_known(dst_reg, var32_off.value);
8676 /* We get our maximum from the var_off, and our minimum is the
8677 * maximum of the operands' minima
8679 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
8680 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
8681 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
8682 /* Lose signed bounds when ORing negative numbers,
8683 * ain't nobody got time for that.
8685 dst_reg->s32_min_value = S32_MIN;
8686 dst_reg->s32_max_value = S32_MAX;
8688 /* ORing two positives gives a positive, so safe to
8689 * cast result into s64.
8691 dst_reg->s32_min_value = dst_reg->u32_min_value;
8692 dst_reg->s32_max_value = dst_reg->u32_max_value;
8696 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
8697 struct bpf_reg_state *src_reg)
8699 bool src_known = tnum_is_const(src_reg->var_off);
8700 bool dst_known = tnum_is_const(dst_reg->var_off);
8701 s64 smin_val = src_reg->smin_value;
8702 u64 umin_val = src_reg->umin_value;
8704 if (src_known && dst_known) {
8705 __mark_reg_known(dst_reg, dst_reg->var_off.value);
8709 /* We get our maximum from the var_off, and our minimum is the
8710 * maximum of the operands' minima
8712 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
8713 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
8714 if (dst_reg->smin_value < 0 || smin_val < 0) {
8715 /* Lose signed bounds when ORing negative numbers,
8716 * ain't nobody got time for that.
8718 dst_reg->smin_value = S64_MIN;
8719 dst_reg->smax_value = S64_MAX;
8721 /* ORing two positives gives a positive, so safe to
8722 * cast result into s64.
8724 dst_reg->smin_value = dst_reg->umin_value;
8725 dst_reg->smax_value = dst_reg->umax_value;
8727 /* We may learn something more from the var_off */
8728 __update_reg_bounds(dst_reg);
8731 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
8732 struct bpf_reg_state *src_reg)
8734 bool src_known = tnum_subreg_is_const(src_reg->var_off);
8735 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
8736 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
8737 s32 smin_val = src_reg->s32_min_value;
8739 if (src_known && dst_known) {
8740 __mark_reg32_known(dst_reg, var32_off.value);
8744 /* We get both minimum and maximum from the var32_off. */
8745 dst_reg->u32_min_value = var32_off.value;
8746 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
8748 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
8749 /* XORing two positive sign numbers gives a positive,
8750 * so safe to cast u32 result into s32.
8752 dst_reg->s32_min_value = dst_reg->u32_min_value;
8753 dst_reg->s32_max_value = dst_reg->u32_max_value;
8755 dst_reg->s32_min_value = S32_MIN;
8756 dst_reg->s32_max_value = S32_MAX;
8760 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
8761 struct bpf_reg_state *src_reg)
8763 bool src_known = tnum_is_const(src_reg->var_off);
8764 bool dst_known = tnum_is_const(dst_reg->var_off);
8765 s64 smin_val = src_reg->smin_value;
8767 if (src_known && dst_known) {
8768 /* dst_reg->var_off.value has been updated earlier */
8769 __mark_reg_known(dst_reg, dst_reg->var_off.value);
8773 /* We get both minimum and maximum from the var_off. */
8774 dst_reg->umin_value = dst_reg->var_off.value;
8775 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
8777 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
8778 /* XORing two positive sign numbers gives a positive,
8779 * so safe to cast u64 result into s64.
8781 dst_reg->smin_value = dst_reg->umin_value;
8782 dst_reg->smax_value = dst_reg->umax_value;
8784 dst_reg->smin_value = S64_MIN;
8785 dst_reg->smax_value = S64_MAX;
8788 __update_reg_bounds(dst_reg);
8791 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
8792 u64 umin_val, u64 umax_val)
8794 /* We lose all sign bit information (except what we can pick
8797 dst_reg->s32_min_value = S32_MIN;
8798 dst_reg->s32_max_value = S32_MAX;
8799 /* If we might shift our top bit out, then we know nothing */
8800 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
8801 dst_reg->u32_min_value = 0;
8802 dst_reg->u32_max_value = U32_MAX;
8804 dst_reg->u32_min_value <<= umin_val;
8805 dst_reg->u32_max_value <<= umax_val;
8809 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
8810 struct bpf_reg_state *src_reg)
8812 u32 umax_val = src_reg->u32_max_value;
8813 u32 umin_val = src_reg->u32_min_value;
8814 /* u32 alu operation will zext upper bits */
8815 struct tnum subreg = tnum_subreg(dst_reg->var_off);
8817 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
8818 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
8819 /* Not required but being careful mark reg64 bounds as unknown so
8820 * that we are forced to pick them up from tnum and zext later and
8821 * if some path skips this step we are still safe.
8823 __mark_reg64_unbounded(dst_reg);
8824 __update_reg32_bounds(dst_reg);
8827 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
8828 u64 umin_val, u64 umax_val)
8830 /* Special case <<32 because it is a common compiler pattern to sign
8831 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
8832 * positive we know this shift will also be positive so we can track
8833 * bounds correctly. Otherwise we lose all sign bit information except
8834 * what we can pick up from var_off. Perhaps we can generalize this
8835 * later to shifts of any length.
8837 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
8838 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
8840 dst_reg->smax_value = S64_MAX;
8842 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
8843 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
8845 dst_reg->smin_value = S64_MIN;
8847 /* If we might shift our top bit out, then we know nothing */
8848 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
8849 dst_reg->umin_value = 0;
8850 dst_reg->umax_value = U64_MAX;
8852 dst_reg->umin_value <<= umin_val;
8853 dst_reg->umax_value <<= umax_val;
8857 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
8858 struct bpf_reg_state *src_reg)
8860 u64 umax_val = src_reg->umax_value;
8861 u64 umin_val = src_reg->umin_value;
8863 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
8864 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
8865 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
8867 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
8868 /* We may learn something more from the var_off */
8869 __update_reg_bounds(dst_reg);
8872 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
8873 struct bpf_reg_state *src_reg)
8875 struct tnum subreg = tnum_subreg(dst_reg->var_off);
8876 u32 umax_val = src_reg->u32_max_value;
8877 u32 umin_val = src_reg->u32_min_value;
8879 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
8880 * be negative, then either:
8881 * 1) src_reg might be zero, so the sign bit of the result is
8882 * unknown, so we lose our signed bounds
8883 * 2) it's known negative, thus the unsigned bounds capture the
8885 * 3) the signed bounds cross zero, so they tell us nothing
8887 * If the value in dst_reg is known nonnegative, then again the
8888 * unsigned bounds capture the signed bounds.
8889 * Thus, in all cases it suffices to blow away our signed bounds
8890 * and rely on inferring new ones from the unsigned bounds and
8891 * var_off of the result.
8893 dst_reg->s32_min_value = S32_MIN;
8894 dst_reg->s32_max_value = S32_MAX;
8896 dst_reg->var_off = tnum_rshift(subreg, umin_val);
8897 dst_reg->u32_min_value >>= umax_val;
8898 dst_reg->u32_max_value >>= umin_val;
8900 __mark_reg64_unbounded(dst_reg);
8901 __update_reg32_bounds(dst_reg);
8904 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
8905 struct bpf_reg_state *src_reg)
8907 u64 umax_val = src_reg->umax_value;
8908 u64 umin_val = src_reg->umin_value;
8910 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
8911 * be negative, then either:
8912 * 1) src_reg might be zero, so the sign bit of the result is
8913 * unknown, so we lose our signed bounds
8914 * 2) it's known negative, thus the unsigned bounds capture the
8916 * 3) the signed bounds cross zero, so they tell us nothing
8918 * If the value in dst_reg is known nonnegative, then again the
8919 * unsigned bounds capture the signed bounds.
8920 * Thus, in all cases it suffices to blow away our signed bounds
8921 * and rely on inferring new ones from the unsigned bounds and
8922 * var_off of the result.
8924 dst_reg->smin_value = S64_MIN;
8925 dst_reg->smax_value = S64_MAX;
8926 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
8927 dst_reg->umin_value >>= umax_val;
8928 dst_reg->umax_value >>= umin_val;
8930 /* Its not easy to operate on alu32 bounds here because it depends
8931 * on bits being shifted in. Take easy way out and mark unbounded
8932 * so we can recalculate later from tnum.
8934 __mark_reg32_unbounded(dst_reg);
8935 __update_reg_bounds(dst_reg);
8938 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
8939 struct bpf_reg_state *src_reg)
8941 u64 umin_val = src_reg->u32_min_value;
8943 /* Upon reaching here, src_known is true and
8944 * umax_val is equal to umin_val.
8946 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
8947 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
8949 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
8951 /* blow away the dst_reg umin_value/umax_value and rely on
8952 * dst_reg var_off to refine the result.
8954 dst_reg->u32_min_value = 0;
8955 dst_reg->u32_max_value = U32_MAX;
8957 __mark_reg64_unbounded(dst_reg);
8958 __update_reg32_bounds(dst_reg);
8961 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
8962 struct bpf_reg_state *src_reg)
8964 u64 umin_val = src_reg->umin_value;
8966 /* Upon reaching here, src_known is true and umax_val is equal
8969 dst_reg->smin_value >>= umin_val;
8970 dst_reg->smax_value >>= umin_val;
8972 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
8974 /* blow away the dst_reg umin_value/umax_value and rely on
8975 * dst_reg var_off to refine the result.
8977 dst_reg->umin_value = 0;
8978 dst_reg->umax_value = U64_MAX;
8980 /* Its not easy to operate on alu32 bounds here because it depends
8981 * on bits being shifted in from upper 32-bits. Take easy way out
8982 * and mark unbounded so we can recalculate later from tnum.
8984 __mark_reg32_unbounded(dst_reg);
8985 __update_reg_bounds(dst_reg);
8988 /* WARNING: This function does calculations on 64-bit values, but the actual
8989 * execution may occur on 32-bit values. Therefore, things like bitshifts
8990 * need extra checks in the 32-bit case.
8992 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
8993 struct bpf_insn *insn,
8994 struct bpf_reg_state *dst_reg,
8995 struct bpf_reg_state src_reg)
8997 struct bpf_reg_state *regs = cur_regs(env);
8998 u8 opcode = BPF_OP(insn->code);
9000 s64 smin_val, smax_val;
9001 u64 umin_val, umax_val;
9002 s32 s32_min_val, s32_max_val;
9003 u32 u32_min_val, u32_max_val;
9004 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
9005 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
9008 smin_val = src_reg.smin_value;
9009 smax_val = src_reg.smax_value;
9010 umin_val = src_reg.umin_value;
9011 umax_val = src_reg.umax_value;
9013 s32_min_val = src_reg.s32_min_value;
9014 s32_max_val = src_reg.s32_max_value;
9015 u32_min_val = src_reg.u32_min_value;
9016 u32_max_val = src_reg.u32_max_value;
9019 src_known = tnum_subreg_is_const(src_reg.var_off);
9021 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
9022 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
9023 /* Taint dst register if offset had invalid bounds
9024 * derived from e.g. dead branches.
9026 __mark_reg_unknown(env, dst_reg);
9030 src_known = tnum_is_const(src_reg.var_off);
9032 (smin_val != smax_val || umin_val != umax_val)) ||
9033 smin_val > smax_val || umin_val > umax_val) {
9034 /* Taint dst register if offset had invalid bounds
9035 * derived from e.g. dead branches.
9037 __mark_reg_unknown(env, dst_reg);
9043 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
9044 __mark_reg_unknown(env, dst_reg);
9048 if (sanitize_needed(opcode)) {
9049 ret = sanitize_val_alu(env, insn);
9051 return sanitize_err(env, insn, ret, NULL, NULL);
9054 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
9055 * There are two classes of instructions: The first class we track both
9056 * alu32 and alu64 sign/unsigned bounds independently this provides the
9057 * greatest amount of precision when alu operations are mixed with jmp32
9058 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
9059 * and BPF_OR. This is possible because these ops have fairly easy to
9060 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
9061 * See alu32 verifier tests for examples. The second class of
9062 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
9063 * with regards to tracking sign/unsigned bounds because the bits may
9064 * cross subreg boundaries in the alu64 case. When this happens we mark
9065 * the reg unbounded in the subreg bound space and use the resulting
9066 * tnum to calculate an approximation of the sign/unsigned bounds.
9070 scalar32_min_max_add(dst_reg, &src_reg);
9071 scalar_min_max_add(dst_reg, &src_reg);
9072 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
9075 scalar32_min_max_sub(dst_reg, &src_reg);
9076 scalar_min_max_sub(dst_reg, &src_reg);
9077 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
9080 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
9081 scalar32_min_max_mul(dst_reg, &src_reg);
9082 scalar_min_max_mul(dst_reg, &src_reg);
9085 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
9086 scalar32_min_max_and(dst_reg, &src_reg);
9087 scalar_min_max_and(dst_reg, &src_reg);
9090 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
9091 scalar32_min_max_or(dst_reg, &src_reg);
9092 scalar_min_max_or(dst_reg, &src_reg);
9095 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
9096 scalar32_min_max_xor(dst_reg, &src_reg);
9097 scalar_min_max_xor(dst_reg, &src_reg);
9100 if (umax_val >= insn_bitness) {
9101 /* Shifts greater than 31 or 63 are undefined.
9102 * This includes shifts by a negative number.
9104 mark_reg_unknown(env, regs, insn->dst_reg);
9108 scalar32_min_max_lsh(dst_reg, &src_reg);
9110 scalar_min_max_lsh(dst_reg, &src_reg);
9113 if (umax_val >= insn_bitness) {
9114 /* Shifts greater than 31 or 63 are undefined.
9115 * This includes shifts by a negative number.
9117 mark_reg_unknown(env, regs, insn->dst_reg);
9121 scalar32_min_max_rsh(dst_reg, &src_reg);
9123 scalar_min_max_rsh(dst_reg, &src_reg);
9126 if (umax_val >= insn_bitness) {
9127 /* Shifts greater than 31 or 63 are undefined.
9128 * This includes shifts by a negative number.
9130 mark_reg_unknown(env, regs, insn->dst_reg);
9134 scalar32_min_max_arsh(dst_reg, &src_reg);
9136 scalar_min_max_arsh(dst_reg, &src_reg);
9139 mark_reg_unknown(env, regs, insn->dst_reg);
9143 /* ALU32 ops are zero extended into 64bit register */
9145 zext_32_to_64(dst_reg);
9146 reg_bounds_sync(dst_reg);
9150 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
9153 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
9154 struct bpf_insn *insn)
9156 struct bpf_verifier_state *vstate = env->cur_state;
9157 struct bpf_func_state *state = vstate->frame[vstate->curframe];
9158 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
9159 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
9160 u8 opcode = BPF_OP(insn->code);
9163 dst_reg = ®s[insn->dst_reg];
9165 if (dst_reg->type != SCALAR_VALUE)
9168 /* Make sure ID is cleared otherwise dst_reg min/max could be
9169 * incorrectly propagated into other registers by find_equal_scalars()
9172 if (BPF_SRC(insn->code) == BPF_X) {
9173 src_reg = ®s[insn->src_reg];
9174 if (src_reg->type != SCALAR_VALUE) {
9175 if (dst_reg->type != SCALAR_VALUE) {
9176 /* Combining two pointers by any ALU op yields
9177 * an arbitrary scalar. Disallow all math except
9178 * pointer subtraction
9180 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
9181 mark_reg_unknown(env, regs, insn->dst_reg);
9184 verbose(env, "R%d pointer %s pointer prohibited\n",
9186 bpf_alu_string[opcode >> 4]);
9189 /* scalar += pointer
9190 * This is legal, but we have to reverse our
9191 * src/dest handling in computing the range
9193 err = mark_chain_precision(env, insn->dst_reg);
9196 return adjust_ptr_min_max_vals(env, insn,
9199 } else if (ptr_reg) {
9200 /* pointer += scalar */
9201 err = mark_chain_precision(env, insn->src_reg);
9204 return adjust_ptr_min_max_vals(env, insn,
9208 /* Pretend the src is a reg with a known value, since we only
9209 * need to be able to read from this state.
9211 off_reg.type = SCALAR_VALUE;
9212 __mark_reg_known(&off_reg, insn->imm);
9214 if (ptr_reg) /* pointer += K */
9215 return adjust_ptr_min_max_vals(env, insn,
9219 /* Got here implies adding two SCALAR_VALUEs */
9220 if (WARN_ON_ONCE(ptr_reg)) {
9221 print_verifier_state(env, state, true);
9222 verbose(env, "verifier internal error: unexpected ptr_reg\n");
9225 if (WARN_ON(!src_reg)) {
9226 print_verifier_state(env, state, true);
9227 verbose(env, "verifier internal error: no src_reg\n");
9230 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
9233 /* check validity of 32-bit and 64-bit arithmetic operations */
9234 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
9236 struct bpf_reg_state *regs = cur_regs(env);
9237 u8 opcode = BPF_OP(insn->code);
9240 if (opcode == BPF_END || opcode == BPF_NEG) {
9241 if (opcode == BPF_NEG) {
9242 if (BPF_SRC(insn->code) != BPF_K ||
9243 insn->src_reg != BPF_REG_0 ||
9244 insn->off != 0 || insn->imm != 0) {
9245 verbose(env, "BPF_NEG uses reserved fields\n");
9249 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
9250 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
9251 BPF_CLASS(insn->code) == BPF_ALU64) {
9252 verbose(env, "BPF_END uses reserved fields\n");
9257 /* check src operand */
9258 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
9262 if (is_pointer_value(env, insn->dst_reg)) {
9263 verbose(env, "R%d pointer arithmetic prohibited\n",
9268 /* check dest operand */
9269 err = check_reg_arg(env, insn->dst_reg, DST_OP);
9273 } else if (opcode == BPF_MOV) {
9275 if (BPF_SRC(insn->code) == BPF_X) {
9276 if (insn->imm != 0 || insn->off != 0) {
9277 verbose(env, "BPF_MOV uses reserved fields\n");
9281 /* check src operand */
9282 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9286 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
9287 verbose(env, "BPF_MOV uses reserved fields\n");
9292 /* check dest operand, mark as required later */
9293 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
9297 if (BPF_SRC(insn->code) == BPF_X) {
9298 struct bpf_reg_state *src_reg = regs + insn->src_reg;
9299 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
9301 if (BPF_CLASS(insn->code) == BPF_ALU64) {
9303 * copy register state to dest reg
9305 if (src_reg->type == SCALAR_VALUE && !src_reg->id)
9306 /* Assign src and dst registers the same ID
9307 * that will be used by find_equal_scalars()
9308 * to propagate min/max range.
9310 src_reg->id = ++env->id_gen;
9311 *dst_reg = *src_reg;
9312 dst_reg->live |= REG_LIVE_WRITTEN;
9313 dst_reg->subreg_def = DEF_NOT_SUBREG;
9316 if (is_pointer_value(env, insn->src_reg)) {
9318 "R%d partial copy of pointer\n",
9321 } else if (src_reg->type == SCALAR_VALUE) {
9322 *dst_reg = *src_reg;
9323 /* Make sure ID is cleared otherwise
9324 * dst_reg min/max could be incorrectly
9325 * propagated into src_reg by find_equal_scalars()
9328 dst_reg->live |= REG_LIVE_WRITTEN;
9329 dst_reg->subreg_def = env->insn_idx + 1;
9331 mark_reg_unknown(env, regs,
9334 zext_32_to_64(dst_reg);
9335 reg_bounds_sync(dst_reg);
9339 * remember the value we stored into this reg
9341 /* clear any state __mark_reg_known doesn't set */
9342 mark_reg_unknown(env, regs, insn->dst_reg);
9343 regs[insn->dst_reg].type = SCALAR_VALUE;
9344 if (BPF_CLASS(insn->code) == BPF_ALU64) {
9345 __mark_reg_known(regs + insn->dst_reg,
9348 __mark_reg_known(regs + insn->dst_reg,
9353 } else if (opcode > BPF_END) {
9354 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
9357 } else { /* all other ALU ops: and, sub, xor, add, ... */
9359 if (BPF_SRC(insn->code) == BPF_X) {
9360 if (insn->imm != 0 || insn->off != 0) {
9361 verbose(env, "BPF_ALU uses reserved fields\n");
9364 /* check src1 operand */
9365 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9369 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
9370 verbose(env, "BPF_ALU uses reserved fields\n");
9375 /* check src2 operand */
9376 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
9380 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
9381 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
9382 verbose(env, "div by zero\n");
9386 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
9387 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
9388 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
9390 if (insn->imm < 0 || insn->imm >= size) {
9391 verbose(env, "invalid shift %d\n", insn->imm);
9396 /* check dest operand */
9397 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
9401 return adjust_reg_min_max_vals(env, insn);
9407 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
9408 struct bpf_reg_state *dst_reg,
9409 enum bpf_reg_type type,
9410 bool range_right_open)
9412 struct bpf_func_state *state;
9413 struct bpf_reg_state *reg;
9416 if (dst_reg->off < 0 ||
9417 (dst_reg->off == 0 && range_right_open))
9418 /* This doesn't give us any range */
9421 if (dst_reg->umax_value > MAX_PACKET_OFF ||
9422 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
9423 /* Risk of overflow. For instance, ptr + (1<<63) may be less
9424 * than pkt_end, but that's because it's also less than pkt.
9428 new_range = dst_reg->off;
9429 if (range_right_open)
9432 /* Examples for register markings:
9434 * pkt_data in dst register:
9438 * if (r2 > pkt_end) goto <handle exception>
9443 * if (r2 < pkt_end) goto <access okay>
9444 * <handle exception>
9447 * r2 == dst_reg, pkt_end == src_reg
9448 * r2=pkt(id=n,off=8,r=0)
9449 * r3=pkt(id=n,off=0,r=0)
9451 * pkt_data in src register:
9455 * if (pkt_end >= r2) goto <access okay>
9456 * <handle exception>
9460 * if (pkt_end <= r2) goto <handle exception>
9464 * pkt_end == dst_reg, r2 == src_reg
9465 * r2=pkt(id=n,off=8,r=0)
9466 * r3=pkt(id=n,off=0,r=0)
9468 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
9469 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
9470 * and [r3, r3 + 8-1) respectively is safe to access depending on
9474 /* If our ids match, then we must have the same max_value. And we
9475 * don't care about the other reg's fixed offset, since if it's too big
9476 * the range won't allow anything.
