1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
3 * Copyright (c) 2016 Facebook
4 * Copyright (c) 2018 Covalent IO, Inc. http://covalent.io
6 #include <uapi/linux/btf.h>
7 #include <linux/kernel.h>
8 #include <linux/types.h>
9 #include <linux/slab.h>
10 #include <linux/bpf.h>
11 #include <linux/btf.h>
12 #include <linux/bpf_verifier.h>
13 #include <linux/filter.h>
14 #include <net/netlink.h>
15 #include <linux/file.h>
16 #include <linux/vmalloc.h>
17 #include <linux/stringify.h>
18 #include <linux/bsearch.h>
19 #include <linux/sort.h>
20 #include <linux/perf_event.h>
21 #include <linux/ctype.h>
22 #include <linux/error-injection.h>
23 #include <linux/bpf_lsm.h>
24 #include <linux/btf_ids.h>
28 static const struct bpf_verifier_ops
* const bpf_verifier_ops
[] = {
29 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
30 [_id] = & _name ## _verifier_ops,
31 #define BPF_MAP_TYPE(_id, _ops)
32 #define BPF_LINK_TYPE(_id, _name)
33 #include <linux/bpf_types.h>
39 /* bpf_check() is a static code analyzer that walks eBPF program
40 * instruction by instruction and updates register/stack state.
41 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
43 * The first pass is depth-first-search to check that the program is a DAG.
44 * It rejects the following programs:
45 * - larger than BPF_MAXINSNS insns
46 * - if loop is present (detected via back-edge)
47 * - unreachable insns exist (shouldn't be a forest. program = one function)
48 * - out of bounds or malformed jumps
49 * The second pass is all possible path descent from the 1st insn.
50 * Since it's analyzing all pathes through the program, the length of the
51 * analysis is limited to 64k insn, which may be hit even if total number of
52 * insn is less then 4K, but there are too many branches that change stack/regs.
53 * Number of 'branches to be analyzed' is limited to 1k
55 * On entry to each instruction, each register has a type, and the instruction
56 * changes the types of the registers depending on instruction semantics.
57 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
60 * All registers are 64-bit.
61 * R0 - return register
62 * R1-R5 argument passing registers
63 * R6-R9 callee saved registers
64 * R10 - frame pointer read-only
66 * At the start of BPF program the register R1 contains a pointer to bpf_context
67 * and has type PTR_TO_CTX.
69 * Verifier tracks arithmetic operations on pointers in case:
70 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
71 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
72 * 1st insn copies R10 (which has FRAME_PTR) type into R1
73 * and 2nd arithmetic instruction is pattern matched to recognize
74 * that it wants to construct a pointer to some element within stack.
75 * So after 2nd insn, the register R1 has type PTR_TO_STACK
76 * (and -20 constant is saved for further stack bounds checking).
77 * Meaning that this reg is a pointer to stack plus known immediate constant.
79 * Most of the time the registers have SCALAR_VALUE type, which
80 * means the register has some value, but it's not a valid pointer.
81 * (like pointer plus pointer becomes SCALAR_VALUE type)
83 * When verifier sees load or store instructions the type of base register
84 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
85 * four pointer types recognized by check_mem_access() function.
87 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
88 * and the range of [ptr, ptr + map's value_size) is accessible.
90 * registers used to pass values to function calls are checked against
91 * function argument constraints.
93 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
94 * It means that the register type passed to this function must be
95 * PTR_TO_STACK and it will be used inside the function as
96 * 'pointer to map element key'
98 * For example the argument constraints for bpf_map_lookup_elem():
99 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
100 * .arg1_type = ARG_CONST_MAP_PTR,
101 * .arg2_type = ARG_PTR_TO_MAP_KEY,
103 * ret_type says that this function returns 'pointer to map elem value or null'
104 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
105 * 2nd argument should be a pointer to stack, which will be used inside
106 * the helper function as a pointer to map element key.
108 * On the kernel side the helper function looks like:
109 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
111 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
112 * void *key = (void *) (unsigned long) r2;
115 * here kernel can access 'key' and 'map' pointers safely, knowing that
116 * [key, key + map->key_size) bytes are valid and were initialized on
117 * the stack of eBPF program.
120 * Corresponding eBPF program may look like:
121 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
122 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
123 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
124 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
125 * here verifier looks at prototype of map_lookup_elem() and sees:
126 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
127 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
129 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
130 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
131 * and were initialized prior to this call.
132 * If it's ok, then verifier allows this BPF_CALL insn and looks at
133 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
134 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
135 * returns ether pointer to map value or NULL.
137 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
138 * insn, the register holding that pointer in the true branch changes state to
139 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
140 * branch. See check_cond_jmp_op().
142 * After the call R0 is set to return type of the function and registers R1-R5
143 * are set to NOT_INIT to indicate that they are no longer readable.
145 * The following reference types represent a potential reference to a kernel
146 * resource which, after first being allocated, must be checked and freed by
148 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
150 * When the verifier sees a helper call return a reference type, it allocates a
151 * pointer id for the reference and stores it in the current function state.
152 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
153 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
154 * passes through a NULL-check conditional. For the branch wherein the state is
155 * changed to CONST_IMM, the verifier releases the reference.
157 * For each helper function that allocates a reference, such as
158 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
159 * bpf_sk_release(). When a reference type passes into the release function,
160 * the verifier also releases the reference. If any unchecked or unreleased
161 * reference remains at the end of the program, the verifier rejects it.
164 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
165 struct bpf_verifier_stack_elem
{
166 /* verifer state is 'st'
167 * before processing instruction 'insn_idx'
168 * and after processing instruction 'prev_insn_idx'
170 struct bpf_verifier_state st
;
173 struct bpf_verifier_stack_elem
*next
;
174 /* length of verifier log at the time this state was pushed on stack */
178 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
179 #define BPF_COMPLEXITY_LIMIT_STATES 64
181 #define BPF_MAP_KEY_POISON (1ULL << 63)
182 #define BPF_MAP_KEY_SEEN (1ULL << 62)
184 #define BPF_MAP_PTR_UNPRIV 1UL
185 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
186 POISON_POINTER_DELTA))
187 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
189 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data
*aux
)
191 return BPF_MAP_PTR(aux
->map_ptr_state
) == BPF_MAP_PTR_POISON
;
194 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data
*aux
)
196 return aux
->map_ptr_state
& BPF_MAP_PTR_UNPRIV
;
199 static void bpf_map_ptr_store(struct bpf_insn_aux_data
*aux
,
200 const struct bpf_map
*map
, bool unpriv
)
202 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON
& BPF_MAP_PTR_UNPRIV
);
203 unpriv
|= bpf_map_ptr_unpriv(aux
);
204 aux
->map_ptr_state
= (unsigned long)map
|
205 (unpriv
? BPF_MAP_PTR_UNPRIV
: 0UL);
208 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data
*aux
)
210 return aux
->map_key_state
& BPF_MAP_KEY_POISON
;
213 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data
*aux
)
215 return !(aux
->map_key_state
& BPF_MAP_KEY_SEEN
);
218 static u64
bpf_map_key_immediate(const struct bpf_insn_aux_data
*aux
)
220 return aux
->map_key_state
& ~(BPF_MAP_KEY_SEEN
| BPF_MAP_KEY_POISON
);
223 static void bpf_map_key_store(struct bpf_insn_aux_data
*aux
, u64 state
)
225 bool poisoned
= bpf_map_key_poisoned(aux
);
227 aux
->map_key_state
= state
| BPF_MAP_KEY_SEEN
|
228 (poisoned
? BPF_MAP_KEY_POISON
: 0ULL);
231 struct bpf_call_arg_meta
{
232 struct bpf_map
*map_ptr
;
247 struct btf
*btf_vmlinux
;
249 static DEFINE_MUTEX(bpf_verifier_lock
);
251 static const struct bpf_line_info
*
252 find_linfo(const struct bpf_verifier_env
*env
, u32 insn_off
)
254 const struct bpf_line_info
*linfo
;
255 const struct bpf_prog
*prog
;
259 nr_linfo
= prog
->aux
->nr_linfo
;
261 if (!nr_linfo
|| insn_off
>= prog
->len
)
264 linfo
= prog
->aux
->linfo
;
265 for (i
= 1; i
< nr_linfo
; i
++)
266 if (insn_off
< linfo
[i
].insn_off
)
269 return &linfo
[i
- 1];
272 void bpf_verifier_vlog(struct bpf_verifier_log
*log
, const char *fmt
,
277 n
= vscnprintf(log
->kbuf
, BPF_VERIFIER_TMP_LOG_SIZE
, fmt
, args
);
279 WARN_ONCE(n
>= BPF_VERIFIER_TMP_LOG_SIZE
- 1,
280 "verifier log line truncated - local buffer too short\n");
282 n
= min(log
->len_total
- log
->len_used
- 1, n
);
285 if (log
->level
== BPF_LOG_KERNEL
) {
286 pr_err("BPF:%s\n", log
->kbuf
);
289 if (!copy_to_user(log
->ubuf
+ log
->len_used
, log
->kbuf
, n
+ 1))
295 static void bpf_vlog_reset(struct bpf_verifier_log
*log
, u32 new_pos
)
299 if (!bpf_verifier_log_needed(log
))
302 log
->len_used
= new_pos
;
303 if (put_user(zero
, log
->ubuf
+ new_pos
))
307 /* log_level controls verbosity level of eBPF verifier.
308 * bpf_verifier_log_write() is used to dump the verification trace to the log,
309 * so the user can figure out what's wrong with the program
311 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env
*env
,
312 const char *fmt
, ...)
316 if (!bpf_verifier_log_needed(&env
->log
))
320 bpf_verifier_vlog(&env
->log
, fmt
, args
);
323 EXPORT_SYMBOL_GPL(bpf_verifier_log_write
);
325 __printf(2, 3) static void verbose(void *private_data
, const char *fmt
, ...)
327 struct bpf_verifier_env
*env
= private_data
;
330 if (!bpf_verifier_log_needed(&env
->log
))
334 bpf_verifier_vlog(&env
->log
, fmt
, args
);
338 __printf(2, 3) void bpf_log(struct bpf_verifier_log
*log
,
339 const char *fmt
, ...)
343 if (!bpf_verifier_log_needed(log
))
347 bpf_verifier_vlog(log
, fmt
, args
);
351 static const char *ltrim(const char *s
)
359 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env
*env
,
361 const char *prefix_fmt
, ...)
363 const struct bpf_line_info
*linfo
;
365 if (!bpf_verifier_log_needed(&env
->log
))
368 linfo
= find_linfo(env
, insn_off
);
369 if (!linfo
|| linfo
== env
->prev_linfo
)
375 va_start(args
, prefix_fmt
);
376 bpf_verifier_vlog(&env
->log
, prefix_fmt
, args
);
381 ltrim(btf_name_by_offset(env
->prog
->aux
->btf
,
384 env
->prev_linfo
= linfo
;
387 static bool type_is_pkt_pointer(enum bpf_reg_type type
)
389 return type
== PTR_TO_PACKET
||
390 type
== PTR_TO_PACKET_META
;
393 static bool type_is_sk_pointer(enum bpf_reg_type type
)
395 return type
== PTR_TO_SOCKET
||
396 type
== PTR_TO_SOCK_COMMON
||
397 type
== PTR_TO_TCP_SOCK
||
398 type
== PTR_TO_XDP_SOCK
;
401 static bool reg_type_not_null(enum bpf_reg_type type
)
403 return type
== PTR_TO_SOCKET
||
404 type
== PTR_TO_TCP_SOCK
||
405 type
== PTR_TO_MAP_VALUE
||
406 type
== PTR_TO_SOCK_COMMON
;
409 static bool reg_type_may_be_null(enum bpf_reg_type type
)
411 return type
== PTR_TO_MAP_VALUE_OR_NULL
||
412 type
== PTR_TO_SOCKET_OR_NULL
||
413 type
== PTR_TO_SOCK_COMMON_OR_NULL
||
414 type
== PTR_TO_TCP_SOCK_OR_NULL
||
415 type
== PTR_TO_BTF_ID_OR_NULL
||
416 type
== PTR_TO_MEM_OR_NULL
||
417 type
== PTR_TO_RDONLY_BUF_OR_NULL
||
418 type
== PTR_TO_RDWR_BUF_OR_NULL
;
421 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state
*reg
)
423 return reg
->type
== PTR_TO_MAP_VALUE
&&
424 map_value_has_spin_lock(reg
->map_ptr
);
427 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type
)
429 return type
== PTR_TO_SOCKET
||
430 type
== PTR_TO_SOCKET_OR_NULL
||
431 type
== PTR_TO_TCP_SOCK
||
432 type
== PTR_TO_TCP_SOCK_OR_NULL
||
433 type
== PTR_TO_MEM
||
434 type
== PTR_TO_MEM_OR_NULL
;
437 static bool arg_type_may_be_refcounted(enum bpf_arg_type type
)
439 return type
== ARG_PTR_TO_SOCK_COMMON
;
442 static bool arg_type_may_be_null(enum bpf_arg_type type
)
444 return type
== ARG_PTR_TO_MAP_VALUE_OR_NULL
||
445 type
== ARG_PTR_TO_MEM_OR_NULL
||
446 type
== ARG_PTR_TO_CTX_OR_NULL
||
447 type
== ARG_PTR_TO_SOCKET_OR_NULL
||
448 type
== ARG_PTR_TO_ALLOC_MEM_OR_NULL
;
451 /* Determine whether the function releases some resources allocated by another
452 * function call. The first reference type argument will be assumed to be
453 * released by release_reference().
