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;
209 static DEFINE_MUTEX(bpf_verifier_lock);
211 static const struct bpf_line_info *
212 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
214 const struct bpf_line_info *linfo;
215 const struct bpf_prog *prog;
219 nr_linfo = prog->aux->nr_linfo;
221 if (!nr_linfo || insn_off >= prog->len)
224 linfo = prog->aux->linfo;
225 for (i = 1; i < nr_linfo; i++)
226 if (insn_off < linfo[i].insn_off)
229 return &linfo[i - 1];
232 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
237 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
239 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
240 "verifier log line truncated - local buffer too short\n");
242 n = min(log->len_total - log->len_used - 1, n);
245 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
251 /* log_level controls verbosity level of eBPF verifier.
252 * bpf_verifier_log_write() is used to dump the verification trace to the log,
253 * so the user can figure out what's wrong with the program
255 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
256 const char *fmt, ...)
260 if (!bpf_verifier_log_needed(&env->log))
264 bpf_verifier_vlog(&env->log, fmt, args);
267 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
269 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
271 struct bpf_verifier_env *env = private_data;
274 if (!bpf_verifier_log_needed(&env->log))
278 bpf_verifier_vlog(&env->log, fmt, args);
282 static const char *ltrim(const char *s)
290 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
292 const char *prefix_fmt, ...)
294 const struct bpf_line_info *linfo;
296 if (!bpf_verifier_log_needed(&env->log))
299 linfo = find_linfo(env, insn_off);
300 if (!linfo || linfo == env->prev_linfo)
306 va_start(args, prefix_fmt);
307 bpf_verifier_vlog(&env->log, prefix_fmt, args);
312 ltrim(btf_name_by_offset(env->prog->aux->btf,
315 env->prev_linfo = linfo;
318 static bool type_is_pkt_pointer(enum bpf_reg_type type)
320 return type == PTR_TO_PACKET ||
321 type == PTR_TO_PACKET_META;
324 static bool type_is_sk_pointer(enum bpf_reg_type type)
326 return type == PTR_TO_SOCKET ||
327 type == PTR_TO_SOCK_COMMON ||
328 type == PTR_TO_TCP_SOCK ||
329 type == PTR_TO_XDP_SOCK;
332 static bool reg_type_may_be_null(enum bpf_reg_type type)
334 return type == PTR_TO_MAP_VALUE_OR_NULL ||
335 type == PTR_TO_SOCKET_OR_NULL ||
336 type == PTR_TO_SOCK_COMMON_OR_NULL ||
337 type == PTR_TO_TCP_SOCK_OR_NULL;
340 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
342 return reg->type == PTR_TO_MAP_VALUE &&
343 map_value_has_spin_lock(reg->map_ptr);
346 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
348 return type == PTR_TO_SOCKET ||
349 type == PTR_TO_SOCKET_OR_NULL ||
350 type == PTR_TO_TCP_SOCK ||
351 type == PTR_TO_TCP_SOCK_OR_NULL;
354 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
356 return type == ARG_PTR_TO_SOCK_COMMON;
359 /* Determine whether the function releases some resources allocated by another
360 * function call. The first reference type argument will be assumed to be
361 * released by release_reference().
363 static bool is_release_function(enum bpf_func_id func_id)
365 return func_id == BPF_FUNC_sk_release;
368 static bool is_acquire_function(enum bpf_func_id func_id)
370 return func_id == BPF_FUNC_sk_lookup_tcp ||
371 func_id == BPF_FUNC_sk_lookup_udp ||
372 func_id == BPF_FUNC_skc_lookup_tcp;
375 static bool is_ptr_cast_function(enum bpf_func_id func_id)
377 return func_id == BPF_FUNC_tcp_sock ||
378 func_id == BPF_FUNC_sk_fullsock;
381 /* string representation of 'enum bpf_reg_type' */
382 static const char * const reg_type_str[] = {
384 [SCALAR_VALUE] = "inv",
385 [PTR_TO_CTX] = "ctx",
386 [CONST_PTR_TO_MAP] = "map_ptr",
387 [PTR_TO_MAP_VALUE] = "map_value",
388 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
389 [PTR_TO_STACK] = "fp",
390 [PTR_TO_PACKET] = "pkt",
391 [PTR_TO_PACKET_META] = "pkt_meta",
392 [PTR_TO_PACKET_END] = "pkt_end",
393 [PTR_TO_FLOW_KEYS] = "flow_keys",
394 [PTR_TO_SOCKET] = "sock",
395 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
396 [PTR_TO_SOCK_COMMON] = "sock_common",
397 [PTR_TO_SOCK_COMMON_OR_NULL] = "sock_common_or_null",
398 [PTR_TO_TCP_SOCK] = "tcp_sock",
399 [PTR_TO_TCP_SOCK_OR_NULL] = "tcp_sock_or_null",
400 [PTR_TO_TP_BUFFER] = "tp_buffer",
401 [PTR_TO_XDP_SOCK] = "xdp_sock",
404 static char slot_type_char[] = {
405 [STACK_INVALID] = '?',
411 static void print_liveness(struct bpf_verifier_env *env,
412 enum bpf_reg_liveness live)
414 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
416 if (live & REG_LIVE_READ)
418 if (live & REG_LIVE_WRITTEN)
420 if (live & REG_LIVE_DONE)
424 static struct bpf_func_state *func(struct bpf_verifier_env *env,
425 const struct bpf_reg_state *reg)
427 struct bpf_verifier_state *cur = env->cur_state;
429 return cur->frame[reg->frameno];
432 static void print_verifier_state(struct bpf_verifier_env *env,
433 const struct bpf_func_state *state)
435 const struct bpf_reg_state *reg;
440 verbose(env, " frame%d:", state->frameno);
441 for (i = 0; i < MAX_BPF_REG; i++) {
442 reg = &state->regs[i];
446 verbose(env, " R%d", i);
447 print_liveness(env, reg->live);
448 verbose(env, "=%s", reg_type_str[t]);
449 if (t == SCALAR_VALUE && reg->precise)
451 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
452 tnum_is_const(reg->var_off)) {
453 /* reg->off should be 0 for SCALAR_VALUE */
454 verbose(env, "%lld", reg->var_off.value + reg->off);
456 verbose(env, "(id=%d", reg->id);
457 if (reg_type_may_be_refcounted_or_null(t))
458 verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
459 if (t != SCALAR_VALUE)
460 verbose(env, ",off=%d", reg->off);
461 if (type_is_pkt_pointer(t))
462 verbose(env, ",r=%d", reg->range);
463 else if (t == CONST_PTR_TO_MAP ||
464 t == PTR_TO_MAP_VALUE ||
465 t == PTR_TO_MAP_VALUE_OR_NULL)
466 verbose(env, ",ks=%d,vs=%d",
467 reg->map_ptr->key_size,
468 reg->map_ptr->value_size);
469 if (tnum_is_const(reg->var_off)) {
470 /* Typically an immediate SCALAR_VALUE, but
471 * could be a pointer whose offset is too big
474 verbose(env, ",imm=%llx", reg->var_off.value);
476 if (reg->smin_value != reg->umin_value &&
477 reg->smin_value != S64_MIN)
478 verbose(env, ",smin_value=%lld",
479 (long long)reg->smin_value);
480 if (reg->smax_value != reg->umax_value &&
481 reg->smax_value != S64_MAX)
482 verbose(env, ",smax_value=%lld",
483 (long long)reg->smax_value);
484 if (reg->umin_value != 0)
485 verbose(env, ",umin_value=%llu",
486 (unsigned long long)reg->umin_value);
487 if (reg->umax_value != U64_MAX)
488 verbose(env, ",umax_value=%llu",
489 (unsigned long long)reg->umax_value);
490 if (!tnum_is_unknown(reg->var_off)) {
493 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
494 verbose(env, ",var_off=%s", tn_buf);
500 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
501 char types_buf[BPF_REG_SIZE + 1];
505 for (j = 0; j < BPF_REG_SIZE; j++) {
506 if (state->stack[i].slot_type[j] != STACK_INVALID)
508 types_buf[j] = slot_type_char[
509 state->stack[i].slot_type[j]];
511 types_buf[BPF_REG_SIZE] = 0;
514 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
515 print_liveness(env, state->stack[i].spilled_ptr.live);
516 if (state->stack[i].slot_type[0] == STACK_SPILL) {
517 reg = &state->stack[i].spilled_ptr;
519 verbose(env, "=%s", reg_type_str[t]);
520 if (t == SCALAR_VALUE && reg->precise)
522 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
523 verbose(env, "%lld", reg->var_off.value + reg->off);
525 verbose(env, "=%s", types_buf);
528 if (state->acquired_refs && state->refs[0].id) {
529 verbose(env, " refs=%d", state->refs[0].id);
530 for (i = 1; i < state->acquired_refs; i++)
531 if (state->refs[i].id)
532 verbose(env, ",%d", state->refs[i].id);
537 #define COPY_STATE_FN(NAME, COUNT, FIELD, SIZE) \
538 static int copy_##NAME##_state(struct bpf_func_state *dst, \
539 const struct bpf_func_state *src) \
543 if (WARN_ON_ONCE(dst->COUNT < src->COUNT)) { \
544 /* internal bug, make state invalid to reject the program */ \
545 memset(dst, 0, sizeof(*dst)); \
548 memcpy(dst->FIELD, src->FIELD, \
549 sizeof(*src->FIELD) * (src->COUNT / SIZE)); \
552 /* copy_reference_state() */
553 COPY_STATE_FN(reference, acquired_refs, refs, 1)
554 /* copy_stack_state() */
555 COPY_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
558 #define REALLOC_STATE_FN(NAME, COUNT, FIELD, SIZE) \
559 static int realloc_##NAME##_state(struct bpf_func_state *state, int size, \
562 u32 old_size = state->COUNT; \
563 struct bpf_##NAME##_state *new_##FIELD; \
564 int slot = size / SIZE; \
566 if (size <= old_size || !size) { \
569 state->COUNT = slot * SIZE; \
570 if (!size && old_size) { \
571 kfree(state->FIELD); \
572 state->FIELD = NULL; \
576 new_##FIELD = kmalloc_array(slot, sizeof(struct bpf_##NAME##_state), \
582 memcpy(new_##FIELD, state->FIELD, \
583 sizeof(*new_##FIELD) * (old_size / SIZE)); \
584 memset(new_##FIELD + old_size / SIZE, 0, \
585 sizeof(*new_##FIELD) * (size - old_size) / SIZE); \
587 state->COUNT = slot * SIZE; \
588 kfree(state->FIELD); \
589 state->FIELD = new_##FIELD; \
592 /* realloc_reference_state() */
593 REALLOC_STATE_FN(reference, acquired_refs, refs, 1)
594 /* realloc_stack_state() */
595 REALLOC_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
596 #undef REALLOC_STATE_FN
598 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
599 * make it consume minimal amount of memory. check_stack_write() access from
600 * the program calls into realloc_func_state() to grow the stack size.
601 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
602 * which realloc_stack_state() copies over. It points to previous
603 * bpf_verifier_state which is never reallocated.
605 static int realloc_func_state(struct bpf_func_state *state, int stack_size,
606 int refs_size, bool copy_old)
608 int err = realloc_reference_state(state, refs_size, copy_old);
611 return realloc_stack_state(state, stack_size, copy_old);
614 /* Acquire a pointer id from the env and update the state->refs to include
615 * this new pointer reference.
616 * On success, returns a valid pointer id to associate with the register
617 * On failure, returns a negative errno.
619 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
621 struct bpf_func_state *state = cur_func(env);
622 int new_ofs = state->acquired_refs;
625 err = realloc_reference_state(state, state->acquired_refs + 1, true);
629 state->refs[new_ofs].id = id;
630 state->refs[new_ofs].insn_idx = insn_idx;
635 /* release function corresponding to acquire_reference_state(). Idempotent. */
636 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
640 last_idx = state->acquired_refs - 1;
641 for (i = 0; i < state->acquired_refs; i++) {
642 if (state->refs[i].id == ptr_id) {
643 if (last_idx && i != last_idx)
644 memcpy(&state->refs[i], &state->refs[last_idx],
645 sizeof(*state->refs));
646 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
647 state->acquired_refs--;
654 static int transfer_reference_state(struct bpf_func_state *dst,
655 struct bpf_func_state *src)
657 int err = realloc_reference_state(dst, src->acquired_refs, false);
660 err = copy_reference_state(dst, src);
666 static void free_func_state(struct bpf_func_state *state)
675 static void clear_jmp_history(struct bpf_verifier_state *state)
677 kfree(state->jmp_history);
678 state->jmp_history = NULL;
679 state->jmp_history_cnt = 0;
682 static void free_verifier_state(struct bpf_verifier_state *state,
687 for (i = 0; i <= state->curframe; i++) {
688 free_func_state(state->frame[i]);
689 state->frame[i] = NULL;
691 clear_jmp_history(state);
696 /* copy verifier state from src to dst growing dst stack space
697 * when necessary to accommodate larger src stack
699 static int copy_func_state(struct bpf_func_state *dst,
700 const struct bpf_func_state *src)
704 err = realloc_func_state(dst, src->allocated_stack, src->acquired_refs,
708 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
709 err = copy_reference_state(dst, src);
712 return copy_stack_state(dst, src);
715 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
716 const struct bpf_verifier_state *src)
718 struct bpf_func_state *dst;
719 u32 jmp_sz = sizeof(struct bpf_idx_pair) * src->jmp_history_cnt;
722 if (dst_state->jmp_history_cnt < src->jmp_history_cnt) {
723 kfree(dst_state->jmp_history);
724 dst_state->jmp_history = kmalloc(jmp_sz, GFP_USER);
725 if (!dst_state->jmp_history)
728 memcpy(dst_state->jmp_history, src->jmp_history, jmp_sz);
729 dst_state->jmp_history_cnt = src->jmp_history_cnt;
731 /* if dst has more stack frames then src frame, free them */
732 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
733 free_func_state(dst_state->frame[i]);
734 dst_state->frame[i] = NULL;
736 dst_state->speculative = src->speculative;
737 dst_state->curframe = src->curframe;
738 dst_state->active_spin_lock = src->active_spin_lock;
739 dst_state->branches = src->branches;
740 dst_state->parent = src->parent;
741 dst_state->first_insn_idx = src->first_insn_idx;
742 dst_state->last_insn_idx = src->last_insn_idx;
743 for (i = 0; i <= src->curframe; i++) {
744 dst = dst_state->frame[i];
746 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
749 dst_state->frame[i] = dst;
751 err = copy_func_state(dst, src->frame[i]);
758 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
761 u32 br = --st->branches;
763 /* WARN_ON(br > 1) technically makes sense here,
764 * but see comment in push_stack(), hence:
766 WARN_ONCE((int)br < 0,
767 "BUG update_branch_counts:branches_to_explore=%d\n",
775 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
778 struct bpf_verifier_state *cur = env->cur_state;
779 struct bpf_verifier_stack_elem *elem, *head = env->head;
782 if (env->head == NULL)
786 err = copy_verifier_state(cur, &head->st);
791 *insn_idx = head->insn_idx;
793 *prev_insn_idx = head->prev_insn_idx;
795 free_verifier_state(&head->st, false);
802 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
803 int insn_idx, int prev_insn_idx,
806 struct bpf_verifier_state *cur = env->cur_state;
807 struct bpf_verifier_stack_elem *elem;
810 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
814 elem->insn_idx = insn_idx;
815 elem->prev_insn_idx = prev_insn_idx;
816 elem->next = env->head;
819 err = copy_verifier_state(&elem->st, cur);
822 elem->st.speculative |= speculative;
823 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
824 verbose(env, "The sequence of %d jumps is too complex.\n",
828 if (elem->st.parent) {
829 ++elem->st.parent->branches;
830 /* WARN_ON(branches > 2) technically makes sense here,
832 * 1. speculative states will bump 'branches' for non-branch
834 * 2. is_state_visited() heuristics may decide not to create
835 * a new state for a sequence of branches and all such current
836 * and cloned states will be pointing to a single parent state
837 * which might have large 'branches' count.
842 free_verifier_state(env->cur_state, true);
843 env->cur_state = NULL;
844 /* pop all elements and return */
845 while (!pop_stack(env, NULL, NULL));
849 #define CALLER_SAVED_REGS 6
850 static const int caller_saved[CALLER_SAVED_REGS] = {
851 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
854 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
855 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(env, 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(const struct bpf_verifier_env *env,
992 struct bpf_reg_state *reg)
995 * Clear type, id, off, and union(map_ptr, range) and
996 * padding between 'type' and union
998 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
999 reg->type = SCALAR_VALUE;
1000 reg->var_off = tnum_unknown;
1002 reg->precise = env->subprog_cnt > 1 || !env->allow_ptr_leaks ?
1004 __mark_reg_unbounded(reg);
1007 static void mark_reg_unknown(struct bpf_verifier_env *env,
1008 struct bpf_reg_state *regs, u32 regno)
1010 if (WARN_ON(regno >= MAX_BPF_REG)) {
1011 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1012 /* Something bad happened, let's kill all regs except FP */
1013 for (regno = 0; regno < BPF_REG_FP; regno++)
1014 __mark_reg_not_init(env, regs + regno);
1017 __mark_reg_unknown(env, regs + regno);
1020 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1021 struct bpf_reg_state *reg)
1023 __mark_reg_unknown(env, 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(env, regs + regno);
1037 __mark_reg_not_init(env, 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 (code == (BPF_JMP | BPF_CALL) &&
1164 insn[i].imm == BPF_FUNC_tail_call &&
1165 insn[i].src_reg != BPF_PSEUDO_CALL)
1166 subprog[cur_subprog].has_tail_call = true;
1167 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
1169 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1171 off = i + insn[i].off + 1;
1172 if (off < subprog_start || off >= subprog_end) {
1173 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1177 if (i == subprog_end - 1) {
1178 /* to avoid fall-through from one subprog into another
1179 * the last insn of the subprog should be either exit
1180 * or unconditional jump back
1182 if (code != (BPF_JMP | BPF_EXIT) &&
1183 code != (BPF_JMP | BPF_JA)) {
1184 verbose(env, "last insn is not an exit or jmp\n");
1187 subprog_start = subprog_end;
1189 if (cur_subprog < env->subprog_cnt)
1190 subprog_end = subprog[cur_subprog + 1].start;
1196 /* Parentage chain of this register (or stack slot) should take care of all
1197 * issues like callee-saved registers, stack slot allocation time, etc.
1199 static int mark_reg_read(struct bpf_verifier_env *env,
1200 const struct bpf_reg_state *state,
1201 struct bpf_reg_state *parent, u8 flag)
1203 bool writes = parent == state->parent; /* Observe write marks */
1207 /* if read wasn't screened by an earlier write ... */
1208 if (writes && state->live & REG_LIVE_WRITTEN)
1210 if (parent->live & REG_LIVE_DONE) {
1211 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1212 reg_type_str[parent->type],
1213 parent->var_off.value, parent->off);
1216 /* The first condition is more likely to be true than the
1217 * second, checked it first.
1219 if ((parent->live & REG_LIVE_READ) == flag ||
1220 parent->live & REG_LIVE_READ64)
1221 /* The parentage chain never changes and
1222 * this parent was already marked as LIVE_READ.
1223 * There is no need to keep walking the chain again and
1224 * keep re-marking all parents as LIVE_READ.
1225 * This case happens when the same register is read
1226 * multiple times without writes into it in-between.
1227 * Also, if parent has the stronger REG_LIVE_READ64 set,
1228 * then no need to set the weak REG_LIVE_READ32.
1231 /* ... then we depend on parent's value */
1232 parent->live |= flag;
1233 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
1234 if (flag == REG_LIVE_READ64)
1235 parent->live &= ~REG_LIVE_READ32;
1237 parent = state->parent;
1242 if (env->longest_mark_read_walk < cnt)
1243 env->longest_mark_read_walk = cnt;
1247 /* This function is supposed to be used by the following 32-bit optimization
1248 * code only. It returns TRUE if the source or destination register operates
1249 * on 64-bit, otherwise return FALSE.
1251 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
1252 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
1257 class = BPF_CLASS(code);
1259 if (class == BPF_JMP) {
1260 /* BPF_EXIT for "main" will reach here. Return TRUE
1265 if (op == BPF_CALL) {
1266 /* BPF to BPF call will reach here because of marking
1267 * caller saved clobber with DST_OP_NO_MARK for which we
1268 * don't care the register def because they are anyway
1269 * marked as NOT_INIT already.
1271 if (insn->src_reg == BPF_PSEUDO_CALL)
1273 /* Helper call will reach here because of arg type
1274 * check, conservatively return TRUE.
1283 if (class == BPF_ALU64 || class == BPF_JMP ||
1284 /* BPF_END always use BPF_ALU class. */
1285 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
1288 if (class == BPF_ALU || class == BPF_JMP32)
1291 if (class == BPF_LDX) {
1293 return BPF_SIZE(code) == BPF_DW;
1294 /* LDX source must be ptr. */
1298 if (class == BPF_STX) {
1299 if (reg->type != SCALAR_VALUE)
1301 return BPF_SIZE(code) == BPF_DW;
1304 if (class == BPF_LD) {
1305 u8 mode = BPF_MODE(code);
1308 if (mode == BPF_IMM)
1311 /* Both LD_IND and LD_ABS return 32-bit data. */
1315 /* Implicit ctx ptr. */
1316 if (regno == BPF_REG_6)
1319 /* Explicit source could be any width. */
1323 if (class == BPF_ST)
1324 /* The only source register for BPF_ST is a ptr. */
1327 /* Conservatively return true at default. */
1331 /* Return TRUE if INSN doesn't have explicit value define. */
1332 static bool insn_no_def(struct bpf_insn *insn)
1334 u8 class = BPF_CLASS(insn->code);
1336 return (class == BPF_JMP || class == BPF_JMP32 ||
1337 class == BPF_STX || class == BPF_ST);
1340 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
1341 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
1343 if (insn_no_def(insn))
1346 return !is_reg64(env, insn, insn->dst_reg, NULL, DST_OP);
1349 static void mark_insn_zext(struct bpf_verifier_env *env,
1350 struct bpf_reg_state *reg)
1352 s32 def_idx = reg->subreg_def;
1354 if (def_idx == DEF_NOT_SUBREG)
1357 env->insn_aux_data[def_idx - 1].zext_dst = true;
1358 /* The dst will be zero extended, so won't be sub-register anymore. */
1359 reg->subreg_def = DEF_NOT_SUBREG;
1362 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
1363 enum reg_arg_type t)
1365 struct bpf_verifier_state *vstate = env->cur_state;
1366 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1367 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
1368 struct bpf_reg_state *reg, *regs = state->regs;
1371 if (regno >= MAX_BPF_REG) {
1372 verbose(env, "R%d is invalid\n", regno);
1377 rw64 = is_reg64(env, insn, regno, reg, t);
1379 /* check whether register used as source operand can be read */
1380 if (reg->type == NOT_INIT) {
1381 verbose(env, "R%d !read_ok\n", regno);
1384 /* We don't need to worry about FP liveness because it's read-only */
1385 if (regno == BPF_REG_FP)
1389 mark_insn_zext(env, reg);
1391 return mark_reg_read(env, reg, reg->parent,
1392 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
1394 /* check whether register used as dest operand can be written to */
1395 if (regno == BPF_REG_FP) {
1396 verbose(env, "frame pointer is read only\n");
1399 reg->live |= REG_LIVE_WRITTEN;
1400 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
1402 mark_reg_unknown(env, regs, regno);
1407 /* for any branch, call, exit record the history of jmps in the given state */
1408 static int push_jmp_history(struct bpf_verifier_env *env,
1409 struct bpf_verifier_state *cur)
1411 u32 cnt = cur->jmp_history_cnt;
1412 struct bpf_idx_pair *p;
1415 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
1418 p[cnt - 1].idx = env->insn_idx;
1419 p[cnt - 1].prev_idx = env->prev_insn_idx;
1420 cur->jmp_history = p;
1421 cur->jmp_history_cnt = cnt;
1425 /* Backtrack one insn at a time. If idx is not at the top of recorded
1426 * history then previous instruction came from straight line execution.
1428 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
1433 if (cnt && st->jmp_history[cnt - 1].idx == i) {
1434 i = st->jmp_history[cnt - 1].prev_idx;
1442 /* For given verifier state backtrack_insn() is called from the last insn to
1443 * the first insn. Its purpose is to compute a bitmask of registers and
1444 * stack slots that needs precision in the parent verifier state.
1446 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
1447 u32 *reg_mask, u64 *stack_mask)
1449 const struct bpf_insn_cbs cbs = {
1450 .cb_print = verbose,
1451 .private_data = env,
1453 struct bpf_insn *insn = env->prog->insnsi + idx;
1454 u8 class = BPF_CLASS(insn->code);
1455 u8 opcode = BPF_OP(insn->code);
1456 u8 mode = BPF_MODE(insn->code);
1457 u32 dreg = 1u << insn->dst_reg;
1458 u32 sreg = 1u << insn->src_reg;
1461 if (insn->code == 0)
1463 if (env->log.level & BPF_LOG_LEVEL) {
1464 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
1465 verbose(env, "%d: ", idx);
1466 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
1469 if (class == BPF_ALU || class == BPF_ALU64) {
1470 if (!(*reg_mask & dreg))
1472 if (opcode == BPF_MOV) {
1473 if (BPF_SRC(insn->code) == BPF_X) {
1475 * dreg needs precision after this insn
1476 * sreg needs precision before this insn
1482 * dreg needs precision after this insn.
1483 * Corresponding register is already marked
1484 * as precise=true in this verifier state.
1485 * No further markings in parent are necessary
1490 if (BPF_SRC(insn->code) == BPF_X) {
1492 * both dreg and sreg need precision
1497 * dreg still needs precision before this insn
1500 } else if (class == BPF_LDX) {
1501 if (!(*reg_mask & dreg))
1505 /* scalars can only be spilled into stack w/o losing precision.
1506 * Load from any other memory can be zero extended.
1507 * The desire to keep that precision is already indicated
1508 * by 'precise' mark in corresponding register of this state.
1509 * No further tracking necessary.
1511 if (insn->src_reg != BPF_REG_FP)
1513 if (BPF_SIZE(insn->code) != BPF_DW)
1516 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
1517 * that [fp - off] slot contains scalar that needs to be
1518 * tracked with precision
1520 spi = (-insn->off - 1) / BPF_REG_SIZE;
1522 verbose(env, "BUG spi %d\n", spi);
1523 WARN_ONCE(1, "verifier backtracking bug");
1526 *stack_mask |= 1ull << spi;
1527 } else if (class == BPF_STX || class == BPF_ST) {
1528 if (*reg_mask & dreg)
1529 /* stx & st shouldn't be using _scalar_ dst_reg
1530 * to access memory. It means backtracking
1531 * encountered a case of pointer subtraction.
1534 /* scalars can only be spilled into stack */
1535 if (insn->dst_reg != BPF_REG_FP)
1537 if (BPF_SIZE(insn->code) != BPF_DW)
1539 spi = (-insn->off - 1) / BPF_REG_SIZE;
1541 verbose(env, "BUG spi %d\n", spi);
1542 WARN_ONCE(1, "verifier backtracking bug");
1545 if (!(*stack_mask & (1ull << spi)))
1547 *stack_mask &= ~(1ull << spi);
1548 if (class == BPF_STX)
1550 } else if (class == BPF_JMP || class == BPF_JMP32) {
1551 if (opcode == BPF_CALL) {
1552 if (insn->src_reg == BPF_PSEUDO_CALL)
1554 /* regular helper call sets R0 */
1556 if (*reg_mask & 0x3f) {
1557 /* if backtracing was looking for registers R1-R5
1558 * they should have been found already.
1560 verbose(env, "BUG regs %x\n", *reg_mask);
1561 WARN_ONCE(1, "verifier backtracking bug");
1564 } else if (opcode == BPF_EXIT) {
1567 } else if (class == BPF_LD) {
1568 if (!(*reg_mask & dreg))
1571 /* It's ld_imm64 or ld_abs or ld_ind.
1572 * For ld_imm64 no further tracking of precision
1573 * into parent is necessary
1575 if (mode == BPF_IND || mode == BPF_ABS)
1576 /* to be analyzed */
1582 /* the scalar precision tracking algorithm:
1583 * . at the start all registers have precise=false.
1584 * . scalar ranges are tracked as normal through alu and jmp insns.
1585 * . once precise value of the scalar register is used in:
1586 * . ptr + scalar alu
1587 * . if (scalar cond K|scalar)
1588 * . helper_call(.., scalar, ...) where ARG_CONST is expected
1589 * backtrack through the verifier states and mark all registers and
1590 * stack slots with spilled constants that these scalar regisers
1591 * should be precise.
1592 * . during state pruning two registers (or spilled stack slots)
1593 * are equivalent if both are not precise.
1595 * Note the verifier cannot simply walk register parentage chain,
1596 * since many different registers and stack slots could have been
1597 * used to compute single precise scalar.
1599 * The approach of starting with precise=true for all registers and then
1600 * backtrack to mark a register as not precise when the verifier detects
1601 * that program doesn't care about specific value (e.g., when helper
1602 * takes register as ARG_ANYTHING parameter) is not safe.
1604 * It's ok to walk single parentage chain of the verifier states.
1605 * It's possible that this backtracking will go all the way till 1st insn.
1606 * All other branches will be explored for needing precision later.
1608 * The backtracking needs to deal with cases like:
1609 * 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)
1612 * if r5 > 0x79f goto pc+7
1613 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
1616 * call bpf_perf_event_output#25
1617 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
1621 * call foo // uses callee's r6 inside to compute r0
1625 * to track above reg_mask/stack_mask needs to be independent for each frame.