9477 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
9479 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
9480 if (reg->type == type && reg->id == dst_reg->id)
9481 /* keep the maximum range already checked */
9482 reg->range = max(reg->range, new_range);
9486 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
9488 struct tnum subreg = tnum_subreg(reg->var_off);
9489 s32 sval = (s32)val;
9493 if (tnum_is_const(subreg))
9494 return !!tnum_equals_const(subreg, val);
9497 if (tnum_is_const(subreg))
9498 return !tnum_equals_const(subreg, val);
9501 if ((~subreg.mask & subreg.value) & val)
9503 if (!((subreg.mask | subreg.value) & val))
9507 if (reg->u32_min_value > val)
9509 else if (reg->u32_max_value <= val)
9513 if (reg->s32_min_value > sval)
9515 else if (reg->s32_max_value <= sval)
9519 if (reg->u32_max_value < val)
9521 else if (reg->u32_min_value >= val)
9525 if (reg->s32_max_value < sval)
9527 else if (reg->s32_min_value >= sval)
9531 if (reg->u32_min_value >= val)
9533 else if (reg->u32_max_value < val)
9537 if (reg->s32_min_value >= sval)
9539 else if (reg->s32_max_value < sval)
9543 if (reg->u32_max_value <= val)
9545 else if (reg->u32_min_value > val)
9549 if (reg->s32_max_value <= sval)
9551 else if (reg->s32_min_value > sval)
9560 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
9562 s64 sval = (s64)val;
9566 if (tnum_is_const(reg->var_off))
9567 return !!tnum_equals_const(reg->var_off, val);
9570 if (tnum_is_const(reg->var_off))
9571 return !tnum_equals_const(reg->var_off, val);
9574 if ((~reg->var_off.mask & reg->var_off.value) & val)
9576 if (!((reg->var_off.mask | reg->var_off.value) & val))
9580 if (reg->umin_value > val)
9582 else if (reg->umax_value <= val)
9586 if (reg->smin_value > sval)
9588 else if (reg->smax_value <= sval)
9592 if (reg->umax_value < val)
9594 else if (reg->umin_value >= val)
9598 if (reg->smax_value < sval)
9600 else if (reg->smin_value >= sval)
9604 if (reg->umin_value >= val)
9606 else if (reg->umax_value < val)
9610 if (reg->smin_value >= sval)
9612 else if (reg->smax_value < sval)
9616 if (reg->umax_value <= val)
9618 else if (reg->umin_value > val)
9622 if (reg->smax_value <= sval)
9624 else if (reg->smin_value > sval)
9632 /* compute branch direction of the expression "if (reg opcode val) goto target;"
9634 * 1 - branch will be taken and "goto target" will be executed
9635 * 0 - branch will not be taken and fall-through to next insn
9636 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
9639 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
9642 if (__is_pointer_value(false, reg)) {
9643 if (!reg_type_not_null(reg->type))
9646 /* If pointer is valid tests against zero will fail so we can
9647 * use this to direct branch taken.
9663 return is_branch32_taken(reg, val, opcode);
9664 return is_branch64_taken(reg, val, opcode);
9667 static int flip_opcode(u32 opcode)
9669 /* How can we transform "a <op> b" into "b <op> a"? */
9670 static const u8 opcode_flip[16] = {
9671 /* these stay the same */
9672 [BPF_JEQ >> 4] = BPF_JEQ,
9673 [BPF_JNE >> 4] = BPF_JNE,
9674 [BPF_JSET >> 4] = BPF_JSET,
9675 /* these swap "lesser" and "greater" (L and G in the opcodes) */
9676 [BPF_JGE >> 4] = BPF_JLE,
9677 [BPF_JGT >> 4] = BPF_JLT,
9678 [BPF_JLE >> 4] = BPF_JGE,
9679 [BPF_JLT >> 4] = BPF_JGT,
9680 [BPF_JSGE >> 4] = BPF_JSLE,
9681 [BPF_JSGT >> 4] = BPF_JSLT,
9682 [BPF_JSLE >> 4] = BPF_JSGE,
9683 [BPF_JSLT >> 4] = BPF_JSGT
9685 return opcode_flip[opcode >> 4];
9688 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
9689 struct bpf_reg_state *src_reg,
9692 struct bpf_reg_state *pkt;
9694 if (src_reg->type == PTR_TO_PACKET_END) {
9696 } else if (dst_reg->type == PTR_TO_PACKET_END) {
9698 opcode = flip_opcode(opcode);
9703 if (pkt->range >= 0)
9708 /* pkt <= pkt_end */
9712 if (pkt->range == BEYOND_PKT_END)
9713 /* pkt has at last one extra byte beyond pkt_end */
9714 return opcode == BPF_JGT;
9720 /* pkt >= pkt_end */
9721 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
9722 return opcode == BPF_JGE;
9728 /* Adjusts the register min/max values in the case that the dst_reg is the
9729 * variable register that we are working on, and src_reg is a constant or we're
9730 * simply doing a BPF_K check.
9731 * In JEQ/JNE cases we also adjust the var_off values.
9733 static void reg_set_min_max(struct bpf_reg_state *true_reg,
9734 struct bpf_reg_state *false_reg,
9736 u8 opcode, bool is_jmp32)
9738 struct tnum false_32off = tnum_subreg(false_reg->var_off);
9739 struct tnum false_64off = false_reg->var_off;
9740 struct tnum true_32off = tnum_subreg(true_reg->var_off);
9741 struct tnum true_64off = true_reg->var_off;
9742 s64 sval = (s64)val;
9743 s32 sval32 = (s32)val32;
9745 /* If the dst_reg is a pointer, we can't learn anything about its
9746 * variable offset from the compare (unless src_reg were a pointer into
9747 * the same object, but we don't bother with that.
9748 * Since false_reg and true_reg have the same type by construction, we
9749 * only need to check one of them for pointerness.
9751 if (__is_pointer_value(false, false_reg))
9755 /* JEQ/JNE comparison doesn't change the register equivalence.
9758 * if (r1 == 42) goto label;
9760 * label: // here both r1 and r2 are known to be 42.
9762 * Hence when marking register as known preserve it's ID.
9766 __mark_reg32_known(true_reg, val32);
9767 true_32off = tnum_subreg(true_reg->var_off);
9769 ___mark_reg_known(true_reg, val);
9770 true_64off = true_reg->var_off;
9775 __mark_reg32_known(false_reg, val32);
9776 false_32off = tnum_subreg(false_reg->var_off);
9778 ___mark_reg_known(false_reg, val);
9779 false_64off = false_reg->var_off;
9784 false_32off = tnum_and(false_32off, tnum_const(~val32));
9785 if (is_power_of_2(val32))
9786 true_32off = tnum_or(true_32off,
9789 false_64off = tnum_and(false_64off, tnum_const(~val));
9790 if (is_power_of_2(val))
9791 true_64off = tnum_or(true_64off,
9799 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
9800 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
9802 false_reg->u32_max_value = min(false_reg->u32_max_value,
9804 true_reg->u32_min_value = max(true_reg->u32_min_value,
9807 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
9808 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
9810 false_reg->umax_value = min(false_reg->umax_value, false_umax);
9811 true_reg->umin_value = max(true_reg->umin_value, true_umin);
9819 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
9820 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
9822 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
9823 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
9825 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
9826 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
9828 false_reg->smax_value = min(false_reg->smax_value, false_smax);
9829 true_reg->smin_value = max(true_reg->smin_value, true_smin);
9837 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
9838 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
9840 false_reg->u32_min_value = max(false_reg->u32_min_value,
9842 true_reg->u32_max_value = min(true_reg->u32_max_value,
9845 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
9846 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
9848 false_reg->umin_value = max(false_reg->umin_value, false_umin);
9849 true_reg->umax_value = min(true_reg->umax_value, true_umax);
9857 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
9858 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
9860 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
9861 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
9863 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
9864 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
9866 false_reg->smin_value = max(false_reg->smin_value, false_smin);
9867 true_reg->smax_value = min(true_reg->smax_value, true_smax);
9876 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
9877 tnum_subreg(false_32off));
9878 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
9879 tnum_subreg(true_32off));
9880 __reg_combine_32_into_64(false_reg);
9881 __reg_combine_32_into_64(true_reg);
9883 false_reg->var_off = false_64off;
9884 true_reg->var_off = true_64off;
9885 __reg_combine_64_into_32(false_reg);
9886 __reg_combine_64_into_32(true_reg);
9890 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
9893 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
9894 struct bpf_reg_state *false_reg,
9896 u8 opcode, bool is_jmp32)
9898 opcode = flip_opcode(opcode);
9899 /* This uses zero as "not present in table"; luckily the zero opcode,
9900 * BPF_JA, can't get here.
9903 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
9906 /* Regs are known to be equal, so intersect their min/max/var_off */
9907 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
9908 struct bpf_reg_state *dst_reg)
9910 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
9911 dst_reg->umin_value);
9912 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
9913 dst_reg->umax_value);
9914 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
9915 dst_reg->smin_value);
9916 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
9917 dst_reg->smax_value);
9918 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
9920 reg_bounds_sync(src_reg);
9921 reg_bounds_sync(dst_reg);
9924 static void reg_combine_min_max(struct bpf_reg_state *true_src,
9925 struct bpf_reg_state *true_dst,
9926 struct bpf_reg_state *false_src,
9927 struct bpf_reg_state *false_dst,
9932 __reg_combine_min_max(true_src, true_dst);
9935 __reg_combine_min_max(false_src, false_dst);
9940 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
9941 struct bpf_reg_state *reg, u32 id,
9944 if (type_may_be_null(reg->type) && reg->id == id &&
9945 !WARN_ON_ONCE(!reg->id)) {
9946 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
9947 !tnum_equals_const(reg->var_off, 0) ||
9949 /* Old offset (both fixed and variable parts) should
9950 * have been known-zero, because we don't allow pointer
9951 * arithmetic on pointers that might be NULL. If we
9952 * see this happening, don't convert the register.
9957 reg->type = SCALAR_VALUE;
9958 /* We don't need id and ref_obj_id from this point
9959 * onwards anymore, thus we should better reset it,
9960 * so that state pruning has chances to take effect.
9963 reg->ref_obj_id = 0;
9968 mark_ptr_not_null_reg(reg);
9970 if (!reg_may_point_to_spin_lock(reg)) {
9971 /* For not-NULL ptr, reg->ref_obj_id will be reset
9972 * in release_reference().
9974 * reg->id is still used by spin_lock ptr. Other
9975 * than spin_lock ptr type, reg->id can be reset.
9982 /* The logic is similar to find_good_pkt_pointers(), both could eventually
9983 * be folded together at some point.
9985 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
9988 struct bpf_func_state *state = vstate->frame[vstate->curframe];
9989 struct bpf_reg_state *regs = state->regs, *reg;
9990 u32 ref_obj_id = regs[regno].ref_obj_id;
9991 u32 id = regs[regno].id;
9993 if (ref_obj_id && ref_obj_id == id && is_null)
9994 /* regs[regno] is in the " == NULL" branch.
9995 * No one could have freed the reference state before
9996 * doing the NULL check.
9998 WARN_ON_ONCE(release_reference_state(state, id));
10000 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
10001 mark_ptr_or_null_reg(state, reg, id, is_null);
10005 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
10006 struct bpf_reg_state *dst_reg,
10007 struct bpf_reg_state *src_reg,
10008 struct bpf_verifier_state *this_branch,
10009 struct bpf_verifier_state *other_branch)
10011 if (BPF_SRC(insn->code) != BPF_X)
10014 /* Pointers are always 64-bit. */
10015 if (BPF_CLASS(insn->code) == BPF_JMP32)
10018 switch (BPF_OP(insn->code)) {
10020 if ((dst_reg->type == PTR_TO_PACKET &&
10021 src_reg->type == PTR_TO_PACKET_END) ||
10022 (dst_reg->type == PTR_TO_PACKET_META &&
10023 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
10024 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
10025 find_good_pkt_pointers(this_branch, dst_reg,
10026 dst_reg->type, false);
10027 mark_pkt_end(other_branch, insn->dst_reg, true);
10028 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
10029 src_reg->type == PTR_TO_PACKET) ||
10030 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
10031 src_reg->type == PTR_TO_PACKET_META)) {
10032 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
10033 find_good_pkt_pointers(other_branch, src_reg,
10034 src_reg->type, true);
10035 mark_pkt_end(this_branch, insn->src_reg, false);
10041 if ((dst_reg->type == PTR_TO_PACKET &&
10042 src_reg->type == PTR_TO_PACKET_END) ||
10043 (dst_reg->type == PTR_TO_PACKET_META &&
10044 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
10045 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
10046 find_good_pkt_pointers(other_branch, dst_reg,
10047 dst_reg->type, true);
10048 mark_pkt_end(this_branch, insn->dst_reg, false);
10049 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
10050 src_reg->type == PTR_TO_PACKET) ||
10051 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
10052 src_reg->type == PTR_TO_PACKET_META)) {
10053 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
10054 find_good_pkt_pointers(this_branch, src_reg,
10055 src_reg->type, false);
10056 mark_pkt_end(other_branch, insn->src_reg, true);
10062 if ((dst_reg->type == PTR_TO_PACKET &&
10063 src_reg->type == PTR_TO_PACKET_END) ||
10064 (dst_reg->type == PTR_TO_PACKET_META &&
10065 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
10066 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
10067 find_good_pkt_pointers(this_branch, dst_reg,
10068 dst_reg->type, true);
10069 mark_pkt_end(other_branch, insn->dst_reg, false);
10070 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
10071 src_reg->type == PTR_TO_PACKET) ||
10072 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
10073 src_reg->type == PTR_TO_PACKET_META)) {
10074 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
10075 find_good_pkt_pointers(other_branch, src_reg,
10076 src_reg->type, false);
10077 mark_pkt_end(this_branch, insn->src_reg, true);
10083 if ((dst_reg->type == PTR_TO_PACKET &&
10084 src_reg->type == PTR_TO_PACKET_END) ||
10085 (dst_reg->type == PTR_TO_PACKET_META &&
10086 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
10087 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
10088 find_good_pkt_pointers(other_branch, dst_reg,
10089 dst_reg->type, false);
10090 mark_pkt_end(this_branch, insn->dst_reg, true);
10091 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
10092 src_reg->type == PTR_TO_PACKET) ||
10093 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
10094 src_reg->type == PTR_TO_PACKET_META)) {
10095 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
10096 find_good_pkt_pointers(this_branch, src_reg,
10097 src_reg->type, true);
10098 mark_pkt_end(other_branch, insn->src_reg, false);
10110 static void find_equal_scalars(struct bpf_verifier_state *vstate,
10111 struct bpf_reg_state *known_reg)
10113 struct bpf_func_state *state;
10114 struct bpf_reg_state *reg;
10116 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
10117 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
10122 static int check_cond_jmp_op(struct bpf_verifier_env *env,
10123 struct bpf_insn *insn, int *insn_idx)
10125 struct bpf_verifier_state *this_branch = env->cur_state;
10126 struct bpf_verifier_state *other_branch;
10127 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
10128 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
10129 u8 opcode = BPF_OP(insn->code);
10134 /* Only conditional jumps are expected to reach here. */
10135 if (opcode == BPF_JA || opcode > BPF_JSLE) {
10136 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
10140 if (BPF_SRC(insn->code) == BPF_X) {
10141 if (insn->imm != 0) {
10142 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
10146 /* check src1 operand */
10147 err = check_reg_arg(env, insn->src_reg, SRC_OP);
10151 if (is_pointer_value(env, insn->src_reg)) {
10152 verbose(env, "R%d pointer comparison prohibited\n",
10156 src_reg = ®s[insn->src_reg];
10158 if (insn->src_reg != BPF_REG_0) {
10159 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
10164 /* check src2 operand */
10165 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
10169 dst_reg = ®s[insn->dst_reg];
10170 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
10172 if (BPF_SRC(insn->code) == BPF_K) {
10173 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
10174 } else if (src_reg->type == SCALAR_VALUE &&
10175 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
10176 pred = is_branch_taken(dst_reg,
10177 tnum_subreg(src_reg->var_off).value,
10180 } else if (src_reg->type == SCALAR_VALUE &&
10181 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
10182 pred = is_branch_taken(dst_reg,
10183 src_reg->var_off.value,
10186 } else if (reg_is_pkt_pointer_any(dst_reg) &&
10187 reg_is_pkt_pointer_any(src_reg) &&
10189 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
10193 /* If we get here with a dst_reg pointer type it is because
10194 * above is_branch_taken() special cased the 0 comparison.
10196 if (!__is_pointer_value(false, dst_reg))
10197 err = mark_chain_precision(env, insn->dst_reg);
10198 if (BPF_SRC(insn->code) == BPF_X && !err &&
10199 !__is_pointer_value(false, src_reg))
10200 err = mark_chain_precision(env, insn->src_reg);
10206 /* Only follow the goto, ignore fall-through. If needed, push
10207 * the fall-through branch for simulation under speculative
10210 if (!env->bypass_spec_v1 &&
10211 !sanitize_speculative_path(env, insn, *insn_idx + 1,
10214 *insn_idx += insn->off;
10216 } else if (pred == 0) {
10217 /* Only follow the fall-through branch, since that's where the
10218 * program will go. If needed, push the goto branch for
10219 * simulation under speculative execution.
10221 if (!env->bypass_spec_v1 &&
10222 !sanitize_speculative_path(env, insn,
10223 *insn_idx + insn->off + 1,
10229 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
10233 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
10235 /* detect if we are comparing against a constant value so we can adjust
10236 * our min/max values for our dst register.
10237 * this is only legit if both are scalars (or pointers to the same
10238 * object, I suppose, but we don't support that right now), because
10239 * otherwise the different base pointers mean the offsets aren't
10242 if (BPF_SRC(insn->code) == BPF_X) {
10243 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
10245 if (dst_reg->type == SCALAR_VALUE &&
10246 src_reg->type == SCALAR_VALUE) {
10247 if (tnum_is_const(src_reg->var_off) ||
10249 tnum_is_const(tnum_subreg(src_reg->var_off))))
10250 reg_set_min_max(&other_branch_regs[insn->dst_reg],
10252 src_reg->var_off.value,
10253 tnum_subreg(src_reg->var_off).value,
10255 else if (tnum_is_const(dst_reg->var_off) ||
10257 tnum_is_const(tnum_subreg(dst_reg->var_off))))
10258 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
10260 dst_reg->var_off.value,
10261 tnum_subreg(dst_reg->var_off).value,
10263 else if (!is_jmp32 &&
10264 (opcode == BPF_JEQ || opcode == BPF_JNE))
10265 /* Comparing for equality, we can combine knowledge */
10266 reg_combine_min_max(&other_branch_regs[insn->src_reg],
10267 &other_branch_regs[insn->dst_reg],
10268 src_reg, dst_reg, opcode);
10270 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
10271 find_equal_scalars(this_branch, src_reg);
10272 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
10276 } else if (dst_reg->type == SCALAR_VALUE) {
10277 reg_set_min_max(&other_branch_regs[insn->dst_reg],
10278 dst_reg, insn->imm, (u32)insn->imm,
10282 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
10283 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
10284 find_equal_scalars(this_branch, dst_reg);
10285 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
10288 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
10289 * NOTE: these optimizations below are related with pointer comparison
10290 * which will never be JMP32.
10292 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
10293 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
10294 type_may_be_null(dst_reg->type)) {
10295 /* Mark all identical registers in each branch as either
10296 * safe or unknown depending R == 0 or R != 0 conditional.
10298 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
10299 opcode == BPF_JNE);
10300 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
10301 opcode == BPF_JEQ);
10302 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
10303 this_branch, other_branch) &&
10304 is_pointer_value(env, insn->dst_reg)) {
10305 verbose(env, "R%d pointer comparison prohibited\n",
10309 if (env->log.level & BPF_LOG_LEVEL)
10310 print_insn_state(env, this_branch->frame[this_branch->curframe]);
10314 /* verify BPF_LD_IMM64 instruction */
10315 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
10317 struct bpf_insn_aux_data *aux = cur_aux(env);
10318 struct bpf_reg_state *regs = cur_regs(env);
10319 struct bpf_reg_state *dst_reg;
10320 struct bpf_map *map;
10323 if (BPF_SIZE(insn->code) != BPF_DW) {
10324 verbose(env, "invalid BPF_LD_IMM insn\n");
10327 if (insn->off != 0) {
10328 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
10332 err = check_reg_arg(env, insn->dst_reg, DST_OP);
10336 dst_reg = ®s[insn->dst_reg];
10337 if (insn->src_reg == 0) {
10338 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
10340 dst_reg->type = SCALAR_VALUE;
10341 __mark_reg_known(®s[insn->dst_reg], imm);
10345 /* All special src_reg cases are listed below. From this point onwards
10346 * we either succeed and assign a corresponding dst_reg->type after
10347 * zeroing the offset, or fail and reject the program.