455 static bool is_release_function(enum bpf_func_id func_id
)
457 return func_id
== BPF_FUNC_sk_release
||
458 func_id
== BPF_FUNC_ringbuf_submit
||
459 func_id
== BPF_FUNC_ringbuf_discard
;
462 static bool may_be_acquire_function(enum bpf_func_id func_id
)
464 return func_id
== BPF_FUNC_sk_lookup_tcp
||
465 func_id
== BPF_FUNC_sk_lookup_udp
||
466 func_id
== BPF_FUNC_skc_lookup_tcp
||
467 func_id
== BPF_FUNC_map_lookup_elem
||
468 func_id
== BPF_FUNC_ringbuf_reserve
;
471 static bool is_acquire_function(enum bpf_func_id func_id
,
472 const struct bpf_map
*map
)
474 enum bpf_map_type map_type
= map
? map
->map_type
: BPF_MAP_TYPE_UNSPEC
;
476 if (func_id
== BPF_FUNC_sk_lookup_tcp
||
477 func_id
== BPF_FUNC_sk_lookup_udp
||
478 func_id
== BPF_FUNC_skc_lookup_tcp
||
479 func_id
== BPF_FUNC_ringbuf_reserve
)
482 if (func_id
== BPF_FUNC_map_lookup_elem
&&
483 (map_type
== BPF_MAP_TYPE_SOCKMAP
||
484 map_type
== BPF_MAP_TYPE_SOCKHASH
))
490 static bool is_ptr_cast_function(enum bpf_func_id func_id
)
492 return func_id
== BPF_FUNC_tcp_sock
||
493 func_id
== BPF_FUNC_sk_fullsock
||
494 func_id
== BPF_FUNC_skc_to_tcp_sock
||
495 func_id
== BPF_FUNC_skc_to_tcp6_sock
||
496 func_id
== BPF_FUNC_skc_to_udp6_sock
||
497 func_id
== BPF_FUNC_skc_to_tcp_timewait_sock
||
498 func_id
== BPF_FUNC_skc_to_tcp_request_sock
;
501 /* string representation of 'enum bpf_reg_type' */
502 static const char * const reg_type_str
[] = {
504 [SCALAR_VALUE
] = "inv",
505 [PTR_TO_CTX
] = "ctx",
506 [CONST_PTR_TO_MAP
] = "map_ptr",
507 [PTR_TO_MAP_VALUE
] = "map_value",
508 [PTR_TO_MAP_VALUE_OR_NULL
] = "map_value_or_null",
509 [PTR_TO_STACK
] = "fp",
510 [PTR_TO_PACKET
] = "pkt",
511 [PTR_TO_PACKET_META
] = "pkt_meta",
512 [PTR_TO_PACKET_END
] = "pkt_end",
513 [PTR_TO_FLOW_KEYS
] = "flow_keys",
514 [PTR_TO_SOCKET
] = "sock",
515 [PTR_TO_SOCKET_OR_NULL
] = "sock_or_null",
516 [PTR_TO_SOCK_COMMON
] = "sock_common",
517 [PTR_TO_SOCK_COMMON_OR_NULL
] = "sock_common_or_null",
518 [PTR_TO_TCP_SOCK
] = "tcp_sock",
519 [PTR_TO_TCP_SOCK_OR_NULL
] = "tcp_sock_or_null",
520 [PTR_TO_TP_BUFFER
] = "tp_buffer",
521 [PTR_TO_XDP_SOCK
] = "xdp_sock",
522 [PTR_TO_BTF_ID
] = "ptr_",
523 [PTR_TO_BTF_ID_OR_NULL
] = "ptr_or_null_",
524 [PTR_TO_PERCPU_BTF_ID
] = "percpu_ptr_",
525 [PTR_TO_MEM
] = "mem",
526 [PTR_TO_MEM_OR_NULL
] = "mem_or_null",
527 [PTR_TO_RDONLY_BUF
] = "rdonly_buf",
528 [PTR_TO_RDONLY_BUF_OR_NULL
] = "rdonly_buf_or_null",
529 [PTR_TO_RDWR_BUF
] = "rdwr_buf",
530 [PTR_TO_RDWR_BUF_OR_NULL
] = "rdwr_buf_or_null",
533 static char slot_type_char
[] = {
534 [STACK_INVALID
] = '?',
540 static void print_liveness(struct bpf_verifier_env
*env
,
541 enum bpf_reg_liveness live
)
543 if (live
& (REG_LIVE_READ
| REG_LIVE_WRITTEN
| REG_LIVE_DONE
))
545 if (live
& REG_LIVE_READ
)
547 if (live
& REG_LIVE_WRITTEN
)
549 if (live
& REG_LIVE_DONE
)
553 static struct bpf_func_state
*func(struct bpf_verifier_env
*env
,
554 const struct bpf_reg_state
*reg
)
556 struct bpf_verifier_state
*cur
= env
->cur_state
;
558 return cur
->frame
[reg
->frameno
];
561 static const char *kernel_type_name(const struct btf
* btf
, u32 id
)
563 return btf_name_by_offset(btf
, btf_type_by_id(btf
, id
)->name_off
);
566 static void print_verifier_state(struct bpf_verifier_env
*env
,
567 const struct bpf_func_state
*state
)
569 const struct bpf_reg_state
*reg
;
574 verbose(env
, " frame%d:", state
->frameno
);
575 for (i
= 0; i
< MAX_BPF_REG
; i
++) {
576 reg
= &state
->regs
[i
];
580 verbose(env
, " R%d", i
);
581 print_liveness(env
, reg
->live
);
582 verbose(env
, "=%s", reg_type_str
[t
]);
583 if (t
== SCALAR_VALUE
&& reg
->precise
)
585 if ((t
== SCALAR_VALUE
|| t
== PTR_TO_STACK
) &&
586 tnum_is_const(reg
->var_off
)) {
587 /* reg->off should be 0 for SCALAR_VALUE */
588 verbose(env
, "%lld", reg
->var_off
.value
+ reg
->off
);
590 if (t
== PTR_TO_BTF_ID
||
591 t
== PTR_TO_BTF_ID_OR_NULL
||
592 t
== PTR_TO_PERCPU_BTF_ID
)
593 verbose(env
, "%s", kernel_type_name(reg
->btf
, reg
->btf_id
));
594 verbose(env
, "(id=%d", reg
->id
);
595 if (reg_type_may_be_refcounted_or_null(t
))
596 verbose(env
, ",ref_obj_id=%d", reg
->ref_obj_id
);
597 if (t
!= SCALAR_VALUE
)
598 verbose(env
, ",off=%d", reg
->off
);
599 if (type_is_pkt_pointer(t
))
600 verbose(env
, ",r=%d", reg
->range
);
601 else if (t
== CONST_PTR_TO_MAP
||
602 t
== PTR_TO_MAP_VALUE
||
603 t
== PTR_TO_MAP_VALUE_OR_NULL
)
604 verbose(env
, ",ks=%d,vs=%d",
605 reg
->map_ptr
->key_size
,
606 reg
->map_ptr
->value_size
);
607 if (tnum_is_const(reg
->var_off
)) {
608 /* Typically an immediate SCALAR_VALUE, but
609 * could be a pointer whose offset is too big
612 verbose(env
, ",imm=%llx", reg
->var_off
.value
);
614 if (reg
->smin_value
!= reg
->umin_value
&&
615 reg
->smin_value
!= S64_MIN
)
616 verbose(env
, ",smin_value=%lld",
617 (long long)reg
->smin_value
);
618 if (reg
->smax_value
!= reg
->umax_value
&&
619 reg
->smax_value
!= S64_MAX
)
620 verbose(env
, ",smax_value=%lld",
621 (long long)reg
->smax_value
);
622 if (reg
->umin_value
!= 0)
623 verbose(env
, ",umin_value=%llu",
624 (unsigned long long)reg
->umin_value
);
625 if (reg
->umax_value
!= U64_MAX
)
626 verbose(env
, ",umax_value=%llu",
627 (unsigned long long)reg
->umax_value
);
628 if (!tnum_is_unknown(reg
->var_off
)) {
631 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
632 verbose(env
, ",var_off=%s", tn_buf
);
634 if (reg
->s32_min_value
!= reg
->smin_value
&&
635 reg
->s32_min_value
!= S32_MIN
)
636 verbose(env
, ",s32_min_value=%d",
637 (int)(reg
->s32_min_value
));
638 if (reg
->s32_max_value
!= reg
->smax_value
&&
639 reg
->s32_max_value
!= S32_MAX
)
640 verbose(env
, ",s32_max_value=%d",
641 (int)(reg
->s32_max_value
));
642 if (reg
->u32_min_value
!= reg
->umin_value
&&
643 reg
->u32_min_value
!= U32_MIN
)
644 verbose(env
, ",u32_min_value=%d",
645 (int)(reg
->u32_min_value
));
646 if (reg
->u32_max_value
!= reg
->umax_value
&&
647 reg
->u32_max_value
!= U32_MAX
)
648 verbose(env
, ",u32_max_value=%d",
649 (int)(reg
->u32_max_value
));
654 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
655 char types_buf
[BPF_REG_SIZE
+ 1];
659 for (j
= 0; j
< BPF_REG_SIZE
; j
++) {
660 if (state
->stack
[i
].slot_type
[j
] != STACK_INVALID
)
662 types_buf
[j
] = slot_type_char
[
663 state
->stack
[i
].slot_type
[j
]];
665 types_buf
[BPF_REG_SIZE
] = 0;
668 verbose(env
, " fp%d", (-i
- 1) * BPF_REG_SIZE
);
669 print_liveness(env
, state
->stack
[i
].spilled_ptr
.live
);
670 if (state
->stack
[i
].slot_type
[0] == STACK_SPILL
) {
671 reg
= &state
->stack
[i
].spilled_ptr
;
673 verbose(env
, "=%s", reg_type_str
[t
]);
674 if (t
== SCALAR_VALUE
&& reg
->precise
)
676 if (t
== SCALAR_VALUE
&& tnum_is_const(reg
->var_off
))
677 verbose(env
, "%lld", reg
->var_off
.value
+ reg
->off
);
679 verbose(env
, "=%s", types_buf
);
682 if (state
->acquired_refs
&& state
->refs
[0].id
) {
683 verbose(env
, " refs=%d", state
->refs
[0].id
);
684 for (i
= 1; i
< state
->acquired_refs
; i
++)
685 if (state
->refs
[i
].id
)
686 verbose(env
, ",%d", state
->refs
[i
].id
);
691 #define COPY_STATE_FN(NAME, COUNT, FIELD, SIZE) \
692 static int copy_##NAME##_state(struct bpf_func_state *dst, \
693 const struct bpf_func_state *src) \
697 if (WARN_ON_ONCE(dst->COUNT < src->COUNT)) { \
698 /* internal bug, make state invalid to reject the program */ \
699 memset(dst, 0, sizeof(*dst)); \
702 memcpy(dst->FIELD, src->FIELD, \
703 sizeof(*src->FIELD) * (src->COUNT / SIZE)); \
706 /* copy_reference_state() */
707 COPY_STATE_FN(reference
, acquired_refs
, refs
, 1)
708 /* copy_stack_state() */
709 COPY_STATE_FN(stack
, allocated_stack
, stack
, BPF_REG_SIZE
)
712 #define REALLOC_STATE_FN(NAME, COUNT, FIELD, SIZE) \
713 static int realloc_##NAME##_state(struct bpf_func_state *state, int size, \
716 u32 old_size = state->COUNT; \
717 struct bpf_##NAME##_state *new_##FIELD; \
718 int slot = size / SIZE; \
720 if (size <= old_size || !size) { \
723 state->COUNT = slot * SIZE; \
724 if (!size && old_size) { \
725 kfree(state->FIELD); \
726 state->FIELD = NULL; \
730 new_##FIELD = kmalloc_array(slot, sizeof(struct bpf_##NAME##_state), \
736 memcpy(new_##FIELD, state->FIELD, \
737 sizeof(*new_##FIELD) * (old_size / SIZE)); \
738 memset(new_##FIELD + old_size / SIZE, 0, \
739 sizeof(*new_##FIELD) * (size - old_size) / SIZE); \
741 state->COUNT = slot * SIZE; \
742 kfree(state->FIELD); \
743 state->FIELD = new_##FIELD; \
746 /* realloc_reference_state() */
747 REALLOC_STATE_FN(reference
, acquired_refs
, refs
, 1)
748 /* realloc_stack_state() */
749 REALLOC_STATE_FN(stack
, allocated_stack
, stack
, BPF_REG_SIZE
)
750 #undef REALLOC_STATE_FN
752 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
753 * make it consume minimal amount of memory. check_stack_write() access from
754 * the program calls into realloc_func_state() to grow the stack size.
755 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
756 * which realloc_stack_state() copies over. It points to previous
757 * bpf_verifier_state which is never reallocated.
759 static int realloc_func_state(struct bpf_func_state
*state
, int stack_size
,
760 int refs_size
, bool copy_old
)
762 int err
= realloc_reference_state(state
, refs_size
, copy_old
);
765 return realloc_stack_state(state
, stack_size
, copy_old
);
768 /* Acquire a pointer id from the env and update the state->refs to include
769 * this new pointer reference.
770 * On success, returns a valid pointer id to associate with the register
771 * On failure, returns a negative errno.
