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>
29 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
30 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
31 [_id] = & _name ## _verifier_ops,
32 #define BPF_MAP_TYPE(_id, _ops)
33 #define BPF_LINK_TYPE(_id, _name)
34 #include <linux/bpf_types.h>
40 /* bpf_check() is a static code analyzer that walks eBPF program
41 * instruction by instruction and updates register/stack state.
42 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
44 * The first pass is depth-first-search to check that the program is a DAG.
45 * It rejects the following programs:
46 * - larger than BPF_MAXINSNS insns
47 * - if loop is present (detected via back-edge)
48 * - unreachable insns exist (shouldn't be a forest. program = one function)
49 * - out of bounds or malformed jumps
50 * The second pass is all possible path descent from the 1st insn.
51 * Since it's analyzing all paths through the program, the length of the
52 * analysis is limited to 64k insn, which may be hit even if total number of
53 * insn is less then 4K, but there are too many branches that change stack/regs.
54 * Number of 'branches to be analyzed' is limited to 1k
56 * On entry to each instruction, each register has a type, and the instruction
57 * changes the types of the registers depending on instruction semantics.
58 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
61 * All registers are 64-bit.
62 * R0 - return register
63 * R1-R5 argument passing registers
64 * R6-R9 callee saved registers
65 * R10 - frame pointer read-only
67 * At the start of BPF program the register R1 contains a pointer to bpf_context
68 * and has type PTR_TO_CTX.
70 * Verifier tracks arithmetic operations on pointers in case:
71 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
72 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
73 * 1st insn copies R10 (which has FRAME_PTR) type into R1
74 * and 2nd arithmetic instruction is pattern matched to recognize
75 * that it wants to construct a pointer to some element within stack.
76 * So after 2nd insn, the register R1 has type PTR_TO_STACK
77 * (and -20 constant is saved for further stack bounds checking).
78 * Meaning that this reg is a pointer to stack plus known immediate constant.
80 * Most of the time the registers have SCALAR_VALUE type, which
81 * means the register has some value, but it's not a valid pointer.
82 * (like pointer plus pointer becomes SCALAR_VALUE type)
84 * When verifier sees load or store instructions the type of base register
85 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
86 * four pointer types recognized by check_mem_access() function.
88 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
89 * and the range of [ptr, ptr + map's value_size) is accessible.
91 * registers used to pass values to function calls are checked against
92 * function argument constraints.
94 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
95 * It means that the register type passed to this function must be
96 * PTR_TO_STACK and it will be used inside the function as
97 * 'pointer to map element key'
99 * For example the argument constraints for bpf_map_lookup_elem():
100 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
101 * .arg1_type = ARG_CONST_MAP_PTR,
102 * .arg2_type = ARG_PTR_TO_MAP_KEY,
104 * ret_type says that this function returns 'pointer to map elem value or null'
105 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
106 * 2nd argument should be a pointer to stack, which will be used inside
107 * the helper function as a pointer to map element key.
109 * On the kernel side the helper function looks like:
110 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
112 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
113 * void *key = (void *) (unsigned long) r2;
116 * here kernel can access 'key' and 'map' pointers safely, knowing that
117 * [key, key + map->key_size) bytes are valid and were initialized on
118 * the stack of eBPF program.
121 * Corresponding eBPF program may look like:
122 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
123 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
124 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
125 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
126 * here verifier looks at prototype of map_lookup_elem() and sees:
127 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
128 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
130 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
131 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
132 * and were initialized prior to this call.
133 * If it's ok, then verifier allows this BPF_CALL insn and looks at
134 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
135 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
136 * returns either pointer to map value or NULL.
138 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
139 * insn, the register holding that pointer in the true branch changes state to
140 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
141 * branch. See check_cond_jmp_op().
143 * After the call R0 is set to return type of the function and registers R1-R5
144 * are set to NOT_INIT to indicate that they are no longer readable.
146 * The following reference types represent a potential reference to a kernel
147 * resource which, after first being allocated, must be checked and freed by
149 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
151 * When the verifier sees a helper call return a reference type, it allocates a
152 * pointer id for the reference and stores it in the current function state.
153 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
154 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
155 * passes through a NULL-check conditional. For the branch wherein the state is
156 * changed to CONST_IMM, the verifier releases the reference.
158 * For each helper function that allocates a reference, such as
159 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
160 * bpf_sk_release(). When a reference type passes into the release function,
161 * the verifier also releases the reference. If any unchecked or unreleased
162 * reference remains at the end of the program, the verifier rejects it.
165 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
166 struct bpf_verifier_stack_elem {
167 /* verifer state is 'st'
168 * before processing instruction 'insn_idx'
169 * and after processing instruction 'prev_insn_idx'
171 struct bpf_verifier_state st;
174 struct bpf_verifier_stack_elem *next;
175 /* length of verifier log at the time this state was pushed on stack */
179 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
180 #define BPF_COMPLEXITY_LIMIT_STATES 64
182 #define BPF_MAP_KEY_POISON (1ULL << 63)
183 #define BPF_MAP_KEY_SEEN (1ULL << 62)
185 #define BPF_MAP_PTR_UNPRIV 1UL
186 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
187 POISON_POINTER_DELTA))
188 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
190 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx);
191 static int release_reference(struct bpf_verifier_env *env, int ref_obj_id);
193 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
195 return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON;
198 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
200 return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV;
203 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
204 const struct bpf_map *map, bool unpriv)
206 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
207 unpriv |= bpf_map_ptr_unpriv(aux);
208 aux->map_ptr_state = (unsigned long)map |
209 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
212 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
214 return aux->map_key_state & BPF_MAP_KEY_POISON;
217 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
219 return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
222 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
224 return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
227 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
229 bool poisoned = bpf_map_key_poisoned(aux);
231 aux->map_key_state = state | BPF_MAP_KEY_SEEN |
232 (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
235 static bool bpf_pseudo_call(const struct bpf_insn *insn)
237 return insn->code == (BPF_JMP | BPF_CALL) &&
238 insn->src_reg == BPF_PSEUDO_CALL;
241 static bool bpf_pseudo_kfunc_call(const struct bpf_insn *insn)
243 return insn->code == (BPF_JMP | BPF_CALL) &&
244 insn->src_reg == BPF_PSEUDO_KFUNC_CALL;
247 struct bpf_call_arg_meta {
248 struct bpf_map *map_ptr;
264 struct bpf_map_value_off_desc *kptr_off_desc;
265 u8 uninit_dynptr_regno;
268 struct btf *btf_vmlinux;
270 static DEFINE_MUTEX(bpf_verifier_lock);
272 static const struct bpf_line_info *
273 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
275 const struct bpf_line_info *linfo;
276 const struct bpf_prog *prog;
280 nr_linfo = prog->aux->nr_linfo;
282 if (!nr_linfo || insn_off >= prog->len)
285 linfo = prog->aux->linfo;
286 for (i = 1; i < nr_linfo; i++)
287 if (insn_off < linfo[i].insn_off)
290 return &linfo[i - 1];
293 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
298 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
300 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
301 "verifier log line truncated - local buffer too short\n");
303 if (log->level == BPF_LOG_KERNEL) {
304 bool newline = n > 0 && log->kbuf[n - 1] == '\n';
306 pr_err("BPF: %s%s", log->kbuf, newline ? "" : "\n");
310 n = min(log->len_total - log->len_used - 1, n);
312 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
318 static void bpf_vlog_reset(struct bpf_verifier_log *log, u32 new_pos)
322 if (!bpf_verifier_log_needed(log))
325 log->len_used = new_pos;
326 if (put_user(zero, log->ubuf + new_pos))
330 /* log_level controls verbosity level of eBPF verifier.
331 * bpf_verifier_log_write() is used to dump the verification trace to the log,
332 * so the user can figure out what's wrong with the program
334 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
335 const char *fmt, ...)
339 if (!bpf_verifier_log_needed(&env->log))
343 bpf_verifier_vlog(&env->log, fmt, args);
346 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
348 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
350 struct bpf_verifier_env *env = private_data;
353 if (!bpf_verifier_log_needed(&env->log))
357 bpf_verifier_vlog(&env->log, fmt, args);
361 __printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
362 const char *fmt, ...)
366 if (!bpf_verifier_log_needed(log))
370 bpf_verifier_vlog(log, fmt, args);
374 static const char *ltrim(const char *s)
382 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
384 const char *prefix_fmt, ...)
386 const struct bpf_line_info *linfo;
388 if (!bpf_verifier_log_needed(&env->log))
391 linfo = find_linfo(env, insn_off);
392 if (!linfo || linfo == env->prev_linfo)
398 va_start(args, prefix_fmt);
399 bpf_verifier_vlog(&env->log, prefix_fmt, args);
404 ltrim(btf_name_by_offset(env->prog->aux->btf,
407 env->prev_linfo = linfo;
410 static void verbose_invalid_scalar(struct bpf_verifier_env *env,
411 struct bpf_reg_state *reg,
412 struct tnum *range, const char *ctx,
413 const char *reg_name)
417 verbose(env, "At %s the register %s ", ctx, reg_name);
418 if (!tnum_is_unknown(reg->var_off)) {
419 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
420 verbose(env, "has value %s", tn_buf);
422 verbose(env, "has unknown scalar value");
424 tnum_strn(tn_buf, sizeof(tn_buf), *range);
425 verbose(env, " should have been in %s\n", tn_buf);
428 static bool type_is_pkt_pointer(enum bpf_reg_type type)
430 return type == PTR_TO_PACKET ||
431 type == PTR_TO_PACKET_META;
434 static bool type_is_sk_pointer(enum bpf_reg_type type)
436 return type == PTR_TO_SOCKET ||
437 type == PTR_TO_SOCK_COMMON ||
438 type == PTR_TO_TCP_SOCK ||
439 type == PTR_TO_XDP_SOCK;
442 static bool reg_type_not_null(enum bpf_reg_type type)
444 return type == PTR_TO_SOCKET ||
445 type == PTR_TO_TCP_SOCK ||
446 type == PTR_TO_MAP_VALUE ||
447 type == PTR_TO_MAP_KEY ||
448 type == PTR_TO_SOCK_COMMON;
451 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
453 return reg->type == PTR_TO_MAP_VALUE &&
454 map_value_has_spin_lock(reg->map_ptr);
457 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
459 return base_type(type) == PTR_TO_SOCKET ||
460 base_type(type) == PTR_TO_TCP_SOCK ||
461 base_type(type) == PTR_TO_MEM ||
462 base_type(type) == PTR_TO_BTF_ID;
465 static bool type_is_rdonly_mem(u32 type)
467 return type & MEM_RDONLY;
470 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
472 return type == ARG_PTR_TO_SOCK_COMMON;
475 static bool type_may_be_null(u32 type)
477 return type & PTR_MAYBE_NULL;
480 static bool may_be_acquire_function(enum bpf_func_id func_id)
482 return 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_map_lookup_elem ||
486 func_id == BPF_FUNC_ringbuf_reserve;
489 static bool is_acquire_function(enum bpf_func_id func_id,
490 const struct bpf_map *map)
492 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
494 if (func_id == BPF_FUNC_sk_lookup_tcp ||
495 func_id == BPF_FUNC_sk_lookup_udp ||
496 func_id == BPF_FUNC_skc_lookup_tcp ||
497 func_id == BPF_FUNC_ringbuf_reserve ||
498 func_id == BPF_FUNC_kptr_xchg)
501 if (func_id == BPF_FUNC_map_lookup_elem &&
502 (map_type == BPF_MAP_TYPE_SOCKMAP ||
503 map_type == BPF_MAP_TYPE_SOCKHASH))
509 static bool is_ptr_cast_function(enum bpf_func_id func_id)
511 return func_id == BPF_FUNC_tcp_sock ||
512 func_id == BPF_FUNC_sk_fullsock ||
513 func_id == BPF_FUNC_skc_to_tcp_sock ||
514 func_id == BPF_FUNC_skc_to_tcp6_sock ||
515 func_id == BPF_FUNC_skc_to_udp6_sock ||
516 func_id == BPF_FUNC_skc_to_mptcp_sock ||
517 func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
518 func_id == BPF_FUNC_skc_to_tcp_request_sock;
521 static bool is_cmpxchg_insn(const struct bpf_insn *insn)
523 return BPF_CLASS(insn->code) == BPF_STX &&
524 BPF_MODE(insn->code) == BPF_ATOMIC &&
525 insn->imm == BPF_CMPXCHG;
528 /* string representation of 'enum bpf_reg_type'
530 * Note that reg_type_str() can not appear more than once in a single verbose()
533 static const char *reg_type_str(struct bpf_verifier_env *env,
534 enum bpf_reg_type type)
536 char postfix[16] = {0}, prefix[32] = {0};
537 static const char * const str[] = {
539 [SCALAR_VALUE] = "scalar",
540 [PTR_TO_CTX] = "ctx",
541 [CONST_PTR_TO_MAP] = "map_ptr",
542 [PTR_TO_MAP_VALUE] = "map_value",
543 [PTR_TO_STACK] = "fp",
544 [PTR_TO_PACKET] = "pkt",
545 [PTR_TO_PACKET_META] = "pkt_meta",
546 [PTR_TO_PACKET_END] = "pkt_end",
547 [PTR_TO_FLOW_KEYS] = "flow_keys",
548 [PTR_TO_SOCKET] = "sock",
549 [PTR_TO_SOCK_COMMON] = "sock_common",
550 [PTR_TO_TCP_SOCK] = "tcp_sock",
551 [PTR_TO_TP_BUFFER] = "tp_buffer",
552 [PTR_TO_XDP_SOCK] = "xdp_sock",
553 [PTR_TO_BTF_ID] = "ptr_",
554 [PTR_TO_MEM] = "mem",
555 [PTR_TO_BUF] = "buf",
556 [PTR_TO_FUNC] = "func",
557 [PTR_TO_MAP_KEY] = "map_key",
560 if (type & PTR_MAYBE_NULL) {
561 if (base_type(type) == PTR_TO_BTF_ID)
562 strncpy(postfix, "or_null_", 16);
564 strncpy(postfix, "_or_null", 16);
567 if (type & MEM_RDONLY)
568 strncpy(prefix, "rdonly_", 32);
569 if (type & MEM_ALLOC)
570 strncpy(prefix, "alloc_", 32);
572 strncpy(prefix, "user_", 32);
573 if (type & MEM_PERCPU)
574 strncpy(prefix, "percpu_", 32);
575 if (type & PTR_UNTRUSTED)
576 strncpy(prefix, "untrusted_", 32);
578 snprintf(env->type_str_buf, TYPE_STR_BUF_LEN, "%s%s%s",
579 prefix, str[base_type(type)], postfix);
580 return env->type_str_buf;
583 static char slot_type_char[] = {
584 [STACK_INVALID] = '?',
588 [STACK_DYNPTR] = 'd',
591 static void print_liveness(struct bpf_verifier_env *env,
592 enum bpf_reg_liveness live)
594 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
596 if (live & REG_LIVE_READ)
598 if (live & REG_LIVE_WRITTEN)
600 if (live & REG_LIVE_DONE)
604 static int get_spi(s32 off)
606 return (-off - 1) / BPF_REG_SIZE;
609 static bool is_spi_bounds_valid(struct bpf_func_state *state, int spi, int nr_slots)
611 int allocated_slots = state->allocated_stack / BPF_REG_SIZE;
613 /* We need to check that slots between [spi - nr_slots + 1, spi] are
614 * within [0, allocated_stack).
616 * Please note that the spi grows downwards. For example, a dynptr
617 * takes the size of two stack slots; the first slot will be at
618 * spi and the second slot will be at spi - 1.
620 return spi - nr_slots + 1 >= 0 && spi < allocated_slots;
623 static struct bpf_func_state *func(struct bpf_verifier_env *env,
624 const struct bpf_reg_state *reg)
626 struct bpf_verifier_state *cur = env->cur_state;
628 return cur->frame[reg->frameno];
631 static const char *kernel_type_name(const struct btf* btf, u32 id)
633 return btf_name_by_offset(btf, btf_type_by_id(btf, id)->name_off);
636 static void mark_reg_scratched(struct bpf_verifier_env *env, u32 regno)
638 env->scratched_regs |= 1U << regno;
641 static void mark_stack_slot_scratched(struct bpf_verifier_env *env, u32 spi)
643 env->scratched_stack_slots |= 1ULL << spi;
646 static bool reg_scratched(const struct bpf_verifier_env *env, u32 regno)
648 return (env->scratched_regs >> regno) & 1;
651 static bool stack_slot_scratched(const struct bpf_verifier_env *env, u64 regno)
653 return (env->scratched_stack_slots >> regno) & 1;
656 static bool verifier_state_scratched(const struct bpf_verifier_env *env)
658 return env->scratched_regs || env->scratched_stack_slots;
661 static void mark_verifier_state_clean(struct bpf_verifier_env *env)
663 env->scratched_regs = 0U;
664 env->scratched_stack_slots = 0ULL;
667 /* Used for printing the entire verifier state. */
668 static void mark_verifier_state_scratched(struct bpf_verifier_env *env)
670 env->scratched_regs = ~0U;
671 env->scratched_stack_slots = ~0ULL;
674 static enum bpf_dynptr_type arg_to_dynptr_type(enum bpf_arg_type arg_type)
676 switch (arg_type & DYNPTR_TYPE_FLAG_MASK) {
677 case DYNPTR_TYPE_LOCAL:
678 return BPF_DYNPTR_TYPE_LOCAL;
679 case DYNPTR_TYPE_RINGBUF:
680 return BPF_DYNPTR_TYPE_RINGBUF;
682 return BPF_DYNPTR_TYPE_INVALID;
686 static bool dynptr_type_refcounted(enum bpf_dynptr_type type)
688 return type == BPF_DYNPTR_TYPE_RINGBUF;
691 static int mark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
692 enum bpf_arg_type arg_type, int insn_idx)
694 struct bpf_func_state *state = func(env, reg);
695 enum bpf_dynptr_type type;
698 spi = get_spi(reg->off);
700 if (!is_spi_bounds_valid(state, spi, BPF_DYNPTR_NR_SLOTS))
703 for (i = 0; i < BPF_REG_SIZE; i++) {
704 state->stack[spi].slot_type[i] = STACK_DYNPTR;
705 state->stack[spi - 1].slot_type[i] = STACK_DYNPTR;
708 type = arg_to_dynptr_type(arg_type);
709 if (type == BPF_DYNPTR_TYPE_INVALID)
712 state->stack[spi].spilled_ptr.dynptr.first_slot = true;
713 state->stack[spi].spilled_ptr.dynptr.type = type;
714 state->stack[spi - 1].spilled_ptr.dynptr.type = type;
716 if (dynptr_type_refcounted(type)) {
717 /* The id is used to track proper releasing */
718 id = acquire_reference_state(env, insn_idx);
722 state->stack[spi].spilled_ptr.id = id;
723 state->stack[spi - 1].spilled_ptr.id = id;
729 static int unmark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
731 struct bpf_func_state *state = func(env, reg);
734 spi = get_spi(reg->off);
736 if (!is_spi_bounds_valid(state, spi, BPF_DYNPTR_NR_SLOTS))
739 for (i = 0; i < BPF_REG_SIZE; i++) {
740 state->stack[spi].slot_type[i] = STACK_INVALID;
741 state->stack[spi - 1].slot_type[i] = STACK_INVALID;
744 /* Invalidate any slices associated with this dynptr */
745 if (dynptr_type_refcounted(state->stack[spi].spilled_ptr.dynptr.type)) {
746 release_reference(env, state->stack[spi].spilled_ptr.id);
747 state->stack[spi].spilled_ptr.id = 0;
748 state->stack[spi - 1].spilled_ptr.id = 0;
751 state->stack[spi].spilled_ptr.dynptr.first_slot = false;
752 state->stack[spi].spilled_ptr.dynptr.type = 0;
753 state->stack[spi - 1].spilled_ptr.dynptr.type = 0;
758 static bool is_dynptr_reg_valid_uninit(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
760 struct bpf_func_state *state = func(env, reg);
761 int spi = get_spi(reg->off);
764 if (!is_spi_bounds_valid(state, spi, BPF_DYNPTR_NR_SLOTS))
767 for (i = 0; i < BPF_REG_SIZE; i++) {
768 if (state->stack[spi].slot_type[i] == STACK_DYNPTR ||
769 state->stack[spi - 1].slot_type[i] == STACK_DYNPTR)
776 static bool is_dynptr_reg_valid_init(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
777 enum bpf_arg_type arg_type)
779 struct bpf_func_state *state = func(env, reg);
780 int spi = get_spi(reg->off);
783 if (!is_spi_bounds_valid(state, spi, BPF_DYNPTR_NR_SLOTS) ||
784 !state->stack[spi].spilled_ptr.dynptr.first_slot)
787 for (i = 0; i < BPF_REG_SIZE; i++) {
788 if (state->stack[spi].slot_type[i] != STACK_DYNPTR ||
789 state->stack[spi - 1].slot_type[i] != STACK_DYNPTR)
793 /* ARG_PTR_TO_DYNPTR takes any type of dynptr */
794 if (arg_type == ARG_PTR_TO_DYNPTR)
797 return state->stack[spi].spilled_ptr.dynptr.type == arg_to_dynptr_type(arg_type);
800 /* The reg state of a pointer or a bounded scalar was saved when
801 * it was spilled to the stack.
803 static bool is_spilled_reg(const struct bpf_stack_state *stack)
805 return stack->slot_type[BPF_REG_SIZE - 1] == STACK_SPILL;
808 static void scrub_spilled_slot(u8 *stype)
810 if (*stype != STACK_INVALID)
814 static void print_verifier_state(struct bpf_verifier_env *env,
815 const struct bpf_func_state *state,
818 const struct bpf_reg_state *reg;
823 verbose(env, " frame%d:", state->frameno);
824 for (i = 0; i < MAX_BPF_REG; i++) {
825 reg = &state->regs[i];
829 if (!print_all && !reg_scratched(env, i))
831 verbose(env, " R%d", i);
832 print_liveness(env, reg->live);
834 if (t == SCALAR_VALUE && reg->precise)
836 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
837 tnum_is_const(reg->var_off)) {
838 /* reg->off should be 0 for SCALAR_VALUE */
839 verbose(env, "%s", t == SCALAR_VALUE ? "" : reg_type_str(env, t));
840 verbose(env, "%lld", reg->var_off.value + reg->off);
842 const char *sep = "";
844 verbose(env, "%s", reg_type_str(env, t));
845 if (base_type(t) == PTR_TO_BTF_ID)
846 verbose(env, "%s", kernel_type_name(reg->btf, reg->btf_id));
849 * _a stands for append, was shortened to avoid multiline statements below.
850 * This macro is used to output a comma separated list of attributes.
852 #define verbose_a(fmt, ...) ({ verbose(env, "%s" fmt, sep, __VA_ARGS__); sep = ","; })
855 verbose_a("id=%d", reg->id);
856 if (reg_type_may_be_refcounted_or_null(t) && reg->ref_obj_id)
857 verbose_a("ref_obj_id=%d", reg->ref_obj_id);
858 if (t != SCALAR_VALUE)
859 verbose_a("off=%d", reg->off);
860 if (type_is_pkt_pointer(t))
861 verbose_a("r=%d", reg->range);
862 else if (base_type(t) == CONST_PTR_TO_MAP ||
863 base_type(t) == PTR_TO_MAP_KEY ||
864 base_type(t) == PTR_TO_MAP_VALUE)
865 verbose_a("ks=%d,vs=%d",
866 reg->map_ptr->key_size,
867 reg->map_ptr->value_size);
868 if (tnum_is_const(reg->var_off)) {
869 /* Typically an immediate SCALAR_VALUE, but
870 * could be a pointer whose offset is too big
873 verbose_a("imm=%llx", reg->var_off.value);
875 if (reg->smin_value != reg->umin_value &&
876 reg->smin_value != S64_MIN)
877 verbose_a("smin=%lld", (long long)reg->smin_value);
878 if (reg->smax_value != reg->umax_value &&
879 reg->smax_value != S64_MAX)
880 verbose_a("smax=%lld", (long long)reg->smax_value);
881 if (reg->umin_value != 0)
882 verbose_a("umin=%llu", (unsigned long long)reg->umin_value);
883 if (reg->umax_value != U64_MAX)
884 verbose_a("umax=%llu", (unsigned long long)reg->umax_value);
885 if (!tnum_is_unknown(reg->var_off)) {
888 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
889 verbose_a("var_off=%s", tn_buf);
891 if (reg->s32_min_value != reg->smin_value &&
892 reg->s32_min_value != S32_MIN)
893 verbose_a("s32_min=%d", (int)(reg->s32_min_value));
894 if (reg->s32_max_value != reg->smax_value &&
895 reg->s32_max_value != S32_MAX)
896 verbose_a("s32_max=%d", (int)(reg->s32_max_value));
897 if (reg->u32_min_value != reg->umin_value &&
898 reg->u32_min_value != U32_MIN)
899 verbose_a("u32_min=%d", (int)(reg->u32_min_value));
900 if (reg->u32_max_value != reg->umax_value &&
901 reg->u32_max_value != U32_MAX)
902 verbose_a("u32_max=%d", (int)(reg->u32_max_value));
909 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
910 char types_buf[BPF_REG_SIZE + 1];
914 for (j = 0; j < BPF_REG_SIZE; j++) {
915 if (state->stack[i].slot_type[j] != STACK_INVALID)
917 types_buf[j] = slot_type_char[
918 state->stack[i].slot_type[j]];
920 types_buf[BPF_REG_SIZE] = 0;
923 if (!print_all && !stack_slot_scratched(env, i))
925 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
926 print_liveness(env, state->stack[i].spilled_ptr.live);
927 if (is_spilled_reg(&state->stack[i])) {
928 reg = &state->stack[i].spilled_ptr;
930 verbose(env, "=%s", t == SCALAR_VALUE ? "" : reg_type_str(env, t));
931 if (t == SCALAR_VALUE && reg->precise)
933 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
934 verbose(env, "%lld", reg->var_off.value + reg->off);
936 verbose(env, "=%s", types_buf);
939 if (state->acquired_refs && state->refs[0].id) {
940 verbose(env, " refs=%d", state->refs[0].id);
941 for (i = 1; i < state->acquired_refs; i++)
942 if (state->refs[i].id)
943 verbose(env, ",%d", state->refs[i].id);
945 if (state->in_callback_fn)
947 if (state->in_async_callback_fn)
948 verbose(env, " async_cb");
950 mark_verifier_state_clean(env);
953 static inline u32 vlog_alignment(u32 pos)
955 return round_up(max(pos + BPF_LOG_MIN_ALIGNMENT / 2, BPF_LOG_ALIGNMENT),
956 BPF_LOG_MIN_ALIGNMENT) - pos - 1;
959 static void print_insn_state(struct bpf_verifier_env *env,
960 const struct bpf_func_state *state)
962 if (env->prev_log_len && env->prev_log_len == env->log.len_used) {
963 /* remove new line character */
964 bpf_vlog_reset(&env->log, env->prev_log_len - 1);
965 verbose(env, "%*c;", vlog_alignment(env->prev_insn_print_len), ' ');
967 verbose(env, "%d:", env->insn_idx);
969 print_verifier_state(env, state, false);
972 /* copy array src of length n * size bytes to dst. dst is reallocated if it's too
973 * small to hold src. This is different from krealloc since we don't want to preserve
974 * the contents of dst.
976 * Leaves dst untouched if src is NULL or length is zero. Returns NULL if memory could
979 static void *copy_array(void *dst, const void *src, size_t n, size_t size, gfp_t flags)
983 if (ZERO_OR_NULL_PTR(src))
986 if (unlikely(check_mul_overflow(n, size, &bytes)))
989 if (ksize(dst) < bytes) {
991 dst = kmalloc_track_caller(bytes, flags);
996 memcpy(dst, src, bytes);
998 return dst ? dst : ZERO_SIZE_PTR;
1001 /* resize an array from old_n items to new_n items. the array is reallocated if it's too
1002 * small to hold new_n items. new items are zeroed out if the array grows.
1004 * Contrary to krealloc_array, does not free arr if new_n is zero.
1006 static void *realloc_array(void *arr, size_t old_n, size_t new_n, size_t size)
1008 if (!new_n || old_n == new_n)
1011 arr = krealloc_array(arr, new_n, size, GFP_KERNEL);
1016 memset(arr + old_n * size, 0, (new_n - old_n) * size);
1019 return arr ? arr : ZERO_SIZE_PTR;
1022 static int copy_reference_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
1024 dst->refs = copy_array(dst->refs, src->refs, src->acquired_refs,
1025 sizeof(struct bpf_reference_state), GFP_KERNEL);
1029 dst->acquired_refs = src->acquired_refs;
1033 static int copy_stack_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
1035 size_t n = src->allocated_stack / BPF_REG_SIZE;
1037 dst->stack = copy_array(dst->stack, src->stack, n, sizeof(struct bpf_stack_state),
1042 dst->allocated_stack = src->allocated_stack;
1046 static int resize_reference_state(struct bpf_func_state *state, size_t n)
1048 state->refs = realloc_array(state->refs, state->acquired_refs, n,
1049 sizeof(struct bpf_reference_state));
1053 state->acquired_refs = n;
1057 static int grow_stack_state(struct bpf_func_state *state, int size)
1059 size_t old_n = state->allocated_stack / BPF_REG_SIZE, n = size / BPF_REG_SIZE;
1064 state->stack = realloc_array(state->stack, old_n, n, sizeof(struct bpf_stack_state));
1068 state->allocated_stack = size;
1072 /* Acquire a pointer id from the env and update the state->refs to include
1073 * this new pointer reference.
1074 * On success, returns a valid pointer id to associate with the register
1075 * On failure, returns a negative errno.
1077 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
1079 struct bpf_func_state *state = cur_func(env);
1080 int new_ofs = state->acquired_refs;
1083 err = resize_reference_state(state, state->acquired_refs + 1);
1087 state->refs[new_ofs].id = id;
1088 state->refs[new_ofs].insn_idx = insn_idx;
1093 /* release function corresponding to acquire_reference_state(). Idempotent. */
1094 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
1098 last_idx = state->acquired_refs - 1;
1099 for (i = 0; i < state->acquired_refs; i++) {
1100 if (state->refs[i].id == ptr_id) {
1101 if (last_idx && i != last_idx)
1102 memcpy(&state->refs[i], &state->refs[last_idx],
1103 sizeof(*state->refs));
1104 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
1105 state->acquired_refs--;
1112 static void free_func_state(struct bpf_func_state *state)
1117 kfree(state->stack);
1121 static void clear_jmp_history(struct bpf_verifier_state *state)
1123 kfree(state->jmp_history);
1124 state->jmp_history = NULL;
1125 state->jmp_history_cnt = 0;
1128 static void free_verifier_state(struct bpf_verifier_state *state,
1133 for (i = 0; i <= state->curframe; i++) {
1134 free_func_state(state->frame[i]);
1135 state->frame[i] = NULL;
1137 clear_jmp_history(state);
1142 /* copy verifier state from src to dst growing dst stack space
1143 * when necessary to accommodate larger src stack
1145 static int copy_func_state(struct bpf_func_state *dst,
1146 const struct bpf_func_state *src)
1150 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
1151 err = copy_reference_state(dst, src);
1154 return copy_stack_state(dst, src);
1157 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
1158 const struct bpf_verifier_state *src)
1160 struct bpf_func_state *dst;
1163 dst_state->jmp_history = copy_array(dst_state->jmp_history, src->jmp_history,
1164 src->jmp_history_cnt, sizeof(struct bpf_idx_pair),
1166 if (!dst_state->jmp_history)
1168 dst_state->jmp_history_cnt = src->jmp_history_cnt;
1170 /* if dst has more stack frames then src frame, free them */
1171 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
1172 free_func_state(dst_state->frame[i]);
1173 dst_state->frame[i] = NULL;
1175 dst_state->speculative = src->speculative;
1176 dst_state->curframe = src->curframe;
1177 dst_state->active_spin_lock = src->active_spin_lock;
1178 dst_state->branches = src->branches;
1179 dst_state->parent = src->parent;
1180 dst_state->first_insn_idx = src->first_insn_idx;
1181 dst_state->last_insn_idx = src->last_insn_idx;
1182 for (i = 0; i <= src->curframe; i++) {
1183 dst = dst_state->frame[i];
1185 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
1188 dst_state->frame[i] = dst;
1190 err = copy_func_state(dst, src->frame[i]);
1197 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
1200 u32 br = --st->branches;
1202 /* WARN_ON(br > 1) technically makes sense here,
1203 * but see comment in push_stack(), hence:
1205 WARN_ONCE((int)br < 0,
1206 "BUG update_branch_counts:branches_to_explore=%d\n",
1214 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
1215 int *insn_idx, bool pop_log)
1217 struct bpf_verifier_state *cur = env->cur_state;
1218 struct bpf_verifier_stack_elem *elem, *head = env->head;
1221 if (env->head == NULL)
1225 err = copy_verifier_state(cur, &head->st);
1230 bpf_vlog_reset(&env->log, head->log_pos);
1232 *insn_idx = head->insn_idx;
1234 *prev_insn_idx = head->prev_insn_idx;
1236 free_verifier_state(&head->st, false);
1243 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
1244 int insn_idx, int prev_insn_idx,
1247 struct bpf_verifier_state *cur = env->cur_state;
1248 struct bpf_verifier_stack_elem *elem;
1251 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1255 elem->insn_idx = insn_idx;
1256 elem->prev_insn_idx = prev_insn_idx;
1257 elem->next = env->head;
1258 elem->log_pos = env->log.len_used;
1261 err = copy_verifier_state(&elem->st, cur);
1264 elem->st.speculative |= speculative;
1265 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1266 verbose(env, "The sequence of %d jumps is too complex.\n",
1270 if (elem->st.parent) {
1271 ++elem->st.parent->branches;
1272 /* WARN_ON(branches > 2) technically makes sense here,
1274 * 1. speculative states will bump 'branches' for non-branch
1276 * 2. is_state_visited() heuristics may decide not to create
1277 * a new state for a sequence of branches and all such current
1278 * and cloned states will be pointing to a single parent state
1279 * which might have large 'branches' count.
1284 free_verifier_state(env->cur_state, true);
1285 env->cur_state = NULL;
1286 /* pop all elements and return */
1287 while (!pop_stack(env, NULL, NULL, false));
1291 #define CALLER_SAVED_REGS 6
1292 static const int caller_saved[CALLER_SAVED_REGS] = {
1293 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
1296 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1297 struct bpf_reg_state *reg);
1299 /* This helper doesn't clear reg->id */
1300 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1302 reg->var_off = tnum_const(imm);
1303 reg->smin_value = (s64)imm;
1304 reg->smax_value = (s64)imm;
1305 reg->umin_value = imm;
1306 reg->umax_value = imm;
1308 reg->s32_min_value = (s32)imm;
1309 reg->s32_max_value = (s32)imm;
1310 reg->u32_min_value = (u32)imm;
1311 reg->u32_max_value = (u32)imm;
1314 /* Mark the unknown part of a register (variable offset or scalar value) as
1315 * known to have the value @imm.
1317 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1319 /* Clear id, off, and union(map_ptr, range) */
1320 memset(((u8 *)reg) + sizeof(reg->type), 0,
1321 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1322 ___mark_reg_known(reg, imm);
1325 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1327 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1328 reg->s32_min_value = (s32)imm;
1329 reg->s32_max_value = (s32)imm;
1330 reg->u32_min_value = (u32)imm;
1331 reg->u32_max_value = (u32)imm;
1334 /* Mark the 'variable offset' part of a register as zero. This should be
1335 * used only on registers holding a pointer type.
1337 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1339 __mark_reg_known(reg, 0);
1342 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1344 __mark_reg_known(reg, 0);
1345 reg->type = SCALAR_VALUE;
1348 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1349 struct bpf_reg_state *regs, u32 regno)
1351 if (WARN_ON(regno >= MAX_BPF_REG)) {
1352 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1353 /* Something bad happened, let's kill all regs */
1354 for (regno = 0; regno < MAX_BPF_REG; regno++)
1355 __mark_reg_not_init(env, regs + regno);
1358 __mark_reg_known_zero(regs + regno);
1361 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
1363 if (base_type(reg->type) == PTR_TO_MAP_VALUE) {
1364 const struct bpf_map *map = reg->map_ptr;
1366 if (map->inner_map_meta) {
1367 reg->type = CONST_PTR_TO_MAP;
1368 reg->map_ptr = map->inner_map_meta;
1369 /* transfer reg's id which is unique for every map_lookup_elem
1370 * as UID of the inner map.
1372 if (map_value_has_timer(map->inner_map_meta))
1373 reg->map_uid = reg->id;
1374 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
1375 reg->type = PTR_TO_XDP_SOCK;
1376 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
1377 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
1378 reg->type = PTR_TO_SOCKET;
1380 reg->type = PTR_TO_MAP_VALUE;
1385 reg->type &= ~PTR_MAYBE_NULL;
1388 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1390 return type_is_pkt_pointer(reg->type);
1393 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
1395 return reg_is_pkt_pointer(reg) ||
1396 reg->type == PTR_TO_PACKET_END;
1399 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
1400 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
1401 enum bpf_reg_type which)
1403 /* The register can already have a range from prior markings.
1404 * This is fine as long as it hasn't been advanced from its
1407 return reg->type == which &&
1410 tnum_equals_const(reg->var_off, 0);
1413 /* Reset the min/max bounds of a register */
1414 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1416 reg->smin_value = S64_MIN;
1417 reg->smax_value = S64_MAX;
1418 reg->umin_value = 0;
1419 reg->umax_value = U64_MAX;
1421 reg->s32_min_value = S32_MIN;
1422 reg->s32_max_value = S32_MAX;
1423 reg->u32_min_value = 0;
1424 reg->u32_max_value = U32_MAX;
1427 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
1429 reg->smin_value = S64_MIN;
1430 reg->smax_value = S64_MAX;
1431 reg->umin_value = 0;
1432 reg->umax_value = U64_MAX;
1435 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
1437 reg->s32_min_value = S32_MIN;
1438 reg->s32_max_value = S32_MAX;
1439 reg->u32_min_value = 0;
1440 reg->u32_max_value = U32_MAX;
1443 static void __update_reg32_bounds(struct bpf_reg_state *reg)
1445 struct tnum var32_off = tnum_subreg(reg->var_off);
1447 /* min signed is max(sign bit) | min(other bits) */
1448 reg->s32_min_value = max_t(s32, reg->s32_min_value,
1449 var32_off.value | (var32_off.mask & S32_MIN));
1450 /* max signed is min(sign bit) | max(other bits) */
1451 reg->s32_max_value = min_t(s32, reg->s32_max_value,
1452 var32_off.value | (var32_off.mask & S32_MAX));
1453 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
1454 reg->u32_max_value = min(reg->u32_max_value,
1455 (u32)(var32_off.value | var32_off.mask));
1458 static void __update_reg64_bounds(struct bpf_reg_state *reg)
1460 /* min signed is max(sign bit) | min(other bits) */
1461 reg->smin_value = max_t(s64, reg->smin_value,
1462 reg->var_off.value | (reg->var_off.mask & S64_MIN));
1463 /* max signed is min(sign bit) | max(other bits) */
1464 reg->smax_value = min_t(s64, reg->smax_value,
1465 reg->var_off.value | (reg->var_off.mask & S64_MAX));
1466 reg->umin_value = max(reg->umin_value, reg->var_off.value);
1467 reg->umax_value = min(reg->umax_value,
1468 reg->var_off.value | reg->var_off.mask);
1471 static void __update_reg_bounds(struct bpf_reg_state *reg)
1473 __update_reg32_bounds(reg);
1474 __update_reg64_bounds(reg);
1477 /* Uses signed min/max values to inform unsigned, and vice-versa */
1478 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
1480 /* Learn sign from signed bounds.
1481 * If we cannot cross the sign boundary, then signed and unsigned bounds
1482 * are the same, so combine. This works even in the negative case, e.g.
1483 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1485 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
1486 reg->s32_min_value = reg->u32_min_value =
1487 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1488 reg->s32_max_value = reg->u32_max_value =
1489 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1492 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1493 * boundary, so we must be careful.
1495 if ((s32)reg->u32_max_value >= 0) {
1496 /* Positive. We can't learn anything from the smin, but smax
1497 * is positive, hence safe.
1499 reg->s32_min_value = reg->u32_min_value;
1500 reg->s32_max_value = reg->u32_max_value =
1501 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1502 } else if ((s32)reg->u32_min_value < 0) {
1503 /* Negative. We can't learn anything from the smax, but smin
1504 * is negative, hence safe.
1506 reg->s32_min_value = reg->u32_min_value =
1507 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1508 reg->s32_max_value = reg->u32_max_value;
1512 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
1514 /* Learn sign from signed bounds.
1515 * If we cannot cross the sign boundary, then signed and unsigned bounds
1516 * are the same, so combine. This works even in the negative case, e.g.
1517 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1519 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1520 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1522 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1526 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1527 * boundary, so we must be careful.
1529 if ((s64)reg->umax_value >= 0) {
1530 /* Positive. We can't learn anything from the smin, but smax
1531 * is positive, hence safe.
1533 reg->smin_value = reg->umin_value;
1534 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1536 } else if ((s64)reg->umin_value < 0) {
1537 /* Negative. We can't learn anything from the smax, but smin
1538 * is negative, hence safe.
1540 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1542 reg->smax_value = reg->umax_value;
1546 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
1548 __reg32_deduce_bounds(reg);
1549 __reg64_deduce_bounds(reg);
1552 /* Attempts to improve var_off based on unsigned min/max information */
1553 static void __reg_bound_offset(struct bpf_reg_state *reg)
1555 struct tnum var64_off = tnum_intersect(reg->var_off,
1556 tnum_range(reg->umin_value,
1558 struct tnum var32_off = tnum_intersect(tnum_subreg(reg->var_off),
1559 tnum_range(reg->u32_min_value,
1560 reg->u32_max_value));
1562 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
1565 static bool __reg32_bound_s64(s32 a)
1567 return a >= 0 && a <= S32_MAX;
1570 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
1572 reg->umin_value = reg->u32_min_value;
1573 reg->umax_value = reg->u32_max_value;
1575 /* Attempt to pull 32-bit signed bounds into 64-bit bounds but must
1576 * be positive otherwise set to worse case bounds and refine later
1579 if (__reg32_bound_s64(reg->s32_min_value) &&
1580 __reg32_bound_s64(reg->s32_max_value)) {
1581 reg->smin_value = reg->s32_min_value;
1582 reg->smax_value = reg->s32_max_value;
1584 reg->smin_value = 0;
1585 reg->smax_value = U32_MAX;
1589 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
1591 /* special case when 64-bit register has upper 32-bit register
1592 * zeroed. Typically happens after zext or <<32, >>32 sequence
1593 * allowing us to use 32-bit bounds directly,
1595 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
1596 __reg_assign_32_into_64(reg);
1598 /* Otherwise the best we can do is push lower 32bit known and
1599 * unknown bits into register (var_off set from jmp logic)
1600 * then learn as much as possible from the 64-bit tnum
1601 * known and unknown bits. The previous smin/smax bounds are
1602 * invalid here because of jmp32 compare so mark them unknown
1603 * so they do not impact tnum bounds calculation.
1605 __mark_reg64_unbounded(reg);
1606 __update_reg_bounds(reg);
1609 /* Intersecting with the old var_off might have improved our bounds
1610 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1611 * then new var_off is (0; 0x7f...fc) which improves our umax.
1613 __reg_deduce_bounds(reg);
1614 __reg_bound_offset(reg);
1615 __update_reg_bounds(reg);
1618 static bool __reg64_bound_s32(s64 a)
1620 return a >= S32_MIN && a <= S32_MAX;
1623 static bool __reg64_bound_u32(u64 a)
1625 return a >= U32_MIN && a <= U32_MAX;
1628 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
1630 __mark_reg32_unbounded(reg);
1632 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
1633 reg->s32_min_value = (s32)reg->smin_value;
1634 reg->s32_max_value = (s32)reg->smax_value;
1636 if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
1637 reg->u32_min_value = (u32)reg->umin_value;
1638 reg->u32_max_value = (u32)reg->umax_value;
1641 /* Intersecting with the old var_off might have improved our bounds
1642 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1643 * then new var_off is (0; 0x7f...fc) which improves our umax.
1645 __reg_deduce_bounds(reg);
1646 __reg_bound_offset(reg);
1647 __update_reg_bounds(reg);
1650 /* Mark a register as having a completely unknown (scalar) value. */
1651 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1652 struct bpf_reg_state *reg)
1655 * Clear type, id, off, and union(map_ptr, range) and
1656 * padding between 'type' and union
1658 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1659 reg->type = SCALAR_VALUE;
1660 reg->var_off = tnum_unknown;
1662 reg->precise = env->subprog_cnt > 1 || !env->bpf_capable;
1663 __mark_reg_unbounded(reg);
1666 static void mark_reg_unknown(struct bpf_verifier_env *env,
1667 struct bpf_reg_state *regs, u32 regno)
1669 if (WARN_ON(regno >= MAX_BPF_REG)) {
1670 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1671 /* Something bad happened, let's kill all regs except FP */
1672 for (regno = 0; regno < BPF_REG_FP; regno++)
1673 __mark_reg_not_init(env, regs + regno);
1676 __mark_reg_unknown(env, regs + regno);
1679 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1680 struct bpf_reg_state *reg)
1682 __mark_reg_unknown(env, reg);
1683 reg->type = NOT_INIT;
1686 static void mark_reg_not_init(struct bpf_verifier_env *env,
1687 struct bpf_reg_state *regs, u32 regno)
1689 if (WARN_ON(regno >= MAX_BPF_REG)) {
1690 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1691 /* Something bad happened, let's kill all regs except FP */
1692 for (regno = 0; regno < BPF_REG_FP; regno++)
1693 __mark_reg_not_init(env, regs + regno);
1696 __mark_reg_not_init(env, regs + regno);
1699 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
1700 struct bpf_reg_state *regs, u32 regno,
1701 enum bpf_reg_type reg_type,
1702 struct btf *btf, u32 btf_id,
1703 enum bpf_type_flag flag)
1705 if (reg_type == SCALAR_VALUE) {
1706 mark_reg_unknown(env, regs, regno);
1709 mark_reg_known_zero(env, regs, regno);
1710 regs[regno].type = PTR_TO_BTF_ID | flag;
1711 regs[regno].btf = btf;
1712 regs[regno].btf_id = btf_id;
1715 #define DEF_NOT_SUBREG (0)
1716 static void init_reg_state(struct bpf_verifier_env *env,
1717 struct bpf_func_state *state)
1719 struct bpf_reg_state *regs = state->regs;
1722 for (i = 0; i < MAX_BPF_REG; i++) {
1723 mark_reg_not_init(env, regs, i);
1724 regs[i].live = REG_LIVE_NONE;
1725 regs[i].parent = NULL;
1726 regs[i].subreg_def = DEF_NOT_SUBREG;
1730 regs[BPF_REG_FP].type = PTR_TO_STACK;
1731 mark_reg_known_zero(env, regs, BPF_REG_FP);
1732 regs[BPF_REG_FP].frameno = state->frameno;
1735 #define BPF_MAIN_FUNC (-1)
1736 static void init_func_state(struct bpf_verifier_env *env,
1737 struct bpf_func_state *state,
1738 int callsite, int frameno, int subprogno)
1740 state->callsite = callsite;
1741 state->frameno = frameno;
1742 state->subprogno = subprogno;
1743 init_reg_state(env, state);
1744 mark_verifier_state_scratched(env);
1747 /* Similar to push_stack(), but for async callbacks */
1748 static struct bpf_verifier_state *push_async_cb(struct bpf_verifier_env *env,
1749 int insn_idx, int prev_insn_idx,
1752 struct bpf_verifier_stack_elem *elem;
1753 struct bpf_func_state *frame;
1755 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1759 elem->insn_idx = insn_idx;
1760 elem->prev_insn_idx = prev_insn_idx;
1761 elem->next = env->head;
1762 elem->log_pos = env->log.len_used;
1765 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1767 "The sequence of %d jumps is too complex for async cb.\n",
1771 /* Unlike push_stack() do not copy_verifier_state().
1772 * The caller state doesn't matter.
1773 * This is async callback. It starts in a fresh stack.
1774 * Initialize it similar to do_check_common().