10349 mark_reg_known_zero(env, regs, insn->dst_reg);
10351 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
10352 dst_reg->type = aux->btf_var.reg_type;
10353 switch (base_type(dst_reg->type)) {
10355 dst_reg->mem_size = aux->btf_var.mem_size;
10357 case PTR_TO_BTF_ID:
10358 dst_reg->btf = aux->btf_var.btf;
10359 dst_reg->btf_id = aux->btf_var.btf_id;
10362 verbose(env, "bpf verifier is misconfigured\n");
10368 if (insn->src_reg == BPF_PSEUDO_FUNC) {
10369 struct bpf_prog_aux *aux = env->prog->aux;
10370 u32 subprogno = find_subprog(env,
10371 env->insn_idx + insn->imm + 1);
10373 if (!aux->func_info) {
10374 verbose(env, "missing btf func_info\n");
10377 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
10378 verbose(env, "callback function not static\n");
10382 dst_reg->type = PTR_TO_FUNC;
10383 dst_reg->subprogno = subprogno;
10387 map = env->used_maps[aux->map_index];
10388 dst_reg->map_ptr = map;
10390 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
10391 insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
10392 dst_reg->type = PTR_TO_MAP_VALUE;
10393 dst_reg->off = aux->map_off;
10394 if (map_value_has_spin_lock(map))
10395 dst_reg->id = ++env->id_gen;
10396 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
10397 insn->src_reg == BPF_PSEUDO_MAP_IDX) {
10398 dst_reg->type = CONST_PTR_TO_MAP;
10400 verbose(env, "bpf verifier is misconfigured\n");
10407 static bool may_access_skb(enum bpf_prog_type type)
10410 case BPF_PROG_TYPE_SOCKET_FILTER:
10411 case BPF_PROG_TYPE_SCHED_CLS:
10412 case BPF_PROG_TYPE_SCHED_ACT:
10419 /* verify safety of LD_ABS|LD_IND instructions:
10420 * - they can only appear in the programs where ctx == skb
10421 * - since they are wrappers of function calls, they scratch R1-R5 registers,
10422 * preserve R6-R9, and store return value into R0
10425 * ctx == skb == R6 == CTX
10428 * SRC == any register
10429 * IMM == 32-bit immediate
10432 * R0 - 8/16/32-bit skb data converted to cpu endianness
10434 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
10436 struct bpf_reg_state *regs = cur_regs(env);
10437 static const int ctx_reg = BPF_REG_6;
10438 u8 mode = BPF_MODE(insn->code);
10441 if (!may_access_skb(resolve_prog_type(env->prog))) {
10442 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
10446 if (!env->ops->gen_ld_abs) {
10447 verbose(env, "bpf verifier is misconfigured\n");
10451 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
10452 BPF_SIZE(insn->code) == BPF_DW ||
10453 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
10454 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
10458 /* check whether implicit source operand (register R6) is readable */
10459 err = check_reg_arg(env, ctx_reg, SRC_OP);
10463 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
10464 * gen_ld_abs() may terminate the program at runtime, leading to
10467 err = check_reference_leak(env);
10469 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
10473 if (env->cur_state->active_spin_lock) {
10474 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
10478 if (regs[ctx_reg].type != PTR_TO_CTX) {
10480 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
10484 if (mode == BPF_IND) {
10485 /* check explicit source operand */
10486 err = check_reg_arg(env, insn->src_reg, SRC_OP);
10491 err = check_ptr_off_reg(env, ®s[ctx_reg], ctx_reg);
10495 /* reset caller saved regs to unreadable */
10496 for (i = 0; i < CALLER_SAVED_REGS; i++) {
10497 mark_reg_not_init(env, regs, caller_saved[i]);
10498 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
10501 /* mark destination R0 register as readable, since it contains
10502 * the value fetched from the packet.
10503 * Already marked as written above.
10505 mark_reg_unknown(env, regs, BPF_REG_0);
10506 /* ld_abs load up to 32-bit skb data. */
10507 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
10511 static int check_return_code(struct bpf_verifier_env *env)
10513 struct tnum enforce_attach_type_range = tnum_unknown;
10514 const struct bpf_prog *prog = env->prog;
10515 struct bpf_reg_state *reg;
10516 struct tnum range = tnum_range(0, 1);
10517 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
10519 struct bpf_func_state *frame = env->cur_state->frame[0];
10520 const bool is_subprog = frame->subprogno;
10522 /* LSM and struct_ops func-ptr's return type could be "void" */
10524 switch (prog_type) {
10525 case BPF_PROG_TYPE_LSM:
10526 if (prog->expected_attach_type == BPF_LSM_CGROUP)
10527 /* See below, can be 0 or 0-1 depending on hook. */
10530 case BPF_PROG_TYPE_STRUCT_OPS:
10531 if (!prog->aux->attach_func_proto->type)
10539 /* eBPF calling convention is such that R0 is used
10540 * to return the value from eBPF program.
10541 * Make sure that it's readable at this time
10542 * of bpf_exit, which means that program wrote
10543 * something into it earlier
10545 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
10549 if (is_pointer_value(env, BPF_REG_0)) {
10550 verbose(env, "R0 leaks addr as return value\n");
10554 reg = cur_regs(env) + BPF_REG_0;
10556 if (frame->in_async_callback_fn) {
10557 /* enforce return zero from async callbacks like timer */
10558 if (reg->type != SCALAR_VALUE) {
10559 verbose(env, "In async callback the register R0 is not a known value (%s)\n",
10560 reg_type_str(env, reg->type));
10564 if (!tnum_in(tnum_const(0), reg->var_off)) {
10565 verbose_invalid_scalar(env, reg, &range, "async callback", "R0");
10572 if (reg->type != SCALAR_VALUE) {
10573 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
10574 reg_type_str(env, reg->type));
10580 switch (prog_type) {
10581 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
10582 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
10583 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
10584 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
10585 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
10586 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
10587 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
10588 range = tnum_range(1, 1);
10589 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
10590 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
10591 range = tnum_range(0, 3);
10593 case BPF_PROG_TYPE_CGROUP_SKB:
10594 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
10595 range = tnum_range(0, 3);
10596 enforce_attach_type_range = tnum_range(2, 3);
10599 case BPF_PROG_TYPE_CGROUP_SOCK:
10600 case BPF_PROG_TYPE_SOCK_OPS:
10601 case BPF_PROG_TYPE_CGROUP_DEVICE:
10602 case BPF_PROG_TYPE_CGROUP_SYSCTL:
10603 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
10605 case BPF_PROG_TYPE_RAW_TRACEPOINT:
10606 if (!env->prog->aux->attach_btf_id)
10608 range = tnum_const(0);
10610 case BPF_PROG_TYPE_TRACING:
10611 switch (env->prog->expected_attach_type) {
10612 case BPF_TRACE_FENTRY:
10613 case BPF_TRACE_FEXIT:
10614 range = tnum_const(0);
10616 case BPF_TRACE_RAW_TP:
10617 case BPF_MODIFY_RETURN:
10619 case BPF_TRACE_ITER:
10625 case BPF_PROG_TYPE_SK_LOOKUP:
10626 range = tnum_range(SK_DROP, SK_PASS);
10629 case BPF_PROG_TYPE_LSM:
10630 if (env->prog->expected_attach_type != BPF_LSM_CGROUP) {
10631 /* Regular BPF_PROG_TYPE_LSM programs can return
10636 if (!env->prog->aux->attach_func_proto->type) {
10637 /* Make sure programs that attach to void
10638 * hooks don't try to modify return value.
10640 range = tnum_range(1, 1);
10644 case BPF_PROG_TYPE_EXT:
10645 /* freplace program can return anything as its return value
10646 * depends on the to-be-replaced kernel func or bpf program.
10652 if (reg->type != SCALAR_VALUE) {
10653 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
10654 reg_type_str(env, reg->type));
10658 if (!tnum_in(range, reg->var_off)) {
10659 verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
10660 if (prog->expected_attach_type == BPF_LSM_CGROUP &&
10661 prog_type == BPF_PROG_TYPE_LSM &&
10662 !prog->aux->attach_func_proto->type)
10663 verbose(env, "Note, BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
10667 if (!tnum_is_unknown(enforce_attach_type_range) &&
10668 tnum_in(enforce_attach_type_range, reg->var_off))
10669 env->prog->enforce_expected_attach_type = 1;
10673 /* non-recursive DFS pseudo code
10674 * 1 procedure DFS-iterative(G,v):
10675 * 2 label v as discovered
10676 * 3 let S be a stack
10678 * 5 while S is not empty
10680 * 7 if t is what we're looking for:
10682 * 9 for all edges e in G.adjacentEdges(t) do
10683 * 10 if edge e is already labelled
10684 * 11 continue with the next edge
10685 * 12 w <- G.adjacentVertex(t,e)
10686 * 13 if vertex w is not discovered and not explored
10687 * 14 label e as tree-edge
10688 * 15 label w as discovered
10691 * 18 else if vertex w is discovered
10692 * 19 label e as back-edge
10694 * 21 // vertex w is explored
10695 * 22 label e as forward- or cross-edge
10696 * 23 label t as explored
10700 * 0x10 - discovered
10701 * 0x11 - discovered and fall-through edge labelled
10702 * 0x12 - discovered and fall-through and branch edges labelled
10713 static u32 state_htab_size(struct bpf_verifier_env *env)
10715 return env->prog->len;
10718 static struct bpf_verifier_state_list **explored_state(
10719 struct bpf_verifier_env *env,
10722 struct bpf_verifier_state *cur = env->cur_state;
10723 struct bpf_func_state *state = cur->frame[cur->curframe];
10725 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
10728 static void init_explored_state(struct bpf_verifier_env *env, int idx)
10730 env->insn_aux_data[idx].prune_point = true;
10734 DONE_EXPLORING = 0,
10735 KEEP_EXPLORING = 1,
10738 /* t, w, e - match pseudo-code above:
10739 * t - index of current instruction
10740 * w - next instruction
10743 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
10746 int *insn_stack = env->cfg.insn_stack;
10747 int *insn_state = env->cfg.insn_state;
10749 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
10750 return DONE_EXPLORING;
10752 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
10753 return DONE_EXPLORING;
10755 if (w < 0 || w >= env->prog->len) {
10756 verbose_linfo(env, t, "%d: ", t);
10757 verbose(env, "jump out of range from insn %d to %d\n", t, w);
10762 /* mark branch target for state pruning */
10763 init_explored_state(env, w);
10765 if (insn_state[w] == 0) {
10767 insn_state[t] = DISCOVERED | e;
10768 insn_state[w] = DISCOVERED;
10769 if (env->cfg.cur_stack >= env->prog->len)
10771 insn_stack[env->cfg.cur_stack++] = w;
10772 return KEEP_EXPLORING;
10773 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
10774 if (loop_ok && env->bpf_capable)
10775 return DONE_EXPLORING;
10776 verbose_linfo(env, t, "%d: ", t);
10777 verbose_linfo(env, w, "%d: ", w);
10778 verbose(env, "back-edge from insn %d to %d\n", t, w);
10780 } else if (insn_state[w] == EXPLORED) {
10781 /* forward- or cross-edge */
10782 insn_state[t] = DISCOVERED | e;
10784 verbose(env, "insn state internal bug\n");
10787 return DONE_EXPLORING;
10790 static int visit_func_call_insn(int t, int insn_cnt,
10791 struct bpf_insn *insns,
10792 struct bpf_verifier_env *env,
10797 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
10801 if (t + 1 < insn_cnt)
10802 init_explored_state(env, t + 1);
10803 if (visit_callee) {
10804 init_explored_state(env, t);
10805 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env,
10806 /* It's ok to allow recursion from CFG point of
10807 * view. __check_func_call() will do the actual
10810 bpf_pseudo_func(insns + t));
10815 /* Visits the instruction at index t and returns one of the following:
10816 * < 0 - an error occurred
10817 * DONE_EXPLORING - the instruction was fully explored
10818 * KEEP_EXPLORING - there is still work to be done before it is fully explored
10820 static int visit_insn(int t, int insn_cnt, struct bpf_verifier_env *env)
10822 struct bpf_insn *insns = env->prog->insnsi;
10825 if (bpf_pseudo_func(insns + t))
10826 return visit_func_call_insn(t, insn_cnt, insns, env, true);
10828 /* All non-branch instructions have a single fall-through edge. */
10829 if (BPF_CLASS(insns[t].code) != BPF_JMP &&
10830 BPF_CLASS(insns[t].code) != BPF_JMP32)
10831 return push_insn(t, t + 1, FALLTHROUGH, env, false);
10833 switch (BPF_OP(insns[t].code)) {
10835 return DONE_EXPLORING;
10838 if (insns[t].imm == BPF_FUNC_timer_set_callback)
10839 /* Mark this call insn to trigger is_state_visited() check
10840 * before call itself is processed by __check_func_call().
10841 * Otherwise new async state will be pushed for further
10844 init_explored_state(env, t);
10845 return visit_func_call_insn(t, insn_cnt, insns, env,
10846 insns[t].src_reg == BPF_PSEUDO_CALL);
10849 if (BPF_SRC(insns[t].code) != BPF_K)
10852 /* unconditional jump with single edge */
10853 ret = push_insn(t, t + insns[t].off + 1, FALLTHROUGH, env,
10858 /* unconditional jmp is not a good pruning point,
10859 * but it's marked, since backtracking needs
10860 * to record jmp history in is_state_visited().
10862 init_explored_state(env, t + insns[t].off + 1);
10863 /* tell verifier to check for equivalent states
10864 * after every call and jump
10866 if (t + 1 < insn_cnt)
10867 init_explored_state(env, t + 1);
10872 /* conditional jump with two edges */
10873 init_explored_state(env, t);
10874 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
10878 return push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
10882 /* non-recursive depth-first-search to detect loops in BPF program
10883 * loop == back-edge in directed graph
10885 static int check_cfg(struct bpf_verifier_env *env)
10887 int insn_cnt = env->prog->len;
10888 int *insn_stack, *insn_state;
10892 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
10896 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
10898 kvfree(insn_state);
10902 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
10903 insn_stack[0] = 0; /* 0 is the first instruction */
10904 env->cfg.cur_stack = 1;
10906 while (env->cfg.cur_stack > 0) {
10907 int t = insn_stack[env->cfg.cur_stack - 1];
10909 ret = visit_insn(t, insn_cnt, env);
10911 case DONE_EXPLORING:
10912 insn_state[t] = EXPLORED;
10913 env->cfg.cur_stack--;
10915 case KEEP_EXPLORING:
10919 verbose(env, "visit_insn internal bug\n");
10926 if (env->cfg.cur_stack < 0) {
10927 verbose(env, "pop stack internal bug\n");
10932 for (i = 0; i < insn_cnt; i++) {
10933 if (insn_state[i] != EXPLORED) {
10934 verbose(env, "unreachable insn %d\n", i);
10939 ret = 0; /* cfg looks good */
10942 kvfree(insn_state);
10943 kvfree(insn_stack);
10944 env->cfg.insn_state = env->cfg.insn_stack = NULL;
10948 static int check_abnormal_return(struct bpf_verifier_env *env)
10952 for (i = 1; i < env->subprog_cnt; i++) {
10953 if (env->subprog_info[i].has_ld_abs) {
10954 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
10957 if (env->subprog_info[i].has_tail_call) {
10958 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
10965 /* The minimum supported BTF func info size */
10966 #define MIN_BPF_FUNCINFO_SIZE 8
10967 #define MAX_FUNCINFO_REC_SIZE 252
10969 static int check_btf_func(struct bpf_verifier_env *env,
10970 const union bpf_attr *attr,
10973 const struct btf_type *type, *func_proto, *ret_type;
10974 u32 i, nfuncs, urec_size, min_size;
10975 u32 krec_size = sizeof(struct bpf_func_info);
10976 struct bpf_func_info *krecord;
10977 struct bpf_func_info_aux *info_aux = NULL;
10978 struct bpf_prog *prog;
10979 const struct btf *btf;
10981 u32 prev_offset = 0;
10982 bool scalar_return;
10985 nfuncs = attr->func_info_cnt;
10987 if (check_abnormal_return(env))
10992 if (nfuncs != env->subprog_cnt) {
10993 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
10997 urec_size = attr->func_info_rec_size;
10998 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
10999 urec_size > MAX_FUNCINFO_REC_SIZE ||
11000 urec_size % sizeof(u32)) {
11001 verbose(env, "invalid func info rec size %u\n", urec_size);
11006 btf = prog->aux->btf;
11008 urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
11009 min_size = min_t(u32, krec_size, urec_size);
11011 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
11014 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
11018 for (i = 0; i < nfuncs; i++) {
11019 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
11021 if (ret == -E2BIG) {
11022 verbose(env, "nonzero tailing record in func info");
11023 /* set the size kernel expects so loader can zero
11024 * out the rest of the record.
11026 if (copy_to_bpfptr_offset(uattr,
11027 offsetof(union bpf_attr, func_info_rec_size),
11028 &min_size, sizeof(min_size)))
11034 if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
11039 /* check insn_off */
11042 if (krecord[i].insn_off) {
11044 "nonzero insn_off %u for the first func info record",
11045 krecord[i].insn_off);
11048 } else if (krecord[i].insn_off <= prev_offset) {
11050 "same or smaller insn offset (%u) than previous func info record (%u)",
11051 krecord[i].insn_off, prev_offset);
11055 if (env->subprog_info[i].start != krecord[i].insn_off) {
11056 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
11060 /* check type_id */
11061 type = btf_type_by_id(btf, krecord[i].type_id);
11062 if (!type || !btf_type_is_func(type)) {
11063 verbose(env, "invalid type id %d in func info",
11064 krecord[i].type_id);
11067 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
11069 func_proto = btf_type_by_id(btf, type->type);
11070 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
11071 /* btf_func_check() already verified it during BTF load */
11073 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
11075 btf_type_is_small_int(ret_type) || btf_is_any_enum(ret_type);
11076 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
11077 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
11080 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
11081 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
11085 prev_offset = krecord[i].insn_off;
11086 bpfptr_add(&urecord, urec_size);
11089 prog->aux->func_info = krecord;
11090 prog->aux->func_info_cnt = nfuncs;
11091 prog->aux->func_info_aux = info_aux;
11100 static void adjust_btf_func(struct bpf_verifier_env *env)
11102 struct bpf_prog_aux *aux = env->prog->aux;
11105 if (!aux->func_info)
11108 for (i = 0; i < env->subprog_cnt; i++)
11109 aux->func_info[i].insn_off = env->subprog_info[i].start;
11112 #define MIN_BPF_LINEINFO_SIZE offsetofend(struct bpf_line_info, line_col)
11113 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
11115 static int check_btf_line(struct bpf_verifier_env *env,
11116 const union bpf_attr *attr,
11119 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
11120 struct bpf_subprog_info *sub;
11121 struct bpf_line_info *linfo;
11122 struct bpf_prog *prog;
11123 const struct btf *btf;
11127 nr_linfo = attr->line_info_cnt;
11130 if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info))
11133 rec_size = attr->line_info_rec_size;
11134 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
11135 rec_size > MAX_LINEINFO_REC_SIZE ||
11136 rec_size & (sizeof(u32) - 1))
11139 /* Need to zero it in case the userspace may
11140 * pass in a smaller bpf_line_info object.
11142 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
11143 GFP_KERNEL | __GFP_NOWARN);
11148 btf = prog->aux->btf;
11151 sub = env->subprog_info;
11152 ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
11153 expected_size = sizeof(struct bpf_line_info);
11154 ncopy = min_t(u32, expected_size, rec_size);
11155 for (i = 0; i < nr_linfo; i++) {
11156 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
11158 if (err == -E2BIG) {
11159 verbose(env, "nonzero tailing record in line_info");
11160 if (copy_to_bpfptr_offset(uattr,
11161 offsetof(union bpf_attr, line_info_rec_size),
11162 &expected_size, sizeof(expected_size)))
11168 if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
11174 * Check insn_off to ensure
11175 * 1) strictly increasing AND
11176 * 2) bounded by prog->len
11178 * The linfo[0].insn_off == 0 check logically falls into
11179 * the later "missing bpf_line_info for func..." case
11180 * because the first linfo[0].insn_off must be the
11181 * first sub also and the first sub must have
11182 * subprog_info[0].start == 0.
11184 if ((i && linfo[i].insn_off <= prev_offset) ||
11185 linfo[i].insn_off >= prog->len) {
11186 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
11187 i, linfo[i].insn_off, prev_offset,
11193 if (!prog->insnsi[linfo[i].insn_off].code) {
11195 "Invalid insn code at line_info[%u].insn_off\n",
11201 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
11202 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
11203 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
11208 if (s != env->subprog_cnt) {
11209 if (linfo[i].insn_off == sub[s].start) {
11210 sub[s].linfo_idx = i;
11212 } else if (sub[s].start < linfo[i].insn_off) {
11213 verbose(env, "missing bpf_line_info for func#%u\n", s);
11219 prev_offset = linfo[i].insn_off;
11220 bpfptr_add(&ulinfo, rec_size);
11223 if (s != env->subprog_cnt) {
11224 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
11225 env->subprog_cnt - s, s);
11230 prog->aux->linfo = linfo;
11231 prog->aux->nr_linfo = nr_linfo;
11240 #define MIN_CORE_RELO_SIZE sizeof(struct bpf_core_relo)
11241 #define MAX_CORE_RELO_SIZE MAX_FUNCINFO_REC_SIZE
11243 static int check_core_relo(struct bpf_verifier_env *env,
11244 const union bpf_attr *attr,
11247 u32 i, nr_core_relo, ncopy, expected_size, rec_size;
11248 struct bpf_core_relo core_relo = {};
11249 struct bpf_prog *prog = env->prog;
11250 const struct btf *btf = prog->aux->btf;
11251 struct bpf_core_ctx ctx = {
11255 bpfptr_t u_core_relo;
11258 nr_core_relo = attr->core_relo_cnt;
11261 if (nr_core_relo > INT_MAX / sizeof(struct bpf_core_relo))
11264 rec_size = attr->core_relo_rec_size;
11265 if (rec_size < MIN_CORE_RELO_SIZE ||
11266 rec_size > MAX_CORE_RELO_SIZE ||
11267 rec_size % sizeof(u32))
11270 u_core_relo = make_bpfptr(attr->core_relos, uattr.is_kernel);
11271 expected_size = sizeof(struct bpf_core_relo);
11272 ncopy = min_t(u32, expected_size, rec_size);
11274 /* Unlike func_info and line_info, copy and apply each CO-RE
11275 * relocation record one at a time.