773 static int acquire_reference_state(struct bpf_verifier_env
*env
, int insn_idx
)
775 struct bpf_func_state
*state
= cur_func(env
);
776 int new_ofs
= state
->acquired_refs
;
779 err
= realloc_reference_state(state
, state
->acquired_refs
+ 1, true);
783 state
->refs
[new_ofs
].id
= id
;
784 state
->refs
[new_ofs
].insn_idx
= insn_idx
;
789 /* release function corresponding to acquire_reference_state(). Idempotent. */
790 static int release_reference_state(struct bpf_func_state
*state
, int ptr_id
)
794 last_idx
= state
->acquired_refs
- 1;
795 for (i
= 0; i
< state
->acquired_refs
; i
++) {
796 if (state
->refs
[i
].id
== ptr_id
) {
797 if (last_idx
&& i
!= last_idx
)
798 memcpy(&state
->refs
[i
], &state
->refs
[last_idx
],
799 sizeof(*state
->refs
));
800 memset(&state
->refs
[last_idx
], 0, sizeof(*state
->refs
));
801 state
->acquired_refs
--;
808 static int transfer_reference_state(struct bpf_func_state
*dst
,
809 struct bpf_func_state
*src
)
811 int err
= realloc_reference_state(dst
, src
->acquired_refs
, false);
814 err
= copy_reference_state(dst
, src
);
820 static void free_func_state(struct bpf_func_state
*state
)
829 static void clear_jmp_history(struct bpf_verifier_state
*state
)
831 kfree(state
->jmp_history
);
832 state
->jmp_history
= NULL
;
833 state
->jmp_history_cnt
= 0;
836 static void free_verifier_state(struct bpf_verifier_state
*state
,
841 for (i
= 0; i
<= state
->curframe
; i
++) {
842 free_func_state(state
->frame
[i
]);
843 state
->frame
[i
] = NULL
;
845 clear_jmp_history(state
);
850 /* copy verifier state from src to dst growing dst stack space
851 * when necessary to accommodate larger src stack
853 static int copy_func_state(struct bpf_func_state
*dst
,
854 const struct bpf_func_state
*src
)
858 err
= realloc_func_state(dst
, src
->allocated_stack
, src
->acquired_refs
,
862 memcpy(dst
, src
, offsetof(struct bpf_func_state
, acquired_refs
));
863 err
= copy_reference_state(dst
, src
);
866 return copy_stack_state(dst
, src
);
869 static int copy_verifier_state(struct bpf_verifier_state
*dst_state
,
870 const struct bpf_verifier_state
*src
)
872 struct bpf_func_state
*dst
;
873 u32 jmp_sz
= sizeof(struct bpf_idx_pair
) * src
->jmp_history_cnt
;
876 if (dst_state
->jmp_history_cnt
< src
->jmp_history_cnt
) {
877 kfree(dst_state
->jmp_history
);
878 dst_state
->jmp_history
= kmalloc(jmp_sz
, GFP_USER
);
879 if (!dst_state
->jmp_history
)
882 memcpy(dst_state
->jmp_history
, src
->jmp_history
, jmp_sz
);
883 dst_state
->jmp_history_cnt
= src
->jmp_history_cnt
;
885 /* if dst has more stack frames then src frame, free them */
886 for (i
= src
->curframe
+ 1; i
<= dst_state
->curframe
; i
++) {
887 free_func_state(dst_state
->frame
[i
]);
888 dst_state
->frame
[i
] = NULL
;
890 dst_state
->speculative
= src
->speculative
;
891 dst_state
->curframe
= src
->curframe
;
892 dst_state
->active_spin_lock
= src
->active_spin_lock
;
893 dst_state
->branches
= src
->branches
;
894 dst_state
->parent
= src
->parent
;
895 dst_state
->first_insn_idx
= src
->first_insn_idx
;
896 dst_state
->last_insn_idx
= src
->last_insn_idx
;
897 for (i
= 0; i
<= src
->curframe
; i
++) {
898 dst
= dst_state
->frame
[i
];
900 dst
= kzalloc(sizeof(*dst
), GFP_KERNEL
);
903 dst_state
->frame
[i
] = dst
;
905 err
= copy_func_state(dst
, src
->frame
[i
]);
912 static void update_branch_counts(struct bpf_verifier_env
*env
, struct bpf_verifier_state
*st
)
915 u32 br
= --st
->branches
;
917 /* WARN_ON(br > 1) technically makes sense here,
918 * but see comment in push_stack(), hence:
920 WARN_ONCE((int)br
< 0,
921 "BUG update_branch_counts:branches_to_explore=%d\n",
929 static int pop_stack(struct bpf_verifier_env
*env
, int *prev_insn_idx
,
930 int *insn_idx
, bool pop_log
)
932 struct bpf_verifier_state
*cur
= env
->cur_state
;
933 struct bpf_verifier_stack_elem
*elem
, *head
= env
->head
;
936 if (env
->head
== NULL
)
940 err
= copy_verifier_state(cur
, &head
->st
);
945 bpf_vlog_reset(&env
->log
, head
->log_pos
);
947 *insn_idx
= head
->insn_idx
;
949 *prev_insn_idx
= head
->prev_insn_idx
;
951 free_verifier_state(&head
->st
, false);
958 static struct bpf_verifier_state
*push_stack(struct bpf_verifier_env
*env
,
959 int insn_idx
, int prev_insn_idx
,
962 struct bpf_verifier_state
*cur
= env
->cur_state
;
963 struct bpf_verifier_stack_elem
*elem
;
966 elem
= kzalloc(sizeof(struct bpf_verifier_stack_elem
), GFP_KERNEL
);
970 elem
->insn_idx
= insn_idx
;
971 elem
->prev_insn_idx
= prev_insn_idx
;
972 elem
->next
= env
->head
;
973 elem
->log_pos
= env
->log
.len_used
;
976 err
= copy_verifier_state(&elem
->st
, cur
);
979 elem
->st
.speculative
|= speculative
;
980 if (env
->stack_size
> BPF_COMPLEXITY_LIMIT_JMP_SEQ
) {
981 verbose(env
, "The sequence of %d jumps is too complex.\n",
985 if (elem
->st
.parent
) {
986 ++elem
->st
.parent
->branches
;
987 /* WARN_ON(branches > 2) technically makes sense here,
989 * 1. speculative states will bump 'branches' for non-branch
991 * 2. is_state_visited() heuristics may decide not to create
992 * a new state for a sequence of branches and all such current
993 * and cloned states will be pointing to a single parent state
994 * which might have large 'branches' count.
999 free_verifier_state(env
->cur_state
, true);
1000 env
->cur_state
= NULL
;
1001 /* pop all elements and return */
1002 while (!pop_stack(env
, NULL
, NULL
, false));
1006 #define CALLER_SAVED_REGS 6
1007 static const int caller_saved
[CALLER_SAVED_REGS
] = {
1008 BPF_REG_0
, BPF_REG_1
, BPF_REG_2
, BPF_REG_3
, BPF_REG_4
, BPF_REG_5
1011 static void __mark_reg_not_init(const struct bpf_verifier_env
*env
,
1012 struct bpf_reg_state
*reg
);
1014 /* This helper doesn't clear reg->id */
1015 static void ___mark_reg_known(struct bpf_reg_state
*reg
, u64 imm
)
1017 reg
->var_off
= tnum_const(imm
);
1018 reg
->smin_value
= (s64
)imm
;
1019 reg
->smax_value
= (s64
)imm
;
1020 reg
->umin_value
= imm
;
1021 reg
->umax_value
= imm
;
1023 reg
->s32_min_value
= (s32
)imm
;
1024 reg
->s32_max_value
= (s32
)imm
;
1025 reg
->u32_min_value
= (u32
)imm
;
1026 reg
->u32_max_value
= (u32
)imm
;
1029 /* Mark the unknown part of a register (variable offset or scalar value) as
1030 * known to have the value @imm.
1032 static void __mark_reg_known(struct bpf_reg_state
*reg
, u64 imm
)
1034 /* Clear id, off, and union(map_ptr, range) */
1035 memset(((u8
*)reg
) + sizeof(reg
->type
), 0,
1036 offsetof(struct bpf_reg_state
, var_off
) - sizeof(reg
->type
));
1037 ___mark_reg_known(reg
, imm
);
1040 static void __mark_reg32_known(struct bpf_reg_state
*reg
, u64 imm
)
1042 reg
->var_off
= tnum_const_subreg(reg
->var_off
, imm
);
1043 reg
->s32_min_value
= (s32
)imm
;
1044 reg
->s32_max_value
= (s32
)imm
;
1045 reg
->u32_min_value
= (u32
)imm
;
1046 reg
->u32_max_value
= (u32
)imm
;
1049 /* Mark the 'variable offset' part of a register as zero. This should be
1050 * used only on registers holding a pointer type.
1052 static void __mark_reg_known_zero(struct bpf_reg_state
*reg
)
1054 __mark_reg_known(reg
, 0);
1057 static void __mark_reg_const_zero(struct bpf_reg_state
*reg
)
1059 __mark_reg_known(reg
, 0);
1060 reg
->type
= SCALAR_VALUE
;
1063 static void mark_reg_known_zero(struct bpf_verifier_env
*env
,
1064 struct bpf_reg_state
*regs
, u32 regno
)
1066 if (WARN_ON(regno
>= MAX_BPF_REG
)) {
1067 verbose(env
, "mark_reg_known_zero(regs, %u)\n", regno
);
1068 /* Something bad happened, let's kill all regs */
1069 for (regno
= 0; regno
< MAX_BPF_REG
; regno
++)
1070 __mark_reg_not_init(env
, regs
+ regno
);
1073 __mark_reg_known_zero(regs
+ regno
);
1076 static bool reg_is_pkt_pointer(const struct bpf_reg_state
*reg
)
1078 return type_is_pkt_pointer(reg
->type
);
1081 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state
*reg
)
1083 return reg_is_pkt_pointer(reg
) ||
1084 reg
->type
== PTR_TO_PACKET_END
;
1087 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
1088 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state
*reg
,
1089 enum bpf_reg_type which
)
1091 /* The register can already have a range from prior markings.
1092 * This is fine as long as it hasn't been advanced from its
1095 return reg
->type
== which
&&
1098 tnum_equals_const(reg
->var_off
, 0);
1101 /* Reset the min/max bounds of a register */
1102 static void __mark_reg_unbounded(struct bpf_reg_state
*reg
)
1104 reg
->smin_value
= S64_MIN
;
1105 reg
->smax_value
= S64_MAX
;
1106 reg
->umin_value
= 0;
1107 reg
->umax_value
= U64_MAX
;
1109 reg
->s32_min_value
= S32_MIN
;
1110 reg
->s32_max_value
= S32_MAX
;
1111 reg
->u32_min_value
= 0;
1112 reg
->u32_max_value
= U32_MAX
;
1115 static void __mark_reg64_unbounded(struct bpf_reg_state
*reg
)
1117 reg
->smin_value
= S64_MIN
;
1118 reg
->smax_value
= S64_MAX
;
1119 reg
->umin_value
= 0;
1120 reg
->umax_value
= U64_MAX
;
1123 static void __mark_reg32_unbounded(struct bpf_reg_state
*reg
)
1125 reg
->s32_min_value
= S32_MIN
;
1126 reg
->s32_max_value
= S32_MAX
;
1127 reg
->u32_min_value
= 0;
1128 reg
->u32_max_value
= U32_MAX
;
1131 static void __update_reg32_bounds(struct bpf_reg_state
*reg
)
1133 struct tnum var32_off
= tnum_subreg(reg
->var_off
);
1135 /* min signed is max(sign bit) | min(other bits) */
1136 reg
->s32_min_value
= max_t(s32
, reg
->s32_min_value
,
1137 var32_off
.value
| (var32_off
.mask
& S32_MIN
));
1138 /* max signed is min(sign bit) | max(other bits) */
1139 reg
->s32_max_value
= min_t(s32
, reg
->s32_max_value
,
1140 var32_off
.value
| (var32_off
.mask
& S32_MAX
));
1141 reg
->u32_min_value
= max_t(u32
, reg
->u32_min_value
, (u32
)var32_off
.value
);
1142 reg
->u32_max_value
= min(reg
->u32_max_value
,
1143 (u32
)(var32_off
.value
| var32_off
.mask
));
1146 static void __update_reg64_bounds(struct bpf_reg_state
*reg
)
1148 /* min signed is max(sign bit) | min(other bits) */
1149 reg
->smin_value
= max_t(s64
, reg
->smin_value
,
1150 reg
->var_off
.value
| (reg
->var_off
.mask
& S64_MIN
));
1151 /* max signed is min(sign bit) | max(other bits) */
1152 reg
->smax_value
= min_t(s64
, reg
->smax_value
,
1153 reg
->var_off
.value
| (reg
->var_off
.mask
& S64_MAX
));
1154 reg
->umin_value
= max(reg
->umin_value
, reg
->var_off
.value
);
1155 reg
->umax_value
= min(reg
->umax_value
,
1156 reg
->var_off
.value
| reg
->var_off
.mask
);
1159 static void __update_reg_bounds(struct bpf_reg_state
*reg
)
1161 __update_reg32_bounds(reg
);
1162 __update_reg64_bounds(reg
);
1165 /* Uses signed min/max values to inform unsigned, and vice-versa */
1166 static void __reg32_deduce_bounds(struct bpf_reg_state
*reg
)
1168 /* Learn sign from signed bounds.
1169 * If we cannot cross the sign boundary, then signed and unsigned bounds
1170 * are the same, so combine. This works even in the negative case, e.g.
1171 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1173 if (reg
->s32_min_value
>= 0 || reg
->s32_max_value
< 0) {
1174 reg
->s32_min_value
= reg
->u32_min_value
=
1175 max_t(u32
, reg
->s32_min_value
, reg
->u32_min_value
);
1176 reg
->s32_max_value
= reg
->u32_max_value
=
1177 min_t(u32
, reg
->s32_max_value
, reg
->u32_max_value
);
1180 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1181 * boundary, so we must be careful.
1183 if ((s32
)reg
->u32_max_value
>= 0) {
1184 /* Positive. We can't learn anything from the smin, but smax
1185 * is positive, hence safe.
1187 reg
->s32_min_value
= reg
->u32_min_value
;
1188 reg
->s32_max_value
= reg
->u32_max_value
=
1189 min_t(u32
, reg
->s32_max_value
, reg
->u32_max_value
);
1190 } else if ((s32
)reg
->u32_min_value
< 0) {
1191 /* Negative. We can't learn anything from the smax, but smin
1192 * is negative, hence safe.
1194 reg
->s32_min_value
= reg
->u32_min_value
=
1195 max_t(u32
, reg
->s32_min_value
, reg
->u32_min_value
);
1196 reg
->s32_max_value
= reg
->u32_max_value
;
1200 static void __reg64_deduce_bounds(struct bpf_reg_state
*reg
)
1202 /* Learn sign from signed bounds.
1203 * If we cannot cross the sign boundary, then signed and unsigned bounds
1204 * are the same, so combine. This works even in the negative case, e.g.
1205 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1207 if (reg
->smin_value
>= 0 || reg
->smax_value
< 0) {
1208 reg
->smin_value
= reg
->umin_value
= max_t(u64
, reg
->smin_value
,
1210 reg
->smax_value
= reg
->umax_value
= min_t(u64
, reg
->smax_value
,
1214 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1215 * boundary, so we must be careful.
1217 if ((s64
)reg
->umax_value
>= 0) {
1218 /* Positive. We can't learn anything from the smin, but smax
1219 * is positive, hence safe.