1776 elem->st.branches = 1;
1777 frame = kzalloc(sizeof(*frame), GFP_KERNEL);
1780 init_func_state(env, frame,
1781 BPF_MAIN_FUNC /* callsite */,
1782 0 /* frameno within this callchain */,
1783 subprog /* subprog number within this prog */);
1784 elem->st.frame[0] = frame;
1787 free_verifier_state(env->cur_state, true);
1788 env->cur_state = NULL;
1789 /* pop all elements and return */
1790 while (!pop_stack(env, NULL, NULL, false));
1796 SRC_OP, /* register is used as source operand */
1797 DST_OP, /* register is used as destination operand */
1798 DST_OP_NO_MARK /* same as above, check only, don't mark */
1801 static int cmp_subprogs(const void *a, const void *b)
1803 return ((struct bpf_subprog_info *)a)->start -
1804 ((struct bpf_subprog_info *)b)->start;
1807 static int find_subprog(struct bpf_verifier_env *env, int off)
1809 struct bpf_subprog_info *p;
1811 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1812 sizeof(env->subprog_info[0]), cmp_subprogs);
1815 return p - env->subprog_info;
1819 static int add_subprog(struct bpf_verifier_env *env, int off)
1821 int insn_cnt = env->prog->len;
1824 if (off >= insn_cnt || off < 0) {
1825 verbose(env, "call to invalid destination\n");
1828 ret = find_subprog(env, off);
1831 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1832 verbose(env, "too many subprograms\n");
1835 /* determine subprog starts. The end is one before the next starts */
1836 env->subprog_info[env->subprog_cnt++].start = off;
1837 sort(env->subprog_info, env->subprog_cnt,
1838 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1839 return env->subprog_cnt - 1;
1842 #define MAX_KFUNC_DESCS 256
1843 #define MAX_KFUNC_BTFS 256
1845 struct bpf_kfunc_desc {
1846 struct btf_func_model func_model;
1852 struct bpf_kfunc_btf {
1854 struct module *module;
1858 struct bpf_kfunc_desc_tab {
1859 struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
1863 struct bpf_kfunc_btf_tab {
1864 struct bpf_kfunc_btf descs[MAX_KFUNC_BTFS];
1868 static int kfunc_desc_cmp_by_id_off(const void *a, const void *b)
1870 const struct bpf_kfunc_desc *d0 = a;
1871 const struct bpf_kfunc_desc *d1 = b;
1873 /* func_id is not greater than BTF_MAX_TYPE */
1874 return d0->func_id - d1->func_id ?: d0->offset - d1->offset;
1877 static int kfunc_btf_cmp_by_off(const void *a, const void *b)
1879 const struct bpf_kfunc_btf *d0 = a;
1880 const struct bpf_kfunc_btf *d1 = b;
1882 return d0->offset - d1->offset;
1885 static const struct bpf_kfunc_desc *
1886 find_kfunc_desc(const struct bpf_prog *prog, u32 func_id, u16 offset)
1888 struct bpf_kfunc_desc desc = {
1892 struct bpf_kfunc_desc_tab *tab;
1894 tab = prog->aux->kfunc_tab;
1895 return bsearch(&desc, tab->descs, tab->nr_descs,
1896 sizeof(tab->descs[0]), kfunc_desc_cmp_by_id_off);
1899 static struct btf *__find_kfunc_desc_btf(struct bpf_verifier_env *env,
1902 struct bpf_kfunc_btf kf_btf = { .offset = offset };
1903 struct bpf_kfunc_btf_tab *tab;
1904 struct bpf_kfunc_btf *b;
1909 tab = env->prog->aux->kfunc_btf_tab;
1910 b = bsearch(&kf_btf, tab->descs, tab->nr_descs,
1911 sizeof(tab->descs[0]), kfunc_btf_cmp_by_off);
1913 if (tab->nr_descs == MAX_KFUNC_BTFS) {
1914 verbose(env, "too many different module BTFs\n");
1915 return ERR_PTR(-E2BIG);
1918 if (bpfptr_is_null(env->fd_array)) {
1919 verbose(env, "kfunc offset > 0 without fd_array is invalid\n");
1920 return ERR_PTR(-EPROTO);
1923 if (copy_from_bpfptr_offset(&btf_fd, env->fd_array,
1924 offset * sizeof(btf_fd),
1926 return ERR_PTR(-EFAULT);
1928 btf = btf_get_by_fd(btf_fd);
1930 verbose(env, "invalid module BTF fd specified\n");
1934 if (!btf_is_module(btf)) {
1935 verbose(env, "BTF fd for kfunc is not a module BTF\n");
1937 return ERR_PTR(-EINVAL);
1940 mod = btf_try_get_module(btf);
1943 return ERR_PTR(-ENXIO);
1946 b = &tab->descs[tab->nr_descs++];
1951 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1952 kfunc_btf_cmp_by_off, NULL);
1957 void bpf_free_kfunc_btf_tab(struct bpf_kfunc_btf_tab *tab)
1962 while (tab->nr_descs--) {
1963 module_put(tab->descs[tab->nr_descs].module);
1964 btf_put(tab->descs[tab->nr_descs].btf);
1969 static struct btf *find_kfunc_desc_btf(struct bpf_verifier_env *env, s16 offset)
1973 /* In the future, this can be allowed to increase limit
1974 * of fd index into fd_array, interpreted as u16.
1976 verbose(env, "negative offset disallowed for kernel module function call\n");
1977 return ERR_PTR(-EINVAL);
1980 return __find_kfunc_desc_btf(env, offset);
1982 return btf_vmlinux ?: ERR_PTR(-ENOENT);
1985 static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id, s16 offset)
1987 const struct btf_type *func, *func_proto;
1988 struct bpf_kfunc_btf_tab *btf_tab;
1989 struct bpf_kfunc_desc_tab *tab;
1990 struct bpf_prog_aux *prog_aux;
1991 struct bpf_kfunc_desc *desc;
1992 const char *func_name;
1993 struct btf *desc_btf;
1994 unsigned long call_imm;
1998 prog_aux = env->prog->aux;
1999 tab = prog_aux->kfunc_tab;
2000 btf_tab = prog_aux->kfunc_btf_tab;
2003 verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
2007 if (!env->prog->jit_requested) {
2008 verbose(env, "JIT is required for calling kernel function\n");
2012 if (!bpf_jit_supports_kfunc_call()) {
2013 verbose(env, "JIT does not support calling kernel function\n");
2017 if (!env->prog->gpl_compatible) {
2018 verbose(env, "cannot call kernel function from non-GPL compatible program\n");
2022 tab = kzalloc(sizeof(*tab), GFP_KERNEL);
2025 prog_aux->kfunc_tab = tab;
2028 /* func_id == 0 is always invalid, but instead of returning an error, be
2029 * conservative and wait until the code elimination pass before returning
2030 * error, so that invalid calls that get pruned out can be in BPF programs
2031 * loaded from userspace. It is also required that offset be untouched
2034 if (!func_id && !offset)
2037 if (!btf_tab && offset) {
2038 btf_tab = kzalloc(sizeof(*btf_tab), GFP_KERNEL);
2041 prog_aux->kfunc_btf_tab = btf_tab;
2044 desc_btf = find_kfunc_desc_btf(env, offset);
2045 if (IS_ERR(desc_btf)) {
2046 verbose(env, "failed to find BTF for kernel function\n");
2047 return PTR_ERR(desc_btf);
2050 if (find_kfunc_desc(env->prog, func_id, offset))
2053 if (tab->nr_descs == MAX_KFUNC_DESCS) {
2054 verbose(env, "too many different kernel function calls\n");
2058 func = btf_type_by_id(desc_btf, func_id);
2059 if (!func || !btf_type_is_func(func)) {
2060 verbose(env, "kernel btf_id %u is not a function\n",
2064 func_proto = btf_type_by_id(desc_btf, func->type);
2065 if (!func_proto || !btf_type_is_func_proto(func_proto)) {
2066 verbose(env, "kernel function btf_id %u does not have a valid func_proto\n",
2071 func_name = btf_name_by_offset(desc_btf, func->name_off);
2072 addr = kallsyms_lookup_name(func_name);
2074 verbose(env, "cannot find address for kernel function %s\n",
2079 call_imm = BPF_CALL_IMM(addr);
2080 /* Check whether or not the relative offset overflows desc->imm */
2081 if ((unsigned long)(s32)call_imm != call_imm) {
2082 verbose(env, "address of kernel function %s is out of range\n",
2087 desc = &tab->descs[tab->nr_descs++];
2088 desc->func_id = func_id;
2089 desc->imm = call_imm;
2090 desc->offset = offset;
2091 err = btf_distill_func_proto(&env->log, desc_btf,
2092 func_proto, func_name,
2095 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2096 kfunc_desc_cmp_by_id_off, NULL);
2100 static int kfunc_desc_cmp_by_imm(const void *a, const void *b)
2102 const struct bpf_kfunc_desc *d0 = a;
2103 const struct bpf_kfunc_desc *d1 = b;
2105 if (d0->imm > d1->imm)
2107 else if (d0->imm < d1->imm)
2112 static void sort_kfunc_descs_by_imm(struct bpf_prog *prog)
2114 struct bpf_kfunc_desc_tab *tab;
2116 tab = prog->aux->kfunc_tab;
2120 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2121 kfunc_desc_cmp_by_imm, NULL);
2124 bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
2126 return !!prog->aux->kfunc_tab;
2129 const struct btf_func_model *
2130 bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
2131 const struct bpf_insn *insn)
2133 const struct bpf_kfunc_desc desc = {
2136 const struct bpf_kfunc_desc *res;
2137 struct bpf_kfunc_desc_tab *tab;
2139 tab = prog->aux->kfunc_tab;
2140 res = bsearch(&desc, tab->descs, tab->nr_descs,
2141 sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm);
2143 return res ? &res->func_model : NULL;
2146 static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
2148 struct bpf_subprog_info *subprog = env->subprog_info;
2149 struct bpf_insn *insn = env->prog->insnsi;
2150 int i, ret, insn_cnt = env->prog->len;
2152 /* Add entry function. */
2153 ret = add_subprog(env, 0);
2157 for (i = 0; i < insn_cnt; i++, insn++) {
2158 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
2159 !bpf_pseudo_kfunc_call(insn))
2162 if (!env->bpf_capable) {
2163 verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
2167 if (bpf_pseudo_func(insn) || bpf_pseudo_call(insn))
2168 ret = add_subprog(env, i + insn->imm + 1);
2170 ret = add_kfunc_call(env, insn->imm, insn->off);
2176 /* Add a fake 'exit' subprog which could simplify subprog iteration
2177 * logic. 'subprog_cnt' should not be increased.
2179 subprog[env->subprog_cnt].start = insn_cnt;
2181 if (env->log.level & BPF_LOG_LEVEL2)
2182 for (i = 0; i < env->subprog_cnt; i++)
2183 verbose(env, "func#%d @%d\n", i, subprog[i].start);
2188 static int check_subprogs(struct bpf_verifier_env *env)
2190 int i, subprog_start, subprog_end, off, cur_subprog = 0;
2191 struct bpf_subprog_info *subprog = env->subprog_info;
2192 struct bpf_insn *insn = env->prog->insnsi;
2193 int insn_cnt = env->prog->len;
2195 /* now check that all jumps are within the same subprog */
2196 subprog_start = subprog[cur_subprog].start;
2197 subprog_end = subprog[cur_subprog + 1].start;
2198 for (i = 0; i < insn_cnt; i++) {
2199 u8 code = insn[i].code;
2201 if (code == (BPF_JMP | BPF_CALL) &&
2202 insn[i].imm == BPF_FUNC_tail_call &&
2203 insn[i].src_reg != BPF_PSEUDO_CALL)
2204 subprog[cur_subprog].has_tail_call = true;
2205 if (BPF_CLASS(code) == BPF_LD &&
2206 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
2207 subprog[cur_subprog].has_ld_abs = true;
2208 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
2210 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
2212 off = i + insn[i].off + 1;
2213 if (off < subprog_start || off >= subprog_end) {
2214 verbose(env, "jump out of range from insn %d to %d\n", i, off);
2218 if (i == subprog_end - 1) {
2219 /* to avoid fall-through from one subprog into another
2220 * the last insn of the subprog should be either exit
2221 * or unconditional jump back
2223 if (code != (BPF_JMP | BPF_EXIT) &&
2224 code != (BPF_JMP | BPF_JA)) {
2225 verbose(env, "last insn is not an exit or jmp\n");
2228 subprog_start = subprog_end;
2230 if (cur_subprog < env->subprog_cnt)
2231 subprog_end = subprog[cur_subprog + 1].start;
2237 /* Parentage chain of this register (or stack slot) should take care of all
2238 * issues like callee-saved registers, stack slot allocation time, etc.
2240 static int mark_reg_read(struct bpf_verifier_env *env,
2241 const struct bpf_reg_state *state,
2242 struct bpf_reg_state *parent, u8 flag)
2244 bool writes = parent == state->parent; /* Observe write marks */
2248 /* if read wasn't screened by an earlier write ... */
2249 if (writes && state->live & REG_LIVE_WRITTEN)
2251 if (parent->live & REG_LIVE_DONE) {
2252 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
2253 reg_type_str(env, parent->type),
2254 parent->var_off.value, parent->off);
2257 /* The first condition is more likely to be true than the
2258 * second, checked it first.
2260 if ((parent->live & REG_LIVE_READ) == flag ||
2261 parent->live & REG_LIVE_READ64)
2262 /* The parentage chain never changes and
2263 * this parent was already marked as LIVE_READ.
2264 * There is no need to keep walking the chain again and
2265 * keep re-marking all parents as LIVE_READ.
2266 * This case happens when the same register is read
2267 * multiple times without writes into it in-between.
2268 * Also, if parent has the stronger REG_LIVE_READ64 set,
2269 * then no need to set the weak REG_LIVE_READ32.
2272 /* ... then we depend on parent's value */
2273 parent->live |= flag;
2274 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
2275 if (flag == REG_LIVE_READ64)
2276 parent->live &= ~REG_LIVE_READ32;
2278 parent = state->parent;
2283 if (env->longest_mark_read_walk < cnt)
2284 env->longest_mark_read_walk = cnt;
2288 /* This function is supposed to be used by the following 32-bit optimization
2289 * code only. It returns TRUE if the source or destination register operates
2290 * on 64-bit, otherwise return FALSE.
2292 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
2293 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
2298 class = BPF_CLASS(code);
2300 if (class == BPF_JMP) {
2301 /* BPF_EXIT for "main" will reach here. Return TRUE
2306 if (op == BPF_CALL) {
2307 /* BPF to BPF call will reach here because of marking
2308 * caller saved clobber with DST_OP_NO_MARK for which we
2309 * don't care the register def because they are anyway
2310 * marked as NOT_INIT already.
2312 if (insn->src_reg == BPF_PSEUDO_CALL)
2314 /* Helper call will reach here because of arg type
2315 * check, conservatively return TRUE.
2324 if (class == BPF_ALU64 || class == BPF_JMP ||
2325 /* BPF_END always use BPF_ALU class. */
2326 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
2329 if (class == BPF_ALU || class == BPF_JMP32)
2332 if (class == BPF_LDX) {
2334 return BPF_SIZE(code) == BPF_DW;
2335 /* LDX source must be ptr. */
2339 if (class == BPF_STX) {
2340 /* BPF_STX (including atomic variants) has multiple source
2341 * operands, one of which is a ptr. Check whether the caller is
2344 if (t == SRC_OP && reg->type != SCALAR_VALUE)
2346 return BPF_SIZE(code) == BPF_DW;
2349 if (class == BPF_LD) {
2350 u8 mode = BPF_MODE(code);
2353 if (mode == BPF_IMM)
2356 /* Both LD_IND and LD_ABS return 32-bit data. */
2360 /* Implicit ctx ptr. */
2361 if (regno == BPF_REG_6)
2364 /* Explicit source could be any width. */
2368 if (class == BPF_ST)
2369 /* The only source register for BPF_ST is a ptr. */
2372 /* Conservatively return true at default. */
2376 /* Return the regno defined by the insn, or -1. */
2377 static int insn_def_regno(const struct bpf_insn *insn)
2379 switch (BPF_CLASS(insn->code)) {
2385 if (BPF_MODE(insn->code) == BPF_ATOMIC &&
2386 (insn->imm & BPF_FETCH)) {
2387 if (insn->imm == BPF_CMPXCHG)
2390 return insn->src_reg;
2395 return insn->dst_reg;
2399 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
2400 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
2402 int dst_reg = insn_def_regno(insn);
2407 return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
2410 static void mark_insn_zext(struct bpf_verifier_env *env,
2411 struct bpf_reg_state *reg)
2413 s32 def_idx = reg->subreg_def;
2415 if (def_idx == DEF_NOT_SUBREG)
2418 env->insn_aux_data[def_idx - 1].zext_dst = true;
2419 /* The dst will be zero extended, so won't be sub-register anymore. */
2420 reg->subreg_def = DEF_NOT_SUBREG;
2423 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
2424 enum reg_arg_type t)
2426 struct bpf_verifier_state *vstate = env->cur_state;
2427 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2428 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
2429 struct bpf_reg_state *reg, *regs = state->regs;
2432 if (regno >= MAX_BPF_REG) {
2433 verbose(env, "R%d is invalid\n", regno);
2437 mark_reg_scratched(env, regno);
2440 rw64 = is_reg64(env, insn, regno, reg, t);
2442 /* check whether register used as source operand can be read */
2443 if (reg->type == NOT_INIT) {
2444 verbose(env, "R%d !read_ok\n", regno);
2447 /* We don't need to worry about FP liveness because it's read-only */
2448 if (regno == BPF_REG_FP)
2452 mark_insn_zext(env, reg);
2454 return mark_reg_read(env, reg, reg->parent,
2455 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
2457 /* check whether register used as dest operand can be written to */
2458 if (regno == BPF_REG_FP) {
2459 verbose(env, "frame pointer is read only\n");
2462 reg->live |= REG_LIVE_WRITTEN;
2463 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
2465 mark_reg_unknown(env, regs, regno);
2470 /* for any branch, call, exit record the history of jmps in the given state */
2471 static int push_jmp_history(struct bpf_verifier_env *env,
2472 struct bpf_verifier_state *cur)
2474 u32 cnt = cur->jmp_history_cnt;
2475 struct bpf_idx_pair *p;
2478 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
2481 p[cnt - 1].idx = env->insn_idx;
2482 p[cnt - 1].prev_idx = env->prev_insn_idx;
2483 cur->jmp_history = p;
2484 cur->jmp_history_cnt = cnt;
2488 /* Backtrack one insn at a time. If idx is not at the top of recorded
2489 * history then previous instruction came from straight line execution.
2491 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
2496 if (cnt && st->jmp_history[cnt - 1].idx == i) {
2497 i = st->jmp_history[cnt - 1].prev_idx;
2505 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
2507 const struct btf_type *func;
2508 struct btf *desc_btf;
2510 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
2513 desc_btf = find_kfunc_desc_btf(data, insn->off);
2514 if (IS_ERR(desc_btf))
2517 func = btf_type_by_id(desc_btf, insn->imm);
2518 return btf_name_by_offset(desc_btf, func->name_off);
2521 /* For given verifier state backtrack_insn() is called from the last insn to
2522 * the first insn. Its purpose is to compute a bitmask of registers and
2523 * stack slots that needs precision in the parent verifier state.
2525 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
2526 u32 *reg_mask, u64 *stack_mask)
2528 const struct bpf_insn_cbs cbs = {
2529 .cb_call = disasm_kfunc_name,
2530 .cb_print = verbose,
2531 .private_data = env,
2533 struct bpf_insn *insn = env->prog->insnsi + idx;
2534 u8 class = BPF_CLASS(insn->code);
2535 u8 opcode = BPF_OP(insn->code);
2536 u8 mode = BPF_MODE(insn->code);
2537 u32 dreg = 1u << insn->dst_reg;
2538 u32 sreg = 1u << insn->src_reg;
2541 if (insn->code == 0)
2543 if (env->log.level & BPF_LOG_LEVEL2) {
2544 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
2545 verbose(env, "%d: ", idx);
2546 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
2549 if (class == BPF_ALU || class == BPF_ALU64) {
2550 if (!(*reg_mask & dreg))
2552 if (opcode == BPF_MOV) {
2553 if (BPF_SRC(insn->code) == BPF_X) {
2555 * dreg needs precision after this insn
2556 * sreg needs precision before this insn
2562 * dreg needs precision after this insn.
2563 * Corresponding register is already marked
2564 * as precise=true in this verifier state.
2565 * No further markings in parent are necessary
2570 if (BPF_SRC(insn->code) == BPF_X) {
2572 * both dreg and sreg need precision
2577 * dreg still needs precision before this insn
2580 } else if (class == BPF_LDX) {
2581 if (!(*reg_mask & dreg))
2585 /* scalars can only be spilled into stack w/o losing precision.
2586 * Load from any other memory can be zero extended.
2587 * The desire to keep that precision is already indicated
2588 * by 'precise' mark in corresponding register of this state.
2589 * No further tracking necessary.
2591 if (insn->src_reg != BPF_REG_FP)
2594 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
2595 * that [fp - off] slot contains scalar that needs to be
2596 * tracked with precision
2598 spi = (-insn->off - 1) / BPF_REG_SIZE;
2600 verbose(env, "BUG spi %d\n", spi);
2601 WARN_ONCE(1, "verifier backtracking bug");
2604 *stack_mask |= 1ull << spi;
2605 } else if (class == BPF_STX || class == BPF_ST) {
2606 if (*reg_mask & dreg)
2607 /* stx & st shouldn't be using _scalar_ dst_reg
2608 * to access memory. It means backtracking
2609 * encountered a case of pointer subtraction.
2612 /* scalars can only be spilled into stack */
2613 if (insn->dst_reg != BPF_REG_FP)
2615 spi = (-insn->off - 1) / BPF_REG_SIZE;
2617 verbose(env, "BUG spi %d\n", spi);
2618 WARN_ONCE(1, "verifier backtracking bug");
2621 if (!(*stack_mask & (1ull << spi)))
2623 *stack_mask &= ~(1ull << spi);
2624 if (class == BPF_STX)
2626 } else if (class == BPF_JMP || class == BPF_JMP32) {
2627 if (opcode == BPF_CALL) {
2628 if (insn->src_reg == BPF_PSEUDO_CALL)
2630 /* regular helper call sets R0 */
2632 if (*reg_mask & 0x3f) {
2633 /* if backtracing was looking for registers R1-R5
2634 * they should have been found already.
2636 verbose(env, "BUG regs %x\n", *reg_mask);
2637 WARN_ONCE(1, "verifier backtracking bug");
2640 } else if (opcode == BPF_EXIT) {
2643 } else if (class == BPF_LD) {
2644 if (!(*reg_mask & dreg))
2647 /* It's ld_imm64 or ld_abs or ld_ind.
2648 * For ld_imm64 no further tracking of precision
2649 * into parent is necessary
2651 if (mode == BPF_IND || mode == BPF_ABS)
2652 /* to be analyzed */
2658 /* the scalar precision tracking algorithm:
2659 * . at the start all registers have precise=false.
2660 * . scalar ranges are tracked as normal through alu and jmp insns.
2661 * . once precise value of the scalar register is used in:
2662 * . ptr + scalar alu
2663 * . if (scalar cond K|scalar)
2664 * . helper_call(.., scalar, ...) where ARG_CONST is expected
2665 * backtrack through the verifier states and mark all registers and
2666 * stack slots with spilled constants that these scalar regisers
2667 * should be precise.
2668 * . during state pruning two registers (or spilled stack slots)
2669 * are equivalent if both are not precise.
2671 * Note the verifier cannot simply walk register parentage chain,
2672 * since many different registers and stack slots could have been
2673 * used to compute single precise scalar.
2675 * The approach of starting with precise=true for all registers and then
2676 * backtrack to mark a register as not precise when the verifier detects
2677 * that program doesn't care about specific value (e.g., when helper
2678 * takes register as ARG_ANYTHING parameter) is not safe.
2680 * It's ok to walk single parentage chain of the verifier states.
2681 * It's possible that this backtracking will go all the way till 1st insn.
2682 * All other branches will be explored for needing precision later.
2684 * The backtracking needs to deal with cases like:
2685 * 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)
2688 * if r5 > 0x79f goto pc+7
2689 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
2692 * call bpf_perf_event_output#25
2693 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
2697 * call foo // uses callee's r6 inside to compute r0
2701 * to track above reg_mask/stack_mask needs to be independent for each frame.
2703 * Also if parent's curframe > frame where backtracking started,
2704 * the verifier need to mark registers in both frames, otherwise callees
2705 * may incorrectly prune callers. This is similar to
2706 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
2708 * For now backtracking falls back into conservative marking.
2710 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
2711 struct bpf_verifier_state *st)
2713 struct bpf_func_state *func;
2714 struct bpf_reg_state *reg;
2717 /* big hammer: mark all scalars precise in this path.
2718 * pop_stack may still get !precise scalars.
2720 for (; st; st = st->parent)
2721 for (i = 0; i <= st->curframe; i++) {
2722 func = st->frame[i];
2723 for (j = 0; j < BPF_REG_FP; j++) {
2724 reg = &func->regs[j];
2725 if (reg->type != SCALAR_VALUE)
2727 reg->precise = true;
2729 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
2730 if (!is_spilled_reg(&func->stack[j]))
2732 reg = &func->stack[j].spilled_ptr;
2733 if (reg->type != SCALAR_VALUE)
2735 reg->precise = true;
2740 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
2743 struct bpf_verifier_state *st = env->cur_state;
2744 int first_idx = st->first_insn_idx;
2745 int last_idx = env->insn_idx;
2746 struct bpf_func_state *func;
2747 struct bpf_reg_state *reg;
2748 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
2749 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
2750 bool skip_first = true;
2751 bool new_marks = false;
2754 if (!env->bpf_capable)
2757 func = st->frame[st->curframe];
2759 reg = &func->regs[regno];
2760 if (reg->type != SCALAR_VALUE) {
2761 WARN_ONCE(1, "backtracing misuse");
2768 reg->precise = true;
2772 if (!is_spilled_reg(&func->stack[spi])) {
2776 reg = &func->stack[spi].spilled_ptr;
2777 if (reg->type != SCALAR_VALUE) {
2785 reg->precise = true;
2791 if (!reg_mask && !stack_mask)
2794 DECLARE_BITMAP(mask, 64);
2795 u32 history = st->jmp_history_cnt;
2797 if (env->log.level & BPF_LOG_LEVEL2)
2798 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
2799 for (i = last_idx;;) {
2804 err = backtrack_insn(env, i, ®_mask, &stack_mask);
2806 if (err == -ENOTSUPP) {
2807 mark_all_scalars_precise(env, st);
2812 if (!reg_mask && !stack_mask)
2813 /* Found assignment(s) into tracked register in this state.
2814 * Since this state is already marked, just return.
2815 * Nothing to be tracked further in the parent state.
2820 i = get_prev_insn_idx(st, i, &history);
2821 if (i >= env->prog->len) {
2822 /* This can happen if backtracking reached insn 0
2823 * and there are still reg_mask or stack_mask
2825 * It means the backtracking missed the spot where
2826 * particular register was initialized with a constant.
2828 verbose(env, "BUG backtracking idx %d\n", i);
2829 WARN_ONCE(1, "verifier backtracking bug");
2838 func = st->frame[st->curframe];
2839 bitmap_from_u64(mask, reg_mask);
2840 for_each_set_bit(i, mask, 32) {
2841 reg = &func->regs[i];
2842 if (reg->type != SCALAR_VALUE) {
2843 reg_mask &= ~(1u << i);
2848 reg->precise = true;
2851 bitmap_from_u64(mask, stack_mask);
2852 for_each_set_bit(i, mask, 64) {
2853 if (i >= func->allocated_stack / BPF_REG_SIZE) {
2854 /* the sequence of instructions:
2856 * 3: (7b) *(u64 *)(r3 -8) = r0
2857 * 4: (79) r4 = *(u64 *)(r10 -8)
2858 * doesn't contain jmps. It's backtracked
2859 * as a single block.
2860 * During backtracking insn 3 is not recognized as
2861 * stack access, so at the end of backtracking
2862 * stack slot fp-8 is still marked in stack_mask.
2863 * However the parent state may not have accessed
2864 * fp-8 and it's "unallocated" stack space.
2865 * In such case fallback to conservative.
2867 mark_all_scalars_precise(env, st);
2871 if (!is_spilled_reg(&func->stack[i])) {
2872 stack_mask &= ~(1ull << i);
2875 reg = &func->stack[i].spilled_ptr;
2876 if (reg->type != SCALAR_VALUE) {
2877 stack_mask &= ~(1ull << i);
2882 reg->precise = true;
2884 if (env->log.level & BPF_LOG_LEVEL2) {
2885 verbose(env, "parent %s regs=%x stack=%llx marks:",
2886 new_marks ? "didn't have" : "already had",
2887 reg_mask, stack_mask);
2888 print_verifier_state(env, func, true);
2891 if (!reg_mask && !stack_mask)
2896 last_idx = st->last_insn_idx;
2897 first_idx = st->first_insn_idx;
2902 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
2904 return __mark_chain_precision(env, regno, -1);
2907 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
2909 return __mark_chain_precision(env, -1, spi);
2912 static bool is_spillable_regtype(enum bpf_reg_type type)
2914 switch (base_type(type)) {
2915 case PTR_TO_MAP_VALUE:
2919 case PTR_TO_PACKET_META:
2920 case PTR_TO_PACKET_END:
2921 case PTR_TO_FLOW_KEYS:
2922 case CONST_PTR_TO_MAP:
2924 case PTR_TO_SOCK_COMMON:
2925 case PTR_TO_TCP_SOCK:
2926 case PTR_TO_XDP_SOCK:
2931 case PTR_TO_MAP_KEY:
2938 /* Does this register contain a constant zero? */
2939 static bool register_is_null(struct bpf_reg_state *reg)
2941 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
2944 static bool register_is_const(struct bpf_reg_state *reg)
2946 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
2949 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
2951 return tnum_is_unknown(reg->var_off) &&
2952 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
2953 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
2954 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
2955 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
2958 static bool register_is_bounded(struct bpf_reg_state *reg)
2960 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
2963 static bool __is_pointer_value(bool allow_ptr_leaks,
2964 const struct bpf_reg_state *reg)
2966 if (allow_ptr_leaks)
2969 return reg->type != SCALAR_VALUE;
2972 static void save_register_state(struct bpf_func_state *state,
2973 int spi, struct bpf_reg_state *reg,
2978 state->stack[spi].spilled_ptr = *reg;
2979 if (size == BPF_REG_SIZE)
2980 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2982 for (i = BPF_REG_SIZE; i > BPF_REG_SIZE - size; i--)
2983 state->stack[spi].slot_type[i - 1] = STACK_SPILL;
2985 /* size < 8 bytes spill */
2987 scrub_spilled_slot(&state->stack[spi].slot_type[i - 1]);
2990 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
2991 * stack boundary and alignment are checked in check_mem_access()
2993 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
2994 /* stack frame we're writing to */
2995 struct bpf_func_state *state,
2996 int off, int size, int value_regno,
2999 struct bpf_func_state *cur; /* state of the current function */
3000 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
3001 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
3002 struct bpf_reg_state *reg = NULL;
3004 err = grow_stack_state(state, round_up(slot + 1, BPF_REG_SIZE));
3007 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
3008 * so it's aligned access and [off, off + size) are within stack limits
3010 if (!env->allow_ptr_leaks &&
3011 state->stack[spi].slot_type[0] == STACK_SPILL &&
3012 size != BPF_REG_SIZE) {
3013 verbose(env, "attempt to corrupt spilled pointer on stack\n");
3017 cur = env->cur_state->frame[env->cur_state->curframe];
3018 if (value_regno >= 0)
3019 reg = &cur->regs[value_regno];
3020 if (!env->bypass_spec_v4) {
3021 bool sanitize = reg && is_spillable_regtype(reg->type);
3023 for (i = 0; i < size; i++) {
3024 if (state->stack[spi].slot_type[i] == STACK_INVALID) {
3031 env->insn_aux_data[insn_idx].sanitize_stack_spill = true;
3034 mark_stack_slot_scratched(env, spi);
3035 if (reg && !(off % BPF_REG_SIZE) && register_is_bounded(reg) &&
3036 !register_is_null(reg) && env->bpf_capable) {
3037 if (dst_reg != BPF_REG_FP) {
3038 /* The backtracking logic can only recognize explicit
3039 * stack slot address like [fp - 8]. Other spill of
3040 * scalar via different register has to be conservative.
3041 * Backtrack from here and mark all registers as precise
3042 * that contributed into 'reg' being a constant.
3044 err = mark_chain_precision(env, value_regno);
3048 save_register_state(state, spi, reg, size);
3049 } else if (reg && is_spillable_regtype(reg->type)) {
3050 /* register containing pointer is being spilled into stack */
3051 if (size != BPF_REG_SIZE) {
3052 verbose_linfo(env, insn_idx, "; ");
3053 verbose(env, "invalid size of register spill\n");
3056 if (state != cur && reg->type == PTR_TO_STACK) {
3057 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
3060 save_register_state(state, spi, reg, size);
3062 u8 type = STACK_MISC;
3064 /* regular write of data into stack destroys any spilled ptr */
3065 state->stack[spi].spilled_ptr.type = NOT_INIT;
3066 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
3067 if (is_spilled_reg(&state->stack[spi]))
3068 for (i = 0; i < BPF_REG_SIZE; i++)
3069 scrub_spilled_slot(&state->stack[spi].slot_type[i]);
3071 /* only mark the slot as written if all 8 bytes were written
3072 * otherwise read propagation may incorrectly stop too soon
3073 * when stack slots are partially written.
3074 * This heuristic means that read propagation will be
3075 * conservative, since it will add reg_live_read marks
3076 * to stack slots all the way to first state when programs
3077 * writes+reads less than 8 bytes
3079 if (size == BPF_REG_SIZE)
3080 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
3082 /* when we zero initialize stack slots mark them as such */
3083 if (reg && register_is_null(reg)) {
3084 /* backtracking doesn't work for STACK_ZERO yet. */
3085 err = mark_chain_precision(env, value_regno);
3091 /* Mark slots affected by this stack write. */
3092 for (i = 0; i < size; i++)
3093 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
3099 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
3100 * known to contain a variable offset.
3101 * This function checks whether the write is permitted and conservatively
3102 * tracks the effects of the write, considering that each stack slot in the
3103 * dynamic range is potentially written to.
3105 * 'off' includes 'regno->off'.
3106 * 'value_regno' can be -1, meaning that an unknown value is being written to
3109 * Spilled pointers in range are not marked as written because we don't know
3110 * what's going to be actually written. This means that read propagation for
3111 * future reads cannot be terminated by this write.
3113 * For privileged programs, uninitialized stack slots are considered
3114 * initialized by this write (even though we don't know exactly what offsets
3115 * are going to be written to). The idea is that we don't want the verifier to
3116 * reject future reads that access slots written to through variable offsets.
3118 static int check_stack_write_var_off(struct bpf_verifier_env *env,
3119 /* func where register points to */
3120 struct bpf_func_state *state,
3121 int ptr_regno, int off, int size,
3122 int value_regno, int insn_idx)
3124 struct bpf_func_state *cur; /* state of the current function */
3125 int min_off, max_off;
3127 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
3128 bool writing_zero = false;
3129 /* set if the fact that we're writing a zero is used to let any
3130 * stack slots remain STACK_ZERO
3132 bool zero_used = false;
3134 cur = env->cur_state->frame[env->cur_state->curframe];
3135 ptr_reg = &cur->regs[ptr_regno];
3136 min_off = ptr_reg->smin_value + off;
3137 max_off = ptr_reg->smax_value + off + size;
3138 if (value_regno >= 0)
3139 value_reg = &cur->regs[value_regno];
3140 if (value_reg && register_is_null(value_reg))
3141 writing_zero = true;
3143 err = grow_stack_state(state, round_up(-min_off, BPF_REG_SIZE));
3148 /* Variable offset writes destroy any spilled pointers in range. */
3149 for (i = min_off; i < max_off; i++) {
3150 u8 new_type, *stype;
3154 spi = slot / BPF_REG_SIZE;
3155 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
3156 mark_stack_slot_scratched(env, spi);
3158 if (!env->allow_ptr_leaks
3159 && *stype != NOT_INIT
3160 && *stype != SCALAR_VALUE) {
3161 /* Reject the write if there's are spilled pointers in
3162 * range. If we didn't reject here, the ptr status
3163 * would be erased below (even though not all slots are
3164 * actually overwritten), possibly opening the door to
3167 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
3172 /* Erase all spilled pointers. */
3173 state->stack[spi].spilled_ptr.type = NOT_INIT;
3175 /* Update the slot type. */
3176 new_type = STACK_MISC;
3177 if (writing_zero && *stype == STACK_ZERO) {
3178 new_type = STACK_ZERO;
3181 /* If the slot is STACK_INVALID, we check whether it's OK to
3182 * pretend that it will be initialized by this write. The slot
3183 * might not actually be written to, and so if we mark it as
3184 * initialized future reads might leak uninitialized memory.
3185 * For privileged programs, we will accept such reads to slots
3186 * that may or may not be written because, if we're reject
3187 * them, the error would be too confusing.
3189 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
3190 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
3197 /* backtracking doesn't work for STACK_ZERO yet. */
3198 err = mark_chain_precision(env, value_regno);
3205 /* When register 'dst_regno' is assigned some values from stack[min_off,
3206 * max_off), we set the register's type according to the types of the
3207 * respective stack slots. If all the stack values are known to be zeros, then
3208 * so is the destination reg. Otherwise, the register is considered to be
3209 * SCALAR. This function does not deal with register filling; the caller must
3210 * ensure that all spilled registers in the stack range have been marked as
3213 static void mark_reg_stack_read(struct bpf_verifier_env *env,
3214 /* func where src register points to */
3215 struct bpf_func_state *ptr_state,
3216 int min_off, int max_off, int dst_regno)
3218 struct bpf_verifier_state *vstate = env->cur_state;
3219 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3224 for (i = min_off; i < max_off; i++) {
3226 spi = slot / BPF_REG_SIZE;
3227 stype = ptr_state->stack[spi].slot_type;
3228 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
3232 if (zeros == max_off - min_off) {
3233 /* any access_size read into register is zero extended,
3234 * so the whole register == const_zero
3236 __mark_reg_const_zero(&state->regs[dst_regno]);
3237 /* backtracking doesn't support STACK_ZERO yet,
3238 * so mark it precise here, so that later
3239 * backtracking can stop here.
3240 * Backtracking may not need this if this register
3241 * doesn't participate in pointer adjustment.
3242 * Forward propagation of precise flag is not
3243 * necessary either. This mark is only to stop
3244 * backtracking. Any register that contributed
3245 * to const 0 was marked precise before spill.
3247 state->regs[dst_regno].precise = true;
3249 /* have read misc data from the stack */
3250 mark_reg_unknown(env, state->regs, dst_regno);
3252 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
3255 /* Read the stack at 'off' and put the results into the register indicated by
3256 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
3259 * 'dst_regno' can be -1, meaning that the read value is not going to a
3262 * The access is assumed to be within the current stack bounds.
3264 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
3265 /* func where src register points to */
3266 struct bpf_func_state *reg_state,
3267 int off, int size, int dst_regno)
3269 struct bpf_verifier_state *vstate = env->cur_state;
3270 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3271 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
3272 struct bpf_reg_state *reg;
3275 stype = reg_state->stack[spi].slot_type;
3276 reg = ®_state->stack[spi].spilled_ptr;
3278 if (is_spilled_reg(®_state->stack[spi])) {
3281 for (i = BPF_REG_SIZE - 1; i > 0 && stype[i - 1] == STACK_SPILL; i--)
3284 if (size != BPF_REG_SIZE || spill_size != BPF_REG_SIZE) {
3285 if (reg->type != SCALAR_VALUE) {
3286 verbose_linfo(env, env->insn_idx, "; ");
3287 verbose(env, "invalid size of register fill\n");
3291 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
3295 if (!(off % BPF_REG_SIZE) && size == spill_size) {
3296 /* The earlier check_reg_arg() has decided the
3297 * subreg_def for this insn. Save it first.
3299 s32 subreg_def = state->regs[dst_regno].subreg_def;
3301 state->regs[dst_regno] = *reg;
3302 state->regs[dst_regno].subreg_def = subreg_def;
3304 for (i = 0; i < size; i++) {
3305 type = stype[(slot - i) % BPF_REG_SIZE];
3306 if (type == STACK_SPILL)
3308 if (type == STACK_MISC)
3310 verbose(env, "invalid read from stack off %d+%d size %d\n",
3314 mark_reg_unknown(env, state->regs, dst_regno);
3316 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
3320 if (dst_regno >= 0) {
3321 /* restore register state from stack */
3322 state->regs[dst_regno] = *reg;
3323 /* mark reg as written since spilled pointer state likely
3324 * has its liveness marks cleared by is_state_visited()
3325 * which resets stack/reg liveness for state transitions
3327 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
3328 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
3329 /* If dst_regno==-1, the caller is asking us whether
3330 * it is acceptable to use this value as a SCALAR_VALUE
3332 * We must not allow unprivileged callers to do that
3333 * with spilled pointers.
3335 verbose(env, "leaking pointer from stack off %d\n",
3339 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
3341 for (i = 0; i < size; i++) {
3342 type = stype[(slot - i) % BPF_REG_SIZE];
3343 if (type == STACK_MISC)
3345 if (type == STACK_ZERO)
3347 verbose(env, "invalid read from stack off %d+%d size %d\n",
3351 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
3353 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
3358 enum bpf_access_src {
3359 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
3360 ACCESS_HELPER = 2, /* the access is performed by a helper */
3363 static int check_stack_range_initialized(struct bpf_verifier_env *env,
3364 int regno, int off, int access_size,
3365 bool zero_size_allowed,
3366 enum bpf_access_src type,
3367 struct bpf_call_arg_meta *meta);
3369 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
3371 return cur_regs(env) + regno;
3374 /* Read the stack at 'ptr_regno + off' and put the result into the register
3376 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
3377 * but not its variable offset.
3378 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
3380 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
3381 * filling registers (i.e. reads of spilled register cannot be detected when
3382 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
3383 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
3384 * offset; for a fixed offset check_stack_read_fixed_off should be used
3387 static int check_stack_read_var_off(struct bpf_verifier_env *env,
3388 int ptr_regno, int off, int size, int dst_regno)
3390 /* The state of the source register. */
3391 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3392 struct bpf_func_state *ptr_state = func(env, reg);
3394 int min_off, max_off;
3396 /* Note that we pass a NULL meta, so raw access will not be permitted.
3398 err = check_stack_range_initialized(env, ptr_regno, off, size,
3399 false, ACCESS_DIRECT, NULL);
3403 min_off = reg->smin_value + off;
3404 max_off = reg->smax_value + off;
3405 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
3409 /* check_stack_read dispatches to check_stack_read_fixed_off or
3410 * check_stack_read_var_off.
3412 * The caller must ensure that the offset falls within the allocated stack
3415 * 'dst_regno' is a register which will receive the value from the stack. It
3416 * can be -1, meaning that the read value is not going to a register.
3418 static int check_stack_read(struct bpf_verifier_env *env,
3419 int ptr_regno, int off, int size,
3422 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3423 struct bpf_func_state *state = func(env, reg);
3425 /* Some accesses are only permitted with a static offset. */
3426 bool var_off = !tnum_is_const(reg->var_off);
3428 /* The offset is required to be static when reads don't go to a
3429 * register, in order to not leak pointers (see
3430 * check_stack_read_fixed_off).
3432 if (dst_regno < 0 && var_off) {
3435 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3436 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
3440 /* Variable offset is prohibited for unprivileged mode for simplicity
3441 * since it requires corresponding support in Spectre masking for stack
3442 * ALU. See also retrieve_ptr_limit().
3444 if (!env->bypass_spec_v1 && var_off) {
3447 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3448 verbose(env, "R%d variable offset stack access prohibited for !root, var_off=%s\n",
3454 off += reg->var_off.value;
3455 err = check_stack_read_fixed_off(env, state, off, size,
3458 /* Variable offset stack reads need more conservative handling
3459 * than fixed offset ones. Note that dst_regno >= 0 on this
3462 err = check_stack_read_var_off(env, ptr_regno, off, size,
3469 /* check_stack_write dispatches to check_stack_write_fixed_off or
3470 * check_stack_write_var_off.
3472 * 'ptr_regno' is the register used as a pointer into the stack.
3473 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
3474 * 'value_regno' is the register whose value we're writing to the stack. It can
3475 * be -1, meaning that we're not writing from a register.
3477 * The caller must ensure that the offset falls within the maximum stack size.
3479 static int check_stack_write(struct bpf_verifier_env *env,
3480 int ptr_regno, int off, int size,
3481 int value_regno, int insn_idx)
3483 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3484 struct bpf_func_state *state = func(env, reg);
3487 if (tnum_is_const(reg->var_off)) {
3488 off += reg->var_off.value;
3489 err = check_stack_write_fixed_off(env, state, off, size,
3490 value_regno, insn_idx);
3492 /* Variable offset stack reads need more conservative handling
3493 * than fixed offset ones.
3495 err = check_stack_write_var_off(env, state,
3496 ptr_regno, off, size,
3497 value_regno, insn_idx);
3502 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
3503 int off, int size, enum bpf_access_type type)
3505 struct bpf_reg_state *regs = cur_regs(env);
3506 struct bpf_map *map = regs[regno].map_ptr;
3507 u32 cap = bpf_map_flags_to_cap(map);
3509 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
3510 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
3511 map->value_size, off, size);
3515 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
3516 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
3517 map->value_size, off, size);
3524 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
3525 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
3526 int off, int size, u32 mem_size,
3527 bool zero_size_allowed)
3529 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
3530 struct bpf_reg_state *reg;
3532 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
3535 reg = &cur_regs(env)[regno];
3536 switch (reg->type) {
3537 case PTR_TO_MAP_KEY:
3538 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
3539 mem_size, off, size);
3541 case PTR_TO_MAP_VALUE:
3542 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
3543 mem_size, off, size);
3546 case PTR_TO_PACKET_META:
3547 case PTR_TO_PACKET_END:
3548 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
3549 off, size, regno, reg->id, off, mem_size);
3553 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
3554 mem_size, off, size);
3560 /* check read/write into a memory region with possible variable offset */
3561 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
3562 int off, int size, u32 mem_size,
3563 bool zero_size_allowed)
3565 struct bpf_verifier_state *vstate = env->cur_state;
3566 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3567 struct bpf_reg_state *reg = &state->regs[regno];
3570 /* We may have adjusted the register pointing to memory region, so we
3571 * need to try adding each of min_value and max_value to off
3572 * to make sure our theoretical access will be safe.
3574 * The minimum value is only important with signed
3575 * comparisons where we can't assume the floor of a
3576 * value is 0. If we are using signed variables for our
3577 * index'es we need to make sure that whatever we use
3578 * will have a set floor within our range.
3580 if (reg->smin_value < 0 &&
3581 (reg->smin_value == S64_MIN ||
3582 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
3583 reg->smin_value + off < 0)) {
3584 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3588 err = __check_mem_access(env, regno, reg->smin_value + off, size,
3589 mem_size, zero_size_allowed);
3591 verbose(env, "R%d min value is outside of the allowed memory range\n",
3596 /* If we haven't set a max value then we need to bail since we can't be
3597 * sure we won't do bad things.
3598 * If reg->umax_value + off could overflow, treat that as unbounded too.
3600 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
3601 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
3605 err = __check_mem_access(env, regno, reg->umax_value + off, size,
3606 mem_size, zero_size_allowed);
3608 verbose(env, "R%d max value is outside of the allowed memory range\n",
3616 static int __check_ptr_off_reg(struct bpf_verifier_env *env,
3617 const struct bpf_reg_state *reg, int regno,
3620 /* Access to this pointer-typed register or passing it to a helper
3621 * is only allowed in its original, unmodified form.
3625 verbose(env, "negative offset %s ptr R%d off=%d disallowed\n",
3626 reg_type_str(env, reg->type), regno, reg->off);
3630 if (!fixed_off_ok && reg->off) {
3631 verbose(env, "dereference of modified %s ptr R%d off=%d disallowed\n",
3632 reg_type_str(env, reg->type), regno, reg->off);
3636 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3639 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3640 verbose(env, "variable %s access var_off=%s disallowed\n",
3641 reg_type_str(env, reg->type), tn_buf);
3648 int check_ptr_off_reg(struct bpf_verifier_env *env,
3649 const struct bpf_reg_state *reg, int regno)
3651 return __check_ptr_off_reg(env, reg, regno, false);
3654 static int map_kptr_match_type(struct bpf_verifier_env *env,
3655 struct bpf_map_value_off_desc *off_desc,
3656 struct bpf_reg_state *reg, u32 regno)
3658 const char *targ_name = kernel_type_name(off_desc->kptr.btf, off_desc->kptr.btf_id);
3659 int perm_flags = PTR_MAYBE_NULL;
3660 const char *reg_name = "";
3662 /* Only unreferenced case accepts untrusted pointers */
3663 if (off_desc->type == BPF_KPTR_UNREF)
3664 perm_flags |= PTR_UNTRUSTED;
3666 if (base_type(reg->type) != PTR_TO_BTF_ID || (type_flag(reg->type) & ~perm_flags))
3669 if (!btf_is_kernel(reg->btf)) {
3670 verbose(env, "R%d must point to kernel BTF\n", regno);
3673 /* We need to verify reg->type and reg->btf, before accessing reg->btf */
3674 reg_name = kernel_type_name(reg->btf, reg->btf_id);
3676 /* For ref_ptr case, release function check should ensure we get one
3677 * referenced PTR_TO_BTF_ID, and that its fixed offset is 0. For the
3678 * normal store of unreferenced kptr, we must ensure var_off is zero.
3679 * Since ref_ptr cannot be accessed directly by BPF insns, checks for
3680 * reg->off and reg->ref_obj_id are not needed here.
3682 if (__check_ptr_off_reg(env, reg, regno, true))
3685 /* A full type match is needed, as BTF can be vmlinux or module BTF, and
3686 * we also need to take into account the reg->off.
3688 * We want to support cases like:
3696 * v = func(); // PTR_TO_BTF_ID
3697 * val->foo = v; // reg->off is zero, btf and btf_id match type
3698 * val->bar = &v->br; // reg->off is still zero, but we need to retry with
3699 * // first member type of struct after comparison fails
3700 * val->baz = &v->bz; // reg->off is non-zero, so struct needs to be walked
3703 * In the kptr_ref case, check_func_arg_reg_off already ensures reg->off
3704 * is zero. We must also ensure that btf_struct_ids_match does not walk
3705 * the struct to match type against first member of struct, i.e. reject
3706 * second case from above. Hence, when type is BPF_KPTR_REF, we set
3707 * strict mode to true for type match.