11277 for (i = 0; i < nr_core_relo; i++) {
11278 /* future proofing when sizeof(bpf_core_relo) changes */
11279 err = bpf_check_uarg_tail_zero(u_core_relo, expected_size, rec_size);
11281 if (err == -E2BIG) {
11282 verbose(env, "nonzero tailing record in core_relo");
11283 if (copy_to_bpfptr_offset(uattr,
11284 offsetof(union bpf_attr, core_relo_rec_size),
11285 &expected_size, sizeof(expected_size)))
11291 if (copy_from_bpfptr(&core_relo, u_core_relo, ncopy)) {
11296 if (core_relo.insn_off % 8 || core_relo.insn_off / 8 >= prog->len) {
11297 verbose(env, "Invalid core_relo[%u].insn_off:%u prog->len:%u\n",
11298 i, core_relo.insn_off, prog->len);
11303 err = bpf_core_apply(&ctx, &core_relo, i,
11304 &prog->insnsi[core_relo.insn_off / 8]);
11307 bpfptr_add(&u_core_relo, rec_size);
11312 static int check_btf_info(struct bpf_verifier_env *env,
11313 const union bpf_attr *attr,
11319 if (!attr->func_info_cnt && !attr->line_info_cnt) {
11320 if (check_abnormal_return(env))
11325 btf = btf_get_by_fd(attr->prog_btf_fd);
11327 return PTR_ERR(btf);
11328 if (btf_is_kernel(btf)) {
11332 env->prog->aux->btf = btf;
11334 err = check_btf_func(env, attr, uattr);
11338 err = check_btf_line(env, attr, uattr);
11342 err = check_core_relo(env, attr, uattr);
11349 /* check %cur's range satisfies %old's */
11350 static bool range_within(struct bpf_reg_state *old,
11351 struct bpf_reg_state *cur)
11353 return old->umin_value <= cur->umin_value &&
11354 old->umax_value >= cur->umax_value &&
11355 old->smin_value <= cur->smin_value &&
11356 old->smax_value >= cur->smax_value &&
11357 old->u32_min_value <= cur->u32_min_value &&
11358 old->u32_max_value >= cur->u32_max_value &&
11359 old->s32_min_value <= cur->s32_min_value &&
11360 old->s32_max_value >= cur->s32_max_value;
11363 /* If in the old state two registers had the same id, then they need to have
11364 * the same id in the new state as well. But that id could be different from
11365 * the old state, so we need to track the mapping from old to new ids.
11366 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
11367 * regs with old id 5 must also have new id 9 for the new state to be safe. But
11368 * regs with a different old id could still have new id 9, we don't care about
11370 * So we look through our idmap to see if this old id has been seen before. If
11371 * so, we require the new id to match; otherwise, we add the id pair to the map.
11373 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_id_pair *idmap)
11377 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
11378 if (!idmap[i].old) {
11379 /* Reached an empty slot; haven't seen this id before */
11380 idmap[i].old = old_id;
11381 idmap[i].cur = cur_id;
11384 if (idmap[i].old == old_id)
11385 return idmap[i].cur == cur_id;
11387 /* We ran out of idmap slots, which should be impossible */
11392 static void clean_func_state(struct bpf_verifier_env *env,
11393 struct bpf_func_state *st)
11395 enum bpf_reg_liveness live;
11398 for (i = 0; i < BPF_REG_FP; i++) {
11399 live = st->regs[i].live;
11400 /* liveness must not touch this register anymore */
11401 st->regs[i].live |= REG_LIVE_DONE;
11402 if (!(live & REG_LIVE_READ))
11403 /* since the register is unused, clear its state
11404 * to make further comparison simpler
11406 __mark_reg_not_init(env, &st->regs[i]);
11409 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
11410 live = st->stack[i].spilled_ptr.live;
11411 /* liveness must not touch this stack slot anymore */
11412 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
11413 if (!(live & REG_LIVE_READ)) {
11414 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
11415 for (j = 0; j < BPF_REG_SIZE; j++)
11416 st->stack[i].slot_type[j] = STACK_INVALID;
11421 static void clean_verifier_state(struct bpf_verifier_env *env,
11422 struct bpf_verifier_state *st)
11426 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
11427 /* all regs in this state in all frames were already marked */
11430 for (i = 0; i <= st->curframe; i++)
11431 clean_func_state(env, st->frame[i]);
11434 /* the parentage chains form a tree.
11435 * the verifier states are added to state lists at given insn and
11436 * pushed into state stack for future exploration.
11437 * when the verifier reaches bpf_exit insn some of the verifer states
11438 * stored in the state lists have their final liveness state already,
11439 * but a lot of states will get revised from liveness point of view when
11440 * the verifier explores other branches.
11443 * 2: if r1 == 100 goto pc+1
11446 * when the verifier reaches exit insn the register r0 in the state list of
11447 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
11448 * of insn 2 and goes exploring further. At the insn 4 it will walk the
11449 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
11451 * Since the verifier pushes the branch states as it sees them while exploring
11452 * the program the condition of walking the branch instruction for the second
11453 * time means that all states below this branch were already explored and
11454 * their final liveness marks are already propagated.
11455 * Hence when the verifier completes the search of state list in is_state_visited()
11456 * we can call this clean_live_states() function to mark all liveness states
11457 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
11458 * will not be used.
11459 * This function also clears the registers and stack for states that !READ
11460 * to simplify state merging.
11462 * Important note here that walking the same branch instruction in the callee
11463 * doesn't meant that the states are DONE. The verifier has to compare
11466 static void clean_live_states(struct bpf_verifier_env *env, int insn,
11467 struct bpf_verifier_state *cur)
11469 struct bpf_verifier_state_list *sl;
11472 sl = *explored_state(env, insn);
11474 if (sl->state.branches)
11476 if (sl->state.insn_idx != insn ||
11477 sl->state.curframe != cur->curframe)
11479 for (i = 0; i <= cur->curframe; i++)
11480 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
11482 clean_verifier_state(env, &sl->state);
11488 /* Returns true if (rold safe implies rcur safe) */
11489 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
11490 struct bpf_reg_state *rcur, struct bpf_id_pair *idmap)
11494 if (!(rold->live & REG_LIVE_READ))
11495 /* explored state didn't use this */
11498 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
11500 if (rold->type == PTR_TO_STACK)
11501 /* two stack pointers are equal only if they're pointing to
11502 * the same stack frame, since fp-8 in foo != fp-8 in bar
11504 return equal && rold->frameno == rcur->frameno;
11509 if (rold->type == NOT_INIT)
11510 /* explored state can't have used this */
11512 if (rcur->type == NOT_INIT)
11514 switch (base_type(rold->type)) {
11516 if (env->explore_alu_limits)
11518 if (rcur->type == SCALAR_VALUE) {
11519 if (!rold->precise && !rcur->precise)
11521 /* new val must satisfy old val knowledge */
11522 return range_within(rold, rcur) &&
11523 tnum_in(rold->var_off, rcur->var_off);
11525 /* We're trying to use a pointer in place of a scalar.
11526 * Even if the scalar was unbounded, this could lead to
11527 * pointer leaks because scalars are allowed to leak
11528 * while pointers are not. We could make this safe in
11529 * special cases if root is calling us, but it's
11530 * probably not worth the hassle.
11534 case PTR_TO_MAP_KEY:
11535 case PTR_TO_MAP_VALUE:
11536 /* a PTR_TO_MAP_VALUE could be safe to use as a
11537 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
11538 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
11539 * checked, doing so could have affected others with the same
11540 * id, and we can't check for that because we lost the id when
11541 * we converted to a PTR_TO_MAP_VALUE.
11543 if (type_may_be_null(rold->type)) {
11544 if (!type_may_be_null(rcur->type))
11546 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
11548 /* Check our ids match any regs they're supposed to */
11549 return check_ids(rold->id, rcur->id, idmap);
11552 /* If the new min/max/var_off satisfy the old ones and
11553 * everything else matches, we are OK.
11554 * 'id' is not compared, since it's only used for maps with
11555 * bpf_spin_lock inside map element and in such cases if
11556 * the rest of the prog is valid for one map element then
11557 * it's valid for all map elements regardless of the key
11558 * used in bpf_map_lookup()
11560 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
11561 range_within(rold, rcur) &&
11562 tnum_in(rold->var_off, rcur->var_off);
11563 case PTR_TO_PACKET_META:
11564 case PTR_TO_PACKET:
11565 if (rcur->type != rold->type)
11567 /* We must have at least as much range as the old ptr
11568 * did, so that any accesses which were safe before are
11569 * still safe. This is true even if old range < old off,
11570 * since someone could have accessed through (ptr - k), or
11571 * even done ptr -= k in a register, to get a safe access.
11573 if (rold->range > rcur->range)
11575 /* If the offsets don't match, we can't trust our alignment;
11576 * nor can we be sure that we won't fall out of range.
11578 if (rold->off != rcur->off)
11580 /* id relations must be preserved */
11581 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
11583 /* new val must satisfy old val knowledge */
11584 return range_within(rold, rcur) &&
11585 tnum_in(rold->var_off, rcur->var_off);
11587 case CONST_PTR_TO_MAP:
11588 case PTR_TO_PACKET_END:
11589 case PTR_TO_FLOW_KEYS:
11590 case PTR_TO_SOCKET:
11591 case PTR_TO_SOCK_COMMON:
11592 case PTR_TO_TCP_SOCK:
11593 case PTR_TO_XDP_SOCK:
11594 /* Only valid matches are exact, which memcmp() above
11595 * would have accepted
11598 /* Don't know what's going on, just say it's not safe */
11602 /* Shouldn't get here; if we do, say it's not safe */
11607 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
11608 struct bpf_func_state *cur, struct bpf_id_pair *idmap)
11612 /* walk slots of the explored stack and ignore any additional
11613 * slots in the current stack, since explored(safe) state
11616 for (i = 0; i < old->allocated_stack; i++) {
11617 spi = i / BPF_REG_SIZE;
11619 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
11620 i += BPF_REG_SIZE - 1;
11621 /* explored state didn't use this */
11625 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
11628 /* explored stack has more populated slots than current stack
11629 * and these slots were used
11631 if (i >= cur->allocated_stack)
11634 /* if old state was safe with misc data in the stack
11635 * it will be safe with zero-initialized stack.
11636 * The opposite is not true
11638 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
11639 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
11641 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
11642 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
11643 /* Ex: old explored (safe) state has STACK_SPILL in
11644 * this stack slot, but current has STACK_MISC ->
11645 * this verifier states are not equivalent,
11646 * return false to continue verification of this path
11649 if (i % BPF_REG_SIZE != BPF_REG_SIZE - 1)
11651 if (!is_spilled_reg(&old->stack[spi]))
11653 if (!regsafe(env, &old->stack[spi].spilled_ptr,
11654 &cur->stack[spi].spilled_ptr, idmap))
11655 /* when explored and current stack slot are both storing
11656 * spilled registers, check that stored pointers types
11657 * are the same as well.
11658 * Ex: explored safe path could have stored
11659 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
11660 * but current path has stored:
11661 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
11662 * such verifier states are not equivalent.
11663 * return false to continue verification of this path
11670 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
11672 if (old->acquired_refs != cur->acquired_refs)
11674 return !memcmp(old->refs, cur->refs,
11675 sizeof(*old->refs) * old->acquired_refs);
11678 /* compare two verifier states
11680 * all states stored in state_list are known to be valid, since
11681 * verifier reached 'bpf_exit' instruction through them
11683 * this function is called when verifier exploring different branches of
11684 * execution popped from the state stack. If it sees an old state that has
11685 * more strict register state and more strict stack state then this execution
11686 * branch doesn't need to be explored further, since verifier already
11687 * concluded that more strict state leads to valid finish.
11689 * Therefore two states are equivalent if register state is more conservative
11690 * and explored stack state is more conservative than the current one.
11693 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
11694 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
11696 * In other words if current stack state (one being explored) has more
11697 * valid slots than old one that already passed validation, it means
11698 * the verifier can stop exploring and conclude that current state is valid too
11700 * Similarly with registers. If explored state has register type as invalid
11701 * whereas register type in current state is meaningful, it means that
11702 * the current state will reach 'bpf_exit' instruction safely
11704 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
11705 struct bpf_func_state *cur)
11709 memset(env->idmap_scratch, 0, sizeof(env->idmap_scratch));
11710 for (i = 0; i < MAX_BPF_REG; i++)
11711 if (!regsafe(env, &old->regs[i], &cur->regs[i],
11712 env->idmap_scratch))
11715 if (!stacksafe(env, old, cur, env->idmap_scratch))
11718 if (!refsafe(old, cur))
11724 static bool states_equal(struct bpf_verifier_env *env,
11725 struct bpf_verifier_state *old,
11726 struct bpf_verifier_state *cur)
11730 if (old->curframe != cur->curframe)
11733 /* Verification state from speculative execution simulation
11734 * must never prune a non-speculative execution one.
11736 if (old->speculative && !cur->speculative)
11739 if (old->active_spin_lock != cur->active_spin_lock)
11742 /* for states to be equal callsites have to be the same
11743 * and all frame states need to be equivalent
11745 for (i = 0; i <= old->curframe; i++) {
11746 if (old->frame[i]->callsite != cur->frame[i]->callsite)
11748 if (!func_states_equal(env, old->frame[i], cur->frame[i]))
11754 /* Return 0 if no propagation happened. Return negative error code if error
11755 * happened. Otherwise, return the propagated bit.
11757 static int propagate_liveness_reg(struct bpf_verifier_env *env,
11758 struct bpf_reg_state *reg,
11759 struct bpf_reg_state *parent_reg)
11761 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
11762 u8 flag = reg->live & REG_LIVE_READ;
11765 /* When comes here, read flags of PARENT_REG or REG could be any of
11766 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
11767 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
11769 if (parent_flag == REG_LIVE_READ64 ||
11770 /* Or if there is no read flag from REG. */
11772 /* Or if the read flag from REG is the same as PARENT_REG. */
11773 parent_flag == flag)
11776 err = mark_reg_read(env, reg, parent_reg, flag);
11783 /* A write screens off any subsequent reads; but write marks come from the
11784 * straight-line code between a state and its parent. When we arrive at an
11785 * equivalent state (jump target or such) we didn't arrive by the straight-line
11786 * code, so read marks in the state must propagate to the parent regardless
11787 * of the state's write marks. That's what 'parent == state->parent' comparison
11788 * in mark_reg_read() is for.
11790 static int propagate_liveness(struct bpf_verifier_env *env,
11791 const struct bpf_verifier_state *vstate,
11792 struct bpf_verifier_state *vparent)
11794 struct bpf_reg_state *state_reg, *parent_reg;
11795 struct bpf_func_state *state, *parent;
11796 int i, frame, err = 0;
11798 if (vparent->curframe != vstate->curframe) {
11799 WARN(1, "propagate_live: parent frame %d current frame %d\n",
11800 vparent->curframe, vstate->curframe);
11803 /* Propagate read liveness of registers... */
11804 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
11805 for (frame = 0; frame <= vstate->curframe; frame++) {
11806 parent = vparent->frame[frame];
11807 state = vstate->frame[frame];
11808 parent_reg = parent->regs;
11809 state_reg = state->regs;
11810 /* We don't need to worry about FP liveness, it's read-only */
11811 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
11812 err = propagate_liveness_reg(env, &state_reg[i],
11816 if (err == REG_LIVE_READ64)
11817 mark_insn_zext(env, &parent_reg[i]);
11820 /* Propagate stack slots. */
11821 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
11822 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
11823 parent_reg = &parent->stack[i].spilled_ptr;
11824 state_reg = &state->stack[i].spilled_ptr;
11825 err = propagate_liveness_reg(env, state_reg,
11834 /* find precise scalars in the previous equivalent state and
11835 * propagate them into the current state
11837 static int propagate_precision(struct bpf_verifier_env *env,
11838 const struct bpf_verifier_state *old)
11840 struct bpf_reg_state *state_reg;
11841 struct bpf_func_state *state;
11844 state = old->frame[old->curframe];
11845 state_reg = state->regs;
11846 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
11847 if (state_reg->type != SCALAR_VALUE ||
11848 !state_reg->precise)
11850 if (env->log.level & BPF_LOG_LEVEL2)
11851 verbose(env, "propagating r%d\n", i);
11852 err = mark_chain_precision(env, i);
11857 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
11858 if (!is_spilled_reg(&state->stack[i]))
11860 state_reg = &state->stack[i].spilled_ptr;
11861 if (state_reg->type != SCALAR_VALUE ||
11862 !state_reg->precise)
11864 if (env->log.level & BPF_LOG_LEVEL2)
11865 verbose(env, "propagating fp%d\n",
11866 (-i - 1) * BPF_REG_SIZE);
11867 err = mark_chain_precision_stack(env, i);
11874 static bool states_maybe_looping(struct bpf_verifier_state *old,
11875 struct bpf_verifier_state *cur)
11877 struct bpf_func_state *fold, *fcur;
11878 int i, fr = cur->curframe;
11880 if (old->curframe != fr)
11883 fold = old->frame[fr];
11884 fcur = cur->frame[fr];
11885 for (i = 0; i < MAX_BPF_REG; i++)
11886 if (memcmp(&fold->regs[i], &fcur->regs[i],
11887 offsetof(struct bpf_reg_state, parent)))
11893 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
11895 struct bpf_verifier_state_list *new_sl;
11896 struct bpf_verifier_state_list *sl, **pprev;
11897 struct bpf_verifier_state *cur = env->cur_state, *new;
11898 int i, j, err, states_cnt = 0;
11899 bool add_new_state = env->test_state_freq ? true : false;
11901 cur->last_insn_idx = env->prev_insn_idx;
11902 if (!env->insn_aux_data[insn_idx].prune_point)
11903 /* this 'insn_idx' instruction wasn't marked, so we will not
11904 * be doing state search here
11908 /* bpf progs typically have pruning point every 4 instructions
11909 * http://vger.kernel.org/bpfconf2019.html#session-1
11910 * Do not add new state for future pruning if the verifier hasn't seen
11911 * at least 2 jumps and at least 8 instructions.
11912 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
11913 * In tests that amounts to up to 50% reduction into total verifier
11914 * memory consumption and 20% verifier time speedup.
11916 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
11917 env->insn_processed - env->prev_insn_processed >= 8)
11918 add_new_state = true;
11920 pprev = explored_state(env, insn_idx);
11923 clean_live_states(env, insn_idx, cur);
11927 if (sl->state.insn_idx != insn_idx)
11930 if (sl->state.branches) {
11931 struct bpf_func_state *frame = sl->state.frame[sl->state.curframe];
11933 if (frame->in_async_callback_fn &&
11934 frame->async_entry_cnt != cur->frame[cur->curframe]->async_entry_cnt) {
11935 /* Different async_entry_cnt means that the verifier is
11936 * processing another entry into async callback.
11937 * Seeing the same state is not an indication of infinite
11938 * loop or infinite recursion.
11939 * But finding the same state doesn't mean that it's safe
11940 * to stop processing the current state. The previous state
11941 * hasn't yet reached bpf_exit, since state.branches > 0.
11942 * Checking in_async_callback_fn alone is not enough either.
11943 * Since the verifier still needs to catch infinite loops
11944 * inside async callbacks.
11946 } else if (states_maybe_looping(&sl->state, cur) &&
11947 states_equal(env, &sl->state, cur)) {
11948 verbose_linfo(env, insn_idx, "; ");
11949 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
11952 /* if the verifier is processing a loop, avoid adding new state
11953 * too often, since different loop iterations have distinct
11954 * states and may not help future pruning.
11955 * This threshold shouldn't be too low to make sure that
11956 * a loop with large bound will be rejected quickly.
11957 * The most abusive loop will be:
11959 * if r1 < 1000000 goto pc-2
11960 * 1M insn_procssed limit / 100 == 10k peak states.
11961 * This threshold shouldn't be too high either, since states
11962 * at the end of the loop are likely to be useful in pruning.
11964 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
11965 env->insn_processed - env->prev_insn_processed < 100)
11966 add_new_state = false;
11969 if (states_equal(env, &sl->state, cur)) {
11971 /* reached equivalent register/stack state,
11972 * prune the search.
11973 * Registers read by the continuation are read by us.
11974 * If we have any write marks in env->cur_state, they
11975 * will prevent corresponding reads in the continuation
11976 * from reaching our parent (an explored_state). Our
11977 * own state will get the read marks recorded, but
11978 * they'll be immediately forgotten as we're pruning
11979 * this state and will pop a new one.
11981 err = propagate_liveness(env, &sl->state, cur);
11983 /* if previous state reached the exit with precision and
11984 * current state is equivalent to it (except precsion marks)
11985 * the precision needs to be propagated back in
11986 * the current state.
11988 err = err ? : push_jmp_history(env, cur);
11989 err = err ? : propagate_precision(env, &sl->state);
11995 /* when new state is not going to be added do not increase miss count.
11996 * Otherwise several loop iterations will remove the state
11997 * recorded earlier. The goal of these heuristics is to have
11998 * states from some iterations of the loop (some in the beginning
11999 * and some at the end) to help pruning.
12003 /* heuristic to determine whether this state is beneficial
12004 * to keep checking from state equivalence point of view.
12005 * Higher numbers increase max_states_per_insn and verification time,
12006 * but do not meaningfully decrease insn_processed.
