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>
25 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
26 #define BPF_PROG_TYPE(_id, _name) \
27 [_id] = & _name ## _verifier_ops,
28 #define BPF_MAP_TYPE(_id, _ops)
29 #include <linux/bpf_types.h>
34 /* bpf_check() is a static code analyzer that walks eBPF program
35 * instruction by instruction and updates register/stack state.
36 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
38 * The first pass is depth-first-search to check that the program is a DAG.
39 * It rejects the following programs:
40 * - larger than BPF_MAXINSNS insns
41 * - if loop is present (detected via back-edge)
42 * - unreachable insns exist (shouldn't be a forest. program = one function)
43 * - out of bounds or malformed jumps
44 * The second pass is all possible path descent from the 1st insn.
45 * Since it's analyzing all pathes through the program, the length of the
46 * analysis is limited to 64k insn, which may be hit even if total number of
47 * insn is less then 4K, but there are too many branches that change stack/regs.
48 * Number of 'branches to be analyzed' is limited to 1k
50 * On entry to each instruction, each register has a type, and the instruction
51 * changes the types of the registers depending on instruction semantics.
52 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
55 * All registers are 64-bit.
56 * R0 - return register
57 * R1-R5 argument passing registers
58 * R6-R9 callee saved registers
59 * R10 - frame pointer read-only
61 * At the start of BPF program the register R1 contains a pointer to bpf_context
62 * and has type PTR_TO_CTX.
64 * Verifier tracks arithmetic operations on pointers in case:
65 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
66 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
67 * 1st insn copies R10 (which has FRAME_PTR) type into R1
68 * and 2nd arithmetic instruction is pattern matched to recognize
69 * that it wants to construct a pointer to some element within stack.
70 * So after 2nd insn, the register R1 has type PTR_TO_STACK
71 * (and -20 constant is saved for further stack bounds checking).
72 * Meaning that this reg is a pointer to stack plus known immediate constant.
74 * Most of the time the registers have SCALAR_VALUE type, which
75 * means the register has some value, but it's not a valid pointer.
76 * (like pointer plus pointer becomes SCALAR_VALUE type)
78 * When verifier sees load or store instructions the type of base register
79 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
80 * four pointer types recognized by check_mem_access() function.
82 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
83 * and the range of [ptr, ptr + map's value_size) is accessible.
85 * registers used to pass values to function calls are checked against
86 * function argument constraints.
88 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
89 * It means that the register type passed to this function must be
90 * PTR_TO_STACK and it will be used inside the function as
91 * 'pointer to map element key'
93 * For example the argument constraints for bpf_map_lookup_elem():
94 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
95 * .arg1_type = ARG_CONST_MAP_PTR,
96 * .arg2_type = ARG_PTR_TO_MAP_KEY,
98 * ret_type says that this function returns 'pointer to map elem value or null'
99 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
100 * 2nd argument should be a pointer to stack, which will be used inside
101 * the helper function as a pointer to map element key.
103 * On the kernel side the helper function looks like:
104 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
106 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
107 * void *key = (void *) (unsigned long) r2;
110 * here kernel can access 'key' and 'map' pointers safely, knowing that
111 * [key, key + map->key_size) bytes are valid and were initialized on
112 * the stack of eBPF program.
115 * Corresponding eBPF program may look like:
116 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
117 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
118 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
119 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
120 * here verifier looks at prototype of map_lookup_elem() and sees:
121 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
122 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
124 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
125 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
126 * and were initialized prior to this call.
127 * If it's ok, then verifier allows this BPF_CALL insn and looks at
128 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
129 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
130 * returns ether pointer to map value or NULL.
132 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
133 * insn, the register holding that pointer in the true branch changes state to
134 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
135 * branch. See check_cond_jmp_op().
137 * After the call R0 is set to return type of the function and registers R1-R5
138 * are set to NOT_INIT to indicate that they are no longer readable.
140 * The following reference types represent a potential reference to a kernel
141 * resource which, after first being allocated, must be checked and freed by
143 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
145 * When the verifier sees a helper call return a reference type, it allocates a
146 * pointer id for the reference and stores it in the current function state.
147 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
148 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
149 * passes through a NULL-check conditional. For the branch wherein the state is
150 * changed to CONST_IMM, the verifier releases the reference.
152 * For each helper function that allocates a reference, such as
153 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
154 * bpf_sk_release(). When a reference type passes into the release function,
155 * the verifier also releases the reference. If any unchecked or unreleased
156 * reference remains at the end of the program, the verifier rejects it.
159 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
160 struct bpf_verifier_stack_elem {
161 /* verifer state is 'st'
162 * before processing instruction 'insn_idx'
163 * and after processing instruction 'prev_insn_idx'
165 struct bpf_verifier_state st;
168 struct bpf_verifier_stack_elem *next;
171 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
172 #define BPF_COMPLEXITY_LIMIT_STATES 64
174 #define BPF_MAP_PTR_UNPRIV 1UL
175 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
176 POISON_POINTER_DELTA))
177 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
179 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
181 return BPF_MAP_PTR(aux->map_state) == BPF_MAP_PTR_POISON;
184 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
186 return aux->map_state & BPF_MAP_PTR_UNPRIV;
189 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
190 const struct bpf_map *map, bool unpriv)
192 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
193 unpriv |= bpf_map_ptr_unpriv(aux);
194 aux->map_state = (unsigned long)map |
195 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
198 struct bpf_call_arg_meta {
199 struct bpf_map *map_ptr;
204 s64 msize_smax_value;
205 u64 msize_umax_value;
210 static DEFINE_MUTEX(bpf_verifier_lock);
212 static const struct bpf_line_info *
213 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
215 const struct bpf_line_info *linfo;
216 const struct bpf_prog *prog;
220 nr_linfo = prog->aux->nr_linfo;
222 if (!nr_linfo || insn_off >= prog->len)
225 linfo = prog->aux->linfo;
226 for (i = 1; i < nr_linfo; i++)
227 if (insn_off < linfo[i].insn_off)
230 return &linfo[i - 1];
233 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
238 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
240 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
241 "verifier log line truncated - local buffer too short\n");
243 n = min(log->len_total - log->len_used - 1, n);
246 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
252 /* log_level controls verbosity level of eBPF verifier.
253 * bpf_verifier_log_write() is used to dump the verification trace to the log,
254 * so the user can figure out what's wrong with the program
256 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
257 const char *fmt, ...)
261 if (!bpf_verifier_log_needed(&env->log))
265 bpf_verifier_vlog(&env->log, fmt, args);
268 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
270 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
272 struct bpf_verifier_env *env = private_data;
275 if (!bpf_verifier_log_needed(&env->log))
279 bpf_verifier_vlog(&env->log, fmt, args);
283 static const char *ltrim(const char *s)
291 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
293 const char *prefix_fmt, ...)
295 const struct bpf_line_info *linfo;
297 if (!bpf_verifier_log_needed(&env->log))
300 linfo = find_linfo(env, insn_off);
301 if (!linfo || linfo == env->prev_linfo)
307 va_start(args, prefix_fmt);
308 bpf_verifier_vlog(&env->log, prefix_fmt, args);
313 ltrim(btf_name_by_offset(env->prog->aux->btf,
316 env->prev_linfo = linfo;
319 static bool type_is_pkt_pointer(enum bpf_reg_type type)
321 return type == PTR_TO_PACKET ||
322 type == PTR_TO_PACKET_META;
325 static bool type_is_sk_pointer(enum bpf_reg_type type)
327 return type == PTR_TO_SOCKET ||
328 type == PTR_TO_SOCK_COMMON ||
329 type == PTR_TO_TCP_SOCK ||
330 type == PTR_TO_XDP_SOCK;
333 static bool reg_type_may_be_null(enum bpf_reg_type type)
335 return type == PTR_TO_MAP_VALUE_OR_NULL ||
336 type == PTR_TO_SOCKET_OR_NULL ||
337 type == PTR_TO_SOCK_COMMON_OR_NULL ||
338 type == PTR_TO_TCP_SOCK_OR_NULL;
341 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
343 return reg->type == PTR_TO_MAP_VALUE &&
344 map_value_has_spin_lock(reg->map_ptr);
347 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
349 return type == PTR_TO_SOCKET ||
350 type == PTR_TO_SOCKET_OR_NULL ||
351 type == PTR_TO_TCP_SOCK ||
352 type == PTR_TO_TCP_SOCK_OR_NULL;
355 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
357 return type == ARG_PTR_TO_SOCK_COMMON;
360 /* Determine whether the function releases some resources allocated by another
361 * function call. The first reference type argument will be assumed to be
362 * released by release_reference().
364 static bool is_release_function(enum bpf_func_id func_id)
366 return func_id == BPF_FUNC_sk_release;
369 static bool is_acquire_function(enum bpf_func_id func_id)
371 return func_id == BPF_FUNC_sk_lookup_tcp ||
372 func_id == BPF_FUNC_sk_lookup_udp ||
373 func_id == BPF_FUNC_skc_lookup_tcp;
376 static bool is_ptr_cast_function(enum bpf_func_id func_id)
378 return func_id == BPF_FUNC_tcp_sock ||
379 func_id == BPF_FUNC_sk_fullsock;
382 /* string representation of 'enum bpf_reg_type' */
383 static const char * const reg_type_str[] = {
385 [SCALAR_VALUE] = "inv",
386 [PTR_TO_CTX] = "ctx",
387 [CONST_PTR_TO_MAP] = "map_ptr",
388 [PTR_TO_MAP_VALUE] = "map_value",
389 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
390 [PTR_TO_STACK] = "fp",
391 [PTR_TO_PACKET] = "pkt",
392 [PTR_TO_PACKET_META] = "pkt_meta",
393 [PTR_TO_PACKET_END] = "pkt_end",
394 [PTR_TO_FLOW_KEYS] = "flow_keys",
395 [PTR_TO_SOCKET] = "sock",
396 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
397 [PTR_TO_SOCK_COMMON] = "sock_common",
398 [PTR_TO_SOCK_COMMON_OR_NULL] = "sock_common_or_null",
399 [PTR_TO_TCP_SOCK] = "tcp_sock",
400 [PTR_TO_TCP_SOCK_OR_NULL] = "tcp_sock_or_null",
401 [PTR_TO_TP_BUFFER] = "tp_buffer",
402 [PTR_TO_XDP_SOCK] = "xdp_sock",
405 static char slot_type_char[] = {
406 [STACK_INVALID] = '?',
412 static void print_liveness(struct bpf_verifier_env *env,
413 enum bpf_reg_liveness live)
415 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
417 if (live & REG_LIVE_READ)
419 if (live & REG_LIVE_WRITTEN)
421 if (live & REG_LIVE_DONE)
425 static struct bpf_func_state *func(struct bpf_verifier_env *env,
426 const struct bpf_reg_state *reg)
428 struct bpf_verifier_state *cur = env->cur_state;
430 return cur->frame[reg->frameno];
433 static void print_verifier_state(struct bpf_verifier_env *env,
434 const struct bpf_func_state *state)
436 const struct bpf_reg_state *reg;
441 verbose(env, " frame%d:", state->frameno);
442 for (i = 0; i < MAX_BPF_REG; i++) {
443 reg = &state->regs[i];
447 verbose(env, " R%d", i);
448 print_liveness(env, reg->live);
449 verbose(env, "=%s", reg_type_str[t]);
450 if (t == SCALAR_VALUE && reg->precise)
452 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
453 tnum_is_const(reg->var_off)) {
454 /* reg->off should be 0 for SCALAR_VALUE */
455 verbose(env, "%lld", reg->var_off.value + reg->off);
457 verbose(env, "(id=%d", reg->id);
458 if (reg_type_may_be_refcounted_or_null(t))
459 verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
460 if (t != SCALAR_VALUE)
461 verbose(env, ",off=%d", reg->off);
462 if (type_is_pkt_pointer(t))
463 verbose(env, ",r=%d", reg->range);
464 else if (t == CONST_PTR_TO_MAP ||
465 t == PTR_TO_MAP_VALUE ||
466 t == PTR_TO_MAP_VALUE_OR_NULL)
467 verbose(env, ",ks=%d,vs=%d",
468 reg->map_ptr->key_size,
469 reg->map_ptr->value_size);
470 if (tnum_is_const(reg->var_off)) {
471 /* Typically an immediate SCALAR_VALUE, but
472 * could be a pointer whose offset is too big
475 verbose(env, ",imm=%llx", reg->var_off.value);
477 if (reg->smin_value != reg->umin_value &&
478 reg->smin_value != S64_MIN)
479 verbose(env, ",smin_value=%lld",
480 (long long)reg->smin_value);
481 if (reg->smax_value != reg->umax_value &&
482 reg->smax_value != S64_MAX)
483 verbose(env, ",smax_value=%lld",
484 (long long)reg->smax_value);
485 if (reg->umin_value != 0)
486 verbose(env, ",umin_value=%llu",
487 (unsigned long long)reg->umin_value);
488 if (reg->umax_value != U64_MAX)
489 verbose(env, ",umax_value=%llu",
490 (unsigned long long)reg->umax_value);
491 if (!tnum_is_unknown(reg->var_off)) {
494 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
495 verbose(env, ",var_off=%s", tn_buf);
501 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
502 char types_buf[BPF_REG_SIZE + 1];
506 for (j = 0; j < BPF_REG_SIZE; j++) {
507 if (state->stack[i].slot_type[j] != STACK_INVALID)
509 types_buf[j] = slot_type_char[
510 state->stack[i].slot_type[j]];
512 types_buf[BPF_REG_SIZE] = 0;
515 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
516 print_liveness(env, state->stack[i].spilled_ptr.live);
517 if (state->stack[i].slot_type[0] == STACK_SPILL) {
518 reg = &state->stack[i].spilled_ptr;
520 verbose(env, "=%s", reg_type_str[t]);
521 if (t == SCALAR_VALUE && reg->precise)
523 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
524 verbose(env, "%lld", reg->var_off.value + reg->off);
526 verbose(env, "=%s", types_buf);
529 if (state->acquired_refs && state->refs[0].id) {
530 verbose(env, " refs=%d", state->refs[0].id);
531 for (i = 1; i < state->acquired_refs; i++)
532 if (state->refs[i].id)
533 verbose(env, ",%d", state->refs[i].id);
538 #define COPY_STATE_FN(NAME, COUNT, FIELD, SIZE) \
539 static int copy_##NAME##_state(struct bpf_func_state *dst, \
540 const struct bpf_func_state *src) \
544 if (WARN_ON_ONCE(dst->COUNT < src->COUNT)) { \
545 /* internal bug, make state invalid to reject the program */ \
546 memset(dst, 0, sizeof(*dst)); \
549 memcpy(dst->FIELD, src->FIELD, \
550 sizeof(*src->FIELD) * (src->COUNT / SIZE)); \
553 /* copy_reference_state() */
554 COPY_STATE_FN(reference, acquired_refs, refs, 1)
555 /* copy_stack_state() */
556 COPY_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
559 #define REALLOC_STATE_FN(NAME, COUNT, FIELD, SIZE) \
560 static int realloc_##NAME##_state(struct bpf_func_state *state, int size, \
563 u32 old_size = state->COUNT; \
564 struct bpf_##NAME##_state *new_##FIELD; \
565 int slot = size / SIZE; \
567 if (size <= old_size || !size) { \
570 state->COUNT = slot * SIZE; \
571 if (!size && old_size) { \
572 kfree(state->FIELD); \
573 state->FIELD = NULL; \
577 new_##FIELD = kmalloc_array(slot, sizeof(struct bpf_##NAME##_state), \
583 memcpy(new_##FIELD, state->FIELD, \
584 sizeof(*new_##FIELD) * (old_size / SIZE)); \
585 memset(new_##FIELD + old_size / SIZE, 0, \
586 sizeof(*new_##FIELD) * (size - old_size) / SIZE); \
588 state->COUNT = slot * SIZE; \
589 kfree(state->FIELD); \
590 state->FIELD = new_##FIELD; \
593 /* realloc_reference_state() */
594 REALLOC_STATE_FN(reference, acquired_refs, refs, 1)
595 /* realloc_stack_state() */
596 REALLOC_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
597 #undef REALLOC_STATE_FN
599 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
600 * make it consume minimal amount of memory. check_stack_write() access from
601 * the program calls into realloc_func_state() to grow the stack size.
602 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
603 * which realloc_stack_state() copies over. It points to previous
604 * bpf_verifier_state which is never reallocated.
606 static int realloc_func_state(struct bpf_func_state *state, int stack_size,
607 int refs_size, bool copy_old)
609 int err = realloc_reference_state(state, refs_size, copy_old);
612 return realloc_stack_state(state, stack_size, copy_old);
615 /* Acquire a pointer id from the env and update the state->refs to include
616 * this new pointer reference.
617 * On success, returns a valid pointer id to associate with the register
618 * On failure, returns a negative errno.
620 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
622 struct bpf_func_state *state = cur_func(env);
623 int new_ofs = state->acquired_refs;
626 err = realloc_reference_state(state, state->acquired_refs + 1, true);
630 state->refs[new_ofs].id = id;
631 state->refs[new_ofs].insn_idx = insn_idx;
636 /* release function corresponding to acquire_reference_state(). Idempotent. */
637 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
641 last_idx = state->acquired_refs - 1;
642 for (i = 0; i < state->acquired_refs; i++) {
643 if (state->refs[i].id == ptr_id) {
644 if (last_idx && i != last_idx)
645 memcpy(&state->refs[i], &state->refs[last_idx],
646 sizeof(*state->refs));
647 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
648 state->acquired_refs--;
655 static int transfer_reference_state(struct bpf_func_state *dst,
656 struct bpf_func_state *src)
658 int err = realloc_reference_state(dst, src->acquired_refs, false);
661 err = copy_reference_state(dst, src);
667 static void free_func_state(struct bpf_func_state *state)
676 static void clear_jmp_history(struct bpf_verifier_state *state)
678 kfree(state->jmp_history);
679 state->jmp_history = NULL;
680 state->jmp_history_cnt = 0;
683 static void free_verifier_state(struct bpf_verifier_state *state,
688 for (i = 0; i <= state->curframe; i++) {
689 free_func_state(state->frame[i]);
690 state->frame[i] = NULL;
692 clear_jmp_history(state);
697 /* copy verifier state from src to dst growing dst stack space
698 * when necessary to accommodate larger src stack
700 static int copy_func_state(struct bpf_func_state *dst,
701 const struct bpf_func_state *src)
705 err = realloc_func_state(dst, src->allocated_stack, src->acquired_refs,
709 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
710 err = copy_reference_state(dst, src);
713 return copy_stack_state(dst, src);
716 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
717 const struct bpf_verifier_state *src)
719 struct bpf_func_state *dst;
720 u32 jmp_sz = sizeof(struct bpf_idx_pair) * src->jmp_history_cnt;
723 if (dst_state->jmp_history_cnt < src->jmp_history_cnt) {
724 kfree(dst_state->jmp_history);
725 dst_state->jmp_history = kmalloc(jmp_sz, GFP_USER);
726 if (!dst_state->jmp_history)
729 memcpy(dst_state->jmp_history, src->jmp_history, jmp_sz);
730 dst_state->jmp_history_cnt = src->jmp_history_cnt;
732 /* if dst has more stack frames then src frame, free them */
733 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
734 free_func_state(dst_state->frame[i]);
735 dst_state->frame[i] = NULL;
737 dst_state->speculative = src->speculative;
738 dst_state->curframe = src->curframe;
739 dst_state->active_spin_lock = src->active_spin_lock;
740 dst_state->branches = src->branches;
741 dst_state->parent = src->parent;
742 dst_state->first_insn_idx = src->first_insn_idx;
743 dst_state->last_insn_idx = src->last_insn_idx;
744 for (i = 0; i <= src->curframe; i++) {
745 dst = dst_state->frame[i];
747 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
750 dst_state->frame[i] = dst;
752 err = copy_func_state(dst, src->frame[i]);
759 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
762 u32 br = --st->branches;
764 /* WARN_ON(br > 1) technically makes sense here,
765 * but see comment in push_stack(), hence:
767 WARN_ONCE((int)br < 0,
768 "BUG update_branch_counts:branches_to_explore=%d\n",
776 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
779 struct bpf_verifier_state *cur = env->cur_state;
780 struct bpf_verifier_stack_elem *elem, *head = env->head;
783 if (env->head == NULL)
787 err = copy_verifier_state(cur, &head->st);
792 *insn_idx = head->insn_idx;
794 *prev_insn_idx = head->prev_insn_idx;
796 free_verifier_state(&head->st, false);
803 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
804 int insn_idx, int prev_insn_idx,
807 struct bpf_verifier_state *cur = env->cur_state;
808 struct bpf_verifier_stack_elem *elem;
811 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
815 elem->insn_idx = insn_idx;
816 elem->prev_insn_idx = prev_insn_idx;
817 elem->next = env->head;
820 err = copy_verifier_state(&elem->st, cur);
823 elem->st.speculative |= speculative;
824 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
825 verbose(env, "The sequence of %d jumps is too complex.\n",
829 if (elem->st.parent) {
830 ++elem->st.parent->branches;
831 /* WARN_ON(branches > 2) technically makes sense here,
833 * 1. speculative states will bump 'branches' for non-branch
835 * 2. is_state_visited() heuristics may decide not to create
836 * a new state for a sequence of branches and all such current
837 * and cloned states will be pointing to a single parent state
838 * which might have large 'branches' count.
843 free_verifier_state(env->cur_state, true);
844 env->cur_state = NULL;
845 /* pop all elements and return */
846 while (!pop_stack(env, NULL, NULL));
850 #define CALLER_SAVED_REGS 6
851 static const int caller_saved[CALLER_SAVED_REGS] = {
852 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
855 static void __mark_reg_not_init(struct bpf_reg_state *reg);
857 /* Mark the unknown part of a register (variable offset or scalar value) as
858 * known to have the value @imm.
860 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
862 /* Clear id, off, and union(map_ptr, range) */
863 memset(((u8 *)reg) + sizeof(reg->type), 0,
864 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
865 reg->var_off = tnum_const(imm);
866 reg->smin_value = (s64)imm;
867 reg->smax_value = (s64)imm;
868 reg->umin_value = imm;
869 reg->umax_value = imm;
872 /* Mark the 'variable offset' part of a register as zero. This should be
873 * used only on registers holding a pointer type.
875 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
877 __mark_reg_known(reg, 0);
880 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
882 __mark_reg_known(reg, 0);
883 reg->type = SCALAR_VALUE;
886 static void mark_reg_known_zero(struct bpf_verifier_env *env,
887 struct bpf_reg_state *regs, u32 regno)
889 if (WARN_ON(regno >= MAX_BPF_REG)) {
890 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
891 /* Something bad happened, let's kill all regs */
892 for (regno = 0; regno < MAX_BPF_REG; regno++)
893 __mark_reg_not_init(regs + regno);
896 __mark_reg_known_zero(regs + regno);
899 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
901 return type_is_pkt_pointer(reg->type);
904 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
906 return reg_is_pkt_pointer(reg) ||
907 reg->type == PTR_TO_PACKET_END;
910 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
911 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
912 enum bpf_reg_type which)
914 /* The register can already have a range from prior markings.
915 * This is fine as long as it hasn't been advanced from its
918 return reg->type == which &&
921 tnum_equals_const(reg->var_off, 0);
924 /* Attempts to improve min/max values based on var_off information */
925 static void __update_reg_bounds(struct bpf_reg_state *reg)
927 /* min signed is max(sign bit) | min(other bits) */
928 reg->smin_value = max_t(s64, reg->smin_value,
929 reg->var_off.value | (reg->var_off.mask & S64_MIN));
930 /* max signed is min(sign bit) | max(other bits) */
931 reg->smax_value = min_t(s64, reg->smax_value,
932 reg->var_off.value | (reg->var_off.mask & S64_MAX));
933 reg->umin_value = max(reg->umin_value, reg->var_off.value);
934 reg->umax_value = min(reg->umax_value,
935 reg->var_off.value | reg->var_off.mask);
938 /* Uses signed min/max values to inform unsigned, and vice-versa */
939 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
941 /* Learn sign from signed bounds.
942 * If we cannot cross the sign boundary, then signed and unsigned bounds
943 * are the same, so combine. This works even in the negative case, e.g.
944 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
946 if (reg->smin_value >= 0 || reg->smax_value < 0) {
947 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
949 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
953 /* Learn sign from unsigned bounds. Signed bounds cross the sign
954 * boundary, so we must be careful.
956 if ((s64)reg->umax_value >= 0) {
957 /* Positive. We can't learn anything from the smin, but smax
958 * is positive, hence safe.
960 reg->smin_value = reg->umin_value;
961 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
963 } else if ((s64)reg->umin_value < 0) {
964 /* Negative. We can't learn anything from the smax, but smin
965 * is negative, hence safe.
967 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
969 reg->smax_value = reg->umax_value;
973 /* Attempts to improve var_off based on unsigned min/max information */
974 static void __reg_bound_offset(struct bpf_reg_state *reg)
976 reg->var_off = tnum_intersect(reg->var_off,
977 tnum_range(reg->umin_value,
981 /* Reset the min/max bounds of a register */
982 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
984 reg->smin_value = S64_MIN;
985 reg->smax_value = S64_MAX;
987 reg->umax_value = U64_MAX;
990 /* Mark a register as having a completely unknown (scalar) value. */
991 static void __mark_reg_unknown(struct bpf_reg_state *reg)
994 * Clear type, id, off, and union(map_ptr, range) and
995 * padding between 'type' and union
997 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
998 reg->type = SCALAR_VALUE;
999 reg->var_off = tnum_unknown;
1001 __mark_reg_unbounded(reg);
1004 static void mark_reg_unknown(struct bpf_verifier_env *env,
1005 struct bpf_reg_state *regs, u32 regno)
1007 if (WARN_ON(regno >= MAX_BPF_REG)) {
1008 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1009 /* Something bad happened, let's kill all regs except FP */
1010 for (regno = 0; regno < BPF_REG_FP; regno++)
1011 __mark_reg_not_init(regs + regno);
1015 __mark_reg_unknown(regs);
1016 /* constant backtracking is enabled for root without bpf2bpf calls */
1017 regs->precise = env->subprog_cnt > 1 || !env->allow_ptr_leaks ?
1021 static void __mark_reg_not_init(struct bpf_reg_state *reg)
1023 __mark_reg_unknown(reg);
1024 reg->type = NOT_INIT;
1027 static void mark_reg_not_init(struct bpf_verifier_env *env,
1028 struct bpf_reg_state *regs, u32 regno)
1030 if (WARN_ON(regno >= MAX_BPF_REG)) {
1031 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1032 /* Something bad happened, let's kill all regs except FP */
1033 for (regno = 0; regno < BPF_REG_FP; regno++)
1034 __mark_reg_not_init(regs + regno);
1037 __mark_reg_not_init(regs + regno);
1040 #define DEF_NOT_SUBREG (0)
1041 static void init_reg_state(struct bpf_verifier_env *env,
1042 struct bpf_func_state *state)
1044 struct bpf_reg_state *regs = state->regs;
1047 for (i = 0; i < MAX_BPF_REG; i++) {
1048 mark_reg_not_init(env, regs, i);
1049 regs[i].live = REG_LIVE_NONE;
1050 regs[i].parent = NULL;
1051 regs[i].subreg_def = DEF_NOT_SUBREG;
1055 regs[BPF_REG_FP].type = PTR_TO_STACK;
1056 mark_reg_known_zero(env, regs, BPF_REG_FP);
1057 regs[BPF_REG_FP].frameno = state->frameno;
1059 /* 1st arg to a function */
1060 regs[BPF_REG_1].type = PTR_TO_CTX;
1061 mark_reg_known_zero(env, regs, BPF_REG_1);
1064 #define BPF_MAIN_FUNC (-1)
1065 static void init_func_state(struct bpf_verifier_env *env,
1066 struct bpf_func_state *state,
1067 int callsite, int frameno, int subprogno)
1069 state->callsite = callsite;
1070 state->frameno = frameno;
1071 state->subprogno = subprogno;
1072 init_reg_state(env, state);
1076 SRC_OP, /* register is used as source operand */
1077 DST_OP, /* register is used as destination operand */
1078 DST_OP_NO_MARK /* same as above, check only, don't mark */
1081 static int cmp_subprogs(const void *a, const void *b)
1083 return ((struct bpf_subprog_info *)a)->start -
1084 ((struct bpf_subprog_info *)b)->start;
1087 static int find_subprog(struct bpf_verifier_env *env, int off)
1089 struct bpf_subprog_info *p;
1091 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1092 sizeof(env->subprog_info[0]), cmp_subprogs);
1095 return p - env->subprog_info;
1099 static int add_subprog(struct bpf_verifier_env *env, int off)
1101 int insn_cnt = env->prog->len;
1104 if (off >= insn_cnt || off < 0) {
1105 verbose(env, "call to invalid destination\n");
1108 ret = find_subprog(env, off);
1111 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1112 verbose(env, "too many subprograms\n");
1115 env->subprog_info[env->subprog_cnt++].start = off;
1116 sort(env->subprog_info, env->subprog_cnt,
1117 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1121 static int check_subprogs(struct bpf_verifier_env *env)
1123 int i, ret, subprog_start, subprog_end, off, cur_subprog = 0;
1124 struct bpf_subprog_info *subprog = env->subprog_info;
1125 struct bpf_insn *insn = env->prog->insnsi;
1126 int insn_cnt = env->prog->len;
1128 /* Add entry function. */
1129 ret = add_subprog(env, 0);
1133 /* determine subprog starts. The end is one before the next starts */
1134 for (i = 0; i < insn_cnt; i++) {
1135 if (insn[i].code != (BPF_JMP | BPF_CALL))
1137 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1139 if (!env->allow_ptr_leaks) {
1140 verbose(env, "function calls to other bpf functions are allowed for root only\n");
1143 ret = add_subprog(env, i + insn[i].imm + 1);
1148 /* Add a fake 'exit' subprog which could simplify subprog iteration
1149 * logic. 'subprog_cnt' should not be increased.
