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 #include <linux/bpf_types.h>
36 /* bpf_check() is a static code analyzer that walks eBPF program
37 * instruction by instruction and updates register/stack state.
38 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
40 * The first pass is depth-first-search to check that the program is a DAG.
41 * It rejects the following programs:
42 * - larger than BPF_MAXINSNS insns
43 * - if loop is present (detected via back-edge)
44 * - unreachable insns exist (shouldn't be a forest. program = one function)
45 * - out of bounds or malformed jumps
46 * The second pass is all possible path descent from the 1st insn.
47 * Since it's analyzing all pathes through the program, the length of the
48 * analysis is limited to 64k insn, which may be hit even if total number of
49 * insn is less then 4K, but there are too many branches that change stack/regs.
50 * Number of 'branches to be analyzed' is limited to 1k
52 * On entry to each instruction, each register has a type, and the instruction
53 * changes the types of the registers depending on instruction semantics.
54 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
57 * All registers are 64-bit.
58 * R0 - return register
59 * R1-R5 argument passing registers
60 * R6-R9 callee saved registers
61 * R10 - frame pointer read-only
63 * At the start of BPF program the register R1 contains a pointer to bpf_context
64 * and has type PTR_TO_CTX.
66 * Verifier tracks arithmetic operations on pointers in case:
67 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
68 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
69 * 1st insn copies R10 (which has FRAME_PTR) type into R1
70 * and 2nd arithmetic instruction is pattern matched to recognize
71 * that it wants to construct a pointer to some element within stack.
72 * So after 2nd insn, the register R1 has type PTR_TO_STACK
73 * (and -20 constant is saved for further stack bounds checking).
74 * Meaning that this reg is a pointer to stack plus known immediate constant.
76 * Most of the time the registers have SCALAR_VALUE type, which
77 * means the register has some value, but it's not a valid pointer.
78 * (like pointer plus pointer becomes SCALAR_VALUE type)
80 * When verifier sees load or store instructions the type of base register
81 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
82 * four pointer types recognized by check_mem_access() function.
84 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
85 * and the range of [ptr, ptr + map's value_size) is accessible.
87 * registers used to pass values to function calls are checked against
88 * function argument constraints.
90 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
91 * It means that the register type passed to this function must be
92 * PTR_TO_STACK and it will be used inside the function as
93 * 'pointer to map element key'
95 * For example the argument constraints for bpf_map_lookup_elem():
96 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
97 * .arg1_type = ARG_CONST_MAP_PTR,
98 * .arg2_type = ARG_PTR_TO_MAP_KEY,
100 * ret_type says that this function returns 'pointer to map elem value or null'
101 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
102 * 2nd argument should be a pointer to stack, which will be used inside
103 * the helper function as a pointer to map element key.
105 * On the kernel side the helper function looks like:
106 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
108 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
109 * void *key = (void *) (unsigned long) r2;
112 * here kernel can access 'key' and 'map' pointers safely, knowing that
113 * [key, key + map->key_size) bytes are valid and were initialized on
114 * the stack of eBPF program.
117 * Corresponding eBPF program may look like:
118 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
119 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
120 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
121 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
122 * here verifier looks at prototype of map_lookup_elem() and sees:
123 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
124 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
126 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
127 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
128 * and were initialized prior to this call.
129 * If it's ok, then verifier allows this BPF_CALL insn and looks at
130 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
131 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
132 * returns ether pointer to map value or NULL.
134 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
135 * insn, the register holding that pointer in the true branch changes state to
136 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
137 * branch. See check_cond_jmp_op().
139 * After the call R0 is set to return type of the function and registers R1-R5
140 * are set to NOT_INIT to indicate that they are no longer readable.
142 * The following reference types represent a potential reference to a kernel
143 * resource which, after first being allocated, must be checked and freed by
145 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
147 * When the verifier sees a helper call return a reference type, it allocates a
148 * pointer id for the reference and stores it in the current function state.
149 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
150 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
151 * passes through a NULL-check conditional. For the branch wherein the state is
152 * changed to CONST_IMM, the verifier releases the reference.
154 * For each helper function that allocates a reference, such as
155 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
156 * bpf_sk_release(). When a reference type passes into the release function,
157 * the verifier also releases the reference. If any unchecked or unreleased
158 * reference remains at the end of the program, the verifier rejects it.
161 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
162 struct bpf_verifier_stack_elem {
163 /* verifer state is 'st'
164 * before processing instruction 'insn_idx'
165 * and after processing instruction 'prev_insn_idx'
167 struct bpf_verifier_state st;
170 struct bpf_verifier_stack_elem *next;
173 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
174 #define BPF_COMPLEXITY_LIMIT_STATES 64
176 #define BPF_MAP_KEY_POISON (1ULL << 63)
177 #define BPF_MAP_KEY_SEEN (1ULL << 62)
179 #define BPF_MAP_PTR_UNPRIV 1UL
180 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
181 POISON_POINTER_DELTA))
182 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
184 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
186 return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON;
189 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
191 return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV;
194 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
195 const struct bpf_map *map, bool unpriv)
197 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
198 unpriv |= bpf_map_ptr_unpriv(aux);
199 aux->map_ptr_state = (unsigned long)map |
200 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
203 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
205 return aux->map_key_state & BPF_MAP_KEY_POISON;
208 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
210 return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
213 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
215 return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
218 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
220 bool poisoned = bpf_map_key_poisoned(aux);
222 aux->map_key_state = state | BPF_MAP_KEY_SEEN |
223 (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
226 struct bpf_call_arg_meta {
227 struct bpf_map *map_ptr;
238 struct btf *btf_vmlinux;
240 static DEFINE_MUTEX(bpf_verifier_lock);
242 static const struct bpf_line_info *
243 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
245 const struct bpf_line_info *linfo;
246 const struct bpf_prog *prog;
250 nr_linfo = prog->aux->nr_linfo;
252 if (!nr_linfo || insn_off >= prog->len)
255 linfo = prog->aux->linfo;
256 for (i = 1; i < nr_linfo; i++)
257 if (insn_off < linfo[i].insn_off)
260 return &linfo[i - 1];
263 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
268 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
270 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
271 "verifier log line truncated - local buffer too short\n");
273 n = min(log->len_total - log->len_used - 1, n);
276 if (log->level == BPF_LOG_KERNEL) {
277 pr_err("BPF:%s\n", log->kbuf);
280 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
286 /* log_level controls verbosity level of eBPF verifier.
287 * bpf_verifier_log_write() is used to dump the verification trace to the log,
288 * so the user can figure out what's wrong with the program
290 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
291 const char *fmt, ...)
295 if (!bpf_verifier_log_needed(&env->log))
299 bpf_verifier_vlog(&env->log, fmt, args);
302 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
304 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
306 struct bpf_verifier_env *env = private_data;
309 if (!bpf_verifier_log_needed(&env->log))
313 bpf_verifier_vlog(&env->log, fmt, args);
317 __printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
318 const char *fmt, ...)
322 if (!bpf_verifier_log_needed(log))
326 bpf_verifier_vlog(log, fmt, args);
330 static const char *ltrim(const char *s)
338 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
340 const char *prefix_fmt, ...)
342 const struct bpf_line_info *linfo;
344 if (!bpf_verifier_log_needed(&env->log))
347 linfo = find_linfo(env, insn_off);
348 if (!linfo || linfo == env->prev_linfo)
354 va_start(args, prefix_fmt);
355 bpf_verifier_vlog(&env->log, prefix_fmt, args);
360 ltrim(btf_name_by_offset(env->prog->aux->btf,
363 env->prev_linfo = linfo;
366 static bool type_is_pkt_pointer(enum bpf_reg_type type)
368 return type == PTR_TO_PACKET ||
369 type == PTR_TO_PACKET_META;
372 static bool type_is_sk_pointer(enum bpf_reg_type type)
374 return type == PTR_TO_SOCKET ||
375 type == PTR_TO_SOCK_COMMON ||
376 type == PTR_TO_TCP_SOCK ||
377 type == PTR_TO_XDP_SOCK;
380 static bool reg_type_may_be_null(enum bpf_reg_type type)
382 return type == PTR_TO_MAP_VALUE_OR_NULL ||
383 type == PTR_TO_SOCKET_OR_NULL ||
384 type == PTR_TO_SOCK_COMMON_OR_NULL ||
385 type == PTR_TO_TCP_SOCK_OR_NULL;
388 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
390 return reg->type == PTR_TO_MAP_VALUE &&
391 map_value_has_spin_lock(reg->map_ptr);
394 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
396 return type == PTR_TO_SOCKET ||
397 type == PTR_TO_SOCKET_OR_NULL ||
398 type == PTR_TO_TCP_SOCK ||
399 type == PTR_TO_TCP_SOCK_OR_NULL;
402 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
404 return type == ARG_PTR_TO_SOCK_COMMON;
407 /* Determine whether the function releases some resources allocated by another
408 * function call. The first reference type argument will be assumed to be
409 * released by release_reference().
411 static bool is_release_function(enum bpf_func_id func_id)
413 return func_id == BPF_FUNC_sk_release;
416 static bool is_acquire_function(enum bpf_func_id func_id)
418 return func_id == BPF_FUNC_sk_lookup_tcp ||
419 func_id == BPF_FUNC_sk_lookup_udp ||
420 func_id == BPF_FUNC_skc_lookup_tcp;
423 static bool is_ptr_cast_function(enum bpf_func_id func_id)
425 return func_id == BPF_FUNC_tcp_sock ||
426 func_id == BPF_FUNC_sk_fullsock;
429 /* string representation of 'enum bpf_reg_type' */
430 static const char * const reg_type_str[] = {
432 [SCALAR_VALUE] = "inv",
433 [PTR_TO_CTX] = "ctx",
434 [CONST_PTR_TO_MAP] = "map_ptr",
435 [PTR_TO_MAP_VALUE] = "map_value",
436 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
437 [PTR_TO_STACK] = "fp",
438 [PTR_TO_PACKET] = "pkt",
439 [PTR_TO_PACKET_META] = "pkt_meta",
440 [PTR_TO_PACKET_END] = "pkt_end",
441 [PTR_TO_FLOW_KEYS] = "flow_keys",
442 [PTR_TO_SOCKET] = "sock",
443 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
444 [PTR_TO_SOCK_COMMON] = "sock_common",
445 [PTR_TO_SOCK_COMMON_OR_NULL] = "sock_common_or_null",
446 [PTR_TO_TCP_SOCK] = "tcp_sock",
447 [PTR_TO_TCP_SOCK_OR_NULL] = "tcp_sock_or_null",
448 [PTR_TO_TP_BUFFER] = "tp_buffer",
449 [PTR_TO_XDP_SOCK] = "xdp_sock",
450 [PTR_TO_BTF_ID] = "ptr_",
453 static char slot_type_char[] = {
454 [STACK_INVALID] = '?',
460 static void print_liveness(struct bpf_verifier_env *env,
461 enum bpf_reg_liveness live)
463 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
465 if (live & REG_LIVE_READ)
467 if (live & REG_LIVE_WRITTEN)
469 if (live & REG_LIVE_DONE)
473 static struct bpf_func_state *func(struct bpf_verifier_env *env,
474 const struct bpf_reg_state *reg)
476 struct bpf_verifier_state *cur = env->cur_state;
478 return cur->frame[reg->frameno];
481 const char *kernel_type_name(u32 id)
483 return btf_name_by_offset(btf_vmlinux,
484 btf_type_by_id(btf_vmlinux, id)->name_off);
487 static void print_verifier_state(struct bpf_verifier_env *env,
488 const struct bpf_func_state *state)
490 const struct bpf_reg_state *reg;
495 verbose(env, " frame%d:", state->frameno);
496 for (i = 0; i < MAX_BPF_REG; i++) {
497 reg = &state->regs[i];
501 verbose(env, " R%d", i);
502 print_liveness(env, reg->live);
503 verbose(env, "=%s", reg_type_str[t]);
504 if (t == SCALAR_VALUE && reg->precise)
506 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
507 tnum_is_const(reg->var_off)) {
508 /* reg->off should be 0 for SCALAR_VALUE */
509 verbose(env, "%lld", reg->var_off.value + reg->off);
511 if (t == PTR_TO_BTF_ID)
512 verbose(env, "%s", kernel_type_name(reg->btf_id));
513 verbose(env, "(id=%d", reg->id);
514 if (reg_type_may_be_refcounted_or_null(t))
515 verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
516 if (t != SCALAR_VALUE)
517 verbose(env, ",off=%d", reg->off);
518 if (type_is_pkt_pointer(t))
519 verbose(env, ",r=%d", reg->range);
520 else if (t == CONST_PTR_TO_MAP ||
521 t == PTR_TO_MAP_VALUE ||
522 t == PTR_TO_MAP_VALUE_OR_NULL)
523 verbose(env, ",ks=%d,vs=%d",
524 reg->map_ptr->key_size,
525 reg->map_ptr->value_size);
526 if (tnum_is_const(reg->var_off)) {
527 /* Typically an immediate SCALAR_VALUE, but
528 * could be a pointer whose offset is too big
531 verbose(env, ",imm=%llx", reg->var_off.value);
533 if (reg->smin_value != reg->umin_value &&
534 reg->smin_value != S64_MIN)
535 verbose(env, ",smin_value=%lld",
536 (long long)reg->smin_value);
537 if (reg->smax_value != reg->umax_value &&
538 reg->smax_value != S64_MAX)
539 verbose(env, ",smax_value=%lld",
540 (long long)reg->smax_value);
541 if (reg->umin_value != 0)
542 verbose(env, ",umin_value=%llu",
543 (unsigned long long)reg->umin_value);
544 if (reg->umax_value != U64_MAX)
545 verbose(env, ",umax_value=%llu",
546 (unsigned long long)reg->umax_value);
547 if (!tnum_is_unknown(reg->var_off)) {
550 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
551 verbose(env, ",var_off=%s", tn_buf);
553 if (reg->s32_min_value != reg->smin_value &&
554 reg->s32_min_value != S32_MIN)
555 verbose(env, ",s32_min_value=%d",
556 (int)(reg->s32_min_value));
557 if (reg->s32_max_value != reg->smax_value &&
558 reg->s32_max_value != S32_MAX)
559 verbose(env, ",s32_max_value=%d",
560 (int)(reg->s32_max_value));
561 if (reg->u32_min_value != reg->umin_value &&
562 reg->u32_min_value != U32_MIN)
563 verbose(env, ",u32_min_value=%d",
564 (int)(reg->u32_min_value));
565 if (reg->u32_max_value != reg->umax_value &&
566 reg->u32_max_value != U32_MAX)
567 verbose(env, ",u32_max_value=%d",
568 (int)(reg->u32_max_value));
573 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
574 char types_buf[BPF_REG_SIZE + 1];
578 for (j = 0; j < BPF_REG_SIZE; j++) {
579 if (state->stack[i].slot_type[j] != STACK_INVALID)
581 types_buf[j] = slot_type_char[
582 state->stack[i].slot_type[j]];
584 types_buf[BPF_REG_SIZE] = 0;
587 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
588 print_liveness(env, state->stack[i].spilled_ptr.live);
589 if (state->stack[i].slot_type[0] == STACK_SPILL) {
590 reg = &state->stack[i].spilled_ptr;
592 verbose(env, "=%s", reg_type_str[t]);
593 if (t == SCALAR_VALUE && reg->precise)
595 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
596 verbose(env, "%lld", reg->var_off.value + reg->off);
598 verbose(env, "=%s", types_buf);
601 if (state->acquired_refs && state->refs[0].id) {
602 verbose(env, " refs=%d", state->refs[0].id);
603 for (i = 1; i < state->acquired_refs; i++)
604 if (state->refs[i].id)
605 verbose(env, ",%d", state->refs[i].id);
610 #define COPY_STATE_FN(NAME, COUNT, FIELD, SIZE) \
611 static int copy_##NAME##_state(struct bpf_func_state *dst, \
612 const struct bpf_func_state *src) \
616 if (WARN_ON_ONCE(dst->COUNT < src->COUNT)) { \
617 /* internal bug, make state invalid to reject the program */ \
618 memset(dst, 0, sizeof(*dst)); \
621 memcpy(dst->FIELD, src->FIELD, \
622 sizeof(*src->FIELD) * (src->COUNT / SIZE)); \
625 /* copy_reference_state() */
626 COPY_STATE_FN(reference, acquired_refs, refs, 1)
627 /* copy_stack_state() */
628 COPY_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
631 #define REALLOC_STATE_FN(NAME, COUNT, FIELD, SIZE) \
632 static int realloc_##NAME##_state(struct bpf_func_state *state, int size, \
635 u32 old_size = state->COUNT; \
636 struct bpf_##NAME##_state *new_##FIELD; \
637 int slot = size / SIZE; \
639 if (size <= old_size || !size) { \
642 state->COUNT = slot * SIZE; \
643 if (!size && old_size) { \
644 kfree(state->FIELD); \
645 state->FIELD = NULL; \
649 new_##FIELD = kmalloc_array(slot, sizeof(struct bpf_##NAME##_state), \
655 memcpy(new_##FIELD, state->FIELD, \
656 sizeof(*new_##FIELD) * (old_size / SIZE)); \
657 memset(new_##FIELD + old_size / SIZE, 0, \
658 sizeof(*new_##FIELD) * (size - old_size) / SIZE); \
660 state->COUNT = slot * SIZE; \
661 kfree(state->FIELD); \
662 state->FIELD = new_##FIELD; \
665 /* realloc_reference_state() */
666 REALLOC_STATE_FN(reference, acquired_refs, refs, 1)
667 /* realloc_stack_state() */
668 REALLOC_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
669 #undef REALLOC_STATE_FN
671 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
672 * make it consume minimal amount of memory. check_stack_write() access from
673 * the program calls into realloc_func_state() to grow the stack size.
674 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
675 * which realloc_stack_state() copies over. It points to previous
676 * bpf_verifier_state which is never reallocated.
678 static int realloc_func_state(struct bpf_func_state *state, int stack_size,
679 int refs_size, bool copy_old)
681 int err = realloc_reference_state(state, refs_size, copy_old);
684 return realloc_stack_state(state, stack_size, copy_old);
687 /* Acquire a pointer id from the env and update the state->refs to include
688 * this new pointer reference.
689 * On success, returns a valid pointer id to associate with the register
690 * On failure, returns a negative errno.
692 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
694 struct bpf_func_state *state = cur_func(env);
695 int new_ofs = state->acquired_refs;
698 err = realloc_reference_state(state, state->acquired_refs + 1, true);
702 state->refs[new_ofs].id = id;
703 state->refs[new_ofs].insn_idx = insn_idx;
708 /* release function corresponding to acquire_reference_state(). Idempotent. */
709 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
713 last_idx = state->acquired_refs - 1;
714 for (i = 0; i < state->acquired_refs; i++) {
715 if (state->refs[i].id == ptr_id) {
716 if (last_idx && i != last_idx)
717 memcpy(&state->refs[i], &state->refs[last_idx],
718 sizeof(*state->refs));
719 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
720 state->acquired_refs--;
727 static int transfer_reference_state(struct bpf_func_state *dst,
728 struct bpf_func_state *src)
730 int err = realloc_reference_state(dst, src->acquired_refs, false);
733 err = copy_reference_state(dst, src);
739 static void free_func_state(struct bpf_func_state *state)
748 static void clear_jmp_history(struct bpf_verifier_state *state)
750 kfree(state->jmp_history);
751 state->jmp_history = NULL;
752 state->jmp_history_cnt = 0;
755 static void free_verifier_state(struct bpf_verifier_state *state,
760 for (i = 0; i <= state->curframe; i++) {
761 free_func_state(state->frame[i]);
762 state->frame[i] = NULL;
764 clear_jmp_history(state);
769 /* copy verifier state from src to dst growing dst stack space
770 * when necessary to accommodate larger src stack
772 static int copy_func_state(struct bpf_func_state *dst,
773 const struct bpf_func_state *src)
777 err = realloc_func_state(dst, src->allocated_stack, src->acquired_refs,
781 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
782 err = copy_reference_state(dst, src);
785 return copy_stack_state(dst, src);
788 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
789 const struct bpf_verifier_state *src)
791 struct bpf_func_state *dst;
792 u32 jmp_sz = sizeof(struct bpf_idx_pair) * src->jmp_history_cnt;
795 if (dst_state->jmp_history_cnt < src->jmp_history_cnt) {
796 kfree(dst_state->jmp_history);
797 dst_state->jmp_history = kmalloc(jmp_sz, GFP_USER);
798 if (!dst_state->jmp_history)
801 memcpy(dst_state->jmp_history, src->jmp_history, jmp_sz);
802 dst_state->jmp_history_cnt = src->jmp_history_cnt;
804 /* if dst has more stack frames then src frame, free them */
805 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
806 free_func_state(dst_state->frame[i]);
807 dst_state->frame[i] = NULL;
809 dst_state->speculative = src->speculative;
810 dst_state->curframe = src->curframe;
811 dst_state->active_spin_lock = src->active_spin_lock;
812 dst_state->branches = src->branches;
813 dst_state->parent = src->parent;
814 dst_state->first_insn_idx = src->first_insn_idx;
815 dst_state->last_insn_idx = src->last_insn_idx;
816 for (i = 0; i <= src->curframe; i++) {
817 dst = dst_state->frame[i];
819 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
822 dst_state->frame[i] = dst;
824 err = copy_func_state(dst, src->frame[i]);
831 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
834 u32 br = --st->branches;
836 /* WARN_ON(br > 1) technically makes sense here,
837 * but see comment in push_stack(), hence:
839 WARN_ONCE((int)br < 0,
840 "BUG update_branch_counts:branches_to_explore=%d\n",
848 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
851 struct bpf_verifier_state *cur = env->cur_state;
852 struct bpf_verifier_stack_elem *elem, *head = env->head;
855 if (env->head == NULL)
859 err = copy_verifier_state(cur, &head->st);
864 *insn_idx = head->insn_idx;
866 *prev_insn_idx = head->prev_insn_idx;
868 free_verifier_state(&head->st, false);
875 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
876 int insn_idx, int prev_insn_idx,
879 struct bpf_verifier_state *cur = env->cur_state;
880 struct bpf_verifier_stack_elem *elem;
883 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
887 elem->insn_idx = insn_idx;
888 elem->prev_insn_idx = prev_insn_idx;
889 elem->next = env->head;
892 err = copy_verifier_state(&elem->st, cur);
895 elem->st.speculative |= speculative;
896 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
897 verbose(env, "The sequence of %d jumps is too complex.\n",
901 if (elem->st.parent) {
902 ++elem->st.parent->branches;
903 /* WARN_ON(branches > 2) technically makes sense here,
905 * 1. speculative states will bump 'branches' for non-branch
907 * 2. is_state_visited() heuristics may decide not to create
908 * a new state for a sequence of branches and all such current
909 * and cloned states will be pointing to a single parent state
910 * which might have large 'branches' count.
915 free_verifier_state(env->cur_state, true);
916 env->cur_state = NULL;
917 /* pop all elements and return */
918 while (!pop_stack(env, NULL, NULL));
922 #define CALLER_SAVED_REGS 6
923 static const int caller_saved[CALLER_SAVED_REGS] = {
924 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
927 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
928 struct bpf_reg_state *reg);
930 /* Mark the unknown part of a register (variable offset or scalar value) as
931 * known to have the value @imm.
933 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
935 /* Clear id, off, and union(map_ptr, range) */
936 memset(((u8 *)reg) + sizeof(reg->type), 0,
937 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
938 reg->var_off = tnum_const(imm);
939 reg->smin_value = (s64)imm;
940 reg->smax_value = (s64)imm;
941 reg->umin_value = imm;
942 reg->umax_value = imm;
944 reg->s32_min_value = (s32)imm;
945 reg->s32_max_value = (s32)imm;
946 reg->u32_min_value = (u32)imm;
947 reg->u32_max_value = (u32)imm;
950 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
952 reg->var_off = tnum_const_subreg(reg->var_off, imm);
953 reg->s32_min_value = (s32)imm;
954 reg->s32_max_value = (s32)imm;
955 reg->u32_min_value = (u32)imm;
956 reg->u32_max_value = (u32)imm;
959 /* Mark the 'variable offset' part of a register as zero. This should be
960 * used only on registers holding a pointer type.
962 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
964 __mark_reg_known(reg, 0);
967 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
969 __mark_reg_known(reg, 0);
970 reg->type = SCALAR_VALUE;
973 static void mark_reg_known_zero(struct bpf_verifier_env *env,
974 struct bpf_reg_state *regs, u32 regno)
976 if (WARN_ON(regno >= MAX_BPF_REG)) {
977 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
978 /* Something bad happened, let's kill all regs */
979 for (regno = 0; regno < MAX_BPF_REG; regno++)
980 __mark_reg_not_init(env, regs + regno);
983 __mark_reg_known_zero(regs + regno);
986 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
988 return type_is_pkt_pointer(reg->type);
991 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
993 return reg_is_pkt_pointer(reg) ||
994 reg->type == PTR_TO_PACKET_END;
997 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
998 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
999 enum bpf_reg_type which)
1001 /* The register can already have a range from prior markings.
1002 * This is fine as long as it hasn't been advanced from its
1005 return reg->type == which &&
1008 tnum_equals_const(reg->var_off, 0);
1011 /* Reset the min/max bounds of a register */
1012 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1014 reg->smin_value = S64_MIN;
1015 reg->smax_value = S64_MAX;
1016 reg->umin_value = 0;
1017 reg->umax_value = U64_MAX;
1019 reg->s32_min_value = S32_MIN;
1020 reg->s32_max_value = S32_MAX;
1021 reg->u32_min_value = 0;
1022 reg->u32_max_value = U32_MAX;
1025 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
1027 reg->smin_value = S64_MIN;
1028 reg->smax_value = S64_MAX;
1029 reg->umin_value = 0;
1030 reg->umax_value = U64_MAX;
1033 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
1035 reg->s32_min_value = S32_MIN;
1036 reg->s32_max_value = S32_MAX;
1037 reg->u32_min_value = 0;
1038 reg->u32_max_value = U32_MAX;
1041 static void __update_reg32_bounds(struct bpf_reg_state *reg)
1043 struct tnum var32_off = tnum_subreg(reg->var_off);
1045 /* min signed is max(sign bit) | min(other bits) */
1046 reg->s32_min_value = max_t(s32, reg->s32_min_value,
1047 var32_off.value | (var32_off.mask & S32_MIN));
1048 /* max signed is min(sign bit) | max(other bits) */
1049 reg->s32_max_value = min_t(s32, reg->s32_max_value,
1050 var32_off.value | (var32_off.mask & S32_MAX));
1051 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
1052 reg->u32_max_value = min(reg->u32_max_value,
1053 (u32)(var32_off.value | var32_off.mask));
1056 static void __update_reg64_bounds(struct bpf_reg_state *reg)
1058 /* min signed is max(sign bit) | min(other bits) */
1059 reg->smin_value = max_t(s64, reg->smin_value,
1060 reg->var_off.value | (reg->var_off.mask & S64_MIN));
1061 /* max signed is min(sign bit) | max(other bits) */
1062 reg->smax_value = min_t(s64, reg->smax_value,
1063 reg->var_off.value | (reg->var_off.mask & S64_MAX));
1064 reg->umin_value = max(reg->umin_value, reg->var_off.value);
1065 reg->umax_value = min(reg->umax_value,
1066 reg->var_off.value | reg->var_off.mask);
1069 static void __update_reg_bounds(struct bpf_reg_state *reg)
1071 __update_reg32_bounds(reg);
1072 __update_reg64_bounds(reg);
1075 /* Uses signed min/max values to inform unsigned, and vice-versa */
1076 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
1078 /* Learn sign from signed bounds.
1079 * If we cannot cross the sign boundary, then signed and unsigned bounds
1080 * are the same, so combine. This works even in the negative case, e.g.
1081 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1083 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
1084 reg->s32_min_value = reg->u32_min_value =
1085 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1086 reg->s32_max_value = reg->u32_max_value =
1087 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1090 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1091 * boundary, so we must be careful.
1093 if ((s32)reg->u32_max_value >= 0) {
1094 /* Positive. We can't learn anything from the smin, but smax
1095 * is positive, hence safe.
1097 reg->s32_min_value = reg->u32_min_value;
1098 reg->s32_max_value = reg->u32_max_value =
1099 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1100 } else if ((s32)reg->u32_min_value < 0) {
1101 /* Negative. We can't learn anything from the smax, but smin
1102 * is negative, hence safe.
1104 reg->s32_min_value = reg->u32_min_value =
1105 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1106 reg->s32_max_value = reg->u32_max_value;
1110 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
1112 /* Learn sign from signed bounds.
1113 * If we cannot cross the sign boundary, then signed and unsigned bounds
1114 * are the same, so combine. This works even in the negative case, e.g.
1115 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1117 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1118 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1120 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1124 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1125 * boundary, so we must be careful.
1127 if ((s64)reg->umax_value >= 0) {
1128 /* Positive. We can't learn anything from the smin, but smax
1129 * is positive, hence safe.
1131 reg->smin_value = reg->umin_value;
1132 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1134 } else if ((s64)reg->umin_value < 0) {
1135 /* Negative. We can't learn anything from the smax, but smin
1136 * is negative, hence safe.
1138 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1140 reg->smax_value = reg->umax_value;
1144 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
1146 __reg32_deduce_bounds(reg);
1147 __reg64_deduce_bounds(reg);
1150 /* Attempts to improve var_off based on unsigned min/max information */
1151 static void __reg_bound_offset(struct bpf_reg_state *reg)
1153 struct tnum var64_off = tnum_intersect(reg->var_off,
1154 tnum_range(reg->umin_value,
1156 struct tnum var32_off = tnum_intersect(tnum_subreg(reg->var_off),
1157 tnum_range(reg->u32_min_value,
1158 reg->u32_max_value));
1160 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
1163 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
1165 reg->umin_value = reg->u32_min_value;
1166 reg->umax_value = reg->u32_max_value;
1167 /* Attempt to pull 32-bit signed bounds into 64-bit bounds
1168 * but must be positive otherwise set to worse case bounds
1169 * and refine later from tnum.
1171 if (reg->s32_min_value > 0)
1172 reg->smin_value = reg->s32_min_value;
1174 reg->smin_value = 0;
1175 if (reg->s32_max_value > 0)
1176 reg->smax_value = reg->s32_max_value;
1178 reg->smax_value = U32_MAX;
1181 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
1183 /* special case when 64-bit register has upper 32-bit register
1184 * zeroed. Typically happens after zext or <<32, >>32 sequence
1185 * allowing us to use 32-bit bounds directly,
1187 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
1188 __reg_assign_32_into_64(reg);
1190 /* Otherwise the best we can do is push lower 32bit known and
1191 * unknown bits into register (var_off set from jmp logic)
1192 * then learn as much as possible from the 64-bit tnum
1193 * known and unknown bits. The previous smin/smax bounds are
1194 * invalid here because of jmp32 compare so mark them unknown
1195 * so they do not impact tnum bounds calculation.
1197 __mark_reg64_unbounded(reg);
1198 __update_reg_bounds(reg);
1201 /* Intersecting with the old var_off might have improved our bounds
1202 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1203 * then new var_off is (0; 0x7f...fc) which improves our umax.
1205 __reg_deduce_bounds(reg);
1206 __reg_bound_offset(reg);
1207 __update_reg_bounds(reg);
1210 static bool __reg64_bound_s32(s64 a)
1212 if (a > S32_MIN && a < S32_MAX)
1217 static bool __reg64_bound_u32(u64 a)
1219 if (a > U32_MIN && a < U32_MAX)
1224 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
1226 __mark_reg32_unbounded(reg);
1228 if (__reg64_bound_s32(reg->smin_value))
1229 reg->s32_min_value = (s32)reg->smin_value;
1230 if (__reg64_bound_s32(reg->smax_value))
1231 reg->s32_max_value = (s32)reg->smax_value;
1232 if (__reg64_bound_u32(reg->umin_value))
1233 reg->u32_min_value = (u32)reg->umin_value;
1234 if (__reg64_bound_u32(reg->umax_value))
1235 reg->u32_max_value = (u32)reg->umax_value;
1237 /* Intersecting with the old var_off might have improved our bounds
1238 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1239 * then new var_off is (0; 0x7f...fc) which improves our umax.
1241 __reg_deduce_bounds(reg);
1242 __reg_bound_offset(reg);
1243 __update_reg_bounds(reg);
1246 /* Mark a register as having a completely unknown (scalar) value. */
1247 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1248 struct bpf_reg_state *reg)
1251 * Clear type, id, off, and union(map_ptr, range) and
1252 * padding between 'type' and union
1254 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1255 reg->type = SCALAR_VALUE;
1256 reg->var_off = tnum_unknown;
1258 reg->precise = env->subprog_cnt > 1 || !env->allow_ptr_leaks ?
1260 __mark_reg_unbounded(reg);
1263 static void mark_reg_unknown(struct bpf_verifier_env *env,
1264 struct bpf_reg_state *regs, u32 regno)
1266 if (WARN_ON(regno >= MAX_BPF_REG)) {
1267 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1268 /* Something bad happened, let's kill all regs except FP */
1269 for (regno = 0; regno < BPF_REG_FP; regno++)
1270 __mark_reg_not_init(env, regs + regno);
1273 __mark_reg_unknown(env, regs + regno);
1276 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1277 struct bpf_reg_state *reg)
1279 __mark_reg_unknown(env, reg);
1280 reg->type = NOT_INIT;
1283 static void mark_reg_not_init(struct bpf_verifier_env *env,
1284 struct bpf_reg_state *regs, u32 regno)
1286 if (WARN_ON(regno >= MAX_BPF_REG)) {
1287 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1288 /* Something bad happened, let's kill all regs except FP */
1289 for (regno = 0; regno < BPF_REG_FP; regno++)
1290 __mark_reg_not_init(env, regs + regno);
1293 __mark_reg_not_init(env, regs + regno);
1296 #define DEF_NOT_SUBREG (0)
1297 static void init_reg_state(struct bpf_verifier_env *env,
1298 struct bpf_func_state *state)
1300 struct bpf_reg_state *regs = state->regs;
1303 for (i = 0; i < MAX_BPF_REG; i++) {
1304 mark_reg_not_init(env, regs, i);
1305 regs[i].live = REG_LIVE_NONE;
1306 regs[i].parent = NULL;
1307 regs[i].subreg_def = DEF_NOT_SUBREG;
1311 regs[BPF_REG_FP].type = PTR_TO_STACK;
1312 mark_reg_known_zero(env, regs, BPF_REG_FP);
1313 regs[BPF_REG_FP].frameno = state->frameno;
1316 #define BPF_MAIN_FUNC (-1)
1317 static void init_func_state(struct bpf_verifier_env *env,
1318 struct bpf_func_state *state,
1319 int callsite, int frameno, int subprogno)
1321 state->callsite = callsite;
1322 state->frameno = frameno;
1323 state->subprogno = subprogno;
1324 init_reg_state(env, state);
1328 SRC_OP, /* register is used as source operand */
1329 DST_OP, /* register is used as destination operand */
1330 DST_OP_NO_MARK /* same as above, check only, don't mark */
1333 static int cmp_subprogs(const void *a, const void *b)
1335 return ((struct bpf_subprog_info *)a)->start -
1336 ((struct bpf_subprog_info *)b)->start;
1339 static int find_subprog(struct bpf_verifier_env *env, int off)
1341 struct bpf_subprog_info *p;
1343 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1344 sizeof(env->subprog_info[0]), cmp_subprogs);
1347 return p - env->subprog_info;
1351 static int add_subprog(struct bpf_verifier_env *env, int off)
1353 int insn_cnt = env->prog->len;
1356 if (off >= insn_cnt || off < 0) {
1357 verbose(env, "call to invalid destination\n");
1360 ret = find_subprog(env, off);
1363 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1364 verbose(env, "too many subprograms\n");
1367 env->subprog_info[env->subprog_cnt++].start = off;
1368 sort(env->subprog_info, env->subprog_cnt,
1369 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1373 static int check_subprogs(struct bpf_verifier_env *env)
1375 int i, ret, subprog_start, subprog_end, off, cur_subprog = 0;
1376 struct bpf_subprog_info *subprog = env->subprog_info;
1377 struct bpf_insn *insn = env->prog->insnsi;
1378 int insn_cnt = env->prog->len;
1380 /* Add entry function. */
1381 ret = add_subprog(env, 0);
1385 /* determine subprog starts. The end is one before the next starts */
1386 for (i = 0; i < insn_cnt; i++) {
1387 if (insn[i].code != (BPF_JMP | BPF_CALL))
1389 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1391 if (!env->allow_ptr_leaks) {
1392 verbose(env, "function calls to other bpf functions are allowed for root only\n");
1395 ret = add_subprog(env, i + insn[i].imm + 1);
1400 /* Add a fake 'exit' subprog which could simplify subprog iteration
1401 * logic. 'subprog_cnt' should not be increased.