12008 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
12009 /* the state is unlikely to be useful. Remove it to
12010 * speed up verification
12013 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
12014 u32 br = sl->state.branches;
12017 "BUG live_done but branches_to_explore %d\n",
12019 free_verifier_state(&sl->state, false);
12021 env->peak_states--;
12023 /* cannot free this state, since parentage chain may
12024 * walk it later. Add it for free_list instead to
12025 * be freed at the end of verification
12027 sl->next = env->free_list;
12028 env->free_list = sl;
12038 if (env->max_states_per_insn < states_cnt)
12039 env->max_states_per_insn = states_cnt;
12041 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
12042 return push_jmp_history(env, cur);
12044 if (!add_new_state)
12045 return push_jmp_history(env, cur);
12047 /* There were no equivalent states, remember the current one.
12048 * Technically the current state is not proven to be safe yet,
12049 * but it will either reach outer most bpf_exit (which means it's safe)
12050 * or it will be rejected. When there are no loops the verifier won't be
12051 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
12052 * again on the way to bpf_exit.
12053 * When looping the sl->state.branches will be > 0 and this state
12054 * will not be considered for equivalence until branches == 0.
12056 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
12059 env->total_states++;
12060 env->peak_states++;
12061 env->prev_jmps_processed = env->jmps_processed;
12062 env->prev_insn_processed = env->insn_processed;
12064 /* add new state to the head of linked list */
12065 new = &new_sl->state;
12066 err = copy_verifier_state(new, cur);
12068 free_verifier_state(new, false);
12072 new->insn_idx = insn_idx;
12073 WARN_ONCE(new->branches != 1,
12074 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
12077 cur->first_insn_idx = insn_idx;
12078 clear_jmp_history(cur);
12079 new_sl->next = *explored_state(env, insn_idx);
12080 *explored_state(env, insn_idx) = new_sl;
12081 /* connect new state to parentage chain. Current frame needs all
12082 * registers connected. Only r6 - r9 of the callers are alive (pushed
12083 * to the stack implicitly by JITs) so in callers' frames connect just
12084 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
12085 * the state of the call instruction (with WRITTEN set), and r0 comes
12086 * from callee with its full parentage chain, anyway.
12088 /* clear write marks in current state: the writes we did are not writes
12089 * our child did, so they don't screen off its reads from us.
12090 * (There are no read marks in current state, because reads always mark
12091 * their parent and current state never has children yet. Only
12092 * explored_states can get read marks.)
12094 for (j = 0; j <= cur->curframe; j++) {
12095 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
12096 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
12097 for (i = 0; i < BPF_REG_FP; i++)
12098 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
12101 /* all stack frames are accessible from callee, clear them all */
12102 for (j = 0; j <= cur->curframe; j++) {
12103 struct bpf_func_state *frame = cur->frame[j];
12104 struct bpf_func_state *newframe = new->frame[j];
12106 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
12107 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
12108 frame->stack[i].spilled_ptr.parent =
12109 &newframe->stack[i].spilled_ptr;
12115 /* Return true if it's OK to have the same insn return a different type. */
12116 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
12118 switch (base_type(type)) {
12120 case PTR_TO_SOCKET:
12121 case PTR_TO_SOCK_COMMON:
12122 case PTR_TO_TCP_SOCK:
12123 case PTR_TO_XDP_SOCK:
12124 case PTR_TO_BTF_ID:
12131 /* If an instruction was previously used with particular pointer types, then we
12132 * need to be careful to avoid cases such as the below, where it may be ok
12133 * for one branch accessing the pointer, but not ok for the other branch:
12138 * R1 = some_other_valid_ptr;
12141 * R2 = *(u32 *)(R1 + 0);
12143 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
12145 return src != prev && (!reg_type_mismatch_ok(src) ||
12146 !reg_type_mismatch_ok(prev));
12149 static int do_check(struct bpf_verifier_env *env)
12151 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
12152 struct bpf_verifier_state *state = env->cur_state;
12153 struct bpf_insn *insns = env->prog->insnsi;
12154 struct bpf_reg_state *regs;
12155 int insn_cnt = env->prog->len;
12156 bool do_print_state = false;
12157 int prev_insn_idx = -1;
12160 struct bpf_insn *insn;
12164 env->prev_insn_idx = prev_insn_idx;
12165 if (env->insn_idx >= insn_cnt) {
12166 verbose(env, "invalid insn idx %d insn_cnt %d\n",
12167 env->insn_idx, insn_cnt);
12171 insn = &insns[env->insn_idx];
12172 class = BPF_CLASS(insn->code);
12174 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
12176 "BPF program is too large. Processed %d insn\n",
12177 env->insn_processed);
12181 err = is_state_visited(env, env->insn_idx);
12185 /* found equivalent state, can prune the search */
12186 if (env->log.level & BPF_LOG_LEVEL) {
12187 if (do_print_state)
12188 verbose(env, "\nfrom %d to %d%s: safe\n",
12189 env->prev_insn_idx, env->insn_idx,
12190 env->cur_state->speculative ?
12191 " (speculative execution)" : "");
12193 verbose(env, "%d: safe\n", env->insn_idx);
12195 goto process_bpf_exit;
12198 if (signal_pending(current))
12201 if (need_resched())
12204 if (env->log.level & BPF_LOG_LEVEL2 && do_print_state) {
12205 verbose(env, "\nfrom %d to %d%s:",
12206 env->prev_insn_idx, env->insn_idx,
12207 env->cur_state->speculative ?
12208 " (speculative execution)" : "");
12209 print_verifier_state(env, state->frame[state->curframe], true);
12210 do_print_state = false;
12213 if (env->log.level & BPF_LOG_LEVEL) {
12214 const struct bpf_insn_cbs cbs = {
12215 .cb_call = disasm_kfunc_name,
12216 .cb_print = verbose,
12217 .private_data = env,
12220 if (verifier_state_scratched(env))
12221 print_insn_state(env, state->frame[state->curframe]);
12223 verbose_linfo(env, env->insn_idx, "; ");
12224 env->prev_log_len = env->log.len_used;
12225 verbose(env, "%d: ", env->insn_idx);
12226 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
12227 env->prev_insn_print_len = env->log.len_used - env->prev_log_len;
12228 env->prev_log_len = env->log.len_used;
12231 if (bpf_prog_is_dev_bound(env->prog->aux)) {
12232 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
12233 env->prev_insn_idx);
12238 regs = cur_regs(env);
12239 sanitize_mark_insn_seen(env);
12240 prev_insn_idx = env->insn_idx;
12242 if (class == BPF_ALU || class == BPF_ALU64) {
12243 err = check_alu_op(env, insn);
12247 } else if (class == BPF_LDX) {
12248 enum bpf_reg_type *prev_src_type, src_reg_type;
12250 /* check for reserved fields is already done */
12252 /* check src operand */
12253 err = check_reg_arg(env, insn->src_reg, SRC_OP);
12257 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
12261 src_reg_type = regs[insn->src_reg].type;
12263 /* check that memory (src_reg + off) is readable,
12264 * the state of dst_reg will be updated by this func
12266 err = check_mem_access(env, env->insn_idx, insn->src_reg,
12267 insn->off, BPF_SIZE(insn->code),
12268 BPF_READ, insn->dst_reg, false);
12272 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
12274 if (*prev_src_type == NOT_INIT) {
12275 /* saw a valid insn
12276 * dst_reg = *(u32 *)(src_reg + off)
12277 * save type to validate intersecting paths
12279 *prev_src_type = src_reg_type;
12281 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
12282 /* ABuser program is trying to use the same insn
12283 * dst_reg = *(u32*) (src_reg + off)
12284 * with different pointer types:
12285 * src_reg == ctx in one branch and
12286 * src_reg == stack|map in some other branch.
12289 verbose(env, "same insn cannot be used with different pointers\n");
12293 } else if (class == BPF_STX) {
12294 enum bpf_reg_type *prev_dst_type, dst_reg_type;
12296 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
12297 err = check_atomic(env, env->insn_idx, insn);
12304 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
12305 verbose(env, "BPF_STX uses reserved fields\n");
12309 /* check src1 operand */
12310 err = check_reg_arg(env, insn->src_reg, SRC_OP);
12313 /* check src2 operand */
12314 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
12318 dst_reg_type = regs[insn->dst_reg].type;
12320 /* check that memory (dst_reg + off) is writeable */
12321 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
12322 insn->off, BPF_SIZE(insn->code),
12323 BPF_WRITE, insn->src_reg, false);
12327 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
12329 if (*prev_dst_type == NOT_INIT) {
12330 *prev_dst_type = dst_reg_type;
12331 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
12332 verbose(env, "same insn cannot be used with different pointers\n");
12336 } else if (class == BPF_ST) {
12337 if (BPF_MODE(insn->code) != BPF_MEM ||
12338 insn->src_reg != BPF_REG_0) {
12339 verbose(env, "BPF_ST uses reserved fields\n");
12342 /* check src operand */
12343 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
12347 if (is_ctx_reg(env, insn->dst_reg)) {
12348 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
12350 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
12354 /* check that memory (dst_reg + off) is writeable */
12355 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
12356 insn->off, BPF_SIZE(insn->code),
12357 BPF_WRITE, -1, false);
12361 } else if (class == BPF_JMP || class == BPF_JMP32) {
12362 u8 opcode = BPF_OP(insn->code);
12364 env->jmps_processed++;
12365 if (opcode == BPF_CALL) {
12366 if (BPF_SRC(insn->code) != BPF_K ||
12367 (insn->src_reg != BPF_PSEUDO_KFUNC_CALL
12368 && insn->off != 0) ||
12369 (insn->src_reg != BPF_REG_0 &&
12370 insn->src_reg != BPF_PSEUDO_CALL &&
12371 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
12372 insn->dst_reg != BPF_REG_0 ||
12373 class == BPF_JMP32) {
12374 verbose(env, "BPF_CALL uses reserved fields\n");
12378 if (env->cur_state->active_spin_lock &&
12379 (insn->src_reg == BPF_PSEUDO_CALL ||
12380 insn->imm != BPF_FUNC_spin_unlock)) {
12381 verbose(env, "function calls are not allowed while holding a lock\n");
12384 if (insn->src_reg == BPF_PSEUDO_CALL)
12385 err = check_func_call(env, insn, &env->insn_idx);
12386 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
12387 err = check_kfunc_call(env, insn, &env->insn_idx);
12389 err = check_helper_call(env, insn, &env->insn_idx);
12392 } else if (opcode == BPF_JA) {
12393 if (BPF_SRC(insn->code) != BPF_K ||
12395 insn->src_reg != BPF_REG_0 ||
12396 insn->dst_reg != BPF_REG_0 ||
12397 class == BPF_JMP32) {
12398 verbose(env, "BPF_JA uses reserved fields\n");
12402 env->insn_idx += insn->off + 1;
12405 } else if (opcode == BPF_EXIT) {
12406 if (BPF_SRC(insn->code) != BPF_K ||
12408 insn->src_reg != BPF_REG_0 ||
12409 insn->dst_reg != BPF_REG_0 ||
12410 class == BPF_JMP32) {
12411 verbose(env, "BPF_EXIT uses reserved fields\n");
12415 if (env->cur_state->active_spin_lock) {
12416 verbose(env, "bpf_spin_unlock is missing\n");
12420 /* We must do check_reference_leak here before
12421 * prepare_func_exit to handle the case when
12422 * state->curframe > 0, it may be a callback
12423 * function, for which reference_state must
12424 * match caller reference state when it exits.
12426 err = check_reference_leak(env);
12430 if (state->curframe) {
12431 /* exit from nested function */
12432 err = prepare_func_exit(env, &env->insn_idx);
12435 do_print_state = true;
12439 err = check_return_code(env);
12443 mark_verifier_state_scratched(env);
12444 update_branch_counts(env, env->cur_state);
12445 err = pop_stack(env, &prev_insn_idx,
12446 &env->insn_idx, pop_log);
12448 if (err != -ENOENT)
12452 do_print_state = true;
12456 err = check_cond_jmp_op(env, insn, &env->insn_idx);
12460 } else if (class == BPF_LD) {
12461 u8 mode = BPF_MODE(insn->code);
12463 if (mode == BPF_ABS || mode == BPF_IND) {
12464 err = check_ld_abs(env, insn);
12468 } else if (mode == BPF_IMM) {
12469 err = check_ld_imm(env, insn);
12474 sanitize_mark_insn_seen(env);
12476 verbose(env, "invalid BPF_LD mode\n");
12480 verbose(env, "unknown insn class %d\n", class);
12490 static int find_btf_percpu_datasec(struct btf *btf)
12492 const struct btf_type *t;
12497 * Both vmlinux and module each have their own ".data..percpu"
12498 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
12499 * types to look at only module's own BTF types.
12501 n = btf_nr_types(btf);
12502 if (btf_is_module(btf))
12503 i = btf_nr_types(btf_vmlinux);
12507 for(; i < n; i++) {
12508 t = btf_type_by_id(btf, i);
12509 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
12512 tname = btf_name_by_offset(btf, t->name_off);
12513 if (!strcmp(tname, ".data..percpu"))
12520 /* replace pseudo btf_id with kernel symbol address */
12521 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
12522 struct bpf_insn *insn,
12523 struct bpf_insn_aux_data *aux)
12525 const struct btf_var_secinfo *vsi;
12526 const struct btf_type *datasec;
12527 struct btf_mod_pair *btf_mod;
12528 const struct btf_type *t;
12529 const char *sym_name;
12530 bool percpu = false;
12531 u32 type, id = insn->imm;
12535 int i, btf_fd, err;
12537 btf_fd = insn[1].imm;
12539 btf = btf_get_by_fd(btf_fd);
12541 verbose(env, "invalid module BTF object FD specified.\n");
12545 if (!btf_vmlinux) {
12546 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
12553 t = btf_type_by_id(btf, id);
12555 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
12560 if (!btf_type_is_var(t)) {
12561 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR.\n", id);
12566 sym_name = btf_name_by_offset(btf, t->name_off);
12567 addr = kallsyms_lookup_name(sym_name);
12569 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
12575 datasec_id = find_btf_percpu_datasec(btf);
12576 if (datasec_id > 0) {
12577 datasec = btf_type_by_id(btf, datasec_id);
12578 for_each_vsi(i, datasec, vsi) {
12579 if (vsi->type == id) {
12586 insn[0].imm = (u32)addr;
12587 insn[1].imm = addr >> 32;
12590 t = btf_type_skip_modifiers(btf, type, NULL);
12592 aux->btf_var.reg_type = PTR_TO_BTF_ID | MEM_PERCPU;
12593 aux->btf_var.btf = btf;
12594 aux->btf_var.btf_id = type;
12595 } else if (!btf_type_is_struct(t)) {
12596 const struct btf_type *ret;
12600 /* resolve the type size of ksym. */
12601 ret = btf_resolve_size(btf, t, &tsize);
12603 tname = btf_name_by_offset(btf, t->name_off);
12604 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
12605 tname, PTR_ERR(ret));
12609 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
12610 aux->btf_var.mem_size = tsize;
12612 aux->btf_var.reg_type = PTR_TO_BTF_ID;
12613 aux->btf_var.btf = btf;
12614 aux->btf_var.btf_id = type;
12617 /* check whether we recorded this BTF (and maybe module) already */
12618 for (i = 0; i < env->used_btf_cnt; i++) {
12619 if (env->used_btfs[i].btf == btf) {
12625 if (env->used_btf_cnt >= MAX_USED_BTFS) {
12630 btf_mod = &env->used_btfs[env->used_btf_cnt];
12631 btf_mod->btf = btf;
12632 btf_mod->module = NULL;
12634 /* if we reference variables from kernel module, bump its refcount */
12635 if (btf_is_module(btf)) {
12636 btf_mod->module = btf_try_get_module(btf);
12637 if (!btf_mod->module) {
12643 env->used_btf_cnt++;
12651 static bool is_tracing_prog_type(enum bpf_prog_type type)
12654 case BPF_PROG_TYPE_KPROBE:
12655 case BPF_PROG_TYPE_TRACEPOINT:
12656 case BPF_PROG_TYPE_PERF_EVENT:
12657 case BPF_PROG_TYPE_RAW_TRACEPOINT:
12658 case BPF_PROG_TYPE_RAW_TRACEPOINT_WRITABLE:
12665 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
12666 struct bpf_map *map,
12667 struct bpf_prog *prog)
12670 enum bpf_prog_type prog_type = resolve_prog_type(prog);
12672 if (map_value_has_spin_lock(map)) {
12673 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
12674 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
12678 if (is_tracing_prog_type(prog_type)) {
12679 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
12683 if (prog->aux->sleepable) {
12684 verbose(env, "sleepable progs cannot use bpf_spin_lock yet\n");
12689 if (map_value_has_timer(map)) {
12690 if (is_tracing_prog_type(prog_type)) {
12691 verbose(env, "tracing progs cannot use bpf_timer yet\n");
12696 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
12697 !bpf_offload_prog_map_match(prog, map)) {
12698 verbose(env, "offload device mismatch between prog and map\n");
12702 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
12703 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
12707 if (prog->aux->sleepable)
12708 switch (map->map_type) {
12709 case BPF_MAP_TYPE_HASH:
12710 case BPF_MAP_TYPE_LRU_HASH:
12711 case BPF_MAP_TYPE_ARRAY:
12712 case BPF_MAP_TYPE_PERCPU_HASH:
12713 case BPF_MAP_TYPE_PERCPU_ARRAY:
12714 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
12715 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
12716 case BPF_MAP_TYPE_HASH_OF_MAPS:
12717 case BPF_MAP_TYPE_RINGBUF:
12718 case BPF_MAP_TYPE_USER_RINGBUF:
12719 case BPF_MAP_TYPE_INODE_STORAGE:
12720 case BPF_MAP_TYPE_SK_STORAGE:
12721 case BPF_MAP_TYPE_TASK_STORAGE:
12725 "Sleepable programs can only use array, hash, and ringbuf maps\n");
12732 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
12734 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
12735 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
12738 /* find and rewrite pseudo imm in ld_imm64 instructions:
12740 * 1. if it accesses map FD, replace it with actual map pointer.
12741 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
12743 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
12745 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
12747 struct bpf_insn *insn = env->prog->insnsi;
12748 int insn_cnt = env->prog->len;
12751 err = bpf_prog_calc_tag(env->prog);
12755 for (i = 0; i < insn_cnt; i++, insn++) {
12756 if (BPF_CLASS(insn->code) == BPF_LDX &&
12757 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
12758 verbose(env, "BPF_LDX uses reserved fields\n");
12762 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
12763 struct bpf_insn_aux_data *aux;
12764 struct bpf_map *map;
12769 if (i == insn_cnt - 1 || insn[1].code != 0 ||
12770 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
12771 insn[1].off != 0) {
12772 verbose(env, "invalid bpf_ld_imm64 insn\n");
12776 if (insn[0].src_reg == 0)
12777 /* valid generic load 64-bit imm */
12780 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
12781 aux = &env->insn_aux_data[i];
12782 err = check_pseudo_btf_id(env, insn, aux);
12788 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
12789 aux = &env->insn_aux_data[i];
12790 aux->ptr_type = PTR_TO_FUNC;
12794 /* In final convert_pseudo_ld_imm64() step, this is
12795 * converted into regular 64-bit imm load insn.
12797 switch (insn[0].src_reg) {
12798 case BPF_PSEUDO_MAP_VALUE:
12799 case BPF_PSEUDO_MAP_IDX_VALUE:
12801 case BPF_PSEUDO_MAP_FD:
12802 case BPF_PSEUDO_MAP_IDX:
12803 if (insn[1].imm == 0)
12807 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
12811 switch (insn[0].src_reg) {
12812 case BPF_PSEUDO_MAP_IDX_VALUE:
12813 case BPF_PSEUDO_MAP_IDX:
12814 if (bpfptr_is_null(env->fd_array)) {
12815 verbose(env, "fd_idx without fd_array is invalid\n");
12818 if (copy_from_bpfptr_offset(&fd, env->fd_array,
12819 insn[0].imm * sizeof(fd),
12829 map = __bpf_map_get(f);
12831 verbose(env, "fd %d is not pointing to valid bpf_map\n",
12833 return PTR_ERR(map);
12836 err = check_map_prog_compatibility(env, map, env->prog);
12842 aux = &env->insn_aux_data[i];
12843 if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
12844 insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
12845 addr = (unsigned long)map;
12847 u32 off = insn[1].imm;
12849 if (off >= BPF_MAX_VAR_OFF) {
12850 verbose(env, "direct value offset of %u is not allowed\n", off);
12855 if (!map->ops->map_direct_value_addr) {
12856 verbose(env, "no direct value access support for this map type\n");
12861 err = map->ops->map_direct_value_addr(map, &addr, off);
12863 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
12864 map->value_size, off);
12869 aux->map_off = off;
12873 insn[0].imm = (u32)addr;
12874 insn[1].imm = addr >> 32;
12876 /* check whether we recorded this map already */
12877 for (j = 0; j < env->used_map_cnt; j++) {
12878 if (env->used_maps[j] == map) {
12879 aux->map_index = j;
12885 if (env->used_map_cnt >= MAX_USED_MAPS) {
12890 /* hold the map. If the program is rejected by verifier,
12891 * the map will be released by release_maps() or it
12892 * will be used by the valid program until it's unloaded
12893 * and all maps are released in free_used_maps()
12897 aux->map_index = env->used_map_cnt;
12898 env->used_maps[env->used_map_cnt++] = map;
12900 if (bpf_map_is_cgroup_storage(map) &&
12901 bpf_cgroup_storage_assign(env->prog->aux, map)) {
12902 verbose(env, "only one cgroup storage of each type is allowed\n");
12914 /* Basic sanity check before we invest more work here. */
12915 if (!bpf_opcode_in_insntable(insn->code)) {
12916 verbose(env, "unknown opcode %02x\n", insn->code);
12921 /* now all pseudo BPF_LD_IMM64 instructions load valid
12922 * 'struct bpf_map *' into a register instead of user map_fd.
12923 * These pointers will be used later by verifier to validate map access.