1151 subprog[env->subprog_cnt].start = insn_cnt;
1153 if (env->log.level & BPF_LOG_LEVEL2)
1154 for (i = 0; i < env->subprog_cnt; i++)
1155 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1157 /* now check that all jumps are within the same subprog */
1158 subprog_start = subprog[cur_subprog].start;
1159 subprog_end = subprog[cur_subprog + 1].start;
1160 for (i = 0; i < insn_cnt; i++) {
1161 u8 code = insn[i].code;
1163 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
1165 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1167 off = i + insn[i].off + 1;
1168 if (off < subprog_start || off >= subprog_end) {
1169 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1173 if (i == subprog_end - 1) {
1174 /* to avoid fall-through from one subprog into another
1175 * the last insn of the subprog should be either exit
1176 * or unconditional jump back
1178 if (code != (BPF_JMP | BPF_EXIT) &&
1179 code != (BPF_JMP | BPF_JA)) {
1180 verbose(env, "last insn is not an exit or jmp\n");
1183 subprog_start = subprog_end;
1185 if (cur_subprog < env->subprog_cnt)
1186 subprog_end = subprog[cur_subprog + 1].start;
1192 /* Parentage chain of this register (or stack slot) should take care of all
1193 * issues like callee-saved registers, stack slot allocation time, etc.
1195 static int mark_reg_read(struct bpf_verifier_env *env,
1196 const struct bpf_reg_state *state,
1197 struct bpf_reg_state *parent, u8 flag)
1199 bool writes = parent == state->parent; /* Observe write marks */
1203 /* if read wasn't screened by an earlier write ... */
1204 if (writes && state->live & REG_LIVE_WRITTEN)
1206 if (parent->live & REG_LIVE_DONE) {
1207 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1208 reg_type_str[parent->type],
1209 parent->var_off.value, parent->off);
1212 /* The first condition is more likely to be true than the
1213 * second, checked it first.
1215 if ((parent->live & REG_LIVE_READ) == flag ||
1216 parent->live & REG_LIVE_READ64)
1217 /* The parentage chain never changes and
1218 * this parent was already marked as LIVE_READ.
1219 * There is no need to keep walking the chain again and
1220 * keep re-marking all parents as LIVE_READ.
1221 * This case happens when the same register is read
1222 * multiple times without writes into it in-between.
1223 * Also, if parent has the stronger REG_LIVE_READ64 set,
1224 * then no need to set the weak REG_LIVE_READ32.
1227 /* ... then we depend on parent's value */
1228 parent->live |= flag;
1229 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
1230 if (flag == REG_LIVE_READ64)
1231 parent->live &= ~REG_LIVE_READ32;
1233 parent = state->parent;
1238 if (env->longest_mark_read_walk < cnt)
1239 env->longest_mark_read_walk = cnt;
1243 /* This function is supposed to be used by the following 32-bit optimization
1244 * code only. It returns TRUE if the source or destination register operates
1245 * on 64-bit, otherwise return FALSE.
1247 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
1248 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
1253 class = BPF_CLASS(code);
1255 if (class == BPF_JMP) {
1256 /* BPF_EXIT for "main" will reach here. Return TRUE
1261 if (op == BPF_CALL) {
1262 /* BPF to BPF call will reach here because of marking
1263 * caller saved clobber with DST_OP_NO_MARK for which we
1264 * don't care the register def because they are anyway
1265 * marked as NOT_INIT already.
1267 if (insn->src_reg == BPF_PSEUDO_CALL)
1269 /* Helper call will reach here because of arg type
1270 * check, conservatively return TRUE.
1279 if (class == BPF_ALU64 || class == BPF_JMP ||
1280 /* BPF_END always use BPF_ALU class. */
1281 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
1284 if (class == BPF_ALU || class == BPF_JMP32)
1287 if (class == BPF_LDX) {
1289 return BPF_SIZE(code) == BPF_DW;
1290 /* LDX source must be ptr. */
1294 if (class == BPF_STX) {
1295 if (reg->type != SCALAR_VALUE)
1297 return BPF_SIZE(code) == BPF_DW;
1300 if (class == BPF_LD) {
1301 u8 mode = BPF_MODE(code);
1304 if (mode == BPF_IMM)
1307 /* Both LD_IND and LD_ABS return 32-bit data. */
1311 /* Implicit ctx ptr. */
1312 if (regno == BPF_REG_6)
1315 /* Explicit source could be any width. */
1319 if (class == BPF_ST)
1320 /* The only source register for BPF_ST is a ptr. */
1323 /* Conservatively return true at default. */
1327 /* Return TRUE if INSN doesn't have explicit value define. */
1328 static bool insn_no_def(struct bpf_insn *insn)
1330 u8 class = BPF_CLASS(insn->code);
1332 return (class == BPF_JMP || class == BPF_JMP32 ||
1333 class == BPF_STX || class == BPF_ST);
1336 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
1337 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
1339 if (insn_no_def(insn))
1342 return !is_reg64(env, insn, insn->dst_reg, NULL, DST_OP);
1345 static void mark_insn_zext(struct bpf_verifier_env *env,
1346 struct bpf_reg_state *reg)
1348 s32 def_idx = reg->subreg_def;
1350 if (def_idx == DEF_NOT_SUBREG)
1353 env->insn_aux_data[def_idx - 1].zext_dst = true;
1354 /* The dst will be zero extended, so won't be sub-register anymore. */
1355 reg->subreg_def = DEF_NOT_SUBREG;
1358 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
1359 enum reg_arg_type t)
1361 struct bpf_verifier_state *vstate = env->cur_state;
1362 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1363 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
1364 struct bpf_reg_state *reg, *regs = state->regs;
1367 if (regno >= MAX_BPF_REG) {
1368 verbose(env, "R%d is invalid\n", regno);
1373 rw64 = is_reg64(env, insn, regno, reg, t);
1375 /* check whether register used as source operand can be read */
1376 if (reg->type == NOT_INIT) {
1377 verbose(env, "R%d !read_ok\n", regno);
1380 /* We don't need to worry about FP liveness because it's read-only */
1381 if (regno == BPF_REG_FP)
1385 mark_insn_zext(env, reg);
1387 return mark_reg_read(env, reg, reg->parent,
1388 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
1390 /* check whether register used as dest operand can be written to */
1391 if (regno == BPF_REG_FP) {
1392 verbose(env, "frame pointer is read only\n");
1395 reg->live |= REG_LIVE_WRITTEN;
1396 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
1398 mark_reg_unknown(env, regs, regno);
1403 /* for any branch, call, exit record the history of jmps in the given state */
1404 static int push_jmp_history(struct bpf_verifier_env *env,
1405 struct bpf_verifier_state *cur)
1407 u32 cnt = cur->jmp_history_cnt;
1408 struct bpf_idx_pair *p;
1411 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
1414 p[cnt - 1].idx = env->insn_idx;
1415 p[cnt - 1].prev_idx = env->prev_insn_idx;
1416 cur->jmp_history = p;
1417 cur->jmp_history_cnt = cnt;
1421 /* Backtrack one insn at a time. If idx is not at the top of recorded
1422 * history then previous instruction came from straight line execution.
1424 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
1429 if (cnt && st->jmp_history[cnt - 1].idx == i) {
1430 i = st->jmp_history[cnt - 1].prev_idx;
1438 /* For given verifier state backtrack_insn() is called from the last insn to
1439 * the first insn. Its purpose is to compute a bitmask of registers and
1440 * stack slots that needs precision in the parent verifier state.
1442 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
1443 u32 *reg_mask, u64 *stack_mask)
1445 const struct bpf_insn_cbs cbs = {
1446 .cb_print = verbose,
1447 .private_data = env,
1449 struct bpf_insn *insn = env->prog->insnsi + idx;
1450 u8 class = BPF_CLASS(insn->code);
1451 u8 opcode = BPF_OP(insn->code);
1452 u8 mode = BPF_MODE(insn->code);
1453 u32 dreg = 1u << insn->dst_reg;
1454 u32 sreg = 1u << insn->src_reg;
1457 if (insn->code == 0)
1459 if (env->log.level & BPF_LOG_LEVEL) {
1460 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
1461 verbose(env, "%d: ", idx);
1462 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
1465 if (class == BPF_ALU || class == BPF_ALU64) {
1466 if (!(*reg_mask & dreg))
1468 if (opcode == BPF_MOV) {
1469 if (BPF_SRC(insn->code) == BPF_X) {
1471 * dreg needs precision after this insn
1472 * sreg needs precision before this insn
1478 * dreg needs precision after this insn.
1479 * Corresponding register is already marked
1480 * as precise=true in this verifier state.
1481 * No further markings in parent are necessary
1486 if (BPF_SRC(insn->code) == BPF_X) {
1488 * both dreg and sreg need precision
1493 * dreg still needs precision before this insn
1496 } else if (class == BPF_LDX) {
1497 if (!(*reg_mask & dreg))
1501 /* scalars can only be spilled into stack w/o losing precision.
1502 * Load from any other memory can be zero extended.
1503 * The desire to keep that precision is already indicated
1504 * by 'precise' mark in corresponding register of this state.
1505 * No further tracking necessary.
1507 if (insn->src_reg != BPF_REG_FP)
1509 if (BPF_SIZE(insn->code) != BPF_DW)
1512 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
1513 * that [fp - off] slot contains scalar that needs to be
1514 * tracked with precision
1516 spi = (-insn->off - 1) / BPF_REG_SIZE;
1518 verbose(env, "BUG spi %d\n", spi);
1519 WARN_ONCE(1, "verifier backtracking bug");
1522 *stack_mask |= 1ull << spi;
1523 } else if (class == BPF_STX || class == BPF_ST) {
1524 if (*reg_mask & dreg)
1525 /* stx & st shouldn't be using _scalar_ dst_reg
1526 * to access memory. It means backtracking
1527 * encountered a case of pointer subtraction.
1530 /* scalars can only be spilled into stack */
1531 if (insn->dst_reg != BPF_REG_FP)
1533 if (BPF_SIZE(insn->code) != BPF_DW)
1535 spi = (-insn->off - 1) / BPF_REG_SIZE;
1537 verbose(env, "BUG spi %d\n", spi);
1538 WARN_ONCE(1, "verifier backtracking bug");
1541 if (!(*stack_mask & (1ull << spi)))
1543 *stack_mask &= ~(1ull << spi);
1544 if (class == BPF_STX)
1546 } else if (class == BPF_JMP || class == BPF_JMP32) {
1547 if (opcode == BPF_CALL) {
1548 if (insn->src_reg == BPF_PSEUDO_CALL)
1550 /* regular helper call sets R0 */
1552 if (*reg_mask & 0x3f) {
1553 /* if backtracing was looking for registers R1-R5
1554 * they should have been found already.
1556 verbose(env, "BUG regs %x\n", *reg_mask);
1557 WARN_ONCE(1, "verifier backtracking bug");
1560 } else if (opcode == BPF_EXIT) {
1563 } else if (class == BPF_LD) {
1564 if (!(*reg_mask & dreg))
1567 /* It's ld_imm64 or ld_abs or ld_ind.
1568 * For ld_imm64 no further tracking of precision
1569 * into parent is necessary
1571 if (mode == BPF_IND || mode == BPF_ABS)
1572 /* to be analyzed */
1578 /* the scalar precision tracking algorithm:
1579 * . at the start all registers have precise=false.
1580 * . scalar ranges are tracked as normal through alu and jmp insns.
1581 * . once precise value of the scalar register is used in:
1582 * . ptr + scalar alu
1583 * . if (scalar cond K|scalar)
1584 * . helper_call(.., scalar, ...) where ARG_CONST is expected
1585 * backtrack through the verifier states and mark all registers and
1586 * stack slots with spilled constants that these scalar regisers
1587 * should be precise.
1588 * . during state pruning two registers (or spilled stack slots)
1589 * are equivalent if both are not precise.
1591 * Note the verifier cannot simply walk register parentage chain,
1592 * since many different registers and stack slots could have been
1593 * used to compute single precise scalar.
1595 * The approach of starting with precise=true for all registers and then
1596 * backtrack to mark a register as not precise when the verifier detects
1597 * that program doesn't care about specific value (e.g., when helper
1598 * takes register as ARG_ANYTHING parameter) is not safe.
1600 * It's ok to walk single parentage chain of the verifier states.
1601 * It's possible that this backtracking will go all the way till 1st insn.
1602 * All other branches will be explored for needing precision later.
1604 * The backtracking needs to deal with cases like:
1605 * 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)
1608 * if r5 > 0x79f goto pc+7
1609 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
1612 * call bpf_perf_event_output#25
1613 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
1617 * call foo // uses callee's r6 inside to compute r0
1621 * to track above reg_mask/stack_mask needs to be independent for each frame.
1623 * Also if parent's curframe > frame where backtracking started,
1624 * the verifier need to mark registers in both frames, otherwise callees
1625 * may incorrectly prune callers. This is similar to
1626 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
1628 * For now backtracking falls back into conservative marking.
1630 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
1631 struct bpf_verifier_state *st)
1633 struct bpf_func_state *func;
1634 struct bpf_reg_state *reg;
1637 /* big hammer: mark all scalars precise in this path.
1638 * pop_stack may still get !precise scalars.
1640 for (; st; st = st->parent)
1641 for (i = 0; i <= st->curframe; i++) {
1642 func = st->frame[i];
1643 for (j = 0; j < BPF_REG_FP; j++) {
1644 reg = &func->regs[j];
1645 if (reg->type != SCALAR_VALUE)
1647 reg->precise = true;
1649 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
1650 if (func->stack[j].slot_type[0] != STACK_SPILL)
1652 reg = &func->stack[j].spilled_ptr;
1653 if (reg->type != SCALAR_VALUE)
1655 reg->precise = true;
1660 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
1663 struct bpf_verifier_state *st = env->cur_state;
1664 int first_idx = st->first_insn_idx;
1665 int last_idx = env->insn_idx;
1666 struct bpf_func_state *func;
1667 struct bpf_reg_state *reg;
1668 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
1669 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
1670 bool skip_first = true;
1671 bool new_marks = false;
1674 if (!env->allow_ptr_leaks)
1675 /* backtracking is root only for now */
1678 func = st->frame[st->curframe];
1680 reg = &func->regs[regno];
1681 if (reg->type != SCALAR_VALUE) {
1682 WARN_ONCE(1, "backtracing misuse");
1689 reg->precise = true;
1693 if (func->stack[spi].slot_type[0] != STACK_SPILL) {
1697 reg = &func->stack[spi].spilled_ptr;
1698 if (reg->type != SCALAR_VALUE) {
1706 reg->precise = true;
1712 if (!reg_mask && !stack_mask)
1715 DECLARE_BITMAP(mask, 64);
1716 u32 history = st->jmp_history_cnt;
1718 if (env->log.level & BPF_LOG_LEVEL)
1719 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
1720 for (i = last_idx;;) {
1725 err = backtrack_insn(env, i, ®_mask, &stack_mask);
1727 if (err == -ENOTSUPP) {
1728 mark_all_scalars_precise(env, st);
1733 if (!reg_mask && !stack_mask)
1734 /* Found assignment(s) into tracked register in this state.
1735 * Since this state is already marked, just return.
1736 * Nothing to be tracked further in the parent state.
1741 i = get_prev_insn_idx(st, i, &history);
1742 if (i >= env->prog->len) {
1743 /* This can happen if backtracking reached insn 0
1744 * and there are still reg_mask or stack_mask
1746 * It means the backtracking missed the spot where
1747 * particular register was initialized with a constant.
1749 verbose(env, "BUG backtracking idx %d\n", i);
1750 WARN_ONCE(1, "verifier backtracking bug");
1759 func = st->frame[st->curframe];
1760 bitmap_from_u64(mask, reg_mask);
1761 for_each_set_bit(i, mask, 32) {
1762 reg = &func->regs[i];
1763 if (reg->type != SCALAR_VALUE) {
1764 reg_mask &= ~(1u << i);
1769 reg->precise = true;
1772 bitmap_from_u64(mask, stack_mask);
1773 for_each_set_bit(i, mask, 64) {
1774 if (i >= func->allocated_stack / BPF_REG_SIZE) {
1775 /* the sequence of instructions:
1777 * 3: (7b) *(u64 *)(r3 -8) = r0
1778 * 4: (79) r4 = *(u64 *)(r10 -8)
1779 * doesn't contain jmps. It's backtracked
1780 * as a single block.
1781 * During backtracking insn 3 is not recognized as
1782 * stack access, so at the end of backtracking
1783 * stack slot fp-8 is still marked in stack_mask.
1784 * However the parent state may not have accessed
1785 * fp-8 and it's "unallocated" stack space.
1786 * In such case fallback to conservative.
1788 mark_all_scalars_precise(env, st);
1792 if (func->stack[i].slot_type[0] != STACK_SPILL) {
1793 stack_mask &= ~(1ull << i);
1796 reg = &func->stack[i].spilled_ptr;
1797 if (reg->type != SCALAR_VALUE) {
1798 stack_mask &= ~(1ull << i);
1803 reg->precise = true;
1805 if (env->log.level & BPF_LOG_LEVEL) {
1806 print_verifier_state(env, func);
1807 verbose(env, "parent %s regs=%x stack=%llx marks\n",
1808 new_marks ? "didn't have" : "already had",
1809 reg_mask, stack_mask);
1812 if (!reg_mask && !stack_mask)
1817 last_idx = st->last_insn_idx;
1818 first_idx = st->first_insn_idx;
1823 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
1825 return __mark_chain_precision(env, regno, -1);
1828 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
1830 return __mark_chain_precision(env, -1, spi);
1833 static bool is_spillable_regtype(enum bpf_reg_type type)
1836 case PTR_TO_MAP_VALUE:
1837 case PTR_TO_MAP_VALUE_OR_NULL:
1841 case PTR_TO_PACKET_META:
1842 case PTR_TO_PACKET_END:
1843 case PTR_TO_FLOW_KEYS:
1844 case CONST_PTR_TO_MAP:
1846 case PTR_TO_SOCKET_OR_NULL:
1847 case PTR_TO_SOCK_COMMON:
1848 case PTR_TO_SOCK_COMMON_OR_NULL:
1849 case PTR_TO_TCP_SOCK:
1850 case PTR_TO_TCP_SOCK_OR_NULL:
1851 case PTR_TO_XDP_SOCK:
1858 /* Does this register contain a constant zero? */
1859 static bool register_is_null(struct bpf_reg_state *reg)
1861 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
1864 static bool register_is_const(struct bpf_reg_state *reg)
1866 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
1869 static void save_register_state(struct bpf_func_state *state,
1870 int spi, struct bpf_reg_state *reg)
1874 state->stack[spi].spilled_ptr = *reg;
1875 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1877 for (i = 0; i < BPF_REG_SIZE; i++)
1878 state->stack[spi].slot_type[i] = STACK_SPILL;
1881 /* check_stack_read/write functions track spill/fill of registers,
1882 * stack boundary and alignment are checked in check_mem_access()
1884 static int check_stack_write(struct bpf_verifier_env *env,
1885 struct bpf_func_state *state, /* func where register points to */
1886 int off, int size, int value_regno, int insn_idx)
1888 struct bpf_func_state *cur; /* state of the current function */
1889 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
1890 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
1891 struct bpf_reg_state *reg = NULL;
1893 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
1894 state->acquired_refs, true);
1897 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
1898 * so it's aligned access and [off, off + size) are within stack limits
1900 if (!env->allow_ptr_leaks &&
1901 state->stack[spi].slot_type[0] == STACK_SPILL &&
1902 size != BPF_REG_SIZE) {
1903 verbose(env, "attempt to corrupt spilled pointer on stack\n");
1907 cur = env->cur_state->frame[env->cur_state->curframe];
1908 if (value_regno >= 0)
1909 reg = &cur->regs[value_regno];
1911 if (reg && size == BPF_REG_SIZE && register_is_const(reg) &&
1912 !register_is_null(reg) && env->allow_ptr_leaks) {
1913 if (dst_reg != BPF_REG_FP) {
1914 /* The backtracking logic can only recognize explicit
1915 * stack slot address like [fp - 8]. Other spill of
1916 * scalar via different register has to be conervative.
1917 * Backtrack from here and mark all registers as precise
1918 * that contributed into 'reg' being a constant.
1920 err = mark_chain_precision(env, value_regno);
1924 save_register_state(state, spi, reg);
1925 } else if (reg && is_spillable_regtype(reg->type)) {
1926 /* register containing pointer is being spilled into stack */
1927 if (size != BPF_REG_SIZE) {
1928 verbose_linfo(env, insn_idx, "; ");
1929 verbose(env, "invalid size of register spill\n");
1933 if (state != cur && reg->type == PTR_TO_STACK) {
1934 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
1938 if (!env->allow_ptr_leaks) {
1939 bool sanitize = false;
1941 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
1942 register_is_const(&state->stack[spi].spilled_ptr))
1944 for (i = 0; i < BPF_REG_SIZE; i++)
1945 if (state->stack[spi].slot_type[i] == STACK_MISC) {
1950 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
1951 int soff = (-spi - 1) * BPF_REG_SIZE;
1953 /* detected reuse of integer stack slot with a pointer
1954 * which means either llvm is reusing stack slot or
1955 * an attacker is trying to exploit CVE-2018-3639
1956 * (speculative store bypass)
1957 * Have to sanitize that slot with preemptive
1960 if (*poff && *poff != soff) {
1961 /* disallow programs where single insn stores
1962 * into two different stack slots, since verifier
1963 * cannot sanitize them
1966 "insn %d cannot access two stack slots fp%d and fp%d",
1967 insn_idx, *poff, soff);
1973 save_register_state(state, spi, reg);
1975 u8 type = STACK_MISC;
1977 /* regular write of data into stack destroys any spilled ptr */
1978 state->stack[spi].spilled_ptr.type = NOT_INIT;
1979 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
1980 if (state->stack[spi].slot_type[0] == STACK_SPILL)
1981 for (i = 0; i < BPF_REG_SIZE; i++)
1982 state->stack[spi].slot_type[i] = STACK_MISC;
1984 /* only mark the slot as written if all 8 bytes were written
1985 * otherwise read propagation may incorrectly stop too soon
1986 * when stack slots are partially written.
1987 * This heuristic means that read propagation will be
1988 * conservative, since it will add reg_live_read marks
1989 * to stack slots all the way to first state when programs
1990 * writes+reads less than 8 bytes
1992 if (size == BPF_REG_SIZE)
1993 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1995 /* when we zero initialize stack slots mark them as such */
1996 if (reg && register_is_null(reg)) {
1997 /* backtracking doesn't work for STACK_ZERO yet. */
1998 err = mark_chain_precision(env, value_regno);
2004 /* Mark slots affected by this stack write. */
2005 for (i = 0; i < size; i++)
2006 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
2012 static int check_stack_read(struct bpf_verifier_env *env,
2013 struct bpf_func_state *reg_state /* func where register points to */,
2014 int off, int size, int value_regno)
2016 struct bpf_verifier_state *vstate = env->cur_state;
2017 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2018 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
2019 struct bpf_reg_state *reg;
2022 if (reg_state->allocated_stack <= slot) {
2023 verbose(env, "invalid read from stack off %d+0 size %d\n",
2027 stype = reg_state->stack[spi].slot_type;
2028 reg = ®_state->stack[spi].spilled_ptr;
2030 if (stype[0] == STACK_SPILL) {
2031 if (size != BPF_REG_SIZE) {
2032 if (reg->type != SCALAR_VALUE) {
2033 verbose_linfo(env, env->insn_idx, "; ");
2034 verbose(env, "invalid size of register fill\n");
2037 if (value_regno >= 0) {
2038 mark_reg_unknown(env, state->regs, value_regno);
2039 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2041 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2044 for (i = 1; i < BPF_REG_SIZE; i++) {
2045 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
2046 verbose(env, "corrupted spill memory\n");
2051 if (value_regno >= 0) {
2052 /* restore register state from stack */
2053 state->regs[value_regno] = *reg;
2054 /* mark reg as written since spilled pointer state likely
2055 * has its liveness marks cleared by is_state_visited()
2056 * which resets stack/reg liveness for state transitions
2058 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2060 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2064 for (i = 0; i < size; i++) {
2065 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
2067 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
2071 verbose(env, "invalid read from stack off %d+%d size %d\n",
2075 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2076 if (value_regno >= 0) {
2077 if (zeros == size) {
2078 /* any size read into register is zero extended,
2079 * so the whole register == const_zero
2081 __mark_reg_const_zero(&state->regs[value_regno]);
2082 /* backtracking doesn't support STACK_ZERO yet,
2083 * so mark it precise here, so that later
2084 * backtracking can stop here.
2085 * Backtracking may not need this if this register
2086 * doesn't participate in pointer adjustment.
2087 * Forward propagation of precise flag is not
2088 * necessary either. This mark is only to stop
2089 * backtracking. Any register that contributed
2090 * to const 0 was marked precise before spill.
2092 state->regs[value_regno].precise = true;
2094 /* have read misc data from the stack */
2095 mark_reg_unknown(env, state->regs, value_regno);
2097 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2103 static int check_stack_access(struct bpf_verifier_env *env,
2104 const struct bpf_reg_state *reg,
2107 /* Stack accesses must be at a fixed offset, so that we
2108 * can determine what type of data were returned. See
2109 * check_stack_read().
2111 if (!tnum_is_const(reg->var_off)) {
2114 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2115 verbose(env, "variable stack access var_off=%s off=%d size=%d\n",
2120 if (off >= 0 || off < -MAX_BPF_STACK) {
2121 verbose(env, "invalid stack off=%d size=%d\n", off, size);
2128 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
2129 int off, int size, enum bpf_access_type type)
2131 struct bpf_reg_state *regs = cur_regs(env);
2132 struct bpf_map *map = regs[regno].map_ptr;
2133 u32 cap = bpf_map_flags_to_cap(map);
2135 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
2136 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
2137 map->value_size, off, size);
2141 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
2142 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
2143 map->value_size, off, size);
2150 /* check read/write into map element returned by bpf_map_lookup_elem() */
2151 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
2152 int size, bool zero_size_allowed)
2154 struct bpf_reg_state *regs = cur_regs(env);
2155 struct bpf_map *map = regs[regno].map_ptr;
2157 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
2158 off + size > map->value_size) {
2159 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
2160 map->value_size, off, size);
2166 /* check read/write into a map element with possible variable offset */
2167 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
2168 int off, int size, bool zero_size_allowed)
2170 struct bpf_verifier_state *vstate = env->cur_state;
2171 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2172 struct bpf_reg_state *reg = &state->regs[regno];
2175 /* We may have adjusted the register to this map value, so we
2176 * need to try adding each of min_value and max_value to off
2177 * to make sure our theoretical access will be safe.
2179 if (env->log.level & BPF_LOG_LEVEL)
2180 print_verifier_state(env, state);
2182 /* The minimum value is only important with signed
2183 * comparisons where we can't assume the floor of a
2184 * value is 0. If we are using signed variables for our
2185 * index'es we need to make sure that whatever we use
2186 * will have a set floor within our range.
2188 if (reg->smin_value < 0 &&
2189 (reg->smin_value == S64_MIN ||
2190 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
2191 reg->smin_value + off < 0)) {
2192 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2196 err = __check_map_access(env, regno, reg->smin_value + off, size,
2199 verbose(env, "R%d min value is outside of the array range\n",
2204 /* If we haven't set a max value then we need to bail since we can't be
2205 * sure we won't do bad things.
2206 * If reg->umax_value + off could overflow, treat that as unbounded too.
2208 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
2209 verbose(env, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
2213 err = __check_map_access(env, regno, reg->umax_value + off, size,
2216 verbose(env, "R%d max value is outside of the array range\n",
2219 if (map_value_has_spin_lock(reg->map_ptr)) {
2220 u32 lock = reg->map_ptr->spin_lock_off;
2222 /* if any part of struct bpf_spin_lock can be touched by
2223 * load/store reject this program.
2224 * To check that [x1, x2) overlaps with [y1, y2)
2225 * it is sufficient to check x1 < y2 && y1 < x2.
2227 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
2228 lock < reg->umax_value + off + size) {
2229 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
2236 #define MAX_PACKET_OFF 0xffff
2238 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
2239 const struct bpf_call_arg_meta *meta,
2240 enum bpf_access_type t)
2242 switch (env->prog->type) {
2243 /* Program types only with direct read access go here! */
2244 case BPF_PROG_TYPE_LWT_IN:
2245 case BPF_PROG_TYPE_LWT_OUT:
2246 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
2247 case BPF_PROG_TYPE_SK_REUSEPORT:
2248 case BPF_PROG_TYPE_FLOW_DISSECTOR:
2249 case BPF_PROG_TYPE_CGROUP_SKB:
2254 /* Program types with direct read + write access go here! */
2255 case BPF_PROG_TYPE_SCHED_CLS:
2256 case BPF_PROG_TYPE_SCHED_ACT:
2257 case BPF_PROG_TYPE_XDP:
2258 case BPF_PROG_TYPE_LWT_XMIT:
2259 case BPF_PROG_TYPE_SK_SKB:
2260 case BPF_PROG_TYPE_SK_MSG:
2262 return meta->pkt_access;
2264 env->seen_direct_write = true;
2267 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
2269 env->seen_direct_write = true;
2278 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
2279 int off, int size, bool zero_size_allowed)
2281 struct bpf_reg_state *regs = cur_regs(env);
2282 struct bpf_reg_state *reg = ®s[regno];
2284 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
2285 (u64)off + size > reg->range) {
2286 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
2287 off, size, regno, reg->id, reg->off, reg->range);
2293 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
2294 int size, bool zero_size_allowed)
2296 struct bpf_reg_state *regs = cur_regs(env);
2297 struct bpf_reg_state *reg = ®s[regno];
2300 /* We may have added a variable offset to the packet pointer; but any
2301 * reg->range we have comes after that. We are only checking the fixed
2305 /* We don't allow negative numbers, because we aren't tracking enough
2306 * detail to prove they're safe.
2308 if (reg->smin_value < 0) {
2309 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2313 err = __check_packet_access(env, regno, off, size, zero_size_allowed);
2315 verbose(env, "R%d offset is outside of the packet\n", regno);
2319 /* __check_packet_access has made sure "off + size - 1" is within u16.
2320 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
2321 * otherwise find_good_pkt_pointers would have refused to set range info
2322 * that __check_packet_access would have rejected this pkt access.