1627 * Also if parent's curframe > frame where backtracking started,
1628 * the verifier need to mark registers in both frames, otherwise callees
1629 * may incorrectly prune callers. This is similar to
1630 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
1632 * For now backtracking falls back into conservative marking.
1634 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
1635 struct bpf_verifier_state *st)
1637 struct bpf_func_state *func;
1638 struct bpf_reg_state *reg;
1641 /* big hammer: mark all scalars precise in this path.
1642 * pop_stack may still get !precise scalars.
1644 for (; st; st = st->parent)
1645 for (i = 0; i <= st->curframe; i++) {
1646 func = st->frame[i];
1647 for (j = 0; j < BPF_REG_FP; j++) {
1648 reg = &func->regs[j];
1649 if (reg->type != SCALAR_VALUE)
1651 reg->precise = true;
1653 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
1654 if (func->stack[j].slot_type[0] != STACK_SPILL)
1656 reg = &func->stack[j].spilled_ptr;
1657 if (reg->type != SCALAR_VALUE)
1659 reg->precise = true;
1664 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
1667 struct bpf_verifier_state *st = env->cur_state;
1668 int first_idx = st->first_insn_idx;
1669 int last_idx = env->insn_idx;
1670 struct bpf_func_state *func;
1671 struct bpf_reg_state *reg;
1672 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
1673 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
1674 bool skip_first = true;
1675 bool new_marks = false;
1678 if (!env->allow_ptr_leaks)
1679 /* backtracking is root only for now */
1682 func = st->frame[st->curframe];
1684 reg = &func->regs[regno];
1685 if (reg->type != SCALAR_VALUE) {
1686 WARN_ONCE(1, "backtracing misuse");
1693 reg->precise = true;
1697 if (func->stack[spi].slot_type[0] != STACK_SPILL) {
1701 reg = &func->stack[spi].spilled_ptr;
1702 if (reg->type != SCALAR_VALUE) {
1710 reg->precise = true;
1716 if (!reg_mask && !stack_mask)
1719 DECLARE_BITMAP(mask, 64);
1720 u32 history = st->jmp_history_cnt;
1722 if (env->log.level & BPF_LOG_LEVEL)
1723 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
1724 for (i = last_idx;;) {
1729 err = backtrack_insn(env, i, ®_mask, &stack_mask);
1731 if (err == -ENOTSUPP) {
1732 mark_all_scalars_precise(env, st);
1737 if (!reg_mask && !stack_mask)
1738 /* Found assignment(s) into tracked register in this state.
1739 * Since this state is already marked, just return.
1740 * Nothing to be tracked further in the parent state.
1745 i = get_prev_insn_idx(st, i, &history);
1746 if (i >= env->prog->len) {
1747 /* This can happen if backtracking reached insn 0
1748 * and there are still reg_mask or stack_mask
1750 * It means the backtracking missed the spot where
1751 * particular register was initialized with a constant.
1753 verbose(env, "BUG backtracking idx %d\n", i);
1754 WARN_ONCE(1, "verifier backtracking bug");
1763 func = st->frame[st->curframe];
1764 bitmap_from_u64(mask, reg_mask);
1765 for_each_set_bit(i, mask, 32) {
1766 reg = &func->regs[i];
1767 if (reg->type != SCALAR_VALUE) {
1768 reg_mask &= ~(1u << i);
1773 reg->precise = true;
1776 bitmap_from_u64(mask, stack_mask);
1777 for_each_set_bit(i, mask, 64) {
1778 if (i >= func->allocated_stack / BPF_REG_SIZE) {
1779 /* the sequence of instructions:
1781 * 3: (7b) *(u64 *)(r3 -8) = r0
1782 * 4: (79) r4 = *(u64 *)(r10 -8)
1783 * doesn't contain jmps. It's backtracked
1784 * as a single block.
1785 * During backtracking insn 3 is not recognized as
1786 * stack access, so at the end of backtracking
1787 * stack slot fp-8 is still marked in stack_mask.
1788 * However the parent state may not have accessed
1789 * fp-8 and it's "unallocated" stack space.
1790 * In such case fallback to conservative.
1792 mark_all_scalars_precise(env, st);
1796 if (func->stack[i].slot_type[0] != STACK_SPILL) {
1797 stack_mask &= ~(1ull << i);
1800 reg = &func->stack[i].spilled_ptr;
1801 if (reg->type != SCALAR_VALUE) {
1802 stack_mask &= ~(1ull << i);
1807 reg->precise = true;
1809 if (env->log.level & BPF_LOG_LEVEL) {
1810 print_verifier_state(env, func);
1811 verbose(env, "parent %s regs=%x stack=%llx marks\n",
1812 new_marks ? "didn't have" : "already had",
1813 reg_mask, stack_mask);
1816 if (!reg_mask && !stack_mask)
1821 last_idx = st->last_insn_idx;
1822 first_idx = st->first_insn_idx;
1827 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
1829 return __mark_chain_precision(env, regno, -1);
1832 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
1834 return __mark_chain_precision(env, -1, spi);
1837 static bool is_spillable_regtype(enum bpf_reg_type type)
1840 case PTR_TO_MAP_VALUE:
1841 case PTR_TO_MAP_VALUE_OR_NULL:
1845 case PTR_TO_PACKET_META:
1846 case PTR_TO_PACKET_END:
1847 case PTR_TO_FLOW_KEYS:
1848 case CONST_PTR_TO_MAP:
1850 case PTR_TO_SOCKET_OR_NULL:
1851 case PTR_TO_SOCK_COMMON:
1852 case PTR_TO_SOCK_COMMON_OR_NULL:
1853 case PTR_TO_TCP_SOCK:
1854 case PTR_TO_TCP_SOCK_OR_NULL:
1855 case PTR_TO_XDP_SOCK:
1862 /* Does this register contain a constant zero? */
1863 static bool register_is_null(struct bpf_reg_state *reg)
1865 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
1868 static bool register_is_const(struct bpf_reg_state *reg)
1870 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
1873 static bool __is_pointer_value(bool allow_ptr_leaks,
1874 const struct bpf_reg_state *reg)
1876 if (allow_ptr_leaks)
1879 return reg->type != SCALAR_VALUE;
1882 static void save_register_state(struct bpf_func_state *state,
1883 int spi, struct bpf_reg_state *reg)
1887 state->stack[spi].spilled_ptr = *reg;
1888 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1890 for (i = 0; i < BPF_REG_SIZE; i++)
1891 state->stack[spi].slot_type[i] = STACK_SPILL;
1894 /* check_stack_read/write functions track spill/fill of registers,
1895 * stack boundary and alignment are checked in check_mem_access()
1897 static int check_stack_write(struct bpf_verifier_env *env,
1898 struct bpf_func_state *state, /* func where register points to */
1899 int off, int size, int value_regno, int insn_idx)
1901 struct bpf_func_state *cur; /* state of the current function */
1902 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
1903 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
1904 struct bpf_reg_state *reg = NULL;
1906 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
1907 state->acquired_refs, true);
1910 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
1911 * so it's aligned access and [off, off + size) are within stack limits
1913 if (!env->allow_ptr_leaks &&
1914 state->stack[spi].slot_type[0] == STACK_SPILL &&
1915 size != BPF_REG_SIZE) {
1916 verbose(env, "attempt to corrupt spilled pointer on stack\n");
1920 cur = env->cur_state->frame[env->cur_state->curframe];
1921 if (value_regno >= 0)
1922 reg = &cur->regs[value_regno];
1923 if (!env->allow_ptr_leaks) {
1924 bool sanitize = reg && is_spillable_regtype(reg->type);
1926 for (i = 0; i < size; i++) {
1927 if (state->stack[spi].slot_type[i] == STACK_INVALID) {
1934 env->insn_aux_data[insn_idx].sanitize_stack_spill = true;
1937 if (reg && size == BPF_REG_SIZE && register_is_const(reg) &&
1938 !register_is_null(reg) && env->allow_ptr_leaks) {
1939 if (dst_reg != BPF_REG_FP) {
1940 /* The backtracking logic can only recognize explicit
1941 * stack slot address like [fp - 8]. Other spill of
1942 * scalar via different register has to be conervative.
1943 * Backtrack from here and mark all registers as precise
1944 * that contributed into 'reg' being a constant.
1946 err = mark_chain_precision(env, value_regno);
1950 save_register_state(state, spi, reg);
1951 } else if (reg && is_spillable_regtype(reg->type)) {
1952 /* register containing pointer is being spilled into stack */
1953 if (size != BPF_REG_SIZE) {
1954 verbose_linfo(env, insn_idx, "; ");
1955 verbose(env, "invalid size of register spill\n");
1958 if (state != cur && reg->type == PTR_TO_STACK) {
1959 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
1962 save_register_state(state, spi, reg);
1964 u8 type = STACK_MISC;
1966 /* regular write of data into stack destroys any spilled ptr */
1967 state->stack[spi].spilled_ptr.type = NOT_INIT;
1968 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
1969 if (state->stack[spi].slot_type[0] == STACK_SPILL)
1970 for (i = 0; i < BPF_REG_SIZE; i++)
1971 state->stack[spi].slot_type[i] = STACK_MISC;
1973 /* only mark the slot as written if all 8 bytes were written
1974 * otherwise read propagation may incorrectly stop too soon
1975 * when stack slots are partially written.
1976 * This heuristic means that read propagation will be
1977 * conservative, since it will add reg_live_read marks
1978 * to stack slots all the way to first state when programs
1979 * writes+reads less than 8 bytes
1981 if (size == BPF_REG_SIZE)
1982 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1984 /* when we zero initialize stack slots mark them as such */
1985 if (reg && register_is_null(reg)) {
1986 /* backtracking doesn't work for STACK_ZERO yet. */
1987 err = mark_chain_precision(env, value_regno);
1993 /* Mark slots affected by this stack write. */
1994 for (i = 0; i < size; i++)
1995 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
2001 static int check_stack_read(struct bpf_verifier_env *env,
2002 struct bpf_func_state *reg_state /* func where register points to */,
2003 int off, int size, int value_regno)
2005 struct bpf_verifier_state *vstate = env->cur_state;
2006 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2007 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
2008 struct bpf_reg_state *reg;
2011 if (reg_state->allocated_stack <= slot) {
2012 verbose(env, "invalid read from stack off %d+0 size %d\n",
2016 stype = reg_state->stack[spi].slot_type;
2017 reg = ®_state->stack[spi].spilled_ptr;
2019 if (stype[0] == STACK_SPILL) {
2020 if (size != BPF_REG_SIZE) {
2021 if (reg->type != SCALAR_VALUE) {
2022 verbose_linfo(env, env->insn_idx, "; ");
2023 verbose(env, "invalid size of register fill\n");
2026 if (value_regno >= 0) {
2027 mark_reg_unknown(env, state->regs, value_regno);
2028 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2030 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2033 for (i = 1; i < BPF_REG_SIZE; i++) {
2034 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
2035 verbose(env, "corrupted spill memory\n");
2040 if (value_regno >= 0) {
2041 /* restore register state from stack */
2042 state->regs[value_regno] = *reg;
2043 /* mark reg as written since spilled pointer state likely
2044 * has its liveness marks cleared by is_state_visited()
2045 * which resets stack/reg liveness for state transitions
2047 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2048 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
2049 /* If value_regno==-1, the caller is asking us whether
2050 * it is acceptable to use this value as a SCALAR_VALUE
2052 * We must not allow unprivileged callers to do that
2053 * with spilled pointers.
2055 verbose(env, "leaking pointer from stack off %d\n",
2059 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2063 for (i = 0; i < size; i++) {
2064 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
2066 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
2070 verbose(env, "invalid read from stack off %d+%d size %d\n",
2074 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2075 if (value_regno >= 0) {
2076 if (zeros == size) {
2077 /* any size read into register is zero extended,
2078 * so the whole register == const_zero
2080 __mark_reg_const_zero(&state->regs[value_regno]);
2081 /* backtracking doesn't support STACK_ZERO yet,
2082 * so mark it precise here, so that later
2083 * backtracking can stop here.
2084 * Backtracking may not need this if this register
2085 * doesn't participate in pointer adjustment.
2086 * Forward propagation of precise flag is not
2087 * necessary either. This mark is only to stop
2088 * backtracking. Any register that contributed
2089 * to const 0 was marked precise before spill.
2091 state->regs[value_regno].precise = true;
2093 /* have read misc data from the stack */
2094 mark_reg_unknown(env, state->regs, value_regno);
2096 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2102 static int check_stack_access(struct bpf_verifier_env *env,
2103 const struct bpf_reg_state *reg,
2106 /* Stack accesses must be at a fixed offset, so that we
2107 * can determine what type of data were returned. See
2108 * check_stack_read().
2110 if (!tnum_is_const(reg->var_off)) {
2113 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2114 verbose(env, "variable stack access var_off=%s off=%d size=%d\n",
2119 if (off >= 0 || off < -MAX_BPF_STACK) {
2120 verbose(env, "invalid stack off=%d size=%d\n", off, size);
2127 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
2128 int off, int size, enum bpf_access_type type)
2130 struct bpf_reg_state *regs = cur_regs(env);
2131 struct bpf_map *map = regs[regno].map_ptr;
2132 u32 cap = bpf_map_flags_to_cap(map);
2134 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
2135 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
2136 map->value_size, off, size);
2140 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
2141 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
2142 map->value_size, off, size);
2149 /* check read/write into map element returned by bpf_map_lookup_elem() */
2150 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
2151 int size, bool zero_size_allowed)
2153 struct bpf_reg_state *regs = cur_regs(env);
2154 struct bpf_map *map = regs[regno].map_ptr;
2156 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
2157 off + size > map->value_size) {
2158 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
2159 map->value_size, off, size);
2165 /* check read/write into a map element with possible variable offset */
2166 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
2167 int off, int size, bool zero_size_allowed)
2169 struct bpf_verifier_state *vstate = env->cur_state;
2170 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2171 struct bpf_reg_state *reg = &state->regs[regno];
2174 /* We may have adjusted the register to this map value, so we
2175 * need to try adding each of min_value and max_value to off
2176 * to make sure our theoretical access will be safe.
2178 if (env->log.level & BPF_LOG_LEVEL)
2179 print_verifier_state(env, state);
2181 /* The minimum value is only important with signed
2182 * comparisons where we can't assume the floor of a
2183 * value is 0. If we are using signed variables for our
2184 * index'es we need to make sure that whatever we use
2185 * will have a set floor within our range.
2187 if (reg->smin_value < 0 &&
2188 (reg->smin_value == S64_MIN ||
2189 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
2190 reg->smin_value + off < 0)) {
2191 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2195 err = __check_map_access(env, regno, reg->smin_value + off, size,
2198 verbose(env, "R%d min value is outside of the array range\n",
2203 /* If we haven't set a max value then we need to bail since we can't be
2204 * sure we won't do bad things.
2205 * If reg->umax_value + off could overflow, treat that as unbounded too.
2207 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
2208 verbose(env, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
2212 err = __check_map_access(env, regno, reg->umax_value + off, size,
2215 verbose(env, "R%d max value is outside of the array range\n",
2218 if (map_value_has_spin_lock(reg->map_ptr)) {
2219 u32 lock = reg->map_ptr->spin_lock_off;
2221 /* if any part of struct bpf_spin_lock can be touched by
2222 * load/store reject this program.
2223 * To check that [x1, x2) overlaps with [y1, y2)
2224 * it is sufficient to check x1 < y2 && y1 < x2.
2226 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
2227 lock < reg->umax_value + off + size) {
2228 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
2235 #define MAX_PACKET_OFF 0xffff
2237 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
2238 const struct bpf_call_arg_meta *meta,
2239 enum bpf_access_type t)
2241 switch (env->prog->type) {
2242 /* Program types only with direct read access go here! */
2243 case BPF_PROG_TYPE_LWT_IN:
2244 case BPF_PROG_TYPE_LWT_OUT:
2245 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
2246 case BPF_PROG_TYPE_SK_REUSEPORT:
2247 case BPF_PROG_TYPE_FLOW_DISSECTOR:
2248 case BPF_PROG_TYPE_CGROUP_SKB:
2253 /* Program types with direct read + write access go here! */
2254 case BPF_PROG_TYPE_SCHED_CLS:
2255 case BPF_PROG_TYPE_SCHED_ACT:
2256 case BPF_PROG_TYPE_XDP:
2257 case BPF_PROG_TYPE_LWT_XMIT:
2258 case BPF_PROG_TYPE_SK_SKB:
2259 case BPF_PROG_TYPE_SK_MSG:
2261 return meta->pkt_access;
2263 env->seen_direct_write = true;
2266 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
2268 env->seen_direct_write = true;
2277 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
2278 int off, int size, bool zero_size_allowed)
2280 struct bpf_reg_state *regs = cur_regs(env);
2281 struct bpf_reg_state *reg = ®s[regno];
2283 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
2284 (u64)off + size > reg->range) {
2285 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
2286 off, size, regno, reg->id, reg->off, reg->range);
2292 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
2293 int size, bool zero_size_allowed)
2295 struct bpf_reg_state *regs = cur_regs(env);
2296 struct bpf_reg_state *reg = ®s[regno];
2299 /* We may have added a variable offset to the packet pointer; but any
2300 * reg->range we have comes after that. We are only checking the fixed
2304 /* We don't allow negative numbers, because we aren't tracking enough
2305 * detail to prove they're safe.
2307 if (reg->smin_value < 0) {
2308 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2312 err = __check_packet_access(env, regno, off, size, zero_size_allowed);
2314 verbose(env, "R%d offset is outside of the packet\n", regno);
2318 /* __check_packet_access has made sure "off + size - 1" is within u16.
2319 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
2320 * otherwise find_good_pkt_pointers would have refused to set range info
2321 * that __check_packet_access would have rejected this pkt access.
2322 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
2324 env->prog->aux->max_pkt_offset =
2325 max_t(u32, env->prog->aux->max_pkt_offset,
2326 off + reg->umax_value + size - 1);
2331 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
2332 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
2333 enum bpf_access_type t, enum bpf_reg_type *reg_type)
2335 struct bpf_insn_access_aux info = {
2336 .reg_type = *reg_type,
2339 if (env->ops->is_valid_access &&
2340 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
2341 /* A non zero info.ctx_field_size indicates that this field is a
2342 * candidate for later verifier transformation to load the whole
2343 * field and then apply a mask when accessed with a narrower
2344 * access than actual ctx access size. A zero info.ctx_field_size
2345 * will only allow for whole field access and rejects any other
2346 * type of narrower access.
2348 *reg_type = info.reg_type;
2350 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
2351 /* remember the offset of last byte accessed in ctx */
2352 if (env->prog->aux->max_ctx_offset < off + size)
2353 env->prog->aux->max_ctx_offset = off + size;
2357 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
2361 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
2364 if (size < 0 || off < 0 ||
2365 (u64)off + size > sizeof(struct bpf_flow_keys)) {
2366 verbose(env, "invalid access to flow keys off=%d size=%d\n",
2373 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
2374 u32 regno, int off, int size,
2375 enum bpf_access_type t)
2377 struct bpf_reg_state *regs = cur_regs(env);
2378 struct bpf_reg_state *reg = ®s[regno];
2379 struct bpf_insn_access_aux info = {};
2382 if (reg->smin_value < 0) {
2383 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2388 switch (reg->type) {
2389 case PTR_TO_SOCK_COMMON:
2390 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
2393 valid = bpf_sock_is_valid_access(off, size, t, &info);
2395 case PTR_TO_TCP_SOCK:
2396 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
2398 case PTR_TO_XDP_SOCK:
2399 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
2407 env->insn_aux_data[insn_idx].ctx_field_size =
2408 info.ctx_field_size;
2412 verbose(env, "R%d invalid %s access off=%d size=%d\n",
2413 regno, reg_type_str[reg->type], off, size);
2418 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
2420 return cur_regs(env) + regno;
2423 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
2425 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
2428 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
2430 const struct bpf_reg_state *reg = reg_state(env, regno);
2432 return reg->type == PTR_TO_CTX;
2435 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
2437 const struct bpf_reg_state *reg = reg_state(env, regno);
2439 return type_is_sk_pointer(reg->type);
2442 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
2444 const struct bpf_reg_state *reg = reg_state(env, regno);
2446 return type_is_pkt_pointer(reg->type);
2449 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
2451 const struct bpf_reg_state *reg = reg_state(env, regno);
2453 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
2454 return reg->type == PTR_TO_FLOW_KEYS;
2457 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
2458 const struct bpf_reg_state *reg,
2459 int off, int size, bool strict)
2461 struct tnum reg_off;
2464 /* Byte size accesses are always allowed. */
2465 if (!strict || size == 1)
2468 /* For platforms that do not have a Kconfig enabling
2469 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
2470 * NET_IP_ALIGN is universally set to '2'. And on platforms
2471 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
2472 * to this code only in strict mode where we want to emulate
2473 * the NET_IP_ALIGN==2 checking. Therefore use an
2474 * unconditional IP align value of '2'.
2478 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
2479 if (!tnum_is_aligned(reg_off, size)) {
2482 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2484 "misaligned packet access off %d+%s+%d+%d size %d\n",
2485 ip_align, tn_buf, reg->off, off, size);
2492 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
2493 const struct bpf_reg_state *reg,
2494 const char *pointer_desc,
2495 int off, int size, bool strict)
2497 struct tnum reg_off;
2499 /* Byte size accesses are always allowed. */
2500 if (!strict || size == 1)
2503 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
2504 if (!tnum_is_aligned(reg_off, size)) {
2507 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2508 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
2509 pointer_desc, tn_buf, reg->off, off, size);
2516 static int check_ptr_alignment(struct bpf_verifier_env *env,
2517 const struct bpf_reg_state *reg, int off,
2518 int size, bool strict_alignment_once)
2520 bool strict = env->strict_alignment || strict_alignment_once;
2521 const char *pointer_desc = "";
2523 switch (reg->type) {
2525 case PTR_TO_PACKET_META:
2526 /* Special case, because of NET_IP_ALIGN. Given metadata sits
2527 * right in front, treat it the very same way.
2529 return check_pkt_ptr_alignment(env, reg, off, size, strict);
2530 case PTR_TO_FLOW_KEYS:
2531 pointer_desc = "flow keys ";
2533 case PTR_TO_MAP_VALUE:
2534 pointer_desc = "value ";
2537 pointer_desc = "context ";
2540 pointer_desc = "stack ";
2541 /* The stack spill tracking logic in check_stack_write()
2542 * and check_stack_read() relies on stack accesses being
2548 pointer_desc = "sock ";
2550 case PTR_TO_SOCK_COMMON:
2551 pointer_desc = "sock_common ";
2553 case PTR_TO_TCP_SOCK:
2554 pointer_desc = "tcp_sock ";
2556 case PTR_TO_XDP_SOCK:
2557 pointer_desc = "xdp_sock ";
2562 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
2566 static int update_stack_depth(struct bpf_verifier_env *env,
2567 const struct bpf_func_state *func,
2570 u16 stack = env->subprog_info[func->subprogno].stack_depth;
2575 /* update known max for given subprogram */
2576 env->subprog_info[func->subprogno].stack_depth = -off;
2580 /* starting from main bpf function walk all instructions of the function
2581 * and recursively walk all callees that given function can call.
2582 * Ignore jump and exit insns.
2583 * Since recursion is prevented by check_cfg() this algorithm
2584 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
2586 static int check_max_stack_depth(struct bpf_verifier_env *env)
2588 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
2589 struct bpf_subprog_info *subprog = env->subprog_info;
2590 struct bpf_insn *insn = env->prog->insnsi;
2591 int ret_insn[MAX_CALL_FRAMES];
2592 int ret_prog[MAX_CALL_FRAMES];
2595 /* protect against potential stack overflow that might happen when
2596 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
2597 * depth for such case down to 256 so that the worst case scenario
2598 * would result in 8k stack size (32 which is tailcall limit * 256 =
2601 * To get the idea what might happen, see an example:
2602 * func1 -> sub rsp, 128
2603 * subfunc1 -> sub rsp, 256
2604 * tailcall1 -> add rsp, 256
2605 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
2606 * subfunc2 -> sub rsp, 64
2607 * subfunc22 -> sub rsp, 128
2608 * tailcall2 -> add rsp, 128
2609 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
2611 * tailcall will unwind the current stack frame but it will not get rid
2612 * of caller's stack as shown on the example above.
2614 if (idx && subprog[idx].has_tail_call && depth >= 256) {
2616 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
2620 /* round up to 32-bytes, since this is granularity
2621 * of interpreter stack size
2623 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
2624 if (depth > MAX_BPF_STACK) {
2625 verbose(env, "combined stack size of %d calls is %d. Too large\n",
2630 subprog_end = subprog[idx + 1].start;
2631 for (; i < subprog_end; i++) {
2632 if (insn[i].code != (BPF_JMP | BPF_CALL))
2634 if (insn[i].src_reg != BPF_PSEUDO_CALL)
2636 /* remember insn and function to return to */
2637 ret_insn[frame] = i + 1;
2638 ret_prog[frame] = idx;
2640 /* find the callee */
2641 i = i + insn[i].imm + 1;
2642 idx = find_subprog(env, i);
2644 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
2649 if (frame >= MAX_CALL_FRAMES) {
2650 verbose(env, "the call stack of %d frames is too deep !\n",
2656 /* end of for() loop means the last insn of the 'subprog'
2657 * was reached. Doesn't matter whether it was JA or EXIT
2661 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
2663 i = ret_insn[frame];
2664 idx = ret_prog[frame];
2668 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
2669 static int get_callee_stack_depth(struct bpf_verifier_env *env,
2670 const struct bpf_insn *insn, int idx)
2672 int start = idx + insn->imm + 1, subprog;
2674 subprog = find_subprog(env, start);
2676 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
2680 return env->subprog_info[subprog].stack_depth;
2684 static int check_ctx_reg(struct bpf_verifier_env *env,
2685 const struct bpf_reg_state *reg, int regno)
2687 /* Access to ctx or passing it to a helper is only allowed in
2688 * its original, unmodified form.
2692 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
2697 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
2700 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2701 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
2708 static int check_tp_buffer_access(struct bpf_verifier_env *env,
2709 const struct bpf_reg_state *reg,
2710 int regno, int off, int size)
2714 "R%d invalid tracepoint buffer access: off=%d, size=%d",
2718 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
2721 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2723 "R%d invalid variable buffer offset: off=%d, var_off=%s",
2724 regno, off, tn_buf);
2727 if (off + size > env->prog->aux->max_tp_access)
2728 env->prog->aux->max_tp_access = off + size;
2734 /* truncate register to smaller size (in bytes)
2735 * must be called with size < BPF_REG_SIZE
2737 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
2741 /* clear high bits in bit representation */
2742 reg->var_off = tnum_cast(reg->var_off, size);
2744 /* fix arithmetic bounds */
2745 mask = ((u64)1 << (size * 8)) - 1;
2746 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
2747 reg->umin_value &= mask;
2748 reg->umax_value &= mask;
2750 reg->umin_value = 0;
2751 reg->umax_value = mask;
2753 reg->smin_value = reg->umin_value;
2754 reg->smax_value = reg->umax_value;
2757 static bool bpf_map_is_rdonly(const struct bpf_map *map)
2759 return (map->map_flags & BPF_F_RDONLY_PROG) && map->frozen;
2762 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
2768 err = map->ops->map_direct_value_addr(map, &addr, off);
2771 ptr = (void *)(long)addr + off;
2775 *val = (u64)*(u8 *)ptr;
2778 *val = (u64)*(u16 *)ptr;
2781 *val = (u64)*(u32 *)ptr;
2792 /* check whether memory at (regno + off) is accessible for t = (read | write)
2793 * if t==write, value_regno is a register which value is stored into memory
2794 * if t==read, value_regno is a register which will receive the value from memory
2795 * if t==write && value_regno==-1, some unknown value is stored into memory
2796 * if t==read && value_regno==-1, don't care what we read from memory
2798 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
2799 int off, int bpf_size, enum bpf_access_type t,
2800 int value_regno, bool strict_alignment_once)
2802 struct bpf_reg_state *regs = cur_regs(env);
2803 struct bpf_reg_state *reg = regs + regno;
2804 struct bpf_func_state *state;
2807 size = bpf_size_to_bytes(bpf_size);
2811 /* alignment checks will add in reg->off themselves */
2812 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
2816 /* for access checks, reg->off is just part of off */
2819 if (reg->type == PTR_TO_MAP_VALUE) {
2820 if (t == BPF_WRITE && value_regno >= 0 &&
2821 is_pointer_value(env, value_regno)) {
2822 verbose(env, "R%d leaks addr into map\n", value_regno);
2825 err = check_map_access_type(env, regno, off, size, t);
2828 err = check_map_access(env, regno, off, size, false);
2829 if (!err && t == BPF_READ && value_regno >= 0) {
2830 struct bpf_map *map = reg->map_ptr;
2832 /* if map is read-only, track its contents as scalars */
2833 if (tnum_is_const(reg->var_off) &&
2834 bpf_map_is_rdonly(map) &&
2835 map->ops->map_direct_value_addr) {
2836 int map_off = off + reg->var_off.value;
2839 err = bpf_map_direct_read(map, map_off, size,
2844 regs[value_regno].type = SCALAR_VALUE;
2845 __mark_reg_known(®s[value_regno], val);
2847 mark_reg_unknown(env, regs, value_regno);
2850 } else if (reg->type == PTR_TO_CTX) {
2851 enum bpf_reg_type reg_type = SCALAR_VALUE;
2853 if (t == BPF_WRITE && value_regno >= 0 &&
2854 is_pointer_value(env, value_regno)) {
2855 verbose(env, "R%d leaks addr into ctx\n", value_regno);
2859 err = check_ctx_reg(env, reg, regno);
2863 err = check_ctx_access(env, insn_idx, off, size, t, ®_type);
2864 if (!err && t == BPF_READ && value_regno >= 0) {
2865 /* ctx access returns either a scalar, or a
2866 * PTR_TO_PACKET[_META,_END]. In the latter
2867 * case, we know the offset is zero.
