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/kernel.h>
8 #include <linux/types.h>
9 #include <linux/slab.h>
10 #include <linux/bpf.h>
11 #include <linux/btf.h>
12 #include <linux/bpf_verifier.h>
13 #include <linux/filter.h>
14 #include <net/netlink.h>
15 #include <linux/file.h>
16 #include <linux/vmalloc.h>
17 #include <linux/stringify.h>
18 #include <linux/bsearch.h>
19 #include <linux/sort.h>
20 #include <linux/perf_event.h>
21 #include <linux/ctype.h>
22 #include <linux/error-injection.h>
23 #include <linux/bpf_lsm.h>
27 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
28 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
29 [_id] = & _name ## _verifier_ops,
30 #define BPF_MAP_TYPE(_id, _ops)
31 #define BPF_LINK_TYPE(_id, _name)
32 #include <linux/bpf_types.h>
38 /* bpf_check() is a static code analyzer that walks eBPF program
39 * instruction by instruction and updates register/stack state.
40 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
42 * The first pass is depth-first-search to check that the program is a DAG.
43 * It rejects the following programs:
44 * - larger than BPF_MAXINSNS insns
45 * - if loop is present (detected via back-edge)
46 * - unreachable insns exist (shouldn't be a forest. program = one function)
47 * - out of bounds or malformed jumps
48 * The second pass is all possible path descent from the 1st insn.
49 * Since it's analyzing all pathes through the program, the length of the
50 * analysis is limited to 64k insn, which may be hit even if total number of
51 * insn is less then 4K, but there are too many branches that change stack/regs.
52 * Number of 'branches to be analyzed' is limited to 1k
54 * On entry to each instruction, each register has a type, and the instruction
55 * changes the types of the registers depending on instruction semantics.
56 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
59 * All registers are 64-bit.
60 * R0 - return register
61 * R1-R5 argument passing registers
62 * R6-R9 callee saved registers
63 * R10 - frame pointer read-only
65 * At the start of BPF program the register R1 contains a pointer to bpf_context
66 * and has type PTR_TO_CTX.
68 * Verifier tracks arithmetic operations on pointers in case:
69 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
70 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
71 * 1st insn copies R10 (which has FRAME_PTR) type into R1
72 * and 2nd arithmetic instruction is pattern matched to recognize
73 * that it wants to construct a pointer to some element within stack.
74 * So after 2nd insn, the register R1 has type PTR_TO_STACK
75 * (and -20 constant is saved for further stack bounds checking).
76 * Meaning that this reg is a pointer to stack plus known immediate constant.
78 * Most of the time the registers have SCALAR_VALUE type, which
79 * means the register has some value, but it's not a valid pointer.
80 * (like pointer plus pointer becomes SCALAR_VALUE type)
82 * When verifier sees load or store instructions the type of base register
83 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
84 * four pointer types recognized by check_mem_access() function.
86 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
87 * and the range of [ptr, ptr + map's value_size) is accessible.
89 * registers used to pass values to function calls are checked against
90 * function argument constraints.
92 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
93 * It means that the register type passed to this function must be
94 * PTR_TO_STACK and it will be used inside the function as
95 * 'pointer to map element key'
97 * For example the argument constraints for bpf_map_lookup_elem():
98 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
99 * .arg1_type = ARG_CONST_MAP_PTR,
100 * .arg2_type = ARG_PTR_TO_MAP_KEY,
102 * ret_type says that this function returns 'pointer to map elem value or null'
103 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
104 * 2nd argument should be a pointer to stack, which will be used inside
105 * the helper function as a pointer to map element key.
107 * On the kernel side the helper function looks like:
108 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
110 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
111 * void *key = (void *) (unsigned long) r2;
114 * here kernel can access 'key' and 'map' pointers safely, knowing that
115 * [key, key + map->key_size) bytes are valid and were initialized on
116 * the stack of eBPF program.
119 * Corresponding eBPF program may look like:
120 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
121 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
122 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
123 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
124 * here verifier looks at prototype of map_lookup_elem() and sees:
125 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
126 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
128 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
129 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
130 * and were initialized prior to this call.
131 * If it's ok, then verifier allows this BPF_CALL insn and looks at
132 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
133 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
134 * returns ether pointer to map value or NULL.
136 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
137 * insn, the register holding that pointer in the true branch changes state to
138 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
139 * branch. See check_cond_jmp_op().
141 * After the call R0 is set to return type of the function and registers R1-R5
142 * are set to NOT_INIT to indicate that they are no longer readable.
144 * The following reference types represent a potential reference to a kernel
145 * resource which, after first being allocated, must be checked and freed by
147 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
149 * When the verifier sees a helper call return a reference type, it allocates a
150 * pointer id for the reference and stores it in the current function state.
151 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
152 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
153 * passes through a NULL-check conditional. For the branch wherein the state is
154 * changed to CONST_IMM, the verifier releases the reference.
156 * For each helper function that allocates a reference, such as
157 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
158 * bpf_sk_release(). When a reference type passes into the release function,
159 * the verifier also releases the reference. If any unchecked or unreleased
160 * reference remains at the end of the program, the verifier rejects it.
163 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
164 struct bpf_verifier_stack_elem {
165 /* verifer state is 'st'
166 * before processing instruction 'insn_idx'
167 * and after processing instruction 'prev_insn_idx'
169 struct bpf_verifier_state st;
172 struct bpf_verifier_stack_elem *next;
173 /* length of verifier log at the time this state was pushed on stack */
177 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
178 #define BPF_COMPLEXITY_LIMIT_STATES 64
180 #define BPF_MAP_KEY_POISON (1ULL << 63)
181 #define BPF_MAP_KEY_SEEN (1ULL << 62)
183 #define BPF_MAP_PTR_UNPRIV 1UL
184 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
185 POISON_POINTER_DELTA))
186 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
188 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
190 return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON;
193 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
195 return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV;
198 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
199 const struct bpf_map *map, bool unpriv)
201 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
202 unpriv |= bpf_map_ptr_unpriv(aux);
203 aux->map_ptr_state = (unsigned long)map |
204 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
207 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
209 return aux->map_key_state & BPF_MAP_KEY_POISON;
212 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
214 return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
217 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
219 return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
222 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
224 bool poisoned = bpf_map_key_poisoned(aux);
226 aux->map_key_state = state | BPF_MAP_KEY_SEEN |
227 (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
230 struct bpf_call_arg_meta {
231 struct bpf_map *map_ptr;
243 struct btf *btf_vmlinux;
245 static DEFINE_MUTEX(bpf_verifier_lock);
247 static const struct bpf_line_info *
248 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
250 const struct bpf_line_info *linfo;
251 const struct bpf_prog *prog;
255 nr_linfo = prog->aux->nr_linfo;
257 if (!nr_linfo || insn_off >= prog->len)
260 linfo = prog->aux->linfo;
261 for (i = 1; i < nr_linfo; i++)
262 if (insn_off < linfo[i].insn_off)
265 return &linfo[i - 1];
268 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
273 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
275 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
276 "verifier log line truncated - local buffer too short\n");
278 n = min(log->len_total - log->len_used - 1, n);
281 if (log->level == BPF_LOG_KERNEL) {
282 pr_err("BPF:%s\n", log->kbuf);
285 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
291 static void bpf_vlog_reset(struct bpf_verifier_log *log, u32 new_pos)
295 if (!bpf_verifier_log_needed(log))
298 log->len_used = new_pos;
299 if (put_user(zero, log->ubuf + new_pos))
303 /* log_level controls verbosity level of eBPF verifier.
304 * bpf_verifier_log_write() is used to dump the verification trace to the log,
305 * so the user can figure out what's wrong with the program
307 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
308 const char *fmt, ...)
312 if (!bpf_verifier_log_needed(&env->log))
316 bpf_verifier_vlog(&env->log, fmt, args);
319 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
321 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
323 struct bpf_verifier_env *env = private_data;
326 if (!bpf_verifier_log_needed(&env->log))
330 bpf_verifier_vlog(&env->log, fmt, args);
334 __printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
335 const char *fmt, ...)
339 if (!bpf_verifier_log_needed(log))
343 bpf_verifier_vlog(log, fmt, args);
347 static const char *ltrim(const char *s)
355 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
357 const char *prefix_fmt, ...)
359 const struct bpf_line_info *linfo;
361 if (!bpf_verifier_log_needed(&env->log))
364 linfo = find_linfo(env, insn_off);
365 if (!linfo || linfo == env->prev_linfo)
371 va_start(args, prefix_fmt);
372 bpf_verifier_vlog(&env->log, prefix_fmt, args);
377 ltrim(btf_name_by_offset(env->prog->aux->btf,
380 env->prev_linfo = linfo;
383 static bool type_is_pkt_pointer(enum bpf_reg_type type)
385 return type == PTR_TO_PACKET ||
386 type == PTR_TO_PACKET_META;
389 static bool type_is_sk_pointer(enum bpf_reg_type type)
391 return type == PTR_TO_SOCKET ||
392 type == PTR_TO_SOCK_COMMON ||
393 type == PTR_TO_TCP_SOCK ||
394 type == PTR_TO_XDP_SOCK;
397 static bool reg_type_not_null(enum bpf_reg_type type)
399 return type == PTR_TO_SOCKET ||
400 type == PTR_TO_TCP_SOCK ||
401 type == PTR_TO_MAP_VALUE ||
402 type == PTR_TO_SOCK_COMMON;
405 static bool reg_type_may_be_null(enum bpf_reg_type type)
407 return type == PTR_TO_MAP_VALUE_OR_NULL ||
408 type == PTR_TO_SOCKET_OR_NULL ||
409 type == PTR_TO_SOCK_COMMON_OR_NULL ||
410 type == PTR_TO_TCP_SOCK_OR_NULL ||
411 type == PTR_TO_BTF_ID_OR_NULL ||
412 type == PTR_TO_MEM_OR_NULL ||
413 type == PTR_TO_RDONLY_BUF_OR_NULL ||
414 type == PTR_TO_RDWR_BUF_OR_NULL;
417 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
419 return reg->type == PTR_TO_MAP_VALUE &&
420 map_value_has_spin_lock(reg->map_ptr);
423 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
425 return type == PTR_TO_SOCKET ||
426 type == PTR_TO_SOCKET_OR_NULL ||
427 type == PTR_TO_TCP_SOCK ||
428 type == PTR_TO_TCP_SOCK_OR_NULL ||
429 type == PTR_TO_MEM ||
430 type == PTR_TO_MEM_OR_NULL;
433 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
435 return type == ARG_PTR_TO_SOCK_COMMON;
438 /* Determine whether the function releases some resources allocated by another
439 * function call. The first reference type argument will be assumed to be
440 * released by release_reference().
442 static bool is_release_function(enum bpf_func_id func_id)
444 return func_id == BPF_FUNC_sk_release ||
445 func_id == BPF_FUNC_ringbuf_submit ||
446 func_id == BPF_FUNC_ringbuf_discard;
449 static bool may_be_acquire_function(enum bpf_func_id func_id)
451 return func_id == BPF_FUNC_sk_lookup_tcp ||
452 func_id == BPF_FUNC_sk_lookup_udp ||
453 func_id == BPF_FUNC_skc_lookup_tcp ||
454 func_id == BPF_FUNC_map_lookup_elem ||
455 func_id == BPF_FUNC_ringbuf_reserve;
458 static bool is_acquire_function(enum bpf_func_id func_id,
459 const struct bpf_map *map)
461 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
463 if (func_id == BPF_FUNC_sk_lookup_tcp ||
464 func_id == BPF_FUNC_sk_lookup_udp ||
465 func_id == BPF_FUNC_skc_lookup_tcp ||
466 func_id == BPF_FUNC_ringbuf_reserve)
469 if (func_id == BPF_FUNC_map_lookup_elem &&
470 (map_type == BPF_MAP_TYPE_SOCKMAP ||
471 map_type == BPF_MAP_TYPE_SOCKHASH))
477 static bool is_ptr_cast_function(enum bpf_func_id func_id)
479 return func_id == BPF_FUNC_tcp_sock ||
480 func_id == BPF_FUNC_sk_fullsock;
483 /* string representation of 'enum bpf_reg_type' */
484 static const char * const reg_type_str[] = {
486 [SCALAR_VALUE] = "inv",
487 [PTR_TO_CTX] = "ctx",
488 [CONST_PTR_TO_MAP] = "map_ptr",
489 [PTR_TO_MAP_VALUE] = "map_value",
490 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
491 [PTR_TO_STACK] = "fp",
492 [PTR_TO_PACKET] = "pkt",
493 [PTR_TO_PACKET_META] = "pkt_meta",
494 [PTR_TO_PACKET_END] = "pkt_end",
495 [PTR_TO_FLOW_KEYS] = "flow_keys",
496 [PTR_TO_SOCKET] = "sock",
497 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
498 [PTR_TO_SOCK_COMMON] = "sock_common",
499 [PTR_TO_SOCK_COMMON_OR_NULL] = "sock_common_or_null",
500 [PTR_TO_TCP_SOCK] = "tcp_sock",
501 [PTR_TO_TCP_SOCK_OR_NULL] = "tcp_sock_or_null",
502 [PTR_TO_TP_BUFFER] = "tp_buffer",
503 [PTR_TO_XDP_SOCK] = "xdp_sock",
504 [PTR_TO_BTF_ID] = "ptr_",
505 [PTR_TO_BTF_ID_OR_NULL] = "ptr_or_null_",
506 [PTR_TO_MEM] = "mem",
507 [PTR_TO_MEM_OR_NULL] = "mem_or_null",
508 [PTR_TO_RDONLY_BUF] = "rdonly_buf",
509 [PTR_TO_RDONLY_BUF_OR_NULL] = "rdonly_buf_or_null",
510 [PTR_TO_RDWR_BUF] = "rdwr_buf",
511 [PTR_TO_RDWR_BUF_OR_NULL] = "rdwr_buf_or_null",
514 static char slot_type_char[] = {
515 [STACK_INVALID] = '?',
521 static void print_liveness(struct bpf_verifier_env *env,
522 enum bpf_reg_liveness live)
524 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
526 if (live & REG_LIVE_READ)
528 if (live & REG_LIVE_WRITTEN)
530 if (live & REG_LIVE_DONE)
534 static struct bpf_func_state *func(struct bpf_verifier_env *env,
535 const struct bpf_reg_state *reg)
537 struct bpf_verifier_state *cur = env->cur_state;
539 return cur->frame[reg->frameno];
542 const char *kernel_type_name(u32 id)
544 return btf_name_by_offset(btf_vmlinux,
545 btf_type_by_id(btf_vmlinux, id)->name_off);
548 static void print_verifier_state(struct bpf_verifier_env *env,
549 const struct bpf_func_state *state)
551 const struct bpf_reg_state *reg;
556 verbose(env, " frame%d:", state->frameno);
557 for (i = 0; i < MAX_BPF_REG; i++) {
558 reg = &state->regs[i];
562 verbose(env, " R%d", i);
563 print_liveness(env, reg->live);
564 verbose(env, "=%s", reg_type_str[t]);
565 if (t == SCALAR_VALUE && reg->precise)
567 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
568 tnum_is_const(reg->var_off)) {
569 /* reg->off should be 0 for SCALAR_VALUE */
570 verbose(env, "%lld", reg->var_off.value + reg->off);
572 if (t == PTR_TO_BTF_ID || t == PTR_TO_BTF_ID_OR_NULL)
573 verbose(env, "%s", kernel_type_name(reg->btf_id));
574 verbose(env, "(id=%d", reg->id);
575 if (reg_type_may_be_refcounted_or_null(t))
576 verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
577 if (t != SCALAR_VALUE)
578 verbose(env, ",off=%d", reg->off);
579 if (type_is_pkt_pointer(t))
580 verbose(env, ",r=%d", reg->range);
581 else if (t == CONST_PTR_TO_MAP ||
582 t == PTR_TO_MAP_VALUE ||
583 t == PTR_TO_MAP_VALUE_OR_NULL)
584 verbose(env, ",ks=%d,vs=%d",
585 reg->map_ptr->key_size,
586 reg->map_ptr->value_size);
587 if (tnum_is_const(reg->var_off)) {
588 /* Typically an immediate SCALAR_VALUE, but
589 * could be a pointer whose offset is too big
592 verbose(env, ",imm=%llx", reg->var_off.value);
594 if (reg->smin_value != reg->umin_value &&
595 reg->smin_value != S64_MIN)
596 verbose(env, ",smin_value=%lld",
597 (long long)reg->smin_value);
598 if (reg->smax_value != reg->umax_value &&
599 reg->smax_value != S64_MAX)
600 verbose(env, ",smax_value=%lld",
601 (long long)reg->smax_value);
602 if (reg->umin_value != 0)
603 verbose(env, ",umin_value=%llu",
604 (unsigned long long)reg->umin_value);
605 if (reg->umax_value != U64_MAX)
606 verbose(env, ",umax_value=%llu",
607 (unsigned long long)reg->umax_value);
608 if (!tnum_is_unknown(reg->var_off)) {
611 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
612 verbose(env, ",var_off=%s", tn_buf);
614 if (reg->s32_min_value != reg->smin_value &&
615 reg->s32_min_value != S32_MIN)
616 verbose(env, ",s32_min_value=%d",
617 (int)(reg->s32_min_value));
618 if (reg->s32_max_value != reg->smax_value &&
619 reg->s32_max_value != S32_MAX)
620 verbose(env, ",s32_max_value=%d",
621 (int)(reg->s32_max_value));
622 if (reg->u32_min_value != reg->umin_value &&
623 reg->u32_min_value != U32_MIN)
624 verbose(env, ",u32_min_value=%d",
625 (int)(reg->u32_min_value));
626 if (reg->u32_max_value != reg->umax_value &&
627 reg->u32_max_value != U32_MAX)
628 verbose(env, ",u32_max_value=%d",
629 (int)(reg->u32_max_value));
634 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
635 char types_buf[BPF_REG_SIZE + 1];
639 for (j = 0; j < BPF_REG_SIZE; j++) {
640 if (state->stack[i].slot_type[j] != STACK_INVALID)
642 types_buf[j] = slot_type_char[
643 state->stack[i].slot_type[j]];
645 types_buf[BPF_REG_SIZE] = 0;
648 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
649 print_liveness(env, state->stack[i].spilled_ptr.live);
650 if (state->stack[i].slot_type[0] == STACK_SPILL) {
651 reg = &state->stack[i].spilled_ptr;
653 verbose(env, "=%s", reg_type_str[t]);
654 if (t == SCALAR_VALUE && reg->precise)
656 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
657 verbose(env, "%lld", reg->var_off.value + reg->off);
659 verbose(env, "=%s", types_buf);
662 if (state->acquired_refs && state->refs[0].id) {
663 verbose(env, " refs=%d", state->refs[0].id);
664 for (i = 1; i < state->acquired_refs; i++)
665 if (state->refs[i].id)
666 verbose(env, ",%d", state->refs[i].id);
671 #define COPY_STATE_FN(NAME, COUNT, FIELD, SIZE) \
672 static int copy_##NAME##_state(struct bpf_func_state *dst, \
673 const struct bpf_func_state *src) \
677 if (WARN_ON_ONCE(dst->COUNT < src->COUNT)) { \
678 /* internal bug, make state invalid to reject the program */ \
679 memset(dst, 0, sizeof(*dst)); \
682 memcpy(dst->FIELD, src->FIELD, \
683 sizeof(*src->FIELD) * (src->COUNT / SIZE)); \
686 /* copy_reference_state() */
687 COPY_STATE_FN(reference, acquired_refs, refs, 1)
688 /* copy_stack_state() */
689 COPY_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
692 #define REALLOC_STATE_FN(NAME, COUNT, FIELD, SIZE) \
693 static int realloc_##NAME##_state(struct bpf_func_state *state, int size, \
696 u32 old_size = state->COUNT; \
697 struct bpf_##NAME##_state *new_##FIELD; \
698 int slot = size / SIZE; \
700 if (size <= old_size || !size) { \
703 state->COUNT = slot * SIZE; \
704 if (!size && old_size) { \
705 kfree(state->FIELD); \
706 state->FIELD = NULL; \
710 new_##FIELD = kmalloc_array(slot, sizeof(struct bpf_##NAME##_state), \
716 memcpy(new_##FIELD, state->FIELD, \
717 sizeof(*new_##FIELD) * (old_size / SIZE)); \
718 memset(new_##FIELD + old_size / SIZE, 0, \
719 sizeof(*new_##FIELD) * (size - old_size) / SIZE); \
721 state->COUNT = slot * SIZE; \
722 kfree(state->FIELD); \
723 state->FIELD = new_##FIELD; \
726 /* realloc_reference_state() */
727 REALLOC_STATE_FN(reference, acquired_refs, refs, 1)
728 /* realloc_stack_state() */
729 REALLOC_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
730 #undef REALLOC_STATE_FN
732 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
733 * make it consume minimal amount of memory. check_stack_write() access from
734 * the program calls into realloc_func_state() to grow the stack size.
735 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
736 * which realloc_stack_state() copies over. It points to previous
737 * bpf_verifier_state which is never reallocated.
739 static int realloc_func_state(struct bpf_func_state *state, int stack_size,
740 int refs_size, bool copy_old)
742 int err = realloc_reference_state(state, refs_size, copy_old);
745 return realloc_stack_state(state, stack_size, copy_old);
748 /* Acquire a pointer id from the env and update the state->refs to include
749 * this new pointer reference.
750 * On success, returns a valid pointer id to associate with the register
751 * On failure, returns a negative errno.
753 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
755 struct bpf_func_state *state = cur_func(env);
756 int new_ofs = state->acquired_refs;
759 err = realloc_reference_state(state, state->acquired_refs + 1, true);
763 state->refs[new_ofs].id = id;
764 state->refs[new_ofs].insn_idx = insn_idx;
769 /* release function corresponding to acquire_reference_state(). Idempotent. */
770 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
774 last_idx = state->acquired_refs - 1;
775 for (i = 0; i < state->acquired_refs; i++) {
776 if (state->refs[i].id == ptr_id) {
777 if (last_idx && i != last_idx)
778 memcpy(&state->refs[i], &state->refs[last_idx],
779 sizeof(*state->refs));
780 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
781 state->acquired_refs--;
788 static int transfer_reference_state(struct bpf_func_state *dst,
789 struct bpf_func_state *src)
791 int err = realloc_reference_state(dst, src->acquired_refs, false);
794 err = copy_reference_state(dst, src);
800 static void free_func_state(struct bpf_func_state *state)
809 static void clear_jmp_history(struct bpf_verifier_state *state)
811 kfree(state->jmp_history);
812 state->jmp_history = NULL;
813 state->jmp_history_cnt = 0;
816 static void free_verifier_state(struct bpf_verifier_state *state,
821 for (i = 0; i <= state->curframe; i++) {
822 free_func_state(state->frame[i]);
823 state->frame[i] = NULL;
825 clear_jmp_history(state);
830 /* copy verifier state from src to dst growing dst stack space
831 * when necessary to accommodate larger src stack
833 static int copy_func_state(struct bpf_func_state *dst,
834 const struct bpf_func_state *src)
838 err = realloc_func_state(dst, src->allocated_stack, src->acquired_refs,
842 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
843 err = copy_reference_state(dst, src);
846 return copy_stack_state(dst, src);
849 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
850 const struct bpf_verifier_state *src)
852 struct bpf_func_state *dst;
853 u32 jmp_sz = sizeof(struct bpf_idx_pair) * src->jmp_history_cnt;
856 if (dst_state->jmp_history_cnt < src->jmp_history_cnt) {
857 kfree(dst_state->jmp_history);
858 dst_state->jmp_history = kmalloc(jmp_sz, GFP_USER);
859 if (!dst_state->jmp_history)
862 memcpy(dst_state->jmp_history, src->jmp_history, jmp_sz);
863 dst_state->jmp_history_cnt = src->jmp_history_cnt;
865 /* if dst has more stack frames then src frame, free them */
866 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
867 free_func_state(dst_state->frame[i]);
868 dst_state->frame[i] = NULL;
870 dst_state->speculative = src->speculative;
871 dst_state->curframe = src->curframe;
872 dst_state->active_spin_lock = src->active_spin_lock;
873 dst_state->branches = src->branches;
874 dst_state->parent = src->parent;
875 dst_state->first_insn_idx = src->first_insn_idx;
876 dst_state->last_insn_idx = src->last_insn_idx;
877 for (i = 0; i <= src->curframe; i++) {
878 dst = dst_state->frame[i];
880 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
883 dst_state->frame[i] = dst;
885 err = copy_func_state(dst, src->frame[i]);
892 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
895 u32 br = --st->branches;
897 /* WARN_ON(br > 1) technically makes sense here,
898 * but see comment in push_stack(), hence:
900 WARN_ONCE((int)br < 0,
901 "BUG update_branch_counts:branches_to_explore=%d\n",
909 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
910 int *insn_idx, bool pop_log)
912 struct bpf_verifier_state *cur = env->cur_state;
913 struct bpf_verifier_stack_elem *elem, *head = env->head;
916 if (env->head == NULL)
920 err = copy_verifier_state(cur, &head->st);
925 bpf_vlog_reset(&env->log, head->log_pos);
927 *insn_idx = head->insn_idx;
929 *prev_insn_idx = head->prev_insn_idx;
931 free_verifier_state(&head->st, false);
938 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
939 int insn_idx, int prev_insn_idx,
942 struct bpf_verifier_state *cur = env->cur_state;
943 struct bpf_verifier_stack_elem *elem;
946 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
950 elem->insn_idx = insn_idx;
951 elem->prev_insn_idx = prev_insn_idx;
952 elem->next = env->head;
953 elem->log_pos = env->log.len_used;
956 err = copy_verifier_state(&elem->st, cur);
959 elem->st.speculative |= speculative;
960 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
961 verbose(env, "The sequence of %d jumps is too complex.\n",
965 if (elem->st.parent) {
966 ++elem->st.parent->branches;
967 /* WARN_ON(branches > 2) technically makes sense here,
969 * 1. speculative states will bump 'branches' for non-branch
971 * 2. is_state_visited() heuristics may decide not to create
972 * a new state for a sequence of branches and all such current
973 * and cloned states will be pointing to a single parent state
974 * which might have large 'branches' count.
979 free_verifier_state(env->cur_state, true);
980 env->cur_state = NULL;
981 /* pop all elements and return */
982 while (!pop_stack(env, NULL, NULL, false));
986 #define CALLER_SAVED_REGS 6
987 static const int caller_saved[CALLER_SAVED_REGS] = {
988 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
991 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
992 struct bpf_reg_state *reg);
994 /* Mark the unknown part of a register (variable offset or scalar value) as
995 * known to have the value @imm.
997 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
999 /* Clear id, off, and union(map_ptr, range) */
1000 memset(((u8 *)reg) + sizeof(reg->type), 0,
1001 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1002 reg->var_off = tnum_const(imm);
1003 reg->smin_value = (s64)imm;
1004 reg->smax_value = (s64)imm;
1005 reg->umin_value = imm;
1006 reg->umax_value = imm;
1008 reg->s32_min_value = (s32)imm;
1009 reg->s32_max_value = (s32)imm;
1010 reg->u32_min_value = (u32)imm;
1011 reg->u32_max_value = (u32)imm;
1014 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1016 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1017 reg->s32_min_value = (s32)imm;
1018 reg->s32_max_value = (s32)imm;
1019 reg->u32_min_value = (u32)imm;
1020 reg->u32_max_value = (u32)imm;
1023 /* Mark the 'variable offset' part of a register as zero. This should be
1024 * used only on registers holding a pointer type.
1026 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1028 __mark_reg_known(reg, 0);
1031 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1033 __mark_reg_known(reg, 0);
1034 reg->type = SCALAR_VALUE;
1037 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1038 struct bpf_reg_state *regs, u32 regno)
1040 if (WARN_ON(regno >= MAX_BPF_REG)) {
1041 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1042 /* Something bad happened, let's kill all regs */
1043 for (regno = 0; regno < MAX_BPF_REG; regno++)
1044 __mark_reg_not_init(env, regs + regno);
1047 __mark_reg_known_zero(regs + regno);
1050 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1052 return type_is_pkt_pointer(reg->type);
1055 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
1057 return reg_is_pkt_pointer(reg) ||
1058 reg->type == PTR_TO_PACKET_END;
1061 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
1062 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
1063 enum bpf_reg_type which)
1065 /* The register can already have a range from prior markings.
1066 * This is fine as long as it hasn't been advanced from its
1069 return reg->type == which &&
1072 tnum_equals_const(reg->var_off, 0);
1075 /* Reset the min/max bounds of a register */
1076 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1078 reg->smin_value = S64_MIN;
1079 reg->smax_value = S64_MAX;
1080 reg->umin_value = 0;
1081 reg->umax_value = U64_MAX;
1083 reg->s32_min_value = S32_MIN;
1084 reg->s32_max_value = S32_MAX;
1085 reg->u32_min_value = 0;
1086 reg->u32_max_value = U32_MAX;
1089 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
1091 reg->smin_value = S64_MIN;
1092 reg->smax_value = S64_MAX;
1093 reg->umin_value = 0;
1094 reg->umax_value = U64_MAX;
1097 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
1099 reg->s32_min_value = S32_MIN;
1100 reg->s32_max_value = S32_MAX;
1101 reg->u32_min_value = 0;
1102 reg->u32_max_value = U32_MAX;
1105 static void __update_reg32_bounds(struct bpf_reg_state *reg)
1107 struct tnum var32_off = tnum_subreg(reg->var_off);
1109 /* min signed is max(sign bit) | min(other bits) */
1110 reg->s32_min_value = max_t(s32, reg->s32_min_value,
1111 var32_off.value | (var32_off.mask & S32_MIN));
1112 /* max signed is min(sign bit) | max(other bits) */
1113 reg->s32_max_value = min_t(s32, reg->s32_max_value,
1114 var32_off.value | (var32_off.mask & S32_MAX));
1115 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
1116 reg->u32_max_value = min(reg->u32_max_value,
1117 (u32)(var32_off.value | var32_off.mask));
1120 static void __update_reg64_bounds(struct bpf_reg_state *reg)
1122 /* min signed is max(sign bit) | min(other bits) */
1123 reg->smin_value = max_t(s64, reg->smin_value,
1124 reg->var_off.value | (reg->var_off.mask & S64_MIN));
1125 /* max signed is min(sign bit) | max(other bits) */
1126 reg->smax_value = min_t(s64, reg->smax_value,
1127 reg->var_off.value | (reg->var_off.mask & S64_MAX));
1128 reg->umin_value = max(reg->umin_value, reg->var_off.value);
1129 reg->umax_value = min(reg->umax_value,
1130 reg->var_off.value | reg->var_off.mask);
1133 static void __update_reg_bounds(struct bpf_reg_state *reg)
1135 __update_reg32_bounds(reg);
1136 __update_reg64_bounds(reg);
1139 /* Uses signed min/max values to inform unsigned, and vice-versa */
1140 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
1142 /* Learn sign from signed bounds.
1143 * If we cannot cross the sign boundary, then signed and unsigned bounds
1144 * are the same, so combine. This works even in the negative case, e.g.
1145 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1147 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
1148 reg->s32_min_value = reg->u32_min_value =
1149 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1150 reg->s32_max_value = reg->u32_max_value =
1151 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1154 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1155 * boundary, so we must be careful.
1157 if ((s32)reg->u32_max_value >= 0) {
1158 /* Positive. We can't learn anything from the smin, but smax
1159 * is positive, hence safe.
1161 reg->s32_min_value = reg->u32_min_value;
1162 reg->s32_max_value = reg->u32_max_value =
1163 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1164 } else if ((s32)reg->u32_min_value < 0) {
1165 /* Negative. We can't learn anything from the smax, but smin
1166 * is negative, hence safe.
1168 reg->s32_min_value = reg->u32_min_value =
1169 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1170 reg->s32_max_value = reg->u32_max_value;
1174 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
1176 /* Learn sign from signed bounds.
1177 * If we cannot cross the sign boundary, then signed and unsigned bounds
1178 * are the same, so combine. This works even in the negative case, e.g.
1179 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1181 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1182 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1184 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1188 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1189 * boundary, so we must be careful.
1191 if ((s64)reg->umax_value >= 0) {
1192 /* Positive. We can't learn anything from the smin, but smax
1193 * is positive, hence safe.
1195 reg->smin_value = reg->umin_value;
1196 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1198 } else if ((s64)reg->umin_value < 0) {
1199 /* Negative. We can't learn anything from the smax, but smin
1200 * is negative, hence safe.
1202 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1204 reg->smax_value = reg->umax_value;
1208 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
1210 __reg32_deduce_bounds(reg);
1211 __reg64_deduce_bounds(reg);
1214 /* Attempts to improve var_off based on unsigned min/max information */
1215 static void __reg_bound_offset(struct bpf_reg_state *reg)
1217 struct tnum var64_off = tnum_intersect(reg->var_off,
1218 tnum_range(reg->umin_value,
1220 struct tnum var32_off = tnum_intersect(tnum_subreg(reg->var_off),
1221 tnum_range(reg->u32_min_value,
1222 reg->u32_max_value));
1224 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
1227 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
1229 reg->umin_value = reg->u32_min_value;
1230 reg->umax_value = reg->u32_max_value;
1231 /* Attempt to pull 32-bit signed bounds into 64-bit bounds
1232 * but must be positive otherwise set to worse case bounds
1233 * and refine later from tnum.
1235 if (reg->s32_min_value >= 0 && reg->s32_max_value >= 0)
1236 reg->smax_value = reg->s32_max_value;
1238 reg->smax_value = U32_MAX;
1239 if (reg->s32_min_value >= 0)
1240 reg->smin_value = reg->s32_min_value;
1242 reg->smin_value = 0;
1245 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
1247 /* special case when 64-bit register has upper 32-bit register
1248 * zeroed. Typically happens after zext or <<32, >>32 sequence
1249 * allowing us to use 32-bit bounds directly,
1251 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
1252 __reg_assign_32_into_64(reg);
1254 /* Otherwise the best we can do is push lower 32bit known and
1255 * unknown bits into register (var_off set from jmp logic)
1256 * then learn as much as possible from the 64-bit tnum
1257 * known and unknown bits. The previous smin/smax bounds are
1258 * invalid here because of jmp32 compare so mark them unknown
1259 * so they do not impact tnum bounds calculation.
1261 __mark_reg64_unbounded(reg);
1262 __update_reg_bounds(reg);
1265 /* Intersecting with the old var_off might have improved our bounds
1266 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1267 * then new var_off is (0; 0x7f...fc) which improves our umax.
1269 __reg_deduce_bounds(reg);
1270 __reg_bound_offset(reg);
1271 __update_reg_bounds(reg);
1274 static bool __reg64_bound_s32(s64 a)
1276 if (a > S32_MIN && a < S32_MAX)
1281 static bool __reg64_bound_u32(u64 a)
1283 if (a > U32_MIN && a < U32_MAX)
1288 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
1290 __mark_reg32_unbounded(reg);
1292 if (__reg64_bound_s32(reg->smin_value))
1293 reg->s32_min_value = (s32)reg->smin_value;
1294 if (__reg64_bound_s32(reg->smax_value))
1295 reg->s32_max_value = (s32)reg->smax_value;
1296 if (__reg64_bound_u32(reg->umin_value))
1297 reg->u32_min_value = (u32)reg->umin_value;
1298 if (__reg64_bound_u32(reg->umax_value))
1299 reg->u32_max_value = (u32)reg->umax_value;
1301 /* Intersecting with the old var_off might have improved our bounds
1302 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1303 * then new var_off is (0; 0x7f...fc) which improves our umax.