2323 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
2325 env->prog->aux->max_pkt_offset =
2326 max_t(u32, env->prog->aux->max_pkt_offset,
2327 off + reg->umax_value + size - 1);
2332 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
2333 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
2334 enum bpf_access_type t, enum bpf_reg_type *reg_type)
2336 struct bpf_insn_access_aux info = {
2337 .reg_type = *reg_type,
2340 if (env->ops->is_valid_access &&
2341 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
2342 /* A non zero info.ctx_field_size indicates that this field is a
2343 * candidate for later verifier transformation to load the whole
2344 * field and then apply a mask when accessed with a narrower
2345 * access than actual ctx access size. A zero info.ctx_field_size
2346 * will only allow for whole field access and rejects any other
2347 * type of narrower access.
2349 *reg_type = info.reg_type;
2351 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
2352 /* remember the offset of last byte accessed in ctx */
2353 if (env->prog->aux->max_ctx_offset < off + size)
2354 env->prog->aux->max_ctx_offset = off + size;
2358 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
2362 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
2365 if (size < 0 || off < 0 ||
2366 (u64)off + size > sizeof(struct bpf_flow_keys)) {
2367 verbose(env, "invalid access to flow keys off=%d size=%d\n",
2374 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
2375 u32 regno, int off, int size,
2376 enum bpf_access_type t)
2378 struct bpf_reg_state *regs = cur_regs(env);
2379 struct bpf_reg_state *reg = ®s[regno];
2380 struct bpf_insn_access_aux info = {};
2383 if (reg->smin_value < 0) {
2384 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2389 switch (reg->type) {
2390 case PTR_TO_SOCK_COMMON:
2391 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
2394 valid = bpf_sock_is_valid_access(off, size, t, &info);
2396 case PTR_TO_TCP_SOCK:
2397 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
2399 case PTR_TO_XDP_SOCK:
2400 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
2408 env->insn_aux_data[insn_idx].ctx_field_size =
2409 info.ctx_field_size;
2413 verbose(env, "R%d invalid %s access off=%d size=%d\n",
2414 regno, reg_type_str[reg->type], off, size);
2419 static bool __is_pointer_value(bool allow_ptr_leaks,
2420 const struct bpf_reg_state *reg)
2422 if (allow_ptr_leaks)
2425 return reg->type != SCALAR_VALUE;
2428 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
2430 return cur_regs(env) + regno;
2433 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
2435 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
2438 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
2440 const struct bpf_reg_state *reg = reg_state(env, regno);
2442 return reg->type == PTR_TO_CTX;
2445 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
2447 const struct bpf_reg_state *reg = reg_state(env, regno);
2449 return type_is_sk_pointer(reg->type);
2452 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
2454 const struct bpf_reg_state *reg = reg_state(env, regno);
2456 return type_is_pkt_pointer(reg->type);
2459 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
2461 const struct bpf_reg_state *reg = reg_state(env, regno);
2463 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
2464 return reg->type == PTR_TO_FLOW_KEYS;
2467 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
2468 const struct bpf_reg_state *reg,
2469 int off, int size, bool strict)
2471 struct tnum reg_off;
2474 /* Byte size accesses are always allowed. */
2475 if (!strict || size == 1)
2478 /* For platforms that do not have a Kconfig enabling
2479 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
2480 * NET_IP_ALIGN is universally set to '2'. And on platforms
2481 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
2482 * to this code only in strict mode where we want to emulate
2483 * the NET_IP_ALIGN==2 checking. Therefore use an
2484 * unconditional IP align value of '2'.
2488 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
2489 if (!tnum_is_aligned(reg_off, size)) {
2492 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2494 "misaligned packet access off %d+%s+%d+%d size %d\n",
2495 ip_align, tn_buf, reg->off, off, size);
2502 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
2503 const struct bpf_reg_state *reg,
2504 const char *pointer_desc,
2505 int off, int size, bool strict)
2507 struct tnum reg_off;
2509 /* Byte size accesses are always allowed. */
2510 if (!strict || size == 1)
2513 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
2514 if (!tnum_is_aligned(reg_off, size)) {
2517 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2518 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
2519 pointer_desc, tn_buf, reg->off, off, size);
2526 static int check_ptr_alignment(struct bpf_verifier_env *env,
2527 const struct bpf_reg_state *reg, int off,
2528 int size, bool strict_alignment_once)
2530 bool strict = env->strict_alignment || strict_alignment_once;
2531 const char *pointer_desc = "";
2533 switch (reg->type) {
2535 case PTR_TO_PACKET_META:
2536 /* Special case, because of NET_IP_ALIGN. Given metadata sits
2537 * right in front, treat it the very same way.
2539 return check_pkt_ptr_alignment(env, reg, off, size, strict);
2540 case PTR_TO_FLOW_KEYS:
2541 pointer_desc = "flow keys ";
2543 case PTR_TO_MAP_VALUE:
2544 pointer_desc = "value ";
2547 pointer_desc = "context ";
2550 pointer_desc = "stack ";
2551 /* The stack spill tracking logic in check_stack_write()
2552 * and check_stack_read() relies on stack accesses being
2558 pointer_desc = "sock ";
2560 case PTR_TO_SOCK_COMMON:
2561 pointer_desc = "sock_common ";
2563 case PTR_TO_TCP_SOCK:
2564 pointer_desc = "tcp_sock ";
2566 case PTR_TO_XDP_SOCK:
2567 pointer_desc = "xdp_sock ";
2572 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
2576 static int update_stack_depth(struct bpf_verifier_env *env,
2577 const struct bpf_func_state *func,
2580 u16 stack = env->subprog_info[func->subprogno].stack_depth;
2585 /* update known max for given subprogram */
2586 env->subprog_info[func->subprogno].stack_depth = -off;
2590 /* starting from main bpf function walk all instructions of the function
2591 * and recursively walk all callees that given function can call.
2592 * Ignore jump and exit insns.
2593 * Since recursion is prevented by check_cfg() this algorithm
2594 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
2596 static int check_max_stack_depth(struct bpf_verifier_env *env)
2598 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
2599 struct bpf_subprog_info *subprog = env->subprog_info;
2600 struct bpf_insn *insn = env->prog->insnsi;
2601 int ret_insn[MAX_CALL_FRAMES];
2602 int ret_prog[MAX_CALL_FRAMES];
2605 /* round up to 32-bytes, since this is granularity
2606 * of interpreter stack size
2608 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
2609 if (depth > MAX_BPF_STACK) {
2610 verbose(env, "combined stack size of %d calls is %d. Too large\n",
2615 subprog_end = subprog[idx + 1].start;
2616 for (; i < subprog_end; i++) {
2617 if (insn[i].code != (BPF_JMP | BPF_CALL))
2619 if (insn[i].src_reg != BPF_PSEUDO_CALL)
2621 /* remember insn and function to return to */
2622 ret_insn[frame] = i + 1;
2623 ret_prog[frame] = idx;
2625 /* find the callee */
2626 i = i + insn[i].imm + 1;
2627 idx = find_subprog(env, i);
2629 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
2634 if (frame >= MAX_CALL_FRAMES) {
2635 verbose(env, "the call stack of %d frames is too deep !\n",
2641 /* end of for() loop means the last insn of the 'subprog'
2642 * was reached. Doesn't matter whether it was JA or EXIT
2646 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
2648 i = ret_insn[frame];
2649 idx = ret_prog[frame];
2653 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
2654 static int get_callee_stack_depth(struct bpf_verifier_env *env,
2655 const struct bpf_insn *insn, int idx)
2657 int start = idx + insn->imm + 1, subprog;
2659 subprog = find_subprog(env, start);
2661 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
2665 return env->subprog_info[subprog].stack_depth;
2669 static int check_ctx_reg(struct bpf_verifier_env *env,
2670 const struct bpf_reg_state *reg, int regno)
2672 /* Access to ctx or passing it to a helper is only allowed in
2673 * its original, unmodified form.
2677 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
2682 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
2685 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2686 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
2693 static int check_tp_buffer_access(struct bpf_verifier_env *env,
2694 const struct bpf_reg_state *reg,
2695 int regno, int off, int size)
2699 "R%d invalid tracepoint buffer access: off=%d, size=%d",
2703 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
2706 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2708 "R%d invalid variable buffer offset: off=%d, var_off=%s",
2709 regno, off, tn_buf);
2712 if (off + size > env->prog->aux->max_tp_access)
2713 env->prog->aux->max_tp_access = off + size;
2719 /* truncate register to smaller size (in bytes)
2720 * must be called with size < BPF_REG_SIZE
2722 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
2726 /* clear high bits in bit representation */
2727 reg->var_off = tnum_cast(reg->var_off, size);
2729 /* fix arithmetic bounds */
2730 mask = ((u64)1 << (size * 8)) - 1;
2731 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
2732 reg->umin_value &= mask;
2733 reg->umax_value &= mask;
2735 reg->umin_value = 0;
2736 reg->umax_value = mask;
2738 reg->smin_value = reg->umin_value;
2739 reg->smax_value = reg->umax_value;
2742 static bool bpf_map_is_rdonly(const struct bpf_map *map)
2744 return (map->map_flags & BPF_F_RDONLY_PROG) && map->frozen;
2747 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
2753 err = map->ops->map_direct_value_addr(map, &addr, off);
2756 ptr = (void *)(long)addr + off;
2760 *val = (u64)*(u8 *)ptr;
2763 *val = (u64)*(u16 *)ptr;
2766 *val = (u64)*(u32 *)ptr;
2777 /* check whether memory at (regno + off) is accessible for t = (read | write)
2778 * if t==write, value_regno is a register which value is stored into memory
2779 * if t==read, value_regno is a register which will receive the value from memory
2780 * if t==write && value_regno==-1, some unknown value is stored into memory
2781 * if t==read && value_regno==-1, don't care what we read from memory
2783 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
2784 int off, int bpf_size, enum bpf_access_type t,
2785 int value_regno, bool strict_alignment_once)
2787 struct bpf_reg_state *regs = cur_regs(env);
2788 struct bpf_reg_state *reg = regs + regno;
2789 struct bpf_func_state *state;
2792 size = bpf_size_to_bytes(bpf_size);
2796 /* alignment checks will add in reg->off themselves */
2797 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
2801 /* for access checks, reg->off is just part of off */
2804 if (reg->type == PTR_TO_MAP_VALUE) {
2805 if (t == BPF_WRITE && value_regno >= 0 &&
2806 is_pointer_value(env, value_regno)) {
2807 verbose(env, "R%d leaks addr into map\n", value_regno);
2810 err = check_map_access_type(env, regno, off, size, t);
2813 err = check_map_access(env, regno, off, size, false);
2814 if (!err && t == BPF_READ && value_regno >= 0) {
2815 struct bpf_map *map = reg->map_ptr;
2817 /* if map is read-only, track its contents as scalars */
2818 if (tnum_is_const(reg->var_off) &&
2819 bpf_map_is_rdonly(map) &&
2820 map->ops->map_direct_value_addr) {
2821 int map_off = off + reg->var_off.value;
2824 err = bpf_map_direct_read(map, map_off, size,
2829 regs[value_regno].type = SCALAR_VALUE;
2830 __mark_reg_known(®s[value_regno], val);
2832 mark_reg_unknown(env, regs, value_regno);
2835 } else if (reg->type == PTR_TO_CTX) {
2836 enum bpf_reg_type reg_type = SCALAR_VALUE;
2838 if (t == BPF_WRITE && value_regno >= 0 &&
2839 is_pointer_value(env, value_regno)) {
2840 verbose(env, "R%d leaks addr into ctx\n", value_regno);
2844 err = check_ctx_reg(env, reg, regno);
2848 err = check_ctx_access(env, insn_idx, off, size, t, ®_type);
2849 if (!err && t == BPF_READ && value_regno >= 0) {
2850 /* ctx access returns either a scalar, or a
2851 * PTR_TO_PACKET[_META,_END]. In the latter
2852 * case, we know the offset is zero.
2854 if (reg_type == SCALAR_VALUE) {
2855 mark_reg_unknown(env, regs, value_regno);
2857 mark_reg_known_zero(env, regs,
2859 if (reg_type_may_be_null(reg_type))
2860 regs[value_regno].id = ++env->id_gen;
2861 /* A load of ctx field could have different
2862 * actual load size with the one encoded in the
2863 * insn. When the dst is PTR, it is for sure not
2866 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
2868 regs[value_regno].type = reg_type;
2871 } else if (reg->type == PTR_TO_STACK) {
2872 off += reg->var_off.value;
2873 err = check_stack_access(env, reg, off, size);
2877 state = func(env, reg);
2878 err = update_stack_depth(env, state, off);
2883 err = check_stack_write(env, state, off, size,
2884 value_regno, insn_idx);
2886 err = check_stack_read(env, state, off, size,
2888 } else if (reg_is_pkt_pointer(reg)) {
2889 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
2890 verbose(env, "cannot write into packet\n");
2893 if (t == BPF_WRITE && value_regno >= 0 &&
2894 is_pointer_value(env, value_regno)) {
2895 verbose(env, "R%d leaks addr into packet\n",
2899 err = check_packet_access(env, regno, off, size, false);
2900 if (!err && t == BPF_READ && value_regno >= 0)
2901 mark_reg_unknown(env, regs, value_regno);
2902 } else if (reg->type == PTR_TO_FLOW_KEYS) {
2903 if (t == BPF_WRITE && value_regno >= 0 &&
2904 is_pointer_value(env, value_regno)) {
2905 verbose(env, "R%d leaks addr into flow keys\n",
2910 err = check_flow_keys_access(env, off, size);
2911 if (!err && t == BPF_READ && value_regno >= 0)
2912 mark_reg_unknown(env, regs, value_regno);
2913 } else if (type_is_sk_pointer(reg->type)) {
2914 if (t == BPF_WRITE) {
2915 verbose(env, "R%d cannot write into %s\n",
2916 regno, reg_type_str[reg->type]);
2919 err = check_sock_access(env, insn_idx, regno, off, size, t);
2920 if (!err && value_regno >= 0)
2921 mark_reg_unknown(env, regs, value_regno);
2922 } else if (reg->type == PTR_TO_TP_BUFFER) {
2923 err = check_tp_buffer_access(env, reg, regno, off, size);
2924 if (!err && t == BPF_READ && value_regno >= 0)
2925 mark_reg_unknown(env, regs, value_regno);
2927 verbose(env, "R%d invalid mem access '%s'\n", regno,
2928 reg_type_str[reg->type]);
2932 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
2933 regs[value_regno].type == SCALAR_VALUE) {
2934 /* b/h/w load zero-extends, mark upper bits as known 0 */
2935 coerce_reg_to_size(®s[value_regno], size);
2940 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
2944 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
2946 verbose(env, "BPF_XADD uses reserved fields\n");
2950 /* check src1 operand */
2951 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2955 /* check src2 operand */
2956 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2960 if (is_pointer_value(env, insn->src_reg)) {
2961 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
2965 if (is_ctx_reg(env, insn->dst_reg) ||
2966 is_pkt_reg(env, insn->dst_reg) ||
2967 is_flow_key_reg(env, insn->dst_reg) ||
2968 is_sk_reg(env, insn->dst_reg)) {
2969 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
2971 reg_type_str[reg_state(env, insn->dst_reg)->type]);
2975 /* check whether atomic_add can read the memory */
2976 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
2977 BPF_SIZE(insn->code), BPF_READ, -1, true);
2981 /* check whether atomic_add can write into the same memory */
2982 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
2983 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
2986 static int __check_stack_boundary(struct bpf_verifier_env *env, u32 regno,
2987 int off, int access_size,
2988 bool zero_size_allowed)
2990 struct bpf_reg_state *reg = reg_state(env, regno);
2992 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
2993 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
2994 if (tnum_is_const(reg->var_off)) {
2995 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
2996 regno, off, access_size);
3000 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3001 verbose(env, "invalid stack type R%d var_off=%s access_size=%d\n",
3002 regno, tn_buf, access_size);
3009 /* when register 'regno' is passed into function that will read 'access_size'
3010 * bytes from that pointer, make sure that it's within stack boundary
3011 * and all elements of stack are initialized.
3012 * Unlike most pointer bounds-checking functions, this one doesn't take an
3013 * 'off' argument, so it has to add in reg->off itself.
3015 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
3016 int access_size, bool zero_size_allowed,
3017 struct bpf_call_arg_meta *meta)
3019 struct bpf_reg_state *reg = reg_state(env, regno);
3020 struct bpf_func_state *state = func(env, reg);
3021 int err, min_off, max_off, i, j, slot, spi;
3023 if (reg->type != PTR_TO_STACK) {
3024 /* Allow zero-byte read from NULL, regardless of pointer type */
3025 if (zero_size_allowed && access_size == 0 &&
3026 register_is_null(reg))
3029 verbose(env, "R%d type=%s expected=%s\n", regno,
3030 reg_type_str[reg->type],
3031 reg_type_str[PTR_TO_STACK]);
3035 if (tnum_is_const(reg->var_off)) {
3036 min_off = max_off = reg->var_off.value + reg->off;
3037 err = __check_stack_boundary(env, regno, min_off, access_size,
3042 /* Variable offset is prohibited for unprivileged mode for
3043 * simplicity since it requires corresponding support in
3044 * Spectre masking for stack ALU.
3045 * See also retrieve_ptr_limit().
3047 if (!env->allow_ptr_leaks) {
3050 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3051 verbose(env, "R%d indirect variable offset stack access prohibited for !root, var_off=%s\n",
3055 /* Only initialized buffer on stack is allowed to be accessed
3056 * with variable offset. With uninitialized buffer it's hard to
3057 * guarantee that whole memory is marked as initialized on
3058 * helper return since specific bounds are unknown what may
3059 * cause uninitialized stack leaking.
3061 if (meta && meta->raw_mode)
3064 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
3065 reg->smax_value <= -BPF_MAX_VAR_OFF) {
3066 verbose(env, "R%d unbounded indirect variable offset stack access\n",
3070 min_off = reg->smin_value + reg->off;
3071 max_off = reg->smax_value + reg->off;
3072 err = __check_stack_boundary(env, regno, min_off, access_size,
3075 verbose(env, "R%d min value is outside of stack bound\n",
3079 err = __check_stack_boundary(env, regno, max_off, access_size,
3082 verbose(env, "R%d max value is outside of stack bound\n",
3088 if (meta && meta->raw_mode) {
3089 meta->access_size = access_size;
3090 meta->regno = regno;
3094 for (i = min_off; i < max_off + access_size; i++) {
3098 spi = slot / BPF_REG_SIZE;
3099 if (state->allocated_stack <= slot)
3101 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
3102 if (*stype == STACK_MISC)
3104 if (*stype == STACK_ZERO) {
3105 /* helper can write anything into the stack */
3106 *stype = STACK_MISC;
3109 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
3110 state->stack[spi].spilled_ptr.type == SCALAR_VALUE) {
3111 __mark_reg_unknown(&state->stack[spi].spilled_ptr);
3112 for (j = 0; j < BPF_REG_SIZE; j++)
3113 state->stack[spi].slot_type[j] = STACK_MISC;
3118 if (tnum_is_const(reg->var_off)) {
3119 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
3120 min_off, i - min_off, access_size);
3124 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3125 verbose(env, "invalid indirect read from stack var_off %s+%d size %d\n",
3126 tn_buf, i - min_off, access_size);
3130 /* reading any byte out of 8-byte 'spill_slot' will cause
3131 * the whole slot to be marked as 'read'
3133 mark_reg_read(env, &state->stack[spi].spilled_ptr,
3134 state->stack[spi].spilled_ptr.parent,
3137 return update_stack_depth(env, state, min_off);
3140 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
3141 int access_size, bool zero_size_allowed,
3142 struct bpf_call_arg_meta *meta)
3144 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3146 switch (reg->type) {
3148 case PTR_TO_PACKET_META:
3149 return check_packet_access(env, regno, reg->off, access_size,
3151 case PTR_TO_MAP_VALUE:
3152 if (check_map_access_type(env, regno, reg->off, access_size,
3153 meta && meta->raw_mode ? BPF_WRITE :
3156 return check_map_access(env, regno, reg->off, access_size,
3158 default: /* scalar_value|ptr_to_stack or invalid ptr */
3159 return check_stack_boundary(env, regno, access_size,
3160 zero_size_allowed, meta);
3164 /* Implementation details:
3165 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
3166 * Two bpf_map_lookups (even with the same key) will have different reg->id.
3167 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
3168 * value_or_null->value transition, since the verifier only cares about
3169 * the range of access to valid map value pointer and doesn't care about actual
3170 * address of the map element.
3171 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
3172 * reg->id > 0 after value_or_null->value transition. By doing so
3173 * two bpf_map_lookups will be considered two different pointers that
3174 * point to different bpf_spin_locks.
3175 * The verifier allows taking only one bpf_spin_lock at a time to avoid
3177 * Since only one bpf_spin_lock is allowed the checks are simpler than
3178 * reg_is_refcounted() logic. The verifier needs to remember only
3179 * one spin_lock instead of array of acquired_refs.
3180 * cur_state->active_spin_lock remembers which map value element got locked
3181 * and clears it after bpf_spin_unlock.
3183 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
3186 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3187 struct bpf_verifier_state *cur = env->cur_state;
3188 bool is_const = tnum_is_const(reg->var_off);
3189 struct bpf_map *map = reg->map_ptr;
3190 u64 val = reg->var_off.value;
3192 if (reg->type != PTR_TO_MAP_VALUE) {
3193 verbose(env, "R%d is not a pointer to map_value\n", regno);
3198 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
3204 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
3208 if (!map_value_has_spin_lock(map)) {
3209 if (map->spin_lock_off == -E2BIG)
3211 "map '%s' has more than one 'struct bpf_spin_lock'\n",
3213 else if (map->spin_lock_off == -ENOENT)
3215 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
3219 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
3223 if (map->spin_lock_off != val + reg->off) {
3224 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
3229 if (cur->active_spin_lock) {
3231 "Locking two bpf_spin_locks are not allowed\n");
3234 cur->active_spin_lock = reg->id;
3236 if (!cur->active_spin_lock) {
3237 verbose(env, "bpf_spin_unlock without taking a lock\n");
3240 if (cur->active_spin_lock != reg->id) {
3241 verbose(env, "bpf_spin_unlock of different lock\n");
3244 cur->active_spin_lock = 0;
3249 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
3251 return type == ARG_PTR_TO_MEM ||
3252 type == ARG_PTR_TO_MEM_OR_NULL ||
3253 type == ARG_PTR_TO_UNINIT_MEM;
3256 static bool arg_type_is_mem_size(enum bpf_arg_type type)
3258 return type == ARG_CONST_SIZE ||
3259 type == ARG_CONST_SIZE_OR_ZERO;
3262 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
3264 return type == ARG_PTR_TO_INT ||
3265 type == ARG_PTR_TO_LONG;
3268 static int int_ptr_type_to_size(enum bpf_arg_type type)
3270 if (type == ARG_PTR_TO_INT)
3272 else if (type == ARG_PTR_TO_LONG)
3278 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
3279 enum bpf_arg_type arg_type,
3280 struct bpf_call_arg_meta *meta)
3282 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3283 enum bpf_reg_type expected_type, type = reg->type;
3286 if (arg_type == ARG_DONTCARE)
3289 err = check_reg_arg(env, regno, SRC_OP);
3293 if (arg_type == ARG_ANYTHING) {
3294 if (is_pointer_value(env, regno)) {
3295 verbose(env, "R%d leaks addr into helper function\n",
3302 if (type_is_pkt_pointer(type) &&
3303 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
3304 verbose(env, "helper access to the packet is not allowed\n");
3308 if (arg_type == ARG_PTR_TO_MAP_KEY ||
3309 arg_type == ARG_PTR_TO_MAP_VALUE ||
3310 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE ||
3311 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL) {
3312 expected_type = PTR_TO_STACK;
3313 if (register_is_null(reg) &&
3314 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL)
3315 /* final test in check_stack_boundary() */;
3316 else if (!type_is_pkt_pointer(type) &&
3317 type != PTR_TO_MAP_VALUE &&
3318 type != expected_type)
3320 } else if (arg_type == ARG_CONST_SIZE ||
3321 arg_type == ARG_CONST_SIZE_OR_ZERO) {
3322 expected_type = SCALAR_VALUE;
3323 if (type != expected_type)
3325 } else if (arg_type == ARG_CONST_MAP_PTR) {
3326 expected_type = CONST_PTR_TO_MAP;
3327 if (type != expected_type)
3329 } else if (arg_type == ARG_PTR_TO_CTX) {
3330 expected_type = PTR_TO_CTX;
3331 if (type != expected_type)
3333 err = check_ctx_reg(env, reg, regno);
3336 } else if (arg_type == ARG_PTR_TO_SOCK_COMMON) {
3337 expected_type = PTR_TO_SOCK_COMMON;
3338 /* Any sk pointer can be ARG_PTR_TO_SOCK_COMMON */
3339 if (!type_is_sk_pointer(type))
3341 if (reg->ref_obj_id) {
3342 if (meta->ref_obj_id) {
3343 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
3344 regno, reg->ref_obj_id,
3348 meta->ref_obj_id = reg->ref_obj_id;
3350 } else if (arg_type == ARG_PTR_TO_SOCKET) {
3351 expected_type = PTR_TO_SOCKET;
3352 if (type != expected_type)
3354 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
3355 if (meta->func_id == BPF_FUNC_spin_lock) {
3356 if (process_spin_lock(env, regno, true))
3358 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
3359 if (process_spin_lock(env, regno, false))
3362 verbose(env, "verifier internal error\n");
3365 } else if (arg_type_is_mem_ptr(arg_type)) {
3366 expected_type = PTR_TO_STACK;
3367 /* One exception here. In case function allows for NULL to be
3368 * passed in as argument, it's a SCALAR_VALUE type. Final test
3369 * happens during stack boundary checking.
3371 if (register_is_null(reg) &&
3372 arg_type == ARG_PTR_TO_MEM_OR_NULL)
3373 /* final test in check_stack_boundary() */;
3374 else if (!type_is_pkt_pointer(type) &&
3375 type != PTR_TO_MAP_VALUE &&
3376 type != expected_type)
3378 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
3379 } else if (arg_type_is_int_ptr(arg_type)) {
3380 expected_type = PTR_TO_STACK;
3381 if (!type_is_pkt_pointer(type) &&
3382 type != PTR_TO_MAP_VALUE &&
3383 type != expected_type)
3386 verbose(env, "unsupported arg_type %d\n", arg_type);
3390 if (arg_type == ARG_CONST_MAP_PTR) {
3391 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
3392 meta->map_ptr = reg->map_ptr;
3393 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
3394 /* bpf_map_xxx(..., map_ptr, ..., key) call:
3395 * check that [key, key + map->key_size) are within
3396 * stack limits and initialized
3398 if (!meta->map_ptr) {
3399 /* in function declaration map_ptr must come before
3400 * map_key, so that it's verified and known before
3401 * we have to check map_key here. Otherwise it means
3402 * that kernel subsystem misconfigured verifier
3404 verbose(env, "invalid map_ptr to access map->key\n");
3407 err = check_helper_mem_access(env, regno,
3408 meta->map_ptr->key_size, false,
3410 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
3411 (arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL &&
3412 !register_is_null(reg)) ||
3413 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
3414 /* bpf_map_xxx(..., map_ptr, ..., value) call:
3415 * check [value, value + map->value_size) validity
3417 if (!meta->map_ptr) {
3418 /* kernel subsystem misconfigured verifier */
3419 verbose(env, "invalid map_ptr to access map->value\n");
3422 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
3423 err = check_helper_mem_access(env, regno,
3424 meta->map_ptr->value_size, false,
3426 } else if (arg_type_is_mem_size(arg_type)) {
3427 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
3429 /* remember the mem_size which may be used later
3430 * to refine return values.
3432 meta->msize_smax_value = reg->smax_value;
3433 meta->msize_umax_value = reg->umax_value;
3435 /* The register is SCALAR_VALUE; the access check
3436 * happens using its boundaries.
3438 if (!tnum_is_const(reg->var_off))
3439 /* For unprivileged variable accesses, disable raw
3440 * mode so that the program is required to
3441 * initialize all the memory that the helper could
3442 * just partially fill up.