2869 if (reg_type == SCALAR_VALUE) {
2870 mark_reg_unknown(env, regs, value_regno);
2872 mark_reg_known_zero(env, regs,
2874 if (reg_type_may_be_null(reg_type))
2875 regs[value_regno].id = ++env->id_gen;
2876 /* A load of ctx field could have different
2877 * actual load size with the one encoded in the
2878 * insn. When the dst is PTR, it is for sure not
2881 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
2883 regs[value_regno].type = reg_type;
2886 } else if (reg->type == PTR_TO_STACK) {
2887 off += reg->var_off.value;
2888 err = check_stack_access(env, reg, off, size);
2892 state = func(env, reg);
2893 err = update_stack_depth(env, state, off);
2898 err = check_stack_write(env, state, off, size,
2899 value_regno, insn_idx);
2901 err = check_stack_read(env, state, off, size,
2903 } else if (reg_is_pkt_pointer(reg)) {
2904 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
2905 verbose(env, "cannot write into packet\n");
2908 if (t == BPF_WRITE && value_regno >= 0 &&
2909 is_pointer_value(env, value_regno)) {
2910 verbose(env, "R%d leaks addr into packet\n",
2914 err = check_packet_access(env, regno, off, size, false);
2915 if (!err && t == BPF_READ && value_regno >= 0)
2916 mark_reg_unknown(env, regs, value_regno);
2917 } else if (reg->type == PTR_TO_FLOW_KEYS) {
2918 if (t == BPF_WRITE && value_regno >= 0 &&
2919 is_pointer_value(env, value_regno)) {
2920 verbose(env, "R%d leaks addr into flow keys\n",
2925 err = check_flow_keys_access(env, off, size);
2926 if (!err && t == BPF_READ && value_regno >= 0)
2927 mark_reg_unknown(env, regs, value_regno);
2928 } else if (type_is_sk_pointer(reg->type)) {
2929 if (t == BPF_WRITE) {
2930 verbose(env, "R%d cannot write into %s\n",
2931 regno, reg_type_str[reg->type]);
2934 err = check_sock_access(env, insn_idx, regno, off, size, t);
2935 if (!err && value_regno >= 0)
2936 mark_reg_unknown(env, regs, value_regno);
2937 } else if (reg->type == PTR_TO_TP_BUFFER) {
2938 err = check_tp_buffer_access(env, reg, regno, off, size);
2939 if (!err && t == BPF_READ && value_regno >= 0)
2940 mark_reg_unknown(env, regs, value_regno);
2942 verbose(env, "R%d invalid mem access '%s'\n", regno,
2943 reg_type_str[reg->type]);
2947 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
2948 regs[value_regno].type == SCALAR_VALUE) {
2949 /* b/h/w load zero-extends, mark upper bits as known 0 */
2950 coerce_reg_to_size(®s[value_regno], size);
2955 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
2959 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
2961 verbose(env, "BPF_XADD uses reserved fields\n");
2965 /* check src1 operand */
2966 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2970 /* check src2 operand */
2971 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2975 if (is_pointer_value(env, insn->src_reg)) {
2976 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
2980 if (is_ctx_reg(env, insn->dst_reg) ||
2981 is_pkt_reg(env, insn->dst_reg) ||
2982 is_flow_key_reg(env, insn->dst_reg) ||
2983 is_sk_reg(env, insn->dst_reg)) {
2984 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
2986 reg_type_str[reg_state(env, insn->dst_reg)->type]);
2990 /* check whether atomic_add can read the memory */
2991 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
2992 BPF_SIZE(insn->code), BPF_READ, -1, true);
2996 /* check whether atomic_add can write into the same memory */
2997 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
2998 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
3001 static int __check_stack_boundary(struct bpf_verifier_env *env, u32 regno,
3002 int off, int access_size,
3003 bool zero_size_allowed)
3005 struct bpf_reg_state *reg = reg_state(env, regno);
3007 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
3008 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
3009 if (tnum_is_const(reg->var_off)) {
3010 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
3011 regno, off, access_size);
3015 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3016 verbose(env, "invalid stack type R%d var_off=%s access_size=%d\n",
3017 regno, tn_buf, access_size);
3024 /* when register 'regno' is passed into function that will read 'access_size'
3025 * bytes from that pointer, make sure that it's within stack boundary
3026 * and all elements of stack are initialized.
3027 * Unlike most pointer bounds-checking functions, this one doesn't take an
3028 * 'off' argument, so it has to add in reg->off itself.
3030 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
3031 int access_size, bool zero_size_allowed,
3032 struct bpf_call_arg_meta *meta)
3034 struct bpf_reg_state *reg = reg_state(env, regno);
3035 struct bpf_func_state *state = func(env, reg);
3036 int err, min_off, max_off, i, j, slot, spi;
3038 if (reg->type != PTR_TO_STACK) {
3039 /* Allow zero-byte read from NULL, regardless of pointer type */
3040 if (zero_size_allowed && access_size == 0 &&
3041 register_is_null(reg))
3044 verbose(env, "R%d type=%s expected=%s\n", regno,
3045 reg_type_str[reg->type],
3046 reg_type_str[PTR_TO_STACK]);
3050 if (tnum_is_const(reg->var_off)) {
3051 min_off = max_off = reg->var_off.value + reg->off;
3052 err = __check_stack_boundary(env, regno, min_off, access_size,
3057 /* Variable offset is prohibited for unprivileged mode for
3058 * simplicity since it requires corresponding support in
3059 * Spectre masking for stack ALU.
3060 * See also retrieve_ptr_limit().
3062 if (!env->allow_ptr_leaks) {
3065 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3066 verbose(env, "R%d indirect variable offset stack access prohibited for !root, var_off=%s\n",
3070 /* Only initialized buffer on stack is allowed to be accessed
3071 * with variable offset. With uninitialized buffer it's hard to
3072 * guarantee that whole memory is marked as initialized on
3073 * helper return since specific bounds are unknown what may
3074 * cause uninitialized stack leaking.
3076 if (meta && meta->raw_mode)
3079 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
3080 reg->smax_value <= -BPF_MAX_VAR_OFF) {
3081 verbose(env, "R%d unbounded indirect variable offset stack access\n",
3085 min_off = reg->smin_value + reg->off;
3086 max_off = reg->smax_value + reg->off;
3087 err = __check_stack_boundary(env, regno, min_off, access_size,
3090 verbose(env, "R%d min value is outside of stack bound\n",
3094 err = __check_stack_boundary(env, regno, max_off, access_size,
3097 verbose(env, "R%d max value is outside of stack bound\n",
3103 if (meta && meta->raw_mode) {
3104 meta->access_size = access_size;
3105 meta->regno = regno;
3109 for (i = min_off; i < max_off + access_size; i++) {
3113 spi = slot / BPF_REG_SIZE;
3114 if (state->allocated_stack <= slot)
3116 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
3117 if (*stype == STACK_MISC)
3119 if (*stype == STACK_ZERO) {
3120 /* helper can write anything into the stack */
3121 *stype = STACK_MISC;
3124 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
3125 state->stack[spi].spilled_ptr.type == SCALAR_VALUE) {
3126 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
3127 for (j = 0; j < BPF_REG_SIZE; j++)
3128 state->stack[spi].slot_type[j] = STACK_MISC;
3133 if (tnum_is_const(reg->var_off)) {
3134 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
3135 min_off, i - min_off, access_size);
3139 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3140 verbose(env, "invalid indirect read from stack var_off %s+%d size %d\n",
3141 tn_buf, i - min_off, access_size);
3145 /* reading any byte out of 8-byte 'spill_slot' will cause
3146 * the whole slot to be marked as 'read'
3148 mark_reg_read(env, &state->stack[spi].spilled_ptr,
3149 state->stack[spi].spilled_ptr.parent,
3152 return update_stack_depth(env, state, min_off);
3155 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
3156 int access_size, bool zero_size_allowed,
3157 struct bpf_call_arg_meta *meta)
3159 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3161 switch (reg->type) {
3163 case PTR_TO_PACKET_META:
3164 return check_packet_access(env, regno, reg->off, access_size,
3166 case PTR_TO_MAP_VALUE:
3167 if (check_map_access_type(env, regno, reg->off, access_size,
3168 meta && meta->raw_mode ? BPF_WRITE :
3171 return check_map_access(env, regno, reg->off, access_size,
3173 default: /* scalar_value|ptr_to_stack or invalid ptr */
3174 return check_stack_boundary(env, regno, access_size,
3175 zero_size_allowed, meta);
3179 /* Implementation details:
3180 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
3181 * Two bpf_map_lookups (even with the same key) will have different reg->id.
3182 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
3183 * value_or_null->value transition, since the verifier only cares about
3184 * the range of access to valid map value pointer and doesn't care about actual
3185 * address of the map element.
3186 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
3187 * reg->id > 0 after value_or_null->value transition. By doing so
3188 * two bpf_map_lookups will be considered two different pointers that
3189 * point to different bpf_spin_locks.
3190 * The verifier allows taking only one bpf_spin_lock at a time to avoid
3192 * Since only one bpf_spin_lock is allowed the checks are simpler than
3193 * reg_is_refcounted() logic. The verifier needs to remember only
3194 * one spin_lock instead of array of acquired_refs.
3195 * cur_state->active_spin_lock remembers which map value element got locked
3196 * and clears it after bpf_spin_unlock.
3198 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
3201 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3202 struct bpf_verifier_state *cur = env->cur_state;
3203 bool is_const = tnum_is_const(reg->var_off);
3204 struct bpf_map *map = reg->map_ptr;
3205 u64 val = reg->var_off.value;
3207 if (reg->type != PTR_TO_MAP_VALUE) {
3208 verbose(env, "R%d is not a pointer to map_value\n", regno);
3213 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
3219 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
3223 if (!map_value_has_spin_lock(map)) {
3224 if (map->spin_lock_off == -E2BIG)
3226 "map '%s' has more than one 'struct bpf_spin_lock'\n",
3228 else if (map->spin_lock_off == -ENOENT)
3230 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
3234 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
3238 if (map->spin_lock_off != val + reg->off) {
3239 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
3244 if (cur->active_spin_lock) {
3246 "Locking two bpf_spin_locks are not allowed\n");
3249 cur->active_spin_lock = reg->id;
3251 if (!cur->active_spin_lock) {
3252 verbose(env, "bpf_spin_unlock without taking a lock\n");
3255 if (cur->active_spin_lock != reg->id) {
3256 verbose(env, "bpf_spin_unlock of different lock\n");
3259 cur->active_spin_lock = 0;
3264 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
3266 return type == ARG_PTR_TO_MEM ||
3267 type == ARG_PTR_TO_MEM_OR_NULL ||
3268 type == ARG_PTR_TO_UNINIT_MEM;
3271 static bool arg_type_is_mem_size(enum bpf_arg_type type)
3273 return type == ARG_CONST_SIZE ||
3274 type == ARG_CONST_SIZE_OR_ZERO;
3277 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
3279 return type == ARG_PTR_TO_INT ||
3280 type == ARG_PTR_TO_LONG;
3283 static int int_ptr_type_to_size(enum bpf_arg_type type)
3285 if (type == ARG_PTR_TO_INT)
3287 else if (type == ARG_PTR_TO_LONG)
3293 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
3294 enum bpf_arg_type arg_type,
3295 struct bpf_call_arg_meta *meta)
3297 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3298 enum bpf_reg_type expected_type, type = reg->type;
3301 if (arg_type == ARG_DONTCARE)
3304 err = check_reg_arg(env, regno, SRC_OP);
3308 if (arg_type == ARG_ANYTHING) {
3309 if (is_pointer_value(env, regno)) {
3310 verbose(env, "R%d leaks addr into helper function\n",
3317 if (type_is_pkt_pointer(type) &&
3318 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
3319 verbose(env, "helper access to the packet is not allowed\n");
3323 if (arg_type == ARG_PTR_TO_MAP_KEY ||
3324 arg_type == ARG_PTR_TO_MAP_VALUE ||
3325 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE ||
3326 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL) {
3327 expected_type = PTR_TO_STACK;
3328 if (register_is_null(reg) &&
3329 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL)
3330 /* final test in check_stack_boundary() */;
3331 else if (!type_is_pkt_pointer(type) &&
3332 type != PTR_TO_MAP_VALUE &&
3333 type != expected_type)
3335 } else if (arg_type == ARG_CONST_SIZE ||
3336 arg_type == ARG_CONST_SIZE_OR_ZERO) {
3337 expected_type = SCALAR_VALUE;
3338 if (type != expected_type)
3340 } else if (arg_type == ARG_CONST_MAP_PTR) {
3341 expected_type = CONST_PTR_TO_MAP;
3342 if (type != expected_type)
3344 } else if (arg_type == ARG_PTR_TO_CTX) {
3345 expected_type = PTR_TO_CTX;
3346 if (type != expected_type)
3348 err = check_ctx_reg(env, reg, regno);
3351 } else if (arg_type == ARG_PTR_TO_SOCK_COMMON) {
3352 expected_type = PTR_TO_SOCK_COMMON;
3353 /* Any sk pointer can be ARG_PTR_TO_SOCK_COMMON */
3354 if (!type_is_sk_pointer(type))
3356 if (reg->ref_obj_id) {
3357 if (meta->ref_obj_id) {
3358 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
3359 regno, reg->ref_obj_id,
3363 meta->ref_obj_id = reg->ref_obj_id;
3365 } else if (arg_type == ARG_PTR_TO_SOCKET) {
3366 expected_type = PTR_TO_SOCKET;
3367 if (type != expected_type)
3369 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
3370 if (meta->func_id == BPF_FUNC_spin_lock) {
3371 if (process_spin_lock(env, regno, true))
3373 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
3374 if (process_spin_lock(env, regno, false))
3377 verbose(env, "verifier internal error\n");
3380 } else if (arg_type_is_mem_ptr(arg_type)) {
3381 expected_type = PTR_TO_STACK;
3382 /* One exception here. In case function allows for NULL to be
3383 * passed in as argument, it's a SCALAR_VALUE type. Final test
3384 * happens during stack boundary checking.
3386 if (register_is_null(reg) &&
3387 arg_type == ARG_PTR_TO_MEM_OR_NULL)
3388 /* final test in check_stack_boundary() */;
3389 else if (!type_is_pkt_pointer(type) &&
3390 type != PTR_TO_MAP_VALUE &&
3391 type != expected_type)
3393 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
3394 } else if (arg_type_is_int_ptr(arg_type)) {
3395 expected_type = PTR_TO_STACK;
3396 if (!type_is_pkt_pointer(type) &&
3397 type != PTR_TO_MAP_VALUE &&
3398 type != expected_type)
3401 verbose(env, "unsupported arg_type %d\n", arg_type);
3405 if (arg_type == ARG_CONST_MAP_PTR) {
3406 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
3407 meta->map_ptr = reg->map_ptr;
3408 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
3409 /* bpf_map_xxx(..., map_ptr, ..., key) call:
3410 * check that [key, key + map->key_size) are within
3411 * stack limits and initialized
3413 if (!meta->map_ptr) {
3414 /* in function declaration map_ptr must come before
3415 * map_key, so that it's verified and known before
3416 * we have to check map_key here. Otherwise it means
3417 * that kernel subsystem misconfigured verifier
3419 verbose(env, "invalid map_ptr to access map->key\n");
3422 err = check_helper_mem_access(env, regno,
3423 meta->map_ptr->key_size, false,
3425 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
3426 (arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL &&
3427 !register_is_null(reg)) ||
3428 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
3429 /* bpf_map_xxx(..., map_ptr, ..., value) call:
3430 * check [value, value + map->value_size) validity
3432 if (!meta->map_ptr) {
3433 /* kernel subsystem misconfigured verifier */
3434 verbose(env, "invalid map_ptr to access map->value\n");
3437 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
3438 err = check_helper_mem_access(env, regno,
3439 meta->map_ptr->value_size, false,
3441 } else if (arg_type_is_mem_size(arg_type)) {
3442 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
3444 /* remember the mem_size which may be used later
3445 * to refine return values.
3447 meta->msize_max_value = reg->umax_value;
3449 /* The register is SCALAR_VALUE; the access check
3450 * happens using its boundaries.
3452 if (!tnum_is_const(reg->var_off))
3453 /* For unprivileged variable accesses, disable raw
3454 * mode so that the program is required to
3455 * initialize all the memory that the helper could
3456 * just partially fill up.
3460 if (reg->smin_value < 0) {
3461 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
3466 if (reg->umin_value == 0) {
3467 err = check_helper_mem_access(env, regno - 1, 0,
3474 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
3475 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
3479 err = check_helper_mem_access(env, regno - 1,
3481 zero_size_allowed, meta);
3483 err = mark_chain_precision(env, regno);
3484 } else if (arg_type_is_int_ptr(arg_type)) {
3485 int size = int_ptr_type_to_size(arg_type);
3487 err = check_helper_mem_access(env, regno, size, false, meta);
3490 err = check_ptr_alignment(env, reg, 0, size, true);
3495 verbose(env, "R%d type=%s expected=%s\n", regno,
3496 reg_type_str[type], reg_type_str[expected_type]);
3500 static int check_map_func_compatibility(struct bpf_verifier_env *env,
3501 struct bpf_map *map, int func_id)
3506 /* We need a two way check, first is from map perspective ... */
3507 switch (map->map_type) {
3508 case BPF_MAP_TYPE_PROG_ARRAY:
3509 if (func_id != BPF_FUNC_tail_call)
3512 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
3513 if (func_id != BPF_FUNC_perf_event_read &&
3514 func_id != BPF_FUNC_perf_event_output &&
3515 func_id != BPF_FUNC_perf_event_read_value)
3518 case BPF_MAP_TYPE_STACK_TRACE:
3519 if (func_id != BPF_FUNC_get_stackid)
3522 case BPF_MAP_TYPE_CGROUP_ARRAY:
3523 if (func_id != BPF_FUNC_skb_under_cgroup &&
3524 func_id != BPF_FUNC_current_task_under_cgroup)
3527 case BPF_MAP_TYPE_CGROUP_STORAGE:
3528 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
3529 if (func_id != BPF_FUNC_get_local_storage)
3532 case BPF_MAP_TYPE_DEVMAP:
3533 case BPF_MAP_TYPE_DEVMAP_HASH:
3534 if (func_id != BPF_FUNC_redirect_map &&
3535 func_id != BPF_FUNC_map_lookup_elem)
3538 /* Restrict bpf side of cpumap and xskmap, open when use-cases
3541 case BPF_MAP_TYPE_CPUMAP:
3542 if (func_id != BPF_FUNC_redirect_map)
3545 case BPF_MAP_TYPE_XSKMAP:
3546 if (func_id != BPF_FUNC_redirect_map &&
3547 func_id != BPF_FUNC_map_lookup_elem)
3550 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
3551 case BPF_MAP_TYPE_HASH_OF_MAPS:
3552 if (func_id != BPF_FUNC_map_lookup_elem)
3555 case BPF_MAP_TYPE_SOCKMAP:
3556 if (func_id != BPF_FUNC_sk_redirect_map &&
3557 func_id != BPF_FUNC_sock_map_update &&
3558 func_id != BPF_FUNC_map_delete_elem &&
3559 func_id != BPF_FUNC_msg_redirect_map)
3562 case BPF_MAP_TYPE_SOCKHASH:
3563 if (func_id != BPF_FUNC_sk_redirect_hash &&
3564 func_id != BPF_FUNC_sock_hash_update &&
3565 func_id != BPF_FUNC_map_delete_elem &&
3566 func_id != BPF_FUNC_msg_redirect_hash)
3569 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
3570 if (func_id != BPF_FUNC_sk_select_reuseport)
3573 case BPF_MAP_TYPE_QUEUE:
3574 case BPF_MAP_TYPE_STACK:
3575 if (func_id != BPF_FUNC_map_peek_elem &&
3576 func_id != BPF_FUNC_map_pop_elem &&
3577 func_id != BPF_FUNC_map_push_elem)
3580 case BPF_MAP_TYPE_SK_STORAGE:
3581 if (func_id != BPF_FUNC_sk_storage_get &&
3582 func_id != BPF_FUNC_sk_storage_delete)
3589 /* ... and second from the function itself. */
3591 case BPF_FUNC_tail_call:
3592 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
3594 if (env->subprog_cnt > 1) {
3595 verbose(env, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
3599 case BPF_FUNC_perf_event_read:
3600 case BPF_FUNC_perf_event_output:
3601 case BPF_FUNC_perf_event_read_value:
3602 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
3605 case BPF_FUNC_get_stackid:
3606 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
3609 case BPF_FUNC_current_task_under_cgroup:
3610 case BPF_FUNC_skb_under_cgroup:
3611 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
3614 case BPF_FUNC_redirect_map:
3615 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
3616 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
3617 map->map_type != BPF_MAP_TYPE_CPUMAP &&
3618 map->map_type != BPF_MAP_TYPE_XSKMAP)
3621 case BPF_FUNC_sk_redirect_map:
3622 case BPF_FUNC_msg_redirect_map:
3623 case BPF_FUNC_sock_map_update:
3624 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
3627 case BPF_FUNC_sk_redirect_hash:
3628 case BPF_FUNC_msg_redirect_hash:
3629 case BPF_FUNC_sock_hash_update:
3630 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
3633 case BPF_FUNC_get_local_storage:
3634 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
3635 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
3638 case BPF_FUNC_sk_select_reuseport:
3639 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY)
3642 case BPF_FUNC_map_peek_elem:
3643 case BPF_FUNC_map_pop_elem:
3644 case BPF_FUNC_map_push_elem:
3645 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
3646 map->map_type != BPF_MAP_TYPE_STACK)
3649 case BPF_FUNC_sk_storage_get:
3650 case BPF_FUNC_sk_storage_delete:
3651 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
3660 verbose(env, "cannot pass map_type %d into func %s#%d\n",
3661 map->map_type, func_id_name(func_id), func_id);
3665 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
3669 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
3671 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
3673 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
3675 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
3677 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
3680 /* We only support one arg being in raw mode at the moment,
3681 * which is sufficient for the helper functions we have
3687 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
3688 enum bpf_arg_type arg_next)
3690 return (arg_type_is_mem_ptr(arg_curr) &&
3691 !arg_type_is_mem_size(arg_next)) ||
3692 (!arg_type_is_mem_ptr(arg_curr) &&
3693 arg_type_is_mem_size(arg_next));
3696 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
3698 /* bpf_xxx(..., buf, len) call will access 'len'
3699 * bytes from memory 'buf'. Both arg types need
3700 * to be paired, so make sure there's no buggy
3701 * helper function specification.
3703 if (arg_type_is_mem_size(fn->arg1_type) ||
3704 arg_type_is_mem_ptr(fn->arg5_type) ||
3705 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
3706 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
3707 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
3708 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
3714 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
3718 if (arg_type_may_be_refcounted(fn->arg1_type))
3720 if (arg_type_may_be_refcounted(fn->arg2_type))
3722 if (arg_type_may_be_refcounted(fn->arg3_type))
3724 if (arg_type_may_be_refcounted(fn->arg4_type))
3726 if (arg_type_may_be_refcounted(fn->arg5_type))
3729 /* A reference acquiring function cannot acquire
3730 * another refcounted ptr.
3732 if (is_acquire_function(func_id) && count)
3735 /* We only support one arg being unreferenced at the moment,
3736 * which is sufficient for the helper functions we have right now.
3741 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
3743 return check_raw_mode_ok(fn) &&
3744 check_arg_pair_ok(fn) &&
3745 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
3748 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
3749 * are now invalid, so turn them into unknown SCALAR_VALUE.
3751 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
3752 struct bpf_func_state *state)
3754 struct bpf_reg_state *regs = state->regs, *reg;
3757 for (i = 0; i < MAX_BPF_REG; i++)
3758 if (reg_is_pkt_pointer_any(®s[i]))
3759 mark_reg_unknown(env, regs, i);
3761 bpf_for_each_spilled_reg(i, state, reg) {
3764 if (reg_is_pkt_pointer_any(reg))
3765 __mark_reg_unknown(env, reg);
3769 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
3771 struct bpf_verifier_state *vstate = env->cur_state;
3774 for (i = 0; i <= vstate->curframe; i++)
3775 __clear_all_pkt_pointers(env, vstate->frame[i]);
3778 static void release_reg_references(struct bpf_verifier_env *env,
3779 struct bpf_func_state *state,
3782 struct bpf_reg_state *regs = state->regs, *reg;
3785 for (i = 0; i < MAX_BPF_REG; i++)
3786 if (regs[i].ref_obj_id == ref_obj_id)
3787 mark_reg_unknown(env, regs, i);
3789 bpf_for_each_spilled_reg(i, state, reg) {
3792 if (reg->ref_obj_id == ref_obj_id)
3793 __mark_reg_unknown(env, reg);
3797 /* The pointer with the specified id has released its reference to kernel
3798 * resources. Identify all copies of the same pointer and clear the reference.
3800 static int release_reference(struct bpf_verifier_env *env,
3803 struct bpf_verifier_state *vstate = env->cur_state;
3807 err = release_reference_state(cur_func(env), ref_obj_id);
3811 for (i = 0; i <= vstate->curframe; i++)
3812 release_reg_references(env, vstate->frame[i], ref_obj_id);
3817 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
3820 struct bpf_verifier_state *state = env->cur_state;
3821 struct bpf_func_state *caller, *callee;
3822 int i, err, subprog, target_insn;
3824 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
3825 verbose(env, "the call stack of %d frames is too deep\n",
3826 state->curframe + 2);
3830 target_insn = *insn_idx + insn->imm;
3831 subprog = find_subprog(env, target_insn + 1);
3833 verbose(env, "verifier bug. No program starts at insn %d\n",
3838 caller = state->frame[state->curframe];
3839 if (state->frame[state->curframe + 1]) {
3840 verbose(env, "verifier bug. Frame %d already allocated\n",
3841 state->curframe + 1);
3845 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
3848 state->frame[state->curframe + 1] = callee;
3850 /* callee cannot access r0, r6 - r9 for reading and has to write
3851 * into its own stack before reading from it.
3852 * callee can read/write into caller's stack
3854 init_func_state(env, callee,
3855 /* remember the callsite, it will be used by bpf_exit */
3856 *insn_idx /* callsite */,
3857 state->curframe + 1 /* frameno within this callchain */,
3858 subprog /* subprog number within this prog */);
3860 /* Transfer references to the callee */
3861 err = transfer_reference_state(callee, caller);
3865 /* copy r1 - r5 args that callee can access. The copy includes parent
3866 * pointers, which connects us up to the liveness chain
3868 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
3869 callee->regs[i] = caller->regs[i];
3871 /* after the call registers r0 - r5 were scratched */
3872 for (i = 0; i < CALLER_SAVED_REGS; i++) {
3873 mark_reg_not_init(env, caller->regs, caller_saved[i]);
3874 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
3877 /* only increment it after check_reg_arg() finished */
3880 /* and go analyze first insn of the callee */
3881 *insn_idx = target_insn;
3883 if (env->log.level & BPF_LOG_LEVEL) {
3884 verbose(env, "caller:\n");
3885 print_verifier_state(env, caller);
3886 verbose(env, "callee:\n");
3887 print_verifier_state(env, callee);
3892 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
3894 struct bpf_verifier_state *state = env->cur_state;
3895 struct bpf_func_state *caller, *callee;
3896 struct bpf_reg_state *r0;
3899 callee = state->frame[state->curframe];
3900 r0 = &callee->regs[BPF_REG_0];
3901 if (r0->type == PTR_TO_STACK) {
3902 /* technically it's ok to return caller's stack pointer
3903 * (or caller's caller's pointer) back to the caller,
3904 * since these pointers are valid. Only current stack
3905 * pointer will be invalid as soon as function exits,
3906 * but let's be conservative
3908 verbose(env, "cannot return stack pointer to the caller\n");
3913 caller = state->frame[state->curframe];
3914 /* return to the caller whatever r0 had in the callee */
3915 caller->regs[BPF_REG_0] = *r0;
3917 /* Transfer references to the caller */
3918 err = transfer_reference_state(caller, callee);
3922 *insn_idx = callee->callsite + 1;
3923 if (env->log.level & BPF_LOG_LEVEL) {
3924 verbose(env, "returning from callee:\n");
3925 print_verifier_state(env, callee);
3926 verbose(env, "to caller at %d:\n", *insn_idx);
3927 print_verifier_state(env, caller);
3929 /* clear everything in the callee */
3930 free_func_state(callee);
3931 state->frame[state->curframe + 1] = NULL;
3935 static int do_refine_retval_range(struct bpf_verifier_env *env,
3936 struct bpf_reg_state *regs, int ret_type,
3937 int func_id, struct bpf_call_arg_meta *meta)
3939 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
3940 struct bpf_reg_state tmp_reg = *ret_reg;
3943 if (ret_type != RET_INTEGER ||
3944 (func_id != BPF_FUNC_get_stack &&
3945 func_id != BPF_FUNC_probe_read_str))
3948 /* Error case where ret is in interval [S32MIN, -1]. */
3949 ret_reg->smin_value = S32_MIN;
3950 ret_reg->smax_value = -1;
3952 __reg_deduce_bounds(ret_reg);
3953 __reg_bound_offset(ret_reg);
3954 __update_reg_bounds(ret_reg);
3956 ret = push_stack(env, env->insn_idx + 1, env->insn_idx, false);
3962 /* Success case where ret is in range [0, msize_max_value]. */
3963 ret_reg->smin_value = 0;
3964 ret_reg->smax_value = meta->msize_max_value;
3965 ret_reg->umin_value = ret_reg->smin_value;
3966 ret_reg->umax_value = ret_reg->smax_value;
3968 __reg_deduce_bounds(ret_reg);
3969 __reg_bound_offset(ret_reg);
3970 __update_reg_bounds(ret_reg);
3976 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
3977 int func_id, int insn_idx)
3979 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
3980 struct bpf_map *map = meta->map_ptr;
3982 if (func_id != BPF_FUNC_tail_call &&
3983 func_id != BPF_FUNC_map_lookup_elem &&
3984 func_id != BPF_FUNC_map_update_elem &&
3985 func_id != BPF_FUNC_map_delete_elem &&
3986 func_id != BPF_FUNC_map_push_elem &&
3987 func_id != BPF_FUNC_map_pop_elem &&
3988 func_id != BPF_FUNC_map_peek_elem)
3992 verbose(env, "kernel subsystem misconfigured verifier\n");
3996 /* In case of read-only, some additional restrictions
3997 * need to be applied in order to prevent altering the
3998 * state of the map from program side.