1305 __reg_deduce_bounds(reg);
1306 __reg_bound_offset(reg);
1307 __update_reg_bounds(reg);
1310 /* Mark a register as having a completely unknown (scalar) value. */
1311 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1312 struct bpf_reg_state *reg)
1315 * Clear type, id, off, and union(map_ptr, range) and
1316 * padding between 'type' and union
1318 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1319 reg->type = SCALAR_VALUE;
1320 reg->var_off = tnum_unknown;
1322 reg->precise = env->subprog_cnt > 1 || !env->bpf_capable;
1323 __mark_reg_unbounded(reg);
1326 static void mark_reg_unknown(struct bpf_verifier_env *env,
1327 struct bpf_reg_state *regs, u32 regno)
1329 if (WARN_ON(regno >= MAX_BPF_REG)) {
1330 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1331 /* Something bad happened, let's kill all regs except FP */
1332 for (regno = 0; regno < BPF_REG_FP; regno++)
1333 __mark_reg_not_init(env, regs + regno);
1336 __mark_reg_unknown(env, regs + regno);
1339 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1340 struct bpf_reg_state *reg)
1342 __mark_reg_unknown(env, reg);
1343 reg->type = NOT_INIT;
1346 static void mark_reg_not_init(struct bpf_verifier_env *env,
1347 struct bpf_reg_state *regs, u32 regno)
1349 if (WARN_ON(regno >= MAX_BPF_REG)) {
1350 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1351 /* Something bad happened, let's kill all regs except FP */
1352 for (regno = 0; regno < BPF_REG_FP; regno++)
1353 __mark_reg_not_init(env, regs + regno);
1356 __mark_reg_not_init(env, regs + regno);
1359 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
1360 struct bpf_reg_state *regs, u32 regno,
1361 enum bpf_reg_type reg_type, u32 btf_id)
1363 if (reg_type == SCALAR_VALUE) {
1364 mark_reg_unknown(env, regs, regno);
1367 mark_reg_known_zero(env, regs, regno);
1368 regs[regno].type = PTR_TO_BTF_ID;
1369 regs[regno].btf_id = btf_id;
1372 #define DEF_NOT_SUBREG (0)
1373 static void init_reg_state(struct bpf_verifier_env *env,
1374 struct bpf_func_state *state)
1376 struct bpf_reg_state *regs = state->regs;
1379 for (i = 0; i < MAX_BPF_REG; i++) {
1380 mark_reg_not_init(env, regs, i);
1381 regs[i].live = REG_LIVE_NONE;
1382 regs[i].parent = NULL;
1383 regs[i].subreg_def = DEF_NOT_SUBREG;
1387 regs[BPF_REG_FP].type = PTR_TO_STACK;
1388 mark_reg_known_zero(env, regs, BPF_REG_FP);
1389 regs[BPF_REG_FP].frameno = state->frameno;
1392 #define BPF_MAIN_FUNC (-1)
1393 static void init_func_state(struct bpf_verifier_env *env,
1394 struct bpf_func_state *state,
1395 int callsite, int frameno, int subprogno)
1397 state->callsite = callsite;
1398 state->frameno = frameno;
1399 state->subprogno = subprogno;
1400 init_reg_state(env, state);
1404 SRC_OP, /* register is used as source operand */
1405 DST_OP, /* register is used as destination operand */
1406 DST_OP_NO_MARK /* same as above, check only, don't mark */
1409 static int cmp_subprogs(const void *a, const void *b)
1411 return ((struct bpf_subprog_info *)a)->start -
1412 ((struct bpf_subprog_info *)b)->start;
1415 static int find_subprog(struct bpf_verifier_env *env, int off)
1417 struct bpf_subprog_info *p;
1419 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1420 sizeof(env->subprog_info[0]), cmp_subprogs);
1423 return p - env->subprog_info;
1427 static int add_subprog(struct bpf_verifier_env *env, int off)
1429 int insn_cnt = env->prog->len;
1432 if (off >= insn_cnt || off < 0) {
1433 verbose(env, "call to invalid destination\n");
1436 ret = find_subprog(env, off);
1439 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1440 verbose(env, "too many subprograms\n");
1443 env->subprog_info[env->subprog_cnt++].start = off;
1444 sort(env->subprog_info, env->subprog_cnt,
1445 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1449 static int check_subprogs(struct bpf_verifier_env *env)
1451 int i, ret, subprog_start, subprog_end, off, cur_subprog = 0;
1452 struct bpf_subprog_info *subprog = env->subprog_info;
1453 struct bpf_insn *insn = env->prog->insnsi;
1454 int insn_cnt = env->prog->len;
1456 /* Add entry function. */
1457 ret = add_subprog(env, 0);
1461 /* determine subprog starts. The end is one before the next starts */
1462 for (i = 0; i < insn_cnt; i++) {
1463 if (insn[i].code != (BPF_JMP | BPF_CALL))
1465 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1467 if (!env->bpf_capable) {
1469 "function calls to other bpf functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
1472 ret = add_subprog(env, i + insn[i].imm + 1);
1477 /* Add a fake 'exit' subprog which could simplify subprog iteration
1478 * logic. 'subprog_cnt' should not be increased.
1480 subprog[env->subprog_cnt].start = insn_cnt;
1482 if (env->log.level & BPF_LOG_LEVEL2)
1483 for (i = 0; i < env->subprog_cnt; i++)
1484 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1486 /* now check that all jumps are within the same subprog */
1487 subprog_start = subprog[cur_subprog].start;
1488 subprog_end = subprog[cur_subprog + 1].start;
1489 for (i = 0; i < insn_cnt; i++) {
1490 u8 code = insn[i].code;
1492 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
1494 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1496 off = i + insn[i].off + 1;
1497 if (off < subprog_start || off >= subprog_end) {
1498 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1502 if (i == subprog_end - 1) {
1503 /* to avoid fall-through from one subprog into another
1504 * the last insn of the subprog should be either exit
1505 * or unconditional jump back
1507 if (code != (BPF_JMP | BPF_EXIT) &&
1508 code != (BPF_JMP | BPF_JA)) {
1509 verbose(env, "last insn is not an exit or jmp\n");
1512 subprog_start = subprog_end;
1514 if (cur_subprog < env->subprog_cnt)
1515 subprog_end = subprog[cur_subprog + 1].start;
1521 /* Parentage chain of this register (or stack slot) should take care of all
1522 * issues like callee-saved registers, stack slot allocation time, etc.
1524 static int mark_reg_read(struct bpf_verifier_env *env,
1525 const struct bpf_reg_state *state,
1526 struct bpf_reg_state *parent, u8 flag)
1528 bool writes = parent == state->parent; /* Observe write marks */
1532 /* if read wasn't screened by an earlier write ... */
1533 if (writes && state->live & REG_LIVE_WRITTEN)
1535 if (parent->live & REG_LIVE_DONE) {
1536 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1537 reg_type_str[parent->type],
1538 parent->var_off.value, parent->off);
1541 /* The first condition is more likely to be true than the
1542 * second, checked it first.
1544 if ((parent->live & REG_LIVE_READ) == flag ||
1545 parent->live & REG_LIVE_READ64)
1546 /* The parentage chain never changes and
1547 * this parent was already marked as LIVE_READ.
1548 * There is no need to keep walking the chain again and
1549 * keep re-marking all parents as LIVE_READ.
1550 * This case happens when the same register is read
1551 * multiple times without writes into it in-between.
1552 * Also, if parent has the stronger REG_LIVE_READ64 set,
1553 * then no need to set the weak REG_LIVE_READ32.
1556 /* ... then we depend on parent's value */
1557 parent->live |= flag;
1558 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
1559 if (flag == REG_LIVE_READ64)
1560 parent->live &= ~REG_LIVE_READ32;
1562 parent = state->parent;
1567 if (env->longest_mark_read_walk < cnt)
1568 env->longest_mark_read_walk = cnt;
1572 /* This function is supposed to be used by the following 32-bit optimization
1573 * code only. It returns TRUE if the source or destination register operates
1574 * on 64-bit, otherwise return FALSE.
1576 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
1577 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
1582 class = BPF_CLASS(code);
1584 if (class == BPF_JMP) {
1585 /* BPF_EXIT for "main" will reach here. Return TRUE
1590 if (op == BPF_CALL) {
1591 /* BPF to BPF call will reach here because of marking
1592 * caller saved clobber with DST_OP_NO_MARK for which we
1593 * don't care the register def because they are anyway
1594 * marked as NOT_INIT already.
1596 if (insn->src_reg == BPF_PSEUDO_CALL)
1598 /* Helper call will reach here because of arg type
1599 * check, conservatively return TRUE.
1608 if (class == BPF_ALU64 || class == BPF_JMP ||
1609 /* BPF_END always use BPF_ALU class. */
1610 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
1613 if (class == BPF_ALU || class == BPF_JMP32)
1616 if (class == BPF_LDX) {
1618 return BPF_SIZE(code) == BPF_DW;
1619 /* LDX source must be ptr. */
1623 if (class == BPF_STX) {
1624 if (reg->type != SCALAR_VALUE)
1626 return BPF_SIZE(code) == BPF_DW;
1629 if (class == BPF_LD) {
1630 u8 mode = BPF_MODE(code);
1633 if (mode == BPF_IMM)
1636 /* Both LD_IND and LD_ABS return 32-bit data. */
1640 /* Implicit ctx ptr. */
1641 if (regno == BPF_REG_6)
1644 /* Explicit source could be any width. */
1648 if (class == BPF_ST)
1649 /* The only source register for BPF_ST is a ptr. */
1652 /* Conservatively return true at default. */
1656 /* Return TRUE if INSN doesn't have explicit value define. */
1657 static bool insn_no_def(struct bpf_insn *insn)
1659 u8 class = BPF_CLASS(insn->code);
1661 return (class == BPF_JMP || class == BPF_JMP32 ||
1662 class == BPF_STX || class == BPF_ST);
1665 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
1666 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
1668 if (insn_no_def(insn))
1671 return !is_reg64(env, insn, insn->dst_reg, NULL, DST_OP);
1674 static void mark_insn_zext(struct bpf_verifier_env *env,
1675 struct bpf_reg_state *reg)
1677 s32 def_idx = reg->subreg_def;
1679 if (def_idx == DEF_NOT_SUBREG)
1682 env->insn_aux_data[def_idx - 1].zext_dst = true;
1683 /* The dst will be zero extended, so won't be sub-register anymore. */
1684 reg->subreg_def = DEF_NOT_SUBREG;
1687 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
1688 enum reg_arg_type t)
1690 struct bpf_verifier_state *vstate = env->cur_state;
1691 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1692 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
1693 struct bpf_reg_state *reg, *regs = state->regs;
1696 if (regno >= MAX_BPF_REG) {
1697 verbose(env, "R%d is invalid\n", regno);
1702 rw64 = is_reg64(env, insn, regno, reg, t);
1704 /* check whether register used as source operand can be read */
1705 if (reg->type == NOT_INIT) {
1706 verbose(env, "R%d !read_ok\n", regno);
1709 /* We don't need to worry about FP liveness because it's read-only */
1710 if (regno == BPF_REG_FP)
1714 mark_insn_zext(env, reg);
1716 return mark_reg_read(env, reg, reg->parent,
1717 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
1719 /* check whether register used as dest operand can be written to */
1720 if (regno == BPF_REG_FP) {
1721 verbose(env, "frame pointer is read only\n");
1724 reg->live |= REG_LIVE_WRITTEN;
1725 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
1727 mark_reg_unknown(env, regs, regno);
1732 /* for any branch, call, exit record the history of jmps in the given state */
1733 static int push_jmp_history(struct bpf_verifier_env *env,
1734 struct bpf_verifier_state *cur)
1736 u32 cnt = cur->jmp_history_cnt;
1737 struct bpf_idx_pair *p;
1740 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
1743 p[cnt - 1].idx = env->insn_idx;
1744 p[cnt - 1].prev_idx = env->prev_insn_idx;
1745 cur->jmp_history = p;
1746 cur->jmp_history_cnt = cnt;
1750 /* Backtrack one insn at a time. If idx is not at the top of recorded
1751 * history then previous instruction came from straight line execution.
1753 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
1758 if (cnt && st->jmp_history[cnt - 1].idx == i) {
1759 i = st->jmp_history[cnt - 1].prev_idx;
1767 /* For given verifier state backtrack_insn() is called from the last insn to
1768 * the first insn. Its purpose is to compute a bitmask of registers and
1769 * stack slots that needs precision in the parent verifier state.
1771 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
1772 u32 *reg_mask, u64 *stack_mask)
1774 const struct bpf_insn_cbs cbs = {
1775 .cb_print = verbose,
1776 .private_data = env,
1778 struct bpf_insn *insn = env->prog->insnsi + idx;
1779 u8 class = BPF_CLASS(insn->code);
1780 u8 opcode = BPF_OP(insn->code);
1781 u8 mode = BPF_MODE(insn->code);
1782 u32 dreg = 1u << insn->dst_reg;
1783 u32 sreg = 1u << insn->src_reg;
1786 if (insn->code == 0)
1788 if (env->log.level & BPF_LOG_LEVEL) {
1789 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
1790 verbose(env, "%d: ", idx);
1791 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
1794 if (class == BPF_ALU || class == BPF_ALU64) {
1795 if (!(*reg_mask & dreg))
1797 if (opcode == BPF_MOV) {
1798 if (BPF_SRC(insn->code) == BPF_X) {
1800 * dreg needs precision after this insn
1801 * sreg needs precision before this insn
1807 * dreg needs precision after this insn.
1808 * Corresponding register is already marked
1809 * as precise=true in this verifier state.
1810 * No further markings in parent are necessary
1815 if (BPF_SRC(insn->code) == BPF_X) {
1817 * both dreg and sreg need precision
1822 * dreg still needs precision before this insn
1825 } else if (class == BPF_LDX) {
1826 if (!(*reg_mask & dreg))
1830 /* scalars can only be spilled into stack w/o losing precision.
1831 * Load from any other memory can be zero extended.
1832 * The desire to keep that precision is already indicated
1833 * by 'precise' mark in corresponding register of this state.
1834 * No further tracking necessary.
1836 if (insn->src_reg != BPF_REG_FP)
1838 if (BPF_SIZE(insn->code) != BPF_DW)
1841 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
1842 * that [fp - off] slot contains scalar that needs to be
1843 * tracked with precision
1845 spi = (-insn->off - 1) / BPF_REG_SIZE;
1847 verbose(env, "BUG spi %d\n", spi);
1848 WARN_ONCE(1, "verifier backtracking bug");
1851 *stack_mask |= 1ull << spi;
1852 } else if (class == BPF_STX || class == BPF_ST) {
1853 if (*reg_mask & dreg)
1854 /* stx & st shouldn't be using _scalar_ dst_reg
1855 * to access memory. It means backtracking
1856 * encountered a case of pointer subtraction.
1859 /* scalars can only be spilled into stack */
1860 if (insn->dst_reg != BPF_REG_FP)
1862 if (BPF_SIZE(insn->code) != BPF_DW)
1864 spi = (-insn->off - 1) / BPF_REG_SIZE;
1866 verbose(env, "BUG spi %d\n", spi);
1867 WARN_ONCE(1, "verifier backtracking bug");
1870 if (!(*stack_mask & (1ull << spi)))
1872 *stack_mask &= ~(1ull << spi);
1873 if (class == BPF_STX)
1875 } else if (class == BPF_JMP || class == BPF_JMP32) {
1876 if (opcode == BPF_CALL) {
1877 if (insn->src_reg == BPF_PSEUDO_CALL)
1879 /* regular helper call sets R0 */
1881 if (*reg_mask & 0x3f) {
1882 /* if backtracing was looking for registers R1-R5
1883 * they should have been found already.
1885 verbose(env, "BUG regs %x\n", *reg_mask);
1886 WARN_ONCE(1, "verifier backtracking bug");
1889 } else if (opcode == BPF_EXIT) {
1892 } else if (class == BPF_LD) {
1893 if (!(*reg_mask & dreg))
1896 /* It's ld_imm64 or ld_abs or ld_ind.
1897 * For ld_imm64 no further tracking of precision
1898 * into parent is necessary
1900 if (mode == BPF_IND || mode == BPF_ABS)
1901 /* to be analyzed */
1907 /* the scalar precision tracking algorithm:
1908 * . at the start all registers have precise=false.
1909 * . scalar ranges are tracked as normal through alu and jmp insns.
1910 * . once precise value of the scalar register is used in:
1911 * . ptr + scalar alu
1912 * . if (scalar cond K|scalar)
1913 * . helper_call(.., scalar, ...) where ARG_CONST is expected
1914 * backtrack through the verifier states and mark all registers and
1915 * stack slots with spilled constants that these scalar regisers
1916 * should be precise.
1917 * . during state pruning two registers (or spilled stack slots)
1918 * are equivalent if both are not precise.
1920 * Note the verifier cannot simply walk register parentage chain,
1921 * since many different registers and stack slots could have been
1922 * used to compute single precise scalar.
1924 * The approach of starting with precise=true for all registers and then
1925 * backtrack to mark a register as not precise when the verifier detects
1926 * that program doesn't care about specific value (e.g., when helper
1927 * takes register as ARG_ANYTHING parameter) is not safe.
1929 * It's ok to walk single parentage chain of the verifier states.
1930 * It's possible that this backtracking will go all the way till 1st insn.
1931 * All other branches will be explored for needing precision later.
1933 * The backtracking needs to deal with cases like:
1934 * 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)
1937 * if r5 > 0x79f goto pc+7
1938 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
1941 * call bpf_perf_event_output#25
1942 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
1946 * call foo // uses callee's r6 inside to compute r0
1950 * to track above reg_mask/stack_mask needs to be independent for each frame.
1952 * Also if parent's curframe > frame where backtracking started,
1953 * the verifier need to mark registers in both frames, otherwise callees
1954 * may incorrectly prune callers. This is similar to
1955 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
1957 * For now backtracking falls back into conservative marking.
1959 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
1960 struct bpf_verifier_state *st)
1962 struct bpf_func_state *func;
1963 struct bpf_reg_state *reg;
1966 /* big hammer: mark all scalars precise in this path.
1967 * pop_stack may still get !precise scalars.
1969 for (; st; st = st->parent)
1970 for (i = 0; i <= st->curframe; i++) {
1971 func = st->frame[i];
1972 for (j = 0; j < BPF_REG_FP; j++) {
1973 reg = &func->regs[j];
1974 if (reg->type != SCALAR_VALUE)
1976 reg->precise = true;
1978 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
1979 if (func->stack[j].slot_type[0] != STACK_SPILL)
1981 reg = &func->stack[j].spilled_ptr;
1982 if (reg->type != SCALAR_VALUE)
1984 reg->precise = true;
1989 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
1992 struct bpf_verifier_state *st = env->cur_state;
1993 int first_idx = st->first_insn_idx;
1994 int last_idx = env->insn_idx;
1995 struct bpf_func_state *func;
1996 struct bpf_reg_state *reg;
1997 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
1998 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
1999 bool skip_first = true;
2000 bool new_marks = false;
2003 if (!env->bpf_capable)
2006 func = st->frame[st->curframe];
2008 reg = &func->regs[regno];
2009 if (reg->type != SCALAR_VALUE) {
2010 WARN_ONCE(1, "backtracing misuse");
2017 reg->precise = true;
2021 if (func->stack[spi].slot_type[0] != STACK_SPILL) {
2025 reg = &func->stack[spi].spilled_ptr;
2026 if (reg->type != SCALAR_VALUE) {
2034 reg->precise = true;
2040 if (!reg_mask && !stack_mask)
2043 DECLARE_BITMAP(mask, 64);
2044 u32 history = st->jmp_history_cnt;
2046 if (env->log.level & BPF_LOG_LEVEL)
2047 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
2048 for (i = last_idx;;) {
2053 err = backtrack_insn(env, i, ®_mask, &stack_mask);
2055 if (err == -ENOTSUPP) {
2056 mark_all_scalars_precise(env, st);
2061 if (!reg_mask && !stack_mask)
2062 /* Found assignment(s) into tracked register in this state.
2063 * Since this state is already marked, just return.
2064 * Nothing to be tracked further in the parent state.
2069 i = get_prev_insn_idx(st, i, &history);
2070 if (i >= env->prog->len) {
2071 /* This can happen if backtracking reached insn 0
2072 * and there are still reg_mask or stack_mask
2074 * It means the backtracking missed the spot where
2075 * particular register was initialized with a constant.
2077 verbose(env, "BUG backtracking idx %d\n", i);
2078 WARN_ONCE(1, "verifier backtracking bug");
2087 func = st->frame[st->curframe];
2088 bitmap_from_u64(mask, reg_mask);
2089 for_each_set_bit(i, mask, 32) {
2090 reg = &func->regs[i];
2091 if (reg->type != SCALAR_VALUE) {
2092 reg_mask &= ~(1u << i);
2097 reg->precise = true;
2100 bitmap_from_u64(mask, stack_mask);
2101 for_each_set_bit(i, mask, 64) {
2102 if (i >= func->allocated_stack / BPF_REG_SIZE) {
2103 /* the sequence of instructions:
2105 * 3: (7b) *(u64 *)(r3 -8) = r0
2106 * 4: (79) r4 = *(u64 *)(r10 -8)
2107 * doesn't contain jmps. It's backtracked
2108 * as a single block.
2109 * During backtracking insn 3 is not recognized as
2110 * stack access, so at the end of backtracking
2111 * stack slot fp-8 is still marked in stack_mask.
2112 * However the parent state may not have accessed
2113 * fp-8 and it's "unallocated" stack space.
2114 * In such case fallback to conservative.
2116 mark_all_scalars_precise(env, st);
2120 if (func->stack[i].slot_type[0] != STACK_SPILL) {
2121 stack_mask &= ~(1ull << i);
2124 reg = &func->stack[i].spilled_ptr;
2125 if (reg->type != SCALAR_VALUE) {
2126 stack_mask &= ~(1ull << i);
2131 reg->precise = true;
2133 if (env->log.level & BPF_LOG_LEVEL) {
2134 print_verifier_state(env, func);
2135 verbose(env, "parent %s regs=%x stack=%llx marks\n",
2136 new_marks ? "didn't have" : "already had",
2137 reg_mask, stack_mask);
2140 if (!reg_mask && !stack_mask)
2145 last_idx = st->last_insn_idx;
2146 first_idx = st->first_insn_idx;
2151 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
2153 return __mark_chain_precision(env, regno, -1);
2156 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
2158 return __mark_chain_precision(env, -1, spi);
2161 static bool is_spillable_regtype(enum bpf_reg_type type)
2164 case PTR_TO_MAP_VALUE:
2165 case PTR_TO_MAP_VALUE_OR_NULL:
2169 case PTR_TO_PACKET_META:
2170 case PTR_TO_PACKET_END:
2171 case PTR_TO_FLOW_KEYS:
2172 case CONST_PTR_TO_MAP:
2174 case PTR_TO_SOCKET_OR_NULL:
2175 case PTR_TO_SOCK_COMMON:
2176 case PTR_TO_SOCK_COMMON_OR_NULL:
2177 case PTR_TO_TCP_SOCK:
2178 case PTR_TO_TCP_SOCK_OR_NULL:
2179 case PTR_TO_XDP_SOCK:
2181 case PTR_TO_BTF_ID_OR_NULL:
2182 case PTR_TO_RDONLY_BUF:
2183 case PTR_TO_RDONLY_BUF_OR_NULL:
2184 case PTR_TO_RDWR_BUF:
2185 case PTR_TO_RDWR_BUF_OR_NULL:
2192 /* Does this register contain a constant zero? */
2193 static bool register_is_null(struct bpf_reg_state *reg)
2195 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
2198 static bool register_is_const(struct bpf_reg_state *reg)
2200 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
2203 static bool __is_pointer_value(bool allow_ptr_leaks,
2204 const struct bpf_reg_state *reg)
2206 if (allow_ptr_leaks)
2209 return reg->type != SCALAR_VALUE;
2212 static void save_register_state(struct bpf_func_state *state,
2213 int spi, struct bpf_reg_state *reg)
2217 state->stack[spi].spilled_ptr = *reg;
2218 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2220 for (i = 0; i < BPF_REG_SIZE; i++)
2221 state->stack[spi].slot_type[i] = STACK_SPILL;
2224 /* check_stack_read/write functions track spill/fill of registers,
2225 * stack boundary and alignment are checked in check_mem_access()
2227 static int check_stack_write(struct bpf_verifier_env *env,
2228 struct bpf_func_state *state, /* func where register points to */
2229 int off, int size, int value_regno, int insn_idx)
2231 struct bpf_func_state *cur; /* state of the current function */
2232 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
2233 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
2234 struct bpf_reg_state *reg = NULL;
2236 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
2237 state->acquired_refs, true);
2240 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
2241 * so it's aligned access and [off, off + size) are within stack limits
2243 if (!env->allow_ptr_leaks &&
2244 state->stack[spi].slot_type[0] == STACK_SPILL &&
2245 size != BPF_REG_SIZE) {
2246 verbose(env, "attempt to corrupt spilled pointer on stack\n");
2250 cur = env->cur_state->frame[env->cur_state->curframe];
2251 if (value_regno >= 0)
2252 reg = &cur->regs[value_regno];
2254 if (reg && size == BPF_REG_SIZE && register_is_const(reg) &&
2255 !register_is_null(reg) && env->bpf_capable) {
2256 if (dst_reg != BPF_REG_FP) {
2257 /* The backtracking logic can only recognize explicit
2258 * stack slot address like [fp - 8]. Other spill of
2259 * scalar via different register has to be conervative.
2260 * Backtrack from here and mark all registers as precise
2261 * that contributed into 'reg' being a constant.
2263 err = mark_chain_precision(env, value_regno);
2267 save_register_state(state, spi, reg);
2268 } else if (reg && is_spillable_regtype(reg->type)) {
2269 /* register containing pointer is being spilled into stack */
2270 if (size != BPF_REG_SIZE) {
2271 verbose_linfo(env, insn_idx, "; ");
2272 verbose(env, "invalid size of register spill\n");
2276 if (state != cur && reg->type == PTR_TO_STACK) {
2277 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
2281 if (!env->bypass_spec_v4) {
2282 bool sanitize = false;
2284 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
2285 register_is_const(&state->stack[spi].spilled_ptr))
2287 for (i = 0; i < BPF_REG_SIZE; i++)
2288 if (state->stack[spi].slot_type[i] == STACK_MISC) {
2293 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
2294 int soff = (-spi - 1) * BPF_REG_SIZE;
2296 /* detected reuse of integer stack slot with a pointer
2297 * which means either llvm is reusing stack slot or
2298 * an attacker is trying to exploit CVE-2018-3639
2299 * (speculative store bypass)
2300 * Have to sanitize that slot with preemptive
2303 if (*poff && *poff != soff) {
2304 /* disallow programs where single insn stores
2305 * into two different stack slots, since verifier
2306 * cannot sanitize them
2309 "insn %d cannot access two stack slots fp%d and fp%d",
2310 insn_idx, *poff, soff);
2316 save_register_state(state, spi, reg);
2318 u8 type = STACK_MISC;
2320 /* regular write of data into stack destroys any spilled ptr */
2321 state->stack[spi].spilled_ptr.type = NOT_INIT;
2322 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
2323 if (state->stack[spi].slot_type[0] == STACK_SPILL)
2324 for (i = 0; i < BPF_REG_SIZE; i++)
2325 state->stack[spi].slot_type[i] = STACK_MISC;
2327 /* only mark the slot as written if all 8 bytes were written
2328 * otherwise read propagation may incorrectly stop too soon
2329 * when stack slots are partially written.
2330 * This heuristic means that read propagation will be
2331 * conservative, since it will add reg_live_read marks
2332 * to stack slots all the way to first state when programs
2333 * writes+reads less than 8 bytes
2335 if (size == BPF_REG_SIZE)
2336 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2338 /* when we zero initialize stack slots mark them as such */
2339 if (reg && register_is_null(reg)) {
2340 /* backtracking doesn't work for STACK_ZERO yet. */
2341 err = mark_chain_precision(env, value_regno);
2347 /* Mark slots affected by this stack write. */
2348 for (i = 0; i < size; i++)
2349 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
2355 static int check_stack_read(struct bpf_verifier_env *env,
2356 struct bpf_func_state *reg_state /* func where register points to */,
2357 int off, int size, int value_regno)
2359 struct bpf_verifier_state *vstate = env->cur_state;
2360 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2361 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
2362 struct bpf_reg_state *reg;
2365 if (reg_state->allocated_stack <= slot) {
2366 verbose(env, "invalid read from stack off %d+0 size %d\n",
2370 stype = reg_state->stack[spi].slot_type;
2371 reg = ®_state->stack[spi].spilled_ptr;
2373 if (stype[0] == STACK_SPILL) {
2374 if (size != BPF_REG_SIZE) {
2375 if (reg->type != SCALAR_VALUE) {
2376 verbose_linfo(env, env->insn_idx, "; ");
2377 verbose(env, "invalid size of register fill\n");
2380 if (value_regno >= 0) {
2381 mark_reg_unknown(env, state->regs, value_regno);
2382 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2384 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2387 for (i = 1; i < BPF_REG_SIZE; i++) {
2388 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
2389 verbose(env, "corrupted spill memory\n");
2394 if (value_regno >= 0) {
2395 /* restore register state from stack */
2396 state->regs[value_regno] = *reg;
2397 /* mark reg as written since spilled pointer state likely
2398 * has its liveness marks cleared by is_state_visited()
2399 * which resets stack/reg liveness for state transitions
2401 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2402 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
2403 /* If value_regno==-1, the caller is asking us whether
2404 * it is acceptable to use this value as a SCALAR_VALUE
2406 * We must not allow unprivileged callers to do that
2407 * with spilled pointers.
2409 verbose(env, "leaking pointer from stack off %d\n",
2413 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2417 for (i = 0; i < size; i++) {
2418 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
2420 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
2424 verbose(env, "invalid read from stack off %d+%d size %d\n",
2428 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2429 if (value_regno >= 0) {
2430 if (zeros == size) {
2431 /* any size read into register is zero extended,
2432 * so the whole register == const_zero
2434 __mark_reg_const_zero(&state->regs[value_regno]);
2435 /* backtracking doesn't support STACK_ZERO yet,
2436 * so mark it precise here, so that later
2437 * backtracking can stop here.
2438 * Backtracking may not need this if this register
2439 * doesn't participate in pointer adjustment.
2440 * Forward propagation of precise flag is not
2441 * necessary either. This mark is only to stop
2442 * backtracking. Any register that contributed
2443 * to const 0 was marked precise before spill.
2445 state->regs[value_regno].precise = true;
2447 /* have read misc data from the stack */
2448 mark_reg_unknown(env, state->regs, value_regno);
2450 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2456 static int check_stack_access(struct bpf_verifier_env *env,
2457 const struct bpf_reg_state *reg,
2460 /* Stack accesses must be at a fixed offset, so that we
2461 * can determine what type of data were returned. See
2462 * check_stack_read().
2464 if (!tnum_is_const(reg->var_off)) {
2467 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2468 verbose(env, "variable stack access var_off=%s off=%d size=%d\n",
2473 if (off >= 0 || off < -MAX_BPF_STACK) {
2474 verbose(env, "invalid stack off=%d size=%d\n", off, size);
2481 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
2482 int off, int size, enum bpf_access_type type)
2484 struct bpf_reg_state *regs = cur_regs(env);
2485 struct bpf_map *map = regs[regno].map_ptr;
2486 u32 cap = bpf_map_flags_to_cap(map);
2488 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
2489 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
2490 map->value_size, off, size);
2494 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
2495 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
2496 map->value_size, off, size);
2503 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
2504 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
2505 int off, int size, u32 mem_size,
2506 bool zero_size_allowed)
2508 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
2509 struct bpf_reg_state *reg;
2511 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
2514 reg = &cur_regs(env)[regno];
2515 switch (reg->type) {
2516 case PTR_TO_MAP_VALUE:
2517 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
2518 mem_size, off, size);
2521 case PTR_TO_PACKET_META:
2522 case PTR_TO_PACKET_END:
2523 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
2524 off, size, regno, reg->id, off, mem_size);
2528 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
2529 mem_size, off, size);
2535 /* check read/write into a memory region with possible variable offset */
2536 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
2537 int off, int size, u32 mem_size,
2538 bool zero_size_allowed)
2540 struct bpf_verifier_state *vstate = env->cur_state;
2541 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2542 struct bpf_reg_state *reg = &state->regs[regno];
2545 /* We may have adjusted the register pointing to memory region, so we
2546 * need to try adding each of min_value and max_value to off
2547 * to make sure our theoretical access will be safe.
2549 if (env->log.level & BPF_LOG_LEVEL)
2550 print_verifier_state(env, state);
2552 /* The minimum value is only important with signed
2553 * comparisons where we can't assume the floor of a
2554 * value is 0. If we are using signed variables for our
2555 * index'es we need to make sure that whatever we use
2556 * will have a set floor within our range.
2558 if (reg->smin_value < 0 &&
2559 (reg->smin_value == S64_MIN ||
2560 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
2561 reg->smin_value + off < 0)) {
2562 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2566 err = __check_mem_access(env, regno, reg->smin_value + off, size,
2567 mem_size, zero_size_allowed);
2569 verbose(env, "R%d min value is outside of the allowed memory range\n",
2574 /* If we haven't set a max value then we need to bail since we can't be
2575 * sure we won't do bad things.
2576 * If reg->umax_value + off could overflow, treat that as unbounded too.
2578 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
2579 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
2583 err = __check_mem_access(env, regno, reg->umax_value + off, size,
2584 mem_size, zero_size_allowed);
2586 verbose(env, "R%d max value is outside of the allowed memory range\n",
2594 /* check read/write into a map element with possible variable offset */
2595 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
2596 int off, int size, bool zero_size_allowed)
2598 struct bpf_verifier_state *vstate = env->cur_state;
2599 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2600 struct bpf_reg_state *reg = &state->regs[regno];
2601 struct bpf_map *map = reg->map_ptr;
2604 err = check_mem_region_access(env, regno, off, size, map->value_size,
2609 if (map_value_has_spin_lock(map)) {
2610 u32 lock = map->spin_lock_off;
2612 /* if any part of struct bpf_spin_lock can be touched by
2613 * load/store reject this program.
2614 * To check that [x1, x2) overlaps with [y1, y2)
2615 * it is sufficient to check x1 < y2 && y1 < x2.
2617 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
2618 lock < reg->umax_value + off + size) {
2619 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
2626 #define MAX_PACKET_OFF 0xffff
2628 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
2629 const struct bpf_call_arg_meta *meta,
2630 enum bpf_access_type t)
2632 switch (env->prog->type) {
2633 /* Program types only with direct read access go here! */
2634 case BPF_PROG_TYPE_LWT_IN:
2635 case BPF_PROG_TYPE_LWT_OUT:
2636 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
2637 case BPF_PROG_TYPE_SK_REUSEPORT:
2638 case BPF_PROG_TYPE_FLOW_DISSECTOR:
2639 case BPF_PROG_TYPE_CGROUP_SKB:
2644 /* Program types with direct read + write access go here! */
2645 case BPF_PROG_TYPE_SCHED_CLS:
2646 case BPF_PROG_TYPE_SCHED_ACT:
2647 case BPF_PROG_TYPE_XDP:
2648 case BPF_PROG_TYPE_LWT_XMIT:
2649 case BPF_PROG_TYPE_SK_SKB:
2650 case BPF_PROG_TYPE_SK_MSG:
2652 return meta->pkt_access;
2654 env->seen_direct_write = true;
2657 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
2659 env->seen_direct_write = true;
2668 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
2669 int size, bool zero_size_allowed)
2671 struct bpf_reg_state *regs = cur_regs(env);
2672 struct bpf_reg_state *reg = ®s[regno];
2675 /* We may have added a variable offset to the packet pointer; but any
2676 * reg->range we have comes after that. We are only checking the fixed
2680 /* We don't allow negative numbers, because we aren't tracking enough
2681 * detail to prove they're safe.
2683 if (reg->smin_value < 0) {
2684 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2688 err = __check_mem_access(env, regno, off, size, reg->range,
2691 verbose(env, "R%d offset is outside of the packet\n", regno);
2695 /* __check_mem_access has made sure "off + size - 1" is within u16.
2696 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
2697 * otherwise find_good_pkt_pointers would have refused to set range info
2698 * that __check_mem_access would have rejected this pkt access.
2699 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
2701 env->prog->aux->max_pkt_offset =
2702 max_t(u32, env->prog->aux->max_pkt_offset,
2703 off + reg->umax_value + size - 1);
2708 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
2709 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
2710 enum bpf_access_type t, enum bpf_reg_type *reg_type,
2713 struct bpf_insn_access_aux info = {
2714 .reg_type = *reg_type,
2718 if (env->ops->is_valid_access &&
2719 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
2720 /* A non zero info.ctx_field_size indicates that this field is a
2721 * candidate for later verifier transformation to load the whole
2722 * field and then apply a mask when accessed with a narrower
2723 * access than actual ctx access size. A zero info.ctx_field_size
2724 * will only allow for whole field access and rejects any other
2725 * type of narrower access.