3709 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
3710 off_desc->kptr.btf, off_desc->kptr.btf_id,
3711 off_desc->type == BPF_KPTR_REF))
3715 verbose(env, "invalid kptr access, R%d type=%s%s ", regno,
3716 reg_type_str(env, reg->type), reg_name);
3717 verbose(env, "expected=%s%s", reg_type_str(env, PTR_TO_BTF_ID), targ_name);
3718 if (off_desc->type == BPF_KPTR_UNREF)
3719 verbose(env, " or %s%s\n", reg_type_str(env, PTR_TO_BTF_ID | PTR_UNTRUSTED),
3726 static int check_map_kptr_access(struct bpf_verifier_env *env, u32 regno,
3727 int value_regno, int insn_idx,
3728 struct bpf_map_value_off_desc *off_desc)
3730 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
3731 int class = BPF_CLASS(insn->code);
3732 struct bpf_reg_state *val_reg;
3734 /* Things we already checked for in check_map_access and caller:
3735 * - Reject cases where variable offset may touch kptr
3736 * - size of access (must be BPF_DW)
3737 * - tnum_is_const(reg->var_off)
3738 * - off_desc->offset == off + reg->var_off.value
3740 /* Only BPF_[LDX,STX,ST] | BPF_MEM | BPF_DW is supported */
3741 if (BPF_MODE(insn->code) != BPF_MEM) {
3742 verbose(env, "kptr in map can only be accessed using BPF_MEM instruction mode\n");
3746 /* We only allow loading referenced kptr, since it will be marked as
3747 * untrusted, similar to unreferenced kptr.
3749 if (class != BPF_LDX && off_desc->type == BPF_KPTR_REF) {
3750 verbose(env, "store to referenced kptr disallowed\n");
3754 if (class == BPF_LDX) {
3755 val_reg = reg_state(env, value_regno);
3756 /* We can simply mark the value_regno receiving the pointer
3757 * value from map as PTR_TO_BTF_ID, with the correct type.
3759 mark_btf_ld_reg(env, cur_regs(env), value_regno, PTR_TO_BTF_ID, off_desc->kptr.btf,
3760 off_desc->kptr.btf_id, PTR_MAYBE_NULL | PTR_UNTRUSTED);
3761 /* For mark_ptr_or_null_reg */
3762 val_reg->id = ++env->id_gen;
3763 } else if (class == BPF_STX) {
3764 val_reg = reg_state(env, value_regno);
3765 if (!register_is_null(val_reg) &&
3766 map_kptr_match_type(env, off_desc, val_reg, value_regno))
3768 } else if (class == BPF_ST) {
3770 verbose(env, "BPF_ST imm must be 0 when storing to kptr at off=%u\n",
3775 verbose(env, "kptr in map can only be accessed using BPF_LDX/BPF_STX/BPF_ST\n");
3781 /* check read/write into a map element with possible variable offset */
3782 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
3783 int off, int size, bool zero_size_allowed,
3784 enum bpf_access_src src)
3786 struct bpf_verifier_state *vstate = env->cur_state;
3787 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3788 struct bpf_reg_state *reg = &state->regs[regno];
3789 struct bpf_map *map = reg->map_ptr;
3792 err = check_mem_region_access(env, regno, off, size, map->value_size,
3797 if (map_value_has_spin_lock(map)) {
3798 u32 lock = map->spin_lock_off;
3800 /* if any part of struct bpf_spin_lock can be touched by
3801 * load/store reject this program.
3802 * To check that [x1, x2) overlaps with [y1, y2)
3803 * it is sufficient to check x1 < y2 && y1 < x2.
3805 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
3806 lock < reg->umax_value + off + size) {
3807 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
3811 if (map_value_has_timer(map)) {
3812 u32 t = map->timer_off;
3814 if (reg->smin_value + off < t + sizeof(struct bpf_timer) &&
3815 t < reg->umax_value + off + size) {
3816 verbose(env, "bpf_timer cannot be accessed directly by load/store\n");
3820 if (map_value_has_kptrs(map)) {
3821 struct bpf_map_value_off *tab = map->kptr_off_tab;
3824 for (i = 0; i < tab->nr_off; i++) {
3825 u32 p = tab->off[i].offset;
3827 if (reg->smin_value + off < p + sizeof(u64) &&
3828 p < reg->umax_value + off + size) {
3829 if (src != ACCESS_DIRECT) {
3830 verbose(env, "kptr cannot be accessed indirectly by helper\n");
3833 if (!tnum_is_const(reg->var_off)) {
3834 verbose(env, "kptr access cannot have variable offset\n");
3837 if (p != off + reg->var_off.value) {
3838 verbose(env, "kptr access misaligned expected=%u off=%llu\n",
3839 p, off + reg->var_off.value);
3842 if (size != bpf_size_to_bytes(BPF_DW)) {
3843 verbose(env, "kptr access size must be BPF_DW\n");
3853 #define MAX_PACKET_OFF 0xffff
3855 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
3856 const struct bpf_call_arg_meta *meta,
3857 enum bpf_access_type t)
3859 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
3861 switch (prog_type) {
3862 /* Program types only with direct read access go here! */
3863 case BPF_PROG_TYPE_LWT_IN:
3864 case BPF_PROG_TYPE_LWT_OUT:
3865 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
3866 case BPF_PROG_TYPE_SK_REUSEPORT:
3867 case BPF_PROG_TYPE_FLOW_DISSECTOR:
3868 case BPF_PROG_TYPE_CGROUP_SKB:
3873 /* Program types with direct read + write access go here! */
3874 case BPF_PROG_TYPE_SCHED_CLS:
3875 case BPF_PROG_TYPE_SCHED_ACT:
3876 case BPF_PROG_TYPE_XDP:
3877 case BPF_PROG_TYPE_LWT_XMIT:
3878 case BPF_PROG_TYPE_SK_SKB:
3879 case BPF_PROG_TYPE_SK_MSG:
3881 return meta->pkt_access;
3883 env->seen_direct_write = true;
3886 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
3888 env->seen_direct_write = true;
3897 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
3898 int size, bool zero_size_allowed)
3900 struct bpf_reg_state *regs = cur_regs(env);
3901 struct bpf_reg_state *reg = ®s[regno];
3904 /* We may have added a variable offset to the packet pointer; but any
3905 * reg->range we have comes after that. We are only checking the fixed
3909 /* We don't allow negative numbers, because we aren't tracking enough
3910 * detail to prove they're safe.
3912 if (reg->smin_value < 0) {
3913 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3918 err = reg->range < 0 ? -EINVAL :
3919 __check_mem_access(env, regno, off, size, reg->range,
3922 verbose(env, "R%d offset is outside of the packet\n", regno);
3926 /* __check_mem_access has made sure "off + size - 1" is within u16.
3927 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
3928 * otherwise find_good_pkt_pointers would have refused to set range info
3929 * that __check_mem_access would have rejected this pkt access.
3930 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
3932 env->prog->aux->max_pkt_offset =
3933 max_t(u32, env->prog->aux->max_pkt_offset,
3934 off + reg->umax_value + size - 1);
3939 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
3940 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
3941 enum bpf_access_type t, enum bpf_reg_type *reg_type,
3942 struct btf **btf, u32 *btf_id)
3944 struct bpf_insn_access_aux info = {
3945 .reg_type = *reg_type,
3949 if (env->ops->is_valid_access &&
3950 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
3951 /* A non zero info.ctx_field_size indicates that this field is a
3952 * candidate for later verifier transformation to load the whole
3953 * field and then apply a mask when accessed with a narrower
3954 * access than actual ctx access size. A zero info.ctx_field_size
3955 * will only allow for whole field access and rejects any other
3956 * type of narrower access.
3958 *reg_type = info.reg_type;
3960 if (base_type(*reg_type) == PTR_TO_BTF_ID) {
3962 *btf_id = info.btf_id;
3964 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
3966 /* remember the offset of last byte accessed in ctx */
3967 if (env->prog->aux->max_ctx_offset < off + size)
3968 env->prog->aux->max_ctx_offset = off + size;
3972 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
3976 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
3979 if (size < 0 || off < 0 ||
3980 (u64)off + size > sizeof(struct bpf_flow_keys)) {
3981 verbose(env, "invalid access to flow keys off=%d size=%d\n",
3988 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
3989 u32 regno, int off, int size,
3990 enum bpf_access_type t)
3992 struct bpf_reg_state *regs = cur_regs(env);
3993 struct bpf_reg_state *reg = ®s[regno];
3994 struct bpf_insn_access_aux info = {};
3997 if (reg->smin_value < 0) {
3998 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
4003 switch (reg->type) {
4004 case PTR_TO_SOCK_COMMON:
4005 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
4008 valid = bpf_sock_is_valid_access(off, size, t, &info);
4010 case PTR_TO_TCP_SOCK:
4011 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
4013 case PTR_TO_XDP_SOCK:
4014 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
4022 env->insn_aux_data[insn_idx].ctx_field_size =
4023 info.ctx_field_size;
4027 verbose(env, "R%d invalid %s access off=%d size=%d\n",
4028 regno, reg_type_str(env, reg->type), off, size);
4033 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
4035 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
4038 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
4040 const struct bpf_reg_state *reg = reg_state(env, regno);
4042 return reg->type == PTR_TO_CTX;
4045 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
4047 const struct bpf_reg_state *reg = reg_state(env, regno);
4049 return type_is_sk_pointer(reg->type);
4052 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
4054 const struct bpf_reg_state *reg = reg_state(env, regno);
4056 return type_is_pkt_pointer(reg->type);
4059 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
4061 const struct bpf_reg_state *reg = reg_state(env, regno);
4063 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
4064 return reg->type == PTR_TO_FLOW_KEYS;
4067 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
4068 const struct bpf_reg_state *reg,
4069 int off, int size, bool strict)
4071 struct tnum reg_off;
4074 /* Byte size accesses are always allowed. */
4075 if (!strict || size == 1)
4078 /* For platforms that do not have a Kconfig enabling
4079 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
4080 * NET_IP_ALIGN is universally set to '2'. And on platforms
4081 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
4082 * to this code only in strict mode where we want to emulate
4083 * the NET_IP_ALIGN==2 checking. Therefore use an
4084 * unconditional IP align value of '2'.
4088 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
4089 if (!tnum_is_aligned(reg_off, size)) {
4092 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4094 "misaligned packet access off %d+%s+%d+%d size %d\n",
4095 ip_align, tn_buf, reg->off, off, size);
4102 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
4103 const struct bpf_reg_state *reg,
4104 const char *pointer_desc,
4105 int off, int size, bool strict)
4107 struct tnum reg_off;
4109 /* Byte size accesses are always allowed. */
4110 if (!strict || size == 1)
4113 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
4114 if (!tnum_is_aligned(reg_off, size)) {
4117 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4118 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
4119 pointer_desc, tn_buf, reg->off, off, size);
4126 static int check_ptr_alignment(struct bpf_verifier_env *env,
4127 const struct bpf_reg_state *reg, int off,
4128 int size, bool strict_alignment_once)
4130 bool strict = env->strict_alignment || strict_alignment_once;
4131 const char *pointer_desc = "";
4133 switch (reg->type) {
4135 case PTR_TO_PACKET_META:
4136 /* Special case, because of NET_IP_ALIGN. Given metadata sits
4137 * right in front, treat it the very same way.
4139 return check_pkt_ptr_alignment(env, reg, off, size, strict);
4140 case PTR_TO_FLOW_KEYS:
4141 pointer_desc = "flow keys ";
4143 case PTR_TO_MAP_KEY:
4144 pointer_desc = "key ";
4146 case PTR_TO_MAP_VALUE:
4147 pointer_desc = "value ";
4150 pointer_desc = "context ";
4153 pointer_desc = "stack ";
4154 /* The stack spill tracking logic in check_stack_write_fixed_off()
4155 * and check_stack_read_fixed_off() relies on stack accesses being
4161 pointer_desc = "sock ";
4163 case PTR_TO_SOCK_COMMON:
4164 pointer_desc = "sock_common ";
4166 case PTR_TO_TCP_SOCK:
4167 pointer_desc = "tcp_sock ";
4169 case PTR_TO_XDP_SOCK:
4170 pointer_desc = "xdp_sock ";
4175 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
4179 static int update_stack_depth(struct bpf_verifier_env *env,
4180 const struct bpf_func_state *func,
4183 u16 stack = env->subprog_info[func->subprogno].stack_depth;
4188 /* update known max for given subprogram */
4189 env->subprog_info[func->subprogno].stack_depth = -off;
4193 /* starting from main bpf function walk all instructions of the function
4194 * and recursively walk all callees that given function can call.
4195 * Ignore jump and exit insns.
4196 * Since recursion is prevented by check_cfg() this algorithm
4197 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
4199 static int check_max_stack_depth(struct bpf_verifier_env *env)
4201 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
4202 struct bpf_subprog_info *subprog = env->subprog_info;
4203 struct bpf_insn *insn = env->prog->insnsi;
4204 bool tail_call_reachable = false;
4205 int ret_insn[MAX_CALL_FRAMES];
4206 int ret_prog[MAX_CALL_FRAMES];
4210 /* protect against potential stack overflow that might happen when
4211 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
4212 * depth for such case down to 256 so that the worst case scenario
4213 * would result in 8k stack size (32 which is tailcall limit * 256 =
4216 * To get the idea what might happen, see an example:
4217 * func1 -> sub rsp, 128
4218 * subfunc1 -> sub rsp, 256
4219 * tailcall1 -> add rsp, 256
4220 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
4221 * subfunc2 -> sub rsp, 64
4222 * subfunc22 -> sub rsp, 128
4223 * tailcall2 -> add rsp, 128
4224 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
4226 * tailcall will unwind the current stack frame but it will not get rid
4227 * of caller's stack as shown on the example above.
4229 if (idx && subprog[idx].has_tail_call && depth >= 256) {
4231 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
4235 /* round up to 32-bytes, since this is granularity
4236 * of interpreter stack size
4238 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
4239 if (depth > MAX_BPF_STACK) {
4240 verbose(env, "combined stack size of %d calls is %d. Too large\n",
4245 subprog_end = subprog[idx + 1].start;
4246 for (; i < subprog_end; i++) {
4249 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
4251 /* remember insn and function to return to */
4252 ret_insn[frame] = i + 1;
4253 ret_prog[frame] = idx;
4255 /* find the callee */
4256 next_insn = i + insn[i].imm + 1;
4257 idx = find_subprog(env, next_insn);
4259 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
4263 if (subprog[idx].is_async_cb) {
4264 if (subprog[idx].has_tail_call) {
4265 verbose(env, "verifier bug. subprog has tail_call and async cb\n");
4268 /* async callbacks don't increase bpf prog stack size */
4273 if (subprog[idx].has_tail_call)
4274 tail_call_reachable = true;
4277 if (frame >= MAX_CALL_FRAMES) {
4278 verbose(env, "the call stack of %d frames is too deep !\n",
4284 /* if tail call got detected across bpf2bpf calls then mark each of the
4285 * currently present subprog frames as tail call reachable subprogs;
4286 * this info will be utilized by JIT so that we will be preserving the
4287 * tail call counter throughout bpf2bpf calls combined with tailcalls
4289 if (tail_call_reachable)
4290 for (j = 0; j < frame; j++)
4291 subprog[ret_prog[j]].tail_call_reachable = true;
4292 if (subprog[0].tail_call_reachable)
4293 env->prog->aux->tail_call_reachable = true;
4295 /* end of for() loop means the last insn of the 'subprog'
4296 * was reached. Doesn't matter whether it was JA or EXIT
4300 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
4302 i = ret_insn[frame];
4303 idx = ret_prog[frame];
4307 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
4308 static int get_callee_stack_depth(struct bpf_verifier_env *env,
4309 const struct bpf_insn *insn, int idx)
4311 int start = idx + insn->imm + 1, subprog;
4313 subprog = find_subprog(env, start);
4315 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
4319 return env->subprog_info[subprog].stack_depth;
4323 static int __check_buffer_access(struct bpf_verifier_env *env,
4324 const char *buf_info,
4325 const struct bpf_reg_state *reg,
4326 int regno, int off, int size)
4330 "R%d invalid %s buffer access: off=%d, size=%d\n",
4331 regno, buf_info, off, size);
4334 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4337 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4339 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
4340 regno, off, tn_buf);
4347 static int check_tp_buffer_access(struct bpf_verifier_env *env,
4348 const struct bpf_reg_state *reg,
4349 int regno, int off, int size)
4353 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
4357 if (off + size > env->prog->aux->max_tp_access)
4358 env->prog->aux->max_tp_access = off + size;
4363 static int check_buffer_access(struct bpf_verifier_env *env,
4364 const struct bpf_reg_state *reg,
4365 int regno, int off, int size,
4366 bool zero_size_allowed,
4369 const char *buf_info = type_is_rdonly_mem(reg->type) ? "rdonly" : "rdwr";
4372 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
4376 if (off + size > *max_access)
4377 *max_access = off + size;
4382 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
4383 static void zext_32_to_64(struct bpf_reg_state *reg)
4385 reg->var_off = tnum_subreg(reg->var_off);
4386 __reg_assign_32_into_64(reg);
4389 /* truncate register to smaller size (in bytes)
4390 * must be called with size < BPF_REG_SIZE
4392 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
4396 /* clear high bits in bit representation */
4397 reg->var_off = tnum_cast(reg->var_off, size);
4399 /* fix arithmetic bounds */
4400 mask = ((u64)1 << (size * 8)) - 1;
4401 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
4402 reg->umin_value &= mask;
4403 reg->umax_value &= mask;
4405 reg->umin_value = 0;
4406 reg->umax_value = mask;
4408 reg->smin_value = reg->umin_value;
4409 reg->smax_value = reg->umax_value;
4411 /* If size is smaller than 32bit register the 32bit register
4412 * values are also truncated so we push 64-bit bounds into
4413 * 32-bit bounds. Above were truncated < 32-bits already.
4417 __reg_combine_64_into_32(reg);
4420 static bool bpf_map_is_rdonly(const struct bpf_map *map)
4422 /* A map is considered read-only if the following condition are true:
4424 * 1) BPF program side cannot change any of the map content. The
4425 * BPF_F_RDONLY_PROG flag is throughout the lifetime of a map
4426 * and was set at map creation time.
4427 * 2) The map value(s) have been initialized from user space by a
4428 * loader and then "frozen", such that no new map update/delete
4429 * operations from syscall side are possible for the rest of
4430 * the map's lifetime from that point onwards.
4431 * 3) Any parallel/pending map update/delete operations from syscall
4432 * side have been completed. Only after that point, it's safe to
4433 * assume that map value(s) are immutable.
4435 return (map->map_flags & BPF_F_RDONLY_PROG) &&
4436 READ_ONCE(map->frozen) &&
4437 !bpf_map_write_active(map);
4440 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
4446 err = map->ops->map_direct_value_addr(map, &addr, off);
4449 ptr = (void *)(long)addr + off;
4453 *val = (u64)*(u8 *)ptr;
4456 *val = (u64)*(u16 *)ptr;
4459 *val = (u64)*(u32 *)ptr;
4470 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
4471 struct bpf_reg_state *regs,
4472 int regno, int off, int size,
4473 enum bpf_access_type atype,
4476 struct bpf_reg_state *reg = regs + regno;
4477 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
4478 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
4479 enum bpf_type_flag flag = 0;
4485 "R%d is ptr_%s invalid negative access: off=%d\n",
4489 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4492 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4494 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
4495 regno, tname, off, tn_buf);
4499 if (reg->type & MEM_USER) {
4501 "R%d is ptr_%s access user memory: off=%d\n",
4506 if (reg->type & MEM_PERCPU) {
4508 "R%d is ptr_%s access percpu memory: off=%d\n",
4513 if (env->ops->btf_struct_access) {
4514 ret = env->ops->btf_struct_access(&env->log, reg->btf, t,
4515 off, size, atype, &btf_id, &flag);
4517 if (atype != BPF_READ) {
4518 verbose(env, "only read is supported\n");
4522 ret = btf_struct_access(&env->log, reg->btf, t, off, size,
4523 atype, &btf_id, &flag);
4529 /* If this is an untrusted pointer, all pointers formed by walking it
4530 * also inherit the untrusted flag.
4532 if (type_flag(reg->type) & PTR_UNTRUSTED)
4533 flag |= PTR_UNTRUSTED;
4535 if (atype == BPF_READ && value_regno >= 0)
4536 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id, flag);
4541 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
4542 struct bpf_reg_state *regs,
4543 int regno, int off, int size,
4544 enum bpf_access_type atype,
4547 struct bpf_reg_state *reg = regs + regno;
4548 struct bpf_map *map = reg->map_ptr;
4549 enum bpf_type_flag flag = 0;
4550 const struct btf_type *t;
4556 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
4560 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
4561 verbose(env, "map_ptr access not supported for map type %d\n",
4566 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
4567 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
4569 if (!env->allow_ptr_to_map_access) {
4571 "%s access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
4577 verbose(env, "R%d is %s invalid negative access: off=%d\n",
4582 if (atype != BPF_READ) {
4583 verbose(env, "only read from %s is supported\n", tname);
4587 ret = btf_struct_access(&env->log, btf_vmlinux, t, off, size, atype, &btf_id, &flag);
4591 if (value_regno >= 0)
4592 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id, flag);
4597 /* Check that the stack access at the given offset is within bounds. The
4598 * maximum valid offset is -1.
4600 * The minimum valid offset is -MAX_BPF_STACK for writes, and
4601 * -state->allocated_stack for reads.
4603 static int check_stack_slot_within_bounds(int off,
4604 struct bpf_func_state *state,
4605 enum bpf_access_type t)
4610 min_valid_off = -MAX_BPF_STACK;
4612 min_valid_off = -state->allocated_stack;
4614 if (off < min_valid_off || off > -1)
4619 /* Check that the stack access at 'regno + off' falls within the maximum stack
4622 * 'off' includes `regno->offset`, but not its dynamic part (if any).
4624 static int check_stack_access_within_bounds(
4625 struct bpf_verifier_env *env,
4626 int regno, int off, int access_size,
4627 enum bpf_access_src src, enum bpf_access_type type)
4629 struct bpf_reg_state *regs = cur_regs(env);
4630 struct bpf_reg_state *reg = regs + regno;
4631 struct bpf_func_state *state = func(env, reg);
4632 int min_off, max_off;
4636 if (src == ACCESS_HELPER)
4637 /* We don't know if helpers are reading or writing (or both). */
4638 err_extra = " indirect access to";
4639 else if (type == BPF_READ)
4640 err_extra = " read from";
4642 err_extra = " write to";
4644 if (tnum_is_const(reg->var_off)) {
4645 min_off = reg->var_off.value + off;
4646 if (access_size > 0)
4647 max_off = min_off + access_size - 1;
4651 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
4652 reg->smin_value <= -BPF_MAX_VAR_OFF) {
4653 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
4657 min_off = reg->smin_value + off;
4658 if (access_size > 0)
4659 max_off = reg->smax_value + off + access_size - 1;
4664 err = check_stack_slot_within_bounds(min_off, state, type);
4666 err = check_stack_slot_within_bounds(max_off, state, type);
4669 if (tnum_is_const(reg->var_off)) {
4670 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
4671 err_extra, regno, off, access_size);
4675 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4676 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
4677 err_extra, regno, tn_buf, access_size);
4683 /* check whether memory at (regno + off) is accessible for t = (read | write)
4684 * if t==write, value_regno is a register which value is stored into memory
4685 * if t==read, value_regno is a register which will receive the value from memory
4686 * if t==write && value_regno==-1, some unknown value is stored into memory
4687 * if t==read && value_regno==-1, don't care what we read from memory
4689 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
4690 int off, int bpf_size, enum bpf_access_type t,
4691 int value_regno, bool strict_alignment_once)
4693 struct bpf_reg_state *regs = cur_regs(env);
4694 struct bpf_reg_state *reg = regs + regno;
4695 struct bpf_func_state *state;
4698 size = bpf_size_to_bytes(bpf_size);
4702 /* alignment checks will add in reg->off themselves */
4703 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
4707 /* for access checks, reg->off is just part of off */
4710 if (reg->type == PTR_TO_MAP_KEY) {
4711 if (t == BPF_WRITE) {
4712 verbose(env, "write to change key R%d not allowed\n", regno);
4716 err = check_mem_region_access(env, regno, off, size,
4717 reg->map_ptr->key_size, false);
4720 if (value_regno >= 0)
4721 mark_reg_unknown(env, regs, value_regno);
4722 } else if (reg->type == PTR_TO_MAP_VALUE) {
4723 struct bpf_map_value_off_desc *kptr_off_desc = NULL;
4725 if (t == BPF_WRITE && value_regno >= 0 &&
4726 is_pointer_value(env, value_regno)) {
4727 verbose(env, "R%d leaks addr into map\n", value_regno);
4730 err = check_map_access_type(env, regno, off, size, t);
4733 err = check_map_access(env, regno, off, size, false, ACCESS_DIRECT);
4736 if (tnum_is_const(reg->var_off))
4737 kptr_off_desc = bpf_map_kptr_off_contains(reg->map_ptr,
4738 off + reg->var_off.value);
4739 if (kptr_off_desc) {
4740 err = check_map_kptr_access(env, regno, value_regno, insn_idx,
4742 } else if (t == BPF_READ && value_regno >= 0) {
4743 struct bpf_map *map = reg->map_ptr;
4745 /* if map is read-only, track its contents as scalars */
4746 if (tnum_is_const(reg->var_off) &&
4747 bpf_map_is_rdonly(map) &&
4748 map->ops->map_direct_value_addr) {
4749 int map_off = off + reg->var_off.value;
4752 err = bpf_map_direct_read(map, map_off, size,
4757 regs[value_regno].type = SCALAR_VALUE;
4758 __mark_reg_known(®s[value_regno], val);
4760 mark_reg_unknown(env, regs, value_regno);
4763 } else if (base_type(reg->type) == PTR_TO_MEM) {
4764 bool rdonly_mem = type_is_rdonly_mem(reg->type);
4766 if (type_may_be_null(reg->type)) {
4767 verbose(env, "R%d invalid mem access '%s'\n", regno,
4768 reg_type_str(env, reg->type));
4772 if (t == BPF_WRITE && rdonly_mem) {
4773 verbose(env, "R%d cannot write into %s\n",
4774 regno, reg_type_str(env, reg->type));
4778 if (t == BPF_WRITE && value_regno >= 0 &&
4779 is_pointer_value(env, value_regno)) {
4780 verbose(env, "R%d leaks addr into mem\n", value_regno);
4784 err = check_mem_region_access(env, regno, off, size,
4785 reg->mem_size, false);
4786 if (!err && value_regno >= 0 && (t == BPF_READ || rdonly_mem))
4787 mark_reg_unknown(env, regs, value_regno);
4788 } else if (reg->type == PTR_TO_CTX) {
4789 enum bpf_reg_type reg_type = SCALAR_VALUE;
4790 struct btf *btf = NULL;
4793 if (t == BPF_WRITE && value_regno >= 0 &&
4794 is_pointer_value(env, value_regno)) {
4795 verbose(env, "R%d leaks addr into ctx\n", value_regno);
4799 err = check_ptr_off_reg(env, reg, regno);
4803 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf,
4806 verbose_linfo(env, insn_idx, "; ");
4807 if (!err && t == BPF_READ && value_regno >= 0) {
4808 /* ctx access returns either a scalar, or a
4809 * PTR_TO_PACKET[_META,_END]. In the latter
4810 * case, we know the offset is zero.
4812 if (reg_type == SCALAR_VALUE) {
4813 mark_reg_unknown(env, regs, value_regno);
4815 mark_reg_known_zero(env, regs,
4817 if (type_may_be_null(reg_type))
4818 regs[value_regno].id = ++env->id_gen;
4819 /* A load of ctx field could have different
4820 * actual load size with the one encoded in the
4821 * insn. When the dst is PTR, it is for sure not
4824 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
4825 if (base_type(reg_type) == PTR_TO_BTF_ID) {
4826 regs[value_regno].btf = btf;
4827 regs[value_regno].btf_id = btf_id;
4830 regs[value_regno].type = reg_type;
4833 } else if (reg->type == PTR_TO_STACK) {
4834 /* Basic bounds checks. */
4835 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
4839 state = func(env, reg);
4840 err = update_stack_depth(env, state, off);
4845 err = check_stack_read(env, regno, off, size,
4848 err = check_stack_write(env, regno, off, size,
4849 value_regno, insn_idx);
4850 } else if (reg_is_pkt_pointer(reg)) {
4851 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
4852 verbose(env, "cannot write into packet\n");
4855 if (t == BPF_WRITE && value_regno >= 0 &&
4856 is_pointer_value(env, value_regno)) {
4857 verbose(env, "R%d leaks addr into packet\n",
4861 err = check_packet_access(env, regno, off, size, false);
4862 if (!err && t == BPF_READ && value_regno >= 0)
4863 mark_reg_unknown(env, regs, value_regno);
4864 } else if (reg->type == PTR_TO_FLOW_KEYS) {
4865 if (t == BPF_WRITE && value_regno >= 0 &&
4866 is_pointer_value(env, value_regno)) {
4867 verbose(env, "R%d leaks addr into flow keys\n",
4872 err = check_flow_keys_access(env, off, size);
4873 if (!err && t == BPF_READ && value_regno >= 0)
4874 mark_reg_unknown(env, regs, value_regno);
4875 } else if (type_is_sk_pointer(reg->type)) {
4876 if (t == BPF_WRITE) {
4877 verbose(env, "R%d cannot write into %s\n",
4878 regno, reg_type_str(env, reg->type));
4881 err = check_sock_access(env, insn_idx, regno, off, size, t);
4882 if (!err && value_regno >= 0)
4883 mark_reg_unknown(env, regs, value_regno);
4884 } else if (reg->type == PTR_TO_TP_BUFFER) {
4885 err = check_tp_buffer_access(env, reg, regno, off, size);
4886 if (!err && t == BPF_READ && value_regno >= 0)
4887 mark_reg_unknown(env, regs, value_regno);
4888 } else if (base_type(reg->type) == PTR_TO_BTF_ID &&
4889 !type_may_be_null(reg->type)) {
4890 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
4892 } else if (reg->type == CONST_PTR_TO_MAP) {
4893 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
4895 } else if (base_type(reg->type) == PTR_TO_BUF) {
4896 bool rdonly_mem = type_is_rdonly_mem(reg->type);
4900 if (t == BPF_WRITE) {
4901 verbose(env, "R%d cannot write into %s\n",
4902 regno, reg_type_str(env, reg->type));
4905 max_access = &env->prog->aux->max_rdonly_access;
4907 max_access = &env->prog->aux->max_rdwr_access;
4910 err = check_buffer_access(env, reg, regno, off, size, false,
4913 if (!err && value_regno >= 0 && (rdonly_mem || t == BPF_READ))
4914 mark_reg_unknown(env, regs, value_regno);
4916 verbose(env, "R%d invalid mem access '%s'\n", regno,
4917 reg_type_str(env, reg->type));
4921 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
4922 regs[value_regno].type == SCALAR_VALUE) {
4923 /* b/h/w load zero-extends, mark upper bits as known 0 */
4924 coerce_reg_to_size(®s[value_regno], size);
4929 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
4934 switch (insn->imm) {
4936 case BPF_ADD | BPF_FETCH:
4938 case BPF_AND | BPF_FETCH:
4940 case BPF_OR | BPF_FETCH:
4942 case BPF_XOR | BPF_FETCH:
4947 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
4951 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
4952 verbose(env, "invalid atomic operand size\n");
4956 /* check src1 operand */
4957 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4961 /* check src2 operand */
4962 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4966 if (insn->imm == BPF_CMPXCHG) {
4967 /* Check comparison of R0 with memory location */
4968 const u32 aux_reg = BPF_REG_0;
4970 err = check_reg_arg(env, aux_reg, SRC_OP);
4974 if (is_pointer_value(env, aux_reg)) {
4975 verbose(env, "R%d leaks addr into mem\n", aux_reg);
4980 if (is_pointer_value(env, insn->src_reg)) {
4981 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
4985 if (is_ctx_reg(env, insn->dst_reg) ||
4986 is_pkt_reg(env, insn->dst_reg) ||
4987 is_flow_key_reg(env, insn->dst_reg) ||
4988 is_sk_reg(env, insn->dst_reg)) {
4989 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
4991 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
4995 if (insn->imm & BPF_FETCH) {
4996 if (insn->imm == BPF_CMPXCHG)
4997 load_reg = BPF_REG_0;
4999 load_reg = insn->src_reg;
5001 /* check and record load of old value */
5002 err = check_reg_arg(env, load_reg, DST_OP);
5006 /* This instruction accesses a memory location but doesn't
5007 * actually load it into a register.
5012 /* Check whether we can read the memory, with second call for fetch
5013 * case to simulate the register fill.
5015 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
5016 BPF_SIZE(insn->code), BPF_READ, -1, true);
5017 if (!err && load_reg >= 0)
5018 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
5019 BPF_SIZE(insn->code), BPF_READ, load_reg,
5024 /* Check whether we can write into the same memory. */
5025 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
5026 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
5033 /* When register 'regno' is used to read the stack (either directly or through
5034 * a helper function) make sure that it's within stack boundary and, depending
5035 * on the access type, that all elements of the stack are initialized.
5037 * 'off' includes 'regno->off', but not its dynamic part (if any).
5039 * All registers that have been spilled on the stack in the slots within the
5040 * read offsets are marked as read.
5042 static int check_stack_range_initialized(
5043 struct bpf_verifier_env *env, int regno, int off,
5044 int access_size, bool zero_size_allowed,
5045 enum bpf_access_src type, struct bpf_call_arg_meta *meta)
5047 struct bpf_reg_state *reg = reg_state(env, regno);
5048 struct bpf_func_state *state = func(env, reg);
5049 int err, min_off, max_off, i, j, slot, spi;
5050 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
5051 enum bpf_access_type bounds_check_type;
5052 /* Some accesses can write anything into the stack, others are
5055 bool clobber = false;
5057 if (access_size == 0 && !zero_size_allowed) {
5058 verbose(env, "invalid zero-sized read\n");
5062 if (type == ACCESS_HELPER) {
5063 /* The bounds checks for writes are more permissive than for
5064 * reads. However, if raw_mode is not set, we'll do extra
5067 bounds_check_type = BPF_WRITE;
5070 bounds_check_type = BPF_READ;
5072 err = check_stack_access_within_bounds(env, regno, off, access_size,
5073 type, bounds_check_type);
5078 if (tnum_is_const(reg->var_off)) {
5079 min_off = max_off = reg->var_off.value + off;
5081 /* Variable offset is prohibited for unprivileged mode for
5082 * simplicity since it requires corresponding support in
5083 * Spectre masking for stack ALU.
5084 * See also retrieve_ptr_limit().
5086 if (!env->bypass_spec_v1) {
5089 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5090 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
5091 regno, err_extra, tn_buf);
5094 /* Only initialized buffer on stack is allowed to be accessed
5095 * with variable offset. With uninitialized buffer it's hard to
5096 * guarantee that whole memory is marked as initialized on
5097 * helper return since specific bounds are unknown what may
5098 * cause uninitialized stack leaking.
5100 if (meta && meta->raw_mode)
5103 min_off = reg->smin_value + off;
5104 max_off = reg->smax_value + off;
5107 if (meta && meta->raw_mode) {
5108 meta->access_size = access_size;
5109 meta->regno = regno;
5113 for (i = min_off; i < max_off + access_size; i++) {
5117 spi = slot / BPF_REG_SIZE;
5118 if (state->allocated_stack <= slot)
5120 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
5121 if (*stype == STACK_MISC)
5123 if (*stype == STACK_ZERO) {
5125 /* helper can write anything into the stack */
5126 *stype = STACK_MISC;
5131 if (is_spilled_reg(&state->stack[spi]) &&
5132 base_type(state->stack[spi].spilled_ptr.type) == PTR_TO_BTF_ID)
5135 if (is_spilled_reg(&state->stack[spi]) &&
5136 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
5137 env->allow_ptr_leaks)) {
5139 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
5140 for (j = 0; j < BPF_REG_SIZE; j++)
5141 scrub_spilled_slot(&state->stack[spi].slot_type[j]);
5147 if (tnum_is_const(reg->var_off)) {
5148 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
5149 err_extra, regno, min_off, i - min_off, access_size);
5153 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5154 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
5155 err_extra, regno, tn_buf, i - min_off, access_size);
5159 /* reading any byte out of 8-byte 'spill_slot' will cause
5160 * the whole slot to be marked as 'read'
5162 mark_reg_read(env, &state->stack[spi].spilled_ptr,
5163 state->stack[spi].spilled_ptr.parent,
5166 return update_stack_depth(env, state, min_off);
5169 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
5170 int access_size, bool zero_size_allowed,
5171 struct bpf_call_arg_meta *meta)
5173 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5176 switch (base_type(reg->type)) {
5178 case PTR_TO_PACKET_META:
5179 return check_packet_access(env, regno, reg->off, access_size,
5181 case PTR_TO_MAP_KEY:
5182 if (meta && meta->raw_mode) {
5183 verbose(env, "R%d cannot write into %s\n", regno,
5184 reg_type_str(env, reg->type));
5187 return check_mem_region_access(env, regno, reg->off, access_size,
5188 reg->map_ptr->key_size, false);
5189 case PTR_TO_MAP_VALUE:
5190 if (check_map_access_type(env, regno, reg->off, access_size,
5191 meta && meta->raw_mode ? BPF_WRITE :
5194 return check_map_access(env, regno, reg->off, access_size,
5195 zero_size_allowed, ACCESS_HELPER);
5197 if (type_is_rdonly_mem(reg->type)) {
5198 if (meta && meta->raw_mode) {
5199 verbose(env, "R%d cannot write into %s\n", regno,
5200 reg_type_str(env, reg->type));
5204 return check_mem_region_access(env, regno, reg->off,
5205 access_size, reg->mem_size,
5208 if (type_is_rdonly_mem(reg->type)) {
5209 if (meta && meta->raw_mode) {
5210 verbose(env, "R%d cannot write into %s\n", regno,
5211 reg_type_str(env, reg->type));
5215 max_access = &env->prog->aux->max_rdonly_access;
5217 max_access = &env->prog->aux->max_rdwr_access;
5219 return check_buffer_access(env, reg, regno, reg->off,
5220 access_size, zero_size_allowed,
5223 return check_stack_range_initialized(
5225 regno, reg->off, access_size,
5226 zero_size_allowed, ACCESS_HELPER, meta);
5227 default: /* scalar_value or invalid ptr */
5228 /* Allow zero-byte read from NULL, regardless of pointer type */
5229 if (zero_size_allowed && access_size == 0 &&
5230 register_is_null(reg))
5233 verbose(env, "R%d type=%s ", regno,
5234 reg_type_str(env, reg->type));
5235 verbose(env, "expected=%s\n", reg_type_str(env, PTR_TO_STACK));
5240 static int check_mem_size_reg(struct bpf_verifier_env *env,
5241 struct bpf_reg_state *reg, u32 regno,
5242 bool zero_size_allowed,
5243 struct bpf_call_arg_meta *meta)
5247 /* This is used to refine r0 return value bounds for helpers
5248 * that enforce this value as an upper bound on return values.
5249 * See do_refine_retval_range() for helpers that can refine
5250 * the return value. C type of helper is u32 so we pull register
5251 * bound from umax_value however, if negative verifier errors
5252 * out. Only upper bounds can be learned because retval is an
5253 * int type and negative retvals are allowed.
5255 meta->msize_max_value = reg->umax_value;
5257 /* The register is SCALAR_VALUE; the access check
5258 * happens using its boundaries.
5260 if (!tnum_is_const(reg->var_off))
5261 /* For unprivileged variable accesses, disable raw
5262 * mode so that the program is required to
5263 * initialize all the memory that the helper could
5264 * just partially fill up.
5268 if (reg->smin_value < 0) {
5269 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
5274 if (reg->umin_value == 0) {
5275 err = check_helper_mem_access(env, regno - 1, 0,
5282 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
5283 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
5287 err = check_helper_mem_access(env, regno - 1,
5289 zero_size_allowed, meta);
5291 err = mark_chain_precision(env, regno);
5295 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
5296 u32 regno, u32 mem_size)
5298 bool may_be_null = type_may_be_null(reg->type);
5299 struct bpf_reg_state saved_reg;
5300 struct bpf_call_arg_meta meta;
5303 if (register_is_null(reg))
5306 memset(&meta, 0, sizeof(meta));
5307 /* Assuming that the register contains a value check if the memory
5308 * access is safe. Temporarily save and restore the register's state as
5309 * the conversion shouldn't be visible to a caller.
5313 mark_ptr_not_null_reg(reg);
5316 err = check_helper_mem_access(env, regno, mem_size, true, &meta);
5317 /* Check access for BPF_WRITE */
5318 meta.raw_mode = true;
5319 err = err ?: check_helper_mem_access(env, regno, mem_size, true, &meta);
5327 int check_kfunc_mem_size_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
5330 struct bpf_reg_state *mem_reg = &cur_regs(env)[regno - 1];
5331 bool may_be_null = type_may_be_null(mem_reg->type);
5332 struct bpf_reg_state saved_reg;
5333 struct bpf_call_arg_meta meta;
5336 WARN_ON_ONCE(regno < BPF_REG_2 || regno > BPF_REG_5);
5338 memset(&meta, 0, sizeof(meta));
5341 saved_reg = *mem_reg;
5342 mark_ptr_not_null_reg(mem_reg);
5345 err = check_mem_size_reg(env, reg, regno, true, &meta);
5346 /* Check access for BPF_WRITE */
5347 meta.raw_mode = true;
5348 err = err ?: check_mem_size_reg(env, reg, regno, true, &meta);
5351 *mem_reg = saved_reg;
5355 /* Implementation details:
5356 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
5357 * Two bpf_map_lookups (even with the same key) will have different reg->id.
5358 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
5359 * value_or_null->value transition, since the verifier only cares about
5360 * the range of access to valid map value pointer and doesn't care about actual
5361 * address of the map element.
5362 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
5363 * reg->id > 0 after value_or_null->value transition. By doing so
5364 * two bpf_map_lookups will be considered two different pointers that
5365 * point to different bpf_spin_locks.
5366 * The verifier allows taking only one bpf_spin_lock at a time to avoid
5368 * Since only one bpf_spin_lock is allowed the checks are simpler than
5369 * reg_is_refcounted() logic. The verifier needs to remember only
5370 * one spin_lock instead of array of acquired_refs.
5371 * cur_state->active_spin_lock remembers which map value element got locked
5372 * and clears it after bpf_spin_unlock.
5374 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
5377 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5378 struct bpf_verifier_state *cur = env->cur_state;
5379 bool is_const = tnum_is_const(reg->var_off);
5380 struct bpf_map *map = reg->map_ptr;
5381 u64 val = reg->var_off.value;
5385 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
5391 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
5395 if (!map_value_has_spin_lock(map)) {
5396 if (map->spin_lock_off == -E2BIG)
5398 "map '%s' has more than one 'struct bpf_spin_lock'\n",
5400 else if (map->spin_lock_off == -ENOENT)
5402 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
5406 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
5410 if (map->spin_lock_off != val + reg->off) {
5411 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
5416 if (cur->active_spin_lock) {
5418 "Locking two bpf_spin_locks are not allowed\n");
5421 cur->active_spin_lock = reg->id;
5423 if (!cur->active_spin_lock) {
5424 verbose(env, "bpf_spin_unlock without taking a lock\n");
5427 if (cur->active_spin_lock != reg->id) {
5428 verbose(env, "bpf_spin_unlock of different lock\n");
5431 cur->active_spin_lock = 0;
5436 static int process_timer_func(struct bpf_verifier_env *env, int regno,
5437 struct bpf_call_arg_meta *meta)
5439 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5440 bool is_const = tnum_is_const(reg->var_off);
5441 struct bpf_map *map = reg->map_ptr;
5442 u64 val = reg->var_off.value;
5446 "R%d doesn't have constant offset. bpf_timer has to be at the constant offset\n",
5451 verbose(env, "map '%s' has to have BTF in order to use bpf_timer\n",
5455 if (!map_value_has_timer(map)) {
5456 if (map->timer_off == -E2BIG)
5458 "map '%s' has more than one 'struct bpf_timer'\n",
5460 else if (map->timer_off == -ENOENT)
5462 "map '%s' doesn't have 'struct bpf_timer'\n",
5466 "map '%s' is not a struct type or bpf_timer is mangled\n",
5470 if (map->timer_off != val + reg->off) {
5471 verbose(env, "off %lld doesn't point to 'struct bpf_timer' that is at %d\n",
5472 val + reg->off, map->timer_off);
5475 if (meta->map_ptr) {
5476 verbose(env, "verifier bug. Two map pointers in a timer helper\n");
5479 meta->map_uid = reg->map_uid;
5480 meta->map_ptr = map;
5484 static int process_kptr_func(struct bpf_verifier_env *env, int regno,
5485 struct bpf_call_arg_meta *meta)
5487 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5488 struct bpf_map_value_off_desc *off_desc;
5489 struct bpf_map *map_ptr = reg->map_ptr;
5493 if (!tnum_is_const(reg->var_off)) {
5495 "R%d doesn't have constant offset. kptr has to be at the constant offset\n",
5499 if (!map_ptr->btf) {
5500 verbose(env, "map '%s' has to have BTF in order to use bpf_kptr_xchg\n",
5504 if (!map_value_has_kptrs(map_ptr)) {
5505 ret = PTR_ERR_OR_ZERO(map_ptr->kptr_off_tab);
5507 verbose(env, "map '%s' has more than %d kptr\n", map_ptr->name,
5508 BPF_MAP_VALUE_OFF_MAX);
5509 else if (ret == -EEXIST)
5510 verbose(env, "map '%s' has repeating kptr BTF tags\n", map_ptr->name);
5512 verbose(env, "map '%s' has no valid kptr\n", map_ptr->name);
5516 meta->map_ptr = map_ptr;
5517 kptr_off = reg->off + reg->var_off.value;
5518 off_desc = bpf_map_kptr_off_contains(map_ptr, kptr_off);
5520 verbose(env, "off=%d doesn't point to kptr\n", kptr_off);
5523 if (off_desc->type != BPF_KPTR_REF) {
5524 verbose(env, "off=%d kptr isn't referenced kptr\n", kptr_off);
5527 meta->kptr_off_desc = off_desc;
5531 static bool arg_type_is_mem_size(enum bpf_arg_type type)
5533 return type == ARG_CONST_SIZE ||
5534 type == ARG_CONST_SIZE_OR_ZERO;
5537 static bool arg_type_is_alloc_size(enum bpf_arg_type type)
5539 return type == ARG_CONST_ALLOC_SIZE_OR_ZERO;
5542 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
5544 return type == ARG_PTR_TO_INT ||
5545 type == ARG_PTR_TO_LONG;
5548 static bool arg_type_is_release(enum bpf_arg_type type)
5550 return type & OBJ_RELEASE;
5553 static bool arg_type_is_dynptr(enum bpf_arg_type type)
5555 return base_type(type) == ARG_PTR_TO_DYNPTR;
5558 static int int_ptr_type_to_size(enum bpf_arg_type type)
5560 if (type == ARG_PTR_TO_INT)
5562 else if (type == ARG_PTR_TO_LONG)
5568 static int resolve_map_arg_type(struct bpf_verifier_env *env,
5569 const struct bpf_call_arg_meta *meta,
5570 enum bpf_arg_type *arg_type)
5572 if (!meta->map_ptr) {
5573 /* kernel subsystem misconfigured verifier */
5574 verbose(env, "invalid map_ptr to access map->type\n");
5578 switch (meta->map_ptr->map_type) {
5579 case BPF_MAP_TYPE_SOCKMAP:
5580 case BPF_MAP_TYPE_SOCKHASH:
5581 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
5582 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
5584 verbose(env, "invalid arg_type for sockmap/sockhash\n");
5588 case BPF_MAP_TYPE_BLOOM_FILTER:
5589 if (meta->func_id == BPF_FUNC_map_peek_elem)
5590 *arg_type = ARG_PTR_TO_MAP_VALUE;
5598 struct bpf_reg_types {
5599 const enum bpf_reg_type types[10];
5603 static const struct bpf_reg_types map_key_value_types = {
5613 static const struct bpf_reg_types sock_types = {
5623 static const struct bpf_reg_types btf_id_sock_common_types = {
5631 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
5635 static const struct bpf_reg_types mem_types = {
5643 PTR_TO_MEM | MEM_ALLOC,
5648 static const struct bpf_reg_types int_ptr_types = {
5658 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
5659 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
5660 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
5661 static const struct bpf_reg_types alloc_mem_types = { .types = { PTR_TO_MEM | MEM_ALLOC } };
5662 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
5663 static const struct bpf_reg_types btf_ptr_types = { .types = { PTR_TO_BTF_ID } };
5664 static const struct bpf_reg_types spin_lock_types = { .types = { PTR_TO_MAP_VALUE } };
5665 static const struct bpf_reg_types percpu_btf_ptr_types = { .types = { PTR_TO_BTF_ID | MEM_PERCPU } };
5666 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
5667 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
5668 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
5669 static const struct bpf_reg_types timer_types = { .types = { PTR_TO_MAP_VALUE } };
5670 static const struct bpf_reg_types kptr_types = { .types = { PTR_TO_MAP_VALUE } };
5672 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
5673 [ARG_PTR_TO_MAP_KEY] = &map_key_value_types,
5674 [ARG_PTR_TO_MAP_VALUE] = &map_key_value_types,
5675 [ARG_CONST_SIZE] = &scalar_types,
5676 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
5677 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
5678 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
5679 [ARG_PTR_TO_CTX] = &context_types,
5680 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
5682 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
5684 [ARG_PTR_TO_SOCKET] = &fullsock_types,
5685 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
5686 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
5687 [ARG_PTR_TO_MEM] = &mem_types,
5688 [ARG_PTR_TO_ALLOC_MEM] = &alloc_mem_types,
5689 [ARG_PTR_TO_INT] = &int_ptr_types,
5690 [ARG_PTR_TO_LONG] = &int_ptr_types,
5691 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
5692 [ARG_PTR_TO_FUNC] = &func_ptr_types,
5693 [ARG_PTR_TO_STACK] = &stack_ptr_types,
5694 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
5695 [ARG_PTR_TO_TIMER] = &timer_types,
5696 [ARG_PTR_TO_KPTR] = &kptr_types,
5697 [ARG_PTR_TO_DYNPTR] = &stack_ptr_types,
5700 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
5701 enum bpf_arg_type arg_type,
5702 const u32 *arg_btf_id,
5703 struct bpf_call_arg_meta *meta)
5705 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5706 enum bpf_reg_type expected, type = reg->type;
5707 const struct bpf_reg_types *compatible;
5710 compatible = compatible_reg_types[base_type(arg_type)];
5712 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
5716 /* ARG_PTR_TO_MEM + RDONLY is compatible with PTR_TO_MEM and PTR_TO_MEM + RDONLY,
5717 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM and NOT with PTR_TO_MEM + RDONLY
5719 * Same for MAYBE_NULL:
5721 * ARG_PTR_TO_MEM + MAYBE_NULL is compatible with PTR_TO_MEM and PTR_TO_MEM + MAYBE_NULL,
5722 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM but NOT with PTR_TO_MEM + MAYBE_NULL
5724 * Therefore we fold these flags depending on the arg_type before comparison.