1403 subprog[env->subprog_cnt].start = insn_cnt;
1405 if (env->log.level & BPF_LOG_LEVEL2)
1406 for (i = 0; i < env->subprog_cnt; i++)
1407 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1409 /* now check that all jumps are within the same subprog */
1410 subprog_start = subprog[cur_subprog].start;
1411 subprog_end = subprog[cur_subprog + 1].start;
1412 for (i = 0; i < insn_cnt; i++) {
1413 u8 code = insn[i].code;
1415 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
1417 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1419 off = i + insn[i].off + 1;
1420 if (off < subprog_start || off >= subprog_end) {
1421 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1425 if (i == subprog_end - 1) {
1426 /* to avoid fall-through from one subprog into another
1427 * the last insn of the subprog should be either exit
1428 * or unconditional jump back
1430 if (code != (BPF_JMP | BPF_EXIT) &&
1431 code != (BPF_JMP | BPF_JA)) {
1432 verbose(env, "last insn is not an exit or jmp\n");
1435 subprog_start = subprog_end;
1437 if (cur_subprog < env->subprog_cnt)
1438 subprog_end = subprog[cur_subprog + 1].start;
1444 /* Parentage chain of this register (or stack slot) should take care of all
1445 * issues like callee-saved registers, stack slot allocation time, etc.
1447 static int mark_reg_read(struct bpf_verifier_env *env,
1448 const struct bpf_reg_state *state,
1449 struct bpf_reg_state *parent, u8 flag)
1451 bool writes = parent == state->parent; /* Observe write marks */
1455 /* if read wasn't screened by an earlier write ... */
1456 if (writes && state->live & REG_LIVE_WRITTEN)
1458 if (parent->live & REG_LIVE_DONE) {
1459 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1460 reg_type_str[parent->type],
1461 parent->var_off.value, parent->off);
1464 /* The first condition is more likely to be true than the
1465 * second, checked it first.
1467 if ((parent->live & REG_LIVE_READ) == flag ||
1468 parent->live & REG_LIVE_READ64)
1469 /* The parentage chain never changes and
1470 * this parent was already marked as LIVE_READ.
1471 * There is no need to keep walking the chain again and
1472 * keep re-marking all parents as LIVE_READ.
1473 * This case happens when the same register is read
1474 * multiple times without writes into it in-between.
1475 * Also, if parent has the stronger REG_LIVE_READ64 set,
1476 * then no need to set the weak REG_LIVE_READ32.
1479 /* ... then we depend on parent's value */
1480 parent->live |= flag;
1481 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
1482 if (flag == REG_LIVE_READ64)
1483 parent->live &= ~REG_LIVE_READ32;
1485 parent = state->parent;
1490 if (env->longest_mark_read_walk < cnt)
1491 env->longest_mark_read_walk = cnt;
1495 /* This function is supposed to be used by the following 32-bit optimization
1496 * code only. It returns TRUE if the source or destination register operates
1497 * on 64-bit, otherwise return FALSE.
1499 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
1500 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
1505 class = BPF_CLASS(code);
1507 if (class == BPF_JMP) {
1508 /* BPF_EXIT for "main" will reach here. Return TRUE
1513 if (op == BPF_CALL) {
1514 /* BPF to BPF call will reach here because of marking
1515 * caller saved clobber with DST_OP_NO_MARK for which we
1516 * don't care the register def because they are anyway
1517 * marked as NOT_INIT already.
1519 if (insn->src_reg == BPF_PSEUDO_CALL)
1521 /* Helper call will reach here because of arg type
1522 * check, conservatively return TRUE.
1531 if (class == BPF_ALU64 || class == BPF_JMP ||
1532 /* BPF_END always use BPF_ALU class. */
1533 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
1536 if (class == BPF_ALU || class == BPF_JMP32)
1539 if (class == BPF_LDX) {
1541 return BPF_SIZE(code) == BPF_DW;
1542 /* LDX source must be ptr. */
1546 if (class == BPF_STX) {
1547 if (reg->type != SCALAR_VALUE)
1549 return BPF_SIZE(code) == BPF_DW;
1552 if (class == BPF_LD) {
1553 u8 mode = BPF_MODE(code);
1556 if (mode == BPF_IMM)
1559 /* Both LD_IND and LD_ABS return 32-bit data. */
1563 /* Implicit ctx ptr. */
1564 if (regno == BPF_REG_6)
1567 /* Explicit source could be any width. */
1571 if (class == BPF_ST)
1572 /* The only source register for BPF_ST is a ptr. */
1575 /* Conservatively return true at default. */
1579 /* Return TRUE if INSN doesn't have explicit value define. */
1580 static bool insn_no_def(struct bpf_insn *insn)
1582 u8 class = BPF_CLASS(insn->code);
1584 return (class == BPF_JMP || class == BPF_JMP32 ||
1585 class == BPF_STX || class == BPF_ST);
1588 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
1589 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
1591 if (insn_no_def(insn))
1594 return !is_reg64(env, insn, insn->dst_reg, NULL, DST_OP);
1597 static void mark_insn_zext(struct bpf_verifier_env *env,
1598 struct bpf_reg_state *reg)
1600 s32 def_idx = reg->subreg_def;
1602 if (def_idx == DEF_NOT_SUBREG)
1605 env->insn_aux_data[def_idx - 1].zext_dst = true;
1606 /* The dst will be zero extended, so won't be sub-register anymore. */
1607 reg->subreg_def = DEF_NOT_SUBREG;
1610 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
1611 enum reg_arg_type t)
1613 struct bpf_verifier_state *vstate = env->cur_state;
1614 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1615 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
1616 struct bpf_reg_state *reg, *regs = state->regs;
1619 if (regno >= MAX_BPF_REG) {
1620 verbose(env, "R%d is invalid\n", regno);
1625 rw64 = is_reg64(env, insn, regno, reg, t);
1627 /* check whether register used as source operand can be read */
1628 if (reg->type == NOT_INIT) {
1629 verbose(env, "R%d !read_ok\n", regno);
1632 /* We don't need to worry about FP liveness because it's read-only */
1633 if (regno == BPF_REG_FP)
1637 mark_insn_zext(env, reg);
1639 return mark_reg_read(env, reg, reg->parent,
1640 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
1642 /* check whether register used as dest operand can be written to */
1643 if (regno == BPF_REG_FP) {
1644 verbose(env, "frame pointer is read only\n");
1647 reg->live |= REG_LIVE_WRITTEN;
1648 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
1650 mark_reg_unknown(env, regs, regno);
1655 /* for any branch, call, exit record the history of jmps in the given state */
1656 static int push_jmp_history(struct bpf_verifier_env *env,
1657 struct bpf_verifier_state *cur)
1659 u32 cnt = cur->jmp_history_cnt;
1660 struct bpf_idx_pair *p;
1663 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
1666 p[cnt - 1].idx = env->insn_idx;
1667 p[cnt - 1].prev_idx = env->prev_insn_idx;
1668 cur->jmp_history = p;
1669 cur->jmp_history_cnt = cnt;
1673 /* Backtrack one insn at a time. If idx is not at the top of recorded
1674 * history then previous instruction came from straight line execution.
1676 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
1681 if (cnt && st->jmp_history[cnt - 1].idx == i) {
1682 i = st->jmp_history[cnt - 1].prev_idx;
1690 /* For given verifier state backtrack_insn() is called from the last insn to
1691 * the first insn. Its purpose is to compute a bitmask of registers and
1692 * stack slots that needs precision in the parent verifier state.
1694 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
1695 u32 *reg_mask, u64 *stack_mask)
1697 const struct bpf_insn_cbs cbs = {
1698 .cb_print = verbose,
1699 .private_data = env,
1701 struct bpf_insn *insn = env->prog->insnsi + idx;
1702 u8 class = BPF_CLASS(insn->code);
1703 u8 opcode = BPF_OP(insn->code);
1704 u8 mode = BPF_MODE(insn->code);
1705 u32 dreg = 1u << insn->dst_reg;
1706 u32 sreg = 1u << insn->src_reg;
1709 if (insn->code == 0)
1711 if (env->log.level & BPF_LOG_LEVEL) {
1712 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
1713 verbose(env, "%d: ", idx);
1714 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
1717 if (class == BPF_ALU || class == BPF_ALU64) {
1718 if (!(*reg_mask & dreg))
1720 if (opcode == BPF_MOV) {
1721 if (BPF_SRC(insn->code) == BPF_X) {
1723 * dreg needs precision after this insn
1724 * sreg needs precision before this insn
1730 * dreg needs precision after this insn.
1731 * Corresponding register is already marked
1732 * as precise=true in this verifier state.
1733 * No further markings in parent are necessary
1738 if (BPF_SRC(insn->code) == BPF_X) {
1740 * both dreg and sreg need precision
1745 * dreg still needs precision before this insn
1748 } else if (class == BPF_LDX) {
1749 if (!(*reg_mask & dreg))
1753 /* scalars can only be spilled into stack w/o losing precision.
1754 * Load from any other memory can be zero extended.
1755 * The desire to keep that precision is already indicated
1756 * by 'precise' mark in corresponding register of this state.
1757 * No further tracking necessary.
1759 if (insn->src_reg != BPF_REG_FP)
1761 if (BPF_SIZE(insn->code) != BPF_DW)
1764 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
1765 * that [fp - off] slot contains scalar that needs to be
1766 * tracked with precision
1768 spi = (-insn->off - 1) / BPF_REG_SIZE;
1770 verbose(env, "BUG spi %d\n", spi);
1771 WARN_ONCE(1, "verifier backtracking bug");
1774 *stack_mask |= 1ull << spi;
1775 } else if (class == BPF_STX || class == BPF_ST) {
1776 if (*reg_mask & dreg)
1777 /* stx & st shouldn't be using _scalar_ dst_reg
1778 * to access memory. It means backtracking
1779 * encountered a case of pointer subtraction.
1782 /* scalars can only be spilled into stack */
1783 if (insn->dst_reg != BPF_REG_FP)
1785 if (BPF_SIZE(insn->code) != BPF_DW)
1787 spi = (-insn->off - 1) / BPF_REG_SIZE;
1789 verbose(env, "BUG spi %d\n", spi);
1790 WARN_ONCE(1, "verifier backtracking bug");
1793 if (!(*stack_mask & (1ull << spi)))
1795 *stack_mask &= ~(1ull << spi);
1796 if (class == BPF_STX)
1798 } else if (class == BPF_JMP || class == BPF_JMP32) {
1799 if (opcode == BPF_CALL) {
1800 if (insn->src_reg == BPF_PSEUDO_CALL)
1802 /* regular helper call sets R0 */
1804 if (*reg_mask & 0x3f) {
1805 /* if backtracing was looking for registers R1-R5
1806 * they should have been found already.
1808 verbose(env, "BUG regs %x\n", *reg_mask);
1809 WARN_ONCE(1, "verifier backtracking bug");
1812 } else if (opcode == BPF_EXIT) {
1815 } else if (class == BPF_LD) {
1816 if (!(*reg_mask & dreg))
1819 /* It's ld_imm64 or ld_abs or ld_ind.
1820 * For ld_imm64 no further tracking of precision
1821 * into parent is necessary
1823 if (mode == BPF_IND || mode == BPF_ABS)
1824 /* to be analyzed */
1830 /* the scalar precision tracking algorithm:
1831 * . at the start all registers have precise=false.
1832 * . scalar ranges are tracked as normal through alu and jmp insns.
1833 * . once precise value of the scalar register is used in:
1834 * . ptr + scalar alu
1835 * . if (scalar cond K|scalar)
1836 * . helper_call(.., scalar, ...) where ARG_CONST is expected
1837 * backtrack through the verifier states and mark all registers and
1838 * stack slots with spilled constants that these scalar regisers
1839 * should be precise.
1840 * . during state pruning two registers (or spilled stack slots)
1841 * are equivalent if both are not precise.
1843 * Note the verifier cannot simply walk register parentage chain,
1844 * since many different registers and stack slots could have been
1845 * used to compute single precise scalar.
1847 * The approach of starting with precise=true for all registers and then
1848 * backtrack to mark a register as not precise when the verifier detects
1849 * that program doesn't care about specific value (e.g., when helper
1850 * takes register as ARG_ANYTHING parameter) is not safe.
1852 * It's ok to walk single parentage chain of the verifier states.
1853 * It's possible that this backtracking will go all the way till 1st insn.
1854 * All other branches will be explored for needing precision later.
1856 * The backtracking needs to deal with cases like:
1857 * 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)
1860 * if r5 > 0x79f goto pc+7
1861 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
1864 * call bpf_perf_event_output#25
1865 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
1869 * call foo // uses callee's r6 inside to compute r0
1873 * to track above reg_mask/stack_mask needs to be independent for each frame.
1875 * Also if parent's curframe > frame where backtracking started,
1876 * the verifier need to mark registers in both frames, otherwise callees
1877 * may incorrectly prune callers. This is similar to
1878 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
1880 * For now backtracking falls back into conservative marking.
1882 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
1883 struct bpf_verifier_state *st)
1885 struct bpf_func_state *func;
1886 struct bpf_reg_state *reg;
1889 /* big hammer: mark all scalars precise in this path.
1890 * pop_stack may still get !precise scalars.
1892 for (; st; st = st->parent)
1893 for (i = 0; i <= st->curframe; i++) {
1894 func = st->frame[i];
1895 for (j = 0; j < BPF_REG_FP; j++) {
1896 reg = &func->regs[j];
1897 if (reg->type != SCALAR_VALUE)
1899 reg->precise = true;
1901 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
1902 if (func->stack[j].slot_type[0] != STACK_SPILL)
1904 reg = &func->stack[j].spilled_ptr;
1905 if (reg->type != SCALAR_VALUE)
1907 reg->precise = true;
1912 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
1915 struct bpf_verifier_state *st = env->cur_state;
1916 int first_idx = st->first_insn_idx;
1917 int last_idx = env->insn_idx;
1918 struct bpf_func_state *func;
1919 struct bpf_reg_state *reg;
1920 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
1921 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
1922 bool skip_first = true;
1923 bool new_marks = false;
1926 if (!env->allow_ptr_leaks)
1927 /* backtracking is root only for now */
1930 func = st->frame[st->curframe];
1932 reg = &func->regs[regno];
1933 if (reg->type != SCALAR_VALUE) {
1934 WARN_ONCE(1, "backtracing misuse");
1941 reg->precise = true;
1945 if (func->stack[spi].slot_type[0] != STACK_SPILL) {
1949 reg = &func->stack[spi].spilled_ptr;
1950 if (reg->type != SCALAR_VALUE) {
1958 reg->precise = true;
1964 if (!reg_mask && !stack_mask)
1967 DECLARE_BITMAP(mask, 64);
1968 u32 history = st->jmp_history_cnt;
1970 if (env->log.level & BPF_LOG_LEVEL)
1971 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
1972 for (i = last_idx;;) {
1977 err = backtrack_insn(env, i, ®_mask, &stack_mask);
1979 if (err == -ENOTSUPP) {
1980 mark_all_scalars_precise(env, st);
1985 if (!reg_mask && !stack_mask)
1986 /* Found assignment(s) into tracked register in this state.
1987 * Since this state is already marked, just return.
1988 * Nothing to be tracked further in the parent state.
1993 i = get_prev_insn_idx(st, i, &history);
1994 if (i >= env->prog->len) {
1995 /* This can happen if backtracking reached insn 0
1996 * and there are still reg_mask or stack_mask
1998 * It means the backtracking missed the spot where
1999 * particular register was initialized with a constant.
2001 verbose(env, "BUG backtracking idx %d\n", i);
2002 WARN_ONCE(1, "verifier backtracking bug");
2011 func = st->frame[st->curframe];
2012 bitmap_from_u64(mask, reg_mask);
2013 for_each_set_bit(i, mask, 32) {
2014 reg = &func->regs[i];
2015 if (reg->type != SCALAR_VALUE) {
2016 reg_mask &= ~(1u << i);
2021 reg->precise = true;
2024 bitmap_from_u64(mask, stack_mask);
2025 for_each_set_bit(i, mask, 64) {
2026 if (i >= func->allocated_stack / BPF_REG_SIZE) {
2027 /* the sequence of instructions:
2029 * 3: (7b) *(u64 *)(r3 -8) = r0
2030 * 4: (79) r4 = *(u64 *)(r10 -8)
2031 * doesn't contain jmps. It's backtracked
2032 * as a single block.
2033 * During backtracking insn 3 is not recognized as
2034 * stack access, so at the end of backtracking
2035 * stack slot fp-8 is still marked in stack_mask.
2036 * However the parent state may not have accessed
2037 * fp-8 and it's "unallocated" stack space.
2038 * In such case fallback to conservative.
2040 mark_all_scalars_precise(env, st);
2044 if (func->stack[i].slot_type[0] != STACK_SPILL) {
2045 stack_mask &= ~(1ull << i);
2048 reg = &func->stack[i].spilled_ptr;
2049 if (reg->type != SCALAR_VALUE) {
2050 stack_mask &= ~(1ull << i);
2055 reg->precise = true;
2057 if (env->log.level & BPF_LOG_LEVEL) {
2058 print_verifier_state(env, func);
2059 verbose(env, "parent %s regs=%x stack=%llx marks\n",
2060 new_marks ? "didn't have" : "already had",
2061 reg_mask, stack_mask);
2064 if (!reg_mask && !stack_mask)
2069 last_idx = st->last_insn_idx;
2070 first_idx = st->first_insn_idx;
2075 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
2077 return __mark_chain_precision(env, regno, -1);
2080 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
2082 return __mark_chain_precision(env, -1, spi);
2085 static bool is_spillable_regtype(enum bpf_reg_type type)
2088 case PTR_TO_MAP_VALUE:
2089 case PTR_TO_MAP_VALUE_OR_NULL:
2093 case PTR_TO_PACKET_META:
2094 case PTR_TO_PACKET_END:
2095 case PTR_TO_FLOW_KEYS:
2096 case CONST_PTR_TO_MAP:
2098 case PTR_TO_SOCKET_OR_NULL:
2099 case PTR_TO_SOCK_COMMON:
2100 case PTR_TO_SOCK_COMMON_OR_NULL:
2101 case PTR_TO_TCP_SOCK:
2102 case PTR_TO_TCP_SOCK_OR_NULL:
2103 case PTR_TO_XDP_SOCK:
2111 /* Does this register contain a constant zero? */
2112 static bool register_is_null(struct bpf_reg_state *reg)
2114 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
2117 static bool register_is_const(struct bpf_reg_state *reg)
2119 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
2122 static void save_register_state(struct bpf_func_state *state,
2123 int spi, struct bpf_reg_state *reg)
2127 state->stack[spi].spilled_ptr = *reg;
2128 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2130 for (i = 0; i < BPF_REG_SIZE; i++)
2131 state->stack[spi].slot_type[i] = STACK_SPILL;
2134 /* check_stack_read/write functions track spill/fill of registers,
2135 * stack boundary and alignment are checked in check_mem_access()
2137 static int check_stack_write(struct bpf_verifier_env *env,
2138 struct bpf_func_state *state, /* func where register points to */
2139 int off, int size, int value_regno, int insn_idx)
2141 struct bpf_func_state *cur; /* state of the current function */
2142 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
2143 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
2144 struct bpf_reg_state *reg = NULL;
2146 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
2147 state->acquired_refs, true);
2150 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
2151 * so it's aligned access and [off, off + size) are within stack limits
2153 if (!env->allow_ptr_leaks &&
2154 state->stack[spi].slot_type[0] == STACK_SPILL &&
2155 size != BPF_REG_SIZE) {
2156 verbose(env, "attempt to corrupt spilled pointer on stack\n");
2160 cur = env->cur_state->frame[env->cur_state->curframe];
2161 if (value_regno >= 0)
2162 reg = &cur->regs[value_regno];
2164 if (reg && size == BPF_REG_SIZE && register_is_const(reg) &&
2165 !register_is_null(reg) && env->allow_ptr_leaks) {
2166 if (dst_reg != BPF_REG_FP) {
2167 /* The backtracking logic can only recognize explicit
2168 * stack slot address like [fp - 8]. Other spill of
2169 * scalar via different register has to be conervative.
2170 * Backtrack from here and mark all registers as precise
2171 * that contributed into 'reg' being a constant.
2173 err = mark_chain_precision(env, value_regno);
2177 save_register_state(state, spi, reg);
2178 } else if (reg && is_spillable_regtype(reg->type)) {
2179 /* register containing pointer is being spilled into stack */
2180 if (size != BPF_REG_SIZE) {
2181 verbose_linfo(env, insn_idx, "; ");
2182 verbose(env, "invalid size of register spill\n");
2186 if (state != cur && reg->type == PTR_TO_STACK) {
2187 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
2191 if (!env->allow_ptr_leaks) {
2192 bool sanitize = false;
2194 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
2195 register_is_const(&state->stack[spi].spilled_ptr))
2197 for (i = 0; i < BPF_REG_SIZE; i++)
2198 if (state->stack[spi].slot_type[i] == STACK_MISC) {
2203 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
2204 int soff = (-spi - 1) * BPF_REG_SIZE;
2206 /* detected reuse of integer stack slot with a pointer
2207 * which means either llvm is reusing stack slot or
2208 * an attacker is trying to exploit CVE-2018-3639
2209 * (speculative store bypass)
2210 * Have to sanitize that slot with preemptive
2213 if (*poff && *poff != soff) {
2214 /* disallow programs where single insn stores
2215 * into two different stack slots, since verifier
2216 * cannot sanitize them
2219 "insn %d cannot access two stack slots fp%d and fp%d",
2220 insn_idx, *poff, soff);
2226 save_register_state(state, spi, reg);
2228 u8 type = STACK_MISC;
2230 /* regular write of data into stack destroys any spilled ptr */
2231 state->stack[spi].spilled_ptr.type = NOT_INIT;
2232 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
2233 if (state->stack[spi].slot_type[0] == STACK_SPILL)
2234 for (i = 0; i < BPF_REG_SIZE; i++)
2235 state->stack[spi].slot_type[i] = STACK_MISC;
2237 /* only mark the slot as written if all 8 bytes were written
2238 * otherwise read propagation may incorrectly stop too soon
2239 * when stack slots are partially written.
2240 * This heuristic means that read propagation will be
2241 * conservative, since it will add reg_live_read marks
2242 * to stack slots all the way to first state when programs
2243 * writes+reads less than 8 bytes
2245 if (size == BPF_REG_SIZE)
2246 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2248 /* when we zero initialize stack slots mark them as such */
2249 if (reg && register_is_null(reg)) {
2250 /* backtracking doesn't work for STACK_ZERO yet. */
2251 err = mark_chain_precision(env, value_regno);
2257 /* Mark slots affected by this stack write. */
2258 for (i = 0; i < size; i++)
2259 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
2265 static int check_stack_read(struct bpf_verifier_env *env,
2266 struct bpf_func_state *reg_state /* func where register points to */,
2267 int off, int size, int value_regno)
2269 struct bpf_verifier_state *vstate = env->cur_state;
2270 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2271 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
2272 struct bpf_reg_state *reg;
2275 if (reg_state->allocated_stack <= slot) {
2276 verbose(env, "invalid read from stack off %d+0 size %d\n",
2280 stype = reg_state->stack[spi].slot_type;
2281 reg = ®_state->stack[spi].spilled_ptr;
2283 if (stype[0] == STACK_SPILL) {
2284 if (size != BPF_REG_SIZE) {
2285 if (reg->type != SCALAR_VALUE) {
2286 verbose_linfo(env, env->insn_idx, "; ");
2287 verbose(env, "invalid size of register fill\n");
2290 if (value_regno >= 0) {
2291 mark_reg_unknown(env, state->regs, value_regno);
2292 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2294 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2297 for (i = 1; i < BPF_REG_SIZE; i++) {
2298 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
2299 verbose(env, "corrupted spill memory\n");
2304 if (value_regno >= 0) {
2305 /* restore register state from stack */
2306 state->regs[value_regno] = *reg;
2307 /* mark reg as written since spilled pointer state likely
2308 * has its liveness marks cleared by is_state_visited()
2309 * which resets stack/reg liveness for state transitions
2311 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2313 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2317 for (i = 0; i < size; i++) {
2318 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
2320 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
2324 verbose(env, "invalid read from stack off %d+%d size %d\n",
2328 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2329 if (value_regno >= 0) {
2330 if (zeros == size) {
2331 /* any size read into register is zero extended,
2332 * so the whole register == const_zero
2334 __mark_reg_const_zero(&state->regs[value_regno]);
2335 /* backtracking doesn't support STACK_ZERO yet,
2336 * so mark it precise here, so that later
2337 * backtracking can stop here.
2338 * Backtracking may not need this if this register
2339 * doesn't participate in pointer adjustment.
2340 * Forward propagation of precise flag is not
2341 * necessary either. This mark is only to stop
2342 * backtracking. Any register that contributed
2343 * to const 0 was marked precise before spill.
2345 state->regs[value_regno].precise = true;
2347 /* have read misc data from the stack */
2348 mark_reg_unknown(env, state->regs, value_regno);
2350 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2356 static int check_stack_access(struct bpf_verifier_env *env,
2357 const struct bpf_reg_state *reg,
2360 /* Stack accesses must be at a fixed offset, so that we
2361 * can determine what type of data were returned. See
2362 * check_stack_read().
2364 if (!tnum_is_const(reg->var_off)) {
2367 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2368 verbose(env, "variable stack access var_off=%s off=%d size=%d\n",
2373 if (off >= 0 || off < -MAX_BPF_STACK) {
2374 verbose(env, "invalid stack off=%d size=%d\n", off, size);
2381 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
2382 int off, int size, enum bpf_access_type type)
2384 struct bpf_reg_state *regs = cur_regs(env);
2385 struct bpf_map *map = regs[regno].map_ptr;
2386 u32 cap = bpf_map_flags_to_cap(map);
2388 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
2389 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
2390 map->value_size, off, size);
2394 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
2395 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
2396 map->value_size, off, size);
2403 /* check read/write into map element returned by bpf_map_lookup_elem() */
2404 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
2405 int size, bool zero_size_allowed)
2407 struct bpf_reg_state *regs = cur_regs(env);
2408 struct bpf_map *map = regs[regno].map_ptr;
2410 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
2411 off + size > map->value_size) {
2412 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
2413 map->value_size, off, size);
2419 /* check read/write into a map element with possible variable offset */
2420 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
2421 int off, int size, bool zero_size_allowed)
2423 struct bpf_verifier_state *vstate = env->cur_state;
2424 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2425 struct bpf_reg_state *reg = &state->regs[regno];
2428 /* We may have adjusted the register to this map value, so we
2429 * need to try adding each of min_value and max_value to off
2430 * to make sure our theoretical access will be safe.
2432 if (env->log.level & BPF_LOG_LEVEL)
2433 print_verifier_state(env, state);
2435 /* The minimum value is only important with signed
2436 * comparisons where we can't assume the floor of a
2437 * value is 0. If we are using signed variables for our
2438 * index'es we need to make sure that whatever we use
2439 * will have a set floor within our range.
2441 if (reg->smin_value < 0 &&
2442 (reg->smin_value == S64_MIN ||
2443 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
2444 reg->smin_value + off < 0)) {
2445 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2449 err = __check_map_access(env, regno, reg->smin_value + off, size,
2452 verbose(env, "R%d min value is outside of the array range\n",
2457 /* If we haven't set a max value then we need to bail since we can't be
2458 * sure we won't do bad things.
2459 * If reg->umax_value + off could overflow, treat that as unbounded too.
2461 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
2462 verbose(env, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
2466 err = __check_map_access(env, regno, reg->umax_value + off, size,
2469 verbose(env, "R%d max value is outside of the array range\n",
2472 if (map_value_has_spin_lock(reg->map_ptr)) {
2473 u32 lock = reg->map_ptr->spin_lock_off;
2475 /* if any part of struct bpf_spin_lock can be touched by
2476 * load/store reject this program.
2477 * To check that [x1, x2) overlaps with [y1, y2)
2478 * it is sufficient to check x1 < y2 && y1 < x2.
2480 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
2481 lock < reg->umax_value + off + size) {
2482 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
2489 #define MAX_PACKET_OFF 0xffff
2491 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
2492 const struct bpf_call_arg_meta *meta,
2493 enum bpf_access_type t)
2495 switch (env->prog->type) {
2496 /* Program types only with direct read access go here! */
2497 case BPF_PROG_TYPE_LWT_IN:
2498 case BPF_PROG_TYPE_LWT_OUT:
2499 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
2500 case BPF_PROG_TYPE_SK_REUSEPORT:
2501 case BPF_PROG_TYPE_FLOW_DISSECTOR:
2502 case BPF_PROG_TYPE_CGROUP_SKB:
2507 /* Program types with direct read + write access go here! */
2508 case BPF_PROG_TYPE_SCHED_CLS:
2509 case BPF_PROG_TYPE_SCHED_ACT:
2510 case BPF_PROG_TYPE_XDP:
2511 case BPF_PROG_TYPE_LWT_XMIT:
2512 case BPF_PROG_TYPE_SK_SKB:
2513 case BPF_PROG_TYPE_SK_MSG:
2515 return meta->pkt_access;
2517 env->seen_direct_write = true;
2520 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
2522 env->seen_direct_write = true;
2531 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
2532 int off, int size, bool zero_size_allowed)
2534 struct bpf_reg_state *regs = cur_regs(env);
2535 struct bpf_reg_state *reg = ®s[regno];
2537 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
2538 (u64)off + size > reg->range) {
2539 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
2540 off, size, regno, reg->id, reg->off, reg->range);
2546 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
2547 int size, bool zero_size_allowed)
2549 struct bpf_reg_state *regs = cur_regs(env);
2550 struct bpf_reg_state *reg = ®s[regno];
2553 /* We may have added a variable offset to the packet pointer; but any
2554 * reg->range we have comes after that. We are only checking the fixed
2558 /* We don't allow negative numbers, because we aren't tracking enough
2559 * detail to prove they're safe.
2561 if (reg->smin_value < 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_packet_access(env, regno, off, size, zero_size_allowed);
2568 verbose(env, "R%d offset is outside of the packet\n", regno);
2572 /* __check_packet_access has made sure "off + size - 1" is within u16.
2573 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
2574 * otherwise find_good_pkt_pointers would have refused to set range info
2575 * that __check_packet_access would have rejected this pkt access.
2576 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
2578 env->prog->aux->max_pkt_offset =
2579 max_t(u32, env->prog->aux->max_pkt_offset,
2580 off + reg->umax_value + size - 1);
2585 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
2586 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
2587 enum bpf_access_type t, enum bpf_reg_type *reg_type,
2590 struct bpf_insn_access_aux info = {
2591 .reg_type = *reg_type,
2595 if (env->ops->is_valid_access &&
2596 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
2597 /* A non zero info.ctx_field_size indicates that this field is a
2598 * candidate for later verifier transformation to load the whole
2599 * field and then apply a mask when accessed with a narrower
2600 * access than actual ctx access size. A zero info.ctx_field_size
2601 * will only allow for whole field access and rejects any other
2602 * type of narrower access.
2604 *reg_type = info.reg_type;
2606 if (*reg_type == PTR_TO_BTF_ID)
2607 *btf_id = info.btf_id;
2609 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
2610 /* remember the offset of last byte accessed in ctx */
2611 if (env->prog->aux->max_ctx_offset < off + size)
2612 env->prog->aux->max_ctx_offset = off + size;
2616 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
2620 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
2623 if (size < 0 || off < 0 ||
2624 (u64)off + size > sizeof(struct bpf_flow_keys)) {
2625 verbose(env, "invalid access to flow keys off=%d size=%d\n",
2632 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
2633 u32 regno, int off, int size,
2634 enum bpf_access_type t)
2636 struct bpf_reg_state *regs = cur_regs(env);
2637 struct bpf_reg_state *reg = ®s[regno];
2638 struct bpf_insn_access_aux info = {};
2641 if (reg->smin_value < 0) {
2642 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2647 switch (reg->type) {
2648 case PTR_TO_SOCK_COMMON:
2649 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
2652 valid = bpf_sock_is_valid_access(off, size, t, &info);
2654 case PTR_TO_TCP_SOCK:
2655 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
2657 case PTR_TO_XDP_SOCK:
2658 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
2666 env->insn_aux_data[insn_idx].ctx_field_size =
2667 info.ctx_field_size;
2671 verbose(env, "R%d invalid %s access off=%d size=%d\n",
2672 regno, reg_type_str[reg->type], off, size);
2677 static bool __is_pointer_value(bool allow_ptr_leaks,
2678 const struct bpf_reg_state *reg)
2680 if (allow_ptr_leaks)
2683 return reg->type != SCALAR_VALUE;
2686 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
2688 return cur_regs(env) + regno;
2691 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
2693 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
2696 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
2698 const struct bpf_reg_state *reg = reg_state(env, regno);
2700 return reg->type == PTR_TO_CTX;
2703 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
2705 const struct bpf_reg_state *reg = reg_state(env, regno);
2707 return type_is_sk_pointer(reg->type);
2710 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
2712 const struct bpf_reg_state *reg = reg_state(env, regno);
2714 return type_is_pkt_pointer(reg->type);
2717 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
2719 const struct bpf_reg_state *reg = reg_state(env, regno);
2721 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
2722 return reg->type == PTR_TO_FLOW_KEYS;
2725 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
2726 const struct bpf_reg_state *reg,
2727 int off, int size, bool strict)
2729 struct tnum reg_off;
2732 /* Byte size accesses are always allowed. */
2733 if (!strict || size == 1)
2736 /* For platforms that do not have a Kconfig enabling
2737 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
2738 * NET_IP_ALIGN is universally set to '2'. And on platforms
2739 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
2740 * to this code only in strict mode where we want to emulate
2741 * the NET_IP_ALIGN==2 checking. Therefore use an
2742 * unconditional IP align value of '2'.