2727 *reg_type = info.reg_type;
2729 if (*reg_type == PTR_TO_BTF_ID || *reg_type == PTR_TO_BTF_ID_OR_NULL)
2730 *btf_id = info.btf_id;
2732 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
2733 /* remember the offset of last byte accessed in ctx */
2734 if (env->prog->aux->max_ctx_offset < off + size)
2735 env->prog->aux->max_ctx_offset = off + size;
2739 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
2743 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
2746 if (size < 0 || off < 0 ||
2747 (u64)off + size > sizeof(struct bpf_flow_keys)) {
2748 verbose(env, "invalid access to flow keys off=%d size=%d\n",
2755 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
2756 u32 regno, int off, int size,
2757 enum bpf_access_type t)
2759 struct bpf_reg_state *regs = cur_regs(env);
2760 struct bpf_reg_state *reg = ®s[regno];
2761 struct bpf_insn_access_aux info = {};
2764 if (reg->smin_value < 0) {
2765 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2770 switch (reg->type) {
2771 case PTR_TO_SOCK_COMMON:
2772 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
2775 valid = bpf_sock_is_valid_access(off, size, t, &info);
2777 case PTR_TO_TCP_SOCK:
2778 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
2780 case PTR_TO_XDP_SOCK:
2781 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
2789 env->insn_aux_data[insn_idx].ctx_field_size =
2790 info.ctx_field_size;
2794 verbose(env, "R%d invalid %s access off=%d size=%d\n",
2795 regno, reg_type_str[reg->type], off, size);
2800 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
2802 return cur_regs(env) + regno;
2805 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
2807 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
2810 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
2812 const struct bpf_reg_state *reg = reg_state(env, regno);
2814 return reg->type == PTR_TO_CTX;
2817 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
2819 const struct bpf_reg_state *reg = reg_state(env, regno);
2821 return type_is_sk_pointer(reg->type);
2824 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
2826 const struct bpf_reg_state *reg = reg_state(env, regno);
2828 return type_is_pkt_pointer(reg->type);
2831 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
2833 const struct bpf_reg_state *reg = reg_state(env, regno);
2835 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
2836 return reg->type == PTR_TO_FLOW_KEYS;
2839 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
2840 const struct bpf_reg_state *reg,
2841 int off, int size, bool strict)
2843 struct tnum reg_off;
2846 /* Byte size accesses are always allowed. */
2847 if (!strict || size == 1)
2850 /* For platforms that do not have a Kconfig enabling
2851 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
2852 * NET_IP_ALIGN is universally set to '2'. And on platforms
2853 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
2854 * to this code only in strict mode where we want to emulate
2855 * the NET_IP_ALIGN==2 checking. Therefore use an
2856 * unconditional IP align value of '2'.
2860 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
2861 if (!tnum_is_aligned(reg_off, size)) {
2864 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2866 "misaligned packet access off %d+%s+%d+%d size %d\n",
2867 ip_align, tn_buf, reg->off, off, size);
2874 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
2875 const struct bpf_reg_state *reg,
2876 const char *pointer_desc,
2877 int off, int size, bool strict)
2879 struct tnum reg_off;
2881 /* Byte size accesses are always allowed. */
2882 if (!strict || size == 1)
2885 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
2886 if (!tnum_is_aligned(reg_off, size)) {
2889 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2890 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
2891 pointer_desc, tn_buf, reg->off, off, size);
2898 static int check_ptr_alignment(struct bpf_verifier_env *env,
2899 const struct bpf_reg_state *reg, int off,
2900 int size, bool strict_alignment_once)
2902 bool strict = env->strict_alignment || strict_alignment_once;
2903 const char *pointer_desc = "";
2905 switch (reg->type) {
2907 case PTR_TO_PACKET_META:
2908 /* Special case, because of NET_IP_ALIGN. Given metadata sits
2909 * right in front, treat it the very same way.
2911 return check_pkt_ptr_alignment(env, reg, off, size, strict);
2912 case PTR_TO_FLOW_KEYS:
2913 pointer_desc = "flow keys ";
2915 case PTR_TO_MAP_VALUE:
2916 pointer_desc = "value ";
2919 pointer_desc = "context ";
2922 pointer_desc = "stack ";
2923 /* The stack spill tracking logic in check_stack_write()
2924 * and check_stack_read() relies on stack accesses being
2930 pointer_desc = "sock ";
2932 case PTR_TO_SOCK_COMMON:
2933 pointer_desc = "sock_common ";
2935 case PTR_TO_TCP_SOCK:
2936 pointer_desc = "tcp_sock ";
2938 case PTR_TO_XDP_SOCK:
2939 pointer_desc = "xdp_sock ";
2944 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
2948 static int update_stack_depth(struct bpf_verifier_env *env,
2949 const struct bpf_func_state *func,
2952 u16 stack = env->subprog_info[func->subprogno].stack_depth;
2957 /* update known max for given subprogram */
2958 env->subprog_info[func->subprogno].stack_depth = -off;
2962 /* starting from main bpf function walk all instructions of the function
2963 * and recursively walk all callees that given function can call.
2964 * Ignore jump and exit insns.
2965 * Since recursion is prevented by check_cfg() this algorithm
2966 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
2968 static int check_max_stack_depth(struct bpf_verifier_env *env)
2970 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
2971 struct bpf_subprog_info *subprog = env->subprog_info;
2972 struct bpf_insn *insn = env->prog->insnsi;
2973 int ret_insn[MAX_CALL_FRAMES];
2974 int ret_prog[MAX_CALL_FRAMES];
2977 /* round up to 32-bytes, since this is granularity
2978 * of interpreter stack size
2980 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
2981 if (depth > MAX_BPF_STACK) {
2982 verbose(env, "combined stack size of %d calls is %d. Too large\n",
2987 subprog_end = subprog[idx + 1].start;
2988 for (; i < subprog_end; i++) {
2989 if (insn[i].code != (BPF_JMP | BPF_CALL))
2991 if (insn[i].src_reg != BPF_PSEUDO_CALL)
2993 /* remember insn and function to return to */
2994 ret_insn[frame] = i + 1;
2995 ret_prog[frame] = idx;
2997 /* find the callee */
2998 i = i + insn[i].imm + 1;
2999 idx = find_subprog(env, i);
3001 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3006 if (frame >= MAX_CALL_FRAMES) {
3007 verbose(env, "the call stack of %d frames is too deep !\n",
3013 /* end of for() loop means the last insn of the 'subprog'
3014 * was reached. Doesn't matter whether it was JA or EXIT
3018 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3020 i = ret_insn[frame];
3021 idx = ret_prog[frame];
3025 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
3026 static int get_callee_stack_depth(struct bpf_verifier_env *env,
3027 const struct bpf_insn *insn, int idx)
3029 int start = idx + insn->imm + 1, subprog;
3031 subprog = find_subprog(env, start);
3033 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3037 return env->subprog_info[subprog].stack_depth;
3041 int check_ctx_reg(struct bpf_verifier_env *env,
3042 const struct bpf_reg_state *reg, int regno)
3044 /* Access to ctx or passing it to a helper is only allowed in
3045 * its original, unmodified form.
3049 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
3054 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3057 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3058 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
3065 static int __check_buffer_access(struct bpf_verifier_env *env,
3066 const char *buf_info,
3067 const struct bpf_reg_state *reg,
3068 int regno, int off, int size)
3072 "R%d invalid %s buffer access: off=%d, size=%d",
3073 regno, buf_info, off, size);
3076 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3079 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3081 "R%d invalid variable buffer offset: off=%d, var_off=%s",
3082 regno, off, tn_buf);
3089 static int check_tp_buffer_access(struct bpf_verifier_env *env,
3090 const struct bpf_reg_state *reg,
3091 int regno, int off, int size)
3095 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
3099 if (off + size > env->prog->aux->max_tp_access)
3100 env->prog->aux->max_tp_access = off + size;
3105 static int check_buffer_access(struct bpf_verifier_env *env,
3106 const struct bpf_reg_state *reg,
3107 int regno, int off, int size,
3108 bool zero_size_allowed,
3109 const char *buf_info,
3114 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
3118 if (off + size > *max_access)
3119 *max_access = off + size;
3124 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
3125 static void zext_32_to_64(struct bpf_reg_state *reg)
3127 reg->var_off = tnum_subreg(reg->var_off);
3128 __reg_assign_32_into_64(reg);
3131 /* truncate register to smaller size (in bytes)
3132 * must be called with size < BPF_REG_SIZE
3134 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
3138 /* clear high bits in bit representation */
3139 reg->var_off = tnum_cast(reg->var_off, size);
3141 /* fix arithmetic bounds */
3142 mask = ((u64)1 << (size * 8)) - 1;
3143 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
3144 reg->umin_value &= mask;
3145 reg->umax_value &= mask;
3147 reg->umin_value = 0;
3148 reg->umax_value = mask;
3150 reg->smin_value = reg->umin_value;
3151 reg->smax_value = reg->umax_value;
3153 /* If size is smaller than 32bit register the 32bit register
3154 * values are also truncated so we push 64-bit bounds into
3155 * 32-bit bounds. Above were truncated < 32-bits already.
3159 __reg_combine_64_into_32(reg);
3162 static bool bpf_map_is_rdonly(const struct bpf_map *map)
3164 return (map->map_flags & BPF_F_RDONLY_PROG) && map->frozen;
3167 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
3173 err = map->ops->map_direct_value_addr(map, &addr, off);
3176 ptr = (void *)(long)addr + off;
3180 *val = (u64)*(u8 *)ptr;
3183 *val = (u64)*(u16 *)ptr;
3186 *val = (u64)*(u32 *)ptr;
3197 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
3198 struct bpf_reg_state *regs,
3199 int regno, int off, int size,
3200 enum bpf_access_type atype,
3203 struct bpf_reg_state *reg = regs + regno;
3204 const struct btf_type *t = btf_type_by_id(btf_vmlinux, reg->btf_id);
3205 const char *tname = btf_name_by_offset(btf_vmlinux, t->name_off);
3211 "R%d is ptr_%s invalid negative access: off=%d\n",
3215 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3218 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3220 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
3221 regno, tname, off, tn_buf);
3225 if (env->ops->btf_struct_access) {
3226 ret = env->ops->btf_struct_access(&env->log, t, off, size,
3229 if (atype != BPF_READ) {
3230 verbose(env, "only read is supported\n");
3234 ret = btf_struct_access(&env->log, t, off, size, atype,
3241 if (atype == BPF_READ && value_regno >= 0)
3242 mark_btf_ld_reg(env, regs, value_regno, ret, btf_id);
3247 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
3248 struct bpf_reg_state *regs,
3249 int regno, int off, int size,
3250 enum bpf_access_type atype,
3253 struct bpf_reg_state *reg = regs + regno;
3254 struct bpf_map *map = reg->map_ptr;
3255 const struct btf_type *t;
3261 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
3265 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
3266 verbose(env, "map_ptr access not supported for map type %d\n",
3271 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
3272 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
3274 if (!env->allow_ptr_to_map_access) {
3276 "%s access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
3282 verbose(env, "R%d is %s invalid negative access: off=%d\n",
3287 if (atype != BPF_READ) {
3288 verbose(env, "only read from %s is supported\n", tname);
3292 ret = btf_struct_access(&env->log, t, off, size, atype, &btf_id);
3296 if (value_regno >= 0)
3297 mark_btf_ld_reg(env, regs, value_regno, ret, btf_id);
3303 /* check whether memory at (regno + off) is accessible for t = (read | write)
3304 * if t==write, value_regno is a register which value is stored into memory
3305 * if t==read, value_regno is a register which will receive the value from memory
3306 * if t==write && value_regno==-1, some unknown value is stored into memory
3307 * if t==read && value_regno==-1, don't care what we read from memory
3309 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
3310 int off, int bpf_size, enum bpf_access_type t,
3311 int value_regno, bool strict_alignment_once)
3313 struct bpf_reg_state *regs = cur_regs(env);
3314 struct bpf_reg_state *reg = regs + regno;
3315 struct bpf_func_state *state;
3318 size = bpf_size_to_bytes(bpf_size);
3322 /* alignment checks will add in reg->off themselves */
3323 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
3327 /* for access checks, reg->off is just part of off */
3330 if (reg->type == PTR_TO_MAP_VALUE) {
3331 if (t == BPF_WRITE && value_regno >= 0 &&
3332 is_pointer_value(env, value_regno)) {
3333 verbose(env, "R%d leaks addr into map\n", value_regno);
3336 err = check_map_access_type(env, regno, off, size, t);
3339 err = check_map_access(env, regno, off, size, false);
3340 if (!err && t == BPF_READ && value_regno >= 0) {
3341 struct bpf_map *map = reg->map_ptr;
3343 /* if map is read-only, track its contents as scalars */
3344 if (tnum_is_const(reg->var_off) &&
3345 bpf_map_is_rdonly(map) &&
3346 map->ops->map_direct_value_addr) {
3347 int map_off = off + reg->var_off.value;
3350 err = bpf_map_direct_read(map, map_off, size,
3355 regs[value_regno].type = SCALAR_VALUE;
3356 __mark_reg_known(®s[value_regno], val);
3358 mark_reg_unknown(env, regs, value_regno);
3361 } else if (reg->type == PTR_TO_MEM) {
3362 if (t == BPF_WRITE && value_regno >= 0 &&
3363 is_pointer_value(env, value_regno)) {
3364 verbose(env, "R%d leaks addr into mem\n", value_regno);
3367 err = check_mem_region_access(env, regno, off, size,
3368 reg->mem_size, false);
3369 if (!err && t == BPF_READ && value_regno >= 0)
3370 mark_reg_unknown(env, regs, value_regno);
3371 } else if (reg->type == PTR_TO_CTX) {
3372 enum bpf_reg_type reg_type = SCALAR_VALUE;
3375 if (t == BPF_WRITE && value_regno >= 0 &&
3376 is_pointer_value(env, value_regno)) {
3377 verbose(env, "R%d leaks addr into ctx\n", value_regno);
3381 err = check_ctx_reg(env, reg, regno);
3385 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf_id);
3387 verbose_linfo(env, insn_idx, "; ");
3388 if (!err && t == BPF_READ && value_regno >= 0) {
3389 /* ctx access returns either a scalar, or a
3390 * PTR_TO_PACKET[_META,_END]. In the latter
3391 * case, we know the offset is zero.
3393 if (reg_type == SCALAR_VALUE) {
3394 mark_reg_unknown(env, regs, value_regno);
3396 mark_reg_known_zero(env, regs,
3398 if (reg_type_may_be_null(reg_type))
3399 regs[value_regno].id = ++env->id_gen;
3400 /* A load of ctx field could have different
3401 * actual load size with the one encoded in the
3402 * insn. When the dst is PTR, it is for sure not
3405 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
3406 if (reg_type == PTR_TO_BTF_ID ||
3407 reg_type == PTR_TO_BTF_ID_OR_NULL)
3408 regs[value_regno].btf_id = btf_id;
3410 regs[value_regno].type = reg_type;
3413 } else if (reg->type == PTR_TO_STACK) {
3414 off += reg->var_off.value;
3415 err = check_stack_access(env, reg, off, size);
3419 state = func(env, reg);
3420 err = update_stack_depth(env, state, off);
3425 err = check_stack_write(env, state, off, size,
3426 value_regno, insn_idx);
3428 err = check_stack_read(env, state, off, size,
3430 } else if (reg_is_pkt_pointer(reg)) {
3431 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
3432 verbose(env, "cannot write into packet\n");
3435 if (t == BPF_WRITE && value_regno >= 0 &&
3436 is_pointer_value(env, value_regno)) {
3437 verbose(env, "R%d leaks addr into packet\n",
3441 err = check_packet_access(env, regno, off, size, false);
3442 if (!err && t == BPF_READ && value_regno >= 0)
3443 mark_reg_unknown(env, regs, value_regno);
3444 } else if (reg->type == PTR_TO_FLOW_KEYS) {
3445 if (t == BPF_WRITE && value_regno >= 0 &&
3446 is_pointer_value(env, value_regno)) {
3447 verbose(env, "R%d leaks addr into flow keys\n",
3452 err = check_flow_keys_access(env, off, size);
3453 if (!err && t == BPF_READ && value_regno >= 0)
3454 mark_reg_unknown(env, regs, value_regno);
3455 } else if (type_is_sk_pointer(reg->type)) {
3456 if (t == BPF_WRITE) {
3457 verbose(env, "R%d cannot write into %s\n",
3458 regno, reg_type_str[reg->type]);
3461 err = check_sock_access(env, insn_idx, regno, off, size, t);
3462 if (!err && value_regno >= 0)
3463 mark_reg_unknown(env, regs, value_regno);
3464 } else if (reg->type == PTR_TO_TP_BUFFER) {
3465 err = check_tp_buffer_access(env, reg, regno, off, size);
3466 if (!err && t == BPF_READ && value_regno >= 0)
3467 mark_reg_unknown(env, regs, value_regno);
3468 } else if (reg->type == PTR_TO_BTF_ID) {
3469 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
3471 } else if (reg->type == CONST_PTR_TO_MAP) {
3472 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
3474 } else if (reg->type == PTR_TO_RDONLY_BUF) {
3475 if (t == BPF_WRITE) {
3476 verbose(env, "R%d cannot write into %s\n",
3477 regno, reg_type_str[reg->type]);
3480 err = check_buffer_access(env, reg, regno, off, size, "rdonly",
3482 &env->prog->aux->max_rdonly_access);
3483 if (!err && value_regno >= 0)
3484 mark_reg_unknown(env, regs, value_regno);
3485 } else if (reg->type == PTR_TO_RDWR_BUF) {
3486 err = check_buffer_access(env, reg, regno, off, size, "rdwr",
3488 &env->prog->aux->max_rdwr_access);
3489 if (!err && t == BPF_READ && value_regno >= 0)
3490 mark_reg_unknown(env, regs, value_regno);
3492 verbose(env, "R%d invalid mem access '%s'\n", regno,
3493 reg_type_str[reg->type]);
3497 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
3498 regs[value_regno].type == SCALAR_VALUE) {
3499 /* b/h/w load zero-extends, mark upper bits as known 0 */
3500 coerce_reg_to_size(®s[value_regno], size);
3505 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
3509 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
3511 verbose(env, "BPF_XADD uses reserved fields\n");
3515 /* check src1 operand */
3516 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3520 /* check src2 operand */
3521 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3525 if (is_pointer_value(env, insn->src_reg)) {
3526 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
3530 if (is_ctx_reg(env, insn->dst_reg) ||
3531 is_pkt_reg(env, insn->dst_reg) ||
3532 is_flow_key_reg(env, insn->dst_reg) ||
3533 is_sk_reg(env, insn->dst_reg)) {
3534 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
3536 reg_type_str[reg_state(env, insn->dst_reg)->type]);
3540 /* check whether atomic_add can read the memory */
3541 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3542 BPF_SIZE(insn->code), BPF_READ, -1, true);
3546 /* check whether atomic_add can write into the same memory */
3547 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3548 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
3551 static int __check_stack_boundary(struct bpf_verifier_env *env, u32 regno,
3552 int off, int access_size,
3553 bool zero_size_allowed)
3555 struct bpf_reg_state *reg = reg_state(env, regno);
3557 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
3558 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
3559 if (tnum_is_const(reg->var_off)) {
3560 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
3561 regno, off, access_size);
3565 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3566 verbose(env, "invalid stack type R%d var_off=%s access_size=%d\n",
3567 regno, tn_buf, access_size);
3574 /* when register 'regno' is passed into function that will read 'access_size'
3575 * bytes from that pointer, make sure that it's within stack boundary
3576 * and all elements of stack are initialized.
3577 * Unlike most pointer bounds-checking functions, this one doesn't take an
3578 * 'off' argument, so it has to add in reg->off itself.
3580 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
3581 int access_size, bool zero_size_allowed,
3582 struct bpf_call_arg_meta *meta)
3584 struct bpf_reg_state *reg = reg_state(env, regno);
3585 struct bpf_func_state *state = func(env, reg);
3586 int err, min_off, max_off, i, j, slot, spi;
3588 if (reg->type != PTR_TO_STACK) {
3589 /* Allow zero-byte read from NULL, regardless of pointer type */
3590 if (zero_size_allowed && access_size == 0 &&
3591 register_is_null(reg))
3594 verbose(env, "R%d type=%s expected=%s\n", regno,
3595 reg_type_str[reg->type],
3596 reg_type_str[PTR_TO_STACK]);
3600 if (tnum_is_const(reg->var_off)) {
3601 min_off = max_off = reg->var_off.value + reg->off;
3602 err = __check_stack_boundary(env, regno, min_off, access_size,
3607 /* Variable offset is prohibited for unprivileged mode for
3608 * simplicity since it requires corresponding support in
3609 * Spectre masking for stack ALU.
3610 * See also retrieve_ptr_limit().
3612 if (!env->bypass_spec_v1) {
3615 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3616 verbose(env, "R%d indirect variable offset stack access prohibited for !root, var_off=%s\n",
3620 /* Only initialized buffer on stack is allowed to be accessed
3621 * with variable offset. With uninitialized buffer it's hard to
3622 * guarantee that whole memory is marked as initialized on
3623 * helper return since specific bounds are unknown what may
3624 * cause uninitialized stack leaking.
3626 if (meta && meta->raw_mode)
3629 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
3630 reg->smax_value <= -BPF_MAX_VAR_OFF) {
3631 verbose(env, "R%d unbounded indirect variable offset stack access\n",
3635 min_off = reg->smin_value + reg->off;
3636 max_off = reg->smax_value + reg->off;
3637 err = __check_stack_boundary(env, regno, min_off, access_size,
3640 verbose(env, "R%d min value is outside of stack bound\n",
3644 err = __check_stack_boundary(env, regno, max_off, access_size,
3647 verbose(env, "R%d max value is outside of stack bound\n",
3653 if (meta && meta->raw_mode) {
3654 meta->access_size = access_size;
3655 meta->regno = regno;
3659 for (i = min_off; i < max_off + access_size; i++) {
3663 spi = slot / BPF_REG_SIZE;
3664 if (state->allocated_stack <= slot)
3666 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
3667 if (*stype == STACK_MISC)
3669 if (*stype == STACK_ZERO) {
3670 /* helper can write anything into the stack */
3671 *stype = STACK_MISC;
3675 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
3676 state->stack[spi].spilled_ptr.type == PTR_TO_BTF_ID)
3679 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
3680 state->stack[spi].spilled_ptr.type == SCALAR_VALUE) {
3681 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
3682 for (j = 0; j < BPF_REG_SIZE; j++)
3683 state->stack[spi].slot_type[j] = STACK_MISC;
3688 if (tnum_is_const(reg->var_off)) {
3689 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
3690 min_off, i - min_off, access_size);
3694 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3695 verbose(env, "invalid indirect read from stack var_off %s+%d size %d\n",
3696 tn_buf, i - min_off, access_size);
3700 /* reading any byte out of 8-byte 'spill_slot' will cause
3701 * the whole slot to be marked as 'read'
3703 mark_reg_read(env, &state->stack[spi].spilled_ptr,
3704 state->stack[spi].spilled_ptr.parent,
3707 return update_stack_depth(env, state, min_off);
3710 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
3711 int access_size, bool zero_size_allowed,
3712 struct bpf_call_arg_meta *meta)
3714 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3716 switch (reg->type) {
3718 case PTR_TO_PACKET_META:
3719 return check_packet_access(env, regno, reg->off, access_size,
3721 case PTR_TO_MAP_VALUE:
3722 if (check_map_access_type(env, regno, reg->off, access_size,
3723 meta && meta->raw_mode ? BPF_WRITE :
3726 return check_map_access(env, regno, reg->off, access_size,
3729 return check_mem_region_access(env, regno, reg->off,
3730 access_size, reg->mem_size,
3732 case PTR_TO_RDONLY_BUF:
3733 if (meta && meta->raw_mode)
3735 return check_buffer_access(env, reg, regno, reg->off,
3736 access_size, zero_size_allowed,
3738 &env->prog->aux->max_rdonly_access);
3739 case PTR_TO_RDWR_BUF:
3740 return check_buffer_access(env, reg, regno, reg->off,
3741 access_size, zero_size_allowed,
3743 &env->prog->aux->max_rdwr_access);
3744 default: /* scalar_value|ptr_to_stack or invalid ptr */
3745 return check_stack_boundary(env, regno, access_size,
3746 zero_size_allowed, meta);
3750 /* Implementation details:
3751 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
3752 * Two bpf_map_lookups (even with the same key) will have different reg->id.
3753 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
3754 * value_or_null->value transition, since the verifier only cares about
3755 * the range of access to valid map value pointer and doesn't care about actual
3756 * address of the map element.
3757 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
3758 * reg->id > 0 after value_or_null->value transition. By doing so
3759 * two bpf_map_lookups will be considered two different pointers that
3760 * point to different bpf_spin_locks.
3761 * The verifier allows taking only one bpf_spin_lock at a time to avoid
3763 * Since only one bpf_spin_lock is allowed the checks are simpler than
3764 * reg_is_refcounted() logic. The verifier needs to remember only
3765 * one spin_lock instead of array of acquired_refs.
3766 * cur_state->active_spin_lock remembers which map value element got locked
3767 * and clears it after bpf_spin_unlock.
3769 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
3772 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3773 struct bpf_verifier_state *cur = env->cur_state;
3774 bool is_const = tnum_is_const(reg->var_off);
3775 struct bpf_map *map = reg->map_ptr;
3776 u64 val = reg->var_off.value;
3778 if (reg->type != PTR_TO_MAP_VALUE) {
3779 verbose(env, "R%d is not a pointer to map_value\n", regno);
3784 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
3790 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
3794 if (!map_value_has_spin_lock(map)) {
3795 if (map->spin_lock_off == -E2BIG)
3797 "map '%s' has more than one 'struct bpf_spin_lock'\n",
3799 else if (map->spin_lock_off == -ENOENT)
3801 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
3805 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
3809 if (map->spin_lock_off != val + reg->off) {
3810 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
3815 if (cur->active_spin_lock) {
3817 "Locking two bpf_spin_locks are not allowed\n");
3820 cur->active_spin_lock = reg->id;
3822 if (!cur->active_spin_lock) {
3823 verbose(env, "bpf_spin_unlock without taking a lock\n");
3826 if (cur->active_spin_lock != reg->id) {
3827 verbose(env, "bpf_spin_unlock of different lock\n");
3830 cur->active_spin_lock = 0;
3835 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
3837 return type == ARG_PTR_TO_MEM ||
3838 type == ARG_PTR_TO_MEM_OR_NULL ||
3839 type == ARG_PTR_TO_UNINIT_MEM;
3842 static bool arg_type_is_mem_size(enum bpf_arg_type type)
3844 return type == ARG_CONST_SIZE ||
3845 type == ARG_CONST_SIZE_OR_ZERO;
3848 static bool arg_type_is_alloc_mem_ptr(enum bpf_arg_type type)
3850 return type == ARG_PTR_TO_ALLOC_MEM ||
3851 type == ARG_PTR_TO_ALLOC_MEM_OR_NULL;
3854 static bool arg_type_is_alloc_size(enum bpf_arg_type type)
3856 return type == ARG_CONST_ALLOC_SIZE_OR_ZERO;
3859 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
3861 return type == ARG_PTR_TO_INT ||
3862 type == ARG_PTR_TO_LONG;
3865 static int int_ptr_type_to_size(enum bpf_arg_type type)
3867 if (type == ARG_PTR_TO_INT)
3869 else if (type == ARG_PTR_TO_LONG)
3875 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
3876 struct bpf_call_arg_meta *meta,
3877 const struct bpf_func_proto *fn)
3879 u32 regno = BPF_REG_1 + arg;
3880 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3881 enum bpf_reg_type expected_type, type = reg->type;
3882 enum bpf_arg_type arg_type = fn->arg_type[arg];
3885 if (arg_type == ARG_DONTCARE)
3888 err = check_reg_arg(env, regno, SRC_OP);
3892 if (arg_type == ARG_ANYTHING) {
3893 if (is_pointer_value(env, regno)) {
3894 verbose(env, "R%d leaks addr into helper function\n",
3901 if (type_is_pkt_pointer(type) &&
3902 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
3903 verbose(env, "helper access to the packet is not allowed\n");
3907 if (arg_type == ARG_PTR_TO_MAP_KEY ||
3908 arg_type == ARG_PTR_TO_MAP_VALUE ||
3909 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE ||
3910 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL) {
3911 expected_type = PTR_TO_STACK;
3912 if (register_is_null(reg) &&
3913 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL)
3914 /* final test in check_stack_boundary() */;
3915 else if (!type_is_pkt_pointer(type) &&
3916 type != PTR_TO_MAP_VALUE &&
3917 type != expected_type)
3919 } else if (arg_type == ARG_CONST_SIZE ||
3920 arg_type == ARG_CONST_SIZE_OR_ZERO ||
3921 arg_type == ARG_CONST_ALLOC_SIZE_OR_ZERO) {
3922 expected_type = SCALAR_VALUE;
3923 if (type != expected_type)
3925 } else if (arg_type == ARG_CONST_MAP_PTR) {
3926 expected_type = CONST_PTR_TO_MAP;
3927 if (type != expected_type)
3929 } else if (arg_type == ARG_PTR_TO_CTX ||
3930 arg_type == ARG_PTR_TO_CTX_OR_NULL) {
3931 expected_type = PTR_TO_CTX;
3932 if (!(register_is_null(reg) &&
3933 arg_type == ARG_PTR_TO_CTX_OR_NULL)) {
3934 if (type != expected_type)
3936 err = check_ctx_reg(env, reg, regno);
3940 } else if (arg_type == ARG_PTR_TO_SOCK_COMMON) {
3941 expected_type = PTR_TO_SOCK_COMMON;
3942 /* Any sk pointer can be ARG_PTR_TO_SOCK_COMMON */
3943 if (!type_is_sk_pointer(type))
3945 if (reg->ref_obj_id) {
3946 if (meta->ref_obj_id) {
3947 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
3948 regno, reg->ref_obj_id,
3952 meta->ref_obj_id = reg->ref_obj_id;
3954 } else if (arg_type == ARG_PTR_TO_SOCKET ||
3955 arg_type == ARG_PTR_TO_SOCKET_OR_NULL) {
3956 expected_type = PTR_TO_SOCKET;
3957 if (!(register_is_null(reg) &&
3958 arg_type == ARG_PTR_TO_SOCKET_OR_NULL)) {
3959 if (type != expected_type)
3962 } else if (arg_type == ARG_PTR_TO_BTF_ID) {
3963 expected_type = PTR_TO_BTF_ID;
3964 if (type != expected_type)
3966 if (!fn->check_btf_id) {
3967 if (reg->btf_id != meta->btf_id) {
3968 verbose(env, "Helper has type %s got %s in R%d\n",
3969 kernel_type_name(meta->btf_id),
3970 kernel_type_name(reg->btf_id), regno);
3974 } else if (!fn->check_btf_id(reg->btf_id, arg)) {
3975 verbose(env, "Helper does not support %s in R%d\n",
3976 kernel_type_name(reg->btf_id), regno);
3980 if (!tnum_is_const(reg->var_off) || reg->var_off.value || reg->off) {
3981 verbose(env, "R%d is a pointer to in-kernel struct with non-zero offset\n",
3985 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
3986 if (meta->func_id == BPF_FUNC_spin_lock) {
3987 if (process_spin_lock(env, regno, true))
3989 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
3990 if (process_spin_lock(env, regno, false))
3993 verbose(env, "verifier internal error\n");
3996 } else if (arg_type_is_mem_ptr(arg_type)) {
3997 expected_type = PTR_TO_STACK;
3998 /* One exception here. In case function allows for NULL to be
3999 * passed in as argument, it's a SCALAR_VALUE type. Final test
4000 * happens during stack boundary checking.
4002 if (register_is_null(reg) &&
4003 (arg_type == ARG_PTR_TO_MEM_OR_NULL ||
4004 arg_type == ARG_PTR_TO_ALLOC_MEM_OR_NULL))
4005 /* final test in check_stack_boundary() */;
4006 else if (!type_is_pkt_pointer(type) &&
4007 type != PTR_TO_MAP_VALUE &&
4008 type != PTR_TO_MEM &&
4009 type != PTR_TO_RDONLY_BUF &&
4010 type != PTR_TO_RDWR_BUF &&
4011 type != expected_type)
4013 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
4014 } else if (arg_type_is_alloc_mem_ptr(arg_type)) {
4015 expected_type = PTR_TO_MEM;
4016 if (register_is_null(reg) &&
4017 arg_type == ARG_PTR_TO_ALLOC_MEM_OR_NULL)
4018 /* final test in check_stack_boundary() */;
4019 else if (type != expected_type)
4021 if (meta->ref_obj_id) {
4022 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
4023 regno, reg->ref_obj_id,
4027 meta->ref_obj_id = reg->ref_obj_id;
4028 } else if (arg_type_is_int_ptr(arg_type)) {
4029 expected_type = PTR_TO_STACK;
4030 if (!type_is_pkt_pointer(type) &&
4031 type != PTR_TO_MAP_VALUE &&
4032 type != expected_type)
4035 verbose(env, "unsupported arg_type %d\n", arg_type);
4039 if (arg_type == ARG_CONST_MAP_PTR) {
4040 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
4041 meta->map_ptr = reg->map_ptr;
4042 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
4043 /* bpf_map_xxx(..., map_ptr, ..., key) call:
4044 * check that [key, key + map->key_size) are within
4045 * stack limits and initialized
4047 if (!meta->map_ptr) {
4048 /* in function declaration map_ptr must come before
4049 * map_key, so that it's verified and known before
4050 * we have to check map_key here. Otherwise it means
4051 * that kernel subsystem misconfigured verifier
4053 verbose(env, "invalid map_ptr to access map->key\n");
4056 err = check_helper_mem_access(env, regno,
4057 meta->map_ptr->key_size, false,
4059 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
4060 (arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL &&
4061 !register_is_null(reg)) ||
4062 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
4063 /* bpf_map_xxx(..., map_ptr, ..., value) call:
4064 * check [value, value + map->value_size) validity
4066 if (!meta->map_ptr) {
4067 /* kernel subsystem misconfigured verifier */
4068 verbose(env, "invalid map_ptr to access map->value\n");
4071 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
4072 err = check_helper_mem_access(env, regno,
4073 meta->map_ptr->value_size, false,
4075 } else if (arg_type_is_mem_size(arg_type)) {
4076 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
4078 /* This is used to refine r0 return value bounds for helpers
4079 * that enforce this value as an upper bound on return values.
4080 * See do_refine_retval_range() for helpers that can refine
4081 * the return value. C type of helper is u32 so we pull register
4082 * bound from umax_value however, if negative verifier errors
4083 * out. Only upper bounds can be learned because retval is an
4084 * int type and negative retvals are allowed.
4086 meta->msize_max_value = reg->umax_value;
4088 /* The register is SCALAR_VALUE; the access check
4089 * happens using its boundaries.
4091 if (!tnum_is_const(reg->var_off))
4092 /* For unprivileged variable accesses, disable raw
4093 * mode so that the program is required to
4094 * initialize all the memory that the helper could
4095 * just partially fill up.