3446 if (reg->smin_value < 0) {
3447 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
3452 if (reg->umin_value == 0) {
3453 err = check_helper_mem_access(env, regno - 1, 0,
3460 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
3461 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
3465 err = check_helper_mem_access(env, regno - 1,
3467 zero_size_allowed, meta);
3469 err = mark_chain_precision(env, regno);
3470 } else if (arg_type_is_int_ptr(arg_type)) {
3471 int size = int_ptr_type_to_size(arg_type);
3473 err = check_helper_mem_access(env, regno, size, false, meta);
3476 err = check_ptr_alignment(env, reg, 0, size, true);
3481 verbose(env, "R%d type=%s expected=%s\n", regno,
3482 reg_type_str[type], reg_type_str[expected_type]);
3486 static int check_map_func_compatibility(struct bpf_verifier_env *env,
3487 struct bpf_map *map, int func_id)
3492 /* We need a two way check, first is from map perspective ... */
3493 switch (map->map_type) {
3494 case BPF_MAP_TYPE_PROG_ARRAY:
3495 if (func_id != BPF_FUNC_tail_call)
3498 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
3499 if (func_id != BPF_FUNC_perf_event_read &&
3500 func_id != BPF_FUNC_perf_event_output &&
3501 func_id != BPF_FUNC_perf_event_read_value)
3504 case BPF_MAP_TYPE_STACK_TRACE:
3505 if (func_id != BPF_FUNC_get_stackid)
3508 case BPF_MAP_TYPE_CGROUP_ARRAY:
3509 if (func_id != BPF_FUNC_skb_under_cgroup &&
3510 func_id != BPF_FUNC_current_task_under_cgroup)
3513 case BPF_MAP_TYPE_CGROUP_STORAGE:
3514 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
3515 if (func_id != BPF_FUNC_get_local_storage)
3518 case BPF_MAP_TYPE_DEVMAP:
3519 case BPF_MAP_TYPE_DEVMAP_HASH:
3520 if (func_id != BPF_FUNC_redirect_map &&
3521 func_id != BPF_FUNC_map_lookup_elem)
3524 /* Restrict bpf side of cpumap and xskmap, open when use-cases
3527 case BPF_MAP_TYPE_CPUMAP:
3528 if (func_id != BPF_FUNC_redirect_map)
3531 case BPF_MAP_TYPE_XSKMAP:
3532 if (func_id != BPF_FUNC_redirect_map &&
3533 func_id != BPF_FUNC_map_lookup_elem)
3536 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
3537 case BPF_MAP_TYPE_HASH_OF_MAPS:
3538 if (func_id != BPF_FUNC_map_lookup_elem)
3541 case BPF_MAP_TYPE_SOCKMAP:
3542 if (func_id != BPF_FUNC_sk_redirect_map &&
3543 func_id != BPF_FUNC_sock_map_update &&
3544 func_id != BPF_FUNC_map_delete_elem &&
3545 func_id != BPF_FUNC_msg_redirect_map)
3548 case BPF_MAP_TYPE_SOCKHASH:
3549 if (func_id != BPF_FUNC_sk_redirect_hash &&
3550 func_id != BPF_FUNC_sock_hash_update &&
3551 func_id != BPF_FUNC_map_delete_elem &&
3552 func_id != BPF_FUNC_msg_redirect_hash)
3555 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
3556 if (func_id != BPF_FUNC_sk_select_reuseport)
3559 case BPF_MAP_TYPE_QUEUE:
3560 case BPF_MAP_TYPE_STACK:
3561 if (func_id != BPF_FUNC_map_peek_elem &&
3562 func_id != BPF_FUNC_map_pop_elem &&
3563 func_id != BPF_FUNC_map_push_elem)
3566 case BPF_MAP_TYPE_SK_STORAGE:
3567 if (func_id != BPF_FUNC_sk_storage_get &&
3568 func_id != BPF_FUNC_sk_storage_delete)
3575 /* ... and second from the function itself. */
3577 case BPF_FUNC_tail_call:
3578 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
3580 if (env->subprog_cnt > 1) {
3581 verbose(env, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
3585 case BPF_FUNC_perf_event_read:
3586 case BPF_FUNC_perf_event_output:
3587 case BPF_FUNC_perf_event_read_value:
3588 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
3591 case BPF_FUNC_get_stackid:
3592 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
3595 case BPF_FUNC_current_task_under_cgroup:
3596 case BPF_FUNC_skb_under_cgroup:
3597 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
3600 case BPF_FUNC_redirect_map:
3601 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
3602 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
3603 map->map_type != BPF_MAP_TYPE_CPUMAP &&
3604 map->map_type != BPF_MAP_TYPE_XSKMAP)
3607 case BPF_FUNC_sk_redirect_map:
3608 case BPF_FUNC_msg_redirect_map:
3609 case BPF_FUNC_sock_map_update:
3610 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
3613 case BPF_FUNC_sk_redirect_hash:
3614 case BPF_FUNC_msg_redirect_hash:
3615 case BPF_FUNC_sock_hash_update:
3616 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
3619 case BPF_FUNC_get_local_storage:
3620 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
3621 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
3624 case BPF_FUNC_sk_select_reuseport:
3625 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY)
3628 case BPF_FUNC_map_peek_elem:
3629 case BPF_FUNC_map_pop_elem:
3630 case BPF_FUNC_map_push_elem:
3631 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
3632 map->map_type != BPF_MAP_TYPE_STACK)
3635 case BPF_FUNC_sk_storage_get:
3636 case BPF_FUNC_sk_storage_delete:
3637 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
3646 verbose(env, "cannot pass map_type %d into func %s#%d\n",
3647 map->map_type, func_id_name(func_id), func_id);
3651 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
3655 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
3657 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
3659 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
3661 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
3663 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
3666 /* We only support one arg being in raw mode at the moment,
3667 * which is sufficient for the helper functions we have
3673 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
3674 enum bpf_arg_type arg_next)
3676 return (arg_type_is_mem_ptr(arg_curr) &&
3677 !arg_type_is_mem_size(arg_next)) ||
3678 (!arg_type_is_mem_ptr(arg_curr) &&
3679 arg_type_is_mem_size(arg_next));
3682 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
3684 /* bpf_xxx(..., buf, len) call will access 'len'
3685 * bytes from memory 'buf'. Both arg types need
3686 * to be paired, so make sure there's no buggy
3687 * helper function specification.
3689 if (arg_type_is_mem_size(fn->arg1_type) ||
3690 arg_type_is_mem_ptr(fn->arg5_type) ||
3691 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
3692 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
3693 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
3694 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
3700 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
3704 if (arg_type_may_be_refcounted(fn->arg1_type))
3706 if (arg_type_may_be_refcounted(fn->arg2_type))
3708 if (arg_type_may_be_refcounted(fn->arg3_type))
3710 if (arg_type_may_be_refcounted(fn->arg4_type))
3712 if (arg_type_may_be_refcounted(fn->arg5_type))
3715 /* A reference acquiring function cannot acquire
3716 * another refcounted ptr.
3718 if (is_acquire_function(func_id) && count)
3721 /* We only support one arg being unreferenced at the moment,
3722 * which is sufficient for the helper functions we have right now.
3727 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
3729 return check_raw_mode_ok(fn) &&
3730 check_arg_pair_ok(fn) &&
3731 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
3734 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
3735 * are now invalid, so turn them into unknown SCALAR_VALUE.
3737 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
3738 struct bpf_func_state *state)
3740 struct bpf_reg_state *regs = state->regs, *reg;
3743 for (i = 0; i < MAX_BPF_REG; i++)
3744 if (reg_is_pkt_pointer_any(®s[i]))
3745 mark_reg_unknown(env, regs, i);
3747 bpf_for_each_spilled_reg(i, state, reg) {
3750 if (reg_is_pkt_pointer_any(reg))
3751 __mark_reg_unknown(reg);
3755 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
3757 struct bpf_verifier_state *vstate = env->cur_state;
3760 for (i = 0; i <= vstate->curframe; i++)
3761 __clear_all_pkt_pointers(env, vstate->frame[i]);
3764 static void release_reg_references(struct bpf_verifier_env *env,
3765 struct bpf_func_state *state,
3768 struct bpf_reg_state *regs = state->regs, *reg;
3771 for (i = 0; i < MAX_BPF_REG; i++)
3772 if (regs[i].ref_obj_id == ref_obj_id)
3773 mark_reg_unknown(env, regs, i);
3775 bpf_for_each_spilled_reg(i, state, reg) {
3778 if (reg->ref_obj_id == ref_obj_id)
3779 __mark_reg_unknown(reg);
3783 /* The pointer with the specified id has released its reference to kernel
3784 * resources. Identify all copies of the same pointer and clear the reference.
3786 static int release_reference(struct bpf_verifier_env *env,
3789 struct bpf_verifier_state *vstate = env->cur_state;
3793 err = release_reference_state(cur_func(env), ref_obj_id);
3797 for (i = 0; i <= vstate->curframe; i++)
3798 release_reg_references(env, vstate->frame[i], ref_obj_id);
3803 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
3806 struct bpf_verifier_state *state = env->cur_state;
3807 struct bpf_func_state *caller, *callee;
3808 int i, err, subprog, target_insn;
3810 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
3811 verbose(env, "the call stack of %d frames is too deep\n",
3812 state->curframe + 2);
3816 target_insn = *insn_idx + insn->imm;
3817 subprog = find_subprog(env, target_insn + 1);
3819 verbose(env, "verifier bug. No program starts at insn %d\n",
3824 caller = state->frame[state->curframe];
3825 if (state->frame[state->curframe + 1]) {
3826 verbose(env, "verifier bug. Frame %d already allocated\n",
3827 state->curframe + 1);
3831 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
3834 state->frame[state->curframe + 1] = callee;
3836 /* callee cannot access r0, r6 - r9 for reading and has to write
3837 * into its own stack before reading from it.
3838 * callee can read/write into caller's stack
3840 init_func_state(env, callee,
3841 /* remember the callsite, it will be used by bpf_exit */
3842 *insn_idx /* callsite */,
3843 state->curframe + 1 /* frameno within this callchain */,
3844 subprog /* subprog number within this prog */);
3846 /* Transfer references to the callee */
3847 err = transfer_reference_state(callee, caller);
3851 /* copy r1 - r5 args that callee can access. The copy includes parent
3852 * pointers, which connects us up to the liveness chain
3854 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
3855 callee->regs[i] = caller->regs[i];
3857 /* after the call registers r0 - r5 were scratched */
3858 for (i = 0; i < CALLER_SAVED_REGS; i++) {
3859 mark_reg_not_init(env, caller->regs, caller_saved[i]);
3860 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
3863 /* only increment it after check_reg_arg() finished */
3866 /* and go analyze first insn of the callee */
3867 *insn_idx = target_insn;
3869 if (env->log.level & BPF_LOG_LEVEL) {
3870 verbose(env, "caller:\n");
3871 print_verifier_state(env, caller);
3872 verbose(env, "callee:\n");
3873 print_verifier_state(env, callee);
3878 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
3880 struct bpf_verifier_state *state = env->cur_state;
3881 struct bpf_func_state *caller, *callee;
3882 struct bpf_reg_state *r0;
3885 callee = state->frame[state->curframe];
3886 r0 = &callee->regs[BPF_REG_0];
3887 if (r0->type == PTR_TO_STACK) {
3888 /* technically it's ok to return caller's stack pointer
3889 * (or caller's caller's pointer) back to the caller,
3890 * since these pointers are valid. Only current stack
3891 * pointer will be invalid as soon as function exits,
3892 * but let's be conservative
3894 verbose(env, "cannot return stack pointer to the caller\n");
3899 caller = state->frame[state->curframe];
3900 /* return to the caller whatever r0 had in the callee */
3901 caller->regs[BPF_REG_0] = *r0;
3903 /* Transfer references to the caller */
3904 err = transfer_reference_state(caller, callee);
3908 *insn_idx = callee->callsite + 1;
3909 if (env->log.level & BPF_LOG_LEVEL) {
3910 verbose(env, "returning from callee:\n");
3911 print_verifier_state(env, callee);
3912 verbose(env, "to caller at %d:\n", *insn_idx);
3913 print_verifier_state(env, caller);
3915 /* clear everything in the callee */
3916 free_func_state(callee);
3917 state->frame[state->curframe + 1] = NULL;
3921 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
3923 struct bpf_call_arg_meta *meta)
3925 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
3927 if (ret_type != RET_INTEGER ||
3928 (func_id != BPF_FUNC_get_stack &&
3929 func_id != BPF_FUNC_probe_read_str))
3932 ret_reg->smax_value = meta->msize_smax_value;
3933 ret_reg->umax_value = meta->msize_umax_value;
3934 __reg_deduce_bounds(ret_reg);
3935 __reg_bound_offset(ret_reg);
3939 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
3940 int func_id, int insn_idx)
3942 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
3943 struct bpf_map *map = meta->map_ptr;
3945 if (func_id != BPF_FUNC_tail_call &&
3946 func_id != BPF_FUNC_map_lookup_elem &&
3947 func_id != BPF_FUNC_map_update_elem &&
3948 func_id != BPF_FUNC_map_delete_elem &&
3949 func_id != BPF_FUNC_map_push_elem &&
3950 func_id != BPF_FUNC_map_pop_elem &&
3951 func_id != BPF_FUNC_map_peek_elem)
3955 verbose(env, "kernel subsystem misconfigured verifier\n");
3959 /* In case of read-only, some additional restrictions
3960 * need to be applied in order to prevent altering the
3961 * state of the map from program side.
3963 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
3964 (func_id == BPF_FUNC_map_delete_elem ||
3965 func_id == BPF_FUNC_map_update_elem ||
3966 func_id == BPF_FUNC_map_push_elem ||
3967 func_id == BPF_FUNC_map_pop_elem)) {
3968 verbose(env, "write into map forbidden\n");
3972 if (!BPF_MAP_PTR(aux->map_state))
3973 bpf_map_ptr_store(aux, meta->map_ptr,
3974 meta->map_ptr->unpriv_array);
3975 else if (BPF_MAP_PTR(aux->map_state) != meta->map_ptr)
3976 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
3977 meta->map_ptr->unpriv_array);
3981 static int check_reference_leak(struct bpf_verifier_env *env)
3983 struct bpf_func_state *state = cur_func(env);
3986 for (i = 0; i < state->acquired_refs; i++) {
3987 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
3988 state->refs[i].id, state->refs[i].insn_idx);
3990 return state->acquired_refs ? -EINVAL : 0;
3993 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
3995 const struct bpf_func_proto *fn = NULL;
3996 struct bpf_reg_state *regs;
3997 struct bpf_call_arg_meta meta;
4001 /* find function prototype */
4002 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
4003 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
4008 if (env->ops->get_func_proto)
4009 fn = env->ops->get_func_proto(func_id, env->prog);
4011 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
4016 /* eBPF programs must be GPL compatible to use GPL-ed functions */
4017 if (!env->prog->gpl_compatible && fn->gpl_only) {
4018 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
4022 /* With LD_ABS/IND some JITs save/restore skb from r1. */
4023 changes_data = bpf_helper_changes_pkt_data(fn->func);
4024 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
4025 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
4026 func_id_name(func_id), func_id);
4030 memset(&meta, 0, sizeof(meta));
4031 meta.pkt_access = fn->pkt_access;
4033 err = check_func_proto(fn, func_id);
4035 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
4036 func_id_name(func_id), func_id);
4040 meta.func_id = func_id;
4042 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
4045 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
4048 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
4051 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
4054 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
4058 err = record_func_map(env, &meta, func_id, insn_idx);
4062 /* Mark slots with STACK_MISC in case of raw mode, stack offset
4063 * is inferred from register state.
4065 for (i = 0; i < meta.access_size; i++) {
4066 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
4067 BPF_WRITE, -1, false);
4072 if (func_id == BPF_FUNC_tail_call) {
4073 err = check_reference_leak(env);
4075 verbose(env, "tail_call would lead to reference leak\n");
4078 } else if (is_release_function(func_id)) {
4079 err = release_reference(env, meta.ref_obj_id);
4081 verbose(env, "func %s#%d reference has not been acquired before\n",
4082 func_id_name(func_id), func_id);
4087 regs = cur_regs(env);
4089 /* check that flags argument in get_local_storage(map, flags) is 0,
4090 * this is required because get_local_storage() can't return an error.
4092 if (func_id == BPF_FUNC_get_local_storage &&
4093 !register_is_null(®s[BPF_REG_2])) {
4094 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
4098 /* reset caller saved regs */
4099 for (i = 0; i < CALLER_SAVED_REGS; i++) {
4100 mark_reg_not_init(env, regs, caller_saved[i]);
4101 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
4104 /* helper call returns 64-bit value. */
4105 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
4107 /* update return register (already marked as written above) */
4108 if (fn->ret_type == RET_INTEGER) {
4109 /* sets type to SCALAR_VALUE */
4110 mark_reg_unknown(env, regs, BPF_REG_0);
4111 } else if (fn->ret_type == RET_VOID) {
4112 regs[BPF_REG_0].type = NOT_INIT;
4113 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
4114 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
4115 /* There is no offset yet applied, variable or fixed */
4116 mark_reg_known_zero(env, regs, BPF_REG_0);
4117 /* remember map_ptr, so that check_map_access()
4118 * can check 'value_size' boundary of memory access
4119 * to map element returned from bpf_map_lookup_elem()
4121 if (meta.map_ptr == NULL) {
4123 "kernel subsystem misconfigured verifier\n");
4126 regs[BPF_REG_0].map_ptr = meta.map_ptr;
4127 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
4128 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
4129 if (map_value_has_spin_lock(meta.map_ptr))
4130 regs[BPF_REG_0].id = ++env->id_gen;
4132 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
4133 regs[BPF_REG_0].id = ++env->id_gen;
4135 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
4136 mark_reg_known_zero(env, regs, BPF_REG_0);
4137 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
4138 regs[BPF_REG_0].id = ++env->id_gen;
4139 } else if (fn->ret_type == RET_PTR_TO_SOCK_COMMON_OR_NULL) {
4140 mark_reg_known_zero(env, regs, BPF_REG_0);
4141 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON_OR_NULL;
4142 regs[BPF_REG_0].id = ++env->id_gen;
4143 } else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
4144 mark_reg_known_zero(env, regs, BPF_REG_0);
4145 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
4146 regs[BPF_REG_0].id = ++env->id_gen;
4148 verbose(env, "unknown return type %d of func %s#%d\n",
4149 fn->ret_type, func_id_name(func_id), func_id);
4153 if (is_ptr_cast_function(func_id)) {
4154 /* For release_reference() */
4155 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
4156 } else if (is_acquire_function(func_id)) {
4157 int id = acquire_reference_state(env, insn_idx);
4161 /* For mark_ptr_or_null_reg() */
4162 regs[BPF_REG_0].id = id;
4163 /* For release_reference() */
4164 regs[BPF_REG_0].ref_obj_id = id;
4167 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
4169 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
4173 if (func_id == BPF_FUNC_get_stack && !env->prog->has_callchain_buf) {
4174 const char *err_str;
4176 #ifdef CONFIG_PERF_EVENTS
4177 err = get_callchain_buffers(sysctl_perf_event_max_stack);
4178 err_str = "cannot get callchain buffer for func %s#%d\n";
4181 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
4184 verbose(env, err_str, func_id_name(func_id), func_id);
4188 env->prog->has_callchain_buf = true;
4192 clear_all_pkt_pointers(env);
4196 static bool signed_add_overflows(s64 a, s64 b)
4198 /* Do the add in u64, where overflow is well-defined */
4199 s64 res = (s64)((u64)a + (u64)b);
4206 static bool signed_sub_overflows(s64 a, s64 b)
4208 /* Do the sub in u64, where overflow is well-defined */
4209 s64 res = (s64)((u64)a - (u64)b);
4216 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
4217 const struct bpf_reg_state *reg,
4218 enum bpf_reg_type type)
4220 bool known = tnum_is_const(reg->var_off);
4221 s64 val = reg->var_off.value;
4222 s64 smin = reg->smin_value;
4224 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
4225 verbose(env, "math between %s pointer and %lld is not allowed\n",
4226 reg_type_str[type], val);
4230 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
4231 verbose(env, "%s pointer offset %d is not allowed\n",
4232 reg_type_str[type], reg->off);
4236 if (smin == S64_MIN) {
4237 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
4238 reg_type_str[type]);
4242 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
4243 verbose(env, "value %lld makes %s pointer be out of bounds\n",
4244 smin, reg_type_str[type]);
4251 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
4253 return &env->insn_aux_data[env->insn_idx];
4256 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
4257 u32 *ptr_limit, u8 opcode, bool off_is_neg)
4259 bool mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
4260 (opcode == BPF_SUB && !off_is_neg);
4263 switch (ptr_reg->type) {
4265 /* Indirect variable offset stack access is prohibited in
4266 * unprivileged mode so it's not handled here.
4268 off = ptr_reg->off + ptr_reg->var_off.value;
4270 *ptr_limit = MAX_BPF_STACK + off;
4274 case PTR_TO_MAP_VALUE:
4276 *ptr_limit = ptr_reg->umax_value + ptr_reg->off;
4278 off = ptr_reg->smin_value + ptr_reg->off;
4279 *ptr_limit = ptr_reg->map_ptr->value_size - off;
4287 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
4288 const struct bpf_insn *insn)
4290 return env->allow_ptr_leaks || BPF_SRC(insn->code) == BPF_K;
4293 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
4294 u32 alu_state, u32 alu_limit)
4296 /* If we arrived here from different branches with different
4297 * state or limits to sanitize, then this won't work.
4299 if (aux->alu_state &&
4300 (aux->alu_state != alu_state ||
4301 aux->alu_limit != alu_limit))
4304 /* Corresponding fixup done in fixup_bpf_calls(). */
4305 aux->alu_state = alu_state;
4306 aux->alu_limit = alu_limit;
4310 static int sanitize_val_alu(struct bpf_verifier_env *env,
4311 struct bpf_insn *insn)
4313 struct bpf_insn_aux_data *aux = cur_aux(env);
4315 if (can_skip_alu_sanitation(env, insn))
4318 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
4321 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
4322 struct bpf_insn *insn,
4323 const struct bpf_reg_state *ptr_reg,
4324 struct bpf_reg_state *dst_reg,
4327 struct bpf_verifier_state *vstate = env->cur_state;
4328 struct bpf_insn_aux_data *aux = cur_aux(env);
4329 bool ptr_is_dst_reg = ptr_reg == dst_reg;
4330 u8 opcode = BPF_OP(insn->code);
4331 u32 alu_state, alu_limit;
4332 struct bpf_reg_state tmp;
4335 if (can_skip_alu_sanitation(env, insn))
4338 /* We already marked aux for masking from non-speculative
4339 * paths, thus we got here in the first place. We only care
4340 * to explore bad access from here.
4342 if (vstate->speculative)
4345 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
4346 alu_state |= ptr_is_dst_reg ?
4347 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
4349 if (retrieve_ptr_limit(ptr_reg, &alu_limit, opcode, off_is_neg))
4351 if (update_alu_sanitation_state(aux, alu_state, alu_limit))
4354 /* Simulate and find potential out-of-bounds access under
4355 * speculative execution from truncation as a result of
4356 * masking when off was not within expected range. If off
4357 * sits in dst, then we temporarily need to move ptr there
4358 * to simulate dst (== 0) +/-= ptr. Needed, for example,
4359 * for cases where we use K-based arithmetic in one direction
4360 * and truncated reg-based in the other in order to explore
4363 if (!ptr_is_dst_reg) {
4365 *dst_reg = *ptr_reg;
4367 ret = push_stack(env, env->insn_idx + 1, env->insn_idx, true);
4368 if (!ptr_is_dst_reg && ret)
4370 return !ret ? -EFAULT : 0;
4373 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
4374 * Caller should also handle BPF_MOV case separately.
4375 * If we return -EACCES, caller may want to try again treating pointer as a
4376 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
4378 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
4379 struct bpf_insn *insn,
4380 const struct bpf_reg_state *ptr_reg,
4381 const struct bpf_reg_state *off_reg)
4383 struct bpf_verifier_state *vstate = env->cur_state;
4384 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4385 struct bpf_reg_state *regs = state->regs, *dst_reg;
4386 bool known = tnum_is_const(off_reg->var_off);
4387 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
4388 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
4389 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
4390 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
4391 u32 dst = insn->dst_reg, src = insn->src_reg;
4392 u8 opcode = BPF_OP(insn->code);
4395 dst_reg = ®s[dst];
4397 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
4398 smin_val > smax_val || umin_val > umax_val) {
4399 /* Taint dst register if offset had invalid bounds derived from
4400 * e.g. dead branches.
4402 __mark_reg_unknown(dst_reg);
4406 if (BPF_CLASS(insn->code) != BPF_ALU64) {
4407 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
4409 "R%d 32-bit pointer arithmetic prohibited\n",
4414 switch (ptr_reg->type) {
4415 case PTR_TO_MAP_VALUE_OR_NULL:
4416 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
4417 dst, reg_type_str[ptr_reg->type]);
4419 case CONST_PTR_TO_MAP:
4420 case PTR_TO_PACKET_END:
4422 case PTR_TO_SOCKET_OR_NULL:
4423 case PTR_TO_SOCK_COMMON:
4424 case PTR_TO_SOCK_COMMON_OR_NULL:
4425 case PTR_TO_TCP_SOCK:
4426 case PTR_TO_TCP_SOCK_OR_NULL:
4427 case PTR_TO_XDP_SOCK:
4428 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
4429 dst, reg_type_str[ptr_reg->type]);
4431 case PTR_TO_MAP_VALUE:
4432 if (!env->allow_ptr_leaks && !known && (smin_val < 0) != (smax_val < 0)) {
4433 verbose(env, "R%d has unknown scalar with mixed signed bounds, pointer arithmetic with it prohibited for !root\n",
4434 off_reg == dst_reg ? dst : src);
4442 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
4443 * The id may be overwritten later if we create a new variable offset.
4445 dst_reg->type = ptr_reg->type;
4446 dst_reg->id = ptr_reg->id;
4448 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
4449 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
4454 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
4456 verbose(env, "R%d tried to add from different maps or paths\n", dst);
4459 /* We can take a fixed offset as long as it doesn't overflow
4460 * the s32 'off' field
4462 if (known && (ptr_reg->off + smin_val ==
4463 (s64)(s32)(ptr_reg->off + smin_val))) {
4464 /* pointer += K. Accumulate it into fixed offset */
4465 dst_reg->smin_value = smin_ptr;
4466 dst_reg->smax_value = smax_ptr;
4467 dst_reg->umin_value = umin_ptr;
4468 dst_reg->umax_value = umax_ptr;
4469 dst_reg->var_off = ptr_reg->var_off;
4470 dst_reg->off = ptr_reg->off + smin_val;
4471 dst_reg->raw = ptr_reg->raw;
4474 /* A new variable offset is created. Note that off_reg->off
4475 * == 0, since it's a scalar.
4476 * dst_reg gets the pointer type and since some positive
4477 * integer value was added to the pointer, give it a new 'id'
4478 * if it's a PTR_TO_PACKET.
4479 * this creates a new 'base' pointer, off_reg (variable) gets
4480 * added into the variable offset, and we copy the fixed offset
4483 if (signed_add_overflows(smin_ptr, smin_val) ||
4484 signed_add_overflows(smax_ptr, smax_val)) {
4485 dst_reg->smin_value = S64_MIN;
4486 dst_reg->smax_value = S64_MAX;
4488 dst_reg->smin_value = smin_ptr + smin_val;
4489 dst_reg->smax_value = smax_ptr + smax_val;
4491 if (umin_ptr + umin_val < umin_ptr ||
4492 umax_ptr + umax_val < umax_ptr) {
4493 dst_reg->umin_value = 0;
4494 dst_reg->umax_value = U64_MAX;
4496 dst_reg->umin_value = umin_ptr + umin_val;
4497 dst_reg->umax_value = umax_ptr + umax_val;
4499 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
4500 dst_reg->off = ptr_reg->off;
4501 dst_reg->raw = ptr_reg->raw;
4502 if (reg_is_pkt_pointer(ptr_reg)) {
4503 dst_reg->id = ++env->id_gen;
4504 /* something was added to pkt_ptr, set range to zero */
4509 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
4511 verbose(env, "R%d tried to sub from different maps or paths\n", dst);
4514 if (dst_reg == off_reg) {
4515 /* scalar -= pointer. Creates an unknown scalar */
4516 verbose(env, "R%d tried to subtract pointer from scalar\n",
4520 /* We don't allow subtraction from FP, because (according to
4521 * test_verifier.c test "invalid fp arithmetic", JITs might not
4522 * be able to deal with it.
4524 if (ptr_reg->type == PTR_TO_STACK) {
4525 verbose(env, "R%d subtraction from stack pointer prohibited\n",
4529 if (known && (ptr_reg->off - smin_val ==
4530 (s64)(s32)(ptr_reg->off - smin_val))) {
4531 /* pointer -= K. Subtract it from fixed offset */
4532 dst_reg->smin_value = smin_ptr;
4533 dst_reg->smax_value = smax_ptr;
4534 dst_reg->umin_value = umin_ptr;
4535 dst_reg->umax_value = umax_ptr;
4536 dst_reg->var_off = ptr_reg->var_off;
4537 dst_reg->id = ptr_reg->id;
4538 dst_reg->off = ptr_reg->off - smin_val;
4539 dst_reg->raw = ptr_reg->raw;
4542 /* A new variable offset is created. If the subtrahend is known
4543 * nonnegative, then any reg->range we had before is still good.
4545 if (signed_sub_overflows(smin_ptr, smax_val) ||
4546 signed_sub_overflows(smax_ptr, smin_val)) {
4547 /* Overflow possible, we know nothing */
4548 dst_reg->smin_value = S64_MIN;
4549 dst_reg->smax_value = S64_MAX;
4551 dst_reg->smin_value = smin_ptr - smax_val;
4552 dst_reg->smax_value = smax_ptr - smin_val;
4554 if (umin_ptr < umax_val) {
4555 /* Overflow possible, we know nothing */
4556 dst_reg->umin_value = 0;
4557 dst_reg->umax_value = U64_MAX;
4559 /* Cannot overflow (as long as bounds are consistent) */
4560 dst_reg->umin_value = umin_ptr - umax_val;
4561 dst_reg->umax_value = umax_ptr - umin_val;
4563 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
4564 dst_reg->off = ptr_reg->off;
4565 dst_reg->raw = ptr_reg->raw;
4566 if (reg_is_pkt_pointer(ptr_reg)) {
4567 dst_reg->id = ++env->id_gen;
4568 /* something was added to pkt_ptr, set range to zero */
4576 /* bitwise ops on pointers are troublesome, prohibit. */
4577 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
4578 dst, bpf_alu_string[opcode >> 4]);
4581 /* other operators (e.g. MUL,LSH) produce non-pointer results */
4582 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
4583 dst, bpf_alu_string[opcode >> 4]);
4587 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
4590 __update_reg_bounds(dst_reg);
4591 __reg_deduce_bounds(dst_reg);
4592 __reg_bound_offset(dst_reg);
4594 /* For unprivileged we require that resulting offset must be in bounds
4595 * in order to be able to sanitize access later on.
4597 if (!env->allow_ptr_leaks) {
4598 if (dst_reg->type == PTR_TO_MAP_VALUE &&
4599 check_map_access(env, dst, dst_reg->off, 1, false)) {
4600 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
4601 "prohibited for !root\n", dst);
4603 } else if (dst_reg->type == PTR_TO_STACK &&
4604 check_stack_access(env, dst_reg, dst_reg->off +
4605 dst_reg->var_off.value, 1)) {
4606 verbose(env, "R%d stack pointer arithmetic goes out of range, "
4607 "prohibited for !root\n", dst);
4615 /* WARNING: This function does calculations on 64-bit values, but the actual
4616 * execution may occur on 32-bit values. Therefore, things like bitshifts
4617 * need extra checks in the 32-bit case.
4619 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
4620 struct bpf_insn *insn,
4621 struct bpf_reg_state *dst_reg,
4622 struct bpf_reg_state src_reg)
4624 struct bpf_reg_state *regs = cur_regs(env);
4625 u8 opcode = BPF_OP(insn->code);
4626 bool src_known, dst_known;
4627 s64 smin_val, smax_val;
4628 u64 umin_val, umax_val;
4629 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
4630 u32 dst = insn->dst_reg;
4633 if (insn_bitness == 32) {
4634 /* Relevant for 32-bit RSH: Information can propagate towards
4635 * LSB, so it isn't sufficient to only truncate the output to
4638 coerce_reg_to_size(dst_reg, 4);
4639 coerce_reg_to_size(&src_reg, 4);
4642 smin_val = src_reg.smin_value;
4643 smax_val = src_reg.smax_value;
4644 umin_val = src_reg.umin_value;
4645 umax_val = src_reg.umax_value;
4646 src_known = tnum_is_const(src_reg.var_off);
4647 dst_known = tnum_is_const(dst_reg->var_off);
4649 if ((src_known && (smin_val != smax_val || umin_val != umax_val)) ||
4650 smin_val > smax_val || umin_val > umax_val) {
4651 /* Taint dst register if offset had invalid bounds derived from
4652 * e.g. dead branches.