2746 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
2747 if (!tnum_is_aligned(reg_off, size)) {
2750 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2752 "misaligned packet access off %d+%s+%d+%d size %d\n",
2753 ip_align, tn_buf, reg->off, off, size);
2760 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
2761 const struct bpf_reg_state *reg,
2762 const char *pointer_desc,
2763 int off, int size, bool strict)
2765 struct tnum reg_off;
2767 /* Byte size accesses are always allowed. */
2768 if (!strict || size == 1)
2771 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
2772 if (!tnum_is_aligned(reg_off, size)) {
2775 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2776 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
2777 pointer_desc, tn_buf, reg->off, off, size);
2784 static int check_ptr_alignment(struct bpf_verifier_env *env,
2785 const struct bpf_reg_state *reg, int off,
2786 int size, bool strict_alignment_once)
2788 bool strict = env->strict_alignment || strict_alignment_once;
2789 const char *pointer_desc = "";
2791 switch (reg->type) {
2793 case PTR_TO_PACKET_META:
2794 /* Special case, because of NET_IP_ALIGN. Given metadata sits
2795 * right in front, treat it the very same way.
2797 return check_pkt_ptr_alignment(env, reg, off, size, strict);
2798 case PTR_TO_FLOW_KEYS:
2799 pointer_desc = "flow keys ";
2801 case PTR_TO_MAP_VALUE:
2802 pointer_desc = "value ";
2805 pointer_desc = "context ";
2808 pointer_desc = "stack ";
2809 /* The stack spill tracking logic in check_stack_write()
2810 * and check_stack_read() relies on stack accesses being
2816 pointer_desc = "sock ";
2818 case PTR_TO_SOCK_COMMON:
2819 pointer_desc = "sock_common ";
2821 case PTR_TO_TCP_SOCK:
2822 pointer_desc = "tcp_sock ";
2824 case PTR_TO_XDP_SOCK:
2825 pointer_desc = "xdp_sock ";
2830 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
2834 static int update_stack_depth(struct bpf_verifier_env *env,
2835 const struct bpf_func_state *func,
2838 u16 stack = env->subprog_info[func->subprogno].stack_depth;
2843 /* update known max for given subprogram */
2844 env->subprog_info[func->subprogno].stack_depth = -off;
2848 /* starting from main bpf function walk all instructions of the function
2849 * and recursively walk all callees that given function can call.
2850 * Ignore jump and exit insns.
2851 * Since recursion is prevented by check_cfg() this algorithm
2852 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
2854 static int check_max_stack_depth(struct bpf_verifier_env *env)
2856 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
2857 struct bpf_subprog_info *subprog = env->subprog_info;
2858 struct bpf_insn *insn = env->prog->insnsi;
2859 int ret_insn[MAX_CALL_FRAMES];
2860 int ret_prog[MAX_CALL_FRAMES];
2863 /* round up to 32-bytes, since this is granularity
2864 * of interpreter stack size
2866 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
2867 if (depth > MAX_BPF_STACK) {
2868 verbose(env, "combined stack size of %d calls is %d. Too large\n",
2873 subprog_end = subprog[idx + 1].start;
2874 for (; i < subprog_end; i++) {
2875 if (insn[i].code != (BPF_JMP | BPF_CALL))
2877 if (insn[i].src_reg != BPF_PSEUDO_CALL)
2879 /* remember insn and function to return to */
2880 ret_insn[frame] = i + 1;
2881 ret_prog[frame] = idx;
2883 /* find the callee */
2884 i = i + insn[i].imm + 1;
2885 idx = find_subprog(env, i);
2887 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
2892 if (frame >= MAX_CALL_FRAMES) {
2893 verbose(env, "the call stack of %d frames is too deep !\n",
2899 /* end of for() loop means the last insn of the 'subprog'
2900 * was reached. Doesn't matter whether it was JA or EXIT
2904 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
2906 i = ret_insn[frame];
2907 idx = ret_prog[frame];
2911 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
2912 static int get_callee_stack_depth(struct bpf_verifier_env *env,
2913 const struct bpf_insn *insn, int idx)
2915 int start = idx + insn->imm + 1, subprog;
2917 subprog = find_subprog(env, start);
2919 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
2923 return env->subprog_info[subprog].stack_depth;
2927 int check_ctx_reg(struct bpf_verifier_env *env,
2928 const struct bpf_reg_state *reg, int regno)
2930 /* Access to ctx or passing it to a helper is only allowed in
2931 * its original, unmodified form.
2935 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
2940 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
2943 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2944 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
2951 static int check_tp_buffer_access(struct bpf_verifier_env *env,
2952 const struct bpf_reg_state *reg,
2953 int regno, int off, int size)
2957 "R%d invalid tracepoint buffer access: off=%d, size=%d",
2961 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
2964 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2966 "R%d invalid variable buffer offset: off=%d, var_off=%s",
2967 regno, off, tn_buf);
2970 if (off + size > env->prog->aux->max_tp_access)
2971 env->prog->aux->max_tp_access = off + size;
2976 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
2977 static void zext_32_to_64(struct bpf_reg_state *reg)
2979 reg->var_off = tnum_subreg(reg->var_off);
2980 __reg_assign_32_into_64(reg);
2983 /* truncate register to smaller size (in bytes)
2984 * must be called with size < BPF_REG_SIZE
2986 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
2990 /* clear high bits in bit representation */
2991 reg->var_off = tnum_cast(reg->var_off, size);
2993 /* fix arithmetic bounds */
2994 mask = ((u64)1 << (size * 8)) - 1;
2995 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
2996 reg->umin_value &= mask;
2997 reg->umax_value &= mask;
2999 reg->umin_value = 0;
3000 reg->umax_value = mask;
3002 reg->smin_value = reg->umin_value;
3003 reg->smax_value = reg->umax_value;
3005 /* If size is smaller than 32bit register the 32bit register
3006 * values are also truncated so we push 64-bit bounds into
3007 * 32-bit bounds. Above were truncated < 32-bits already.
3011 __reg_combine_64_into_32(reg);
3014 static bool bpf_map_is_rdonly(const struct bpf_map *map)
3016 return (map->map_flags & BPF_F_RDONLY_PROG) && map->frozen;
3019 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
3025 err = map->ops->map_direct_value_addr(map, &addr, off);
3028 ptr = (void *)(long)addr + off;
3032 *val = (u64)*(u8 *)ptr;
3035 *val = (u64)*(u16 *)ptr;
3038 *val = (u64)*(u32 *)ptr;
3049 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
3050 struct bpf_reg_state *regs,
3051 int regno, int off, int size,
3052 enum bpf_access_type atype,
3055 struct bpf_reg_state *reg = regs + regno;
3056 const struct btf_type *t = btf_type_by_id(btf_vmlinux, reg->btf_id);
3057 const char *tname = btf_name_by_offset(btf_vmlinux, t->name_off);
3063 "R%d is ptr_%s invalid negative access: off=%d\n",
3067 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3070 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3072 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
3073 regno, tname, off, tn_buf);
3077 if (env->ops->btf_struct_access) {
3078 ret = env->ops->btf_struct_access(&env->log, t, off, size,
3081 if (atype != BPF_READ) {
3082 verbose(env, "only read is supported\n");
3086 ret = btf_struct_access(&env->log, t, off, size, atype,
3093 if (atype == BPF_READ) {
3094 if (ret == SCALAR_VALUE) {
3095 mark_reg_unknown(env, regs, value_regno);
3098 mark_reg_known_zero(env, regs, value_regno);
3099 regs[value_regno].type = PTR_TO_BTF_ID;
3100 regs[value_regno].btf_id = btf_id;
3106 /* check whether memory at (regno + off) is accessible for t = (read | write)
3107 * if t==write, value_regno is a register which value is stored into memory
3108 * if t==read, value_regno is a register which will receive the value from memory
3109 * if t==write && value_regno==-1, some unknown value is stored into memory
3110 * if t==read && value_regno==-1, don't care what we read from memory
3112 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
3113 int off, int bpf_size, enum bpf_access_type t,
3114 int value_regno, bool strict_alignment_once)
3116 struct bpf_reg_state *regs = cur_regs(env);
3117 struct bpf_reg_state *reg = regs + regno;
3118 struct bpf_func_state *state;
3121 size = bpf_size_to_bytes(bpf_size);
3125 /* alignment checks will add in reg->off themselves */
3126 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
3130 /* for access checks, reg->off is just part of off */
3133 if (reg->type == PTR_TO_MAP_VALUE) {
3134 if (t == BPF_WRITE && value_regno >= 0 &&
3135 is_pointer_value(env, value_regno)) {
3136 verbose(env, "R%d leaks addr into map\n", value_regno);
3139 err = check_map_access_type(env, regno, off, size, t);
3142 err = check_map_access(env, regno, off, size, false);
3143 if (!err && t == BPF_READ && value_regno >= 0) {
3144 struct bpf_map *map = reg->map_ptr;
3146 /* if map is read-only, track its contents as scalars */
3147 if (tnum_is_const(reg->var_off) &&
3148 bpf_map_is_rdonly(map) &&
3149 map->ops->map_direct_value_addr) {
3150 int map_off = off + reg->var_off.value;
3153 err = bpf_map_direct_read(map, map_off, size,
3158 regs[value_regno].type = SCALAR_VALUE;
3159 __mark_reg_known(®s[value_regno], val);
3161 mark_reg_unknown(env, regs, value_regno);
3164 } else if (reg->type == PTR_TO_CTX) {
3165 enum bpf_reg_type reg_type = SCALAR_VALUE;
3168 if (t == BPF_WRITE && value_regno >= 0 &&
3169 is_pointer_value(env, value_regno)) {
3170 verbose(env, "R%d leaks addr into ctx\n", value_regno);
3174 err = check_ctx_reg(env, reg, regno);
3178 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf_id);
3180 verbose_linfo(env, insn_idx, "; ");
3181 if (!err && t == BPF_READ && value_regno >= 0) {
3182 /* ctx access returns either a scalar, or a
3183 * PTR_TO_PACKET[_META,_END]. In the latter
3184 * case, we know the offset is zero.
3186 if (reg_type == SCALAR_VALUE) {
3187 mark_reg_unknown(env, regs, value_regno);
3189 mark_reg_known_zero(env, regs,
3191 if (reg_type_may_be_null(reg_type))
3192 regs[value_regno].id = ++env->id_gen;
3193 /* A load of ctx field could have different
3194 * actual load size with the one encoded in the
3195 * insn. When the dst is PTR, it is for sure not
3198 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
3199 if (reg_type == PTR_TO_BTF_ID)
3200 regs[value_regno].btf_id = btf_id;
3202 regs[value_regno].type = reg_type;
3205 } else if (reg->type == PTR_TO_STACK) {
3206 off += reg->var_off.value;
3207 err = check_stack_access(env, reg, off, size);
3211 state = func(env, reg);
3212 err = update_stack_depth(env, state, off);
3217 err = check_stack_write(env, state, off, size,
3218 value_regno, insn_idx);
3220 err = check_stack_read(env, state, off, size,
3222 } else if (reg_is_pkt_pointer(reg)) {
3223 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
3224 verbose(env, "cannot write into packet\n");
3227 if (t == BPF_WRITE && value_regno >= 0 &&
3228 is_pointer_value(env, value_regno)) {
3229 verbose(env, "R%d leaks addr into packet\n",
3233 err = check_packet_access(env, regno, off, size, false);
3234 if (!err && t == BPF_READ && value_regno >= 0)
3235 mark_reg_unknown(env, regs, value_regno);
3236 } else if (reg->type == PTR_TO_FLOW_KEYS) {
3237 if (t == BPF_WRITE && value_regno >= 0 &&
3238 is_pointer_value(env, value_regno)) {
3239 verbose(env, "R%d leaks addr into flow keys\n",
3244 err = check_flow_keys_access(env, off, size);
3245 if (!err && t == BPF_READ && value_regno >= 0)
3246 mark_reg_unknown(env, regs, value_regno);
3247 } else if (type_is_sk_pointer(reg->type)) {
3248 if (t == BPF_WRITE) {
3249 verbose(env, "R%d cannot write into %s\n",
3250 regno, reg_type_str[reg->type]);
3253 err = check_sock_access(env, insn_idx, regno, off, size, t);
3254 if (!err && value_regno >= 0)
3255 mark_reg_unknown(env, regs, value_regno);
3256 } else if (reg->type == PTR_TO_TP_BUFFER) {
3257 err = check_tp_buffer_access(env, reg, regno, off, size);
3258 if (!err && t == BPF_READ && value_regno >= 0)
3259 mark_reg_unknown(env, regs, value_regno);
3260 } else if (reg->type == PTR_TO_BTF_ID) {
3261 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
3264 verbose(env, "R%d invalid mem access '%s'\n", regno,
3265 reg_type_str[reg->type]);
3269 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
3270 regs[value_regno].type == SCALAR_VALUE) {
3271 /* b/h/w load zero-extends, mark upper bits as known 0 */
3272 coerce_reg_to_size(®s[value_regno], size);
3277 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
3281 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
3283 verbose(env, "BPF_XADD uses reserved fields\n");
3287 /* check src1 operand */
3288 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3292 /* check src2 operand */
3293 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3297 if (is_pointer_value(env, insn->src_reg)) {
3298 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
3302 if (is_ctx_reg(env, insn->dst_reg) ||
3303 is_pkt_reg(env, insn->dst_reg) ||
3304 is_flow_key_reg(env, insn->dst_reg) ||
3305 is_sk_reg(env, insn->dst_reg)) {
3306 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
3308 reg_type_str[reg_state(env, insn->dst_reg)->type]);
3312 /* check whether atomic_add can read the memory */
3313 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3314 BPF_SIZE(insn->code), BPF_READ, -1, true);
3318 /* check whether atomic_add can write into the same memory */
3319 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3320 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
3323 static int __check_stack_boundary(struct bpf_verifier_env *env, u32 regno,
3324 int off, int access_size,
3325 bool zero_size_allowed)
3327 struct bpf_reg_state *reg = reg_state(env, regno);
3329 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
3330 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
3331 if (tnum_is_const(reg->var_off)) {
3332 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
3333 regno, off, access_size);
3337 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3338 verbose(env, "invalid stack type R%d var_off=%s access_size=%d\n",
3339 regno, tn_buf, access_size);
3346 /* when register 'regno' is passed into function that will read 'access_size'
3347 * bytes from that pointer, make sure that it's within stack boundary
3348 * and all elements of stack are initialized.
3349 * Unlike most pointer bounds-checking functions, this one doesn't take an
3350 * 'off' argument, so it has to add in reg->off itself.
3352 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
3353 int access_size, bool zero_size_allowed,
3354 struct bpf_call_arg_meta *meta)
3356 struct bpf_reg_state *reg = reg_state(env, regno);
3357 struct bpf_func_state *state = func(env, reg);
3358 int err, min_off, max_off, i, j, slot, spi;
3360 if (reg->type != PTR_TO_STACK) {
3361 /* Allow zero-byte read from NULL, regardless of pointer type */
3362 if (zero_size_allowed && access_size == 0 &&
3363 register_is_null(reg))
3366 verbose(env, "R%d type=%s expected=%s\n", regno,
3367 reg_type_str[reg->type],
3368 reg_type_str[PTR_TO_STACK]);
3372 if (tnum_is_const(reg->var_off)) {
3373 min_off = max_off = reg->var_off.value + reg->off;
3374 err = __check_stack_boundary(env, regno, min_off, access_size,
3379 /* Variable offset is prohibited for unprivileged mode for
3380 * simplicity since it requires corresponding support in
3381 * Spectre masking for stack ALU.
3382 * See also retrieve_ptr_limit().
3384 if (!env->allow_ptr_leaks) {
3387 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3388 verbose(env, "R%d indirect variable offset stack access prohibited for !root, var_off=%s\n",
3392 /* Only initialized buffer on stack is allowed to be accessed
3393 * with variable offset. With uninitialized buffer it's hard to
3394 * guarantee that whole memory is marked as initialized on
3395 * helper return since specific bounds are unknown what may
3396 * cause uninitialized stack leaking.
3398 if (meta && meta->raw_mode)
3401 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
3402 reg->smax_value <= -BPF_MAX_VAR_OFF) {
3403 verbose(env, "R%d unbounded indirect variable offset stack access\n",
3407 min_off = reg->smin_value + reg->off;
3408 max_off = reg->smax_value + reg->off;
3409 err = __check_stack_boundary(env, regno, min_off, access_size,
3412 verbose(env, "R%d min value is outside of stack bound\n",
3416 err = __check_stack_boundary(env, regno, max_off, access_size,
3419 verbose(env, "R%d max value is outside of stack bound\n",
3425 if (meta && meta->raw_mode) {
3426 meta->access_size = access_size;
3427 meta->regno = regno;
3431 for (i = min_off; i < max_off + access_size; i++) {
3435 spi = slot / BPF_REG_SIZE;
3436 if (state->allocated_stack <= slot)
3438 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
3439 if (*stype == STACK_MISC)
3441 if (*stype == STACK_ZERO) {
3442 /* helper can write anything into the stack */
3443 *stype = STACK_MISC;
3446 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
3447 state->stack[spi].spilled_ptr.type == SCALAR_VALUE) {
3448 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
3449 for (j = 0; j < BPF_REG_SIZE; j++)
3450 state->stack[spi].slot_type[j] = STACK_MISC;
3455 if (tnum_is_const(reg->var_off)) {
3456 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
3457 min_off, i - min_off, access_size);
3461 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3462 verbose(env, "invalid indirect read from stack var_off %s+%d size %d\n",
3463 tn_buf, i - min_off, access_size);
3467 /* reading any byte out of 8-byte 'spill_slot' will cause
3468 * the whole slot to be marked as 'read'
3470 mark_reg_read(env, &state->stack[spi].spilled_ptr,
3471 state->stack[spi].spilled_ptr.parent,
3474 return update_stack_depth(env, state, min_off);
3477 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
3478 int access_size, bool zero_size_allowed,
3479 struct bpf_call_arg_meta *meta)
3481 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3483 switch (reg->type) {
3485 case PTR_TO_PACKET_META:
3486 return check_packet_access(env, regno, reg->off, access_size,
3488 case PTR_TO_MAP_VALUE:
3489 if (check_map_access_type(env, regno, reg->off, access_size,
3490 meta && meta->raw_mode ? BPF_WRITE :
3493 return check_map_access(env, regno, reg->off, access_size,
3495 default: /* scalar_value|ptr_to_stack or invalid ptr */
3496 return check_stack_boundary(env, regno, access_size,
3497 zero_size_allowed, meta);
3501 /* Implementation details:
3502 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
3503 * Two bpf_map_lookups (even with the same key) will have different reg->id.
3504 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
3505 * value_or_null->value transition, since the verifier only cares about
3506 * the range of access to valid map value pointer and doesn't care about actual
3507 * address of the map element.
3508 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
3509 * reg->id > 0 after value_or_null->value transition. By doing so
3510 * two bpf_map_lookups will be considered two different pointers that
3511 * point to different bpf_spin_locks.
3512 * The verifier allows taking only one bpf_spin_lock at a time to avoid
3514 * Since only one bpf_spin_lock is allowed the checks are simpler than
3515 * reg_is_refcounted() logic. The verifier needs to remember only
3516 * one spin_lock instead of array of acquired_refs.
3517 * cur_state->active_spin_lock remembers which map value element got locked
3518 * and clears it after bpf_spin_unlock.
3520 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
3523 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3524 struct bpf_verifier_state *cur = env->cur_state;
3525 bool is_const = tnum_is_const(reg->var_off);
3526 struct bpf_map *map = reg->map_ptr;
3527 u64 val = reg->var_off.value;
3529 if (reg->type != PTR_TO_MAP_VALUE) {
3530 verbose(env, "R%d is not a pointer to map_value\n", regno);
3535 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
3541 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
3545 if (!map_value_has_spin_lock(map)) {
3546 if (map->spin_lock_off == -E2BIG)
3548 "map '%s' has more than one 'struct bpf_spin_lock'\n",
3550 else if (map->spin_lock_off == -ENOENT)
3552 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
3556 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
3560 if (map->spin_lock_off != val + reg->off) {
3561 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
3566 if (cur->active_spin_lock) {
3568 "Locking two bpf_spin_locks are not allowed\n");
3571 cur->active_spin_lock = reg->id;
3573 if (!cur->active_spin_lock) {
3574 verbose(env, "bpf_spin_unlock without taking a lock\n");
3577 if (cur->active_spin_lock != reg->id) {
3578 verbose(env, "bpf_spin_unlock of different lock\n");
3581 cur->active_spin_lock = 0;
3586 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
3588 return type == ARG_PTR_TO_MEM ||
3589 type == ARG_PTR_TO_MEM_OR_NULL ||
3590 type == ARG_PTR_TO_UNINIT_MEM;
3593 static bool arg_type_is_mem_size(enum bpf_arg_type type)
3595 return type == ARG_CONST_SIZE ||
3596 type == ARG_CONST_SIZE_OR_ZERO;
3599 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
3601 return type == ARG_PTR_TO_INT ||
3602 type == ARG_PTR_TO_LONG;
3605 static int int_ptr_type_to_size(enum bpf_arg_type type)
3607 if (type == ARG_PTR_TO_INT)
3609 else if (type == ARG_PTR_TO_LONG)
3615 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
3616 enum bpf_arg_type arg_type,
3617 struct bpf_call_arg_meta *meta)
3619 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3620 enum bpf_reg_type expected_type, type = reg->type;
3623 if (arg_type == ARG_DONTCARE)
3626 err = check_reg_arg(env, regno, SRC_OP);
3630 if (arg_type == ARG_ANYTHING) {
3631 if (is_pointer_value(env, regno)) {
3632 verbose(env, "R%d leaks addr into helper function\n",
3639 if (type_is_pkt_pointer(type) &&
3640 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
3641 verbose(env, "helper access to the packet is not allowed\n");
3645 if (arg_type == ARG_PTR_TO_MAP_KEY ||
3646 arg_type == ARG_PTR_TO_MAP_VALUE ||
3647 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE ||
3648 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL) {
3649 expected_type = PTR_TO_STACK;
3650 if (register_is_null(reg) &&
3651 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL)
3652 /* final test in check_stack_boundary() */;
3653 else if (!type_is_pkt_pointer(type) &&
3654 type != PTR_TO_MAP_VALUE &&
3655 type != expected_type)
3657 } else if (arg_type == ARG_CONST_SIZE ||
3658 arg_type == ARG_CONST_SIZE_OR_ZERO) {
3659 expected_type = SCALAR_VALUE;
3660 if (type != expected_type)
3662 } else if (arg_type == ARG_CONST_MAP_PTR) {
3663 expected_type = CONST_PTR_TO_MAP;
3664 if (type != expected_type)
3666 } else if (arg_type == ARG_PTR_TO_CTX ||
3667 arg_type == ARG_PTR_TO_CTX_OR_NULL) {
3668 expected_type = PTR_TO_CTX;
3669 if (!(register_is_null(reg) &&
3670 arg_type == ARG_PTR_TO_CTX_OR_NULL)) {
3671 if (type != expected_type)
3673 err = check_ctx_reg(env, reg, regno);
3677 } else if (arg_type == ARG_PTR_TO_SOCK_COMMON) {
3678 expected_type = PTR_TO_SOCK_COMMON;
3679 /* Any sk pointer can be ARG_PTR_TO_SOCK_COMMON */
3680 if (!type_is_sk_pointer(type))
3682 if (reg->ref_obj_id) {
3683 if (meta->ref_obj_id) {
3684 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
3685 regno, reg->ref_obj_id,
3689 meta->ref_obj_id = reg->ref_obj_id;
3691 } else if (arg_type == ARG_PTR_TO_SOCKET) {
3692 expected_type = PTR_TO_SOCKET;
3693 if (type != expected_type)
3695 } else if (arg_type == ARG_PTR_TO_BTF_ID) {
3696 expected_type = PTR_TO_BTF_ID;
3697 if (type != expected_type)
3699 if (reg->btf_id != meta->btf_id) {
3700 verbose(env, "Helper has type %s got %s in R%d\n",
3701 kernel_type_name(meta->btf_id),
3702 kernel_type_name(reg->btf_id), regno);
3706 if (!tnum_is_const(reg->var_off) || reg->var_off.value || reg->off) {
3707 verbose(env, "R%d is a pointer to in-kernel struct with non-zero offset\n",
3711 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
3712 if (meta->func_id == BPF_FUNC_spin_lock) {
3713 if (process_spin_lock(env, regno, true))
3715 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
3716 if (process_spin_lock(env, regno, false))
3719 verbose(env, "verifier internal error\n");
3722 } else if (arg_type_is_mem_ptr(arg_type)) {
3723 expected_type = PTR_TO_STACK;
3724 /* One exception here. In case function allows for NULL to be
3725 * passed in as argument, it's a SCALAR_VALUE type. Final test
3726 * happens during stack boundary checking.
3728 if (register_is_null(reg) &&
3729 arg_type == ARG_PTR_TO_MEM_OR_NULL)
3730 /* final test in check_stack_boundary() */;
3731 else if (!type_is_pkt_pointer(type) &&
3732 type != PTR_TO_MAP_VALUE &&
3733 type != expected_type)
3735 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
3736 } else if (arg_type_is_int_ptr(arg_type)) {
3737 expected_type = PTR_TO_STACK;
3738 if (!type_is_pkt_pointer(type) &&
3739 type != PTR_TO_MAP_VALUE &&
3740 type != expected_type)
3743 verbose(env, "unsupported arg_type %d\n", arg_type);
3747 if (arg_type == ARG_CONST_MAP_PTR) {
3748 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
3749 meta->map_ptr = reg->map_ptr;
3750 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
3751 /* bpf_map_xxx(..., map_ptr, ..., key) call:
3752 * check that [key, key + map->key_size) are within
3753 * stack limits and initialized
3755 if (!meta->map_ptr) {
3756 /* in function declaration map_ptr must come before
3757 * map_key, so that it's verified and known before
3758 * we have to check map_key here. Otherwise it means
3759 * that kernel subsystem misconfigured verifier
3761 verbose(env, "invalid map_ptr to access map->key\n");
3764 err = check_helper_mem_access(env, regno,
3765 meta->map_ptr->key_size, false,
3767 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
3768 (arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL &&
3769 !register_is_null(reg)) ||
3770 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
3771 /* bpf_map_xxx(..., map_ptr, ..., value) call:
3772 * check [value, value + map->value_size) validity
3774 if (!meta->map_ptr) {
3775 /* kernel subsystem misconfigured verifier */
3776 verbose(env, "invalid map_ptr to access map->value\n");
3779 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
3780 err = check_helper_mem_access(env, regno,
3781 meta->map_ptr->value_size, false,
3783 } else if (arg_type_is_mem_size(arg_type)) {
3784 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
3786 /* This is used to refine r0 return value bounds for helpers
3787 * that enforce this value as an upper bound on return values.
3788 * See do_refine_retval_range() for helpers that can refine
3789 * the return value. C type of helper is u32 so we pull register
3790 * bound from umax_value however, if negative verifier errors
3791 * out. Only upper bounds can be learned because retval is an
3792 * int type and negative retvals are allowed.
3794 meta->msize_max_value = reg->umax_value;
3796 /* The register is SCALAR_VALUE; the access check
3797 * happens using its boundaries.
3799 if (!tnum_is_const(reg->var_off))
3800 /* For unprivileged variable accesses, disable raw
3801 * mode so that the program is required to
3802 * initialize all the memory that the helper could
3803 * just partially fill up.
3807 if (reg->smin_value < 0) {
3808 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
3813 if (reg->umin_value == 0) {
3814 err = check_helper_mem_access(env, regno - 1, 0,
3821 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
3822 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
3826 err = check_helper_mem_access(env, regno - 1,
3828 zero_size_allowed, meta);
3830 err = mark_chain_precision(env, regno);
3831 } else if (arg_type_is_int_ptr(arg_type)) {
3832 int size = int_ptr_type_to_size(arg_type);
3834 err = check_helper_mem_access(env, regno, size, false, meta);
3837 err = check_ptr_alignment(env, reg, 0, size, true);
3842 verbose(env, "R%d type=%s expected=%s\n", regno,
3843 reg_type_str[type], reg_type_str[expected_type]);
3847 static int check_map_func_compatibility(struct bpf_verifier_env *env,
3848 struct bpf_map *map, int func_id)
3853 /* We need a two way check, first is from map perspective ... */
3854 switch (map->map_type) {
3855 case BPF_MAP_TYPE_PROG_ARRAY:
3856 if (func_id != BPF_FUNC_tail_call)
3859 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
3860 if (func_id != BPF_FUNC_perf_event_read &&
3861 func_id != BPF_FUNC_perf_event_output &&
3862 func_id != BPF_FUNC_skb_output &&
3863 func_id != BPF_FUNC_perf_event_read_value &&
3864 func_id != BPF_FUNC_xdp_output)
3867 case BPF_MAP_TYPE_STACK_TRACE:
3868 if (func_id != BPF_FUNC_get_stackid)
3871 case BPF_MAP_TYPE_CGROUP_ARRAY:
3872 if (func_id != BPF_FUNC_skb_under_cgroup &&
3873 func_id != BPF_FUNC_current_task_under_cgroup)
3876 case BPF_MAP_TYPE_CGROUP_STORAGE:
3877 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
3878 if (func_id != BPF_FUNC_get_local_storage)
3881 case BPF_MAP_TYPE_DEVMAP:
3882 case BPF_MAP_TYPE_DEVMAP_HASH:
3883 if (func_id != BPF_FUNC_redirect_map &&
3884 func_id != BPF_FUNC_map_lookup_elem)
3887 /* Restrict bpf side of cpumap and xskmap, open when use-cases
3890 case BPF_MAP_TYPE_CPUMAP:
3891 if (func_id != BPF_FUNC_redirect_map)
3894 case BPF_MAP_TYPE_XSKMAP:
3895 if (func_id != BPF_FUNC_redirect_map &&
3896 func_id != BPF_FUNC_map_lookup_elem)
3899 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
3900 case BPF_MAP_TYPE_HASH_OF_MAPS:
3901 if (func_id != BPF_FUNC_map_lookup_elem)
3904 case BPF_MAP_TYPE_SOCKMAP:
3905 if (func_id != BPF_FUNC_sk_redirect_map &&
3906 func_id != BPF_FUNC_sock_map_update &&
3907 func_id != BPF_FUNC_map_delete_elem &&
3908 func_id != BPF_FUNC_msg_redirect_map &&
3909 func_id != BPF_FUNC_sk_select_reuseport)
3912 case BPF_MAP_TYPE_SOCKHASH:
3913 if (func_id != BPF_FUNC_sk_redirect_hash &&
3914 func_id != BPF_FUNC_sock_hash_update &&
3915 func_id != BPF_FUNC_map_delete_elem &&
3916 func_id != BPF_FUNC_msg_redirect_hash &&
3917 func_id != BPF_FUNC_sk_select_reuseport)
3920 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
3921 if (func_id != BPF_FUNC_sk_select_reuseport)
3924 case BPF_MAP_TYPE_QUEUE:
3925 case BPF_MAP_TYPE_STACK:
3926 if (func_id != BPF_FUNC_map_peek_elem &&
3927 func_id != BPF_FUNC_map_pop_elem &&
3928 func_id != BPF_FUNC_map_push_elem)
3931 case BPF_MAP_TYPE_SK_STORAGE:
3932 if (func_id != BPF_FUNC_sk_storage_get &&
3933 func_id != BPF_FUNC_sk_storage_delete)
3940 /* ... and second from the function itself. */
3942 case BPF_FUNC_tail_call:
3943 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
3945 if (env->subprog_cnt > 1) {
3946 verbose(env, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
3950 case BPF_FUNC_perf_event_read:
3951 case BPF_FUNC_perf_event_output:
3952 case BPF_FUNC_perf_event_read_value:
3953 case BPF_FUNC_skb_output:
3954 case BPF_FUNC_xdp_output:
3955 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
3958 case BPF_FUNC_get_stackid:
3959 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
3962 case BPF_FUNC_current_task_under_cgroup:
3963 case BPF_FUNC_skb_under_cgroup:
3964 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
3967 case BPF_FUNC_redirect_map:
3968 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
3969 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
3970 map->map_type != BPF_MAP_TYPE_CPUMAP &&
3971 map->map_type != BPF_MAP_TYPE_XSKMAP)
3974 case BPF_FUNC_sk_redirect_map:
3975 case BPF_FUNC_msg_redirect_map:
3976 case BPF_FUNC_sock_map_update:
3977 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
3980 case BPF_FUNC_sk_redirect_hash:
3981 case BPF_FUNC_msg_redirect_hash:
3982 case BPF_FUNC_sock_hash_update:
3983 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
3986 case BPF_FUNC_get_local_storage:
3987 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
3988 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
3991 case BPF_FUNC_sk_select_reuseport:
3992 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
3993 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
3994 map->map_type != BPF_MAP_TYPE_SOCKHASH)
3997 case BPF_FUNC_map_peek_elem:
3998 case BPF_FUNC_map_pop_elem:
3999 case BPF_FUNC_map_push_elem:
4000 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
4001 map->map_type != BPF_MAP_TYPE_STACK)
4004 case BPF_FUNC_sk_storage_get:
4005 case BPF_FUNC_sk_storage_delete:
4006 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
4015 verbose(env, "cannot pass map_type %d into func %s#%d\n",
4016 map->map_type, func_id_name(func_id), func_id);
4020 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
4024 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
4026 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
4028 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
4030 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
4032 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
4035 /* We only support one arg being in raw mode at the moment,
4036 * which is sufficient for the helper functions we have
4042 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
4043 enum bpf_arg_type arg_next)
4045 return (arg_type_is_mem_ptr(arg_curr) &&
4046 !arg_type_is_mem_size(arg_next)) ||
4047 (!arg_type_is_mem_ptr(arg_curr) &&
4048 arg_type_is_mem_size(arg_next));
4051 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
4053 /* bpf_xxx(..., buf, len) call will access 'len'
4054 * bytes from memory 'buf'. Both arg types need
4055 * to be paired, so make sure there's no buggy
4056 * helper function specification.