4099 if (reg->smin_value < 0) {
4100 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
4105 if (reg->umin_value == 0) {
4106 err = check_helper_mem_access(env, regno - 1, 0,
4113 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
4114 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
4118 err = check_helper_mem_access(env, regno - 1,
4120 zero_size_allowed, meta);
4122 err = mark_chain_precision(env, regno);
4123 } else if (arg_type_is_alloc_size(arg_type)) {
4124 if (!tnum_is_const(reg->var_off)) {
4125 verbose(env, "R%d unbounded size, use 'var &= const' or 'if (var < const)'\n",
4129 meta->mem_size = reg->var_off.value;
4130 } else if (arg_type_is_int_ptr(arg_type)) {
4131 int size = int_ptr_type_to_size(arg_type);
4133 err = check_helper_mem_access(env, regno, size, false, meta);
4136 err = check_ptr_alignment(env, reg, 0, size, true);
4141 verbose(env, "R%d type=%s expected=%s\n", regno,
4142 reg_type_str[type], reg_type_str[expected_type]);
4146 static int check_map_func_compatibility(struct bpf_verifier_env *env,
4147 struct bpf_map *map, int func_id)
4152 /* We need a two way check, first is from map perspective ... */
4153 switch (map->map_type) {
4154 case BPF_MAP_TYPE_PROG_ARRAY:
4155 if (func_id != BPF_FUNC_tail_call)
4158 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
4159 if (func_id != BPF_FUNC_perf_event_read &&
4160 func_id != BPF_FUNC_perf_event_output &&
4161 func_id != BPF_FUNC_skb_output &&
4162 func_id != BPF_FUNC_perf_event_read_value &&
4163 func_id != BPF_FUNC_xdp_output)
4166 case BPF_MAP_TYPE_RINGBUF:
4167 if (func_id != BPF_FUNC_ringbuf_output &&
4168 func_id != BPF_FUNC_ringbuf_reserve &&
4169 func_id != BPF_FUNC_ringbuf_submit &&
4170 func_id != BPF_FUNC_ringbuf_discard &&
4171 func_id != BPF_FUNC_ringbuf_query)
4174 case BPF_MAP_TYPE_STACK_TRACE:
4175 if (func_id != BPF_FUNC_get_stackid)
4178 case BPF_MAP_TYPE_CGROUP_ARRAY:
4179 if (func_id != BPF_FUNC_skb_under_cgroup &&
4180 func_id != BPF_FUNC_current_task_under_cgroup)
4183 case BPF_MAP_TYPE_CGROUP_STORAGE:
4184 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
4185 if (func_id != BPF_FUNC_get_local_storage)
4188 case BPF_MAP_TYPE_DEVMAP:
4189 case BPF_MAP_TYPE_DEVMAP_HASH:
4190 if (func_id != BPF_FUNC_redirect_map &&
4191 func_id != BPF_FUNC_map_lookup_elem)
4194 /* Restrict bpf side of cpumap and xskmap, open when use-cases
4197 case BPF_MAP_TYPE_CPUMAP:
4198 if (func_id != BPF_FUNC_redirect_map)
4201 case BPF_MAP_TYPE_XSKMAP:
4202 if (func_id != BPF_FUNC_redirect_map &&
4203 func_id != BPF_FUNC_map_lookup_elem)
4206 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
4207 case BPF_MAP_TYPE_HASH_OF_MAPS:
4208 if (func_id != BPF_FUNC_map_lookup_elem)
4211 case BPF_MAP_TYPE_SOCKMAP:
4212 if (func_id != BPF_FUNC_sk_redirect_map &&
4213 func_id != BPF_FUNC_sock_map_update &&
4214 func_id != BPF_FUNC_map_delete_elem &&
4215 func_id != BPF_FUNC_msg_redirect_map &&
4216 func_id != BPF_FUNC_sk_select_reuseport &&
4217 func_id != BPF_FUNC_map_lookup_elem)
4220 case BPF_MAP_TYPE_SOCKHASH:
4221 if (func_id != BPF_FUNC_sk_redirect_hash &&
4222 func_id != BPF_FUNC_sock_hash_update &&
4223 func_id != BPF_FUNC_map_delete_elem &&
4224 func_id != BPF_FUNC_msg_redirect_hash &&
4225 func_id != BPF_FUNC_sk_select_reuseport &&
4226 func_id != BPF_FUNC_map_lookup_elem)
4229 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
4230 if (func_id != BPF_FUNC_sk_select_reuseport)
4233 case BPF_MAP_TYPE_QUEUE:
4234 case BPF_MAP_TYPE_STACK:
4235 if (func_id != BPF_FUNC_map_peek_elem &&
4236 func_id != BPF_FUNC_map_pop_elem &&
4237 func_id != BPF_FUNC_map_push_elem)
4240 case BPF_MAP_TYPE_SK_STORAGE:
4241 if (func_id != BPF_FUNC_sk_storage_get &&
4242 func_id != BPF_FUNC_sk_storage_delete)
4249 /* ... and second from the function itself. */
4251 case BPF_FUNC_tail_call:
4252 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
4254 if (env->subprog_cnt > 1) {
4255 verbose(env, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
4259 case BPF_FUNC_perf_event_read:
4260 case BPF_FUNC_perf_event_output:
4261 case BPF_FUNC_perf_event_read_value:
4262 case BPF_FUNC_skb_output:
4263 case BPF_FUNC_xdp_output:
4264 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
4267 case BPF_FUNC_get_stackid:
4268 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
4271 case BPF_FUNC_current_task_under_cgroup:
4272 case BPF_FUNC_skb_under_cgroup:
4273 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
4276 case BPF_FUNC_redirect_map:
4277 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
4278 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
4279 map->map_type != BPF_MAP_TYPE_CPUMAP &&
4280 map->map_type != BPF_MAP_TYPE_XSKMAP)
4283 case BPF_FUNC_sk_redirect_map:
4284 case BPF_FUNC_msg_redirect_map:
4285 case BPF_FUNC_sock_map_update:
4286 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
4289 case BPF_FUNC_sk_redirect_hash:
4290 case BPF_FUNC_msg_redirect_hash:
4291 case BPF_FUNC_sock_hash_update:
4292 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
4295 case BPF_FUNC_get_local_storage:
4296 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
4297 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
4300 case BPF_FUNC_sk_select_reuseport:
4301 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
4302 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
4303 map->map_type != BPF_MAP_TYPE_SOCKHASH)
4306 case BPF_FUNC_map_peek_elem:
4307 case BPF_FUNC_map_pop_elem:
4308 case BPF_FUNC_map_push_elem:
4309 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
4310 map->map_type != BPF_MAP_TYPE_STACK)
4313 case BPF_FUNC_sk_storage_get:
4314 case BPF_FUNC_sk_storage_delete:
4315 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
4324 verbose(env, "cannot pass map_type %d into func %s#%d\n",
4325 map->map_type, func_id_name(func_id), func_id);
4329 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
4333 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
4335 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
4337 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
4339 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
4341 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
4344 /* We only support one arg being in raw mode at the moment,
4345 * which is sufficient for the helper functions we have
4351 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
4352 enum bpf_arg_type arg_next)
4354 return (arg_type_is_mem_ptr(arg_curr) &&
4355 !arg_type_is_mem_size(arg_next)) ||
4356 (!arg_type_is_mem_ptr(arg_curr) &&
4357 arg_type_is_mem_size(arg_next));
4360 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
4362 /* bpf_xxx(..., buf, len) call will access 'len'
4363 * bytes from memory 'buf'. Both arg types need
4364 * to be paired, so make sure there's no buggy
4365 * helper function specification.
4367 if (arg_type_is_mem_size(fn->arg1_type) ||
4368 arg_type_is_mem_ptr(fn->arg5_type) ||
4369 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
4370 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
4371 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
4372 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
4378 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
4382 if (arg_type_may_be_refcounted(fn->arg1_type))
4384 if (arg_type_may_be_refcounted(fn->arg2_type))
4386 if (arg_type_may_be_refcounted(fn->arg3_type))
4388 if (arg_type_may_be_refcounted(fn->arg4_type))
4390 if (arg_type_may_be_refcounted(fn->arg5_type))
4393 /* A reference acquiring function cannot acquire
4394 * another refcounted ptr.
4396 if (may_be_acquire_function(func_id) && count)
4399 /* We only support one arg being unreferenced at the moment,
4400 * which is sufficient for the helper functions we have right now.
4405 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
4407 return check_raw_mode_ok(fn) &&
4408 check_arg_pair_ok(fn) &&
4409 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
4412 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
4413 * are now invalid, so turn them into unknown SCALAR_VALUE.
4415 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
4416 struct bpf_func_state *state)
4418 struct bpf_reg_state *regs = state->regs, *reg;
4421 for (i = 0; i < MAX_BPF_REG; i++)
4422 if (reg_is_pkt_pointer_any(®s[i]))
4423 mark_reg_unknown(env, regs, i);
4425 bpf_for_each_spilled_reg(i, state, reg) {
4428 if (reg_is_pkt_pointer_any(reg))
4429 __mark_reg_unknown(env, reg);
4433 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
4435 struct bpf_verifier_state *vstate = env->cur_state;
4438 for (i = 0; i <= vstate->curframe; i++)
4439 __clear_all_pkt_pointers(env, vstate->frame[i]);
4442 static void release_reg_references(struct bpf_verifier_env *env,
4443 struct bpf_func_state *state,
4446 struct bpf_reg_state *regs = state->regs, *reg;
4449 for (i = 0; i < MAX_BPF_REG; i++)
4450 if (regs[i].ref_obj_id == ref_obj_id)
4451 mark_reg_unknown(env, regs, i);
4453 bpf_for_each_spilled_reg(i, state, reg) {
4456 if (reg->ref_obj_id == ref_obj_id)
4457 __mark_reg_unknown(env, reg);
4461 /* The pointer with the specified id has released its reference to kernel
4462 * resources. Identify all copies of the same pointer and clear the reference.
4464 static int release_reference(struct bpf_verifier_env *env,
4467 struct bpf_verifier_state *vstate = env->cur_state;
4471 err = release_reference_state(cur_func(env), ref_obj_id);
4475 for (i = 0; i <= vstate->curframe; i++)
4476 release_reg_references(env, vstate->frame[i], ref_obj_id);
4481 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
4482 struct bpf_reg_state *regs)
4486 /* after the call registers r0 - r5 were scratched */
4487 for (i = 0; i < CALLER_SAVED_REGS; i++) {
4488 mark_reg_not_init(env, regs, caller_saved[i]);
4489 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
4493 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
4496 struct bpf_verifier_state *state = env->cur_state;
4497 struct bpf_func_info_aux *func_info_aux;
4498 struct bpf_func_state *caller, *callee;
4499 int i, err, subprog, target_insn;
4500 bool is_global = false;
4502 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
4503 verbose(env, "the call stack of %d frames is too deep\n",
4504 state->curframe + 2);
4508 target_insn = *insn_idx + insn->imm;
4509 subprog = find_subprog(env, target_insn + 1);
4511 verbose(env, "verifier bug. No program starts at insn %d\n",
4516 caller = state->frame[state->curframe];
4517 if (state->frame[state->curframe + 1]) {
4518 verbose(env, "verifier bug. Frame %d already allocated\n",
4519 state->curframe + 1);
4523 func_info_aux = env->prog->aux->func_info_aux;
4525 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
4526 err = btf_check_func_arg_match(env, subprog, caller->regs);
4531 verbose(env, "Caller passes invalid args into func#%d\n",
4535 if (env->log.level & BPF_LOG_LEVEL)
4537 "Func#%d is global and valid. Skipping.\n",
4539 clear_caller_saved_regs(env, caller->regs);
4541 /* All global functions return SCALAR_VALUE */
4542 mark_reg_unknown(env, caller->regs, BPF_REG_0);
4544 /* continue with next insn after call */
4549 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
4552 state->frame[state->curframe + 1] = callee;
4554 /* callee cannot access r0, r6 - r9 for reading and has to write
4555 * into its own stack before reading from it.
4556 * callee can read/write into caller's stack
4558 init_func_state(env, callee,
4559 /* remember the callsite, it will be used by bpf_exit */
4560 *insn_idx /* callsite */,
4561 state->curframe + 1 /* frameno within this callchain */,
4562 subprog /* subprog number within this prog */);
4564 /* Transfer references to the callee */
4565 err = transfer_reference_state(callee, caller);
4569 /* copy r1 - r5 args that callee can access. The copy includes parent
4570 * pointers, which connects us up to the liveness chain
4572 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
4573 callee->regs[i] = caller->regs[i];
4575 clear_caller_saved_regs(env, caller->regs);
4577 /* only increment it after check_reg_arg() finished */
4580 /* and go analyze first insn of the callee */
4581 *insn_idx = target_insn;
4583 if (env->log.level & BPF_LOG_LEVEL) {
4584 verbose(env, "caller:\n");
4585 print_verifier_state(env, caller);
4586 verbose(env, "callee:\n");
4587 print_verifier_state(env, callee);
4592 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
4594 struct bpf_verifier_state *state = env->cur_state;
4595 struct bpf_func_state *caller, *callee;
4596 struct bpf_reg_state *r0;
4599 callee = state->frame[state->curframe];
4600 r0 = &callee->regs[BPF_REG_0];
4601 if (r0->type == PTR_TO_STACK) {
4602 /* technically it's ok to return caller's stack pointer
4603 * (or caller's caller's pointer) back to the caller,
4604 * since these pointers are valid. Only current stack
4605 * pointer will be invalid as soon as function exits,
4606 * but let's be conservative
4608 verbose(env, "cannot return stack pointer to the caller\n");
4613 caller = state->frame[state->curframe];
4614 /* return to the caller whatever r0 had in the callee */
4615 caller->regs[BPF_REG_0] = *r0;
4617 /* Transfer references to the caller */
4618 err = transfer_reference_state(caller, callee);
4622 *insn_idx = callee->callsite + 1;
4623 if (env->log.level & BPF_LOG_LEVEL) {
4624 verbose(env, "returning from callee:\n");
4625 print_verifier_state(env, callee);
4626 verbose(env, "to caller at %d:\n", *insn_idx);
4627 print_verifier_state(env, caller);
4629 /* clear everything in the callee */
4630 free_func_state(callee);
4631 state->frame[state->curframe + 1] = NULL;
4635 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
4637 struct bpf_call_arg_meta *meta)
4639 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
4641 if (ret_type != RET_INTEGER ||
4642 (func_id != BPF_FUNC_get_stack &&
4643 func_id != BPF_FUNC_probe_read_str &&
4644 func_id != BPF_FUNC_probe_read_kernel_str &&
4645 func_id != BPF_FUNC_probe_read_user_str))
4648 ret_reg->smax_value = meta->msize_max_value;
4649 ret_reg->s32_max_value = meta->msize_max_value;
4650 __reg_deduce_bounds(ret_reg);
4651 __reg_bound_offset(ret_reg);
4652 __update_reg_bounds(ret_reg);
4656 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
4657 int func_id, int insn_idx)
4659 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
4660 struct bpf_map *map = meta->map_ptr;
4662 if (func_id != BPF_FUNC_tail_call &&
4663 func_id != BPF_FUNC_map_lookup_elem &&
4664 func_id != BPF_FUNC_map_update_elem &&
4665 func_id != BPF_FUNC_map_delete_elem &&
4666 func_id != BPF_FUNC_map_push_elem &&
4667 func_id != BPF_FUNC_map_pop_elem &&
4668 func_id != BPF_FUNC_map_peek_elem)
4672 verbose(env, "kernel subsystem misconfigured verifier\n");
4676 /* In case of read-only, some additional restrictions
4677 * need to be applied in order to prevent altering the
4678 * state of the map from program side.
4680 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
4681 (func_id == BPF_FUNC_map_delete_elem ||
4682 func_id == BPF_FUNC_map_update_elem ||
4683 func_id == BPF_FUNC_map_push_elem ||
4684 func_id == BPF_FUNC_map_pop_elem)) {
4685 verbose(env, "write into map forbidden\n");
4689 if (!BPF_MAP_PTR(aux->map_ptr_state))
4690 bpf_map_ptr_store(aux, meta->map_ptr,
4691 !meta->map_ptr->bypass_spec_v1);
4692 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
4693 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
4694 !meta->map_ptr->bypass_spec_v1);
4699 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
4700 int func_id, int insn_idx)
4702 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
4703 struct bpf_reg_state *regs = cur_regs(env), *reg;
4704 struct bpf_map *map = meta->map_ptr;
4709 if (func_id != BPF_FUNC_tail_call)
4711 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
4712 verbose(env, "kernel subsystem misconfigured verifier\n");
4716 range = tnum_range(0, map->max_entries - 1);
4717 reg = ®s[BPF_REG_3];
4719 if (!register_is_const(reg) || !tnum_in(range, reg->var_off)) {
4720 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
4724 err = mark_chain_precision(env, BPF_REG_3);
4728 val = reg->var_off.value;
4729 if (bpf_map_key_unseen(aux))
4730 bpf_map_key_store(aux, val);
4731 else if (!bpf_map_key_poisoned(aux) &&
4732 bpf_map_key_immediate(aux) != val)
4733 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
4737 static int check_reference_leak(struct bpf_verifier_env *env)
4739 struct bpf_func_state *state = cur_func(env);
4742 for (i = 0; i < state->acquired_refs; i++) {
4743 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
4744 state->refs[i].id, state->refs[i].insn_idx);
4746 return state->acquired_refs ? -EINVAL : 0;
4749 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
4751 const struct bpf_func_proto *fn = NULL;
4752 struct bpf_reg_state *regs;
4753 struct bpf_call_arg_meta meta;
4757 /* find function prototype */
4758 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
4759 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
4764 if (env->ops->get_func_proto)
4765 fn = env->ops->get_func_proto(func_id, env->prog);
4767 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
4772 /* eBPF programs must be GPL compatible to use GPL-ed functions */
4773 if (!env->prog->gpl_compatible && fn->gpl_only) {
4774 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
4778 /* With LD_ABS/IND some JITs save/restore skb from r1. */
4779 changes_data = bpf_helper_changes_pkt_data(fn->func);
4780 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
4781 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
4782 func_id_name(func_id), func_id);
4786 memset(&meta, 0, sizeof(meta));
4787 meta.pkt_access = fn->pkt_access;
4789 err = check_func_proto(fn, func_id);
4791 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
4792 func_id_name(func_id), func_id);
4796 meta.func_id = func_id;
4798 for (i = 0; i < 5; i++) {
4799 if (!fn->check_btf_id) {
4800 err = btf_resolve_helper_id(&env->log, fn, i);
4804 err = check_func_arg(env, i, &meta, fn);
4809 err = record_func_map(env, &meta, func_id, insn_idx);
4813 err = record_func_key(env, &meta, func_id, insn_idx);
4817 /* Mark slots with STACK_MISC in case of raw mode, stack offset
4818 * is inferred from register state.
4820 for (i = 0; i < meta.access_size; i++) {
4821 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
4822 BPF_WRITE, -1, false);
4827 if (func_id == BPF_FUNC_tail_call) {
4828 err = check_reference_leak(env);
4830 verbose(env, "tail_call would lead to reference leak\n");
4833 } else if (is_release_function(func_id)) {
4834 err = release_reference(env, meta.ref_obj_id);
4836 verbose(env, "func %s#%d reference has not been acquired before\n",
4837 func_id_name(func_id), func_id);
4842 regs = cur_regs(env);
4844 /* check that flags argument in get_local_storage(map, flags) is 0,
4845 * this is required because get_local_storage() can't return an error.
4847 if (func_id == BPF_FUNC_get_local_storage &&
4848 !register_is_null(®s[BPF_REG_2])) {
4849 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
4853 /* reset caller saved regs */
4854 for (i = 0; i < CALLER_SAVED_REGS; i++) {
4855 mark_reg_not_init(env, regs, caller_saved[i]);
4856 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
4859 /* helper call returns 64-bit value. */
4860 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
4862 /* update return register (already marked as written above) */
4863 if (fn->ret_type == RET_INTEGER) {
4864 /* sets type to SCALAR_VALUE */
4865 mark_reg_unknown(env, regs, BPF_REG_0);
4866 } else if (fn->ret_type == RET_VOID) {
4867 regs[BPF_REG_0].type = NOT_INIT;
4868 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
4869 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
4870 /* There is no offset yet applied, variable or fixed */
4871 mark_reg_known_zero(env, regs, BPF_REG_0);
4872 /* remember map_ptr, so that check_map_access()
4873 * can check 'value_size' boundary of memory access
4874 * to map element returned from bpf_map_lookup_elem()
4876 if (meta.map_ptr == NULL) {
4878 "kernel subsystem misconfigured verifier\n");
4881 regs[BPF_REG_0].map_ptr = meta.map_ptr;
4882 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
4883 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
4884 if (map_value_has_spin_lock(meta.map_ptr))
4885 regs[BPF_REG_0].id = ++env->id_gen;
4887 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
4888 regs[BPF_REG_0].id = ++env->id_gen;
4890 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
4891 mark_reg_known_zero(env, regs, BPF_REG_0);
4892 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
4893 regs[BPF_REG_0].id = ++env->id_gen;
4894 } else if (fn->ret_type == RET_PTR_TO_SOCK_COMMON_OR_NULL) {
4895 mark_reg_known_zero(env, regs, BPF_REG_0);
4896 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON_OR_NULL;
4897 regs[BPF_REG_0].id = ++env->id_gen;
4898 } else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
4899 mark_reg_known_zero(env, regs, BPF_REG_0);
4900 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
4901 regs[BPF_REG_0].id = ++env->id_gen;
4902 } else if (fn->ret_type == RET_PTR_TO_ALLOC_MEM_OR_NULL) {
4903 mark_reg_known_zero(env, regs, BPF_REG_0);
4904 regs[BPF_REG_0].type = PTR_TO_MEM_OR_NULL;
4905 regs[BPF_REG_0].id = ++env->id_gen;
4906 regs[BPF_REG_0].mem_size = meta.mem_size;
4907 } else if (fn->ret_type == RET_PTR_TO_BTF_ID_OR_NULL) {
4910 mark_reg_known_zero(env, regs, BPF_REG_0);
4911 regs[BPF_REG_0].type = PTR_TO_BTF_ID_OR_NULL;
4912 ret_btf_id = *fn->ret_btf_id;
4913 if (ret_btf_id == 0) {
4914 verbose(env, "invalid return type %d of func %s#%d\n",
4915 fn->ret_type, func_id_name(func_id), func_id);
4918 regs[BPF_REG_0].btf_id = ret_btf_id;
4920 verbose(env, "unknown return type %d of func %s#%d\n",
4921 fn->ret_type, func_id_name(func_id), func_id);
4925 if (is_ptr_cast_function(func_id)) {
4926 /* For release_reference() */
4927 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
4928 } else if (is_acquire_function(func_id, meta.map_ptr)) {
4929 int id = acquire_reference_state(env, insn_idx);
4933 /* For mark_ptr_or_null_reg() */
4934 regs[BPF_REG_0].id = id;
4935 /* For release_reference() */
4936 regs[BPF_REG_0].ref_obj_id = id;
4939 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
4941 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
4945 if ((func_id == BPF_FUNC_get_stack ||
4946 func_id == BPF_FUNC_get_task_stack) &&
4947 !env->prog->has_callchain_buf) {
4948 const char *err_str;
4950 #ifdef CONFIG_PERF_EVENTS
4951 err = get_callchain_buffers(sysctl_perf_event_max_stack);
4952 err_str = "cannot get callchain buffer for func %s#%d\n";
4955 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
4958 verbose(env, err_str, func_id_name(func_id), func_id);
4962 env->prog->has_callchain_buf = true;
4965 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
4966 env->prog->call_get_stack = true;
4969 clear_all_pkt_pointers(env);
4973 static bool signed_add_overflows(s64 a, s64 b)
4975 /* Do the add in u64, where overflow is well-defined */
4976 s64 res = (s64)((u64)a + (u64)b);
4983 static bool signed_add32_overflows(s64 a, s64 b)
4985 /* Do the add in u32, where overflow is well-defined */
4986 s32 res = (s32)((u32)a + (u32)b);
4993 static bool signed_sub_overflows(s32 a, s32 b)
4995 /* Do the sub in u64, where overflow is well-defined */
4996 s64 res = (s64)((u64)a - (u64)b);
5003 static bool signed_sub32_overflows(s32 a, s32 b)
5005 /* Do the sub in u64, where overflow is well-defined */
5006 s32 res = (s32)((u32)a - (u32)b);
5013 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
5014 const struct bpf_reg_state *reg,
5015 enum bpf_reg_type type)
5017 bool known = tnum_is_const(reg->var_off);
5018 s64 val = reg->var_off.value;
5019 s64 smin = reg->smin_value;
5021 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
5022 verbose(env, "math between %s pointer and %lld is not allowed\n",
5023 reg_type_str[type], val);
5027 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
5028 verbose(env, "%s pointer offset %d is not allowed\n",
5029 reg_type_str[type], reg->off);
5033 if (smin == S64_MIN) {
5034 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
5035 reg_type_str[type]);
5039 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
5040 verbose(env, "value %lld makes %s pointer be out of bounds\n",
5041 smin, reg_type_str[type]);
5048 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
5050 return &env->insn_aux_data[env->insn_idx];
5053 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
5054 u32 *ptr_limit, u8 opcode, bool off_is_neg)
5056 bool mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
5057 (opcode == BPF_SUB && !off_is_neg);
5060 switch (ptr_reg->type) {
5062 /* Indirect variable offset stack access is prohibited in
5063 * unprivileged mode so it's not handled here.
5065 off = ptr_reg->off + ptr_reg->var_off.value;
5067 *ptr_limit = MAX_BPF_STACK + off;
5071 case PTR_TO_MAP_VALUE:
5073 *ptr_limit = ptr_reg->umax_value + ptr_reg->off;
5075 off = ptr_reg->smin_value + ptr_reg->off;
5076 *ptr_limit = ptr_reg->map_ptr->value_size - off;
5084 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
5085 const struct bpf_insn *insn)
5087 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
5090 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
5091 u32 alu_state, u32 alu_limit)
5093 /* If we arrived here from different branches with different
5094 * state or limits to sanitize, then this won't work.
5096 if (aux->alu_state &&
5097 (aux->alu_state != alu_state ||
5098 aux->alu_limit != alu_limit))
5101 /* Corresponding fixup done in fixup_bpf_calls(). */
5102 aux->alu_state = alu_state;
5103 aux->alu_limit = alu_limit;
5107 static int sanitize_val_alu(struct bpf_verifier_env *env,
5108 struct bpf_insn *insn)
5110 struct bpf_insn_aux_data *aux = cur_aux(env);
5112 if (can_skip_alu_sanitation(env, insn))
5115 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
5118 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
5119 struct bpf_insn *insn,
5120 const struct bpf_reg_state *ptr_reg,
5121 struct bpf_reg_state *dst_reg,
5124 struct bpf_verifier_state *vstate = env->cur_state;
5125 struct bpf_insn_aux_data *aux = cur_aux(env);
5126 bool ptr_is_dst_reg = ptr_reg == dst_reg;
5127 u8 opcode = BPF_OP(insn->code);
5128 u32 alu_state, alu_limit;
5129 struct bpf_reg_state tmp;
5132 if (can_skip_alu_sanitation(env, insn))
5135 /* We already marked aux for masking from non-speculative
5136 * paths, thus we got here in the first place. We only care
5137 * to explore bad access from here.
5139 if (vstate->speculative)
5142 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
5143 alu_state |= ptr_is_dst_reg ?
5144 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
5146 if (retrieve_ptr_limit(ptr_reg, &alu_limit, opcode, off_is_neg))
5148 if (update_alu_sanitation_state(aux, alu_state, alu_limit))
5151 /* Simulate and find potential out-of-bounds access under
5152 * speculative execution from truncation as a result of
5153 * masking when off was not within expected range. If off
5154 * sits in dst, then we temporarily need to move ptr there
5155 * to simulate dst (== 0) +/-= ptr. Needed, for example,
5156 * for cases where we use K-based arithmetic in one direction
5157 * and truncated reg-based in the other in order to explore
5160 if (!ptr_is_dst_reg) {
5162 *dst_reg = *ptr_reg;
5164 ret = push_stack(env, env->insn_idx + 1, env->insn_idx, true);
5165 if (!ptr_is_dst_reg && ret)
5167 return !ret ? -EFAULT : 0;
5170 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
5171 * Caller should also handle BPF_MOV case separately.
5172 * If we return -EACCES, caller may want to try again treating pointer as a
5173 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
5175 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
5176 struct bpf_insn *insn,
5177 const struct bpf_reg_state *ptr_reg,
5178 const struct bpf_reg_state *off_reg)
5180 struct bpf_verifier_state *vstate = env->cur_state;
5181 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5182 struct bpf_reg_state *regs = state->regs, *dst_reg;
5183 bool known = tnum_is_const(off_reg->var_off);
5184 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
5185 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
5186 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
5187 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
5188 u32 dst = insn->dst_reg, src = insn->src_reg;
5189 u8 opcode = BPF_OP(insn->code);
5192 dst_reg = ®s[dst];
5194 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
5195 smin_val > smax_val || umin_val > umax_val) {
5196 /* Taint dst register if offset had invalid bounds derived from
5197 * e.g. dead branches.
5199 __mark_reg_unknown(env, dst_reg);
5203 if (BPF_CLASS(insn->code) != BPF_ALU64) {
5204 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
5205 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
5206 __mark_reg_unknown(env, dst_reg);
5211 "R%d 32-bit pointer arithmetic prohibited\n",
5216 switch (ptr_reg->type) {
5217 case PTR_TO_MAP_VALUE_OR_NULL:
5218 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
5219 dst, reg_type_str[ptr_reg->type]);
5221 case CONST_PTR_TO_MAP:
5222 case PTR_TO_PACKET_END:
5224 case PTR_TO_SOCKET_OR_NULL:
5225 case PTR_TO_SOCK_COMMON:
5226 case PTR_TO_SOCK_COMMON_OR_NULL:
5227 case PTR_TO_TCP_SOCK:
5228 case PTR_TO_TCP_SOCK_OR_NULL:
5229 case PTR_TO_XDP_SOCK:
5230 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
5231 dst, reg_type_str[ptr_reg->type]);
5233 case PTR_TO_MAP_VALUE:
5234 if (!env->allow_ptr_leaks && !known && (smin_val < 0) != (smax_val < 0)) {
5235 verbose(env, "R%d has unknown scalar with mixed signed bounds, pointer arithmetic with it prohibited for !root\n",
5236 off_reg == dst_reg ? dst : src);
5244 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
5245 * The id may be overwritten later if we create a new variable offset.
5247 dst_reg->type = ptr_reg->type;
5248 dst_reg->id = ptr_reg->id;
5250 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
5251 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
5254 /* pointer types do not carry 32-bit bounds at the moment. */
5255 __mark_reg32_unbounded(dst_reg);
5259 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
5261 verbose(env, "R%d tried to add from different maps or paths\n", dst);
5264 /* We can take a fixed offset as long as it doesn't overflow
5265 * the s32 'off' field
5267 if (known && (ptr_reg->off + smin_val ==
5268 (s64)(s32)(ptr_reg->off + smin_val))) {
5269 /* pointer += K. Accumulate it into fixed offset */
5270 dst_reg->smin_value = smin_ptr;
5271 dst_reg->smax_value = smax_ptr;
5272 dst_reg->umin_value = umin_ptr;
5273 dst_reg->umax_value = umax_ptr;
5274 dst_reg->var_off = ptr_reg->var_off;
5275 dst_reg->off = ptr_reg->off + smin_val;
5276 dst_reg->raw = ptr_reg->raw;
5279 /* A new variable offset is created. Note that off_reg->off
5280 * == 0, since it's a scalar.
5281 * dst_reg gets the pointer type and since some positive
5282 * integer value was added to the pointer, give it a new 'id'
5283 * if it's a PTR_TO_PACKET.
5284 * this creates a new 'base' pointer, off_reg (variable) gets
5285 * added into the variable offset, and we copy the fixed offset
5288 if (signed_add_overflows(smin_ptr, smin_val) ||
5289 signed_add_overflows(smax_ptr, smax_val)) {
5290 dst_reg->smin_value = S64_MIN;
5291 dst_reg->smax_value = S64_MAX;
5293 dst_reg->smin_value = smin_ptr + smin_val;
5294 dst_reg->smax_value = smax_ptr + smax_val;
5296 if (umin_ptr + umin_val < umin_ptr ||
5297 umax_ptr + umax_val < umax_ptr) {
5298 dst_reg->umin_value = 0;
5299 dst_reg->umax_value = U64_MAX;
5301 dst_reg->umin_value = umin_ptr + umin_val;
5302 dst_reg->umax_value = umax_ptr + umax_val;
5304 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
5305 dst_reg->off = ptr_reg->off;
5306 dst_reg->raw = ptr_reg->raw;
5307 if (reg_is_pkt_pointer(ptr_reg)) {
5308 dst_reg->id = ++env->id_gen;
5309 /* something was added to pkt_ptr, set range to zero */
5314 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
5316 verbose(env, "R%d tried to sub from different maps or paths\n", dst);
5319 if (dst_reg == off_reg) {
5320 /* scalar -= pointer. Creates an unknown scalar */
5321 verbose(env, "R%d tried to subtract pointer from scalar\n",
5325 /* We don't allow subtraction from FP, because (according to
5326 * test_verifier.c test "invalid fp arithmetic", JITs might not
5327 * be able to deal with it.
5329 if (ptr_reg->type == PTR_TO_STACK) {
5330 verbose(env, "R%d subtraction from stack pointer prohibited\n",
5334 if (known && (ptr_reg->off - smin_val ==
5335 (s64)(s32)(ptr_reg->off - smin_val))) {
5336 /* pointer -= K. Subtract it from fixed offset */
5337 dst_reg->smin_value = smin_ptr;
5338 dst_reg->smax_value = smax_ptr;
5339 dst_reg->umin_value = umin_ptr;
5340 dst_reg->umax_value = umax_ptr;
5341 dst_reg->var_off = ptr_reg->var_off;
5342 dst_reg->id = ptr_reg->id;
5343 dst_reg->off = ptr_reg->off - smin_val;
5344 dst_reg->raw = ptr_reg->raw;
5347 /* A new variable offset is created. If the subtrahend is known
5348 * nonnegative, then any reg->range we had before is still good.
5350 if (signed_sub_overflows(smin_ptr, smax_val) ||
5351 signed_sub_overflows(smax_ptr, smin_val)) {
5352 /* Overflow possible, we know nothing */
5353 dst_reg->smin_value = S64_MIN;
5354 dst_reg->smax_value = S64_MAX;
5356 dst_reg->smin_value = smin_ptr - smax_val;
5357 dst_reg->smax_value = smax_ptr - smin_val;
5359 if (umin_ptr < umax_val) {
5360 /* Overflow possible, we know nothing */
5361 dst_reg->umin_value = 0;
5362 dst_reg->umax_value = U64_MAX;
5364 /* Cannot overflow (as long as bounds are consistent) */
5365 dst_reg->umin_value = umin_ptr - umax_val;
5366 dst_reg->umax_value = umax_ptr - umin_val;
5368 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
5369 dst_reg->off = ptr_reg->off;
5370 dst_reg->raw = ptr_reg->raw;
5371 if (reg_is_pkt_pointer(ptr_reg)) {
5372 dst_reg->id = ++env->id_gen;
5373 /* something was added to pkt_ptr, set range to zero */
5381 /* bitwise ops on pointers are troublesome, prohibit. */
5382 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
5383 dst, bpf_alu_string[opcode >> 4]);
5386 /* other operators (e.g. MUL,LSH) produce non-pointer results */
5387 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
5388 dst, bpf_alu_string[opcode >> 4]);
5392 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
5395 __update_reg_bounds(dst_reg);
5396 __reg_deduce_bounds(dst_reg);
5397 __reg_bound_offset(dst_reg);
5399 /* For unprivileged we require that resulting offset must be in bounds
5400 * in order to be able to sanitize access later on.
5402 if (!env->bypass_spec_v1) {
5403 if (dst_reg->type == PTR_TO_MAP_VALUE &&
5404 check_map_access(env, dst, dst_reg->off, 1, false)) {
5405 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
5406 "prohibited for !root\n", dst);
5408 } else if (dst_reg->type == PTR_TO_STACK &&
5409 check_stack_access(env, dst_reg, dst_reg->off +
5410 dst_reg->var_off.value, 1)) {
5411 verbose(env, "R%d stack pointer arithmetic goes out of range, "
5412 "prohibited for !root\n", dst);
5420 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
5421 struct bpf_reg_state *src_reg)
5423 s32 smin_val = src_reg->s32_min_value;
5424 s32 smax_val = src_reg->s32_max_value;
5425 u32 umin_val = src_reg->u32_min_value;
5426 u32 umax_val = src_reg->u32_max_value;
5428 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
5429 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
5430 dst_reg->s32_min_value = S32_MIN;
5431 dst_reg->s32_max_value = S32_MAX;
5433 dst_reg->s32_min_value += smin_val;
5434 dst_reg->s32_max_value += smax_val;
5436 if (dst_reg->u32_min_value + umin_val < umin_val ||
5437 dst_reg->u32_max_value + umax_val < umax_val) {
5438 dst_reg->u32_min_value = 0;
5439 dst_reg->u32_max_value = U32_MAX;
5441 dst_reg->u32_min_value += umin_val;
5442 dst_reg->u32_max_value += umax_val;
5446 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
5447 struct bpf_reg_state *src_reg)
5449 s64 smin_val = src_reg->smin_value;
5450 s64 smax_val = src_reg->smax_value;
5451 u64 umin_val = src_reg->umin_value;
5452 u64 umax_val = src_reg->umax_value;
5454 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
5455 signed_add_overflows(dst_reg->smax_value, smax_val)) {
5456 dst_reg->smin_value = S64_MIN;
5457 dst_reg->smax_value = S64_MAX;
5459 dst_reg->smin_value += smin_val;
5460 dst_reg->smax_value += smax_val;
5462 if (dst_reg->umin_value + umin_val < umin_val ||
5463 dst_reg->umax_value + umax_val < umax_val) {
5464 dst_reg->umin_value = 0;
5465 dst_reg->umax_value = U64_MAX;
5467 dst_reg->umin_value += umin_val;
5468 dst_reg->umax_value += umax_val;
5472 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
5473 struct bpf_reg_state *src_reg)
5475 s32 smin_val = src_reg->s32_min_value;
5476 s32 smax_val = src_reg->s32_max_value;
5477 u32 umin_val = src_reg->u32_min_value;
5478 u32 umax_val = src_reg->u32_max_value;
5480 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
5481 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
5482 /* Overflow possible, we know nothing */
5483 dst_reg->s32_min_value = S32_MIN;
5484 dst_reg->s32_max_value = S32_MAX;
5486 dst_reg->s32_min_value -= smax_val;
5487 dst_reg->s32_max_value -= smin_val;
5489 if (dst_reg->u32_min_value < umax_val) {
5490 /* Overflow possible, we know nothing */
5491 dst_reg->u32_min_value = 0;
5492 dst_reg->u32_max_value = U32_MAX;
5494 /* Cannot overflow (as long as bounds are consistent) */
5495 dst_reg->u32_min_value -= umax_val;
5496 dst_reg->u32_max_value -= umin_val;
5500 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
5501 struct bpf_reg_state *src_reg)
5503 s64 smin_val = src_reg->smin_value;
5504 s64 smax_val = src_reg->smax_value;
5505 u64 umin_val = src_reg->umin_value;
5506 u64 umax_val = src_reg->umax_value;
5508 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
5509 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
5510 /* Overflow possible, we know nothing */
5511 dst_reg->smin_value = S64_MIN;
5512 dst_reg->smax_value = S64_MAX;
5514 dst_reg->smin_value -= smax_val;
5515 dst_reg->smax_value -= smin_val;
5517 if (dst_reg->umin_value < umax_val) {
5518 /* Overflow possible, we know nothing */
5519 dst_reg->umin_value = 0;
5520 dst_reg->umax_value = U64_MAX;
5522 /* Cannot overflow (as long as bounds are consistent) */
5523 dst_reg->umin_value -= umax_val;
5524 dst_reg->umax_value -= umin_val;
5528 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
5529 struct bpf_reg_state *src_reg)
5531 s32 smin_val = src_reg->s32_min_value;
5532 u32 umin_val = src_reg->u32_min_value;
5533 u32 umax_val = src_reg->u32_max_value;
5535 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
5536 /* Ain't nobody got time to multiply that sign */
5537 __mark_reg32_unbounded(dst_reg);
5540 /* Both values are positive, so we can work with unsigned and
5541 * copy the result to signed (unless it exceeds S32_MAX).