1221 reg
->smin_value
= reg
->umin_value
;
1222 reg
->smax_value
= reg
->umax_value
= min_t(u64
, reg
->smax_value
,
1224 } else if ((s64
)reg
->umin_value
< 0) {
1225 /* Negative. We can't learn anything from the smax, but smin
1226 * is negative, hence safe.
1228 reg
->smin_value
= reg
->umin_value
= max_t(u64
, reg
->smin_value
,
1230 reg
->smax_value
= reg
->umax_value
;
1234 static void __reg_deduce_bounds(struct bpf_reg_state
*reg
)
1236 __reg32_deduce_bounds(reg
);
1237 __reg64_deduce_bounds(reg
);
1240 /* Attempts to improve var_off based on unsigned min/max information */
1241 static void __reg_bound_offset(struct bpf_reg_state
*reg
)
1243 struct tnum var64_off
= tnum_intersect(reg
->var_off
,
1244 tnum_range(reg
->umin_value
,
1246 struct tnum var32_off
= tnum_intersect(tnum_subreg(reg
->var_off
),
1247 tnum_range(reg
->u32_min_value
,
1248 reg
->u32_max_value
));
1250 reg
->var_off
= tnum_or(tnum_clear_subreg(var64_off
), var32_off
);
1253 static void __reg_assign_32_into_64(struct bpf_reg_state
*reg
)
1255 reg
->umin_value
= reg
->u32_min_value
;
1256 reg
->umax_value
= reg
->u32_max_value
;
1257 /* Attempt to pull 32-bit signed bounds into 64-bit bounds
1258 * but must be positive otherwise set to worse case bounds
1259 * and refine later from tnum.
1261 if (reg
->s32_min_value
>= 0 && reg
->s32_max_value
>= 0)
1262 reg
->smax_value
= reg
->s32_max_value
;
1264 reg
->smax_value
= U32_MAX
;
1265 if (reg
->s32_min_value
>= 0)
1266 reg
->smin_value
= reg
->s32_min_value
;
1268 reg
->smin_value
= 0;
1271 static void __reg_combine_32_into_64(struct bpf_reg_state
*reg
)
1273 /* special case when 64-bit register has upper 32-bit register
1274 * zeroed. Typically happens after zext or <<32, >>32 sequence
1275 * allowing us to use 32-bit bounds directly,
1277 if (tnum_equals_const(tnum_clear_subreg(reg
->var_off
), 0)) {
1278 __reg_assign_32_into_64(reg
);
1280 /* Otherwise the best we can do is push lower 32bit known and
1281 * unknown bits into register (var_off set from jmp logic)
1282 * then learn as much as possible from the 64-bit tnum
1283 * known and unknown bits. The previous smin/smax bounds are
1284 * invalid here because of jmp32 compare so mark them unknown
1285 * so they do not impact tnum bounds calculation.
1287 __mark_reg64_unbounded(reg
);
1288 __update_reg_bounds(reg
);
1291 /* Intersecting with the old var_off might have improved our bounds
1292 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1293 * then new var_off is (0; 0x7f...fc) which improves our umax.
1295 __reg_deduce_bounds(reg
);
1296 __reg_bound_offset(reg
);
1297 __update_reg_bounds(reg
);
1300 static bool __reg64_bound_s32(s64 a
)
1302 return a
> S32_MIN
&& a
< S32_MAX
;
1305 static bool __reg64_bound_u32(u64 a
)
1307 if (a
> U32_MIN
&& a
< U32_MAX
)
1312 static void __reg_combine_64_into_32(struct bpf_reg_state
*reg
)
1314 __mark_reg32_unbounded(reg
);
1316 if (__reg64_bound_s32(reg
->smin_value
) && __reg64_bound_s32(reg
->smax_value
)) {
1317 reg
->s32_min_value
= (s32
)reg
->smin_value
;
1318 reg
->s32_max_value
= (s32
)reg
->smax_value
;
1320 if (__reg64_bound_u32(reg
->umin_value
))
1321 reg
->u32_min_value
= (u32
)reg
->umin_value
;
1322 if (__reg64_bound_u32(reg
->umax_value
))
1323 reg
->u32_max_value
= (u32
)reg
->umax_value
;
1325 /* Intersecting with the old var_off might have improved our bounds
1326 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1327 * then new var_off is (0; 0x7f...fc) which improves our umax.
1329 __reg_deduce_bounds(reg
);
1330 __reg_bound_offset(reg
);
1331 __update_reg_bounds(reg
);
1334 /* Mark a register as having a completely unknown (scalar) value. */
1335 static void __mark_reg_unknown(const struct bpf_verifier_env
*env
,
1336 struct bpf_reg_state
*reg
)
1339 * Clear type, id, off, and union(map_ptr, range) and
1340 * padding between 'type' and union
1342 memset(reg
, 0, offsetof(struct bpf_reg_state
, var_off
));
1343 reg
->type
= SCALAR_VALUE
;
1344 reg
->var_off
= tnum_unknown
;
1346 reg
->precise
= env
->subprog_cnt
> 1 || !env
->bpf_capable
;
1347 __mark_reg_unbounded(reg
);
1350 static void mark_reg_unknown(struct bpf_verifier_env
*env
,
1351 struct bpf_reg_state
*regs
, u32 regno
)
1353 if (WARN_ON(regno
>= MAX_BPF_REG
)) {
1354 verbose(env
, "mark_reg_unknown(regs, %u)\n", regno
);
1355 /* Something bad happened, let's kill all regs except FP */
1356 for (regno
= 0; regno
< BPF_REG_FP
; regno
++)
1357 __mark_reg_not_init(env
, regs
+ regno
);
1360 __mark_reg_unknown(env
, regs
+ regno
);
1363 static void __mark_reg_not_init(const struct bpf_verifier_env
*env
,
1364 struct bpf_reg_state
*reg
)
1366 __mark_reg_unknown(env
, reg
);
1367 reg
->type
= NOT_INIT
;
1370 static void mark_reg_not_init(struct bpf_verifier_env
*env
,
1371 struct bpf_reg_state
*regs
, u32 regno
)
1373 if (WARN_ON(regno
>= MAX_BPF_REG
)) {
1374 verbose(env
, "mark_reg_not_init(regs, %u)\n", regno
);
1375 /* Something bad happened, let's kill all regs except FP */
1376 for (regno
= 0; regno
< BPF_REG_FP
; regno
++)
1377 __mark_reg_not_init(env
, regs
+ regno
);
1380 __mark_reg_not_init(env
, regs
+ regno
);
1383 static void mark_btf_ld_reg(struct bpf_verifier_env
*env
,
1384 struct bpf_reg_state
*regs
, u32 regno
,
1385 enum bpf_reg_type reg_type
,
1386 struct btf
*btf
, u32 btf_id
)
1388 if (reg_type
== SCALAR_VALUE
) {
1389 mark_reg_unknown(env
, regs
, regno
);
1392 mark_reg_known_zero(env
, regs
, regno
);
1393 regs
[regno
].type
= PTR_TO_BTF_ID
;
1394 regs
[regno
].btf
= btf
;
1395 regs
[regno
].btf_id
= btf_id
;
1398 #define DEF_NOT_SUBREG (0)
1399 static void init_reg_state(struct bpf_verifier_env
*env
,
1400 struct bpf_func_state
*state
)
1402 struct bpf_reg_state
*regs
= state
->regs
;
1405 for (i
= 0; i
< MAX_BPF_REG
; i
++) {
1406 mark_reg_not_init(env
, regs
, i
);
1407 regs
[i
].live
= REG_LIVE_NONE
;
1408 regs
[i
].parent
= NULL
;
1409 regs
[i
].subreg_def
= DEF_NOT_SUBREG
;
1413 regs
[BPF_REG_FP
].type
= PTR_TO_STACK
;
1414 mark_reg_known_zero(env
, regs
, BPF_REG_FP
);
1415 regs
[BPF_REG_FP
].frameno
= state
->frameno
;
1418 #define BPF_MAIN_FUNC (-1)
1419 static void init_func_state(struct bpf_verifier_env
*env
,
1420 struct bpf_func_state
*state
,
1421 int callsite
, int frameno
, int subprogno
)
1423 state
->callsite
= callsite
;
1424 state
->frameno
= frameno
;
1425 state
->subprogno
= subprogno
;
1426 init_reg_state(env
, state
);
1430 SRC_OP
, /* register is used as source operand */
1431 DST_OP
, /* register is used as destination operand */
1432 DST_OP_NO_MARK
/* same as above, check only, don't mark */
1435 static int cmp_subprogs(const void *a
, const void *b
)
1437 return ((struct bpf_subprog_info
*)a
)->start
-
1438 ((struct bpf_subprog_info
*)b
)->start
;
1441 static int find_subprog(struct bpf_verifier_env
*env
, int off
)
1443 struct bpf_subprog_info
*p
;
1445 p
= bsearch(&off
, env
->subprog_info
, env
->subprog_cnt
,
1446 sizeof(env
->subprog_info
[0]), cmp_subprogs
);
1449 return p
- env
->subprog_info
;
1453 static int add_subprog(struct bpf_verifier_env
*env
, int off
)
1455 int insn_cnt
= env
->prog
->len
;
1458 if (off
>= insn_cnt
|| off
< 0) {
1459 verbose(env
, "call to invalid destination\n");
1462 ret
= find_subprog(env
, off
);
1465 if (env
->subprog_cnt
>= BPF_MAX_SUBPROGS
) {
1466 verbose(env
, "too many subprograms\n");
1469 env
->subprog_info
[env
->subprog_cnt
++].start
= off
;
1470 sort(env
->subprog_info
, env
->subprog_cnt
,
1471 sizeof(env
->subprog_info
[0]), cmp_subprogs
, NULL
);
1475 static int check_subprogs(struct bpf_verifier_env
*env
)
1477 int i
, ret
, subprog_start
, subprog_end
, off
, cur_subprog
= 0;
1478 struct bpf_subprog_info
*subprog
= env
->subprog_info
;
1479 struct bpf_insn
*insn
= env
->prog
->insnsi
;
1480 int insn_cnt
= env
->prog
->len
;
1482 /* Add entry function. */
1483 ret
= add_subprog(env
, 0);
1487 /* determine subprog starts. The end is one before the next starts */
1488 for (i
= 0; i
< insn_cnt
; i
++) {
1489 if (insn
[i
].code
!= (BPF_JMP
| BPF_CALL
))
1491 if (insn
[i
].src_reg
!= BPF_PSEUDO_CALL
)
1493 if (!env
->bpf_capable
) {
1495 "function calls to other bpf functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
1498 ret
= add_subprog(env
, i
+ insn
[i
].imm
+ 1);
1503 /* Add a fake 'exit' subprog which could simplify subprog iteration
1504 * logic. 'subprog_cnt' should not be increased.
1506 subprog
[env
->subprog_cnt
].start
= insn_cnt
;
1508 if (env
->log
.level
& BPF_LOG_LEVEL2
)
1509 for (i
= 0; i
< env
->subprog_cnt
; i
++)
1510 verbose(env
, "func#%d @%d\n", i
, subprog
[i
].start
);
1512 /* now check that all jumps are within the same subprog */
1513 subprog_start
= subprog
[cur_subprog
].start
;
1514 subprog_end
= subprog
[cur_subprog
+ 1].start
;
1515 for (i
= 0; i
< insn_cnt
; i
++) {
1516 u8 code
= insn
[i
].code
;
1518 if (code
== (BPF_JMP
| BPF_CALL
) &&
1519 insn
[i
].imm
== BPF_FUNC_tail_call
&&
1520 insn
[i
].src_reg
!= BPF_PSEUDO_CALL
)
1521 subprog
[cur_subprog
].has_tail_call
= true;
1522 if (BPF_CLASS(code
) == BPF_LD
&&
1523 (BPF_MODE(code
) == BPF_ABS
|| BPF_MODE(code
) == BPF_IND
))
1524 subprog
[cur_subprog
].has_ld_abs
= true;
1525 if (BPF_CLASS(code
) != BPF_JMP
&& BPF_CLASS(code
) != BPF_JMP32
)
1527 if (BPF_OP(code
) == BPF_EXIT
|| BPF_OP(code
) == BPF_CALL
)
1529 off
= i
+ insn
[i
].off
+ 1;
1530 if (off
< subprog_start
|| off
>= subprog_end
) {
1531 verbose(env
, "jump out of range from insn %d to %d\n", i
, off
);
1535 if (i
== subprog_end
- 1) {
1536 /* to avoid fall-through from one subprog into another
1537 * the last insn of the subprog should be either exit
1538 * or unconditional jump back
1540 if (code
!= (BPF_JMP
| BPF_EXIT
) &&
1541 code
!= (BPF_JMP
| BPF_JA
)) {
1542 verbose(env
, "last insn is not an exit or jmp\n");
1545 subprog_start
= subprog_end
;
1547 if (cur_subprog
< env
->subprog_cnt
)
1548 subprog_end
= subprog
[cur_subprog
+ 1].start
;
1554 /* Parentage chain of this register (or stack slot) should take care of all
1555 * issues like callee-saved registers, stack slot allocation time, etc.
1557 static int mark_reg_read(struct bpf_verifier_env
*env
,
1558 const struct bpf_reg_state
*state
,
1559 struct bpf_reg_state
*parent
, u8 flag
)
1561 bool writes
= parent
== state
->parent
; /* Observe write marks */
1565 /* if read wasn't screened by an earlier write ... */
1566 if (writes
&& state
->live
& REG_LIVE_WRITTEN
)
1568 if (parent
->live
& REG_LIVE_DONE
) {
1569 verbose(env
, "verifier BUG type %s var_off %lld off %d\n",
1570 reg_type_str
[parent
->type
],
1571 parent
->var_off
.value
, parent
->off
);
1574 /* The first condition is more likely to be true than the
1575 * second, checked it first.
1577 if ((parent
->live
& REG_LIVE_READ
) == flag
||
1578 parent
->live
& REG_LIVE_READ64
)
1579 /* The parentage chain never changes and
1580 * this parent was already marked as LIVE_READ.