4058 if (arg_type_is_mem_size(fn->arg1_type) ||
4059 arg_type_is_mem_ptr(fn->arg5_type) ||
4060 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
4061 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
4062 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
4063 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
4069 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
4073 if (arg_type_may_be_refcounted(fn->arg1_type))
4075 if (arg_type_may_be_refcounted(fn->arg2_type))
4077 if (arg_type_may_be_refcounted(fn->arg3_type))
4079 if (arg_type_may_be_refcounted(fn->arg4_type))
4081 if (arg_type_may_be_refcounted(fn->arg5_type))
4084 /* A reference acquiring function cannot acquire
4085 * another refcounted ptr.
4087 if (is_acquire_function(func_id) && count)
4090 /* We only support one arg being unreferenced at the moment,
4091 * which is sufficient for the helper functions we have right now.
4096 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
4098 return check_raw_mode_ok(fn) &&
4099 check_arg_pair_ok(fn) &&
4100 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
4103 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
4104 * are now invalid, so turn them into unknown SCALAR_VALUE.
4106 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
4107 struct bpf_func_state *state)
4109 struct bpf_reg_state *regs = state->regs, *reg;
4112 for (i = 0; i < MAX_BPF_REG; i++)
4113 if (reg_is_pkt_pointer_any(®s[i]))
4114 mark_reg_unknown(env, regs, i);
4116 bpf_for_each_spilled_reg(i, state, reg) {
4119 if (reg_is_pkt_pointer_any(reg))
4120 __mark_reg_unknown(env, reg);
4124 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
4126 struct bpf_verifier_state *vstate = env->cur_state;
4129 for (i = 0; i <= vstate->curframe; i++)
4130 __clear_all_pkt_pointers(env, vstate->frame[i]);
4133 static void release_reg_references(struct bpf_verifier_env *env,
4134 struct bpf_func_state *state,
4137 struct bpf_reg_state *regs = state->regs, *reg;
4140 for (i = 0; i < MAX_BPF_REG; i++)
4141 if (regs[i].ref_obj_id == ref_obj_id)
4142 mark_reg_unknown(env, regs, i);
4144 bpf_for_each_spilled_reg(i, state, reg) {
4147 if (reg->ref_obj_id == ref_obj_id)
4148 __mark_reg_unknown(env, reg);
4152 /* The pointer with the specified id has released its reference to kernel
4153 * resources. Identify all copies of the same pointer and clear the reference.
4155 static int release_reference(struct bpf_verifier_env *env,
4158 struct bpf_verifier_state *vstate = env->cur_state;
4162 err = release_reference_state(cur_func(env), ref_obj_id);
4166 for (i = 0; i <= vstate->curframe; i++)
4167 release_reg_references(env, vstate->frame[i], ref_obj_id);
4172 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
4173 struct bpf_reg_state *regs)
4177 /* after the call registers r0 - r5 were scratched */
4178 for (i = 0; i < CALLER_SAVED_REGS; i++) {
4179 mark_reg_not_init(env, regs, caller_saved[i]);
4180 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
4184 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
4187 struct bpf_verifier_state *state = env->cur_state;
4188 struct bpf_func_info_aux *func_info_aux;
4189 struct bpf_func_state *caller, *callee;
4190 int i, err, subprog, target_insn;
4191 bool is_global = false;
4193 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
4194 verbose(env, "the call stack of %d frames is too deep\n",
4195 state->curframe + 2);
4199 target_insn = *insn_idx + insn->imm;
4200 subprog = find_subprog(env, target_insn + 1);
4202 verbose(env, "verifier bug. No program starts at insn %d\n",
4207 caller = state->frame[state->curframe];
4208 if (state->frame[state->curframe + 1]) {
4209 verbose(env, "verifier bug. Frame %d already allocated\n",
4210 state->curframe + 1);
4214 func_info_aux = env->prog->aux->func_info_aux;
4216 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
4217 err = btf_check_func_arg_match(env, subprog, caller->regs);
4222 verbose(env, "Caller passes invalid args into func#%d\n",
4226 if (env->log.level & BPF_LOG_LEVEL)
4228 "Func#%d is global and valid. Skipping.\n",
4230 clear_caller_saved_regs(env, caller->regs);
4232 /* All global functions return SCALAR_VALUE */
4233 mark_reg_unknown(env, caller->regs, BPF_REG_0);
4235 /* continue with next insn after call */
4240 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
4243 state->frame[state->curframe + 1] = callee;
4245 /* callee cannot access r0, r6 - r9 for reading and has to write
4246 * into its own stack before reading from it.
4247 * callee can read/write into caller's stack
4249 init_func_state(env, callee,
4250 /* remember the callsite, it will be used by bpf_exit */
4251 *insn_idx /* callsite */,
4252 state->curframe + 1 /* frameno within this callchain */,
4253 subprog /* subprog number within this prog */);
4255 /* Transfer references to the callee */
4256 err = transfer_reference_state(callee, caller);
4260 /* copy r1 - r5 args that callee can access. The copy includes parent
4261 * pointers, which connects us up to the liveness chain
4263 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
4264 callee->regs[i] = caller->regs[i];
4266 clear_caller_saved_regs(env, caller->regs);
4268 /* only increment it after check_reg_arg() finished */
4271 /* and go analyze first insn of the callee */
4272 *insn_idx = target_insn;
4274 if (env->log.level & BPF_LOG_LEVEL) {
4275 verbose(env, "caller:\n");
4276 print_verifier_state(env, caller);
4277 verbose(env, "callee:\n");
4278 print_verifier_state(env, callee);
4283 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
4285 struct bpf_verifier_state *state = env->cur_state;
4286 struct bpf_func_state *caller, *callee;
4287 struct bpf_reg_state *r0;
4290 callee = state->frame[state->curframe];
4291 r0 = &callee->regs[BPF_REG_0];
4292 if (r0->type == PTR_TO_STACK) {
4293 /* technically it's ok to return caller's stack pointer
4294 * (or caller's caller's pointer) back to the caller,
4295 * since these pointers are valid. Only current stack
4296 * pointer will be invalid as soon as function exits,
4297 * but let's be conservative
4299 verbose(env, "cannot return stack pointer to the caller\n");
4304 caller = state->frame[state->curframe];
4305 /* return to the caller whatever r0 had in the callee */
4306 caller->regs[BPF_REG_0] = *r0;
4308 /* Transfer references to the caller */
4309 err = transfer_reference_state(caller, callee);
4313 *insn_idx = callee->callsite + 1;
4314 if (env->log.level & BPF_LOG_LEVEL) {
4315 verbose(env, "returning from callee:\n");
4316 print_verifier_state(env, callee);
4317 verbose(env, "to caller at %d:\n", *insn_idx);
4318 print_verifier_state(env, caller);
4320 /* clear everything in the callee */
4321 free_func_state(callee);
4322 state->frame[state->curframe + 1] = NULL;
4326 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
4328 struct bpf_call_arg_meta *meta)
4330 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
4332 if (ret_type != RET_INTEGER ||
4333 (func_id != BPF_FUNC_get_stack &&
4334 func_id != BPF_FUNC_probe_read_str))
4337 ret_reg->smax_value = meta->msize_max_value;
4338 __reg_deduce_bounds(ret_reg);
4339 __reg_bound_offset(ret_reg);
4340 __update_reg_bounds(ret_reg);
4344 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
4345 int func_id, int insn_idx)
4347 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
4348 struct bpf_map *map = meta->map_ptr;
4350 if (func_id != BPF_FUNC_tail_call &&
4351 func_id != BPF_FUNC_map_lookup_elem &&
4352 func_id != BPF_FUNC_map_update_elem &&
4353 func_id != BPF_FUNC_map_delete_elem &&
4354 func_id != BPF_FUNC_map_push_elem &&
4355 func_id != BPF_FUNC_map_pop_elem &&
4356 func_id != BPF_FUNC_map_peek_elem)
4360 verbose(env, "kernel subsystem misconfigured verifier\n");
4364 /* In case of read-only, some additional restrictions
4365 * need to be applied in order to prevent altering the
4366 * state of the map from program side.
4368 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
4369 (func_id == BPF_FUNC_map_delete_elem ||
4370 func_id == BPF_FUNC_map_update_elem ||
4371 func_id == BPF_FUNC_map_push_elem ||
4372 func_id == BPF_FUNC_map_pop_elem)) {
4373 verbose(env, "write into map forbidden\n");
4377 if (!BPF_MAP_PTR(aux->map_ptr_state))
4378 bpf_map_ptr_store(aux, meta->map_ptr,
4379 meta->map_ptr->unpriv_array);
4380 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
4381 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
4382 meta->map_ptr->unpriv_array);
4387 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
4388 int func_id, int insn_idx)
4390 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
4391 struct bpf_reg_state *regs = cur_regs(env), *reg;
4392 struct bpf_map *map = meta->map_ptr;
4397 if (func_id != BPF_FUNC_tail_call)
4399 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
4400 verbose(env, "kernel subsystem misconfigured verifier\n");
4404 range = tnum_range(0, map->max_entries - 1);
4405 reg = ®s[BPF_REG_3];
4407 if (!register_is_const(reg) || !tnum_in(range, reg->var_off)) {
4408 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
4412 err = mark_chain_precision(env, BPF_REG_3);
4416 val = reg->var_off.value;
4417 if (bpf_map_key_unseen(aux))
4418 bpf_map_key_store(aux, val);
4419 else if (!bpf_map_key_poisoned(aux) &&
4420 bpf_map_key_immediate(aux) != val)
4421 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
4425 static int check_reference_leak(struct bpf_verifier_env *env)
4427 struct bpf_func_state *state = cur_func(env);
4430 for (i = 0; i < state->acquired_refs; i++) {
4431 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
4432 state->refs[i].id, state->refs[i].insn_idx);
4434 return state->acquired_refs ? -EINVAL : 0;
4437 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
4439 const struct bpf_func_proto *fn = NULL;
4440 struct bpf_reg_state *regs;
4441 struct bpf_call_arg_meta meta;
4445 /* find function prototype */
4446 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
4447 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
4452 if (env->ops->get_func_proto)
4453 fn = env->ops->get_func_proto(func_id, env->prog);
4455 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
4460 /* eBPF programs must be GPL compatible to use GPL-ed functions */
4461 if (!env->prog->gpl_compatible && fn->gpl_only) {
4462 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
4466 /* With LD_ABS/IND some JITs save/restore skb from r1. */
4467 changes_data = bpf_helper_changes_pkt_data(fn->func);
4468 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
4469 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
4470 func_id_name(func_id), func_id);
4474 memset(&meta, 0, sizeof(meta));
4475 meta.pkt_access = fn->pkt_access;
4477 err = check_func_proto(fn, func_id);
4479 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
4480 func_id_name(func_id), func_id);
4484 meta.func_id = func_id;
4486 for (i = 0; i < 5; i++) {
4487 err = btf_resolve_helper_id(&env->log, fn, i);
4490 err = check_func_arg(env, BPF_REG_1 + i, fn->arg_type[i], &meta);
4495 err = record_func_map(env, &meta, func_id, insn_idx);
4499 err = record_func_key(env, &meta, func_id, insn_idx);
4503 /* Mark slots with STACK_MISC in case of raw mode, stack offset
4504 * is inferred from register state.
4506 for (i = 0; i < meta.access_size; i++) {
4507 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
4508 BPF_WRITE, -1, false);
4513 if (func_id == BPF_FUNC_tail_call) {
4514 err = check_reference_leak(env);
4516 verbose(env, "tail_call would lead to reference leak\n");
4519 } else if (is_release_function(func_id)) {
4520 err = release_reference(env, meta.ref_obj_id);
4522 verbose(env, "func %s#%d reference has not been acquired before\n",
4523 func_id_name(func_id), func_id);
4528 regs = cur_regs(env);
4530 /* check that flags argument in get_local_storage(map, flags) is 0,
4531 * this is required because get_local_storage() can't return an error.
4533 if (func_id == BPF_FUNC_get_local_storage &&
4534 !register_is_null(®s[BPF_REG_2])) {
4535 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
4539 /* reset caller saved regs */
4540 for (i = 0; i < CALLER_SAVED_REGS; i++) {
4541 mark_reg_not_init(env, regs, caller_saved[i]);
4542 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
4545 /* helper call returns 64-bit value. */
4546 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
4548 /* update return register (already marked as written above) */
4549 if (fn->ret_type == RET_INTEGER) {
4550 /* sets type to SCALAR_VALUE */
4551 mark_reg_unknown(env, regs, BPF_REG_0);
4552 } else if (fn->ret_type == RET_VOID) {
4553 regs[BPF_REG_0].type = NOT_INIT;
4554 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
4555 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
4556 /* There is no offset yet applied, variable or fixed */
4557 mark_reg_known_zero(env, regs, BPF_REG_0);
4558 /* remember map_ptr, so that check_map_access()
4559 * can check 'value_size' boundary of memory access
4560 * to map element returned from bpf_map_lookup_elem()
4562 if (meta.map_ptr == NULL) {
4564 "kernel subsystem misconfigured verifier\n");
4567 regs[BPF_REG_0].map_ptr = meta.map_ptr;
4568 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
4569 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
4570 if (map_value_has_spin_lock(meta.map_ptr))
4571 regs[BPF_REG_0].id = ++env->id_gen;
4573 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
4574 regs[BPF_REG_0].id = ++env->id_gen;
4576 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
4577 mark_reg_known_zero(env, regs, BPF_REG_0);
4578 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
4579 regs[BPF_REG_0].id = ++env->id_gen;
4580 } else if (fn->ret_type == RET_PTR_TO_SOCK_COMMON_OR_NULL) {
4581 mark_reg_known_zero(env, regs, BPF_REG_0);
4582 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON_OR_NULL;
4583 regs[BPF_REG_0].id = ++env->id_gen;
4584 } else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
4585 mark_reg_known_zero(env, regs, BPF_REG_0);
4586 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
4587 regs[BPF_REG_0].id = ++env->id_gen;
4589 verbose(env, "unknown return type %d of func %s#%d\n",
4590 fn->ret_type, func_id_name(func_id), func_id);
4594 if (is_ptr_cast_function(func_id)) {
4595 /* For release_reference() */
4596 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
4597 } else if (is_acquire_function(func_id)) {
4598 int id = acquire_reference_state(env, insn_idx);
4602 /* For mark_ptr_or_null_reg() */
4603 regs[BPF_REG_0].id = id;
4604 /* For release_reference() */
4605 regs[BPF_REG_0].ref_obj_id = id;
4608 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
4610 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
4614 if (func_id == BPF_FUNC_get_stack && !env->prog->has_callchain_buf) {
4615 const char *err_str;
4617 #ifdef CONFIG_PERF_EVENTS
4618 err = get_callchain_buffers(sysctl_perf_event_max_stack);
4619 err_str = "cannot get callchain buffer for func %s#%d\n";
4622 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
4625 verbose(env, err_str, func_id_name(func_id), func_id);
4629 env->prog->has_callchain_buf = true;
4633 clear_all_pkt_pointers(env);
4637 static bool signed_add_overflows(s64 a, s64 b)
4639 /* Do the add in u64, where overflow is well-defined */
4640 s64 res = (s64)((u64)a + (u64)b);
4647 static bool signed_add32_overflows(s64 a, s64 b)
4649 /* Do the add in u32, where overflow is well-defined */
4650 s32 res = (s32)((u32)a + (u32)b);
4657 static bool signed_sub_overflows(s32 a, s32 b)
4659 /* Do the sub in u64, where overflow is well-defined */
4660 s64 res = (s64)((u64)a - (u64)b);
4667 static bool signed_sub32_overflows(s32 a, s32 b)
4669 /* Do the sub in u64, where overflow is well-defined */
4670 s32 res = (s32)((u32)a - (u32)b);
4677 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
4678 const struct bpf_reg_state *reg,
4679 enum bpf_reg_type type)
4681 bool known = tnum_is_const(reg->var_off);
4682 s64 val = reg->var_off.value;
4683 s64 smin = reg->smin_value;
4685 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
4686 verbose(env, "math between %s pointer and %lld is not allowed\n",
4687 reg_type_str[type], val);
4691 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
4692 verbose(env, "%s pointer offset %d is not allowed\n",
4693 reg_type_str[type], reg->off);
4697 if (smin == S64_MIN) {
4698 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
4699 reg_type_str[type]);
4703 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
4704 verbose(env, "value %lld makes %s pointer be out of bounds\n",
4705 smin, reg_type_str[type]);
4712 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
4714 return &env->insn_aux_data[env->insn_idx];
4717 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
4718 u32 *ptr_limit, u8 opcode, bool off_is_neg)
4720 bool mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
4721 (opcode == BPF_SUB && !off_is_neg);
4724 switch (ptr_reg->type) {
4726 /* Indirect variable offset stack access is prohibited in
4727 * unprivileged mode so it's not handled here.
4729 off = ptr_reg->off + ptr_reg->var_off.value;
4731 *ptr_limit = MAX_BPF_STACK + off;
4735 case PTR_TO_MAP_VALUE:
4737 *ptr_limit = ptr_reg->umax_value + ptr_reg->off;
4739 off = ptr_reg->smin_value + ptr_reg->off;
4740 *ptr_limit = ptr_reg->map_ptr->value_size - off;
4748 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
4749 const struct bpf_insn *insn)
4751 return env->allow_ptr_leaks || BPF_SRC(insn->code) == BPF_K;
4754 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
4755 u32 alu_state, u32 alu_limit)
4757 /* If we arrived here from different branches with different
4758 * state or limits to sanitize, then this won't work.
4760 if (aux->alu_state &&
4761 (aux->alu_state != alu_state ||
4762 aux->alu_limit != alu_limit))
4765 /* Corresponding fixup done in fixup_bpf_calls(). */
4766 aux->alu_state = alu_state;
4767 aux->alu_limit = alu_limit;
4771 static int sanitize_val_alu(struct bpf_verifier_env *env,
4772 struct bpf_insn *insn)
4774 struct bpf_insn_aux_data *aux = cur_aux(env);
4776 if (can_skip_alu_sanitation(env, insn))
4779 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
4782 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
4783 struct bpf_insn *insn,
4784 const struct bpf_reg_state *ptr_reg,
4785 struct bpf_reg_state *dst_reg,
4788 struct bpf_verifier_state *vstate = env->cur_state;
4789 struct bpf_insn_aux_data *aux = cur_aux(env);
4790 bool ptr_is_dst_reg = ptr_reg == dst_reg;
4791 u8 opcode = BPF_OP(insn->code);
4792 u32 alu_state, alu_limit;
4793 struct bpf_reg_state tmp;
4796 if (can_skip_alu_sanitation(env, insn))
4799 /* We already marked aux for masking from non-speculative
4800 * paths, thus we got here in the first place. We only care
4801 * to explore bad access from here.
4803 if (vstate->speculative)
4806 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
4807 alu_state |= ptr_is_dst_reg ?
4808 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
4810 if (retrieve_ptr_limit(ptr_reg, &alu_limit, opcode, off_is_neg))
4812 if (update_alu_sanitation_state(aux, alu_state, alu_limit))
4815 /* Simulate and find potential out-of-bounds access under
4816 * speculative execution from truncation as a result of
4817 * masking when off was not within expected range. If off
4818 * sits in dst, then we temporarily need to move ptr there
4819 * to simulate dst (== 0) +/-= ptr. Needed, for example,
4820 * for cases where we use K-based arithmetic in one direction
4821 * and truncated reg-based in the other in order to explore
4824 if (!ptr_is_dst_reg) {
4826 *dst_reg = *ptr_reg;
4828 ret = push_stack(env, env->insn_idx + 1, env->insn_idx, true);
4829 if (!ptr_is_dst_reg && ret)
4831 return !ret ? -EFAULT : 0;
4834 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
4835 * Caller should also handle BPF_MOV case separately.
4836 * If we return -EACCES, caller may want to try again treating pointer as a
4837 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
4839 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
4840 struct bpf_insn *insn,
4841 const struct bpf_reg_state *ptr_reg,
4842 const struct bpf_reg_state *off_reg)
4844 struct bpf_verifier_state *vstate = env->cur_state;
4845 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4846 struct bpf_reg_state *regs = state->regs, *dst_reg;
4847 bool known = tnum_is_const(off_reg->var_off);
4848 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
4849 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
4850 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
4851 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
4852 u32 dst = insn->dst_reg, src = insn->src_reg;
4853 u8 opcode = BPF_OP(insn->code);
4856 dst_reg = ®s[dst];
4858 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
4859 smin_val > smax_val || umin_val > umax_val) {
4860 /* Taint dst register if offset had invalid bounds derived from
4861 * e.g. dead branches.
4863 __mark_reg_unknown(env, dst_reg);
4867 if (BPF_CLASS(insn->code) != BPF_ALU64) {
4868 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
4870 "R%d 32-bit pointer arithmetic prohibited\n",
4875 switch (ptr_reg->type) {
4876 case PTR_TO_MAP_VALUE_OR_NULL:
4877 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
4878 dst, reg_type_str[ptr_reg->type]);
4880 case CONST_PTR_TO_MAP:
4881 case PTR_TO_PACKET_END:
4883 case PTR_TO_SOCKET_OR_NULL:
4884 case PTR_TO_SOCK_COMMON:
4885 case PTR_TO_SOCK_COMMON_OR_NULL:
4886 case PTR_TO_TCP_SOCK:
4887 case PTR_TO_TCP_SOCK_OR_NULL:
4888 case PTR_TO_XDP_SOCK:
4889 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
4890 dst, reg_type_str[ptr_reg->type]);
4892 case PTR_TO_MAP_VALUE:
4893 if (!env->allow_ptr_leaks && !known && (smin_val < 0) != (smax_val < 0)) {
4894 verbose(env, "R%d has unknown scalar with mixed signed bounds, pointer arithmetic with it prohibited for !root\n",
4895 off_reg == dst_reg ? dst : src);
4903 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
4904 * The id may be overwritten later if we create a new variable offset.
4906 dst_reg->type = ptr_reg->type;
4907 dst_reg->id = ptr_reg->id;
4909 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
4910 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
4913 /* pointer types do not carry 32-bit bounds at the moment. */
4914 __mark_reg32_unbounded(dst_reg);
4918 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
4920 verbose(env, "R%d tried to add from different maps or paths\n", dst);
4923 /* We can take a fixed offset as long as it doesn't overflow
4924 * the s32 'off' field
4926 if (known && (ptr_reg->off + smin_val ==
4927 (s64)(s32)(ptr_reg->off + smin_val))) {
4928 /* pointer += K. Accumulate it into fixed offset */
4929 dst_reg->smin_value = smin_ptr;
4930 dst_reg->smax_value = smax_ptr;
4931 dst_reg->umin_value = umin_ptr;
4932 dst_reg->umax_value = umax_ptr;
4933 dst_reg->var_off = ptr_reg->var_off;
4934 dst_reg->off = ptr_reg->off + smin_val;
4935 dst_reg->raw = ptr_reg->raw;
4938 /* A new variable offset is created. Note that off_reg->off
4939 * == 0, since it's a scalar.
4940 * dst_reg gets the pointer type and since some positive
4941 * integer value was added to the pointer, give it a new 'id'
4942 * if it's a PTR_TO_PACKET.
4943 * this creates a new 'base' pointer, off_reg (variable) gets
4944 * added into the variable offset, and we copy the fixed offset
4947 if (signed_add_overflows(smin_ptr, smin_val) ||
4948 signed_add_overflows(smax_ptr, smax_val)) {
4949 dst_reg->smin_value = S64_MIN;
4950 dst_reg->smax_value = S64_MAX;
4952 dst_reg->smin_value = smin_ptr + smin_val;
4953 dst_reg->smax_value = smax_ptr + smax_val;
4955 if (umin_ptr + umin_val < umin_ptr ||
4956 umax_ptr + umax_val < umax_ptr) {
4957 dst_reg->umin_value = 0;
4958 dst_reg->umax_value = U64_MAX;
4960 dst_reg->umin_value = umin_ptr + umin_val;
4961 dst_reg->umax_value = umax_ptr + umax_val;
4963 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
4964 dst_reg->off = ptr_reg->off;
4965 dst_reg->raw = ptr_reg->raw;
4966 if (reg_is_pkt_pointer(ptr_reg)) {
4967 dst_reg->id = ++env->id_gen;
4968 /* something was added to pkt_ptr, set range to zero */
4973 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
4975 verbose(env, "R%d tried to sub from different maps or paths\n", dst);
4978 if (dst_reg == off_reg) {
4979 /* scalar -= pointer. Creates an unknown scalar */
4980 verbose(env, "R%d tried to subtract pointer from scalar\n",
4984 /* We don't allow subtraction from FP, because (according to
4985 * test_verifier.c test "invalid fp arithmetic", JITs might not
4986 * be able to deal with it.
4988 if (ptr_reg->type == PTR_TO_STACK) {
4989 verbose(env, "R%d subtraction from stack pointer prohibited\n",
4993 if (known && (ptr_reg->off - smin_val ==
4994 (s64)(s32)(ptr_reg->off - smin_val))) {
4995 /* pointer -= K. Subtract it from fixed offset */
4996 dst_reg->smin_value = smin_ptr;
4997 dst_reg->smax_value = smax_ptr;
4998 dst_reg->umin_value = umin_ptr;
4999 dst_reg->umax_value = umax_ptr;
5000 dst_reg->var_off = ptr_reg->var_off;
5001 dst_reg->id = ptr_reg->id;
5002 dst_reg->off = ptr_reg->off - smin_val;
5003 dst_reg->raw = ptr_reg->raw;
5006 /* A new variable offset is created. If the subtrahend is known
5007 * nonnegative, then any reg->range we had before is still good.
5009 if (signed_sub_overflows(smin_ptr, smax_val) ||
5010 signed_sub_overflows(smax_ptr, smin_val)) {
5011 /* Overflow possible, we know nothing */
5012 dst_reg->smin_value = S64_MIN;
5013 dst_reg->smax_value = S64_MAX;
5015 dst_reg->smin_value = smin_ptr - smax_val;
5016 dst_reg->smax_value = smax_ptr - smin_val;
5018 if (umin_ptr < umax_val) {
5019 /* Overflow possible, we know nothing */
5020 dst_reg->umin_value = 0;
5021 dst_reg->umax_value = U64_MAX;
5023 /* Cannot overflow (as long as bounds are consistent) */
5024 dst_reg->umin_value = umin_ptr - umax_val;
5025 dst_reg->umax_value = umax_ptr - umin_val;
5027 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
5028 dst_reg->off = ptr_reg->off;
5029 dst_reg->raw = ptr_reg->raw;
5030 if (reg_is_pkt_pointer(ptr_reg)) {
5031 dst_reg->id = ++env->id_gen;
5032 /* something was added to pkt_ptr, set range to zero */
5040 /* bitwise ops on pointers are troublesome, prohibit. */
5041 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
5042 dst, bpf_alu_string[opcode >> 4]);
5045 /* other operators (e.g. MUL,LSH) produce non-pointer results */
5046 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
5047 dst, bpf_alu_string[opcode >> 4]);
5051 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
5054 __update_reg_bounds(dst_reg);
5055 __reg_deduce_bounds(dst_reg);
5056 __reg_bound_offset(dst_reg);
5058 /* For unprivileged we require that resulting offset must be in bounds
5059 * in order to be able to sanitize access later on.
5061 if (!env->allow_ptr_leaks) {
5062 if (dst_reg->type == PTR_TO_MAP_VALUE &&
5063 check_map_access(env, dst, dst_reg->off, 1, false)) {
5064 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
5065 "prohibited for !root\n", dst);
5067 } else if (dst_reg->type == PTR_TO_STACK &&
5068 check_stack_access(env, dst_reg, dst_reg->off +
5069 dst_reg->var_off.value, 1)) {
5070 verbose(env, "R%d stack pointer arithmetic goes out of range, "
5071 "prohibited for !root\n", dst);
5079 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
5080 struct bpf_reg_state *src_reg)
5082 s32 smin_val = src_reg->s32_min_value;
5083 s32 smax_val = src_reg->s32_max_value;
5084 u32 umin_val = src_reg->u32_min_value;
5085 u32 umax_val = src_reg->u32_max_value;
5087 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
5088 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
5089 dst_reg->s32_min_value = S32_MIN;
5090 dst_reg->s32_max_value = S32_MAX;
5092 dst_reg->s32_min_value += smin_val;
5093 dst_reg->s32_max_value += smax_val;
5095 if (dst_reg->u32_min_value + umin_val < umin_val ||
5096 dst_reg->u32_max_value + umax_val < umax_val) {
5097 dst_reg->u32_min_value = 0;
5098 dst_reg->u32_max_value = U32_MAX;
5100 dst_reg->u32_min_value += umin_val;
5101 dst_reg->u32_max_value += umax_val;
5105 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
5106 struct bpf_reg_state *src_reg)
5108 s64 smin_val = src_reg->smin_value;
5109 s64 smax_val = src_reg->smax_value;
5110 u64 umin_val = src_reg->umin_value;
5111 u64 umax_val = src_reg->umax_value;
5113 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
5114 signed_add_overflows(dst_reg->smax_value, smax_val)) {
5115 dst_reg->smin_value = S64_MIN;
5116 dst_reg->smax_value = S64_MAX;
5118 dst_reg->smin_value += smin_val;
5119 dst_reg->smax_value += smax_val;
5121 if (dst_reg->umin_value + umin_val < umin_val ||
5122 dst_reg->umax_value + umax_val < umax_val) {
5123 dst_reg->umin_value = 0;
5124 dst_reg->umax_value = U64_MAX;
5126 dst_reg->umin_value += umin_val;
5127 dst_reg->umax_value += umax_val;
5131 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
5132 struct bpf_reg_state *src_reg)
5134 s32 smin_val = src_reg->s32_min_value;
5135 s32 smax_val = src_reg->s32_max_value;
5136 u32 umin_val = src_reg->u32_min_value;
5137 u32 umax_val = src_reg->u32_max_value;
5139 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
5140 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
5141 /* Overflow possible, we know nothing */
5142 dst_reg->s32_min_value = S32_MIN;
5143 dst_reg->s32_max_value = S32_MAX;
5145 dst_reg->s32_min_value -= smax_val;
5146 dst_reg->s32_max_value -= smin_val;
5148 if (dst_reg->u32_min_value < umax_val) {
5149 /* Overflow possible, we know nothing */
5150 dst_reg->u32_min_value = 0;
5151 dst_reg->u32_max_value = U32_MAX;
5153 /* Cannot overflow (as long as bounds are consistent) */
5154 dst_reg->u32_min_value -= umax_val;
5155 dst_reg->u32_max_value -= umin_val;
5159 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
5160 struct bpf_reg_state *src_reg)
5162 s64 smin_val = src_reg->smin_value;
5163 s64 smax_val = src_reg->smax_value;
5164 u64 umin_val = src_reg->umin_value;
5165 u64 umax_val = src_reg->umax_value;
5167 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
5168 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
5169 /* Overflow possible, we know nothing */
5170 dst_reg->smin_value = S64_MIN;
5171 dst_reg->smax_value = S64_MAX;
5173 dst_reg->smin_value -= smax_val;
5174 dst_reg->smax_value -= smin_val;
5176 if (dst_reg->umin_value < umax_val) {
5177 /* Overflow possible, we know nothing */
5178 dst_reg->umin_value = 0;
5179 dst_reg->umax_value = U64_MAX;
5181 /* Cannot overflow (as long as bounds are consistent) */
5182 dst_reg->umin_value -= umax_val;
5183 dst_reg->umax_value -= umin_val;
5187 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
5188 struct bpf_reg_state *src_reg)
5190 s32 smin_val = src_reg->s32_min_value;
5191 u32 umin_val = src_reg->u32_min_value;
5192 u32 umax_val = src_reg->u32_max_value;
5194 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
5195 /* Ain't nobody got time to multiply that sign */
5196 __mark_reg32_unbounded(dst_reg);
5199 /* Both values are positive, so we can work with unsigned and
5200 * copy the result to signed (unless it exceeds S32_MAX).
5202 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
5203 /* Potential overflow, we know nothing */
5204 __mark_reg32_unbounded(dst_reg);
5207 dst_reg->u32_min_value *= umin_val;
5208 dst_reg->u32_max_value *= umax_val;
5209 if (dst_reg->u32_max_value > S32_MAX) {
5210 /* Overflow possible, we know nothing */
5211 dst_reg->s32_min_value = S32_MIN;
5212 dst_reg->s32_max_value = S32_MAX;
5214 dst_reg->s32_min_value = dst_reg->u32_min_value;
5215 dst_reg->s32_max_value = dst_reg->u32_max_value;
5219 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
5220 struct bpf_reg_state *src_reg)
5222 s64 smin_val = src_reg->smin_value;
5223 u64 umin_val = src_reg->umin_value;
5224 u64 umax_val = src_reg->umax_value;
5226 if (smin_val < 0 || dst_reg->smin_value < 0) {
5227 /* Ain't nobody got time to multiply that sign */
5228 __mark_reg64_unbounded(dst_reg);
5231 /* Both values are positive, so we can work with unsigned and
5232 * copy the result to signed (unless it exceeds S64_MAX).
5234 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
5235 /* Potential overflow, we know nothing */
5236 __mark_reg64_unbounded(dst_reg);
5239 dst_reg->umin_value *= umin_val;
5240 dst_reg->umax_value *= umax_val;
5241 if (dst_reg->umax_value > S64_MAX) {
5242 /* Overflow possible, we know nothing */
5243 dst_reg->smin_value = S64_MIN;
5244 dst_reg->smax_value = S64_MAX;
5246 dst_reg->smin_value = dst_reg->umin_value;
5247 dst_reg->smax_value = dst_reg->umax_value;
5251 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
5252 struct bpf_reg_state *src_reg)
5254 bool src_known = tnum_subreg_is_const(src_reg->var_off);
5255 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
5256 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
5257 s32 smin_val = src_reg->s32_min_value;
5258 u32 umax_val = src_reg->u32_max_value;
5260 /* Assuming scalar64_min_max_and will be called so its safe
5261 * to skip updating register for known 32-bit case.
5263 if (src_known && dst_known)
5266 /* We get our minimum from the var_off, since that's inherently
5267 * bitwise. Our maximum is the minimum of the operands' maxima.
5269 dst_reg->u32_min_value = var32_off.value;
5270 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
5271 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
5272 /* Lose signed bounds when ANDing negative numbers,
5273 * ain't nobody got time for that.
5275 dst_reg->s32_min_value = S32_MIN;
5276 dst_reg->s32_max_value = S32_MAX;
5278 /* ANDing two positives gives a positive, so safe to
5279 * cast result into s64.
5281 dst_reg->s32_min_value = dst_reg->u32_min_value;
5282 dst_reg->s32_max_value = dst_reg->u32_max_value;
5287 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
5288 struct bpf_reg_state *src_reg)
5290 bool src_known = tnum_is_const(src_reg->var_off);
5291 bool dst_known = tnum_is_const(dst_reg->var_off);
5292 s64 smin_val = src_reg->smin_value;
5293 u64 umax_val = src_reg->umax_value;
5295 if (src_known && dst_known) {
5296 __mark_reg_known(dst_reg, dst_reg->var_off.value &
5297 src_reg->var_off.value);
5301 /* We get our minimum from the var_off, since that's inherently
5302 * bitwise. Our maximum is the minimum of the operands' maxima.