5543 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
5544 /* Potential overflow, we know nothing */
5545 __mark_reg32_unbounded(dst_reg);
5548 dst_reg->u32_min_value *= umin_val;
5549 dst_reg->u32_max_value *= umax_val;
5550 if (dst_reg->u32_max_value > S32_MAX) {
5551 /* Overflow possible, we know nothing */
5552 dst_reg->s32_min_value = S32_MIN;
5553 dst_reg->s32_max_value = S32_MAX;
5555 dst_reg->s32_min_value = dst_reg->u32_min_value;
5556 dst_reg->s32_max_value = dst_reg->u32_max_value;
5560 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
5561 struct bpf_reg_state *src_reg)
5563 s64 smin_val = src_reg->smin_value;
5564 u64 umin_val = src_reg->umin_value;
5565 u64 umax_val = src_reg->umax_value;
5567 if (smin_val < 0 || dst_reg->smin_value < 0) {
5568 /* Ain't nobody got time to multiply that sign */
5569 __mark_reg64_unbounded(dst_reg);
5572 /* Both values are positive, so we can work with unsigned and
5573 * copy the result to signed (unless it exceeds S64_MAX).
5575 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
5576 /* Potential overflow, we know nothing */
5577 __mark_reg64_unbounded(dst_reg);
5580 dst_reg->umin_value *= umin_val;
5581 dst_reg->umax_value *= umax_val;
5582 if (dst_reg->umax_value > S64_MAX) {
5583 /* Overflow possible, we know nothing */
5584 dst_reg->smin_value = S64_MIN;
5585 dst_reg->smax_value = S64_MAX;
5587 dst_reg->smin_value = dst_reg->umin_value;
5588 dst_reg->smax_value = dst_reg->umax_value;
5592 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
5593 struct bpf_reg_state *src_reg)
5595 bool src_known = tnum_subreg_is_const(src_reg->var_off);
5596 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
5597 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
5598 s32 smin_val = src_reg->s32_min_value;
5599 u32 umax_val = src_reg->u32_max_value;
5601 /* Assuming scalar64_min_max_and will be called so its safe
5602 * to skip updating register for known 32-bit case.
5604 if (src_known && dst_known)
5607 /* We get our minimum from the var_off, since that's inherently
5608 * bitwise. Our maximum is the minimum of the operands' maxima.
5610 dst_reg->u32_min_value = var32_off.value;
5611 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
5612 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
5613 /* Lose signed bounds when ANDing negative numbers,
5614 * ain't nobody got time for that.
5616 dst_reg->s32_min_value = S32_MIN;
5617 dst_reg->s32_max_value = S32_MAX;
5619 /* ANDing two positives gives a positive, so safe to
5620 * cast result into s64.
5622 dst_reg->s32_min_value = dst_reg->u32_min_value;
5623 dst_reg->s32_max_value = dst_reg->u32_max_value;
5628 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
5629 struct bpf_reg_state *src_reg)
5631 bool src_known = tnum_is_const(src_reg->var_off);
5632 bool dst_known = tnum_is_const(dst_reg->var_off);
5633 s64 smin_val = src_reg->smin_value;
5634 u64 umax_val = src_reg->umax_value;
5636 if (src_known && dst_known) {
5637 __mark_reg_known(dst_reg, dst_reg->var_off.value &
5638 src_reg->var_off.value);
5642 /* We get our minimum from the var_off, since that's inherently
5643 * bitwise. Our maximum is the minimum of the operands' maxima.
5645 dst_reg->umin_value = dst_reg->var_off.value;
5646 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
5647 if (dst_reg->smin_value < 0 || smin_val < 0) {
5648 /* Lose signed bounds when ANDing negative numbers,
5649 * ain't nobody got time for that.
5651 dst_reg->smin_value = S64_MIN;
5652 dst_reg->smax_value = S64_MAX;
5654 /* ANDing two positives gives a positive, so safe to
5655 * cast result into s64.
5657 dst_reg->smin_value = dst_reg->umin_value;
5658 dst_reg->smax_value = dst_reg->umax_value;
5660 /* We may learn something more from the var_off */
5661 __update_reg_bounds(dst_reg);
5664 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
5665 struct bpf_reg_state *src_reg)
5667 bool src_known = tnum_subreg_is_const(src_reg->var_off);
5668 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
5669 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
5670 s32 smin_val = src_reg->smin_value;
5671 u32 umin_val = src_reg->umin_value;
5673 /* Assuming scalar64_min_max_or will be called so it is safe
5674 * to skip updating register for known case.
5676 if (src_known && dst_known)
5679 /* We get our maximum from the var_off, and our minimum is the
5680 * maximum of the operands' minima
5682 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
5683 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
5684 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
5685 /* Lose signed bounds when ORing negative numbers,
5686 * ain't nobody got time for that.
5688 dst_reg->s32_min_value = S32_MIN;
5689 dst_reg->s32_max_value = S32_MAX;
5691 /* ORing two positives gives a positive, so safe to
5692 * cast result into s64.
5694 dst_reg->s32_min_value = dst_reg->umin_value;
5695 dst_reg->s32_max_value = dst_reg->umax_value;
5699 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
5700 struct bpf_reg_state *src_reg)
5702 bool src_known = tnum_is_const(src_reg->var_off);
5703 bool dst_known = tnum_is_const(dst_reg->var_off);
5704 s64 smin_val = src_reg->smin_value;
5705 u64 umin_val = src_reg->umin_value;
5707 if (src_known && dst_known) {
5708 __mark_reg_known(dst_reg, dst_reg->var_off.value |
5709 src_reg->var_off.value);
5713 /* We get our maximum from the var_off, and our minimum is the
5714 * maximum of the operands' minima
5716 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
5717 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
5718 if (dst_reg->smin_value < 0 || smin_val < 0) {
5719 /* Lose signed bounds when ORing negative numbers,
5720 * ain't nobody got time for that.
5722 dst_reg->smin_value = S64_MIN;
5723 dst_reg->smax_value = S64_MAX;
5725 /* ORing two positives gives a positive, so safe to
5726 * cast result into s64.
5728 dst_reg->smin_value = dst_reg->umin_value;
5729 dst_reg->smax_value = dst_reg->umax_value;
5731 /* We may learn something more from the var_off */
5732 __update_reg_bounds(dst_reg);
5735 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
5736 u64 umin_val, u64 umax_val)
5738 /* We lose all sign bit information (except what we can pick
5741 dst_reg->s32_min_value = S32_MIN;
5742 dst_reg->s32_max_value = S32_MAX;
5743 /* If we might shift our top bit out, then we know nothing */
5744 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
5745 dst_reg->u32_min_value = 0;
5746 dst_reg->u32_max_value = U32_MAX;
5748 dst_reg->u32_min_value <<= umin_val;
5749 dst_reg->u32_max_value <<= umax_val;
5753 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
5754 struct bpf_reg_state *src_reg)
5756 u32 umax_val = src_reg->u32_max_value;
5757 u32 umin_val = src_reg->u32_min_value;
5758 /* u32 alu operation will zext upper bits */
5759 struct tnum subreg = tnum_subreg(dst_reg->var_off);
5761 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
5762 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
5763 /* Not required but being careful mark reg64 bounds as unknown so
5764 * that we are forced to pick them up from tnum and zext later and
5765 * if some path skips this step we are still safe.
5767 __mark_reg64_unbounded(dst_reg);
5768 __update_reg32_bounds(dst_reg);
5771 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
5772 u64 umin_val, u64 umax_val)
5774 /* Special case <<32 because it is a common compiler pattern to sign
5775 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
5776 * positive we know this shift will also be positive so we can track
5777 * bounds correctly. Otherwise we lose all sign bit information except
5778 * what we can pick up from var_off. Perhaps we can generalize this
5779 * later to shifts of any length.
5781 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
5782 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
5784 dst_reg->smax_value = S64_MAX;
5786 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
5787 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
5789 dst_reg->smin_value = S64_MIN;
5791 /* If we might shift our top bit out, then we know nothing */
5792 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
5793 dst_reg->umin_value = 0;
5794 dst_reg->umax_value = U64_MAX;
5796 dst_reg->umin_value <<= umin_val;
5797 dst_reg->umax_value <<= umax_val;
5801 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
5802 struct bpf_reg_state *src_reg)
5804 u64 umax_val = src_reg->umax_value;
5805 u64 umin_val = src_reg->umin_value;
5807 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
5808 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
5809 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
5811 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
5812 /* We may learn something more from the var_off */
5813 __update_reg_bounds(dst_reg);
5816 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
5817 struct bpf_reg_state *src_reg)
5819 struct tnum subreg = tnum_subreg(dst_reg->var_off);
5820 u32 umax_val = src_reg->u32_max_value;
5821 u32 umin_val = src_reg->u32_min_value;
5823 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
5824 * be negative, then either:
5825 * 1) src_reg might be zero, so the sign bit of the result is
5826 * unknown, so we lose our signed bounds
5827 * 2) it's known negative, thus the unsigned bounds capture the
5829 * 3) the signed bounds cross zero, so they tell us nothing
5831 * If the value in dst_reg is known nonnegative, then again the
5832 * unsigned bounts capture the signed bounds.
5833 * Thus, in all cases it suffices to blow away our signed bounds
5834 * and rely on inferring new ones from the unsigned bounds and
5835 * var_off of the result.
5837 dst_reg->s32_min_value = S32_MIN;
5838 dst_reg->s32_max_value = S32_MAX;
5840 dst_reg->var_off = tnum_rshift(subreg, umin_val);
5841 dst_reg->u32_min_value >>= umax_val;
5842 dst_reg->u32_max_value >>= umin_val;
5844 __mark_reg64_unbounded(dst_reg);
5845 __update_reg32_bounds(dst_reg);
5848 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
5849 struct bpf_reg_state *src_reg)
5851 u64 umax_val = src_reg->umax_value;
5852 u64 umin_val = src_reg->umin_value;
5854 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
5855 * be negative, then either:
5856 * 1) src_reg might be zero, so the sign bit of the result is
5857 * unknown, so we lose our signed bounds
5858 * 2) it's known negative, thus the unsigned bounds capture the
5860 * 3) the signed bounds cross zero, so they tell us nothing
5862 * If the value in dst_reg is known nonnegative, then again the
5863 * unsigned bounts capture the signed bounds.
5864 * Thus, in all cases it suffices to blow away our signed bounds
5865 * and rely on inferring new ones from the unsigned bounds and
5866 * var_off of the result.
5868 dst_reg->smin_value = S64_MIN;
5869 dst_reg->smax_value = S64_MAX;
5870 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
5871 dst_reg->umin_value >>= umax_val;
5872 dst_reg->umax_value >>= umin_val;
5874 /* Its not easy to operate on alu32 bounds here because it depends
5875 * on bits being shifted in. Take easy way out and mark unbounded
5876 * so we can recalculate later from tnum.
5878 __mark_reg32_unbounded(dst_reg);
5879 __update_reg_bounds(dst_reg);
5882 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
5883 struct bpf_reg_state *src_reg)
5885 u64 umin_val = src_reg->u32_min_value;
5887 /* Upon reaching here, src_known is true and
5888 * umax_val is equal to umin_val.
5890 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
5891 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
5893 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
5895 /* blow away the dst_reg umin_value/umax_value and rely on
5896 * dst_reg var_off to refine the result.
5898 dst_reg->u32_min_value = 0;
5899 dst_reg->u32_max_value = U32_MAX;
5901 __mark_reg64_unbounded(dst_reg);
5902 __update_reg32_bounds(dst_reg);
5905 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
5906 struct bpf_reg_state *src_reg)
5908 u64 umin_val = src_reg->umin_value;
5910 /* Upon reaching here, src_known is true and umax_val is equal
5913 dst_reg->smin_value >>= umin_val;
5914 dst_reg->smax_value >>= umin_val;
5916 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
5918 /* blow away the dst_reg umin_value/umax_value and rely on
5919 * dst_reg var_off to refine the result.
5921 dst_reg->umin_value = 0;
5922 dst_reg->umax_value = U64_MAX;
5924 /* Its not easy to operate on alu32 bounds here because it depends
5925 * on bits being shifted in from upper 32-bits. Take easy way out
5926 * and mark unbounded so we can recalculate later from tnum.
5928 __mark_reg32_unbounded(dst_reg);
5929 __update_reg_bounds(dst_reg);
5932 /* WARNING: This function does calculations on 64-bit values, but the actual
5933 * execution may occur on 32-bit values. Therefore, things like bitshifts
5934 * need extra checks in the 32-bit case.
5936 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
5937 struct bpf_insn *insn,
5938 struct bpf_reg_state *dst_reg,
5939 struct bpf_reg_state src_reg)
5941 struct bpf_reg_state *regs = cur_regs(env);
5942 u8 opcode = BPF_OP(insn->code);
5944 s64 smin_val, smax_val;
5945 u64 umin_val, umax_val;
5946 s32 s32_min_val, s32_max_val;
5947 u32 u32_min_val, u32_max_val;
5948 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
5949 u32 dst = insn->dst_reg;
5951 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
5953 smin_val = src_reg.smin_value;
5954 smax_val = src_reg.smax_value;
5955 umin_val = src_reg.umin_value;
5956 umax_val = src_reg.umax_value;
5958 s32_min_val = src_reg.s32_min_value;
5959 s32_max_val = src_reg.s32_max_value;
5960 u32_min_val = src_reg.u32_min_value;
5961 u32_max_val = src_reg.u32_max_value;
5964 src_known = tnum_subreg_is_const(src_reg.var_off);
5966 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
5967 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
5968 /* Taint dst register if offset had invalid bounds
5969 * derived from e.g. dead branches.
5971 __mark_reg_unknown(env, dst_reg);
5975 src_known = tnum_is_const(src_reg.var_off);
5977 (smin_val != smax_val || umin_val != umax_val)) ||
5978 smin_val > smax_val || umin_val > umax_val) {
5979 /* Taint dst register if offset had invalid bounds
5980 * derived from e.g. dead branches.
5982 __mark_reg_unknown(env, dst_reg);
5988 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
5989 __mark_reg_unknown(env, dst_reg);
5993 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
5994 * There are two classes of instructions: The first class we track both
5995 * alu32 and alu64 sign/unsigned bounds independently this provides the
5996 * greatest amount of precision when alu operations are mixed with jmp32
5997 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
5998 * and BPF_OR. This is possible because these ops have fairly easy to
5999 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
6000 * See alu32 verifier tests for examples. The second class of
6001 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
6002 * with regards to tracking sign/unsigned bounds because the bits may
6003 * cross subreg boundaries in the alu64 case. When this happens we mark
6004 * the reg unbounded in the subreg bound space and use the resulting
6005 * tnum to calculate an approximation of the sign/unsigned bounds.
6009 ret = sanitize_val_alu(env, insn);
6011 verbose(env, "R%d tried to add from different pointers or scalars\n", dst);
6014 scalar32_min_max_add(dst_reg, &src_reg);
6015 scalar_min_max_add(dst_reg, &src_reg);
6016 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
6019 ret = sanitize_val_alu(env, insn);
6021 verbose(env, "R%d tried to sub from different pointers or scalars\n", dst);
6024 scalar32_min_max_sub(dst_reg, &src_reg);
6025 scalar_min_max_sub(dst_reg, &src_reg);
6026 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
6029 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
6030 scalar32_min_max_mul(dst_reg, &src_reg);
6031 scalar_min_max_mul(dst_reg, &src_reg);
6034 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
6035 scalar32_min_max_and(dst_reg, &src_reg);
6036 scalar_min_max_and(dst_reg, &src_reg);
6039 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
6040 scalar32_min_max_or(dst_reg, &src_reg);
6041 scalar_min_max_or(dst_reg, &src_reg);
6044 if (umax_val >= insn_bitness) {
6045 /* Shifts greater than 31 or 63 are undefined.
6046 * This includes shifts by a negative number.
6048 mark_reg_unknown(env, regs, insn->dst_reg);
6052 scalar32_min_max_lsh(dst_reg, &src_reg);
6054 scalar_min_max_lsh(dst_reg, &src_reg);
6057 if (umax_val >= insn_bitness) {
6058 /* Shifts greater than 31 or 63 are undefined.
6059 * This includes shifts by a negative number.
6061 mark_reg_unknown(env, regs, insn->dst_reg);
6065 scalar32_min_max_rsh(dst_reg, &src_reg);
6067 scalar_min_max_rsh(dst_reg, &src_reg);
6070 if (umax_val >= insn_bitness) {
6071 /* Shifts greater than 31 or 63 are undefined.
6072 * This includes shifts by a negative number.
6074 mark_reg_unknown(env, regs, insn->dst_reg);
6078 scalar32_min_max_arsh(dst_reg, &src_reg);
6080 scalar_min_max_arsh(dst_reg, &src_reg);
6083 mark_reg_unknown(env, regs, insn->dst_reg);
6087 /* ALU32 ops are zero extended into 64bit register */
6089 zext_32_to_64(dst_reg);
6091 __update_reg_bounds(dst_reg);
6092 __reg_deduce_bounds(dst_reg);
6093 __reg_bound_offset(dst_reg);
6097 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
6100 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
6101 struct bpf_insn *insn)
6103 struct bpf_verifier_state *vstate = env->cur_state;
6104 struct bpf_func_state *state = vstate->frame[vstate->curframe];
6105 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
6106 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
6107 u8 opcode = BPF_OP(insn->code);
6110 dst_reg = ®s[insn->dst_reg];
6112 if (dst_reg->type != SCALAR_VALUE)
6114 if (BPF_SRC(insn->code) == BPF_X) {
6115 src_reg = ®s[insn->src_reg];
6116 if (src_reg->type != SCALAR_VALUE) {
6117 if (dst_reg->type != SCALAR_VALUE) {
6118 /* Combining two pointers by any ALU op yields
6119 * an arbitrary scalar. Disallow all math except
6120 * pointer subtraction
6122 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
6123 mark_reg_unknown(env, regs, insn->dst_reg);
6126 verbose(env, "R%d pointer %s pointer prohibited\n",
6128 bpf_alu_string[opcode >> 4]);
6131 /* scalar += pointer
6132 * This is legal, but we have to reverse our
6133 * src/dest handling in computing the range
6135 err = mark_chain_precision(env, insn->dst_reg);
6138 return adjust_ptr_min_max_vals(env, insn,
6141 } else if (ptr_reg) {
6142 /* pointer += scalar */
6143 err = mark_chain_precision(env, insn->src_reg);
6146 return adjust_ptr_min_max_vals(env, insn,
6150 /* Pretend the src is a reg with a known value, since we only
6151 * need to be able to read from this state.
6153 off_reg.type = SCALAR_VALUE;
6154 __mark_reg_known(&off_reg, insn->imm);
6156 if (ptr_reg) /* pointer += K */
6157 return adjust_ptr_min_max_vals(env, insn,
6161 /* Got here implies adding two SCALAR_VALUEs */
6162 if (WARN_ON_ONCE(ptr_reg)) {
6163 print_verifier_state(env, state);
6164 verbose(env, "verifier internal error: unexpected ptr_reg\n");
6167 if (WARN_ON(!src_reg)) {
6168 print_verifier_state(env, state);
6169 verbose(env, "verifier internal error: no src_reg\n");
6172 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
6175 /* check validity of 32-bit and 64-bit arithmetic operations */
6176 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
6178 struct bpf_reg_state *regs = cur_regs(env);
6179 u8 opcode = BPF_OP(insn->code);
6182 if (opcode == BPF_END || opcode == BPF_NEG) {
6183 if (opcode == BPF_NEG) {
6184 if (BPF_SRC(insn->code) != 0 ||
6185 insn->src_reg != BPF_REG_0 ||
6186 insn->off != 0 || insn->imm != 0) {
6187 verbose(env, "BPF_NEG uses reserved fields\n");
6191 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
6192 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
6193 BPF_CLASS(insn->code) == BPF_ALU64) {
6194 verbose(env, "BPF_END uses reserved fields\n");
6199 /* check src operand */
6200 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6204 if (is_pointer_value(env, insn->dst_reg)) {
6205 verbose(env, "R%d pointer arithmetic prohibited\n",
6210 /* check dest operand */
6211 err = check_reg_arg(env, insn->dst_reg, DST_OP);
6215 } else if (opcode == BPF_MOV) {
6217 if (BPF_SRC(insn->code) == BPF_X) {
6218 if (insn->imm != 0 || insn->off != 0) {
6219 verbose(env, "BPF_MOV uses reserved fields\n");
6223 /* check src operand */
6224 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6228 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
6229 verbose(env, "BPF_MOV uses reserved fields\n");
6234 /* check dest operand, mark as required later */
6235 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
6239 if (BPF_SRC(insn->code) == BPF_X) {
6240 struct bpf_reg_state *src_reg = regs + insn->src_reg;
6241 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
6243 if (BPF_CLASS(insn->code) == BPF_ALU64) {
6245 * copy register state to dest reg
6247 *dst_reg = *src_reg;
6248 dst_reg->live |= REG_LIVE_WRITTEN;
6249 dst_reg->subreg_def = DEF_NOT_SUBREG;
6252 if (is_pointer_value(env, insn->src_reg)) {
6254 "R%d partial copy of pointer\n",
6257 } else if (src_reg->type == SCALAR_VALUE) {
6258 *dst_reg = *src_reg;
6259 dst_reg->live |= REG_LIVE_WRITTEN;
6260 dst_reg->subreg_def = env->insn_idx + 1;
6262 mark_reg_unknown(env, regs,
6265 zext_32_to_64(dst_reg);
6269 * remember the value we stored into this reg
6271 /* clear any state __mark_reg_known doesn't set */
6272 mark_reg_unknown(env, regs, insn->dst_reg);
6273 regs[insn->dst_reg].type = SCALAR_VALUE;
6274 if (BPF_CLASS(insn->code) == BPF_ALU64) {
6275 __mark_reg_known(regs + insn->dst_reg,
6278 __mark_reg_known(regs + insn->dst_reg,
6283 } else if (opcode > BPF_END) {
6284 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
6287 } else { /* all other ALU ops: and, sub, xor, add, ... */
6289 if (BPF_SRC(insn->code) == BPF_X) {
6290 if (insn->imm != 0 || insn->off != 0) {
6291 verbose(env, "BPF_ALU uses reserved fields\n");
6294 /* check src1 operand */
6295 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6299 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
6300 verbose(env, "BPF_ALU uses reserved fields\n");
6305 /* check src2 operand */
6306 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6310 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
6311 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
6312 verbose(env, "div by zero\n");
6316 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
6317 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
6318 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
6320 if (insn->imm < 0 || insn->imm >= size) {
6321 verbose(env, "invalid shift %d\n", insn->imm);
6326 /* check dest operand */
6327 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
6331 return adjust_reg_min_max_vals(env, insn);
6337 static void __find_good_pkt_pointers(struct bpf_func_state *state,
6338 struct bpf_reg_state *dst_reg,
6339 enum bpf_reg_type type, u16 new_range)
6341 struct bpf_reg_state *reg;
6344 for (i = 0; i < MAX_BPF_REG; i++) {
6345 reg = &state->regs[i];
6346 if (reg->type == type && reg->id == dst_reg->id)
6347 /* keep the maximum range already checked */
6348 reg->range = max(reg->range, new_range);
6351 bpf_for_each_spilled_reg(i, state, reg) {
6354 if (reg->type == type && reg->id == dst_reg->id)
6355 reg->range = max(reg->range, new_range);
6359 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
6360 struct bpf_reg_state *dst_reg,
6361 enum bpf_reg_type type,
6362 bool range_right_open)
6367 if (dst_reg->off < 0 ||
6368 (dst_reg->off == 0 && range_right_open))
6369 /* This doesn't give us any range */
6372 if (dst_reg->umax_value > MAX_PACKET_OFF ||
6373 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
6374 /* Risk of overflow. For instance, ptr + (1<<63) may be less
6375 * than pkt_end, but that's because it's also less than pkt.
6379 new_range = dst_reg->off;
6380 if (range_right_open)
6383 /* Examples for register markings:
6385 * pkt_data in dst register:
6389 * if (r2 > pkt_end) goto <handle exception>
6394 * if (r2 < pkt_end) goto <access okay>
6395 * <handle exception>
6398 * r2 == dst_reg, pkt_end == src_reg
6399 * r2=pkt(id=n,off=8,r=0)
6400 * r3=pkt(id=n,off=0,r=0)
6402 * pkt_data in src register:
6406 * if (pkt_end >= r2) goto <access okay>
6407 * <handle exception>
6411 * if (pkt_end <= r2) goto <handle exception>
6415 * pkt_end == dst_reg, r2 == src_reg
6416 * r2=pkt(id=n,off=8,r=0)
6417 * r3=pkt(id=n,off=0,r=0)
6419 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
6420 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
6421 * and [r3, r3 + 8-1) respectively is safe to access depending on
6425 /* If our ids match, then we must have the same max_value. And we
6426 * don't care about the other reg's fixed offset, since if it's too big
6427 * the range won't allow anything.
6428 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
6430 for (i = 0; i <= vstate->curframe; i++)
6431 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
6435 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
6437 struct tnum subreg = tnum_subreg(reg->var_off);
6438 s32 sval = (s32)val;
6442 if (tnum_is_const(subreg))
6443 return !!tnum_equals_const(subreg, val);
6446 if (tnum_is_const(subreg))
6447 return !tnum_equals_const(subreg, val);
6450 if ((~subreg.mask & subreg.value) & val)
6452 if (!((subreg.mask | subreg.value) & val))
6456 if (reg->u32_min_value > val)
6458 else if (reg->u32_max_value <= val)
6462 if (reg->s32_min_value > sval)
6464 else if (reg->s32_max_value < sval)
6468 if (reg->u32_max_value < val)
6470 else if (reg->u32_min_value >= val)
6474 if (reg->s32_max_value < sval)
6476 else if (reg->s32_min_value >= sval)
6480 if (reg->u32_min_value >= val)
6482 else if (reg->u32_max_value < val)
6486 if (reg->s32_min_value >= sval)
6488 else if (reg->s32_max_value < sval)
6492 if (reg->u32_max_value <= val)
6494 else if (reg->u32_min_value > val)
6498 if (reg->s32_max_value <= sval)
6500 else if (reg->s32_min_value > sval)
6509 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
6511 s64 sval = (s64)val;
6515 if (tnum_is_const(reg->var_off))
6516 return !!tnum_equals_const(reg->var_off, val);
6519 if (tnum_is_const(reg->var_off))
6520 return !tnum_equals_const(reg->var_off, val);
6523 if ((~reg->var_off.mask & reg->var_off.value) & val)
6525 if (!((reg->var_off.mask | reg->var_off.value) & val))
6529 if (reg->umin_value > val)
6531 else if (reg->umax_value <= val)
6535 if (reg->smin_value > sval)
6537 else if (reg->smax_value < sval)
6541 if (reg->umax_value < val)
6543 else if (reg->umin_value >= val)
6547 if (reg->smax_value < sval)
6549 else if (reg->smin_value >= sval)
6553 if (reg->umin_value >= val)
6555 else if (reg->umax_value < val)
6559 if (reg->smin_value >= sval)
6561 else if (reg->smax_value < sval)
6565 if (reg->umax_value <= val)
6567 else if (reg->umin_value > val)
6571 if (reg->smax_value <= sval)
6573 else if (reg->smin_value > sval)
6581 /* compute branch direction of the expression "if (reg opcode val) goto target;"
6583 * 1 - branch will be taken and "goto target" will be executed
6584 * 0 - branch will not be taken and fall-through to next insn
6585 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
6588 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
6591 if (__is_pointer_value(false, reg)) {
6592 if (!reg_type_not_null(reg->type))
6595 /* If pointer is valid tests against zero will fail so we can
6596 * use this to direct branch taken.
6612 return is_branch32_taken(reg, val, opcode);
6613 return is_branch64_taken(reg, val, opcode);
6616 /* Adjusts the register min/max values in the case that the dst_reg is the
6617 * variable register that we are working on, and src_reg is a constant or we're
6618 * simply doing a BPF_K check.
6619 * In JEQ/JNE cases we also adjust the var_off values.
6621 static void reg_set_min_max(struct bpf_reg_state *true_reg,
6622 struct bpf_reg_state *false_reg,
6624 u8 opcode, bool is_jmp32)
6626 struct tnum false_32off = tnum_subreg(false_reg->var_off);
6627 struct tnum false_64off = false_reg->var_off;
6628 struct tnum true_32off = tnum_subreg(true_reg->var_off);
6629 struct tnum true_64off = true_reg->var_off;
6630 s64 sval = (s64)val;
6631 s32 sval32 = (s32)val32;
6633 /* If the dst_reg is a pointer, we can't learn anything about its
6634 * variable offset from the compare (unless src_reg were a pointer into
6635 * the same object, but we don't bother with that.
6636 * Since false_reg and true_reg have the same type by construction, we
6637 * only need to check one of them for pointerness.
6639 if (__is_pointer_value(false, false_reg))
6646 struct bpf_reg_state *reg =
6647 opcode == BPF_JEQ ? true_reg : false_reg;
6649 /* For BPF_JEQ, if this is false we know nothing Jon Snow, but
6650 * if it is true we know the value for sure. Likewise for
6654 __mark_reg32_known(reg, val32);
6656 __mark_reg_known(reg, val);
6661 false_32off = tnum_and(false_32off, tnum_const(~val32));
6662 if (is_power_of_2(val32))
6663 true_32off = tnum_or(true_32off,
6666 false_64off = tnum_and(false_64off, tnum_const(~val));
6667 if (is_power_of_2(val))
6668 true_64off = tnum_or(true_64off,
6676 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
6677 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
6679 false_reg->u32_max_value = min(false_reg->u32_max_value,
6681 true_reg->u32_min_value = max(true_reg->u32_min_value,
6684 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
6685 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
6687 false_reg->umax_value = min(false_reg->umax_value, false_umax);
6688 true_reg->umin_value = max(true_reg->umin_value, true_umin);
6696 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
6697 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
6699 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
6700 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
6702 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
6703 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
6705 false_reg->smax_value = min(false_reg->smax_value, false_smax);
6706 true_reg->smin_value = max(true_reg->smin_value, true_smin);
6714 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
6715 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
6717 false_reg->u32_min_value = max(false_reg->u32_min_value,
6719 true_reg->u32_max_value = min(true_reg->u32_max_value,
6722 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
6723 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
6725 false_reg->umin_value = max(false_reg->umin_value, false_umin);
6726 true_reg->umax_value = min(true_reg->umax_value, true_umax);
6734 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
6735 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
6737 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
6738 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
6740 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
6741 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
6743 false_reg->smin_value = max(false_reg->smin_value, false_smin);
6744 true_reg->smax_value = min(true_reg->smax_value, true_smax);
6753 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
6754 tnum_subreg(false_32off));
6755 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
6756 tnum_subreg(true_32off));
6757 __reg_combine_32_into_64(false_reg);
6758 __reg_combine_32_into_64(true_reg);
6760 false_reg->var_off = false_64off;
6761 true_reg->var_off = true_64off;
6762 __reg_combine_64_into_32(false_reg);
6763 __reg_combine_64_into_32(true_reg);
6767 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
6770 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
6771 struct bpf_reg_state *false_reg,
6773 u8 opcode, bool is_jmp32)
6775 /* How can we transform "a <op> b" into "b <op> a"? */
6776 static const u8 opcode_flip[16] = {
6777 /* these stay the same */
6778 [BPF_JEQ >> 4] = BPF_JEQ,
6779 [BPF_JNE >> 4] = BPF_JNE,
6780 [BPF_JSET >> 4] = BPF_JSET,
6781 /* these swap "lesser" and "greater" (L and G in the opcodes) */
6782 [BPF_JGE >> 4] = BPF_JLE,
6783 [BPF_JGT >> 4] = BPF_JLT,
6784 [BPF_JLE >> 4] = BPF_JGE,
6785 [BPF_JLT >> 4] = BPF_JGT,
6786 [BPF_JSGE >> 4] = BPF_JSLE,
6787 [BPF_JSGT >> 4] = BPF_JSLT,
6788 [BPF_JSLE >> 4] = BPF_JSGE,
6789 [BPF_JSLT >> 4] = BPF_JSGT
6791 opcode = opcode_flip[opcode >> 4];
6792 /* This uses zero as "not present in table"; luckily the zero opcode,
6793 * BPF_JA, can't get here.
6796 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
6799 /* Regs are known to be equal, so intersect their min/max/var_off */
6800 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
6801 struct bpf_reg_state *dst_reg)
6803 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
6804 dst_reg->umin_value);
6805 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
6806 dst_reg->umax_value);
6807 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
6808 dst_reg->smin_value);
6809 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
6810 dst_reg->smax_value);
6811 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
6813 /* We might have learned new bounds from the var_off. */
6814 __update_reg_bounds(src_reg);
6815 __update_reg_bounds(dst_reg);
6816 /* We might have learned something about the sign bit. */
6817 __reg_deduce_bounds(src_reg);
6818 __reg_deduce_bounds(dst_reg);
6819 /* We might have learned some bits from the bounds. */
6820 __reg_bound_offset(src_reg);
6821 __reg_bound_offset(dst_reg);
6822 /* Intersecting with the old var_off might have improved our bounds
6823 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
6824 * then new var_off is (0; 0x7f...fc) which improves our umax.
6826 __update_reg_bounds(src_reg);
6827 __update_reg_bounds(dst_reg);
6830 static void reg_combine_min_max(struct bpf_reg_state *true_src,
6831 struct bpf_reg_state *true_dst,
6832 struct bpf_reg_state *false_src,
6833 struct bpf_reg_state *false_dst,
6838 __reg_combine_min_max(true_src, true_dst);
6841 __reg_combine_min_max(false_src, false_dst);
6846 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
6847 struct bpf_reg_state *reg, u32 id,
6850 if (reg_type_may_be_null(reg->type) && reg->id == id) {
6851 /* Old offset (both fixed and variable parts) should
6852 * have been known-zero, because we don't allow pointer
6853 * arithmetic on pointers that might be NULL.
6855 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
6856 !tnum_equals_const(reg->var_off, 0) ||
6858 __mark_reg_known_zero(reg);
6862 reg->type = SCALAR_VALUE;
6863 } else if (reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
6864 const struct bpf_map *map = reg->map_ptr;
6866 if (map->inner_map_meta) {
6867 reg->type = CONST_PTR_TO_MAP;
6868 reg->map_ptr = map->inner_map_meta;
6869 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
6870 reg->type = PTR_TO_XDP_SOCK;
6871 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
6872 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
6873 reg->type = PTR_TO_SOCKET;
6875 reg->type = PTR_TO_MAP_VALUE;
6877 } else if (reg->type == PTR_TO_SOCKET_OR_NULL) {
6878 reg->type = PTR_TO_SOCKET;
6879 } else if (reg->type == PTR_TO_SOCK_COMMON_OR_NULL) {
6880 reg->type = PTR_TO_SOCK_COMMON;
6881 } else if (reg->type == PTR_TO_TCP_SOCK_OR_NULL) {
6882 reg->type = PTR_TO_TCP_SOCK;
6883 } else if (reg->type == PTR_TO_BTF_ID_OR_NULL) {
6884 reg->type = PTR_TO_BTF_ID;
6885 } else if (reg->type == PTR_TO_MEM_OR_NULL) {
6886 reg->type = PTR_TO_MEM;
6887 } else if (reg->type == PTR_TO_RDONLY_BUF_OR_NULL) {
6888 reg->type = PTR_TO_RDONLY_BUF;
6889 } else if (reg->type == PTR_TO_RDWR_BUF_OR_NULL) {
6890 reg->type = PTR_TO_RDWR_BUF;
6893 /* We don't need id and ref_obj_id from this point
6894 * onwards anymore, thus we should better reset it,
6895 * so that state pruning has chances to take effect.