12928 /* drop refcnt of maps used by the rejected program */
12929 static void release_maps(struct bpf_verifier_env *env)
12931 __bpf_free_used_maps(env->prog->aux, env->used_maps,
12932 env->used_map_cnt);
12935 /* drop refcnt of maps used by the rejected program */
12936 static void release_btfs(struct bpf_verifier_env *env)
12938 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
12939 env->used_btf_cnt);
12942 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
12943 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
12945 struct bpf_insn *insn = env->prog->insnsi;
12946 int insn_cnt = env->prog->len;
12949 for (i = 0; i < insn_cnt; i++, insn++) {
12950 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
12952 if (insn->src_reg == BPF_PSEUDO_FUNC)
12958 /* single env->prog->insni[off] instruction was replaced with the range
12959 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
12960 * [0, off) and [off, end) to new locations, so the patched range stays zero
12962 static void adjust_insn_aux_data(struct bpf_verifier_env *env,
12963 struct bpf_insn_aux_data *new_data,
12964 struct bpf_prog *new_prog, u32 off, u32 cnt)
12966 struct bpf_insn_aux_data *old_data = env->insn_aux_data;
12967 struct bpf_insn *insn = new_prog->insnsi;
12968 u32 old_seen = old_data[off].seen;
12972 /* aux info at OFF always needs adjustment, no matter fast path
12973 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
12974 * original insn at old prog.
12976 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
12980 prog_len = new_prog->len;
12982 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
12983 memcpy(new_data + off + cnt - 1, old_data + off,
12984 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
12985 for (i = off; i < off + cnt - 1; i++) {
12986 /* Expand insni[off]'s seen count to the patched range. */
12987 new_data[i].seen = old_seen;
12988 new_data[i].zext_dst = insn_has_def32(env, insn + i);
12990 env->insn_aux_data = new_data;
12994 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
13000 /* NOTE: fake 'exit' subprog should be updated as well. */
13001 for (i = 0; i <= env->subprog_cnt; i++) {
13002 if (env->subprog_info[i].start <= off)
13004 env->subprog_info[i].start += len - 1;
13008 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
13010 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
13011 int i, sz = prog->aux->size_poke_tab;
13012 struct bpf_jit_poke_descriptor *desc;
13014 for (i = 0; i < sz; i++) {
13016 if (desc->insn_idx <= off)
13018 desc->insn_idx += len - 1;
13022 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
13023 const struct bpf_insn *patch, u32 len)
13025 struct bpf_prog *new_prog;
13026 struct bpf_insn_aux_data *new_data = NULL;
13029 new_data = vzalloc(array_size(env->prog->len + len - 1,
13030 sizeof(struct bpf_insn_aux_data)));
13035 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
13036 if (IS_ERR(new_prog)) {
13037 if (PTR_ERR(new_prog) == -ERANGE)
13039 "insn %d cannot be patched due to 16-bit range\n",
13040 env->insn_aux_data[off].orig_idx);
13044 adjust_insn_aux_data(env, new_data, new_prog, off, len);
13045 adjust_subprog_starts(env, off, len);
13046 adjust_poke_descs(new_prog, off, len);
13050 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
13055 /* find first prog starting at or after off (first to remove) */
13056 for (i = 0; i < env->subprog_cnt; i++)
13057 if (env->subprog_info[i].start >= off)
13059 /* find first prog starting at or after off + cnt (first to stay) */
13060 for (j = i; j < env->subprog_cnt; j++)
13061 if (env->subprog_info[j].start >= off + cnt)
13063 /* if j doesn't start exactly at off + cnt, we are just removing
13064 * the front of previous prog
13066 if (env->subprog_info[j].start != off + cnt)
13070 struct bpf_prog_aux *aux = env->prog->aux;
13073 /* move fake 'exit' subprog as well */
13074 move = env->subprog_cnt + 1 - j;
13076 memmove(env->subprog_info + i,
13077 env->subprog_info + j,
13078 sizeof(*env->subprog_info) * move);
13079 env->subprog_cnt -= j - i;
13081 /* remove func_info */
13082 if (aux->func_info) {
13083 move = aux->func_info_cnt - j;
13085 memmove(aux->func_info + i,
13086 aux->func_info + j,
13087 sizeof(*aux->func_info) * move);
13088 aux->func_info_cnt -= j - i;
13089 /* func_info->insn_off is set after all code rewrites,
13090 * in adjust_btf_func() - no need to adjust
13094 /* convert i from "first prog to remove" to "first to adjust" */
13095 if (env->subprog_info[i].start == off)
13099 /* update fake 'exit' subprog as well */
13100 for (; i <= env->subprog_cnt; i++)
13101 env->subprog_info[i].start -= cnt;
13106 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
13109 struct bpf_prog *prog = env->prog;
13110 u32 i, l_off, l_cnt, nr_linfo;
13111 struct bpf_line_info *linfo;
13113 nr_linfo = prog->aux->nr_linfo;
13117 linfo = prog->aux->linfo;
13119 /* find first line info to remove, count lines to be removed */
13120 for (i = 0; i < nr_linfo; i++)
13121 if (linfo[i].insn_off >= off)
13126 for (; i < nr_linfo; i++)
13127 if (linfo[i].insn_off < off + cnt)
13132 /* First live insn doesn't match first live linfo, it needs to "inherit"
13133 * last removed linfo. prog is already modified, so prog->len == off
13134 * means no live instructions after (tail of the program was removed).
13136 if (prog->len != off && l_cnt &&
13137 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
13139 linfo[--i].insn_off = off + cnt;
13142 /* remove the line info which refer to the removed instructions */
13144 memmove(linfo + l_off, linfo + i,
13145 sizeof(*linfo) * (nr_linfo - i));
13147 prog->aux->nr_linfo -= l_cnt;
13148 nr_linfo = prog->aux->nr_linfo;
13151 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
13152 for (i = l_off; i < nr_linfo; i++)
13153 linfo[i].insn_off -= cnt;
13155 /* fix up all subprogs (incl. 'exit') which start >= off */
13156 for (i = 0; i <= env->subprog_cnt; i++)
13157 if (env->subprog_info[i].linfo_idx > l_off) {
13158 /* program may have started in the removed region but
13159 * may not be fully removed
13161 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
13162 env->subprog_info[i].linfo_idx -= l_cnt;
13164 env->subprog_info[i].linfo_idx = l_off;
13170 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
13172 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
13173 unsigned int orig_prog_len = env->prog->len;
13176 if (bpf_prog_is_dev_bound(env->prog->aux))
13177 bpf_prog_offload_remove_insns(env, off, cnt);
13179 err = bpf_remove_insns(env->prog, off, cnt);
13183 err = adjust_subprog_starts_after_remove(env, off, cnt);
13187 err = bpf_adj_linfo_after_remove(env, off, cnt);
13191 memmove(aux_data + off, aux_data + off + cnt,
13192 sizeof(*aux_data) * (orig_prog_len - off - cnt));
13197 /* The verifier does more data flow analysis than llvm and will not
13198 * explore branches that are dead at run time. Malicious programs can
13199 * have dead code too. Therefore replace all dead at-run-time code
13202 * Just nops are not optimal, e.g. if they would sit at the end of the
13203 * program and through another bug we would manage to jump there, then
13204 * we'd execute beyond program memory otherwise. Returning exception
13205 * code also wouldn't work since we can have subprogs where the dead
13206 * code could be located.
13208 static void sanitize_dead_code(struct bpf_verifier_env *env)
13210 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
13211 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
13212 struct bpf_insn *insn = env->prog->insnsi;
13213 const int insn_cnt = env->prog->len;
13216 for (i = 0; i < insn_cnt; i++) {
13217 if (aux_data[i].seen)
13219 memcpy(insn + i, &trap, sizeof(trap));
13220 aux_data[i].zext_dst = false;
13224 static bool insn_is_cond_jump(u8 code)
13228 if (BPF_CLASS(code) == BPF_JMP32)
13231 if (BPF_CLASS(code) != BPF_JMP)
13235 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
13238 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
13240 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
13241 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
13242 struct bpf_insn *insn = env->prog->insnsi;
13243 const int insn_cnt = env->prog->len;
13246 for (i = 0; i < insn_cnt; i++, insn++) {
13247 if (!insn_is_cond_jump(insn->code))
13250 if (!aux_data[i + 1].seen)
13251 ja.off = insn->off;
13252 else if (!aux_data[i + 1 + insn->off].seen)
13257 if (bpf_prog_is_dev_bound(env->prog->aux))
13258 bpf_prog_offload_replace_insn(env, i, &ja);
13260 memcpy(insn, &ja, sizeof(ja));
13264 static int opt_remove_dead_code(struct bpf_verifier_env *env)
13266 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
13267 int insn_cnt = env->prog->len;
13270 for (i = 0; i < insn_cnt; i++) {
13274 while (i + j < insn_cnt && !aux_data[i + j].seen)
13279 err = verifier_remove_insns(env, i, j);
13282 insn_cnt = env->prog->len;
13288 static int opt_remove_nops(struct bpf_verifier_env *env)
13290 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
13291 struct bpf_insn *insn = env->prog->insnsi;
13292 int insn_cnt = env->prog->len;
13295 for (i = 0; i < insn_cnt; i++) {
13296 if (memcmp(&insn[i], &ja, sizeof(ja)))
13299 err = verifier_remove_insns(env, i, 1);
13309 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
13310 const union bpf_attr *attr)
13312 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
13313 struct bpf_insn_aux_data *aux = env->insn_aux_data;
13314 int i, patch_len, delta = 0, len = env->prog->len;
13315 struct bpf_insn *insns = env->prog->insnsi;
13316 struct bpf_prog *new_prog;
13319 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
13320 zext_patch[1] = BPF_ZEXT_REG(0);
13321 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
13322 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
13323 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
13324 for (i = 0; i < len; i++) {
13325 int adj_idx = i + delta;
13326 struct bpf_insn insn;
13329 insn = insns[adj_idx];
13330 load_reg = insn_def_regno(&insn);
13331 if (!aux[adj_idx].zext_dst) {
13339 class = BPF_CLASS(code);
13340 if (load_reg == -1)
13343 /* NOTE: arg "reg" (the fourth one) is only used for
13344 * BPF_STX + SRC_OP, so it is safe to pass NULL
13347 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
13348 if (class == BPF_LD &&
13349 BPF_MODE(code) == BPF_IMM)
13354 /* ctx load could be transformed into wider load. */
13355 if (class == BPF_LDX &&
13356 aux[adj_idx].ptr_type == PTR_TO_CTX)
13359 imm_rnd = get_random_u32();
13360 rnd_hi32_patch[0] = insn;
13361 rnd_hi32_patch[1].imm = imm_rnd;
13362 rnd_hi32_patch[3].dst_reg = load_reg;
13363 patch = rnd_hi32_patch;
13365 goto apply_patch_buffer;
13368 /* Add in an zero-extend instruction if a) the JIT has requested
13369 * it or b) it's a CMPXCHG.
13371 * The latter is because: BPF_CMPXCHG always loads a value into
13372 * R0, therefore always zero-extends. However some archs'
13373 * equivalent instruction only does this load when the
13374 * comparison is successful. This detail of CMPXCHG is
13375 * orthogonal to the general zero-extension behaviour of the
13376 * CPU, so it's treated independently of bpf_jit_needs_zext.
13378 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
13381 if (WARN_ON(load_reg == -1)) {
13382 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
13386 zext_patch[0] = insn;
13387 zext_patch[1].dst_reg = load_reg;
13388 zext_patch[1].src_reg = load_reg;
13389 patch = zext_patch;
13391 apply_patch_buffer:
13392 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
13395 env->prog = new_prog;
13396 insns = new_prog->insnsi;
13397 aux = env->insn_aux_data;
13398 delta += patch_len - 1;
13404 /* convert load instructions that access fields of a context type into a
13405 * sequence of instructions that access fields of the underlying structure:
13406 * struct __sk_buff -> struct sk_buff
13407 * struct bpf_sock_ops -> struct sock
13409 static int convert_ctx_accesses(struct bpf_verifier_env *env)
13411 const struct bpf_verifier_ops *ops = env->ops;
13412 int i, cnt, size, ctx_field_size, delta = 0;
13413 const int insn_cnt = env->prog->len;
13414 struct bpf_insn insn_buf[16], *insn;
13415 u32 target_size, size_default, off;
13416 struct bpf_prog *new_prog;
13417 enum bpf_access_type type;
13418 bool is_narrower_load;
13420 if (ops->gen_prologue || env->seen_direct_write) {
13421 if (!ops->gen_prologue) {
13422 verbose(env, "bpf verifier is misconfigured\n");
13425 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
13427 if (cnt >= ARRAY_SIZE(insn_buf)) {
13428 verbose(env, "bpf verifier is misconfigured\n");
13431 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
13435 env->prog = new_prog;
13440 if (bpf_prog_is_dev_bound(env->prog->aux))
13443 insn = env->prog->insnsi + delta;
13445 for (i = 0; i < insn_cnt; i++, insn++) {
13446 bpf_convert_ctx_access_t convert_ctx_access;
13449 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
13450 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
13451 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
13452 insn->code == (BPF_LDX | BPF_MEM | BPF_DW)) {
13455 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
13456 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
13457 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
13458 insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
13459 insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
13460 insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
13461 insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
13462 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
13464 ctx_access = BPF_CLASS(insn->code) == BPF_STX;
13469 if (type == BPF_WRITE &&
13470 env->insn_aux_data[i + delta].sanitize_stack_spill) {
13471 struct bpf_insn patch[] = {
13476 cnt = ARRAY_SIZE(patch);
13477 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
13482 env->prog = new_prog;
13483 insn = new_prog->insnsi + i + delta;
13490 switch ((int)env->insn_aux_data[i + delta].ptr_type) {
13492 if (!ops->convert_ctx_access)
13494 convert_ctx_access = ops->convert_ctx_access;
13496 case PTR_TO_SOCKET:
13497 case PTR_TO_SOCK_COMMON:
13498 convert_ctx_access = bpf_sock_convert_ctx_access;
13500 case PTR_TO_TCP_SOCK:
13501 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
13503 case PTR_TO_XDP_SOCK:
13504 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
13506 case PTR_TO_BTF_ID:
13507 case PTR_TO_BTF_ID | PTR_UNTRUSTED:
13508 if (type == BPF_READ) {
13509 insn->code = BPF_LDX | BPF_PROBE_MEM |
13510 BPF_SIZE((insn)->code);
13511 env->prog->aux->num_exentries++;
13518 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
13519 size = BPF_LDST_BYTES(insn);
13521 /* If the read access is a narrower load of the field,
13522 * convert to a 4/8-byte load, to minimum program type specific
13523 * convert_ctx_access changes. If conversion is successful,
13524 * we will apply proper mask to the result.
13526 is_narrower_load = size < ctx_field_size;
13527 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
13529 if (is_narrower_load) {
13532 if (type == BPF_WRITE) {
13533 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
13538 if (ctx_field_size == 4)
13540 else if (ctx_field_size == 8)
13541 size_code = BPF_DW;
13543 insn->off = off & ~(size_default - 1);
13544 insn->code = BPF_LDX | BPF_MEM | size_code;
13548 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
13550 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
13551 (ctx_field_size && !target_size)) {
13552 verbose(env, "bpf verifier is misconfigured\n");
13556 if (is_narrower_load && size < target_size) {
13557 u8 shift = bpf_ctx_narrow_access_offset(
13558 off, size, size_default) * 8;
13559 if (shift && cnt + 1 >= ARRAY_SIZE(insn_buf)) {
13560 verbose(env, "bpf verifier narrow ctx load misconfigured\n");
13563 if (ctx_field_size <= 4) {
13565 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
13568 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
13569 (1 << size * 8) - 1);
13572 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
13575 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
13576 (1ULL << size * 8) - 1);
13580 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
13586 /* keep walking new program and skip insns we just inserted */
13587 env->prog = new_prog;
13588 insn = new_prog->insnsi + i + delta;
13594 static int jit_subprogs(struct bpf_verifier_env *env)
13596 struct bpf_prog *prog = env->prog, **func, *tmp;
13597 int i, j, subprog_start, subprog_end = 0, len, subprog;
13598 struct bpf_map *map_ptr;
13599 struct bpf_insn *insn;
13600 void *old_bpf_func;
13601 int err, num_exentries;
13603 if (env->subprog_cnt <= 1)
13606 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
13607 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn))
13610 /* Upon error here we cannot fall back to interpreter but
13611 * need a hard reject of the program. Thus -EFAULT is
13612 * propagated in any case.
13614 subprog = find_subprog(env, i + insn->imm + 1);
13616 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
13617 i + insn->imm + 1);
13620 /* temporarily remember subprog id inside insn instead of
13621 * aux_data, since next loop will split up all insns into funcs
13623 insn->off = subprog;
13624 /* remember original imm in case JIT fails and fallback
13625 * to interpreter will be needed
13627 env->insn_aux_data[i].call_imm = insn->imm;
13628 /* point imm to __bpf_call_base+1 from JITs point of view */
13630 if (bpf_pseudo_func(insn))
13631 /* jit (e.g. x86_64) may emit fewer instructions
13632 * if it learns a u32 imm is the same as a u64 imm.
13633 * Force a non zero here.
13638 err = bpf_prog_alloc_jited_linfo(prog);
13640 goto out_undo_insn;
13643 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
13645 goto out_undo_insn;
13647 for (i = 0; i < env->subprog_cnt; i++) {
13648 subprog_start = subprog_end;
13649 subprog_end = env->subprog_info[i + 1].start;
13651 len = subprog_end - subprog_start;
13652 /* bpf_prog_run() doesn't call subprogs directly,
13653 * hence main prog stats include the runtime of subprogs.
13654 * subprogs don't have IDs and not reachable via prog_get_next_id
13655 * func[i]->stats will never be accessed and stays NULL
13657 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
13660 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
13661 len * sizeof(struct bpf_insn));
13662 func[i]->type = prog->type;
13663 func[i]->len = len;
13664 if (bpf_prog_calc_tag(func[i]))
13666 func[i]->is_func = 1;
13667 func[i]->aux->func_idx = i;
13668 /* Below members will be freed only at prog->aux */
13669 func[i]->aux->btf = prog->aux->btf;
13670 func[i]->aux->func_info = prog->aux->func_info;
13671 func[i]->aux->func_info_cnt = prog->aux->func_info_cnt;
13672 func[i]->aux->poke_tab = prog->aux->poke_tab;
13673 func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
13675 for (j = 0; j < prog->aux->size_poke_tab; j++) {
13676 struct bpf_jit_poke_descriptor *poke;
13678 poke = &prog->aux->poke_tab[j];
13679 if (poke->insn_idx < subprog_end &&
13680 poke->insn_idx >= subprog_start)
13681 poke->aux = func[i]->aux;
13684 func[i]->aux->name[0] = 'F';
13685 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
13686 func[i]->jit_requested = 1;
13687 func[i]->blinding_requested = prog->blinding_requested;
13688 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
13689 func[i]->aux->kfunc_btf_tab = prog->aux->kfunc_btf_tab;
13690 func[i]->aux->linfo = prog->aux->linfo;
13691 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
13692 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
13693 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
13695 insn = func[i]->insnsi;
13696 for (j = 0; j < func[i]->len; j++, insn++) {
13697 if (BPF_CLASS(insn->code) == BPF_LDX &&
13698 BPF_MODE(insn->code) == BPF_PROBE_MEM)
13701 func[i]->aux->num_exentries = num_exentries;
13702 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
13703 func[i] = bpf_int_jit_compile(func[i]);
13704 if (!func[i]->jited) {
13711 /* at this point all bpf functions were successfully JITed
13712 * now populate all bpf_calls with correct addresses and
13713 * run last pass of JIT
13715 for (i = 0; i < env->subprog_cnt; i++) {
13716 insn = func[i]->insnsi;
13717 for (j = 0; j < func[i]->len; j++, insn++) {
13718 if (bpf_pseudo_func(insn)) {
13719 subprog = insn->off;
13720 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
13721 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
13724 if (!bpf_pseudo_call(insn))
13726 subprog = insn->off;
13727 insn->imm = BPF_CALL_IMM(func[subprog]->bpf_func);
13730 /* we use the aux data to keep a list of the start addresses
13731 * of the JITed images for each function in the program
13733 * for some architectures, such as powerpc64, the imm field
13734 * might not be large enough to hold the offset of the start
13735 * address of the callee's JITed image from __bpf_call_base
13737 * in such cases, we can lookup the start address of a callee
13738 * by using its subprog id, available from the off field of
13739 * the call instruction, as an index for this list
13741 func[i]->aux->func = func;
13742 func[i]->aux->func_cnt = env->subprog_cnt;
13744 for (i = 0; i < env->subprog_cnt; i++) {
13745 old_bpf_func = func[i]->bpf_func;
13746 tmp = bpf_int_jit_compile(func[i]);
13747 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
13748 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
13755 /* finally lock prog and jit images for all functions and
13756 * populate kallsysm
13758 for (i = 0; i < env->subprog_cnt; i++) {
13759 bpf_prog_lock_ro(func[i]);
13760 bpf_prog_kallsyms_add(func[i]);
13763 /* Last step: make now unused interpreter insns from main
13764 * prog consistent for later dump requests, so they can
13765 * later look the same as if they were interpreted only.
13767 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
13768 if (bpf_pseudo_func(insn)) {
13769 insn[0].imm = env->insn_aux_data[i].call_imm;
13770 insn[1].imm = insn->off;
13774 if (!bpf_pseudo_call(insn))
13776 insn->off = env->insn_aux_data[i].call_imm;
13777 subprog = find_subprog(env, i + insn->off + 1);
13778 insn->imm = subprog;
13782 prog->bpf_func = func[0]->bpf_func;
13783 prog->jited_len = func[0]->jited_len;
13784 prog->aux->func = func;
13785 prog->aux->func_cnt = env->subprog_cnt;
13786 bpf_prog_jit_attempt_done(prog);
13789 /* We failed JIT'ing, so at this point we need to unregister poke
13790 * descriptors from subprogs, so that kernel is not attempting to
13791 * patch it anymore as we're freeing the subprog JIT memory.