4654 __mark_reg_unknown(dst_reg);
4659 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
4660 __mark_reg_unknown(dst_reg);
4666 ret = sanitize_val_alu(env, insn);
4668 verbose(env, "R%d tried to add from different pointers or scalars\n", dst);
4671 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
4672 signed_add_overflows(dst_reg->smax_value, smax_val)) {
4673 dst_reg->smin_value = S64_MIN;
4674 dst_reg->smax_value = S64_MAX;
4676 dst_reg->smin_value += smin_val;
4677 dst_reg->smax_value += smax_val;
4679 if (dst_reg->umin_value + umin_val < umin_val ||
4680 dst_reg->umax_value + umax_val < umax_val) {
4681 dst_reg->umin_value = 0;
4682 dst_reg->umax_value = U64_MAX;
4684 dst_reg->umin_value += umin_val;
4685 dst_reg->umax_value += umax_val;
4687 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
4690 ret = sanitize_val_alu(env, insn);
4692 verbose(env, "R%d tried to sub from different pointers or scalars\n", dst);
4695 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
4696 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
4697 /* Overflow possible, we know nothing */
4698 dst_reg->smin_value = S64_MIN;
4699 dst_reg->smax_value = S64_MAX;
4701 dst_reg->smin_value -= smax_val;
4702 dst_reg->smax_value -= smin_val;
4704 if (dst_reg->umin_value < umax_val) {
4705 /* Overflow possible, we know nothing */
4706 dst_reg->umin_value = 0;
4707 dst_reg->umax_value = U64_MAX;
4709 /* Cannot overflow (as long as bounds are consistent) */
4710 dst_reg->umin_value -= umax_val;
4711 dst_reg->umax_value -= umin_val;
4713 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
4716 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
4717 if (smin_val < 0 || dst_reg->smin_value < 0) {
4718 /* Ain't nobody got time to multiply that sign */
4719 __mark_reg_unbounded(dst_reg);
4720 __update_reg_bounds(dst_reg);
4723 /* Both values are positive, so we can work with unsigned and
4724 * copy the result to signed (unless it exceeds S64_MAX).
4726 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
4727 /* Potential overflow, we know nothing */
4728 __mark_reg_unbounded(dst_reg);
4729 /* (except what we can learn from the var_off) */
4730 __update_reg_bounds(dst_reg);
4733 dst_reg->umin_value *= umin_val;
4734 dst_reg->umax_value *= umax_val;
4735 if (dst_reg->umax_value > S64_MAX) {
4736 /* Overflow possible, we know nothing */
4737 dst_reg->smin_value = S64_MIN;
4738 dst_reg->smax_value = S64_MAX;
4740 dst_reg->smin_value = dst_reg->umin_value;
4741 dst_reg->smax_value = dst_reg->umax_value;
4745 if (src_known && dst_known) {
4746 __mark_reg_known(dst_reg, dst_reg->var_off.value &
4747 src_reg.var_off.value);
4750 /* We get our minimum from the var_off, since that's inherently
4751 * bitwise. Our maximum is the minimum of the operands' maxima.
4753 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
4754 dst_reg->umin_value = dst_reg->var_off.value;
4755 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
4756 if (dst_reg->smin_value < 0 || smin_val < 0) {
4757 /* Lose signed bounds when ANDing negative numbers,
4758 * ain't nobody got time for that.
4760 dst_reg->smin_value = S64_MIN;
4761 dst_reg->smax_value = S64_MAX;
4763 /* ANDing two positives gives a positive, so safe to
4764 * cast result into s64.
4766 dst_reg->smin_value = dst_reg->umin_value;
4767 dst_reg->smax_value = dst_reg->umax_value;
4769 /* We may learn something more from the var_off */
4770 __update_reg_bounds(dst_reg);
4773 if (src_known && dst_known) {
4774 __mark_reg_known(dst_reg, dst_reg->var_off.value |
4775 src_reg.var_off.value);
4778 /* We get our maximum from the var_off, and our minimum is the
4779 * maximum of the operands' minima
4781 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
4782 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
4783 dst_reg->umax_value = dst_reg->var_off.value |
4784 dst_reg->var_off.mask;
4785 if (dst_reg->smin_value < 0 || smin_val < 0) {
4786 /* Lose signed bounds when ORing negative numbers,
4787 * ain't nobody got time for that.
4789 dst_reg->smin_value = S64_MIN;
4790 dst_reg->smax_value = S64_MAX;
4792 /* ORing two positives gives a positive, so safe to
4793 * cast result into s64.
4795 dst_reg->smin_value = dst_reg->umin_value;
4796 dst_reg->smax_value = dst_reg->umax_value;
4798 /* We may learn something more from the var_off */
4799 __update_reg_bounds(dst_reg);
4802 if (umax_val >= insn_bitness) {
4803 /* Shifts greater than 31 or 63 are undefined.
4804 * This includes shifts by a negative number.
4806 mark_reg_unknown(env, regs, insn->dst_reg);
4809 /* We lose all sign bit information (except what we can pick
4812 dst_reg->smin_value = S64_MIN;
4813 dst_reg->smax_value = S64_MAX;
4814 /* If we might shift our top bit out, then we know nothing */
4815 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
4816 dst_reg->umin_value = 0;
4817 dst_reg->umax_value = U64_MAX;
4819 dst_reg->umin_value <<= umin_val;
4820 dst_reg->umax_value <<= umax_val;
4822 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
4823 /* We may learn something more from the var_off */
4824 __update_reg_bounds(dst_reg);
4827 if (umax_val >= insn_bitness) {
4828 /* Shifts greater than 31 or 63 are undefined.
4829 * This includes shifts by a negative number.
4831 mark_reg_unknown(env, regs, insn->dst_reg);
4834 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
4835 * be negative, then either:
4836 * 1) src_reg might be zero, so the sign bit of the result is
4837 * unknown, so we lose our signed bounds
4838 * 2) it's known negative, thus the unsigned bounds capture the
4840 * 3) the signed bounds cross zero, so they tell us nothing
4842 * If the value in dst_reg is known nonnegative, then again the
4843 * unsigned bounts capture the signed bounds.
4844 * Thus, in all cases it suffices to blow away our signed bounds
4845 * and rely on inferring new ones from the unsigned bounds and
4846 * var_off of the result.
4848 dst_reg->smin_value = S64_MIN;
4849 dst_reg->smax_value = S64_MAX;
4850 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
4851 dst_reg->umin_value >>= umax_val;
4852 dst_reg->umax_value >>= umin_val;
4853 /* We may learn something more from the var_off */
4854 __update_reg_bounds(dst_reg);
4857 if (umax_val >= insn_bitness) {
4858 /* Shifts greater than 31 or 63 are undefined.
4859 * This includes shifts by a negative number.
4861 mark_reg_unknown(env, regs, insn->dst_reg);
4865 /* Upon reaching here, src_known is true and
4866 * umax_val is equal to umin_val.
4868 dst_reg->smin_value >>= umin_val;
4869 dst_reg->smax_value >>= umin_val;
4870 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val);
4872 /* blow away the dst_reg umin_value/umax_value and rely on
4873 * dst_reg var_off to refine the result.
4875 dst_reg->umin_value = 0;
4876 dst_reg->umax_value = U64_MAX;
4877 __update_reg_bounds(dst_reg);
4880 mark_reg_unknown(env, regs, insn->dst_reg);
4884 if (BPF_CLASS(insn->code) != BPF_ALU64) {
4885 /* 32-bit ALU ops are (32,32)->32 */
4886 coerce_reg_to_size(dst_reg, 4);
4889 __reg_deduce_bounds(dst_reg);
4890 __reg_bound_offset(dst_reg);
4894 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
4897 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
4898 struct bpf_insn *insn)
4900 struct bpf_verifier_state *vstate = env->cur_state;
4901 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4902 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
4903 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
4904 u8 opcode = BPF_OP(insn->code);
4907 dst_reg = ®s[insn->dst_reg];
4909 if (dst_reg->type != SCALAR_VALUE)
4911 if (BPF_SRC(insn->code) == BPF_X) {
4912 src_reg = ®s[insn->src_reg];
4913 if (src_reg->type != SCALAR_VALUE) {
4914 if (dst_reg->type != SCALAR_VALUE) {
4915 /* Combining two pointers by any ALU op yields
4916 * an arbitrary scalar. Disallow all math except
4917 * pointer subtraction
4919 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
4920 mark_reg_unknown(env, regs, insn->dst_reg);
4923 verbose(env, "R%d pointer %s pointer prohibited\n",
4925 bpf_alu_string[opcode >> 4]);
4928 /* scalar += pointer
4929 * This is legal, but we have to reverse our
4930 * src/dest handling in computing the range
4932 err = mark_chain_precision(env, insn->dst_reg);
4935 return adjust_ptr_min_max_vals(env, insn,
4938 } else if (ptr_reg) {
4939 /* pointer += scalar */
4940 err = mark_chain_precision(env, insn->src_reg);
4943 return adjust_ptr_min_max_vals(env, insn,
4947 /* Pretend the src is a reg with a known value, since we only
4948 * need to be able to read from this state.
4950 off_reg.type = SCALAR_VALUE;
4951 __mark_reg_known(&off_reg, insn->imm);
4953 if (ptr_reg) /* pointer += K */
4954 return adjust_ptr_min_max_vals(env, insn,
4958 /* Got here implies adding two SCALAR_VALUEs */
4959 if (WARN_ON_ONCE(ptr_reg)) {
4960 print_verifier_state(env, state);
4961 verbose(env, "verifier internal error: unexpected ptr_reg\n");
4964 if (WARN_ON(!src_reg)) {
4965 print_verifier_state(env, state);
4966 verbose(env, "verifier internal error: no src_reg\n");
4969 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
4972 /* check validity of 32-bit and 64-bit arithmetic operations */
4973 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
4975 struct bpf_reg_state *regs = cur_regs(env);
4976 u8 opcode = BPF_OP(insn->code);
4979 if (opcode == BPF_END || opcode == BPF_NEG) {
4980 if (opcode == BPF_NEG) {
4981 if (BPF_SRC(insn->code) != 0 ||
4982 insn->src_reg != BPF_REG_0 ||
4983 insn->off != 0 || insn->imm != 0) {
4984 verbose(env, "BPF_NEG uses reserved fields\n");
4988 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
4989 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
4990 BPF_CLASS(insn->code) == BPF_ALU64) {
4991 verbose(env, "BPF_END uses reserved fields\n");
4996 /* check src operand */
4997 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5001 if (is_pointer_value(env, insn->dst_reg)) {
5002 verbose(env, "R%d pointer arithmetic prohibited\n",
5007 /* check dest operand */
5008 err = check_reg_arg(env, insn->dst_reg, DST_OP);
5012 } else if (opcode == BPF_MOV) {
5014 if (BPF_SRC(insn->code) == BPF_X) {
5015 if (insn->imm != 0 || insn->off != 0) {
5016 verbose(env, "BPF_MOV uses reserved fields\n");
5020 /* check src operand */
5021 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5025 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
5026 verbose(env, "BPF_MOV uses reserved fields\n");
5031 /* check dest operand, mark as required later */
5032 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
5036 if (BPF_SRC(insn->code) == BPF_X) {
5037 struct bpf_reg_state *src_reg = regs + insn->src_reg;
5038 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
5040 if (BPF_CLASS(insn->code) == BPF_ALU64) {
5042 * copy register state to dest reg
5044 *dst_reg = *src_reg;
5045 dst_reg->live |= REG_LIVE_WRITTEN;
5046 dst_reg->subreg_def = DEF_NOT_SUBREG;
5049 if (is_pointer_value(env, insn->src_reg)) {
5051 "R%d partial copy of pointer\n",
5054 } else if (src_reg->type == SCALAR_VALUE) {
5055 *dst_reg = *src_reg;
5056 dst_reg->live |= REG_LIVE_WRITTEN;
5057 dst_reg->subreg_def = env->insn_idx + 1;
5059 mark_reg_unknown(env, regs,
5062 coerce_reg_to_size(dst_reg, 4);
5066 * remember the value we stored into this reg
5068 /* clear any state __mark_reg_known doesn't set */
5069 mark_reg_unknown(env, regs, insn->dst_reg);
5070 regs[insn->dst_reg].type = SCALAR_VALUE;
5071 if (BPF_CLASS(insn->code) == BPF_ALU64) {
5072 __mark_reg_known(regs + insn->dst_reg,
5075 __mark_reg_known(regs + insn->dst_reg,
5080 } else if (opcode > BPF_END) {
5081 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
5084 } else { /* all other ALU ops: and, sub, xor, add, ... */
5086 if (BPF_SRC(insn->code) == BPF_X) {
5087 if (insn->imm != 0 || insn->off != 0) {
5088 verbose(env, "BPF_ALU uses reserved fields\n");
5091 /* check src1 operand */
5092 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5096 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
5097 verbose(env, "BPF_ALU uses reserved fields\n");
5102 /* check src2 operand */
5103 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5107 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
5108 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
5109 verbose(env, "div by zero\n");
5113 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
5114 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
5115 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
5117 if (insn->imm < 0 || insn->imm >= size) {
5118 verbose(env, "invalid shift %d\n", insn->imm);
5123 /* check dest operand */
5124 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
5128 return adjust_reg_min_max_vals(env, insn);
5134 static void __find_good_pkt_pointers(struct bpf_func_state *state,
5135 struct bpf_reg_state *dst_reg,
5136 enum bpf_reg_type type, u16 new_range)
5138 struct bpf_reg_state *reg;
5141 for (i = 0; i < MAX_BPF_REG; i++) {
5142 reg = &state->regs[i];
5143 if (reg->type == type && reg->id == dst_reg->id)
5144 /* keep the maximum range already checked */
5145 reg->range = max(reg->range, new_range);
5148 bpf_for_each_spilled_reg(i, state, reg) {
5151 if (reg->type == type && reg->id == dst_reg->id)
5152 reg->range = max(reg->range, new_range);
5156 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
5157 struct bpf_reg_state *dst_reg,
5158 enum bpf_reg_type type,
5159 bool range_right_open)
5164 if (dst_reg->off < 0 ||
5165 (dst_reg->off == 0 && range_right_open))
5166 /* This doesn't give us any range */
5169 if (dst_reg->umax_value > MAX_PACKET_OFF ||
5170 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
5171 /* Risk of overflow. For instance, ptr + (1<<63) may be less
5172 * than pkt_end, but that's because it's also less than pkt.
5176 new_range = dst_reg->off;
5177 if (range_right_open)
5180 /* Examples for register markings:
5182 * pkt_data in dst register:
5186 * if (r2 > pkt_end) goto <handle exception>
5191 * if (r2 < pkt_end) goto <access okay>
5192 * <handle exception>
5195 * r2 == dst_reg, pkt_end == src_reg
5196 * r2=pkt(id=n,off=8,r=0)
5197 * r3=pkt(id=n,off=0,r=0)
5199 * pkt_data in src register:
5203 * if (pkt_end >= r2) goto <access okay>
5204 * <handle exception>
5208 * if (pkt_end <= r2) goto <handle exception>
5212 * pkt_end == dst_reg, r2 == src_reg
5213 * r2=pkt(id=n,off=8,r=0)
5214 * r3=pkt(id=n,off=0,r=0)
5216 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
5217 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
5218 * and [r3, r3 + 8-1) respectively is safe to access depending on
5222 /* If our ids match, then we must have the same max_value. And we
5223 * don't care about the other reg's fixed offset, since if it's too big
5224 * the range won't allow anything.
5225 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
5227 for (i = 0; i <= vstate->curframe; i++)
5228 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
5232 /* compute branch direction of the expression "if (reg opcode val) goto target;"
5234 * 1 - branch will be taken and "goto target" will be executed
5235 * 0 - branch will not be taken and fall-through to next insn
5236 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value range [0,10]
5238 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
5241 struct bpf_reg_state reg_lo;
5244 if (__is_pointer_value(false, reg))
5250 /* For JMP32, only low 32 bits are compared, coerce_reg_to_size
5251 * could truncate high bits and update umin/umax according to
5252 * information of low bits.
5254 coerce_reg_to_size(reg, 4);
5255 /* smin/smax need special handling. For example, after coerce,
5256 * if smin_value is 0x00000000ffffffffLL, the value is -1 when
5257 * used as operand to JMP32. It is a negative number from s32's
5258 * point of view, while it is a positive number when seen as
5259 * s64. The smin/smax are kept as s64, therefore, when used with
5260 * JMP32, they need to be transformed into s32, then sign
5261 * extended back to s64.
5263 * Also, smin/smax were copied from umin/umax. If umin/umax has
5264 * different sign bit, then min/max relationship doesn't
5265 * maintain after casting into s32, for this case, set smin/smax
5268 if ((reg->umax_value ^ reg->umin_value) &
5270 reg->smin_value = S32_MIN;
5271 reg->smax_value = S32_MAX;
5273 reg->smin_value = (s64)(s32)reg->smin_value;
5274 reg->smax_value = (s64)(s32)reg->smax_value;
5277 sval = (s64)(s32)val;
5284 if (tnum_is_const(reg->var_off))
5285 return !!tnum_equals_const(reg->var_off, val);
5288 if (tnum_is_const(reg->var_off))
5289 return !tnum_equals_const(reg->var_off, val);
5292 if ((~reg->var_off.mask & reg->var_off.value) & val)
5294 if (!((reg->var_off.mask | reg->var_off.value) & val))
5298 if (reg->umin_value > val)
5300 else if (reg->umax_value <= val)
5304 if (reg->smin_value > sval)
5306 else if (reg->smax_value < sval)
5310 if (reg->umax_value < val)
5312 else if (reg->umin_value >= val)
5316 if (reg->smax_value < sval)
5318 else if (reg->smin_value >= sval)
5322 if (reg->umin_value >= val)
5324 else if (reg->umax_value < val)
5328 if (reg->smin_value >= sval)
5330 else if (reg->smax_value < sval)
5334 if (reg->umax_value <= val)
5336 else if (reg->umin_value > val)
5340 if (reg->smax_value <= sval)
5342 else if (reg->smin_value > sval)
5350 /* Generate min value of the high 32-bit from TNUM info. */
5351 static u64 gen_hi_min(struct tnum var)
5353 return var.value & ~0xffffffffULL;
5356 /* Generate max value of the high 32-bit from TNUM info. */
5357 static u64 gen_hi_max(struct tnum var)
5359 return (var.value | var.mask) & ~0xffffffffULL;
5362 /* Return true if VAL is compared with a s64 sign extended from s32, and they
5363 * are with the same signedness.
5365 static bool cmp_val_with_extended_s64(s64 sval, struct bpf_reg_state *reg)
5367 return ((s32)sval >= 0 &&
5368 reg->smin_value >= 0 && reg->smax_value <= S32_MAX) ||
5370 reg->smax_value <= 0 && reg->smin_value >= S32_MIN);
5373 /* Adjusts the register min/max values in the case that the dst_reg is the
5374 * variable register that we are working on, and src_reg is a constant or we're
5375 * simply doing a BPF_K check.
5376 * In JEQ/JNE cases we also adjust the var_off values.
5378 static void reg_set_min_max(struct bpf_reg_state *true_reg,
5379 struct bpf_reg_state *false_reg, u64 val,
5380 u8 opcode, bool is_jmp32)
5384 /* If the dst_reg is a pointer, we can't learn anything about its
5385 * variable offset from the compare (unless src_reg were a pointer into
5386 * the same object, but we don't bother with that.
5387 * Since false_reg and true_reg have the same type by construction, we
5388 * only need to check one of them for pointerness.
5390 if (__is_pointer_value(false, false_reg))
5393 val = is_jmp32 ? (u32)val : val;
5394 sval = is_jmp32 ? (s64)(s32)val : (s64)val;
5400 struct bpf_reg_state *reg =
5401 opcode == BPF_JEQ ? true_reg : false_reg;
5403 /* For BPF_JEQ, if this is false we know nothing Jon Snow, but
5404 * if it is true we know the value for sure. Likewise for
5408 u64 old_v = reg->var_off.value;
5409 u64 hi_mask = ~0xffffffffULL;
5411 reg->var_off.value = (old_v & hi_mask) | val;
5412 reg->var_off.mask &= hi_mask;
5414 __mark_reg_known(reg, val);
5419 false_reg->var_off = tnum_and(false_reg->var_off,
5421 if (is_power_of_2(val))
5422 true_reg->var_off = tnum_or(true_reg->var_off,
5428 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
5429 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
5432 false_umax += gen_hi_max(false_reg->var_off);
5433 true_umin += gen_hi_min(true_reg->var_off);
5435 false_reg->umax_value = min(false_reg->umax_value, false_umax);
5436 true_reg->umin_value = max(true_reg->umin_value, true_umin);
5442 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
5443 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
5445 /* If the full s64 was not sign-extended from s32 then don't
5446 * deduct further info.
5448 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
5450 false_reg->smax_value = min(false_reg->smax_value, false_smax);
5451 true_reg->smin_value = max(true_reg->smin_value, true_smin);
5457 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
5458 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
5461 false_umin += gen_hi_min(false_reg->var_off);
5462 true_umax += gen_hi_max(true_reg->var_off);
5464 false_reg->umin_value = max(false_reg->umin_value, false_umin);
5465 true_reg->umax_value = min(true_reg->umax_value, true_umax);
5471 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
5472 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
5474 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
5476 false_reg->smin_value = max(false_reg->smin_value, false_smin);
5477 true_reg->smax_value = min(true_reg->smax_value, true_smax);
5484 __reg_deduce_bounds(false_reg);
5485 __reg_deduce_bounds(true_reg);
5486 /* We might have learned some bits from the bounds. */
5487 __reg_bound_offset(false_reg);
5488 __reg_bound_offset(true_reg);
5489 /* Intersecting with the old var_off might have improved our bounds
5490 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
5491 * then new var_off is (0; 0x7f...fc) which improves our umax.
5493 __update_reg_bounds(false_reg);
5494 __update_reg_bounds(true_reg);
5497 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
5500 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
5501 struct bpf_reg_state *false_reg, u64 val,
5502 u8 opcode, bool is_jmp32)
5506 if (__is_pointer_value(false, false_reg))
5509 val = is_jmp32 ? (u32)val : val;
5510 sval = is_jmp32 ? (s64)(s32)val : (s64)val;
5516 struct bpf_reg_state *reg =
5517 opcode == BPF_JEQ ? true_reg : false_reg;
5520 u64 old_v = reg->var_off.value;
5521 u64 hi_mask = ~0xffffffffULL;
5523 reg->var_off.value = (old_v & hi_mask) | val;
5524 reg->var_off.mask &= hi_mask;
5526 __mark_reg_known(reg, val);
5531 false_reg->var_off = tnum_and(false_reg->var_off,
5533 if (is_power_of_2(val))
5534 true_reg->var_off = tnum_or(true_reg->var_off,
5540 u64 false_umin = opcode == BPF_JGT ? val : val + 1;
5541 u64 true_umax = opcode == BPF_JGT ? val - 1 : val;
5544 false_umin += gen_hi_min(false_reg->var_off);
5545 true_umax += gen_hi_max(true_reg->var_off);
5547 false_reg->umin_value = max(false_reg->umin_value, false_umin);
5548 true_reg->umax_value = min(true_reg->umax_value, true_umax);
5554 s64 false_smin = opcode == BPF_JSGT ? sval : sval + 1;
5555 s64 true_smax = opcode == BPF_JSGT ? sval - 1 : sval;
5557 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
5559 false_reg->smin_value = max(false_reg->smin_value, false_smin);
5560 true_reg->smax_value = min(true_reg->smax_value, true_smax);
5566 u64 false_umax = opcode == BPF_JLT ? val : val - 1;
5567 u64 true_umin = opcode == BPF_JLT ? val + 1 : val;
5570 false_umax += gen_hi_max(false_reg->var_off);
5571 true_umin += gen_hi_min(true_reg->var_off);
5573 false_reg->umax_value = min(false_reg->umax_value, false_umax);
5574 true_reg->umin_value = max(true_reg->umin_value, true_umin);
5580 s64 false_smax = opcode == BPF_JSLT ? sval : sval - 1;
5581 s64 true_smin = opcode == BPF_JSLT ? sval + 1 : sval;
5583 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
5585 false_reg->smax_value = min(false_reg->smax_value, false_smax);
5586 true_reg->smin_value = max(true_reg->smin_value, true_smin);
5593 __reg_deduce_bounds(false_reg);
5594 __reg_deduce_bounds(true_reg);
5595 /* We might have learned some bits from the bounds. */
5596 __reg_bound_offset(false_reg);
5597 __reg_bound_offset(true_reg);
5598 /* Intersecting with the old var_off might have improved our bounds
5599 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
5600 * then new var_off is (0; 0x7f...fc) which improves our umax.
5602 __update_reg_bounds(false_reg);
5603 __update_reg_bounds(true_reg);
5606 /* Regs are known to be equal, so intersect their min/max/var_off */
5607 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
5608 struct bpf_reg_state *dst_reg)
5610 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
5611 dst_reg->umin_value);
5612 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
5613 dst_reg->umax_value);
5614 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
5615 dst_reg->smin_value);
5616 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
5617 dst_reg->smax_value);
5618 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
5620 /* We might have learned new bounds from the var_off. */
5621 __update_reg_bounds(src_reg);
5622 __update_reg_bounds(dst_reg);
5623 /* We might have learned something about the sign bit. */
5624 __reg_deduce_bounds(src_reg);
5625 __reg_deduce_bounds(dst_reg);
5626 /* We might have learned some bits from the bounds. */
5627 __reg_bound_offset(src_reg);
5628 __reg_bound_offset(dst_reg);
5629 /* Intersecting with the old var_off might have improved our bounds
5630 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
5631 * then new var_off is (0; 0x7f...fc) which improves our umax.
5633 __update_reg_bounds(src_reg);
5634 __update_reg_bounds(dst_reg);
5637 static void reg_combine_min_max(struct bpf_reg_state *true_src,
5638 struct bpf_reg_state *true_dst,
5639 struct bpf_reg_state *false_src,
5640 struct bpf_reg_state *false_dst,
5645 __reg_combine_min_max(true_src, true_dst);
5648 __reg_combine_min_max(false_src, false_dst);
5653 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
5654 struct bpf_reg_state *reg, u32 id,
5657 if (reg_type_may_be_null(reg->type) && reg->id == id) {
5658 /* Old offset (both fixed and variable parts) should
5659 * have been known-zero, because we don't allow pointer
5660 * arithmetic on pointers that might be NULL.
5662 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
5663 !tnum_equals_const(reg->var_off, 0) ||
5665 __mark_reg_known_zero(reg);
5669 reg->type = SCALAR_VALUE;
5670 } else if (reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
5671 if (reg->map_ptr->inner_map_meta) {
5672 reg->type = CONST_PTR_TO_MAP;
5673 reg->map_ptr = reg->map_ptr->inner_map_meta;
5674 } else if (reg->map_ptr->map_type ==
5675 BPF_MAP_TYPE_XSKMAP) {
5676 reg->type = PTR_TO_XDP_SOCK;
5678 reg->type = PTR_TO_MAP_VALUE;
5680 } else if (reg->type == PTR_TO_SOCKET_OR_NULL) {
5681 reg->type = PTR_TO_SOCKET;
5682 } else if (reg->type == PTR_TO_SOCK_COMMON_OR_NULL) {
5683 reg->type = PTR_TO_SOCK_COMMON;
5684 } else if (reg->type == PTR_TO_TCP_SOCK_OR_NULL) {
5685 reg->type = PTR_TO_TCP_SOCK;
5688 /* We don't need id and ref_obj_id from this point
5689 * onwards anymore, thus we should better reset it,
5690 * so that state pruning has chances to take effect.
5693 reg->ref_obj_id = 0;
5694 } else if (!reg_may_point_to_spin_lock(reg)) {
5695 /* For not-NULL ptr, reg->ref_obj_id will be reset
5696 * in release_reg_references().
5698 * reg->id is still used by spin_lock ptr. Other
5699 * than spin_lock ptr type, reg->id can be reset.
5706 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
5709 struct bpf_reg_state *reg;
5712 for (i = 0; i < MAX_BPF_REG; i++)
5713 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
5715 bpf_for_each_spilled_reg(i, state, reg) {
5718 mark_ptr_or_null_reg(state, reg, id, is_null);
5722 /* The logic is similar to find_good_pkt_pointers(), both could eventually
5723 * be folded together at some point.
5725 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
5728 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5729 struct bpf_reg_state *regs = state->regs;
5730 u32 ref_obj_id = regs[regno].ref_obj_id;
5731 u32 id = regs[regno].id;
5734 if (ref_obj_id && ref_obj_id == id && is_null)
5735 /* regs[regno] is in the " == NULL" branch.
5736 * No one could have freed the reference state before
5737 * doing the NULL check.