6898 reg->ref_obj_id = 0;
6899 } else if (!reg_may_point_to_spin_lock(reg)) {
6900 /* For not-NULL ptr, reg->ref_obj_id will be reset
6901 * in release_reg_references().
6903 * reg->id is still used by spin_lock ptr. Other
6904 * than spin_lock ptr type, reg->id can be reset.
6911 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
6914 struct bpf_reg_state *reg;
6917 for (i = 0; i < MAX_BPF_REG; i++)
6918 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
6920 bpf_for_each_spilled_reg(i, state, reg) {
6923 mark_ptr_or_null_reg(state, reg, id, is_null);
6927 /* The logic is similar to find_good_pkt_pointers(), both could eventually
6928 * be folded together at some point.
6930 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
6933 struct bpf_func_state *state = vstate->frame[vstate->curframe];
6934 struct bpf_reg_state *regs = state->regs;
6935 u32 ref_obj_id = regs[regno].ref_obj_id;
6936 u32 id = regs[regno].id;
6939 if (ref_obj_id && ref_obj_id == id && is_null)
6940 /* regs[regno] is in the " == NULL" branch.
6941 * No one could have freed the reference state before
6942 * doing the NULL check.
6944 WARN_ON_ONCE(release_reference_state(state, id));
6946 for (i = 0; i <= vstate->curframe; i++)
6947 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
6950 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
6951 struct bpf_reg_state *dst_reg,
6952 struct bpf_reg_state *src_reg,
6953 struct bpf_verifier_state *this_branch,
6954 struct bpf_verifier_state *other_branch)
6956 if (BPF_SRC(insn->code) != BPF_X)
6959 /* Pointers are always 64-bit. */
6960 if (BPF_CLASS(insn->code) == BPF_JMP32)
6963 switch (BPF_OP(insn->code)) {
6965 if ((dst_reg->type == PTR_TO_PACKET &&
6966 src_reg->type == PTR_TO_PACKET_END) ||
6967 (dst_reg->type == PTR_TO_PACKET_META &&
6968 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
6969 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
6970 find_good_pkt_pointers(this_branch, dst_reg,
6971 dst_reg->type, false);
6972 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6973 src_reg->type == PTR_TO_PACKET) ||
6974 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6975 src_reg->type == PTR_TO_PACKET_META)) {
6976 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
6977 find_good_pkt_pointers(other_branch, src_reg,
6978 src_reg->type, true);
6984 if ((dst_reg->type == PTR_TO_PACKET &&
6985 src_reg->type == PTR_TO_PACKET_END) ||
6986 (dst_reg->type == PTR_TO_PACKET_META &&
6987 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
6988 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
6989 find_good_pkt_pointers(other_branch, dst_reg,
6990 dst_reg->type, true);
6991 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6992 src_reg->type == PTR_TO_PACKET) ||
6993 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6994 src_reg->type == PTR_TO_PACKET_META)) {
6995 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
6996 find_good_pkt_pointers(this_branch, src_reg,
6997 src_reg->type, false);
7003 if ((dst_reg->type == PTR_TO_PACKET &&
7004 src_reg->type == PTR_TO_PACKET_END) ||
7005 (dst_reg->type == PTR_TO_PACKET_META &&
7006 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
7007 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
7008 find_good_pkt_pointers(this_branch, dst_reg,
7009 dst_reg->type, true);
7010 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
7011 src_reg->type == PTR_TO_PACKET) ||
7012 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
7013 src_reg->type == PTR_TO_PACKET_META)) {
7014 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
7015 find_good_pkt_pointers(other_branch, src_reg,
7016 src_reg->type, false);
7022 if ((dst_reg->type == PTR_TO_PACKET &&
7023 src_reg->type == PTR_TO_PACKET_END) ||
7024 (dst_reg->type == PTR_TO_PACKET_META &&
7025 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
7026 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
7027 find_good_pkt_pointers(other_branch, dst_reg,
7028 dst_reg->type, false);
7029 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
7030 src_reg->type == PTR_TO_PACKET) ||
7031 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
7032 src_reg->type == PTR_TO_PACKET_META)) {
7033 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
7034 find_good_pkt_pointers(this_branch, src_reg,
7035 src_reg->type, true);
7047 static int check_cond_jmp_op(struct bpf_verifier_env *env,
7048 struct bpf_insn *insn, int *insn_idx)
7050 struct bpf_verifier_state *this_branch = env->cur_state;
7051 struct bpf_verifier_state *other_branch;
7052 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
7053 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
7054 u8 opcode = BPF_OP(insn->code);
7059 /* Only conditional jumps are expected to reach here. */
7060 if (opcode == BPF_JA || opcode > BPF_JSLE) {
7061 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
7065 if (BPF_SRC(insn->code) == BPF_X) {
7066 if (insn->imm != 0) {
7067 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
7071 /* check src1 operand */
7072 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7076 if (is_pointer_value(env, insn->src_reg)) {
7077 verbose(env, "R%d pointer comparison prohibited\n",
7081 src_reg = ®s[insn->src_reg];
7083 if (insn->src_reg != BPF_REG_0) {
7084 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
7089 /* check src2 operand */
7090 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7094 dst_reg = ®s[insn->dst_reg];
7095 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
7097 if (BPF_SRC(insn->code) == BPF_K) {
7098 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
7099 } else if (src_reg->type == SCALAR_VALUE &&
7100 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
7101 pred = is_branch_taken(dst_reg,
7102 tnum_subreg(src_reg->var_off).value,
7105 } else if (src_reg->type == SCALAR_VALUE &&
7106 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
7107 pred = is_branch_taken(dst_reg,
7108 src_reg->var_off.value,
7114 /* If we get here with a dst_reg pointer type it is because
7115 * above is_branch_taken() special cased the 0 comparison.
7117 if (!__is_pointer_value(false, dst_reg))
7118 err = mark_chain_precision(env, insn->dst_reg);
7119 if (BPF_SRC(insn->code) == BPF_X && !err)
7120 err = mark_chain_precision(env, insn->src_reg);
7125 /* only follow the goto, ignore fall-through */
7126 *insn_idx += insn->off;
7128 } else if (pred == 0) {
7129 /* only follow fall-through branch, since
7130 * that's where the program will go
7135 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
7139 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
7141 /* detect if we are comparing against a constant value so we can adjust
7142 * our min/max values for our dst register.
7143 * this is only legit if both are scalars (or pointers to the same
7144 * object, I suppose, but we don't support that right now), because
7145 * otherwise the different base pointers mean the offsets aren't
7148 if (BPF_SRC(insn->code) == BPF_X) {
7149 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
7151 if (dst_reg->type == SCALAR_VALUE &&
7152 src_reg->type == SCALAR_VALUE) {
7153 if (tnum_is_const(src_reg->var_off) ||
7155 tnum_is_const(tnum_subreg(src_reg->var_off))))
7156 reg_set_min_max(&other_branch_regs[insn->dst_reg],
7158 src_reg->var_off.value,
7159 tnum_subreg(src_reg->var_off).value,
7161 else if (tnum_is_const(dst_reg->var_off) ||
7163 tnum_is_const(tnum_subreg(dst_reg->var_off))))
7164 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
7166 dst_reg->var_off.value,
7167 tnum_subreg(dst_reg->var_off).value,
7169 else if (!is_jmp32 &&
7170 (opcode == BPF_JEQ || opcode == BPF_JNE))
7171 /* Comparing for equality, we can combine knowledge */
7172 reg_combine_min_max(&other_branch_regs[insn->src_reg],
7173 &other_branch_regs[insn->dst_reg],
7174 src_reg, dst_reg, opcode);
7176 } else if (dst_reg->type == SCALAR_VALUE) {
7177 reg_set_min_max(&other_branch_regs[insn->dst_reg],
7178 dst_reg, insn->imm, (u32)insn->imm,
7182 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
7183 * NOTE: these optimizations below are related with pointer comparison
7184 * which will never be JMP32.
7186 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
7187 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
7188 reg_type_may_be_null(dst_reg->type)) {
7189 /* Mark all identical registers in each branch as either
7190 * safe or unknown depending R == 0 or R != 0 conditional.
7192 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
7194 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
7196 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
7197 this_branch, other_branch) &&
7198 is_pointer_value(env, insn->dst_reg)) {
7199 verbose(env, "R%d pointer comparison prohibited\n",
7203 if (env->log.level & BPF_LOG_LEVEL)
7204 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
7208 /* verify BPF_LD_IMM64 instruction */
7209 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
7211 struct bpf_insn_aux_data *aux = cur_aux(env);
7212 struct bpf_reg_state *regs = cur_regs(env);
7213 struct bpf_map *map;
7216 if (BPF_SIZE(insn->code) != BPF_DW) {
7217 verbose(env, "invalid BPF_LD_IMM insn\n");
7220 if (insn->off != 0) {
7221 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
7225 err = check_reg_arg(env, insn->dst_reg, DST_OP);
7229 if (insn->src_reg == 0) {
7230 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
7232 regs[insn->dst_reg].type = SCALAR_VALUE;
7233 __mark_reg_known(®s[insn->dst_reg], imm);
7237 map = env->used_maps[aux->map_index];
7238 mark_reg_known_zero(env, regs, insn->dst_reg);
7239 regs[insn->dst_reg].map_ptr = map;
7241 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE) {
7242 regs[insn->dst_reg].type = PTR_TO_MAP_VALUE;
7243 regs[insn->dst_reg].off = aux->map_off;
7244 if (map_value_has_spin_lock(map))
7245 regs[insn->dst_reg].id = ++env->id_gen;
7246 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
7247 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
7249 verbose(env, "bpf verifier is misconfigured\n");
7256 static bool may_access_skb(enum bpf_prog_type type)
7259 case BPF_PROG_TYPE_SOCKET_FILTER:
7260 case BPF_PROG_TYPE_SCHED_CLS:
7261 case BPF_PROG_TYPE_SCHED_ACT:
7268 /* verify safety of LD_ABS|LD_IND instructions:
7269 * - they can only appear in the programs where ctx == skb
7270 * - since they are wrappers of function calls, they scratch R1-R5 registers,
7271 * preserve R6-R9, and store return value into R0
7274 * ctx == skb == R6 == CTX
7277 * SRC == any register
7278 * IMM == 32-bit immediate
7281 * R0 - 8/16/32-bit skb data converted to cpu endianness
7283 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
7285 struct bpf_reg_state *regs = cur_regs(env);
7286 static const int ctx_reg = BPF_REG_6;
7287 u8 mode = BPF_MODE(insn->code);
7290 if (!may_access_skb(env->prog->type)) {
7291 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
7295 if (!env->ops->gen_ld_abs) {
7296 verbose(env, "bpf verifier is misconfigured\n");
7300 if (env->subprog_cnt > 1) {
7301 /* when program has LD_ABS insn JITs and interpreter assume
7302 * that r1 == ctx == skb which is not the case for callees
7303 * that can have arbitrary arguments. It's problematic
7304 * for main prog as well since JITs would need to analyze
7305 * all functions in order to make proper register save/restore
7306 * decisions in the main prog. Hence disallow LD_ABS with calls
7308 verbose(env, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
7312 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
7313 BPF_SIZE(insn->code) == BPF_DW ||
7314 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
7315 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
7319 /* check whether implicit source operand (register R6) is readable */
7320 err = check_reg_arg(env, ctx_reg, SRC_OP);
7324 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
7325 * gen_ld_abs() may terminate the program at runtime, leading to
7328 err = check_reference_leak(env);
7330 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
7334 if (env->cur_state->active_spin_lock) {
7335 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
7339 if (regs[ctx_reg].type != PTR_TO_CTX) {
7341 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
7345 if (mode == BPF_IND) {
7346 /* check explicit source operand */
7347 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7352 err = check_ctx_reg(env, ®s[ctx_reg], ctx_reg);
7356 /* reset caller saved regs to unreadable */
7357 for (i = 0; i < CALLER_SAVED_REGS; i++) {
7358 mark_reg_not_init(env, regs, caller_saved[i]);
7359 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
7362 /* mark destination R0 register as readable, since it contains
7363 * the value fetched from the packet.
7364 * Already marked as written above.
7366 mark_reg_unknown(env, regs, BPF_REG_0);
7367 /* ld_abs load up to 32-bit skb data. */
7368 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
7372 static int check_return_code(struct bpf_verifier_env *env)
7374 struct tnum enforce_attach_type_range = tnum_unknown;
7375 const struct bpf_prog *prog = env->prog;
7376 struct bpf_reg_state *reg;
7377 struct tnum range = tnum_range(0, 1);
7380 /* LSM and struct_ops func-ptr's return type could be "void" */
7381 if ((env->prog->type == BPF_PROG_TYPE_STRUCT_OPS ||
7382 env->prog->type == BPF_PROG_TYPE_LSM) &&
7383 !prog->aux->attach_func_proto->type)
7386 /* eBPF calling convetion is such that R0 is used
7387 * to return the value from eBPF program.
7388 * Make sure that it's readable at this time
7389 * of bpf_exit, which means that program wrote
7390 * something into it earlier
7392 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
7396 if (is_pointer_value(env, BPF_REG_0)) {
7397 verbose(env, "R0 leaks addr as return value\n");
7401 switch (env->prog->type) {
7402 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
7403 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
7404 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
7405 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
7406 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
7407 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
7408 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
7409 range = tnum_range(1, 1);
7411 case BPF_PROG_TYPE_CGROUP_SKB:
7412 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
7413 range = tnum_range(0, 3);
7414 enforce_attach_type_range = tnum_range(2, 3);
7417 case BPF_PROG_TYPE_CGROUP_SOCK:
7418 case BPF_PROG_TYPE_SOCK_OPS:
7419 case BPF_PROG_TYPE_CGROUP_DEVICE:
7420 case BPF_PROG_TYPE_CGROUP_SYSCTL:
7421 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
7423 case BPF_PROG_TYPE_RAW_TRACEPOINT:
7424 if (!env->prog->aux->attach_btf_id)
7426 range = tnum_const(0);
7428 case BPF_PROG_TYPE_TRACING:
7429 switch (env->prog->expected_attach_type) {
7430 case BPF_TRACE_FENTRY:
7431 case BPF_TRACE_FEXIT:
7432 range = tnum_const(0);
7434 case BPF_TRACE_RAW_TP:
7435 case BPF_MODIFY_RETURN:
7437 case BPF_TRACE_ITER:
7443 case BPF_PROG_TYPE_SK_LOOKUP:
7444 range = tnum_range(SK_DROP, SK_PASS);
7446 case BPF_PROG_TYPE_EXT:
7447 /* freplace program can return anything as its return value
7448 * depends on the to-be-replaced kernel func or bpf program.
7454 reg = cur_regs(env) + BPF_REG_0;
7455 if (reg->type != SCALAR_VALUE) {
7456 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
7457 reg_type_str[reg->type]);
7461 if (!tnum_in(range, reg->var_off)) {
7464 verbose(env, "At program exit the register R0 ");
7465 if (!tnum_is_unknown(reg->var_off)) {
7466 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
7467 verbose(env, "has value %s", tn_buf);
7469 verbose(env, "has unknown scalar value");
7471 tnum_strn(tn_buf, sizeof(tn_buf), range);
7472 verbose(env, " should have been in %s\n", tn_buf);
7476 if (!tnum_is_unknown(enforce_attach_type_range) &&
7477 tnum_in(enforce_attach_type_range, reg->var_off))
7478 env->prog->enforce_expected_attach_type = 1;
7482 /* non-recursive DFS pseudo code
7483 * 1 procedure DFS-iterative(G,v):
7484 * 2 label v as discovered
7485 * 3 let S be a stack
7487 * 5 while S is not empty
7489 * 7 if t is what we're looking for:
7491 * 9 for all edges e in G.adjacentEdges(t) do
7492 * 10 if edge e is already labelled
7493 * 11 continue with the next edge
7494 * 12 w <- G.adjacentVertex(t,e)
7495 * 13 if vertex w is not discovered and not explored
7496 * 14 label e as tree-edge
7497 * 15 label w as discovered
7500 * 18 else if vertex w is discovered
7501 * 19 label e as back-edge
7503 * 21 // vertex w is explored
7504 * 22 label e as forward- or cross-edge
7505 * 23 label t as explored
7510 * 0x11 - discovered and fall-through edge labelled
7511 * 0x12 - discovered and fall-through and branch edges labelled
7522 static u32 state_htab_size(struct bpf_verifier_env *env)
7524 return env->prog->len;
7527 static struct bpf_verifier_state_list **explored_state(
7528 struct bpf_verifier_env *env,
7531 struct bpf_verifier_state *cur = env->cur_state;
7532 struct bpf_func_state *state = cur->frame[cur->curframe];
7534 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
7537 static void init_explored_state(struct bpf_verifier_env *env, int idx)
7539 env->insn_aux_data[idx].prune_point = true;
7542 /* t, w, e - match pseudo-code above:
7543 * t - index of current instruction
7544 * w - next instruction
7547 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
7550 int *insn_stack = env->cfg.insn_stack;
7551 int *insn_state = env->cfg.insn_state;
7553 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
7556 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
7559 if (w < 0 || w >= env->prog->len) {
7560 verbose_linfo(env, t, "%d: ", t);
7561 verbose(env, "jump out of range from insn %d to %d\n", t, w);
7566 /* mark branch target for state pruning */
7567 init_explored_state(env, w);
7569 if (insn_state[w] == 0) {
7571 insn_state[t] = DISCOVERED | e;
7572 insn_state[w] = DISCOVERED;
7573 if (env->cfg.cur_stack >= env->prog->len)
7575 insn_stack[env->cfg.cur_stack++] = w;
7577 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
7578 if (loop_ok && env->bpf_capable)
7580 verbose_linfo(env, t, "%d: ", t);
7581 verbose_linfo(env, w, "%d: ", w);
7582 verbose(env, "back-edge from insn %d to %d\n", t, w);
7584 } else if (insn_state[w] == EXPLORED) {
7585 /* forward- or cross-edge */
7586 insn_state[t] = DISCOVERED | e;
7588 verbose(env, "insn state internal bug\n");
7594 /* non-recursive depth-first-search to detect loops in BPF program
7595 * loop == back-edge in directed graph
7597 static int check_cfg(struct bpf_verifier_env *env)
7599 struct bpf_insn *insns = env->prog->insnsi;
7600 int insn_cnt = env->prog->len;
7601 int *insn_stack, *insn_state;
7605 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
7609 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
7615 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
7616 insn_stack[0] = 0; /* 0 is the first instruction */
7617 env->cfg.cur_stack = 1;
7620 if (env->cfg.cur_stack == 0)
7622 t = insn_stack[env->cfg.cur_stack - 1];
7624 if (BPF_CLASS(insns[t].code) == BPF_JMP ||
7625 BPF_CLASS(insns[t].code) == BPF_JMP32) {
7626 u8 opcode = BPF_OP(insns[t].code);
7628 if (opcode == BPF_EXIT) {
7630 } else if (opcode == BPF_CALL) {
7631 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
7636 if (t + 1 < insn_cnt)
7637 init_explored_state(env, t + 1);
7638 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
7639 init_explored_state(env, t);
7640 ret = push_insn(t, t + insns[t].imm + 1, BRANCH,
7647 } else if (opcode == BPF_JA) {
7648 if (BPF_SRC(insns[t].code) != BPF_K) {
7652 /* unconditional jump with single edge */
7653 ret = push_insn(t, t + insns[t].off + 1,
7654 FALLTHROUGH, env, true);
7659 /* unconditional jmp is not a good pruning point,
7660 * but it's marked, since backtracking needs
7661 * to record jmp history in is_state_visited().
7663 init_explored_state(env, t + insns[t].off + 1);
7664 /* tell verifier to check for equivalent states
7665 * after every call and jump
7667 if (t + 1 < insn_cnt)
7668 init_explored_state(env, t + 1);
7670 /* conditional jump with two edges */
7671 init_explored_state(env, t);
7672 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
7678 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
7685 /* all other non-branch instructions with single
7688 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
7696 insn_state[t] = EXPLORED;
7697 if (env->cfg.cur_stack-- <= 0) {
7698 verbose(env, "pop stack internal bug\n");
7705 for (i = 0; i < insn_cnt; i++) {
7706 if (insn_state[i] != EXPLORED) {
7707 verbose(env, "unreachable insn %d\n", i);
7712 ret = 0; /* cfg looks good */
7717 env->cfg.insn_state = env->cfg.insn_stack = NULL;
7721 /* The minimum supported BTF func info size */
7722 #define MIN_BPF_FUNCINFO_SIZE 8
7723 #define MAX_FUNCINFO_REC_SIZE 252
7725 static int check_btf_func(struct bpf_verifier_env *env,
7726 const union bpf_attr *attr,
7727 union bpf_attr __user *uattr)
7729 u32 i, nfuncs, urec_size, min_size;
7730 u32 krec_size = sizeof(struct bpf_func_info);
7731 struct bpf_func_info *krecord;
7732 struct bpf_func_info_aux *info_aux = NULL;
7733 const struct btf_type *type;
7734 struct bpf_prog *prog;
7735 const struct btf *btf;
7736 void __user *urecord;
7737 u32 prev_offset = 0;
7740 nfuncs = attr->func_info_cnt;
7744 if (nfuncs != env->subprog_cnt) {
7745 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
7749 urec_size = attr->func_info_rec_size;
7750 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
7751 urec_size > MAX_FUNCINFO_REC_SIZE ||
7752 urec_size % sizeof(u32)) {
7753 verbose(env, "invalid func info rec size %u\n", urec_size);
7758 btf = prog->aux->btf;
7760 urecord = u64_to_user_ptr(attr->func_info);
7761 min_size = min_t(u32, krec_size, urec_size);
7763 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
7766 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
7770 for (i = 0; i < nfuncs; i++) {
7771 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
7773 if (ret == -E2BIG) {
7774 verbose(env, "nonzero tailing record in func info");
7775 /* set the size kernel expects so loader can zero
7776 * out the rest of the record.
7778 if (put_user(min_size, &uattr->func_info_rec_size))
7784 if (copy_from_user(&krecord[i], urecord, min_size)) {
7789 /* check insn_off */
7791 if (krecord[i].insn_off) {
7793 "nonzero insn_off %u for the first func info record",
7794 krecord[i].insn_off);
7798 } else if (krecord[i].insn_off <= prev_offset) {
7800 "same or smaller insn offset (%u) than previous func info record (%u)",
7801 krecord[i].insn_off, prev_offset);
7806 if (env->subprog_info[i].start != krecord[i].insn_off) {
7807 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
7813 type = btf_type_by_id(btf, krecord[i].type_id);
7814 if (!type || !btf_type_is_func(type)) {
7815 verbose(env, "invalid type id %d in func info",
7816 krecord[i].type_id);
7820 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
7821 prev_offset = krecord[i].insn_off;
7822 urecord += urec_size;
7825 prog->aux->func_info = krecord;
7826 prog->aux->func_info_cnt = nfuncs;
7827 prog->aux->func_info_aux = info_aux;
7836 static void adjust_btf_func(struct bpf_verifier_env *env)
7838 struct bpf_prog_aux *aux = env->prog->aux;
7841 if (!aux->func_info)
7844 for (i = 0; i < env->subprog_cnt; i++)
7845 aux->func_info[i].insn_off = env->subprog_info[i].start;
7848 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
7849 sizeof(((struct bpf_line_info *)(0))->line_col))
7850 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
7852 static int check_btf_line(struct bpf_verifier_env *env,
7853 const union bpf_attr *attr,
7854 union bpf_attr __user *uattr)
7856 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
7857 struct bpf_subprog_info *sub;
7858 struct bpf_line_info *linfo;
7859 struct bpf_prog *prog;
7860 const struct btf *btf;
7861 void __user *ulinfo;
7864 nr_linfo = attr->line_info_cnt;
7868 rec_size = attr->line_info_rec_size;
7869 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
7870 rec_size > MAX_LINEINFO_REC_SIZE ||
7871 rec_size & (sizeof(u32) - 1))
7874 /* Need to zero it in case the userspace may
7875 * pass in a smaller bpf_line_info object.
7877 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
7878 GFP_KERNEL | __GFP_NOWARN);
7883 btf = prog->aux->btf;
7886 sub = env->subprog_info;
7887 ulinfo = u64_to_user_ptr(attr->line_info);
7888 expected_size = sizeof(struct bpf_line_info);
7889 ncopy = min_t(u32, expected_size, rec_size);
7890 for (i = 0; i < nr_linfo; i++) {
7891 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
7893 if (err == -E2BIG) {
7894 verbose(env, "nonzero tailing record in line_info");
7895 if (put_user(expected_size,
7896 &uattr->line_info_rec_size))
7902 if (copy_from_user(&linfo[i], ulinfo, ncopy)) {
7908 * Check insn_off to ensure
7909 * 1) strictly increasing AND
7910 * 2) bounded by prog->len
7912 * The linfo[0].insn_off == 0 check logically falls into
7913 * the later "missing bpf_line_info for func..." case
7914 * because the first linfo[0].insn_off must be the
7915 * first sub also and the first sub must have
7916 * subprog_info[0].start == 0.
7918 if ((i && linfo[i].insn_off <= prev_offset) ||
7919 linfo[i].insn_off >= prog->len) {
7920 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
7921 i, linfo[i].insn_off, prev_offset,
7927 if (!prog->insnsi[linfo[i].insn_off].code) {
7929 "Invalid insn code at line_info[%u].insn_off\n",
7935 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
7936 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
7937 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
7942 if (s != env->subprog_cnt) {
7943 if (linfo[i].insn_off == sub[s].start) {
7944 sub[s].linfo_idx = i;
7946 } else if (sub[s].start < linfo[i].insn_off) {
7947 verbose(env, "missing bpf_line_info for func#%u\n", s);
7953 prev_offset = linfo[i].insn_off;
7957 if (s != env->subprog_cnt) {
7958 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
7959 env->subprog_cnt - s, s);
7964 prog->aux->linfo = linfo;
7965 prog->aux->nr_linfo = nr_linfo;
7974 static int check_btf_info(struct bpf_verifier_env *env,
7975 const union bpf_attr *attr,
7976 union bpf_attr __user *uattr)
7981 if (!attr->func_info_cnt && !attr->line_info_cnt)
7984 btf = btf_get_by_fd(attr->prog_btf_fd);
7986 return PTR_ERR(btf);
7987 env->prog->aux->btf = btf;
7989 err = check_btf_func(env, attr, uattr);
7993 err = check_btf_line(env, attr, uattr);
8000 /* check %cur's range satisfies %old's */
8001 static bool range_within(struct bpf_reg_state *old,
8002 struct bpf_reg_state *cur)
8004 return old->umin_value <= cur->umin_value &&
8005 old->umax_value >= cur->umax_value &&
8006 old->smin_value <= cur->smin_value &&
8007 old->smax_value >= cur->smax_value;
8010 /* Maximum number of register states that can exist at once */
8011 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
8017 /* If in the old state two registers had the same id, then they need to have
8018 * the same id in the new state as well. But that id could be different from
8019 * the old state, so we need to track the mapping from old to new ids.
8020 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
8021 * regs with old id 5 must also have new id 9 for the new state to be safe. But
8022 * regs with a different old id could still have new id 9, we don't care about
8024 * So we look through our idmap to see if this old id has been seen before. If
8025 * so, we require the new id to match; otherwise, we add the id pair to the map.
8027 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
8031 for (i = 0; i < ID_MAP_SIZE; i++) {
8032 if (!idmap[i].old) {
8033 /* Reached an empty slot; haven't seen this id before */
8034 idmap[i].old = old_id;
8035 idmap[i].cur = cur_id;
8038 if (idmap[i].old == old_id)
8039 return idmap[i].cur == cur_id;
8041 /* We ran out of idmap slots, which should be impossible */
8046 static void clean_func_state(struct bpf_verifier_env *env,
8047 struct bpf_func_state *st)
8049 enum bpf_reg_liveness live;
8052 for (i = 0; i < BPF_REG_FP; i++) {
8053 live = st->regs[i].live;
8054 /* liveness must not touch this register anymore */
8055 st->regs[i].live |= REG_LIVE_DONE;
8056 if (!(live & REG_LIVE_READ))
8057 /* since the register is unused, clear its state
8058 * to make further comparison simpler
8060 __mark_reg_not_init(env, &st->regs[i]);
8063 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
8064 live = st->stack[i].spilled_ptr.live;
8065 /* liveness must not touch this stack slot anymore */
8066 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
8067 if (!(live & REG_LIVE_READ)) {
8068 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
8069 for (j = 0; j < BPF_REG_SIZE; j++)
8070 st->stack[i].slot_type[j] = STACK_INVALID;
8075 static void clean_verifier_state(struct bpf_verifier_env *env,
8076 struct bpf_verifier_state *st)
8080 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
8081 /* all regs in this state in all frames were already marked */
8084 for (i = 0; i <= st->curframe; i++)
8085 clean_func_state(env, st->frame[i]);
8088 /* the parentage chains form a tree.
8089 * the verifier states are added to state lists at given insn and
8090 * pushed into state stack for future exploration.
8091 * when the verifier reaches bpf_exit insn some of the verifer states
8092 * stored in the state lists have their final liveness state already,
8093 * but a lot of states will get revised from liveness point of view when
8094 * the verifier explores other branches.
8097 * 2: if r1 == 100 goto pc+1
8100 * when the verifier reaches exit insn the register r0 in the state list of
8101 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
8102 * of insn 2 and goes exploring further. At the insn 4 it will walk the
8103 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
8105 * Since the verifier pushes the branch states as it sees them while exploring
8106 * the program the condition of walking the branch instruction for the second
8107 * time means that all states below this branch were already explored and
8108 * their final liveness markes are already propagated.
8109 * Hence when the verifier completes the search of state list in is_state_visited()
8110 * we can call this clean_live_states() function to mark all liveness states
8111 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
8113 * This function also clears the registers and stack for states that !READ
8114 * to simplify state merging.
8116 * Important note here that walking the same branch instruction in the callee
8117 * doesn't meant that the states are DONE. The verifier has to compare
8120 static void clean_live_states(struct bpf_verifier_env *env, int insn,
8121 struct bpf_verifier_state *cur)
8123 struct bpf_verifier_state_list *sl;
8126 sl = *explored_state(env, insn);
8128 if (sl->state.branches)
8130 if (sl->state.insn_idx != insn ||
8131 sl->state.curframe != cur->curframe)
8133 for (i = 0; i <= cur->curframe; i++)
8134 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
8136 clean_verifier_state(env, &sl->state);
8142 /* Returns true if (rold safe implies rcur safe) */
8143 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
8144 struct idpair *idmap)
8148 if (!(rold->live & REG_LIVE_READ))
8149 /* explored state didn't use this */
8152 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
8154 if (rold->type == PTR_TO_STACK)
8155 /* two stack pointers are equal only if they're pointing to
8156 * the same stack frame, since fp-8 in foo != fp-8 in bar
8158 return equal && rold->frameno == rcur->frameno;
8163 if (rold->type == NOT_INIT)
8164 /* explored state can't have used this */
8166 if (rcur->type == NOT_INIT)
8168 switch (rold->type) {
8170 if (rcur->type == SCALAR_VALUE) {
8171 if (!rold->precise && !rcur->precise)
8173 /* new val must satisfy old val knowledge */
8174 return range_within(rold, rcur) &&
8175 tnum_in(rold->var_off, rcur->var_off);
8177 /* We're trying to use a pointer in place of a scalar.
8178 * Even if the scalar was unbounded, this could lead to
8179 * pointer leaks because scalars are allowed to leak
8180 * while pointers are not. We could make this safe in
8181 * special cases if root is calling us, but it's
8182 * probably not worth the hassle.
8186 case PTR_TO_MAP_VALUE:
8187 /* If the new min/max/var_off satisfy the old ones and
8188 * everything else matches, we are OK.
8189 * 'id' is not compared, since it's only used for maps with
8190 * bpf_spin_lock inside map element and in such cases if
8191 * the rest of the prog is valid for one map element then
8192 * it's valid for all map elements regardless of the key
8193 * used in bpf_map_lookup()
8195 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
8196 range_within(rold, rcur) &&
8197 tnum_in(rold->var_off, rcur->var_off);
8198 case PTR_TO_MAP_VALUE_OR_NULL:
8199 /* a PTR_TO_MAP_VALUE could be safe to use as a
8200 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
8201 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
8202 * checked, doing so could have affected others with the same
8203 * id, and we can't check for that because we lost the id when
8204 * we converted to a PTR_TO_MAP_VALUE.
8206 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
8208 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
8210 /* Check our ids match any regs they're supposed to */
8211 return check_ids(rold->id, rcur->id, idmap);
8212 case PTR_TO_PACKET_META:
8214 if (rcur->type != rold->type)
8216 /* We must have at least as much range as the old ptr
8217 * did, so that any accesses which were safe before are
8218 * still safe. This is true even if old range < old off,
8219 * since someone could have accessed through (ptr - k), or
8220 * even done ptr -= k in a register, to get a safe access.
8222 if (rold->range > rcur->range)
8224 /* If the offsets don't match, we can't trust our alignment;
8225 * nor can we be sure that we won't fall out of range.
8227 if (rold->off != rcur->off)
8229 /* id relations must be preserved */
8230 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
8232 /* new val must satisfy old val knowledge */
8233 return range_within(rold, rcur) &&
8234 tnum_in(rold->var_off, rcur->var_off);
8236 case CONST_PTR_TO_MAP:
8237 case PTR_TO_PACKET_END:
8238 case PTR_TO_FLOW_KEYS:
8240 case PTR_TO_SOCKET_OR_NULL:
8241 case PTR_TO_SOCK_COMMON:
8242 case PTR_TO_SOCK_COMMON_OR_NULL:
8243 case PTR_TO_TCP_SOCK:
8244 case PTR_TO_TCP_SOCK_OR_NULL:
8245 case PTR_TO_XDP_SOCK:
8246 /* Only valid matches are exact, which memcmp() above
8247 * would have accepted
8250 /* Don't know what's going on, just say it's not safe */
8254 /* Shouldn't get here; if we do, say it's not safe */
8259 static bool stacksafe(struct bpf_func_state *old,
8260 struct bpf_func_state *cur,
8261 struct idpair *idmap)
8265 /* walk slots of the explored stack and ignore any additional
8266 * slots in the current stack, since explored(safe) state
8269 for (i = 0; i < old->allocated_stack; i++) {
8270 spi = i / BPF_REG_SIZE;
8272 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
8273 i += BPF_REG_SIZE - 1;
8274 /* explored state didn't use this */
8278 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
8281 /* explored stack has more populated slots than current stack
8282 * and these slots were used
8284 if (i >= cur->allocated_stack)
8287 /* if old state was safe with misc data in the stack
8288 * it will be safe with zero-initialized stack.
8289 * The opposite is not true
8291 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
8292 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
8294 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
8295 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
8296 /* Ex: old explored (safe) state has STACK_SPILL in
8297 * this stack slot, but current has has STACK_MISC ->
8298 * this verifier states are not equivalent,
8299 * return false to continue verification of this path
8302 if (i % BPF_REG_SIZE)
8304 if (old->stack[spi].slot_type[0] != STACK_SPILL)
8306 if (!regsafe(&old->stack[spi].spilled_ptr,
8307 &cur->stack[spi].spilled_ptr,
8309 /* when explored and current stack slot are both storing
8310 * spilled registers, check that stored pointers types
8311 * are the same as well.
8312 * Ex: explored safe path could have stored
8313 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
8314 * but current path has stored:
8315 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
8316 * such verifier states are not equivalent.
8317 * return false to continue verification of this path
8324 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
8326 if (old->acquired_refs != cur->acquired_refs)
8328 return !memcmp(old->refs, cur->refs,
8329 sizeof(*old->refs) * old->acquired_refs);
8332 /* compare two verifier states
8334 * all states stored in state_list are known to be valid, since
8335 * verifier reached 'bpf_exit' instruction through them
8337 * this function is called when verifier exploring different branches of
8338 * execution popped from the state stack. If it sees an old state that has
8339 * more strict register state and more strict stack state then this execution
8340 * branch doesn't need to be explored further, since verifier already
8341 * concluded that more strict state leads to valid finish.
8343 * Therefore two states are equivalent if register state is more conservative
8344 * and explored stack state is more conservative than the current one.