13793 for (i = 0; i < prog->aux->size_poke_tab; i++) {
13794 map_ptr = prog->aux->poke_tab[i].tail_call.map;
13795 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
13797 /* At this point we're guaranteed that poke descriptors are not
13798 * live anymore. We can just unlink its descriptor table as it's
13799 * released with the main prog.
13801 for (i = 0; i < env->subprog_cnt; i++) {
13804 func[i]->aux->poke_tab = NULL;
13805 bpf_jit_free(func[i]);
13809 /* cleanup main prog to be interpreted */
13810 prog->jit_requested = 0;
13811 prog->blinding_requested = 0;
13812 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
13813 if (!bpf_pseudo_call(insn))
13816 insn->imm = env->insn_aux_data[i].call_imm;
13818 bpf_prog_jit_attempt_done(prog);
13822 static int fixup_call_args(struct bpf_verifier_env *env)
13824 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
13825 struct bpf_prog *prog = env->prog;
13826 struct bpf_insn *insn = prog->insnsi;
13827 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
13832 if (env->prog->jit_requested &&
13833 !bpf_prog_is_dev_bound(env->prog->aux)) {
13834 err = jit_subprogs(env);
13837 if (err == -EFAULT)
13840 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
13841 if (has_kfunc_call) {
13842 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
13845 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
13846 /* When JIT fails the progs with bpf2bpf calls and tail_calls
13847 * have to be rejected, since interpreter doesn't support them yet.
13849 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
13852 for (i = 0; i < prog->len; i++, insn++) {
13853 if (bpf_pseudo_func(insn)) {
13854 /* When JIT fails the progs with callback calls
13855 * have to be rejected, since interpreter doesn't support them yet.
13857 verbose(env, "callbacks are not allowed in non-JITed programs\n");
13861 if (!bpf_pseudo_call(insn))
13863 depth = get_callee_stack_depth(env, insn, i);
13866 bpf_patch_call_args(insn, depth);
13873 static int fixup_kfunc_call(struct bpf_verifier_env *env,
13874 struct bpf_insn *insn)
13876 const struct bpf_kfunc_desc *desc;
13879 verbose(env, "invalid kernel function call not eliminated in verifier pass\n");
13883 /* insn->imm has the btf func_id. Replace it with
13884 * an address (relative to __bpf_base_call).
13886 desc = find_kfunc_desc(env->prog, insn->imm, insn->off);
13888 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
13893 insn->imm = desc->imm;
13898 /* Do various post-verification rewrites in a single program pass.
13899 * These rewrites simplify JIT and interpreter implementations.
13901 static int do_misc_fixups(struct bpf_verifier_env *env)
13903 struct bpf_prog *prog = env->prog;
13904 enum bpf_attach_type eatype = prog->expected_attach_type;
13905 enum bpf_prog_type prog_type = resolve_prog_type(prog);
13906 struct bpf_insn *insn = prog->insnsi;
13907 const struct bpf_func_proto *fn;
13908 const int insn_cnt = prog->len;
13909 const struct bpf_map_ops *ops;
13910 struct bpf_insn_aux_data *aux;
13911 struct bpf_insn insn_buf[16];
13912 struct bpf_prog *new_prog;
13913 struct bpf_map *map_ptr;
13914 int i, ret, cnt, delta = 0;
13916 for (i = 0; i < insn_cnt; i++, insn++) {
13917 /* Make divide-by-zero exceptions impossible. */
13918 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
13919 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
13920 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
13921 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
13922 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
13923 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
13924 struct bpf_insn *patchlet;
13925 struct bpf_insn chk_and_div[] = {
13926 /* [R,W]x div 0 -> 0 */
13927 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
13928 BPF_JNE | BPF_K, insn->src_reg,
13930 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
13931 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
13934 struct bpf_insn chk_and_mod[] = {
13935 /* [R,W]x mod 0 -> [R,W]x */
13936 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
13937 BPF_JEQ | BPF_K, insn->src_reg,
13938 0, 1 + (is64 ? 0 : 1), 0),
13940 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
13941 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
13944 patchlet = isdiv ? chk_and_div : chk_and_mod;
13945 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
13946 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
13948 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
13953 env->prog = prog = new_prog;
13954 insn = new_prog->insnsi + i + delta;
13958 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
13959 if (BPF_CLASS(insn->code) == BPF_LD &&
13960 (BPF_MODE(insn->code) == BPF_ABS ||
13961 BPF_MODE(insn->code) == BPF_IND)) {
13962 cnt = env->ops->gen_ld_abs(insn, insn_buf);
13963 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
13964 verbose(env, "bpf verifier is misconfigured\n");
13968 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
13973 env->prog = prog = new_prog;
13974 insn = new_prog->insnsi + i + delta;
13978 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
13979 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
13980 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
13981 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
13982 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
13983 struct bpf_insn *patch = &insn_buf[0];
13984 bool issrc, isneg, isimm;
13987 aux = &env->insn_aux_data[i + delta];
13988 if (!aux->alu_state ||
13989 aux->alu_state == BPF_ALU_NON_POINTER)
13992 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
13993 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
13994 BPF_ALU_SANITIZE_SRC;
13995 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
13997 off_reg = issrc ? insn->src_reg : insn->dst_reg;
13999 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
14002 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
14003 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
14004 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
14005 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
14006 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
14007 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
14008 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
14011 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
14012 insn->src_reg = BPF_REG_AX;
14014 insn->code = insn->code == code_add ?
14015 code_sub : code_add;
14017 if (issrc && isneg && !isimm)
14018 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
14019 cnt = patch - insn_buf;
14021 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
14026 env->prog = prog = new_prog;
14027 insn = new_prog->insnsi + i + delta;
14031 if (insn->code != (BPF_JMP | BPF_CALL))
14033 if (insn->src_reg == BPF_PSEUDO_CALL)
14035 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
14036 ret = fixup_kfunc_call(env, insn);
14042 if (insn->imm == BPF_FUNC_get_route_realm)
14043 prog->dst_needed = 1;
14044 if (insn->imm == BPF_FUNC_get_prandom_u32)
14045 bpf_user_rnd_init_once();
14046 if (insn->imm == BPF_FUNC_override_return)
14047 prog->kprobe_override = 1;
14048 if (insn->imm == BPF_FUNC_tail_call) {
14049 /* If we tail call into other programs, we
14050 * cannot make any assumptions since they can
14051 * be replaced dynamically during runtime in
14052 * the program array.
14054 prog->cb_access = 1;
14055 if (!allow_tail_call_in_subprogs(env))
14056 prog->aux->stack_depth = MAX_BPF_STACK;
14057 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
14059 /* mark bpf_tail_call as different opcode to avoid
14060 * conditional branch in the interpreter for every normal
14061 * call and to prevent accidental JITing by JIT compiler
14062 * that doesn't support bpf_tail_call yet
14065 insn->code = BPF_JMP | BPF_TAIL_CALL;
14067 aux = &env->insn_aux_data[i + delta];
14068 if (env->bpf_capable && !prog->blinding_requested &&
14069 prog->jit_requested &&
14070 !bpf_map_key_poisoned(aux) &&
14071 !bpf_map_ptr_poisoned(aux) &&
14072 !bpf_map_ptr_unpriv(aux)) {
14073 struct bpf_jit_poke_descriptor desc = {
14074 .reason = BPF_POKE_REASON_TAIL_CALL,
14075 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
14076 .tail_call.key = bpf_map_key_immediate(aux),
14077 .insn_idx = i + delta,
14080 ret = bpf_jit_add_poke_descriptor(prog, &desc);
14082 verbose(env, "adding tail call poke descriptor failed\n");
14086 insn->imm = ret + 1;
14090 if (!bpf_map_ptr_unpriv(aux))
14093 /* instead of changing every JIT dealing with tail_call
14094 * emit two extra insns:
14095 * if (index >= max_entries) goto out;
14096 * index &= array->index_mask;
14097 * to avoid out-of-bounds cpu speculation
14099 if (bpf_map_ptr_poisoned(aux)) {
14100 verbose(env, "tail_call abusing map_ptr\n");
14104 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
14105 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
14106 map_ptr->max_entries, 2);
14107 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
14108 container_of(map_ptr,
14111 insn_buf[2] = *insn;
14113 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
14118 env->prog = prog = new_prog;
14119 insn = new_prog->insnsi + i + delta;
14123 if (insn->imm == BPF_FUNC_timer_set_callback) {
14124 /* The verifier will process callback_fn as many times as necessary
14125 * with different maps and the register states prepared by
14126 * set_timer_callback_state will be accurate.
14128 * The following use case is valid:
14129 * map1 is shared by prog1, prog2, prog3.
14130 * prog1 calls bpf_timer_init for some map1 elements
14131 * prog2 calls bpf_timer_set_callback for some map1 elements.
14132 * Those that were not bpf_timer_init-ed will return -EINVAL.
14133 * prog3 calls bpf_timer_start for some map1 elements.
14134 * Those that were not both bpf_timer_init-ed and
14135 * bpf_timer_set_callback-ed will return -EINVAL.
14137 struct bpf_insn ld_addrs[2] = {
14138 BPF_LD_IMM64(BPF_REG_3, (long)prog->aux),
14141 insn_buf[0] = ld_addrs[0];
14142 insn_buf[1] = ld_addrs[1];
14143 insn_buf[2] = *insn;
14146 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
14151 env->prog = prog = new_prog;
14152 insn = new_prog->insnsi + i + delta;
14153 goto patch_call_imm;
14156 if (insn->imm == BPF_FUNC_task_storage_get ||
14157 insn->imm == BPF_FUNC_sk_storage_get ||
14158 insn->imm == BPF_FUNC_inode_storage_get) {
14159 if (env->prog->aux->sleepable)
14160 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_KERNEL);
14162 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_ATOMIC);
14163 insn_buf[1] = *insn;
14166 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
14171 env->prog = prog = new_prog;
14172 insn = new_prog->insnsi + i + delta;
14173 goto patch_call_imm;
14176 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
14177 * and other inlining handlers are currently limited to 64 bit
14180 if (prog->jit_requested && BITS_PER_LONG == 64 &&
14181 (insn->imm == BPF_FUNC_map_lookup_elem ||
14182 insn->imm == BPF_FUNC_map_update_elem ||
14183 insn->imm == BPF_FUNC_map_delete_elem ||
14184 insn->imm == BPF_FUNC_map_push_elem ||
14185 insn->imm == BPF_FUNC_map_pop_elem ||
14186 insn->imm == BPF_FUNC_map_peek_elem ||
14187 insn->imm == BPF_FUNC_redirect_map ||
14188 insn->imm == BPF_FUNC_for_each_map_elem ||
14189 insn->imm == BPF_FUNC_map_lookup_percpu_elem)) {
14190 aux = &env->insn_aux_data[i + delta];
14191 if (bpf_map_ptr_poisoned(aux))
14192 goto patch_call_imm;
14194 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
14195 ops = map_ptr->ops;
14196 if (insn->imm == BPF_FUNC_map_lookup_elem &&
14197 ops->map_gen_lookup) {
14198 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
14199 if (cnt == -EOPNOTSUPP)
14200 goto patch_map_ops_generic;
14201 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
14202 verbose(env, "bpf verifier is misconfigured\n");
14206 new_prog = bpf_patch_insn_data(env, i + delta,
14212 env->prog = prog = new_prog;
14213 insn = new_prog->insnsi + i + delta;
14217 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
14218 (void *(*)(struct bpf_map *map, void *key))NULL));
14219 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
14220 (int (*)(struct bpf_map *map, void *key))NULL));
14221 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
14222 (int (*)(struct bpf_map *map, void *key, void *value,
14224 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
14225 (int (*)(struct bpf_map *map, void *value,
14227 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
14228 (int (*)(struct bpf_map *map, void *value))NULL));
14229 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
14230 (int (*)(struct bpf_map *map, void *value))NULL));
14231 BUILD_BUG_ON(!__same_type(ops->map_redirect,
14232 (int (*)(struct bpf_map *map, u32 ifindex, u64 flags))NULL));
14233 BUILD_BUG_ON(!__same_type(ops->map_for_each_callback,
14234 (int (*)(struct bpf_map *map,
14235 bpf_callback_t callback_fn,
14236 void *callback_ctx,
14238 BUILD_BUG_ON(!__same_type(ops->map_lookup_percpu_elem,
14239 (void *(*)(struct bpf_map *map, void *key, u32 cpu))NULL));
14241 patch_map_ops_generic:
14242 switch (insn->imm) {
14243 case BPF_FUNC_map_lookup_elem:
14244 insn->imm = BPF_CALL_IMM(ops->map_lookup_elem);
14246 case BPF_FUNC_map_update_elem:
14247 insn->imm = BPF_CALL_IMM(ops->map_update_elem);
14249 case BPF_FUNC_map_delete_elem:
14250 insn->imm = BPF_CALL_IMM(ops->map_delete_elem);
14252 case BPF_FUNC_map_push_elem:
14253 insn->imm = BPF_CALL_IMM(ops->map_push_elem);
14255 case BPF_FUNC_map_pop_elem:
14256 insn->imm = BPF_CALL_IMM(ops->map_pop_elem);
14258 case BPF_FUNC_map_peek_elem:
14259 insn->imm = BPF_CALL_IMM(ops->map_peek_elem);
14261 case BPF_FUNC_redirect_map:
14262 insn->imm = BPF_CALL_IMM(ops->map_redirect);
14264 case BPF_FUNC_for_each_map_elem:
14265 insn->imm = BPF_CALL_IMM(ops->map_for_each_callback);
14267 case BPF_FUNC_map_lookup_percpu_elem:
14268 insn->imm = BPF_CALL_IMM(ops->map_lookup_percpu_elem);
14272 goto patch_call_imm;
14275 /* Implement bpf_jiffies64 inline. */
14276 if (prog->jit_requested && BITS_PER_LONG == 64 &&
14277 insn->imm == BPF_FUNC_jiffies64) {
14278 struct bpf_insn ld_jiffies_addr[2] = {
14279 BPF_LD_IMM64(BPF_REG_0,
14280 (unsigned long)&jiffies),
14283 insn_buf[0] = ld_jiffies_addr[0];
14284 insn_buf[1] = ld_jiffies_addr[1];
14285 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
14289 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
14295 env->prog = prog = new_prog;
14296 insn = new_prog->insnsi + i + delta;
14300 /* Implement bpf_get_func_arg inline. */
14301 if (prog_type == BPF_PROG_TYPE_TRACING &&
14302 insn->imm == BPF_FUNC_get_func_arg) {
14303 /* Load nr_args from ctx - 8 */
14304 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
14305 insn_buf[1] = BPF_JMP32_REG(BPF_JGE, BPF_REG_2, BPF_REG_0, 6);
14306 insn_buf[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_2, 3);
14307 insn_buf[3] = BPF_ALU64_REG(BPF_ADD, BPF_REG_2, BPF_REG_1);
14308 insn_buf[4] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_2, 0);
14309 insn_buf[5] = BPF_STX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
14310 insn_buf[6] = BPF_MOV64_IMM(BPF_REG_0, 0);
14311 insn_buf[7] = BPF_JMP_A(1);
14312 insn_buf[8] = BPF_MOV64_IMM(BPF_REG_0, -EINVAL);
14315 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
14320 env->prog = prog = new_prog;
14321 insn = new_prog->insnsi + i + delta;
14325 /* Implement bpf_get_func_ret inline. */
14326 if (prog_type == BPF_PROG_TYPE_TRACING &&
14327 insn->imm == BPF_FUNC_get_func_ret) {
14328 if (eatype == BPF_TRACE_FEXIT ||
14329 eatype == BPF_MODIFY_RETURN) {
14330 /* Load nr_args from ctx - 8 */
14331 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
14332 insn_buf[1] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_0, 3);
14333 insn_buf[2] = BPF_ALU64_REG(BPF_ADD, BPF_REG_0, BPF_REG_1);
14334 insn_buf[3] = BPF_LDX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
14335 insn_buf[4] = BPF_STX_MEM(BPF_DW, BPF_REG_2, BPF_REG_3, 0);
14336 insn_buf[5] = BPF_MOV64_IMM(BPF_REG_0, 0);
14339 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, -EOPNOTSUPP);
14343 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
14348 env->prog = prog = new_prog;
14349 insn = new_prog->insnsi + i + delta;
14353 /* Implement get_func_arg_cnt inline. */
14354 if (prog_type == BPF_PROG_TYPE_TRACING &&
14355 insn->imm == BPF_FUNC_get_func_arg_cnt) {
14356 /* Load nr_args from ctx - 8 */
14357 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
14359 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
14363 env->prog = prog = new_prog;
14364 insn = new_prog->insnsi + i + delta;
14368 /* Implement bpf_get_func_ip inline. */
14369 if (prog_type == BPF_PROG_TYPE_TRACING &&
14370 insn->imm == BPF_FUNC_get_func_ip) {
14371 /* Load IP address from ctx - 16 */
14372 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -16);
14374 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
14378 env->prog = prog = new_prog;
14379 insn = new_prog->insnsi + i + delta;
14384 fn = env->ops->get_func_proto(insn->imm, env->prog);
14385 /* all functions that have prototype and verifier allowed
14386 * programs to call them, must be real in-kernel functions
14390 "kernel subsystem misconfigured func %s#%d\n",
14391 func_id_name(insn->imm), insn->imm);
14394 insn->imm = fn->func - __bpf_call_base;
14397 /* Since poke tab is now finalized, publish aux to tracker. */
14398 for (i = 0; i < prog->aux->size_poke_tab; i++) {
14399 map_ptr = prog->aux->poke_tab[i].tail_call.map;
14400 if (!map_ptr->ops->map_poke_track ||
14401 !map_ptr->ops->map_poke_untrack ||
14402 !map_ptr->ops->map_poke_run) {
14403 verbose(env, "bpf verifier is misconfigured\n");
14407 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
14409 verbose(env, "tracking tail call prog failed\n");
14414 sort_kfunc_descs_by_imm(env->prog);
14419 static struct bpf_prog *inline_bpf_loop(struct bpf_verifier_env *env,
14422 u32 callback_subprogno,
14425 s32 r6_offset = stack_base + 0 * BPF_REG_SIZE;
14426 s32 r7_offset = stack_base + 1 * BPF_REG_SIZE;
14427 s32 r8_offset = stack_base + 2 * BPF_REG_SIZE;
14428 int reg_loop_max = BPF_REG_6;
14429 int reg_loop_cnt = BPF_REG_7;
14430 int reg_loop_ctx = BPF_REG_8;
14432 struct bpf_prog *new_prog;
14433 u32 callback_start;
14434 u32 call_insn_offset;
14435 s32 callback_offset;
14437 /* This represents an inlined version of bpf_iter.c:bpf_loop,
14438 * be careful to modify this code in sync.
14440 struct bpf_insn insn_buf[] = {
14441 /* Return error and jump to the end of the patch if
14442 * expected number of iterations is too big.
14444 BPF_JMP_IMM(BPF_JLE, BPF_REG_1, BPF_MAX_LOOPS, 2),
14445 BPF_MOV32_IMM(BPF_REG_0, -E2BIG),
14446 BPF_JMP_IMM(BPF_JA, 0, 0, 16),
14447 /* spill R6, R7, R8 to use these as loop vars */
14448 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_6, r6_offset),
14449 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_7, r7_offset),
14450 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_8, r8_offset),
14451 /* initialize loop vars */
14452 BPF_MOV64_REG(reg_loop_max, BPF_REG_1),
14453 BPF_MOV32_IMM(reg_loop_cnt, 0),
14454 BPF_MOV64_REG(reg_loop_ctx, BPF_REG_3),
14456 * if reg_loop_cnt >= reg_loop_max skip the loop body
14458 BPF_JMP_REG(BPF_JGE, reg_loop_cnt, reg_loop_max, 5),
14460 * correct callback offset would be set after patching
14462 BPF_MOV64_REG(BPF_REG_1, reg_loop_cnt),
14463 BPF_MOV64_REG(BPF_REG_2, reg_loop_ctx),
14465 /* increment loop counter */
14466 BPF_ALU64_IMM(BPF_ADD, reg_loop_cnt, 1),
14467 /* jump to loop header if callback returned 0 */
14468 BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, -6),
14469 /* return value of bpf_loop,
14470 * set R0 to the number of iterations
14472 BPF_MOV64_REG(BPF_REG_0, reg_loop_cnt),
14473 /* restore original values of R6, R7, R8 */
14474 BPF_LDX_MEM(BPF_DW, BPF_REG_6, BPF_REG_10, r6_offset),
14475 BPF_LDX_MEM(BPF_DW, BPF_REG_7, BPF_REG_10, r7_offset),
14476 BPF_LDX_MEM(BPF_DW, BPF_REG_8, BPF_REG_10, r8_offset),
14479 *cnt = ARRAY_SIZE(insn_buf);
14480 new_prog = bpf_patch_insn_data(env, position, insn_buf, *cnt);
14484 /* callback start is known only after patching */
14485 callback_start = env->subprog_info[callback_subprogno].start;
14486 /* Note: insn_buf[12] is an offset of BPF_CALL_REL instruction */
14487 call_insn_offset = position + 12;
14488 callback_offset = callback_start - call_insn_offset - 1;
14489 new_prog->insnsi[call_insn_offset].imm = callback_offset;
14494 static bool is_bpf_loop_call(struct bpf_insn *insn)
14496 return insn->code == (BPF_JMP | BPF_CALL) &&
14497 insn->src_reg == 0 &&
14498 insn->imm == BPF_FUNC_loop;
14501 /* For all sub-programs in the program (including main) check
14502 * insn_aux_data to see if there are bpf_loop calls that require
14503 * inlining. If such calls are found the calls are replaced with a
14504 * sequence of instructions produced by `inline_bpf_loop` function and
14505 * subprog stack_depth is increased by the size of 3 registers.