4000 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
4001 (func_id == BPF_FUNC_map_delete_elem ||
4002 func_id == BPF_FUNC_map_update_elem ||
4003 func_id == BPF_FUNC_map_push_elem ||
4004 func_id == BPF_FUNC_map_pop_elem)) {
4005 verbose(env, "write into map forbidden\n");
4009 if (!BPF_MAP_PTR(aux->map_state))
4010 bpf_map_ptr_store(aux, meta->map_ptr,
4011 meta->map_ptr->unpriv_array);
4012 else if (BPF_MAP_PTR(aux->map_state) != meta->map_ptr)
4013 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
4014 meta->map_ptr->unpriv_array);
4018 static int check_reference_leak(struct bpf_verifier_env *env)
4020 struct bpf_func_state *state = cur_func(env);
4023 for (i = 0; i < state->acquired_refs; i++) {
4024 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
4025 state->refs[i].id, state->refs[i].insn_idx);
4027 return state->acquired_refs ? -EINVAL : 0;
4030 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
4032 const struct bpf_func_proto *fn = NULL;
4033 struct bpf_reg_state *regs;
4034 struct bpf_call_arg_meta meta;
4038 /* find function prototype */
4039 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
4040 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
4045 if (env->ops->get_func_proto)
4046 fn = env->ops->get_func_proto(func_id, env->prog);
4048 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
4053 /* eBPF programs must be GPL compatible to use GPL-ed functions */
4054 if (!env->prog->gpl_compatible && fn->gpl_only) {
4055 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
4059 /* With LD_ABS/IND some JITs save/restore skb from r1. */
4060 changes_data = bpf_helper_changes_pkt_data(fn->func);
4061 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
4062 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
4063 func_id_name(func_id), func_id);
4067 memset(&meta, 0, sizeof(meta));
4068 meta.pkt_access = fn->pkt_access;
4070 err = check_func_proto(fn, func_id);
4072 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
4073 func_id_name(func_id), func_id);
4077 meta.func_id = func_id;
4079 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
4082 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
4085 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
4088 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
4091 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
4095 err = record_func_map(env, &meta, func_id, insn_idx);
4099 /* Mark slots with STACK_MISC in case of raw mode, stack offset
4100 * is inferred from register state.
4102 for (i = 0; i < meta.access_size; i++) {
4103 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
4104 BPF_WRITE, -1, false);
4109 if (func_id == BPF_FUNC_tail_call) {
4110 err = check_reference_leak(env);
4112 verbose(env, "tail_call would lead to reference leak\n");
4115 } else if (is_release_function(func_id)) {
4116 err = release_reference(env, meta.ref_obj_id);
4118 verbose(env, "func %s#%d reference has not been acquired before\n",
4119 func_id_name(func_id), func_id);
4124 regs = cur_regs(env);
4126 /* check that flags argument in get_local_storage(map, flags) is 0,
4127 * this is required because get_local_storage() can't return an error.
4129 if (func_id == BPF_FUNC_get_local_storage &&
4130 !register_is_null(®s[BPF_REG_2])) {
4131 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
4135 /* reset caller saved regs */
4136 for (i = 0; i < CALLER_SAVED_REGS; i++) {
4137 mark_reg_not_init(env, regs, caller_saved[i]);
4138 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
4141 /* helper call returns 64-bit value. */
4142 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
4144 /* update return register (already marked as written above) */
4145 if (fn->ret_type == RET_INTEGER) {
4146 /* sets type to SCALAR_VALUE */
4147 mark_reg_unknown(env, regs, BPF_REG_0);
4148 } else if (fn->ret_type == RET_VOID) {
4149 regs[BPF_REG_0].type = NOT_INIT;
4150 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
4151 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
4152 /* There is no offset yet applied, variable or fixed */
4153 mark_reg_known_zero(env, regs, BPF_REG_0);
4154 /* remember map_ptr, so that check_map_access()
4155 * can check 'value_size' boundary of memory access
4156 * to map element returned from bpf_map_lookup_elem()
4158 if (meta.map_ptr == NULL) {
4160 "kernel subsystem misconfigured verifier\n");
4163 regs[BPF_REG_0].map_ptr = meta.map_ptr;
4164 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
4165 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
4166 if (map_value_has_spin_lock(meta.map_ptr))
4167 regs[BPF_REG_0].id = ++env->id_gen;
4169 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
4170 regs[BPF_REG_0].id = ++env->id_gen;
4172 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
4173 mark_reg_known_zero(env, regs, BPF_REG_0);
4174 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
4175 regs[BPF_REG_0].id = ++env->id_gen;
4176 } else if (fn->ret_type == RET_PTR_TO_SOCK_COMMON_OR_NULL) {
4177 mark_reg_known_zero(env, regs, BPF_REG_0);
4178 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON_OR_NULL;
4179 regs[BPF_REG_0].id = ++env->id_gen;
4180 } else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
4181 mark_reg_known_zero(env, regs, BPF_REG_0);
4182 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
4183 regs[BPF_REG_0].id = ++env->id_gen;
4185 verbose(env, "unknown return type %d of func %s#%d\n",
4186 fn->ret_type, func_id_name(func_id), func_id);
4190 if (is_ptr_cast_function(func_id)) {
4191 /* For release_reference() */
4192 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
4193 } else if (is_acquire_function(func_id)) {
4194 int id = acquire_reference_state(env, insn_idx);
4198 /* For mark_ptr_or_null_reg() */
4199 regs[BPF_REG_0].id = id;
4200 /* For release_reference() */
4201 regs[BPF_REG_0].ref_obj_id = id;
4204 err = do_refine_retval_range(env, regs, fn->ret_type, func_id, &meta);
4208 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
4212 if (func_id == BPF_FUNC_get_stack && !env->prog->has_callchain_buf) {
4213 const char *err_str;
4215 #ifdef CONFIG_PERF_EVENTS
4216 err = get_callchain_buffers(sysctl_perf_event_max_stack);
4217 err_str = "cannot get callchain buffer for func %s#%d\n";
4220 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
4223 verbose(env, err_str, func_id_name(func_id), func_id);
4227 env->prog->has_callchain_buf = true;
4231 clear_all_pkt_pointers(env);
4235 static bool signed_add_overflows(s64 a, s64 b)
4237 /* Do the add in u64, where overflow is well-defined */
4238 s64 res = (s64)((u64)a + (u64)b);
4245 static bool signed_sub_overflows(s64 a, s64 b)
4247 /* Do the sub in u64, where overflow is well-defined */
4248 s64 res = (s64)((u64)a - (u64)b);
4255 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
4256 const struct bpf_reg_state *reg,
4257 enum bpf_reg_type type)
4259 bool known = tnum_is_const(reg->var_off);
4260 s64 val = reg->var_off.value;
4261 s64 smin = reg->smin_value;
4263 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
4264 verbose(env, "math between %s pointer and %lld is not allowed\n",
4265 reg_type_str[type], val);
4269 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
4270 verbose(env, "%s pointer offset %d is not allowed\n",
4271 reg_type_str[type], reg->off);
4275 if (smin == S64_MIN) {
4276 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
4277 reg_type_str[type]);
4281 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
4282 verbose(env, "value %lld makes %s pointer be out of bounds\n",
4283 smin, reg_type_str[type]);
4290 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
4292 return &env->insn_aux_data[env->insn_idx];
4303 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
4304 u32 *alu_limit, bool mask_to_left)
4306 u32 max = 0, ptr_limit = 0;
4308 switch (ptr_reg->type) {
4310 /* Offset 0 is out-of-bounds, but acceptable start for the
4311 * left direction, see BPF_REG_FP. Also, unknown scalar
4312 * offset where we would need to deal with min/max bounds is
4313 * currently prohibited for unprivileged.
4315 max = MAX_BPF_STACK + mask_to_left;
4316 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
4318 case PTR_TO_MAP_VALUE:
4319 max = ptr_reg->map_ptr->value_size;
4320 ptr_limit = (mask_to_left ?
4321 ptr_reg->smin_value :
4322 ptr_reg->umax_value) + ptr_reg->off;
4328 if (ptr_limit >= max)
4329 return REASON_LIMIT;
4330 *alu_limit = ptr_limit;
4334 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
4335 const struct bpf_insn *insn)
4337 return env->allow_ptr_leaks || BPF_SRC(insn->code) == BPF_K;
4340 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
4341 u32 alu_state, u32 alu_limit)
4343 /* If we arrived here from different branches with different
4344 * state or limits to sanitize, then this won't work.
4346 if (aux->alu_state &&
4347 (aux->alu_state != alu_state ||
4348 aux->alu_limit != alu_limit))
4349 return REASON_PATHS;
4351 /* Corresponding fixup done in fixup_bpf_calls(). */
4352 aux->alu_state = alu_state;
4353 aux->alu_limit = alu_limit;
4357 static int sanitize_val_alu(struct bpf_verifier_env *env,
4358 struct bpf_insn *insn)
4360 struct bpf_insn_aux_data *aux = cur_aux(env);
4362 if (can_skip_alu_sanitation(env, insn))
4365 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
4368 static bool sanitize_needed(u8 opcode)
4370 return opcode == BPF_ADD || opcode == BPF_SUB;
4373 struct bpf_sanitize_info {
4374 struct bpf_insn_aux_data aux;
4378 static struct bpf_verifier_state *
4379 sanitize_speculative_path(struct bpf_verifier_env *env,
4380 const struct bpf_insn *insn,
4381 u32 next_idx, u32 curr_idx)
4383 struct bpf_verifier_state *branch;
4384 struct bpf_reg_state *regs;
4386 branch = push_stack(env, next_idx, curr_idx, true);
4387 if (branch && insn) {
4388 regs = branch->frame[branch->curframe]->regs;
4389 if (BPF_SRC(insn->code) == BPF_K) {
4390 mark_reg_unknown(env, regs, insn->dst_reg);
4391 } else if (BPF_SRC(insn->code) == BPF_X) {
4392 mark_reg_unknown(env, regs, insn->dst_reg);
4393 mark_reg_unknown(env, regs, insn->src_reg);
4399 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
4400 struct bpf_insn *insn,
4401 const struct bpf_reg_state *ptr_reg,
4402 const struct bpf_reg_state *off_reg,
4403 struct bpf_reg_state *dst_reg,
4404 struct bpf_sanitize_info *info,
4405 const bool commit_window)
4407 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
4408 struct bpf_verifier_state *vstate = env->cur_state;
4409 bool off_is_imm = tnum_is_const(off_reg->var_off);
4410 bool off_is_neg = off_reg->smin_value < 0;
4411 bool ptr_is_dst_reg = ptr_reg == dst_reg;
4412 u8 opcode = BPF_OP(insn->code);
4413 u32 alu_state, alu_limit;
4414 struct bpf_reg_state tmp;
4418 if (can_skip_alu_sanitation(env, insn))
4421 /* We already marked aux for masking from non-speculative
4422 * paths, thus we got here in the first place. We only care
4423 * to explore bad access from here.
4425 if (vstate->speculative)
4428 if (!commit_window) {
4429 if (!tnum_is_const(off_reg->var_off) &&
4430 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
4431 return REASON_BOUNDS;
4433 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
4434 (opcode == BPF_SUB && !off_is_neg);
4437 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
4441 if (commit_window) {
4442 /* In commit phase we narrow the masking window based on
4443 * the observed pointer move after the simulated operation.
4445 alu_state = info->aux.alu_state;
4446 alu_limit = abs(info->aux.alu_limit - alu_limit);
4448 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
4449 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
4450 alu_state |= ptr_is_dst_reg ?
4451 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
4453 /* Limit pruning on unknown scalars to enable deep search for
4454 * potential masking differences from other program paths.
4457 env->explore_alu_limits = true;
4460 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
4464 /* If we're in commit phase, we're done here given we already
4465 * pushed the truncated dst_reg into the speculative verification
4468 * Also, when register is a known constant, we rewrite register-based
4469 * operation to immediate-based, and thus do not need masking (and as
4470 * a consequence, do not need to simulate the zero-truncation either).
4472 if (commit_window || off_is_imm)
4475 /* Simulate and find potential out-of-bounds access under
4476 * speculative execution from truncation as a result of
4477 * masking when off was not within expected range. If off
4478 * sits in dst, then we temporarily need to move ptr there
4479 * to simulate dst (== 0) +/-= ptr. Needed, for example,
4480 * for cases where we use K-based arithmetic in one direction
4481 * and truncated reg-based in the other in order to explore
4484 if (!ptr_is_dst_reg) {
4486 *dst_reg = *ptr_reg;
4488 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
4490 if (!ptr_is_dst_reg && ret)
4492 return !ret ? REASON_STACK : 0;
4495 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
4497 struct bpf_verifier_state *vstate = env->cur_state;
4499 /* If we simulate paths under speculation, we don't update the
4500 * insn as 'seen' such that when we verify unreachable paths in
4501 * the non-speculative domain, sanitize_dead_code() can still
4502 * rewrite/sanitize them.
4504 if (!vstate->speculative)
4505 env->insn_aux_data[env->insn_idx].seen = true;
4508 static int sanitize_err(struct bpf_verifier_env *env,
4509 const struct bpf_insn *insn, int reason,
4510 const struct bpf_reg_state *off_reg,
4511 const struct bpf_reg_state *dst_reg)
4513 static const char *err = "pointer arithmetic with it prohibited for !root";
4514 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
4515 u32 dst = insn->dst_reg, src = insn->src_reg;
4519 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
4520 off_reg == dst_reg ? dst : src, err);
4523 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
4524 off_reg == dst_reg ? src : dst, err);
4527 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
4531 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
4535 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
4539 verbose(env, "verifier internal error: unknown reason (%d)\n",
4547 static int sanitize_check_bounds(struct bpf_verifier_env *env,
4548 const struct bpf_insn *insn,
4549 const struct bpf_reg_state *dst_reg)
4551 u32 dst = insn->dst_reg;
4553 /* For unprivileged we require that resulting offset must be in bounds
4554 * in order to be able to sanitize access later on.
4556 if (env->allow_ptr_leaks)
4559 switch (dst_reg->type) {
4561 if (check_stack_access(env, dst_reg, dst_reg->off +
4562 dst_reg->var_off.value, 1)) {
4563 verbose(env, "R%d stack pointer arithmetic goes out of range, "
4564 "prohibited for !root\n", dst);
4568 case PTR_TO_MAP_VALUE:
4569 if (check_map_access(env, dst, dst_reg->off, 1, false)) {
4570 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
4571 "prohibited for !root\n", dst);
4582 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
4583 * Caller should also handle BPF_MOV case separately.
4584 * If we return -EACCES, caller may want to try again treating pointer as a
4585 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
4587 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
4588 struct bpf_insn *insn,
4589 const struct bpf_reg_state *ptr_reg,
4590 const struct bpf_reg_state *off_reg)
4592 struct bpf_verifier_state *vstate = env->cur_state;
4593 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4594 struct bpf_reg_state *regs = state->regs, *dst_reg;
4595 bool known = tnum_is_const(off_reg->var_off);
4596 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
4597 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
4598 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
4599 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
4600 struct bpf_sanitize_info info = {};
4601 u8 opcode = BPF_OP(insn->code);
4602 u32 dst = insn->dst_reg;
4605 dst_reg = ®s[dst];
4607 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
4608 smin_val > smax_val || umin_val > umax_val) {
4609 /* Taint dst register if offset had invalid bounds derived from
4610 * e.g. dead branches.
4612 __mark_reg_unknown(env, dst_reg);
4616 if (BPF_CLASS(insn->code) != BPF_ALU64) {
4617 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
4619 "R%d 32-bit pointer arithmetic prohibited\n",
4624 switch (ptr_reg->type) {
4625 case PTR_TO_MAP_VALUE_OR_NULL:
4626 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
4627 dst, reg_type_str[ptr_reg->type]);
4629 case CONST_PTR_TO_MAP:
4630 /* smin_val represents the known value */
4631 if (known && smin_val == 0 && opcode == BPF_ADD)
4634 case PTR_TO_PACKET_END:
4636 case PTR_TO_SOCKET_OR_NULL:
4637 case PTR_TO_SOCK_COMMON:
4638 case PTR_TO_SOCK_COMMON_OR_NULL:
4639 case PTR_TO_TCP_SOCK:
4640 case PTR_TO_TCP_SOCK_OR_NULL:
4641 case PTR_TO_XDP_SOCK:
4642 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
4643 dst, reg_type_str[ptr_reg->type]);
4649 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
4650 * The id may be overwritten later if we create a new variable offset.
4652 dst_reg->type = ptr_reg->type;
4653 dst_reg->id = ptr_reg->id;
4655 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
4656 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
4659 if (sanitize_needed(opcode)) {
4660 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
4663 return sanitize_err(env, insn, ret, off_reg, dst_reg);
4668 /* We can take a fixed offset as long as it doesn't overflow
4669 * the s32 'off' field
4671 if (known && (ptr_reg->off + smin_val ==
4672 (s64)(s32)(ptr_reg->off + smin_val))) {
4673 /* pointer += K. Accumulate it into fixed offset */
4674 dst_reg->smin_value = smin_ptr;
4675 dst_reg->smax_value = smax_ptr;
4676 dst_reg->umin_value = umin_ptr;
4677 dst_reg->umax_value = umax_ptr;
4678 dst_reg->var_off = ptr_reg->var_off;
4679 dst_reg->off = ptr_reg->off + smin_val;
4680 dst_reg->raw = ptr_reg->raw;
4683 /* A new variable offset is created. Note that off_reg->off
4684 * == 0, since it's a scalar.
4685 * dst_reg gets the pointer type and since some positive
4686 * integer value was added to the pointer, give it a new 'id'
4687 * if it's a PTR_TO_PACKET.
4688 * this creates a new 'base' pointer, off_reg (variable) gets
4689 * added into the variable offset, and we copy the fixed offset
4692 if (signed_add_overflows(smin_ptr, smin_val) ||
4693 signed_add_overflows(smax_ptr, smax_val)) {
4694 dst_reg->smin_value = S64_MIN;
4695 dst_reg->smax_value = S64_MAX;
4697 dst_reg->smin_value = smin_ptr + smin_val;
4698 dst_reg->smax_value = smax_ptr + smax_val;
4700 if (umin_ptr + umin_val < umin_ptr ||
4701 umax_ptr + umax_val < umax_ptr) {
4702 dst_reg->umin_value = 0;
4703 dst_reg->umax_value = U64_MAX;
4705 dst_reg->umin_value = umin_ptr + umin_val;
4706 dst_reg->umax_value = umax_ptr + umax_val;
4708 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
4709 dst_reg->off = ptr_reg->off;
4710 dst_reg->raw = ptr_reg->raw;
4711 if (reg_is_pkt_pointer(ptr_reg)) {
4712 dst_reg->id = ++env->id_gen;
4713 /* something was added to pkt_ptr, set range to zero */
4718 if (dst_reg == off_reg) {
4719 /* scalar -= pointer. Creates an unknown scalar */
4720 verbose(env, "R%d tried to subtract pointer from scalar\n",
4724 /* We don't allow subtraction from FP, because (according to
4725 * test_verifier.c test "invalid fp arithmetic", JITs might not
4726 * be able to deal with it.
4728 if (ptr_reg->type == PTR_TO_STACK) {
4729 verbose(env, "R%d subtraction from stack pointer prohibited\n",
4733 if (known && (ptr_reg->off - smin_val ==
4734 (s64)(s32)(ptr_reg->off - smin_val))) {
4735 /* pointer -= K. Subtract it from fixed offset */
4736 dst_reg->smin_value = smin_ptr;
4737 dst_reg->smax_value = smax_ptr;
4738 dst_reg->umin_value = umin_ptr;
4739 dst_reg->umax_value = umax_ptr;
4740 dst_reg->var_off = ptr_reg->var_off;
4741 dst_reg->id = ptr_reg->id;
4742 dst_reg->off = ptr_reg->off - smin_val;
4743 dst_reg->raw = ptr_reg->raw;
4746 /* A new variable offset is created. If the subtrahend is known
4747 * nonnegative, then any reg->range we had before is still good.
4749 if (signed_sub_overflows(smin_ptr, smax_val) ||
4750 signed_sub_overflows(smax_ptr, smin_val)) {
4751 /* Overflow possible, we know nothing */
4752 dst_reg->smin_value = S64_MIN;
4753 dst_reg->smax_value = S64_MAX;
4755 dst_reg->smin_value = smin_ptr - smax_val;
4756 dst_reg->smax_value = smax_ptr - smin_val;
4758 if (umin_ptr < umax_val) {
4759 /* Overflow possible, we know nothing */
4760 dst_reg->umin_value = 0;
4761 dst_reg->umax_value = U64_MAX;
4763 /* Cannot overflow (as long as bounds are consistent) */
4764 dst_reg->umin_value = umin_ptr - umax_val;
4765 dst_reg->umax_value = umax_ptr - umin_val;
4767 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
4768 dst_reg->off = ptr_reg->off;
4769 dst_reg->raw = ptr_reg->raw;
4770 if (reg_is_pkt_pointer(ptr_reg)) {
4771 dst_reg->id = ++env->id_gen;
4772 /* something was added to pkt_ptr, set range to zero */
4780 /* bitwise ops on pointers are troublesome, prohibit. */
4781 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
4782 dst, bpf_alu_string[opcode >> 4]);
4785 /* other operators (e.g. MUL,LSH) produce non-pointer results */
4786 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
4787 dst, bpf_alu_string[opcode >> 4]);
4791 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
4794 __update_reg_bounds(dst_reg);
4795 __reg_deduce_bounds(dst_reg);
4796 __reg_bound_offset(dst_reg);
4798 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
4800 if (sanitize_needed(opcode)) {
4801 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
4804 return sanitize_err(env, insn, ret, off_reg, dst_reg);
4810 /* WARNING: This function does calculations on 64-bit values, but the actual
4811 * execution may occur on 32-bit values. Therefore, things like bitshifts
4812 * need extra checks in the 32-bit case.
4814 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
4815 struct bpf_insn *insn,
4816 struct bpf_reg_state *dst_reg,
4817 struct bpf_reg_state src_reg)
4819 struct bpf_reg_state *regs = cur_regs(env);
4820 u8 opcode = BPF_OP(insn->code);
4821 bool src_known, dst_known;
4822 s64 smin_val, smax_val;
4823 u64 umin_val, umax_val;
4824 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
4827 if (insn_bitness == 32) {
4828 /* Relevant for 32-bit RSH: Information can propagate towards
4829 * LSB, so it isn't sufficient to only truncate the output to
4832 coerce_reg_to_size(dst_reg, 4);
4833 coerce_reg_to_size(&src_reg, 4);
4836 smin_val = src_reg.smin_value;
4837 smax_val = src_reg.smax_value;
4838 umin_val = src_reg.umin_value;
4839 umax_val = src_reg.umax_value;
4840 src_known = tnum_is_const(src_reg.var_off);
4841 dst_known = tnum_is_const(dst_reg->var_off);
4843 if ((src_known && (smin_val != smax_val || umin_val != umax_val)) ||
4844 smin_val > smax_val || umin_val > umax_val) {
4845 /* Taint dst register if offset had invalid bounds derived from
4846 * e.g. dead branches.
4848 __mark_reg_unknown(env, dst_reg);
4853 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
4854 __mark_reg_unknown(env, dst_reg);
4858 if (sanitize_needed(opcode)) {
4859 ret = sanitize_val_alu(env, insn);
4861 return sanitize_err(env, insn, ret, NULL, NULL);
4866 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
4867 signed_add_overflows(dst_reg->smax_value, smax_val)) {
4868 dst_reg->smin_value = S64_MIN;
4869 dst_reg->smax_value = S64_MAX;
4871 dst_reg->smin_value += smin_val;
4872 dst_reg->smax_value += smax_val;
4874 if (dst_reg->umin_value + umin_val < umin_val ||
4875 dst_reg->umax_value + umax_val < umax_val) {
4876 dst_reg->umin_value = 0;
4877 dst_reg->umax_value = U64_MAX;
4879 dst_reg->umin_value += umin_val;
4880 dst_reg->umax_value += umax_val;
4882 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
4885 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
4886 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
4887 /* Overflow possible, we know nothing */
4888 dst_reg->smin_value = S64_MIN;
4889 dst_reg->smax_value = S64_MAX;
4891 dst_reg->smin_value -= smax_val;
4892 dst_reg->smax_value -= smin_val;
4894 if (dst_reg->umin_value < umax_val) {
4895 /* Overflow possible, we know nothing */
4896 dst_reg->umin_value = 0;
4897 dst_reg->umax_value = U64_MAX;
4899 /* Cannot overflow (as long as bounds are consistent) */
4900 dst_reg->umin_value -= umax_val;
4901 dst_reg->umax_value -= umin_val;
4903 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
4906 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
4907 if (smin_val < 0 || dst_reg->smin_value < 0) {
4908 /* Ain't nobody got time to multiply that sign */
4909 __mark_reg_unbounded(dst_reg);
4910 __update_reg_bounds(dst_reg);
4913 /* Both values are positive, so we can work with unsigned and
4914 * copy the result to signed (unless it exceeds S64_MAX).
4916 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
4917 /* Potential overflow, we know nothing */
4918 __mark_reg_unbounded(dst_reg);
4919 /* (except what we can learn from the var_off) */
4920 __update_reg_bounds(dst_reg);
4923 dst_reg->umin_value *= umin_val;
4924 dst_reg->umax_value *= umax_val;
4925 if (dst_reg->umax_value > S64_MAX) {
4926 /* Overflow possible, we know nothing */
4927 dst_reg->smin_value = S64_MIN;
4928 dst_reg->smax_value = S64_MAX;
4930 dst_reg->smin_value = dst_reg->umin_value;
4931 dst_reg->smax_value = dst_reg->umax_value;
4935 if (src_known && dst_known) {
4936 __mark_reg_known(dst_reg, dst_reg->var_off.value &
4937 src_reg.var_off.value);
4940 /* We get our minimum from the var_off, since that's inherently
4941 * bitwise. Our maximum is the minimum of the operands' maxima.
4943 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
4944 dst_reg->umin_value = dst_reg->var_off.value;
4945 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
4946 if (dst_reg->smin_value < 0 || smin_val < 0) {
4947 /* Lose signed bounds when ANDing negative numbers,
4948 * ain't nobody got time for that.
4950 dst_reg->smin_value = S64_MIN;
4951 dst_reg->smax_value = S64_MAX;
4953 /* ANDing two positives gives a positive, so safe to
4954 * cast result into s64.
4956 dst_reg->smin_value = dst_reg->umin_value;
4957 dst_reg->smax_value = dst_reg->umax_value;
4959 /* We may learn something more from the var_off */
4960 __update_reg_bounds(dst_reg);
4963 if (src_known && dst_known) {
4964 __mark_reg_known(dst_reg, dst_reg->var_off.value |
4965 src_reg.var_off.value);
4968 /* We get our maximum from the var_off, and our minimum is the
4969 * maximum of the operands' minima
4971 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
4972 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
4973 dst_reg->umax_value = dst_reg->var_off.value |
4974 dst_reg->var_off.mask;
4975 if (dst_reg->smin_value < 0 || smin_val < 0) {
4976 /* Lose signed bounds when ORing negative numbers,
4977 * ain't nobody got time for that.
4979 dst_reg->smin_value = S64_MIN;
4980 dst_reg->smax_value = S64_MAX;
4982 /* ORing two positives gives a positive, so safe to
4983 * cast result into s64.
4985 dst_reg->smin_value = dst_reg->umin_value;
4986 dst_reg->smax_value = dst_reg->umax_value;
4988 /* We may learn something more from the var_off */
4989 __update_reg_bounds(dst_reg);
4992 if (umax_val >= insn_bitness) {
4993 /* Shifts greater than 31 or 63 are undefined.
4994 * This includes shifts by a negative number.