5304 dst_reg->umin_value = dst_reg->var_off.value;
5305 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
5306 if (dst_reg->smin_value < 0 || smin_val < 0) {
5307 /* Lose signed bounds when ANDing negative numbers,
5308 * ain't nobody got time for that.
5310 dst_reg->smin_value = S64_MIN;
5311 dst_reg->smax_value = S64_MAX;
5313 /* ANDing two positives gives a positive, so safe to
5314 * cast result into s64.
5316 dst_reg->smin_value = dst_reg->umin_value;
5317 dst_reg->smax_value = dst_reg->umax_value;
5319 /* We may learn something more from the var_off */
5320 __update_reg_bounds(dst_reg);
5323 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
5324 struct bpf_reg_state *src_reg)
5326 bool src_known = tnum_subreg_is_const(src_reg->var_off);
5327 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
5328 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
5329 s32 smin_val = src_reg->smin_value;
5330 u32 umin_val = src_reg->umin_value;
5332 /* Assuming scalar64_min_max_or will be called so it is safe
5333 * to skip updating register for known case.
5335 if (src_known && dst_known)
5338 /* We get our maximum from the var_off, and our minimum is the
5339 * maximum of the operands' minima
5341 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
5342 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
5343 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
5344 /* Lose signed bounds when ORing negative numbers,
5345 * ain't nobody got time for that.
5347 dst_reg->s32_min_value = S32_MIN;
5348 dst_reg->s32_max_value = S32_MAX;
5350 /* ORing two positives gives a positive, so safe to
5351 * cast result into s64.
5353 dst_reg->s32_min_value = dst_reg->umin_value;
5354 dst_reg->s32_max_value = dst_reg->umax_value;
5358 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
5359 struct bpf_reg_state *src_reg)
5361 bool src_known = tnum_is_const(src_reg->var_off);
5362 bool dst_known = tnum_is_const(dst_reg->var_off);
5363 s64 smin_val = src_reg->smin_value;
5364 u64 umin_val = src_reg->umin_value;
5366 if (src_known && dst_known) {
5367 __mark_reg_known(dst_reg, dst_reg->var_off.value |
5368 src_reg->var_off.value);
5372 /* We get our maximum from the var_off, and our minimum is the
5373 * maximum of the operands' minima
5375 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
5376 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
5377 if (dst_reg->smin_value < 0 || smin_val < 0) {
5378 /* Lose signed bounds when ORing negative numbers,
5379 * ain't nobody got time for that.
5381 dst_reg->smin_value = S64_MIN;
5382 dst_reg->smax_value = S64_MAX;
5384 /* ORing two positives gives a positive, so safe to
5385 * cast result into s64.
5387 dst_reg->smin_value = dst_reg->umin_value;
5388 dst_reg->smax_value = dst_reg->umax_value;
5390 /* We may learn something more from the var_off */
5391 __update_reg_bounds(dst_reg);
5394 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
5395 u64 umin_val, u64 umax_val)
5397 /* We lose all sign bit information (except what we can pick
5400 dst_reg->s32_min_value = S32_MIN;
5401 dst_reg->s32_max_value = S32_MAX;
5402 /* If we might shift our top bit out, then we know nothing */
5403 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
5404 dst_reg->u32_min_value = 0;
5405 dst_reg->u32_max_value = U32_MAX;
5407 dst_reg->u32_min_value <<= umin_val;
5408 dst_reg->u32_max_value <<= umax_val;
5412 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
5413 struct bpf_reg_state *src_reg)
5415 u32 umax_val = src_reg->u32_max_value;
5416 u32 umin_val = src_reg->u32_min_value;
5417 /* u32 alu operation will zext upper bits */
5418 struct tnum subreg = tnum_subreg(dst_reg->var_off);
5420 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
5421 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
5422 /* Not required but being careful mark reg64 bounds as unknown so
5423 * that we are forced to pick them up from tnum and zext later and
5424 * if some path skips this step we are still safe.
5426 __mark_reg64_unbounded(dst_reg);
5427 __update_reg32_bounds(dst_reg);
5430 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
5431 u64 umin_val, u64 umax_val)
5433 /* Special case <<32 because it is a common compiler pattern to sign
5434 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
5435 * positive we know this shift will also be positive so we can track
5436 * bounds correctly. Otherwise we lose all sign bit information except
5437 * what we can pick up from var_off. Perhaps we can generalize this
5438 * later to shifts of any length.
5440 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
5441 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
5443 dst_reg->smax_value = S64_MAX;
5445 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
5446 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
5448 dst_reg->smin_value = S64_MIN;
5450 /* If we might shift our top bit out, then we know nothing */
5451 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
5452 dst_reg->umin_value = 0;
5453 dst_reg->umax_value = U64_MAX;
5455 dst_reg->umin_value <<= umin_val;
5456 dst_reg->umax_value <<= umax_val;
5460 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
5461 struct bpf_reg_state *src_reg)
5463 u64 umax_val = src_reg->umax_value;
5464 u64 umin_val = src_reg->umin_value;
5466 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
5467 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
5468 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
5470 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
5471 /* We may learn something more from the var_off */
5472 __update_reg_bounds(dst_reg);
5475 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
5476 struct bpf_reg_state *src_reg)
5478 struct tnum subreg = tnum_subreg(dst_reg->var_off);
5479 u32 umax_val = src_reg->u32_max_value;
5480 u32 umin_val = src_reg->u32_min_value;
5482 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
5483 * be negative, then either:
5484 * 1) src_reg might be zero, so the sign bit of the result is
5485 * unknown, so we lose our signed bounds
5486 * 2) it's known negative, thus the unsigned bounds capture the
5488 * 3) the signed bounds cross zero, so they tell us nothing
5490 * If the value in dst_reg is known nonnegative, then again the
5491 * unsigned bounts capture the signed bounds.
5492 * Thus, in all cases it suffices to blow away our signed bounds
5493 * and rely on inferring new ones from the unsigned bounds and
5494 * var_off of the result.
5496 dst_reg->s32_min_value = S32_MIN;
5497 dst_reg->s32_max_value = S32_MAX;
5499 dst_reg->var_off = tnum_rshift(subreg, umin_val);
5500 dst_reg->u32_min_value >>= umax_val;
5501 dst_reg->u32_max_value >>= umin_val;
5503 __mark_reg64_unbounded(dst_reg);
5504 __update_reg32_bounds(dst_reg);
5507 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
5508 struct bpf_reg_state *src_reg)
5510 u64 umax_val = src_reg->umax_value;
5511 u64 umin_val = src_reg->umin_value;
5513 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
5514 * be negative, then either:
5515 * 1) src_reg might be zero, so the sign bit of the result is
5516 * unknown, so we lose our signed bounds
5517 * 2) it's known negative, thus the unsigned bounds capture the
5519 * 3) the signed bounds cross zero, so they tell us nothing
5521 * If the value in dst_reg is known nonnegative, then again the
5522 * unsigned bounts capture the signed bounds.
5523 * Thus, in all cases it suffices to blow away our signed bounds
5524 * and rely on inferring new ones from the unsigned bounds and
5525 * var_off of the result.
5527 dst_reg->smin_value = S64_MIN;
5528 dst_reg->smax_value = S64_MAX;
5529 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
5530 dst_reg->umin_value >>= umax_val;
5531 dst_reg->umax_value >>= umin_val;
5533 /* Its not easy to operate on alu32 bounds here because it depends
5534 * on bits being shifted in. Take easy way out and mark unbounded
5535 * so we can recalculate later from tnum.
5537 __mark_reg32_unbounded(dst_reg);
5538 __update_reg_bounds(dst_reg);
5541 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
5542 struct bpf_reg_state *src_reg)
5544 u64 umin_val = src_reg->u32_min_value;
5546 /* Upon reaching here, src_known is true and
5547 * umax_val is equal to umin_val.
5549 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
5550 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
5552 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
5554 /* blow away the dst_reg umin_value/umax_value and rely on
5555 * dst_reg var_off to refine the result.
5557 dst_reg->u32_min_value = 0;
5558 dst_reg->u32_max_value = U32_MAX;
5560 __mark_reg64_unbounded(dst_reg);
5561 __update_reg32_bounds(dst_reg);
5564 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
5565 struct bpf_reg_state *src_reg)
5567 u64 umin_val = src_reg->umin_value;
5569 /* Upon reaching here, src_known is true and umax_val is equal
5572 dst_reg->smin_value >>= umin_val;
5573 dst_reg->smax_value >>= umin_val;
5575 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
5577 /* blow away the dst_reg umin_value/umax_value and rely on
5578 * dst_reg var_off to refine the result.
5580 dst_reg->umin_value = 0;
5581 dst_reg->umax_value = U64_MAX;
5583 /* Its not easy to operate on alu32 bounds here because it depends
5584 * on bits being shifted in from upper 32-bits. Take easy way out
5585 * and mark unbounded so we can recalculate later from tnum.
5587 __mark_reg32_unbounded(dst_reg);
5588 __update_reg_bounds(dst_reg);
5591 /* WARNING: This function does calculations on 64-bit values, but the actual
5592 * execution may occur on 32-bit values. Therefore, things like bitshifts
5593 * need extra checks in the 32-bit case.
5595 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
5596 struct bpf_insn *insn,
5597 struct bpf_reg_state *dst_reg,
5598 struct bpf_reg_state src_reg)
5600 struct bpf_reg_state *regs = cur_regs(env);
5601 u8 opcode = BPF_OP(insn->code);
5602 bool src_known, dst_known;
5603 s64 smin_val, smax_val;
5604 u64 umin_val, umax_val;
5605 s32 s32_min_val, s32_max_val;
5606 u32 u32_min_val, u32_max_val;
5607 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
5608 u32 dst = insn->dst_reg;
5610 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
5612 smin_val = src_reg.smin_value;
5613 smax_val = src_reg.smax_value;
5614 umin_val = src_reg.umin_value;
5615 umax_val = src_reg.umax_value;
5617 s32_min_val = src_reg.s32_min_value;
5618 s32_max_val = src_reg.s32_max_value;
5619 u32_min_val = src_reg.u32_min_value;
5620 u32_max_val = src_reg.u32_max_value;
5623 src_known = tnum_subreg_is_const(src_reg.var_off);
5624 dst_known = tnum_subreg_is_const(dst_reg->var_off);
5626 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
5627 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
5628 /* Taint dst register if offset had invalid bounds
5629 * derived from e.g. dead branches.
5631 __mark_reg_unknown(env, dst_reg);
5635 src_known = tnum_is_const(src_reg.var_off);
5636 dst_known = tnum_is_const(dst_reg->var_off);
5638 (smin_val != smax_val || umin_val != umax_val)) ||
5639 smin_val > smax_val || umin_val > umax_val) {
5640 /* Taint dst register if offset had invalid bounds
5641 * derived from e.g. dead branches.
5643 __mark_reg_unknown(env, dst_reg);
5649 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
5650 __mark_reg_unknown(env, dst_reg);
5654 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
5655 * There are two classes of instructions: The first class we track both
5656 * alu32 and alu64 sign/unsigned bounds independently this provides the
5657 * greatest amount of precision when alu operations are mixed with jmp32
5658 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
5659 * and BPF_OR. This is possible because these ops have fairly easy to
5660 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
5661 * See alu32 verifier tests for examples. The second class of
5662 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
5663 * with regards to tracking sign/unsigned bounds because the bits may
5664 * cross subreg boundaries in the alu64 case. When this happens we mark
5665 * the reg unbounded in the subreg bound space and use the resulting
5666 * tnum to calculate an approximation of the sign/unsigned bounds.
5670 ret = sanitize_val_alu(env, insn);
5672 verbose(env, "R%d tried to add from different pointers or scalars\n", dst);
5675 scalar32_min_max_add(dst_reg, &src_reg);
5676 scalar_min_max_add(dst_reg, &src_reg);
5677 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
5680 ret = sanitize_val_alu(env, insn);
5682 verbose(env, "R%d tried to sub from different pointers or scalars\n", dst);
5685 scalar32_min_max_sub(dst_reg, &src_reg);
5686 scalar_min_max_sub(dst_reg, &src_reg);
5687 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
5690 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
5691 scalar32_min_max_mul(dst_reg, &src_reg);
5692 scalar_min_max_mul(dst_reg, &src_reg);
5695 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
5696 scalar32_min_max_and(dst_reg, &src_reg);
5697 scalar_min_max_and(dst_reg, &src_reg);
5700 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
5701 scalar32_min_max_or(dst_reg, &src_reg);
5702 scalar_min_max_or(dst_reg, &src_reg);
5705 if (umax_val >= insn_bitness) {
5706 /* Shifts greater than 31 or 63 are undefined.
5707 * This includes shifts by a negative number.
5709 mark_reg_unknown(env, regs, insn->dst_reg);
5713 scalar32_min_max_lsh(dst_reg, &src_reg);
5715 scalar_min_max_lsh(dst_reg, &src_reg);
5718 if (umax_val >= insn_bitness) {
5719 /* Shifts greater than 31 or 63 are undefined.
5720 * This includes shifts by a negative number.
5722 mark_reg_unknown(env, regs, insn->dst_reg);
5726 scalar32_min_max_rsh(dst_reg, &src_reg);
5728 scalar_min_max_rsh(dst_reg, &src_reg);
5731 if (umax_val >= insn_bitness) {
5732 /* Shifts greater than 31 or 63 are undefined.
5733 * This includes shifts by a negative number.
5735 mark_reg_unknown(env, regs, insn->dst_reg);
5739 scalar32_min_max_arsh(dst_reg, &src_reg);
5741 scalar_min_max_arsh(dst_reg, &src_reg);
5744 mark_reg_unknown(env, regs, insn->dst_reg);
5748 /* ALU32 ops are zero extended into 64bit register */
5750 zext_32_to_64(dst_reg);
5752 __update_reg_bounds(dst_reg);
5753 __reg_deduce_bounds(dst_reg);
5754 __reg_bound_offset(dst_reg);
5758 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
5761 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
5762 struct bpf_insn *insn)
5764 struct bpf_verifier_state *vstate = env->cur_state;
5765 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5766 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
5767 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
5768 u8 opcode = BPF_OP(insn->code);
5771 dst_reg = ®s[insn->dst_reg];
5773 if (dst_reg->type != SCALAR_VALUE)
5775 if (BPF_SRC(insn->code) == BPF_X) {
5776 src_reg = ®s[insn->src_reg];
5777 if (src_reg->type != SCALAR_VALUE) {
5778 if (dst_reg->type != SCALAR_VALUE) {
5779 /* Combining two pointers by any ALU op yields
5780 * an arbitrary scalar. Disallow all math except
5781 * pointer subtraction
5783 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
5784 mark_reg_unknown(env, regs, insn->dst_reg);
5787 verbose(env, "R%d pointer %s pointer prohibited\n",
5789 bpf_alu_string[opcode >> 4]);
5792 /* scalar += pointer
5793 * This is legal, but we have to reverse our
5794 * src/dest handling in computing the range
5796 err = mark_chain_precision(env, insn->dst_reg);
5799 return adjust_ptr_min_max_vals(env, insn,
5802 } else if (ptr_reg) {
5803 /* pointer += scalar */
5804 err = mark_chain_precision(env, insn->src_reg);
5807 return adjust_ptr_min_max_vals(env, insn,
5811 /* Pretend the src is a reg with a known value, since we only
5812 * need to be able to read from this state.
5814 off_reg.type = SCALAR_VALUE;
5815 __mark_reg_known(&off_reg, insn->imm);
5817 if (ptr_reg) /* pointer += K */
5818 return adjust_ptr_min_max_vals(env, insn,
5822 /* Got here implies adding two SCALAR_VALUEs */
5823 if (WARN_ON_ONCE(ptr_reg)) {
5824 print_verifier_state(env, state);
5825 verbose(env, "verifier internal error: unexpected ptr_reg\n");
5828 if (WARN_ON(!src_reg)) {
5829 print_verifier_state(env, state);
5830 verbose(env, "verifier internal error: no src_reg\n");
5833 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
5836 /* check validity of 32-bit and 64-bit arithmetic operations */
5837 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
5839 struct bpf_reg_state *regs = cur_regs(env);
5840 u8 opcode = BPF_OP(insn->code);
5843 if (opcode == BPF_END || opcode == BPF_NEG) {
5844 if (opcode == BPF_NEG) {
5845 if (BPF_SRC(insn->code) != 0 ||
5846 insn->src_reg != BPF_REG_0 ||
5847 insn->off != 0 || insn->imm != 0) {
5848 verbose(env, "BPF_NEG uses reserved fields\n");
5852 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
5853 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
5854 BPF_CLASS(insn->code) == BPF_ALU64) {
5855 verbose(env, "BPF_END uses reserved fields\n");
5860 /* check src operand */
5861 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5865 if (is_pointer_value(env, insn->dst_reg)) {
5866 verbose(env, "R%d pointer arithmetic prohibited\n",
5871 /* check dest operand */
5872 err = check_reg_arg(env, insn->dst_reg, DST_OP);
5876 } else if (opcode == BPF_MOV) {
5878 if (BPF_SRC(insn->code) == BPF_X) {
5879 if (insn->imm != 0 || insn->off != 0) {
5880 verbose(env, "BPF_MOV uses reserved fields\n");
5884 /* check src operand */
5885 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5889 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
5890 verbose(env, "BPF_MOV uses reserved fields\n");
5895 /* check dest operand, mark as required later */
5896 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
5900 if (BPF_SRC(insn->code) == BPF_X) {
5901 struct bpf_reg_state *src_reg = regs + insn->src_reg;
5902 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
5904 if (BPF_CLASS(insn->code) == BPF_ALU64) {
5906 * copy register state to dest reg
5908 *dst_reg = *src_reg;
5909 dst_reg->live |= REG_LIVE_WRITTEN;
5910 dst_reg->subreg_def = DEF_NOT_SUBREG;
5913 if (is_pointer_value(env, insn->src_reg)) {
5915 "R%d partial copy of pointer\n",
5918 } else if (src_reg->type == SCALAR_VALUE) {
5919 *dst_reg = *src_reg;
5920 dst_reg->live |= REG_LIVE_WRITTEN;
5921 dst_reg->subreg_def = env->insn_idx + 1;
5923 mark_reg_unknown(env, regs,
5926 zext_32_to_64(dst_reg);
5930 * remember the value we stored into this reg
5932 /* clear any state __mark_reg_known doesn't set */
5933 mark_reg_unknown(env, regs, insn->dst_reg);
5934 regs[insn->dst_reg].type = SCALAR_VALUE;
5935 if (BPF_CLASS(insn->code) == BPF_ALU64) {
5936 __mark_reg_known(regs + insn->dst_reg,
5939 __mark_reg_known(regs + insn->dst_reg,
5944 } else if (opcode > BPF_END) {
5945 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
5948 } else { /* all other ALU ops: and, sub, xor, add, ... */
5950 if (BPF_SRC(insn->code) == BPF_X) {
5951 if (insn->imm != 0 || insn->off != 0) {
5952 verbose(env, "BPF_ALU uses reserved fields\n");
5955 /* check src1 operand */
5956 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5960 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
5961 verbose(env, "BPF_ALU uses reserved fields\n");
5966 /* check src2 operand */
5967 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5971 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
5972 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
5973 verbose(env, "div by zero\n");
5977 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
5978 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
5979 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
5981 if (insn->imm < 0 || insn->imm >= size) {
5982 verbose(env, "invalid shift %d\n", insn->imm);
5987 /* check dest operand */
5988 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
5992 return adjust_reg_min_max_vals(env, insn);
5998 static void __find_good_pkt_pointers(struct bpf_func_state *state,
5999 struct bpf_reg_state *dst_reg,
6000 enum bpf_reg_type type, u16 new_range)
6002 struct bpf_reg_state *reg;
6005 for (i = 0; i < MAX_BPF_REG; i++) {
6006 reg = &state->regs[i];
6007 if (reg->type == type && reg->id == dst_reg->id)
6008 /* keep the maximum range already checked */
6009 reg->range = max(reg->range, new_range);
6012 bpf_for_each_spilled_reg(i, state, reg) {
6015 if (reg->type == type && reg->id == dst_reg->id)
6016 reg->range = max(reg->range, new_range);
6020 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
6021 struct bpf_reg_state *dst_reg,
6022 enum bpf_reg_type type,
6023 bool range_right_open)
6028 if (dst_reg->off < 0 ||
6029 (dst_reg->off == 0 && range_right_open))
6030 /* This doesn't give us any range */
6033 if (dst_reg->umax_value > MAX_PACKET_OFF ||
6034 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
6035 /* Risk of overflow. For instance, ptr + (1<<63) may be less
6036 * than pkt_end, but that's because it's also less than pkt.
6040 new_range = dst_reg->off;
6041 if (range_right_open)
6044 /* Examples for register markings:
6046 * pkt_data in dst register:
6050 * if (r2 > pkt_end) goto <handle exception>
6055 * if (r2 < pkt_end) goto <access okay>
6056 * <handle exception>
6059 * r2 == dst_reg, pkt_end == src_reg
6060 * r2=pkt(id=n,off=8,r=0)
6061 * r3=pkt(id=n,off=0,r=0)
6063 * pkt_data in src register:
6067 * if (pkt_end >= r2) goto <access okay>
6068 * <handle exception>
6072 * if (pkt_end <= r2) goto <handle exception>
6076 * pkt_end == dst_reg, r2 == src_reg
6077 * r2=pkt(id=n,off=8,r=0)
6078 * r3=pkt(id=n,off=0,r=0)
6080 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
6081 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
6082 * and [r3, r3 + 8-1) respectively is safe to access depending on
6086 /* If our ids match, then we must have the same max_value. And we
6087 * don't care about the other reg's fixed offset, since if it's too big
6088 * the range won't allow anything.
6089 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
6091 for (i = 0; i <= vstate->curframe; i++)
6092 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
6096 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
6098 struct tnum subreg = tnum_subreg(reg->var_off);
6099 s32 sval = (s32)val;
6103 if (tnum_is_const(subreg))
6104 return !!tnum_equals_const(subreg, val);
6107 if (tnum_is_const(subreg))
6108 return !tnum_equals_const(subreg, val);
6111 if ((~subreg.mask & subreg.value) & val)
6113 if (!((subreg.mask | subreg.value) & val))
6117 if (reg->u32_min_value > val)
6119 else if (reg->u32_max_value <= val)
6123 if (reg->s32_min_value > sval)
6125 else if (reg->s32_max_value < sval)
6129 if (reg->u32_max_value < val)
6131 else if (reg->u32_min_value >= val)
6135 if (reg->s32_max_value < sval)
6137 else if (reg->s32_min_value >= sval)
6141 if (reg->u32_min_value >= val)
6143 else if (reg->u32_max_value < val)
6147 if (reg->s32_min_value >= sval)
6149 else if (reg->s32_max_value < sval)
6153 if (reg->u32_max_value <= val)
6155 else if (reg->u32_min_value > val)
6159 if (reg->s32_max_value <= sval)
6161 else if (reg->s32_min_value > sval)
6170 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
6172 s64 sval = (s64)val;
6176 if (tnum_is_const(reg->var_off))
6177 return !!tnum_equals_const(reg->var_off, val);
6180 if (tnum_is_const(reg->var_off))
6181 return !tnum_equals_const(reg->var_off, val);
6184 if ((~reg->var_off.mask & reg->var_off.value) & val)
6186 if (!((reg->var_off.mask | reg->var_off.value) & val))
6190 if (reg->umin_value > val)
6192 else if (reg->umax_value <= val)
6196 if (reg->smin_value > sval)
6198 else if (reg->smax_value < sval)
6202 if (reg->umax_value < val)
6204 else if (reg->umin_value >= val)
6208 if (reg->smax_value < sval)
6210 else if (reg->smin_value >= sval)
6214 if (reg->umin_value >= val)
6216 else if (reg->umax_value < val)
6220 if (reg->smin_value >= sval)
6222 else if (reg->smax_value < sval)
6226 if (reg->umax_value <= val)
6228 else if (reg->umin_value > val)
6232 if (reg->smax_value <= sval)
6234 else if (reg->smin_value > sval)
6242 /* compute branch direction of the expression "if (reg opcode val) goto target;"
6244 * 1 - branch will be taken and "goto target" will be executed
6245 * 0 - branch will not be taken and fall-through to next insn
6246 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
6249 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
6252 if (__is_pointer_value(false, reg))
6256 return is_branch32_taken(reg, val, opcode);
6257 return is_branch64_taken(reg, val, opcode);
6260 /* Adjusts the register min/max values in the case that the dst_reg is the
6261 * variable register that we are working on, and src_reg is a constant or we're
6262 * simply doing a BPF_K check.
6263 * In JEQ/JNE cases we also adjust the var_off values.
6265 static void reg_set_min_max(struct bpf_reg_state *true_reg,
6266 struct bpf_reg_state *false_reg,
6268 u8 opcode, bool is_jmp32)
6270 struct tnum false_32off = tnum_subreg(false_reg->var_off);
6271 struct tnum false_64off = false_reg->var_off;
6272 struct tnum true_32off = tnum_subreg(true_reg->var_off);
6273 struct tnum true_64off = true_reg->var_off;
6274 s64 sval = (s64)val;
6275 s32 sval32 = (s32)val32;
6277 /* If the dst_reg is a pointer, we can't learn anything about its
6278 * variable offset from the compare (unless src_reg were a pointer into
6279 * the same object, but we don't bother with that.
6280 * Since false_reg and true_reg have the same type by construction, we
6281 * only need to check one of them for pointerness.
6283 if (__is_pointer_value(false, false_reg))
6290 struct bpf_reg_state *reg =
6291 opcode == BPF_JEQ ? true_reg : false_reg;
6293 /* For BPF_JEQ, if this is false we know nothing Jon Snow, but
6294 * if it is true we know the value for sure. Likewise for
6298 __mark_reg32_known(reg, val32);
6300 __mark_reg_known(reg, val);
6305 false_32off = tnum_and(false_32off, tnum_const(~val32));
6306 if (is_power_of_2(val32))
6307 true_32off = tnum_or(true_32off,
6310 false_64off = tnum_and(false_64off, tnum_const(~val));
6311 if (is_power_of_2(val))
6312 true_64off = tnum_or(true_64off,
6320 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
6321 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
6323 false_reg->u32_max_value = min(false_reg->u32_max_value,
6325 true_reg->u32_min_value = max(true_reg->u32_min_value,
6328 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
6329 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
6331 false_reg->umax_value = min(false_reg->umax_value, false_umax);
6332 true_reg->umin_value = max(true_reg->umin_value, true_umin);
6340 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
6341 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
6343 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
6344 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
6346 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
6347 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
6349 false_reg->smax_value = min(false_reg->smax_value, false_smax);
6350 true_reg->smin_value = max(true_reg->smin_value, true_smin);
6358 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
6359 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
6361 false_reg->u32_min_value = max(false_reg->u32_min_value,
6363 true_reg->u32_max_value = min(true_reg->u32_max_value,
6366 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
6367 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
6369 false_reg->umin_value = max(false_reg->umin_value, false_umin);
6370 true_reg->umax_value = min(true_reg->umax_value, true_umax);
6378 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
6379 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
6381 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
6382 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
6384 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
6385 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
6387 false_reg->smin_value = max(false_reg->smin_value, false_smin);
6388 true_reg->smax_value = min(true_reg->smax_value, true_smax);
6397 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
6398 tnum_subreg(false_32off));
6399 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
6400 tnum_subreg(true_32off));
6401 __reg_combine_32_into_64(false_reg);
6402 __reg_combine_32_into_64(true_reg);
6404 false_reg->var_off = false_64off;
6405 true_reg->var_off = true_64off;
6406 __reg_combine_64_into_32(false_reg);
6407 __reg_combine_64_into_32(true_reg);
6411 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
6414 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
6415 struct bpf_reg_state *false_reg,
6417 u8 opcode, bool is_jmp32)
6419 /* How can we transform "a <op> b" into "b <op> a"? */
6420 static const u8 opcode_flip[16] = {
6421 /* these stay the same */
6422 [BPF_JEQ >> 4] = BPF_JEQ,
6423 [BPF_JNE >> 4] = BPF_JNE,
6424 [BPF_JSET >> 4] = BPF_JSET,
6425 /* these swap "lesser" and "greater" (L and G in the opcodes) */
6426 [BPF_JGE >> 4] = BPF_JLE,
6427 [BPF_JGT >> 4] = BPF_JLT,
6428 [BPF_JLE >> 4] = BPF_JGE,
6429 [BPF_JLT >> 4] = BPF_JGT,
6430 [BPF_JSGE >> 4] = BPF_JSLE,
6431 [BPF_JSGT >> 4] = BPF_JSLT,
6432 [BPF_JSLE >> 4] = BPF_JSGE,
6433 [BPF_JSLT >> 4] = BPF_JSGT
6435 opcode = opcode_flip[opcode >> 4];
6436 /* This uses zero as "not present in table"; luckily the zero opcode,
6437 * BPF_JA, can't get here.
6440 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
6443 /* Regs are known to be equal, so intersect their min/max/var_off */
6444 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
6445 struct bpf_reg_state *dst_reg)
6447 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
6448 dst_reg->umin_value);
6449 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
6450 dst_reg->umax_value);
6451 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
6452 dst_reg->smin_value);
6453 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
6454 dst_reg->smax_value);
6455 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
6457 /* We might have learned new bounds from the var_off. */
6458 __update_reg_bounds(src_reg);
6459 __update_reg_bounds(dst_reg);
6460 /* We might have learned something about the sign bit. */
6461 __reg_deduce_bounds(src_reg);
6462 __reg_deduce_bounds(dst_reg);
6463 /* We might have learned some bits from the bounds. */
6464 __reg_bound_offset(src_reg);
6465 __reg_bound_offset(dst_reg);
6466 /* Intersecting with the old var_off might have improved our bounds
6467 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
6468 * then new var_off is (0; 0x7f...fc) which improves our umax.
6470 __update_reg_bounds(src_reg);
6471 __update_reg_bounds(dst_reg);
6474 static void reg_combine_min_max(struct bpf_reg_state *true_src,
6475 struct bpf_reg_state *true_dst,
6476 struct bpf_reg_state *false_src,
6477 struct bpf_reg_state *false_dst,
6482 __reg_combine_min_max(true_src, true_dst);
6485 __reg_combine_min_max(false_src, false_dst);
6490 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
6491 struct bpf_reg_state *reg, u32 id,
6494 if (reg_type_may_be_null(reg->type) && reg->id == id) {
6495 /* Old offset (both fixed and variable parts) should
6496 * have been known-zero, because we don't allow pointer
6497 * arithmetic on pointers that might be NULL.
6499 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
6500 !tnum_equals_const(reg->var_off, 0) ||
6502 __mark_reg_known_zero(reg);
6506 reg->type = SCALAR_VALUE;
6507 } else if (reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
6508 if (reg->map_ptr->inner_map_meta) {
6509 reg->type = CONST_PTR_TO_MAP;
6510 reg->map_ptr = reg->map_ptr->inner_map_meta;
6511 } else if (reg->map_ptr->map_type ==
6512 BPF_MAP_TYPE_XSKMAP) {
6513 reg->type = PTR_TO_XDP_SOCK;
6515 reg->type = PTR_TO_MAP_VALUE;
6517 } else if (reg->type == PTR_TO_SOCKET_OR_NULL) {
6518 reg->type = PTR_TO_SOCKET;
6519 } else if (reg->type == PTR_TO_SOCK_COMMON_OR_NULL) {
6520 reg->type = PTR_TO_SOCK_COMMON;
6521 } else if (reg->type == PTR_TO_TCP_SOCK_OR_NULL) {
6522 reg->type = PTR_TO_TCP_SOCK;
6525 /* We don't need id and ref_obj_id from this point
6526 * onwards anymore, thus we should better reset it,
6527 * so that state pruning has chances to take effect.
6530 reg->ref_obj_id = 0;
6531 } else if (!reg_may_point_to_spin_lock(reg)) {
6532 /* For not-NULL ptr, reg->ref_obj_id will be reset
6533 * in release_reg_references().
6535 * reg->id is still used by spin_lock ptr. Other
6536 * than spin_lock ptr type, reg->id can be reset.
6543 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
6546 struct bpf_reg_state *reg;
6549 for (i = 0; i < MAX_BPF_REG; i++)
6550 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
6552 bpf_for_each_spilled_reg(i, state, reg) {
6555 mark_ptr_or_null_reg(state, reg, id, is_null);
6559 /* The logic is similar to find_good_pkt_pointers(), both could eventually
6560 * be folded together at some point.
6562 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
6565 struct bpf_func_state *state = vstate->frame[vstate->curframe];
6566 struct bpf_reg_state *regs = state->regs;
6567 u32 ref_obj_id = regs[regno].ref_obj_id;
6568 u32 id = regs[regno].id;
6571 if (ref_obj_id && ref_obj_id == id && is_null)
6572 /* regs[regno] is in the " == NULL" branch.
6573 * No one could have freed the reference state before
6574 * doing the NULL check.