1581 * There is no need to keep walking the chain again and
1582 * keep re-marking all parents as LIVE_READ.
1583 * This case happens when the same register is read
1584 * multiple times without writes into it in-between.
1585 * Also, if parent has the stronger REG_LIVE_READ64 set,
1586 * then no need to set the weak REG_LIVE_READ32.
1589 /* ... then we depend on parent's value */
1590 parent
->live
|= flag
;
1591 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
1592 if (flag
== REG_LIVE_READ64
)
1593 parent
->live
&= ~REG_LIVE_READ32
;
1595 parent
= state
->parent
;
1600 if (env
->longest_mark_read_walk
< cnt
)
1601 env
->longest_mark_read_walk
= cnt
;
1605 /* This function is supposed to be used by the following 32-bit optimization
1606 * code only. It returns TRUE if the source or destination register operates
1607 * on 64-bit, otherwise return FALSE.
1609 static bool is_reg64(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
,
1610 u32 regno
, struct bpf_reg_state
*reg
, enum reg_arg_type t
)
1615 class = BPF_CLASS(code
);
1617 if (class == BPF_JMP
) {
1618 /* BPF_EXIT for "main" will reach here. Return TRUE
1623 if (op
== BPF_CALL
) {
1624 /* BPF to BPF call will reach here because of marking
1625 * caller saved clobber with DST_OP_NO_MARK for which we
1626 * don't care the register def because they are anyway
1627 * marked as NOT_INIT already.
1629 if (insn
->src_reg
== BPF_PSEUDO_CALL
)
1631 /* Helper call will reach here because of arg type
1632 * check, conservatively return TRUE.
1641 if (class == BPF_ALU64
|| class == BPF_JMP
||
1642 /* BPF_END always use BPF_ALU class. */
1643 (class == BPF_ALU
&& op
== BPF_END
&& insn
->imm
== 64))
1646 if (class == BPF_ALU
|| class == BPF_JMP32
)
1649 if (class == BPF_LDX
) {
1651 return BPF_SIZE(code
) == BPF_DW
;
1652 /* LDX source must be ptr. */
1656 if (class == BPF_STX
) {
1657 if (reg
->type
!= SCALAR_VALUE
)
1659 return BPF_SIZE(code
) == BPF_DW
;
1662 if (class == BPF_LD
) {
1663 u8 mode
= BPF_MODE(code
);
1666 if (mode
== BPF_IMM
)
1669 /* Both LD_IND and LD_ABS return 32-bit data. */
1673 /* Implicit ctx ptr. */
1674 if (regno
== BPF_REG_6
)
1677 /* Explicit source could be any width. */
1681 if (class == BPF_ST
)
1682 /* The only source register for BPF_ST is a ptr. */
1685 /* Conservatively return true at default. */
1689 /* Return TRUE if INSN doesn't have explicit value define. */
1690 static bool insn_no_def(struct bpf_insn
*insn
)
1692 u8
class = BPF_CLASS(insn
->code
);
1694 return (class == BPF_JMP
|| class == BPF_JMP32
||
1695 class == BPF_STX
|| class == BPF_ST
);
1698 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
1699 static bool insn_has_def32(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
1701 if (insn_no_def(insn
))
1704 return !is_reg64(env
, insn
, insn
->dst_reg
, NULL
, DST_OP
);
1707 static void mark_insn_zext(struct bpf_verifier_env
*env
,
1708 struct bpf_reg_state
*reg
)
1710 s32 def_idx
= reg
->subreg_def
;
1712 if (def_idx
== DEF_NOT_SUBREG
)
1715 env
->insn_aux_data
[def_idx
- 1].zext_dst
= true;
1716 /* The dst will be zero extended, so won't be sub-register anymore. */
1717 reg
->subreg_def
= DEF_NOT_SUBREG
;
1720 static int check_reg_arg(struct bpf_verifier_env
*env
, u32 regno
,
1721 enum reg_arg_type t
)
1723 struct bpf_verifier_state
*vstate
= env
->cur_state
;
1724 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
1725 struct bpf_insn
*insn
= env
->prog
->insnsi
+ env
->insn_idx
;
1726 struct bpf_reg_state
*reg
, *regs
= state
->regs
;
1729 if (regno
>= MAX_BPF_REG
) {
1730 verbose(env
, "R%d is invalid\n", regno
);
1735 rw64
= is_reg64(env
, insn
, regno
, reg
, t
);
1737 /* check whether register used as source operand can be read */
1738 if (reg
->type
== NOT_INIT
) {
1739 verbose(env
, "R%d !read_ok\n", regno
);
1742 /* We don't need to worry about FP liveness because it's read-only */
1743 if (regno
== BPF_REG_FP
)
1747 mark_insn_zext(env
, reg
);
1749 return mark_reg_read(env
, reg
, reg
->parent
,
1750 rw64
? REG_LIVE_READ64
: REG_LIVE_READ32
);
1752 /* check whether register used as dest operand can be written to */
1753 if (regno
== BPF_REG_FP
) {
1754 verbose(env
, "frame pointer is read only\n");
1757 reg
->live
|= REG_LIVE_WRITTEN
;
1758 reg
->subreg_def
= rw64
? DEF_NOT_SUBREG
: env
->insn_idx
+ 1;
1760 mark_reg_unknown(env
, regs
, regno
);
1765 /* for any branch, call, exit record the history of jmps in the given state */
1766 static int push_jmp_history(struct bpf_verifier_env
*env
,
1767 struct bpf_verifier_state
*cur
)
1769 u32 cnt
= cur
->jmp_history_cnt
;
1770 struct bpf_idx_pair
*p
;
1773 p
= krealloc(cur
->jmp_history
, cnt
* sizeof(*p
), GFP_USER
);
1776 p
[cnt
- 1].idx
= env
->insn_idx
;
1777 p
[cnt
- 1].prev_idx
= env
->prev_insn_idx
;
1778 cur
->jmp_history
= p
;
1779 cur
->jmp_history_cnt
= cnt
;
1783 /* Backtrack one insn at a time. If idx is not at the top of recorded
1784 * history then previous instruction came from straight line execution.
1786 static int get_prev_insn_idx(struct bpf_verifier_state
*st
, int i
,
1791 if (cnt
&& st
->jmp_history
[cnt
- 1].idx
== i
) {
1792 i
= st
->jmp_history
[cnt
- 1].prev_idx
;
1800 /* For given verifier state backtrack_insn() is called from the last insn to
1801 * the first insn. Its purpose is to compute a bitmask of registers and
1802 * stack slots that needs precision in the parent verifier state.
1804 static int backtrack_insn(struct bpf_verifier_env
*env
, int idx
,
1805 u32
*reg_mask
, u64
*stack_mask
)
1807 const struct bpf_insn_cbs cbs
= {
1808 .cb_print
= verbose
,
1809 .private_data
= env
,
1811 struct bpf_insn
*insn
= env
->prog
->insnsi
+ idx
;
1812 u8
class = BPF_CLASS(insn
->code
);
1813 u8 opcode
= BPF_OP(insn
->code
);
1814 u8 mode
= BPF_MODE(insn
->code
);
1815 u32 dreg
= 1u << insn
->dst_reg
;
1816 u32 sreg
= 1u << insn
->src_reg
;
1819 if (insn
->code
== 0)
1821 if (env
->log
.level
& BPF_LOG_LEVEL
) {
1822 verbose(env
, "regs=%x stack=%llx before ", *reg_mask
, *stack_mask
);
1823 verbose(env
, "%d: ", idx
);
1824 print_bpf_insn(&cbs
, insn
, env
->allow_ptr_leaks
);
1827 if (class == BPF_ALU
|| class == BPF_ALU64
) {
1828 if (!(*reg_mask
& dreg
))
1830 if (opcode
== BPF_MOV
) {
1831 if (BPF_SRC(insn
->code
) == BPF_X
) {
1833 * dreg needs precision after this insn
1834 * sreg needs precision before this insn
1840 * dreg needs precision after this insn.
1841 * Corresponding register is already marked
1842 * as precise=true in this verifier state.
1843 * No further markings in parent are necessary
1848 if (BPF_SRC(insn
->code
) == BPF_X
) {
1850 * both dreg and sreg need precision
1855 * dreg still needs precision before this insn
1858 } else if (class == BPF_LDX
) {
1859 if (!(*reg_mask
& dreg
))
1863 /* scalars can only be spilled into stack w/o losing precision.
1864 * Load from any other memory can be zero extended.
1865 * The desire to keep that precision is already indicated
1866 * by 'precise' mark in corresponding register of this state.
1867 * No further tracking necessary.
1869 if (insn
->src_reg
!= BPF_REG_FP
)
1871 if (BPF_SIZE(insn
->code
) != BPF_DW
)
1874 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
1875 * that [fp - off] slot contains scalar that needs to be
1876 * tracked with precision
1878 spi
= (-insn
->off
- 1) / BPF_REG_SIZE
;
1880 verbose(env
, "BUG spi %d\n", spi
);
1881 WARN_ONCE(1, "verifier backtracking bug");
1884 *stack_mask
|= 1ull << spi
;
1885 } else if (class == BPF_STX
|| class == BPF_ST
) {
1886 if (*reg_mask
& dreg
)
1887 /* stx & st shouldn't be using _scalar_ dst_reg
1888 * to access memory. It means backtracking
1889 * encountered a case of pointer subtraction.
1892 /* scalars can only be spilled into stack */
1893 if (insn
->dst_reg
!= BPF_REG_FP
)
1895 if (BPF_SIZE(insn
->code
) != BPF_DW
)
1897 spi
= (-insn
->off
- 1) / BPF_REG_SIZE
;
1899 verbose(env
, "BUG spi %d\n", spi
);
1900 WARN_ONCE(1, "verifier backtracking bug");
1903 if (!(*stack_mask
& (1ull << spi
)))
1905 *stack_mask
&= ~(1ull << spi
);
1906 if (class == BPF_STX
)
1908 } else if (class == BPF_JMP
|| class == BPF_JMP32
) {
1909 if (opcode
== BPF_CALL
) {
1910 if (insn
->src_reg
== BPF_PSEUDO_CALL
)
1912 /* regular helper call sets R0 */
1914 if (*reg_mask
& 0x3f) {
1915 /* if backtracing was looking for registers R1-R5
1916 * they should have been found already.
1918 verbose(env
, "BUG regs %x\n", *reg_mask
);
1919 WARN_ONCE(1, "verifier backtracking bug");
1922 } else if (opcode
== BPF_EXIT
) {
1925 } else if (class == BPF_LD
) {
1926 if (!(*reg_mask
& dreg
))
1929 /* It's ld_imm64 or ld_abs or ld_ind.
1930 * For ld_imm64 no further tracking of precision
1931 * into parent is necessary
1933 if (mode
== BPF_IND
|| mode
== BPF_ABS
)
1934 /* to be analyzed */
1940 /* the scalar precision tracking algorithm:
1941 * . at the start all registers have precise=false.
1942 * . scalar ranges are tracked as normal through alu and jmp insns.
1943 * . once precise value of the scalar register is used in:
1944 * . ptr + scalar alu
1945 * . if (scalar cond K|scalar)
1946 * . helper_call(.., scalar, ...) where ARG_CONST is expected
1947 * backtrack through the verifier states and mark all registers and
1948 * stack slots with spilled constants that these scalar regisers
1949 * should be precise.
1950 * . during state pruning two registers (or spilled stack slots)
1951 * are equivalent if both are not precise.
1953 * Note the verifier cannot simply walk register parentage chain,
1954 * since many different registers and stack slots could have been
1955 * used to compute single precise scalar.
1957 * The approach of starting with precise=true for all registers and then
1958 * backtrack to mark a register as not precise when the verifier detects
1959 * that program doesn't care about specific value (e.g., when helper
1960 * takes register as ARG_ANYTHING parameter) is not safe.
1962 * It's ok to walk single parentage chain of the verifier states.
1963 * It's possible that this backtracking will go all the way till 1st insn.
1964 * All other branches will be explored for needing precision later.
1966 * The backtracking needs to deal with cases like:
1967 * 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)
1970 * if r5 > 0x79f goto pc+7
1971 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
1974 * call bpf_perf_event_output#25
1975 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
1979 * call foo // uses callee's r6 inside to compute r0
1983 * to track above reg_mask/stack_mask needs to be independent for each frame.
1985 * Also if parent's curframe > frame where backtracking started,
1986 * the verifier need to mark registers in both frames, otherwise callees
1987 * may incorrectly prune callers. This is similar to
1988 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
1990 * For now backtracking falls back into conservative marking.
1992 static void mark_all_scalars_precise(struct bpf_verifier_env
*env
,
1993 struct bpf_verifier_state
*st
)
1995 struct bpf_func_state
*func
;
1996 struct bpf_reg_state
*reg
;
1999 /* big hammer: mark all scalars precise in this path.
2000 * pop_stack may still get !precise scalars.