4996 mark_reg_unknown(env, regs, insn->dst_reg);
4999 /* We lose all sign bit information (except what we can pick
5002 dst_reg->smin_value = S64_MIN;
5003 dst_reg->smax_value = S64_MAX;
5004 /* If we might shift our top bit out, then we know nothing */
5005 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
5006 dst_reg->umin_value = 0;
5007 dst_reg->umax_value = U64_MAX;
5009 dst_reg->umin_value <<= umin_val;
5010 dst_reg->umax_value <<= umax_val;
5012 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
5013 /* We may learn something more from the var_off */
5014 __update_reg_bounds(dst_reg);
5017 if (umax_val >= insn_bitness) {
5018 /* Shifts greater than 31 or 63 are undefined.
5019 * This includes shifts by a negative number.
5021 mark_reg_unknown(env, regs, insn->dst_reg);
5024 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
5025 * be negative, then either:
5026 * 1) src_reg might be zero, so the sign bit of the result is
5027 * unknown, so we lose our signed bounds
5028 * 2) it's known negative, thus the unsigned bounds capture the
5030 * 3) the signed bounds cross zero, so they tell us nothing
5032 * If the value in dst_reg is known nonnegative, then again the
5033 * unsigned bounts capture the signed bounds.
5034 * Thus, in all cases it suffices to blow away our signed bounds
5035 * and rely on inferring new ones from the unsigned bounds and
5036 * var_off of the result.
5038 dst_reg->smin_value = S64_MIN;
5039 dst_reg->smax_value = S64_MAX;
5040 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
5041 dst_reg->umin_value >>= umax_val;
5042 dst_reg->umax_value >>= umin_val;
5043 /* We may learn something more from the var_off */
5044 __update_reg_bounds(dst_reg);
5047 if (umax_val >= insn_bitness) {
5048 /* Shifts greater than 31 or 63 are undefined.
5049 * This includes shifts by a negative number.
5051 mark_reg_unknown(env, regs, insn->dst_reg);
5055 /* Upon reaching here, src_known is true and
5056 * umax_val is equal to umin_val.
5058 if (insn_bitness == 32) {
5059 dst_reg->smin_value = (u32)(((s32)dst_reg->smin_value) >> umin_val);
5060 dst_reg->smax_value = (u32)(((s32)dst_reg->smax_value) >> umin_val);
5062 dst_reg->smin_value >>= umin_val;
5063 dst_reg->smax_value >>= umin_val;
5066 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val,
5069 /* blow away the dst_reg umin_value/umax_value and rely on
5070 * dst_reg var_off to refine the result.
5072 dst_reg->umin_value = 0;
5073 dst_reg->umax_value = U64_MAX;
5074 __update_reg_bounds(dst_reg);
5077 mark_reg_unknown(env, regs, insn->dst_reg);
5081 if (BPF_CLASS(insn->code) != BPF_ALU64) {
5082 /* 32-bit ALU ops are (32,32)->32 */
5083 coerce_reg_to_size(dst_reg, 4);
5086 __reg_deduce_bounds(dst_reg);
5087 __reg_bound_offset(dst_reg);
5091 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
5094 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
5095 struct bpf_insn *insn)
5097 struct bpf_verifier_state *vstate = env->cur_state;
5098 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5099 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
5100 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
5101 u8 opcode = BPF_OP(insn->code);
5104 dst_reg = ®s[insn->dst_reg];
5106 if (dst_reg->type != SCALAR_VALUE)
5108 if (BPF_SRC(insn->code) == BPF_X) {
5109 src_reg = ®s[insn->src_reg];
5110 if (src_reg->type != SCALAR_VALUE) {
5111 if (dst_reg->type != SCALAR_VALUE) {
5112 /* Combining two pointers by any ALU op yields
5113 * an arbitrary scalar. Disallow all math except
5114 * pointer subtraction
5116 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
5117 mark_reg_unknown(env, regs, insn->dst_reg);
5120 verbose(env, "R%d pointer %s pointer prohibited\n",
5122 bpf_alu_string[opcode >> 4]);
5125 /* scalar += pointer
5126 * This is legal, but we have to reverse our
5127 * src/dest handling in computing the range
5129 err = mark_chain_precision(env, insn->dst_reg);
5132 return adjust_ptr_min_max_vals(env, insn,
5135 } else if (ptr_reg) {
5136 /* pointer += scalar */
5137 err = mark_chain_precision(env, insn->src_reg);
5140 return adjust_ptr_min_max_vals(env, insn,
5144 /* Pretend the src is a reg with a known value, since we only
5145 * need to be able to read from this state.
5147 off_reg.type = SCALAR_VALUE;
5148 __mark_reg_known(&off_reg, insn->imm);
5150 if (ptr_reg) /* pointer += K */
5151 return adjust_ptr_min_max_vals(env, insn,
5155 /* Got here implies adding two SCALAR_VALUEs */
5156 if (WARN_ON_ONCE(ptr_reg)) {
5157 print_verifier_state(env, state);
5158 verbose(env, "verifier internal error: unexpected ptr_reg\n");
5161 if (WARN_ON(!src_reg)) {
5162 print_verifier_state(env, state);
5163 verbose(env, "verifier internal error: no src_reg\n");
5166 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
5169 /* check validity of 32-bit and 64-bit arithmetic operations */
5170 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
5172 struct bpf_reg_state *regs = cur_regs(env);
5173 u8 opcode = BPF_OP(insn->code);
5176 if (opcode == BPF_END || opcode == BPF_NEG) {
5177 if (opcode == BPF_NEG) {
5178 if (BPF_SRC(insn->code) != 0 ||
5179 insn->src_reg != BPF_REG_0 ||
5180 insn->off != 0 || insn->imm != 0) {
5181 verbose(env, "BPF_NEG uses reserved fields\n");
5185 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
5186 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
5187 BPF_CLASS(insn->code) == BPF_ALU64) {
5188 verbose(env, "BPF_END uses reserved fields\n");
5193 /* check src operand */
5194 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5198 if (is_pointer_value(env, insn->dst_reg)) {
5199 verbose(env, "R%d pointer arithmetic prohibited\n",
5204 /* check dest operand */
5205 err = check_reg_arg(env, insn->dst_reg, DST_OP);
5209 } else if (opcode == BPF_MOV) {
5211 if (BPF_SRC(insn->code) == BPF_X) {
5212 if (insn->imm != 0 || insn->off != 0) {
5213 verbose(env, "BPF_MOV uses reserved fields\n");
5217 /* check src operand */
5218 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5222 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
5223 verbose(env, "BPF_MOV uses reserved fields\n");
5228 /* check dest operand, mark as required later */
5229 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
5233 if (BPF_SRC(insn->code) == BPF_X) {
5234 struct bpf_reg_state *src_reg = regs + insn->src_reg;
5235 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
5237 if (BPF_CLASS(insn->code) == BPF_ALU64) {
5239 * copy register state to dest reg
5241 *dst_reg = *src_reg;
5242 dst_reg->live |= REG_LIVE_WRITTEN;
5243 dst_reg->subreg_def = DEF_NOT_SUBREG;
5246 if (is_pointer_value(env, insn->src_reg)) {
5248 "R%d partial copy of pointer\n",
5251 } else if (src_reg->type == SCALAR_VALUE) {
5252 *dst_reg = *src_reg;
5253 dst_reg->live |= REG_LIVE_WRITTEN;
5254 dst_reg->subreg_def = env->insn_idx + 1;
5256 mark_reg_unknown(env, regs,
5259 coerce_reg_to_size(dst_reg, 4);
5263 * remember the value we stored into this reg
5265 /* clear any state __mark_reg_known doesn't set */
5266 mark_reg_unknown(env, regs, insn->dst_reg);
5267 regs[insn->dst_reg].type = SCALAR_VALUE;
5268 if (BPF_CLASS(insn->code) == BPF_ALU64) {
5269 __mark_reg_known(regs + insn->dst_reg,
5272 __mark_reg_known(regs + insn->dst_reg,
5277 } else if (opcode > BPF_END) {
5278 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
5281 } else { /* all other ALU ops: and, sub, xor, add, ... */
5283 if (BPF_SRC(insn->code) == BPF_X) {
5284 if (insn->imm != 0 || insn->off != 0) {
5285 verbose(env, "BPF_ALU uses reserved fields\n");
5288 /* check src1 operand */
5289 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5293 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
5294 verbose(env, "BPF_ALU uses reserved fields\n");
5299 /* check src2 operand */
5300 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5304 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
5305 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
5306 verbose(env, "div by zero\n");
5310 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
5311 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
5312 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
5314 if (insn->imm < 0 || insn->imm >= size) {
5315 verbose(env, "invalid shift %d\n", insn->imm);
5320 /* check dest operand */
5321 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
5325 return adjust_reg_min_max_vals(env, insn);
5331 static void __find_good_pkt_pointers(struct bpf_func_state *state,
5332 struct bpf_reg_state *dst_reg,
5333 enum bpf_reg_type type, u16 new_range)
5335 struct bpf_reg_state *reg;
5338 for (i = 0; i < MAX_BPF_REG; i++) {
5339 reg = &state->regs[i];
5340 if (reg->type == type && reg->id == dst_reg->id)
5341 /* keep the maximum range already checked */
5342 reg->range = max(reg->range, new_range);
5345 bpf_for_each_spilled_reg(i, state, reg) {
5348 if (reg->type == type && reg->id == dst_reg->id)
5349 reg->range = max(reg->range, new_range);
5353 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
5354 struct bpf_reg_state *dst_reg,
5355 enum bpf_reg_type type,
5356 bool range_right_open)
5361 if (dst_reg->off < 0 ||
5362 (dst_reg->off == 0 && range_right_open))
5363 /* This doesn't give us any range */
5366 if (dst_reg->umax_value > MAX_PACKET_OFF ||
5367 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
5368 /* Risk of overflow. For instance, ptr + (1<<63) may be less
5369 * than pkt_end, but that's because it's also less than pkt.
5373 new_range = dst_reg->off;
5374 if (range_right_open)
5377 /* Examples for register markings:
5379 * pkt_data in dst register:
5383 * if (r2 > pkt_end) goto <handle exception>
5388 * if (r2 < pkt_end) goto <access okay>
5389 * <handle exception>
5392 * r2 == dst_reg, pkt_end == src_reg
5393 * r2=pkt(id=n,off=8,r=0)
5394 * r3=pkt(id=n,off=0,r=0)
5396 * pkt_data in src register:
5400 * if (pkt_end >= r2) goto <access okay>
5401 * <handle exception>
5405 * if (pkt_end <= r2) goto <handle exception>
5409 * pkt_end == dst_reg, r2 == src_reg
5410 * r2=pkt(id=n,off=8,r=0)
5411 * r3=pkt(id=n,off=0,r=0)
5413 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
5414 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
5415 * and [r3, r3 + 8-1) respectively is safe to access depending on
5419 /* If our ids match, then we must have the same max_value. And we
5420 * don't care about the other reg's fixed offset, since if it's too big
5421 * the range won't allow anything.
5422 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
5424 for (i = 0; i <= vstate->curframe; i++)
5425 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
5429 /* compute branch direction of the expression "if (reg opcode val) goto target;"
5431 * 1 - branch will be taken and "goto target" will be executed
5432 * 0 - branch will not be taken and fall-through to next insn
5433 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value range [0,10]
5435 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
5438 struct bpf_reg_state reg_lo;
5441 if (__is_pointer_value(false, reg))
5447 /* For JMP32, only low 32 bits are compared, coerce_reg_to_size
5448 * could truncate high bits and update umin/umax according to
5449 * information of low bits.
5451 coerce_reg_to_size(reg, 4);
5452 /* smin/smax need special handling. For example, after coerce,
5453 * if smin_value is 0x00000000ffffffffLL, the value is -1 when
5454 * used as operand to JMP32. It is a negative number from s32's
5455 * point of view, while it is a positive number when seen as
5456 * s64. The smin/smax are kept as s64, therefore, when used with
5457 * JMP32, they need to be transformed into s32, then sign
5458 * extended back to s64.
5460 * Also, smin/smax were copied from umin/umax. If umin/umax has
5461 * different sign bit, then min/max relationship doesn't
5462 * maintain after casting into s32, for this case, set smin/smax
5465 if ((reg->umax_value ^ reg->umin_value) &
5467 reg->smin_value = S32_MIN;
5468 reg->smax_value = S32_MAX;
5470 reg->smin_value = (s64)(s32)reg->smin_value;
5471 reg->smax_value = (s64)(s32)reg->smax_value;
5474 sval = (s64)(s32)val;
5481 if (tnum_is_const(reg->var_off))
5482 return !!tnum_equals_const(reg->var_off, val);
5485 if (tnum_is_const(reg->var_off))
5486 return !tnum_equals_const(reg->var_off, val);
5489 if ((~reg->var_off.mask & reg->var_off.value) & val)
5491 if (!((reg->var_off.mask | reg->var_off.value) & val))
5495 if (reg->umin_value > val)
5497 else if (reg->umax_value <= val)
5501 if (reg->smin_value > sval)
5503 else if (reg->smax_value < sval)
5507 if (reg->umax_value < val)
5509 else if (reg->umin_value >= val)
5513 if (reg->smax_value < sval)
5515 else if (reg->smin_value >= sval)
5519 if (reg->umin_value >= val)
5521 else if (reg->umax_value < val)
5525 if (reg->smin_value >= sval)
5527 else if (reg->smax_value < sval)
5531 if (reg->umax_value <= val)
5533 else if (reg->umin_value > val)
5537 if (reg->smax_value <= sval)
5539 else if (reg->smin_value > sval)
5547 /* Generate min value of the high 32-bit from TNUM info. */
5548 static u64 gen_hi_min(struct tnum var)
5550 return var.value & ~0xffffffffULL;
5553 /* Generate max value of the high 32-bit from TNUM info. */
5554 static u64 gen_hi_max(struct tnum var)
5556 return (var.value | var.mask) & ~0xffffffffULL;
5559 /* Return true if VAL is compared with a s64 sign extended from s32, and they
5560 * are with the same signedness.
5562 static bool cmp_val_with_extended_s64(s64 sval, struct bpf_reg_state *reg)
5564 return ((s32)sval >= 0 &&
5565 reg->smin_value >= 0 && reg->smax_value <= S32_MAX) ||
5567 reg->smax_value <= 0 && reg->smin_value >= S32_MIN);
5570 /* Constrain the possible values of @reg with unsigned upper bound @bound.
5571 * If @is_exclusive, @bound is an exclusive limit, otherwise it is inclusive.
5572 * If @is_jmp32, @bound is a 32-bit value that only constrains the low 32 bits
5575 static void set_upper_bound(struct bpf_reg_state *reg, u64 bound, bool is_jmp32,
5579 /* There are no values for `reg` that make `reg<0` true. */
5585 /* Constrain the register's value in the tnum representation.
5586 * For 64-bit comparisons this happens later in
5587 * __reg_bound_offset(), but for 32-bit comparisons, we can be
5588 * more precise than what can be derived from the updated
5591 struct tnum t = tnum_range(0, bound);
5593 t.mask |= ~0xffffffffULL; /* upper half is unknown */
5594 reg->var_off = tnum_intersect(reg->var_off, t);
5596 /* Compute the 64-bit bound from the 32-bit bound. */
5597 bound += gen_hi_max(reg->var_off);
5599 reg->umax_value = min(reg->umax_value, bound);
5602 /* Constrain the possible values of @reg with unsigned lower bound @bound.
5603 * If @is_exclusive, @bound is an exclusive limit, otherwise it is inclusive.
5604 * If @is_jmp32, @bound is a 32-bit value that only constrains the low 32 bits
5607 static void set_lower_bound(struct bpf_reg_state *reg, u64 bound, bool is_jmp32,
5611 /* There are no values for `reg` that make `reg>MAX` true. */
5612 if (bound == (is_jmp32 ? U32_MAX : U64_MAX))
5617 /* Constrain the register's value in the tnum representation.
5618 * For 64-bit comparisons this happens later in
5619 * __reg_bound_offset(), but for 32-bit comparisons, we can be
5620 * more precise than what can be derived from the updated
5623 struct tnum t = tnum_range(bound, U32_MAX);
5625 t.mask |= ~0xffffffffULL; /* upper half is unknown */
5626 reg->var_off = tnum_intersect(reg->var_off, t);
5628 /* Compute the 64-bit bound from the 32-bit bound. */
5629 bound += gen_hi_min(reg->var_off);
5631 reg->umin_value = max(reg->umin_value, bound);
5634 /* Adjusts the register min/max values in the case that the dst_reg is the
5635 * variable register that we are working on, and src_reg is a constant or we're
5636 * simply doing a BPF_K check.
5637 * In JEQ/JNE cases we also adjust the var_off values.
5639 static void reg_set_min_max(struct bpf_reg_state *true_reg,
5640 struct bpf_reg_state *false_reg, u64 val,
5641 u8 opcode, bool is_jmp32)
5645 /* If the dst_reg is a pointer, we can't learn anything about its
5646 * variable offset from the compare (unless src_reg were a pointer into
5647 * the same object, but we don't bother with that.
5648 * Since false_reg and true_reg have the same type by construction, we
5649 * only need to check one of them for pointerness.
5651 if (__is_pointer_value(false, false_reg))
5654 val = is_jmp32 ? (u32)val : val;
5655 sval = is_jmp32 ? (s64)(s32)val : (s64)val;
5661 struct bpf_reg_state *reg =
5662 opcode == BPF_JEQ ? true_reg : false_reg;
5664 /* For BPF_JEQ, if this is false we know nothing Jon Snow, but
5665 * if it is true we know the value for sure. Likewise for
5669 u64 old_v = reg->var_off.value;
5670 u64 hi_mask = ~0xffffffffULL;
5672 reg->var_off.value = (old_v & hi_mask) | val;
5673 reg->var_off.mask &= hi_mask;
5675 __mark_reg_known(reg, val);
5680 false_reg->var_off = tnum_and(false_reg->var_off,
5682 if (is_power_of_2(val))
5683 true_reg->var_off = tnum_or(true_reg->var_off,
5689 set_upper_bound(false_reg, val, is_jmp32, opcode == BPF_JGE);
5690 set_lower_bound(true_reg, val, is_jmp32, opcode == BPF_JGT);
5696 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
5697 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
5699 /* If the full s64 was not sign-extended from s32 then don't
5700 * deduct further info.
5702 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
5704 false_reg->smax_value = min(false_reg->smax_value, false_smax);
5705 true_reg->smin_value = max(true_reg->smin_value, true_smin);
5711 set_lower_bound(false_reg, val, is_jmp32, opcode == BPF_JLE);
5712 set_upper_bound(true_reg, val, is_jmp32, opcode == BPF_JLT);
5718 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
5719 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
5721 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
5723 false_reg->smin_value = max(false_reg->smin_value, false_smin);
5724 true_reg->smax_value = min(true_reg->smax_value, true_smax);
5731 __reg_deduce_bounds(false_reg);
5732 __reg_deduce_bounds(true_reg);
5733 /* We might have learned some bits from the bounds. */
5734 __reg_bound_offset(false_reg);
5735 __reg_bound_offset(true_reg);
5736 /* Intersecting with the old var_off might have improved our bounds
5737 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
5738 * then new var_off is (0; 0x7f...fc) which improves our umax.
5740 __update_reg_bounds(false_reg);
5741 __update_reg_bounds(true_reg);
5744 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
5747 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
5748 struct bpf_reg_state *false_reg, u64 val,
5749 u8 opcode, bool is_jmp32)
5753 if (__is_pointer_value(false, false_reg))
5756 val = is_jmp32 ? (u32)val : val;
5757 sval = is_jmp32 ? (s64)(s32)val : (s64)val;
5763 struct bpf_reg_state *reg =
5764 opcode == BPF_JEQ ? true_reg : false_reg;
5767 u64 old_v = reg->var_off.value;
5768 u64 hi_mask = ~0xffffffffULL;
5770 reg->var_off.value = (old_v & hi_mask) | val;
5771 reg->var_off.mask &= hi_mask;
5773 __mark_reg_known(reg, val);
5778 false_reg->var_off = tnum_and(false_reg->var_off,
5780 if (is_power_of_2(val))
5781 true_reg->var_off = tnum_or(true_reg->var_off,
5787 set_lower_bound(false_reg, val, is_jmp32, opcode == BPF_JGE);
5788 set_upper_bound(true_reg, val, is_jmp32, opcode == BPF_JGT);
5794 s64 false_smin = opcode == BPF_JSGT ? sval : sval + 1;
5795 s64 true_smax = opcode == BPF_JSGT ? sval - 1 : sval;
5797 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
5799 false_reg->smin_value = max(false_reg->smin_value, false_smin);
5800 true_reg->smax_value = min(true_reg->smax_value, true_smax);
5806 set_upper_bound(false_reg, val, is_jmp32, opcode == BPF_JLE);
5807 set_lower_bound(true_reg, val, is_jmp32, opcode == BPF_JLT);
5813 s64 false_smax = opcode == BPF_JSLT ? sval : sval - 1;
5814 s64 true_smin = opcode == BPF_JSLT ? sval + 1 : sval;
5816 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
5818 false_reg->smax_value = min(false_reg->smax_value, false_smax);
5819 true_reg->smin_value = max(true_reg->smin_value, true_smin);
5826 __reg_deduce_bounds(false_reg);
5827 __reg_deduce_bounds(true_reg);
5828 /* We might have learned some bits from the bounds. */
5829 __reg_bound_offset(false_reg);
5830 __reg_bound_offset(true_reg);
5831 /* Intersecting with the old var_off might have improved our bounds
5832 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
5833 * then new var_off is (0; 0x7f...fc) which improves our umax.
5835 __update_reg_bounds(false_reg);
5836 __update_reg_bounds(true_reg);
5839 /* Regs are known to be equal, so intersect their min/max/var_off */
5840 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
5841 struct bpf_reg_state *dst_reg)
5843 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
5844 dst_reg->umin_value);
5845 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
5846 dst_reg->umax_value);
5847 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
5848 dst_reg->smin_value);
5849 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
5850 dst_reg->smax_value);
5851 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
5853 /* We might have learned new bounds from the var_off. */
5854 __update_reg_bounds(src_reg);
5855 __update_reg_bounds(dst_reg);
5856 /* We might have learned something about the sign bit. */
5857 __reg_deduce_bounds(src_reg);
5858 __reg_deduce_bounds(dst_reg);
5859 /* We might have learned some bits from the bounds. */
5860 __reg_bound_offset(src_reg);
5861 __reg_bound_offset(dst_reg);
5862 /* Intersecting with the old var_off might have improved our bounds
5863 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
5864 * then new var_off is (0; 0x7f...fc) which improves our umax.
5866 __update_reg_bounds(src_reg);
5867 __update_reg_bounds(dst_reg);
5870 static void reg_combine_min_max(struct bpf_reg_state *true_src,
5871 struct bpf_reg_state *true_dst,
5872 struct bpf_reg_state *false_src,
5873 struct bpf_reg_state *false_dst,
5878 __reg_combine_min_max(true_src, true_dst);
5881 __reg_combine_min_max(false_src, false_dst);
5886 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
5887 struct bpf_reg_state *reg, u32 id,
5890 if (reg_type_may_be_null(reg->type) && reg->id == id) {
5891 /* Old offset (both fixed and variable parts) should
5892 * have been known-zero, because we don't allow pointer
5893 * arithmetic on pointers that might be NULL.
5895 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
5896 !tnum_equals_const(reg->var_off, 0) ||
5898 __mark_reg_known_zero(reg);
5902 reg->type = SCALAR_VALUE;
5903 } else if (reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
5904 if (reg->map_ptr->inner_map_meta) {
5905 reg->type = CONST_PTR_TO_MAP;
5906 reg->map_ptr = reg->map_ptr->inner_map_meta;
5907 } else if (reg->map_ptr->map_type ==
5908 BPF_MAP_TYPE_XSKMAP) {
5909 reg->type = PTR_TO_XDP_SOCK;
5911 reg->type = PTR_TO_MAP_VALUE;
5913 } else if (reg->type == PTR_TO_SOCKET_OR_NULL) {
5914 reg->type = PTR_TO_SOCKET;
5915 } else if (reg->type == PTR_TO_SOCK_COMMON_OR_NULL) {
5916 reg->type = PTR_TO_SOCK_COMMON;
5917 } else if (reg->type == PTR_TO_TCP_SOCK_OR_NULL) {
5918 reg->type = PTR_TO_TCP_SOCK;
5921 /* We don't need id and ref_obj_id from this point
5922 * onwards anymore, thus we should better reset it,
5923 * so that state pruning has chances to take effect.
5926 reg->ref_obj_id = 0;
5927 } else if (!reg_may_point_to_spin_lock(reg)) {
5928 /* For not-NULL ptr, reg->ref_obj_id will be reset
5929 * in release_reg_references().
5931 * reg->id is still used by spin_lock ptr. Other
5932 * than spin_lock ptr type, reg->id can be reset.
5939 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
5942 struct bpf_reg_state *reg;
5945 for (i = 0; i < MAX_BPF_REG; i++)
5946 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
5948 bpf_for_each_spilled_reg(i, state, reg) {
5951 mark_ptr_or_null_reg(state, reg, id, is_null);
5955 /* The logic is similar to find_good_pkt_pointers(), both could eventually
5956 * be folded together at some point.
5958 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
5961 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5962 struct bpf_reg_state *regs = state->regs;
5963 u32 ref_obj_id = regs[regno].ref_obj_id;
5964 u32 id = regs[regno].id;
5967 if (ref_obj_id && ref_obj_id == id && is_null)
5968 /* regs[regno] is in the " == NULL" branch.
5969 * No one could have freed the reference state before
5970 * doing the NULL check.
5972 WARN_ON_ONCE(release_reference_state(state, id));
5974 for (i = 0; i <= vstate->curframe; i++)
5975 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
5978 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
5979 struct bpf_reg_state *dst_reg,
5980 struct bpf_reg_state *src_reg,
5981 struct bpf_verifier_state *this_branch,
5982 struct bpf_verifier_state *other_branch)
5984 if (BPF_SRC(insn->code) != BPF_X)
5987 /* Pointers are always 64-bit. */
5988 if (BPF_CLASS(insn->code) == BPF_JMP32)
5991 switch (BPF_OP(insn->code)) {
5993 if ((dst_reg->type == PTR_TO_PACKET &&
5994 src_reg->type == PTR_TO_PACKET_END) ||
5995 (dst_reg->type == PTR_TO_PACKET_META &&
5996 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
5997 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
5998 find_good_pkt_pointers(this_branch, dst_reg,
5999 dst_reg->type, false);
6000 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6001 src_reg->type == PTR_TO_PACKET) ||
6002 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6003 src_reg->type == PTR_TO_PACKET_META)) {
6004 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
6005 find_good_pkt_pointers(other_branch, src_reg,
6006 src_reg->type, true);
6012 if ((dst_reg->type == PTR_TO_PACKET &&
6013 src_reg->type == PTR_TO_PACKET_END) ||
6014 (dst_reg->type == PTR_TO_PACKET_META &&
6015 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
6016 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
6017 find_good_pkt_pointers(other_branch, dst_reg,
6018 dst_reg->type, true);
6019 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6020 src_reg->type == PTR_TO_PACKET) ||
6021 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6022 src_reg->type == PTR_TO_PACKET_META)) {
6023 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
6024 find_good_pkt_pointers(this_branch, src_reg,
6025 src_reg->type, false);
6031 if ((dst_reg->type == PTR_TO_PACKET &&
6032 src_reg->type == PTR_TO_PACKET_END) ||
6033 (dst_reg->type == PTR_TO_PACKET_META &&
6034 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
6035 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
6036 find_good_pkt_pointers(this_branch, dst_reg,
6037 dst_reg->type, true);
6038 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6039 src_reg->type == PTR_TO_PACKET) ||
6040 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6041 src_reg->type == PTR_TO_PACKET_META)) {
6042 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
6043 find_good_pkt_pointers(other_branch, src_reg,
6044 src_reg->type, false);
6050 if ((dst_reg->type == PTR_TO_PACKET &&
6051 src_reg->type == PTR_TO_PACKET_END) ||
6052 (dst_reg->type == PTR_TO_PACKET_META &&
6053 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
6054 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
6055 find_good_pkt_pointers(other_branch, dst_reg,
6056 dst_reg->type, false);
6057 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6058 src_reg->type == PTR_TO_PACKET) ||
6059 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6060 src_reg->type == PTR_TO_PACKET_META)) {
6061 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
6062 find_good_pkt_pointers(this_branch, src_reg,
6063 src_reg->type, true);
6075 static int check_cond_jmp_op(struct bpf_verifier_env *env,
6076 struct bpf_insn *insn, int *insn_idx)
6078 struct bpf_verifier_state *this_branch = env->cur_state;
6079 struct bpf_verifier_state *other_branch;
6080 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
6081 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
6082 u8 opcode = BPF_OP(insn->code);
6087 /* Only conditional jumps are expected to reach here. */
6088 if (opcode == BPF_JA || opcode > BPF_JSLE) {
6089 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
6093 if (BPF_SRC(insn->code) == BPF_X) {
6094 if (insn->imm != 0) {
6095 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
6099 /* check src1 operand */
6100 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6104 if (is_pointer_value(env, insn->src_reg)) {
6105 verbose(env, "R%d pointer comparison prohibited\n",
6109 src_reg = ®s[insn->src_reg];
6111 if (insn->src_reg != BPF_REG_0) {
6112 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
6117 /* check src2 operand */
6118 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6122 dst_reg = ®s[insn->dst_reg];
6123 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
6125 if (BPF_SRC(insn->code) == BPF_K)
6126 pred = is_branch_taken(dst_reg, insn->imm,
6128 else if (src_reg->type == SCALAR_VALUE &&
6129 tnum_is_const(src_reg->var_off))
6130 pred = is_branch_taken(dst_reg, src_reg->var_off.value,
6133 err = mark_chain_precision(env, insn->dst_reg);
6134 if (BPF_SRC(insn->code) == BPF_X && !err)
6135 err = mark_chain_precision(env, insn->src_reg);
6141 /* Only follow the goto, ignore fall-through. If needed, push
6142 * the fall-through branch for simulation under speculative
6145 if (!env->allow_ptr_leaks &&
6146 !sanitize_speculative_path(env, insn, *insn_idx + 1,
6149 *insn_idx += insn->off;
6151 } else if (pred == 0) {
6152 /* Only follow the fall-through branch, since that's where the
6153 * program will go. If needed, push the goto branch for
6154 * simulation under speculative execution.