5726 if (arg_type & MEM_RDONLY)
5727 type &= ~MEM_RDONLY;
5728 if (arg_type & PTR_MAYBE_NULL)
5729 type &= ~PTR_MAYBE_NULL;
5731 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
5732 expected = compatible->types[i];
5733 if (expected == NOT_INIT)
5736 if (type == expected)
5740 verbose(env, "R%d type=%s expected=", regno, reg_type_str(env, reg->type));
5741 for (j = 0; j + 1 < i; j++)
5742 verbose(env, "%s, ", reg_type_str(env, compatible->types[j]));
5743 verbose(env, "%s\n", reg_type_str(env, compatible->types[j]));
5747 if (reg->type == PTR_TO_BTF_ID) {
5748 /* For bpf_sk_release, it needs to match against first member
5749 * 'struct sock_common', hence make an exception for it. This
5750 * allows bpf_sk_release to work for multiple socket types.
5752 bool strict_type_match = arg_type_is_release(arg_type) &&
5753 meta->func_id != BPF_FUNC_sk_release;
5756 if (!compatible->btf_id) {
5757 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
5760 arg_btf_id = compatible->btf_id;
5763 if (meta->func_id == BPF_FUNC_kptr_xchg) {
5764 if (map_kptr_match_type(env, meta->kptr_off_desc, reg, regno))
5766 } else if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
5767 btf_vmlinux, *arg_btf_id,
5768 strict_type_match)) {
5769 verbose(env, "R%d is of type %s but %s is expected\n",
5770 regno, kernel_type_name(reg->btf, reg->btf_id),
5771 kernel_type_name(btf_vmlinux, *arg_btf_id));
5779 int check_func_arg_reg_off(struct bpf_verifier_env *env,
5780 const struct bpf_reg_state *reg, int regno,
5781 enum bpf_arg_type arg_type)
5783 enum bpf_reg_type type = reg->type;
5784 bool fixed_off_ok = false;
5786 switch ((u32)type) {
5787 /* Pointer types where reg offset is explicitly allowed: */
5789 if (arg_type_is_dynptr(arg_type) && reg->off % BPF_REG_SIZE) {
5790 verbose(env, "cannot pass in dynptr at an offset\n");
5795 case PTR_TO_PACKET_META:
5796 case PTR_TO_MAP_KEY:
5797 case PTR_TO_MAP_VALUE:
5799 case PTR_TO_MEM | MEM_RDONLY:
5800 case PTR_TO_MEM | MEM_ALLOC:
5802 case PTR_TO_BUF | MEM_RDONLY:
5804 /* Some of the argument types nevertheless require a
5805 * zero register offset.
5807 if (base_type(arg_type) != ARG_PTR_TO_ALLOC_MEM)
5810 /* All the rest must be rejected, except PTR_TO_BTF_ID which allows
5814 /* When referenced PTR_TO_BTF_ID is passed to release function,
5815 * it's fixed offset must be 0. In the other cases, fixed offset
5818 if (arg_type_is_release(arg_type) && reg->off) {
5819 verbose(env, "R%d must have zero offset when passed to release func\n",
5823 /* For arg is release pointer, fixed_off_ok must be false, but
5824 * we already checked and rejected reg->off != 0 above, so set
5825 * to true to allow fixed offset for all other cases.
5827 fixed_off_ok = true;
5832 return __check_ptr_off_reg(env, reg, regno, fixed_off_ok);
5835 static u32 stack_slot_get_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
5837 struct bpf_func_state *state = func(env, reg);
5838 int spi = get_spi(reg->off);
5840 return state->stack[spi].spilled_ptr.id;
5843 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
5844 struct bpf_call_arg_meta *meta,
5845 const struct bpf_func_proto *fn)
5847 u32 regno = BPF_REG_1 + arg;
5848 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5849 enum bpf_arg_type arg_type = fn->arg_type[arg];
5850 enum bpf_reg_type type = reg->type;
5853 if (arg_type == ARG_DONTCARE)
5856 err = check_reg_arg(env, regno, SRC_OP);
5860 if (arg_type == ARG_ANYTHING) {
5861 if (is_pointer_value(env, regno)) {
5862 verbose(env, "R%d leaks addr into helper function\n",
5869 if (type_is_pkt_pointer(type) &&
5870 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
5871 verbose(env, "helper access to the packet is not allowed\n");
5875 if (base_type(arg_type) == ARG_PTR_TO_MAP_VALUE) {
5876 err = resolve_map_arg_type(env, meta, &arg_type);
5881 if (register_is_null(reg) && type_may_be_null(arg_type))
5882 /* A NULL register has a SCALAR_VALUE type, so skip
5885 goto skip_type_check;
5887 err = check_reg_type(env, regno, arg_type, fn->arg_btf_id[arg], meta);
5891 err = check_func_arg_reg_off(env, reg, regno, arg_type);
5896 if (arg_type_is_release(arg_type)) {
5897 if (arg_type_is_dynptr(arg_type)) {
5898 struct bpf_func_state *state = func(env, reg);
5899 int spi = get_spi(reg->off);
5901 if (!is_spi_bounds_valid(state, spi, BPF_DYNPTR_NR_SLOTS) ||
5902 !state->stack[spi].spilled_ptr.id) {
5903 verbose(env, "arg %d is an unacquired reference\n", regno);
5906 } else if (!reg->ref_obj_id && !register_is_null(reg)) {
5907 verbose(env, "R%d must be referenced when passed to release function\n",
5911 if (meta->release_regno) {
5912 verbose(env, "verifier internal error: more than one release argument\n");
5915 meta->release_regno = regno;
5918 if (reg->ref_obj_id) {
5919 if (meta->ref_obj_id) {
5920 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
5921 regno, reg->ref_obj_id,
5925 meta->ref_obj_id = reg->ref_obj_id;
5928 if (arg_type == ARG_CONST_MAP_PTR) {
5929 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
5930 if (meta->map_ptr) {
5931 /* Use map_uid (which is unique id of inner map) to reject:
5932 * inner_map1 = bpf_map_lookup_elem(outer_map, key1)
5933 * inner_map2 = bpf_map_lookup_elem(outer_map, key2)
5934 * if (inner_map1 && inner_map2) {
5935 * timer = bpf_map_lookup_elem(inner_map1);
5937 * // mismatch would have been allowed
5938 * bpf_timer_init(timer, inner_map2);
5941 * Comparing map_ptr is enough to distinguish normal and outer maps.
5943 if (meta->map_ptr != reg->map_ptr ||
5944 meta->map_uid != reg->map_uid) {
5946 "timer pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
5947 meta->map_uid, reg->map_uid);
5951 meta->map_ptr = reg->map_ptr;
5952 meta->map_uid = reg->map_uid;
5953 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
5954 /* bpf_map_xxx(..., map_ptr, ..., key) call:
5955 * check that [key, key + map->key_size) are within
5956 * stack limits and initialized
5958 if (!meta->map_ptr) {
5959 /* in function declaration map_ptr must come before
5960 * map_key, so that it's verified and known before
5961 * we have to check map_key here. Otherwise it means
5962 * that kernel subsystem misconfigured verifier
5964 verbose(env, "invalid map_ptr to access map->key\n");
5967 err = check_helper_mem_access(env, regno,
5968 meta->map_ptr->key_size, false,
5970 } else if (base_type(arg_type) == ARG_PTR_TO_MAP_VALUE) {
5971 if (type_may_be_null(arg_type) && register_is_null(reg))
5974 /* bpf_map_xxx(..., map_ptr, ..., value) call:
5975 * check [value, value + map->value_size) validity
5977 if (!meta->map_ptr) {
5978 /* kernel subsystem misconfigured verifier */
5979 verbose(env, "invalid map_ptr to access map->value\n");
5982 meta->raw_mode = arg_type & MEM_UNINIT;
5983 err = check_helper_mem_access(env, regno,
5984 meta->map_ptr->value_size, false,
5986 } else if (arg_type == ARG_PTR_TO_PERCPU_BTF_ID) {
5988 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
5991 meta->ret_btf = reg->btf;
5992 meta->ret_btf_id = reg->btf_id;
5993 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
5994 if (meta->func_id == BPF_FUNC_spin_lock) {
5995 if (process_spin_lock(env, regno, true))
5997 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
5998 if (process_spin_lock(env, regno, false))
6001 verbose(env, "verifier internal error\n");
6004 } else if (arg_type == ARG_PTR_TO_TIMER) {
6005 if (process_timer_func(env, regno, meta))
6007 } else if (arg_type == ARG_PTR_TO_FUNC) {
6008 meta->subprogno = reg->subprogno;
6009 } else if (base_type(arg_type) == ARG_PTR_TO_MEM) {
6010 /* The access to this pointer is only checked when we hit the
6011 * next is_mem_size argument below.
6013 meta->raw_mode = arg_type & MEM_UNINIT;
6014 } else if (arg_type_is_mem_size(arg_type)) {
6015 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
6017 err = check_mem_size_reg(env, reg, regno, zero_size_allowed, meta);
6018 } else if (arg_type_is_dynptr(arg_type)) {
6019 if (arg_type & MEM_UNINIT) {
6020 if (!is_dynptr_reg_valid_uninit(env, reg)) {
6021 verbose(env, "Dynptr has to be an uninitialized dynptr\n");
6025 /* We only support one dynptr being uninitialized at the moment,
6026 * which is sufficient for the helper functions we have right now.
6028 if (meta->uninit_dynptr_regno) {
6029 verbose(env, "verifier internal error: multiple uninitialized dynptr args\n");
6033 meta->uninit_dynptr_regno = regno;
6034 } else if (!is_dynptr_reg_valid_init(env, reg, arg_type)) {
6035 const char *err_extra = "";
6037 switch (arg_type & DYNPTR_TYPE_FLAG_MASK) {
6038 case DYNPTR_TYPE_LOCAL:
6039 err_extra = "local ";
6041 case DYNPTR_TYPE_RINGBUF:
6042 err_extra = "ringbuf ";
6048 verbose(env, "Expected an initialized %sdynptr as arg #%d\n",
6049 err_extra, arg + 1);
6052 } else if (arg_type_is_alloc_size(arg_type)) {
6053 if (!tnum_is_const(reg->var_off)) {
6054 verbose(env, "R%d is not a known constant'\n",
6058 meta->mem_size = reg->var_off.value;
6059 } else if (arg_type_is_int_ptr(arg_type)) {
6060 int size = int_ptr_type_to_size(arg_type);
6062 err = check_helper_mem_access(env, regno, size, false, meta);
6065 err = check_ptr_alignment(env, reg, 0, size, true);
6066 } else if (arg_type == ARG_PTR_TO_CONST_STR) {
6067 struct bpf_map *map = reg->map_ptr;
6072 if (!bpf_map_is_rdonly(map)) {
6073 verbose(env, "R%d does not point to a readonly map'\n", regno);
6077 if (!tnum_is_const(reg->var_off)) {
6078 verbose(env, "R%d is not a constant address'\n", regno);
6082 if (!map->ops->map_direct_value_addr) {
6083 verbose(env, "no direct value access support for this map type\n");
6087 err = check_map_access(env, regno, reg->off,
6088 map->value_size - reg->off, false,
6093 map_off = reg->off + reg->var_off.value;
6094 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
6096 verbose(env, "direct value access on string failed\n");
6100 str_ptr = (char *)(long)(map_addr);
6101 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
6102 verbose(env, "string is not zero-terminated\n");
6105 } else if (arg_type == ARG_PTR_TO_KPTR) {
6106 if (process_kptr_func(env, regno, meta))
6113 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
6115 enum bpf_attach_type eatype = env->prog->expected_attach_type;
6116 enum bpf_prog_type type = resolve_prog_type(env->prog);
6118 if (func_id != BPF_FUNC_map_update_elem)
6121 /* It's not possible to get access to a locked struct sock in these
6122 * contexts, so updating is safe.
6125 case BPF_PROG_TYPE_TRACING:
6126 if (eatype == BPF_TRACE_ITER)
6129 case BPF_PROG_TYPE_SOCKET_FILTER:
6130 case BPF_PROG_TYPE_SCHED_CLS:
6131 case BPF_PROG_TYPE_SCHED_ACT:
6132 case BPF_PROG_TYPE_XDP:
6133 case BPF_PROG_TYPE_SK_REUSEPORT:
6134 case BPF_PROG_TYPE_FLOW_DISSECTOR:
6135 case BPF_PROG_TYPE_SK_LOOKUP:
6141 verbose(env, "cannot update sockmap in this context\n");
6145 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
6147 return env->prog->jit_requested && IS_ENABLED(CONFIG_X86_64);
6150 static int check_map_func_compatibility(struct bpf_verifier_env *env,
6151 struct bpf_map *map, int func_id)
6156 /* We need a two way check, first is from map perspective ... */
6157 switch (map->map_type) {
6158 case BPF_MAP_TYPE_PROG_ARRAY:
6159 if (func_id != BPF_FUNC_tail_call)
6162 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
6163 if (func_id != BPF_FUNC_perf_event_read &&
6164 func_id != BPF_FUNC_perf_event_output &&
6165 func_id != BPF_FUNC_skb_output &&
6166 func_id != BPF_FUNC_perf_event_read_value &&
6167 func_id != BPF_FUNC_xdp_output)
6170 case BPF_MAP_TYPE_RINGBUF:
6171 if (func_id != BPF_FUNC_ringbuf_output &&
6172 func_id != BPF_FUNC_ringbuf_reserve &&
6173 func_id != BPF_FUNC_ringbuf_query &&
6174 func_id != BPF_FUNC_ringbuf_reserve_dynptr &&
6175 func_id != BPF_FUNC_ringbuf_submit_dynptr &&
6176 func_id != BPF_FUNC_ringbuf_discard_dynptr)
6179 case BPF_MAP_TYPE_STACK_TRACE:
6180 if (func_id != BPF_FUNC_get_stackid)
6183 case BPF_MAP_TYPE_CGROUP_ARRAY:
6184 if (func_id != BPF_FUNC_skb_under_cgroup &&
6185 func_id != BPF_FUNC_current_task_under_cgroup)
6188 case BPF_MAP_TYPE_CGROUP_STORAGE:
6189 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
6190 if (func_id != BPF_FUNC_get_local_storage)
6193 case BPF_MAP_TYPE_DEVMAP:
6194 case BPF_MAP_TYPE_DEVMAP_HASH:
6195 if (func_id != BPF_FUNC_redirect_map &&
6196 func_id != BPF_FUNC_map_lookup_elem)
6199 /* Restrict bpf side of cpumap and xskmap, open when use-cases
6202 case BPF_MAP_TYPE_CPUMAP:
6203 if (func_id != BPF_FUNC_redirect_map)
6206 case BPF_MAP_TYPE_XSKMAP:
6207 if (func_id != BPF_FUNC_redirect_map &&
6208 func_id != BPF_FUNC_map_lookup_elem)
6211 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
6212 case BPF_MAP_TYPE_HASH_OF_MAPS:
6213 if (func_id != BPF_FUNC_map_lookup_elem)
6216 case BPF_MAP_TYPE_SOCKMAP:
6217 if (func_id != BPF_FUNC_sk_redirect_map &&
6218 func_id != BPF_FUNC_sock_map_update &&
6219 func_id != BPF_FUNC_map_delete_elem &&
6220 func_id != BPF_FUNC_msg_redirect_map &&
6221 func_id != BPF_FUNC_sk_select_reuseport &&
6222 func_id != BPF_FUNC_map_lookup_elem &&
6223 !may_update_sockmap(env, func_id))
6226 case BPF_MAP_TYPE_SOCKHASH:
6227 if (func_id != BPF_FUNC_sk_redirect_hash &&
6228 func_id != BPF_FUNC_sock_hash_update &&
6229 func_id != BPF_FUNC_map_delete_elem &&
6230 func_id != BPF_FUNC_msg_redirect_hash &&
6231 func_id != BPF_FUNC_sk_select_reuseport &&
6232 func_id != BPF_FUNC_map_lookup_elem &&
6233 !may_update_sockmap(env, func_id))
6236 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
6237 if (func_id != BPF_FUNC_sk_select_reuseport)
6240 case BPF_MAP_TYPE_QUEUE:
6241 case BPF_MAP_TYPE_STACK:
6242 if (func_id != BPF_FUNC_map_peek_elem &&
6243 func_id != BPF_FUNC_map_pop_elem &&
6244 func_id != BPF_FUNC_map_push_elem)
6247 case BPF_MAP_TYPE_SK_STORAGE:
6248 if (func_id != BPF_FUNC_sk_storage_get &&
6249 func_id != BPF_FUNC_sk_storage_delete)
6252 case BPF_MAP_TYPE_INODE_STORAGE:
6253 if (func_id != BPF_FUNC_inode_storage_get &&
6254 func_id != BPF_FUNC_inode_storage_delete)
6257 case BPF_MAP_TYPE_TASK_STORAGE:
6258 if (func_id != BPF_FUNC_task_storage_get &&
6259 func_id != BPF_FUNC_task_storage_delete)
6262 case BPF_MAP_TYPE_BLOOM_FILTER:
6263 if (func_id != BPF_FUNC_map_peek_elem &&
6264 func_id != BPF_FUNC_map_push_elem)
6271 /* ... and second from the function itself. */
6273 case BPF_FUNC_tail_call:
6274 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
6276 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
6277 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
6281 case BPF_FUNC_perf_event_read:
6282 case BPF_FUNC_perf_event_output:
6283 case BPF_FUNC_perf_event_read_value:
6284 case BPF_FUNC_skb_output:
6285 case BPF_FUNC_xdp_output:
6286 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
6289 case BPF_FUNC_ringbuf_output:
6290 case BPF_FUNC_ringbuf_reserve:
6291 case BPF_FUNC_ringbuf_query:
6292 case BPF_FUNC_ringbuf_reserve_dynptr:
6293 case BPF_FUNC_ringbuf_submit_dynptr:
6294 case BPF_FUNC_ringbuf_discard_dynptr:
6295 if (map->map_type != BPF_MAP_TYPE_RINGBUF)
6298 case BPF_FUNC_get_stackid:
6299 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
6302 case BPF_FUNC_current_task_under_cgroup:
6303 case BPF_FUNC_skb_under_cgroup:
6304 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
6307 case BPF_FUNC_redirect_map:
6308 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
6309 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
6310 map->map_type != BPF_MAP_TYPE_CPUMAP &&
6311 map->map_type != BPF_MAP_TYPE_XSKMAP)
6314 case BPF_FUNC_sk_redirect_map:
6315 case BPF_FUNC_msg_redirect_map:
6316 case BPF_FUNC_sock_map_update:
6317 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
6320 case BPF_FUNC_sk_redirect_hash:
6321 case BPF_FUNC_msg_redirect_hash:
6322 case BPF_FUNC_sock_hash_update:
6323 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
6326 case BPF_FUNC_get_local_storage:
6327 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
6328 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
6331 case BPF_FUNC_sk_select_reuseport:
6332 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
6333 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
6334 map->map_type != BPF_MAP_TYPE_SOCKHASH)
6337 case BPF_FUNC_map_pop_elem:
6338 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
6339 map->map_type != BPF_MAP_TYPE_STACK)
6342 case BPF_FUNC_map_peek_elem:
6343 case BPF_FUNC_map_push_elem:
6344 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
6345 map->map_type != BPF_MAP_TYPE_STACK &&
6346 map->map_type != BPF_MAP_TYPE_BLOOM_FILTER)
6349 case BPF_FUNC_map_lookup_percpu_elem:
6350 if (map->map_type != BPF_MAP_TYPE_PERCPU_ARRAY &&
6351 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
6352 map->map_type != BPF_MAP_TYPE_LRU_PERCPU_HASH)
6355 case BPF_FUNC_sk_storage_get:
6356 case BPF_FUNC_sk_storage_delete:
6357 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
6360 case BPF_FUNC_inode_storage_get:
6361 case BPF_FUNC_inode_storage_delete:
6362 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
6365 case BPF_FUNC_task_storage_get:
6366 case BPF_FUNC_task_storage_delete:
6367 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
6376 verbose(env, "cannot pass map_type %d into func %s#%d\n",
6377 map->map_type, func_id_name(func_id), func_id);
6381 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
6385 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
6387 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
6389 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
6391 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
6393 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
6396 /* We only support one arg being in raw mode at the moment,
6397 * which is sufficient for the helper functions we have
6403 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
6404 enum bpf_arg_type arg_next)
6406 return (base_type(arg_curr) == ARG_PTR_TO_MEM) !=
6407 arg_type_is_mem_size(arg_next);
6410 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
6412 /* bpf_xxx(..., buf, len) call will access 'len'
6413 * bytes from memory 'buf'. Both arg types need
6414 * to be paired, so make sure there's no buggy
6415 * helper function specification.
6417 if (arg_type_is_mem_size(fn->arg1_type) ||
6418 base_type(fn->arg5_type) == ARG_PTR_TO_MEM ||
6419 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
6420 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
6421 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
6422 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
6428 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
6432 if (arg_type_may_be_refcounted(fn->arg1_type))
6434 if (arg_type_may_be_refcounted(fn->arg2_type))
6436 if (arg_type_may_be_refcounted(fn->arg3_type))
6438 if (arg_type_may_be_refcounted(fn->arg4_type))
6440 if (arg_type_may_be_refcounted(fn->arg5_type))
6443 /* A reference acquiring function cannot acquire
6444 * another refcounted ptr.
6446 if (may_be_acquire_function(func_id) && count)
6449 /* We only support one arg being unreferenced at the moment,
6450 * which is sufficient for the helper functions we have right now.
6455 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
6459 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
6460 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_BTF_ID && !fn->arg_btf_id[i])
6463 if (base_type(fn->arg_type[i]) != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i])
6470 static int check_func_proto(const struct bpf_func_proto *fn, int func_id,
6471 struct bpf_call_arg_meta *meta)
6473 return check_raw_mode_ok(fn) &&
6474 check_arg_pair_ok(fn) &&
6475 check_btf_id_ok(fn) &&
6476 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
6479 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
6480 * are now invalid, so turn them into unknown SCALAR_VALUE.
6482 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
6483 struct bpf_func_state *state)
6485 struct bpf_reg_state *regs = state->regs, *reg;
6488 for (i = 0; i < MAX_BPF_REG; i++)
6489 if (reg_is_pkt_pointer_any(®s[i]))
6490 mark_reg_unknown(env, regs, i);
6492 bpf_for_each_spilled_reg(i, state, reg) {
6495 if (reg_is_pkt_pointer_any(reg))
6496 __mark_reg_unknown(env, reg);
6500 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
6502 struct bpf_verifier_state *vstate = env->cur_state;
6505 for (i = 0; i <= vstate->curframe; i++)
6506 __clear_all_pkt_pointers(env, vstate->frame[i]);
6511 BEYOND_PKT_END = -2,
6514 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
6516 struct bpf_func_state *state = vstate->frame[vstate->curframe];
6517 struct bpf_reg_state *reg = &state->regs[regn];
6519 if (reg->type != PTR_TO_PACKET)
6520 /* PTR_TO_PACKET_META is not supported yet */
6523 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
6524 * How far beyond pkt_end it goes is unknown.
6525 * if (!range_open) it's the case of pkt >= pkt_end
6526 * if (range_open) it's the case of pkt > pkt_end
6527 * hence this pointer is at least 1 byte bigger than pkt_end
6530 reg->range = BEYOND_PKT_END;
6532 reg->range = AT_PKT_END;
6535 static void release_reg_references(struct bpf_verifier_env *env,
6536 struct bpf_func_state *state,
6539 struct bpf_reg_state *regs = state->regs, *reg;
6542 for (i = 0; i < MAX_BPF_REG; i++)
6543 if (regs[i].ref_obj_id == ref_obj_id)
6544 mark_reg_unknown(env, regs, i);
6546 bpf_for_each_spilled_reg(i, state, reg) {
6549 if (reg->ref_obj_id == ref_obj_id)
6550 __mark_reg_unknown(env, reg);
6554 /* The pointer with the specified id has released its reference to kernel
6555 * resources. Identify all copies of the same pointer and clear the reference.
6557 static int release_reference(struct bpf_verifier_env *env,
6560 struct bpf_verifier_state *vstate = env->cur_state;
6564 err = release_reference_state(cur_func(env), ref_obj_id);
6568 for (i = 0; i <= vstate->curframe; i++)
6569 release_reg_references(env, vstate->frame[i], ref_obj_id);
6574 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
6575 struct bpf_reg_state *regs)
6579 /* after the call registers r0 - r5 were scratched */
6580 for (i = 0; i < CALLER_SAVED_REGS; i++) {
6581 mark_reg_not_init(env, regs, caller_saved[i]);
6582 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
6586 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
6587 struct bpf_func_state *caller,
6588 struct bpf_func_state *callee,
6591 static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
6592 int *insn_idx, int subprog,
6593 set_callee_state_fn set_callee_state_cb)
6595 struct bpf_verifier_state *state = env->cur_state;
6596 struct bpf_func_info_aux *func_info_aux;
6597 struct bpf_func_state *caller, *callee;
6599 bool is_global = false;
6601 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
6602 verbose(env, "the call stack of %d frames is too deep\n",
6603 state->curframe + 2);
6607 caller = state->frame[state->curframe];
6608 if (state->frame[state->curframe + 1]) {
6609 verbose(env, "verifier bug. Frame %d already allocated\n",
6610 state->curframe + 1);
6614 func_info_aux = env->prog->aux->func_info_aux;
6616 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
6617 err = btf_check_subprog_arg_match(env, subprog, caller->regs);
6622 verbose(env, "Caller passes invalid args into func#%d\n",
6626 if (env->log.level & BPF_LOG_LEVEL)
6628 "Func#%d is global and valid. Skipping.\n",
6630 clear_caller_saved_regs(env, caller->regs);
6632 /* All global functions return a 64-bit SCALAR_VALUE */
6633 mark_reg_unknown(env, caller->regs, BPF_REG_0);
6634 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
6636 /* continue with next insn after call */
6641 if (insn->code == (BPF_JMP | BPF_CALL) &&
6642 insn->src_reg == 0 &&
6643 insn->imm == BPF_FUNC_timer_set_callback) {
6644 struct bpf_verifier_state *async_cb;
6646 /* there is no real recursion here. timer callbacks are async */
6647 env->subprog_info[subprog].is_async_cb = true;
6648 async_cb = push_async_cb(env, env->subprog_info[subprog].start,
6649 *insn_idx, subprog);
6652 callee = async_cb->frame[0];
6653 callee->async_entry_cnt = caller->async_entry_cnt + 1;
6655 /* Convert bpf_timer_set_callback() args into timer callback args */
6656 err = set_callee_state_cb(env, caller, callee, *insn_idx);
6660 clear_caller_saved_regs(env, caller->regs);
6661 mark_reg_unknown(env, caller->regs, BPF_REG_0);
6662 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
6663 /* continue with next insn after call */
6667 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
6670 state->frame[state->curframe + 1] = callee;
6672 /* callee cannot access r0, r6 - r9 for reading and has to write
6673 * into its own stack before reading from it.
6674 * callee can read/write into caller's stack
6676 init_func_state(env, callee,
6677 /* remember the callsite, it will be used by bpf_exit */
6678 *insn_idx /* callsite */,
6679 state->curframe + 1 /* frameno within this callchain */,
6680 subprog /* subprog number within this prog */);
6682 /* Transfer references to the callee */
6683 err = copy_reference_state(callee, caller);
6687 err = set_callee_state_cb(env, caller, callee, *insn_idx);
6691 clear_caller_saved_regs(env, caller->regs);
6693 /* only increment it after check_reg_arg() finished */
6696 /* and go analyze first insn of the callee */
6697 *insn_idx = env->subprog_info[subprog].start - 1;
6699 if (env->log.level & BPF_LOG_LEVEL) {
6700 verbose(env, "caller:\n");
6701 print_verifier_state(env, caller, true);
6702 verbose(env, "callee:\n");
6703 print_verifier_state(env, callee, true);
6708 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
6709 struct bpf_func_state *caller,
6710 struct bpf_func_state *callee)
6712 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
6713 * void *callback_ctx, u64 flags);
6714 * callback_fn(struct bpf_map *map, void *key, void *value,
6715 * void *callback_ctx);
6717 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
6719 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
6720 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
6721 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
6723 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
6724 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
6725 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
6727 /* pointer to stack or null */
6728 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
6731 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
6735 static int set_callee_state(struct bpf_verifier_env *env,
6736 struct bpf_func_state *caller,
6737 struct bpf_func_state *callee, int insn_idx)
6741 /* copy r1 - r5 args that callee can access. The copy includes parent
6742 * pointers, which connects us up to the liveness chain
6744 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
6745 callee->regs[i] = caller->regs[i];
6749 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
6752 int subprog, target_insn;
6754 target_insn = *insn_idx + insn->imm + 1;
6755 subprog = find_subprog(env, target_insn);
6757 verbose(env, "verifier bug. No program starts at insn %d\n",
6762 return __check_func_call(env, insn, insn_idx, subprog, set_callee_state);
6765 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
6766 struct bpf_func_state *caller,
6767 struct bpf_func_state *callee,
6770 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
6771 struct bpf_map *map;
6774 if (bpf_map_ptr_poisoned(insn_aux)) {
6775 verbose(env, "tail_call abusing map_ptr\n");
6779 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
6780 if (!map->ops->map_set_for_each_callback_args ||
6781 !map->ops->map_for_each_callback) {
6782 verbose(env, "callback function not allowed for map\n");
6786 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
6790 callee->in_callback_fn = true;
6794 static int set_loop_callback_state(struct bpf_verifier_env *env,
6795 struct bpf_func_state *caller,
6796 struct bpf_func_state *callee,
6799 /* bpf_loop(u32 nr_loops, void *callback_fn, void *callback_ctx,
6801 * callback_fn(u32 index, void *callback_ctx);
6803 callee->regs[BPF_REG_1].type = SCALAR_VALUE;
6804 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
6807 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
6808 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
6809 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
6811 callee->in_callback_fn = true;
6815 static int set_timer_callback_state(struct bpf_verifier_env *env,
6816 struct bpf_func_state *caller,
6817 struct bpf_func_state *callee,
6820 struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
6822 /* bpf_timer_set_callback(struct bpf_timer *timer, void *callback_fn);
6823 * callback_fn(struct bpf_map *map, void *key, void *value);
6825 callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP;
6826 __mark_reg_known_zero(&callee->regs[BPF_REG_1]);
6827 callee->regs[BPF_REG_1].map_ptr = map_ptr;
6829 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
6830 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
6831 callee->regs[BPF_REG_2].map_ptr = map_ptr;
6833 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
6834 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
6835 callee->regs[BPF_REG_3].map_ptr = map_ptr;
6838 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
6839 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
6840 callee->in_async_callback_fn = true;
6844 static int set_find_vma_callback_state(struct bpf_verifier_env *env,
6845 struct bpf_func_state *caller,
6846 struct bpf_func_state *callee,
6849 /* bpf_find_vma(struct task_struct *task, u64 addr,
6850 * void *callback_fn, void *callback_ctx, u64 flags)
6851 * (callback_fn)(struct task_struct *task,
6852 * struct vm_area_struct *vma, void *callback_ctx);
6854 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
6856 callee->regs[BPF_REG_2].type = PTR_TO_BTF_ID;
6857 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
6858 callee->regs[BPF_REG_2].btf = btf_vmlinux;
6859 callee->regs[BPF_REG_2].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_VMA],
6861 /* pointer to stack or null */
6862 callee->regs[BPF_REG_3] = caller->regs[BPF_REG_4];
6865 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
6866 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
6867 callee->in_callback_fn = true;
6871 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
6873 struct bpf_verifier_state *state = env->cur_state;
6874 struct bpf_func_state *caller, *callee;
6875 struct bpf_reg_state *r0;
6878 callee = state->frame[state->curframe];
6879 r0 = &callee->regs[BPF_REG_0];
6880 if (r0->type == PTR_TO_STACK) {
6881 /* technically it's ok to return caller's stack pointer
6882 * (or caller's caller's pointer) back to the caller,
6883 * since these pointers are valid. Only current stack
6884 * pointer will be invalid as soon as function exits,
6885 * but let's be conservative
6887 verbose(env, "cannot return stack pointer to the caller\n");
6892 caller = state->frame[state->curframe];
6893 if (callee->in_callback_fn) {
6894 /* enforce R0 return value range [0, 1]. */
6895 struct tnum range = tnum_range(0, 1);
6897 if (r0->type != SCALAR_VALUE) {
6898 verbose(env, "R0 not a scalar value\n");
6901 if (!tnum_in(range, r0->var_off)) {
6902 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
6906 /* return to the caller whatever r0 had in the callee */
6907 caller->regs[BPF_REG_0] = *r0;
6910 /* Transfer references to the caller */
6911 err = copy_reference_state(caller, callee);
6915 *insn_idx = callee->callsite + 1;
6916 if (env->log.level & BPF_LOG_LEVEL) {
6917 verbose(env, "returning from callee:\n");
6918 print_verifier_state(env, callee, true);
6919 verbose(env, "to caller at %d:\n", *insn_idx);
6920 print_verifier_state(env, caller, true);
6922 /* clear everything in the callee */
6923 free_func_state(callee);
6924 state->frame[state->curframe + 1] = NULL;
6928 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
6930 struct bpf_call_arg_meta *meta)
6932 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
6934 if (ret_type != RET_INTEGER ||
6935 (func_id != BPF_FUNC_get_stack &&
6936 func_id != BPF_FUNC_get_task_stack &&
6937 func_id != BPF_FUNC_probe_read_str &&
6938 func_id != BPF_FUNC_probe_read_kernel_str &&
6939 func_id != BPF_FUNC_probe_read_user_str))
6942 ret_reg->smax_value = meta->msize_max_value;
6943 ret_reg->s32_max_value = meta->msize_max_value;
6944 ret_reg->smin_value = -MAX_ERRNO;
6945 ret_reg->s32_min_value = -MAX_ERRNO;
6946 __reg_deduce_bounds(ret_reg);
6947 __reg_bound_offset(ret_reg);
6948 __update_reg_bounds(ret_reg);
6952 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
6953 int func_id, int insn_idx)
6955 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
6956 struct bpf_map *map = meta->map_ptr;
6958 if (func_id != BPF_FUNC_tail_call &&
6959 func_id != BPF_FUNC_map_lookup_elem &&
6960 func_id != BPF_FUNC_map_update_elem &&
6961 func_id != BPF_FUNC_map_delete_elem &&
6962 func_id != BPF_FUNC_map_push_elem &&
6963 func_id != BPF_FUNC_map_pop_elem &&
6964 func_id != BPF_FUNC_map_peek_elem &&
6965 func_id != BPF_FUNC_for_each_map_elem &&
6966 func_id != BPF_FUNC_redirect_map &&
6967 func_id != BPF_FUNC_map_lookup_percpu_elem)
6971 verbose(env, "kernel subsystem misconfigured verifier\n");
6975 /* In case of read-only, some additional restrictions
6976 * need to be applied in order to prevent altering the
6977 * state of the map from program side.
6979 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
6980 (func_id == BPF_FUNC_map_delete_elem ||
6981 func_id == BPF_FUNC_map_update_elem ||
6982 func_id == BPF_FUNC_map_push_elem ||
6983 func_id == BPF_FUNC_map_pop_elem)) {
6984 verbose(env, "write into map forbidden\n");
6988 if (!BPF_MAP_PTR(aux->map_ptr_state))
6989 bpf_map_ptr_store(aux, meta->map_ptr,
6990 !meta->map_ptr->bypass_spec_v1);
6991 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
6992 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
6993 !meta->map_ptr->bypass_spec_v1);
6998 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
6999 int func_id, int insn_idx)
7001 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
7002 struct bpf_reg_state *regs = cur_regs(env), *reg;
7003 struct bpf_map *map = meta->map_ptr;
7008 if (func_id != BPF_FUNC_tail_call)
7010 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
7011 verbose(env, "kernel subsystem misconfigured verifier\n");
7015 range = tnum_range(0, map->max_entries - 1);
7016 reg = ®s[BPF_REG_3];
7018 if (!register_is_const(reg) || !tnum_in(range, reg->var_off)) {
7019 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
7023 err = mark_chain_precision(env, BPF_REG_3);
7027 val = reg->var_off.value;
7028 if (bpf_map_key_unseen(aux))
7029 bpf_map_key_store(aux, val);
7030 else if (!bpf_map_key_poisoned(aux) &&
7031 bpf_map_key_immediate(aux) != val)
7032 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
7036 static int check_reference_leak(struct bpf_verifier_env *env)
7038 struct bpf_func_state *state = cur_func(env);
7041 for (i = 0; i < state->acquired_refs; i++) {
7042 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
7043 state->refs[i].id, state->refs[i].insn_idx);
7045 return state->acquired_refs ? -EINVAL : 0;
7048 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
7049 struct bpf_reg_state *regs)
7051 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
7052 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
7053 struct bpf_map *fmt_map = fmt_reg->map_ptr;
7054 int err, fmt_map_off, num_args;
7058 /* data must be an array of u64 */
7059 if (data_len_reg->var_off.value % 8)
7061 num_args = data_len_reg->var_off.value / 8;
7063 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
7064 * and map_direct_value_addr is set.
7066 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
7067 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
7070 verbose(env, "verifier bug\n");
7073 fmt = (char *)(long)fmt_addr + fmt_map_off;
7075 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
7076 * can focus on validating the format specifiers.
7078 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, NULL, num_args);
7080 verbose(env, "Invalid format string\n");
7085 static int check_get_func_ip(struct bpf_verifier_env *env)
7087 enum bpf_prog_type type = resolve_prog_type(env->prog);
7088 int func_id = BPF_FUNC_get_func_ip;
7090 if (type == BPF_PROG_TYPE_TRACING) {
7091 if (!bpf_prog_has_trampoline(env->prog)) {
7092 verbose(env, "func %s#%d supported only for fentry/fexit/fmod_ret programs\n",
7093 func_id_name(func_id), func_id);
7097 } else if (type == BPF_PROG_TYPE_KPROBE) {
7101 verbose(env, "func %s#%d not supported for program type %d\n",
7102 func_id_name(func_id), func_id, type);
7106 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
7109 const struct bpf_func_proto *fn = NULL;
7110 enum bpf_return_type ret_type;
7111 enum bpf_type_flag ret_flag;
7112 struct bpf_reg_state *regs;
7113 struct bpf_call_arg_meta meta;
7114 int insn_idx = *insn_idx_p;
7116 int i, err, func_id;
7118 /* find function prototype */
7119 func_id = insn->imm;
7120 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
7121 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
7126 if (env->ops->get_func_proto)
7127 fn = env->ops->get_func_proto(func_id, env->prog);
7129 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
7134 /* eBPF programs must be GPL compatible to use GPL-ed functions */
7135 if (!env->prog->gpl_compatible && fn->gpl_only) {
7136 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
7140 if (fn->allowed && !fn->allowed(env->prog)) {
7141 verbose(env, "helper call is not allowed in probe\n");
7145 /* With LD_ABS/IND some JITs save/restore skb from r1. */
7146 changes_data = bpf_helper_changes_pkt_data(fn->func);
7147 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
7148 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
7149 func_id_name(func_id), func_id);
7153 memset(&meta, 0, sizeof(meta));
7154 meta.pkt_access = fn->pkt_access;
7156 err = check_func_proto(fn, func_id, &meta);
7158 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
7159 func_id_name(func_id), func_id);
7163 meta.func_id = func_id;
7165 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
7166 err = check_func_arg(env, i, &meta, fn);
7171 err = record_func_map(env, &meta, func_id, insn_idx);
7175 err = record_func_key(env, &meta, func_id, insn_idx);
7179 /* Mark slots with STACK_MISC in case of raw mode, stack offset
7180 * is inferred from register state.
7182 for (i = 0; i < meta.access_size; i++) {
7183 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
7184 BPF_WRITE, -1, false);
7189 regs = cur_regs(env);
7191 if (meta.uninit_dynptr_regno) {
7192 /* we write BPF_DW bits (8 bytes) at a time */
7193 for (i = 0; i < BPF_DYNPTR_SIZE; i += 8) {
7194 err = check_mem_access(env, insn_idx, meta.uninit_dynptr_regno,
7195 i, BPF_DW, BPF_WRITE, -1, false);
7200 err = mark_stack_slots_dynptr(env, ®s[meta.uninit_dynptr_regno],
7201 fn->arg_type[meta.uninit_dynptr_regno - BPF_REG_1],
7207 if (meta.release_regno) {
7209 if (arg_type_is_dynptr(fn->arg_type[meta.release_regno - BPF_REG_1]))
7210 err = unmark_stack_slots_dynptr(env, ®s[meta.release_regno]);
7211 else if (meta.ref_obj_id)
7212 err = release_reference(env, meta.ref_obj_id);
7213 /* meta.ref_obj_id can only be 0 if register that is meant to be
7214 * released is NULL, which must be > R0.
7216 else if (register_is_null(®s[meta.release_regno]))
7219 verbose(env, "func %s#%d reference has not been acquired before\n",
7220 func_id_name(func_id), func_id);
7226 case BPF_FUNC_tail_call:
7227 err = check_reference_leak(env);
7229 verbose(env, "tail_call would lead to reference leak\n");
7233 case BPF_FUNC_get_local_storage:
7234 /* check that flags argument in get_local_storage(map, flags) is 0,
7235 * this is required because get_local_storage() can't return an error.
7237 if (!register_is_null(®s[BPF_REG_2])) {
7238 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
7242 case BPF_FUNC_for_each_map_elem:
7243 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
7244 set_map_elem_callback_state);
7246 case BPF_FUNC_timer_set_callback:
7247 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
7248 set_timer_callback_state);
7250 case BPF_FUNC_find_vma:
7251 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
7252 set_find_vma_callback_state);
7254 case BPF_FUNC_snprintf:
7255 err = check_bpf_snprintf_call(env, regs);
7258 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
7259 set_loop_callback_state);
7261 case BPF_FUNC_dynptr_from_mem:
7262 if (regs[BPF_REG_1].type != PTR_TO_MAP_VALUE) {
7263 verbose(env, "Unsupported reg type %s for bpf_dynptr_from_mem data\n",
7264 reg_type_str(env, regs[BPF_REG_1].type));
7272 /* reset caller saved regs */
7273 for (i = 0; i < CALLER_SAVED_REGS; i++) {
7274 mark_reg_not_init(env, regs, caller_saved[i]);
7275 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
7278 /* helper call returns 64-bit value. */
7279 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
7281 /* update return register (already marked as written above) */
7282 ret_type = fn->ret_type;
7283 ret_flag = type_flag(fn->ret_type);
7284 if (ret_type == RET_INTEGER) {
7285 /* sets type to SCALAR_VALUE */
7286 mark_reg_unknown(env, regs, BPF_REG_0);
7287 } else if (ret_type == RET_VOID) {
7288 regs[BPF_REG_0].type = NOT_INIT;
7289 } else if (base_type(ret_type) == RET_PTR_TO_MAP_VALUE) {
7290 /* There is no offset yet applied, variable or fixed */
7291 mark_reg_known_zero(env, regs, BPF_REG_0);
7292 /* remember map_ptr, so that check_map_access()
7293 * can check 'value_size' boundary of memory access
7294 * to map element returned from bpf_map_lookup_elem()
7296 if (meta.map_ptr == NULL) {
7298 "kernel subsystem misconfigured verifier\n");
7301 regs[BPF_REG_0].map_ptr = meta.map_ptr;
7302 regs[BPF_REG_0].map_uid = meta.map_uid;
7303 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE | ret_flag;
7304 if (!type_may_be_null(ret_type) &&
7305 map_value_has_spin_lock(meta.map_ptr)) {
7306 regs[BPF_REG_0].id = ++env->id_gen;
7308 } else if (base_type(ret_type) == RET_PTR_TO_SOCKET) {
7309 mark_reg_known_zero(env, regs, BPF_REG_0);
7310 regs[BPF_REG_0].type = PTR_TO_SOCKET | ret_flag;
7311 } else if (base_type(ret_type) == RET_PTR_TO_SOCK_COMMON) {
7312 mark_reg_known_zero(env, regs, BPF_REG_0);
7313 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON | ret_flag;
7314 } else if (base_type(ret_type) == RET_PTR_TO_TCP_SOCK) {
7315 mark_reg_known_zero(env, regs, BPF_REG_0);
7316 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK | ret_flag;
7317 } else if (base_type(ret_type) == RET_PTR_TO_ALLOC_MEM) {
7318 mark_reg_known_zero(env, regs, BPF_REG_0);
7319 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
7320 regs[BPF_REG_0].mem_size = meta.mem_size;
7321 } else if (base_type(ret_type) == RET_PTR_TO_MEM_OR_BTF_ID) {
7322 const struct btf_type *t;
7324 mark_reg_known_zero(env, regs, BPF_REG_0);
7325 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
7326 if (!btf_type_is_struct(t)) {
7328 const struct btf_type *ret;
7331 /* resolve the type size of ksym. */
7332 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
7334 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
7335 verbose(env, "unable to resolve the size of type '%s': %ld\n",
7336 tname, PTR_ERR(ret));
7339 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
7340 regs[BPF_REG_0].mem_size = tsize;
7342 /* MEM_RDONLY may be carried from ret_flag, but it
7343 * doesn't apply on PTR_TO_BTF_ID. Fold it, otherwise
7344 * it will confuse the check of PTR_TO_BTF_ID in
7345 * check_mem_access().
7347 ret_flag &= ~MEM_RDONLY;
7349 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
7350 regs[BPF_REG_0].btf = meta.ret_btf;
7351 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
7353 } else if (base_type(ret_type) == RET_PTR_TO_BTF_ID) {
7354 struct btf *ret_btf;
7357 mark_reg_known_zero(env, regs, BPF_REG_0);
7358 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
7359 if (func_id == BPF_FUNC_kptr_xchg) {
7360 ret_btf = meta.kptr_off_desc->kptr.btf;
7361 ret_btf_id = meta.kptr_off_desc->kptr.btf_id;
7363 ret_btf = btf_vmlinux;
7364 ret_btf_id = *fn->ret_btf_id;
7366 if (ret_btf_id == 0) {
7367 verbose(env, "invalid return type %u of func %s#%d\n",
7368 base_type(ret_type), func_id_name(func_id),
7372 regs[BPF_REG_0].btf = ret_btf;
7373 regs[BPF_REG_0].btf_id = ret_btf_id;
7375 verbose(env, "unknown return type %u of func %s#%d\n",
7376 base_type(ret_type), func_id_name(func_id), func_id);
7380 if (type_may_be_null(regs[BPF_REG_0].type))
7381 regs[BPF_REG_0].id = ++env->id_gen;
7383 if (is_ptr_cast_function(func_id)) {
7384 /* For release_reference() */
7385 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
7386 } else if (is_acquire_function(func_id, meta.map_ptr)) {
7387 int id = acquire_reference_state(env, insn_idx);
7391 /* For mark_ptr_or_null_reg() */
7392 regs[BPF_REG_0].id = id;
7393 /* For release_reference() */
7394 regs[BPF_REG_0].ref_obj_id = id;
7395 } else if (func_id == BPF_FUNC_dynptr_data) {
7396 int dynptr_id = 0, i;
7398 /* Find the id of the dynptr we're acquiring a reference to */
7399 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
7400 if (arg_type_is_dynptr(fn->arg_type[i])) {
7402 verbose(env, "verifier internal error: multiple dynptr args in func\n");
7405 dynptr_id = stack_slot_get_id(env, ®s[BPF_REG_1 + i]);
7408 /* For release_reference() */
7409 regs[BPF_REG_0].ref_obj_id = dynptr_id;
7412 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
7414 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
7418 if ((func_id == BPF_FUNC_get_stack ||
7419 func_id == BPF_FUNC_get_task_stack) &&
7420 !env->prog->has_callchain_buf) {
7421 const char *err_str;
7423 #ifdef CONFIG_PERF_EVENTS
7424 err = get_callchain_buffers(sysctl_perf_event_max_stack);
7425 err_str = "cannot get callchain buffer for func %s#%d\n";
7428 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
7431 verbose(env, err_str, func_id_name(func_id), func_id);
7435 env->prog->has_callchain_buf = true;
7438 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
7439 env->prog->call_get_stack = true;
7441 if (func_id == BPF_FUNC_get_func_ip) {
7442 if (check_get_func_ip(env))
7444 env->prog->call_get_func_ip = true;
7448 clear_all_pkt_pointers(env);
7452 /* mark_btf_func_reg_size() is used when the reg size is determined by
7453 * the BTF func_proto's return value size and argument.
7455 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
7458 struct bpf_reg_state *reg = &cur_regs(env)[regno];
7460 if (regno == BPF_REG_0) {
7461 /* Function return value */
7462 reg->live |= REG_LIVE_WRITTEN;
7463 reg->subreg_def = reg_size == sizeof(u64) ?