2002 for (; st
; st
= st
->parent
)
2003 for (i
= 0; i
<= st
->curframe
; i
++) {
2004 func
= st
->frame
[i
];
2005 for (j
= 0; j
< BPF_REG_FP
; j
++) {
2006 reg
= &func
->regs
[j
];
2007 if (reg
->type
!= SCALAR_VALUE
)
2009 reg
->precise
= true;
2011 for (j
= 0; j
< func
->allocated_stack
/ BPF_REG_SIZE
; j
++) {
2012 if (func
->stack
[j
].slot_type
[0] != STACK_SPILL
)
2014 reg
= &func
->stack
[j
].spilled_ptr
;
2015 if (reg
->type
!= SCALAR_VALUE
)
2017 reg
->precise
= true;
2022 static int __mark_chain_precision(struct bpf_verifier_env
*env
, int regno
,
2025 struct bpf_verifier_state
*st
= env
->cur_state
;
2026 int first_idx
= st
->first_insn_idx
;
2027 int last_idx
= env
->insn_idx
;
2028 struct bpf_func_state
*func
;
2029 struct bpf_reg_state
*reg
;
2030 u32 reg_mask
= regno
>= 0 ? 1u << regno
: 0;
2031 u64 stack_mask
= spi
>= 0 ? 1ull << spi
: 0;
2032 bool skip_first
= true;
2033 bool new_marks
= false;
2036 if (!env
->bpf_capable
)
2039 func
= st
->frame
[st
->curframe
];
2041 reg
= &func
->regs
[regno
];
2042 if (reg
->type
!= SCALAR_VALUE
) {
2043 WARN_ONCE(1, "backtracing misuse");
2050 reg
->precise
= true;
2054 if (func
->stack
[spi
].slot_type
[0] != STACK_SPILL
) {
2058 reg
= &func
->stack
[spi
].spilled_ptr
;
2059 if (reg
->type
!= SCALAR_VALUE
) {
2067 reg
->precise
= true;
2073 if (!reg_mask
&& !stack_mask
)
2076 DECLARE_BITMAP(mask
, 64);
2077 u32 history
= st
->jmp_history_cnt
;
2079 if (env
->log
.level
& BPF_LOG_LEVEL
)
2080 verbose(env
, "last_idx %d first_idx %d\n", last_idx
, first_idx
);
2081 for (i
= last_idx
;;) {
2086 err
= backtrack_insn(env
, i
, ®_mask
, &stack_mask
);
2088 if (err
== -ENOTSUPP
) {
2089 mark_all_scalars_precise(env
, st
);
2094 if (!reg_mask
&& !stack_mask
)
2095 /* Found assignment(s) into tracked register in this state.
2096 * Since this state is already marked, just return.
2097 * Nothing to be tracked further in the parent state.
2102 i
= get_prev_insn_idx(st
, i
, &history
);
2103 if (i
>= env
->prog
->len
) {
2104 /* This can happen if backtracking reached insn 0
2105 * and there are still reg_mask or stack_mask
2107 * It means the backtracking missed the spot where
2108 * particular register was initialized with a constant.
2110 verbose(env
, "BUG backtracking idx %d\n", i
);
2111 WARN_ONCE(1, "verifier backtracking bug");
2120 func
= st
->frame
[st
->curframe
];
2121 bitmap_from_u64(mask
, reg_mask
);
2122 for_each_set_bit(i
, mask
, 32) {
2123 reg
= &func
->regs
[i
];
2124 if (reg
->type
!= SCALAR_VALUE
) {
2125 reg_mask
&= ~(1u << i
);
2130 reg
->precise
= true;
2133 bitmap_from_u64(mask
, stack_mask
);
2134 for_each_set_bit(i
, mask
, 64) {
2135 if (i
>= func
->allocated_stack
/ BPF_REG_SIZE
) {
2136 /* the sequence of instructions:
2138 * 3: (7b) *(u64 *)(r3 -8) = r0
2139 * 4: (79) r4 = *(u64 *)(r10 -8)
2140 * doesn't contain jmps. It's backtracked
2141 * as a single block.
2142 * During backtracking insn 3 is not recognized as
2143 * stack access, so at the end of backtracking
2144 * stack slot fp-8 is still marked in stack_mask.
2145 * However the parent state may not have accessed
2146 * fp-8 and it's "unallocated" stack space.
2147 * In such case fallback to conservative.
2149 mark_all_scalars_precise(env
, st
);
2153 if (func
->stack
[i
].slot_type
[0] != STACK_SPILL
) {
2154 stack_mask
&= ~(1ull << i
);
2157 reg
= &func
->stack
[i
].spilled_ptr
;
2158 if (reg
->type
!= SCALAR_VALUE
) {
2159 stack_mask
&= ~(1ull << i
);
2164 reg
->precise
= true;
2166 if (env
->log
.level
& BPF_LOG_LEVEL
) {
2167 print_verifier_state(env
, func
);
2168 verbose(env
, "parent %s regs=%x stack=%llx marks\n",
2169 new_marks
? "didn't have" : "already had",
2170 reg_mask
, stack_mask
);
2173 if (!reg_mask
&& !stack_mask
)
2178 last_idx
= st
->last_insn_idx
;
2179 first_idx
= st
->first_insn_idx
;
2184 static int mark_chain_precision(struct bpf_verifier_env
*env
, int regno
)
2186 return __mark_chain_precision(env
, regno
, -1);
2189 static int mark_chain_precision_stack(struct bpf_verifier_env
*env
, int spi
)
2191 return __mark_chain_precision(env
, -1, spi
);
2194 static bool is_spillable_regtype(enum bpf_reg_type type
)
2197 case PTR_TO_MAP_VALUE
:
2198 case PTR_TO_MAP_VALUE_OR_NULL
:
2202 case PTR_TO_PACKET_META
:
2203 case PTR_TO_PACKET_END
:
2204 case PTR_TO_FLOW_KEYS
:
2205 case CONST_PTR_TO_MAP
:
2207 case PTR_TO_SOCKET_OR_NULL
:
2208 case PTR_TO_SOCK_COMMON
:
2209 case PTR_TO_SOCK_COMMON_OR_NULL
:
2210 case PTR_TO_TCP_SOCK
:
2211 case PTR_TO_TCP_SOCK_OR_NULL
:
2212 case PTR_TO_XDP_SOCK
:
2214 case PTR_TO_BTF_ID_OR_NULL
:
2215 case PTR_TO_RDONLY_BUF
:
2216 case PTR_TO_RDONLY_BUF_OR_NULL
:
2217 case PTR_TO_RDWR_BUF
:
2218 case PTR_TO_RDWR_BUF_OR_NULL
:
2219 case PTR_TO_PERCPU_BTF_ID
:
2221 case PTR_TO_MEM_OR_NULL
:
2228 /* Does this register contain a constant zero? */
2229 static bool register_is_null(struct bpf_reg_state
*reg
)
2231 return reg
->type
== SCALAR_VALUE
&& tnum_equals_const(reg
->var_off
, 0);
2234 static bool register_is_const(struct bpf_reg_state
*reg
)
2236 return reg
->type
== SCALAR_VALUE
&& tnum_is_const(reg
->var_off
);
2239 static bool __is_scalar_unbounded(struct bpf_reg_state
*reg
)
2241 return tnum_is_unknown(reg
->var_off
) &&
2242 reg
->smin_value
== S64_MIN
&& reg
->smax_value
== S64_MAX
&&
2243 reg
->umin_value
== 0 && reg
->umax_value
== U64_MAX
&&
2244 reg
->s32_min_value
== S32_MIN
&& reg
->s32_max_value
== S32_MAX
&&
2245 reg
->u32_min_value
== 0 && reg
->u32_max_value
== U32_MAX
;
2248 static bool register_is_bounded(struct bpf_reg_state
*reg
)
2250 return reg
->type
== SCALAR_VALUE
&& !__is_scalar_unbounded(reg
);
2253 static bool __is_pointer_value(bool allow_ptr_leaks
,
2254 const struct bpf_reg_state
*reg
)
2256 if (allow_ptr_leaks
)
2259 return reg
->type
!= SCALAR_VALUE
;
2262 static void save_register_state(struct bpf_func_state
*state
,
2263 int spi
, struct bpf_reg_state
*reg
)
2267 state
->stack
[spi
].spilled_ptr
= *reg
;
2268 state
->stack
[spi
].spilled_ptr
.live
|= REG_LIVE_WRITTEN
;
2270 for (i
= 0; i
< BPF_REG_SIZE
; i
++)
2271 state
->stack
[spi
].slot_type
[i
] = STACK_SPILL
;
2274 /* check_stack_read/write functions track spill/fill of registers,
2275 * stack boundary and alignment are checked in check_mem_access()
2277 static int check_stack_write(struct bpf_verifier_env
*env
,
2278 struct bpf_func_state
*state
, /* func where register points to */
2279 int off
, int size
, int value_regno
, int insn_idx
)
2281 struct bpf_func_state
*cur
; /* state of the current function */
2282 int i
, slot
= -off
- 1, spi
= slot
/ BPF_REG_SIZE
, err
;
2283 u32 dst_reg
= env
->prog
->insnsi
[insn_idx
].dst_reg
;
2284 struct bpf_reg_state
*reg
= NULL
;
2286 err
= realloc_func_state(state
, round_up(slot
+ 1, BPF_REG_SIZE
),
2287 state
->acquired_refs
, true);
2290 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
2291 * so it's aligned access and [off, off + size) are within stack limits
2293 if (!env
->allow_ptr_leaks
&&
2294 state
->stack
[spi
].slot_type
[0] == STACK_SPILL
&&
2295 size
!= BPF_REG_SIZE
) {
2296 verbose(env
, "attempt to corrupt spilled pointer on stack\n");
2300 cur
= env
->cur_state
->frame
[env
->cur_state
->curframe
];
2301 if (value_regno
>= 0)
2302 reg
= &cur
->regs
[value_regno
];
2304 if (reg
&& size
== BPF_REG_SIZE
&& register_is_bounded(reg
) &&
2305 !register_is_null(reg
) && env
->bpf_capable
) {
2306 if (dst_reg
!= BPF_REG_FP
) {
2307 /* The backtracking logic can only recognize explicit
2308 * stack slot address like [fp - 8]. Other spill of
2309 * scalar via different register has to be conervative.
2310 * Backtrack from here and mark all registers as precise
2311 * that contributed into 'reg' being a constant.
2313 err
= mark_chain_precision(env
, value_regno
);
2317 save_register_state(state
, spi
, reg
);
2318 } else if (reg
&& is_spillable_regtype(reg
->type
)) {
2319 /* register containing pointer is being spilled into stack */
2320 if (size
!= BPF_REG_SIZE
) {
2321 verbose_linfo(env
, insn_idx
, "; ");
2322 verbose(env
, "invalid size of register spill\n");
2326 if (state
!= cur
&& reg
->type
== PTR_TO_STACK
) {
2327 verbose(env
, "cannot spill pointers to stack into stack frame of the caller\n");
2331 if (!env
->bypass_spec_v4
) {
2332 bool sanitize
= false;
2334 if (state
->stack
[spi
].slot_type
[0] == STACK_SPILL
&&
2335 register_is_const(&state
->stack
[spi
].spilled_ptr
))
2337 for (i
= 0; i
< BPF_REG_SIZE
; i
++)
2338 if (state
->stack
[spi
].slot_type
[i
] == STACK_MISC
) {
2343 int *poff
= &env
->insn_aux_data
[insn_idx
].sanitize_stack_off
;
2344 int soff
= (-spi
- 1) * BPF_REG_SIZE
;
2346 /* detected reuse of integer stack slot with a pointer
2347 * which means either llvm is reusing stack slot or
2348 * an attacker is trying to exploit CVE-2018-3639
2349 * (speculative store bypass)
2350 * Have to sanitize that slot with preemptive
2353 if (*poff
&& *poff
!= soff
) {
2354 /* disallow programs where single insn stores
2355 * into two different stack slots, since verifier
2356 * cannot sanitize them
2359 "insn %d cannot access two stack slots fp%d and fp%d",
2360 insn_idx
, *poff
, soff
);
2366 save_register_state(state
, spi
, reg
);
2368 u8 type
= STACK_MISC
;
2370 /* regular write of data into stack destroys any spilled ptr */
2371 state
->stack
[spi
].spilled_ptr
.type
= NOT_INIT
;
2372 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
2373 if (state
->stack
[spi
].slot_type
[0] == STACK_SPILL
)
2374 for (i
= 0; i
< BPF_REG_SIZE
; i
++)
2375 state
->stack
[spi
].slot_type
[i
] = STACK_MISC
;
2377 /* only mark the slot as written if all 8 bytes were written
2378 * otherwise read propagation may incorrectly stop too soon
2379 * when stack slots are partially written.
2380 * This heuristic means that read propagation will be
2381 * conservative, since it will add reg_live_read marks
2382 * to stack slots all the way to first state when programs
2383 * writes+reads less than 8 bytes
2385 if (size
== BPF_REG_SIZE
)
2386 state
->stack
[spi
].spilled_ptr
.live
|= REG_LIVE_WRITTEN
;
2388 /* when we zero initialize stack slots mark them as such */
2389 if (reg
&& register_is_null(reg
)) {
2390 /* backtracking doesn't work for STACK_ZERO yet. */
2391 err
= mark_chain_precision(env
, value_regno
);
2397 /* Mark slots affected by this stack write. */
2398 for (i
= 0; i
< size
; i
++)
2399 state
->stack
[spi
].slot_type
[(slot
- i
) % BPF_REG_SIZE
] =
2405 static int check_stack_read(struct bpf_verifier_env
*env
,
2406 struct bpf_func_state
*reg_state
/* func where register points to */,
2407 int off
, int size
, int value_regno
)
2409 struct bpf_verifier_state
*vstate
= env
->cur_state
;
2410 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
2411 int i
, slot
= -off
- 1, spi
= slot
/ BPF_REG_SIZE
;
2412 struct bpf_reg_state
*reg
;
2415 if (reg_state
->allocated_stack
<= slot
) {
2416 verbose(env
, "invalid read from stack off %d+0 size %d\n",
2420 stype
= reg_state
->stack
[spi
].slot_type
;
2421 reg
= ®_state
->stack
[spi
].spilled_ptr
;
2423 if (stype
[0] == STACK_SPILL
) {
2424 if (size
!= BPF_REG_SIZE
) {
2425 if (reg
->type
!= SCALAR_VALUE
) {
2426 verbose_linfo(env
, env
->insn_idx
, "; ");
2427 verbose(env
, "invalid size of register fill\n");
2430 if (value_regno
>= 0) {
2431 mark_reg_unknown(env
, state
->regs
, value_regno
);
2432 state
->regs
[value_regno
].live
|= REG_LIVE_WRITTEN
;
2434 mark_reg_read(env
, reg
, reg
->parent
, REG_LIVE_READ64
);
2437 for (i
= 1; i
< BPF_REG_SIZE
; i
++) {
2438 if (stype
[(slot
- i
) % BPF_REG_SIZE
] != STACK_SPILL
) {
2439 verbose(env
, "corrupted spill memory\n");
2444 if (value_regno
>= 0) {
2445 /* restore register state from stack */
2446 state
->regs
[value_regno
] = *reg
;
2447 /* mark reg as written since spilled pointer state likely
2448 * has its liveness marks cleared by is_state_visited()
2449 * which resets stack/reg liveness for state transitions
2451 state
->regs
[value_regno
].live
|= REG_LIVE_WRITTEN
;
2452 } else if (__is_pointer_value(env
->allow_ptr_leaks
, reg
)) {
2453 /* If value_regno==-1, the caller is asking us whether
2454 * it is acceptable to use this value as a SCALAR_VALUE
2456 * We must not allow unprivileged callers to do that
2457 * with spilled pointers.