6156 if (!env->allow_ptr_leaks &&
6157 !sanitize_speculative_path(env, insn,
6158 *insn_idx + insn->off + 1,
6164 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
6168 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
6170 /* detect if we are comparing against a constant value so we can adjust
6171 * our min/max values for our dst register.
6172 * this is only legit if both are scalars (or pointers to the same
6173 * object, I suppose, but we don't support that right now), because
6174 * otherwise the different base pointers mean the offsets aren't
6177 if (BPF_SRC(insn->code) == BPF_X) {
6178 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
6179 struct bpf_reg_state lo_reg0 = *dst_reg;
6180 struct bpf_reg_state lo_reg1 = *src_reg;
6181 struct bpf_reg_state *src_lo, *dst_lo;
6185 coerce_reg_to_size(dst_lo, 4);
6186 coerce_reg_to_size(src_lo, 4);
6188 if (dst_reg->type == SCALAR_VALUE &&
6189 src_reg->type == SCALAR_VALUE) {
6190 if (tnum_is_const(src_reg->var_off) ||
6191 (is_jmp32 && tnum_is_const(src_lo->var_off)))
6192 reg_set_min_max(&other_branch_regs[insn->dst_reg],
6195 ? src_lo->var_off.value
6196 : src_reg->var_off.value,
6198 else if (tnum_is_const(dst_reg->var_off) ||
6199 (is_jmp32 && tnum_is_const(dst_lo->var_off)))
6200 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
6203 ? dst_lo->var_off.value
6204 : dst_reg->var_off.value,
6206 else if (!is_jmp32 &&
6207 (opcode == BPF_JEQ || opcode == BPF_JNE))
6208 /* Comparing for equality, we can combine knowledge */
6209 reg_combine_min_max(&other_branch_regs[insn->src_reg],
6210 &other_branch_regs[insn->dst_reg],
6211 src_reg, dst_reg, opcode);
6213 } else if (dst_reg->type == SCALAR_VALUE) {
6214 reg_set_min_max(&other_branch_regs[insn->dst_reg],
6215 dst_reg, insn->imm, opcode, is_jmp32);
6218 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
6219 * NOTE: these optimizations below are related with pointer comparison
6220 * which will never be JMP32.
6222 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
6223 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
6224 reg_type_may_be_null(dst_reg->type)) {
6225 /* Mark all identical registers in each branch as either
6226 * safe or unknown depending R == 0 or R != 0 conditional.
6228 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
6230 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
6232 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
6233 this_branch, other_branch) &&
6234 is_pointer_value(env, insn->dst_reg)) {
6235 verbose(env, "R%d pointer comparison prohibited\n",
6239 if (env->log.level & BPF_LOG_LEVEL)
6240 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
6244 /* verify BPF_LD_IMM64 instruction */
6245 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
6247 struct bpf_insn_aux_data *aux = cur_aux(env);
6248 struct bpf_reg_state *regs = cur_regs(env);
6249 struct bpf_map *map;
6252 if (BPF_SIZE(insn->code) != BPF_DW) {
6253 verbose(env, "invalid BPF_LD_IMM insn\n");
6256 if (insn->off != 0) {
6257 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
6261 err = check_reg_arg(env, insn->dst_reg, DST_OP);
6265 if (insn->src_reg == 0) {
6266 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
6268 regs[insn->dst_reg].type = SCALAR_VALUE;
6269 __mark_reg_known(®s[insn->dst_reg], imm);
6273 map = env->used_maps[aux->map_index];
6274 mark_reg_known_zero(env, regs, insn->dst_reg);
6275 regs[insn->dst_reg].map_ptr = map;
6277 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE) {
6278 regs[insn->dst_reg].type = PTR_TO_MAP_VALUE;
6279 regs[insn->dst_reg].off = aux->map_off;
6280 if (map_value_has_spin_lock(map))
6281 regs[insn->dst_reg].id = ++env->id_gen;
6282 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
6283 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
6285 verbose(env, "bpf verifier is misconfigured\n");
6292 static bool may_access_skb(enum bpf_prog_type type)
6295 case BPF_PROG_TYPE_SOCKET_FILTER:
6296 case BPF_PROG_TYPE_SCHED_CLS:
6297 case BPF_PROG_TYPE_SCHED_ACT:
6304 /* verify safety of LD_ABS|LD_IND instructions:
6305 * - they can only appear in the programs where ctx == skb
6306 * - since they are wrappers of function calls, they scratch R1-R5 registers,
6307 * preserve R6-R9, and store return value into R0
6310 * ctx == skb == R6 == CTX
6313 * SRC == any register
6314 * IMM == 32-bit immediate
6317 * R0 - 8/16/32-bit skb data converted to cpu endianness
6319 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
6321 struct bpf_reg_state *regs = cur_regs(env);
6322 static const int ctx_reg = BPF_REG_6;
6323 u8 mode = BPF_MODE(insn->code);
6326 if (!may_access_skb(env->prog->type)) {
6327 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
6331 if (!env->ops->gen_ld_abs) {
6332 verbose(env, "bpf verifier is misconfigured\n");
6336 if (env->subprog_cnt > 1) {
6337 /* when program has LD_ABS insn JITs and interpreter assume
6338 * that r1 == ctx == skb which is not the case for callees
6339 * that can have arbitrary arguments. It's problematic
6340 * for main prog as well since JITs would need to analyze
6341 * all functions in order to make proper register save/restore
6342 * decisions in the main prog. Hence disallow LD_ABS with calls
6344 verbose(env, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
6348 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
6349 BPF_SIZE(insn->code) == BPF_DW ||
6350 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
6351 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
6355 /* check whether implicit source operand (register R6) is readable */
6356 err = check_reg_arg(env, ctx_reg, SRC_OP);
6360 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
6361 * gen_ld_abs() may terminate the program at runtime, leading to
6364 err = check_reference_leak(env);
6366 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
6370 if (env->cur_state->active_spin_lock) {
6371 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
6375 if (regs[ctx_reg].type != PTR_TO_CTX) {
6377 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
6381 if (mode == BPF_IND) {
6382 /* check explicit source operand */
6383 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6388 err = check_ctx_reg(env, ®s[ctx_reg], ctx_reg);
6392 /* reset caller saved regs to unreadable */
6393 for (i = 0; i < CALLER_SAVED_REGS; i++) {
6394 mark_reg_not_init(env, regs, caller_saved[i]);
6395 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
6398 /* mark destination R0 register as readable, since it contains
6399 * the value fetched from the packet.
6400 * Already marked as written above.
6402 mark_reg_unknown(env, regs, BPF_REG_0);
6403 /* ld_abs load up to 32-bit skb data. */
6404 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
6408 static int check_return_code(struct bpf_verifier_env *env)
6410 struct tnum enforce_attach_type_range = tnum_unknown;
6411 struct bpf_reg_state *reg;
6412 struct tnum range = tnum_range(0, 1);
6414 switch (env->prog->type) {
6415 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
6416 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
6417 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG)
6418 range = tnum_range(1, 1);
6420 case BPF_PROG_TYPE_CGROUP_SKB:
6421 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
6422 range = tnum_range(0, 3);
6423 enforce_attach_type_range = tnum_range(2, 3);
6426 case BPF_PROG_TYPE_CGROUP_SOCK:
6427 case BPF_PROG_TYPE_SOCK_OPS:
6428 case BPF_PROG_TYPE_CGROUP_DEVICE:
6429 case BPF_PROG_TYPE_CGROUP_SYSCTL:
6430 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
6436 reg = cur_regs(env) + BPF_REG_0;
6437 if (reg->type != SCALAR_VALUE) {
6438 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
6439 reg_type_str[reg->type]);
6443 if (!tnum_in(range, reg->var_off)) {
6446 verbose(env, "At program exit the register R0 ");
6447 if (!tnum_is_unknown(reg->var_off)) {
6448 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6449 verbose(env, "has value %s", tn_buf);
6451 verbose(env, "has unknown scalar value");
6453 tnum_strn(tn_buf, sizeof(tn_buf), range);
6454 verbose(env, " should have been in %s\n", tn_buf);
6458 if (!tnum_is_unknown(enforce_attach_type_range) &&
6459 tnum_in(enforce_attach_type_range, reg->var_off))
6460 env->prog->enforce_expected_attach_type = 1;
6464 /* non-recursive DFS pseudo code
6465 * 1 procedure DFS-iterative(G,v):
6466 * 2 label v as discovered
6467 * 3 let S be a stack
6469 * 5 while S is not empty
6471 * 7 if t is what we're looking for:
6473 * 9 for all edges e in G.adjacentEdges(t) do
6474 * 10 if edge e is already labelled
6475 * 11 continue with the next edge
6476 * 12 w <- G.adjacentVertex(t,e)
6477 * 13 if vertex w is not discovered and not explored
6478 * 14 label e as tree-edge
6479 * 15 label w as discovered
6482 * 18 else if vertex w is discovered
6483 * 19 label e as back-edge
6485 * 21 // vertex w is explored
6486 * 22 label e as forward- or cross-edge
6487 * 23 label t as explored
6492 * 0x11 - discovered and fall-through edge labelled
6493 * 0x12 - discovered and fall-through and branch edges labelled
6504 static u32 state_htab_size(struct bpf_verifier_env *env)
6506 return env->prog->len;
6509 static struct bpf_verifier_state_list **explored_state(
6510 struct bpf_verifier_env *env,
6513 struct bpf_verifier_state *cur = env->cur_state;
6514 struct bpf_func_state *state = cur->frame[cur->curframe];
6516 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
6519 static void init_explored_state(struct bpf_verifier_env *env, int idx)
6521 env->insn_aux_data[idx].prune_point = true;
6524 /* t, w, e - match pseudo-code above:
6525 * t - index of current instruction
6526 * w - next instruction
6529 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
6532 int *insn_stack = env->cfg.insn_stack;
6533 int *insn_state = env->cfg.insn_state;
6535 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
6538 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
6541 if (w < 0 || w >= env->prog->len) {
6542 verbose_linfo(env, t, "%d: ", t);
6543 verbose(env, "jump out of range from insn %d to %d\n", t, w);
6548 /* mark branch target for state pruning */
6549 init_explored_state(env, w);
6551 if (insn_state[w] == 0) {
6553 insn_state[t] = DISCOVERED | e;
6554 insn_state[w] = DISCOVERED;
6555 if (env->cfg.cur_stack >= env->prog->len)
6557 insn_stack[env->cfg.cur_stack++] = w;
6559 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
6560 if (loop_ok && env->allow_ptr_leaks)
6562 verbose_linfo(env, t, "%d: ", t);
6563 verbose_linfo(env, w, "%d: ", w);
6564 verbose(env, "back-edge from insn %d to %d\n", t, w);
6566 } else if (insn_state[w] == EXPLORED) {
6567 /* forward- or cross-edge */
6568 insn_state[t] = DISCOVERED | e;
6570 verbose(env, "insn state internal bug\n");
6576 /* non-recursive depth-first-search to detect loops in BPF program
6577 * loop == back-edge in directed graph
6579 static int check_cfg(struct bpf_verifier_env *env)
6581 struct bpf_insn *insns = env->prog->insnsi;
6582 int insn_cnt = env->prog->len;
6583 int *insn_stack, *insn_state;
6587 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
6591 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
6597 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
6598 insn_stack[0] = 0; /* 0 is the first instruction */
6599 env->cfg.cur_stack = 1;
6602 if (env->cfg.cur_stack == 0)
6604 t = insn_stack[env->cfg.cur_stack - 1];
6606 if (BPF_CLASS(insns[t].code) == BPF_JMP ||
6607 BPF_CLASS(insns[t].code) == BPF_JMP32) {
6608 u8 opcode = BPF_OP(insns[t].code);
6610 if (opcode == BPF_EXIT) {
6612 } else if (opcode == BPF_CALL) {
6613 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
6618 if (t + 1 < insn_cnt)
6619 init_explored_state(env, t + 1);
6620 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
6621 init_explored_state(env, t);
6622 ret = push_insn(t, t + insns[t].imm + 1, BRANCH,
6629 } else if (opcode == BPF_JA) {
6630 if (BPF_SRC(insns[t].code) != BPF_K) {
6634 /* unconditional jump with single edge */
6635 ret = push_insn(t, t + insns[t].off + 1,
6636 FALLTHROUGH, env, true);
6641 /* unconditional jmp is not a good pruning point,
6642 * but it's marked, since backtracking needs
6643 * to record jmp history in is_state_visited().
6645 init_explored_state(env, t + insns[t].off + 1);
6646 /* tell verifier to check for equivalent states
6647 * after every call and jump
6649 if (t + 1 < insn_cnt)
6650 init_explored_state(env, t + 1);
6652 /* conditional jump with two edges */
6653 init_explored_state(env, t);
6654 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
6660 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
6667 /* all other non-branch instructions with single
6670 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
6678 insn_state[t] = EXPLORED;
6679 if (env->cfg.cur_stack-- <= 0) {
6680 verbose(env, "pop stack internal bug\n");
6687 for (i = 0; i < insn_cnt; i++) {
6688 if (insn_state[i] != EXPLORED) {
6689 verbose(env, "unreachable insn %d\n", i);
6694 ret = 0; /* cfg looks good */
6699 env->cfg.insn_state = env->cfg.insn_stack = NULL;
6703 /* The minimum supported BTF func info size */
6704 #define MIN_BPF_FUNCINFO_SIZE 8
6705 #define MAX_FUNCINFO_REC_SIZE 252
6707 static int check_btf_func(struct bpf_verifier_env *env,
6708 const union bpf_attr *attr,
6709 union bpf_attr __user *uattr)
6711 u32 i, nfuncs, urec_size, min_size;
6712 u32 krec_size = sizeof(struct bpf_func_info);
6713 struct bpf_func_info *krecord;
6714 const struct btf_type *type;
6715 struct bpf_prog *prog;
6716 const struct btf *btf;
6717 void __user *urecord;
6718 u32 prev_offset = 0;
6721 nfuncs = attr->func_info_cnt;
6725 if (nfuncs != env->subprog_cnt) {
6726 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
6730 urec_size = attr->func_info_rec_size;
6731 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
6732 urec_size > MAX_FUNCINFO_REC_SIZE ||
6733 urec_size % sizeof(u32)) {
6734 verbose(env, "invalid func info rec size %u\n", urec_size);
6739 btf = prog->aux->btf;
6741 urecord = u64_to_user_ptr(attr->func_info);
6742 min_size = min_t(u32, krec_size, urec_size);
6744 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
6748 for (i = 0; i < nfuncs; i++) {
6749 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
6751 if (ret == -E2BIG) {
6752 verbose(env, "nonzero tailing record in func info");
6753 /* set the size kernel expects so loader can zero
6754 * out the rest of the record.
6756 if (put_user(min_size, &uattr->func_info_rec_size))
6762 if (copy_from_user(&krecord[i], urecord, min_size)) {
6767 /* check insn_off */
6769 if (krecord[i].insn_off) {
6771 "nonzero insn_off %u for the first func info record",
6772 krecord[i].insn_off);
6776 } else if (krecord[i].insn_off <= prev_offset) {
6778 "same or smaller insn offset (%u) than previous func info record (%u)",
6779 krecord[i].insn_off, prev_offset);
6784 if (env->subprog_info[i].start != krecord[i].insn_off) {
6785 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
6791 type = btf_type_by_id(btf, krecord[i].type_id);
6792 if (!type || BTF_INFO_KIND(type->info) != BTF_KIND_FUNC) {
6793 verbose(env, "invalid type id %d in func info",
6794 krecord[i].type_id);
6799 prev_offset = krecord[i].insn_off;
6800 urecord += urec_size;
6803 prog->aux->func_info = krecord;
6804 prog->aux->func_info_cnt = nfuncs;
6812 static void adjust_btf_func(struct bpf_verifier_env *env)
6816 if (!env->prog->aux->func_info)
6819 for (i = 0; i < env->subprog_cnt; i++)
6820 env->prog->aux->func_info[i].insn_off = env->subprog_info[i].start;
6823 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
6824 sizeof(((struct bpf_line_info *)(0))->line_col))
6825 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
6827 static int check_btf_line(struct bpf_verifier_env *env,
6828 const union bpf_attr *attr,
6829 union bpf_attr __user *uattr)
6831 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
6832 struct bpf_subprog_info *sub;
6833 struct bpf_line_info *linfo;
6834 struct bpf_prog *prog;
6835 const struct btf *btf;
6836 void __user *ulinfo;
6839 nr_linfo = attr->line_info_cnt;
6842 if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info))
6845 rec_size = attr->line_info_rec_size;
6846 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
6847 rec_size > MAX_LINEINFO_REC_SIZE ||
6848 rec_size & (sizeof(u32) - 1))
6851 /* Need to zero it in case the userspace may
6852 * pass in a smaller bpf_line_info object.
6854 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
6855 GFP_KERNEL | __GFP_NOWARN);
6860 btf = prog->aux->btf;
6863 sub = env->subprog_info;
6864 ulinfo = u64_to_user_ptr(attr->line_info);
6865 expected_size = sizeof(struct bpf_line_info);
6866 ncopy = min_t(u32, expected_size, rec_size);
6867 for (i = 0; i < nr_linfo; i++) {
6868 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
6870 if (err == -E2BIG) {
6871 verbose(env, "nonzero tailing record in line_info");
6872 if (put_user(expected_size,
6873 &uattr->line_info_rec_size))
6879 if (copy_from_user(&linfo[i], ulinfo, ncopy)) {
6885 * Check insn_off to ensure
6886 * 1) strictly increasing AND
6887 * 2) bounded by prog->len
6889 * The linfo[0].insn_off == 0 check logically falls into
6890 * the later "missing bpf_line_info for func..." case
6891 * because the first linfo[0].insn_off must be the
6892 * first sub also and the first sub must have
6893 * subprog_info[0].start == 0.
6895 if ((i && linfo[i].insn_off <= prev_offset) ||
6896 linfo[i].insn_off >= prog->len) {
6897 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
6898 i, linfo[i].insn_off, prev_offset,
6904 if (!prog->insnsi[linfo[i].insn_off].code) {
6906 "Invalid insn code at line_info[%u].insn_off\n",
6912 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
6913 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
6914 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
6919 if (s != env->subprog_cnt) {
6920 if (linfo[i].insn_off == sub[s].start) {
6921 sub[s].linfo_idx = i;
6923 } else if (sub[s].start < linfo[i].insn_off) {
6924 verbose(env, "missing bpf_line_info for func#%u\n", s);
6930 prev_offset = linfo[i].insn_off;
6934 if (s != env->subprog_cnt) {
6935 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
6936 env->subprog_cnt - s, s);
6941 prog->aux->linfo = linfo;
6942 prog->aux->nr_linfo = nr_linfo;
6951 static int check_btf_info(struct bpf_verifier_env *env,
6952 const union bpf_attr *attr,
6953 union bpf_attr __user *uattr)
6958 if (!attr->func_info_cnt && !attr->line_info_cnt)
6961 btf = btf_get_by_fd(attr->prog_btf_fd);
6963 return PTR_ERR(btf);
6964 env->prog->aux->btf = btf;
6966 err = check_btf_func(env, attr, uattr);
6970 err = check_btf_line(env, attr, uattr);
6977 /* check %cur's range satisfies %old's */
6978 static bool range_within(struct bpf_reg_state *old,
6979 struct bpf_reg_state *cur)
6981 return old->umin_value <= cur->umin_value &&
6982 old->umax_value >= cur->umax_value &&
6983 old->smin_value <= cur->smin_value &&
6984 old->smax_value >= cur->smax_value;
6987 /* If in the old state two registers had the same id, then they need to have
6988 * the same id in the new state as well. But that id could be different from
6989 * the old state, so we need to track the mapping from old to new ids.
6990 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
6991 * regs with old id 5 must also have new id 9 for the new state to be safe. But
6992 * regs with a different old id could still have new id 9, we don't care about
6994 * So we look through our idmap to see if this old id has been seen before. If
6995 * so, we require the new id to match; otherwise, we add the id pair to the map.
6997 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_id_pair *idmap)
7001 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
7002 if (!idmap[i].old) {
7003 /* Reached an empty slot; haven't seen this id before */
7004 idmap[i].old = old_id;
7005 idmap[i].cur = cur_id;
7008 if (idmap[i].old == old_id)
7009 return idmap[i].cur == cur_id;
7011 /* We ran out of idmap slots, which should be impossible */
7016 static void clean_func_state(struct bpf_verifier_env *env,
7017 struct bpf_func_state *st)
7019 enum bpf_reg_liveness live;
7022 for (i = 0; i < BPF_REG_FP; i++) {
7023 live = st->regs[i].live;
7024 /* liveness must not touch this register anymore */
7025 st->regs[i].live |= REG_LIVE_DONE;
7026 if (!(live & REG_LIVE_READ))
7027 /* since the register is unused, clear its state
7028 * to make further comparison simpler
7030 __mark_reg_not_init(env, &st->regs[i]);
7033 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
7034 live = st->stack[i].spilled_ptr.live;
7035 /* liveness must not touch this stack slot anymore */
7036 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
7037 if (!(live & REG_LIVE_READ)) {
7038 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
7039 for (j = 0; j < BPF_REG_SIZE; j++)
7040 st->stack[i].slot_type[j] = STACK_INVALID;
7045 static void clean_verifier_state(struct bpf_verifier_env *env,
7046 struct bpf_verifier_state *st)
7050 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
7051 /* all regs in this state in all frames were already marked */
7054 for (i = 0; i <= st->curframe; i++)
7055 clean_func_state(env, st->frame[i]);
7058 /* the parentage chains form a tree.
7059 * the verifier states are added to state lists at given insn and
7060 * pushed into state stack for future exploration.
7061 * when the verifier reaches bpf_exit insn some of the verifer states
7062 * stored in the state lists have their final liveness state already,
7063 * but a lot of states will get revised from liveness point of view when
7064 * the verifier explores other branches.
7067 * 2: if r1 == 100 goto pc+1
7070 * when the verifier reaches exit insn the register r0 in the state list of
7071 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
7072 * of insn 2 and goes exploring further. At the insn 4 it will walk the
7073 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
7075 * Since the verifier pushes the branch states as it sees them while exploring
7076 * the program the condition of walking the branch instruction for the second
7077 * time means that all states below this branch were already explored and
7078 * their final liveness markes are already propagated.
7079 * Hence when the verifier completes the search of state list in is_state_visited()
7080 * we can call this clean_live_states() function to mark all liveness states
7081 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
7083 * This function also clears the registers and stack for states that !READ
7084 * to simplify state merging.
7086 * Important note here that walking the same branch instruction in the callee
7087 * doesn't meant that the states are DONE. The verifier has to compare
7090 static void clean_live_states(struct bpf_verifier_env *env, int insn,
7091 struct bpf_verifier_state *cur)
7093 struct bpf_verifier_state_list *sl;
7096 sl = *explored_state(env, insn);
7098 if (sl->state.branches)
7100 if (sl->state.insn_idx != insn ||
7101 sl->state.curframe != cur->curframe)
7103 for (i = 0; i <= cur->curframe; i++)
7104 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
7106 clean_verifier_state(env, &sl->state);
7112 /* Returns true if (rold safe implies rcur safe) */
7113 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
7114 struct bpf_reg_state *rcur, struct bpf_id_pair *idmap)
7118 if (!(rold->live & REG_LIVE_READ))
7119 /* explored state didn't use this */
7122 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
7124 if (rold->type == PTR_TO_STACK)
7125 /* two stack pointers are equal only if they're pointing to
7126 * the same stack frame, since fp-8 in foo != fp-8 in bar
7128 return equal && rold->frameno == rcur->frameno;
7133 if (rold->type == NOT_INIT)
7134 /* explored state can't have used this */
7136 if (rcur->type == NOT_INIT)
7138 switch (rold->type) {
7140 if (env->explore_alu_limits)
7142 if (rcur->type == SCALAR_VALUE) {
7143 if (!rold->precise && !rcur->precise)
7145 /* new val must satisfy old val knowledge */
7146 return range_within(rold, rcur) &&
7147 tnum_in(rold->var_off, rcur->var_off);
7149 /* We're trying to use a pointer in place of a scalar.
7150 * Even if the scalar was unbounded, this could lead to
7151 * pointer leaks because scalars are allowed to leak
7152 * while pointers are not. We could make this safe in
7153 * special cases if root is calling us, but it's
7154 * probably not worth the hassle.
7158 case PTR_TO_MAP_VALUE:
7159 /* If the new min/max/var_off satisfy the old ones and
7160 * everything else matches, we are OK.
7161 * 'id' is not compared, since it's only used for maps with
7162 * bpf_spin_lock inside map element and in such cases if
7163 * the rest of the prog is valid for one map element then
7164 * it's valid for all map elements regardless of the key
7165 * used in bpf_map_lookup()
7167 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
7168 range_within(rold, rcur) &&
7169 tnum_in(rold->var_off, rcur->var_off);
7170 case PTR_TO_MAP_VALUE_OR_NULL:
7171 /* a PTR_TO_MAP_VALUE could be safe to use as a
7172 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
7173 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
7174 * checked, doing so could have affected others with the same
7175 * id, and we can't check for that because we lost the id when
7176 * we converted to a PTR_TO_MAP_VALUE.
7178 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
7180 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
7182 /* Check our ids match any regs they're supposed to */
7183 return check_ids(rold->id, rcur->id, idmap);
7184 case PTR_TO_PACKET_META:
7186 if (rcur->type != rold->type)
7188 /* We must have at least as much range as the old ptr
7189 * did, so that any accesses which were safe before are
7190 * still safe. This is true even if old range < old off,
7191 * since someone could have accessed through (ptr - k), or
7192 * even done ptr -= k in a register, to get a safe access.
7194 if (rold->range > rcur->range)
7196 /* If the offsets don't match, we can't trust our alignment;
7197 * nor can we be sure that we won't fall out of range.
7199 if (rold->off != rcur->off)
7201 /* id relations must be preserved */
7202 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
7204 /* new val must satisfy old val knowledge */
7205 return range_within(rold, rcur) &&
7206 tnum_in(rold->var_off, rcur->var_off);
7208 case CONST_PTR_TO_MAP:
7209 case PTR_TO_PACKET_END:
7210 case PTR_TO_FLOW_KEYS:
7212 case PTR_TO_SOCKET_OR_NULL:
7213 case PTR_TO_SOCK_COMMON:
7214 case PTR_TO_SOCK_COMMON_OR_NULL:
7215 case PTR_TO_TCP_SOCK:
7216 case PTR_TO_TCP_SOCK_OR_NULL:
7217 case PTR_TO_XDP_SOCK:
7218 /* Only valid matches are exact, which memcmp() above
7219 * would have accepted
7222 /* Don't know what's going on, just say it's not safe */
7226 /* Shouldn't get here; if we do, say it's not safe */
7231 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
7232 struct bpf_func_state *cur, struct bpf_id_pair *idmap)
7236 /* walk slots of the explored stack and ignore any additional
7237 * slots in the current stack, since explored(safe) state
7240 for (i = 0; i < old->allocated_stack; i++) {
7241 spi = i / BPF_REG_SIZE;
7243 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
7244 i += BPF_REG_SIZE - 1;
7245 /* explored state didn't use this */
7249 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
7252 /* explored stack has more populated slots than current stack
7253 * and these slots were used
7255 if (i >= cur->allocated_stack)
7258 /* if old state was safe with misc data in the stack
7259 * it will be safe with zero-initialized stack.
7260 * The opposite is not true
7262 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
7263 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
7265 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
7266 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
7267 /* Ex: old explored (safe) state has STACK_SPILL in
7268 * this stack slot, but current has has STACK_MISC ->
7269 * this verifier states are not equivalent,
7270 * return false to continue verification of this path
7273 if (i % BPF_REG_SIZE)
7275 if (old->stack[spi].slot_type[0] != STACK_SPILL)
7277 if (!regsafe(env, &old->stack[spi].spilled_ptr,
7278 &cur->stack[spi].spilled_ptr, idmap))
7279 /* when explored and current stack slot are both storing
7280 * spilled registers, check that stored pointers types
7281 * are the same as well.
7282 * Ex: explored safe path could have stored
7283 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
7284 * but current path has stored:
7285 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
7286 * such verifier states are not equivalent.
7287 * return false to continue verification of this path
7294 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
7296 if (old->acquired_refs != cur->acquired_refs)
7298 return !memcmp(old->refs, cur->refs,
7299 sizeof(*old->refs) * old->acquired_refs);
7302 /* compare two verifier states
7304 * all states stored in state_list are known to be valid, since
7305 * verifier reached 'bpf_exit' instruction through them
7307 * this function is called when verifier exploring different branches of
7308 * execution popped from the state stack. If it sees an old state that has
7309 * more strict register state and more strict stack state then this execution
7310 * branch doesn't need to be explored further, since verifier already
7311 * concluded that more strict state leads to valid finish.
7313 * Therefore two states are equivalent if register state is more conservative
7314 * and explored stack state is more conservative than the current one.