7464 DEF_NOT_SUBREG : env->insn_idx + 1;
7466 /* Function argument */
7467 if (reg_size == sizeof(u64)) {
7468 mark_insn_zext(env, reg);
7469 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
7471 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
7476 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
7479 const struct btf_type *t, *func, *func_proto, *ptr_type;
7480 struct bpf_reg_state *regs = cur_regs(env);
7481 const char *func_name, *ptr_type_name;
7482 u32 i, nargs, func_id, ptr_type_id;
7483 int err, insn_idx = *insn_idx_p;
7484 const struct btf_param *args;
7485 struct btf *desc_btf;
7488 /* skip for now, but return error when we find this in fixup_kfunc_call */
7492 desc_btf = find_kfunc_desc_btf(env, insn->off);
7493 if (IS_ERR(desc_btf))
7494 return PTR_ERR(desc_btf);
7496 func_id = insn->imm;
7497 func = btf_type_by_id(desc_btf, func_id);
7498 func_name = btf_name_by_offset(desc_btf, func->name_off);
7499 func_proto = btf_type_by_id(desc_btf, func->type);
7501 if (!btf_kfunc_id_set_contains(desc_btf, resolve_prog_type(env->prog),
7502 BTF_KFUNC_TYPE_CHECK, func_id)) {
7503 verbose(env, "calling kernel function %s is not allowed\n",
7508 acq = btf_kfunc_id_set_contains(desc_btf, resolve_prog_type(env->prog),
7509 BTF_KFUNC_TYPE_ACQUIRE, func_id);
7511 /* Check the arguments */
7512 err = btf_check_kfunc_arg_match(env, desc_btf, func_id, regs);
7515 /* In case of release function, we get register number of refcounted
7516 * PTR_TO_BTF_ID back from btf_check_kfunc_arg_match, do the release now
7519 err = release_reference(env, regs[err].ref_obj_id);
7521 verbose(env, "kfunc %s#%d reference has not been acquired before\n",
7522 func_name, func_id);
7527 for (i = 0; i < CALLER_SAVED_REGS; i++)
7528 mark_reg_not_init(env, regs, caller_saved[i]);
7530 /* Check return type */
7531 t = btf_type_skip_modifiers(desc_btf, func_proto->type, NULL);
7533 if (acq && !btf_type_is_ptr(t)) {
7534 verbose(env, "acquire kernel function does not return PTR_TO_BTF_ID\n");
7538 if (btf_type_is_scalar(t)) {
7539 mark_reg_unknown(env, regs, BPF_REG_0);
7540 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
7541 } else if (btf_type_is_ptr(t)) {
7542 ptr_type = btf_type_skip_modifiers(desc_btf, t->type,
7544 if (!btf_type_is_struct(ptr_type)) {
7545 ptr_type_name = btf_name_by_offset(desc_btf,
7546 ptr_type->name_off);
7547 verbose(env, "kernel function %s returns pointer type %s %s is not supported\n",
7548 func_name, btf_type_str(ptr_type),
7552 mark_reg_known_zero(env, regs, BPF_REG_0);
7553 regs[BPF_REG_0].btf = desc_btf;
7554 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
7555 regs[BPF_REG_0].btf_id = ptr_type_id;
7556 if (btf_kfunc_id_set_contains(desc_btf, resolve_prog_type(env->prog),
7557 BTF_KFUNC_TYPE_RET_NULL, func_id)) {
7558 regs[BPF_REG_0].type |= PTR_MAYBE_NULL;
7559 /* For mark_ptr_or_null_reg, see 93c230e3f5bd6 */
7560 regs[BPF_REG_0].id = ++env->id_gen;
7562 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
7564 int id = acquire_reference_state(env, insn_idx);
7568 regs[BPF_REG_0].id = id;
7569 regs[BPF_REG_0].ref_obj_id = id;
7571 } /* else { add_kfunc_call() ensures it is btf_type_is_void(t) } */
7573 nargs = btf_type_vlen(func_proto);
7574 args = (const struct btf_param *)(func_proto + 1);
7575 for (i = 0; i < nargs; i++) {
7578 t = btf_type_skip_modifiers(desc_btf, args[i].type, NULL);
7579 if (btf_type_is_ptr(t))
7580 mark_btf_func_reg_size(env, regno, sizeof(void *));
7582 /* scalar. ensured by btf_check_kfunc_arg_match() */
7583 mark_btf_func_reg_size(env, regno, t->size);
7589 static bool signed_add_overflows(s64 a, s64 b)
7591 /* Do the add in u64, where overflow is well-defined */
7592 s64 res = (s64)((u64)a + (u64)b);
7599 static bool signed_add32_overflows(s32 a, s32 b)
7601 /* Do the add in u32, where overflow is well-defined */
7602 s32 res = (s32)((u32)a + (u32)b);
7609 static bool signed_sub_overflows(s64 a, s64 b)
7611 /* Do the sub in u64, where overflow is well-defined */
7612 s64 res = (s64)((u64)a - (u64)b);
7619 static bool signed_sub32_overflows(s32 a, s32 b)
7621 /* Do the sub in u32, where overflow is well-defined */
7622 s32 res = (s32)((u32)a - (u32)b);
7629 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
7630 const struct bpf_reg_state *reg,
7631 enum bpf_reg_type type)
7633 bool known = tnum_is_const(reg->var_off);
7634 s64 val = reg->var_off.value;
7635 s64 smin = reg->smin_value;
7637 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
7638 verbose(env, "math between %s pointer and %lld is not allowed\n",
7639 reg_type_str(env, type), val);
7643 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
7644 verbose(env, "%s pointer offset %d is not allowed\n",
7645 reg_type_str(env, type), reg->off);
7649 if (smin == S64_MIN) {
7650 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
7651 reg_type_str(env, type));
7655 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
7656 verbose(env, "value %lld makes %s pointer be out of bounds\n",
7657 smin, reg_type_str(env, type));
7664 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
7666 return &env->insn_aux_data[env->insn_idx];
7677 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
7678 u32 *alu_limit, bool mask_to_left)
7680 u32 max = 0, ptr_limit = 0;
7682 switch (ptr_reg->type) {
7684 /* Offset 0 is out-of-bounds, but acceptable start for the
7685 * left direction, see BPF_REG_FP. Also, unknown scalar
7686 * offset where we would need to deal with min/max bounds is
7687 * currently prohibited for unprivileged.
7689 max = MAX_BPF_STACK + mask_to_left;
7690 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
7692 case PTR_TO_MAP_VALUE:
7693 max = ptr_reg->map_ptr->value_size;
7694 ptr_limit = (mask_to_left ?
7695 ptr_reg->smin_value :
7696 ptr_reg->umax_value) + ptr_reg->off;
7702 if (ptr_limit >= max)
7703 return REASON_LIMIT;
7704 *alu_limit = ptr_limit;
7708 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
7709 const struct bpf_insn *insn)
7711 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
7714 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
7715 u32 alu_state, u32 alu_limit)
7717 /* If we arrived here from different branches with different
7718 * state or limits to sanitize, then this won't work.
7720 if (aux->alu_state &&
7721 (aux->alu_state != alu_state ||
7722 aux->alu_limit != alu_limit))
7723 return REASON_PATHS;
7725 /* Corresponding fixup done in do_misc_fixups(). */
7726 aux->alu_state = alu_state;
7727 aux->alu_limit = alu_limit;
7731 static int sanitize_val_alu(struct bpf_verifier_env *env,
7732 struct bpf_insn *insn)
7734 struct bpf_insn_aux_data *aux = cur_aux(env);
7736 if (can_skip_alu_sanitation(env, insn))
7739 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
7742 static bool sanitize_needed(u8 opcode)
7744 return opcode == BPF_ADD || opcode == BPF_SUB;
7747 struct bpf_sanitize_info {
7748 struct bpf_insn_aux_data aux;
7752 static struct bpf_verifier_state *
7753 sanitize_speculative_path(struct bpf_verifier_env *env,
7754 const struct bpf_insn *insn,
7755 u32 next_idx, u32 curr_idx)
7757 struct bpf_verifier_state *branch;
7758 struct bpf_reg_state *regs;
7760 branch = push_stack(env, next_idx, curr_idx, true);
7761 if (branch && insn) {
7762 regs = branch->frame[branch->curframe]->regs;
7763 if (BPF_SRC(insn->code) == BPF_K) {
7764 mark_reg_unknown(env, regs, insn->dst_reg);
7765 } else if (BPF_SRC(insn->code) == BPF_X) {
7766 mark_reg_unknown(env, regs, insn->dst_reg);
7767 mark_reg_unknown(env, regs, insn->src_reg);
7773 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
7774 struct bpf_insn *insn,
7775 const struct bpf_reg_state *ptr_reg,
7776 const struct bpf_reg_state *off_reg,
7777 struct bpf_reg_state *dst_reg,
7778 struct bpf_sanitize_info *info,
7779 const bool commit_window)
7781 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
7782 struct bpf_verifier_state *vstate = env->cur_state;
7783 bool off_is_imm = tnum_is_const(off_reg->var_off);
7784 bool off_is_neg = off_reg->smin_value < 0;
7785 bool ptr_is_dst_reg = ptr_reg == dst_reg;
7786 u8 opcode = BPF_OP(insn->code);
7787 u32 alu_state, alu_limit;
7788 struct bpf_reg_state tmp;
7792 if (can_skip_alu_sanitation(env, insn))
7795 /* We already marked aux for masking from non-speculative
7796 * paths, thus we got here in the first place. We only care
7797 * to explore bad access from here.
7799 if (vstate->speculative)
7802 if (!commit_window) {
7803 if (!tnum_is_const(off_reg->var_off) &&
7804 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
7805 return REASON_BOUNDS;
7807 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
7808 (opcode == BPF_SUB && !off_is_neg);
7811 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
7815 if (commit_window) {
7816 /* In commit phase we narrow the masking window based on
7817 * the observed pointer move after the simulated operation.
7819 alu_state = info->aux.alu_state;
7820 alu_limit = abs(info->aux.alu_limit - alu_limit);
7822 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
7823 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
7824 alu_state |= ptr_is_dst_reg ?
7825 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
7827 /* Limit pruning on unknown scalars to enable deep search for
7828 * potential masking differences from other program paths.
7831 env->explore_alu_limits = true;
7834 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
7838 /* If we're in commit phase, we're done here given we already
7839 * pushed the truncated dst_reg into the speculative verification
7842 * Also, when register is a known constant, we rewrite register-based
7843 * operation to immediate-based, and thus do not need masking (and as
7844 * a consequence, do not need to simulate the zero-truncation either).
7846 if (commit_window || off_is_imm)
7849 /* Simulate and find potential out-of-bounds access under
7850 * speculative execution from truncation as a result of
7851 * masking when off was not within expected range. If off
7852 * sits in dst, then we temporarily need to move ptr there
7853 * to simulate dst (== 0) +/-= ptr. Needed, for example,
7854 * for cases where we use K-based arithmetic in one direction
7855 * and truncated reg-based in the other in order to explore
7858 if (!ptr_is_dst_reg) {
7860 *dst_reg = *ptr_reg;
7862 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
7864 if (!ptr_is_dst_reg && ret)
7866 return !ret ? REASON_STACK : 0;
7869 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
7871 struct bpf_verifier_state *vstate = env->cur_state;
7873 /* If we simulate paths under speculation, we don't update the
7874 * insn as 'seen' such that when we verify unreachable paths in
7875 * the non-speculative domain, sanitize_dead_code() can still
7876 * rewrite/sanitize them.
7878 if (!vstate->speculative)
7879 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
7882 static int sanitize_err(struct bpf_verifier_env *env,
7883 const struct bpf_insn *insn, int reason,
7884 const struct bpf_reg_state *off_reg,
7885 const struct bpf_reg_state *dst_reg)
7887 static const char *err = "pointer arithmetic with it prohibited for !root";
7888 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
7889 u32 dst = insn->dst_reg, src = insn->src_reg;
7893 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
7894 off_reg == dst_reg ? dst : src, err);
7897 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
7898 off_reg == dst_reg ? src : dst, err);
7901 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
7905 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
7909 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
7913 verbose(env, "verifier internal error: unknown reason (%d)\n",
7921 /* check that stack access falls within stack limits and that 'reg' doesn't
7922 * have a variable offset.
7924 * Variable offset is prohibited for unprivileged mode for simplicity since it
7925 * requires corresponding support in Spectre masking for stack ALU. See also
7926 * retrieve_ptr_limit().
7929 * 'off' includes 'reg->off'.
7931 static int check_stack_access_for_ptr_arithmetic(
7932 struct bpf_verifier_env *env,
7934 const struct bpf_reg_state *reg,
7937 if (!tnum_is_const(reg->var_off)) {
7940 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
7941 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
7942 regno, tn_buf, off);
7946 if (off >= 0 || off < -MAX_BPF_STACK) {
7947 verbose(env, "R%d stack pointer arithmetic goes out of range, "
7948 "prohibited for !root; off=%d\n", regno, off);
7955 static int sanitize_check_bounds(struct bpf_verifier_env *env,
7956 const struct bpf_insn *insn,
7957 const struct bpf_reg_state *dst_reg)
7959 u32 dst = insn->dst_reg;
7961 /* For unprivileged we require that resulting offset must be in bounds
7962 * in order to be able to sanitize access later on.
7964 if (env->bypass_spec_v1)
7967 switch (dst_reg->type) {
7969 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
7970 dst_reg->off + dst_reg->var_off.value))
7973 case PTR_TO_MAP_VALUE:
7974 if (check_map_access(env, dst, dst_reg->off, 1, false, ACCESS_HELPER)) {
7975 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
7976 "prohibited for !root\n", dst);
7987 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
7988 * Caller should also handle BPF_MOV case separately.
7989 * If we return -EACCES, caller may want to try again treating pointer as a
7990 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
7992 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
7993 struct bpf_insn *insn,
7994 const struct bpf_reg_state *ptr_reg,
7995 const struct bpf_reg_state *off_reg)
7997 struct bpf_verifier_state *vstate = env->cur_state;
7998 struct bpf_func_state *state = vstate->frame[vstate->curframe];
7999 struct bpf_reg_state *regs = state->regs, *dst_reg;
8000 bool known = tnum_is_const(off_reg->var_off);
8001 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
8002 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
8003 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
8004 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
8005 struct bpf_sanitize_info info = {};
8006 u8 opcode = BPF_OP(insn->code);
8007 u32 dst = insn->dst_reg;
8010 dst_reg = ®s[dst];
8012 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
8013 smin_val > smax_val || umin_val > umax_val) {
8014 /* Taint dst register if offset had invalid bounds derived from
8015 * e.g. dead branches.
8017 __mark_reg_unknown(env, dst_reg);
8021 if (BPF_CLASS(insn->code) != BPF_ALU64) {
8022 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
8023 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
8024 __mark_reg_unknown(env, dst_reg);
8029 "R%d 32-bit pointer arithmetic prohibited\n",
8034 if (ptr_reg->type & PTR_MAYBE_NULL) {
8035 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
8036 dst, reg_type_str(env, ptr_reg->type));
8040 switch (base_type(ptr_reg->type)) {
8041 case CONST_PTR_TO_MAP:
8042 /* smin_val represents the known value */
8043 if (known && smin_val == 0 && opcode == BPF_ADD)
8046 case PTR_TO_PACKET_END:
8048 case PTR_TO_SOCK_COMMON:
8049 case PTR_TO_TCP_SOCK:
8050 case PTR_TO_XDP_SOCK:
8051 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
8052 dst, reg_type_str(env, ptr_reg->type));
8058 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
8059 * The id may be overwritten later if we create a new variable offset.
8061 dst_reg->type = ptr_reg->type;
8062 dst_reg->id = ptr_reg->id;
8064 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
8065 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
8068 /* pointer types do not carry 32-bit bounds at the moment. */
8069 __mark_reg32_unbounded(dst_reg);
8071 if (sanitize_needed(opcode)) {
8072 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
8075 return sanitize_err(env, insn, ret, off_reg, dst_reg);
8080 /* We can take a fixed offset as long as it doesn't overflow
8081 * the s32 'off' field
8083 if (known && (ptr_reg->off + smin_val ==
8084 (s64)(s32)(ptr_reg->off + smin_val))) {
8085 /* pointer += K. Accumulate it into fixed offset */
8086 dst_reg->smin_value = smin_ptr;
8087 dst_reg->smax_value = smax_ptr;
8088 dst_reg->umin_value = umin_ptr;
8089 dst_reg->umax_value = umax_ptr;
8090 dst_reg->var_off = ptr_reg->var_off;
8091 dst_reg->off = ptr_reg->off + smin_val;
8092 dst_reg->raw = ptr_reg->raw;
8095 /* A new variable offset is created. Note that off_reg->off
8096 * == 0, since it's a scalar.
8097 * dst_reg gets the pointer type and since some positive
8098 * integer value was added to the pointer, give it a new 'id'
8099 * if it's a PTR_TO_PACKET.
8100 * this creates a new 'base' pointer, off_reg (variable) gets
8101 * added into the variable offset, and we copy the fixed offset
8104 if (signed_add_overflows(smin_ptr, smin_val) ||
8105 signed_add_overflows(smax_ptr, smax_val)) {
8106 dst_reg->smin_value = S64_MIN;
8107 dst_reg->smax_value = S64_MAX;
8109 dst_reg->smin_value = smin_ptr + smin_val;
8110 dst_reg->smax_value = smax_ptr + smax_val;
8112 if (umin_ptr + umin_val < umin_ptr ||
8113 umax_ptr + umax_val < umax_ptr) {
8114 dst_reg->umin_value = 0;
8115 dst_reg->umax_value = U64_MAX;
8117 dst_reg->umin_value = umin_ptr + umin_val;
8118 dst_reg->umax_value = umax_ptr + umax_val;
8120 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
8121 dst_reg->off = ptr_reg->off;
8122 dst_reg->raw = ptr_reg->raw;
8123 if (reg_is_pkt_pointer(ptr_reg)) {
8124 dst_reg->id = ++env->id_gen;
8125 /* something was added to pkt_ptr, set range to zero */
8126 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
8130 if (dst_reg == off_reg) {
8131 /* scalar -= pointer. Creates an unknown scalar */
8132 verbose(env, "R%d tried to subtract pointer from scalar\n",
8136 /* We don't allow subtraction from FP, because (according to
8137 * test_verifier.c test "invalid fp arithmetic", JITs might not
8138 * be able to deal with it.
8140 if (ptr_reg->type == PTR_TO_STACK) {
8141 verbose(env, "R%d subtraction from stack pointer prohibited\n",
8145 if (known && (ptr_reg->off - smin_val ==
8146 (s64)(s32)(ptr_reg->off - smin_val))) {
8147 /* pointer -= K. Subtract it from fixed offset */
8148 dst_reg->smin_value = smin_ptr;
8149 dst_reg->smax_value = smax_ptr;
8150 dst_reg->umin_value = umin_ptr;
8151 dst_reg->umax_value = umax_ptr;
8152 dst_reg->var_off = ptr_reg->var_off;
8153 dst_reg->id = ptr_reg->id;
8154 dst_reg->off = ptr_reg->off - smin_val;
8155 dst_reg->raw = ptr_reg->raw;
8158 /* A new variable offset is created. If the subtrahend is known
8159 * nonnegative, then any reg->range we had before is still good.
8161 if (signed_sub_overflows(smin_ptr, smax_val) ||
8162 signed_sub_overflows(smax_ptr, smin_val)) {
8163 /* Overflow possible, we know nothing */
8164 dst_reg->smin_value = S64_MIN;
8165 dst_reg->smax_value = S64_MAX;
8167 dst_reg->smin_value = smin_ptr - smax_val;
8168 dst_reg->smax_value = smax_ptr - smin_val;
8170 if (umin_ptr < umax_val) {
8171 /* Overflow possible, we know nothing */
8172 dst_reg->umin_value = 0;
8173 dst_reg->umax_value = U64_MAX;
8175 /* Cannot overflow (as long as bounds are consistent) */
8176 dst_reg->umin_value = umin_ptr - umax_val;
8177 dst_reg->umax_value = umax_ptr - umin_val;
8179 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
8180 dst_reg->off = ptr_reg->off;
8181 dst_reg->raw = ptr_reg->raw;
8182 if (reg_is_pkt_pointer(ptr_reg)) {
8183 dst_reg->id = ++env->id_gen;
8184 /* something was added to pkt_ptr, set range to zero */
8186 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
8192 /* bitwise ops on pointers are troublesome, prohibit. */
8193 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
8194 dst, bpf_alu_string[opcode >> 4]);
8197 /* other operators (e.g. MUL,LSH) produce non-pointer results */
8198 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
8199 dst, bpf_alu_string[opcode >> 4]);
8203 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
8206 __update_reg_bounds(dst_reg);
8207 __reg_deduce_bounds(dst_reg);
8208 __reg_bound_offset(dst_reg);
8210 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
8212 if (sanitize_needed(opcode)) {
8213 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
8216 return sanitize_err(env, insn, ret, off_reg, dst_reg);
8222 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
8223 struct bpf_reg_state *src_reg)
8225 s32 smin_val = src_reg->s32_min_value;
8226 s32 smax_val = src_reg->s32_max_value;
8227 u32 umin_val = src_reg->u32_min_value;
8228 u32 umax_val = src_reg->u32_max_value;
8230 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
8231 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
8232 dst_reg->s32_min_value = S32_MIN;
8233 dst_reg->s32_max_value = S32_MAX;
8235 dst_reg->s32_min_value += smin_val;
8236 dst_reg->s32_max_value += smax_val;
8238 if (dst_reg->u32_min_value + umin_val < umin_val ||
8239 dst_reg->u32_max_value + umax_val < umax_val) {
8240 dst_reg->u32_min_value = 0;
8241 dst_reg->u32_max_value = U32_MAX;
8243 dst_reg->u32_min_value += umin_val;
8244 dst_reg->u32_max_value += umax_val;
8248 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
8249 struct bpf_reg_state *src_reg)
8251 s64 smin_val = src_reg->smin_value;
8252 s64 smax_val = src_reg->smax_value;
8253 u64 umin_val = src_reg->umin_value;
8254 u64 umax_val = src_reg->umax_value;
8256 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
8257 signed_add_overflows(dst_reg->smax_value, smax_val)) {
8258 dst_reg->smin_value = S64_MIN;
8259 dst_reg->smax_value = S64_MAX;
8261 dst_reg->smin_value += smin_val;
8262 dst_reg->smax_value += smax_val;
8264 if (dst_reg->umin_value + umin_val < umin_val ||
8265 dst_reg->umax_value + umax_val < umax_val) {
8266 dst_reg->umin_value = 0;
8267 dst_reg->umax_value = U64_MAX;
8269 dst_reg->umin_value += umin_val;
8270 dst_reg->umax_value += umax_val;
8274 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
8275 struct bpf_reg_state *src_reg)
8277 s32 smin_val = src_reg->s32_min_value;
8278 s32 smax_val = src_reg->s32_max_value;
8279 u32 umin_val = src_reg->u32_min_value;
8280 u32 umax_val = src_reg->u32_max_value;
8282 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
8283 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
8284 /* Overflow possible, we know nothing */
8285 dst_reg->s32_min_value = S32_MIN;
8286 dst_reg->s32_max_value = S32_MAX;
8288 dst_reg->s32_min_value -= smax_val;
8289 dst_reg->s32_max_value -= smin_val;
8291 if (dst_reg->u32_min_value < umax_val) {
8292 /* Overflow possible, we know nothing */
8293 dst_reg->u32_min_value = 0;
8294 dst_reg->u32_max_value = U32_MAX;
8296 /* Cannot overflow (as long as bounds are consistent) */
8297 dst_reg->u32_min_value -= umax_val;
8298 dst_reg->u32_max_value -= umin_val;
8302 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
8303 struct bpf_reg_state *src_reg)
8305 s64 smin_val = src_reg->smin_value;
8306 s64 smax_val = src_reg->smax_value;
8307 u64 umin_val = src_reg->umin_value;
8308 u64 umax_val = src_reg->umax_value;
8310 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
8311 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
8312 /* Overflow possible, we know nothing */
8313 dst_reg->smin_value = S64_MIN;
8314 dst_reg->smax_value = S64_MAX;
8316 dst_reg->smin_value -= smax_val;
8317 dst_reg->smax_value -= smin_val;
8319 if (dst_reg->umin_value < umax_val) {
8320 /* Overflow possible, we know nothing */
8321 dst_reg->umin_value = 0;
8322 dst_reg->umax_value = U64_MAX;
8324 /* Cannot overflow (as long as bounds are consistent) */
8325 dst_reg->umin_value -= umax_val;
8326 dst_reg->umax_value -= umin_val;
8330 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
8331 struct bpf_reg_state *src_reg)
8333 s32 smin_val = src_reg->s32_min_value;
8334 u32 umin_val = src_reg->u32_min_value;
8335 u32 umax_val = src_reg->u32_max_value;
8337 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
8338 /* Ain't nobody got time to multiply that sign */
8339 __mark_reg32_unbounded(dst_reg);
8342 /* Both values are positive, so we can work with unsigned and
8343 * copy the result to signed (unless it exceeds S32_MAX).
8345 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
8346 /* Potential overflow, we know nothing */
8347 __mark_reg32_unbounded(dst_reg);
8350 dst_reg->u32_min_value *= umin_val;
8351 dst_reg->u32_max_value *= umax_val;
8352 if (dst_reg->u32_max_value > S32_MAX) {
8353 /* Overflow possible, we know nothing */
8354 dst_reg->s32_min_value = S32_MIN;
8355 dst_reg->s32_max_value = S32_MAX;
8357 dst_reg->s32_min_value = dst_reg->u32_min_value;
8358 dst_reg->s32_max_value = dst_reg->u32_max_value;
8362 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
8363 struct bpf_reg_state *src_reg)
8365 s64 smin_val = src_reg->smin_value;
8366 u64 umin_val = src_reg->umin_value;
8367 u64 umax_val = src_reg->umax_value;
8369 if (smin_val < 0 || dst_reg->smin_value < 0) {
8370 /* Ain't nobody got time to multiply that sign */
8371 __mark_reg64_unbounded(dst_reg);
8374 /* Both values are positive, so we can work with unsigned and
8375 * copy the result to signed (unless it exceeds S64_MAX).
8377 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
8378 /* Potential overflow, we know nothing */
8379 __mark_reg64_unbounded(dst_reg);
8382 dst_reg->umin_value *= umin_val;
8383 dst_reg->umax_value *= umax_val;
8384 if (dst_reg->umax_value > S64_MAX) {
8385 /* Overflow possible, we know nothing */
8386 dst_reg->smin_value = S64_MIN;
8387 dst_reg->smax_value = S64_MAX;
8389 dst_reg->smin_value = dst_reg->umin_value;
8390 dst_reg->smax_value = dst_reg->umax_value;
8394 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
8395 struct bpf_reg_state *src_reg)
8397 bool src_known = tnum_subreg_is_const(src_reg->var_off);
8398 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
8399 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
8400 s32 smin_val = src_reg->s32_min_value;
8401 u32 umax_val = src_reg->u32_max_value;
8403 if (src_known && dst_known) {
8404 __mark_reg32_known(dst_reg, var32_off.value);
8408 /* We get our minimum from the var_off, since that's inherently
8409 * bitwise. Our maximum is the minimum of the operands' maxima.
8411 dst_reg->u32_min_value = var32_off.value;
8412 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
8413 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
8414 /* Lose signed bounds when ANDing negative numbers,
8415 * ain't nobody got time for that.
8417 dst_reg->s32_min_value = S32_MIN;
8418 dst_reg->s32_max_value = S32_MAX;
8420 /* ANDing two positives gives a positive, so safe to
8421 * cast result into s64.
8423 dst_reg->s32_min_value = dst_reg->u32_min_value;
8424 dst_reg->s32_max_value = dst_reg->u32_max_value;
8428 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
8429 struct bpf_reg_state *src_reg)
8431 bool src_known = tnum_is_const(src_reg->var_off);
8432 bool dst_known = tnum_is_const(dst_reg->var_off);
8433 s64 smin_val = src_reg->smin_value;
8434 u64 umax_val = src_reg->umax_value;
8436 if (src_known && dst_known) {
8437 __mark_reg_known(dst_reg, dst_reg->var_off.value);
8441 /* We get our minimum from the var_off, since that's inherently
8442 * bitwise. Our maximum is the minimum of the operands' maxima.
8444 dst_reg->umin_value = dst_reg->var_off.value;
8445 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
8446 if (dst_reg->smin_value < 0 || smin_val < 0) {
8447 /* Lose signed bounds when ANDing negative numbers,
8448 * ain't nobody got time for that.
8450 dst_reg->smin_value = S64_MIN;
8451 dst_reg->smax_value = S64_MAX;
8453 /* ANDing two positives gives a positive, so safe to
8454 * cast result into s64.
8456 dst_reg->smin_value = dst_reg->umin_value;
8457 dst_reg->smax_value = dst_reg->umax_value;
8459 /* We may learn something more from the var_off */
8460 __update_reg_bounds(dst_reg);
8463 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
8464 struct bpf_reg_state *src_reg)
8466 bool src_known = tnum_subreg_is_const(src_reg->var_off);
8467 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
8468 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
8469 s32 smin_val = src_reg->s32_min_value;
8470 u32 umin_val = src_reg->u32_min_value;
8472 if (src_known && dst_known) {
8473 __mark_reg32_known(dst_reg, var32_off.value);
8477 /* We get our maximum from the var_off, and our minimum is the
8478 * maximum of the operands' minima
8480 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
8481 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
8482 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
8483 /* Lose signed bounds when ORing negative numbers,
8484 * ain't nobody got time for that.
8486 dst_reg->s32_min_value = S32_MIN;
8487 dst_reg->s32_max_value = S32_MAX;
8489 /* ORing two positives gives a positive, so safe to
8490 * cast result into s64.
8492 dst_reg->s32_min_value = dst_reg->u32_min_value;
8493 dst_reg->s32_max_value = dst_reg->u32_max_value;
8497 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
8498 struct bpf_reg_state *src_reg)
8500 bool src_known = tnum_is_const(src_reg->var_off);
8501 bool dst_known = tnum_is_const(dst_reg->var_off);
8502 s64 smin_val = src_reg->smin_value;
8503 u64 umin_val = src_reg->umin_value;
8505 if (src_known && dst_known) {
8506 __mark_reg_known(dst_reg, dst_reg->var_off.value);
8510 /* We get our maximum from the var_off, and our minimum is the
8511 * maximum of the operands' minima
8513 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
8514 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
8515 if (dst_reg->smin_value < 0 || smin_val < 0) {
8516 /* Lose signed bounds when ORing negative numbers,
8517 * ain't nobody got time for that.
8519 dst_reg->smin_value = S64_MIN;
8520 dst_reg->smax_value = S64_MAX;
8522 /* ORing two positives gives a positive, so safe to
8523 * cast result into s64.
8525 dst_reg->smin_value = dst_reg->umin_value;
8526 dst_reg->smax_value = dst_reg->umax_value;
8528 /* We may learn something more from the var_off */
8529 __update_reg_bounds(dst_reg);
8532 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
8533 struct bpf_reg_state *src_reg)
8535 bool src_known = tnum_subreg_is_const(src_reg->var_off);
8536 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
8537 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
8538 s32 smin_val = src_reg->s32_min_value;
8540 if (src_known && dst_known) {
8541 __mark_reg32_known(dst_reg, var32_off.value);
8545 /* We get both minimum and maximum from the var32_off. */
8546 dst_reg->u32_min_value = var32_off.value;
8547 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
8549 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
8550 /* XORing two positive sign numbers gives a positive,
8551 * so safe to cast u32 result into s32.
8553 dst_reg->s32_min_value = dst_reg->u32_min_value;
8554 dst_reg->s32_max_value = dst_reg->u32_max_value;
8556 dst_reg->s32_min_value = S32_MIN;
8557 dst_reg->s32_max_value = S32_MAX;
8561 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
8562 struct bpf_reg_state *src_reg)
8564 bool src_known = tnum_is_const(src_reg->var_off);
8565 bool dst_known = tnum_is_const(dst_reg->var_off);
8566 s64 smin_val = src_reg->smin_value;
8568 if (src_known && dst_known) {
8569 /* dst_reg->var_off.value has been updated earlier */
8570 __mark_reg_known(dst_reg, dst_reg->var_off.value);
8574 /* We get both minimum and maximum from the var_off. */
8575 dst_reg->umin_value = dst_reg->var_off.value;
8576 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
8578 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
8579 /* XORing two positive sign numbers gives a positive,
8580 * so safe to cast u64 result into s64.
8582 dst_reg->smin_value = dst_reg->umin_value;
8583 dst_reg->smax_value = dst_reg->umax_value;
8585 dst_reg->smin_value = S64_MIN;
8586 dst_reg->smax_value = S64_MAX;
8589 __update_reg_bounds(dst_reg);
8592 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
8593 u64 umin_val, u64 umax_val)
8595 /* We lose all sign bit information (except what we can pick
8598 dst_reg->s32_min_value = S32_MIN;
8599 dst_reg->s32_max_value = S32_MAX;
8600 /* If we might shift our top bit out, then we know nothing */
8601 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
8602 dst_reg->u32_min_value = 0;
8603 dst_reg->u32_max_value = U32_MAX;
8605 dst_reg->u32_min_value <<= umin_val;
8606 dst_reg->u32_max_value <<= umax_val;
8610 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
8611 struct bpf_reg_state *src_reg)
8613 u32 umax_val = src_reg->u32_max_value;
8614 u32 umin_val = src_reg->u32_min_value;
8615 /* u32 alu operation will zext upper bits */
8616 struct tnum subreg = tnum_subreg(dst_reg->var_off);
8618 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
8619 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
8620 /* Not required but being careful mark reg64 bounds as unknown so
8621 * that we are forced to pick them up from tnum and zext later and
8622 * if some path skips this step we are still safe.
8624 __mark_reg64_unbounded(dst_reg);
8625 __update_reg32_bounds(dst_reg);
8628 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
8629 u64 umin_val, u64 umax_val)
8631 /* Special case <<32 because it is a common compiler pattern to sign
8632 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
8633 * positive we know this shift will also be positive so we can track
8634 * bounds correctly. Otherwise we lose all sign bit information except
8635 * what we can pick up from var_off. Perhaps we can generalize this
8636 * later to shifts of any length.
8638 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
8639 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
8641 dst_reg->smax_value = S64_MAX;
8643 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
8644 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
8646 dst_reg->smin_value = S64_MIN;
8648 /* If we might shift our top bit out, then we know nothing */
8649 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
8650 dst_reg->umin_value = 0;
8651 dst_reg->umax_value = U64_MAX;
8653 dst_reg->umin_value <<= umin_val;
8654 dst_reg->umax_value <<= umax_val;
8658 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
8659 struct bpf_reg_state *src_reg)
8661 u64 umax_val = src_reg->umax_value;
8662 u64 umin_val = src_reg->umin_value;
8664 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
8665 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
8666 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
8668 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
8669 /* We may learn something more from the var_off */
8670 __update_reg_bounds(dst_reg);
8673 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
8674 struct bpf_reg_state *src_reg)
8676 struct tnum subreg = tnum_subreg(dst_reg->var_off);
8677 u32 umax_val = src_reg->u32_max_value;
8678 u32 umin_val = src_reg->u32_min_value;
8680 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
8681 * be negative, then either:
8682 * 1) src_reg might be zero, so the sign bit of the result is
8683 * unknown, so we lose our signed bounds
8684 * 2) it's known negative, thus the unsigned bounds capture the
8686 * 3) the signed bounds cross zero, so they tell us nothing
8688 * If the value in dst_reg is known nonnegative, then again the
8689 * unsigned bounds capture the signed bounds.
8690 * Thus, in all cases it suffices to blow away our signed bounds
8691 * and rely on inferring new ones from the unsigned bounds and
8692 * var_off of the result.
8694 dst_reg->s32_min_value = S32_MIN;
8695 dst_reg->s32_max_value = S32_MAX;
8697 dst_reg->var_off = tnum_rshift(subreg, umin_val);
8698 dst_reg->u32_min_value >>= umax_val;
8699 dst_reg->u32_max_value >>= umin_val;
8701 __mark_reg64_unbounded(dst_reg);
8702 __update_reg32_bounds(dst_reg);
8705 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
8706 struct bpf_reg_state *src_reg)
8708 u64 umax_val = src_reg->umax_value;
8709 u64 umin_val = src_reg->umin_value;
8711 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
8712 * be negative, then either:
8713 * 1) src_reg might be zero, so the sign bit of the result is
8714 * unknown, so we lose our signed bounds
8715 * 2) it's known negative, thus the unsigned bounds capture the
8717 * 3) the signed bounds cross zero, so they tell us nothing
8719 * If the value in dst_reg is known nonnegative, then again the
8720 * unsigned bounds capture the signed bounds.
8721 * Thus, in all cases it suffices to blow away our signed bounds
8722 * and rely on inferring new ones from the unsigned bounds and
8723 * var_off of the result.
8725 dst_reg->smin_value = S64_MIN;
8726 dst_reg->smax_value = S64_MAX;
8727 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
8728 dst_reg->umin_value >>= umax_val;
8729 dst_reg->umax_value >>= umin_val;
8731 /* Its not easy to operate on alu32 bounds here because it depends
8732 * on bits being shifted in. Take easy way out and mark unbounded
8733 * so we can recalculate later from tnum.
8735 __mark_reg32_unbounded(dst_reg);
8736 __update_reg_bounds(dst_reg);
8739 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
8740 struct bpf_reg_state *src_reg)
8742 u64 umin_val = src_reg->u32_min_value;
8744 /* Upon reaching here, src_known is true and
8745 * umax_val is equal to umin_val.
8747 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
8748 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
8750 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
8752 /* blow away the dst_reg umin_value/umax_value and rely on
8753 * dst_reg var_off to refine the result.
8755 dst_reg->u32_min_value = 0;
8756 dst_reg->u32_max_value = U32_MAX;
8758 __mark_reg64_unbounded(dst_reg);
8759 __update_reg32_bounds(dst_reg);
8762 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
8763 struct bpf_reg_state *src_reg)
8765 u64 umin_val = src_reg->umin_value;
8767 /* Upon reaching here, src_known is true and umax_val is equal
8770 dst_reg->smin_value >>= umin_val;
8771 dst_reg->smax_value >>= umin_val;
8773 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
8775 /* blow away the dst_reg umin_value/umax_value and rely on
8776 * dst_reg var_off to refine the result.
8778 dst_reg->umin_value = 0;
8779 dst_reg->umax_value = U64_MAX;
8781 /* Its not easy to operate on alu32 bounds here because it depends
8782 * on bits being shifted in from upper 32-bits. Take easy way out
8783 * and mark unbounded so we can recalculate later from tnum.
8785 __mark_reg32_unbounded(dst_reg);
8786 __update_reg_bounds(dst_reg);
8789 /* WARNING: This function does calculations on 64-bit values, but the actual
8790 * execution may occur on 32-bit values. Therefore, things like bitshifts
8791 * need extra checks in the 32-bit case.
8793 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
8794 struct bpf_insn *insn,
8795 struct bpf_reg_state *dst_reg,
8796 struct bpf_reg_state src_reg)
8798 struct bpf_reg_state *regs = cur_regs(env);
8799 u8 opcode = BPF_OP(insn->code);
8801 s64 smin_val, smax_val;
8802 u64 umin_val, umax_val;
8803 s32 s32_min_val, s32_max_val;
8804 u32 u32_min_val, u32_max_val;
8805 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
8806 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
8809 smin_val = src_reg.smin_value;
8810 smax_val = src_reg.smax_value;
8811 umin_val = src_reg.umin_value;
8812 umax_val = src_reg.umax_value;
8814 s32_min_val = src_reg.s32_min_value;
8815 s32_max_val = src_reg.s32_max_value;
8816 u32_min_val = src_reg.u32_min_value;
8817 u32_max_val = src_reg.u32_max_value;
8820 src_known = tnum_subreg_is_const(src_reg.var_off);
8822 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
8823 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
8824 /* Taint dst register if offset had invalid bounds
8825 * derived from e.g. dead branches.
8827 __mark_reg_unknown(env, dst_reg);
8831 src_known = tnum_is_const(src_reg.var_off);
8833 (smin_val != smax_val || umin_val != umax_val)) ||
8834 smin_val > smax_val || umin_val > umax_val) {
8835 /* Taint dst register if offset had invalid bounds
8836 * derived from e.g. dead branches.
8838 __mark_reg_unknown(env, dst_reg);
8844 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
8845 __mark_reg_unknown(env, dst_reg);
8849 if (sanitize_needed(opcode)) {
8850 ret = sanitize_val_alu(env, insn);
8852 return sanitize_err(env, insn, ret, NULL, NULL);
8855 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
8856 * There are two classes of instructions: The first class we track both
8857 * alu32 and alu64 sign/unsigned bounds independently this provides the
8858 * greatest amount of precision when alu operations are mixed with jmp32
8859 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
8860 * and BPF_OR. This is possible because these ops have fairly easy to
8861 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
8862 * See alu32 verifier tests for examples. The second class of
8863 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
8864 * with regards to tracking sign/unsigned bounds because the bits may
8865 * cross subreg boundaries in the alu64 case. When this happens we mark
8866 * the reg unbounded in the subreg bound space and use the resulting
8867 * tnum to calculate an approximation of the sign/unsigned bounds.
8871 scalar32_min_max_add(dst_reg, &src_reg);
8872 scalar_min_max_add(dst_reg, &src_reg);
8873 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
8876 scalar32_min_max_sub(dst_reg, &src_reg);
8877 scalar_min_max_sub(dst_reg, &src_reg);
8878 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
8881 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
8882 scalar32_min_max_mul(dst_reg, &src_reg);
8883 scalar_min_max_mul(dst_reg, &src_reg);
8886 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
8887 scalar32_min_max_and(dst_reg, &src_reg);
8888 scalar_min_max_and(dst_reg, &src_reg);
8891 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
8892 scalar32_min_max_or(dst_reg, &src_reg);
8893 scalar_min_max_or(dst_reg, &src_reg);
8896 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
8897 scalar32_min_max_xor(dst_reg, &src_reg);
8898 scalar_min_max_xor(dst_reg, &src_reg);
8901 if (umax_val >= insn_bitness) {
8902 /* Shifts greater than 31 or 63 are undefined.
8903 * This includes shifts by a negative number.
8905 mark_reg_unknown(env, regs, insn->dst_reg);
8909 scalar32_min_max_lsh(dst_reg, &src_reg);
8911 scalar_min_max_lsh(dst_reg, &src_reg);
8914 if (umax_val >= insn_bitness) {
8915 /* Shifts greater than 31 or 63 are undefined.
8916 * This includes shifts by a negative number.
8918 mark_reg_unknown(env, regs, insn->dst_reg);
8922 scalar32_min_max_rsh(dst_reg, &src_reg);
8924 scalar_min_max_rsh(dst_reg, &src_reg);
8927 if (umax_val >= insn_bitness) {
8928 /* Shifts greater than 31 or 63 are undefined.
8929 * This includes shifts by a negative number.
8931 mark_reg_unknown(env, regs, insn->dst_reg);
8935 scalar32_min_max_arsh(dst_reg, &src_reg);
8937 scalar_min_max_arsh(dst_reg, &src_reg);
8940 mark_reg_unknown(env, regs, insn->dst_reg);
8944 /* ALU32 ops are zero extended into 64bit register */
8946 zext_32_to_64(dst_reg);
8948 __update_reg_bounds(dst_reg);
8949 __reg_deduce_bounds(dst_reg);
8950 __reg_bound_offset(dst_reg);
8954 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
8957 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
8958 struct bpf_insn *insn)
8960 struct bpf_verifier_state *vstate = env->cur_state;
8961 struct bpf_func_state *state = vstate->frame[vstate->curframe];
8962 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
8963 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
8964 u8 opcode = BPF_OP(insn->code);
8967 dst_reg = ®s[insn->dst_reg];
8969 if (dst_reg->type != SCALAR_VALUE)
8972 /* Make sure ID is cleared otherwise dst_reg min/max could be
8973 * incorrectly propagated into other registers by find_equal_scalars()
8976 if (BPF_SRC(insn->code) == BPF_X) {
8977 src_reg = ®s[insn->src_reg];
8978 if (src_reg->type != SCALAR_VALUE) {
8979 if (dst_reg->type != SCALAR_VALUE) {
8980 /* Combining two pointers by any ALU op yields
8981 * an arbitrary scalar. Disallow all math except
8982 * pointer subtraction
8984 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
8985 mark_reg_unknown(env, regs, insn->dst_reg);
8988 verbose(env, "R%d pointer %s pointer prohibited\n",
8990 bpf_alu_string[opcode >> 4]);
8993 /* scalar += pointer
8994 * This is legal, but we have to reverse our
8995 * src/dest handling in computing the range
8997 err = mark_chain_precision(env, insn->dst_reg);
9000 return adjust_ptr_min_max_vals(env, insn,
9003 } else if (ptr_reg) {
9004 /* pointer += scalar */
9005 err = mark_chain_precision(env, insn->src_reg);
9008 return adjust_ptr_min_max_vals(env, insn,
9012 /* Pretend the src is a reg with a known value, since we only
9013 * need to be able to read from this state.
9015 off_reg.type = SCALAR_VALUE;
9016 __mark_reg_known(&off_reg, insn->imm);
9018 if (ptr_reg) /* pointer += K */
9019 return adjust_ptr_min_max_vals(env, insn,
9023 /* Got here implies adding two SCALAR_VALUEs */
9024 if (WARN_ON_ONCE(ptr_reg)) {
9025 print_verifier_state(env, state, true);
9026 verbose(env, "verifier internal error: unexpected ptr_reg\n");
9029 if (WARN_ON(!src_reg)) {
9030 print_verifier_state(env, state, true);
9031 verbose(env, "verifier internal error: no src_reg\n");
9034 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
9037 /* check validity of 32-bit and 64-bit arithmetic operations */
9038 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
9040 struct bpf_reg_state *regs = cur_regs(env);
9041 u8 opcode = BPF_OP(insn->code);
9044 if (opcode == BPF_END || opcode == BPF_NEG) {
9045 if (opcode == BPF_NEG) {
9046 if (BPF_SRC(insn->code) != 0 ||
9047 insn->src_reg != BPF_REG_0 ||
9048 insn->off != 0 || insn->imm != 0) {
9049 verbose(env, "BPF_NEG uses reserved fields\n");
9053 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
9054 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
9055 BPF_CLASS(insn->code) == BPF_ALU64) {
9056 verbose(env, "BPF_END uses reserved fields\n");
9061 /* check src operand */
9062 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
9066 if (is_pointer_value(env, insn->dst_reg)) {
9067 verbose(env, "R%d pointer arithmetic prohibited\n",
9072 /* check dest operand */
9073 err = check_reg_arg(env, insn->dst_reg, DST_OP);
9077 } else if (opcode == BPF_MOV) {
9079 if (BPF_SRC(insn->code) == BPF_X) {
9080 if (insn->imm != 0 || insn->off != 0) {
9081 verbose(env, "BPF_MOV uses reserved fields\n");
9085 /* check src operand */
9086 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9090 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
9091 verbose(env, "BPF_MOV uses reserved fields\n");
9096 /* check dest operand, mark as required later */
9097 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
9101 if (BPF_SRC(insn->code) == BPF_X) {
9102 struct bpf_reg_state *src_reg = regs + insn->src_reg;
9103 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
9105 if (BPF_CLASS(insn->code) == BPF_ALU64) {
9107 * copy register state to dest reg
9109 if (src_reg->type == SCALAR_VALUE && !src_reg->id)
9110 /* Assign src and dst registers the same ID
9111 * that will be used by find_equal_scalars()
9112 * to propagate min/max range.
9114 src_reg->id = ++env->id_gen;
9115 *dst_reg = *src_reg;
9116 dst_reg->live |= REG_LIVE_WRITTEN;
9117 dst_reg->subreg_def = DEF_NOT_SUBREG;
9120 if (is_pointer_value(env, insn->src_reg)) {
9122 "R%d partial copy of pointer\n",
9125 } else if (src_reg->type == SCALAR_VALUE) {
9126 *dst_reg = *src_reg;
9127 /* Make sure ID is cleared otherwise
9128 * dst_reg min/max could be incorrectly
9129 * propagated into src_reg by find_equal_scalars()
9132 dst_reg->live |= REG_LIVE_WRITTEN;
9133 dst_reg->subreg_def = env->insn_idx + 1;
9135 mark_reg_unknown(env, regs,
9138 zext_32_to_64(dst_reg);
9140 __update_reg_bounds(dst_reg);
9141 __reg_deduce_bounds(dst_reg);
9142 __reg_bound_offset(dst_reg);
9146 * remember the value we stored into this reg
9148 /* clear any state __mark_reg_known doesn't set */
9149 mark_reg_unknown(env, regs, insn->dst_reg);
9150 regs[insn->dst_reg].type = SCALAR_VALUE;
9151 if (BPF_CLASS(insn->code) == BPF_ALU64) {
9152 __mark_reg_known(regs + insn->dst_reg,
9155 __mark_reg_known(regs + insn->dst_reg,
9160 } else if (opcode > BPF_END) {
9161 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
9164 } else { /* all other ALU ops: and, sub, xor, add, ... */
9166 if (BPF_SRC(insn->code) == BPF_X) {
9167 if (insn->imm != 0 || insn->off != 0) {
9168 verbose(env, "BPF_ALU uses reserved fields\n");
9171 /* check src1 operand */
9172 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9176 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
9177 verbose(env, "BPF_ALU uses reserved fields\n");
9182 /* check src2 operand */
9183 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
9187 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
9188 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
9189 verbose(env, "div by zero\n");
9193 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
9194 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
9195 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
9197 if (insn->imm < 0 || insn->imm >= size) {
9198 verbose(env, "invalid shift %d\n", insn->imm);
9203 /* check dest operand */
9204 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
9208 return adjust_reg_min_max_vals(env, insn);
9214 static void __find_good_pkt_pointers(struct bpf_func_state *state,
9215 struct bpf_reg_state *dst_reg,
9216 enum bpf_reg_type type, int new_range)
9218 struct bpf_reg_state *reg;
9221 for (i = 0; i < MAX_BPF_REG; i++) {
9222 reg = &state->regs[i];
9223 if (reg->type == type && reg->id == dst_reg->id)
9224 /* keep the maximum range already checked */
9225 reg->range = max(reg->range, new_range);
9228 bpf_for_each_spilled_reg(i, state, reg) {
9231 if (reg->type == type && reg->id == dst_reg->id)
9232 reg->range = max(reg->range, new_range);
9236 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
9237 struct bpf_reg_state *dst_reg,
9238 enum bpf_reg_type type,
9239 bool range_right_open)
9243 if (dst_reg->off < 0 ||
9244 (dst_reg->off == 0 && range_right_open))
9245 /* This doesn't give us any range */
9248 if (dst_reg->umax_value > MAX_PACKET_OFF ||
9249 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
9250 /* Risk of overflow. For instance, ptr + (1<<63) may be less
9251 * than pkt_end, but that's because it's also less than pkt.