2459 verbose(env
, "leaking pointer from stack off %d\n",
2463 mark_reg_read(env
, reg
, reg
->parent
, REG_LIVE_READ64
);
2467 for (i
= 0; i
< size
; i
++) {
2468 if (stype
[(slot
- i
) % BPF_REG_SIZE
] == STACK_MISC
)
2470 if (stype
[(slot
- i
) % BPF_REG_SIZE
] == STACK_ZERO
) {
2474 verbose(env
, "invalid read from stack off %d+%d size %d\n",
2478 mark_reg_read(env
, reg
, reg
->parent
, REG_LIVE_READ64
);
2479 if (value_regno
>= 0) {
2480 if (zeros
== size
) {
2481 /* any size read into register is zero extended,
2482 * so the whole register == const_zero
2484 __mark_reg_const_zero(&state
->regs
[value_regno
]);
2485 /* backtracking doesn't support STACK_ZERO yet,
2486 * so mark it precise here, so that later
2487 * backtracking can stop here.
2488 * Backtracking may not need this if this register
2489 * doesn't participate in pointer adjustment.
2490 * Forward propagation of precise flag is not
2491 * necessary either. This mark is only to stop
2492 * backtracking. Any register that contributed
2493 * to const 0 was marked precise before spill.
2495 state
->regs
[value_regno
].precise
= true;
2497 /* have read misc data from the stack */
2498 mark_reg_unknown(env
, state
->regs
, value_regno
);
2500 state
->regs
[value_regno
].live
|= REG_LIVE_WRITTEN
;
2506 static int check_stack_access(struct bpf_verifier_env
*env
,
2507 const struct bpf_reg_state
*reg
,
2510 /* Stack accesses must be at a fixed offset, so that we
2511 * can determine what type of data were returned. See
2512 * check_stack_read().
2514 if (!tnum_is_const(reg
->var_off
)) {
2517 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
2518 verbose(env
, "variable stack access var_off=%s off=%d size=%d\n",
2523 if (off
>= 0 || off
< -MAX_BPF_STACK
) {
2524 verbose(env
, "invalid stack off=%d size=%d\n", off
, size
);
2531 static int check_map_access_type(struct bpf_verifier_env
*env
, u32 regno
,
2532 int off
, int size
, enum bpf_access_type type
)
2534 struct bpf_reg_state
*regs
= cur_regs(env
);
2535 struct bpf_map
*map
= regs
[regno
].map_ptr
;
2536 u32 cap
= bpf_map_flags_to_cap(map
);
2538 if (type
== BPF_WRITE
&& !(cap
& BPF_MAP_CAN_WRITE
)) {
2539 verbose(env
, "write into map forbidden, value_size=%d off=%d size=%d\n",
2540 map
->value_size
, off
, size
);
2544 if (type
== BPF_READ
&& !(cap
& BPF_MAP_CAN_READ
)) {
2545 verbose(env
, "read from map forbidden, value_size=%d off=%d size=%d\n",
2546 map
->value_size
, off
, size
);
2553 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
2554 static int __check_mem_access(struct bpf_verifier_env
*env
, int regno
,
2555 int off
, int size
, u32 mem_size
,
2556 bool zero_size_allowed
)
2558 bool size_ok
= size
> 0 || (size
== 0 && zero_size_allowed
);
2559 struct bpf_reg_state
*reg
;
2561 if (off
>= 0 && size_ok
&& (u64
)off
+ size
<= mem_size
)
2564 reg
= &cur_regs(env
)[regno
];
2565 switch (reg
->type
) {
2566 case PTR_TO_MAP_VALUE
:
2567 verbose(env
, "invalid access to map value, value_size=%d off=%d size=%d\n",
2568 mem_size
, off
, size
);
2571 case PTR_TO_PACKET_META
:
2572 case PTR_TO_PACKET_END
:
2573 verbose(env
, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
2574 off
, size
, regno
, reg
->id
, off
, mem_size
);
2578 verbose(env
, "invalid access to memory, mem_size=%u off=%d size=%d\n",
2579 mem_size
, off
, size
);
2585 /* check read/write into a memory region with possible variable offset */
2586 static int check_mem_region_access(struct bpf_verifier_env
*env
, u32 regno
,
2587 int off
, int size
, u32 mem_size
,
2588 bool zero_size_allowed
)
2590 struct bpf_verifier_state
*vstate
= env
->cur_state
;
2591 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
2592 struct bpf_reg_state
*reg
= &state
->regs
[regno
];
2595 /* We may have adjusted the register pointing to memory region, so we
2596 * need to try adding each of min_value and max_value to off
2597 * to make sure our theoretical access will be safe.
2599 if (env
->log
.level
& BPF_LOG_LEVEL
)
2600 print_verifier_state(env
, state
);
2602 /* The minimum value is only important with signed
2603 * comparisons where we can't assume the floor of a
2604 * value is 0. If we are using signed variables for our
2605 * index'es we need to make sure that whatever we use
2606 * will have a set floor within our range.
2608 if (reg
->smin_value
< 0 &&
2609 (reg
->smin_value
== S64_MIN
||
2610 (off
+ reg
->smin_value
!= (s64
)(s32
)(off
+ reg
->smin_value
)) ||
2611 reg
->smin_value
+ off
< 0)) {
2612 verbose(env
, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2616 err
= __check_mem_access(env
, regno
, reg
->smin_value
+ off
, size
,
2617 mem_size
, zero_size_allowed
);
2619 verbose(env
, "R%d min value is outside of the allowed memory range\n",
2624 /* If we haven't set a max value then we need to bail since we can't be
2625 * sure we won't do bad things.
2626 * If reg->umax_value + off could overflow, treat that as unbounded too.
2628 if (reg
->umax_value
>= BPF_MAX_VAR_OFF
) {
2629 verbose(env
, "R%d unbounded memory access, make sure to bounds check any such access\n",
2633 err
= __check_mem_access(env
, regno
, reg
->umax_value
+ off
, size
,
2634 mem_size
, zero_size_allowed
);
2636 verbose(env
, "R%d max value is outside of the allowed memory range\n",
2644 /* check read/write into a map element with possible variable offset */
2645 static int check_map_access(struct bpf_verifier_env
*env
, u32 regno
,
2646 int off
, int size
, bool zero_size_allowed
)
2648 struct bpf_verifier_state
*vstate
= env
->cur_state
;
2649 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
2650 struct bpf_reg_state
*reg
= &state
->regs
[regno
];
2651 struct bpf_map
*map
= reg
->map_ptr
;
2654 err
= check_mem_region_access(env
, regno
, off
, size
, map
->value_size
,
2659 if (map_value_has_spin_lock(map
)) {
2660 u32 lock
= map
->spin_lock_off
;
2662 /* if any part of struct bpf_spin_lock can be touched by
2663 * load/store reject this program.
2664 * To check that [x1, x2) overlaps with [y1, y2)
2665 * it is sufficient to check x1 < y2 && y1 < x2.
2667 if (reg
->smin_value
+ off
< lock
+ sizeof(struct bpf_spin_lock
) &&
2668 lock
< reg
->umax_value
+ off
+ size
) {
2669 verbose(env
, "bpf_spin_lock cannot be accessed directly by load/store\n");
2676 #define MAX_PACKET_OFF 0xffff
2678 static enum bpf_prog_type
resolve_prog_type(struct bpf_prog
*prog
)
2680 return prog
->aux
->dst_prog
? prog
->aux
->dst_prog
->type
: prog
->type
;
2683 static bool may_access_direct_pkt_data(struct bpf_verifier_env
*env
,
2684 const struct bpf_call_arg_meta
*meta
,
2685 enum bpf_access_type t
)
2687 enum bpf_prog_type prog_type
= resolve_prog_type(env
->prog
);
2689 switch (prog_type
) {
2690 /* Program types only with direct read access go here! */
2691 case BPF_PROG_TYPE_LWT_IN
:
2692 case BPF_PROG_TYPE_LWT_OUT
:
2693 case BPF_PROG_TYPE_LWT_SEG6LOCAL
:
2694 case BPF_PROG_TYPE_SK_REUSEPORT
:
2695 case BPF_PROG_TYPE_FLOW_DISSECTOR
:
2696 case BPF_PROG_TYPE_CGROUP_SKB
:
2701 /* Program types with direct read + write access go here! */
2702 case BPF_PROG_TYPE_SCHED_CLS
:
2703 case BPF_PROG_TYPE_SCHED_ACT
:
2704 case BPF_PROG_TYPE_XDP
:
2705 case BPF_PROG_TYPE_LWT_XMIT
:
2706 case BPF_PROG_TYPE_SK_SKB
:
2707 case BPF_PROG_TYPE_SK_MSG
:
2709 return meta
->pkt_access
;
2711 env
->seen_direct_write
= true;
2714 case BPF_PROG_TYPE_CGROUP_SOCKOPT
:
2716 env
->seen_direct_write
= true;
2725 static int check_packet_access(struct bpf_verifier_env
*env
, u32 regno
, int off
,
2726 int size
, bool zero_size_allowed
)
2728 struct bpf_reg_state
*regs
= cur_regs(env
);
2729 struct bpf_reg_state
*reg
= ®s
[regno
];
2732 /* We may have added a variable offset to the packet pointer; but any
2733 * reg->range we have comes after that. We are only checking the fixed
2737 /* We don't allow negative numbers, because we aren't tracking enough
2738 * detail to prove they're safe.
2740 if (reg
->smin_value
< 0) {
2741 verbose(env
, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2746 err
= reg
->range
< 0 ? -EINVAL
:
2747 __check_mem_access(env
, regno
, off
, size
, reg
->range
,
2750 verbose(env
, "R%d offset is outside of the packet\n", regno
);
2754 /* __check_mem_access has made sure "off + size - 1" is within u16.
2755 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
2756 * otherwise find_good_pkt_pointers would have refused to set range info
2757 * that __check_mem_access would have rejected this pkt access.
2758 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
2760 env
->prog
->aux
->max_pkt_offset
=
2761 max_t(u32
, env
->prog
->aux
->max_pkt_offset
,
2762 off
+ reg
->umax_value
+ size
- 1);
2767 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
2768 static int check_ctx_access(struct bpf_verifier_env
*env
, int insn_idx
, int off
, int size
,
2769 enum bpf_access_type t
, enum bpf_reg_type
*reg_type
,
2770 struct btf
**btf
, u32
*btf_id
)
2772 struct bpf_insn_access_aux info
= {
2773 .reg_type
= *reg_type
,
2777 if (env
->ops
->is_valid_access
&&
2778 env
->ops
->is_valid_access(off
, size
, t
, env
->prog
, &info
)) {
2779 /* A non zero info.ctx_field_size indicates that this field is a
2780 * candidate for later verifier transformation to load the whole
2781 * field and then apply a mask when accessed with a narrower
2782 * access than actual ctx access size. A zero info.ctx_field_size
2783 * will only allow for whole field access and rejects any other
2784 * type of narrower access.
2786 *reg_type
= info
.reg_type
;
2788 if (*reg_type
== PTR_TO_BTF_ID
|| *reg_type
== PTR_TO_BTF_ID_OR_NULL
) {
2790 *btf_id
= info
.btf_id
;
2792 env
->insn_aux_data
[insn_idx
].ctx_field_size
= info
.ctx_field_size
;
2794 /* remember the offset of last byte accessed in ctx */
2795 if (env
->prog
->aux
->max_ctx_offset
< off
+ size
)
2796 env
->prog
->aux
->max_ctx_offset
= off
+ size
;
2800 verbose(env
, "invalid bpf_context access off=%d size=%d\n", off
, size
);
2804 static int check_flow_keys_access(struct bpf_verifier_env
*env
, int off
,
2807 if (size
< 0 || off
< 0 ||
2808 (u64
)off
+ size
> sizeof(struct bpf_flow_keys
)) {
2809 verbose(env
, "invalid access to flow keys off=%d size=%d\n",
2816 static int check_sock_access(struct bpf_verifier_env
*env
, int insn_idx
,
2817 u32 regno
, int off
, int size
,
2818 enum bpf_access_type t
)
2820 struct bpf_reg_state
*regs
= cur_regs(env
);
2821 struct bpf_reg_state
*reg
= ®s
[regno
];
2822 struct bpf_insn_access_aux info
= {};
2825 if (reg
->smin_value
< 0) {
2826 verbose(env
, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2831 switch (reg
->type
) {
2832 case PTR_TO_SOCK_COMMON
:
2833 valid
= bpf_sock_common_is_valid_access(off
, size
, t
, &info
);
2836 valid
= bpf_sock_is_valid_access(off
, size
, t
, &info
);
2838 case PTR_TO_TCP_SOCK
:
2839 valid
= bpf_tcp_sock_is_valid_access(off
, size
, t
, &info
);
2841 case PTR_TO_XDP_SOCK
:
2842 valid
= bpf_xdp_sock_is_valid_access(off
, size
, t
, &info
);
2850 env
->insn_aux_data
[insn_idx
].ctx_field_size
=
2851 info
.ctx_field_size
;
2855 verbose(env
, "R%d invalid %s access off=%d size=%d\n",
2856 regno
, reg_type_str
[reg
->type
], off
, size
);
2861 static struct bpf_reg_state
*reg_state(struct bpf_verifier_env
*env
, int regno
)
2863 return cur_regs(env
) + regno
;
2866 static bool is_pointer_value(struct bpf_verifier_env
*env
, int regno
)
2868 return __is_pointer_value(env
->allow_ptr_leaks
, reg_state(env
, regno
));
2871 static bool is_ctx_reg(struct bpf_verifier_env
*env
, int regno
)
2873 const struct bpf_reg_state
*reg
= reg_state(env
, regno
);
2875 return reg
->type
== PTR_TO_CTX
;
2878 static bool is_sk_reg(struct bpf_verifier_env
*env
, int regno
)
2880 const struct bpf_reg_state
*reg
= reg_state(env
, regno
);
2882 return type_is_sk_pointer(reg
->type
);
2885 static bool is_pkt_reg(struct bpf_verifier_env
*env
, int regno
)
2887 const struct bpf_reg_state
*reg
= reg_state(env
, regno
);
2889 return type_is_pkt_pointer(reg
->type
);
2892 static bool is_flow_key_reg(struct bpf_verifier_env
*env
, int regno
)
2894 const struct bpf_reg_state
*reg
= reg_state(env
, regno
);
2896 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
2897 return reg
->type
== PTR_TO_FLOW_KEYS
;
2900 static int check_pkt_ptr_alignment(struct bpf_verifier_env
*env
,
2901 const struct bpf_reg_state
*reg
,
2902 int off
, int size
, bool strict
)
2904 struct tnum reg_off
;
2907 /* Byte size accesses are always allowed. */
2908 if (!strict
|| size
== 1)
2911 /* For platforms that do not have a Kconfig enabling
2912 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
2913 * NET_IP_ALIGN is universally set to '2'. And on platforms
2914 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
2915 * to this code only in strict mode where we want to emulate
2916 * the NET_IP_ALIGN==2 checking. Therefore use an
2917 * unconditional IP align value of '2'.