5739 WARN_ON_ONCE(release_reference_state(state, id));
5741 for (i = 0; i <= vstate->curframe; i++)
5742 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
5745 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
5746 struct bpf_reg_state *dst_reg,
5747 struct bpf_reg_state *src_reg,
5748 struct bpf_verifier_state *this_branch,
5749 struct bpf_verifier_state *other_branch)
5751 if (BPF_SRC(insn->code) != BPF_X)
5754 /* Pointers are always 64-bit. */
5755 if (BPF_CLASS(insn->code) == BPF_JMP32)
5758 switch (BPF_OP(insn->code)) {
5760 if ((dst_reg->type == PTR_TO_PACKET &&
5761 src_reg->type == PTR_TO_PACKET_END) ||
5762 (dst_reg->type == PTR_TO_PACKET_META &&
5763 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
5764 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
5765 find_good_pkt_pointers(this_branch, dst_reg,
5766 dst_reg->type, false);
5767 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
5768 src_reg->type == PTR_TO_PACKET) ||
5769 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
5770 src_reg->type == PTR_TO_PACKET_META)) {
5771 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
5772 find_good_pkt_pointers(other_branch, src_reg,
5773 src_reg->type, true);
5779 if ((dst_reg->type == PTR_TO_PACKET &&
5780 src_reg->type == PTR_TO_PACKET_END) ||
5781 (dst_reg->type == PTR_TO_PACKET_META &&
5782 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
5783 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
5784 find_good_pkt_pointers(other_branch, dst_reg,
5785 dst_reg->type, true);
5786 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
5787 src_reg->type == PTR_TO_PACKET) ||
5788 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
5789 src_reg->type == PTR_TO_PACKET_META)) {
5790 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
5791 find_good_pkt_pointers(this_branch, src_reg,
5792 src_reg->type, false);
5798 if ((dst_reg->type == PTR_TO_PACKET &&
5799 src_reg->type == PTR_TO_PACKET_END) ||
5800 (dst_reg->type == PTR_TO_PACKET_META &&
5801 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
5802 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
5803 find_good_pkt_pointers(this_branch, dst_reg,
5804 dst_reg->type, true);
5805 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
5806 src_reg->type == PTR_TO_PACKET) ||
5807 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
5808 src_reg->type == PTR_TO_PACKET_META)) {
5809 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
5810 find_good_pkt_pointers(other_branch, src_reg,
5811 src_reg->type, false);
5817 if ((dst_reg->type == PTR_TO_PACKET &&
5818 src_reg->type == PTR_TO_PACKET_END) ||
5819 (dst_reg->type == PTR_TO_PACKET_META &&
5820 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
5821 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
5822 find_good_pkt_pointers(other_branch, dst_reg,
5823 dst_reg->type, false);
5824 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
5825 src_reg->type == PTR_TO_PACKET) ||
5826 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
5827 src_reg->type == PTR_TO_PACKET_META)) {
5828 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
5829 find_good_pkt_pointers(this_branch, src_reg,
5830 src_reg->type, true);
5842 static int check_cond_jmp_op(struct bpf_verifier_env *env,
5843 struct bpf_insn *insn, int *insn_idx)
5845 struct bpf_verifier_state *this_branch = env->cur_state;
5846 struct bpf_verifier_state *other_branch;
5847 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
5848 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
5849 u8 opcode = BPF_OP(insn->code);
5854 /* Only conditional jumps are expected to reach here. */
5855 if (opcode == BPF_JA || opcode > BPF_JSLE) {
5856 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
5860 if (BPF_SRC(insn->code) == BPF_X) {
5861 if (insn->imm != 0) {
5862 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
5866 /* check src1 operand */
5867 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5871 if (is_pointer_value(env, insn->src_reg)) {
5872 verbose(env, "R%d pointer comparison prohibited\n",
5876 src_reg = ®s[insn->src_reg];
5878 if (insn->src_reg != BPF_REG_0) {
5879 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
5884 /* check src2 operand */
5885 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5889 dst_reg = ®s[insn->dst_reg];
5890 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
5892 if (BPF_SRC(insn->code) == BPF_K)
5893 pred = is_branch_taken(dst_reg, insn->imm,
5895 else if (src_reg->type == SCALAR_VALUE &&
5896 tnum_is_const(src_reg->var_off))
5897 pred = is_branch_taken(dst_reg, src_reg->var_off.value,
5900 err = mark_chain_precision(env, insn->dst_reg);
5901 if (BPF_SRC(insn->code) == BPF_X && !err)
5902 err = mark_chain_precision(env, insn->src_reg);
5907 /* only follow the goto, ignore fall-through */
5908 *insn_idx += insn->off;
5910 } else if (pred == 0) {
5911 /* only follow fall-through branch, since
5912 * that's where the program will go
5917 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
5921 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
5923 /* detect if we are comparing against a constant value so we can adjust
5924 * our min/max values for our dst register.
5925 * this is only legit if both are scalars (or pointers to the same
5926 * object, I suppose, but we don't support that right now), because
5927 * otherwise the different base pointers mean the offsets aren't
5930 if (BPF_SRC(insn->code) == BPF_X) {
5931 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
5932 struct bpf_reg_state lo_reg0 = *dst_reg;
5933 struct bpf_reg_state lo_reg1 = *src_reg;
5934 struct bpf_reg_state *src_lo, *dst_lo;
5938 coerce_reg_to_size(dst_lo, 4);
5939 coerce_reg_to_size(src_lo, 4);
5941 if (dst_reg->type == SCALAR_VALUE &&
5942 src_reg->type == SCALAR_VALUE) {
5943 if (tnum_is_const(src_reg->var_off) ||
5944 (is_jmp32 && tnum_is_const(src_lo->var_off)))
5945 reg_set_min_max(&other_branch_regs[insn->dst_reg],
5948 ? src_lo->var_off.value
5949 : src_reg->var_off.value,
5951 else if (tnum_is_const(dst_reg->var_off) ||
5952 (is_jmp32 && tnum_is_const(dst_lo->var_off)))
5953 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
5956 ? dst_lo->var_off.value
5957 : dst_reg->var_off.value,
5959 else if (!is_jmp32 &&
5960 (opcode == BPF_JEQ || opcode == BPF_JNE))
5961 /* Comparing for equality, we can combine knowledge */
5962 reg_combine_min_max(&other_branch_regs[insn->src_reg],
5963 &other_branch_regs[insn->dst_reg],
5964 src_reg, dst_reg, opcode);
5966 } else if (dst_reg->type == SCALAR_VALUE) {
5967 reg_set_min_max(&other_branch_regs[insn->dst_reg],
5968 dst_reg, insn->imm, opcode, is_jmp32);
5971 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
5972 * NOTE: these optimizations below are related with pointer comparison
5973 * which will never be JMP32.
5975 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
5976 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
5977 reg_type_may_be_null(dst_reg->type)) {
5978 /* Mark all identical registers in each branch as either
5979 * safe or unknown depending R == 0 or R != 0 conditional.
5981 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
5983 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
5985 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
5986 this_branch, other_branch) &&
5987 is_pointer_value(env, insn->dst_reg)) {
5988 verbose(env, "R%d pointer comparison prohibited\n",
5992 if (env->log.level & BPF_LOG_LEVEL)
5993 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
5997 /* verify BPF_LD_IMM64 instruction */
5998 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
6000 struct bpf_insn_aux_data *aux = cur_aux(env);
6001 struct bpf_reg_state *regs = cur_regs(env);
6002 struct bpf_map *map;
6005 if (BPF_SIZE(insn->code) != BPF_DW) {
6006 verbose(env, "invalid BPF_LD_IMM insn\n");
6009 if (insn->off != 0) {
6010 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
6014 err = check_reg_arg(env, insn->dst_reg, DST_OP);
6018 if (insn->src_reg == 0) {
6019 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
6021 regs[insn->dst_reg].type = SCALAR_VALUE;
6022 __mark_reg_known(®s[insn->dst_reg], imm);
6026 map = env->used_maps[aux->map_index];
6027 mark_reg_known_zero(env, regs, insn->dst_reg);
6028 regs[insn->dst_reg].map_ptr = map;
6030 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE) {
6031 regs[insn->dst_reg].type = PTR_TO_MAP_VALUE;
6032 regs[insn->dst_reg].off = aux->map_off;
6033 if (map_value_has_spin_lock(map))
6034 regs[insn->dst_reg].id = ++env->id_gen;
6035 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
6036 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
6038 verbose(env, "bpf verifier is misconfigured\n");
6045 static bool may_access_skb(enum bpf_prog_type type)
6048 case BPF_PROG_TYPE_SOCKET_FILTER:
6049 case BPF_PROG_TYPE_SCHED_CLS:
6050 case BPF_PROG_TYPE_SCHED_ACT:
6057 /* verify safety of LD_ABS|LD_IND instructions:
6058 * - they can only appear in the programs where ctx == skb
6059 * - since they are wrappers of function calls, they scratch R1-R5 registers,
6060 * preserve R6-R9, and store return value into R0
6063 * ctx == skb == R6 == CTX
6066 * SRC == any register
6067 * IMM == 32-bit immediate
6070 * R0 - 8/16/32-bit skb data converted to cpu endianness
6072 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
6074 struct bpf_reg_state *regs = cur_regs(env);
6075 u8 mode = BPF_MODE(insn->code);
6078 if (!may_access_skb(env->prog->type)) {
6079 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
6083 if (!env->ops->gen_ld_abs) {
6084 verbose(env, "bpf verifier is misconfigured\n");
6088 if (env->subprog_cnt > 1) {
6089 /* when program has LD_ABS insn JITs and interpreter assume
6090 * that r1 == ctx == skb which is not the case for callees
6091 * that can have arbitrary arguments. It's problematic
6092 * for main prog as well since JITs would need to analyze
6093 * all functions in order to make proper register save/restore
6094 * decisions in the main prog. Hence disallow LD_ABS with calls
6096 verbose(env, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
6100 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
6101 BPF_SIZE(insn->code) == BPF_DW ||
6102 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
6103 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
6107 /* check whether implicit source operand (register R6) is readable */
6108 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
6112 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
6113 * gen_ld_abs() may terminate the program at runtime, leading to
6116 err = check_reference_leak(env);
6118 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
6122 if (env->cur_state->active_spin_lock) {
6123 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
6127 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
6129 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
6133 if (mode == BPF_IND) {
6134 /* check explicit source operand */
6135 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6140 /* reset caller saved regs to unreadable */
6141 for (i = 0; i < CALLER_SAVED_REGS; i++) {
6142 mark_reg_not_init(env, regs, caller_saved[i]);
6143 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
6146 /* mark destination R0 register as readable, since it contains
6147 * the value fetched from the packet.
6148 * Already marked as written above.
6150 mark_reg_unknown(env, regs, BPF_REG_0);
6151 /* ld_abs load up to 32-bit skb data. */
6152 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
6156 static int check_return_code(struct bpf_verifier_env *env)
6158 struct tnum enforce_attach_type_range = tnum_unknown;
6159 struct bpf_reg_state *reg;
6160 struct tnum range = tnum_range(0, 1);
6162 switch (env->prog->type) {
6163 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
6164 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
6165 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG)
6166 range = tnum_range(1, 1);
6168 case BPF_PROG_TYPE_CGROUP_SKB:
6169 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
6170 range = tnum_range(0, 3);
6171 enforce_attach_type_range = tnum_range(2, 3);
6174 case BPF_PROG_TYPE_CGROUP_SOCK:
6175 case BPF_PROG_TYPE_SOCK_OPS:
6176 case BPF_PROG_TYPE_CGROUP_DEVICE:
6177 case BPF_PROG_TYPE_CGROUP_SYSCTL:
6178 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
6184 reg = cur_regs(env) + BPF_REG_0;
6185 if (reg->type != SCALAR_VALUE) {
6186 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
6187 reg_type_str[reg->type]);
6191 if (!tnum_in(range, reg->var_off)) {
6194 verbose(env, "At program exit the register R0 ");
6195 if (!tnum_is_unknown(reg->var_off)) {
6196 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6197 verbose(env, "has value %s", tn_buf);
6199 verbose(env, "has unknown scalar value");
6201 tnum_strn(tn_buf, sizeof(tn_buf), range);
6202 verbose(env, " should have been in %s\n", tn_buf);
6206 if (!tnum_is_unknown(enforce_attach_type_range) &&
6207 tnum_in(enforce_attach_type_range, reg->var_off))
6208 env->prog->enforce_expected_attach_type = 1;
6212 /* non-recursive DFS pseudo code
6213 * 1 procedure DFS-iterative(G,v):
6214 * 2 label v as discovered
6215 * 3 let S be a stack
6217 * 5 while S is not empty
6219 * 7 if t is what we're looking for:
6221 * 9 for all edges e in G.adjacentEdges(t) do
6222 * 10 if edge e is already labelled
6223 * 11 continue with the next edge
6224 * 12 w <- G.adjacentVertex(t,e)
6225 * 13 if vertex w is not discovered and not explored
6226 * 14 label e as tree-edge
6227 * 15 label w as discovered
6230 * 18 else if vertex w is discovered
6231 * 19 label e as back-edge
6233 * 21 // vertex w is explored
6234 * 22 label e as forward- or cross-edge
6235 * 23 label t as explored
6240 * 0x11 - discovered and fall-through edge labelled
6241 * 0x12 - discovered and fall-through and branch edges labelled
6252 static u32 state_htab_size(struct bpf_verifier_env *env)
6254 return env->prog->len;
6257 static struct bpf_verifier_state_list **explored_state(
6258 struct bpf_verifier_env *env,
6261 struct bpf_verifier_state *cur = env->cur_state;
6262 struct bpf_func_state *state = cur->frame[cur->curframe];
6264 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
6267 static void init_explored_state(struct bpf_verifier_env *env, int idx)
6269 env->insn_aux_data[idx].prune_point = true;
6272 /* t, w, e - match pseudo-code above:
6273 * t - index of current instruction
6274 * w - next instruction
6277 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
6280 int *insn_stack = env->cfg.insn_stack;
6281 int *insn_state = env->cfg.insn_state;
6283 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
6286 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
6289 if (w < 0 || w >= env->prog->len) {
6290 verbose_linfo(env, t, "%d: ", t);
6291 verbose(env, "jump out of range from insn %d to %d\n", t, w);
6296 /* mark branch target for state pruning */
6297 init_explored_state(env, w);
6299 if (insn_state[w] == 0) {
6301 insn_state[t] = DISCOVERED | e;
6302 insn_state[w] = DISCOVERED;
6303 if (env->cfg.cur_stack >= env->prog->len)
6305 insn_stack[env->cfg.cur_stack++] = w;
6307 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
6308 if (loop_ok && env->allow_ptr_leaks)
6310 verbose_linfo(env, t, "%d: ", t);
6311 verbose_linfo(env, w, "%d: ", w);
6312 verbose(env, "back-edge from insn %d to %d\n", t, w);
6314 } else if (insn_state[w] == EXPLORED) {
6315 /* forward- or cross-edge */
6316 insn_state[t] = DISCOVERED | e;
6318 verbose(env, "insn state internal bug\n");
6324 /* non-recursive depth-first-search to detect loops in BPF program
6325 * loop == back-edge in directed graph
6327 static int check_cfg(struct bpf_verifier_env *env)
6329 struct bpf_insn *insns = env->prog->insnsi;
6330 int insn_cnt = env->prog->len;
6331 int *insn_stack, *insn_state;
6335 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
6339 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
6345 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
6346 insn_stack[0] = 0; /* 0 is the first instruction */
6347 env->cfg.cur_stack = 1;
6350 if (env->cfg.cur_stack == 0)
6352 t = insn_stack[env->cfg.cur_stack - 1];
6354 if (BPF_CLASS(insns[t].code) == BPF_JMP ||
6355 BPF_CLASS(insns[t].code) == BPF_JMP32) {
6356 u8 opcode = BPF_OP(insns[t].code);
6358 if (opcode == BPF_EXIT) {
6360 } else if (opcode == BPF_CALL) {
6361 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
6366 if (t + 1 < insn_cnt)
6367 init_explored_state(env, t + 1);
6368 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
6369 init_explored_state(env, t);
6370 ret = push_insn(t, t + insns[t].imm + 1, BRANCH,
6377 } else if (opcode == BPF_JA) {
6378 if (BPF_SRC(insns[t].code) != BPF_K) {
6382 /* unconditional jump with single edge */
6383 ret = push_insn(t, t + insns[t].off + 1,
6384 FALLTHROUGH, env, true);
6389 /* unconditional jmp is not a good pruning point,
6390 * but it's marked, since backtracking needs
6391 * to record jmp history in is_state_visited().
6393 init_explored_state(env, t + insns[t].off + 1);
6394 /* tell verifier to check for equivalent states
6395 * after every call and jump
6397 if (t + 1 < insn_cnt)
6398 init_explored_state(env, t + 1);
6400 /* conditional jump with two edges */
6401 init_explored_state(env, t);
6402 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
6408 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
6415 /* all other non-branch instructions with single
6418 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
6426 insn_state[t] = EXPLORED;
6427 if (env->cfg.cur_stack-- <= 0) {
6428 verbose(env, "pop stack internal bug\n");
6435 for (i = 0; i < insn_cnt; i++) {
6436 if (insn_state[i] != EXPLORED) {
6437 verbose(env, "unreachable insn %d\n", i);
6442 ret = 0; /* cfg looks good */
6447 env->cfg.insn_state = env->cfg.insn_stack = NULL;
6451 /* The minimum supported BTF func info size */
6452 #define MIN_BPF_FUNCINFO_SIZE 8
6453 #define MAX_FUNCINFO_REC_SIZE 252
6455 static int check_btf_func(struct bpf_verifier_env *env,
6456 const union bpf_attr *attr,
6457 union bpf_attr __user *uattr)
6459 u32 i, nfuncs, urec_size, min_size;
6460 u32 krec_size = sizeof(struct bpf_func_info);
6461 struct bpf_func_info *krecord;
6462 const struct btf_type *type;
6463 struct bpf_prog *prog;
6464 const struct btf *btf;
6465 void __user *urecord;
6466 u32 prev_offset = 0;
6469 nfuncs = attr->func_info_cnt;
6473 if (nfuncs != env->subprog_cnt) {
6474 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
6478 urec_size = attr->func_info_rec_size;
6479 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
6480 urec_size > MAX_FUNCINFO_REC_SIZE ||
6481 urec_size % sizeof(u32)) {
6482 verbose(env, "invalid func info rec size %u\n", urec_size);
6487 btf = prog->aux->btf;
6489 urecord = u64_to_user_ptr(attr->func_info);
6490 min_size = min_t(u32, krec_size, urec_size);
6492 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
6496 for (i = 0; i < nfuncs; i++) {
6497 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
6499 if (ret == -E2BIG) {
6500 verbose(env, "nonzero tailing record in func info");
6501 /* set the size kernel expects so loader can zero
6502 * out the rest of the record.
6504 if (put_user(min_size, &uattr->func_info_rec_size))
6510 if (copy_from_user(&krecord[i], urecord, min_size)) {
6515 /* check insn_off */
6517 if (krecord[i].insn_off) {
6519 "nonzero insn_off %u for the first func info record",
6520 krecord[i].insn_off);
6524 } else if (krecord[i].insn_off <= prev_offset) {
6526 "same or smaller insn offset (%u) than previous func info record (%u)",
6527 krecord[i].insn_off, prev_offset);
6532 if (env->subprog_info[i].start != krecord[i].insn_off) {
6533 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
6539 type = btf_type_by_id(btf, krecord[i].type_id);
6540 if (!type || BTF_INFO_KIND(type->info) != BTF_KIND_FUNC) {
6541 verbose(env, "invalid type id %d in func info",
6542 krecord[i].type_id);
6547 prev_offset = krecord[i].insn_off;
6548 urecord += urec_size;
6551 prog->aux->func_info = krecord;
6552 prog->aux->func_info_cnt = nfuncs;
6560 static void adjust_btf_func(struct bpf_verifier_env *env)
6564 if (!env->prog->aux->func_info)
6567 for (i = 0; i < env->subprog_cnt; i++)
6568 env->prog->aux->func_info[i].insn_off = env->subprog_info[i].start;
6571 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
6572 sizeof(((struct bpf_line_info *)(0))->line_col))
6573 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
6575 static int check_btf_line(struct bpf_verifier_env *env,
6576 const union bpf_attr *attr,
6577 union bpf_attr __user *uattr)
6579 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
6580 struct bpf_subprog_info *sub;
6581 struct bpf_line_info *linfo;
6582 struct bpf_prog *prog;
6583 const struct btf *btf;
6584 void __user *ulinfo;
6587 nr_linfo = attr->line_info_cnt;
6591 rec_size = attr->line_info_rec_size;
6592 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
6593 rec_size > MAX_LINEINFO_REC_SIZE ||
6594 rec_size & (sizeof(u32) - 1))
6597 /* Need to zero it in case the userspace may
6598 * pass in a smaller bpf_line_info object.
6600 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
6601 GFP_KERNEL | __GFP_NOWARN);
6606 btf = prog->aux->btf;
6609 sub = env->subprog_info;
6610 ulinfo = u64_to_user_ptr(attr->line_info);
6611 expected_size = sizeof(struct bpf_line_info);
6612 ncopy = min_t(u32, expected_size, rec_size);
6613 for (i = 0; i < nr_linfo; i++) {
6614 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
6616 if (err == -E2BIG) {
6617 verbose(env, "nonzero tailing record in line_info");
6618 if (put_user(expected_size,
6619 &uattr->line_info_rec_size))
6625 if (copy_from_user(&linfo[i], ulinfo, ncopy)) {
6631 * Check insn_off to ensure
6632 * 1) strictly increasing AND
6633 * 2) bounded by prog->len
6635 * The linfo[0].insn_off == 0 check logically falls into
6636 * the later "missing bpf_line_info for func..." case
6637 * because the first linfo[0].insn_off must be the
6638 * first sub also and the first sub must have
6639 * subprog_info[0].start == 0.
6641 if ((i && linfo[i].insn_off <= prev_offset) ||
6642 linfo[i].insn_off >= prog->len) {
6643 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
6644 i, linfo[i].insn_off, prev_offset,
6650 if (!prog->insnsi[linfo[i].insn_off].code) {
6652 "Invalid insn code at line_info[%u].insn_off\n",
6658 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
6659 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
6660 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
6665 if (s != env->subprog_cnt) {
6666 if (linfo[i].insn_off == sub[s].start) {
6667 sub[s].linfo_idx = i;
6669 } else if (sub[s].start < linfo[i].insn_off) {
6670 verbose(env, "missing bpf_line_info for func#%u\n", s);
6676 prev_offset = linfo[i].insn_off;
6680 if (s != env->subprog_cnt) {
6681 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
6682 env->subprog_cnt - s, s);
6687 prog->aux->linfo = linfo;
6688 prog->aux->nr_linfo = nr_linfo;
6697 static int check_btf_info(struct bpf_verifier_env *env,
6698 const union bpf_attr *attr,
6699 union bpf_attr __user *uattr)
6704 if (!attr->func_info_cnt && !attr->line_info_cnt)
6707 btf = btf_get_by_fd(attr->prog_btf_fd);
6709 return PTR_ERR(btf);
6710 env->prog->aux->btf = btf;
6712 err = check_btf_func(env, attr, uattr);
6716 err = check_btf_line(env, attr, uattr);
6723 /* check %cur's range satisfies %old's */
6724 static bool range_within(struct bpf_reg_state *old,
6725 struct bpf_reg_state *cur)
6727 return old->umin_value <= cur->umin_value &&
6728 old->umax_value >= cur->umax_value &&
6729 old->smin_value <= cur->smin_value &&
6730 old->smax_value >= cur->smax_value;
6733 /* Maximum number of register states that can exist at once */
6734 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
6740 /* If in the old state two registers had the same id, then they need to have
6741 * the same id in the new state as well. But that id could be different from
6742 * the old state, so we need to track the mapping from old to new ids.
6743 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
6744 * regs with old id 5 must also have new id 9 for the new state to be safe. But
6745 * regs with a different old id could still have new id 9, we don't care about
6747 * So we look through our idmap to see if this old id has been seen before. If
6748 * so, we require the new id to match; otherwise, we add the id pair to the map.
6750 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
6754 for (i = 0; i < ID_MAP_SIZE; i++) {
6755 if (!idmap[i].old) {
6756 /* Reached an empty slot; haven't seen this id before */
6757 idmap[i].old = old_id;
6758 idmap[i].cur = cur_id;
6761 if (idmap[i].old == old_id)
6762 return idmap[i].cur == cur_id;
6764 /* We ran out of idmap slots, which should be impossible */
6769 static void clean_func_state(struct bpf_verifier_env *env,
6770 struct bpf_func_state *st)
6772 enum bpf_reg_liveness live;
6775 for (i = 0; i < BPF_REG_FP; i++) {
6776 live = st->regs[i].live;
6777 /* liveness must not touch this register anymore */
6778 st->regs[i].live |= REG_LIVE_DONE;
6779 if (!(live & REG_LIVE_READ))
6780 /* since the register is unused, clear its state
6781 * to make further comparison simpler
6783 __mark_reg_not_init(&st->regs[i]);
6786 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
6787 live = st->stack[i].spilled_ptr.live;
6788 /* liveness must not touch this stack slot anymore */
6789 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
6790 if (!(live & REG_LIVE_READ)) {
6791 __mark_reg_not_init(&st->stack[i].spilled_ptr);
6792 for (j = 0; j < BPF_REG_SIZE; j++)
6793 st->stack[i].slot_type[j] = STACK_INVALID;
6798 static void clean_verifier_state(struct bpf_verifier_env *env,
6799 struct bpf_verifier_state *st)
6803 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
6804 /* all regs in this state in all frames were already marked */
6807 for (i = 0; i <= st->curframe; i++)
6808 clean_func_state(env, st->frame[i]);
6811 /* the parentage chains form a tree.
6812 * the verifier states are added to state lists at given insn and
6813 * pushed into state stack for future exploration.
6814 * when the verifier reaches bpf_exit insn some of the verifer states
6815 * stored in the state lists have their final liveness state already,
6816 * but a lot of states will get revised from liveness point of view when
6817 * the verifier explores other branches.
6820 * 2: if r1 == 100 goto pc+1
6823 * when the verifier reaches exit insn the register r0 in the state list of
6824 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
6825 * of insn 2 and goes exploring further. At the insn 4 it will walk the
6826 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
6828 * Since the verifier pushes the branch states as it sees them while exploring
6829 * the program the condition of walking the branch instruction for the second
6830 * time means that all states below this branch were already explored and
6831 * their final liveness markes are already propagated.
6832 * Hence when the verifier completes the search of state list in is_state_visited()
6833 * we can call this clean_live_states() function to mark all liveness states
6834 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
6836 * This function also clears the registers and stack for states that !READ
6837 * to simplify state merging.
6839 * Important note here that walking the same branch instruction in the callee
6840 * doesn't meant that the states are DONE. The verifier has to compare
6843 static void clean_live_states(struct bpf_verifier_env *env, int insn,
6844 struct bpf_verifier_state *cur)
6846 struct bpf_verifier_state_list *sl;
6849 sl = *explored_state(env, insn);
6851 if (sl->state.branches)
6853 if (sl->state.insn_idx != insn ||
6854 sl->state.curframe != cur->curframe)
6856 for (i = 0; i <= cur->curframe; i++)
6857 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
6859 clean_verifier_state(env, &sl->state);
6865 /* Returns true if (rold safe implies rcur safe) */
6866 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
6867 struct idpair *idmap)
6871 if (!(rold->live & REG_LIVE_READ))
6872 /* explored state didn't use this */
6875 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
6877 if (rold->type == PTR_TO_STACK)
6878 /* two stack pointers are equal only if they're pointing to
6879 * the same stack frame, since fp-8 in foo != fp-8 in bar
6881 return equal && rold->frameno == rcur->frameno;
6886 if (rold->type == NOT_INIT)
6887 /* explored state can't have used this */
6889 if (rcur->type == NOT_INIT)
6891 switch (rold->type) {
6893 if (rcur->type == SCALAR_VALUE) {
6894 if (!rold->precise && !rcur->precise)
6896 /* new val must satisfy old val knowledge */
6897 return range_within(rold, rcur) &&
6898 tnum_in(rold->var_off, rcur->var_off);
6900 /* We're trying to use a pointer in place of a scalar.
6901 * Even if the scalar was unbounded, this could lead to
6902 * pointer leaks because scalars are allowed to leak
6903 * while pointers are not. We could make this safe in
6904 * special cases if root is calling us, but it's
6905 * probably not worth the hassle.
6909 case PTR_TO_MAP_VALUE:
6910 /* If the new min/max/var_off satisfy the old ones and
6911 * everything else matches, we are OK.
6912 * 'id' is not compared, since it's only used for maps with
6913 * bpf_spin_lock inside map element and in such cases if
6914 * the rest of the prog is valid for one map element then
6915 * it's valid for all map elements regardless of the key
6916 * used in bpf_map_lookup()
6918 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
6919 range_within(rold, rcur) &&
6920 tnum_in(rold->var_off, rcur->var_off);
6921 case PTR_TO_MAP_VALUE_OR_NULL:
6922 /* a PTR_TO_MAP_VALUE could be safe to use as a
6923 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
6924 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
6925 * checked, doing so could have affected others with the same
6926 * id, and we can't check for that because we lost the id when
6927 * we converted to a PTR_TO_MAP_VALUE.
6929 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
6931 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
6933 /* Check our ids match any regs they're supposed to */
6934 return check_ids(rold->id, rcur->id, idmap);
6935 case PTR_TO_PACKET_META:
6937 if (rcur->type != rold->type)
6939 /* We must have at least as much range as the old ptr
6940 * did, so that any accesses which were safe before are
6941 * still safe. This is true even if old range < old off,
6942 * since someone could have accessed through (ptr - k), or
6943 * even done ptr -= k in a register, to get a safe access.
6945 if (rold->range > rcur->range)
6947 /* If the offsets don't match, we can't trust our alignment;
6948 * nor can we be sure that we won't fall out of range.
6950 if (rold->off != rcur->off)
6952 /* id relations must be preserved */
6953 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
6955 /* new val must satisfy old val knowledge */
6956 return range_within(rold, rcur) &&
6957 tnum_in(rold->var_off, rcur->var_off);
6959 case CONST_PTR_TO_MAP:
6960 case PTR_TO_PACKET_END:
6961 case PTR_TO_FLOW_KEYS:
6963 case PTR_TO_SOCKET_OR_NULL:
6964 case PTR_TO_SOCK_COMMON:
6965 case PTR_TO_SOCK_COMMON_OR_NULL:
6966 case PTR_TO_TCP_SOCK:
6967 case PTR_TO_TCP_SOCK_OR_NULL:
6968 case PTR_TO_XDP_SOCK:
6969 /* Only valid matches are exact, which memcmp() above
6970 * would have accepted
6973 /* Don't know what's going on, just say it's not safe */
6977 /* Shouldn't get here; if we do, say it's not safe */
6982 static bool stacksafe(struct bpf_func_state *old,
6983 struct bpf_func_state *cur,
6984 struct idpair *idmap)
6988 /* walk slots of the explored stack and ignore any additional
6989 * slots in the current stack, since explored(safe) state
6992 for (i = 0; i < old->allocated_stack; i++) {
6993 spi = i / BPF_REG_SIZE;
6995 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
6996 i += BPF_REG_SIZE - 1;
6997 /* explored state didn't use this */
7001 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
7004 /* explored stack has more populated slots than current stack
7005 * and these slots were used
7007 if (i >= cur->allocated_stack)
7010 /* if old state was safe with misc data in the stack
7011 * it will be safe with zero-initialized stack.