7317 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
7318 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
7320 * In other words if current stack state (one being explored) has more
7321 * valid slots than old one that already passed validation, it means
7322 * the verifier can stop exploring and conclude that current state is valid too
7324 * Similarly with registers. If explored state has register type as invalid
7325 * whereas register type in current state is meaningful, it means that
7326 * the current state will reach 'bpf_exit' instruction safely
7328 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
7329 struct bpf_func_state *cur)
7333 memset(env->idmap_scratch, 0, sizeof(env->idmap_scratch));
7334 for (i = 0; i < MAX_BPF_REG; i++)
7335 if (!regsafe(env, &old->regs[i], &cur->regs[i],
7336 env->idmap_scratch))
7339 if (!stacksafe(env, old, cur, env->idmap_scratch))
7342 if (!refsafe(old, cur))
7348 static bool states_equal(struct bpf_verifier_env *env,
7349 struct bpf_verifier_state *old,
7350 struct bpf_verifier_state *cur)
7354 if (old->curframe != cur->curframe)
7357 /* Verification state from speculative execution simulation
7358 * must never prune a non-speculative execution one.
7360 if (old->speculative && !cur->speculative)
7363 if (old->active_spin_lock != cur->active_spin_lock)
7366 /* for states to be equal callsites have to be the same
7367 * and all frame states need to be equivalent
7369 for (i = 0; i <= old->curframe; i++) {
7370 if (old->frame[i]->callsite != cur->frame[i]->callsite)
7372 if (!func_states_equal(env, old->frame[i], cur->frame[i]))
7378 /* Return 0 if no propagation happened. Return negative error code if error
7379 * happened. Otherwise, return the propagated bit.
7381 static int propagate_liveness_reg(struct bpf_verifier_env *env,
7382 struct bpf_reg_state *reg,
7383 struct bpf_reg_state *parent_reg)
7385 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
7386 u8 flag = reg->live & REG_LIVE_READ;
7389 /* When comes here, read flags of PARENT_REG or REG could be any of
7390 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
7391 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
7393 if (parent_flag == REG_LIVE_READ64 ||
7394 /* Or if there is no read flag from REG. */
7396 /* Or if the read flag from REG is the same as PARENT_REG. */
7397 parent_flag == flag)
7400 err = mark_reg_read(env, reg, parent_reg, flag);
7407 /* A write screens off any subsequent reads; but write marks come from the
7408 * straight-line code between a state and its parent. When we arrive at an
7409 * equivalent state (jump target or such) we didn't arrive by the straight-line
7410 * code, so read marks in the state must propagate to the parent regardless
7411 * of the state's write marks. That's what 'parent == state->parent' comparison
7412 * in mark_reg_read() is for.
7414 static int propagate_liveness(struct bpf_verifier_env *env,
7415 const struct bpf_verifier_state *vstate,
7416 struct bpf_verifier_state *vparent)
7418 struct bpf_reg_state *state_reg, *parent_reg;
7419 struct bpf_func_state *state, *parent;
7420 int i, frame, err = 0;
7422 if (vparent->curframe != vstate->curframe) {
7423 WARN(1, "propagate_live: parent frame %d current frame %d\n",
7424 vparent->curframe, vstate->curframe);
7427 /* Propagate read liveness of registers... */
7428 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
7429 for (frame = 0; frame <= vstate->curframe; frame++) {
7430 parent = vparent->frame[frame];
7431 state = vstate->frame[frame];
7432 parent_reg = parent->regs;
7433 state_reg = state->regs;
7434 /* We don't need to worry about FP liveness, it's read-only */
7435 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
7436 err = propagate_liveness_reg(env, &state_reg[i],
7440 if (err == REG_LIVE_READ64)
7441 mark_insn_zext(env, &parent_reg[i]);
7444 /* Propagate stack slots. */
7445 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
7446 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
7447 parent_reg = &parent->stack[i].spilled_ptr;
7448 state_reg = &state->stack[i].spilled_ptr;
7449 err = propagate_liveness_reg(env, state_reg,
7458 /* find precise scalars in the previous equivalent state and
7459 * propagate them into the current state
7461 static int propagate_precision(struct bpf_verifier_env *env,
7462 const struct bpf_verifier_state *old)
7464 struct bpf_reg_state *state_reg;
7465 struct bpf_func_state *state;
7468 state = old->frame[old->curframe];
7469 state_reg = state->regs;
7470 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
7471 if (state_reg->type != SCALAR_VALUE ||
7472 !state_reg->precise)
7474 if (env->log.level & BPF_LOG_LEVEL2)
7475 verbose(env, "propagating r%d\n", i);
7476 err = mark_chain_precision(env, i);
7481 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
7482 if (state->stack[i].slot_type[0] != STACK_SPILL)
7484 state_reg = &state->stack[i].spilled_ptr;
7485 if (state_reg->type != SCALAR_VALUE ||
7486 !state_reg->precise)
7488 if (env->log.level & BPF_LOG_LEVEL2)
7489 verbose(env, "propagating fp%d\n",
7490 (-i - 1) * BPF_REG_SIZE);
7491 err = mark_chain_precision_stack(env, i);
7498 static bool states_maybe_looping(struct bpf_verifier_state *old,
7499 struct bpf_verifier_state *cur)
7501 struct bpf_func_state *fold, *fcur;
7502 int i, fr = cur->curframe;
7504 if (old->curframe != fr)
7507 fold = old->frame[fr];
7508 fcur = cur->frame[fr];
7509 for (i = 0; i < MAX_BPF_REG; i++)
7510 if (memcmp(&fold->regs[i], &fcur->regs[i],
7511 offsetof(struct bpf_reg_state, parent)))
7517 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
7519 struct bpf_verifier_state_list *new_sl;
7520 struct bpf_verifier_state_list *sl, **pprev;
7521 struct bpf_verifier_state *cur = env->cur_state, *new;
7522 int i, j, err, states_cnt = 0;
7523 bool add_new_state = env->test_state_freq ? true : false;
7525 cur->last_insn_idx = env->prev_insn_idx;
7526 if (!env->insn_aux_data[insn_idx].prune_point)
7527 /* this 'insn_idx' instruction wasn't marked, so we will not
7528 * be doing state search here
7532 /* bpf progs typically have pruning point every 4 instructions
7533 * http://vger.kernel.org/bpfconf2019.html#session-1
7534 * Do not add new state for future pruning if the verifier hasn't seen
7535 * at least 2 jumps and at least 8 instructions.
7536 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
7537 * In tests that amounts to up to 50% reduction into total verifier
7538 * memory consumption and 20% verifier time speedup.
7540 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
7541 env->insn_processed - env->prev_insn_processed >= 8)
7542 add_new_state = true;
7544 pprev = explored_state(env, insn_idx);
7547 clean_live_states(env, insn_idx, cur);
7551 if (sl->state.insn_idx != insn_idx)
7553 if (sl->state.branches) {
7554 if (states_maybe_looping(&sl->state, cur) &&
7555 states_equal(env, &sl->state, cur)) {
7556 verbose_linfo(env, insn_idx, "; ");
7557 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
7560 /* if the verifier is processing a loop, avoid adding new state
7561 * too often, since different loop iterations have distinct
7562 * states and may not help future pruning.
7563 * This threshold shouldn't be too low to make sure that
7564 * a loop with large bound will be rejected quickly.
7565 * The most abusive loop will be:
7567 * if r1 < 1000000 goto pc-2
7568 * 1M insn_procssed limit / 100 == 10k peak states.
7569 * This threshold shouldn't be too high either, since states
7570 * at the end of the loop are likely to be useful in pruning.
7572 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
7573 env->insn_processed - env->prev_insn_processed < 100)
7574 add_new_state = false;
7577 if (states_equal(env, &sl->state, cur)) {
7579 /* reached equivalent register/stack state,
7581 * Registers read by the continuation are read by us.
7582 * If we have any write marks in env->cur_state, they
7583 * will prevent corresponding reads in the continuation
7584 * from reaching our parent (an explored_state). Our
7585 * own state will get the read marks recorded, but
7586 * they'll be immediately forgotten as we're pruning
7587 * this state and will pop a new one.
7589 err = propagate_liveness(env, &sl->state, cur);
7591 /* if previous state reached the exit with precision and
7592 * current state is equivalent to it (except precsion marks)
7593 * the precision needs to be propagated back in
7594 * the current state.
7596 err = err ? : push_jmp_history(env, cur);
7597 err = err ? : propagate_precision(env, &sl->state);
7603 /* when new state is not going to be added do not increase miss count.
7604 * Otherwise several loop iterations will remove the state
7605 * recorded earlier. The goal of these heuristics is to have
7606 * states from some iterations of the loop (some in the beginning
7607 * and some at the end) to help pruning.
7611 /* heuristic to determine whether this state is beneficial
7612 * to keep checking from state equivalence point of view.
7613 * Higher numbers increase max_states_per_insn and verification time,
7614 * but do not meaningfully decrease insn_processed.
7616 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
7617 /* the state is unlikely to be useful. Remove it to
7618 * speed up verification
7621 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
7622 u32 br = sl->state.branches;
7625 "BUG live_done but branches_to_explore %d\n",
7627 free_verifier_state(&sl->state, false);
7631 /* cannot free this state, since parentage chain may
7632 * walk it later. Add it for free_list instead to
7633 * be freed at the end of verification
7635 sl->next = env->free_list;
7636 env->free_list = sl;
7646 if (env->max_states_per_insn < states_cnt)
7647 env->max_states_per_insn = states_cnt;
7649 if (!env->allow_ptr_leaks && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
7650 return push_jmp_history(env, cur);
7653 return push_jmp_history(env, cur);
7655 /* There were no equivalent states, remember the current one.
7656 * Technically the current state is not proven to be safe yet,
7657 * but it will either reach outer most bpf_exit (which means it's safe)
7658 * or it will be rejected. When there are no loops the verifier won't be
7659 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
7660 * again on the way to bpf_exit.
7661 * When looping the sl->state.branches will be > 0 and this state
7662 * will not be considered for equivalence until branches == 0.
7664 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
7667 env->total_states++;
7669 env->prev_jmps_processed = env->jmps_processed;
7670 env->prev_insn_processed = env->insn_processed;
7672 /* add new state to the head of linked list */
7673 new = &new_sl->state;
7674 err = copy_verifier_state(new, cur);
7676 free_verifier_state(new, false);
7680 new->insn_idx = insn_idx;
7681 WARN_ONCE(new->branches != 1,
7682 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
7685 cur->first_insn_idx = insn_idx;
7686 clear_jmp_history(cur);
7687 new_sl->next = *explored_state(env, insn_idx);
7688 *explored_state(env, insn_idx) = new_sl;
7689 /* connect new state to parentage chain. Current frame needs all
7690 * registers connected. Only r6 - r9 of the callers are alive (pushed
7691 * to the stack implicitly by JITs) so in callers' frames connect just
7692 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
7693 * the state of the call instruction (with WRITTEN set), and r0 comes
7694 * from callee with its full parentage chain, anyway.
7696 /* clear write marks in current state: the writes we did are not writes
7697 * our child did, so they don't screen off its reads from us.
7698 * (There are no read marks in current state, because reads always mark
7699 * their parent and current state never has children yet. Only
7700 * explored_states can get read marks.)
7702 for (j = 0; j <= cur->curframe; j++) {
7703 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
7704 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
7705 for (i = 0; i < BPF_REG_FP; i++)
7706 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
7709 /* all stack frames are accessible from callee, clear them all */
7710 for (j = 0; j <= cur->curframe; j++) {
7711 struct bpf_func_state *frame = cur->frame[j];
7712 struct bpf_func_state *newframe = new->frame[j];
7714 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
7715 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
7716 frame->stack[i].spilled_ptr.parent =
7717 &newframe->stack[i].spilled_ptr;
7723 /* Return true if it's OK to have the same insn return a different type. */
7724 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
7729 case PTR_TO_SOCKET_OR_NULL:
7730 case PTR_TO_SOCK_COMMON:
7731 case PTR_TO_SOCK_COMMON_OR_NULL:
7732 case PTR_TO_TCP_SOCK:
7733 case PTR_TO_TCP_SOCK_OR_NULL:
7734 case PTR_TO_XDP_SOCK:
7741 /* If an instruction was previously used with particular pointer types, then we
7742 * need to be careful to avoid cases such as the below, where it may be ok
7743 * for one branch accessing the pointer, but not ok for the other branch:
7748 * R1 = some_other_valid_ptr;
7751 * R2 = *(u32 *)(R1 + 0);
7753 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
7755 return src != prev && (!reg_type_mismatch_ok(src) ||
7756 !reg_type_mismatch_ok(prev));
7759 static int do_check(struct bpf_verifier_env *env)
7761 struct bpf_verifier_state *state;
7762 struct bpf_insn *insns = env->prog->insnsi;
7763 struct bpf_reg_state *regs;
7764 int insn_cnt = env->prog->len;
7765 bool do_print_state = false;
7766 int prev_insn_idx = -1;
7768 env->prev_linfo = NULL;
7770 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
7773 state->curframe = 0;
7774 state->speculative = false;
7775 state->branches = 1;
7776 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
7777 if (!state->frame[0]) {
7781 env->cur_state = state;
7782 init_func_state(env, state->frame[0],
7783 BPF_MAIN_FUNC /* callsite */,
7785 0 /* subprogno, zero == main subprog */);
7788 struct bpf_insn *insn;
7792 env->prev_insn_idx = prev_insn_idx;
7793 if (env->insn_idx >= insn_cnt) {
7794 verbose(env, "invalid insn idx %d insn_cnt %d\n",
7795 env->insn_idx, insn_cnt);
7799 insn = &insns[env->insn_idx];
7800 class = BPF_CLASS(insn->code);
7802 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
7804 "BPF program is too large. Processed %d insn\n",
7805 env->insn_processed);
7809 err = is_state_visited(env, env->insn_idx);
7813 /* found equivalent state, can prune the search */
7814 if (env->log.level & BPF_LOG_LEVEL) {
7816 verbose(env, "\nfrom %d to %d%s: safe\n",
7817 env->prev_insn_idx, env->insn_idx,
7818 env->cur_state->speculative ?
7819 " (speculative execution)" : "");
7821 verbose(env, "%d: safe\n", env->insn_idx);
7823 goto process_bpf_exit;
7826 if (signal_pending(current))
7832 if (env->log.level & BPF_LOG_LEVEL2 ||
7833 (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
7834 if (env->log.level & BPF_LOG_LEVEL2)
7835 verbose(env, "%d:", env->insn_idx);
7837 verbose(env, "\nfrom %d to %d%s:",
7838 env->prev_insn_idx, env->insn_idx,
7839 env->cur_state->speculative ?
7840 " (speculative execution)" : "");
7841 print_verifier_state(env, state->frame[state->curframe]);
7842 do_print_state = false;
7845 if (env->log.level & BPF_LOG_LEVEL) {
7846 const struct bpf_insn_cbs cbs = {
7847 .cb_print = verbose,
7848 .private_data = env,
7851 verbose_linfo(env, env->insn_idx, "; ");
7852 verbose(env, "%d: ", env->insn_idx);
7853 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
7856 if (bpf_prog_is_dev_bound(env->prog->aux)) {
7857 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
7858 env->prev_insn_idx);
7863 regs = cur_regs(env);
7864 sanitize_mark_insn_seen(env);
7865 prev_insn_idx = env->insn_idx;
7867 if (class == BPF_ALU || class == BPF_ALU64) {
7868 err = check_alu_op(env, insn);
7872 } else if (class == BPF_LDX) {
7873 enum bpf_reg_type *prev_src_type, src_reg_type;
7875 /* check for reserved fields is already done */
7877 /* check src operand */
7878 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7882 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
7886 src_reg_type = regs[insn->src_reg].type;
7888 /* check that memory (src_reg + off) is readable,
7889 * the state of dst_reg will be updated by this func
7891 err = check_mem_access(env, env->insn_idx, insn->src_reg,
7892 insn->off, BPF_SIZE(insn->code),
7893 BPF_READ, insn->dst_reg, false);
7897 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
7899 if (*prev_src_type == NOT_INIT) {
7901 * dst_reg = *(u32 *)(src_reg + off)
7902 * save type to validate intersecting paths
7904 *prev_src_type = src_reg_type;
7906 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
7907 /* ABuser program is trying to use the same insn
7908 * dst_reg = *(u32*) (src_reg + off)
7909 * with different pointer types:
7910 * src_reg == ctx in one branch and
7911 * src_reg == stack|map in some other branch.
7914 verbose(env, "same insn cannot be used with different pointers\n");
7918 } else if (class == BPF_STX) {
7919 enum bpf_reg_type *prev_dst_type, dst_reg_type;
7921 if (BPF_MODE(insn->code) == BPF_XADD) {
7922 err = check_xadd(env, env->insn_idx, insn);
7929 /* check src1 operand */
7930 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7933 /* check src2 operand */
7934 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7938 dst_reg_type = regs[insn->dst_reg].type;
7940 /* check that memory (dst_reg + off) is writeable */
7941 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
7942 insn->off, BPF_SIZE(insn->code),
7943 BPF_WRITE, insn->src_reg, false);
7947 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
7949 if (*prev_dst_type == NOT_INIT) {
7950 *prev_dst_type = dst_reg_type;
7951 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
7952 verbose(env, "same insn cannot be used with different pointers\n");
7956 } else if (class == BPF_ST) {
7957 if (BPF_MODE(insn->code) != BPF_MEM ||
7958 insn->src_reg != BPF_REG_0) {
7959 verbose(env, "BPF_ST uses reserved fields\n");
7962 /* check src operand */
7963 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7967 if (is_ctx_reg(env, insn->dst_reg)) {
7968 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
7970 reg_type_str[reg_state(env, insn->dst_reg)->type]);
7974 /* check that memory (dst_reg + off) is writeable */
7975 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
7976 insn->off, BPF_SIZE(insn->code),
7977 BPF_WRITE, -1, false);
7981 } else if (class == BPF_JMP || class == BPF_JMP32) {
7982 u8 opcode = BPF_OP(insn->code);
7984 env->jmps_processed++;
7985 if (opcode == BPF_CALL) {
7986 if (BPF_SRC(insn->code) != BPF_K ||
7988 (insn->src_reg != BPF_REG_0 &&
7989 insn->src_reg != BPF_PSEUDO_CALL) ||
7990 insn->dst_reg != BPF_REG_0 ||
7991 class == BPF_JMP32) {
7992 verbose(env, "BPF_CALL uses reserved fields\n");
7996 if (env->cur_state->active_spin_lock &&
7997 (insn->src_reg == BPF_PSEUDO_CALL ||
7998 insn->imm != BPF_FUNC_spin_unlock)) {
7999 verbose(env, "function calls are not allowed while holding a lock\n");
8002 if (insn->src_reg == BPF_PSEUDO_CALL)
8003 err = check_func_call(env, insn, &env->insn_idx);
8005 err = check_helper_call(env, insn->imm, env->insn_idx);
8009 } else if (opcode == BPF_JA) {
8010 if (BPF_SRC(insn->code) != BPF_K ||
8012 insn->src_reg != BPF_REG_0 ||
8013 insn->dst_reg != BPF_REG_0 ||
8014 class == BPF_JMP32) {
8015 verbose(env, "BPF_JA uses reserved fields\n");
8019 env->insn_idx += insn->off + 1;
8022 } else if (opcode == BPF_EXIT) {
8023 if (BPF_SRC(insn->code) != BPF_K ||
8025 insn->src_reg != BPF_REG_0 ||
8026 insn->dst_reg != BPF_REG_0 ||
8027 class == BPF_JMP32) {
8028 verbose(env, "BPF_EXIT uses reserved fields\n");
8032 if (env->cur_state->active_spin_lock) {
8033 verbose(env, "bpf_spin_unlock is missing\n");
8037 if (state->curframe) {
8038 /* exit from nested function */
8039 err = prepare_func_exit(env, &env->insn_idx);
8042 do_print_state = true;
8046 err = check_reference_leak(env);
8050 /* eBPF calling convetion is such that R0 is used
8051 * to return the value from eBPF program.
8052 * Make sure that it's readable at this time
8053 * of bpf_exit, which means that program wrote
8054 * something into it earlier
8056 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
8060 if (is_pointer_value(env, BPF_REG_0)) {
8061 verbose(env, "R0 leaks addr as return value\n");
8065 err = check_return_code(env);
8069 update_branch_counts(env, env->cur_state);
8070 err = pop_stack(env, &prev_insn_idx,
8077 do_print_state = true;
8081 err = check_cond_jmp_op(env, insn, &env->insn_idx);
8085 } else if (class == BPF_LD) {
8086 u8 mode = BPF_MODE(insn->code);
8088 if (mode == BPF_ABS || mode == BPF_IND) {
8089 err = check_ld_abs(env, insn);
8093 } else if (mode == BPF_IMM) {
8094 err = check_ld_imm(env, insn);
8099 sanitize_mark_insn_seen(env);
8101 verbose(env, "invalid BPF_LD mode\n");
8105 verbose(env, "unknown insn class %d\n", class);
8112 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
8116 static int check_map_prealloc(struct bpf_map *map)
8118 return (map->map_type != BPF_MAP_TYPE_HASH &&
8119 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
8120 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
8121 !(map->map_flags & BPF_F_NO_PREALLOC);
8124 static bool is_tracing_prog_type(enum bpf_prog_type type)
8127 case BPF_PROG_TYPE_KPROBE:
8128 case BPF_PROG_TYPE_TRACEPOINT:
8129 case BPF_PROG_TYPE_PERF_EVENT:
8130 case BPF_PROG_TYPE_RAW_TRACEPOINT:
8137 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
8138 struct bpf_map *map,
8139 struct bpf_prog *prog)
8142 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
8143 * preallocated hash maps, since doing memory allocation
8144 * in overflow_handler can crash depending on where nmi got
8147 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
8148 if (!check_map_prealloc(map)) {
8149 verbose(env, "perf_event programs can only use preallocated hash map\n");
8152 if (map->inner_map_meta &&
8153 !check_map_prealloc(map->inner_map_meta)) {
8154 verbose(env, "perf_event programs can only use preallocated inner hash map\n");
8159 if ((is_tracing_prog_type(prog->type) ||
8160 prog->type == BPF_PROG_TYPE_SOCKET_FILTER) &&
8161 map_value_has_spin_lock(map)) {
8162 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
8166 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
8167 !bpf_offload_prog_map_match(prog, map)) {
8168 verbose(env, "offload device mismatch between prog and map\n");
8175 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
8177 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
8178 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
8181 /* look for pseudo eBPF instructions that access map FDs and
8182 * replace them with actual map pointers
8184 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
8186 struct bpf_insn *insn = env->prog->insnsi;
8187 int insn_cnt = env->prog->len;
8190 err = bpf_prog_calc_tag(env->prog);
8194 for (i = 0; i < insn_cnt; i++, insn++) {
8195 if (BPF_CLASS(insn->code) == BPF_LDX &&
8196 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
8197 verbose(env, "BPF_LDX uses reserved fields\n");
8201 if (BPF_CLASS(insn->code) == BPF_STX &&
8202 ((BPF_MODE(insn->code) != BPF_MEM &&
8203 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
8204 verbose(env, "BPF_STX uses reserved fields\n");
8208 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
8209 struct bpf_insn_aux_data *aux;
8210 struct bpf_map *map;
8214 if (i == insn_cnt - 1 || insn[1].code != 0 ||
8215 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
8217 verbose(env, "invalid bpf_ld_imm64 insn\n");
8221 if (insn[0].src_reg == 0)
8222 /* valid generic load 64-bit imm */
8225 /* In final convert_pseudo_ld_imm64() step, this is
8226 * converted into regular 64-bit imm load insn.
8228 if ((insn[0].src_reg != BPF_PSEUDO_MAP_FD &&
8229 insn[0].src_reg != BPF_PSEUDO_MAP_VALUE) ||
8230 (insn[0].src_reg == BPF_PSEUDO_MAP_FD &&
8231 insn[1].imm != 0)) {
8233 "unrecognized bpf_ld_imm64 insn\n");
8237 f = fdget(insn[0].imm);
8238 map = __bpf_map_get(f);
8240 verbose(env, "fd %d is not pointing to valid bpf_map\n",
8242 return PTR_ERR(map);
8245 err = check_map_prog_compatibility(env, map, env->prog);
8251 aux = &env->insn_aux_data[i];
8252 if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
8253 addr = (unsigned long)map;
8255 u32 off = insn[1].imm;
8257 if (off >= BPF_MAX_VAR_OFF) {
8258 verbose(env, "direct value offset of %u is not allowed\n", off);
8263 if (!map->ops->map_direct_value_addr) {
8264 verbose(env, "no direct value access support for this map type\n");
8269 err = map->ops->map_direct_value_addr(map, &addr, off);
8271 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
8272 map->value_size, off);
8281 insn[0].imm = (u32)addr;
8282 insn[1].imm = addr >> 32;
8284 /* check whether we recorded this map already */
8285 for (j = 0; j < env->used_map_cnt; j++) {
8286 if (env->used_maps[j] == map) {
8293 if (env->used_map_cnt >= MAX_USED_MAPS) {
8298 /* hold the map. If the program is rejected by verifier,
8299 * the map will be released by release_maps() or it
8300 * will be used by the valid program until it's unloaded
8301 * and all maps are released in free_used_maps()
8303 map = bpf_map_inc(map, false);
8306 return PTR_ERR(map);
8309 aux->map_index = env->used_map_cnt;
8310 env->used_maps[env->used_map_cnt++] = map;
8312 if (bpf_map_is_cgroup_storage(map) &&
8313 bpf_cgroup_storage_assign(env->prog, map)) {
8314 verbose(env, "only one cgroup storage of each type is allowed\n");
8326 /* Basic sanity check before we invest more work here. */
8327 if (!bpf_opcode_in_insntable(insn->code)) {
8328 verbose(env, "unknown opcode %02x\n", insn->code);
8333 /* now all pseudo BPF_LD_IMM64 instructions load valid
8334 * 'struct bpf_map *' into a register instead of user map_fd.
8335 * These pointers will be used later by verifier to validate map access.
8340 /* drop refcnt of maps used by the rejected program */
8341 static void release_maps(struct bpf_verifier_env *env)
8343 enum bpf_cgroup_storage_type stype;
8346 for_each_cgroup_storage_type(stype) {
8347 if (!env->prog->aux->cgroup_storage[stype])
8349 bpf_cgroup_storage_release(env->prog,
8350 env->prog->aux->cgroup_storage[stype]);
8353 for (i = 0; i < env->used_map_cnt; i++)
8354 bpf_map_put(env->used_maps[i]);
8357 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
8358 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
8360 struct bpf_insn *insn = env->prog->insnsi;
8361 int insn_cnt = env->prog->len;
8364 for (i = 0; i < insn_cnt; i++, insn++)
8365 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
8369 /* single env->prog->insni[off] instruction was replaced with the range
8370 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
8371 * [0, off) and [off, end) to new locations, so the patched range stays zero
8373 static void adjust_insn_aux_data(struct bpf_verifier_env *env,
8374 struct bpf_insn_aux_data *new_data,
8375 struct bpf_prog *new_prog, u32 off, u32 cnt)
8377 struct bpf_insn_aux_data *old_data = env->insn_aux_data;
8378 struct bpf_insn *insn = new_prog->insnsi;
8379 bool old_seen = old_data[off].seen;
8383 /* aux info at OFF always needs adjustment, no matter fast path
8384 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
8385 * original insn at old prog.
8387 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
8391 prog_len = new_prog->len;
8393 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
8394 memcpy(new_data + off + cnt - 1, old_data + off,
8395 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
8396 for (i = off; i < off + cnt - 1; i++) {
8397 /* Expand insni[off]'s seen count to the patched range. */
8398 new_data[i].seen = old_seen;
8399 new_data[i].zext_dst = insn_has_def32(env, insn + i);
8401 env->insn_aux_data = new_data;
8405 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
8411 /* NOTE: fake 'exit' subprog should be updated as well. */
8412 for (i = 0; i <= env->subprog_cnt; i++) {
8413 if (env->subprog_info[i].start <= off)
8415 env->subprog_info[i].start += len - 1;
8419 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
8420 const struct bpf_insn *patch, u32 len)
8422 struct bpf_prog *new_prog;
8423 struct bpf_insn_aux_data *new_data = NULL;
8426 new_data = vzalloc(array_size(env->prog->len + len - 1,
8427 sizeof(struct bpf_insn_aux_data)));
8432 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
8433 if (IS_ERR(new_prog)) {
8434 if (PTR_ERR(new_prog) == -ERANGE)
8436 "insn %d cannot be patched due to 16-bit range\n",
8437 env->insn_aux_data[off].orig_idx);
8441 adjust_insn_aux_data(env, new_data, new_prog, off, len);
8442 adjust_subprog_starts(env, off, len);
8446 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
8451 /* find first prog starting at or after off (first to remove) */
8452 for (i = 0; i < env->subprog_cnt; i++)
8453 if (env->subprog_info[i].start >= off)
8455 /* find first prog starting at or after off + cnt (first to stay) */
8456 for (j = i; j < env->subprog_cnt; j++)
8457 if (env->subprog_info[j].start >= off + cnt)
8459 /* if j doesn't start exactly at off + cnt, we are just removing
8460 * the front of previous prog
8462 if (env->subprog_info[j].start != off + cnt)
8466 struct bpf_prog_aux *aux = env->prog->aux;
8469 /* move fake 'exit' subprog as well */
8470 move = env->subprog_cnt + 1 - j;
8472 memmove(env->subprog_info + i,
8473 env->subprog_info + j,
8474 sizeof(*env->subprog_info) * move);
8475 env->subprog_cnt -= j - i;
8477 /* remove func_info */
8478 if (aux->func_info) {
8479 move = aux->func_info_cnt - j;
8481 memmove(aux->func_info + i,
8483 sizeof(*aux->func_info) * move);
8484 aux->func_info_cnt -= j - i;
8485 /* func_info->insn_off is set after all code rewrites,
8486 * in adjust_btf_func() - no need to adjust
8490 /* convert i from "first prog to remove" to "first to adjust" */
8491 if (env->subprog_info[i].start == off)
8495 /* update fake 'exit' subprog as well */
8496 for (; i <= env->subprog_cnt; i++)
8497 env->subprog_info[i].start -= cnt;
8502 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
8505 struct bpf_prog *prog = env->prog;
8506 u32 i, l_off, l_cnt, nr_linfo;
8507 struct bpf_line_info *linfo;
8509 nr_linfo = prog->aux->nr_linfo;
8513 linfo = prog->aux->linfo;
8515 /* find first line info to remove, count lines to be removed */
8516 for (i = 0; i < nr_linfo; i++)
8517 if (linfo[i].insn_off >= off)
8522 for (; i < nr_linfo; i++)
8523 if (linfo[i].insn_off < off + cnt)
8528 /* First live insn doesn't match first live linfo, it needs to "inherit"
8529 * last removed linfo. prog is already modified, so prog->len == off
8530 * means no live instructions after (tail of the program was removed).