9255 new_range = dst_reg->off;
9256 if (range_right_open)
9259 /* Examples for register markings:
9261 * pkt_data in dst register:
9265 * if (r2 > pkt_end) goto <handle exception>
9270 * if (r2 < pkt_end) goto <access okay>
9271 * <handle exception>
9274 * r2 == dst_reg, pkt_end == src_reg
9275 * r2=pkt(id=n,off=8,r=0)
9276 * r3=pkt(id=n,off=0,r=0)
9278 * pkt_data in src register:
9282 * if (pkt_end >= r2) goto <access okay>
9283 * <handle exception>
9287 * if (pkt_end <= r2) goto <handle exception>
9291 * pkt_end == dst_reg, r2 == src_reg
9292 * r2=pkt(id=n,off=8,r=0)
9293 * r3=pkt(id=n,off=0,r=0)
9295 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
9296 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
9297 * and [r3, r3 + 8-1) respectively is safe to access depending on
9301 /* If our ids match, then we must have the same max_value. And we
9302 * don't care about the other reg's fixed offset, since if it's too big
9303 * the range won't allow anything.
9304 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
9306 for (i = 0; i <= vstate->curframe; i++)
9307 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
9311 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
9313 struct tnum subreg = tnum_subreg(reg->var_off);
9314 s32 sval = (s32)val;
9318 if (tnum_is_const(subreg))
9319 return !!tnum_equals_const(subreg, val);
9322 if (tnum_is_const(subreg))
9323 return !tnum_equals_const(subreg, val);
9326 if ((~subreg.mask & subreg.value) & val)
9328 if (!((subreg.mask | subreg.value) & val))
9332 if (reg->u32_min_value > val)
9334 else if (reg->u32_max_value <= val)
9338 if (reg->s32_min_value > sval)
9340 else if (reg->s32_max_value <= sval)
9344 if (reg->u32_max_value < val)
9346 else if (reg->u32_min_value >= val)
9350 if (reg->s32_max_value < sval)
9352 else if (reg->s32_min_value >= sval)
9356 if (reg->u32_min_value >= val)
9358 else if (reg->u32_max_value < val)
9362 if (reg->s32_min_value >= sval)
9364 else if (reg->s32_max_value < sval)
9368 if (reg->u32_max_value <= val)
9370 else if (reg->u32_min_value > val)
9374 if (reg->s32_max_value <= sval)
9376 else if (reg->s32_min_value > sval)
9385 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
9387 s64 sval = (s64)val;
9391 if (tnum_is_const(reg->var_off))
9392 return !!tnum_equals_const(reg->var_off, val);
9395 if (tnum_is_const(reg->var_off))
9396 return !tnum_equals_const(reg->var_off, val);
9399 if ((~reg->var_off.mask & reg->var_off.value) & val)
9401 if (!((reg->var_off.mask | reg->var_off.value) & val))
9405 if (reg->umin_value > val)
9407 else if (reg->umax_value <= val)
9411 if (reg->smin_value > sval)
9413 else if (reg->smax_value <= sval)
9417 if (reg->umax_value < val)
9419 else if (reg->umin_value >= val)
9423 if (reg->smax_value < sval)
9425 else if (reg->smin_value >= sval)
9429 if (reg->umin_value >= val)
9431 else if (reg->umax_value < val)
9435 if (reg->smin_value >= sval)
9437 else if (reg->smax_value < sval)
9441 if (reg->umax_value <= val)
9443 else if (reg->umin_value > val)
9447 if (reg->smax_value <= sval)
9449 else if (reg->smin_value > sval)
9457 /* compute branch direction of the expression "if (reg opcode val) goto target;"
9459 * 1 - branch will be taken and "goto target" will be executed
9460 * 0 - branch will not be taken and fall-through to next insn
9461 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
9464 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
9467 if (__is_pointer_value(false, reg)) {
9468 if (!reg_type_not_null(reg->type))
9471 /* If pointer is valid tests against zero will fail so we can
9472 * use this to direct branch taken.
9488 return is_branch32_taken(reg, val, opcode);
9489 return is_branch64_taken(reg, val, opcode);
9492 static int flip_opcode(u32 opcode)
9494 /* How can we transform "a <op> b" into "b <op> a"? */
9495 static const u8 opcode_flip[16] = {
9496 /* these stay the same */
9497 [BPF_JEQ >> 4] = BPF_JEQ,
9498 [BPF_JNE >> 4] = BPF_JNE,
9499 [BPF_JSET >> 4] = BPF_JSET,
9500 /* these swap "lesser" and "greater" (L and G in the opcodes) */
9501 [BPF_JGE >> 4] = BPF_JLE,
9502 [BPF_JGT >> 4] = BPF_JLT,
9503 [BPF_JLE >> 4] = BPF_JGE,
9504 [BPF_JLT >> 4] = BPF_JGT,
9505 [BPF_JSGE >> 4] = BPF_JSLE,
9506 [BPF_JSGT >> 4] = BPF_JSLT,
9507 [BPF_JSLE >> 4] = BPF_JSGE,
9508 [BPF_JSLT >> 4] = BPF_JSGT
9510 return opcode_flip[opcode >> 4];
9513 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
9514 struct bpf_reg_state *src_reg,
9517 struct bpf_reg_state *pkt;
9519 if (src_reg->type == PTR_TO_PACKET_END) {
9521 } else if (dst_reg->type == PTR_TO_PACKET_END) {
9523 opcode = flip_opcode(opcode);
9528 if (pkt->range >= 0)
9533 /* pkt <= pkt_end */
9537 if (pkt->range == BEYOND_PKT_END)
9538 /* pkt has at last one extra byte beyond pkt_end */
9539 return opcode == BPF_JGT;
9545 /* pkt >= pkt_end */
9546 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
9547 return opcode == BPF_JGE;
9553 /* Adjusts the register min/max values in the case that the dst_reg is the
9554 * variable register that we are working on, and src_reg is a constant or we're
9555 * simply doing a BPF_K check.
9556 * In JEQ/JNE cases we also adjust the var_off values.
9558 static void reg_set_min_max(struct bpf_reg_state *true_reg,
9559 struct bpf_reg_state *false_reg,
9561 u8 opcode, bool is_jmp32)
9563 struct tnum false_32off = tnum_subreg(false_reg->var_off);
9564 struct tnum false_64off = false_reg->var_off;
9565 struct tnum true_32off = tnum_subreg(true_reg->var_off);
9566 struct tnum true_64off = true_reg->var_off;
9567 s64 sval = (s64)val;
9568 s32 sval32 = (s32)val32;
9570 /* If the dst_reg is a pointer, we can't learn anything about its
9571 * variable offset from the compare (unless src_reg were a pointer into
9572 * the same object, but we don't bother with that.
9573 * Since false_reg and true_reg have the same type by construction, we
9574 * only need to check one of them for pointerness.
9576 if (__is_pointer_value(false, false_reg))
9583 struct bpf_reg_state *reg =
9584 opcode == BPF_JEQ ? true_reg : false_reg;
9586 /* JEQ/JNE comparison doesn't change the register equivalence.
9588 * if (r1 == 42) goto label;
9590 * label: // here both r1 and r2 are known to be 42.
9592 * Hence when marking register as known preserve it's ID.
9595 __mark_reg32_known(reg, val32);
9597 ___mark_reg_known(reg, val);
9602 false_32off = tnum_and(false_32off, tnum_const(~val32));
9603 if (is_power_of_2(val32))
9604 true_32off = tnum_or(true_32off,
9607 false_64off = tnum_and(false_64off, tnum_const(~val));
9608 if (is_power_of_2(val))
9609 true_64off = tnum_or(true_64off,
9617 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
9618 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
9620 false_reg->u32_max_value = min(false_reg->u32_max_value,
9622 true_reg->u32_min_value = max(true_reg->u32_min_value,
9625 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
9626 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
9628 false_reg->umax_value = min(false_reg->umax_value, false_umax);
9629 true_reg->umin_value = max(true_reg->umin_value, true_umin);
9637 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
9638 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
9640 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
9641 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
9643 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
9644 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
9646 false_reg->smax_value = min(false_reg->smax_value, false_smax);
9647 true_reg->smin_value = max(true_reg->smin_value, true_smin);
9655 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
9656 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
9658 false_reg->u32_min_value = max(false_reg->u32_min_value,
9660 true_reg->u32_max_value = min(true_reg->u32_max_value,
9663 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
9664 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
9666 false_reg->umin_value = max(false_reg->umin_value, false_umin);
9667 true_reg->umax_value = min(true_reg->umax_value, true_umax);
9675 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
9676 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
9678 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
9679 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
9681 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
9682 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
9684 false_reg->smin_value = max(false_reg->smin_value, false_smin);
9685 true_reg->smax_value = min(true_reg->smax_value, true_smax);
9694 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
9695 tnum_subreg(false_32off));
9696 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
9697 tnum_subreg(true_32off));
9698 __reg_combine_32_into_64(false_reg);
9699 __reg_combine_32_into_64(true_reg);
9701 false_reg->var_off = false_64off;
9702 true_reg->var_off = true_64off;
9703 __reg_combine_64_into_32(false_reg);
9704 __reg_combine_64_into_32(true_reg);
9708 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
9711 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
9712 struct bpf_reg_state *false_reg,
9714 u8 opcode, bool is_jmp32)
9716 opcode = flip_opcode(opcode);
9717 /* This uses zero as "not present in table"; luckily the zero opcode,
9718 * BPF_JA, can't get here.
9721 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
9724 /* Regs are known to be equal, so intersect their min/max/var_off */
9725 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
9726 struct bpf_reg_state *dst_reg)
9728 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
9729 dst_reg->umin_value);
9730 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
9731 dst_reg->umax_value);
9732 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
9733 dst_reg->smin_value);
9734 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
9735 dst_reg->smax_value);
9736 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
9738 /* We might have learned new bounds from the var_off. */
9739 __update_reg_bounds(src_reg);
9740 __update_reg_bounds(dst_reg);
9741 /* We might have learned something about the sign bit. */
9742 __reg_deduce_bounds(src_reg);
9743 __reg_deduce_bounds(dst_reg);
9744 /* We might have learned some bits from the bounds. */
9745 __reg_bound_offset(src_reg);
9746 __reg_bound_offset(dst_reg);
9747 /* Intersecting with the old var_off might have improved our bounds
9748 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
9749 * then new var_off is (0; 0x7f...fc) which improves our umax.
9751 __update_reg_bounds(src_reg);
9752 __update_reg_bounds(dst_reg);
9755 static void reg_combine_min_max(struct bpf_reg_state *true_src,
9756 struct bpf_reg_state *true_dst,
9757 struct bpf_reg_state *false_src,
9758 struct bpf_reg_state *false_dst,
9763 __reg_combine_min_max(true_src, true_dst);
9766 __reg_combine_min_max(false_src, false_dst);
9771 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
9772 struct bpf_reg_state *reg, u32 id,
9775 if (type_may_be_null(reg->type) && reg->id == id &&
9776 !WARN_ON_ONCE(!reg->id)) {
9777 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
9778 !tnum_equals_const(reg->var_off, 0) ||
9780 /* Old offset (both fixed and variable parts) should
9781 * have been known-zero, because we don't allow pointer
9782 * arithmetic on pointers that might be NULL. If we
9783 * see this happening, don't convert the register.
9788 reg->type = SCALAR_VALUE;
9789 /* We don't need id and ref_obj_id from this point
9790 * onwards anymore, thus we should better reset it,
9791 * so that state pruning has chances to take effect.
9794 reg->ref_obj_id = 0;
9799 mark_ptr_not_null_reg(reg);
9801 if (!reg_may_point_to_spin_lock(reg)) {
9802 /* For not-NULL ptr, reg->ref_obj_id will be reset
9803 * in release_reg_references().
9805 * reg->id is still used by spin_lock ptr. Other
9806 * than spin_lock ptr type, reg->id can be reset.
9813 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
9816 struct bpf_reg_state *reg;
9819 for (i = 0; i < MAX_BPF_REG; i++)
9820 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
9822 bpf_for_each_spilled_reg(i, state, reg) {
9825 mark_ptr_or_null_reg(state, reg, id, is_null);
9829 /* The logic is similar to find_good_pkt_pointers(), both could eventually
9830 * be folded together at some point.
9832 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
9835 struct bpf_func_state *state = vstate->frame[vstate->curframe];
9836 struct bpf_reg_state *regs = state->regs;
9837 u32 ref_obj_id = regs[regno].ref_obj_id;
9838 u32 id = regs[regno].id;
9841 if (ref_obj_id && ref_obj_id == id && is_null)
9842 /* regs[regno] is in the " == NULL" branch.
9843 * No one could have freed the reference state before
9844 * doing the NULL check.
9846 WARN_ON_ONCE(release_reference_state(state, id));
9848 for (i = 0; i <= vstate->curframe; i++)
9849 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
9852 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
9853 struct bpf_reg_state *dst_reg,
9854 struct bpf_reg_state *src_reg,
9855 struct bpf_verifier_state *this_branch,
9856 struct bpf_verifier_state *other_branch)
9858 if (BPF_SRC(insn->code) != BPF_X)
9861 /* Pointers are always 64-bit. */
9862 if (BPF_CLASS(insn->code) == BPF_JMP32)
9865 switch (BPF_OP(insn->code)) {
9867 if ((dst_reg->type == PTR_TO_PACKET &&
9868 src_reg->type == PTR_TO_PACKET_END) ||
9869 (dst_reg->type == PTR_TO_PACKET_META &&
9870 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
9871 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
9872 find_good_pkt_pointers(this_branch, dst_reg,
9873 dst_reg->type, false);
9874 mark_pkt_end(other_branch, insn->dst_reg, true);
9875 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
9876 src_reg->type == PTR_TO_PACKET) ||
9877 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
9878 src_reg->type == PTR_TO_PACKET_META)) {
9879 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
9880 find_good_pkt_pointers(other_branch, src_reg,
9881 src_reg->type, true);
9882 mark_pkt_end(this_branch, insn->src_reg, false);
9888 if ((dst_reg->type == PTR_TO_PACKET &&
9889 src_reg->type == PTR_TO_PACKET_END) ||
9890 (dst_reg->type == PTR_TO_PACKET_META &&
9891 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
9892 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
9893 find_good_pkt_pointers(other_branch, dst_reg,
9894 dst_reg->type, true);
9895 mark_pkt_end(this_branch, insn->dst_reg, false);
9896 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
9897 src_reg->type == PTR_TO_PACKET) ||
9898 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
9899 src_reg->type == PTR_TO_PACKET_META)) {
9900 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
9901 find_good_pkt_pointers(this_branch, src_reg,
9902 src_reg->type, false);
9903 mark_pkt_end(other_branch, insn->src_reg, true);
9909 if ((dst_reg->type == PTR_TO_PACKET &&
9910 src_reg->type == PTR_TO_PACKET_END) ||
9911 (dst_reg->type == PTR_TO_PACKET_META &&
9912 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
9913 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
9914 find_good_pkt_pointers(this_branch, dst_reg,
9915 dst_reg->type, true);
9916 mark_pkt_end(other_branch, insn->dst_reg, false);
9917 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
9918 src_reg->type == PTR_TO_PACKET) ||
9919 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
9920 src_reg->type == PTR_TO_PACKET_META)) {
9921 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
9922 find_good_pkt_pointers(other_branch, src_reg,
9923 src_reg->type, false);
9924 mark_pkt_end(this_branch, insn->src_reg, true);
9930 if ((dst_reg->type == PTR_TO_PACKET &&
9931 src_reg->type == PTR_TO_PACKET_END) ||
9932 (dst_reg->type == PTR_TO_PACKET_META &&
9933 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
9934 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
9935 find_good_pkt_pointers(other_branch, dst_reg,
9936 dst_reg->type, false);
9937 mark_pkt_end(this_branch, insn->dst_reg, true);
9938 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
9939 src_reg->type == PTR_TO_PACKET) ||
9940 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
9941 src_reg->type == PTR_TO_PACKET_META)) {
9942 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
9943 find_good_pkt_pointers(this_branch, src_reg,
9944 src_reg->type, true);
9945 mark_pkt_end(other_branch, insn->src_reg, false);
9957 static void find_equal_scalars(struct bpf_verifier_state *vstate,
9958 struct bpf_reg_state *known_reg)
9960 struct bpf_func_state *state;
9961 struct bpf_reg_state *reg;
9964 for (i = 0; i <= vstate->curframe; i++) {
9965 state = vstate->frame[i];
9966 for (j = 0; j < MAX_BPF_REG; j++) {
9967 reg = &state->regs[j];
9968 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
9972 bpf_for_each_spilled_reg(j, state, reg) {
9975 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
9981 static int check_cond_jmp_op(struct bpf_verifier_env *env,
9982 struct bpf_insn *insn, int *insn_idx)
9984 struct bpf_verifier_state *this_branch = env->cur_state;
9985 struct bpf_verifier_state *other_branch;
9986 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
9987 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
9988 u8 opcode = BPF_OP(insn->code);
9993 /* Only conditional jumps are expected to reach here. */
9994 if (opcode == BPF_JA || opcode > BPF_JSLE) {
9995 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
9999 if (BPF_SRC(insn->code) == BPF_X) {
10000 if (insn->imm != 0) {
10001 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
10005 /* check src1 operand */
10006 err = check_reg_arg(env, insn->src_reg, SRC_OP);
10010 if (is_pointer_value(env, insn->src_reg)) {
10011 verbose(env, "R%d pointer comparison prohibited\n",
10015 src_reg = ®s[insn->src_reg];
10017 if (insn->src_reg != BPF_REG_0) {
10018 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
10023 /* check src2 operand */
10024 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
10028 dst_reg = ®s[insn->dst_reg];
10029 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
10031 if (BPF_SRC(insn->code) == BPF_K) {
10032 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
10033 } else if (src_reg->type == SCALAR_VALUE &&
10034 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
10035 pred = is_branch_taken(dst_reg,
10036 tnum_subreg(src_reg->var_off).value,
10039 } else if (src_reg->type == SCALAR_VALUE &&
10040 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
10041 pred = is_branch_taken(dst_reg,
10042 src_reg->var_off.value,
10045 } else if (reg_is_pkt_pointer_any(dst_reg) &&
10046 reg_is_pkt_pointer_any(src_reg) &&
10048 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
10052 /* If we get here with a dst_reg pointer type it is because
10053 * above is_branch_taken() special cased the 0 comparison.
10055 if (!__is_pointer_value(false, dst_reg))
10056 err = mark_chain_precision(env, insn->dst_reg);
10057 if (BPF_SRC(insn->code) == BPF_X && !err &&
10058 !__is_pointer_value(false, src_reg))
10059 err = mark_chain_precision(env, insn->src_reg);
10065 /* Only follow the goto, ignore fall-through. If needed, push
10066 * the fall-through branch for simulation under speculative
10069 if (!env->bypass_spec_v1 &&
10070 !sanitize_speculative_path(env, insn, *insn_idx + 1,
10073 *insn_idx += insn->off;
10075 } else if (pred == 0) {
10076 /* Only follow the fall-through branch, since that's where the
10077 * program will go. If needed, push the goto branch for
10078 * simulation under speculative execution.
10080 if (!env->bypass_spec_v1 &&
10081 !sanitize_speculative_path(env, insn,
10082 *insn_idx + insn->off + 1,
10088 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
10092 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
10094 /* detect if we are comparing against a constant value so we can adjust
10095 * our min/max values for our dst register.
10096 * this is only legit if both are scalars (or pointers to the same
10097 * object, I suppose, but we don't support that right now), because
10098 * otherwise the different base pointers mean the offsets aren't
10101 if (BPF_SRC(insn->code) == BPF_X) {
10102 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
10104 if (dst_reg->type == SCALAR_VALUE &&
10105 src_reg->type == SCALAR_VALUE) {
10106 if (tnum_is_const(src_reg->var_off) ||
10108 tnum_is_const(tnum_subreg(src_reg->var_off))))
10109 reg_set_min_max(&other_branch_regs[insn->dst_reg],
10111 src_reg->var_off.value,
10112 tnum_subreg(src_reg->var_off).value,
10114 else if (tnum_is_const(dst_reg->var_off) ||
10116 tnum_is_const(tnum_subreg(dst_reg->var_off))))
10117 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
10119 dst_reg->var_off.value,
10120 tnum_subreg(dst_reg->var_off).value,
10122 else if (!is_jmp32 &&
10123 (opcode == BPF_JEQ || opcode == BPF_JNE))
10124 /* Comparing for equality, we can combine knowledge */
10125 reg_combine_min_max(&other_branch_regs[insn->src_reg],
10126 &other_branch_regs[insn->dst_reg],
10127 src_reg, dst_reg, opcode);
10129 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
10130 find_equal_scalars(this_branch, src_reg);
10131 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
10135 } else if (dst_reg->type == SCALAR_VALUE) {
10136 reg_set_min_max(&other_branch_regs[insn->dst_reg],
10137 dst_reg, insn->imm, (u32)insn->imm,
10141 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
10142 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
10143 find_equal_scalars(this_branch, dst_reg);
10144 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
10147 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
10148 * NOTE: these optimizations below are related with pointer comparison
10149 * which will never be JMP32.
10151 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
10152 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
10153 type_may_be_null(dst_reg->type)) {
10154 /* Mark all identical registers in each branch as either
10155 * safe or unknown depending R == 0 or R != 0 conditional.
10157 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
10158 opcode == BPF_JNE);
10159 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
10160 opcode == BPF_JEQ);
10161 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
10162 this_branch, other_branch) &&
10163 is_pointer_value(env, insn->dst_reg)) {
10164 verbose(env, "R%d pointer comparison prohibited\n",
10168 if (env->log.level & BPF_LOG_LEVEL)
10169 print_insn_state(env, this_branch->frame[this_branch->curframe]);
10173 /* verify BPF_LD_IMM64 instruction */
10174 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
10176 struct bpf_insn_aux_data *aux = cur_aux(env);
10177 struct bpf_reg_state *regs = cur_regs(env);
10178 struct bpf_reg_state *dst_reg;
10179 struct bpf_map *map;
10182 if (BPF_SIZE(insn->code) != BPF_DW) {
10183 verbose(env, "invalid BPF_LD_IMM insn\n");
10186 if (insn->off != 0) {
10187 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
10191 err = check_reg_arg(env, insn->dst_reg, DST_OP);
10195 dst_reg = ®s[insn->dst_reg];
10196 if (insn->src_reg == 0) {
10197 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
10199 dst_reg->type = SCALAR_VALUE;
10200 __mark_reg_known(®s[insn->dst_reg], imm);
10204 /* All special src_reg cases are listed below. From this point onwards
10205 * we either succeed and assign a corresponding dst_reg->type after
10206 * zeroing the offset, or fail and reject the program.
10208 mark_reg_known_zero(env, regs, insn->dst_reg);
10210 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
10211 dst_reg->type = aux->btf_var.reg_type;
10212 switch (base_type(dst_reg->type)) {
10214 dst_reg->mem_size = aux->btf_var.mem_size;
10216 case PTR_TO_BTF_ID:
10217 dst_reg->btf = aux->btf_var.btf;
10218 dst_reg->btf_id = aux->btf_var.btf_id;
10221 verbose(env, "bpf verifier is misconfigured\n");
10227 if (insn->src_reg == BPF_PSEUDO_FUNC) {
10228 struct bpf_prog_aux *aux = env->prog->aux;
10229 u32 subprogno = find_subprog(env,
10230 env->insn_idx + insn->imm + 1);
10232 if (!aux->func_info) {
10233 verbose(env, "missing btf func_info\n");
10236 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
10237 verbose(env, "callback function not static\n");
10241 dst_reg->type = PTR_TO_FUNC;
10242 dst_reg->subprogno = subprogno;
10246 map = env->used_maps[aux->map_index];
10247 dst_reg->map_ptr = map;
10249 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
10250 insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
10251 dst_reg->type = PTR_TO_MAP_VALUE;
10252 dst_reg->off = aux->map_off;
10253 if (map_value_has_spin_lock(map))
10254 dst_reg->id = ++env->id_gen;
10255 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
10256 insn->src_reg == BPF_PSEUDO_MAP_IDX) {
10257 dst_reg->type = CONST_PTR_TO_MAP;
10259 verbose(env, "bpf verifier is misconfigured\n");
10266 static bool may_access_skb(enum bpf_prog_type type)
10269 case BPF_PROG_TYPE_SOCKET_FILTER:
10270 case BPF_PROG_TYPE_SCHED_CLS:
10271 case BPF_PROG_TYPE_SCHED_ACT:
10278 /* verify safety of LD_ABS|LD_IND instructions:
10279 * - they can only appear in the programs where ctx == skb
10280 * - since they are wrappers of function calls, they scratch R1-R5 registers,
10281 * preserve R6-R9, and store return value into R0
10284 * ctx == skb == R6 == CTX
10287 * SRC == any register
10288 * IMM == 32-bit immediate
10291 * R0 - 8/16/32-bit skb data converted to cpu endianness
10293 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
10295 struct bpf_reg_state *regs = cur_regs(env);
10296 static const int ctx_reg = BPF_REG_6;
10297 u8 mode = BPF_MODE(insn->code);
10300 if (!may_access_skb(resolve_prog_type(env->prog))) {
10301 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
10305 if (!env->ops->gen_ld_abs) {
10306 verbose(env, "bpf verifier is misconfigured\n");
10310 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
10311 BPF_SIZE(insn->code) == BPF_DW ||
10312 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
10313 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
10317 /* check whether implicit source operand (register R6) is readable */
10318 err = check_reg_arg(env, ctx_reg, SRC_OP);
10322 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
10323 * gen_ld_abs() may terminate the program at runtime, leading to
10326 err = check_reference_leak(env);
10328 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
10332 if (env->cur_state->active_spin_lock) {
10333 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
10337 if (regs[ctx_reg].type != PTR_TO_CTX) {
10339 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
10343 if (mode == BPF_IND) {
10344 /* check explicit source operand */
10345 err = check_reg_arg(env, insn->src_reg, SRC_OP);
10350 err = check_ptr_off_reg(env, ®s[ctx_reg], ctx_reg);
10354 /* reset caller saved regs to unreadable */
10355 for (i = 0; i < CALLER_SAVED_REGS; i++) {
10356 mark_reg_not_init(env, regs, caller_saved[i]);
10357 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
10360 /* mark destination R0 register as readable, since it contains
10361 * the value fetched from the packet.
10362 * Already marked as written above.
10364 mark_reg_unknown(env, regs, BPF_REG_0);
10365 /* ld_abs load up to 32-bit skb data. */
10366 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
10370 static int check_return_code(struct bpf_verifier_env *env)
10372 struct tnum enforce_attach_type_range = tnum_unknown;
10373 const struct bpf_prog *prog = env->prog;
10374 struct bpf_reg_state *reg;
10375 struct tnum range = tnum_range(0, 1);
10376 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
10378 struct bpf_func_state *frame = env->cur_state->frame[0];
10379 const bool is_subprog = frame->subprogno;
10381 /* LSM and struct_ops func-ptr's return type could be "void" */
10383 (prog_type == BPF_PROG_TYPE_STRUCT_OPS ||
10384 prog_type == BPF_PROG_TYPE_LSM) &&
10385 !prog->aux->attach_func_proto->type)
10388 /* eBPF calling convention is such that R0 is used
10389 * to return the value from eBPF program.
10390 * Make sure that it's readable at this time
10391 * of bpf_exit, which means that program wrote
10392 * something into it earlier
10394 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
10398 if (is_pointer_value(env, BPF_REG_0)) {
10399 verbose(env, "R0 leaks addr as return value\n");
10403 reg = cur_regs(env) + BPF_REG_0;
10405 if (frame->in_async_callback_fn) {
10406 /* enforce return zero from async callbacks like timer */
10407 if (reg->type != SCALAR_VALUE) {
10408 verbose(env, "In async callback the register R0 is not a known value (%s)\n",
10409 reg_type_str(env, reg->type));
10413 if (!tnum_in(tnum_const(0), reg->var_off)) {
10414 verbose_invalid_scalar(env, reg, &range, "async callback", "R0");
10421 if (reg->type != SCALAR_VALUE) {
10422 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
10423 reg_type_str(env, reg->type));
10429 switch (prog_type) {
10430 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
10431 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
10432 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
10433 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
10434 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
10435 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
10436 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
10437 range = tnum_range(1, 1);
10438 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
10439 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
10440 range = tnum_range(0, 3);
10442 case BPF_PROG_TYPE_CGROUP_SKB:
10443 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
10444 range = tnum_range(0, 3);
10445 enforce_attach_type_range = tnum_range(2, 3);
10448 case BPF_PROG_TYPE_CGROUP_SOCK:
10449 case BPF_PROG_TYPE_SOCK_OPS:
10450 case BPF_PROG_TYPE_CGROUP_DEVICE:
10451 case BPF_PROG_TYPE_CGROUP_SYSCTL:
10452 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
10454 case BPF_PROG_TYPE_RAW_TRACEPOINT:
10455 if (!env->prog->aux->attach_btf_id)
10457 range = tnum_const(0);
10459 case BPF_PROG_TYPE_TRACING:
10460 switch (env->prog->expected_attach_type) {
10461 case BPF_TRACE_FENTRY:
10462 case BPF_TRACE_FEXIT:
10463 range = tnum_const(0);
10465 case BPF_TRACE_RAW_TP:
10466 case BPF_MODIFY_RETURN:
10468 case BPF_TRACE_ITER:
10474 case BPF_PROG_TYPE_SK_LOOKUP:
10475 range = tnum_range(SK_DROP, SK_PASS);
10477 case BPF_PROG_TYPE_EXT:
10478 /* freplace program can return anything as its return value
10479 * depends on the to-be-replaced kernel func or bpf program.
10485 if (reg->type != SCALAR_VALUE) {
10486 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
10487 reg_type_str(env, reg->type));
10491 if (!tnum_in(range, reg->var_off)) {
10492 verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
10496 if (!tnum_is_unknown(enforce_attach_type_range) &&
10497 tnum_in(enforce_attach_type_range, reg->var_off))
10498 env->prog->enforce_expected_attach_type = 1;
10502 /* non-recursive DFS pseudo code
10503 * 1 procedure DFS-iterative(G,v):
10504 * 2 label v as discovered
10505 * 3 let S be a stack
10507 * 5 while S is not empty
10509 * 7 if t is what we're looking for:
10511 * 9 for all edges e in G.adjacentEdges(t) do
10512 * 10 if edge e is already labelled
10513 * 11 continue with the next edge
10514 * 12 w <- G.adjacentVertex(t,e)
10515 * 13 if vertex w is not discovered and not explored
10516 * 14 label e as tree-edge
10517 * 15 label w as discovered
10520 * 18 else if vertex w is discovered
10521 * 19 label e as back-edge
10523 * 21 // vertex w is explored
10524 * 22 label e as forward- or cross-edge
10525 * 23 label t as explored
10529 * 0x10 - discovered
10530 * 0x11 - discovered and fall-through edge labelled
10531 * 0x12 - discovered and fall-through and branch edges labelled
10542 static u32 state_htab_size(struct bpf_verifier_env *env)
10544 return env->prog->len;
10547 static struct bpf_verifier_state_list **explored_state(
10548 struct bpf_verifier_env *env,
10551 struct bpf_verifier_state *cur = env->cur_state;
10552 struct bpf_func_state *state = cur->frame[cur->curframe];
10554 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
10557 static void init_explored_state(struct bpf_verifier_env *env, int idx)
10559 env->insn_aux_data[idx].prune_point = true;
10563 DONE_EXPLORING = 0,
10564 KEEP_EXPLORING = 1,
10567 /* t, w, e - match pseudo-code above:
10568 * t - index of current instruction
10569 * w - next instruction
10572 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
10575 int *insn_stack = env->cfg.insn_stack;
10576 int *insn_state = env->cfg.insn_state;
10578 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
10579 return DONE_EXPLORING;
10581 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
10582 return DONE_EXPLORING;
10584 if (w < 0 || w >= env->prog->len) {
10585 verbose_linfo(env, t, "%d: ", t);
10586 verbose(env, "jump out of range from insn %d to %d\n", t, w);
10591 /* mark branch target for state pruning */
10592 init_explored_state(env, w);
10594 if (insn_state[w] == 0) {
10596 insn_state[t] = DISCOVERED | e;
10597 insn_state[w] = DISCOVERED;
10598 if (env->cfg.cur_stack >= env->prog->len)
10600 insn_stack[env->cfg.cur_stack++] = w;
10601 return KEEP_EXPLORING;
10602 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
10603 if (loop_ok && env->bpf_capable)
10604 return DONE_EXPLORING;
10605 verbose_linfo(env, t, "%d: ", t);
10606 verbose_linfo(env, w, "%d: ", w);
10607 verbose(env, "back-edge from insn %d to %d\n", t, w);
10609 } else if (insn_state[w] == EXPLORED) {
10610 /* forward- or cross-edge */
10611 insn_state[t] = DISCOVERED | e;
10613 verbose(env, "insn state internal bug\n");
10616 return DONE_EXPLORING;
10619 static int visit_func_call_insn(int t, int insn_cnt,
10620 struct bpf_insn *insns,
10621 struct bpf_verifier_env *env,
10626 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
10630 if (t + 1 < insn_cnt)
10631 init_explored_state(env, t + 1);
10632 if (visit_callee) {
10633 init_explored_state(env, t);
10634 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env,
10635 /* It's ok to allow recursion from CFG point of
10636 * view. __check_func_call() will do the actual
10639 bpf_pseudo_func(insns + t));
10644 /* Visits the instruction at index t and returns one of the following:
10645 * < 0 - an error occurred
10646 * DONE_EXPLORING - the instruction was fully explored
10647 * KEEP_EXPLORING - there is still work to be done before it is fully explored
10649 static int visit_insn(int t, int insn_cnt, struct bpf_verifier_env *env)
10651 struct bpf_insn *insns = env->prog->insnsi;
10654 if (bpf_pseudo_func(insns + t))
10655 return visit_func_call_insn(t, insn_cnt, insns, env, true);
10657 /* All non-branch instructions have a single fall-through edge. */
10658 if (BPF_CLASS(insns[t].code) != BPF_JMP &&
10659 BPF_CLASS(insns[t].code) != BPF_JMP32)
10660 return push_insn(t, t + 1, FALLTHROUGH, env, false);
10662 switch (BPF_OP(insns[t].code)) {
10664 return DONE_EXPLORING;
10667 if (insns[t].imm == BPF_FUNC_timer_set_callback)
10668 /* Mark this call insn to trigger is_state_visited() check
10669 * before call itself is processed by __check_func_call().
10670 * Otherwise new async state will be pushed for further
10673 init_explored_state(env, t);
10674 return visit_func_call_insn(t, insn_cnt, insns, env,
10675 insns[t].src_reg == BPF_PSEUDO_CALL);
10678 if (BPF_SRC(insns[t].code) != BPF_K)
10681 /* unconditional jump with single edge */
10682 ret = push_insn(t, t + insns[t].off + 1, FALLTHROUGH, env,
10687 /* unconditional jmp is not a good pruning point,
10688 * but it's marked, since backtracking needs
10689 * to record jmp history in is_state_visited().
10691 init_explored_state(env, t + insns[t].off + 1);
10692 /* tell verifier to check for equivalent states
10693 * after every call and jump
10695 if (t + 1 < insn_cnt)
10696 init_explored_state(env, t + 1);
10701 /* conditional jump with two edges */
10702 init_explored_state(env, t);
10703 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
10707 return push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
10711 /* non-recursive depth-first-search to detect loops in BPF program
10712 * loop == back-edge in directed graph
10714 static int check_cfg(struct bpf_verifier_env *env)
10716 int insn_cnt = env->prog->len;
10717 int *insn_stack, *insn_state;
10721 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
10725 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
10727 kvfree(insn_state);
10731 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
10732 insn_stack[0] = 0; /* 0 is the first instruction */
10733 env->cfg.cur_stack = 1;
10735 while (env->cfg.cur_stack > 0) {
10736 int t = insn_stack[env->cfg.cur_stack - 1];
10738 ret = visit_insn(t, insn_cnt, env);
10740 case DONE_EXPLORING:
10741 insn_state[t] = EXPLORED;
10742 env->cfg.cur_stack--;
10744 case KEEP_EXPLORING:
10748 verbose(env, "visit_insn internal bug\n");
10755 if (env->cfg.cur_stack < 0) {
10756 verbose(env, "pop stack internal bug\n");
10761 for (i = 0; i < insn_cnt; i++) {
10762 if (insn_state[i] != EXPLORED) {
10763 verbose(env, "unreachable insn %d\n", i);
10768 ret = 0; /* cfg looks good */
10771 kvfree(insn_state);
10772 kvfree(insn_stack);
10773 env->cfg.insn_state = env->cfg.insn_stack = NULL;
10777 static int check_abnormal_return(struct bpf_verifier_env *env)
10781 for (i = 1; i < env->subprog_cnt; i++) {
10782 if (env->subprog_info[i].has_ld_abs) {
10783 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
10786 if (env->subprog_info[i].has_tail_call) {
10787 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
10794 /* The minimum supported BTF func info size */
10795 #define MIN_BPF_FUNCINFO_SIZE 8
10796 #define MAX_FUNCINFO_REC_SIZE 252
10798 static int check_btf_func(struct bpf_verifier_env *env,
10799 const union bpf_attr *attr,
10802 const struct btf_type *type, *func_proto, *ret_type;
10803 u32 i, nfuncs, urec_size, min_size;
10804 u32 krec_size = sizeof(struct bpf_func_info);
10805 struct bpf_func_info *krecord;
10806 struct bpf_func_info_aux *info_aux = NULL;
10807 struct bpf_prog *prog;
10808 const struct btf *btf;
10810 u32 prev_offset = 0;
10811 bool scalar_return;
10814 nfuncs = attr->func_info_cnt;
10816 if (check_abnormal_return(env))
10821 if (nfuncs != env->subprog_cnt) {
10822 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
10826 urec_size = attr->func_info_rec_size;
10827 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
10828 urec_size > MAX_FUNCINFO_REC_SIZE ||
10829 urec_size % sizeof(u32)) {
10830 verbose(env, "invalid func info rec size %u\n", urec_size);
10835 btf = prog->aux->btf;
10837 urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
10838 min_size = min_t(u32, krec_size, urec_size);
10840 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
10843 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
10847 for (i = 0; i < nfuncs; i++) {
10848 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
10850 if (ret == -E2BIG) {
10851 verbose(env, "nonzero tailing record in func info");
10852 /* set the size kernel expects so loader can zero
10853 * out the rest of the record.
10855 if (copy_to_bpfptr_offset(uattr,
10856 offsetof(union bpf_attr, func_info_rec_size),
10857 &min_size, sizeof(min_size)))
10863 if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
10868 /* check insn_off */
10871 if (krecord[i].insn_off) {
10873 "nonzero insn_off %u for the first func info record",
10874 krecord[i].insn_off);
10877 } else if (krecord[i].insn_off <= prev_offset) {
10879 "same or smaller insn offset (%u) than previous func info record (%u)",
10880 krecord[i].insn_off, prev_offset);
10884 if (env->subprog_info[i].start != krecord[i].insn_off) {
10885 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
10889 /* check type_id */
10890 type = btf_type_by_id(btf, krecord[i].type_id);
10891 if (!type || !btf_type_is_func(type)) {
10892 verbose(env, "invalid type id %d in func info",
10893 krecord[i].type_id);
10896 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
10898 func_proto = btf_type_by_id(btf, type->type);
10899 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
10900 /* btf_func_check() already verified it during BTF load */
10902 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
10904 btf_type_is_small_int(ret_type) || btf_type_is_enum(ret_type);
10905 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
10906 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
10909 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
10910 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
10914 prev_offset = krecord[i].insn_off;
10915 bpfptr_add(&urecord, urec_size);
10918 prog->aux->func_info = krecord;
10919 prog->aux->func_info_cnt = nfuncs;
10920 prog->aux->func_info_aux = info_aux;
10929 static void adjust_btf_func(struct bpf_verifier_env *env)
10931 struct bpf_prog_aux *aux = env->prog->aux;
10934 if (!aux->func_info)
10937 for (i = 0; i < env->subprog_cnt; i++)
10938 aux->func_info[i].insn_off = env->subprog_info[i].start;
10941 #define MIN_BPF_LINEINFO_SIZE offsetofend(struct bpf_line_info, line_col)
10942 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
10944 static int check_btf_line(struct bpf_verifier_env *env,
10945 const union bpf_attr *attr,
10948 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
10949 struct bpf_subprog_info *sub;
10950 struct bpf_line_info *linfo;
10951 struct bpf_prog *prog;
10952 const struct btf *btf;
10956 nr_linfo = attr->line_info_cnt;
10959 if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info))
10962 rec_size = attr->line_info_rec_size;
10963 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
10964 rec_size > MAX_LINEINFO_REC_SIZE ||
10965 rec_size & (sizeof(u32) - 1))
10968 /* Need to zero it in case the userspace may
10969 * pass in a smaller bpf_line_info object.
10971 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
10972 GFP_KERNEL | __GFP_NOWARN);
10977 btf = prog->aux->btf;
10980 sub = env->subprog_info;
10981 ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
10982 expected_size = sizeof(struct bpf_line_info);
10983 ncopy = min_t(u32, expected_size, rec_size);
10984 for (i = 0; i < nr_linfo; i++) {
10985 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
10987 if (err == -E2BIG) {
10988 verbose(env, "nonzero tailing record in line_info");
10989 if (copy_to_bpfptr_offset(uattr,
10990 offsetof(union bpf_attr, line_info_rec_size),
10991 &expected_size, sizeof(expected_size)))
10997 if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
11003 * Check insn_off to ensure
11004 * 1) strictly increasing AND
11005 * 2) bounded by prog->len
11007 * The linfo[0].insn_off == 0 check logically falls into
11008 * the later "missing bpf_line_info for func..." case
11009 * because the first linfo[0].insn_off must be the
11010 * first sub also and the first sub must have
11011 * subprog_info[0].start == 0.
11013 if ((i && linfo[i].insn_off <= prev_offset) ||
11014 linfo[i].insn_off >= prog->len) {
11015 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
11016 i, linfo[i].insn_off, prev_offset,
11022 if (!prog->insnsi[linfo[i].insn_off].code) {
11024 "Invalid insn code at line_info[%u].insn_off\n",
11030 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
11031 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
11032 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
11037 if (s != env->subprog_cnt) {
11038 if (linfo[i].insn_off == sub[s].start) {
11039 sub[s].linfo_idx = i;
11041 } else if (sub[s].start < linfo[i].insn_off) {
11042 verbose(env, "missing bpf_line_info for func#%u\n", s);
11048 prev_offset = linfo[i].insn_off;
11049 bpfptr_add(&ulinfo, rec_size);
11052 if (s != env->subprog_cnt) {
11053 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
11054 env->subprog_cnt - s, s);
11059 prog->aux->linfo = linfo;
11060 prog->aux->nr_linfo = nr_linfo;
11069 #define MIN_CORE_RELO_SIZE sizeof(struct bpf_core_relo)
11070 #define MAX_CORE_RELO_SIZE MAX_FUNCINFO_REC_SIZE
11072 static int check_core_relo(struct bpf_verifier_env *env,
11073 const union bpf_attr *attr,
11076 u32 i, nr_core_relo, ncopy, expected_size, rec_size;
11077 struct bpf_core_relo core_relo = {};
11078 struct bpf_prog *prog = env->prog;
11079 const struct btf *btf = prog->aux->btf;
11080 struct bpf_core_ctx ctx = {
11084 bpfptr_t u_core_relo;
11087 nr_core_relo = attr->core_relo_cnt;
11090 if (nr_core_relo > INT_MAX / sizeof(struct bpf_core_relo))
11093 rec_size = attr->core_relo_rec_size;
11094 if (rec_size < MIN_CORE_RELO_SIZE ||
11095 rec_size > MAX_CORE_RELO_SIZE ||
11096 rec_size % sizeof(u32))
11099 u_core_relo = make_bpfptr(attr->core_relos, uattr.is_kernel);
11100 expected_size = sizeof(struct bpf_core_relo);
11101 ncopy = min_t(u32, expected_size, rec_size);
11103 /* Unlike func_info and line_info, copy and apply each CO-RE
11104 * relocation record one at a time.
11106 for (i = 0; i < nr_core_relo; i++) {
11107 /* future proofing when sizeof(bpf_core_relo) changes */
11108 err = bpf_check_uarg_tail_zero(u_core_relo, expected_size, rec_size);
11110 if (err == -E2BIG) {
11111 verbose(env, "nonzero tailing record in core_relo");
11112 if (copy_to_bpfptr_offset(uattr,
11113 offsetof(union bpf_attr, core_relo_rec_size),
11114 &expected_size, sizeof(expected_size)))
11120 if (copy_from_bpfptr(&core_relo, u_core_relo, ncopy)) {
11125 if (core_relo.insn_off % 8 || core_relo.insn_off / 8 >= prog->len) {
11126 verbose(env, "Invalid core_relo[%u].insn_off:%u prog->len:%u\n",
11127 i, core_relo.insn_off, prog->len);
11132 err = bpf_core_apply(&ctx, &core_relo, i,
11133 &prog->insnsi[core_relo.insn_off / 8]);
11136 bpfptr_add(&u_core_relo, rec_size);
11141 static int check_btf_info(struct bpf_verifier_env *env,
11142 const union bpf_attr *attr,
11148 if (!attr->func_info_cnt && !attr->line_info_cnt) {
11149 if (check_abnormal_return(env))
11154 btf = btf_get_by_fd(attr->prog_btf_fd);
11156 return PTR_ERR(btf);
11157 if (btf_is_kernel(btf)) {
11161 env->prog->aux->btf = btf;
11163 err = check_btf_func(env, attr, uattr);
11167 err = check_btf_line(env, attr, uattr);
11171 err = check_core_relo(env, attr, uattr);
11178 /* check %cur's range satisfies %old's */
11179 static bool range_within(struct bpf_reg_state *old,
11180 struct bpf_reg_state *cur)
11182 return old->umin_value <= cur->umin_value &&
11183 old->umax_value >= cur->umax_value &&
11184 old->smin_value <= cur->smin_value &&
11185 old->smax_value >= cur->smax_value &&
11186 old->u32_min_value <= cur->u32_min_value &&
11187 old->u32_max_value >= cur->u32_max_value &&
11188 old->s32_min_value <= cur->s32_min_value &&
11189 old->s32_max_value >= cur->s32_max_value;
11192 /* If in the old state two registers had the same id, then they need to have
11193 * the same id in the new state as well. But that id could be different from
11194 * the old state, so we need to track the mapping from old to new ids.
11195 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
11196 * regs with old id 5 must also have new id 9 for the new state to be safe. But
11197 * regs with a different old id could still have new id 9, we don't care about
11199 * So we look through our idmap to see if this old id has been seen before. If
11200 * so, we require the new id to match; otherwise, we add the id pair to the map.
11202 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_id_pair *idmap)
11206 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
11207 if (!idmap[i].old) {
11208 /* Reached an empty slot; haven't seen this id before */
11209 idmap[i].old = old_id;
11210 idmap[i].cur = cur_id;
11213 if (idmap[i].old == old_id)
11214 return idmap[i].cur == cur_id;
11216 /* We ran out of idmap slots, which should be impossible */
11221 static void clean_func_state(struct bpf_verifier_env *env,
11222 struct bpf_func_state *st)
11224 enum bpf_reg_liveness live;
11227 for (i = 0; i < BPF_REG_FP; i++) {
11228 live = st->regs[i].live;
11229 /* liveness must not touch this register anymore */
11230 st->regs[i].live |= REG_LIVE_DONE;
11231 if (!(live & REG_LIVE_READ))
11232 /* since the register is unused, clear its state
11233 * to make further comparison simpler
11235 __mark_reg_not_init(env, &st->regs[i]);
11238 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
11239 live = st->stack[i].spilled_ptr.live;
11240 /* liveness must not touch this stack slot anymore */
11241 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
11242 if (!(live & REG_LIVE_READ)) {
11243 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
11244 for (j = 0; j < BPF_REG_SIZE; j++)
11245 st->stack[i].slot_type[j] = STACK_INVALID;
11250 static void clean_verifier_state(struct bpf_verifier_env *env,
11251 struct bpf_verifier_state *st)
11255 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
11256 /* all regs in this state in all frames were already marked */
11259 for (i = 0; i <= st->curframe; i++)
11260 clean_func_state(env, st->frame[i]);
11263 /* the parentage chains form a tree.
11264 * the verifier states are added to state lists at given insn and
11265 * pushed into state stack for future exploration.