7012 * The opposite is not true
7014 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
7015 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
7017 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
7018 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
7019 /* Ex: old explored (safe) state has STACK_SPILL in
7020 * this stack slot, but current has has STACK_MISC ->
7021 * this verifier states are not equivalent,
7022 * return false to continue verification of this path
7025 if (i % BPF_REG_SIZE)
7027 if (old->stack[spi].slot_type[0] != STACK_SPILL)
7029 if (!regsafe(&old->stack[spi].spilled_ptr,
7030 &cur->stack[spi].spilled_ptr,
7032 /* when explored and current stack slot are both storing
7033 * spilled registers, check that stored pointers types
7034 * are the same as well.
7035 * Ex: explored safe path could have stored
7036 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
7037 * but current path has stored:
7038 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
7039 * such verifier states are not equivalent.
7040 * return false to continue verification of this path
7047 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
7049 if (old->acquired_refs != cur->acquired_refs)
7051 return !memcmp(old->refs, cur->refs,
7052 sizeof(*old->refs) * old->acquired_refs);
7055 /* compare two verifier states
7057 * all states stored in state_list are known to be valid, since
7058 * verifier reached 'bpf_exit' instruction through them
7060 * this function is called when verifier exploring different branches of
7061 * execution popped from the state stack. If it sees an old state that has
7062 * more strict register state and more strict stack state then this execution
7063 * branch doesn't need to be explored further, since verifier already
7064 * concluded that more strict state leads to valid finish.
7066 * Therefore two states are equivalent if register state is more conservative
7067 * and explored stack state is more conservative than the current one.
7070 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
7071 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
7073 * In other words if current stack state (one being explored) has more
7074 * valid slots than old one that already passed validation, it means
7075 * the verifier can stop exploring and conclude that current state is valid too
7077 * Similarly with registers. If explored state has register type as invalid
7078 * whereas register type in current state is meaningful, it means that
7079 * the current state will reach 'bpf_exit' instruction safely
7081 static bool func_states_equal(struct bpf_func_state *old,
7082 struct bpf_func_state *cur)
7084 struct idpair *idmap;
7088 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
7089 /* If we failed to allocate the idmap, just say it's not safe */
7093 for (i = 0; i < MAX_BPF_REG; i++) {
7094 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
7098 if (!stacksafe(old, cur, idmap))
7101 if (!refsafe(old, cur))
7109 static bool states_equal(struct bpf_verifier_env *env,
7110 struct bpf_verifier_state *old,
7111 struct bpf_verifier_state *cur)
7115 if (old->curframe != cur->curframe)
7118 /* Verification state from speculative execution simulation
7119 * must never prune a non-speculative execution one.
7121 if (old->speculative && !cur->speculative)
7124 if (old->active_spin_lock != cur->active_spin_lock)
7127 /* for states to be equal callsites have to be the same
7128 * and all frame states need to be equivalent
7130 for (i = 0; i <= old->curframe; i++) {
7131 if (old->frame[i]->callsite != cur->frame[i]->callsite)
7133 if (!func_states_equal(old->frame[i], cur->frame[i]))
7139 /* Return 0 if no propagation happened. Return negative error code if error
7140 * happened. Otherwise, return the propagated bit.
7142 static int propagate_liveness_reg(struct bpf_verifier_env *env,
7143 struct bpf_reg_state *reg,
7144 struct bpf_reg_state *parent_reg)
7146 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
7147 u8 flag = reg->live & REG_LIVE_READ;
7150 /* When comes here, read flags of PARENT_REG or REG could be any of
7151 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
7152 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
7154 if (parent_flag == REG_LIVE_READ64 ||
7155 /* Or if there is no read flag from REG. */
7157 /* Or if the read flag from REG is the same as PARENT_REG. */
7158 parent_flag == flag)
7161 err = mark_reg_read(env, reg, parent_reg, flag);
7168 /* A write screens off any subsequent reads; but write marks come from the
7169 * straight-line code between a state and its parent. When we arrive at an
7170 * equivalent state (jump target or such) we didn't arrive by the straight-line
7171 * code, so read marks in the state must propagate to the parent regardless
7172 * of the state's write marks. That's what 'parent == state->parent' comparison
7173 * in mark_reg_read() is for.
7175 static int propagate_liveness(struct bpf_verifier_env *env,
7176 const struct bpf_verifier_state *vstate,
7177 struct bpf_verifier_state *vparent)
7179 struct bpf_reg_state *state_reg, *parent_reg;
7180 struct bpf_func_state *state, *parent;
7181 int i, frame, err = 0;
7183 if (vparent->curframe != vstate->curframe) {
7184 WARN(1, "propagate_live: parent frame %d current frame %d\n",
7185 vparent->curframe, vstate->curframe);
7188 /* Propagate read liveness of registers... */
7189 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
7190 for (frame = 0; frame <= vstate->curframe; frame++) {
7191 parent = vparent->frame[frame];
7192 state = vstate->frame[frame];
7193 parent_reg = parent->regs;
7194 state_reg = state->regs;
7195 /* We don't need to worry about FP liveness, it's read-only */
7196 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
7197 err = propagate_liveness_reg(env, &state_reg[i],
7201 if (err == REG_LIVE_READ64)
7202 mark_insn_zext(env, &parent_reg[i]);
7205 /* Propagate stack slots. */
7206 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
7207 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
7208 parent_reg = &parent->stack[i].spilled_ptr;
7209 state_reg = &state->stack[i].spilled_ptr;
7210 err = propagate_liveness_reg(env, state_reg,
7219 /* find precise scalars in the previous equivalent state and
7220 * propagate them into the current state
7222 static int propagate_precision(struct bpf_verifier_env *env,
7223 const struct bpf_verifier_state *old)
7225 struct bpf_reg_state *state_reg;
7226 struct bpf_func_state *state;
7229 state = old->frame[old->curframe];
7230 state_reg = state->regs;
7231 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
7232 if (state_reg->type != SCALAR_VALUE ||
7233 !state_reg->precise)
7235 if (env->log.level & BPF_LOG_LEVEL2)
7236 verbose(env, "propagating r%d\n", i);
7237 err = mark_chain_precision(env, i);
7242 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
7243 if (state->stack[i].slot_type[0] != STACK_SPILL)
7245 state_reg = &state->stack[i].spilled_ptr;
7246 if (state_reg->type != SCALAR_VALUE ||
7247 !state_reg->precise)
7249 if (env->log.level & BPF_LOG_LEVEL2)
7250 verbose(env, "propagating fp%d\n",
7251 (-i - 1) * BPF_REG_SIZE);
7252 err = mark_chain_precision_stack(env, i);
7259 static bool states_maybe_looping(struct bpf_verifier_state *old,
7260 struct bpf_verifier_state *cur)
7262 struct bpf_func_state *fold, *fcur;
7263 int i, fr = cur->curframe;
7265 if (old->curframe != fr)
7268 fold = old->frame[fr];
7269 fcur = cur->frame[fr];
7270 for (i = 0; i < MAX_BPF_REG; i++)
7271 if (memcmp(&fold->regs[i], &fcur->regs[i],
7272 offsetof(struct bpf_reg_state, parent)))
7278 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
7280 struct bpf_verifier_state_list *new_sl;
7281 struct bpf_verifier_state_list *sl, **pprev;
7282 struct bpf_verifier_state *cur = env->cur_state, *new;
7283 int i, j, err, states_cnt = 0;
7284 bool add_new_state = env->test_state_freq ? true : false;
7286 cur->last_insn_idx = env->prev_insn_idx;
7287 if (!env->insn_aux_data[insn_idx].prune_point)
7288 /* this 'insn_idx' instruction wasn't marked, so we will not
7289 * be doing state search here
7293 /* bpf progs typically have pruning point every 4 instructions
7294 * http://vger.kernel.org/bpfconf2019.html#session-1
7295 * Do not add new state for future pruning if the verifier hasn't seen
7296 * at least 2 jumps and at least 8 instructions.
7297 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
7298 * In tests that amounts to up to 50% reduction into total verifier
7299 * memory consumption and 20% verifier time speedup.
7301 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
7302 env->insn_processed - env->prev_insn_processed >= 8)
7303 add_new_state = true;
7305 pprev = explored_state(env, insn_idx);
7308 clean_live_states(env, insn_idx, cur);
7312 if (sl->state.insn_idx != insn_idx)
7314 if (sl->state.branches) {
7315 if (states_maybe_looping(&sl->state, cur) &&
7316 states_equal(env, &sl->state, cur)) {
7317 verbose_linfo(env, insn_idx, "; ");
7318 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
7321 /* if the verifier is processing a loop, avoid adding new state
7322 * too often, since different loop iterations have distinct
7323 * states and may not help future pruning.
7324 * This threshold shouldn't be too low to make sure that
7325 * a loop with large bound will be rejected quickly.
7326 * The most abusive loop will be:
7328 * if r1 < 1000000 goto pc-2
7329 * 1M insn_procssed limit / 100 == 10k peak states.
7330 * This threshold shouldn't be too high either, since states
7331 * at the end of the loop are likely to be useful in pruning.
7333 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
7334 env->insn_processed - env->prev_insn_processed < 100)
7335 add_new_state = false;
7338 if (states_equal(env, &sl->state, cur)) {
7340 /* reached equivalent register/stack state,
7342 * Registers read by the continuation are read by us.
7343 * If we have any write marks in env->cur_state, they
7344 * will prevent corresponding reads in the continuation
7345 * from reaching our parent (an explored_state). Our
7346 * own state will get the read marks recorded, but
7347 * they'll be immediately forgotten as we're pruning
7348 * this state and will pop a new one.
7350 err = propagate_liveness(env, &sl->state, cur);
7352 /* if previous state reached the exit with precision and
7353 * current state is equivalent to it (except precsion marks)
7354 * the precision needs to be propagated back in
7355 * the current state.
7357 err = err ? : push_jmp_history(env, cur);
7358 err = err ? : propagate_precision(env, &sl->state);
7364 /* when new state is not going to be added do not increase miss count.
7365 * Otherwise several loop iterations will remove the state
7366 * recorded earlier. The goal of these heuristics is to have
7367 * states from some iterations of the loop (some in the beginning
7368 * and some at the end) to help pruning.
7372 /* heuristic to determine whether this state is beneficial
7373 * to keep checking from state equivalence point of view.
7374 * Higher numbers increase max_states_per_insn and verification time,
7375 * but do not meaningfully decrease insn_processed.
7377 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
7378 /* the state is unlikely to be useful. Remove it to
7379 * speed up verification
7382 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
7383 u32 br = sl->state.branches;
7386 "BUG live_done but branches_to_explore %d\n",
7388 free_verifier_state(&sl->state, false);
7392 /* cannot free this state, since parentage chain may
7393 * walk it later. Add it for free_list instead to
7394 * be freed at the end of verification
7396 sl->next = env->free_list;
7397 env->free_list = sl;
7407 if (env->max_states_per_insn < states_cnt)
7408 env->max_states_per_insn = states_cnt;
7410 if (!env->allow_ptr_leaks && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
7411 return push_jmp_history(env, cur);
7414 return push_jmp_history(env, cur);
7416 /* There were no equivalent states, remember the current one.
7417 * Technically the current state is not proven to be safe yet,
7418 * but it will either reach outer most bpf_exit (which means it's safe)
7419 * or it will be rejected. When there are no loops the verifier won't be
7420 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
7421 * again on the way to bpf_exit.
7422 * When looping the sl->state.branches will be > 0 and this state
7423 * will not be considered for equivalence until branches == 0.
7425 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
7428 env->total_states++;
7430 env->prev_jmps_processed = env->jmps_processed;
7431 env->prev_insn_processed = env->insn_processed;
7433 /* add new state to the head of linked list */
7434 new = &new_sl->state;
7435 err = copy_verifier_state(new, cur);
7437 free_verifier_state(new, false);
7441 new->insn_idx = insn_idx;
7442 WARN_ONCE(new->branches != 1,
7443 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
7446 cur->first_insn_idx = insn_idx;
7447 clear_jmp_history(cur);
7448 new_sl->next = *explored_state(env, insn_idx);
7449 *explored_state(env, insn_idx) = new_sl;
7450 /* connect new state to parentage chain. Current frame needs all
7451 * registers connected. Only r6 - r9 of the callers are alive (pushed
7452 * to the stack implicitly by JITs) so in callers' frames connect just
7453 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
7454 * the state of the call instruction (with WRITTEN set), and r0 comes
7455 * from callee with its full parentage chain, anyway.
7457 /* clear write marks in current state: the writes we did are not writes
7458 * our child did, so they don't screen off its reads from us.
7459 * (There are no read marks in current state, because reads always mark
7460 * their parent and current state never has children yet. Only
7461 * explored_states can get read marks.)
7463 for (j = 0; j <= cur->curframe; j++) {
7464 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
7465 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
7466 for (i = 0; i < BPF_REG_FP; i++)
7467 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
7470 /* all stack frames are accessible from callee, clear them all */
7471 for (j = 0; j <= cur->curframe; j++) {
7472 struct bpf_func_state *frame = cur->frame[j];
7473 struct bpf_func_state *newframe = new->frame[j];
7475 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
7476 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
7477 frame->stack[i].spilled_ptr.parent =
7478 &newframe->stack[i].spilled_ptr;
7484 /* Return true if it's OK to have the same insn return a different type. */
7485 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
7490 case PTR_TO_SOCKET_OR_NULL:
7491 case PTR_TO_SOCK_COMMON:
7492 case PTR_TO_SOCK_COMMON_OR_NULL:
7493 case PTR_TO_TCP_SOCK:
7494 case PTR_TO_TCP_SOCK_OR_NULL:
7495 case PTR_TO_XDP_SOCK:
7502 /* If an instruction was previously used with particular pointer types, then we
7503 * need to be careful to avoid cases such as the below, where it may be ok
7504 * for one branch accessing the pointer, but not ok for the other branch:
7509 * R1 = some_other_valid_ptr;
7512 * R2 = *(u32 *)(R1 + 0);
7514 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
7516 return src != prev && (!reg_type_mismatch_ok(src) ||
7517 !reg_type_mismatch_ok(prev));
7520 static int do_check(struct bpf_verifier_env *env)
7522 struct bpf_verifier_state *state;
7523 struct bpf_insn *insns = env->prog->insnsi;
7524 struct bpf_reg_state *regs;
7525 int insn_cnt = env->prog->len;
7526 bool do_print_state = false;
7527 int prev_insn_idx = -1;
7529 env->prev_linfo = NULL;
7531 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
7534 state->curframe = 0;
7535 state->speculative = false;
7536 state->branches = 1;
7537 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
7538 if (!state->frame[0]) {
7542 env->cur_state = state;
7543 init_func_state(env, state->frame[0],
7544 BPF_MAIN_FUNC /* callsite */,
7546 0 /* subprogno, zero == main subprog */);
7549 struct bpf_insn *insn;
7553 env->prev_insn_idx = prev_insn_idx;
7554 if (env->insn_idx >= insn_cnt) {
7555 verbose(env, "invalid insn idx %d insn_cnt %d\n",
7556 env->insn_idx, insn_cnt);
7560 insn = &insns[env->insn_idx];
7561 class = BPF_CLASS(insn->code);
7563 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
7565 "BPF program is too large. Processed %d insn\n",
7566 env->insn_processed);
7570 err = is_state_visited(env, env->insn_idx);
7574 /* found equivalent state, can prune the search */
7575 if (env->log.level & BPF_LOG_LEVEL) {
7577 verbose(env, "\nfrom %d to %d%s: safe\n",
7578 env->prev_insn_idx, env->insn_idx,
7579 env->cur_state->speculative ?
7580 " (speculative execution)" : "");
7582 verbose(env, "%d: safe\n", env->insn_idx);
7584 goto process_bpf_exit;
7587 if (signal_pending(current))
7593 if (env->log.level & BPF_LOG_LEVEL2 ||
7594 (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
7595 if (env->log.level & BPF_LOG_LEVEL2)
7596 verbose(env, "%d:", env->insn_idx);
7598 verbose(env, "\nfrom %d to %d%s:",
7599 env->prev_insn_idx, env->insn_idx,
7600 env->cur_state->speculative ?
7601 " (speculative execution)" : "");
7602 print_verifier_state(env, state->frame[state->curframe]);
7603 do_print_state = false;
7606 if (env->log.level & BPF_LOG_LEVEL) {
7607 const struct bpf_insn_cbs cbs = {
7608 .cb_print = verbose,
7609 .private_data = env,
7612 verbose_linfo(env, env->insn_idx, "; ");
7613 verbose(env, "%d: ", env->insn_idx);
7614 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
7617 if (bpf_prog_is_dev_bound(env->prog->aux)) {
7618 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
7619 env->prev_insn_idx);
7624 regs = cur_regs(env);
7625 env->insn_aux_data[env->insn_idx].seen = true;
7626 prev_insn_idx = env->insn_idx;
7628 if (class == BPF_ALU || class == BPF_ALU64) {
7629 err = check_alu_op(env, insn);
7633 } else if (class == BPF_LDX) {
7634 enum bpf_reg_type *prev_src_type, src_reg_type;
7636 /* check for reserved fields is already done */
7638 /* check src operand */
7639 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7643 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
7647 src_reg_type = regs[insn->src_reg].type;
7649 /* check that memory (src_reg + off) is readable,
7650 * the state of dst_reg will be updated by this func
7652 err = check_mem_access(env, env->insn_idx, insn->src_reg,
7653 insn->off, BPF_SIZE(insn->code),
7654 BPF_READ, insn->dst_reg, false);
7658 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
7660 if (*prev_src_type == NOT_INIT) {
7662 * dst_reg = *(u32 *)(src_reg + off)
7663 * save type to validate intersecting paths
7665 *prev_src_type = src_reg_type;
7667 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
7668 /* ABuser program is trying to use the same insn
7669 * dst_reg = *(u32*) (src_reg + off)
7670 * with different pointer types:
7671 * src_reg == ctx in one branch and
7672 * src_reg == stack|map in some other branch.
7675 verbose(env, "same insn cannot be used with different pointers\n");
7679 } else if (class == BPF_STX) {
7680 enum bpf_reg_type *prev_dst_type, dst_reg_type;
7682 if (BPF_MODE(insn->code) == BPF_XADD) {
7683 err = check_xadd(env, env->insn_idx, insn);
7690 /* check src1 operand */
7691 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7694 /* check src2 operand */
7695 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7699 dst_reg_type = regs[insn->dst_reg].type;
7701 /* check that memory (dst_reg + off) is writeable */
7702 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
7703 insn->off, BPF_SIZE(insn->code),
7704 BPF_WRITE, insn->src_reg, false);
7708 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
7710 if (*prev_dst_type == NOT_INIT) {
7711 *prev_dst_type = dst_reg_type;
7712 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
7713 verbose(env, "same insn cannot be used with different pointers\n");
7717 } else if (class == BPF_ST) {
7718 if (BPF_MODE(insn->code) != BPF_MEM ||
7719 insn->src_reg != BPF_REG_0) {
7720 verbose(env, "BPF_ST uses reserved fields\n");
7723 /* check src operand */
7724 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7728 if (is_ctx_reg(env, insn->dst_reg)) {
7729 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
7731 reg_type_str[reg_state(env, insn->dst_reg)->type]);
7735 /* check that memory (dst_reg + off) is writeable */
7736 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
7737 insn->off, BPF_SIZE(insn->code),
7738 BPF_WRITE, -1, false);
7742 } else if (class == BPF_JMP || class == BPF_JMP32) {
7743 u8 opcode = BPF_OP(insn->code);
7745 env->jmps_processed++;
7746 if (opcode == BPF_CALL) {
7747 if (BPF_SRC(insn->code) != BPF_K ||
7749 (insn->src_reg != BPF_REG_0 &&
7750 insn->src_reg != BPF_PSEUDO_CALL) ||
7751 insn->dst_reg != BPF_REG_0 ||
7752 class == BPF_JMP32) {
7753 verbose(env, "BPF_CALL uses reserved fields\n");
7757 if (env->cur_state->active_spin_lock &&
7758 (insn->src_reg == BPF_PSEUDO_CALL ||
7759 insn->imm != BPF_FUNC_spin_unlock)) {
7760 verbose(env, "function calls are not allowed while holding a lock\n");
7763 if (insn->src_reg == BPF_PSEUDO_CALL)
7764 err = check_func_call(env, insn, &env->insn_idx);
7766 err = check_helper_call(env, insn->imm, env->insn_idx);
7770 } else if (opcode == BPF_JA) {
7771 if (BPF_SRC(insn->code) != BPF_K ||
7773 insn->src_reg != BPF_REG_0 ||
7774 insn->dst_reg != BPF_REG_0 ||
7775 class == BPF_JMP32) {
7776 verbose(env, "BPF_JA uses reserved fields\n");
7780 env->insn_idx += insn->off + 1;
7783 } else if (opcode == BPF_EXIT) {
7784 if (BPF_SRC(insn->code) != BPF_K ||
7786 insn->src_reg != BPF_REG_0 ||
7787 insn->dst_reg != BPF_REG_0 ||
7788 class == BPF_JMP32) {
7789 verbose(env, "BPF_EXIT uses reserved fields\n");
7793 if (env->cur_state->active_spin_lock) {
7794 verbose(env, "bpf_spin_unlock is missing\n");
7798 if (state->curframe) {
7799 /* exit from nested function */
7800 err = prepare_func_exit(env, &env->insn_idx);
7803 do_print_state = true;
7807 err = check_reference_leak(env);
7811 /* eBPF calling convetion is such that R0 is used
7812 * to return the value from eBPF program.
7813 * Make sure that it's readable at this time
7814 * of bpf_exit, which means that program wrote
7815 * something into it earlier
7817 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
7821 if (is_pointer_value(env, BPF_REG_0)) {
7822 verbose(env, "R0 leaks addr as return value\n");
7826 err = check_return_code(env);
7830 update_branch_counts(env, env->cur_state);
7831 err = pop_stack(env, &prev_insn_idx,
7838 do_print_state = true;
7842 err = check_cond_jmp_op(env, insn, &env->insn_idx);
7846 } else if (class == BPF_LD) {
7847 u8 mode = BPF_MODE(insn->code);
7849 if (mode == BPF_ABS || mode == BPF_IND) {
7850 err = check_ld_abs(env, insn);
7854 } else if (mode == BPF_IMM) {
7855 err = check_ld_imm(env, insn);
7860 env->insn_aux_data[env->insn_idx].seen = true;
7862 verbose(env, "invalid BPF_LD mode\n");
7866 verbose(env, "unknown insn class %d\n", class);
7873 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
7877 static int check_map_prealloc(struct bpf_map *map)
7879 return (map->map_type != BPF_MAP_TYPE_HASH &&
7880 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
7881 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
7882 !(map->map_flags & BPF_F_NO_PREALLOC);
7885 static bool is_tracing_prog_type(enum bpf_prog_type type)
7888 case BPF_PROG_TYPE_KPROBE:
7889 case BPF_PROG_TYPE_TRACEPOINT:
7890 case BPF_PROG_TYPE_PERF_EVENT:
7891 case BPF_PROG_TYPE_RAW_TRACEPOINT:
7898 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
7899 struct bpf_map *map,
7900 struct bpf_prog *prog)
7903 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
7904 * preallocated hash maps, since doing memory allocation
7905 * in overflow_handler can crash depending on where nmi got
7908 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
7909 if (!check_map_prealloc(map)) {
7910 verbose(env, "perf_event programs can only use preallocated hash map\n");
7913 if (map->inner_map_meta &&
7914 !check_map_prealloc(map->inner_map_meta)) {
7915 verbose(env, "perf_event programs can only use preallocated inner hash map\n");
7920 if ((is_tracing_prog_type(prog->type) ||
7921 prog->type == BPF_PROG_TYPE_SOCKET_FILTER) &&
7922 map_value_has_spin_lock(map)) {
7923 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
7927 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
7928 !bpf_offload_prog_map_match(prog, map)) {
7929 verbose(env, "offload device mismatch between prog and map\n");
7936 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
7938 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
7939 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
7942 /* look for pseudo eBPF instructions that access map FDs and
7943 * replace them with actual map pointers
7945 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
7947 struct bpf_insn *insn = env->prog->insnsi;
7948 int insn_cnt = env->prog->len;
7951 err = bpf_prog_calc_tag(env->prog);
7955 for (i = 0; i < insn_cnt; i++, insn++) {
7956 if (BPF_CLASS(insn->code) == BPF_LDX &&
7957 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
7958 verbose(env, "BPF_LDX uses reserved fields\n");
7962 if (BPF_CLASS(insn->code) == BPF_STX &&
7963 ((BPF_MODE(insn->code) != BPF_MEM &&
7964 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
7965 verbose(env, "BPF_STX uses reserved fields\n");
7969 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
7970 struct bpf_insn_aux_data *aux;
7971 struct bpf_map *map;
7975 if (i == insn_cnt - 1 || insn[1].code != 0 ||
7976 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
7978 verbose(env, "invalid bpf_ld_imm64 insn\n");
7982 if (insn[0].src_reg == 0)
7983 /* valid generic load 64-bit imm */
7986 /* In final convert_pseudo_ld_imm64() step, this is
7987 * converted into regular 64-bit imm load insn.
7989 if ((insn[0].src_reg != BPF_PSEUDO_MAP_FD &&
7990 insn[0].src_reg != BPF_PSEUDO_MAP_VALUE) ||
7991 (insn[0].src_reg == BPF_PSEUDO_MAP_FD &&
7992 insn[1].imm != 0)) {
7994 "unrecognized bpf_ld_imm64 insn\n");
7998 f = fdget(insn[0].imm);
7999 map = __bpf_map_get(f);
8001 verbose(env, "fd %d is not pointing to valid bpf_map\n",
8003 return PTR_ERR(map);
8006 err = check_map_prog_compatibility(env, map, env->prog);
8012 aux = &env->insn_aux_data[i];
8013 if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
8014 addr = (unsigned long)map;
8016 u32 off = insn[1].imm;
8018 if (off >= BPF_MAX_VAR_OFF) {
8019 verbose(env, "direct value offset of %u is not allowed\n", off);
8024 if (!map->ops->map_direct_value_addr) {
8025 verbose(env, "no direct value access support for this map type\n");
8030 err = map->ops->map_direct_value_addr(map, &addr, off);
8032 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
8033 map->value_size, off);
8042 insn[0].imm = (u32)addr;
8043 insn[1].imm = addr >> 32;
8045 /* check whether we recorded this map already */
8046 for (j = 0; j < env->used_map_cnt; j++) {
8047 if (env->used_maps[j] == map) {
8054 if (env->used_map_cnt >= MAX_USED_MAPS) {
8059 /* hold the map. If the program is rejected by verifier,
8060 * the map will be released by release_maps() or it
8061 * will be used by the valid program until it's unloaded
8062 * and all maps are released in free_used_maps()
8064 map = bpf_map_inc(map, false);
8067 return PTR_ERR(map);
8070 aux->map_index = env->used_map_cnt;
8071 env->used_maps[env->used_map_cnt++] = map;
8073 if (bpf_map_is_cgroup_storage(map) &&
8074 bpf_cgroup_storage_assign(env->prog, map)) {
8075 verbose(env, "only one cgroup storage of each type is allowed\n");
8087 /* Basic sanity check before we invest more work here. */
8088 if (!bpf_opcode_in_insntable(insn->code)) {
8089 verbose(env, "unknown opcode %02x\n", insn->code);
8094 /* now all pseudo BPF_LD_IMM64 instructions load valid
8095 * 'struct bpf_map *' into a register instead of user map_fd.
8096 * These pointers will be used later by verifier to validate map access.
8101 /* drop refcnt of maps used by the rejected program */
8102 static void release_maps(struct bpf_verifier_env *env)
8104 enum bpf_cgroup_storage_type stype;
8107 for_each_cgroup_storage_type(stype) {
8108 if (!env->prog->aux->cgroup_storage[stype])
8110 bpf_cgroup_storage_release(env->prog,
8111 env->prog->aux->cgroup_storage[stype]);
8114 for (i = 0; i < env->used_map_cnt; i++)
8115 bpf_map_put(env->used_maps[i]);
8118 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
8119 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
8121 struct bpf_insn *insn = env->prog->insnsi;
8122 int insn_cnt = env->prog->len;
8125 for (i = 0; i < insn_cnt; i++, insn++)
8126 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
8130 /* single env->prog->insni[off] instruction was replaced with the range
8131 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
8132 * [0, off) and [off, end) to new locations, so the patched range stays zero
8134 static int adjust_insn_aux_data(struct bpf_verifier_env *env,
8135 struct bpf_prog *new_prog, u32 off, u32 cnt)
8137 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
8138 struct bpf_insn *insn = new_prog->insnsi;
8142 /* aux info at OFF always needs adjustment, no matter fast path
8143 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
8144 * original insn at old prog.