8532 if (prog->len != off && l_cnt &&
8533 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
8535 linfo[--i].insn_off = off + cnt;
8538 /* remove the line info which refer to the removed instructions */
8540 memmove(linfo + l_off, linfo + i,
8541 sizeof(*linfo) * (nr_linfo - i));
8543 prog->aux->nr_linfo -= l_cnt;
8544 nr_linfo = prog->aux->nr_linfo;
8547 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
8548 for (i = l_off; i < nr_linfo; i++)
8549 linfo[i].insn_off -= cnt;
8551 /* fix up all subprogs (incl. 'exit') which start >= off */
8552 for (i = 0; i <= env->subprog_cnt; i++)
8553 if (env->subprog_info[i].linfo_idx > l_off) {
8554 /* program may have started in the removed region but
8555 * may not be fully removed
8557 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
8558 env->subprog_info[i].linfo_idx -= l_cnt;
8560 env->subprog_info[i].linfo_idx = l_off;
8566 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
8568 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
8569 unsigned int orig_prog_len = env->prog->len;
8572 if (bpf_prog_is_dev_bound(env->prog->aux))
8573 bpf_prog_offload_remove_insns(env, off, cnt);
8575 err = bpf_remove_insns(env->prog, off, cnt);
8579 err = adjust_subprog_starts_after_remove(env, off, cnt);
8583 err = bpf_adj_linfo_after_remove(env, off, cnt);
8587 memmove(aux_data + off, aux_data + off + cnt,
8588 sizeof(*aux_data) * (orig_prog_len - off - cnt));
8593 /* The verifier does more data flow analysis than llvm and will not
8594 * explore branches that are dead at run time. Malicious programs can
8595 * have dead code too. Therefore replace all dead at-run-time code
8598 * Just nops are not optimal, e.g. if they would sit at the end of the
8599 * program and through another bug we would manage to jump there, then
8600 * we'd execute beyond program memory otherwise. Returning exception
8601 * code also wouldn't work since we can have subprogs where the dead
8602 * code could be located.
8604 static void sanitize_dead_code(struct bpf_verifier_env *env)
8606 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
8607 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
8608 struct bpf_insn *insn = env->prog->insnsi;
8609 const int insn_cnt = env->prog->len;
8612 for (i = 0; i < insn_cnt; i++) {
8613 if (aux_data[i].seen)
8615 memcpy(insn + i, &trap, sizeof(trap));
8616 aux_data[i].zext_dst = false;
8620 static bool insn_is_cond_jump(u8 code)
8624 if (BPF_CLASS(code) == BPF_JMP32)
8627 if (BPF_CLASS(code) != BPF_JMP)
8631 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
8634 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
8636 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
8637 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
8638 struct bpf_insn *insn = env->prog->insnsi;
8639 const int insn_cnt = env->prog->len;
8642 for (i = 0; i < insn_cnt; i++, insn++) {
8643 if (!insn_is_cond_jump(insn->code))
8646 if (!aux_data[i + 1].seen)
8648 else if (!aux_data[i + 1 + insn->off].seen)
8653 if (bpf_prog_is_dev_bound(env->prog->aux))
8654 bpf_prog_offload_replace_insn(env, i, &ja);
8656 memcpy(insn, &ja, sizeof(ja));
8660 static int opt_remove_dead_code(struct bpf_verifier_env *env)
8662 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
8663 int insn_cnt = env->prog->len;
8666 for (i = 0; i < insn_cnt; i++) {
8670 while (i + j < insn_cnt && !aux_data[i + j].seen)
8675 err = verifier_remove_insns(env, i, j);
8678 insn_cnt = env->prog->len;
8684 static int opt_remove_nops(struct bpf_verifier_env *env)
8686 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
8687 struct bpf_insn *insn = env->prog->insnsi;
8688 int insn_cnt = env->prog->len;
8691 for (i = 0; i < insn_cnt; i++) {
8692 if (memcmp(&insn[i], &ja, sizeof(ja)))
8695 err = verifier_remove_insns(env, i, 1);
8705 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
8706 const union bpf_attr *attr)
8708 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
8709 struct bpf_insn_aux_data *aux = env->insn_aux_data;
8710 int i, patch_len, delta = 0, len = env->prog->len;
8711 struct bpf_insn *insns = env->prog->insnsi;
8712 struct bpf_prog *new_prog;
8715 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
8716 zext_patch[1] = BPF_ZEXT_REG(0);
8717 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
8718 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
8719 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
8720 for (i = 0; i < len; i++) {
8721 int adj_idx = i + delta;
8722 struct bpf_insn insn;
8724 insn = insns[adj_idx];
8725 if (!aux[adj_idx].zext_dst) {
8733 class = BPF_CLASS(code);
8734 if (insn_no_def(&insn))
8737 /* NOTE: arg "reg" (the fourth one) is only used for
8738 * BPF_STX which has been ruled out in above
8739 * check, it is safe to pass NULL here.
8741 if (is_reg64(env, &insn, insn.dst_reg, NULL, DST_OP)) {
8742 if (class == BPF_LD &&
8743 BPF_MODE(code) == BPF_IMM)
8748 /* ctx load could be transformed into wider load. */
8749 if (class == BPF_LDX &&
8750 aux[adj_idx].ptr_type == PTR_TO_CTX)
8753 imm_rnd = get_random_int();
8754 rnd_hi32_patch[0] = insn;
8755 rnd_hi32_patch[1].imm = imm_rnd;
8756 rnd_hi32_patch[3].dst_reg = insn.dst_reg;
8757 patch = rnd_hi32_patch;
8759 goto apply_patch_buffer;
8762 if (!bpf_jit_needs_zext())
8765 zext_patch[0] = insn;
8766 zext_patch[1].dst_reg = insn.dst_reg;
8767 zext_patch[1].src_reg = insn.dst_reg;
8771 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
8774 env->prog = new_prog;
8775 insns = new_prog->insnsi;
8776 aux = env->insn_aux_data;
8777 delta += patch_len - 1;
8783 /* convert load instructions that access fields of a context type into a
8784 * sequence of instructions that access fields of the underlying structure:
8785 * struct __sk_buff -> struct sk_buff
8786 * struct bpf_sock_ops -> struct sock
8788 static int convert_ctx_accesses(struct bpf_verifier_env *env)
8790 const struct bpf_verifier_ops *ops = env->ops;
8791 int i, cnt, size, ctx_field_size, delta = 0;
8792 const int insn_cnt = env->prog->len;
8793 struct bpf_insn insn_buf[16], *insn;
8794 u32 target_size, size_default, off;
8795 struct bpf_prog *new_prog;
8796 enum bpf_access_type type;
8797 bool is_narrower_load;
8799 if (ops->gen_prologue || env->seen_direct_write) {
8800 if (!ops->gen_prologue) {
8801 verbose(env, "bpf verifier is misconfigured\n");
8804 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
8806 if (cnt >= ARRAY_SIZE(insn_buf)) {
8807 verbose(env, "bpf verifier is misconfigured\n");
8810 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
8814 env->prog = new_prog;
8819 if (bpf_prog_is_dev_bound(env->prog->aux))
8822 insn = env->prog->insnsi + delta;
8824 for (i = 0; i < insn_cnt; i++, insn++) {
8825 bpf_convert_ctx_access_t convert_ctx_access;
8828 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
8829 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
8830 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
8831 insn->code == (BPF_LDX | BPF_MEM | BPF_DW)) {
8834 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
8835 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
8836 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
8837 insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
8838 insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
8839 insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
8840 insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
8841 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
8843 ctx_access = BPF_CLASS(insn->code) == BPF_STX;
8848 if (type == BPF_WRITE &&
8849 env->insn_aux_data[i + delta].sanitize_stack_spill) {
8850 struct bpf_insn patch[] = {
8855 cnt = ARRAY_SIZE(patch);
8856 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
8861 env->prog = new_prog;
8862 insn = new_prog->insnsi + i + delta;
8869 switch (env->insn_aux_data[i + delta].ptr_type) {
8871 if (!ops->convert_ctx_access)
8873 convert_ctx_access = ops->convert_ctx_access;
8876 case PTR_TO_SOCK_COMMON:
8877 convert_ctx_access = bpf_sock_convert_ctx_access;
8879 case PTR_TO_TCP_SOCK:
8880 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
8882 case PTR_TO_XDP_SOCK:
8883 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
8889 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
8890 size = BPF_LDST_BYTES(insn);
8892 /* If the read access is a narrower load of the field,
8893 * convert to a 4/8-byte load, to minimum program type specific
8894 * convert_ctx_access changes. If conversion is successful,
8895 * we will apply proper mask to the result.
8897 is_narrower_load = size < ctx_field_size;
8898 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
8900 if (is_narrower_load) {
8903 if (type == BPF_WRITE) {
8904 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
8909 if (ctx_field_size == 4)
8911 else if (ctx_field_size == 8)
8914 insn->off = off & ~(size_default - 1);
8915 insn->code = BPF_LDX | BPF_MEM | size_code;
8919 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
8921 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
8922 (ctx_field_size && !target_size)) {
8923 verbose(env, "bpf verifier is misconfigured\n");
8927 if (is_narrower_load && size < target_size) {
8928 u8 shift = bpf_ctx_narrow_access_offset(
8929 off, size, size_default) * 8;
8930 if (shift && cnt + 1 >= ARRAY_SIZE(insn_buf)) {
8931 verbose(env, "bpf verifier narrow ctx load misconfigured\n");
8934 if (ctx_field_size <= 4) {
8936 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
8939 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
8940 (1 << size * 8) - 1);
8943 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
8946 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
8947 (1ULL << size * 8) - 1);
8951 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
8957 /* keep walking new program and skip insns we just inserted */
8958 env->prog = new_prog;
8959 insn = new_prog->insnsi + i + delta;
8965 static int jit_subprogs(struct bpf_verifier_env *env)
8967 struct bpf_prog *prog = env->prog, **func, *tmp;
8968 int i, j, subprog_start, subprog_end = 0, len, subprog;
8969 struct bpf_insn *insn;
8973 if (env->subprog_cnt <= 1)
8976 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
8977 if (insn->code != (BPF_JMP | BPF_CALL) ||
8978 insn->src_reg != BPF_PSEUDO_CALL)
8980 /* Upon error here we cannot fall back to interpreter but
8981 * need a hard reject of the program. Thus -EFAULT is
8982 * propagated in any case.
8984 subprog = find_subprog(env, i + insn->imm + 1);
8986 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
8990 /* temporarily remember subprog id inside insn instead of
8991 * aux_data, since next loop will split up all insns into funcs
8993 insn->off = subprog;
8994 /* remember original imm in case JIT fails and fallback
8995 * to interpreter will be needed
8997 env->insn_aux_data[i].call_imm = insn->imm;
8998 /* point imm to __bpf_call_base+1 from JITs point of view */
9002 err = bpf_prog_alloc_jited_linfo(prog);
9007 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
9011 for (i = 0; i < env->subprog_cnt; i++) {
9012 subprog_start = subprog_end;
9013 subprog_end = env->subprog_info[i + 1].start;
9015 len = subprog_end - subprog_start;
9016 /* BPF_PROG_RUN doesn't call subprogs directly,
9017 * hence main prog stats include the runtime of subprogs.
9018 * subprogs don't have IDs and not reachable via prog_get_next_id
9019 * func[i]->aux->stats will never be accessed and stays NULL
9021 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
9024 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
9025 len * sizeof(struct bpf_insn));
9026 func[i]->type = prog->type;
9028 if (bpf_prog_calc_tag(func[i]))
9030 func[i]->is_func = 1;
9031 func[i]->aux->func_idx = i;
9032 /* the btf and func_info will be freed only at prog->aux */
9033 func[i]->aux->btf = prog->aux->btf;
9034 func[i]->aux->func_info = prog->aux->func_info;
9036 /* Use bpf_prog_F_tag to indicate functions in stack traces.
9037 * Long term would need debug info to populate names
9039 func[i]->aux->name[0] = 'F';
9040 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
9041 func[i]->jit_requested = 1;
9042 func[i]->aux->linfo = prog->aux->linfo;
9043 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
9044 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
9045 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
9046 func[i] = bpf_int_jit_compile(func[i]);
9047 if (!func[i]->jited) {
9053 /* at this point all bpf functions were successfully JITed
9054 * now populate all bpf_calls with correct addresses and
9055 * run last pass of JIT
9057 for (i = 0; i < env->subprog_cnt; i++) {
9058 insn = func[i]->insnsi;
9059 for (j = 0; j < func[i]->len; j++, insn++) {
9060 if (insn->code != (BPF_JMP | BPF_CALL) ||
9061 insn->src_reg != BPF_PSEUDO_CALL)
9063 subprog = insn->off;
9064 insn->imm = BPF_CAST_CALL(func[subprog]->bpf_func) -
9068 /* we use the aux data to keep a list of the start addresses
9069 * of the JITed images for each function in the program
9071 * for some architectures, such as powerpc64, the imm field
9072 * might not be large enough to hold the offset of the start
9073 * address of the callee's JITed image from __bpf_call_base
9075 * in such cases, we can lookup the start address of a callee
9076 * by using its subprog id, available from the off field of
9077 * the call instruction, as an index for this list
9079 func[i]->aux->func = func;
9080 func[i]->aux->func_cnt = env->subprog_cnt;
9082 for (i = 0; i < env->subprog_cnt; i++) {
9083 old_bpf_func = func[i]->bpf_func;
9084 tmp = bpf_int_jit_compile(func[i]);
9085 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
9086 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
9093 /* finally lock prog and jit images for all functions and
9096 for (i = 0; i < env->subprog_cnt; i++) {
9097 bpf_prog_lock_ro(func[i]);
9098 bpf_prog_kallsyms_add(func[i]);
9101 /* Last step: make now unused interpreter insns from main
9102 * prog consistent for later dump requests, so they can
9103 * later look the same as if they were interpreted only.
9105 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
9106 if (insn->code != (BPF_JMP | BPF_CALL) ||
9107 insn->src_reg != BPF_PSEUDO_CALL)
9109 insn->off = env->insn_aux_data[i].call_imm;
9110 subprog = find_subprog(env, i + insn->off + 1);
9111 insn->imm = subprog;
9115 prog->bpf_func = func[0]->bpf_func;
9116 prog->aux->func = func;
9117 prog->aux->func_cnt = env->subprog_cnt;
9118 bpf_prog_free_unused_jited_linfo(prog);
9121 for (i = 0; i < env->subprog_cnt; i++)
9123 bpf_jit_free(func[i]);
9126 /* cleanup main prog to be interpreted */
9127 prog->jit_requested = 0;
9128 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
9129 if (insn->code != (BPF_JMP | BPF_CALL) ||
9130 insn->src_reg != BPF_PSEUDO_CALL)
9133 insn->imm = env->insn_aux_data[i].call_imm;
9135 bpf_prog_free_jited_linfo(prog);
9139 static int fixup_call_args(struct bpf_verifier_env *env)
9141 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
9142 struct bpf_prog *prog = env->prog;
9143 struct bpf_insn *insn = prog->insnsi;
9148 if (env->prog->jit_requested &&
9149 !bpf_prog_is_dev_bound(env->prog->aux)) {
9150 err = jit_subprogs(env);
9156 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
9157 for (i = 0; i < prog->len; i++, insn++) {
9158 if (insn->code != (BPF_JMP | BPF_CALL) ||
9159 insn->src_reg != BPF_PSEUDO_CALL)
9161 depth = get_callee_stack_depth(env, insn, i);
9164 bpf_patch_call_args(insn, depth);
9171 /* fixup insn->imm field of bpf_call instructions
9172 * and inline eligible helpers as explicit sequence of BPF instructions
9174 * this function is called after eBPF program passed verification
9176 static int fixup_bpf_calls(struct bpf_verifier_env *env)
9178 struct bpf_prog *prog = env->prog;
9179 struct bpf_insn *insn = prog->insnsi;
9180 const struct bpf_func_proto *fn;
9181 const int insn_cnt = prog->len;
9182 const struct bpf_map_ops *ops;
9183 struct bpf_insn_aux_data *aux;
9184 struct bpf_insn insn_buf[16];
9185 struct bpf_prog *new_prog;
9186 struct bpf_map *map_ptr;
9187 int i, cnt, delta = 0;
9189 for (i = 0; i < insn_cnt; i++, insn++) {
9190 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
9191 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
9192 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
9193 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
9194 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
9195 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
9196 struct bpf_insn *patchlet;
9197 struct bpf_insn chk_and_div[] = {
9198 /* [R,W]x div 0 -> 0 */
9199 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
9200 BPF_JNE | BPF_K, insn->src_reg,
9202 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
9203 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
9206 struct bpf_insn chk_and_mod[] = {
9207 /* [R,W]x mod 0 -> [R,W]x */
9208 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
9209 BPF_JEQ | BPF_K, insn->src_reg,
9210 0, 1 + (is64 ? 0 : 1), 0),
9212 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
9213 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
9216 patchlet = isdiv ? chk_and_div : chk_and_mod;
9217 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
9218 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
9220 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
9225 env->prog = prog = new_prog;
9226 insn = new_prog->insnsi + i + delta;
9230 if (BPF_CLASS(insn->code) == BPF_LD &&
9231 (BPF_MODE(insn->code) == BPF_ABS ||
9232 BPF_MODE(insn->code) == BPF_IND)) {
9233 cnt = env->ops->gen_ld_abs(insn, insn_buf);
9234 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
9235 verbose(env, "bpf verifier is misconfigured\n");
9239 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9244 env->prog = prog = new_prog;
9245 insn = new_prog->insnsi + i + delta;
9249 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
9250 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
9251 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
9252 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
9253 struct bpf_insn insn_buf[16];
9254 struct bpf_insn *patch = &insn_buf[0];
9255 bool issrc, isneg, isimm;
9258 aux = &env->insn_aux_data[i + delta];
9259 if (!aux->alu_state ||
9260 aux->alu_state == BPF_ALU_NON_POINTER)
9263 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
9264 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
9265 BPF_ALU_SANITIZE_SRC;
9266 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
9268 off_reg = issrc ? insn->src_reg : insn->dst_reg;
9270 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
9273 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
9274 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
9275 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
9276 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
9277 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
9278 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
9279 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
9282 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
9283 insn->src_reg = BPF_REG_AX;
9285 insn->code = insn->code == code_add ?
9286 code_sub : code_add;
9288 if (issrc && isneg && !isimm)
9289 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
9290 cnt = patch - insn_buf;
9292 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9297 env->prog = prog = new_prog;
9298 insn = new_prog->insnsi + i + delta;
9302 if (insn->code != (BPF_JMP | BPF_CALL))
9304 if (insn->src_reg == BPF_PSEUDO_CALL)
9307 if (insn->imm == BPF_FUNC_get_route_realm)
9308 prog->dst_needed = 1;
9309 if (insn->imm == BPF_FUNC_get_prandom_u32)
9310 bpf_user_rnd_init_once();
9311 if (insn->imm == BPF_FUNC_override_return)
9312 prog->kprobe_override = 1;
9313 if (insn->imm == BPF_FUNC_tail_call) {
9314 /* If we tail call into other programs, we
9315 * cannot make any assumptions since they can
9316 * be replaced dynamically during runtime in
9317 * the program array.
9319 prog->cb_access = 1;
9320 env->prog->aux->stack_depth = MAX_BPF_STACK;
9321 env->prog->aux->max_pkt_offset = MAX_PACKET_OFF;
9323 /* mark bpf_tail_call as different opcode to avoid
9324 * conditional branch in the interpeter for every normal
9325 * call and to prevent accidental JITing by JIT compiler
9326 * that doesn't support bpf_tail_call yet
9329 insn->code = BPF_JMP | BPF_TAIL_CALL;
9331 aux = &env->insn_aux_data[i + delta];
9332 if (!bpf_map_ptr_unpriv(aux))
9335 /* instead of changing every JIT dealing with tail_call
9336 * emit two extra insns:
9337 * if (index >= max_entries) goto out;
9338 * index &= array->index_mask;
9339 * to avoid out-of-bounds cpu speculation
9341 if (bpf_map_ptr_poisoned(aux)) {
9342 verbose(env, "tail_call abusing map_ptr\n");
9346 map_ptr = BPF_MAP_PTR(aux->map_state);
9347 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
9348 map_ptr->max_entries, 2);
9349 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
9350 container_of(map_ptr,
9353 insn_buf[2] = *insn;
9355 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9360 env->prog = prog = new_prog;
9361 insn = new_prog->insnsi + i + delta;
9365 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
9366 * and other inlining handlers are currently limited to 64 bit
9369 if (prog->jit_requested && BITS_PER_LONG == 64 &&
9370 (insn->imm == BPF_FUNC_map_lookup_elem ||
9371 insn->imm == BPF_FUNC_map_update_elem ||
9372 insn->imm == BPF_FUNC_map_delete_elem ||
9373 insn->imm == BPF_FUNC_map_push_elem ||
9374 insn->imm == BPF_FUNC_map_pop_elem ||
9375 insn->imm == BPF_FUNC_map_peek_elem)) {
9376 aux = &env->insn_aux_data[i + delta];
9377 if (bpf_map_ptr_poisoned(aux))
9378 goto patch_call_imm;
9380 map_ptr = BPF_MAP_PTR(aux->map_state);
9382 if (insn->imm == BPF_FUNC_map_lookup_elem &&
9383 ops->map_gen_lookup) {
9384 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
9385 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
9386 verbose(env, "bpf verifier is misconfigured\n");
9390 new_prog = bpf_patch_insn_data(env, i + delta,
9396 env->prog = prog = new_prog;
9397 insn = new_prog->insnsi + i + delta;
9401 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
9402 (void *(*)(struct bpf_map *map, void *key))NULL));
9403 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
9404 (int (*)(struct bpf_map *map, void *key))NULL));
9405 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
9406 (int (*)(struct bpf_map *map, void *key, void *value,
9408 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
9409 (int (*)(struct bpf_map *map, void *value,
9411 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
9412 (int (*)(struct bpf_map *map, void *value))NULL));
9413 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
9414 (int (*)(struct bpf_map *map, void *value))NULL));
9416 switch (insn->imm) {
9417 case BPF_FUNC_map_lookup_elem:
9418 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
9421 case BPF_FUNC_map_update_elem:
9422 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
9425 case BPF_FUNC_map_delete_elem:
9426 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
9429 case BPF_FUNC_map_push_elem:
9430 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
9433 case BPF_FUNC_map_pop_elem:
9434 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
9437 case BPF_FUNC_map_peek_elem:
9438 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
9443 goto patch_call_imm;
9447 fn = env->ops->get_func_proto(insn->imm, env->prog);
9448 /* all functions that have prototype and verifier allowed
9449 * programs to call them, must be real in-kernel functions
9453 "kernel subsystem misconfigured func %s#%d\n",
9454 func_id_name(insn->imm), insn->imm);
9457 insn->imm = fn->func - __bpf_call_base;
9463 static void free_states(struct bpf_verifier_env *env)
9465 struct bpf_verifier_state_list *sl, *sln;
9468 sl = env->free_list;
9471 free_verifier_state(&sl->state, false);
9476 if (!env->explored_states)
9479 for (i = 0; i < state_htab_size(env); i++) {
9480 sl = env->explored_states[i];
9484 free_verifier_state(&sl->state, false);
9490 kvfree(env->explored_states);
9493 static void print_verification_stats(struct bpf_verifier_env *env)
9497 if (env->log.level & BPF_LOG_STATS) {
9498 verbose(env, "verification time %lld usec\n",
9499 div_u64(env->verification_time, 1000));
9500 verbose(env, "stack depth ");
9501 for (i = 0; i < env->subprog_cnt; i++) {
9502 u32 depth = env->subprog_info[i].stack_depth;
9504 verbose(env, "%d", depth);
9505 if (i + 1 < env->subprog_cnt)
9510 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
9511 "total_states %d peak_states %d mark_read %d\n",
9512 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
9513 env->max_states_per_insn, env->total_states,
9514 env->peak_states, env->longest_mark_read_walk);
9517 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr,
9518 union bpf_attr __user *uattr)
9520 u64 start_time = ktime_get_ns();
9521 struct bpf_verifier_env *env;
9522 struct bpf_verifier_log *log;
9523 int i, len, ret = -EINVAL;
9526 /* no program is valid */
9527 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
9530 /* 'struct bpf_verifier_env' can be global, but since it's not small,
9531 * allocate/free it every time bpf_check() is called
9533 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
9539 env->insn_aux_data =
9540 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
9542 if (!env->insn_aux_data)
9544 for (i = 0; i < len; i++)
9545 env->insn_aux_data[i].orig_idx = i;
9547 env->ops = bpf_verifier_ops[env->prog->type];
9548 is_priv = capable(CAP_SYS_ADMIN);
9550 /* grab the mutex to protect few globals used by verifier */
9552 mutex_lock(&bpf_verifier_lock);
9554 if (attr->log_level || attr->log_buf || attr->log_size) {
9555 /* user requested verbose verifier output
9556 * and supplied buffer to store the verification trace
9558 log->level = attr->log_level;
9559 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
9560 log->len_total = attr->log_size;
9563 /* log attributes have to be sane */
9564 if (log->len_total < 128 || log->len_total > UINT_MAX >> 2 ||
9565 !log->level || !log->ubuf || log->level & ~BPF_LOG_MASK)
9569 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
9570 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
9571 env->strict_alignment = true;
9572 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
9573 env->strict_alignment = false;
9575 env->allow_ptr_leaks = is_priv;
9578 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
9580 ret = replace_map_fd_with_map_ptr(env);
9582 goto skip_full_check;
9584 if (bpf_prog_is_dev_bound(env->prog->aux)) {
9585 ret = bpf_prog_offload_verifier_prep(env->prog);
9587 goto skip_full_check;
9590 env->explored_states = kvcalloc(state_htab_size(env),
9591 sizeof(struct bpf_verifier_state_list *),
9594 if (!env->explored_states)
9595 goto skip_full_check;
9597 ret = check_subprogs(env);
9599 goto skip_full_check;
9601 ret = check_btf_info(env, attr, uattr);
9603 goto skip_full_check;
9605 ret = check_cfg(env);
9607 goto skip_full_check;
9609 ret = do_check(env);
9610 if (env->cur_state) {
9611 free_verifier_state(env->cur_state, true);
9612 env->cur_state = NULL;
9615 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
9616 ret = bpf_prog_offload_finalize(env);
9619 while (!pop_stack(env, NULL, NULL));
9623 ret = check_max_stack_depth(env);
9625 /* instruction rewrites happen after this point */
9628 opt_hard_wire_dead_code_branches(env);
9630 ret = opt_remove_dead_code(env);
9632 ret = opt_remove_nops(env);
9635 sanitize_dead_code(env);
9639 /* program is valid, convert *(u32*)(ctx + off) accesses */
9640 ret = convert_ctx_accesses(env);
9643 ret = fixup_bpf_calls(env);
9645 /* do 32-bit optimization after insn patching has done so those patched
9646 * insns could be handled correctly.
9648 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
9649 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
9650 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
9655 ret = fixup_call_args(env);
9657 env->verification_time = ktime_get_ns() - start_time;
9658 print_verification_stats(env);
9660 if (log->level && bpf_verifier_log_full(log))
9662 if (log->level && !log->ubuf) {
9664 goto err_release_maps;
9667 if (ret == 0 && env->used_map_cnt) {
9668 /* if program passed verifier, update used_maps in bpf_prog_info */
9669 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
9670 sizeof(env->used_maps[0]),
9673 if (!env->prog->aux->used_maps) {
9675 goto err_release_maps;
9678 memcpy(env->prog->aux->used_maps, env->used_maps,
9679 sizeof(env->used_maps[0]) * env->used_map_cnt);
9680 env->prog->aux->used_map_cnt = env->used_map_cnt;
9682 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
9683 * bpf_ld_imm64 instructions
9685 convert_pseudo_ld_imm64(env);
9689 adjust_btf_func(env);
9692 if (!env->prog->aux->used_maps)
9693 /* if we didn't copy map pointers into bpf_prog_info, release
9694 * them now. Otherwise free_used_maps() will release them.
9700 mutex_unlock(&bpf_verifier_lock);
9701 vfree(env->insn_aux_data);