8347 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
8348 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
8350 * In other words if current stack state (one being explored) has more
8351 * valid slots than old one that already passed validation, it means
8352 * the verifier can stop exploring and conclude that current state is valid too
8354 * Similarly with registers. If explored state has register type as invalid
8355 * whereas register type in current state is meaningful, it means that
8356 * the current state will reach 'bpf_exit' instruction safely
8358 static bool func_states_equal(struct bpf_func_state *old,
8359 struct bpf_func_state *cur)
8361 struct idpair *idmap;
8365 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
8366 /* If we failed to allocate the idmap, just say it's not safe */
8370 for (i = 0; i < MAX_BPF_REG; i++) {
8371 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
8375 if (!stacksafe(old, cur, idmap))
8378 if (!refsafe(old, cur))
8386 static bool states_equal(struct bpf_verifier_env *env,
8387 struct bpf_verifier_state *old,
8388 struct bpf_verifier_state *cur)
8392 if (old->curframe != cur->curframe)
8395 /* Verification state from speculative execution simulation
8396 * must never prune a non-speculative execution one.
8398 if (old->speculative && !cur->speculative)
8401 if (old->active_spin_lock != cur->active_spin_lock)
8404 /* for states to be equal callsites have to be the same
8405 * and all frame states need to be equivalent
8407 for (i = 0; i <= old->curframe; i++) {
8408 if (old->frame[i]->callsite != cur->frame[i]->callsite)
8410 if (!func_states_equal(old->frame[i], cur->frame[i]))
8416 /* Return 0 if no propagation happened. Return negative error code if error
8417 * happened. Otherwise, return the propagated bit.
8419 static int propagate_liveness_reg(struct bpf_verifier_env *env,
8420 struct bpf_reg_state *reg,
8421 struct bpf_reg_state *parent_reg)
8423 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
8424 u8 flag = reg->live & REG_LIVE_READ;
8427 /* When comes here, read flags of PARENT_REG or REG could be any of
8428 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
8429 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
8431 if (parent_flag == REG_LIVE_READ64 ||
8432 /* Or if there is no read flag from REG. */
8434 /* Or if the read flag from REG is the same as PARENT_REG. */
8435 parent_flag == flag)
8438 err = mark_reg_read(env, reg, parent_reg, flag);
8445 /* A write screens off any subsequent reads; but write marks come from the
8446 * straight-line code between a state and its parent. When we arrive at an
8447 * equivalent state (jump target or such) we didn't arrive by the straight-line
8448 * code, so read marks in the state must propagate to the parent regardless
8449 * of the state's write marks. That's what 'parent == state->parent' comparison
8450 * in mark_reg_read() is for.
8452 static int propagate_liveness(struct bpf_verifier_env *env,
8453 const struct bpf_verifier_state *vstate,
8454 struct bpf_verifier_state *vparent)
8456 struct bpf_reg_state *state_reg, *parent_reg;
8457 struct bpf_func_state *state, *parent;
8458 int i, frame, err = 0;
8460 if (vparent->curframe != vstate->curframe) {
8461 WARN(1, "propagate_live: parent frame %d current frame %d\n",
8462 vparent->curframe, vstate->curframe);
8465 /* Propagate read liveness of registers... */
8466 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
8467 for (frame = 0; frame <= vstate->curframe; frame++) {
8468 parent = vparent->frame[frame];
8469 state = vstate->frame[frame];
8470 parent_reg = parent->regs;
8471 state_reg = state->regs;
8472 /* We don't need to worry about FP liveness, it's read-only */
8473 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
8474 err = propagate_liveness_reg(env, &state_reg[i],
8478 if (err == REG_LIVE_READ64)
8479 mark_insn_zext(env, &parent_reg[i]);
8482 /* Propagate stack slots. */
8483 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
8484 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
8485 parent_reg = &parent->stack[i].spilled_ptr;
8486 state_reg = &state->stack[i].spilled_ptr;
8487 err = propagate_liveness_reg(env, state_reg,
8496 /* find precise scalars in the previous equivalent state and
8497 * propagate them into the current state
8499 static int propagate_precision(struct bpf_verifier_env *env,
8500 const struct bpf_verifier_state *old)
8502 struct bpf_reg_state *state_reg;
8503 struct bpf_func_state *state;
8506 state = old->frame[old->curframe];
8507 state_reg = state->regs;
8508 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
8509 if (state_reg->type != SCALAR_VALUE ||
8510 !state_reg->precise)
8512 if (env->log.level & BPF_LOG_LEVEL2)
8513 verbose(env, "propagating r%d\n", i);
8514 err = mark_chain_precision(env, i);
8519 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
8520 if (state->stack[i].slot_type[0] != STACK_SPILL)
8522 state_reg = &state->stack[i].spilled_ptr;
8523 if (state_reg->type != SCALAR_VALUE ||
8524 !state_reg->precise)
8526 if (env->log.level & BPF_LOG_LEVEL2)
8527 verbose(env, "propagating fp%d\n",
8528 (-i - 1) * BPF_REG_SIZE);
8529 err = mark_chain_precision_stack(env, i);
8536 static bool states_maybe_looping(struct bpf_verifier_state *old,
8537 struct bpf_verifier_state *cur)
8539 struct bpf_func_state *fold, *fcur;
8540 int i, fr = cur->curframe;
8542 if (old->curframe != fr)
8545 fold = old->frame[fr];
8546 fcur = cur->frame[fr];
8547 for (i = 0; i < MAX_BPF_REG; i++)
8548 if (memcmp(&fold->regs[i], &fcur->regs[i],
8549 offsetof(struct bpf_reg_state, parent)))
8555 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
8557 struct bpf_verifier_state_list *new_sl;
8558 struct bpf_verifier_state_list *sl, **pprev;
8559 struct bpf_verifier_state *cur = env->cur_state, *new;
8560 int i, j, err, states_cnt = 0;
8561 bool add_new_state = env->test_state_freq ? true : false;
8563 cur->last_insn_idx = env->prev_insn_idx;
8564 if (!env->insn_aux_data[insn_idx].prune_point)
8565 /* this 'insn_idx' instruction wasn't marked, so we will not
8566 * be doing state search here
8570 /* bpf progs typically have pruning point every 4 instructions
8571 * http://vger.kernel.org/bpfconf2019.html#session-1
8572 * Do not add new state for future pruning if the verifier hasn't seen
8573 * at least 2 jumps and at least 8 instructions.
8574 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
8575 * In tests that amounts to up to 50% reduction into total verifier
8576 * memory consumption and 20% verifier time speedup.
8578 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
8579 env->insn_processed - env->prev_insn_processed >= 8)
8580 add_new_state = true;
8582 pprev = explored_state(env, insn_idx);
8585 clean_live_states(env, insn_idx, cur);
8589 if (sl->state.insn_idx != insn_idx)
8591 if (sl->state.branches) {
8592 if (states_maybe_looping(&sl->state, cur) &&
8593 states_equal(env, &sl->state, cur)) {
8594 verbose_linfo(env, insn_idx, "; ");
8595 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
8598 /* if the verifier is processing a loop, avoid adding new state
8599 * too often, since different loop iterations have distinct
8600 * states and may not help future pruning.
8601 * This threshold shouldn't be too low to make sure that
8602 * a loop with large bound will be rejected quickly.
8603 * The most abusive loop will be:
8605 * if r1 < 1000000 goto pc-2
8606 * 1M insn_procssed limit / 100 == 10k peak states.
8607 * This threshold shouldn't be too high either, since states
8608 * at the end of the loop are likely to be useful in pruning.
8610 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
8611 env->insn_processed - env->prev_insn_processed < 100)
8612 add_new_state = false;
8615 if (states_equal(env, &sl->state, cur)) {
8617 /* reached equivalent register/stack state,
8619 * Registers read by the continuation are read by us.
8620 * If we have any write marks in env->cur_state, they
8621 * will prevent corresponding reads in the continuation
8622 * from reaching our parent (an explored_state). Our
8623 * own state will get the read marks recorded, but
8624 * they'll be immediately forgotten as we're pruning
8625 * this state and will pop a new one.
8627 err = propagate_liveness(env, &sl->state, cur);
8629 /* if previous state reached the exit with precision and
8630 * current state is equivalent to it (except precsion marks)
8631 * the precision needs to be propagated back in
8632 * the current state.
8634 err = err ? : push_jmp_history(env, cur);
8635 err = err ? : propagate_precision(env, &sl->state);
8641 /* when new state is not going to be added do not increase miss count.
8642 * Otherwise several loop iterations will remove the state
8643 * recorded earlier. The goal of these heuristics is to have
8644 * states from some iterations of the loop (some in the beginning
8645 * and some at the end) to help pruning.
8649 /* heuristic to determine whether this state is beneficial
8650 * to keep checking from state equivalence point of view.
8651 * Higher numbers increase max_states_per_insn and verification time,
8652 * but do not meaningfully decrease insn_processed.
8654 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
8655 /* the state is unlikely to be useful. Remove it to
8656 * speed up verification
8659 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
8660 u32 br = sl->state.branches;
8663 "BUG live_done but branches_to_explore %d\n",
8665 free_verifier_state(&sl->state, false);
8669 /* cannot free this state, since parentage chain may
8670 * walk it later. Add it for free_list instead to
8671 * be freed at the end of verification
8673 sl->next = env->free_list;
8674 env->free_list = sl;
8684 if (env->max_states_per_insn < states_cnt)
8685 env->max_states_per_insn = states_cnt;
8687 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
8688 return push_jmp_history(env, cur);
8691 return push_jmp_history(env, cur);
8693 /* There were no equivalent states, remember the current one.
8694 * Technically the current state is not proven to be safe yet,
8695 * but it will either reach outer most bpf_exit (which means it's safe)
8696 * or it will be rejected. When there are no loops the verifier won't be
8697 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
8698 * again on the way to bpf_exit.
8699 * When looping the sl->state.branches will be > 0 and this state
8700 * will not be considered for equivalence until branches == 0.
8702 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
8705 env->total_states++;
8707 env->prev_jmps_processed = env->jmps_processed;
8708 env->prev_insn_processed = env->insn_processed;
8710 /* add new state to the head of linked list */
8711 new = &new_sl->state;
8712 err = copy_verifier_state(new, cur);
8714 free_verifier_state(new, false);
8718 new->insn_idx = insn_idx;
8719 WARN_ONCE(new->branches != 1,
8720 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
8723 cur->first_insn_idx = insn_idx;
8724 clear_jmp_history(cur);
8725 new_sl->next = *explored_state(env, insn_idx);
8726 *explored_state(env, insn_idx) = new_sl;
8727 /* connect new state to parentage chain. Current frame needs all
8728 * registers connected. Only r6 - r9 of the callers are alive (pushed
8729 * to the stack implicitly by JITs) so in callers' frames connect just
8730 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
8731 * the state of the call instruction (with WRITTEN set), and r0 comes
8732 * from callee with its full parentage chain, anyway.
8734 /* clear write marks in current state: the writes we did are not writes
8735 * our child did, so they don't screen off its reads from us.
8736 * (There are no read marks in current state, because reads always mark
8737 * their parent and current state never has children yet. Only
8738 * explored_states can get read marks.)
8740 for (j = 0; j <= cur->curframe; j++) {
8741 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
8742 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
8743 for (i = 0; i < BPF_REG_FP; i++)
8744 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
8747 /* all stack frames are accessible from callee, clear them all */
8748 for (j = 0; j <= cur->curframe; j++) {
8749 struct bpf_func_state *frame = cur->frame[j];
8750 struct bpf_func_state *newframe = new->frame[j];
8752 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
8753 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
8754 frame->stack[i].spilled_ptr.parent =
8755 &newframe->stack[i].spilled_ptr;
8761 /* Return true if it's OK to have the same insn return a different type. */
8762 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
8767 case PTR_TO_SOCKET_OR_NULL:
8768 case PTR_TO_SOCK_COMMON:
8769 case PTR_TO_SOCK_COMMON_OR_NULL:
8770 case PTR_TO_TCP_SOCK:
8771 case PTR_TO_TCP_SOCK_OR_NULL:
8772 case PTR_TO_XDP_SOCK:
8774 case PTR_TO_BTF_ID_OR_NULL:
8781 /* If an instruction was previously used with particular pointer types, then we
8782 * need to be careful to avoid cases such as the below, where it may be ok
8783 * for one branch accessing the pointer, but not ok for the other branch:
8788 * R1 = some_other_valid_ptr;
8791 * R2 = *(u32 *)(R1 + 0);
8793 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
8795 return src != prev && (!reg_type_mismatch_ok(src) ||
8796 !reg_type_mismatch_ok(prev));
8799 static int do_check(struct bpf_verifier_env *env)
8801 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
8802 struct bpf_verifier_state *state = env->cur_state;
8803 struct bpf_insn *insns = env->prog->insnsi;
8804 struct bpf_reg_state *regs;
8805 int insn_cnt = env->prog->len;
8806 bool do_print_state = false;
8807 int prev_insn_idx = -1;
8810 struct bpf_insn *insn;
8814 env->prev_insn_idx = prev_insn_idx;
8815 if (env->insn_idx >= insn_cnt) {
8816 verbose(env, "invalid insn idx %d insn_cnt %d\n",
8817 env->insn_idx, insn_cnt);
8821 insn = &insns[env->insn_idx];
8822 class = BPF_CLASS(insn->code);
8824 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
8826 "BPF program is too large. Processed %d insn\n",
8827 env->insn_processed);
8831 err = is_state_visited(env, env->insn_idx);
8835 /* found equivalent state, can prune the search */
8836 if (env->log.level & BPF_LOG_LEVEL) {
8838 verbose(env, "\nfrom %d to %d%s: safe\n",
8839 env->prev_insn_idx, env->insn_idx,
8840 env->cur_state->speculative ?
8841 " (speculative execution)" : "");
8843 verbose(env, "%d: safe\n", env->insn_idx);
8845 goto process_bpf_exit;
8848 if (signal_pending(current))
8854 if (env->log.level & BPF_LOG_LEVEL2 ||
8855 (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
8856 if (env->log.level & BPF_LOG_LEVEL2)
8857 verbose(env, "%d:", env->insn_idx);
8859 verbose(env, "\nfrom %d to %d%s:",
8860 env->prev_insn_idx, env->insn_idx,
8861 env->cur_state->speculative ?
8862 " (speculative execution)" : "");
8863 print_verifier_state(env, state->frame[state->curframe]);
8864 do_print_state = false;
8867 if (env->log.level & BPF_LOG_LEVEL) {
8868 const struct bpf_insn_cbs cbs = {
8869 .cb_print = verbose,
8870 .private_data = env,
8873 verbose_linfo(env, env->insn_idx, "; ");
8874 verbose(env, "%d: ", env->insn_idx);
8875 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
8878 if (bpf_prog_is_dev_bound(env->prog->aux)) {
8879 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
8880 env->prev_insn_idx);
8885 regs = cur_regs(env);
8886 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
8887 prev_insn_idx = env->insn_idx;
8889 if (class == BPF_ALU || class == BPF_ALU64) {
8890 err = check_alu_op(env, insn);
8894 } else if (class == BPF_LDX) {
8895 enum bpf_reg_type *prev_src_type, src_reg_type;
8897 /* check for reserved fields is already done */
8899 /* check src operand */
8900 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8904 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
8908 src_reg_type = regs[insn->src_reg].type;
8910 /* check that memory (src_reg + off) is readable,
8911 * the state of dst_reg will be updated by this func
8913 err = check_mem_access(env, env->insn_idx, insn->src_reg,
8914 insn->off, BPF_SIZE(insn->code),
8915 BPF_READ, insn->dst_reg, false);
8919 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
8921 if (*prev_src_type == NOT_INIT) {
8923 * dst_reg = *(u32 *)(src_reg + off)
8924 * save type to validate intersecting paths
8926 *prev_src_type = src_reg_type;
8928 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
8929 /* ABuser program is trying to use the same insn
8930 * dst_reg = *(u32*) (src_reg + off)
8931 * with different pointer types:
8932 * src_reg == ctx in one branch and
8933 * src_reg == stack|map in some other branch.
8936 verbose(env, "same insn cannot be used with different pointers\n");
8940 } else if (class == BPF_STX) {
8941 enum bpf_reg_type *prev_dst_type, dst_reg_type;
8943 if (BPF_MODE(insn->code) == BPF_XADD) {
8944 err = check_xadd(env, env->insn_idx, insn);
8951 /* check src1 operand */
8952 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8955 /* check src2 operand */
8956 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8960 dst_reg_type = regs[insn->dst_reg].type;
8962 /* check that memory (dst_reg + off) is writeable */
8963 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
8964 insn->off, BPF_SIZE(insn->code),
8965 BPF_WRITE, insn->src_reg, false);
8969 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
8971 if (*prev_dst_type == NOT_INIT) {
8972 *prev_dst_type = dst_reg_type;
8973 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
8974 verbose(env, "same insn cannot be used with different pointers\n");
8978 } else if (class == BPF_ST) {
8979 if (BPF_MODE(insn->code) != BPF_MEM ||
8980 insn->src_reg != BPF_REG_0) {
8981 verbose(env, "BPF_ST uses reserved fields\n");
8984 /* check src operand */
8985 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8989 if (is_ctx_reg(env, insn->dst_reg)) {
8990 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
8992 reg_type_str[reg_state(env, insn->dst_reg)->type]);
8996 /* check that memory (dst_reg + off) is writeable */
8997 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
8998 insn->off, BPF_SIZE(insn->code),
8999 BPF_WRITE, -1, false);
9003 } else if (class == BPF_JMP || class == BPF_JMP32) {
9004 u8 opcode = BPF_OP(insn->code);
9006 env->jmps_processed++;
9007 if (opcode == BPF_CALL) {
9008 if (BPF_SRC(insn->code) != BPF_K ||
9010 (insn->src_reg != BPF_REG_0 &&
9011 insn->src_reg != BPF_PSEUDO_CALL) ||
9012 insn->dst_reg != BPF_REG_0 ||
9013 class == BPF_JMP32) {
9014 verbose(env, "BPF_CALL uses reserved fields\n");
9018 if (env->cur_state->active_spin_lock &&
9019 (insn->src_reg == BPF_PSEUDO_CALL ||
9020 insn->imm != BPF_FUNC_spin_unlock)) {
9021 verbose(env, "function calls are not allowed while holding a lock\n");
9024 if (insn->src_reg == BPF_PSEUDO_CALL)
9025 err = check_func_call(env, insn, &env->insn_idx);
9027 err = check_helper_call(env, insn->imm, env->insn_idx);
9031 } else if (opcode == BPF_JA) {
9032 if (BPF_SRC(insn->code) != BPF_K ||
9034 insn->src_reg != BPF_REG_0 ||
9035 insn->dst_reg != BPF_REG_0 ||
9036 class == BPF_JMP32) {
9037 verbose(env, "BPF_JA uses reserved fields\n");
9041 env->insn_idx += insn->off + 1;
9044 } else if (opcode == BPF_EXIT) {
9045 if (BPF_SRC(insn->code) != BPF_K ||
9047 insn->src_reg != BPF_REG_0 ||
9048 insn->dst_reg != BPF_REG_0 ||
9049 class == BPF_JMP32) {
9050 verbose(env, "BPF_EXIT uses reserved fields\n");
9054 if (env->cur_state->active_spin_lock) {
9055 verbose(env, "bpf_spin_unlock is missing\n");
9059 if (state->curframe) {
9060 /* exit from nested function */
9061 err = prepare_func_exit(env, &env->insn_idx);
9064 do_print_state = true;
9068 err = check_reference_leak(env);
9072 err = check_return_code(env);
9076 update_branch_counts(env, env->cur_state);
9077 err = pop_stack(env, &prev_insn_idx,
9078 &env->insn_idx, pop_log);
9084 do_print_state = true;
9088 err = check_cond_jmp_op(env, insn, &env->insn_idx);
9092 } else if (class == BPF_LD) {
9093 u8 mode = BPF_MODE(insn->code);
9095 if (mode == BPF_ABS || mode == BPF_IND) {
9096 err = check_ld_abs(env, insn);
9100 } else if (mode == BPF_IMM) {
9101 err = check_ld_imm(env, insn);
9106 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
9108 verbose(env, "invalid BPF_LD mode\n");
9112 verbose(env, "unknown insn class %d\n", class);
9122 static int check_map_prealloc(struct bpf_map *map)
9124 return (map->map_type != BPF_MAP_TYPE_HASH &&
9125 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
9126 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
9127 !(map->map_flags & BPF_F_NO_PREALLOC);
9130 static bool is_tracing_prog_type(enum bpf_prog_type type)
9133 case BPF_PROG_TYPE_KPROBE:
9134 case BPF_PROG_TYPE_TRACEPOINT:
9135 case BPF_PROG_TYPE_PERF_EVENT:
9136 case BPF_PROG_TYPE_RAW_TRACEPOINT:
9143 static bool is_preallocated_map(struct bpf_map *map)
9145 if (!check_map_prealloc(map))
9147 if (map->inner_map_meta && !check_map_prealloc(map->inner_map_meta))
9152 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
9153 struct bpf_map *map,
9154 struct bpf_prog *prog)
9158 * Validate that trace type programs use preallocated hash maps.
9160 * For programs attached to PERF events this is mandatory as the
9161 * perf NMI can hit any arbitrary code sequence.
9163 * All other trace types using preallocated hash maps are unsafe as
9164 * well because tracepoint or kprobes can be inside locked regions
9165 * of the memory allocator or at a place where a recursion into the
9166 * memory allocator would see inconsistent state.
9168 * On RT enabled kernels run-time allocation of all trace type
9169 * programs is strictly prohibited due to lock type constraints. On
9170 * !RT kernels it is allowed for backwards compatibility reasons for
9171 * now, but warnings are emitted so developers are made aware of
9172 * the unsafety and can fix their programs before this is enforced.
9174 if (is_tracing_prog_type(prog->type) && !is_preallocated_map(map)) {
9175 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
9176 verbose(env, "perf_event programs can only use preallocated hash map\n");
9179 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
9180 verbose(env, "trace type programs can only use preallocated hash map\n");
9183 WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
9184 verbose(env, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
9187 if ((is_tracing_prog_type(prog->type) ||
9188 prog->type == BPF_PROG_TYPE_SOCKET_FILTER) &&
9189 map_value_has_spin_lock(map)) {
9190 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
9194 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
9195 !bpf_offload_prog_map_match(prog, map)) {
9196 verbose(env, "offload device mismatch between prog and map\n");
9200 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
9201 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
9208 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
9210 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
9211 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
9214 /* look for pseudo eBPF instructions that access map FDs and
9215 * replace them with actual map pointers
9217 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
9219 struct bpf_insn *insn = env->prog->insnsi;
9220 int insn_cnt = env->prog->len;
9223 err = bpf_prog_calc_tag(env->prog);
9227 for (i = 0; i < insn_cnt; i++, insn++) {
9228 if (BPF_CLASS(insn->code) == BPF_LDX &&
9229 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
9230 verbose(env, "BPF_LDX uses reserved fields\n");
9234 if (BPF_CLASS(insn->code) == BPF_STX &&
9235 ((BPF_MODE(insn->code) != BPF_MEM &&
9236 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
9237 verbose(env, "BPF_STX uses reserved fields\n");
9241 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
9242 struct bpf_insn_aux_data *aux;
9243 struct bpf_map *map;
9247 if (i == insn_cnt - 1 || insn[1].code != 0 ||
9248 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
9250 verbose(env, "invalid bpf_ld_imm64 insn\n");
9254 if (insn[0].src_reg == 0)
9255 /* valid generic load 64-bit imm */
9258 /* In final convert_pseudo_ld_imm64() step, this is
9259 * converted into regular 64-bit imm load insn.
9261 if ((insn[0].src_reg != BPF_PSEUDO_MAP_FD &&
9262 insn[0].src_reg != BPF_PSEUDO_MAP_VALUE) ||
9263 (insn[0].src_reg == BPF_PSEUDO_MAP_FD &&
9264 insn[1].imm != 0)) {
9266 "unrecognized bpf_ld_imm64 insn\n");
9270 f = fdget(insn[0].imm);
9271 map = __bpf_map_get(f);
9273 verbose(env, "fd %d is not pointing to valid bpf_map\n",
9275 return PTR_ERR(map);
9278 err = check_map_prog_compatibility(env, map, env->prog);
9284 aux = &env->insn_aux_data[i];
9285 if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
9286 addr = (unsigned long)map;
9288 u32 off = insn[1].imm;
9290 if (off >= BPF_MAX_VAR_OFF) {
9291 verbose(env, "direct value offset of %u is not allowed\n", off);
9296 if (!map->ops->map_direct_value_addr) {
9297 verbose(env, "no direct value access support for this map type\n");
9302 err = map->ops->map_direct_value_addr(map, &addr, off);
9304 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
9305 map->value_size, off);
9314 insn[0].imm = (u32)addr;
9315 insn[1].imm = addr >> 32;
9317 /* check whether we recorded this map already */
9318 for (j = 0; j < env->used_map_cnt; j++) {
9319 if (env->used_maps[j] == map) {
9326 if (env->used_map_cnt >= MAX_USED_MAPS) {
9331 /* hold the map. If the program is rejected by verifier,
9332 * the map will be released by release_maps() or it
9333 * will be used by the valid program until it's unloaded
9334 * and all maps are released in free_used_maps()
9338 aux->map_index = env->used_map_cnt;
9339 env->used_maps[env->used_map_cnt++] = map;
9341 if (bpf_map_is_cgroup_storage(map) &&
9342 bpf_cgroup_storage_assign(env->prog->aux, map)) {
9343 verbose(env, "only one cgroup storage of each type is allowed\n");
9355 /* Basic sanity check before we invest more work here. */
9356 if (!bpf_opcode_in_insntable(insn->code)) {
9357 verbose(env, "unknown opcode %02x\n", insn->code);
9362 /* now all pseudo BPF_LD_IMM64 instructions load valid
9363 * 'struct bpf_map *' into a register instead of user map_fd.
9364 * These pointers will be used later by verifier to validate map access.
9369 /* drop refcnt of maps used by the rejected program */
9370 static void release_maps(struct bpf_verifier_env *env)
9372 __bpf_free_used_maps(env->prog->aux, env->used_maps,
9376 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
9377 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
9379 struct bpf_insn *insn = env->prog->insnsi;
9380 int insn_cnt = env->prog->len;
9383 for (i = 0; i < insn_cnt; i++, insn++)
9384 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
9388 /* single env->prog->insni[off] instruction was replaced with the range
9389 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
9390 * [0, off) and [off, end) to new locations, so the patched range stays zero
9392 static int adjust_insn_aux_data(struct bpf_verifier_env *env,
9393 struct bpf_prog *new_prog, u32 off, u32 cnt)
9395 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
9396 struct bpf_insn *insn = new_prog->insnsi;
9400 /* aux info at OFF always needs adjustment, no matter fast path
9401 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
9402 * original insn at old prog.
9404 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
9408 prog_len = new_prog->len;
9409 new_data = vzalloc(array_size(prog_len,
9410 sizeof(struct bpf_insn_aux_data)));
9413 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
9414 memcpy(new_data + off + cnt - 1, old_data + off,
9415 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
9416 for (i = off; i < off + cnt - 1; i++) {
9417 new_data[i].seen = env->pass_cnt;
9418 new_data[i].zext_dst = insn_has_def32(env, insn + i);
9420 env->insn_aux_data = new_data;
9425 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
9431 /* NOTE: fake 'exit' subprog should be updated as well. */
9432 for (i = 0; i <= env->subprog_cnt; i++) {
9433 if (env->subprog_info[i].start <= off)
9435 env->subprog_info[i].start += len - 1;
9439 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
9440 const struct bpf_insn *patch, u32 len)
9442 struct bpf_prog *new_prog;
9444 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
9445 if (IS_ERR(new_prog)) {
9446 if (PTR_ERR(new_prog) == -ERANGE)
9448 "insn %d cannot be patched due to 16-bit range\n",
9449 env->insn_aux_data[off].orig_idx);
9452 if (adjust_insn_aux_data(env, new_prog, off, len))
9454 adjust_subprog_starts(env, off, len);
9458 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
9463 /* find first prog starting at or after off (first to remove) */
9464 for (i = 0; i < env->subprog_cnt; i++)
9465 if (env->subprog_info[i].start >= off)
9467 /* find first prog starting at or after off + cnt (first to stay) */
9468 for (j = i; j < env->subprog_cnt; j++)
9469 if (env->subprog_info[j].start >= off + cnt)
9471 /* if j doesn't start exactly at off + cnt, we are just removing
9472 * the front of previous prog
9474 if (env->subprog_info[j].start != off + cnt)
9478 struct bpf_prog_aux *aux = env->prog->aux;
9481 /* move fake 'exit' subprog as well */
9482 move = env->subprog_cnt + 1 - j;
9484 memmove(env->subprog_info + i,
9485 env->subprog_info + j,
9486 sizeof(*env->subprog_info) * move);
9487 env->subprog_cnt -= j - i;
9489 /* remove func_info */
9490 if (aux->func_info) {
9491 move = aux->func_info_cnt - j;
9493 memmove(aux->func_info + i,
9495 sizeof(*aux->func_info) * move);
9496 aux->func_info_cnt -= j - i;
9497 /* func_info->insn_off is set after all code rewrites,
9498 * in adjust_btf_func() - no need to adjust
9502 /* convert i from "first prog to remove" to "first to adjust" */
9503 if (env->subprog_info[i].start == off)
9507 /* update fake 'exit' subprog as well */
9508 for (; i <= env->subprog_cnt; i++)
9509 env->subprog_info[i].start -= cnt;
9514 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
9517 struct bpf_prog *prog = env->prog;
9518 u32 i, l_off, l_cnt, nr_linfo;
9519 struct bpf_line_info *linfo;
9521 nr_linfo = prog->aux->nr_linfo;
9525 linfo = prog->aux->linfo;
9527 /* find first line info to remove, count lines to be removed */
9528 for (i = 0; i < nr_linfo; i++)
9529 if (linfo[i].insn_off >= off)
9534 for (; i < nr_linfo; i++)
9535 if (linfo[i].insn_off < off + cnt)
9540 /* First live insn doesn't match first live linfo, it needs to "inherit"
9541 * last removed linfo. prog is already modified, so prog->len == off
9542 * means no live instructions after (tail of the program was removed).
9544 if (prog->len != off && l_cnt &&
9545 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
9547 linfo[--i].insn_off = off + cnt;
9550 /* remove the line info which refer to the removed instructions */
9552 memmove(linfo + l_off, linfo + i,
9553 sizeof(*linfo) * (nr_linfo - i));
9555 prog->aux->nr_linfo -= l_cnt;
9556 nr_linfo = prog->aux->nr_linfo;
9559 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
9560 for (i = l_off; i < nr_linfo; i++)
9561 linfo[i].insn_off -= cnt;
9563 /* fix up all subprogs (incl. 'exit') which start >= off */
9564 for (i = 0; i <= env->subprog_cnt; i++)
9565 if (env->subprog_info[i].linfo_idx > l_off) {
9566 /* program may have started in the removed region but
9567 * may not be fully removed
9569 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
9570 env->subprog_info[i].linfo_idx -= l_cnt;
9572 env->subprog_info[i].linfo_idx = l_off;
9578 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
9580 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
9581 unsigned int orig_prog_len = env->prog->len;
9584 if (bpf_prog_is_dev_bound(env->prog->aux))
9585 bpf_prog_offload_remove_insns(env, off, cnt);
9587 err = bpf_remove_insns(env->prog, off, cnt);
9591 err = adjust_subprog_starts_after_remove(env, off, cnt);
9595 err = bpf_adj_linfo_after_remove(env, off, cnt);
9599 memmove(aux_data + off, aux_data + off + cnt,
9600 sizeof(*aux_data) * (orig_prog_len - off - cnt));
9605 /* The verifier does more data flow analysis than llvm and will not
9606 * explore branches that are dead at run time. Malicious programs can
9607 * have dead code too. Therefore replace all dead at-run-time code
9610 * Just nops are not optimal, e.g. if they would sit at the end of the
9611 * program and through another bug we would manage to jump there, then
9612 * we'd execute beyond program memory otherwise. Returning exception
9613 * code also wouldn't work since we can have subprogs where the dead
9614 * code could be located.
9616 static void sanitize_dead_code(struct bpf_verifier_env *env)
9618 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
9619 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
9620 struct bpf_insn *insn = env->prog->insnsi;
9621 const int insn_cnt = env->prog->len;
9624 for (i = 0; i < insn_cnt; i++) {
9625 if (aux_data[i].seen)
9627 memcpy(insn + i, &trap, sizeof(trap));
9631 static bool insn_is_cond_jump(u8 code)
9635 if (BPF_CLASS(code) == BPF_JMP32)
9638 if (BPF_CLASS(code) != BPF_JMP)
9642 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
9645 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
9647 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
9648 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
9649 struct bpf_insn *insn = env->prog->insnsi;
9650 const int insn_cnt = env->prog->len;
9653 for (i = 0; i < insn_cnt; i++, insn++) {
9654 if (!insn_is_cond_jump(insn->code))
9657 if (!aux_data[i + 1].seen)
9659 else if (!aux_data[i + 1 + insn->off].seen)
9664 if (bpf_prog_is_dev_bound(env->prog->aux))
9665 bpf_prog_offload_replace_insn(env, i, &ja);
9667 memcpy(insn, &ja, sizeof(ja));
9671 static int opt_remove_dead_code(struct bpf_verifier_env *env)
9673 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
9674 int insn_cnt = env->prog->len;
9677 for (i = 0; i < insn_cnt; i++) {
9681 while (i + j < insn_cnt && !aux_data[i + j].seen)
9686 err = verifier_remove_insns(env, i, j);
9689 insn_cnt = env->prog->len;
9695 static int opt_remove_nops(struct bpf_verifier_env *env)
9697 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
9698 struct bpf_insn *insn = env->prog->insnsi;
9699 int insn_cnt = env->prog->len;
9702 for (i = 0; i < insn_cnt; i++) {
9703 if (memcmp(&insn[i], &ja, sizeof(ja)))
9706 err = verifier_remove_insns(env, i, 1);
9716 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
9717 const union bpf_attr *attr)
9719 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
9720 struct bpf_insn_aux_data *aux = env->insn_aux_data;
9721 int i, patch_len, delta = 0, len = env->prog->len;
9722 struct bpf_insn *insns = env->prog->insnsi;
9723 struct bpf_prog *new_prog;
9726 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
9727 zext_patch[1] = BPF_ZEXT_REG(0);
9728 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
9729 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
9730 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
9731 for (i = 0; i < len; i++) {
9732 int adj_idx = i + delta;
9733 struct bpf_insn insn;
9735 insn = insns[adj_idx];
9736 if (!aux[adj_idx].zext_dst) {
9744 class = BPF_CLASS(code);
9745 if (insn_no_def(&insn))
9748 /* NOTE: arg "reg" (the fourth one) is only used for
9749 * BPF_STX which has been ruled out in above
9750 * check, it is safe to pass NULL here.
9752 if (is_reg64(env, &insn, insn.dst_reg, NULL, DST_OP)) {
9753 if (class == BPF_LD &&
9754 BPF_MODE(code) == BPF_IMM)
9759 /* ctx load could be transformed into wider load. */
9760 if (class == BPF_LDX &&
9761 aux[adj_idx].ptr_type == PTR_TO_CTX)
9764 imm_rnd = get_random_int();
9765 rnd_hi32_patch[0] = insn;
9766 rnd_hi32_patch[1].imm = imm_rnd;
9767 rnd_hi32_patch[3].dst_reg = insn.dst_reg;
9768 patch = rnd_hi32_patch;
9770 goto apply_patch_buffer;
9773 if (!bpf_jit_needs_zext())
9776 zext_patch[0] = insn;
9777 zext_patch[1].dst_reg = insn.dst_reg;
9778 zext_patch[1].src_reg = insn.dst_reg;
9782 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
9785 env->prog = new_prog;
9786 insns = new_prog->insnsi;
9787 aux = env->insn_aux_data;
9788 delta += patch_len - 1;
9794 /* convert load instructions that access fields of a context type into a
9795 * sequence of instructions that access fields of the underlying structure:
9796 * struct __sk_buff -> struct sk_buff
9797 * struct bpf_sock_ops -> struct sock
9799 static int convert_ctx_accesses(struct bpf_verifier_env *env)
9801 const struct bpf_verifier_ops *ops = env->ops;
9802 int i, cnt, size, ctx_field_size, delta = 0;
9803 const int insn_cnt = env->prog->len;
9804 struct bpf_insn insn_buf[16], *insn;
9805 u32 target_size, size_default, off;
9806 struct bpf_prog *new_prog;
9807 enum bpf_access_type type;
9808 bool is_narrower_load;
9810 if (ops->gen_prologue || env->seen_direct_write) {
9811 if (!ops->gen_prologue) {
9812 verbose(env, "bpf verifier is misconfigured\n");
9815 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
9817 if (cnt >= ARRAY_SIZE(insn_buf)) {
9818 verbose(env, "bpf verifier is misconfigured\n");
9821 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
9825 env->prog = new_prog;
9830 if (bpf_prog_is_dev_bound(env->prog->aux))
9833 insn = env->prog->insnsi + delta;
9835 for (i = 0; i < insn_cnt; i++, insn++) {
9836 bpf_convert_ctx_access_t convert_ctx_access;
9838 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
9839 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
9840 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
9841 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
9843 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
9844 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
9845 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
9846 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
9851 if (type == BPF_WRITE &&
9852 env->insn_aux_data[i + delta].sanitize_stack_off) {
9853 struct bpf_insn patch[] = {
9854 /* Sanitize suspicious stack slot with zero.