14506 * This stack space is used to spill values of the R6, R7, R8. These
14507 * registers are used to store the loop bound, counter and context
14510 static int optimize_bpf_loop(struct bpf_verifier_env *env)
14512 struct bpf_subprog_info *subprogs = env->subprog_info;
14513 int i, cur_subprog = 0, cnt, delta = 0;
14514 struct bpf_insn *insn = env->prog->insnsi;
14515 int insn_cnt = env->prog->len;
14516 u16 stack_depth = subprogs[cur_subprog].stack_depth;
14517 u16 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
14518 u16 stack_depth_extra = 0;
14520 for (i = 0; i < insn_cnt; i++, insn++) {
14521 struct bpf_loop_inline_state *inline_state =
14522 &env->insn_aux_data[i + delta].loop_inline_state;
14524 if (is_bpf_loop_call(insn) && inline_state->fit_for_inline) {
14525 struct bpf_prog *new_prog;
14527 stack_depth_extra = BPF_REG_SIZE * 3 + stack_depth_roundup;
14528 new_prog = inline_bpf_loop(env,
14530 -(stack_depth + stack_depth_extra),
14531 inline_state->callback_subprogno,
14537 env->prog = new_prog;
14538 insn = new_prog->insnsi + i + delta;
14541 if (subprogs[cur_subprog + 1].start == i + delta + 1) {
14542 subprogs[cur_subprog].stack_depth += stack_depth_extra;
14544 stack_depth = subprogs[cur_subprog].stack_depth;
14545 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
14546 stack_depth_extra = 0;
14550 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
14555 static void free_states(struct bpf_verifier_env *env)
14557 struct bpf_verifier_state_list *sl, *sln;
14560 sl = env->free_list;
14563 free_verifier_state(&sl->state, false);
14567 env->free_list = NULL;
14569 if (!env->explored_states)
14572 for (i = 0; i < state_htab_size(env); i++) {
14573 sl = env->explored_states[i];
14577 free_verifier_state(&sl->state, false);
14581 env->explored_states[i] = NULL;
14585 static int do_check_common(struct bpf_verifier_env *env, int subprog)
14587 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
14588 struct bpf_verifier_state *state;
14589 struct bpf_reg_state *regs;
14592 env->prev_linfo = NULL;
14595 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
14598 state->curframe = 0;
14599 state->speculative = false;
14600 state->branches = 1;
14601 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
14602 if (!state->frame[0]) {
14606 env->cur_state = state;
14607 init_func_state(env, state->frame[0],
14608 BPF_MAIN_FUNC /* callsite */,
14612 regs = state->frame[state->curframe]->regs;
14613 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
14614 ret = btf_prepare_func_args(env, subprog, regs);
14617 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
14618 if (regs[i].type == PTR_TO_CTX)
14619 mark_reg_known_zero(env, regs, i);
14620 else if (regs[i].type == SCALAR_VALUE)
14621 mark_reg_unknown(env, regs, i);
14622 else if (base_type(regs[i].type) == PTR_TO_MEM) {
14623 const u32 mem_size = regs[i].mem_size;
14625 mark_reg_known_zero(env, regs, i);
14626 regs[i].mem_size = mem_size;
14627 regs[i].id = ++env->id_gen;
14631 /* 1st arg to a function */
14632 regs[BPF_REG_1].type = PTR_TO_CTX;
14633 mark_reg_known_zero(env, regs, BPF_REG_1);
14634 ret = btf_check_subprog_arg_match(env, subprog, regs);
14635 if (ret == -EFAULT)
14636 /* unlikely verifier bug. abort.
14637 * ret == 0 and ret < 0 are sadly acceptable for
14638 * main() function due to backward compatibility.
14639 * Like socket filter program may be written as:
14640 * int bpf_prog(struct pt_regs *ctx)
14641 * and never dereference that ctx in the program.
14642 * 'struct pt_regs' is a type mismatch for socket
14643 * filter that should be using 'struct __sk_buff'.
14648 ret = do_check(env);
14650 /* check for NULL is necessary, since cur_state can be freed inside
14651 * do_check() under memory pressure.
14653 if (env->cur_state) {
14654 free_verifier_state(env->cur_state, true);
14655 env->cur_state = NULL;
14657 while (!pop_stack(env, NULL, NULL, false));
14658 if (!ret && pop_log)
14659 bpf_vlog_reset(&env->log, 0);
14664 /* Verify all global functions in a BPF program one by one based on their BTF.
14665 * All global functions must pass verification. Otherwise the whole program is rejected.
14676 * foo() will be verified first for R1=any_scalar_value. During verification it
14677 * will be assumed that bar() already verified successfully and call to bar()
14678 * from foo() will be checked for type match only. Later bar() will be verified
14679 * independently to check that it's safe for R1=any_scalar_value.
14681 static int do_check_subprogs(struct bpf_verifier_env *env)
14683 struct bpf_prog_aux *aux = env->prog->aux;
14686 if (!aux->func_info)
14689 for (i = 1; i < env->subprog_cnt; i++) {
14690 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
14692 env->insn_idx = env->subprog_info[i].start;
14693 WARN_ON_ONCE(env->insn_idx == 0);
14694 ret = do_check_common(env, i);
14697 } else if (env->log.level & BPF_LOG_LEVEL) {
14699 "Func#%d is safe for any args that match its prototype\n",
14706 static int do_check_main(struct bpf_verifier_env *env)
14711 ret = do_check_common(env, 0);
14713 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
14718 static void print_verification_stats(struct bpf_verifier_env *env)
14722 if (env->log.level & BPF_LOG_STATS) {
14723 verbose(env, "verification time %lld usec\n",
14724 div_u64(env->verification_time, 1000));
14725 verbose(env, "stack depth ");
14726 for (i = 0; i < env->subprog_cnt; i++) {
14727 u32 depth = env->subprog_info[i].stack_depth;
14729 verbose(env, "%d", depth);
14730 if (i + 1 < env->subprog_cnt)
14733 verbose(env, "\n");
14735 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
14736 "total_states %d peak_states %d mark_read %d\n",
14737 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
14738 env->max_states_per_insn, env->total_states,
14739 env->peak_states, env->longest_mark_read_walk);
14742 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
14744 const struct btf_type *t, *func_proto;
14745 const struct bpf_struct_ops *st_ops;
14746 const struct btf_member *member;
14747 struct bpf_prog *prog = env->prog;
14748 u32 btf_id, member_idx;
14751 if (!prog->gpl_compatible) {
14752 verbose(env, "struct ops programs must have a GPL compatible license\n");
14756 btf_id = prog->aux->attach_btf_id;
14757 st_ops = bpf_struct_ops_find(btf_id);
14759 verbose(env, "attach_btf_id %u is not a supported struct\n",
14765 member_idx = prog->expected_attach_type;
14766 if (member_idx >= btf_type_vlen(t)) {
14767 verbose(env, "attach to invalid member idx %u of struct %s\n",
14768 member_idx, st_ops->name);
14772 member = &btf_type_member(t)[member_idx];
14773 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
14774 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
14777 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
14778 mname, member_idx, st_ops->name);
14782 if (st_ops->check_member) {
14783 int err = st_ops->check_member(t, member);
14786 verbose(env, "attach to unsupported member %s of struct %s\n",
14787 mname, st_ops->name);
14792 prog->aux->attach_func_proto = func_proto;
14793 prog->aux->attach_func_name = mname;
14794 env->ops = st_ops->verifier_ops;
14798 #define SECURITY_PREFIX "security_"
14800 static int check_attach_modify_return(unsigned long addr, const char *func_name)
14802 if (within_error_injection_list(addr) ||
14803 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
14809 /* list of non-sleepable functions that are otherwise on
14810 * ALLOW_ERROR_INJECTION list
14812 BTF_SET_START(btf_non_sleepable_error_inject)
14813 /* Three functions below can be called from sleepable and non-sleepable context.
14814 * Assume non-sleepable from bpf safety point of view.
14816 BTF_ID(func, __filemap_add_folio)
14817 BTF_ID(func, should_fail_alloc_page)
14818 BTF_ID(func, should_failslab)
14819 BTF_SET_END(btf_non_sleepable_error_inject)
14821 static int check_non_sleepable_error_inject(u32 btf_id)
14823 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
14826 int bpf_check_attach_target(struct bpf_verifier_log *log,
14827 const struct bpf_prog *prog,
14828 const struct bpf_prog *tgt_prog,
14830 struct bpf_attach_target_info *tgt_info)
14832 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
14833 const char prefix[] = "btf_trace_";
14834 int ret = 0, subprog = -1, i;
14835 const struct btf_type *t;
14836 bool conservative = true;
14842 bpf_log(log, "Tracing programs must provide btf_id\n");
14845 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
14848 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
14851 t = btf_type_by_id(btf, btf_id);
14853 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
14856 tname = btf_name_by_offset(btf, t->name_off);
14858 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
14862 struct bpf_prog_aux *aux = tgt_prog->aux;
14864 for (i = 0; i < aux->func_info_cnt; i++)
14865 if (aux->func_info[i].type_id == btf_id) {
14869 if (subprog == -1) {
14870 bpf_log(log, "Subprog %s doesn't exist\n", tname);
14873 conservative = aux->func_info_aux[subprog].unreliable;
14874 if (prog_extension) {
14875 if (conservative) {
14877 "Cannot replace static functions\n");
14880 if (!prog->jit_requested) {
14882 "Extension programs should be JITed\n");
14886 if (!tgt_prog->jited) {
14887 bpf_log(log, "Can attach to only JITed progs\n");
14890 if (tgt_prog->type == prog->type) {
14891 /* Cannot fentry/fexit another fentry/fexit program.
14892 * Cannot attach program extension to another extension.
14893 * It's ok to attach fentry/fexit to extension program.
14895 bpf_log(log, "Cannot recursively attach\n");
14898 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
14900 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
14901 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
14902 /* Program extensions can extend all program types
14903 * except fentry/fexit. The reason is the following.
14904 * The fentry/fexit programs are used for performance
14905 * analysis, stats and can be attached to any program
14906 * type except themselves. When extension program is
14907 * replacing XDP function it is necessary to allow
14908 * performance analysis of all functions. Both original
14909 * XDP program and its program extension. Hence
14910 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
14911 * allowed. If extending of fentry/fexit was allowed it
14912 * would be possible to create long call chain
14913 * fentry->extension->fentry->extension beyond
14914 * reasonable stack size. Hence extending fentry is not
14917 bpf_log(log, "Cannot extend fentry/fexit\n");
14921 if (prog_extension) {
14922 bpf_log(log, "Cannot replace kernel functions\n");
14927 switch (prog->expected_attach_type) {
14928 case BPF_TRACE_RAW_TP:
14931 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
14934 if (!btf_type_is_typedef(t)) {
14935 bpf_log(log, "attach_btf_id %u is not a typedef\n",
14939 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
14940 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
14944 tname += sizeof(prefix) - 1;
14945 t = btf_type_by_id(btf, t->type);
14946 if (!btf_type_is_ptr(t))
14947 /* should never happen in valid vmlinux build */
14949 t = btf_type_by_id(btf, t->type);
14950 if (!btf_type_is_func_proto(t))
14951 /* should never happen in valid vmlinux build */
14955 case BPF_TRACE_ITER:
14956 if (!btf_type_is_func(t)) {
14957 bpf_log(log, "attach_btf_id %u is not a function\n",
14961 t = btf_type_by_id(btf, t->type);
14962 if (!btf_type_is_func_proto(t))
14964 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
14969 if (!prog_extension)
14972 case BPF_MODIFY_RETURN:
14974 case BPF_LSM_CGROUP:
14975 case BPF_TRACE_FENTRY:
14976 case BPF_TRACE_FEXIT:
14977 if (!btf_type_is_func(t)) {
14978 bpf_log(log, "attach_btf_id %u is not a function\n",
14982 if (prog_extension &&
14983 btf_check_type_match(log, prog, btf, t))
14985 t = btf_type_by_id(btf, t->type);
14986 if (!btf_type_is_func_proto(t))
14989 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
14990 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
14991 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
14994 if (tgt_prog && conservative)
14997 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
15003 addr = (long) tgt_prog->bpf_func;
15005 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
15007 addr = kallsyms_lookup_name(tname);
15010 "The address of function %s cannot be found\n",
15016 if (prog->aux->sleepable) {
15018 switch (prog->type) {
15019 case BPF_PROG_TYPE_TRACING:
15020 /* fentry/fexit/fmod_ret progs can be sleepable only if they are
15021 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
15023 if (!check_non_sleepable_error_inject(btf_id) &&
15024 within_error_injection_list(addr))
15027 case BPF_PROG_TYPE_LSM:
15028 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
15029 * Only some of them are sleepable.
15031 if (bpf_lsm_is_sleepable_hook(btf_id))
15038 bpf_log(log, "%s is not sleepable\n", tname);
15041 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
15043 bpf_log(log, "can't modify return codes of BPF programs\n");
15046 ret = check_attach_modify_return(addr, tname);
15048 bpf_log(log, "%s() is not modifiable\n", tname);
15055 tgt_info->tgt_addr = addr;
15056 tgt_info->tgt_name = tname;
15057 tgt_info->tgt_type = t;
15061 BTF_SET_START(btf_id_deny)
15064 BTF_ID(func, migrate_disable)
15065 BTF_ID(func, migrate_enable)
15067 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
15068 BTF_ID(func, rcu_read_unlock_strict)
15070 BTF_SET_END(btf_id_deny)
15072 static int check_attach_btf_id(struct bpf_verifier_env *env)
15074 struct bpf_prog *prog = env->prog;
15075 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
15076 struct bpf_attach_target_info tgt_info = {};
15077 u32 btf_id = prog->aux->attach_btf_id;
15078 struct bpf_trampoline *tr;
15082 if (prog->type == BPF_PROG_TYPE_SYSCALL) {
15083 if (prog->aux->sleepable)
15084 /* attach_btf_id checked to be zero already */
15086 verbose(env, "Syscall programs can only be sleepable\n");
15090 if (prog->aux->sleepable && prog->type != BPF_PROG_TYPE_TRACING &&
15091 prog->type != BPF_PROG_TYPE_LSM && prog->type != BPF_PROG_TYPE_KPROBE) {
15092 verbose(env, "Only fentry/fexit/fmod_ret, lsm, and kprobe/uprobe programs can be sleepable\n");
15096 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
15097 return check_struct_ops_btf_id(env);
15099 if (prog->type != BPF_PROG_TYPE_TRACING &&
15100 prog->type != BPF_PROG_TYPE_LSM &&
15101 prog->type != BPF_PROG_TYPE_EXT)
15104 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
15108 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
15109 /* to make freplace equivalent to their targets, they need to
15110 * inherit env->ops and expected_attach_type for the rest of the
15113 env->ops = bpf_verifier_ops[tgt_prog->type];
15114 prog->expected_attach_type = tgt_prog->expected_attach_type;
15117 /* store info about the attachment target that will be used later */
15118 prog->aux->attach_func_proto = tgt_info.tgt_type;
15119 prog->aux->attach_func_name = tgt_info.tgt_name;
15122 prog->aux->saved_dst_prog_type = tgt_prog->type;
15123 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
15126 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
15127 prog->aux->attach_btf_trace = true;
15129 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
15130 if (!bpf_iter_prog_supported(prog))
15135 if (prog->type == BPF_PROG_TYPE_LSM) {
15136 ret = bpf_lsm_verify_prog(&env->log, prog);
15139 } else if (prog->type == BPF_PROG_TYPE_TRACING &&
15140 btf_id_set_contains(&btf_id_deny, btf_id)) {
15144 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
15145 tr = bpf_trampoline_get(key, &tgt_info);
15149 prog->aux->dst_trampoline = tr;
15153 struct btf *bpf_get_btf_vmlinux(void)
15155 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
15156 mutex_lock(&bpf_verifier_lock);
15158 btf_vmlinux = btf_parse_vmlinux();
15159 mutex_unlock(&bpf_verifier_lock);
15161 return btf_vmlinux;
15164 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr)
15166 u64 start_time = ktime_get_ns();
15167 struct bpf_verifier_env *env;
15168 struct bpf_verifier_log *log;
15169 int i, len, ret = -EINVAL;
15172 /* no program is valid */
15173 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
15176 /* 'struct bpf_verifier_env' can be global, but since it's not small,
15177 * allocate/free it every time bpf_check() is called
15179 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
15184 len = (*prog)->len;
15185 env->insn_aux_data =
15186 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
15188 if (!env->insn_aux_data)
15190 for (i = 0; i < len; i++)
15191 env->insn_aux_data[i].orig_idx = i;
15193 env->ops = bpf_verifier_ops[env->prog->type];
15194 env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
15195 is_priv = bpf_capable();
15197 bpf_get_btf_vmlinux();
15199 /* grab the mutex to protect few globals used by verifier */
15201 mutex_lock(&bpf_verifier_lock);
15203 if (attr->log_level || attr->log_buf || attr->log_size) {
15204 /* user requested verbose verifier output
15205 * and supplied buffer to store the verification trace
15207 log->level = attr->log_level;
15208 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
15209 log->len_total = attr->log_size;
15211 /* log attributes have to be sane */
15212 if (!bpf_verifier_log_attr_valid(log)) {
15218 mark_verifier_state_clean(env);
15220 if (IS_ERR(btf_vmlinux)) {
15221 /* Either gcc or pahole or kernel are broken. */
15222 verbose(env, "in-kernel BTF is malformed\n");
15223 ret = PTR_ERR(btf_vmlinux);
15224 goto skip_full_check;
15227 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
15228 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
15229 env->strict_alignment = true;
15230 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
15231 env->strict_alignment = false;
15233 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
15234 env->allow_uninit_stack = bpf_allow_uninit_stack();
15235 env->allow_ptr_to_map_access = bpf_allow_ptr_to_map_access();
15236 env->bypass_spec_v1 = bpf_bypass_spec_v1();
15237 env->bypass_spec_v4 = bpf_bypass_spec_v4();
15238 env->bpf_capable = bpf_capable();
15241 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
15243 env->explored_states = kvcalloc(state_htab_size(env),
15244 sizeof(struct bpf_verifier_state_list *),
15247 if (!env->explored_states)
15248 goto skip_full_check;
15250 ret = add_subprog_and_kfunc(env);
15252 goto skip_full_check;
15254 ret = check_subprogs(env);
15256 goto skip_full_check;
15258 ret = check_btf_info(env, attr, uattr);
15260 goto skip_full_check;
15262 ret = check_attach_btf_id(env);
15264 goto skip_full_check;
15266 ret = resolve_pseudo_ldimm64(env);
15268 goto skip_full_check;
15270 if (bpf_prog_is_dev_bound(env->prog->aux)) {
15271 ret = bpf_prog_offload_verifier_prep(env->prog);
15273 goto skip_full_check;
15276 ret = check_cfg(env);
15278 goto skip_full_check;
15280 ret = do_check_subprogs(env);
15281 ret = ret ?: do_check_main(env);
15283 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
15284 ret = bpf_prog_offload_finalize(env);
15287 kvfree(env->explored_states);
15290 ret = check_max_stack_depth(env);
15292 /* instruction rewrites happen after this point */
15294 ret = optimize_bpf_loop(env);
15298 opt_hard_wire_dead_code_branches(env);
15300 ret = opt_remove_dead_code(env);
15302 ret = opt_remove_nops(env);
15305 sanitize_dead_code(env);
15309 /* program is valid, convert *(u32*)(ctx + off) accesses */
15310 ret = convert_ctx_accesses(env);
15313 ret = do_misc_fixups(env);
15315 /* do 32-bit optimization after insn patching has done so those patched
15316 * insns could be handled correctly.
15318 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
15319 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
15320 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
15325 ret = fixup_call_args(env);
15327 env->verification_time = ktime_get_ns() - start_time;
15328 print_verification_stats(env);
15329 env->prog->aux->verified_insns = env->insn_processed;
15331 if (log->level && bpf_verifier_log_full(log))
15333 if (log->level && !log->ubuf) {
15335 goto err_release_maps;
15339 goto err_release_maps;
15341 if (env->used_map_cnt) {
15342 /* if program passed verifier, update used_maps in bpf_prog_info */
15343 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
15344 sizeof(env->used_maps[0]),
15347 if (!env->prog->aux->used_maps) {
15349 goto err_release_maps;
15352 memcpy(env->prog->aux->used_maps, env->used_maps,
15353 sizeof(env->used_maps[0]) * env->used_map_cnt);
15354 env->prog->aux->used_map_cnt = env->used_map_cnt;
15356 if (env->used_btf_cnt) {
15357 /* if program passed verifier, update used_btfs in bpf_prog_aux */
15358 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
15359 sizeof(env->used_btfs[0]),
15361 if (!env->prog->aux->used_btfs) {
15363 goto err_release_maps;
15366 memcpy(env->prog->aux->used_btfs, env->used_btfs,
15367 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
15368 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
15370 if (env->used_map_cnt || env->used_btf_cnt) {
15371 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
15372 * bpf_ld_imm64 instructions
15374 convert_pseudo_ld_imm64(env);
15377 adjust_btf_func(env);
15380 if (!env->prog->aux->used_maps)
15381 /* if we didn't copy map pointers into bpf_prog_info, release
15382 * them now. Otherwise free_used_maps() will release them.
15385 if (!env->prog->aux->used_btfs)
15388 /* extension progs temporarily inherit the attach_type of their targets
15389 for verification purposes, so set it back to zero before returning
15391 if (env->prog->type == BPF_PROG_TYPE_EXT)
15392 env->prog->expected_attach_type = 0;
15397 mutex_unlock(&bpf_verifier_lock);
15398 vfree(env->insn_aux_data);