6576 WARN_ON_ONCE(release_reference_state(state, id));
6578 for (i = 0; i <= vstate->curframe; i++)
6579 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
6582 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
6583 struct bpf_reg_state *dst_reg,
6584 struct bpf_reg_state *src_reg,
6585 struct bpf_verifier_state *this_branch,
6586 struct bpf_verifier_state *other_branch)
6588 if (BPF_SRC(insn->code) != BPF_X)
6591 /* Pointers are always 64-bit. */
6592 if (BPF_CLASS(insn->code) == BPF_JMP32)
6595 switch (BPF_OP(insn->code)) {
6597 if ((dst_reg->type == PTR_TO_PACKET &&
6598 src_reg->type == PTR_TO_PACKET_END) ||
6599 (dst_reg->type == PTR_TO_PACKET_META &&
6600 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
6601 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
6602 find_good_pkt_pointers(this_branch, dst_reg,
6603 dst_reg->type, false);
6604 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6605 src_reg->type == PTR_TO_PACKET) ||
6606 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6607 src_reg->type == PTR_TO_PACKET_META)) {
6608 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
6609 find_good_pkt_pointers(other_branch, src_reg,
6610 src_reg->type, true);
6616 if ((dst_reg->type == PTR_TO_PACKET &&
6617 src_reg->type == PTR_TO_PACKET_END) ||
6618 (dst_reg->type == PTR_TO_PACKET_META &&
6619 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
6620 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
6621 find_good_pkt_pointers(other_branch, dst_reg,
6622 dst_reg->type, true);
6623 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6624 src_reg->type == PTR_TO_PACKET) ||
6625 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6626 src_reg->type == PTR_TO_PACKET_META)) {
6627 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
6628 find_good_pkt_pointers(this_branch, src_reg,
6629 src_reg->type, false);
6635 if ((dst_reg->type == PTR_TO_PACKET &&
6636 src_reg->type == PTR_TO_PACKET_END) ||
6637 (dst_reg->type == PTR_TO_PACKET_META &&
6638 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
6639 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
6640 find_good_pkt_pointers(this_branch, dst_reg,
6641 dst_reg->type, true);
6642 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6643 src_reg->type == PTR_TO_PACKET) ||
6644 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6645 src_reg->type == PTR_TO_PACKET_META)) {
6646 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
6647 find_good_pkt_pointers(other_branch, src_reg,
6648 src_reg->type, false);
6654 if ((dst_reg->type == PTR_TO_PACKET &&
6655 src_reg->type == PTR_TO_PACKET_END) ||
6656 (dst_reg->type == PTR_TO_PACKET_META &&
6657 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
6658 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
6659 find_good_pkt_pointers(other_branch, dst_reg,
6660 dst_reg->type, false);
6661 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6662 src_reg->type == PTR_TO_PACKET) ||
6663 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6664 src_reg->type == PTR_TO_PACKET_META)) {
6665 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
6666 find_good_pkt_pointers(this_branch, src_reg,
6667 src_reg->type, true);
6679 static int check_cond_jmp_op(struct bpf_verifier_env *env,
6680 struct bpf_insn *insn, int *insn_idx)
6682 struct bpf_verifier_state *this_branch = env->cur_state;
6683 struct bpf_verifier_state *other_branch;
6684 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
6685 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
6686 u8 opcode = BPF_OP(insn->code);
6691 /* Only conditional jumps are expected to reach here. */
6692 if (opcode == BPF_JA || opcode > BPF_JSLE) {
6693 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
6697 if (BPF_SRC(insn->code) == BPF_X) {
6698 if (insn->imm != 0) {
6699 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
6703 /* check src1 operand */
6704 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6708 if (is_pointer_value(env, insn->src_reg)) {
6709 verbose(env, "R%d pointer comparison prohibited\n",
6713 src_reg = ®s[insn->src_reg];
6715 if (insn->src_reg != BPF_REG_0) {
6716 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
6721 /* check src2 operand */
6722 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6726 dst_reg = ®s[insn->dst_reg];
6727 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
6729 if (BPF_SRC(insn->code) == BPF_K) {
6730 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
6731 } else if (src_reg->type == SCALAR_VALUE &&
6732 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
6733 pred = is_branch_taken(dst_reg,
6734 tnum_subreg(src_reg->var_off).value,
6737 } else if (src_reg->type == SCALAR_VALUE &&
6738 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
6739 pred = is_branch_taken(dst_reg,
6740 src_reg->var_off.value,
6746 err = mark_chain_precision(env, insn->dst_reg);
6747 if (BPF_SRC(insn->code) == BPF_X && !err)
6748 err = mark_chain_precision(env, insn->src_reg);
6753 /* only follow the goto, ignore fall-through */
6754 *insn_idx += insn->off;
6756 } else if (pred == 0) {
6757 /* only follow fall-through branch, since
6758 * that's where the program will go
6763 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
6767 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
6769 /* detect if we are comparing against a constant value so we can adjust
6770 * our min/max values for our dst register.
6771 * this is only legit if both are scalars (or pointers to the same
6772 * object, I suppose, but we don't support that right now), because
6773 * otherwise the different base pointers mean the offsets aren't
6776 if (BPF_SRC(insn->code) == BPF_X) {
6777 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
6779 if (dst_reg->type == SCALAR_VALUE &&
6780 src_reg->type == SCALAR_VALUE) {
6781 if (tnum_is_const(src_reg->var_off) ||
6783 tnum_is_const(tnum_subreg(src_reg->var_off))))
6784 reg_set_min_max(&other_branch_regs[insn->dst_reg],
6786 src_reg->var_off.value,
6787 tnum_subreg(src_reg->var_off).value,
6789 else if (tnum_is_const(dst_reg->var_off) ||
6791 tnum_is_const(tnum_subreg(dst_reg->var_off))))
6792 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
6794 dst_reg->var_off.value,
6795 tnum_subreg(dst_reg->var_off).value,
6797 else if (!is_jmp32 &&
6798 (opcode == BPF_JEQ || opcode == BPF_JNE))
6799 /* Comparing for equality, we can combine knowledge */
6800 reg_combine_min_max(&other_branch_regs[insn->src_reg],
6801 &other_branch_regs[insn->dst_reg],
6802 src_reg, dst_reg, opcode);
6804 } else if (dst_reg->type == SCALAR_VALUE) {
6805 reg_set_min_max(&other_branch_regs[insn->dst_reg],
6806 dst_reg, insn->imm, (u32)insn->imm,
6810 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
6811 * NOTE: these optimizations below are related with pointer comparison
6812 * which will never be JMP32.
6814 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
6815 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
6816 reg_type_may_be_null(dst_reg->type)) {
6817 /* Mark all identical registers in each branch as either
6818 * safe or unknown depending R == 0 or R != 0 conditional.
6820 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
6822 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
6824 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
6825 this_branch, other_branch) &&
6826 is_pointer_value(env, insn->dst_reg)) {
6827 verbose(env, "R%d pointer comparison prohibited\n",
6831 if (env->log.level & BPF_LOG_LEVEL)
6832 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
6836 /* verify BPF_LD_IMM64 instruction */
6837 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
6839 struct bpf_insn_aux_data *aux = cur_aux(env);
6840 struct bpf_reg_state *regs = cur_regs(env);
6841 struct bpf_map *map;
6844 if (BPF_SIZE(insn->code) != BPF_DW) {
6845 verbose(env, "invalid BPF_LD_IMM insn\n");
6848 if (insn->off != 0) {
6849 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
6853 err = check_reg_arg(env, insn->dst_reg, DST_OP);
6857 if (insn->src_reg == 0) {
6858 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
6860 regs[insn->dst_reg].type = SCALAR_VALUE;
6861 __mark_reg_known(®s[insn->dst_reg], imm);
6865 map = env->used_maps[aux->map_index];
6866 mark_reg_known_zero(env, regs, insn->dst_reg);
6867 regs[insn->dst_reg].map_ptr = map;
6869 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE) {
6870 regs[insn->dst_reg].type = PTR_TO_MAP_VALUE;
6871 regs[insn->dst_reg].off = aux->map_off;
6872 if (map_value_has_spin_lock(map))
6873 regs[insn->dst_reg].id = ++env->id_gen;
6874 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
6875 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
6877 verbose(env, "bpf verifier is misconfigured\n");
6884 static bool may_access_skb(enum bpf_prog_type type)
6887 case BPF_PROG_TYPE_SOCKET_FILTER:
6888 case BPF_PROG_TYPE_SCHED_CLS:
6889 case BPF_PROG_TYPE_SCHED_ACT:
6896 /* verify safety of LD_ABS|LD_IND instructions:
6897 * - they can only appear in the programs where ctx == skb
6898 * - since they are wrappers of function calls, they scratch R1-R5 registers,
6899 * preserve R6-R9, and store return value into R0
6902 * ctx == skb == R6 == CTX
6905 * SRC == any register
6906 * IMM == 32-bit immediate
6909 * R0 - 8/16/32-bit skb data converted to cpu endianness
6911 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
6913 struct bpf_reg_state *regs = cur_regs(env);
6914 static const int ctx_reg = BPF_REG_6;
6915 u8 mode = BPF_MODE(insn->code);
6918 if (!may_access_skb(env->prog->type)) {
6919 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
6923 if (!env->ops->gen_ld_abs) {
6924 verbose(env, "bpf verifier is misconfigured\n");
6928 if (env->subprog_cnt > 1) {
6929 /* when program has LD_ABS insn JITs and interpreter assume
6930 * that r1 == ctx == skb which is not the case for callees
6931 * that can have arbitrary arguments. It's problematic
6932 * for main prog as well since JITs would need to analyze
6933 * all functions in order to make proper register save/restore
6934 * decisions in the main prog. Hence disallow LD_ABS with calls
6936 verbose(env, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
6940 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
6941 BPF_SIZE(insn->code) == BPF_DW ||
6942 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
6943 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
6947 /* check whether implicit source operand (register R6) is readable */
6948 err = check_reg_arg(env, ctx_reg, SRC_OP);
6952 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
6953 * gen_ld_abs() may terminate the program at runtime, leading to
6956 err = check_reference_leak(env);
6958 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
6962 if (env->cur_state->active_spin_lock) {
6963 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
6967 if (regs[ctx_reg].type != PTR_TO_CTX) {
6969 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
6973 if (mode == BPF_IND) {
6974 /* check explicit source operand */
6975 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6980 err = check_ctx_reg(env, ®s[ctx_reg], ctx_reg);
6984 /* reset caller saved regs to unreadable */
6985 for (i = 0; i < CALLER_SAVED_REGS; i++) {
6986 mark_reg_not_init(env, regs, caller_saved[i]);
6987 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
6990 /* mark destination R0 register as readable, since it contains
6991 * the value fetched from the packet.
6992 * Already marked as written above.
6994 mark_reg_unknown(env, regs, BPF_REG_0);
6995 /* ld_abs load up to 32-bit skb data. */
6996 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
7000 static int check_return_code(struct bpf_verifier_env *env)
7002 struct tnum enforce_attach_type_range = tnum_unknown;
7003 const struct bpf_prog *prog = env->prog;
7004 struct bpf_reg_state *reg;
7005 struct tnum range = tnum_range(0, 1);
7008 /* LSM and struct_ops func-ptr's return type could be "void" */
7009 if ((env->prog->type == BPF_PROG_TYPE_STRUCT_OPS ||
7010 env->prog->type == BPF_PROG_TYPE_LSM) &&
7011 !prog->aux->attach_func_proto->type)
7014 /* eBPF calling convetion is such that R0 is used
7015 * to return the value from eBPF program.
7016 * Make sure that it's readable at this time
7017 * of bpf_exit, which means that program wrote
7018 * something into it earlier
7020 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
7024 if (is_pointer_value(env, BPF_REG_0)) {
7025 verbose(env, "R0 leaks addr as return value\n");
7029 switch (env->prog->type) {
7030 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
7031 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
7032 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG)
7033 range = tnum_range(1, 1);
7035 case BPF_PROG_TYPE_CGROUP_SKB:
7036 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
7037 range = tnum_range(0, 3);
7038 enforce_attach_type_range = tnum_range(2, 3);
7041 case BPF_PROG_TYPE_CGROUP_SOCK:
7042 case BPF_PROG_TYPE_SOCK_OPS:
7043 case BPF_PROG_TYPE_CGROUP_DEVICE:
7044 case BPF_PROG_TYPE_CGROUP_SYSCTL:
7045 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
7047 case BPF_PROG_TYPE_RAW_TRACEPOINT:
7048 if (!env->prog->aux->attach_btf_id)
7050 range = tnum_const(0);
7056 reg = cur_regs(env) + BPF_REG_0;
7057 if (reg->type != SCALAR_VALUE) {
7058 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
7059 reg_type_str[reg->type]);
7063 if (!tnum_in(range, reg->var_off)) {
7066 verbose(env, "At program exit the register R0 ");
7067 if (!tnum_is_unknown(reg->var_off)) {
7068 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
7069 verbose(env, "has value %s", tn_buf);
7071 verbose(env, "has unknown scalar value");
7073 tnum_strn(tn_buf, sizeof(tn_buf), range);
7074 verbose(env, " should have been in %s\n", tn_buf);
7078 if (!tnum_is_unknown(enforce_attach_type_range) &&
7079 tnum_in(enforce_attach_type_range, reg->var_off))
7080 env->prog->enforce_expected_attach_type = 1;
7084 /* non-recursive DFS pseudo code
7085 * 1 procedure DFS-iterative(G,v):
7086 * 2 label v as discovered
7087 * 3 let S be a stack
7089 * 5 while S is not empty
7091 * 7 if t is what we're looking for:
7093 * 9 for all edges e in G.adjacentEdges(t) do
7094 * 10 if edge e is already labelled
7095 * 11 continue with the next edge
7096 * 12 w <- G.adjacentVertex(t,e)
7097 * 13 if vertex w is not discovered and not explored
7098 * 14 label e as tree-edge
7099 * 15 label w as discovered
7102 * 18 else if vertex w is discovered
7103 * 19 label e as back-edge
7105 * 21 // vertex w is explored
7106 * 22 label e as forward- or cross-edge
7107 * 23 label t as explored
7112 * 0x11 - discovered and fall-through edge labelled
7113 * 0x12 - discovered and fall-through and branch edges labelled
7124 static u32 state_htab_size(struct bpf_verifier_env *env)
7126 return env->prog->len;
7129 static struct bpf_verifier_state_list **explored_state(
7130 struct bpf_verifier_env *env,
7133 struct bpf_verifier_state *cur = env->cur_state;
7134 struct bpf_func_state *state = cur->frame[cur->curframe];
7136 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
7139 static void init_explored_state(struct bpf_verifier_env *env, int idx)
7141 env->insn_aux_data[idx].prune_point = true;
7144 /* t, w, e - match pseudo-code above:
7145 * t - index of current instruction
7146 * w - next instruction
7149 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
7152 int *insn_stack = env->cfg.insn_stack;
7153 int *insn_state = env->cfg.insn_state;
7155 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
7158 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
7161 if (w < 0 || w >= env->prog->len) {
7162 verbose_linfo(env, t, "%d: ", t);
7163 verbose(env, "jump out of range from insn %d to %d\n", t, w);
7168 /* mark branch target for state pruning */
7169 init_explored_state(env, w);
7171 if (insn_state[w] == 0) {
7173 insn_state[t] = DISCOVERED | e;
7174 insn_state[w] = DISCOVERED;
7175 if (env->cfg.cur_stack >= env->prog->len)
7177 insn_stack[env->cfg.cur_stack++] = w;
7179 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
7180 if (loop_ok && env->allow_ptr_leaks)
7182 verbose_linfo(env, t, "%d: ", t);
7183 verbose_linfo(env, w, "%d: ", w);
7184 verbose(env, "back-edge from insn %d to %d\n", t, w);
7186 } else if (insn_state[w] == EXPLORED) {
7187 /* forward- or cross-edge */
7188 insn_state[t] = DISCOVERED | e;
7190 verbose(env, "insn state internal bug\n");
7196 /* non-recursive depth-first-search to detect loops in BPF program
7197 * loop == back-edge in directed graph
7199 static int check_cfg(struct bpf_verifier_env *env)
7201 struct bpf_insn *insns = env->prog->insnsi;
7202 int insn_cnt = env->prog->len;
7203 int *insn_stack, *insn_state;
7207 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
7211 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
7217 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
7218 insn_stack[0] = 0; /* 0 is the first instruction */
7219 env->cfg.cur_stack = 1;
7222 if (env->cfg.cur_stack == 0)
7224 t = insn_stack[env->cfg.cur_stack - 1];
7226 if (BPF_CLASS(insns[t].code) == BPF_JMP ||
7227 BPF_CLASS(insns[t].code) == BPF_JMP32) {
7228 u8 opcode = BPF_OP(insns[t].code);
7230 if (opcode == BPF_EXIT) {
7232 } else if (opcode == BPF_CALL) {
7233 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
7238 if (t + 1 < insn_cnt)
7239 init_explored_state(env, t + 1);
7240 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
7241 init_explored_state(env, t);
7242 ret = push_insn(t, t + insns[t].imm + 1, BRANCH,
7249 } else if (opcode == BPF_JA) {
7250 if (BPF_SRC(insns[t].code) != BPF_K) {
7254 /* unconditional jump with single edge */
7255 ret = push_insn(t, t + insns[t].off + 1,
7256 FALLTHROUGH, env, true);
7261 /* unconditional jmp is not a good pruning point,
7262 * but it's marked, since backtracking needs
7263 * to record jmp history in is_state_visited().
7265 init_explored_state(env, t + insns[t].off + 1);
7266 /* tell verifier to check for equivalent states
7267 * after every call and jump
7269 if (t + 1 < insn_cnt)
7270 init_explored_state(env, t + 1);
7272 /* conditional jump with two edges */
7273 init_explored_state(env, t);
7274 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
7280 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
7287 /* all other non-branch instructions with single
7290 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
7298 insn_state[t] = EXPLORED;
7299 if (env->cfg.cur_stack-- <= 0) {
7300 verbose(env, "pop stack internal bug\n");
7307 for (i = 0; i < insn_cnt; i++) {
7308 if (insn_state[i] != EXPLORED) {
7309 verbose(env, "unreachable insn %d\n", i);
7314 ret = 0; /* cfg looks good */
7319 env->cfg.insn_state = env->cfg.insn_stack = NULL;
7323 /* The minimum supported BTF func info size */
7324 #define MIN_BPF_FUNCINFO_SIZE 8
7325 #define MAX_FUNCINFO_REC_SIZE 252
7327 static int check_btf_func(struct bpf_verifier_env *env,
7328 const union bpf_attr *attr,
7329 union bpf_attr __user *uattr)
7331 u32 i, nfuncs, urec_size, min_size;
7332 u32 krec_size = sizeof(struct bpf_func_info);
7333 struct bpf_func_info *krecord;
7334 struct bpf_func_info_aux *info_aux = NULL;
7335 const struct btf_type *type;
7336 struct bpf_prog *prog;
7337 const struct btf *btf;
7338 void __user *urecord;
7339 u32 prev_offset = 0;
7342 nfuncs = attr->func_info_cnt;
7346 if (nfuncs != env->subprog_cnt) {
7347 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
7351 urec_size = attr->func_info_rec_size;
7352 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
7353 urec_size > MAX_FUNCINFO_REC_SIZE ||
7354 urec_size % sizeof(u32)) {
7355 verbose(env, "invalid func info rec size %u\n", urec_size);
7360 btf = prog->aux->btf;
7362 urecord = u64_to_user_ptr(attr->func_info);
7363 min_size = min_t(u32, krec_size, urec_size);
7365 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
7368 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
7372 for (i = 0; i < nfuncs; i++) {
7373 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
7375 if (ret == -E2BIG) {
7376 verbose(env, "nonzero tailing record in func info");
7377 /* set the size kernel expects so loader can zero
7378 * out the rest of the record.
7380 if (put_user(min_size, &uattr->func_info_rec_size))
7386 if (copy_from_user(&krecord[i], urecord, min_size)) {
7391 /* check insn_off */
7393 if (krecord[i].insn_off) {
7395 "nonzero insn_off %u for the first func info record",
7396 krecord[i].insn_off);
7400 } else if (krecord[i].insn_off <= prev_offset) {
7402 "same or smaller insn offset (%u) than previous func info record (%u)",
7403 krecord[i].insn_off, prev_offset);
7408 if (env->subprog_info[i].start != krecord[i].insn_off) {
7409 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
7415 type = btf_type_by_id(btf, krecord[i].type_id);
7416 if (!type || !btf_type_is_func(type)) {
7417 verbose(env, "invalid type id %d in func info",
7418 krecord[i].type_id);
7422 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
7423 prev_offset = krecord[i].insn_off;
7424 urecord += urec_size;
7427 prog->aux->func_info = krecord;
7428 prog->aux->func_info_cnt = nfuncs;
7429 prog->aux->func_info_aux = info_aux;
7438 static void adjust_btf_func(struct bpf_verifier_env *env)
7440 struct bpf_prog_aux *aux = env->prog->aux;
7443 if (!aux->func_info)
7446 for (i = 0; i < env->subprog_cnt; i++)
7447 aux->func_info[i].insn_off = env->subprog_info[i].start;
7450 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
7451 sizeof(((struct bpf_line_info *)(0))->line_col))
7452 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
7454 static int check_btf_line(struct bpf_verifier_env *env,
7455 const union bpf_attr *attr,
7456 union bpf_attr __user *uattr)
7458 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
7459 struct bpf_subprog_info *sub;
7460 struct bpf_line_info *linfo;
7461 struct bpf_prog *prog;
7462 const struct btf *btf;
7463 void __user *ulinfo;
7466 nr_linfo = attr->line_info_cnt;
7470 rec_size = attr->line_info_rec_size;
7471 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
7472 rec_size > MAX_LINEINFO_REC_SIZE ||
7473 rec_size & (sizeof(u32) - 1))
7476 /* Need to zero it in case the userspace may
7477 * pass in a smaller bpf_line_info object.
7479 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
7480 GFP_KERNEL | __GFP_NOWARN);
7485 btf = prog->aux->btf;
7488 sub = env->subprog_info;
7489 ulinfo = u64_to_user_ptr(attr->line_info);
7490 expected_size = sizeof(struct bpf_line_info);
7491 ncopy = min_t(u32, expected_size, rec_size);
7492 for (i = 0; i < nr_linfo; i++) {
7493 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
7495 if (err == -E2BIG) {
7496 verbose(env, "nonzero tailing record in line_info");
7497 if (put_user(expected_size,
7498 &uattr->line_info_rec_size))
7504 if (copy_from_user(&linfo[i], ulinfo, ncopy)) {
7510 * Check insn_off to ensure
7511 * 1) strictly increasing AND
7512 * 2) bounded by prog->len
7514 * The linfo[0].insn_off == 0 check logically falls into
7515 * the later "missing bpf_line_info for func..." case
7516 * because the first linfo[0].insn_off must be the
7517 * first sub also and the first sub must have
7518 * subprog_info[0].start == 0.
7520 if ((i && linfo[i].insn_off <= prev_offset) ||
7521 linfo[i].insn_off >= prog->len) {
7522 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
7523 i, linfo[i].insn_off, prev_offset,
7529 if (!prog->insnsi[linfo[i].insn_off].code) {
7531 "Invalid insn code at line_info[%u].insn_off\n",
7537 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
7538 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
7539 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
7544 if (s != env->subprog_cnt) {
7545 if (linfo[i].insn_off == sub[s].start) {
7546 sub[s].linfo_idx = i;
7548 } else if (sub[s].start < linfo[i].insn_off) {
7549 verbose(env, "missing bpf_line_info for func#%u\n", s);
7555 prev_offset = linfo[i].insn_off;
7559 if (s != env->subprog_cnt) {
7560 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
7561 env->subprog_cnt - s, s);
7566 prog->aux->linfo = linfo;
7567 prog->aux->nr_linfo = nr_linfo;
7576 static int check_btf_info(struct bpf_verifier_env *env,
7577 const union bpf_attr *attr,
7578 union bpf_attr __user *uattr)
7583 if (!attr->func_info_cnt && !attr->line_info_cnt)
7586 btf = btf_get_by_fd(attr->prog_btf_fd);
7588 return PTR_ERR(btf);
7589 env->prog->aux->btf = btf;
7591 err = check_btf_func(env, attr, uattr);
7595 err = check_btf_line(env, attr, uattr);
7602 /* check %cur's range satisfies %old's */
7603 static bool range_within(struct bpf_reg_state *old,
7604 struct bpf_reg_state *cur)
7606 return old->umin_value <= cur->umin_value &&
7607 old->umax_value >= cur->umax_value &&
7608 old->smin_value <= cur->smin_value &&
7609 old->smax_value >= cur->smax_value;
7612 /* Maximum number of register states that can exist at once */
7613 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
7619 /* If in the old state two registers had the same id, then they need to have
7620 * the same id in the new state as well. But that id could be different from
7621 * the old state, so we need to track the mapping from old to new ids.
7622 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
7623 * regs with old id 5 must also have new id 9 for the new state to be safe. But
7624 * regs with a different old id could still have new id 9, we don't care about
7626 * So we look through our idmap to see if this old id has been seen before. If
7627 * so, we require the new id to match; otherwise, we add the id pair to the map.
7629 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
7633 for (i = 0; i < ID_MAP_SIZE; i++) {
7634 if (!idmap[i].old) {
7635 /* Reached an empty slot; haven't seen this id before */
7636 idmap[i].old = old_id;
7637 idmap[i].cur = cur_id;
7640 if (idmap[i].old == old_id)
7641 return idmap[i].cur == cur_id;
7643 /* We ran out of idmap slots, which should be impossible */
7648 static void clean_func_state(struct bpf_verifier_env *env,
7649 struct bpf_func_state *st)
7651 enum bpf_reg_liveness live;
7654 for (i = 0; i < BPF_REG_FP; i++) {
7655 live = st->regs[i].live;
7656 /* liveness must not touch this register anymore */
7657 st->regs[i].live |= REG_LIVE_DONE;
7658 if (!(live & REG_LIVE_READ))
7659 /* since the register is unused, clear its state
7660 * to make further comparison simpler
7662 __mark_reg_not_init(env, &st->regs[i]);
7665 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
7666 live = st->stack[i].spilled_ptr.live;
7667 /* liveness must not touch this stack slot anymore */
7668 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
7669 if (!(live & REG_LIVE_READ)) {
7670 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
7671 for (j = 0; j < BPF_REG_SIZE; j++)
7672 st->stack[i].slot_type[j] = STACK_INVALID;
7677 static void clean_verifier_state(struct bpf_verifier_env *env,
7678 struct bpf_verifier_state *st)
7682 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
7683 /* all regs in this state in all frames were already marked */
7686 for (i = 0; i <= st->curframe; i++)
7687 clean_func_state(env, st->frame[i]);
7690 /* the parentage chains form a tree.
7691 * the verifier states are added to state lists at given insn and
7692 * pushed into state stack for future exploration.
7693 * when the verifier reaches bpf_exit insn some of the verifer states
7694 * stored in the state lists have their final liveness state already,
7695 * but a lot of states will get revised from liveness point of view when
7696 * the verifier explores other branches.
7699 * 2: if r1 == 100 goto pc+1
7702 * when the verifier reaches exit insn the register r0 in the state list of
7703 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
7704 * of insn 2 and goes exploring further. At the insn 4 it will walk the
7705 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
7707 * Since the verifier pushes the branch states as it sees them while exploring
7708 * the program the condition of walking the branch instruction for the second
7709 * time means that all states below this branch were already explored and
7710 * their final liveness markes are already propagated.
7711 * Hence when the verifier completes the search of state list in is_state_visited()
7712 * we can call this clean_live_states() function to mark all liveness states
7713 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
7715 * This function also clears the registers and stack for states that !READ
7716 * to simplify state merging.
7718 * Important note here that walking the same branch instruction in the callee
7719 * doesn't meant that the states are DONE. The verifier has to compare
7722 static void clean_live_states(struct bpf_verifier_env *env, int insn,
7723 struct bpf_verifier_state *cur)
7725 struct bpf_verifier_state_list *sl;
7728 sl = *explored_state(env, insn);
7730 if (sl->state.branches)
7732 if (sl->state.insn_idx != insn ||
7733 sl->state.curframe != cur->curframe)
7735 for (i = 0; i <= cur->curframe; i++)
7736 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
7738 clean_verifier_state(env, &sl->state);
7744 /* Returns true if (rold safe implies rcur safe) */
7745 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
7746 struct idpair *idmap)
7750 if (!(rold->live & REG_LIVE_READ))
7751 /* explored state didn't use this */
7754 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
7756 if (rold->type == PTR_TO_STACK)
7757 /* two stack pointers are equal only if they're pointing to
7758 * the same stack frame, since fp-8 in foo != fp-8 in bar
7760 return equal && rold->frameno == rcur->frameno;
7765 if (rold->type == NOT_INIT)
7766 /* explored state can't have used this */
7768 if (rcur->type == NOT_INIT)
7770 switch (rold->type) {
7772 if (rcur->type == SCALAR_VALUE) {
7773 if (!rold->precise && !rcur->precise)
7775 /* new val must satisfy old val knowledge */
7776 return range_within(rold, rcur) &&
7777 tnum_in(rold->var_off, rcur->var_off);
7779 /* We're trying to use a pointer in place of a scalar.
7780 * Even if the scalar was unbounded, this could lead to
7781 * pointer leaks because scalars are allowed to leak
7782 * while pointers are not. We could make this safe in
7783 * special cases if root is calling us, but it's
7784 * probably not worth the hassle.
7788 case PTR_TO_MAP_VALUE:
7789 /* If the new min/max/var_off satisfy the old ones and
7790 * everything else matches, we are OK.
7791 * 'id' is not compared, since it's only used for maps with
7792 * bpf_spin_lock inside map element and in such cases if
7793 * the rest of the prog is valid for one map element then
7794 * it's valid for all map elements regardless of the key
7795 * used in bpf_map_lookup()
7797 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
7798 range_within(rold, rcur) &&
7799 tnum_in(rold->var_off, rcur->var_off);
7800 case PTR_TO_MAP_VALUE_OR_NULL:
7801 /* a PTR_TO_MAP_VALUE could be safe to use as a
7802 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
7803 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
7804 * checked, doing so could have affected others with the same
7805 * id, and we can't check for that because we lost the id when
7806 * we converted to a PTR_TO_MAP_VALUE.
7808 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
7810 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
7812 /* Check our ids match any regs they're supposed to */
7813 return check_ids(rold->id, rcur->id, idmap);
7814 case PTR_TO_PACKET_META:
7816 if (rcur->type != rold->type)
7818 /* We must have at least as much range as the old ptr
7819 * did, so that any accesses which were safe before are
7820 * still safe. This is true even if old range < old off,
7821 * since someone could have accessed through (ptr - k), or
7822 * even done ptr -= k in a register, to get a safe access.
7824 if (rold->range > rcur->range)
7826 /* If the offsets don't match, we can't trust our alignment;
7827 * nor can we be sure that we won't fall out of range.
7829 if (rold->off != rcur->off)
7831 /* id relations must be preserved */
7832 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
7834 /* new val must satisfy old val knowledge */
7835 return range_within(rold, rcur) &&
7836 tnum_in(rold->var_off, rcur->var_off);
7838 case CONST_PTR_TO_MAP:
7839 case PTR_TO_PACKET_END:
7840 case PTR_TO_FLOW_KEYS:
7842 case PTR_TO_SOCKET_OR_NULL:
7843 case PTR_TO_SOCK_COMMON:
7844 case PTR_TO_SOCK_COMMON_OR_NULL:
7845 case PTR_TO_TCP_SOCK:
7846 case PTR_TO_TCP_SOCK_OR_NULL:
7847 case PTR_TO_XDP_SOCK:
7848 /* Only valid matches are exact, which memcmp() above
7849 * would have accepted
7852 /* Don't know what's going on, just say it's not safe */
7856 /* Shouldn't get here; if we do, say it's not safe */
7861 static bool stacksafe(struct bpf_func_state *old,
7862 struct bpf_func_state *cur,
7863 struct idpair *idmap)
7867 /* walk slots of the explored stack and ignore any additional
7868 * slots in the current stack, since explored(safe) state
7871 for (i = 0; i < old->allocated_stack; i++) {
7872 spi = i / BPF_REG_SIZE;
7874 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
7875 i += BPF_REG_SIZE - 1;
7876 /* explored state didn't use this */
7880 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
7883 /* explored stack has more populated slots than current stack
7884 * and these slots were used
7886 if (i >= cur->allocated_stack)
7889 /* if old state was safe with misc data in the stack
7890 * it will be safe with zero-initialized stack.
7891 * The opposite is not true
7893 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
7894 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
7896 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
7897 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
7898 /* Ex: old explored (safe) state has STACK_SPILL in
7899 * this stack slot, but current has has STACK_MISC ->
7900 * this verifier states are not equivalent,
7901 * return false to continue verification of this path
7904 if (i % BPF_REG_SIZE)
7906 if (old->stack[spi].slot_type[0] != STACK_SPILL)
7908 if (!regsafe(&old->stack[spi].spilled_ptr,
7909 &cur->stack[spi].spilled_ptr,
7911 /* when explored and current stack slot are both storing
7912 * spilled registers, check that stored pointers types
7913 * are the same as well.
7914 * Ex: explored safe path could have stored
7915 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
7916 * but current path has stored:
7917 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
7918 * such verifier states are not equivalent.
7919 * return false to continue verification of this path
7926 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
7928 if (old->acquired_refs != cur->acquired_refs)
7930 return !memcmp(old->refs, cur->refs,
7931 sizeof(*old->refs) * old->acquired_refs);
7934 /* compare two verifier states
7936 * all states stored in state_list are known to be valid, since
7937 * verifier reached 'bpf_exit' instruction through them
7939 * this function is called when verifier exploring different branches of
7940 * execution popped from the state stack. If it sees an old state that has
7941 * more strict register state and more strict stack state then this execution
7942 * branch doesn't need to be explored further, since verifier already
7943 * concluded that more strict state leads to valid finish.
7945 * Therefore two states are equivalent if register state is more conservative
7946 * and explored stack state is more conservative than the current one.
7949 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
7950 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
7952 * In other words if current stack state (one being explored) has more
7953 * valid slots than old one that already passed validation, it means
7954 * the verifier can stop exploring and conclude that current state is valid too
7956 * Similarly with registers. If explored state has register type as invalid
7957 * whereas register type in current state is meaningful, it means that
7958 * the current state will reach 'bpf_exit' instruction safely
7960 static bool func_states_equal(struct bpf_func_state *old,
7961 struct bpf_func_state *cur)
7963 struct idpair *idmap;
7967 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
7968 /* If we failed to allocate the idmap, just say it's not safe */
7972 for (i = 0; i < MAX_BPF_REG; i++) {
7973 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
7977 if (!stacksafe(old, cur, idmap))
7980 if (!refsafe(old, cur))
7988 static bool states_equal(struct bpf_verifier_env *env,
7989 struct bpf_verifier_state *old,
7990 struct bpf_verifier_state *cur)
7994 if (old->curframe != cur->curframe)
7997 /* Verification state from speculative execution simulation
7998 * must never prune a non-speculative execution one.
8000 if (old->speculative && !cur->speculative)
8003 if (old->active_spin_lock != cur->active_spin_lock)
8006 /* for states to be equal callsites have to be the same
8007 * and all frame states need to be equivalent
8009 for (i = 0; i <= old->curframe; i++) {
8010 if (old->frame[i]->callsite != cur->frame[i]->callsite)
8012 if (!func_states_equal(old->frame[i], cur->frame[i]))
8018 /* Return 0 if no propagation happened. Return negative error code if error
8019 * happened. Otherwise, return the propagated bit.
8021 static int propagate_liveness_reg(struct bpf_verifier_env *env,
8022 struct bpf_reg_state *reg,
8023 struct bpf_reg_state *parent_reg)
8025 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
8026 u8 flag = reg->live & REG_LIVE_READ;
8029 /* When comes here, read flags of PARENT_REG or REG could be any of
8030 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
8031 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
8033 if (parent_flag == REG_LIVE_READ64 ||
8034 /* Or if there is no read flag from REG. */
8036 /* Or if the read flag from REG is the same as PARENT_REG. */
8037 parent_flag == flag)
8040 err = mark_reg_read(env, reg, parent_reg, flag);
8047 /* A write screens off any subsequent reads; but write marks come from the
8048 * straight-line code between a state and its parent. When we arrive at an
8049 * equivalent state (jump target or such) we didn't arrive by the straight-line
8050 * code, so read marks in the state must propagate to the parent regardless
8051 * of the state's write marks. That's what 'parent == state->parent' comparison
8052 * in mark_reg_read() is for.