11266 * when the verifier reaches bpf_exit insn some of the verifer states
11267 * stored in the state lists have their final liveness state already,
11268 * but a lot of states will get revised from liveness point of view when
11269 * the verifier explores other branches.
11272 * 2: if r1 == 100 goto pc+1
11275 * when the verifier reaches exit insn the register r0 in the state list of
11276 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
11277 * of insn 2 and goes exploring further. At the insn 4 it will walk the
11278 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
11280 * Since the verifier pushes the branch states as it sees them while exploring
11281 * the program the condition of walking the branch instruction for the second
11282 * time means that all states below this branch were already explored and
11283 * their final liveness marks are already propagated.
11284 * Hence when the verifier completes the search of state list in is_state_visited()
11285 * we can call this clean_live_states() function to mark all liveness states
11286 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
11287 * will not be used.
11288 * This function also clears the registers and stack for states that !READ
11289 * to simplify state merging.
11291 * Important note here that walking the same branch instruction in the callee
11292 * doesn't meant that the states are DONE. The verifier has to compare
11295 static void clean_live_states(struct bpf_verifier_env *env, int insn,
11296 struct bpf_verifier_state *cur)
11298 struct bpf_verifier_state_list *sl;
11301 sl = *explored_state(env, insn);
11303 if (sl->state.branches)
11305 if (sl->state.insn_idx != insn ||
11306 sl->state.curframe != cur->curframe)
11308 for (i = 0; i <= cur->curframe; i++)
11309 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
11311 clean_verifier_state(env, &sl->state);
11317 /* Returns true if (rold safe implies rcur safe) */
11318 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
11319 struct bpf_reg_state *rcur, struct bpf_id_pair *idmap)
11323 if (!(rold->live & REG_LIVE_READ))
11324 /* explored state didn't use this */
11327 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
11329 if (rold->type == PTR_TO_STACK)
11330 /* two stack pointers are equal only if they're pointing to
11331 * the same stack frame, since fp-8 in foo != fp-8 in bar
11333 return equal && rold->frameno == rcur->frameno;
11338 if (rold->type == NOT_INIT)
11339 /* explored state can't have used this */
11341 if (rcur->type == NOT_INIT)
11343 switch (base_type(rold->type)) {
11345 if (env->explore_alu_limits)
11347 if (rcur->type == SCALAR_VALUE) {
11348 if (!rold->precise && !rcur->precise)
11350 /* new val must satisfy old val knowledge */
11351 return range_within(rold, rcur) &&
11352 tnum_in(rold->var_off, rcur->var_off);
11354 /* We're trying to use a pointer in place of a scalar.
11355 * Even if the scalar was unbounded, this could lead to
11356 * pointer leaks because scalars are allowed to leak
11357 * while pointers are not. We could make this safe in
11358 * special cases if root is calling us, but it's
11359 * probably not worth the hassle.
11363 case PTR_TO_MAP_KEY:
11364 case PTR_TO_MAP_VALUE:
11365 /* a PTR_TO_MAP_VALUE could be safe to use as a
11366 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
11367 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
11368 * checked, doing so could have affected others with the same
11369 * id, and we can't check for that because we lost the id when
11370 * we converted to a PTR_TO_MAP_VALUE.
11372 if (type_may_be_null(rold->type)) {
11373 if (!type_may_be_null(rcur->type))
11375 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
11377 /* Check our ids match any regs they're supposed to */
11378 return check_ids(rold->id, rcur->id, idmap);
11381 /* If the new min/max/var_off satisfy the old ones and
11382 * everything else matches, we are OK.
11383 * 'id' is not compared, since it's only used for maps with
11384 * bpf_spin_lock inside map element and in such cases if
11385 * the rest of the prog is valid for one map element then
11386 * it's valid for all map elements regardless of the key
11387 * used in bpf_map_lookup()
11389 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
11390 range_within(rold, rcur) &&
11391 tnum_in(rold->var_off, rcur->var_off);
11392 case PTR_TO_PACKET_META:
11393 case PTR_TO_PACKET:
11394 if (rcur->type != rold->type)
11396 /* We must have at least as much range as the old ptr
11397 * did, so that any accesses which were safe before are
11398 * still safe. This is true even if old range < old off,
11399 * since someone could have accessed through (ptr - k), or
11400 * even done ptr -= k in a register, to get a safe access.
11402 if (rold->range > rcur->range)
11404 /* If the offsets don't match, we can't trust our alignment;
11405 * nor can we be sure that we won't fall out of range.
11407 if (rold->off != rcur->off)
11409 /* id relations must be preserved */
11410 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
11412 /* new val must satisfy old val knowledge */
11413 return range_within(rold, rcur) &&
11414 tnum_in(rold->var_off, rcur->var_off);
11416 case CONST_PTR_TO_MAP:
11417 case PTR_TO_PACKET_END:
11418 case PTR_TO_FLOW_KEYS:
11419 case PTR_TO_SOCKET:
11420 case PTR_TO_SOCK_COMMON:
11421 case PTR_TO_TCP_SOCK:
11422 case PTR_TO_XDP_SOCK:
11423 /* Only valid matches are exact, which memcmp() above
11424 * would have accepted
11427 /* Don't know what's going on, just say it's not safe */
11431 /* Shouldn't get here; if we do, say it's not safe */
11436 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
11437 struct bpf_func_state *cur, struct bpf_id_pair *idmap)
11441 /* walk slots of the explored stack and ignore any additional
11442 * slots in the current stack, since explored(safe) state
11445 for (i = 0; i < old->allocated_stack; i++) {
11446 spi = i / BPF_REG_SIZE;
11448 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
11449 i += BPF_REG_SIZE - 1;
11450 /* explored state didn't use this */
11454 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
11457 /* explored stack has more populated slots than current stack
11458 * and these slots were used
11460 if (i >= cur->allocated_stack)
11463 /* if old state was safe with misc data in the stack
11464 * it will be safe with zero-initialized stack.
11465 * The opposite is not true
11467 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
11468 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
11470 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
11471 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
11472 /* Ex: old explored (safe) state has STACK_SPILL in
11473 * this stack slot, but current has STACK_MISC ->
11474 * this verifier states are not equivalent,
11475 * return false to continue verification of this path
11478 if (i % BPF_REG_SIZE != BPF_REG_SIZE - 1)
11480 if (!is_spilled_reg(&old->stack[spi]))
11482 if (!regsafe(env, &old->stack[spi].spilled_ptr,
11483 &cur->stack[spi].spilled_ptr, idmap))
11484 /* when explored and current stack slot are both storing
11485 * spilled registers, check that stored pointers types
11486 * are the same as well.
11487 * Ex: explored safe path could have stored
11488 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
11489 * but current path has stored:
11490 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
11491 * such verifier states are not equivalent.
11492 * return false to continue verification of this path
11499 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
11501 if (old->acquired_refs != cur->acquired_refs)
11503 return !memcmp(old->refs, cur->refs,
11504 sizeof(*old->refs) * old->acquired_refs);
11507 /* compare two verifier states
11509 * all states stored in state_list are known to be valid, since
11510 * verifier reached 'bpf_exit' instruction through them
11512 * this function is called when verifier exploring different branches of
11513 * execution popped from the state stack. If it sees an old state that has
11514 * more strict register state and more strict stack state then this execution
11515 * branch doesn't need to be explored further, since verifier already
11516 * concluded that more strict state leads to valid finish.
11518 * Therefore two states are equivalent if register state is more conservative
11519 * and explored stack state is more conservative than the current one.
11522 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
11523 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
11525 * In other words if current stack state (one being explored) has more
11526 * valid slots than old one that already passed validation, it means
11527 * the verifier can stop exploring and conclude that current state is valid too
11529 * Similarly with registers. If explored state has register type as invalid
11530 * whereas register type in current state is meaningful, it means that
11531 * the current state will reach 'bpf_exit' instruction safely
11533 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
11534 struct bpf_func_state *cur)
11538 memset(env->idmap_scratch, 0, sizeof(env->idmap_scratch));
11539 for (i = 0; i < MAX_BPF_REG; i++)
11540 if (!regsafe(env, &old->regs[i], &cur->regs[i],
11541 env->idmap_scratch))
11544 if (!stacksafe(env, old, cur, env->idmap_scratch))
11547 if (!refsafe(old, cur))
11553 static bool states_equal(struct bpf_verifier_env *env,
11554 struct bpf_verifier_state *old,
11555 struct bpf_verifier_state *cur)
11559 if (old->curframe != cur->curframe)
11562 /* Verification state from speculative execution simulation
11563 * must never prune a non-speculative execution one.
11565 if (old->speculative && !cur->speculative)
11568 if (old->active_spin_lock != cur->active_spin_lock)
11571 /* for states to be equal callsites have to be the same
11572 * and all frame states need to be equivalent
11574 for (i = 0; i <= old->curframe; i++) {
11575 if (old->frame[i]->callsite != cur->frame[i]->callsite)
11577 if (!func_states_equal(env, old->frame[i], cur->frame[i]))
11583 /* Return 0 if no propagation happened. Return negative error code if error
11584 * happened. Otherwise, return the propagated bit.
11586 static int propagate_liveness_reg(struct bpf_verifier_env *env,
11587 struct bpf_reg_state *reg,
11588 struct bpf_reg_state *parent_reg)
11590 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
11591 u8 flag = reg->live & REG_LIVE_READ;
11594 /* When comes here, read flags of PARENT_REG or REG could be any of
11595 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
11596 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
11598 if (parent_flag == REG_LIVE_READ64 ||
11599 /* Or if there is no read flag from REG. */
11601 /* Or if the read flag from REG is the same as PARENT_REG. */
11602 parent_flag == flag)
11605 err = mark_reg_read(env, reg, parent_reg, flag);
11612 /* A write screens off any subsequent reads; but write marks come from the
11613 * straight-line code between a state and its parent. When we arrive at an
11614 * equivalent state (jump target or such) we didn't arrive by the straight-line
11615 * code, so read marks in the state must propagate to the parent regardless
11616 * of the state's write marks. That's what 'parent == state->parent' comparison
11617 * in mark_reg_read() is for.
11619 static int propagate_liveness(struct bpf_verifier_env *env,
11620 const struct bpf_verifier_state *vstate,
11621 struct bpf_verifier_state *vparent)
11623 struct bpf_reg_state *state_reg, *parent_reg;
11624 struct bpf_func_state *state, *parent;
11625 int i, frame, err = 0;
11627 if (vparent->curframe != vstate->curframe) {
11628 WARN(1, "propagate_live: parent frame %d current frame %d\n",
11629 vparent->curframe, vstate->curframe);
11632 /* Propagate read liveness of registers... */
11633 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
11634 for (frame = 0; frame <= vstate->curframe; frame++) {
11635 parent = vparent->frame[frame];
11636 state = vstate->frame[frame];
11637 parent_reg = parent->regs;
11638 state_reg = state->regs;
11639 /* We don't need to worry about FP liveness, it's read-only */
11640 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
11641 err = propagate_liveness_reg(env, &state_reg[i],
11645 if (err == REG_LIVE_READ64)
11646 mark_insn_zext(env, &parent_reg[i]);
11649 /* Propagate stack slots. */
11650 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
11651 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
11652 parent_reg = &parent->stack[i].spilled_ptr;
11653 state_reg = &state->stack[i].spilled_ptr;
11654 err = propagate_liveness_reg(env, state_reg,
11663 /* find precise scalars in the previous equivalent state and
11664 * propagate them into the current state
11666 static int propagate_precision(struct bpf_verifier_env *env,
11667 const struct bpf_verifier_state *old)
11669 struct bpf_reg_state *state_reg;
11670 struct bpf_func_state *state;
11673 state = old->frame[old->curframe];
11674 state_reg = state->regs;
11675 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
11676 if (state_reg->type != SCALAR_VALUE ||
11677 !state_reg->precise)
11679 if (env->log.level & BPF_LOG_LEVEL2)
11680 verbose(env, "propagating r%d\n", i);
11681 err = mark_chain_precision(env, i);
11686 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
11687 if (!is_spilled_reg(&state->stack[i]))
11689 state_reg = &state->stack[i].spilled_ptr;
11690 if (state_reg->type != SCALAR_VALUE ||
11691 !state_reg->precise)
11693 if (env->log.level & BPF_LOG_LEVEL2)
11694 verbose(env, "propagating fp%d\n",
11695 (-i - 1) * BPF_REG_SIZE);
11696 err = mark_chain_precision_stack(env, i);
11703 static bool states_maybe_looping(struct bpf_verifier_state *old,
11704 struct bpf_verifier_state *cur)
11706 struct bpf_func_state *fold, *fcur;
11707 int i, fr = cur->curframe;
11709 if (old->curframe != fr)
11712 fold = old->frame[fr];
11713 fcur = cur->frame[fr];
11714 for (i = 0; i < MAX_BPF_REG; i++)
11715 if (memcmp(&fold->regs[i], &fcur->regs[i],
11716 offsetof(struct bpf_reg_state, parent)))
11722 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
11724 struct bpf_verifier_state_list *new_sl;
11725 struct bpf_verifier_state_list *sl, **pprev;
11726 struct bpf_verifier_state *cur = env->cur_state, *new;
11727 int i, j, err, states_cnt = 0;
11728 bool add_new_state = env->test_state_freq ? true : false;
11730 cur->last_insn_idx = env->prev_insn_idx;
11731 if (!env->insn_aux_data[insn_idx].prune_point)
11732 /* this 'insn_idx' instruction wasn't marked, so we will not
11733 * be doing state search here
11737 /* bpf progs typically have pruning point every 4 instructions
11738 * http://vger.kernel.org/bpfconf2019.html#session-1
11739 * Do not add new state for future pruning if the verifier hasn't seen
11740 * at least 2 jumps and at least 8 instructions.
11741 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
11742 * In tests that amounts to up to 50% reduction into total verifier
11743 * memory consumption and 20% verifier time speedup.
11745 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
11746 env->insn_processed - env->prev_insn_processed >= 8)
11747 add_new_state = true;
11749 pprev = explored_state(env, insn_idx);
11752 clean_live_states(env, insn_idx, cur);
11756 if (sl->state.insn_idx != insn_idx)
11759 if (sl->state.branches) {
11760 struct bpf_func_state *frame = sl->state.frame[sl->state.curframe];
11762 if (frame->in_async_callback_fn &&
11763 frame->async_entry_cnt != cur->frame[cur->curframe]->async_entry_cnt) {
11764 /* Different async_entry_cnt means that the verifier is
11765 * processing another entry into async callback.
11766 * Seeing the same state is not an indication of infinite
11767 * loop or infinite recursion.
11768 * But finding the same state doesn't mean that it's safe
11769 * to stop processing the current state. The previous state
11770 * hasn't yet reached bpf_exit, since state.branches > 0.
11771 * Checking in_async_callback_fn alone is not enough either.
11772 * Since the verifier still needs to catch infinite loops
11773 * inside async callbacks.
11775 } else if (states_maybe_looping(&sl->state, cur) &&
11776 states_equal(env, &sl->state, cur)) {
11777 verbose_linfo(env, insn_idx, "; ");
11778 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
11781 /* if the verifier is processing a loop, avoid adding new state
11782 * too often, since different loop iterations have distinct
11783 * states and may not help future pruning.
11784 * This threshold shouldn't be too low to make sure that
11785 * a loop with large bound will be rejected quickly.
11786 * The most abusive loop will be:
11788 * if r1 < 1000000 goto pc-2
11789 * 1M insn_procssed limit / 100 == 10k peak states.
11790 * This threshold shouldn't be too high either, since states
11791 * at the end of the loop are likely to be useful in pruning.
11793 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
11794 env->insn_processed - env->prev_insn_processed < 100)
11795 add_new_state = false;
11798 if (states_equal(env, &sl->state, cur)) {
11800 /* reached equivalent register/stack state,
11801 * prune the search.
11802 * Registers read by the continuation are read by us.
11803 * If we have any write marks in env->cur_state, they
11804 * will prevent corresponding reads in the continuation
11805 * from reaching our parent (an explored_state). Our
11806 * own state will get the read marks recorded, but
11807 * they'll be immediately forgotten as we're pruning
11808 * this state and will pop a new one.
11810 err = propagate_liveness(env, &sl->state, cur);
11812 /* if previous state reached the exit with precision and
11813 * current state is equivalent to it (except precsion marks)
11814 * the precision needs to be propagated back in
11815 * the current state.
11817 err = err ? : push_jmp_history(env, cur);
11818 err = err ? : propagate_precision(env, &sl->state);
11824 /* when new state is not going to be added do not increase miss count.
11825 * Otherwise several loop iterations will remove the state
11826 * recorded earlier. The goal of these heuristics is to have
11827 * states from some iterations of the loop (some in the beginning
11828 * and some at the end) to help pruning.
11832 /* heuristic to determine whether this state is beneficial
11833 * to keep checking from state equivalence point of view.
11834 * Higher numbers increase max_states_per_insn and verification time,
11835 * but do not meaningfully decrease insn_processed.
11837 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
11838 /* the state is unlikely to be useful. Remove it to
11839 * speed up verification
11842 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
11843 u32 br = sl->state.branches;
11846 "BUG live_done but branches_to_explore %d\n",
11848 free_verifier_state(&sl->state, false);
11850 env->peak_states--;
11852 /* cannot free this state, since parentage chain may
11853 * walk it later. Add it for free_list instead to
11854 * be freed at the end of verification
11856 sl->next = env->free_list;
11857 env->free_list = sl;
11867 if (env->max_states_per_insn < states_cnt)
11868 env->max_states_per_insn = states_cnt;
11870 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
11871 return push_jmp_history(env, cur);
11873 if (!add_new_state)
11874 return push_jmp_history(env, cur);
11876 /* There were no equivalent states, remember the current one.
11877 * Technically the current state is not proven to be safe yet,
11878 * but it will either reach outer most bpf_exit (which means it's safe)
11879 * or it will be rejected. When there are no loops the verifier won't be
11880 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
11881 * again on the way to bpf_exit.
11882 * When looping the sl->state.branches will be > 0 and this state
11883 * will not be considered for equivalence until branches == 0.
11885 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
11888 env->total_states++;
11889 env->peak_states++;
11890 env->prev_jmps_processed = env->jmps_processed;
11891 env->prev_insn_processed = env->insn_processed;
11893 /* add new state to the head of linked list */
11894 new = &new_sl->state;
11895 err = copy_verifier_state(new, cur);
11897 free_verifier_state(new, false);
11901 new->insn_idx = insn_idx;
11902 WARN_ONCE(new->branches != 1,
11903 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
11906 cur->first_insn_idx = insn_idx;
11907 clear_jmp_history(cur);
11908 new_sl->next = *explored_state(env, insn_idx);
11909 *explored_state(env, insn_idx) = new_sl;
11910 /* connect new state to parentage chain. Current frame needs all
11911 * registers connected. Only r6 - r9 of the callers are alive (pushed
11912 * to the stack implicitly by JITs) so in callers' frames connect just
11913 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
11914 * the state of the call instruction (with WRITTEN set), and r0 comes
11915 * from callee with its full parentage chain, anyway.
11917 /* clear write marks in current state: the writes we did are not writes
11918 * our child did, so they don't screen off its reads from us.
11919 * (There are no read marks in current state, because reads always mark
11920 * their parent and current state never has children yet. Only
11921 * explored_states can get read marks.)
11923 for (j = 0; j <= cur->curframe; j++) {
11924 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
11925 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
11926 for (i = 0; i < BPF_REG_FP; i++)
11927 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
11930 /* all stack frames are accessible from callee, clear them all */
11931 for (j = 0; j <= cur->curframe; j++) {
11932 struct bpf_func_state *frame = cur->frame[j];
11933 struct bpf_func_state *newframe = new->frame[j];
11935 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
11936 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
11937 frame->stack[i].spilled_ptr.parent =
11938 &newframe->stack[i].spilled_ptr;
11944 /* Return true if it's OK to have the same insn return a different type. */
11945 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
11947 switch (base_type(type)) {
11949 case PTR_TO_SOCKET:
11950 case PTR_TO_SOCK_COMMON:
11951 case PTR_TO_TCP_SOCK:
11952 case PTR_TO_XDP_SOCK:
11953 case PTR_TO_BTF_ID:
11960 /* If an instruction was previously used with particular pointer types, then we
11961 * need to be careful to avoid cases such as the below, where it may be ok
11962 * for one branch accessing the pointer, but not ok for the other branch:
11967 * R1 = some_other_valid_ptr;
11970 * R2 = *(u32 *)(R1 + 0);
11972 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
11974 return src != prev && (!reg_type_mismatch_ok(src) ||
11975 !reg_type_mismatch_ok(prev));
11978 static int do_check(struct bpf_verifier_env *env)
11980 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
11981 struct bpf_verifier_state *state = env->cur_state;
11982 struct bpf_insn *insns = env->prog->insnsi;
11983 struct bpf_reg_state *regs;
11984 int insn_cnt = env->prog->len;
11985 bool do_print_state = false;
11986 int prev_insn_idx = -1;
11989 struct bpf_insn *insn;
11993 env->prev_insn_idx = prev_insn_idx;
11994 if (env->insn_idx >= insn_cnt) {
11995 verbose(env, "invalid insn idx %d insn_cnt %d\n",
11996 env->insn_idx, insn_cnt);
12000 insn = &insns[env->insn_idx];
12001 class = BPF_CLASS(insn->code);
12003 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
12005 "BPF program is too large. Processed %d insn\n",
12006 env->insn_processed);
12010 err = is_state_visited(env, env->insn_idx);
12014 /* found equivalent state, can prune the search */
12015 if (env->log.level & BPF_LOG_LEVEL) {
12016 if (do_print_state)
12017 verbose(env, "\nfrom %d to %d%s: safe\n",
12018 env->prev_insn_idx, env->insn_idx,
12019 env->cur_state->speculative ?
12020 " (speculative execution)" : "");
12022 verbose(env, "%d: safe\n", env->insn_idx);
12024 goto process_bpf_exit;
12027 if (signal_pending(current))
12030 if (need_resched())
12033 if (env->log.level & BPF_LOG_LEVEL2 && do_print_state) {
12034 verbose(env, "\nfrom %d to %d%s:",
12035 env->prev_insn_idx, env->insn_idx,
12036 env->cur_state->speculative ?
12037 " (speculative execution)" : "");
12038 print_verifier_state(env, state->frame[state->curframe], true);
12039 do_print_state = false;
12042 if (env->log.level & BPF_LOG_LEVEL) {
12043 const struct bpf_insn_cbs cbs = {
12044 .cb_call = disasm_kfunc_name,
12045 .cb_print = verbose,
12046 .private_data = env,
12049 if (verifier_state_scratched(env))
12050 print_insn_state(env, state->frame[state->curframe]);
12052 verbose_linfo(env, env->insn_idx, "; ");
12053 env->prev_log_len = env->log.len_used;
12054 verbose(env, "%d: ", env->insn_idx);
12055 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
12056 env->prev_insn_print_len = env->log.len_used - env->prev_log_len;
12057 env->prev_log_len = env->log.len_used;
12060 if (bpf_prog_is_dev_bound(env->prog->aux)) {
12061 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
12062 env->prev_insn_idx);
12067 regs = cur_regs(env);
12068 sanitize_mark_insn_seen(env);
12069 prev_insn_idx = env->insn_idx;
12071 if (class == BPF_ALU || class == BPF_ALU64) {
12072 err = check_alu_op(env, insn);
12076 } else if (class == BPF_LDX) {
12077 enum bpf_reg_type *prev_src_type, src_reg_type;
12079 /* check for reserved fields is already done */
12081 /* check src operand */
12082 err = check_reg_arg(env, insn->src_reg, SRC_OP);
12086 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
12090 src_reg_type = regs[insn->src_reg].type;
12092 /* check that memory (src_reg + off) is readable,
12093 * the state of dst_reg will be updated by this func
12095 err = check_mem_access(env, env->insn_idx, insn->src_reg,
12096 insn->off, BPF_SIZE(insn->code),
12097 BPF_READ, insn->dst_reg, false);
12101 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
12103 if (*prev_src_type == NOT_INIT) {
12104 /* saw a valid insn
12105 * dst_reg = *(u32 *)(src_reg + off)
12106 * save type to validate intersecting paths
12108 *prev_src_type = src_reg_type;
12110 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
12111 /* ABuser program is trying to use the same insn
12112 * dst_reg = *(u32*) (src_reg + off)
12113 * with different pointer types:
12114 * src_reg == ctx in one branch and
12115 * src_reg == stack|map in some other branch.
12118 verbose(env, "same insn cannot be used with different pointers\n");
12122 } else if (class == BPF_STX) {
12123 enum bpf_reg_type *prev_dst_type, dst_reg_type;
12125 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
12126 err = check_atomic(env, env->insn_idx, insn);
12133 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
12134 verbose(env, "BPF_STX uses reserved fields\n");
12138 /* check src1 operand */
12139 err = check_reg_arg(env, insn->src_reg, SRC_OP);
12142 /* check src2 operand */
12143 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
12147 dst_reg_type = regs[insn->dst_reg].type;
12149 /* check that memory (dst_reg + off) is writeable */
12150 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
12151 insn->off, BPF_SIZE(insn->code),
12152 BPF_WRITE, insn->src_reg, false);
12156 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
12158 if (*prev_dst_type == NOT_INIT) {
12159 *prev_dst_type = dst_reg_type;
12160 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
12161 verbose(env, "same insn cannot be used with different pointers\n");
12165 } else if (class == BPF_ST) {
12166 if (BPF_MODE(insn->code) != BPF_MEM ||
12167 insn->src_reg != BPF_REG_0) {
12168 verbose(env, "BPF_ST uses reserved fields\n");
12171 /* check src operand */
12172 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
12176 if (is_ctx_reg(env, insn->dst_reg)) {
12177 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
12179 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
12183 /* check that memory (dst_reg + off) is writeable */
12184 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
12185 insn->off, BPF_SIZE(insn->code),
12186 BPF_WRITE, -1, false);
12190 } else if (class == BPF_JMP || class == BPF_JMP32) {
12191 u8 opcode = BPF_OP(insn->code);
12193 env->jmps_processed++;
12194 if (opcode == BPF_CALL) {
12195 if (BPF_SRC(insn->code) != BPF_K ||
12196 (insn->src_reg != BPF_PSEUDO_KFUNC_CALL
12197 && insn->off != 0) ||
12198 (insn->src_reg != BPF_REG_0 &&
12199 insn->src_reg != BPF_PSEUDO_CALL &&
12200 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
12201 insn->dst_reg != BPF_REG_0 ||
12202 class == BPF_JMP32) {
12203 verbose(env, "BPF_CALL uses reserved fields\n");
12207 if (env->cur_state->active_spin_lock &&
12208 (insn->src_reg == BPF_PSEUDO_CALL ||
12209 insn->imm != BPF_FUNC_spin_unlock)) {
12210 verbose(env, "function calls are not allowed while holding a lock\n");
12213 if (insn->src_reg == BPF_PSEUDO_CALL)
12214 err = check_func_call(env, insn, &env->insn_idx);
12215 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
12216 err = check_kfunc_call(env, insn, &env->insn_idx);
12218 err = check_helper_call(env, insn, &env->insn_idx);
12221 } else if (opcode == BPF_JA) {
12222 if (BPF_SRC(insn->code) != BPF_K ||
12224 insn->src_reg != BPF_REG_0 ||
12225 insn->dst_reg != BPF_REG_0 ||
12226 class == BPF_JMP32) {
12227 verbose(env, "BPF_JA uses reserved fields\n");
12231 env->insn_idx += insn->off + 1;
12234 } else if (opcode == BPF_EXIT) {
12235 if (BPF_SRC(insn->code) != BPF_K ||
12237 insn->src_reg != BPF_REG_0 ||
12238 insn->dst_reg != BPF_REG_0 ||
12239 class == BPF_JMP32) {
12240 verbose(env, "BPF_EXIT uses reserved fields\n");
12244 if (env->cur_state->active_spin_lock) {
12245 verbose(env, "bpf_spin_unlock is missing\n");
12249 if (state->curframe) {
12250 /* exit from nested function */
12251 err = prepare_func_exit(env, &env->insn_idx);
12254 do_print_state = true;
12258 err = check_reference_leak(env);
12262 err = check_return_code(env);
12266 mark_verifier_state_scratched(env);
12267 update_branch_counts(env, env->cur_state);
12268 err = pop_stack(env, &prev_insn_idx,
12269 &env->insn_idx, pop_log);
12271 if (err != -ENOENT)
12275 do_print_state = true;
12279 err = check_cond_jmp_op(env, insn, &env->insn_idx);
12283 } else if (class == BPF_LD) {
12284 u8 mode = BPF_MODE(insn->code);
12286 if (mode == BPF_ABS || mode == BPF_IND) {
12287 err = check_ld_abs(env, insn);
12291 } else if (mode == BPF_IMM) {
12292 err = check_ld_imm(env, insn);
12297 sanitize_mark_insn_seen(env);
12299 verbose(env, "invalid BPF_LD mode\n");
12303 verbose(env, "unknown insn class %d\n", class);
12313 static int find_btf_percpu_datasec(struct btf *btf)
12315 const struct btf_type *t;
12320 * Both vmlinux and module each have their own ".data..percpu"
12321 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
12322 * types to look at only module's own BTF types.
12324 n = btf_nr_types(btf);
12325 if (btf_is_module(btf))
12326 i = btf_nr_types(btf_vmlinux);
12330 for(; i < n; i++) {
12331 t = btf_type_by_id(btf, i);
12332 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
12335 tname = btf_name_by_offset(btf, t->name_off);
12336 if (!strcmp(tname, ".data..percpu"))
12343 /* replace pseudo btf_id with kernel symbol address */
12344 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
12345 struct bpf_insn *insn,
12346 struct bpf_insn_aux_data *aux)
12348 const struct btf_var_secinfo *vsi;
12349 const struct btf_type *datasec;
12350 struct btf_mod_pair *btf_mod;
12351 const struct btf_type *t;
12352 const char *sym_name;
12353 bool percpu = false;
12354 u32 type, id = insn->imm;
12358 int i, btf_fd, err;
12360 btf_fd = insn[1].imm;
12362 btf = btf_get_by_fd(btf_fd);
12364 verbose(env, "invalid module BTF object FD specified.\n");
12368 if (!btf_vmlinux) {
12369 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
12376 t = btf_type_by_id(btf, id);
12378 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
12383 if (!btf_type_is_var(t)) {
12384 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR.\n", id);
12389 sym_name = btf_name_by_offset(btf, t->name_off);
12390 addr = kallsyms_lookup_name(sym_name);
12392 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
12398 datasec_id = find_btf_percpu_datasec(btf);
12399 if (datasec_id > 0) {
12400 datasec = btf_type_by_id(btf, datasec_id);
12401 for_each_vsi(i, datasec, vsi) {
12402 if (vsi->type == id) {
12409 insn[0].imm = (u32)addr;
12410 insn[1].imm = addr >> 32;
12413 t = btf_type_skip_modifiers(btf, type, NULL);
12415 aux->btf_var.reg_type = PTR_TO_BTF_ID | MEM_PERCPU;
12416 aux->btf_var.btf = btf;
12417 aux->btf_var.btf_id = type;
12418 } else if (!btf_type_is_struct(t)) {
12419 const struct btf_type *ret;
12423 /* resolve the type size of ksym. */
12424 ret = btf_resolve_size(btf, t, &tsize);
12426 tname = btf_name_by_offset(btf, t->name_off);
12427 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
12428 tname, PTR_ERR(ret));
12432 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
12433 aux->btf_var.mem_size = tsize;
12435 aux->btf_var.reg_type = PTR_TO_BTF_ID;
12436 aux->btf_var.btf = btf;
12437 aux->btf_var.btf_id = type;
12440 /* check whether we recorded this BTF (and maybe module) already */
12441 for (i = 0; i < env->used_btf_cnt; i++) {
12442 if (env->used_btfs[i].btf == btf) {
12448 if (env->used_btf_cnt >= MAX_USED_BTFS) {
12453 btf_mod = &env->used_btfs[env->used_btf_cnt];
12454 btf_mod->btf = btf;
12455 btf_mod->module = NULL;
12457 /* if we reference variables from kernel module, bump its refcount */
12458 if (btf_is_module(btf)) {
12459 btf_mod->module = btf_try_get_module(btf);
12460 if (!btf_mod->module) {
12466 env->used_btf_cnt++;
12474 static int check_map_prealloc(struct bpf_map *map)
12476 return (map->map_type != BPF_MAP_TYPE_HASH &&
12477 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
12478 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
12479 !(map->map_flags & BPF_F_NO_PREALLOC);
12482 static bool is_tracing_prog_type(enum bpf_prog_type type)
12485 case BPF_PROG_TYPE_KPROBE:
12486 case BPF_PROG_TYPE_TRACEPOINT:
12487 case BPF_PROG_TYPE_PERF_EVENT:
12488 case BPF_PROG_TYPE_RAW_TRACEPOINT:
12495 static bool is_preallocated_map(struct bpf_map *map)
12497 if (!check_map_prealloc(map))
12499 if (map->inner_map_meta && !check_map_prealloc(map->inner_map_meta))
12504 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
12505 struct bpf_map *map,
12506 struct bpf_prog *prog)
12509 enum bpf_prog_type prog_type = resolve_prog_type(prog);
12511 * Validate that trace type programs use preallocated hash maps.
12513 * For programs attached to PERF events this is mandatory as the
12514 * perf NMI can hit any arbitrary code sequence.
12516 * All other trace types using preallocated hash maps are unsafe as
12517 * well because tracepoint or kprobes can be inside locked regions
12518 * of the memory allocator or at a place where a recursion into the
12519 * memory allocator would see inconsistent state.
12521 * On RT enabled kernels run-time allocation of all trace type
12522 * programs is strictly prohibited due to lock type constraints. On
12523 * !RT kernels it is allowed for backwards compatibility reasons for
12524 * now, but warnings are emitted so developers are made aware of
12525 * the unsafety and can fix their programs before this is enforced.
12527 if (is_tracing_prog_type(prog_type) && !is_preallocated_map(map)) {
12528 if (prog_type == BPF_PROG_TYPE_PERF_EVENT) {
12529 verbose(env, "perf_event programs can only use preallocated hash map\n");
12532 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
12533 verbose(env, "trace type programs can only use preallocated hash map\n");
12536 WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
12537 verbose(env, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
12540 if (map_value_has_spin_lock(map)) {
12541 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
12542 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
12546 if (is_tracing_prog_type(prog_type)) {
12547 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
12551 if (prog->aux->sleepable) {
12552 verbose(env, "sleepable progs cannot use bpf_spin_lock yet\n");
12557 if (map_value_has_timer(map)) {
12558 if (is_tracing_prog_type(prog_type)) {
12559 verbose(env, "tracing progs cannot use bpf_timer yet\n");
12564 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
12565 !bpf_offload_prog_map_match(prog, map)) {
12566 verbose(env, "offload device mismatch between prog and map\n");
12570 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
12571 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
12575 if (prog->aux->sleepable)
12576 switch (map->map_type) {
12577 case BPF_MAP_TYPE_HASH:
12578 case BPF_MAP_TYPE_LRU_HASH:
12579 case BPF_MAP_TYPE_ARRAY:
12580 case BPF_MAP_TYPE_PERCPU_HASH:
12581 case BPF_MAP_TYPE_PERCPU_ARRAY:
12582 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
12583 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
12584 case BPF_MAP_TYPE_HASH_OF_MAPS:
12585 if (!is_preallocated_map(map)) {
12587 "Sleepable programs can only use preallocated maps\n");
12591 case BPF_MAP_TYPE_RINGBUF:
12592 case BPF_MAP_TYPE_INODE_STORAGE:
12593 case BPF_MAP_TYPE_SK_STORAGE:
12594 case BPF_MAP_TYPE_TASK_STORAGE:
12598 "Sleepable programs can only use array, hash, and ringbuf maps\n");
12605 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
12607 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
12608 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
12611 /* find and rewrite pseudo imm in ld_imm64 instructions:
12613 * 1. if it accesses map FD, replace it with actual map pointer.
12614 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
12616 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
12618 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
12620 struct bpf_insn *insn = env->prog->insnsi;
12621 int insn_cnt = env->prog->len;
12624 err = bpf_prog_calc_tag(env->prog);
12628 for (i = 0; i < insn_cnt; i++, insn++) {
12629 if (BPF_CLASS(insn->code) == BPF_LDX &&
12630 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
12631 verbose(env, "BPF_LDX uses reserved fields\n");
12635 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
12636 struct bpf_insn_aux_data *aux;
12637 struct bpf_map *map;
12642 if (i == insn_cnt - 1 || insn[1].code != 0 ||
12643 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
12644 insn[1].off != 0) {
12645 verbose(env, "invalid bpf_ld_imm64 insn\n");
12649 if (insn[0].src_reg == 0)
12650 /* valid generic load 64-bit imm */
12653 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
12654 aux = &env->insn_aux_data[i];
12655 err = check_pseudo_btf_id(env, insn, aux);
12661 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
12662 aux = &env->insn_aux_data[i];
12663 aux->ptr_type = PTR_TO_FUNC;
12667 /* In final convert_pseudo_ld_imm64() step, this is
12668 * converted into regular 64-bit imm load insn.
12670 switch (insn[0].src_reg) {
12671 case BPF_PSEUDO_MAP_VALUE:
12672 case BPF_PSEUDO_MAP_IDX_VALUE:
12674 case BPF_PSEUDO_MAP_FD:
12675 case BPF_PSEUDO_MAP_IDX:
12676 if (insn[1].imm == 0)
12680 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
12684 switch (insn[0].src_reg) {
12685 case BPF_PSEUDO_MAP_IDX_VALUE:
12686 case BPF_PSEUDO_MAP_IDX:
12687 if (bpfptr_is_null(env->fd_array)) {
12688 verbose(env, "fd_idx without fd_array is invalid\n");
12691 if (copy_from_bpfptr_offset(&fd, env->fd_array,
12692 insn[0].imm * sizeof(fd),
12702 map = __bpf_map_get(f);
12704 verbose(env, "fd %d is not pointing to valid bpf_map\n",
12706 return PTR_ERR(map);
12709 err = check_map_prog_compatibility(env, map, env->prog);
12715 aux = &env->insn_aux_data[i];
12716 if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
12717 insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
12718 addr = (unsigned long)map;
12720 u32 off = insn[1].imm;
12722 if (off >= BPF_MAX_VAR_OFF) {
12723 verbose(env, "direct value offset of %u is not allowed\n", off);
12728 if (!map->ops->map_direct_value_addr) {
12729 verbose(env, "no direct value access support for this map type\n");
12734 err = map->ops->map_direct_value_addr(map, &addr, off);
12736 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
12737 map->value_size, off);
12742 aux->map_off = off;
12746 insn[0].imm = (u32)addr;
12747 insn[1].imm = addr >> 32;
12749 /* check whether we recorded this map already */
12750 for (j = 0; j < env->used_map_cnt; j++) {
12751 if (env->used_maps[j] == map) {
12752 aux->map_index = j;
12758 if (env->used_map_cnt >= MAX_USED_MAPS) {
12763 /* hold the map. If the program is rejected by verifier,
12764 * the map will be released by release_maps() or it
12765 * will be used by the valid program until it's unloaded
12766 * and all maps are released in free_used_maps()
12770 aux->map_index = env->used_map_cnt;
12771 env->used_maps[env->used_map_cnt++] = map;
12773 if (bpf_map_is_cgroup_storage(map) &&
12774 bpf_cgroup_storage_assign(env->prog->aux, map)) {
12775 verbose(env, "only one cgroup storage of each type is allowed\n");
12787 /* Basic sanity check before we invest more work here. */
12788 if (!bpf_opcode_in_insntable(insn->code)) {
12789 verbose(env, "unknown opcode %02x\n", insn->code);
12794 /* now all pseudo BPF_LD_IMM64 instructions load valid
12795 * 'struct bpf_map *' into a register instead of user map_fd.
12796 * These pointers will be used later by verifier to validate map access.
12801 /* drop refcnt of maps used by the rejected program */
12802 static void release_maps(struct bpf_verifier_env *env)
12804 __bpf_free_used_maps(env->prog->aux, env->used_maps,
12805 env->used_map_cnt);
12808 /* drop refcnt of maps used by the rejected program */
12809 static void release_btfs(struct bpf_verifier_env *env)
12811 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
12812 env->used_btf_cnt);
12815 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
12816 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
12818 struct bpf_insn *insn = env->prog->insnsi;
12819 int insn_cnt = env->prog->len;
12822 for (i = 0; i < insn_cnt; i++, insn++) {
12823 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
12825 if (insn->src_reg == BPF_PSEUDO_FUNC)
12831 /* single env->prog->insni[off] instruction was replaced with the range
12832 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
12833 * [0, off) and [off, end) to new locations, so the patched range stays zero
12835 static void adjust_insn_aux_data(struct bpf_verifier_env *env,
12836 struct bpf_insn_aux_data *new_data,
12837 struct bpf_prog *new_prog, u32 off, u32 cnt)
12839 struct bpf_insn_aux_data *old_data = env->insn_aux_data;
12840 struct bpf_insn *insn = new_prog->insnsi;
12841 u32 old_seen = old_data[off].seen;
12845 /* aux info at OFF always needs adjustment, no matter fast path
12846 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
12847 * original insn at old prog.
12849 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
12853 prog_len = new_prog->len;
12855 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
12856 memcpy(new_data + off + cnt - 1, old_data + off,
12857 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
12858 for (i = off; i < off + cnt - 1; i++) {
12859 /* Expand insni[off]'s seen count to the patched range. */
12860 new_data[i].seen = old_seen;
12861 new_data[i].zext_dst = insn_has_def32(env, insn + i);
12863 env->insn_aux_data = new_data;
12867 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
12873 /* NOTE: fake 'exit' subprog should be updated as well. */
12874 for (i = 0; i <= env->subprog_cnt; i++) {
12875 if (env->subprog_info[i].start <= off)
12877 env->subprog_info[i].start += len - 1;
12881 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
12883 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
12884 int i, sz = prog->aux->size_poke_tab;
12885 struct bpf_jit_poke_descriptor *desc;
12887 for (i = 0; i < sz; i++) {
12889 if (desc->insn_idx <= off)
12891 desc->insn_idx += len - 1;
12895 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
12896 const struct bpf_insn *patch, u32 len)
12898 struct bpf_prog *new_prog;
12899 struct bpf_insn_aux_data *new_data = NULL;
12902 new_data = vzalloc(array_size(env->prog->len + len - 1,
12903 sizeof(struct bpf_insn_aux_data)));
12908 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
12909 if (IS_ERR(new_prog)) {
12910 if (PTR_ERR(new_prog) == -ERANGE)
12912 "insn %d cannot be patched due to 16-bit range\n",
12913 env->insn_aux_data[off].orig_idx);
12917 adjust_insn_aux_data(env, new_data, new_prog, off, len);
12918 adjust_subprog_starts(env, off, len);
12919 adjust_poke_descs(new_prog, off, len);
12923 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
12928 /* find first prog starting at or after off (first to remove) */
12929 for (i = 0; i < env->subprog_cnt; i++)
12930 if (env->subprog_info[i].start >= off)
12932 /* find first prog starting at or after off + cnt (first to stay) */
12933 for (j = i; j < env->subprog_cnt; j++)
12934 if (env->subprog_info[j].start >= off + cnt)
12936 /* if j doesn't start exactly at off + cnt, we are just removing
12937 * the front of previous prog
12939 if (env->subprog_info[j].start != off + cnt)
12943 struct bpf_prog_aux *aux = env->prog->aux;
12946 /* move fake 'exit' subprog as well */
12947 move = env->subprog_cnt + 1 - j;
12949 memmove(env->subprog_info + i,
12950 env->subprog_info + j,
12951 sizeof(*env->subprog_info) * move);
12952 env->subprog_cnt -= j - i;
12954 /* remove func_info */
12955 if (aux->func_info) {
12956 move = aux->func_info_cnt - j;
12958 memmove(aux->func_info + i,
12959 aux->func_info + j,
12960 sizeof(*aux->func_info) * move);
12961 aux->func_info_cnt -= j - i;
12962 /* func_info->insn_off is set after all code rewrites,
12963 * in adjust_btf_func() - no need to adjust
12967 /* convert i from "first prog to remove" to "first to adjust" */
12968 if (env->subprog_info[i].start == off)
12972 /* update fake 'exit' subprog as well */
12973 for (; i <= env->subprog_cnt; i++)
12974 env->subprog_info[i].start -= cnt;
12979 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
12982 struct bpf_prog *prog = env->prog;
12983 u32 i, l_off, l_cnt, nr_linfo;
12984 struct bpf_line_info *linfo;
12986 nr_linfo = prog->aux->nr_linfo;
12990 linfo = prog->aux->linfo;
12992 /* find first line info to remove, count lines to be removed */
12993 for (i = 0; i < nr_linfo; i++)
12994 if (linfo[i].insn_off >= off)
12999 for (; i < nr_linfo; i++)
13000 if (linfo[i].insn_off < off + cnt)
13005 /* First live insn doesn't match first live linfo, it needs to "inherit"
13006 * last removed linfo. prog is already modified, so prog->len == off
13007 * means no live instructions after (tail of the program was removed).
13009 if (prog->len != off && l_cnt &&
13010 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
13012 linfo[--i].insn_off = off + cnt;
13015 /* remove the line info which refer to the removed instructions */
13017 memmove(linfo + l_off, linfo + i,
13018 sizeof(*linfo) * (nr_linfo - i));
13020 prog->aux->nr_linfo -= l_cnt;
13021 nr_linfo = prog->aux->nr_linfo;
13024 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
13025 for (i = l_off; i < nr_linfo; i++)
13026 linfo[i].insn_off -= cnt;
13028 /* fix up all subprogs (incl. 'exit') which start >= off */
13029 for (i = 0; i <= env->subprog_cnt; i++)
13030 if (env->subprog_info[i].linfo_idx > l_off) {
13031 /* program may have started in the removed region but
13032 * may not be fully removed
13034 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
13035 env->subprog_info[i].linfo_idx -= l_cnt;
13037 env->subprog_info[i].linfo_idx = l_off;
13043 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
13045 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
13046 unsigned int orig_prog_len = env->prog->len;
13049 if (bpf_prog_is_dev_bound(env->prog->aux))
13050 bpf_prog_offload_remove_insns(env, off, cnt);
13052 err = bpf_remove_insns(env->prog, off, cnt);
13056 err = adjust_subprog_starts_after_remove(env, off, cnt);
13060 err = bpf_adj_linfo_after_remove(env, off, cnt);
13064 memmove(aux_data + off, aux_data + off + cnt,
13065 sizeof(*aux_data) * (orig_prog_len - off - cnt));
13070 /* The verifier does more data flow analysis than llvm and will not
13071 * explore branches that are dead at run time. Malicious programs can
13072 * have dead code too. Therefore replace all dead at-run-time code
13075 * Just nops are not optimal, e.g. if they would sit at the end of the
13076 * program and through another bug we would manage to jump there, then
13077 * we'd execute beyond program memory otherwise. Returning exception
13078 * code also wouldn't work since we can have subprogs where the dead
13079 * code could be located.
13081 static void sanitize_dead_code(struct bpf_verifier_env *env)
13083 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
13084 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
13085 struct bpf_insn *insn = env->prog->insnsi;
13086 const int insn_cnt = env->prog->len;
13089 for (i = 0; i < insn_cnt; i++) {
13090 if (aux_data[i].seen)
13092 memcpy(insn + i, &trap, sizeof(trap));
13093 aux_data[i].zext_dst = false;
13097 static bool insn_is_cond_jump(u8 code)
13101 if (BPF_CLASS(code) == BPF_JMP32)
13104 if (BPF_CLASS(code) != BPF_JMP)
13108 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
13111 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
13113 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
13114 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
13115 struct bpf_insn *insn = env->prog->insnsi;
13116 const int insn_cnt = env->prog->len;
13119 for (i = 0; i < insn_cnt; i++, insn++) {
13120 if (!insn_is_cond_jump(insn->code))
13123 if (!aux_data[i + 1].seen)
13124 ja.off = insn->off;
13125 else if (!aux_data[i + 1 + insn->off].seen)
13130 if (bpf_prog_is_dev_bound(env->prog->aux))
13131 bpf_prog_offload_replace_insn(env, i, &ja);
13133 memcpy(insn, &ja, sizeof(ja));
13137 static int opt_remove_dead_code(struct bpf_verifier_env *env)
13139 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
13140 int insn_cnt = env->prog->len;
13143 for (i = 0; i < insn_cnt; i++) {
13147 while (i + j < insn_cnt && !aux_data[i + j].seen)
13152 err = verifier_remove_insns(env, i, j);
13155 insn_cnt = env->prog->len;
13161 static int opt_remove_nops(struct bpf_verifier_env *env)
13163 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
13164 struct bpf_insn *insn = env->prog->insnsi;
13165 int insn_cnt = env->prog->len;
13168 for (i = 0; i < insn_cnt; i++) {
13169 if (memcmp(&insn[i], &ja, sizeof(ja)))
13172 err = verifier_remove_insns(env, i, 1);
13182 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
13183 const union bpf_attr *attr)
13185 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
13186 struct bpf_insn_aux_data *aux = env->insn_aux_data;
13187 int i, patch_len, delta = 0, len = env->prog->len;
13188 struct bpf_insn *insns = env->prog->insnsi;
13189 struct bpf_prog *new_prog;
13192 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
13193 zext_patch[1] = BPF_ZEXT_REG(0);
13194 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
13195 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
13196 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
13197 for (i = 0; i < len; i++) {
13198 int adj_idx = i + delta;
13199 struct bpf_insn insn;
13202 insn = insns[adj_idx];
13203 load_reg = insn_def_regno(&insn);
13204 if (!aux[adj_idx].zext_dst) {
13212 class = BPF_CLASS(code);
13213 if (load_reg == -1)
13216 /* NOTE: arg "reg" (the fourth one) is only used for
13217 * BPF_STX + SRC_OP, so it is safe to pass NULL
13220 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
13221 if (class == BPF_LD &&
13222 BPF_MODE(code) == BPF_IMM)
13227 /* ctx load could be transformed into wider load. */
13228 if (class == BPF_LDX &&
13229 aux[adj_idx].ptr_type == PTR_TO_CTX)
13232 imm_rnd = get_random_int();
13233 rnd_hi32_patch[0] = insn;
13234 rnd_hi32_patch[1].imm = imm_rnd;
13235 rnd_hi32_patch[3].dst_reg = load_reg;
13236 patch = rnd_hi32_patch;
13238 goto apply_patch_buffer;
13241 /* Add in an zero-extend instruction if a) the JIT has requested
13242 * it or b) it's a CMPXCHG.