8146 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
8150 prog_len = new_prog->len;
8151 new_data = vzalloc(array_size(prog_len,
8152 sizeof(struct bpf_insn_aux_data)));
8155 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
8156 memcpy(new_data + off + cnt - 1, old_data + off,
8157 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
8158 for (i = off; i < off + cnt - 1; i++) {
8159 new_data[i].seen = true;
8160 new_data[i].zext_dst = insn_has_def32(env, insn + i);
8162 env->insn_aux_data = new_data;
8167 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
8173 /* NOTE: fake 'exit' subprog should be updated as well. */
8174 for (i = 0; i <= env->subprog_cnt; i++) {
8175 if (env->subprog_info[i].start <= off)
8177 env->subprog_info[i].start += len - 1;
8181 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
8182 const struct bpf_insn *patch, u32 len)
8184 struct bpf_prog *new_prog;
8186 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
8187 if (IS_ERR(new_prog)) {
8188 if (PTR_ERR(new_prog) == -ERANGE)
8190 "insn %d cannot be patched due to 16-bit range\n",
8191 env->insn_aux_data[off].orig_idx);
8194 if (adjust_insn_aux_data(env, new_prog, off, len))
8196 adjust_subprog_starts(env, off, len);
8200 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
8205 /* find first prog starting at or after off (first to remove) */
8206 for (i = 0; i < env->subprog_cnt; i++)
8207 if (env->subprog_info[i].start >= off)
8209 /* find first prog starting at or after off + cnt (first to stay) */
8210 for (j = i; j < env->subprog_cnt; j++)
8211 if (env->subprog_info[j].start >= off + cnt)
8213 /* if j doesn't start exactly at off + cnt, we are just removing
8214 * the front of previous prog
8216 if (env->subprog_info[j].start != off + cnt)
8220 struct bpf_prog_aux *aux = env->prog->aux;
8223 /* move fake 'exit' subprog as well */
8224 move = env->subprog_cnt + 1 - j;
8226 memmove(env->subprog_info + i,
8227 env->subprog_info + j,
8228 sizeof(*env->subprog_info) * move);
8229 env->subprog_cnt -= j - i;
8231 /* remove func_info */
8232 if (aux->func_info) {
8233 move = aux->func_info_cnt - j;
8235 memmove(aux->func_info + i,
8237 sizeof(*aux->func_info) * move);
8238 aux->func_info_cnt -= j - i;
8239 /* func_info->insn_off is set after all code rewrites,
8240 * in adjust_btf_func() - no need to adjust
8244 /* convert i from "first prog to remove" to "first to adjust" */
8245 if (env->subprog_info[i].start == off)
8249 /* update fake 'exit' subprog as well */
8250 for (; i <= env->subprog_cnt; i++)
8251 env->subprog_info[i].start -= cnt;
8256 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
8259 struct bpf_prog *prog = env->prog;
8260 u32 i, l_off, l_cnt, nr_linfo;
8261 struct bpf_line_info *linfo;
8263 nr_linfo = prog->aux->nr_linfo;
8267 linfo = prog->aux->linfo;
8269 /* find first line info to remove, count lines to be removed */
8270 for (i = 0; i < nr_linfo; i++)
8271 if (linfo[i].insn_off >= off)
8276 for (; i < nr_linfo; i++)
8277 if (linfo[i].insn_off < off + cnt)
8282 /* First live insn doesn't match first live linfo, it needs to "inherit"
8283 * last removed linfo. prog is already modified, so prog->len == off
8284 * means no live instructions after (tail of the program was removed).
8286 if (prog->len != off && l_cnt &&
8287 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
8289 linfo[--i].insn_off = off + cnt;
8292 /* remove the line info which refer to the removed instructions */
8294 memmove(linfo + l_off, linfo + i,
8295 sizeof(*linfo) * (nr_linfo - i));
8297 prog->aux->nr_linfo -= l_cnt;
8298 nr_linfo = prog->aux->nr_linfo;
8301 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
8302 for (i = l_off; i < nr_linfo; i++)
8303 linfo[i].insn_off -= cnt;
8305 /* fix up all subprogs (incl. 'exit') which start >= off */
8306 for (i = 0; i <= env->subprog_cnt; i++)
8307 if (env->subprog_info[i].linfo_idx > l_off) {
8308 /* program may have started in the removed region but
8309 * may not be fully removed
8311 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
8312 env->subprog_info[i].linfo_idx -= l_cnt;
8314 env->subprog_info[i].linfo_idx = l_off;
8320 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
8322 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
8323 unsigned int orig_prog_len = env->prog->len;
8326 if (bpf_prog_is_dev_bound(env->prog->aux))
8327 bpf_prog_offload_remove_insns(env, off, cnt);
8329 err = bpf_remove_insns(env->prog, off, cnt);
8333 err = adjust_subprog_starts_after_remove(env, off, cnt);
8337 err = bpf_adj_linfo_after_remove(env, off, cnt);
8341 memmove(aux_data + off, aux_data + off + cnt,
8342 sizeof(*aux_data) * (orig_prog_len - off - cnt));
8347 /* The verifier does more data flow analysis than llvm and will not
8348 * explore branches that are dead at run time. Malicious programs can
8349 * have dead code too. Therefore replace all dead at-run-time code
8352 * Just nops are not optimal, e.g. if they would sit at the end of the
8353 * program and through another bug we would manage to jump there, then
8354 * we'd execute beyond program memory otherwise. Returning exception
8355 * code also wouldn't work since we can have subprogs where the dead
8356 * code could be located.
8358 static void sanitize_dead_code(struct bpf_verifier_env *env)
8360 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
8361 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
8362 struct bpf_insn *insn = env->prog->insnsi;
8363 const int insn_cnt = env->prog->len;
8366 for (i = 0; i < insn_cnt; i++) {
8367 if (aux_data[i].seen)
8369 memcpy(insn + i, &trap, sizeof(trap));
8373 static bool insn_is_cond_jump(u8 code)
8377 if (BPF_CLASS(code) == BPF_JMP32)
8380 if (BPF_CLASS(code) != BPF_JMP)
8384 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
8387 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
8389 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
8390 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
8391 struct bpf_insn *insn = env->prog->insnsi;
8392 const int insn_cnt = env->prog->len;
8395 for (i = 0; i < insn_cnt; i++, insn++) {
8396 if (!insn_is_cond_jump(insn->code))
8399 if (!aux_data[i + 1].seen)
8401 else if (!aux_data[i + 1 + insn->off].seen)
8406 if (bpf_prog_is_dev_bound(env->prog->aux))
8407 bpf_prog_offload_replace_insn(env, i, &ja);
8409 memcpy(insn, &ja, sizeof(ja));
8413 static int opt_remove_dead_code(struct bpf_verifier_env *env)
8415 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
8416 int insn_cnt = env->prog->len;
8419 for (i = 0; i < insn_cnt; i++) {
8423 while (i + j < insn_cnt && !aux_data[i + j].seen)
8428 err = verifier_remove_insns(env, i, j);
8431 insn_cnt = env->prog->len;
8437 static int opt_remove_nops(struct bpf_verifier_env *env)
8439 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
8440 struct bpf_insn *insn = env->prog->insnsi;
8441 int insn_cnt = env->prog->len;
8444 for (i = 0; i < insn_cnt; i++) {
8445 if (memcmp(&insn[i], &ja, sizeof(ja)))
8448 err = verifier_remove_insns(env, i, 1);
8458 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
8459 const union bpf_attr *attr)
8461 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
8462 struct bpf_insn_aux_data *aux = env->insn_aux_data;
8463 int i, patch_len, delta = 0, len = env->prog->len;
8464 struct bpf_insn *insns = env->prog->insnsi;
8465 struct bpf_prog *new_prog;
8468 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
8469 zext_patch[1] = BPF_ZEXT_REG(0);
8470 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
8471 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
8472 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
8473 for (i = 0; i < len; i++) {
8474 int adj_idx = i + delta;
8475 struct bpf_insn insn;
8477 insn = insns[adj_idx];
8478 if (!aux[adj_idx].zext_dst) {
8486 class = BPF_CLASS(code);
8487 if (insn_no_def(&insn))
8490 /* NOTE: arg "reg" (the fourth one) is only used for
8491 * BPF_STX which has been ruled out in above
8492 * check, it is safe to pass NULL here.
8494 if (is_reg64(env, &insn, insn.dst_reg, NULL, DST_OP)) {
8495 if (class == BPF_LD &&
8496 BPF_MODE(code) == BPF_IMM)
8501 /* ctx load could be transformed into wider load. */
8502 if (class == BPF_LDX &&
8503 aux[adj_idx].ptr_type == PTR_TO_CTX)
8506 imm_rnd = get_random_int();
8507 rnd_hi32_patch[0] = insn;
8508 rnd_hi32_patch[1].imm = imm_rnd;
8509 rnd_hi32_patch[3].dst_reg = insn.dst_reg;
8510 patch = rnd_hi32_patch;
8512 goto apply_patch_buffer;
8515 if (!bpf_jit_needs_zext())
8518 zext_patch[0] = insn;
8519 zext_patch[1].dst_reg = insn.dst_reg;
8520 zext_patch[1].src_reg = insn.dst_reg;
8524 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
8527 env->prog = new_prog;
8528 insns = new_prog->insnsi;
8529 aux = env->insn_aux_data;
8530 delta += patch_len - 1;
8536 /* convert load instructions that access fields of a context type into a
8537 * sequence of instructions that access fields of the underlying structure:
8538 * struct __sk_buff -> struct sk_buff
8539 * struct bpf_sock_ops -> struct sock
8541 static int convert_ctx_accesses(struct bpf_verifier_env *env)
8543 const struct bpf_verifier_ops *ops = env->ops;
8544 int i, cnt, size, ctx_field_size, delta = 0;
8545 const int insn_cnt = env->prog->len;
8546 struct bpf_insn insn_buf[16], *insn;
8547 u32 target_size, size_default, off;
8548 struct bpf_prog *new_prog;
8549 enum bpf_access_type type;
8550 bool is_narrower_load;
8552 if (ops->gen_prologue || env->seen_direct_write) {
8553 if (!ops->gen_prologue) {
8554 verbose(env, "bpf verifier is misconfigured\n");
8557 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
8559 if (cnt >= ARRAY_SIZE(insn_buf)) {
8560 verbose(env, "bpf verifier is misconfigured\n");
8563 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
8567 env->prog = new_prog;
8572 if (bpf_prog_is_dev_bound(env->prog->aux))
8575 insn = env->prog->insnsi + delta;
8577 for (i = 0; i < insn_cnt; i++, insn++) {
8578 bpf_convert_ctx_access_t convert_ctx_access;
8580 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
8581 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
8582 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
8583 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
8585 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
8586 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
8587 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
8588 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
8593 if (type == BPF_WRITE &&
8594 env->insn_aux_data[i + delta].sanitize_stack_off) {
8595 struct bpf_insn patch[] = {
8596 /* Sanitize suspicious stack slot with zero.
8597 * There are no memory dependencies for this store,
8598 * since it's only using frame pointer and immediate
8601 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
8602 env->insn_aux_data[i + delta].sanitize_stack_off,
8604 /* the original STX instruction will immediately
8605 * overwrite the same stack slot with appropriate value
8610 cnt = ARRAY_SIZE(patch);
8611 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
8616 env->prog = new_prog;
8617 insn = new_prog->insnsi + i + delta;
8621 switch (env->insn_aux_data[i + delta].ptr_type) {
8623 if (!ops->convert_ctx_access)
8625 convert_ctx_access = ops->convert_ctx_access;
8628 case PTR_TO_SOCK_COMMON:
8629 convert_ctx_access = bpf_sock_convert_ctx_access;
8631 case PTR_TO_TCP_SOCK:
8632 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
8634 case PTR_TO_XDP_SOCK:
8635 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
8641 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
8642 size = BPF_LDST_BYTES(insn);
8644 /* If the read access is a narrower load of the field,
8645 * convert to a 4/8-byte load, to minimum program type specific
8646 * convert_ctx_access changes. If conversion is successful,
8647 * we will apply proper mask to the result.
8649 is_narrower_load = size < ctx_field_size;
8650 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
8652 if (is_narrower_load) {
8655 if (type == BPF_WRITE) {
8656 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
8661 if (ctx_field_size == 4)
8663 else if (ctx_field_size == 8)
8666 insn->off = off & ~(size_default - 1);
8667 insn->code = BPF_LDX | BPF_MEM | size_code;
8671 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
8673 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
8674 (ctx_field_size && !target_size)) {
8675 verbose(env, "bpf verifier is misconfigured\n");
8679 if (is_narrower_load && size < target_size) {
8680 u8 shift = bpf_ctx_narrow_access_offset(
8681 off, size, size_default) * 8;
8682 if (ctx_field_size <= 4) {
8684 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
8687 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
8688 (1 << size * 8) - 1);
8691 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
8694 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
8695 (1ULL << size * 8) - 1);
8699 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
8705 /* keep walking new program and skip insns we just inserted */
8706 env->prog = new_prog;
8707 insn = new_prog->insnsi + i + delta;
8713 static int jit_subprogs(struct bpf_verifier_env *env)
8715 struct bpf_prog *prog = env->prog, **func, *tmp;
8716 int i, j, subprog_start, subprog_end = 0, len, subprog;
8717 struct bpf_insn *insn;
8721 if (env->subprog_cnt <= 1)
8724 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
8725 if (insn->code != (BPF_JMP | BPF_CALL) ||
8726 insn->src_reg != BPF_PSEUDO_CALL)
8728 /* Upon error here we cannot fall back to interpreter but
8729 * need a hard reject of the program. Thus -EFAULT is
8730 * propagated in any case.
8732 subprog = find_subprog(env, i + insn->imm + 1);
8734 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
8738 /* temporarily remember subprog id inside insn instead of
8739 * aux_data, since next loop will split up all insns into funcs
8741 insn->off = subprog;
8742 /* remember original imm in case JIT fails and fallback
8743 * to interpreter will be needed
8745 env->insn_aux_data[i].call_imm = insn->imm;
8746 /* point imm to __bpf_call_base+1 from JITs point of view */
8750 err = bpf_prog_alloc_jited_linfo(prog);
8755 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
8759 for (i = 0; i < env->subprog_cnt; i++) {
8760 subprog_start = subprog_end;
8761 subprog_end = env->subprog_info[i + 1].start;
8763 len = subprog_end - subprog_start;
8764 /* BPF_PROG_RUN doesn't call subprogs directly,
8765 * hence main prog stats include the runtime of subprogs.
8766 * subprogs don't have IDs and not reachable via prog_get_next_id
8767 * func[i]->aux->stats will never be accessed and stays NULL
8769 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
8772 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
8773 len * sizeof(struct bpf_insn));
8774 func[i]->type = prog->type;
8776 if (bpf_prog_calc_tag(func[i]))
8778 func[i]->is_func = 1;
8779 func[i]->aux->func_idx = i;
8780 /* the btf and func_info will be freed only at prog->aux */
8781 func[i]->aux->btf = prog->aux->btf;
8782 func[i]->aux->func_info = prog->aux->func_info;
8784 /* Use bpf_prog_F_tag to indicate functions in stack traces.
8785 * Long term would need debug info to populate names
8787 func[i]->aux->name[0] = 'F';
8788 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
8789 func[i]->jit_requested = 1;
8790 func[i]->aux->linfo = prog->aux->linfo;
8791 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
8792 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
8793 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
8794 func[i] = bpf_int_jit_compile(func[i]);
8795 if (!func[i]->jited) {
8801 /* at this point all bpf functions were successfully JITed
8802 * now populate all bpf_calls with correct addresses and
8803 * run last pass of JIT
8805 for (i = 0; i < env->subprog_cnt; i++) {
8806 insn = func[i]->insnsi;
8807 for (j = 0; j < func[i]->len; j++, insn++) {
8808 if (insn->code != (BPF_JMP | BPF_CALL) ||
8809 insn->src_reg != BPF_PSEUDO_CALL)
8811 subprog = insn->off;
8812 insn->imm = BPF_CAST_CALL(func[subprog]->bpf_func) -
8816 /* we use the aux data to keep a list of the start addresses
8817 * of the JITed images for each function in the program
8819 * for some architectures, such as powerpc64, the imm field
8820 * might not be large enough to hold the offset of the start
8821 * address of the callee's JITed image from __bpf_call_base
8823 * in such cases, we can lookup the start address of a callee
8824 * by using its subprog id, available from the off field of
8825 * the call instruction, as an index for this list
8827 func[i]->aux->func = func;
8828 func[i]->aux->func_cnt = env->subprog_cnt;
8830 for (i = 0; i < env->subprog_cnt; i++) {
8831 old_bpf_func = func[i]->bpf_func;
8832 tmp = bpf_int_jit_compile(func[i]);
8833 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
8834 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
8841 /* finally lock prog and jit images for all functions and
8844 for (i = 0; i < env->subprog_cnt; i++) {
8845 bpf_prog_lock_ro(func[i]);
8846 bpf_prog_kallsyms_add(func[i]);
8849 /* Last step: make now unused interpreter insns from main
8850 * prog consistent for later dump requests, so they can
8851 * later look the same as if they were interpreted only.
8853 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
8854 if (insn->code != (BPF_JMP | BPF_CALL) ||
8855 insn->src_reg != BPF_PSEUDO_CALL)
8857 insn->off = env->insn_aux_data[i].call_imm;
8858 subprog = find_subprog(env, i + insn->off + 1);
8859 insn->imm = subprog;
8863 prog->bpf_func = func[0]->bpf_func;
8864 prog->aux->func = func;
8865 prog->aux->func_cnt = env->subprog_cnt;
8866 bpf_prog_free_unused_jited_linfo(prog);
8869 for (i = 0; i < env->subprog_cnt; i++)
8871 bpf_jit_free(func[i]);
8874 /* cleanup main prog to be interpreted */
8875 prog->jit_requested = 0;
8876 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
8877 if (insn->code != (BPF_JMP | BPF_CALL) ||
8878 insn->src_reg != BPF_PSEUDO_CALL)
8881 insn->imm = env->insn_aux_data[i].call_imm;
8883 bpf_prog_free_jited_linfo(prog);
8887 static int fixup_call_args(struct bpf_verifier_env *env)
8889 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
8890 struct bpf_prog *prog = env->prog;
8891 struct bpf_insn *insn = prog->insnsi;
8896 if (env->prog->jit_requested &&
8897 !bpf_prog_is_dev_bound(env->prog->aux)) {
8898 err = jit_subprogs(env);
8904 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
8905 for (i = 0; i < prog->len; i++, insn++) {
8906 if (insn->code != (BPF_JMP | BPF_CALL) ||
8907 insn->src_reg != BPF_PSEUDO_CALL)
8909 depth = get_callee_stack_depth(env, insn, i);
8912 bpf_patch_call_args(insn, depth);
8919 /* fixup insn->imm field of bpf_call instructions
8920 * and inline eligible helpers as explicit sequence of BPF instructions
8922 * this function is called after eBPF program passed verification
8924 static int fixup_bpf_calls(struct bpf_verifier_env *env)
8926 struct bpf_prog *prog = env->prog;
8927 struct bpf_insn *insn = prog->insnsi;
8928 const struct bpf_func_proto *fn;
8929 const int insn_cnt = prog->len;
8930 const struct bpf_map_ops *ops;
8931 struct bpf_insn_aux_data *aux;
8932 struct bpf_insn insn_buf[16];
8933 struct bpf_prog *new_prog;
8934 struct bpf_map *map_ptr;
8935 int i, cnt, delta = 0;
8937 for (i = 0; i < insn_cnt; i++, insn++) {
8938 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
8939 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
8940 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
8941 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
8942 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
8943 struct bpf_insn mask_and_div[] = {
8944 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
8946 BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2),
8947 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
8948 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
8951 struct bpf_insn mask_and_mod[] = {
8952 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
8953 /* Rx mod 0 -> Rx */
8954 BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1),
8957 struct bpf_insn *patchlet;
8959 if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
8960 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
8961 patchlet = mask_and_div + (is64 ? 1 : 0);
8962 cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0);
8964 patchlet = mask_and_mod + (is64 ? 1 : 0);
8965 cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0);
8968 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
8973 env->prog = prog = new_prog;
8974 insn = new_prog->insnsi + i + delta;
8978 if (BPF_CLASS(insn->code) == BPF_LD &&
8979 (BPF_MODE(insn->code) == BPF_ABS ||
8980 BPF_MODE(insn->code) == BPF_IND)) {
8981 cnt = env->ops->gen_ld_abs(insn, insn_buf);
8982 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
8983 verbose(env, "bpf verifier is misconfigured\n");
8987 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
8992 env->prog = prog = new_prog;
8993 insn = new_prog->insnsi + i + delta;
8997 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
8998 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
8999 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
9000 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
9001 struct bpf_insn insn_buf[16];
9002 struct bpf_insn *patch = &insn_buf[0];
9006 aux = &env->insn_aux_data[i + delta];
9007 if (!aux->alu_state ||
9008 aux->alu_state == BPF_ALU_NON_POINTER)
9011 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
9012 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
9013 BPF_ALU_SANITIZE_SRC;
9015 off_reg = issrc ? insn->src_reg : insn->dst_reg;
9017 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
9018 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit - 1);
9019 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
9020 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
9021 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
9022 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
9024 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX,
9026 insn->src_reg = BPF_REG_AX;
9028 *patch++ = BPF_ALU64_REG(BPF_AND, off_reg,
9032 insn->code = insn->code == code_add ?
9033 code_sub : code_add;
9036 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
9037 cnt = patch - insn_buf;
9039 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9044 env->prog = prog = new_prog;
9045 insn = new_prog->insnsi + i + delta;
9049 if (insn->code != (BPF_JMP | BPF_CALL))
9051 if (insn->src_reg == BPF_PSEUDO_CALL)
9054 if (insn->imm == BPF_FUNC_get_route_realm)
9055 prog->dst_needed = 1;
9056 if (insn->imm == BPF_FUNC_get_prandom_u32)
9057 bpf_user_rnd_init_once();
9058 if (insn->imm == BPF_FUNC_override_return)
9059 prog->kprobe_override = 1;
9060 if (insn->imm == BPF_FUNC_tail_call) {
9061 /* If we tail call into other programs, we
9062 * cannot make any assumptions since they can
9063 * be replaced dynamically during runtime in
9064 * the program array.
9066 prog->cb_access = 1;
9067 env->prog->aux->stack_depth = MAX_BPF_STACK;
9068 env->prog->aux->max_pkt_offset = MAX_PACKET_OFF;
9070 /* mark bpf_tail_call as different opcode to avoid
9071 * conditional branch in the interpeter for every normal
9072 * call and to prevent accidental JITing by JIT compiler
9073 * that doesn't support bpf_tail_call yet
9076 insn->code = BPF_JMP | BPF_TAIL_CALL;
9078 aux = &env->insn_aux_data[i + delta];
9079 if (!bpf_map_ptr_unpriv(aux))
9082 /* instead of changing every JIT dealing with tail_call
9083 * emit two extra insns:
9084 * if (index >= max_entries) goto out;
9085 * index &= array->index_mask;
9086 * to avoid out-of-bounds cpu speculation
9088 if (bpf_map_ptr_poisoned(aux)) {
9089 verbose(env, "tail_call abusing map_ptr\n");
9093 map_ptr = BPF_MAP_PTR(aux->map_state);
9094 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
9095 map_ptr->max_entries, 2);
9096 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
9097 container_of(map_ptr,
9100 insn_buf[2] = *insn;
9102 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9107 env->prog = prog = new_prog;
9108 insn = new_prog->insnsi + i + delta;
9112 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
9113 * and other inlining handlers are currently limited to 64 bit
9116 if (prog->jit_requested && BITS_PER_LONG == 64 &&
9117 (insn->imm == BPF_FUNC_map_lookup_elem ||
9118 insn->imm == BPF_FUNC_map_update_elem ||
9119 insn->imm == BPF_FUNC_map_delete_elem ||
9120 insn->imm == BPF_FUNC_map_push_elem ||
9121 insn->imm == BPF_FUNC_map_pop_elem ||
9122 insn->imm == BPF_FUNC_map_peek_elem)) {
9123 aux = &env->insn_aux_data[i + delta];
9124 if (bpf_map_ptr_poisoned(aux))
9125 goto patch_call_imm;
9127 map_ptr = BPF_MAP_PTR(aux->map_state);
9129 if (insn->imm == BPF_FUNC_map_lookup_elem &&
9130 ops->map_gen_lookup) {
9131 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
9132 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
9133 verbose(env, "bpf verifier is misconfigured\n");
9137 new_prog = bpf_patch_insn_data(env, i + delta,
9143 env->prog = prog = new_prog;
9144 insn = new_prog->insnsi + i + delta;
9148 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
9149 (void *(*)(struct bpf_map *map, void *key))NULL));
9150 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
9151 (int (*)(struct bpf_map *map, void *key))NULL));
9152 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
9153 (int (*)(struct bpf_map *map, void *key, void *value,
9155 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
9156 (int (*)(struct bpf_map *map, void *value,
9158 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
9159 (int (*)(struct bpf_map *map, void *value))NULL));
9160 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
9161 (int (*)(struct bpf_map *map, void *value))NULL));
9163 switch (insn->imm) {
9164 case BPF_FUNC_map_lookup_elem:
9165 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
9168 case BPF_FUNC_map_update_elem:
9169 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
9172 case BPF_FUNC_map_delete_elem:
9173 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
9176 case BPF_FUNC_map_push_elem:
9177 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
9180 case BPF_FUNC_map_pop_elem:
9181 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
9184 case BPF_FUNC_map_peek_elem:
9185 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
9190 goto patch_call_imm;
9194 fn = env->ops->get_func_proto(insn->imm, env->prog);
9195 /* all functions that have prototype and verifier allowed
9196 * programs to call them, must be real in-kernel functions
9200 "kernel subsystem misconfigured func %s#%d\n",
9201 func_id_name(insn->imm), insn->imm);
9204 insn->imm = fn->func - __bpf_call_base;
9210 static void free_states(struct bpf_verifier_env *env)
9212 struct bpf_verifier_state_list *sl, *sln;
9215 sl = env->free_list;
9218 free_verifier_state(&sl->state, false);
9223 if (!env->explored_states)
9226 for (i = 0; i < state_htab_size(env); i++) {
9227 sl = env->explored_states[i];
9231 free_verifier_state(&sl->state, false);
9237 kvfree(env->explored_states);
9240 static void print_verification_stats(struct bpf_verifier_env *env)
9244 if (env->log.level & BPF_LOG_STATS) {
9245 verbose(env, "verification time %lld usec\n",
9246 div_u64(env->verification_time, 1000));
9247 verbose(env, "stack depth ");
9248 for (i = 0; i < env->subprog_cnt; i++) {
9249 u32 depth = env->subprog_info[i].stack_depth;
9251 verbose(env, "%d", depth);
9252 if (i + 1 < env->subprog_cnt)
9257 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
9258 "total_states %d peak_states %d mark_read %d\n",
9259 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
9260 env->max_states_per_insn, env->total_states,
9261 env->peak_states, env->longest_mark_read_walk);
9264 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr,
9265 union bpf_attr __user *uattr)
9267 u64 start_time = ktime_get_ns();
9268 struct bpf_verifier_env *env;
9269 struct bpf_verifier_log *log;
9270 int i, len, ret = -EINVAL;
9273 /* no program is valid */
9274 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
9277 /* 'struct bpf_verifier_env' can be global, but since it's not small,
9278 * allocate/free it every time bpf_check() is called
9280 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
9286 env->insn_aux_data =
9287 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
9289 if (!env->insn_aux_data)
9291 for (i = 0; i < len; i++)
9292 env->insn_aux_data[i].orig_idx = i;
9294 env->ops = bpf_verifier_ops[env->prog->type];
9295 is_priv = capable(CAP_SYS_ADMIN);
9297 /* grab the mutex to protect few globals used by verifier */
9299 mutex_lock(&bpf_verifier_lock);
9301 if (attr->log_level || attr->log_buf || attr->log_size) {
9302 /* user requested verbose verifier output
9303 * and supplied buffer to store the verification trace
9305 log->level = attr->log_level;
9306 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
9307 log->len_total = attr->log_size;
9310 /* log attributes have to be sane */
9311 if (log->len_total < 128 || log->len_total > UINT_MAX >> 2 ||
9312 !log->level || !log->ubuf || log->level & ~BPF_LOG_MASK)
9316 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
9317 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
9318 env->strict_alignment = true;
9319 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
9320 env->strict_alignment = false;
9322 env->allow_ptr_leaks = is_priv;
9325 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
9327 ret = replace_map_fd_with_map_ptr(env);
9329 goto skip_full_check;
9331 if (bpf_prog_is_dev_bound(env->prog->aux)) {
9332 ret = bpf_prog_offload_verifier_prep(env->prog);
9334 goto skip_full_check;
9337 env->explored_states = kvcalloc(state_htab_size(env),
9338 sizeof(struct bpf_verifier_state_list *),
9341 if (!env->explored_states)
9342 goto skip_full_check;
9344 ret = check_subprogs(env);
9346 goto skip_full_check;
9348 ret = check_btf_info(env, attr, uattr);
9350 goto skip_full_check;
9352 ret = check_cfg(env);
9354 goto skip_full_check;
9356 ret = do_check(env);
9357 if (env->cur_state) {
9358 free_verifier_state(env->cur_state, true);
9359 env->cur_state = NULL;
9362 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
9363 ret = bpf_prog_offload_finalize(env);
9366 while (!pop_stack(env, NULL, NULL));
9370 ret = check_max_stack_depth(env);
9372 /* instruction rewrites happen after this point */
9375 opt_hard_wire_dead_code_branches(env);
9377 ret = opt_remove_dead_code(env);
9379 ret = opt_remove_nops(env);
9382 sanitize_dead_code(env);
9386 /* program is valid, convert *(u32*)(ctx + off) accesses */
9387 ret = convert_ctx_accesses(env);
9390 ret = fixup_bpf_calls(env);
9392 /* do 32-bit optimization after insn patching has done so those patched
9393 * insns could be handled correctly.
9395 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
9396 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
9397 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
9402 ret = fixup_call_args(env);
9404 env->verification_time = ktime_get_ns() - start_time;
9405 print_verification_stats(env);
9407 if (log->level && bpf_verifier_log_full(log))
9409 if (log->level && !log->ubuf) {
9411 goto err_release_maps;
9414 if (ret == 0 && env->used_map_cnt) {
9415 /* if program passed verifier, update used_maps in bpf_prog_info */
9416 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
9417 sizeof(env->used_maps[0]),
9420 if (!env->prog->aux->used_maps) {
9422 goto err_release_maps;
9425 memcpy(env->prog->aux->used_maps, env->used_maps,
9426 sizeof(env->used_maps[0]) * env->used_map_cnt);
9427 env->prog->aux->used_map_cnt = env->used_map_cnt;
9429 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
9430 * bpf_ld_imm64 instructions
9432 convert_pseudo_ld_imm64(env);
9436 adjust_btf_func(env);
9439 if (!env->prog->aux->used_maps)
9440 /* if we didn't copy map pointers into bpf_prog_info, release
9441 * them now. Otherwise free_used_maps() will release them.
9447 mutex_unlock(&bpf_verifier_lock);
9448 vfree(env->insn_aux_data);