8054 static int propagate_liveness(struct bpf_verifier_env *env,
8055 const struct bpf_verifier_state *vstate,
8056 struct bpf_verifier_state *vparent)
8058 struct bpf_reg_state *state_reg, *parent_reg;
8059 struct bpf_func_state *state, *parent;
8060 int i, frame, err = 0;
8062 if (vparent->curframe != vstate->curframe) {
8063 WARN(1, "propagate_live: parent frame %d current frame %d\n",
8064 vparent->curframe, vstate->curframe);
8067 /* Propagate read liveness of registers... */
8068 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
8069 for (frame = 0; frame <= vstate->curframe; frame++) {
8070 parent = vparent->frame[frame];
8071 state = vstate->frame[frame];
8072 parent_reg = parent->regs;
8073 state_reg = state->regs;
8074 /* We don't need to worry about FP liveness, it's read-only */
8075 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
8076 err = propagate_liveness_reg(env, &state_reg[i],
8080 if (err == REG_LIVE_READ64)
8081 mark_insn_zext(env, &parent_reg[i]);
8084 /* Propagate stack slots. */
8085 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
8086 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
8087 parent_reg = &parent->stack[i].spilled_ptr;
8088 state_reg = &state->stack[i].spilled_ptr;
8089 err = propagate_liveness_reg(env, state_reg,
8098 /* find precise scalars in the previous equivalent state and
8099 * propagate them into the current state
8101 static int propagate_precision(struct bpf_verifier_env *env,
8102 const struct bpf_verifier_state *old)
8104 struct bpf_reg_state *state_reg;
8105 struct bpf_func_state *state;
8108 state = old->frame[old->curframe];
8109 state_reg = state->regs;
8110 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
8111 if (state_reg->type != SCALAR_VALUE ||
8112 !state_reg->precise)
8114 if (env->log.level & BPF_LOG_LEVEL2)
8115 verbose(env, "propagating r%d\n", i);
8116 err = mark_chain_precision(env, i);
8121 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
8122 if (state->stack[i].slot_type[0] != STACK_SPILL)
8124 state_reg = &state->stack[i].spilled_ptr;
8125 if (state_reg->type != SCALAR_VALUE ||
8126 !state_reg->precise)
8128 if (env->log.level & BPF_LOG_LEVEL2)
8129 verbose(env, "propagating fp%d\n",
8130 (-i - 1) * BPF_REG_SIZE);
8131 err = mark_chain_precision_stack(env, i);
8138 static bool states_maybe_looping(struct bpf_verifier_state *old,
8139 struct bpf_verifier_state *cur)
8141 struct bpf_func_state *fold, *fcur;
8142 int i, fr = cur->curframe;
8144 if (old->curframe != fr)
8147 fold = old->frame[fr];
8148 fcur = cur->frame[fr];
8149 for (i = 0; i < MAX_BPF_REG; i++)
8150 if (memcmp(&fold->regs[i], &fcur->regs[i],
8151 offsetof(struct bpf_reg_state, parent)))
8157 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
8159 struct bpf_verifier_state_list *new_sl;
8160 struct bpf_verifier_state_list *sl, **pprev;
8161 struct bpf_verifier_state *cur = env->cur_state, *new;
8162 int i, j, err, states_cnt = 0;
8163 bool add_new_state = env->test_state_freq ? true : false;
8165 cur->last_insn_idx = env->prev_insn_idx;
8166 if (!env->insn_aux_data[insn_idx].prune_point)
8167 /* this 'insn_idx' instruction wasn't marked, so we will not
8168 * be doing state search here
8172 /* bpf progs typically have pruning point every 4 instructions
8173 * http://vger.kernel.org/bpfconf2019.html#session-1
8174 * Do not add new state for future pruning if the verifier hasn't seen
8175 * at least 2 jumps and at least 8 instructions.
8176 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
8177 * In tests that amounts to up to 50% reduction into total verifier
8178 * memory consumption and 20% verifier time speedup.
8180 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
8181 env->insn_processed - env->prev_insn_processed >= 8)
8182 add_new_state = true;
8184 pprev = explored_state(env, insn_idx);
8187 clean_live_states(env, insn_idx, cur);
8191 if (sl->state.insn_idx != insn_idx)
8193 if (sl->state.branches) {
8194 if (states_maybe_looping(&sl->state, cur) &&
8195 states_equal(env, &sl->state, cur)) {
8196 verbose_linfo(env, insn_idx, "; ");
8197 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
8200 /* if the verifier is processing a loop, avoid adding new state
8201 * too often, since different loop iterations have distinct
8202 * states and may not help future pruning.
8203 * This threshold shouldn't be too low to make sure that
8204 * a loop with large bound will be rejected quickly.
8205 * The most abusive loop will be:
8207 * if r1 < 1000000 goto pc-2
8208 * 1M insn_procssed limit / 100 == 10k peak states.
8209 * This threshold shouldn't be too high either, since states
8210 * at the end of the loop are likely to be useful in pruning.
8212 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
8213 env->insn_processed - env->prev_insn_processed < 100)
8214 add_new_state = false;
8217 if (states_equal(env, &sl->state, cur)) {
8219 /* reached equivalent register/stack state,
8221 * Registers read by the continuation are read by us.
8222 * If we have any write marks in env->cur_state, they
8223 * will prevent corresponding reads in the continuation
8224 * from reaching our parent (an explored_state). Our
8225 * own state will get the read marks recorded, but
8226 * they'll be immediately forgotten as we're pruning
8227 * this state and will pop a new one.
8229 err = propagate_liveness(env, &sl->state, cur);
8231 /* if previous state reached the exit with precision and
8232 * current state is equivalent to it (except precsion marks)
8233 * the precision needs to be propagated back in
8234 * the current state.
8236 err = err ? : push_jmp_history(env, cur);
8237 err = err ? : propagate_precision(env, &sl->state);
8243 /* when new state is not going to be added do not increase miss count.
8244 * Otherwise several loop iterations will remove the state
8245 * recorded earlier. The goal of these heuristics is to have
8246 * states from some iterations of the loop (some in the beginning
8247 * and some at the end) to help pruning.
8251 /* heuristic to determine whether this state is beneficial
8252 * to keep checking from state equivalence point of view.
8253 * Higher numbers increase max_states_per_insn and verification time,
8254 * but do not meaningfully decrease insn_processed.
8256 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
8257 /* the state is unlikely to be useful. Remove it to
8258 * speed up verification
8261 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
8262 u32 br = sl->state.branches;
8265 "BUG live_done but branches_to_explore %d\n",
8267 free_verifier_state(&sl->state, false);
8271 /* cannot free this state, since parentage chain may
8272 * walk it later. Add it for free_list instead to
8273 * be freed at the end of verification
8275 sl->next = env->free_list;
8276 env->free_list = sl;
8286 if (env->max_states_per_insn < states_cnt)
8287 env->max_states_per_insn = states_cnt;
8289 if (!env->allow_ptr_leaks && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
8290 return push_jmp_history(env, cur);
8293 return push_jmp_history(env, cur);
8295 /* There were no equivalent states, remember the current one.
8296 * Technically the current state is not proven to be safe yet,
8297 * but it will either reach outer most bpf_exit (which means it's safe)
8298 * or it will be rejected. When there are no loops the verifier won't be
8299 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
8300 * again on the way to bpf_exit.
8301 * When looping the sl->state.branches will be > 0 and this state
8302 * will not be considered for equivalence until branches == 0.
8304 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
8307 env->total_states++;
8309 env->prev_jmps_processed = env->jmps_processed;
8310 env->prev_insn_processed = env->insn_processed;
8312 /* add new state to the head of linked list */
8313 new = &new_sl->state;
8314 err = copy_verifier_state(new, cur);
8316 free_verifier_state(new, false);
8320 new->insn_idx = insn_idx;
8321 WARN_ONCE(new->branches != 1,
8322 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
8325 cur->first_insn_idx = insn_idx;
8326 clear_jmp_history(cur);
8327 new_sl->next = *explored_state(env, insn_idx);
8328 *explored_state(env, insn_idx) = new_sl;
8329 /* connect new state to parentage chain. Current frame needs all
8330 * registers connected. Only r6 - r9 of the callers are alive (pushed
8331 * to the stack implicitly by JITs) so in callers' frames connect just
8332 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
8333 * the state of the call instruction (with WRITTEN set), and r0 comes
8334 * from callee with its full parentage chain, anyway.
8336 /* clear write marks in current state: the writes we did are not writes
8337 * our child did, so they don't screen off its reads from us.
8338 * (There are no read marks in current state, because reads always mark
8339 * their parent and current state never has children yet. Only
8340 * explored_states can get read marks.)
8342 for (j = 0; j <= cur->curframe; j++) {
8343 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
8344 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
8345 for (i = 0; i < BPF_REG_FP; i++)
8346 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
8349 /* all stack frames are accessible from callee, clear them all */
8350 for (j = 0; j <= cur->curframe; j++) {
8351 struct bpf_func_state *frame = cur->frame[j];
8352 struct bpf_func_state *newframe = new->frame[j];
8354 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
8355 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
8356 frame->stack[i].spilled_ptr.parent =
8357 &newframe->stack[i].spilled_ptr;
8363 /* Return true if it's OK to have the same insn return a different type. */
8364 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
8369 case PTR_TO_SOCKET_OR_NULL:
8370 case PTR_TO_SOCK_COMMON:
8371 case PTR_TO_SOCK_COMMON_OR_NULL:
8372 case PTR_TO_TCP_SOCK:
8373 case PTR_TO_TCP_SOCK_OR_NULL:
8374 case PTR_TO_XDP_SOCK:
8382 /* If an instruction was previously used with particular pointer types, then we
8383 * need to be careful to avoid cases such as the below, where it may be ok
8384 * for one branch accessing the pointer, but not ok for the other branch:
8389 * R1 = some_other_valid_ptr;
8392 * R2 = *(u32 *)(R1 + 0);
8394 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
8396 return src != prev && (!reg_type_mismatch_ok(src) ||
8397 !reg_type_mismatch_ok(prev));
8400 static int do_check(struct bpf_verifier_env *env)
8402 struct bpf_verifier_state *state = env->cur_state;
8403 struct bpf_insn *insns = env->prog->insnsi;
8404 struct bpf_reg_state *regs;
8405 int insn_cnt = env->prog->len;
8406 bool do_print_state = false;
8407 int prev_insn_idx = -1;
8410 struct bpf_insn *insn;
8414 env->prev_insn_idx = prev_insn_idx;
8415 if (env->insn_idx >= insn_cnt) {
8416 verbose(env, "invalid insn idx %d insn_cnt %d\n",
8417 env->insn_idx, insn_cnt);
8421 insn = &insns[env->insn_idx];
8422 class = BPF_CLASS(insn->code);
8424 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
8426 "BPF program is too large. Processed %d insn\n",
8427 env->insn_processed);
8431 err = is_state_visited(env, env->insn_idx);
8435 /* found equivalent state, can prune the search */
8436 if (env->log.level & BPF_LOG_LEVEL) {
8438 verbose(env, "\nfrom %d to %d%s: safe\n",
8439 env->prev_insn_idx, env->insn_idx,
8440 env->cur_state->speculative ?
8441 " (speculative execution)" : "");
8443 verbose(env, "%d: safe\n", env->insn_idx);
8445 goto process_bpf_exit;
8448 if (signal_pending(current))
8454 if (env->log.level & BPF_LOG_LEVEL2 ||
8455 (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
8456 if (env->log.level & BPF_LOG_LEVEL2)
8457 verbose(env, "%d:", env->insn_idx);
8459 verbose(env, "\nfrom %d to %d%s:",
8460 env->prev_insn_idx, env->insn_idx,
8461 env->cur_state->speculative ?
8462 " (speculative execution)" : "");
8463 print_verifier_state(env, state->frame[state->curframe]);
8464 do_print_state = false;
8467 if (env->log.level & BPF_LOG_LEVEL) {
8468 const struct bpf_insn_cbs cbs = {
8469 .cb_print = verbose,
8470 .private_data = env,
8473 verbose_linfo(env, env->insn_idx, "; ");
8474 verbose(env, "%d: ", env->insn_idx);
8475 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
8478 if (bpf_prog_is_dev_bound(env->prog->aux)) {
8479 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
8480 env->prev_insn_idx);
8485 regs = cur_regs(env);
8486 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
8487 prev_insn_idx = env->insn_idx;
8489 if (class == BPF_ALU || class == BPF_ALU64) {
8490 err = check_alu_op(env, insn);
8494 } else if (class == BPF_LDX) {
8495 enum bpf_reg_type *prev_src_type, src_reg_type;
8497 /* check for reserved fields is already done */
8499 /* check src operand */
8500 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8504 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
8508 src_reg_type = regs[insn->src_reg].type;
8510 /* check that memory (src_reg + off) is readable,
8511 * the state of dst_reg will be updated by this func
8513 err = check_mem_access(env, env->insn_idx, insn->src_reg,
8514 insn->off, BPF_SIZE(insn->code),
8515 BPF_READ, insn->dst_reg, false);
8519 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
8521 if (*prev_src_type == NOT_INIT) {
8523 * dst_reg = *(u32 *)(src_reg + off)
8524 * save type to validate intersecting paths
8526 *prev_src_type = src_reg_type;
8528 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
8529 /* ABuser program is trying to use the same insn
8530 * dst_reg = *(u32*) (src_reg + off)
8531 * with different pointer types:
8532 * src_reg == ctx in one branch and
8533 * src_reg == stack|map in some other branch.
8536 verbose(env, "same insn cannot be used with different pointers\n");
8540 } else if (class == BPF_STX) {
8541 enum bpf_reg_type *prev_dst_type, dst_reg_type;
8543 if (BPF_MODE(insn->code) == BPF_XADD) {
8544 err = check_xadd(env, env->insn_idx, insn);
8551 /* check src1 operand */
8552 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8555 /* check src2 operand */
8556 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8560 dst_reg_type = regs[insn->dst_reg].type;
8562 /* check that memory (dst_reg + off) is writeable */
8563 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
8564 insn->off, BPF_SIZE(insn->code),
8565 BPF_WRITE, insn->src_reg, false);
8569 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
8571 if (*prev_dst_type == NOT_INIT) {
8572 *prev_dst_type = dst_reg_type;
8573 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
8574 verbose(env, "same insn cannot be used with different pointers\n");
8578 } else if (class == BPF_ST) {
8579 if (BPF_MODE(insn->code) != BPF_MEM ||
8580 insn->src_reg != BPF_REG_0) {
8581 verbose(env, "BPF_ST uses reserved fields\n");
8584 /* check src operand */
8585 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8589 if (is_ctx_reg(env, insn->dst_reg)) {
8590 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
8592 reg_type_str[reg_state(env, insn->dst_reg)->type]);
8596 /* check that memory (dst_reg + off) is writeable */
8597 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
8598 insn->off, BPF_SIZE(insn->code),
8599 BPF_WRITE, -1, false);
8603 } else if (class == BPF_JMP || class == BPF_JMP32) {
8604 u8 opcode = BPF_OP(insn->code);
8606 env->jmps_processed++;
8607 if (opcode == BPF_CALL) {
8608 if (BPF_SRC(insn->code) != BPF_K ||
8610 (insn->src_reg != BPF_REG_0 &&
8611 insn->src_reg != BPF_PSEUDO_CALL) ||
8612 insn->dst_reg != BPF_REG_0 ||
8613 class == BPF_JMP32) {
8614 verbose(env, "BPF_CALL uses reserved fields\n");
8618 if (env->cur_state->active_spin_lock &&
8619 (insn->src_reg == BPF_PSEUDO_CALL ||
8620 insn->imm != BPF_FUNC_spin_unlock)) {
8621 verbose(env, "function calls are not allowed while holding a lock\n");
8624 if (insn->src_reg == BPF_PSEUDO_CALL)
8625 err = check_func_call(env, insn, &env->insn_idx);
8627 err = check_helper_call(env, insn->imm, env->insn_idx);
8631 } else if (opcode == BPF_JA) {
8632 if (BPF_SRC(insn->code) != BPF_K ||
8634 insn->src_reg != BPF_REG_0 ||
8635 insn->dst_reg != BPF_REG_0 ||
8636 class == BPF_JMP32) {
8637 verbose(env, "BPF_JA uses reserved fields\n");
8641 env->insn_idx += insn->off + 1;
8644 } else if (opcode == BPF_EXIT) {
8645 if (BPF_SRC(insn->code) != BPF_K ||
8647 insn->src_reg != BPF_REG_0 ||
8648 insn->dst_reg != BPF_REG_0 ||
8649 class == BPF_JMP32) {
8650 verbose(env, "BPF_EXIT uses reserved fields\n");
8654 if (env->cur_state->active_spin_lock) {
8655 verbose(env, "bpf_spin_unlock is missing\n");
8659 if (state->curframe) {
8660 /* exit from nested function */
8661 err = prepare_func_exit(env, &env->insn_idx);
8664 do_print_state = true;
8668 err = check_reference_leak(env);
8672 err = check_return_code(env);
8676 update_branch_counts(env, env->cur_state);
8677 err = pop_stack(env, &prev_insn_idx,
8684 do_print_state = true;
8688 err = check_cond_jmp_op(env, insn, &env->insn_idx);
8692 } else if (class == BPF_LD) {
8693 u8 mode = BPF_MODE(insn->code);
8695 if (mode == BPF_ABS || mode == BPF_IND) {
8696 err = check_ld_abs(env, insn);
8700 } else if (mode == BPF_IMM) {
8701 err = check_ld_imm(env, insn);
8706 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
8708 verbose(env, "invalid BPF_LD mode\n");
8712 verbose(env, "unknown insn class %d\n", class);
8722 static int check_map_prealloc(struct bpf_map *map)
8724 return (map->map_type != BPF_MAP_TYPE_HASH &&
8725 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
8726 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
8727 !(map->map_flags & BPF_F_NO_PREALLOC);
8730 static bool is_tracing_prog_type(enum bpf_prog_type type)
8733 case BPF_PROG_TYPE_KPROBE:
8734 case BPF_PROG_TYPE_TRACEPOINT:
8735 case BPF_PROG_TYPE_PERF_EVENT:
8736 case BPF_PROG_TYPE_RAW_TRACEPOINT:
8743 static bool is_preallocated_map(struct bpf_map *map)
8745 if (!check_map_prealloc(map))
8747 if (map->inner_map_meta && !check_map_prealloc(map->inner_map_meta))
8752 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
8753 struct bpf_map *map,
8754 struct bpf_prog *prog)
8758 * Validate that trace type programs use preallocated hash maps.
8760 * For programs attached to PERF events this is mandatory as the
8761 * perf NMI can hit any arbitrary code sequence.
8763 * All other trace types using preallocated hash maps are unsafe as
8764 * well because tracepoint or kprobes can be inside locked regions
8765 * of the memory allocator or at a place where a recursion into the
8766 * memory allocator would see inconsistent state.
8768 * On RT enabled kernels run-time allocation of all trace type
8769 * programs is strictly prohibited due to lock type constraints. On
8770 * !RT kernels it is allowed for backwards compatibility reasons for
8771 * now, but warnings are emitted so developers are made aware of
8772 * the unsafety and can fix their programs before this is enforced.
8774 if (is_tracing_prog_type(prog->type) && !is_preallocated_map(map)) {
8775 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
8776 verbose(env, "perf_event programs can only use preallocated hash map\n");
8779 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
8780 verbose(env, "trace type programs can only use preallocated hash map\n");
8783 WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
8784 verbose(env, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
8787 if ((is_tracing_prog_type(prog->type) ||
8788 prog->type == BPF_PROG_TYPE_SOCKET_FILTER) &&
8789 map_value_has_spin_lock(map)) {
8790 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
8794 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
8795 !bpf_offload_prog_map_match(prog, map)) {
8796 verbose(env, "offload device mismatch between prog and map\n");
8800 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
8801 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
8808 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
8810 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
8811 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
8814 /* look for pseudo eBPF instructions that access map FDs and
8815 * replace them with actual map pointers
8817 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
8819 struct bpf_insn *insn = env->prog->insnsi;
8820 int insn_cnt = env->prog->len;
8823 err = bpf_prog_calc_tag(env->prog);
8827 for (i = 0; i < insn_cnt; i++, insn++) {
8828 if (BPF_CLASS(insn->code) == BPF_LDX &&
8829 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
8830 verbose(env, "BPF_LDX uses reserved fields\n");
8834 if (BPF_CLASS(insn->code) == BPF_STX &&
8835 ((BPF_MODE(insn->code) != BPF_MEM &&
8836 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
8837 verbose(env, "BPF_STX uses reserved fields\n");
8841 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
8842 struct bpf_insn_aux_data *aux;
8843 struct bpf_map *map;
8847 if (i == insn_cnt - 1 || insn[1].code != 0 ||
8848 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
8850 verbose(env, "invalid bpf_ld_imm64 insn\n");
8854 if (insn[0].src_reg == 0)
8855 /* valid generic load 64-bit imm */
8858 /* In final convert_pseudo_ld_imm64() step, this is
8859 * converted into regular 64-bit imm load insn.
8861 if ((insn[0].src_reg != BPF_PSEUDO_MAP_FD &&
8862 insn[0].src_reg != BPF_PSEUDO_MAP_VALUE) ||
8863 (insn[0].src_reg == BPF_PSEUDO_MAP_FD &&
8864 insn[1].imm != 0)) {
8866 "unrecognized bpf_ld_imm64 insn\n");
8870 f = fdget(insn[0].imm);
8871 map = __bpf_map_get(f);
8873 verbose(env, "fd %d is not pointing to valid bpf_map\n",
8875 return PTR_ERR(map);
8878 err = check_map_prog_compatibility(env, map, env->prog);
8884 aux = &env->insn_aux_data[i];
8885 if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
8886 addr = (unsigned long)map;
8888 u32 off = insn[1].imm;
8890 if (off >= BPF_MAX_VAR_OFF) {
8891 verbose(env, "direct value offset of %u is not allowed\n", off);
8896 if (!map->ops->map_direct_value_addr) {
8897 verbose(env, "no direct value access support for this map type\n");
8902 err = map->ops->map_direct_value_addr(map, &addr, off);
8904 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
8905 map->value_size, off);
8914 insn[0].imm = (u32)addr;
8915 insn[1].imm = addr >> 32;
8917 /* check whether we recorded this map already */
8918 for (j = 0; j < env->used_map_cnt; j++) {
8919 if (env->used_maps[j] == map) {
8926 if (env->used_map_cnt >= MAX_USED_MAPS) {
8931 /* hold the map. If the program is rejected by verifier,
8932 * the map will be released by release_maps() or it
8933 * will be used by the valid program until it's unloaded
8934 * and all maps are released in free_used_maps()
8938 aux->map_index = env->used_map_cnt;
8939 env->used_maps[env->used_map_cnt++] = map;
8941 if (bpf_map_is_cgroup_storage(map) &&
8942 bpf_cgroup_storage_assign(env->prog->aux, map)) {
8943 verbose(env, "only one cgroup storage of each type is allowed\n");
8955 /* Basic sanity check before we invest more work here. */
8956 if (!bpf_opcode_in_insntable(insn->code)) {
8957 verbose(env, "unknown opcode %02x\n", insn->code);
8962 /* now all pseudo BPF_LD_IMM64 instructions load valid
8963 * 'struct bpf_map *' into a register instead of user map_fd.
8964 * These pointers will be used later by verifier to validate map access.
8969 /* drop refcnt of maps used by the rejected program */
8970 static void release_maps(struct bpf_verifier_env *env)
8972 __bpf_free_used_maps(env->prog->aux, env->used_maps,
8976 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
8977 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
8979 struct bpf_insn *insn = env->prog->insnsi;
8980 int insn_cnt = env->prog->len;
8983 for (i = 0; i < insn_cnt; i++, insn++)
8984 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
8988 /* single env->prog->insni[off] instruction was replaced with the range
8989 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
8990 * [0, off) and [off, end) to new locations, so the patched range stays zero
8992 static int adjust_insn_aux_data(struct bpf_verifier_env *env,
8993 struct bpf_prog *new_prog, u32 off, u32 cnt)
8995 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
8996 struct bpf_insn *insn = new_prog->insnsi;
9000 /* aux info at OFF always needs adjustment, no matter fast path
9001 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
9002 * original insn at old prog.
9004 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
9008 prog_len = new_prog->len;
9009 new_data = vzalloc(array_size(prog_len,
9010 sizeof(struct bpf_insn_aux_data)));
9013 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
9014 memcpy(new_data + off + cnt - 1, old_data + off,
9015 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
9016 for (i = off; i < off + cnt - 1; i++) {
9017 new_data[i].seen = env->pass_cnt;
9018 new_data[i].zext_dst = insn_has_def32(env, insn + i);
9020 env->insn_aux_data = new_data;
9025 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
9031 /* NOTE: fake 'exit' subprog should be updated as well. */
9032 for (i = 0; i <= env->subprog_cnt; i++) {
9033 if (env->subprog_info[i].start <= off)
9035 env->subprog_info[i].start += len - 1;
9039 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
9040 const struct bpf_insn *patch, u32 len)
9042 struct bpf_prog *new_prog;
9044 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
9045 if (IS_ERR(new_prog)) {
9046 if (PTR_ERR(new_prog) == -ERANGE)
9048 "insn %d cannot be patched due to 16-bit range\n",
9049 env->insn_aux_data[off].orig_idx);
9052 if (adjust_insn_aux_data(env, new_prog, off, len))
9054 adjust_subprog_starts(env, off, len);
9058 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
9063 /* find first prog starting at or after off (first to remove) */
9064 for (i = 0; i < env->subprog_cnt; i++)
9065 if (env->subprog_info[i].start >= off)
9067 /* find first prog starting at or after off + cnt (first to stay) */
9068 for (j = i; j < env->subprog_cnt; j++)
9069 if (env->subprog_info[j].start >= off + cnt)
9071 /* if j doesn't start exactly at off + cnt, we are just removing
9072 * the front of previous prog
9074 if (env->subprog_info[j].start != off + cnt)
9078 struct bpf_prog_aux *aux = env->prog->aux;
9081 /* move fake 'exit' subprog as well */
9082 move = env->subprog_cnt + 1 - j;
9084 memmove(env->subprog_info + i,
9085 env->subprog_info + j,
9086 sizeof(*env->subprog_info) * move);
9087 env->subprog_cnt -= j - i;
9089 /* remove func_info */
9090 if (aux->func_info) {
9091 move = aux->func_info_cnt - j;
9093 memmove(aux->func_info + i,
9095 sizeof(*aux->func_info) * move);
9096 aux->func_info_cnt -= j - i;
9097 /* func_info->insn_off is set after all code rewrites,
9098 * in adjust_btf_func() - no need to adjust
9102 /* convert i from "first prog to remove" to "first to adjust" */
9103 if (env->subprog_info[i].start == off)
9107 /* update fake 'exit' subprog as well */
9108 for (; i <= env->subprog_cnt; i++)
9109 env->subprog_info[i].start -= cnt;
9114 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
9117 struct bpf_prog *prog = env->prog;
9118 u32 i, l_off, l_cnt, nr_linfo;
9119 struct bpf_line_info *linfo;
9121 nr_linfo = prog->aux->nr_linfo;
9125 linfo = prog->aux->linfo;
9127 /* find first line info to remove, count lines to be removed */
9128 for (i = 0; i < nr_linfo; i++)
9129 if (linfo[i].insn_off >= off)
9134 for (; i < nr_linfo; i++)
9135 if (linfo[i].insn_off < off + cnt)
9140 /* First live insn doesn't match first live linfo, it needs to "inherit"
9141 * last removed linfo. prog is already modified, so prog->len == off
9142 * means no live instructions after (tail of the program was removed).
9144 if (prog->len != off && l_cnt &&
9145 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
9147 linfo[--i].insn_off = off + cnt;
9150 /* remove the line info which refer to the removed instructions */
9152 memmove(linfo + l_off, linfo + i,
9153 sizeof(*linfo) * (nr_linfo - i));
9155 prog->aux->nr_linfo -= l_cnt;
9156 nr_linfo = prog->aux->nr_linfo;
9159 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
9160 for (i = l_off; i < nr_linfo; i++)
9161 linfo[i].insn_off -= cnt;
9163 /* fix up all subprogs (incl. 'exit') which start >= off */
9164 for (i = 0; i <= env->subprog_cnt; i++)
9165 if (env->subprog_info[i].linfo_idx > l_off) {
9166 /* program may have started in the removed region but
9167 * may not be fully removed
9169 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
9170 env->subprog_info[i].linfo_idx -= l_cnt;
9172 env->subprog_info[i].linfo_idx = l_off;
9178 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
9180 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
9181 unsigned int orig_prog_len = env->prog->len;
9184 if (bpf_prog_is_dev_bound(env->prog->aux))
9185 bpf_prog_offload_remove_insns(env, off, cnt);
9187 err = bpf_remove_insns(env->prog, off, cnt);
9191 err = adjust_subprog_starts_after_remove(env, off, cnt);
9195 err = bpf_adj_linfo_after_remove(env, off, cnt);
9199 memmove(aux_data + off, aux_data + off + cnt,
9200 sizeof(*aux_data) * (orig_prog_len - off - cnt));
9205 /* The verifier does more data flow analysis than llvm and will not
9206 * explore branches that are dead at run time. Malicious programs can
9207 * have dead code too. Therefore replace all dead at-run-time code
9210 * Just nops are not optimal, e.g. if they would sit at the end of the
9211 * program and through another bug we would manage to jump there, then
9212 * we'd execute beyond program memory otherwise. Returning exception
9213 * code also wouldn't work since we can have subprogs where the dead
9214 * code could be located.
9216 static void sanitize_dead_code(struct bpf_verifier_env *env)
9218 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
9219 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
9220 struct bpf_insn *insn = env->prog->insnsi;
9221 const int insn_cnt = env->prog->len;
9224 for (i = 0; i < insn_cnt; i++) {
9225 if (aux_data[i].seen)
9227 memcpy(insn + i, &trap, sizeof(trap));
9231 static bool insn_is_cond_jump(u8 code)
9235 if (BPF_CLASS(code) == BPF_JMP32)
9238 if (BPF_CLASS(code) != BPF_JMP)
9242 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
9245 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
9247 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
9248 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
9249 struct bpf_insn *insn = env->prog->insnsi;
9250 const int insn_cnt = env->prog->len;
9253 for (i = 0; i < insn_cnt; i++, insn++) {
9254 if (!insn_is_cond_jump(insn->code))
9257 if (!aux_data[i + 1].seen)
9259 else if (!aux_data[i + 1 + insn->off].seen)
9264 if (bpf_prog_is_dev_bound(env->prog->aux))
9265 bpf_prog_offload_replace_insn(env, i, &ja);
9267 memcpy(insn, &ja, sizeof(ja));
9271 static int opt_remove_dead_code(struct bpf_verifier_env *env)
9273 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
9274 int insn_cnt = env->prog->len;
9277 for (i = 0; i < insn_cnt; i++) {
9281 while (i + j < insn_cnt && !aux_data[i + j].seen)
9286 err = verifier_remove_insns(env, i, j);
9289 insn_cnt = env->prog->len;
9295 static int opt_remove_nops(struct bpf_verifier_env *env)
9297 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
9298 struct bpf_insn *insn = env->prog->insnsi;
9299 int insn_cnt = env->prog->len;
9302 for (i = 0; i < insn_cnt; i++) {
9303 if (memcmp(&insn[i], &ja, sizeof(ja)))
9306 err = verifier_remove_insns(env, i, 1);
9316 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
9317 const union bpf_attr *attr)
9319 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
9320 struct bpf_insn_aux_data *aux = env->insn_aux_data;
9321 int i, patch_len, delta = 0, len = env->prog->len;
9322 struct bpf_insn *insns = env->prog->insnsi;
9323 struct bpf_prog *new_prog;
9326 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
9327 zext_patch[1] = BPF_ZEXT_REG(0);
9328 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
9329 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
9330 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
9331 for (i = 0; i < len; i++) {
9332 int adj_idx = i + delta;
9333 struct bpf_insn insn;
9335 insn = insns[adj_idx];
9336 if (!aux[adj_idx].zext_dst) {
9344 class = BPF_CLASS(code);
9345 if (insn_no_def(&insn))
9348 /* NOTE: arg "reg" (the fourth one) is only used for
9349 * BPF_STX which has been ruled out in above
9350 * check, it is safe to pass NULL here.
9352 if (is_reg64(env, &insn, insn.dst_reg, NULL, DST_OP)) {
9353 if (class == BPF_LD &&
9354 BPF_MODE(code) == BPF_IMM)
9359 /* ctx load could be transformed into wider load. */
9360 if (class == BPF_LDX &&
9361 aux[adj_idx].ptr_type == PTR_TO_CTX)
9364 imm_rnd = get_random_int();
9365 rnd_hi32_patch[0] = insn;
9366 rnd_hi32_patch[1].imm = imm_rnd;
9367 rnd_hi32_patch[3].dst_reg = insn.dst_reg;
9368 patch = rnd_hi32_patch;
9370 goto apply_patch_buffer;
9373 if (!bpf_jit_needs_zext())
9376 zext_patch[0] = insn;
9377 zext_patch[1].dst_reg = insn.dst_reg;
9378 zext_patch[1].src_reg = insn.dst_reg;
9382 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
9385 env->prog = new_prog;
9386 insns = new_prog->insnsi;
9387 aux = env->insn_aux_data;
9388 delta += patch_len - 1;
9394 /* convert load instructions that access fields of a context type into a
9395 * sequence of instructions that access fields of the underlying structure:
9396 * struct __sk_buff -> struct sk_buff
9397 * struct bpf_sock_ops -> struct sock
9399 static int convert_ctx_accesses(struct bpf_verifier_env *env)
9401 const struct bpf_verifier_ops *ops = env->ops;
9402 int i, cnt, size, ctx_field_size, delta = 0;
9403 const int insn_cnt = env->prog->len;
9404 struct bpf_insn insn_buf[16], *insn;
9405 u32 target_size, size_default, off;
9406 struct bpf_prog *new_prog;
9407 enum bpf_access_type type;
9408 bool is_narrower_load;
9410 if (ops->gen_prologue || env->seen_direct_write) {
9411 if (!ops->gen_prologue) {
9412 verbose(env, "bpf verifier is misconfigured\n");
9415 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
9417 if (cnt >= ARRAY_SIZE(insn_buf)) {
9418 verbose(env, "bpf verifier is misconfigured\n");
9421 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
9425 env->prog = new_prog;
9430 if (bpf_prog_is_dev_bound(env->prog->aux))
9433 insn = env->prog->insnsi + delta;
9435 for (i = 0; i < insn_cnt; i++, insn++) {
9436 bpf_convert_ctx_access_t convert_ctx_access;
9438 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
9439 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
9440 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
9441 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
9443 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
9444 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
9445 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
9446 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
9451 if (type == BPF_WRITE &&
9452 env->insn_aux_data[i + delta].sanitize_stack_off) {
9453 struct bpf_insn patch[] = {
9454 /* Sanitize suspicious stack slot with zero.