13244 * The latter is because: BPF_CMPXCHG always loads a value into
13245 * R0, therefore always zero-extends. However some archs'
13246 * equivalent instruction only does this load when the
13247 * comparison is successful. This detail of CMPXCHG is
13248 * orthogonal to the general zero-extension behaviour of the
13249 * CPU, so it's treated independently of bpf_jit_needs_zext.
13251 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
13254 if (WARN_ON(load_reg == -1)) {
13255 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
13259 zext_patch[0] = insn;
13260 zext_patch[1].dst_reg = load_reg;
13261 zext_patch[1].src_reg = load_reg;
13262 patch = zext_patch;
13264 apply_patch_buffer:
13265 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
13268 env->prog = new_prog;
13269 insns = new_prog->insnsi;
13270 aux = env->insn_aux_data;
13271 delta += patch_len - 1;
13277 /* convert load instructions that access fields of a context type into a
13278 * sequence of instructions that access fields of the underlying structure:
13279 * struct __sk_buff -> struct sk_buff
13280 * struct bpf_sock_ops -> struct sock
13282 static int convert_ctx_accesses(struct bpf_verifier_env *env)
13284 const struct bpf_verifier_ops *ops = env->ops;
13285 int i, cnt, size, ctx_field_size, delta = 0;
13286 const int insn_cnt = env->prog->len;
13287 struct bpf_insn insn_buf[16], *insn;
13288 u32 target_size, size_default, off;
13289 struct bpf_prog *new_prog;
13290 enum bpf_access_type type;
13291 bool is_narrower_load;
13293 if (ops->gen_prologue || env->seen_direct_write) {
13294 if (!ops->gen_prologue) {
13295 verbose(env, "bpf verifier is misconfigured\n");
13298 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
13300 if (cnt >= ARRAY_SIZE(insn_buf)) {
13301 verbose(env, "bpf verifier is misconfigured\n");
13304 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
13308 env->prog = new_prog;
13313 if (bpf_prog_is_dev_bound(env->prog->aux))
13316 insn = env->prog->insnsi + delta;
13318 for (i = 0; i < insn_cnt; i++, insn++) {
13319 bpf_convert_ctx_access_t convert_ctx_access;
13322 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
13323 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
13324 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
13325 insn->code == (BPF_LDX | BPF_MEM | BPF_DW)) {
13328 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
13329 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
13330 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
13331 insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
13332 insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
13333 insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
13334 insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
13335 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
13337 ctx_access = BPF_CLASS(insn->code) == BPF_STX;
13342 if (type == BPF_WRITE &&
13343 env->insn_aux_data[i + delta].sanitize_stack_spill) {
13344 struct bpf_insn patch[] = {
13349 cnt = ARRAY_SIZE(patch);
13350 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
13355 env->prog = new_prog;
13356 insn = new_prog->insnsi + i + delta;
13363 switch ((int)env->insn_aux_data[i + delta].ptr_type) {
13365 if (!ops->convert_ctx_access)
13367 convert_ctx_access = ops->convert_ctx_access;
13369 case PTR_TO_SOCKET:
13370 case PTR_TO_SOCK_COMMON:
13371 convert_ctx_access = bpf_sock_convert_ctx_access;
13373 case PTR_TO_TCP_SOCK:
13374 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
13376 case PTR_TO_XDP_SOCK:
13377 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
13379 case PTR_TO_BTF_ID:
13380 case PTR_TO_BTF_ID | PTR_UNTRUSTED:
13381 if (type == BPF_READ) {
13382 insn->code = BPF_LDX | BPF_PROBE_MEM |
13383 BPF_SIZE((insn)->code);
13384 env->prog->aux->num_exentries++;
13385 } else if (resolve_prog_type(env->prog) != BPF_PROG_TYPE_STRUCT_OPS) {
13386 verbose(env, "Writes through BTF pointers are not allowed\n");
13394 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
13395 size = BPF_LDST_BYTES(insn);
13397 /* If the read access is a narrower load of the field,
13398 * convert to a 4/8-byte load, to minimum program type specific
13399 * convert_ctx_access changes. If conversion is successful,
13400 * we will apply proper mask to the result.
13402 is_narrower_load = size < ctx_field_size;
13403 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
13405 if (is_narrower_load) {
13408 if (type == BPF_WRITE) {
13409 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
13414 if (ctx_field_size == 4)
13416 else if (ctx_field_size == 8)
13417 size_code = BPF_DW;
13419 insn->off = off & ~(size_default - 1);
13420 insn->code = BPF_LDX | BPF_MEM | size_code;
13424 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
13426 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
13427 (ctx_field_size && !target_size)) {
13428 verbose(env, "bpf verifier is misconfigured\n");
13432 if (is_narrower_load && size < target_size) {
13433 u8 shift = bpf_ctx_narrow_access_offset(
13434 off, size, size_default) * 8;
13435 if (shift && cnt + 1 >= ARRAY_SIZE(insn_buf)) {
13436 verbose(env, "bpf verifier narrow ctx load misconfigured\n");
13439 if (ctx_field_size <= 4) {
13441 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
13444 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
13445 (1 << size * 8) - 1);
13448 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
13451 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
13452 (1ULL << size * 8) - 1);
13456 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
13462 /* keep walking new program and skip insns we just inserted */
13463 env->prog = new_prog;
13464 insn = new_prog->insnsi + i + delta;
13470 static int jit_subprogs(struct bpf_verifier_env *env)
13472 struct bpf_prog *prog = env->prog, **func, *tmp;
13473 int i, j, subprog_start, subprog_end = 0, len, subprog;
13474 struct bpf_map *map_ptr;
13475 struct bpf_insn *insn;
13476 void *old_bpf_func;
13477 int err, num_exentries;
13479 if (env->subprog_cnt <= 1)
13482 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
13483 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn))
13486 /* Upon error here we cannot fall back to interpreter but
13487 * need a hard reject of the program. Thus -EFAULT is
13488 * propagated in any case.
13490 subprog = find_subprog(env, i + insn->imm + 1);
13492 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
13493 i + insn->imm + 1);
13496 /* temporarily remember subprog id inside insn instead of
13497 * aux_data, since next loop will split up all insns into funcs
13499 insn->off = subprog;
13500 /* remember original imm in case JIT fails and fallback
13501 * to interpreter will be needed
13503 env->insn_aux_data[i].call_imm = insn->imm;
13504 /* point imm to __bpf_call_base+1 from JITs point of view */
13506 if (bpf_pseudo_func(insn))
13507 /* jit (e.g. x86_64) may emit fewer instructions
13508 * if it learns a u32 imm is the same as a u64 imm.
13509 * Force a non zero here.
13514 err = bpf_prog_alloc_jited_linfo(prog);
13516 goto out_undo_insn;
13519 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
13521 goto out_undo_insn;
13523 for (i = 0; i < env->subprog_cnt; i++) {
13524 subprog_start = subprog_end;
13525 subprog_end = env->subprog_info[i + 1].start;
13527 len = subprog_end - subprog_start;
13528 /* bpf_prog_run() doesn't call subprogs directly,
13529 * hence main prog stats include the runtime of subprogs.
13530 * subprogs don't have IDs and not reachable via prog_get_next_id
13531 * func[i]->stats will never be accessed and stays NULL
13533 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
13536 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
13537 len * sizeof(struct bpf_insn));
13538 func[i]->type = prog->type;
13539 func[i]->len = len;
13540 if (bpf_prog_calc_tag(func[i]))
13542 func[i]->is_func = 1;
13543 func[i]->aux->func_idx = i;
13544 /* Below members will be freed only at prog->aux */
13545 func[i]->aux->btf = prog->aux->btf;
13546 func[i]->aux->func_info = prog->aux->func_info;
13547 func[i]->aux->poke_tab = prog->aux->poke_tab;
13548 func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
13550 for (j = 0; j < prog->aux->size_poke_tab; j++) {
13551 struct bpf_jit_poke_descriptor *poke;
13553 poke = &prog->aux->poke_tab[j];
13554 if (poke->insn_idx < subprog_end &&
13555 poke->insn_idx >= subprog_start)
13556 poke->aux = func[i]->aux;
13559 /* Use bpf_prog_F_tag to indicate functions in stack traces.
13560 * Long term would need debug info to populate names
13562 func[i]->aux->name[0] = 'F';
13563 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
13564 func[i]->jit_requested = 1;
13565 func[i]->blinding_requested = prog->blinding_requested;
13566 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
13567 func[i]->aux->kfunc_btf_tab = prog->aux->kfunc_btf_tab;
13568 func[i]->aux->linfo = prog->aux->linfo;
13569 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
13570 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
13571 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
13573 insn = func[i]->insnsi;
13574 for (j = 0; j < func[i]->len; j++, insn++) {
13575 if (BPF_CLASS(insn->code) == BPF_LDX &&
13576 BPF_MODE(insn->code) == BPF_PROBE_MEM)
13579 func[i]->aux->num_exentries = num_exentries;
13580 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
13581 func[i] = bpf_int_jit_compile(func[i]);
13582 if (!func[i]->jited) {
13589 /* at this point all bpf functions were successfully JITed
13590 * now populate all bpf_calls with correct addresses and
13591 * run last pass of JIT
13593 for (i = 0; i < env->subprog_cnt; i++) {
13594 insn = func[i]->insnsi;
13595 for (j = 0; j < func[i]->len; j++, insn++) {
13596 if (bpf_pseudo_func(insn)) {
13597 subprog = insn->off;
13598 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
13599 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
13602 if (!bpf_pseudo_call(insn))
13604 subprog = insn->off;
13605 insn->imm = BPF_CALL_IMM(func[subprog]->bpf_func);
13608 /* we use the aux data to keep a list of the start addresses
13609 * of the JITed images for each function in the program
13611 * for some architectures, such as powerpc64, the imm field
13612 * might not be large enough to hold the offset of the start
13613 * address of the callee's JITed image from __bpf_call_base
13615 * in such cases, we can lookup the start address of a callee
13616 * by using its subprog id, available from the off field of
13617 * the call instruction, as an index for this list
13619 func[i]->aux->func = func;
13620 func[i]->aux->func_cnt = env->subprog_cnt;
13622 for (i = 0; i < env->subprog_cnt; i++) {
13623 old_bpf_func = func[i]->bpf_func;
13624 tmp = bpf_int_jit_compile(func[i]);
13625 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
13626 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
13633 /* finally lock prog and jit images for all functions and
13634 * populate kallsysm
13636 for (i = 0; i < env->subprog_cnt; i++) {
13637 bpf_prog_lock_ro(func[i]);
13638 bpf_prog_kallsyms_add(func[i]);
13641 /* Last step: make now unused interpreter insns from main
13642 * prog consistent for later dump requests, so they can
13643 * later look the same as if they were interpreted only.
13645 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
13646 if (bpf_pseudo_func(insn)) {
13647 insn[0].imm = env->insn_aux_data[i].call_imm;
13648 insn[1].imm = insn->off;
13652 if (!bpf_pseudo_call(insn))
13654 insn->off = env->insn_aux_data[i].call_imm;
13655 subprog = find_subprog(env, i + insn->off + 1);
13656 insn->imm = subprog;
13660 prog->bpf_func = func[0]->bpf_func;
13661 prog->jited_len = func[0]->jited_len;
13662 prog->aux->func = func;
13663 prog->aux->func_cnt = env->subprog_cnt;
13664 bpf_prog_jit_attempt_done(prog);
13667 /* We failed JIT'ing, so at this point we need to unregister poke
13668 * descriptors from subprogs, so that kernel is not attempting to
13669 * patch it anymore as we're freeing the subprog JIT memory.
13671 for (i = 0; i < prog->aux->size_poke_tab; i++) {
13672 map_ptr = prog->aux->poke_tab[i].tail_call.map;
13673 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
13675 /* At this point we're guaranteed that poke descriptors are not
13676 * live anymore. We can just unlink its descriptor table as it's
13677 * released with the main prog.
13679 for (i = 0; i < env->subprog_cnt; i++) {
13682 func[i]->aux->poke_tab = NULL;
13683 bpf_jit_free(func[i]);
13687 /* cleanup main prog to be interpreted */
13688 prog->jit_requested = 0;
13689 prog->blinding_requested = 0;
13690 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
13691 if (!bpf_pseudo_call(insn))
13694 insn->imm = env->insn_aux_data[i].call_imm;
13696 bpf_prog_jit_attempt_done(prog);
13700 static int fixup_call_args(struct bpf_verifier_env *env)
13702 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
13703 struct bpf_prog *prog = env->prog;
13704 struct bpf_insn *insn = prog->insnsi;
13705 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
13710 if (env->prog->jit_requested &&
13711 !bpf_prog_is_dev_bound(env->prog->aux)) {
13712 err = jit_subprogs(env);
13715 if (err == -EFAULT)
13718 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
13719 if (has_kfunc_call) {
13720 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
13723 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
13724 /* When JIT fails the progs with bpf2bpf calls and tail_calls
13725 * have to be rejected, since interpreter doesn't support them yet.
13727 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
13730 for (i = 0; i < prog->len; i++, insn++) {
13731 if (bpf_pseudo_func(insn)) {
13732 /* When JIT fails the progs with callback calls
13733 * have to be rejected, since interpreter doesn't support them yet.
13735 verbose(env, "callbacks are not allowed in non-JITed programs\n");
13739 if (!bpf_pseudo_call(insn))
13741 depth = get_callee_stack_depth(env, insn, i);
13744 bpf_patch_call_args(insn, depth);
13751 static int fixup_kfunc_call(struct bpf_verifier_env *env,
13752 struct bpf_insn *insn)
13754 const struct bpf_kfunc_desc *desc;
13757 verbose(env, "invalid kernel function call not eliminated in verifier pass\n");
13761 /* insn->imm has the btf func_id. Replace it with
13762 * an address (relative to __bpf_base_call).
13764 desc = find_kfunc_desc(env->prog, insn->imm, insn->off);
13766 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
13771 insn->imm = desc->imm;
13776 /* Do various post-verification rewrites in a single program pass.
13777 * These rewrites simplify JIT and interpreter implementations.
13779 static int do_misc_fixups(struct bpf_verifier_env *env)
13781 struct bpf_prog *prog = env->prog;
13782 enum bpf_attach_type eatype = prog->expected_attach_type;
13783 enum bpf_prog_type prog_type = resolve_prog_type(prog);
13784 struct bpf_insn *insn = prog->insnsi;
13785 const struct bpf_func_proto *fn;
13786 const int insn_cnt = prog->len;
13787 const struct bpf_map_ops *ops;
13788 struct bpf_insn_aux_data *aux;
13789 struct bpf_insn insn_buf[16];
13790 struct bpf_prog *new_prog;
13791 struct bpf_map *map_ptr;
13792 int i, ret, cnt, delta = 0;
13794 for (i = 0; i < insn_cnt; i++, insn++) {
13795 /* Make divide-by-zero exceptions impossible. */
13796 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
13797 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
13798 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
13799 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
13800 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
13801 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
13802 struct bpf_insn *patchlet;
13803 struct bpf_insn chk_and_div[] = {
13804 /* [R,W]x div 0 -> 0 */
13805 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
13806 BPF_JNE | BPF_K, insn->src_reg,
13808 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
13809 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
13812 struct bpf_insn chk_and_mod[] = {
13813 /* [R,W]x mod 0 -> [R,W]x */
13814 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
13815 BPF_JEQ | BPF_K, insn->src_reg,
13816 0, 1 + (is64 ? 0 : 1), 0),
13818 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
13819 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
13822 patchlet = isdiv ? chk_and_div : chk_and_mod;
13823 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
13824 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
13826 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
13831 env->prog = prog = new_prog;
13832 insn = new_prog->insnsi + i + delta;
13836 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
13837 if (BPF_CLASS(insn->code) == BPF_LD &&
13838 (BPF_MODE(insn->code) == BPF_ABS ||
13839 BPF_MODE(insn->code) == BPF_IND)) {
13840 cnt = env->ops->gen_ld_abs(insn, insn_buf);
13841 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
13842 verbose(env, "bpf verifier is misconfigured\n");
13846 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
13851 env->prog = prog = new_prog;
13852 insn = new_prog->insnsi + i + delta;
13856 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
13857 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
13858 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
13859 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
13860 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
13861 struct bpf_insn *patch = &insn_buf[0];
13862 bool issrc, isneg, isimm;
13865 aux = &env->insn_aux_data[i + delta];
13866 if (!aux->alu_state ||
13867 aux->alu_state == BPF_ALU_NON_POINTER)
13870 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
13871 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
13872 BPF_ALU_SANITIZE_SRC;
13873 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
13875 off_reg = issrc ? insn->src_reg : insn->dst_reg;
13877 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
13880 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
13881 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
13882 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
13883 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
13884 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
13885 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
13886 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
13889 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
13890 insn->src_reg = BPF_REG_AX;
13892 insn->code = insn->code == code_add ?
13893 code_sub : code_add;
13895 if (issrc && isneg && !isimm)
13896 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
13897 cnt = patch - insn_buf;
13899 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
13904 env->prog = prog = new_prog;
13905 insn = new_prog->insnsi + i + delta;
13909 if (insn->code != (BPF_JMP | BPF_CALL))
13911 if (insn->src_reg == BPF_PSEUDO_CALL)
13913 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
13914 ret = fixup_kfunc_call(env, insn);
13920 if (insn->imm == BPF_FUNC_get_route_realm)
13921 prog->dst_needed = 1;
13922 if (insn->imm == BPF_FUNC_get_prandom_u32)
13923 bpf_user_rnd_init_once();
13924 if (insn->imm == BPF_FUNC_override_return)
13925 prog->kprobe_override = 1;
13926 if (insn->imm == BPF_FUNC_tail_call) {
13927 /* If we tail call into other programs, we
13928 * cannot make any assumptions since they can
13929 * be replaced dynamically during runtime in
13930 * the program array.
13932 prog->cb_access = 1;
13933 if (!allow_tail_call_in_subprogs(env))
13934 prog->aux->stack_depth = MAX_BPF_STACK;
13935 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
13937 /* mark bpf_tail_call as different opcode to avoid
13938 * conditional branch in the interpreter for every normal
13939 * call and to prevent accidental JITing by JIT compiler
13940 * that doesn't support bpf_tail_call yet
13943 insn->code = BPF_JMP | BPF_TAIL_CALL;
13945 aux = &env->insn_aux_data[i + delta];
13946 if (env->bpf_capable && !prog->blinding_requested &&
13947 prog->jit_requested &&
13948 !bpf_map_key_poisoned(aux) &&
13949 !bpf_map_ptr_poisoned(aux) &&
13950 !bpf_map_ptr_unpriv(aux)) {
13951 struct bpf_jit_poke_descriptor desc = {
13952 .reason = BPF_POKE_REASON_TAIL_CALL,
13953 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
13954 .tail_call.key = bpf_map_key_immediate(aux),
13955 .insn_idx = i + delta,
13958 ret = bpf_jit_add_poke_descriptor(prog, &desc);
13960 verbose(env, "adding tail call poke descriptor failed\n");
13964 insn->imm = ret + 1;
13968 if (!bpf_map_ptr_unpriv(aux))
13971 /* instead of changing every JIT dealing with tail_call
13972 * emit two extra insns:
13973 * if (index >= max_entries) goto out;
13974 * index &= array->index_mask;
13975 * to avoid out-of-bounds cpu speculation
13977 if (bpf_map_ptr_poisoned(aux)) {
13978 verbose(env, "tail_call abusing map_ptr\n");
13982 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
13983 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
13984 map_ptr->max_entries, 2);
13985 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
13986 container_of(map_ptr,
13989 insn_buf[2] = *insn;
13991 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
13996 env->prog = prog = new_prog;
13997 insn = new_prog->insnsi + i + delta;
14001 if (insn->imm == BPF_FUNC_timer_set_callback) {
14002 /* The verifier will process callback_fn as many times as necessary
14003 * with different maps and the register states prepared by
14004 * set_timer_callback_state will be accurate.
14006 * The following use case is valid:
14007 * map1 is shared by prog1, prog2, prog3.
14008 * prog1 calls bpf_timer_init for some map1 elements
14009 * prog2 calls bpf_timer_set_callback for some map1 elements.
14010 * Those that were not bpf_timer_init-ed will return -EINVAL.
14011 * prog3 calls bpf_timer_start for some map1 elements.
14012 * Those that were not both bpf_timer_init-ed and
14013 * bpf_timer_set_callback-ed will return -EINVAL.
14015 struct bpf_insn ld_addrs[2] = {
14016 BPF_LD_IMM64(BPF_REG_3, (long)prog->aux),
14019 insn_buf[0] = ld_addrs[0];
14020 insn_buf[1] = ld_addrs[1];
14021 insn_buf[2] = *insn;
14024 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
14029 env->prog = prog = new_prog;
14030 insn = new_prog->insnsi + i + delta;
14031 goto patch_call_imm;
14034 if (insn->imm == BPF_FUNC_task_storage_get ||
14035 insn->imm == BPF_FUNC_sk_storage_get ||
14036 insn->imm == BPF_FUNC_inode_storage_get) {
14037 if (env->prog->aux->sleepable)
14038 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_KERNEL);
14040 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_ATOMIC);
14041 insn_buf[1] = *insn;
14044 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
14049 env->prog = prog = new_prog;
14050 insn = new_prog->insnsi + i + delta;
14051 goto patch_call_imm;
14054 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
14055 * and other inlining handlers are currently limited to 64 bit
14058 if (prog->jit_requested && BITS_PER_LONG == 64 &&
14059 (insn->imm == BPF_FUNC_map_lookup_elem ||
14060 insn->imm == BPF_FUNC_map_update_elem ||
14061 insn->imm == BPF_FUNC_map_delete_elem ||
14062 insn->imm == BPF_FUNC_map_push_elem ||
14063 insn->imm == BPF_FUNC_map_pop_elem ||
14064 insn->imm == BPF_FUNC_map_peek_elem ||
14065 insn->imm == BPF_FUNC_redirect_map ||
14066 insn->imm == BPF_FUNC_for_each_map_elem ||
14067 insn->imm == BPF_FUNC_map_lookup_percpu_elem)) {
14068 aux = &env->insn_aux_data[i + delta];
14069 if (bpf_map_ptr_poisoned(aux))
14070 goto patch_call_imm;
14072 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
14073 ops = map_ptr->ops;
14074 if (insn->imm == BPF_FUNC_map_lookup_elem &&
14075 ops->map_gen_lookup) {
14076 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
14077 if (cnt == -EOPNOTSUPP)
14078 goto patch_map_ops_generic;
14079 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
14080 verbose(env, "bpf verifier is misconfigured\n");
14084 new_prog = bpf_patch_insn_data(env, i + delta,
14090 env->prog = prog = new_prog;
14091 insn = new_prog->insnsi + i + delta;
14095 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
14096 (void *(*)(struct bpf_map *map, void *key))NULL));
14097 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
14098 (int (*)(struct bpf_map *map, void *key))NULL));
14099 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
14100 (int (*)(struct bpf_map *map, void *key, void *value,
14102 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
14103 (int (*)(struct bpf_map *map, void *value,
14105 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
14106 (int (*)(struct bpf_map *map, void *value))NULL));
14107 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
14108 (int (*)(struct bpf_map *map, void *value))NULL));
14109 BUILD_BUG_ON(!__same_type(ops->map_redirect,
14110 (int (*)(struct bpf_map *map, u32 ifindex, u64 flags))NULL));
14111 BUILD_BUG_ON(!__same_type(ops->map_for_each_callback,
14112 (int (*)(struct bpf_map *map,
14113 bpf_callback_t callback_fn,
14114 void *callback_ctx,
14116 BUILD_BUG_ON(!__same_type(ops->map_lookup_percpu_elem,
14117 (void *(*)(struct bpf_map *map, void *key, u32 cpu))NULL));
14119 patch_map_ops_generic:
14120 switch (insn->imm) {
14121 case BPF_FUNC_map_lookup_elem:
14122 insn->imm = BPF_CALL_IMM(ops->map_lookup_elem);
14124 case BPF_FUNC_map_update_elem:
14125 insn->imm = BPF_CALL_IMM(ops->map_update_elem);
14127 case BPF_FUNC_map_delete_elem:
14128 insn->imm = BPF_CALL_IMM(ops->map_delete_elem);
14130 case BPF_FUNC_map_push_elem:
14131 insn->imm = BPF_CALL_IMM(ops->map_push_elem);
14133 case BPF_FUNC_map_pop_elem:
14134 insn->imm = BPF_CALL_IMM(ops->map_pop_elem);
14136 case BPF_FUNC_map_peek_elem:
14137 insn->imm = BPF_CALL_IMM(ops->map_peek_elem);
14139 case BPF_FUNC_redirect_map:
14140 insn->imm = BPF_CALL_IMM(ops->map_redirect);
14142 case BPF_FUNC_for_each_map_elem:
14143 insn->imm = BPF_CALL_IMM(ops->map_for_each_callback);
14145 case BPF_FUNC_map_lookup_percpu_elem:
14146 insn->imm = BPF_CALL_IMM(ops->map_lookup_percpu_elem);
14150 goto patch_call_imm;
14153 /* Implement bpf_jiffies64 inline. */
14154 if (prog->jit_requested && BITS_PER_LONG == 64 &&
14155 insn->imm == BPF_FUNC_jiffies64) {
14156 struct bpf_insn ld_jiffies_addr[2] = {
14157 BPF_LD_IMM64(BPF_REG_0,
14158 (unsigned long)&jiffies),
14161 insn_buf[0] = ld_jiffies_addr[0];
14162 insn_buf[1] = ld_jiffies_addr[1];
14163 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
14167 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
14173 env->prog = prog = new_prog;
14174 insn = new_prog->insnsi + i + delta;
14178 /* Implement bpf_get_func_arg inline. */
14179 if (prog_type == BPF_PROG_TYPE_TRACING &&
14180 insn->imm == BPF_FUNC_get_func_arg) {
14181 /* Load nr_args from ctx - 8 */
14182 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
14183 insn_buf[1] = BPF_JMP32_REG(BPF_JGE, BPF_REG_2, BPF_REG_0, 6);
14184 insn_buf[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_2, 3);
14185 insn_buf[3] = BPF_ALU64_REG(BPF_ADD, BPF_REG_2, BPF_REG_1);
14186 insn_buf[4] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_2, 0);
14187 insn_buf[5] = BPF_STX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
14188 insn_buf[6] = BPF_MOV64_IMM(BPF_REG_0, 0);
14189 insn_buf[7] = BPF_JMP_A(1);
14190 insn_buf[8] = BPF_MOV64_IMM(BPF_REG_0, -EINVAL);
14193 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
14198 env->prog = prog = new_prog;
14199 insn = new_prog->insnsi + i + delta;
14203 /* Implement bpf_get_func_ret inline. */
14204 if (prog_type == BPF_PROG_TYPE_TRACING &&
14205 insn->imm == BPF_FUNC_get_func_ret) {
14206 if (eatype == BPF_TRACE_FEXIT ||
14207 eatype == BPF_MODIFY_RETURN) {
14208 /* Load nr_args from ctx - 8 */
14209 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
14210 insn_buf[1] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_0, 3);
14211 insn_buf[2] = BPF_ALU64_REG(BPF_ADD, BPF_REG_0, BPF_REG_1);
14212 insn_buf[3] = BPF_LDX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
14213 insn_buf[4] = BPF_STX_MEM(BPF_DW, BPF_REG_2, BPF_REG_3, 0);
14214 insn_buf[5] = BPF_MOV64_IMM(BPF_REG_0, 0);
14217 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, -EOPNOTSUPP);
14221 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
14226 env->prog = prog = new_prog;
14227 insn = new_prog->insnsi + i + delta;
14231 /* Implement get_func_arg_cnt inline. */
14232 if (prog_type == BPF_PROG_TYPE_TRACING &&
14233 insn->imm == BPF_FUNC_get_func_arg_cnt) {
14234 /* Load nr_args from ctx - 8 */
14235 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
14237 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
14241 env->prog = prog = new_prog;
14242 insn = new_prog->insnsi + i + delta;
14246 /* Implement bpf_get_func_ip inline. */
14247 if (prog_type == BPF_PROG_TYPE_TRACING &&
14248 insn->imm == BPF_FUNC_get_func_ip) {
14249 /* Load IP address from ctx - 16 */
14250 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -16);
14252 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
14256 env->prog = prog = new_prog;
14257 insn = new_prog->insnsi + i + delta;
14262 fn = env->ops->get_func_proto(insn->imm, env->prog);
14263 /* all functions that have prototype and verifier allowed
14264 * programs to call them, must be real in-kernel functions
14268 "kernel subsystem misconfigured func %s#%d\n",
14269 func_id_name(insn->imm), insn->imm);
14272 insn->imm = fn->func - __bpf_call_base;
14275 /* Since poke tab is now finalized, publish aux to tracker. */
14276 for (i = 0; i < prog->aux->size_poke_tab; i++) {
14277 map_ptr = prog->aux->poke_tab[i].tail_call.map;
14278 if (!map_ptr->ops->map_poke_track ||
14279 !map_ptr->ops->map_poke_untrack ||
14280 !map_ptr->ops->map_poke_run) {
14281 verbose(env, "bpf verifier is misconfigured\n");
14285 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
14287 verbose(env, "tracking tail call prog failed\n");
14292 sort_kfunc_descs_by_imm(env->prog);
14297 static void free_states(struct bpf_verifier_env *env)
14299 struct bpf_verifier_state_list *sl, *sln;
14302 sl = env->free_list;
14305 free_verifier_state(&sl->state, false);
14309 env->free_list = NULL;
14311 if (!env->explored_states)
14314 for (i = 0; i < state_htab_size(env); i++) {
14315 sl = env->explored_states[i];
14319 free_verifier_state(&sl->state, false);
14323 env->explored_states[i] = NULL;
14327 static int do_check_common(struct bpf_verifier_env *env, int subprog)
14329 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
14330 struct bpf_verifier_state *state;
14331 struct bpf_reg_state *regs;
14334 env->prev_linfo = NULL;
14337 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
14340 state->curframe = 0;
14341 state->speculative = false;
14342 state->branches = 1;
14343 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
14344 if (!state->frame[0]) {
14348 env->cur_state = state;
14349 init_func_state(env, state->frame[0],
14350 BPF_MAIN_FUNC /* callsite */,
14354 regs = state->frame[state->curframe]->regs;
14355 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
14356 ret = btf_prepare_func_args(env, subprog, regs);
14359 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
14360 if (regs[i].type == PTR_TO_CTX)
14361 mark_reg_known_zero(env, regs, i);
14362 else if (regs[i].type == SCALAR_VALUE)
14363 mark_reg_unknown(env, regs, i);
14364 else if (base_type(regs[i].type) == PTR_TO_MEM) {
14365 const u32 mem_size = regs[i].mem_size;
14367 mark_reg_known_zero(env, regs, i);
14368 regs[i].mem_size = mem_size;
14369 regs[i].id = ++env->id_gen;
14373 /* 1st arg to a function */
14374 regs[BPF_REG_1].type = PTR_TO_CTX;
14375 mark_reg_known_zero(env, regs, BPF_REG_1);
14376 ret = btf_check_subprog_arg_match(env, subprog, regs);
14377 if (ret == -EFAULT)
14378 /* unlikely verifier bug. abort.
14379 * ret == 0 and ret < 0 are sadly acceptable for
14380 * main() function due to backward compatibility.
14381 * Like socket filter program may be written as:
14382 * int bpf_prog(struct pt_regs *ctx)
14383 * and never dereference that ctx in the program.
14384 * 'struct pt_regs' is a type mismatch for socket
14385 * filter that should be using 'struct __sk_buff'.
14390 ret = do_check(env);
14392 /* check for NULL is necessary, since cur_state can be freed inside
14393 * do_check() under memory pressure.
14395 if (env->cur_state) {
14396 free_verifier_state(env->cur_state, true);
14397 env->cur_state = NULL;
14399 while (!pop_stack(env, NULL, NULL, false));
14400 if (!ret && pop_log)
14401 bpf_vlog_reset(&env->log, 0);
14406 /* Verify all global functions in a BPF program one by one based on their BTF.
14407 * All global functions must pass verification. Otherwise the whole program is rejected.
14418 * foo() will be verified first for R1=any_scalar_value. During verification it
14419 * will be assumed that bar() already verified successfully and call to bar()
14420 * from foo() will be checked for type match only. Later bar() will be verified
14421 * independently to check that it's safe for R1=any_scalar_value.
14423 static int do_check_subprogs(struct bpf_verifier_env *env)
14425 struct bpf_prog_aux *aux = env->prog->aux;
14428 if (!aux->func_info)
14431 for (i = 1; i < env->subprog_cnt; i++) {
14432 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
14434 env->insn_idx = env->subprog_info[i].start;
14435 WARN_ON_ONCE(env->insn_idx == 0);
14436 ret = do_check_common(env, i);
14439 } else if (env->log.level & BPF_LOG_LEVEL) {
14441 "Func#%d is safe for any args that match its prototype\n",
14448 static int do_check_main(struct bpf_verifier_env *env)
14453 ret = do_check_common(env, 0);
14455 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
14460 static void print_verification_stats(struct bpf_verifier_env *env)
14464 if (env->log.level & BPF_LOG_STATS) {
14465 verbose(env, "verification time %lld usec\n",
14466 div_u64(env->verification_time, 1000));
14467 verbose(env, "stack depth ");
14468 for (i = 0; i < env->subprog_cnt; i++) {
14469 u32 depth = env->subprog_info[i].stack_depth;
14471 verbose(env, "%d", depth);
14472 if (i + 1 < env->subprog_cnt)
14475 verbose(env, "\n");
14477 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
14478 "total_states %d peak_states %d mark_read %d\n",
14479 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
14480 env->max_states_per_insn, env->total_states,
14481 env->peak_states, env->longest_mark_read_walk);
14484 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
14486 const struct btf_type *t, *func_proto;
14487 const struct bpf_struct_ops *st_ops;
14488 const struct btf_member *member;
14489 struct bpf_prog *prog = env->prog;
14490 u32 btf_id, member_idx;
14493 if (!prog->gpl_compatible) {
14494 verbose(env, "struct ops programs must have a GPL compatible license\n");
14498 btf_id = prog->aux->attach_btf_id;
14499 st_ops = bpf_struct_ops_find(btf_id);
14501 verbose(env, "attach_btf_id %u is not a supported struct\n",
14507 member_idx = prog->expected_attach_type;
14508 if (member_idx >= btf_type_vlen(t)) {
14509 verbose(env, "attach to invalid member idx %u of struct %s\n",
14510 member_idx, st_ops->name);
14514 member = &btf_type_member(t)[member_idx];
14515 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
14516 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
14519 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
14520 mname, member_idx, st_ops->name);
14524 if (st_ops->check_member) {
14525 int err = st_ops->check_member(t, member);
14528 verbose(env, "attach to unsupported member %s of struct %s\n",
14529 mname, st_ops->name);
14534 prog->aux->attach_func_proto = func_proto;
14535 prog->aux->attach_func_name = mname;
14536 env->ops = st_ops->verifier_ops;
14540 #define SECURITY_PREFIX "security_"
14542 static int check_attach_modify_return(unsigned long addr, const char *func_name)
14544 if (within_error_injection_list(addr) ||
14545 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
14551 /* list of non-sleepable functions that are otherwise on
14552 * ALLOW_ERROR_INJECTION list
14554 BTF_SET_START(btf_non_sleepable_error_inject)
14555 /* Three functions below can be called from sleepable and non-sleepable context.
14556 * Assume non-sleepable from bpf safety point of view.
14558 BTF_ID(func, __filemap_add_folio)
14559 BTF_ID(func, should_fail_alloc_page)
14560 BTF_ID(func, should_failslab)
14561 BTF_SET_END(btf_non_sleepable_error_inject)
14563 static int check_non_sleepable_error_inject(u32 btf_id)
14565 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
14568 int bpf_check_attach_target(struct bpf_verifier_log *log,
14569 const struct bpf_prog *prog,
14570 const struct bpf_prog *tgt_prog,
14572 struct bpf_attach_target_info *tgt_info)
14574 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
14575 const char prefix[] = "btf_trace_";
14576 int ret = 0, subprog = -1, i;
14577 const struct btf_type *t;
14578 bool conservative = true;
14584 bpf_log(log, "Tracing programs must provide btf_id\n");
14587 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
14590 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
14593 t = btf_type_by_id(btf, btf_id);
14595 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
14598 tname = btf_name_by_offset(btf, t->name_off);
14600 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
14604 struct bpf_prog_aux *aux = tgt_prog->aux;
14606 for (i = 0; i < aux->func_info_cnt; i++)
14607 if (aux->func_info[i].type_id == btf_id) {
14611 if (subprog == -1) {
14612 bpf_log(log, "Subprog %s doesn't exist\n", tname);
14615 conservative = aux->func_info_aux[subprog].unreliable;
14616 if (prog_extension) {
14617 if (conservative) {
14619 "Cannot replace static functions\n");
14622 if (!prog->jit_requested) {
14624 "Extension programs should be JITed\n");
14628 if (!tgt_prog->jited) {
14629 bpf_log(log, "Can attach to only JITed progs\n");
14632 if (tgt_prog->type == prog->type) {
14633 /* Cannot fentry/fexit another fentry/fexit program.
14634 * Cannot attach program extension to another extension.
14635 * It's ok to attach fentry/fexit to extension program.
14637 bpf_log(log, "Cannot recursively attach\n");
14640 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
14642 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
14643 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
14644 /* Program extensions can extend all program types
14645 * except fentry/fexit. The reason is the following.
14646 * The fentry/fexit programs are used for performance
14647 * analysis, stats and can be attached to any program
14648 * type except themselves. When extension program is
14649 * replacing XDP function it is necessary to allow
14650 * performance analysis of all functions. Both original
14651 * XDP program and its program extension. Hence
14652 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
14653 * allowed. If extending of fentry/fexit was allowed it
14654 * would be possible to create long call chain
14655 * fentry->extension->fentry->extension beyond
14656 * reasonable stack size. Hence extending fentry is not
14659 bpf_log(log, "Cannot extend fentry/fexit\n");
14663 if (prog_extension) {
14664 bpf_log(log, "Cannot replace kernel functions\n");
14669 switch (prog->expected_attach_type) {
14670 case BPF_TRACE_RAW_TP:
14673 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
14676 if (!btf_type_is_typedef(t)) {
14677 bpf_log(log, "attach_btf_id %u is not a typedef\n",
14681 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
14682 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
14686 tname += sizeof(prefix) - 1;
14687 t = btf_type_by_id(btf, t->type);
14688 if (!btf_type_is_ptr(t))
14689 /* should never happen in valid vmlinux build */
14691 t = btf_type_by_id(btf, t->type);
14692 if (!btf_type_is_func_proto(t))
14693 /* should never happen in valid vmlinux build */
14697 case BPF_TRACE_ITER:
14698 if (!btf_type_is_func(t)) {
14699 bpf_log(log, "attach_btf_id %u is not a function\n",
14703 t = btf_type_by_id(btf, t->type);
14704 if (!btf_type_is_func_proto(t))
14706 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
14711 if (!prog_extension)
14714 case BPF_MODIFY_RETURN:
14716 case BPF_TRACE_FENTRY:
14717 case BPF_TRACE_FEXIT:
14718 if (!btf_type_is_func(t)) {
14719 bpf_log(log, "attach_btf_id %u is not a function\n",
14723 if (prog_extension &&
14724 btf_check_type_match(log, prog, btf, t))
14726 t = btf_type_by_id(btf, t->type);
14727 if (!btf_type_is_func_proto(t))
14730 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
14731 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
14732 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
14735 if (tgt_prog && conservative)
14738 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
14744 addr = (long) tgt_prog->bpf_func;
14746 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
14748 addr = kallsyms_lookup_name(tname);
14751 "The address of function %s cannot be found\n",
14757 if (prog->aux->sleepable) {
14759 switch (prog->type) {
14760 case BPF_PROG_TYPE_TRACING:
14761 /* fentry/fexit/fmod_ret progs can be sleepable only if they are
14762 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
14764 if (!check_non_sleepable_error_inject(btf_id) &&
14765 within_error_injection_list(addr))
14768 case BPF_PROG_TYPE_LSM:
14769 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
14770 * Only some of them are sleepable.
14772 if (bpf_lsm_is_sleepable_hook(btf_id))
14779 bpf_log(log, "%s is not sleepable\n", tname);
14782 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
14784 bpf_log(log, "can't modify return codes of BPF programs\n");
14787 ret = check_attach_modify_return(addr, tname);
14789 bpf_log(log, "%s() is not modifiable\n", tname);
14796 tgt_info->tgt_addr = addr;
14797 tgt_info->tgt_name = tname;
14798 tgt_info->tgt_type = t;
14802 BTF_SET_START(btf_id_deny)
14805 BTF_ID(func, migrate_disable)
14806 BTF_ID(func, migrate_enable)
14808 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
14809 BTF_ID(func, rcu_read_unlock_strict)
14811 BTF_SET_END(btf_id_deny)
14813 static int check_attach_btf_id(struct bpf_verifier_env *env)
14815 struct bpf_prog *prog = env->prog;
14816 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
14817 struct bpf_attach_target_info tgt_info = {};
14818 u32 btf_id = prog->aux->attach_btf_id;
14819 struct bpf_trampoline *tr;
14823 if (prog->type == BPF_PROG_TYPE_SYSCALL) {
14824 if (prog->aux->sleepable)
14825 /* attach_btf_id checked to be zero already */
14827 verbose(env, "Syscall programs can only be sleepable\n");
14831 if (prog->aux->sleepable && prog->type != BPF_PROG_TYPE_TRACING &&
14832 prog->type != BPF_PROG_TYPE_LSM) {
14833 verbose(env, "Only fentry/fexit/fmod_ret and lsm programs can be sleepable\n");
14837 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
14838 return check_struct_ops_btf_id(env);
14840 if (prog->type != BPF_PROG_TYPE_TRACING &&
14841 prog->type != BPF_PROG_TYPE_LSM &&
14842 prog->type != BPF_PROG_TYPE_EXT)
14845 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
14849 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
14850 /* to make freplace equivalent to their targets, they need to
14851 * inherit env->ops and expected_attach_type for the rest of the
14854 env->ops = bpf_verifier_ops[tgt_prog->type];
14855 prog->expected_attach_type = tgt_prog->expected_attach_type;
14858 /* store info about the attachment target that will be used later */
14859 prog->aux->attach_func_proto = tgt_info.tgt_type;
14860 prog->aux->attach_func_name = tgt_info.tgt_name;
14863 prog->aux->saved_dst_prog_type = tgt_prog->type;
14864 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
14867 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
14868 prog->aux->attach_btf_trace = true;
14870 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
14871 if (!bpf_iter_prog_supported(prog))
14876 if (prog->type == BPF_PROG_TYPE_LSM) {
14877 ret = bpf_lsm_verify_prog(&env->log, prog);
14880 } else if (prog->type == BPF_PROG_TYPE_TRACING &&
14881 btf_id_set_contains(&btf_id_deny, btf_id)) {
14885 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
14886 tr = bpf_trampoline_get(key, &tgt_info);
14890 prog->aux->dst_trampoline = tr;
14894 struct btf *bpf_get_btf_vmlinux(void)
14896 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
14897 mutex_lock(&bpf_verifier_lock);
14899 btf_vmlinux = btf_parse_vmlinux();
14900 mutex_unlock(&bpf_verifier_lock);
14902 return btf_vmlinux;
14905 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr)
14907 u64 start_time = ktime_get_ns();
14908 struct bpf_verifier_env *env;
14909 struct bpf_verifier_log *log;
14910 int i, len, ret = -EINVAL;
14913 /* no program is valid */
14914 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
14917 /* 'struct bpf_verifier_env' can be global, but since it's not small,
14918 * allocate/free it every time bpf_check() is called
14920 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
14925 len = (*prog)->len;
14926 env->insn_aux_data =
14927 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
14929 if (!env->insn_aux_data)
14931 for (i = 0; i < len; i++)
14932 env->insn_aux_data[i].orig_idx = i;
14934 env->ops = bpf_verifier_ops[env->prog->type];
14935 env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
14936 is_priv = bpf_capable();
14938 bpf_get_btf_vmlinux();
14940 /* grab the mutex to protect few globals used by verifier */
14942 mutex_lock(&bpf_verifier_lock);
14944 if (attr->log_level || attr->log_buf || attr->log_size) {
14945 /* user requested verbose verifier output
14946 * and supplied buffer to store the verification trace
14948 log->level = attr->log_level;
14949 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
14950 log->len_total = attr->log_size;
14952 /* log attributes have to be sane */
14953 if (!bpf_verifier_log_attr_valid(log)) {
14959 mark_verifier_state_clean(env);
14961 if (IS_ERR(btf_vmlinux)) {
14962 /* Either gcc or pahole or kernel are broken. */
14963 verbose(env, "in-kernel BTF is malformed\n");
14964 ret = PTR_ERR(btf_vmlinux);
14965 goto skip_full_check;
14968 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
14969 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
14970 env->strict_alignment = true;
14971 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
14972 env->strict_alignment = false;
14974 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
14975 env->allow_uninit_stack = bpf_allow_uninit_stack();
14976 env->allow_ptr_to_map_access = bpf_allow_ptr_to_map_access();
14977 env->bypass_spec_v1 = bpf_bypass_spec_v1();
14978 env->bypass_spec_v4 = bpf_bypass_spec_v4();
14979 env->bpf_capable = bpf_capable();
14982 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
14984 env->explored_states = kvcalloc(state_htab_size(env),
14985 sizeof(struct bpf_verifier_state_list *),
14988 if (!env->explored_states)
14989 goto skip_full_check;
14991 ret = add_subprog_and_kfunc(env);
14993 goto skip_full_check;
14995 ret = check_subprogs(env);
14997 goto skip_full_check;
14999 ret = check_btf_info(env, attr, uattr);
15001 goto skip_full_check;
15003 ret = check_attach_btf_id(env);
15005 goto skip_full_check;
15007 ret = resolve_pseudo_ldimm64(env);
15009 goto skip_full_check;
15011 if (bpf_prog_is_dev_bound(env->prog->aux)) {
15012 ret = bpf_prog_offload_verifier_prep(env->prog);
15014 goto skip_full_check;
15017 ret = check_cfg(env);
15019 goto skip_full_check;
15021 ret = do_check_subprogs(env);
15022 ret = ret ?: do_check_main(env);
15024 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
15025 ret = bpf_prog_offload_finalize(env);
15028 kvfree(env->explored_states);
15031 ret = check_max_stack_depth(env);
15033 /* instruction rewrites happen after this point */
15036 opt_hard_wire_dead_code_branches(env);
15038 ret = opt_remove_dead_code(env);
15040 ret = opt_remove_nops(env);
15043 sanitize_dead_code(env);
15047 /* program is valid, convert *(u32*)(ctx + off) accesses */
15048 ret = convert_ctx_accesses(env);
15051 ret = do_misc_fixups(env);
15053 /* do 32-bit optimization after insn patching has done so those patched
15054 * insns could be handled correctly.
15056 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
15057 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
15058 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
15063 ret = fixup_call_args(env);
15065 env->verification_time = ktime_get_ns() - start_time;
15066 print_verification_stats(env);
15067 env->prog->aux->verified_insns = env->insn_processed;
15069 if (log->level && bpf_verifier_log_full(log))
15071 if (log->level && !log->ubuf) {
15073 goto err_release_maps;
15077 goto err_release_maps;
15079 if (env->used_map_cnt) {
15080 /* if program passed verifier, update used_maps in bpf_prog_info */
15081 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
15082 sizeof(env->used_maps[0]),
15085 if (!env->prog->aux->used_maps) {
15087 goto err_release_maps;
15090 memcpy(env->prog->aux->used_maps, env->used_maps,
15091 sizeof(env->used_maps[0]) * env->used_map_cnt);
15092 env->prog->aux->used_map_cnt = env->used_map_cnt;
15094 if (env->used_btf_cnt) {
15095 /* if program passed verifier, update used_btfs in bpf_prog_aux */
15096 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
15097 sizeof(env->used_btfs[0]),
15099 if (!env->prog->aux->used_btfs) {
15101 goto err_release_maps;
15104 memcpy(env->prog->aux->used_btfs, env->used_btfs,
15105 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
15106 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
15108 if (env->used_map_cnt || env->used_btf_cnt) {
15109 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
15110 * bpf_ld_imm64 instructions
15112 convert_pseudo_ld_imm64(env);
15115 adjust_btf_func(env);
15118 if (!env->prog->aux->used_maps)
15119 /* if we didn't copy map pointers into bpf_prog_info, release
15120 * them now. Otherwise free_used_maps() will release them.
15123 if (!env->prog->aux->used_btfs)
15126 /* extension progs temporarily inherit the attach_type of their targets
15127 for verification purposes, so set it back to zero before returning
15129 if (env->prog->type == BPF_PROG_TYPE_EXT)
15130 env->prog->expected_attach_type = 0;
15135 mutex_unlock(&bpf_verifier_lock);
15136 vfree(env->insn_aux_data);