9855 * There are no memory dependencies for this store,
9856 * since it's only using frame pointer and immediate
9859 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
9860 env->insn_aux_data[i + delta].sanitize_stack_off,
9862 /* the original STX instruction will immediately
9863 * overwrite the same stack slot with appropriate value
9868 cnt = ARRAY_SIZE(patch);
9869 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
9874 env->prog = new_prog;
9875 insn = new_prog->insnsi + i + delta;
9879 switch (env->insn_aux_data[i + delta].ptr_type) {
9881 if (!ops->convert_ctx_access)
9883 convert_ctx_access = ops->convert_ctx_access;
9886 case PTR_TO_SOCK_COMMON:
9887 convert_ctx_access = bpf_sock_convert_ctx_access;
9889 case PTR_TO_TCP_SOCK:
9890 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
9892 case PTR_TO_XDP_SOCK:
9893 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
9896 if (type == BPF_READ) {
9897 insn->code = BPF_LDX | BPF_PROBE_MEM |
9898 BPF_SIZE((insn)->code);
9899 env->prog->aux->num_exentries++;
9900 } else if (env->prog->type != BPF_PROG_TYPE_STRUCT_OPS) {
9901 verbose(env, "Writes through BTF pointers are not allowed\n");
9909 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
9910 size = BPF_LDST_BYTES(insn);
9912 /* If the read access is a narrower load of the field,
9913 * convert to a 4/8-byte load, to minimum program type specific
9914 * convert_ctx_access changes. If conversion is successful,
9915 * we will apply proper mask to the result.
9917 is_narrower_load = size < ctx_field_size;
9918 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
9920 if (is_narrower_load) {
9923 if (type == BPF_WRITE) {
9924 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
9929 if (ctx_field_size == 4)
9931 else if (ctx_field_size == 8)
9934 insn->off = off & ~(size_default - 1);
9935 insn->code = BPF_LDX | BPF_MEM | size_code;
9939 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
9941 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
9942 (ctx_field_size && !target_size)) {
9943 verbose(env, "bpf verifier is misconfigured\n");
9947 if (is_narrower_load && size < target_size) {
9948 u8 shift = bpf_ctx_narrow_access_offset(
9949 off, size, size_default) * 8;
9950 if (ctx_field_size <= 4) {
9952 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
9955 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
9956 (1 << size * 8) - 1);
9959 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
9962 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
9963 (1ULL << size * 8) - 1);
9967 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9973 /* keep walking new program and skip insns we just inserted */
9974 env->prog = new_prog;
9975 insn = new_prog->insnsi + i + delta;
9981 static int jit_subprogs(struct bpf_verifier_env *env)
9983 struct bpf_prog *prog = env->prog, **func, *tmp;
9984 int i, j, subprog_start, subprog_end = 0, len, subprog;
9985 struct bpf_insn *insn;
9987 int err, num_exentries;
9989 if (env->subprog_cnt <= 1)
9992 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
9993 if (insn->code != (BPF_JMP | BPF_CALL) ||
9994 insn->src_reg != BPF_PSEUDO_CALL)
9996 /* Upon error here we cannot fall back to interpreter but
9997 * need a hard reject of the program. Thus -EFAULT is
9998 * propagated in any case.
10000 subprog = find_subprog(env, i + insn->imm + 1);
10002 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
10003 i + insn->imm + 1);
10006 /* temporarily remember subprog id inside insn instead of
10007 * aux_data, since next loop will split up all insns into funcs
10009 insn->off = subprog;
10010 /* remember original imm in case JIT fails and fallback
10011 * to interpreter will be needed
10013 env->insn_aux_data[i].call_imm = insn->imm;
10014 /* point imm to __bpf_call_base+1 from JITs point of view */
10018 err = bpf_prog_alloc_jited_linfo(prog);
10020 goto out_undo_insn;
10023 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
10025 goto out_undo_insn;
10027 for (i = 0; i < env->subprog_cnt; i++) {
10028 subprog_start = subprog_end;
10029 subprog_end = env->subprog_info[i + 1].start;
10031 len = subprog_end - subprog_start;
10032 /* BPF_PROG_RUN doesn't call subprogs directly,
10033 * hence main prog stats include the runtime of subprogs.
10034 * subprogs don't have IDs and not reachable via prog_get_next_id
10035 * func[i]->aux->stats will never be accessed and stays NULL
10037 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
10040 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
10041 len * sizeof(struct bpf_insn));
10042 func[i]->type = prog->type;
10043 func[i]->len = len;
10044 if (bpf_prog_calc_tag(func[i]))
10046 func[i]->is_func = 1;
10047 func[i]->aux->func_idx = i;
10048 /* the btf and func_info will be freed only at prog->aux */
10049 func[i]->aux->btf = prog->aux->btf;
10050 func[i]->aux->func_info = prog->aux->func_info;
10052 /* Use bpf_prog_F_tag to indicate functions in stack traces.
10053 * Long term would need debug info to populate names
10055 func[i]->aux->name[0] = 'F';
10056 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
10057 func[i]->jit_requested = 1;
10058 func[i]->aux->linfo = prog->aux->linfo;
10059 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
10060 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
10061 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
10063 insn = func[i]->insnsi;
10064 for (j = 0; j < func[i]->len; j++, insn++) {
10065 if (BPF_CLASS(insn->code) == BPF_LDX &&
10066 BPF_MODE(insn->code) == BPF_PROBE_MEM)
10069 func[i]->aux->num_exentries = num_exentries;
10070 func[i] = bpf_int_jit_compile(func[i]);
10071 if (!func[i]->jited) {
10077 /* at this point all bpf functions were successfully JITed
10078 * now populate all bpf_calls with correct addresses and
10079 * run last pass of JIT
10081 for (i = 0; i < env->subprog_cnt; i++) {
10082 insn = func[i]->insnsi;
10083 for (j = 0; j < func[i]->len; j++, insn++) {
10084 if (insn->code != (BPF_JMP | BPF_CALL) ||
10085 insn->src_reg != BPF_PSEUDO_CALL)
10087 subprog = insn->off;
10088 insn->imm = BPF_CAST_CALL(func[subprog]->bpf_func) -
10092 /* we use the aux data to keep a list of the start addresses
10093 * of the JITed images for each function in the program
10095 * for some architectures, such as powerpc64, the imm field
10096 * might not be large enough to hold the offset of the start
10097 * address of the callee's JITed image from __bpf_call_base
10099 * in such cases, we can lookup the start address of a callee
10100 * by using its subprog id, available from the off field of
10101 * the call instruction, as an index for this list
10103 func[i]->aux->func = func;
10104 func[i]->aux->func_cnt = env->subprog_cnt;
10106 for (i = 0; i < env->subprog_cnt; i++) {
10107 old_bpf_func = func[i]->bpf_func;
10108 tmp = bpf_int_jit_compile(func[i]);
10109 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
10110 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
10117 /* finally lock prog and jit images for all functions and
10118 * populate kallsysm
10120 for (i = 0; i < env->subprog_cnt; i++) {
10121 bpf_prog_lock_ro(func[i]);
10122 bpf_prog_kallsyms_add(func[i]);
10125 /* Last step: make now unused interpreter insns from main
10126 * prog consistent for later dump requests, so they can
10127 * later look the same as if they were interpreted only.
10129 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
10130 if (insn->code != (BPF_JMP | BPF_CALL) ||
10131 insn->src_reg != BPF_PSEUDO_CALL)
10133 insn->off = env->insn_aux_data[i].call_imm;
10134 subprog = find_subprog(env, i + insn->off + 1);
10135 insn->imm = subprog;
10139 prog->bpf_func = func[0]->bpf_func;
10140 prog->aux->func = func;
10141 prog->aux->func_cnt = env->subprog_cnt;
10142 bpf_prog_free_unused_jited_linfo(prog);
10145 for (i = 0; i < env->subprog_cnt; i++)
10147 bpf_jit_free(func[i]);
10150 /* cleanup main prog to be interpreted */
10151 prog->jit_requested = 0;
10152 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
10153 if (insn->code != (BPF_JMP | BPF_CALL) ||
10154 insn->src_reg != BPF_PSEUDO_CALL)
10157 insn->imm = env->insn_aux_data[i].call_imm;
10159 bpf_prog_free_jited_linfo(prog);
10163 static int fixup_call_args(struct bpf_verifier_env *env)
10165 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
10166 struct bpf_prog *prog = env->prog;
10167 struct bpf_insn *insn = prog->insnsi;
10172 if (env->prog->jit_requested &&
10173 !bpf_prog_is_dev_bound(env->prog->aux)) {
10174 err = jit_subprogs(env);
10177 if (err == -EFAULT)
10180 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
10181 for (i = 0; i < prog->len; i++, insn++) {
10182 if (insn->code != (BPF_JMP | BPF_CALL) ||
10183 insn->src_reg != BPF_PSEUDO_CALL)
10185 depth = get_callee_stack_depth(env, insn, i);
10188 bpf_patch_call_args(insn, depth);
10195 /* fixup insn->imm field of bpf_call instructions
10196 * and inline eligible helpers as explicit sequence of BPF instructions
10198 * this function is called after eBPF program passed verification
10200 static int fixup_bpf_calls(struct bpf_verifier_env *env)
10202 struct bpf_prog *prog = env->prog;
10203 bool expect_blinding = bpf_jit_blinding_enabled(prog);
10204 struct bpf_insn *insn = prog->insnsi;
10205 const struct bpf_func_proto *fn;
10206 const int insn_cnt = prog->len;
10207 const struct bpf_map_ops *ops;
10208 struct bpf_insn_aux_data *aux;
10209 struct bpf_insn insn_buf[16];
10210 struct bpf_prog *new_prog;
10211 struct bpf_map *map_ptr;
10212 int i, ret, cnt, delta = 0;
10214 for (i = 0; i < insn_cnt; i++, insn++) {
10215 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
10216 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
10217 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
10218 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
10219 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
10220 struct bpf_insn mask_and_div[] = {
10221 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
10222 /* Rx div 0 -> 0 */
10223 BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2),
10224 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
10225 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
10228 struct bpf_insn mask_and_mod[] = {
10229 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
10230 /* Rx mod 0 -> Rx */
10231 BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1),
10234 struct bpf_insn *patchlet;
10236 if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
10237 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
10238 patchlet = mask_and_div + (is64 ? 1 : 0);
10239 cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0);
10241 patchlet = mask_and_mod + (is64 ? 1 : 0);
10242 cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0);
10245 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
10250 env->prog = prog = new_prog;
10251 insn = new_prog->insnsi + i + delta;
10255 if (BPF_CLASS(insn->code) == BPF_LD &&
10256 (BPF_MODE(insn->code) == BPF_ABS ||
10257 BPF_MODE(insn->code) == BPF_IND)) {
10258 cnt = env->ops->gen_ld_abs(insn, insn_buf);
10259 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
10260 verbose(env, "bpf verifier is misconfigured\n");
10264 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
10269 env->prog = prog = new_prog;
10270 insn = new_prog->insnsi + i + delta;
10274 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
10275 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
10276 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
10277 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
10278 struct bpf_insn insn_buf[16];
10279 struct bpf_insn *patch = &insn_buf[0];
10283 aux = &env->insn_aux_data[i + delta];
10284 if (!aux->alu_state ||
10285 aux->alu_state == BPF_ALU_NON_POINTER)
10288 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
10289 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
10290 BPF_ALU_SANITIZE_SRC;
10292 off_reg = issrc ? insn->src_reg : insn->dst_reg;
10294 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
10295 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit - 1);
10296 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
10297 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
10298 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
10299 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
10301 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX,
10303 insn->src_reg = BPF_REG_AX;
10305 *patch++ = BPF_ALU64_REG(BPF_AND, off_reg,
10309 insn->code = insn->code == code_add ?
10310 code_sub : code_add;
10312 if (issrc && isneg)
10313 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
10314 cnt = patch - insn_buf;
10316 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
10321 env->prog = prog = new_prog;
10322 insn = new_prog->insnsi + i + delta;
10326 if (insn->code != (BPF_JMP | BPF_CALL))
10328 if (insn->src_reg == BPF_PSEUDO_CALL)
10331 if (insn->imm == BPF_FUNC_get_route_realm)
10332 prog->dst_needed = 1;
10333 if (insn->imm == BPF_FUNC_get_prandom_u32)
10334 bpf_user_rnd_init_once();
10335 if (insn->imm == BPF_FUNC_override_return)
10336 prog->kprobe_override = 1;
10337 if (insn->imm == BPF_FUNC_tail_call) {
10338 /* If we tail call into other programs, we
10339 * cannot make any assumptions since they can
10340 * be replaced dynamically during runtime in
10341 * the program array.
10343 prog->cb_access = 1;
10344 env->prog->aux->stack_depth = MAX_BPF_STACK;
10345 env->prog->aux->max_pkt_offset = MAX_PACKET_OFF;
10347 /* mark bpf_tail_call as different opcode to avoid
10348 * conditional branch in the interpeter for every normal
10349 * call and to prevent accidental JITing by JIT compiler
10350 * that doesn't support bpf_tail_call yet
10353 insn->code = BPF_JMP | BPF_TAIL_CALL;
10355 aux = &env->insn_aux_data[i + delta];
10356 if (env->bpf_capable && !expect_blinding &&
10357 prog->jit_requested &&
10358 !bpf_map_key_poisoned(aux) &&
10359 !bpf_map_ptr_poisoned(aux) &&
10360 !bpf_map_ptr_unpriv(aux)) {
10361 struct bpf_jit_poke_descriptor desc = {
10362 .reason = BPF_POKE_REASON_TAIL_CALL,
10363 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
10364 .tail_call.key = bpf_map_key_immediate(aux),
10367 ret = bpf_jit_add_poke_descriptor(prog, &desc);
10369 verbose(env, "adding tail call poke descriptor failed\n");
10373 insn->imm = ret + 1;
10377 if (!bpf_map_ptr_unpriv(aux))
10380 /* instead of changing every JIT dealing with tail_call
10381 * emit two extra insns:
10382 * if (index >= max_entries) goto out;
10383 * index &= array->index_mask;
10384 * to avoid out-of-bounds cpu speculation
10386 if (bpf_map_ptr_poisoned(aux)) {
10387 verbose(env, "tail_call abusing map_ptr\n");
10391 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
10392 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
10393 map_ptr->max_entries, 2);
10394 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
10395 container_of(map_ptr,
10398 insn_buf[2] = *insn;
10400 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
10405 env->prog = prog = new_prog;
10406 insn = new_prog->insnsi + i + delta;
10410 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
10411 * and other inlining handlers are currently limited to 64 bit
10414 if (prog->jit_requested && BITS_PER_LONG == 64 &&
10415 (insn->imm == BPF_FUNC_map_lookup_elem ||
10416 insn->imm == BPF_FUNC_map_update_elem ||
10417 insn->imm == BPF_FUNC_map_delete_elem ||
10418 insn->imm == BPF_FUNC_map_push_elem ||
10419 insn->imm == BPF_FUNC_map_pop_elem ||
10420 insn->imm == BPF_FUNC_map_peek_elem)) {
10421 aux = &env->insn_aux_data[i + delta];
10422 if (bpf_map_ptr_poisoned(aux))
10423 goto patch_call_imm;
10425 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
10426 ops = map_ptr->ops;
10427 if (insn->imm == BPF_FUNC_map_lookup_elem &&
10428 ops->map_gen_lookup) {
10429 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
10430 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
10431 verbose(env, "bpf verifier is misconfigured\n");
10435 new_prog = bpf_patch_insn_data(env, i + delta,
10441 env->prog = prog = new_prog;
10442 insn = new_prog->insnsi + i + delta;
10446 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
10447 (void *(*)(struct bpf_map *map, void *key))NULL));
10448 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
10449 (int (*)(struct bpf_map *map, void *key))NULL));
10450 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
10451 (int (*)(struct bpf_map *map, void *key, void *value,
10453 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
10454 (int (*)(struct bpf_map *map, void *value,
10456 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
10457 (int (*)(struct bpf_map *map, void *value))NULL));
10458 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
10459 (int (*)(struct bpf_map *map, void *value))NULL));
10461 switch (insn->imm) {
10462 case BPF_FUNC_map_lookup_elem:
10463 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
10466 case BPF_FUNC_map_update_elem:
10467 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
10470 case BPF_FUNC_map_delete_elem:
10471 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
10474 case BPF_FUNC_map_push_elem:
10475 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
10478 case BPF_FUNC_map_pop_elem:
10479 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
10482 case BPF_FUNC_map_peek_elem:
10483 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
10488 goto patch_call_imm;
10491 if (prog->jit_requested && BITS_PER_LONG == 64 &&
10492 insn->imm == BPF_FUNC_jiffies64) {
10493 struct bpf_insn ld_jiffies_addr[2] = {
10494 BPF_LD_IMM64(BPF_REG_0,
10495 (unsigned long)&jiffies),
10498 insn_buf[0] = ld_jiffies_addr[0];
10499 insn_buf[1] = ld_jiffies_addr[1];
10500 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
10504 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
10510 env->prog = prog = new_prog;
10511 insn = new_prog->insnsi + i + delta;
10516 fn = env->ops->get_func_proto(insn->imm, env->prog);
10517 /* all functions that have prototype and verifier allowed
10518 * programs to call them, must be real in-kernel functions
10522 "kernel subsystem misconfigured func %s#%d\n",
10523 func_id_name(insn->imm), insn->imm);
10526 insn->imm = fn->func - __bpf_call_base;
10529 /* Since poke tab is now finalized, publish aux to tracker. */
10530 for (i = 0; i < prog->aux->size_poke_tab; i++) {
10531 map_ptr = prog->aux->poke_tab[i].tail_call.map;
10532 if (!map_ptr->ops->map_poke_track ||
10533 !map_ptr->ops->map_poke_untrack ||
10534 !map_ptr->ops->map_poke_run) {
10535 verbose(env, "bpf verifier is misconfigured\n");
10539 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
10541 verbose(env, "tracking tail call prog failed\n");
10549 static void free_states(struct bpf_verifier_env *env)
10551 struct bpf_verifier_state_list *sl, *sln;
10554 sl = env->free_list;
10557 free_verifier_state(&sl->state, false);
10561 env->free_list = NULL;
10563 if (!env->explored_states)
10566 for (i = 0; i < state_htab_size(env); i++) {
10567 sl = env->explored_states[i];
10571 free_verifier_state(&sl->state, false);
10575 env->explored_states[i] = NULL;
10579 /* The verifier is using insn_aux_data[] to store temporary data during
10580 * verification and to store information for passes that run after the
10581 * verification like dead code sanitization. do_check_common() for subprogram N
10582 * may analyze many other subprograms. sanitize_insn_aux_data() clears all
10583 * temporary data after do_check_common() finds that subprogram N cannot be
10584 * verified independently. pass_cnt counts the number of times
10585 * do_check_common() was run and insn->aux->seen tells the pass number
10586 * insn_aux_data was touched. These variables are compared to clear temporary
10587 * data from failed pass. For testing and experiments do_check_common() can be
10588 * run multiple times even when prior attempt to verify is unsuccessful.
10590 static void sanitize_insn_aux_data(struct bpf_verifier_env *env)
10592 struct bpf_insn *insn = env->prog->insnsi;
10593 struct bpf_insn_aux_data *aux;
10596 for (i = 0; i < env->prog->len; i++) {
10597 class = BPF_CLASS(insn[i].code);
10598 if (class != BPF_LDX && class != BPF_STX)
10600 aux = &env->insn_aux_data[i];
10601 if (aux->seen != env->pass_cnt)
10603 memset(aux, 0, offsetof(typeof(*aux), orig_idx));
10607 static int do_check_common(struct bpf_verifier_env *env, int subprog)
10609 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
10610 struct bpf_verifier_state *state;
10611 struct bpf_reg_state *regs;
10614 env->prev_linfo = NULL;
10617 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
10620 state->curframe = 0;
10621 state->speculative = false;
10622 state->branches = 1;
10623 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
10624 if (!state->frame[0]) {
10628 env->cur_state = state;
10629 init_func_state(env, state->frame[0],
10630 BPF_MAIN_FUNC /* callsite */,
10634 regs = state->frame[state->curframe]->regs;
10635 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
10636 ret = btf_prepare_func_args(env, subprog, regs);
10639 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
10640 if (regs[i].type == PTR_TO_CTX)
10641 mark_reg_known_zero(env, regs, i);
10642 else if (regs[i].type == SCALAR_VALUE)
10643 mark_reg_unknown(env, regs, i);
10646 /* 1st arg to a function */
10647 regs[BPF_REG_1].type = PTR_TO_CTX;
10648 mark_reg_known_zero(env, regs, BPF_REG_1);
10649 ret = btf_check_func_arg_match(env, subprog, regs);
10650 if (ret == -EFAULT)
10651 /* unlikely verifier bug. abort.
10652 * ret == 0 and ret < 0 are sadly acceptable for
10653 * main() function due to backward compatibility.
10654 * Like socket filter program may be written as:
10655 * int bpf_prog(struct pt_regs *ctx)
10656 * and never dereference that ctx in the program.
10657 * 'struct pt_regs' is a type mismatch for socket
10658 * filter that should be using 'struct __sk_buff'.
10663 ret = do_check(env);
10665 /* check for NULL is necessary, since cur_state can be freed inside
10666 * do_check() under memory pressure.
10668 if (env->cur_state) {
10669 free_verifier_state(env->cur_state, true);
10670 env->cur_state = NULL;
10672 while (!pop_stack(env, NULL, NULL, false));
10673 if (!ret && pop_log)
10674 bpf_vlog_reset(&env->log, 0);
10677 /* clean aux data in case subprog was rejected */
10678 sanitize_insn_aux_data(env);
10682 /* Verify all global functions in a BPF program one by one based on their BTF.
10683 * All global functions must pass verification. Otherwise the whole program is rejected.
10694 * foo() will be verified first for R1=any_scalar_value. During verification it
10695 * will be assumed that bar() already verified successfully and call to bar()
10696 * from foo() will be checked for type match only. Later bar() will be verified
10697 * independently to check that it's safe for R1=any_scalar_value.
10699 static int do_check_subprogs(struct bpf_verifier_env *env)
10701 struct bpf_prog_aux *aux = env->prog->aux;
10704 if (!aux->func_info)
10707 for (i = 1; i < env->subprog_cnt; i++) {
10708 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
10710 env->insn_idx = env->subprog_info[i].start;
10711 WARN_ON_ONCE(env->insn_idx == 0);
10712 ret = do_check_common(env, i);
10715 } else if (env->log.level & BPF_LOG_LEVEL) {
10717 "Func#%d is safe for any args that match its prototype\n",
10724 static int do_check_main(struct bpf_verifier_env *env)
10729 ret = do_check_common(env, 0);
10731 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
10736 static void print_verification_stats(struct bpf_verifier_env *env)
10740 if (env->log.level & BPF_LOG_STATS) {
10741 verbose(env, "verification time %lld usec\n",
10742 div_u64(env->verification_time, 1000));
10743 verbose(env, "stack depth ");
10744 for (i = 0; i < env->subprog_cnt; i++) {
10745 u32 depth = env->subprog_info[i].stack_depth;
10747 verbose(env, "%d", depth);
10748 if (i + 1 < env->subprog_cnt)
10751 verbose(env, "\n");
10753 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
10754 "total_states %d peak_states %d mark_read %d\n",
10755 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
10756 env->max_states_per_insn, env->total_states,
10757 env->peak_states, env->longest_mark_read_walk);
10760 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
10762 const struct btf_type *t, *func_proto;
10763 const struct bpf_struct_ops *st_ops;
10764 const struct btf_member *member;
10765 struct bpf_prog *prog = env->prog;
10766 u32 btf_id, member_idx;
10769 btf_id = prog->aux->attach_btf_id;
10770 st_ops = bpf_struct_ops_find(btf_id);
10772 verbose(env, "attach_btf_id %u is not a supported struct\n",
10778 member_idx = prog->expected_attach_type;
10779 if (member_idx >= btf_type_vlen(t)) {
10780 verbose(env, "attach to invalid member idx %u of struct %s\n",
10781 member_idx, st_ops->name);
10785 member = &btf_type_member(t)[member_idx];
10786 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
10787 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
10790 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
10791 mname, member_idx, st_ops->name);
10795 if (st_ops->check_member) {
10796 int err = st_ops->check_member(t, member);
10799 verbose(env, "attach to unsupported member %s of struct %s\n",
10800 mname, st_ops->name);
10805 prog->aux->attach_func_proto = func_proto;
10806 prog->aux->attach_func_name = mname;
10807 env->ops = st_ops->verifier_ops;
10811 #define SECURITY_PREFIX "security_"
10813 static int check_attach_modify_return(struct bpf_prog *prog, unsigned long addr)
10815 if (within_error_injection_list(addr) ||
10816 !strncmp(SECURITY_PREFIX, prog->aux->attach_func_name,
10817 sizeof(SECURITY_PREFIX) - 1))
10823 static int check_attach_btf_id(struct bpf_verifier_env *env)
10825 struct bpf_prog *prog = env->prog;
10826 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
10827 struct bpf_prog *tgt_prog = prog->aux->linked_prog;
10828 u32 btf_id = prog->aux->attach_btf_id;
10829 const char prefix[] = "btf_trace_";
10830 struct btf_func_model fmodel;
10831 int ret = 0, subprog = -1, i;
10832 struct bpf_trampoline *tr;
10833 const struct btf_type *t;
10834 bool conservative = true;
10840 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
10841 return check_struct_ops_btf_id(env);
10843 if (prog->type != BPF_PROG_TYPE_TRACING &&
10844 prog->type != BPF_PROG_TYPE_LSM &&
10849 verbose(env, "Tracing programs must provide btf_id\n");
10852 btf = bpf_prog_get_target_btf(prog);
10855 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
10858 t = btf_type_by_id(btf, btf_id);
10860 verbose(env, "attach_btf_id %u is invalid\n", btf_id);
10863 tname = btf_name_by_offset(btf, t->name_off);
10865 verbose(env, "attach_btf_id %u doesn't have a name\n", btf_id);
10869 struct bpf_prog_aux *aux = tgt_prog->aux;
10871 for (i = 0; i < aux->func_info_cnt; i++)
10872 if (aux->func_info[i].type_id == btf_id) {
10876 if (subprog == -1) {
10877 verbose(env, "Subprog %s doesn't exist\n", tname);
10880 conservative = aux->func_info_aux[subprog].unreliable;
10881 if (prog_extension) {
10882 if (conservative) {
10884 "Cannot replace static functions\n");
10887 if (!prog->jit_requested) {
10889 "Extension programs should be JITed\n");
10892 env->ops = bpf_verifier_ops[tgt_prog->type];
10893 prog->expected_attach_type = tgt_prog->expected_attach_type;
10895 if (!tgt_prog->jited) {
10896 verbose(env, "Can attach to only JITed progs\n");
10899 if (tgt_prog->type == prog->type) {
10900 /* Cannot fentry/fexit another fentry/fexit program.
10901 * Cannot attach program extension to another extension.
10902 * It's ok to attach fentry/fexit to extension program.
10904 verbose(env, "Cannot recursively attach\n");
10907 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
10909 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
10910 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
10911 /* Program extensions can extend all program types
10912 * except fentry/fexit. The reason is the following.
10913 * The fentry/fexit programs are used for performance
10914 * analysis, stats and can be attached to any program
10915 * type except themselves. When extension program is
10916 * replacing XDP function it is necessary to allow
10917 * performance analysis of all functions. Both original
10918 * XDP program and its program extension. Hence
10919 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
10920 * allowed. If extending of fentry/fexit was allowed it
10921 * would be possible to create long call chain
10922 * fentry->extension->fentry->extension beyond
10923 * reasonable stack size. Hence extending fentry is not
10926 verbose(env, "Cannot extend fentry/fexit\n");
10929 key = ((u64)aux->id) << 32 | btf_id;
10931 if (prog_extension) {
10932 verbose(env, "Cannot replace kernel functions\n");
10938 switch (prog->expected_attach_type) {
10939 case BPF_TRACE_RAW_TP:
10942 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
10945 if (!btf_type_is_typedef(t)) {
10946 verbose(env, "attach_btf_id %u is not a typedef\n",
10950 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
10951 verbose(env, "attach_btf_id %u points to wrong type name %s\n",
10955 tname += sizeof(prefix) - 1;
10956 t = btf_type_by_id(btf, t->type);
10957 if (!btf_type_is_ptr(t))
10958 /* should never happen in valid vmlinux build */
10960 t = btf_type_by_id(btf, t->type);
10961 if (!btf_type_is_func_proto(t))
10962 /* should never happen in valid vmlinux build */
10965 /* remember two read only pointers that are valid for
10966 * the life time of the kernel
10968 prog->aux->attach_func_name = tname;
10969 prog->aux->attach_func_proto = t;
10970 prog->aux->attach_btf_trace = true;
10972 case BPF_TRACE_ITER:
10973 if (!btf_type_is_func(t)) {
10974 verbose(env, "attach_btf_id %u is not a function\n",
10978 t = btf_type_by_id(btf, t->type);
10979 if (!btf_type_is_func_proto(t))
10981 prog->aux->attach_func_name = tname;
10982 prog->aux->attach_func_proto = t;
10983 if (!bpf_iter_prog_supported(prog))
10985 ret = btf_distill_func_proto(&env->log, btf, t,
10989 if (!prog_extension)
10992 case BPF_MODIFY_RETURN:
10994 case BPF_TRACE_FENTRY:
10995 case BPF_TRACE_FEXIT:
10996 prog->aux->attach_func_name = tname;
10997 if (prog->type == BPF_PROG_TYPE_LSM) {
10998 ret = bpf_lsm_verify_prog(&env->log, prog);
11003 if (!btf_type_is_func(t)) {
11004 verbose(env, "attach_btf_id %u is not a function\n",
11008 if (prog_extension &&
11009 btf_check_type_match(env, prog, btf, t))
11011 t = btf_type_by_id(btf, t->type);
11012 if (!btf_type_is_func_proto(t))
11014 tr = bpf_trampoline_lookup(key);
11017 /* t is either vmlinux type or another program's type */
11018 prog->aux->attach_func_proto = t;
11019 mutex_lock(&tr->mutex);
11020 if (tr->func.addr) {
11021 prog->aux->trampoline = tr;
11024 if (tgt_prog && conservative) {
11025 prog->aux->attach_func_proto = NULL;
11028 ret = btf_distill_func_proto(&env->log, btf, t,
11029 tname, &tr->func.model);
11034 addr = (long) tgt_prog->bpf_func;
11036 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
11038 addr = kallsyms_lookup_name(tname);
11041 "The address of function %s cannot be found\n",
11048 if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
11049 ret = check_attach_modify_return(prog, addr);
11051 verbose(env, "%s() is not modifiable\n",
11052 prog->aux->attach_func_name);
11057 tr->func.addr = (void *)addr;
11058 prog->aux->trampoline = tr;
11060 mutex_unlock(&tr->mutex);
11062 bpf_trampoline_put(tr);
11067 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr,
11068 union bpf_attr __user *uattr)
11070 u64 start_time = ktime_get_ns();
11071 struct bpf_verifier_env *env;
11072 struct bpf_verifier_log *log;
11073 int i, len, ret = -EINVAL;
11076 /* no program is valid */
11077 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
11080 /* 'struct bpf_verifier_env' can be global, but since it's not small,
11081 * allocate/free it every time bpf_check() is called
11083 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
11088 len = (*prog)->len;
11089 env->insn_aux_data =
11090 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
11092 if (!env->insn_aux_data)
11094 for (i = 0; i < len; i++)
11095 env->insn_aux_data[i].orig_idx = i;
11097 env->ops = bpf_verifier_ops[env->prog->type];
11098 is_priv = bpf_capable();
11100 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
11101 mutex_lock(&bpf_verifier_lock);
11103 btf_vmlinux = btf_parse_vmlinux();
11104 mutex_unlock(&bpf_verifier_lock);
11107 /* grab the mutex to protect few globals used by verifier */
11109 mutex_lock(&bpf_verifier_lock);
11111 if (attr->log_level || attr->log_buf || attr->log_size) {
11112 /* user requested verbose verifier output
11113 * and supplied buffer to store the verification trace
11115 log->level = attr->log_level;
11116 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
11117 log->len_total = attr->log_size;
11120 /* log attributes have to be sane */
11121 if (log->len_total < 128 || log->len_total > UINT_MAX >> 2 ||
11122 !log->level || !log->ubuf || log->level & ~BPF_LOG_MASK)
11126 if (IS_ERR(btf_vmlinux)) {
11127 /* Either gcc or pahole or kernel are broken. */
11128 verbose(env, "in-kernel BTF is malformed\n");
11129 ret = PTR_ERR(btf_vmlinux);
11130 goto skip_full_check;
11133 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
11134 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
11135 env->strict_alignment = true;
11136 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
11137 env->strict_alignment = false;
11139 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
11140 env->allow_ptr_to_map_access = bpf_allow_ptr_to_map_access();
11141 env->bypass_spec_v1 = bpf_bypass_spec_v1();
11142 env->bypass_spec_v4 = bpf_bypass_spec_v4();
11143 env->bpf_capable = bpf_capable();
11146 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
11148 ret = replace_map_fd_with_map_ptr(env);
11150 goto skip_full_check;
11152 if (bpf_prog_is_dev_bound(env->prog->aux)) {
11153 ret = bpf_prog_offload_verifier_prep(env->prog);
11155 goto skip_full_check;
11158 env->explored_states = kvcalloc(state_htab_size(env),
11159 sizeof(struct bpf_verifier_state_list *),
11162 if (!env->explored_states)
11163 goto skip_full_check;
11165 ret = check_subprogs(env);
11167 goto skip_full_check;
11169 ret = check_btf_info(env, attr, uattr);
11171 goto skip_full_check;
11173 ret = check_attach_btf_id(env);
11175 goto skip_full_check;
11177 ret = check_cfg(env);
11179 goto skip_full_check;
11181 ret = do_check_subprogs(env);
11182 ret = ret ?: do_check_main(env);
11184 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
11185 ret = bpf_prog_offload_finalize(env);
11188 kvfree(env->explored_states);
11191 ret = check_max_stack_depth(env);
11193 /* instruction rewrites happen after this point */
11196 opt_hard_wire_dead_code_branches(env);
11198 ret = opt_remove_dead_code(env);
11200 ret = opt_remove_nops(env);
11203 sanitize_dead_code(env);
11207 /* program is valid, convert *(u32*)(ctx + off) accesses */
11208 ret = convert_ctx_accesses(env);
11211 ret = fixup_bpf_calls(env);
11213 /* do 32-bit optimization after insn patching has done so those patched
11214 * insns could be handled correctly.
11216 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
11217 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
11218 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
11223 ret = fixup_call_args(env);
11225 env->verification_time = ktime_get_ns() - start_time;
11226 print_verification_stats(env);
11228 if (log->level && bpf_verifier_log_full(log))
11230 if (log->level && !log->ubuf) {
11232 goto err_release_maps;
11235 if (ret == 0 && env->used_map_cnt) {
11236 /* if program passed verifier, update used_maps in bpf_prog_info */
11237 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
11238 sizeof(env->used_maps[0]),
11241 if (!env->prog->aux->used_maps) {
11243 goto err_release_maps;
11246 memcpy(env->prog->aux->used_maps, env->used_maps,
11247 sizeof(env->used_maps[0]) * env->used_map_cnt);
11248 env->prog->aux->used_map_cnt = env->used_map_cnt;
11250 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
11251 * bpf_ld_imm64 instructions
11253 convert_pseudo_ld_imm64(env);
11257 adjust_btf_func(env);
11260 if (!env->prog->aux->used_maps)
11261 /* if we didn't copy map pointers into bpf_prog_info, release
11262 * them now. Otherwise free_used_maps() will release them.
11266 /* extension progs temporarily inherit the attach_type of their targets
11267 for verification purposes, so set it back to zero before returning
11269 if (env->prog->type == BPF_PROG_TYPE_EXT)
11270 env->prog->expected_attach_type = 0;
11275 mutex_unlock(&bpf_verifier_lock);
11276 vfree(env->insn_aux_data);