9455 * There are no memory dependencies for this store,
9456 * since it's only using frame pointer and immediate
9459 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
9460 env->insn_aux_data[i + delta].sanitize_stack_off,
9462 /* the original STX instruction will immediately
9463 * overwrite the same stack slot with appropriate value
9468 cnt = ARRAY_SIZE(patch);
9469 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
9474 env->prog = new_prog;
9475 insn = new_prog->insnsi + i + delta;
9479 switch (env->insn_aux_data[i + delta].ptr_type) {
9481 if (!ops->convert_ctx_access)
9483 convert_ctx_access = ops->convert_ctx_access;
9486 case PTR_TO_SOCK_COMMON:
9487 convert_ctx_access = bpf_sock_convert_ctx_access;
9489 case PTR_TO_TCP_SOCK:
9490 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
9492 case PTR_TO_XDP_SOCK:
9493 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
9496 if (type == BPF_READ) {
9497 insn->code = BPF_LDX | BPF_PROBE_MEM |
9498 BPF_SIZE((insn)->code);
9499 env->prog->aux->num_exentries++;
9500 } else if (env->prog->type != BPF_PROG_TYPE_STRUCT_OPS) {
9501 verbose(env, "Writes through BTF pointers are not allowed\n");
9509 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
9510 size = BPF_LDST_BYTES(insn);
9512 /* If the read access is a narrower load of the field,
9513 * convert to a 4/8-byte load, to minimum program type specific
9514 * convert_ctx_access changes. If conversion is successful,
9515 * we will apply proper mask to the result.
9517 is_narrower_load = size < ctx_field_size;
9518 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
9520 if (is_narrower_load) {
9523 if (type == BPF_WRITE) {
9524 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
9529 if (ctx_field_size == 4)
9531 else if (ctx_field_size == 8)
9534 insn->off = off & ~(size_default - 1);
9535 insn->code = BPF_LDX | BPF_MEM | size_code;
9539 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
9541 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
9542 (ctx_field_size && !target_size)) {
9543 verbose(env, "bpf verifier is misconfigured\n");
9547 if (is_narrower_load && size < target_size) {
9548 u8 shift = bpf_ctx_narrow_access_offset(
9549 off, size, size_default) * 8;
9550 if (ctx_field_size <= 4) {
9552 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
9555 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
9556 (1 << size * 8) - 1);
9559 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
9562 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
9563 (1ULL << size * 8) - 1);
9567 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9573 /* keep walking new program and skip insns we just inserted */
9574 env->prog = new_prog;
9575 insn = new_prog->insnsi + i + delta;
9581 static int jit_subprogs(struct bpf_verifier_env *env)
9583 struct bpf_prog *prog = env->prog, **func, *tmp;
9584 int i, j, subprog_start, subprog_end = 0, len, subprog;
9585 struct bpf_insn *insn;
9589 if (env->subprog_cnt <= 1)
9592 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
9593 if (insn->code != (BPF_JMP | BPF_CALL) ||
9594 insn->src_reg != BPF_PSEUDO_CALL)
9596 /* Upon error here we cannot fall back to interpreter but
9597 * need a hard reject of the program. Thus -EFAULT is
9598 * propagated in any case.
9600 subprog = find_subprog(env, i + insn->imm + 1);
9602 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
9606 /* temporarily remember subprog id inside insn instead of
9607 * aux_data, since next loop will split up all insns into funcs
9609 insn->off = subprog;
9610 /* remember original imm in case JIT fails and fallback
9611 * to interpreter will be needed
9613 env->insn_aux_data[i].call_imm = insn->imm;
9614 /* point imm to __bpf_call_base+1 from JITs point of view */
9618 err = bpf_prog_alloc_jited_linfo(prog);
9623 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
9627 for (i = 0; i < env->subprog_cnt; i++) {
9628 subprog_start = subprog_end;
9629 subprog_end = env->subprog_info[i + 1].start;
9631 len = subprog_end - subprog_start;
9632 /* BPF_PROG_RUN doesn't call subprogs directly,
9633 * hence main prog stats include the runtime of subprogs.
9634 * subprogs don't have IDs and not reachable via prog_get_next_id
9635 * func[i]->aux->stats will never be accessed and stays NULL
9637 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
9640 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
9641 len * sizeof(struct bpf_insn));
9642 func[i]->type = prog->type;
9644 if (bpf_prog_calc_tag(func[i]))
9646 func[i]->is_func = 1;
9647 func[i]->aux->func_idx = i;
9648 /* the btf and func_info will be freed only at prog->aux */
9649 func[i]->aux->btf = prog->aux->btf;
9650 func[i]->aux->func_info = prog->aux->func_info;
9652 /* Use bpf_prog_F_tag to indicate functions in stack traces.
9653 * Long term would need debug info to populate names
9655 func[i]->aux->name[0] = 'F';
9656 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
9657 func[i]->jit_requested = 1;
9658 func[i]->aux->linfo = prog->aux->linfo;
9659 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
9660 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
9661 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
9662 func[i] = bpf_int_jit_compile(func[i]);
9663 if (!func[i]->jited) {
9669 /* at this point all bpf functions were successfully JITed
9670 * now populate all bpf_calls with correct addresses and
9671 * run last pass of JIT
9673 for (i = 0; i < env->subprog_cnt; i++) {
9674 insn = func[i]->insnsi;
9675 for (j = 0; j < func[i]->len; j++, insn++) {
9676 if (insn->code != (BPF_JMP | BPF_CALL) ||
9677 insn->src_reg != BPF_PSEUDO_CALL)
9679 subprog = insn->off;
9680 insn->imm = BPF_CAST_CALL(func[subprog]->bpf_func) -
9684 /* we use the aux data to keep a list of the start addresses
9685 * of the JITed images for each function in the program
9687 * for some architectures, such as powerpc64, the imm field
9688 * might not be large enough to hold the offset of the start
9689 * address of the callee's JITed image from __bpf_call_base
9691 * in such cases, we can lookup the start address of a callee
9692 * by using its subprog id, available from the off field of
9693 * the call instruction, as an index for this list
9695 func[i]->aux->func = func;
9696 func[i]->aux->func_cnt = env->subprog_cnt;
9698 for (i = 0; i < env->subprog_cnt; i++) {
9699 old_bpf_func = func[i]->bpf_func;
9700 tmp = bpf_int_jit_compile(func[i]);
9701 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
9702 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
9709 /* finally lock prog and jit images for all functions and
9712 for (i = 0; i < env->subprog_cnt; i++) {
9713 bpf_prog_lock_ro(func[i]);
9714 bpf_prog_kallsyms_add(func[i]);
9717 /* Last step: make now unused interpreter insns from main
9718 * prog consistent for later dump requests, so they can
9719 * later look the same as if they were interpreted only.
9721 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
9722 if (insn->code != (BPF_JMP | BPF_CALL) ||
9723 insn->src_reg != BPF_PSEUDO_CALL)
9725 insn->off = env->insn_aux_data[i].call_imm;
9726 subprog = find_subprog(env, i + insn->off + 1);
9727 insn->imm = subprog;
9731 prog->bpf_func = func[0]->bpf_func;
9732 prog->aux->func = func;
9733 prog->aux->func_cnt = env->subprog_cnt;
9734 bpf_prog_free_unused_jited_linfo(prog);
9737 for (i = 0; i < env->subprog_cnt; i++)
9739 bpf_jit_free(func[i]);
9742 /* cleanup main prog to be interpreted */
9743 prog->jit_requested = 0;
9744 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
9745 if (insn->code != (BPF_JMP | BPF_CALL) ||
9746 insn->src_reg != BPF_PSEUDO_CALL)
9749 insn->imm = env->insn_aux_data[i].call_imm;
9751 bpf_prog_free_jited_linfo(prog);
9755 static int fixup_call_args(struct bpf_verifier_env *env)
9757 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
9758 struct bpf_prog *prog = env->prog;
9759 struct bpf_insn *insn = prog->insnsi;
9764 if (env->prog->jit_requested &&
9765 !bpf_prog_is_dev_bound(env->prog->aux)) {
9766 err = jit_subprogs(env);
9772 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
9773 for (i = 0; i < prog->len; i++, insn++) {
9774 if (insn->code != (BPF_JMP | BPF_CALL) ||
9775 insn->src_reg != BPF_PSEUDO_CALL)
9777 depth = get_callee_stack_depth(env, insn, i);
9780 bpf_patch_call_args(insn, depth);
9787 /* fixup insn->imm field of bpf_call instructions
9788 * and inline eligible helpers as explicit sequence of BPF instructions
9790 * this function is called after eBPF program passed verification
9792 static int fixup_bpf_calls(struct bpf_verifier_env *env)
9794 struct bpf_prog *prog = env->prog;
9795 bool expect_blinding = bpf_jit_blinding_enabled(prog);
9796 struct bpf_insn *insn = prog->insnsi;
9797 const struct bpf_func_proto *fn;
9798 const int insn_cnt = prog->len;
9799 const struct bpf_map_ops *ops;
9800 struct bpf_insn_aux_data *aux;
9801 struct bpf_insn insn_buf[16];
9802 struct bpf_prog *new_prog;
9803 struct bpf_map *map_ptr;
9804 int i, ret, cnt, delta = 0;
9806 for (i = 0; i < insn_cnt; i++, insn++) {
9807 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
9808 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
9809 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
9810 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
9811 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
9812 struct bpf_insn mask_and_div[] = {
9813 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
9815 BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2),
9816 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
9817 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
9820 struct bpf_insn mask_and_mod[] = {
9821 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
9822 /* Rx mod 0 -> Rx */
9823 BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1),
9826 struct bpf_insn *patchlet;
9828 if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
9829 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
9830 patchlet = mask_and_div + (is64 ? 1 : 0);
9831 cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0);
9833 patchlet = mask_and_mod + (is64 ? 1 : 0);
9834 cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0);
9837 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
9842 env->prog = prog = new_prog;
9843 insn = new_prog->insnsi + i + delta;
9847 if (BPF_CLASS(insn->code) == BPF_LD &&
9848 (BPF_MODE(insn->code) == BPF_ABS ||
9849 BPF_MODE(insn->code) == BPF_IND)) {
9850 cnt = env->ops->gen_ld_abs(insn, insn_buf);
9851 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
9852 verbose(env, "bpf verifier is misconfigured\n");
9856 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9861 env->prog = prog = new_prog;
9862 insn = new_prog->insnsi + i + delta;
9866 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
9867 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
9868 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
9869 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
9870 struct bpf_insn insn_buf[16];
9871 struct bpf_insn *patch = &insn_buf[0];
9875 aux = &env->insn_aux_data[i + delta];
9876 if (!aux->alu_state ||
9877 aux->alu_state == BPF_ALU_NON_POINTER)
9880 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
9881 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
9882 BPF_ALU_SANITIZE_SRC;
9884 off_reg = issrc ? insn->src_reg : insn->dst_reg;
9886 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
9887 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit - 1);
9888 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
9889 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
9890 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
9891 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
9893 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX,
9895 insn->src_reg = BPF_REG_AX;
9897 *patch++ = BPF_ALU64_REG(BPF_AND, off_reg,
9901 insn->code = insn->code == code_add ?
9902 code_sub : code_add;
9905 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
9906 cnt = patch - insn_buf;
9908 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9913 env->prog = prog = new_prog;
9914 insn = new_prog->insnsi + i + delta;
9918 if (insn->code != (BPF_JMP | BPF_CALL))
9920 if (insn->src_reg == BPF_PSEUDO_CALL)
9923 if (insn->imm == BPF_FUNC_get_route_realm)
9924 prog->dst_needed = 1;
9925 if (insn->imm == BPF_FUNC_get_prandom_u32)
9926 bpf_user_rnd_init_once();
9927 if (insn->imm == BPF_FUNC_override_return)
9928 prog->kprobe_override = 1;
9929 if (insn->imm == BPF_FUNC_tail_call) {
9930 /* If we tail call into other programs, we
9931 * cannot make any assumptions since they can
9932 * be replaced dynamically during runtime in
9933 * the program array.
9935 prog->cb_access = 1;
9936 env->prog->aux->stack_depth = MAX_BPF_STACK;
9937 env->prog->aux->max_pkt_offset = MAX_PACKET_OFF;
9939 /* mark bpf_tail_call as different opcode to avoid
9940 * conditional branch in the interpeter for every normal
9941 * call and to prevent accidental JITing by JIT compiler
9942 * that doesn't support bpf_tail_call yet
9945 insn->code = BPF_JMP | BPF_TAIL_CALL;
9947 aux = &env->insn_aux_data[i + delta];
9948 if (env->allow_ptr_leaks && !expect_blinding &&
9949 prog->jit_requested &&
9950 !bpf_map_key_poisoned(aux) &&
9951 !bpf_map_ptr_poisoned(aux) &&
9952 !bpf_map_ptr_unpriv(aux)) {
9953 struct bpf_jit_poke_descriptor desc = {
9954 .reason = BPF_POKE_REASON_TAIL_CALL,
9955 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
9956 .tail_call.key = bpf_map_key_immediate(aux),
9959 ret = bpf_jit_add_poke_descriptor(prog, &desc);
9961 verbose(env, "adding tail call poke descriptor failed\n");
9965 insn->imm = ret + 1;
9969 if (!bpf_map_ptr_unpriv(aux))
9972 /* instead of changing every JIT dealing with tail_call
9973 * emit two extra insns:
9974 * if (index >= max_entries) goto out;
9975 * index &= array->index_mask;
9976 * to avoid out-of-bounds cpu speculation
9978 if (bpf_map_ptr_poisoned(aux)) {
9979 verbose(env, "tail_call abusing map_ptr\n");
9983 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
9984 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
9985 map_ptr->max_entries, 2);
9986 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
9987 container_of(map_ptr,
9990 insn_buf[2] = *insn;
9992 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9997 env->prog = prog = new_prog;
9998 insn = new_prog->insnsi + i + delta;
10002 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
10003 * and other inlining handlers are currently limited to 64 bit
10006 if (prog->jit_requested && BITS_PER_LONG == 64 &&
10007 (insn->imm == BPF_FUNC_map_lookup_elem ||
10008 insn->imm == BPF_FUNC_map_update_elem ||
10009 insn->imm == BPF_FUNC_map_delete_elem ||
10010 insn->imm == BPF_FUNC_map_push_elem ||
10011 insn->imm == BPF_FUNC_map_pop_elem ||
10012 insn->imm == BPF_FUNC_map_peek_elem)) {
10013 aux = &env->insn_aux_data[i + delta];
10014 if (bpf_map_ptr_poisoned(aux))
10015 goto patch_call_imm;
10017 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
10018 ops = map_ptr->ops;
10019 if (insn->imm == BPF_FUNC_map_lookup_elem &&
10020 ops->map_gen_lookup) {
10021 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
10022 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
10023 verbose(env, "bpf verifier is misconfigured\n");
10027 new_prog = bpf_patch_insn_data(env, i + delta,
10033 env->prog = prog = new_prog;
10034 insn = new_prog->insnsi + i + delta;
10038 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
10039 (void *(*)(struct bpf_map *map, void *key))NULL));
10040 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
10041 (int (*)(struct bpf_map *map, void *key))NULL));
10042 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
10043 (int (*)(struct bpf_map *map, void *key, void *value,
10045 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
10046 (int (*)(struct bpf_map *map, void *value,
10048 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
10049 (int (*)(struct bpf_map *map, void *value))NULL));
10050 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
10051 (int (*)(struct bpf_map *map, void *value))NULL));
10053 switch (insn->imm) {
10054 case BPF_FUNC_map_lookup_elem:
10055 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
10058 case BPF_FUNC_map_update_elem:
10059 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
10062 case BPF_FUNC_map_delete_elem:
10063 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
10066 case BPF_FUNC_map_push_elem:
10067 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
10070 case BPF_FUNC_map_pop_elem:
10071 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
10074 case BPF_FUNC_map_peek_elem:
10075 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
10080 goto patch_call_imm;
10083 if (prog->jit_requested && BITS_PER_LONG == 64 &&
10084 insn->imm == BPF_FUNC_jiffies64) {
10085 struct bpf_insn ld_jiffies_addr[2] = {
10086 BPF_LD_IMM64(BPF_REG_0,
10087 (unsigned long)&jiffies),
10090 insn_buf[0] = ld_jiffies_addr[0];
10091 insn_buf[1] = ld_jiffies_addr[1];
10092 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
10096 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
10102 env->prog = prog = new_prog;
10103 insn = new_prog->insnsi + i + delta;
10108 fn = env->ops->get_func_proto(insn->imm, env->prog);
10109 /* all functions that have prototype and verifier allowed
10110 * programs to call them, must be real in-kernel functions
10114 "kernel subsystem misconfigured func %s#%d\n",
10115 func_id_name(insn->imm), insn->imm);
10118 insn->imm = fn->func - __bpf_call_base;
10121 /* Since poke tab is now finalized, publish aux to tracker. */
10122 for (i = 0; i < prog->aux->size_poke_tab; i++) {
10123 map_ptr = prog->aux->poke_tab[i].tail_call.map;
10124 if (!map_ptr->ops->map_poke_track ||
10125 !map_ptr->ops->map_poke_untrack ||
10126 !map_ptr->ops->map_poke_run) {
10127 verbose(env, "bpf verifier is misconfigured\n");
10131 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
10133 verbose(env, "tracking tail call prog failed\n");
10141 static void free_states(struct bpf_verifier_env *env)
10143 struct bpf_verifier_state_list *sl, *sln;
10146 sl = env->free_list;
10149 free_verifier_state(&sl->state, false);
10153 env->free_list = NULL;
10155 if (!env->explored_states)
10158 for (i = 0; i < state_htab_size(env); i++) {
10159 sl = env->explored_states[i];
10163 free_verifier_state(&sl->state, false);
10167 env->explored_states[i] = NULL;
10171 /* The verifier is using insn_aux_data[] to store temporary data during
10172 * verification and to store information for passes that run after the
10173 * verification like dead code sanitization. do_check_common() for subprogram N
10174 * may analyze many other subprograms. sanitize_insn_aux_data() clears all
10175 * temporary data after do_check_common() finds that subprogram N cannot be
10176 * verified independently. pass_cnt counts the number of times
10177 * do_check_common() was run and insn->aux->seen tells the pass number
10178 * insn_aux_data was touched. These variables are compared to clear temporary
10179 * data from failed pass. For testing and experiments do_check_common() can be
10180 * run multiple times even when prior attempt to verify is unsuccessful.
10182 static void sanitize_insn_aux_data(struct bpf_verifier_env *env)
10184 struct bpf_insn *insn = env->prog->insnsi;
10185 struct bpf_insn_aux_data *aux;
10188 for (i = 0; i < env->prog->len; i++) {
10189 class = BPF_CLASS(insn[i].code);
10190 if (class != BPF_LDX && class != BPF_STX)
10192 aux = &env->insn_aux_data[i];
10193 if (aux->seen != env->pass_cnt)
10195 memset(aux, 0, offsetof(typeof(*aux), orig_idx));
10199 static int do_check_common(struct bpf_verifier_env *env, int subprog)
10201 struct bpf_verifier_state *state;
10202 struct bpf_reg_state *regs;
10205 env->prev_linfo = NULL;
10208 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
10211 state->curframe = 0;
10212 state->speculative = false;
10213 state->branches = 1;
10214 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
10215 if (!state->frame[0]) {
10219 env->cur_state = state;
10220 init_func_state(env, state->frame[0],
10221 BPF_MAIN_FUNC /* callsite */,
10225 regs = state->frame[state->curframe]->regs;
10226 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
10227 ret = btf_prepare_func_args(env, subprog, regs);
10230 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
10231 if (regs[i].type == PTR_TO_CTX)
10232 mark_reg_known_zero(env, regs, i);
10233 else if (regs[i].type == SCALAR_VALUE)
10234 mark_reg_unknown(env, regs, i);
10237 /* 1st arg to a function */
10238 regs[BPF_REG_1].type = PTR_TO_CTX;
10239 mark_reg_known_zero(env, regs, BPF_REG_1);
10240 ret = btf_check_func_arg_match(env, subprog, regs);
10241 if (ret == -EFAULT)
10242 /* unlikely verifier bug. abort.
10243 * ret == 0 and ret < 0 are sadly acceptable for
10244 * main() function due to backward compatibility.
10245 * Like socket filter program may be written as:
10246 * int bpf_prog(struct pt_regs *ctx)
10247 * and never dereference that ctx in the program.
10248 * 'struct pt_regs' is a type mismatch for socket
10249 * filter that should be using 'struct __sk_buff'.
10254 ret = do_check(env);
10256 /* check for NULL is necessary, since cur_state can be freed inside
10257 * do_check() under memory pressure.
10259 if (env->cur_state) {
10260 free_verifier_state(env->cur_state, true);
10261 env->cur_state = NULL;
10263 while (!pop_stack(env, NULL, NULL));
10266 /* clean aux data in case subprog was rejected */
10267 sanitize_insn_aux_data(env);
10271 /* Verify all global functions in a BPF program one by one based on their BTF.
10272 * All global functions must pass verification. Otherwise the whole program is rejected.
10283 * foo() will be verified first for R1=any_scalar_value. During verification it
10284 * will be assumed that bar() already verified successfully and call to bar()
10285 * from foo() will be checked for type match only. Later bar() will be verified
10286 * independently to check that it's safe for R1=any_scalar_value.
10288 static int do_check_subprogs(struct bpf_verifier_env *env)
10290 struct bpf_prog_aux *aux = env->prog->aux;
10293 if (!aux->func_info)
10296 for (i = 1; i < env->subprog_cnt; i++) {
10297 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
10299 env->insn_idx = env->subprog_info[i].start;
10300 WARN_ON_ONCE(env->insn_idx == 0);
10301 ret = do_check_common(env, i);
10304 } else if (env->log.level & BPF_LOG_LEVEL) {
10306 "Func#%d is safe for any args that match its prototype\n",
10313 static int do_check_main(struct bpf_verifier_env *env)
10318 ret = do_check_common(env, 0);
10320 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
10325 static void print_verification_stats(struct bpf_verifier_env *env)
10329 if (env->log.level & BPF_LOG_STATS) {
10330 verbose(env, "verification time %lld usec\n",
10331 div_u64(env->verification_time, 1000));
10332 verbose(env, "stack depth ");
10333 for (i = 0; i < env->subprog_cnt; i++) {
10334 u32 depth = env->subprog_info[i].stack_depth;
10336 verbose(env, "%d", depth);
10337 if (i + 1 < env->subprog_cnt)
10340 verbose(env, "\n");
10342 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
10343 "total_states %d peak_states %d mark_read %d\n",
10344 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
10345 env->max_states_per_insn, env->total_states,
10346 env->peak_states, env->longest_mark_read_walk);
10349 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
10351 const struct btf_type *t, *func_proto;
10352 const struct bpf_struct_ops *st_ops;
10353 const struct btf_member *member;
10354 struct bpf_prog *prog = env->prog;
10355 u32 btf_id, member_idx;
10358 btf_id = prog->aux->attach_btf_id;
10359 st_ops = bpf_struct_ops_find(btf_id);
10361 verbose(env, "attach_btf_id %u is not a supported struct\n",
10367 member_idx = prog->expected_attach_type;
10368 if (member_idx >= btf_type_vlen(t)) {
10369 verbose(env, "attach to invalid member idx %u of struct %s\n",
10370 member_idx, st_ops->name);
10374 member = &btf_type_member(t)[member_idx];
10375 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
10376 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
10379 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
10380 mname, member_idx, st_ops->name);
10384 if (st_ops->check_member) {
10385 int err = st_ops->check_member(t, member);
10388 verbose(env, "attach to unsupported member %s of struct %s\n",
10389 mname, st_ops->name);
10394 prog->aux->attach_func_proto = func_proto;
10395 prog->aux->attach_func_name = mname;
10396 env->ops = st_ops->verifier_ops;
10400 #define SECURITY_PREFIX "security_"
10402 static int check_attach_modify_return(struct bpf_verifier_env *env)
10404 struct bpf_prog *prog = env->prog;
10405 unsigned long addr = (unsigned long) prog->aux->trampoline->func.addr;
10407 /* This is expected to be cleaned up in the future with the KRSI effort
10408 * introducing the LSM_HOOK macro for cleaning up lsm_hooks.h.
10410 if (within_error_injection_list(addr) ||
10411 !strncmp(SECURITY_PREFIX, prog->aux->attach_func_name,
10412 sizeof(SECURITY_PREFIX) - 1))
10415 verbose(env, "fmod_ret attach_btf_id %u (%s) is not modifiable\n",
10416 prog->aux->attach_btf_id, prog->aux->attach_func_name);
10421 static int check_attach_btf_id(struct bpf_verifier_env *env)
10423 struct bpf_prog *prog = env->prog;
10424 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
10425 struct bpf_prog *tgt_prog = prog->aux->linked_prog;
10426 u32 btf_id = prog->aux->attach_btf_id;
10427 const char prefix[] = "btf_trace_";
10428 int ret = 0, subprog = -1, i;
10429 struct bpf_trampoline *tr;
10430 const struct btf_type *t;
10431 bool conservative = true;
10437 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
10438 return check_struct_ops_btf_id(env);
10440 if (prog->type != BPF_PROG_TYPE_TRACING &&
10441 prog->type != BPF_PROG_TYPE_LSM &&
10446 verbose(env, "Tracing programs must provide btf_id\n");
10449 btf = bpf_prog_get_target_btf(prog);
10452 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
10455 t = btf_type_by_id(btf, btf_id);
10457 verbose(env, "attach_btf_id %u is invalid\n", btf_id);
10460 tname = btf_name_by_offset(btf, t->name_off);
10462 verbose(env, "attach_btf_id %u doesn't have a name\n", btf_id);
10466 struct bpf_prog_aux *aux = tgt_prog->aux;
10468 for (i = 0; i < aux->func_info_cnt; i++)
10469 if (aux->func_info[i].type_id == btf_id) {
10473 if (subprog == -1) {
10474 verbose(env, "Subprog %s doesn't exist\n", tname);
10477 conservative = aux->func_info_aux[subprog].unreliable;
10478 if (prog_extension) {
10479 if (conservative) {
10481 "Cannot replace static functions\n");
10484 if (!prog->jit_requested) {
10486 "Extension programs should be JITed\n");
10489 env->ops = bpf_verifier_ops[tgt_prog->type];
10491 if (!tgt_prog->jited) {
10492 verbose(env, "Can attach to only JITed progs\n");
10495 if (tgt_prog->type == prog->type) {
10496 /* Cannot fentry/fexit another fentry/fexit program.
10497 * Cannot attach program extension to another extension.
10498 * It's ok to attach fentry/fexit to extension program.
10500 verbose(env, "Cannot recursively attach\n");
10503 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
10505 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
10506 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
10507 /* Program extensions can extend all program types
10508 * except fentry/fexit. The reason is the following.
10509 * The fentry/fexit programs are used for performance
10510 * analysis, stats and can be attached to any program
10511 * type except themselves. When extension program is
10512 * replacing XDP function it is necessary to allow
10513 * performance analysis of all functions. Both original
10514 * XDP program and its program extension. Hence
10515 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
10516 * allowed. If extending of fentry/fexit was allowed it
10517 * would be possible to create long call chain
10518 * fentry->extension->fentry->extension beyond
10519 * reasonable stack size. Hence extending fentry is not
10522 verbose(env, "Cannot extend fentry/fexit\n");
10525 key = ((u64)aux->id) << 32 | btf_id;
10527 if (prog_extension) {
10528 verbose(env, "Cannot replace kernel functions\n");
10534 switch (prog->expected_attach_type) {
10535 case BPF_TRACE_RAW_TP:
10538 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
10541 if (!btf_type_is_typedef(t)) {
10542 verbose(env, "attach_btf_id %u is not a typedef\n",
10546 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
10547 verbose(env, "attach_btf_id %u points to wrong type name %s\n",
10551 tname += sizeof(prefix) - 1;
10552 t = btf_type_by_id(btf, t->type);
10553 if (!btf_type_is_ptr(t))
10554 /* should never happen in valid vmlinux build */
10556 t = btf_type_by_id(btf, t->type);
10557 if (!btf_type_is_func_proto(t))
10558 /* should never happen in valid vmlinux build */
10561 /* remember two read only pointers that are valid for
10562 * the life time of the kernel
10564 prog->aux->attach_func_name = tname;
10565 prog->aux->attach_func_proto = t;
10566 prog->aux->attach_btf_trace = true;
10569 if (!prog_extension)
10572 case BPF_MODIFY_RETURN:
10574 case BPF_TRACE_FENTRY:
10575 case BPF_TRACE_FEXIT:
10576 prog->aux->attach_func_name = tname;
10577 if (prog->type == BPF_PROG_TYPE_LSM) {
10578 ret = bpf_lsm_verify_prog(&env->log, prog);
10583 if (!btf_type_is_func(t)) {
10584 verbose(env, "attach_btf_id %u is not a function\n",
10588 if (prog_extension &&
10589 btf_check_type_match(env, prog, btf, t))
10591 t = btf_type_by_id(btf, t->type);
10592 if (!btf_type_is_func_proto(t))
10594 tr = bpf_trampoline_lookup(key);
10597 /* t is either vmlinux type or another program's type */
10598 prog->aux->attach_func_proto = t;
10599 mutex_lock(&tr->mutex);
10600 if (tr->func.addr) {
10601 prog->aux->trampoline = tr;
10604 if (tgt_prog && conservative) {
10605 prog->aux->attach_func_proto = NULL;
10608 ret = btf_distill_func_proto(&env->log, btf, t,
10609 tname, &tr->func.model);
10614 addr = (long) tgt_prog->bpf_func;
10616 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
10618 addr = kallsyms_lookup_name(tname);
10621 "The address of function %s cannot be found\n",
10627 tr->func.addr = (void *)addr;
10628 prog->aux->trampoline = tr;
10630 if (prog->expected_attach_type == BPF_MODIFY_RETURN)
10631 ret = check_attach_modify_return(env);
10633 mutex_unlock(&tr->mutex);
10635 bpf_trampoline_put(tr);
10640 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr,
10641 union bpf_attr __user *uattr)
10643 u64 start_time = ktime_get_ns();
10644 struct bpf_verifier_env *env;
10645 struct bpf_verifier_log *log;
10646 int i, len, ret = -EINVAL;
10649 /* no program is valid */
10650 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
10653 /* 'struct bpf_verifier_env' can be global, but since it's not small,
10654 * allocate/free it every time bpf_check() is called
10656 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
10661 len = (*prog)->len;
10662 env->insn_aux_data =
10663 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
10665 if (!env->insn_aux_data)
10667 for (i = 0; i < len; i++)
10668 env->insn_aux_data[i].orig_idx = i;
10670 env->ops = bpf_verifier_ops[env->prog->type];
10671 is_priv = capable(CAP_SYS_ADMIN);
10673 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
10674 mutex_lock(&bpf_verifier_lock);
10676 btf_vmlinux = btf_parse_vmlinux();
10677 mutex_unlock(&bpf_verifier_lock);
10680 /* grab the mutex to protect few globals used by verifier */
10682 mutex_lock(&bpf_verifier_lock);
10684 if (attr->log_level || attr->log_buf || attr->log_size) {
10685 /* user requested verbose verifier output
10686 * and supplied buffer to store the verification trace
10688 log->level = attr->log_level;
10689 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
10690 log->len_total = attr->log_size;
10693 /* log attributes have to be sane */
10694 if (log->len_total < 128 || log->len_total > UINT_MAX >> 2 ||
10695 !log->level || !log->ubuf || log->level & ~BPF_LOG_MASK)
10699 if (IS_ERR(btf_vmlinux)) {
10700 /* Either gcc or pahole or kernel are broken. */
10701 verbose(env, "in-kernel BTF is malformed\n");
10702 ret = PTR_ERR(btf_vmlinux);
10703 goto skip_full_check;
10706 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
10707 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
10708 env->strict_alignment = true;
10709 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
10710 env->strict_alignment = false;
10712 env->allow_ptr_leaks = is_priv;
10715 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
10717 ret = replace_map_fd_with_map_ptr(env);
10719 goto skip_full_check;
10721 if (bpf_prog_is_dev_bound(env->prog->aux)) {
10722 ret = bpf_prog_offload_verifier_prep(env->prog);
10724 goto skip_full_check;
10727 env->explored_states = kvcalloc(state_htab_size(env),
10728 sizeof(struct bpf_verifier_state_list *),
10731 if (!env->explored_states)
10732 goto skip_full_check;
10734 ret = check_subprogs(env);
10736 goto skip_full_check;
10738 ret = check_btf_info(env, attr, uattr);
10740 goto skip_full_check;
10742 ret = check_attach_btf_id(env);
10744 goto skip_full_check;
10746 ret = check_cfg(env);
10748 goto skip_full_check;
10750 ret = do_check_subprogs(env);
10751 ret = ret ?: do_check_main(env);
10753 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
10754 ret = bpf_prog_offload_finalize(env);
10757 kvfree(env->explored_states);
10760 ret = check_max_stack_depth(env);
10762 /* instruction rewrites happen after this point */
10765 opt_hard_wire_dead_code_branches(env);
10767 ret = opt_remove_dead_code(env);
10769 ret = opt_remove_nops(env);
10772 sanitize_dead_code(env);
10776 /* program is valid, convert *(u32*)(ctx + off) accesses */
10777 ret = convert_ctx_accesses(env);
10780 ret = fixup_bpf_calls(env);
10782 /* do 32-bit optimization after insn patching has done so those patched
10783 * insns could be handled correctly.
10785 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
10786 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
10787 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
10792 ret = fixup_call_args(env);
10794 env->verification_time = ktime_get_ns() - start_time;
10795 print_verification_stats(env);
10797 if (log->level && bpf_verifier_log_full(log))
10799 if (log->level && !log->ubuf) {
10801 goto err_release_maps;
10804 if (ret == 0 && env->used_map_cnt) {
10805 /* if program passed verifier, update used_maps in bpf_prog_info */
10806 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
10807 sizeof(env->used_maps[0]),
10810 if (!env->prog->aux->used_maps) {
10812 goto err_release_maps;
10815 memcpy(env->prog->aux->used_maps, env->used_maps,
10816 sizeof(env->used_maps[0]) * env->used_map_cnt);
10817 env->prog->aux->used_map_cnt = env->used_map_cnt;
10819 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
10820 * bpf_ld_imm64 instructions
10822 convert_pseudo_ld_imm64(env);
10826 adjust_btf_func(env);
10829 if (!env->prog->aux->used_maps)
10830 /* if we didn't copy map pointers into bpf_prog_info, release
10831 * them now. Otherwise free_used_maps() will release them.
10837 mutex_unlock(&bpf_verifier_lock);
10838 vfree(env->insn_aux_data);