2921 reg_off
= tnum_add(reg
->var_off
, tnum_const(ip_align
+ reg
->off
+ off
));
2922 if (!tnum_is_aligned(reg_off
, size
)) {
2925 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
2927 "misaligned packet access off %d+%s+%d+%d size %d\n",
2928 ip_align
, tn_buf
, reg
->off
, off
, size
);
2935 static int check_generic_ptr_alignment(struct bpf_verifier_env
*env
,
2936 const struct bpf_reg_state
*reg
,
2937 const char *pointer_desc
,
2938 int off
, int size
, bool strict
)
2940 struct tnum reg_off
;
2942 /* Byte size accesses are always allowed. */
2943 if (!strict
|| size
== 1)
2946 reg_off
= tnum_add(reg
->var_off
, tnum_const(reg
->off
+ off
));
2947 if (!tnum_is_aligned(reg_off
, size
)) {
2950 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
2951 verbose(env
, "misaligned %saccess off %s+%d+%d size %d\n",
2952 pointer_desc
, tn_buf
, reg
->off
, off
, size
);
2959 static int check_ptr_alignment(struct bpf_verifier_env
*env
,
2960 const struct bpf_reg_state
*reg
, int off
,
2961 int size
, bool strict_alignment_once
)
2963 bool strict
= env
->strict_alignment
|| strict_alignment_once
;
2964 const char *pointer_desc
= "";
2966 switch (reg
->type
) {
2968 case PTR_TO_PACKET_META
:
2969 /* Special case, because of NET_IP_ALIGN. Given metadata sits
2970 * right in front, treat it the very same way.
2972 return check_pkt_ptr_alignment(env
, reg
, off
, size
, strict
);
2973 case PTR_TO_FLOW_KEYS
:
2974 pointer_desc
= "flow keys ";
2976 case PTR_TO_MAP_VALUE
:
2977 pointer_desc
= "value ";
2980 pointer_desc
= "context ";
2983 pointer_desc
= "stack ";
2984 /* The stack spill tracking logic in check_stack_write()
2985 * and check_stack_read() relies on stack accesses being
2991 pointer_desc
= "sock ";
2993 case PTR_TO_SOCK_COMMON
:
2994 pointer_desc
= "sock_common ";
2996 case PTR_TO_TCP_SOCK
:
2997 pointer_desc
= "tcp_sock ";
2999 case PTR_TO_XDP_SOCK
:
3000 pointer_desc
= "xdp_sock ";
3005 return check_generic_ptr_alignment(env
, reg
, pointer_desc
, off
, size
,
3009 static int update_stack_depth(struct bpf_verifier_env
*env
,
3010 const struct bpf_func_state
*func
,
3013 u16 stack
= env
->subprog_info
[func
->subprogno
].stack_depth
;
3018 /* update known max for given subprogram */
3019 env
->subprog_info
[func
->subprogno
].stack_depth
= -off
;
3023 /* starting from main bpf function walk all instructions of the function
3024 * and recursively walk all callees that given function can call.
3025 * Ignore jump and exit insns.
3026 * Since recursion is prevented by check_cfg() this algorithm
3027 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
3029 static int check_max_stack_depth(struct bpf_verifier_env
*env
)
3031 int depth
= 0, frame
= 0, idx
= 0, i
= 0, subprog_end
;
3032 struct bpf_subprog_info
*subprog
= env
->subprog_info
;
3033 struct bpf_insn
*insn
= env
->prog
->insnsi
;
3034 bool tail_call_reachable
= false;
3035 int ret_insn
[MAX_CALL_FRAMES
];
3036 int ret_prog
[MAX_CALL_FRAMES
];
3040 /* protect against potential stack overflow that might happen when
3041 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
3042 * depth for such case down to 256 so that the worst case scenario
3043 * would result in 8k stack size (32 which is tailcall limit * 256 =
3046 * To get the idea what might happen, see an example:
3047 * func1 -> sub rsp, 128
3048 * subfunc1 -> sub rsp, 256
3049 * tailcall1 -> add rsp, 256
3050 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
3051 * subfunc2 -> sub rsp, 64
3052 * subfunc22 -> sub rsp, 128
3053 * tailcall2 -> add rsp, 128
3054 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
3056 * tailcall will unwind the current stack frame but it will not get rid
3057 * of caller's stack as shown on the example above.
3059 if (idx
&& subprog
[idx
].has_tail_call
&& depth
>= 256) {
3061 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
3065 /* round up to 32-bytes, since this is granularity
3066 * of interpreter stack size
3068 depth
+= round_up(max_t(u32
, subprog
[idx
].stack_depth
, 1), 32);
3069 if (depth
> MAX_BPF_STACK
) {
3070 verbose(env
, "combined stack size of %d calls is %d. Too large\n",
3075 subprog_end
= subprog
[idx
+ 1].start
;
3076 for (; i
< subprog_end
; i
++) {
3077 if (insn
[i
].code
!= (BPF_JMP
| BPF_CALL
))
3079 if (insn
[i
].src_reg
!= BPF_PSEUDO_CALL
)
3081 /* remember insn and function to return to */
3082 ret_insn
[frame
] = i
+ 1;
3083 ret_prog
[frame
] = idx
;
3085 /* find the callee */
3086 i
= i
+ insn
[i
].imm
+ 1;
3087 idx
= find_subprog(env
, i
);
3089 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3094 if (subprog
[idx
].has_tail_call
)
3095 tail_call_reachable
= true;
3098 if (frame
>= MAX_CALL_FRAMES
) {
3099 verbose(env
, "the call stack of %d frames is too deep !\n",
3105 /* if tail call got detected across bpf2bpf calls then mark each of the
3106 * currently present subprog frames as tail call reachable subprogs;
3107 * this info will be utilized by JIT so that we will be preserving the
3108 * tail call counter throughout bpf2bpf calls combined with tailcalls
3110 if (tail_call_reachable
)
3111 for (j
= 0; j
< frame
; j
++)
3112 subprog
[ret_prog
[j
]].tail_call_reachable
= true;
3114 /* end of for() loop means the last insn of the 'subprog'
3115 * was reached. Doesn't matter whether it was JA or EXIT
3119 depth
-= round_up(max_t(u32
, subprog
[idx
].stack_depth
, 1), 32);
3121 i
= ret_insn
[frame
];
3122 idx
= ret_prog
[frame
];
3126 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
3127 static int get_callee_stack_depth(struct bpf_verifier_env
*env
,
3128 const struct bpf_insn
*insn
, int idx
)
3130 int start
= idx
+ insn
->imm
+ 1, subprog
;
3132 subprog
= find_subprog(env
, start
);
3134 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3138 return env
->subprog_info
[subprog
].stack_depth
;
3142 int check_ctx_reg(struct bpf_verifier_env
*env
,
3143 const struct bpf_reg_state
*reg
, int regno
)
3145 /* Access to ctx or passing it to a helper is only allowed in
3146 * its original, unmodified form.
3150 verbose(env
, "dereference of modified ctx ptr R%d off=%d disallowed\n",
3155 if (!tnum_is_const(reg
->var_off
) || reg
->var_off
.value
) {
3158 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
3159 verbose(env
, "variable ctx access var_off=%s disallowed\n", tn_buf
);
3166 static int __check_buffer_access(struct bpf_verifier_env
*env
,
3167 const char *buf_info
,
3168 const struct bpf_reg_state
*reg
,
3169 int regno
, int off
, int size
)
3173 "R%d invalid %s buffer access: off=%d, size=%d\n",
3174 regno
, buf_info
, off
, size
);
3177 if (!tnum_is_const(reg
->var_off
) || reg
->var_off
.value
) {
3180 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
3182 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
3183 regno
, off
, tn_buf
);
3190 static int check_tp_buffer_access(struct bpf_verifier_env
*env
,
3191 const struct bpf_reg_state
*reg
,
3192 int regno
, int off
, int size
)
3196 err
= __check_buffer_access(env
, "tracepoint", reg
, regno
, off
, size
);
3200 if (off
+ size
> env
->prog
->aux
->max_tp_access
)
3201 env
->prog
->aux
->max_tp_access
= off
+ size
;
3206 static int check_buffer_access(struct bpf_verifier_env
*env
,
3207 const struct bpf_reg_state
*reg
,
3208 int regno
, int off
, int size
,
3209 bool zero_size_allowed
,
3210 const char *buf_info
,
3215 err
= __check_buffer_access(env
, buf_info
, reg
, regno
, off
, size
);
3219 if (off
+ size
> *max_access
)
3220 *max_access
= off
+ size
;
3225 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
3226 static void zext_32_to_64(struct bpf_reg_state
*reg
)
3228 reg
->var_off
= tnum_subreg(reg
->var_off
);
3229 __reg_assign_32_into_64(reg
);
3232 /* truncate register to smaller size (in bytes)
3233 * must be called with size < BPF_REG_SIZE
3235 static void coerce_reg_to_size(struct bpf_reg_state
*reg
, int size
)
3239 /* clear high bits in bit representation */
3240 reg
->var_off
= tnum_cast(reg
->var_off
, size
);
3242 /* fix arithmetic bounds */
3243 mask
= ((u64
)1 << (size
* 8)) - 1;
3244 if ((reg
->umin_value
& ~mask
) == (reg
->umax_value
& ~mask
)) {
3245 reg
->umin_value
&= mask
;
3246 reg
->umax_value
&= mask
;
3248 reg
->umin_value
= 0;
3249 reg
->umax_value
= mask
;
3251 reg
->smin_value
= reg
->umin_value
;
3252 reg
->smax_value
= reg
->umax_value
;
3254 /* If size is smaller than 32bit register the 32bit register
3255 * values are also truncated so we push 64-bit bounds into
3256 * 32-bit bounds. Above were truncated < 32-bits already.
3260 __reg_combine_64_into_32(reg
);
3263 static bool bpf_map_is_rdonly(const struct bpf_map
*map
)
3265 return (map
->map_flags
& BPF_F_RDONLY_PROG
) && map
->frozen
;
3268 static int bpf_map_direct_read(struct bpf_map
*map
, int off
, int size
, u64
*val
)
3274 err
= map
->ops
->map_direct_value_addr(map
, &addr
, off
);
3277 ptr
= (void *)(long)addr
+ off
;
3281 *val
= (u64
)*(u8
*)ptr
;
3284 *val
= (u64
)*(u16
*)ptr
;
3287 *val
= (u64
)*(u32
*)ptr
;
3298 static int check_ptr_to_btf_access(struct bpf_verifier_env
*env
,
3299 struct bpf_reg_state
*regs
,
3300 int regno
, int off
, int size
,
3301 enum bpf_access_type atype
,
3304 struct bpf_reg_state
*reg
= regs
+ regno
;
3305 const struct btf_type
*t
= btf_type_by_id(reg
->btf
, reg
->btf_id
);
3306 const char *tname
= btf_name_by_offset(reg
->btf
, t
->name_off
);
3312 "R%d is ptr_%s invalid negative access: off=%d\n",
3316 if (!tnum_is_const(reg
->var_off
) || reg
->var_off
.value
) {
3319 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
3321 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
3322 regno
, tname
, off
, tn_buf
);
3326 if (env
->ops
->btf_struct_access
) {
3327 ret
= env
->ops
->btf_struct_access(&env
->log
, reg
->btf
, t
,
3328 off
, size
, atype
, &btf_id
);
3330 if (atype
!= BPF_READ
) {
3331 verbose(env
, "only read is supported\n");
3335 ret
= btf_struct_access(&env
->log
, reg
->btf
, t
, off
, size
,
3342 if (atype
== BPF_READ
&& value_regno
>= 0)
3343 mark_btf_ld_reg(env
, regs
, value_regno
, ret
, reg
->btf
, btf_id
);
3348 static int check_ptr_to_map_access(struct bpf_verifier_env
*env
,
3349 struct bpf_reg_state
*regs
,
3350 int regno
, int off
, int size
,
3351 enum bpf_access_type atype
,
3354 struct bpf_reg_state
*reg
= regs
+ regno
;
3355 struct bpf_map
*map
= reg
->map_ptr
;
3356 const struct btf_type
*t
;
3362 verbose(env
, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
3366 if (!map
->ops
->map_btf_id
|| !*map
->ops
->map_btf_id
) {
3367 verbose(env
, "map_ptr access not supported for map type %d\n",
3372 t
= btf_type_by_id(btf_vmlinux
, *map
->ops
->map_btf_id
);
3373 tname
= btf_name_by_offset(btf_vmlinux
, t
->name_off
);
3375 if (!env
->allow_ptr_to_map_access
) {
3377 "%s access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
3383 verbose(env
, "R%d is %s invalid negative access: off=%d\n",
3388 if (atype
!= BPF_READ
) {
3389 verbose(env
, "only read from %s is supported\n", tname
);
3393 ret
= btf_struct_access(&env
->log
, btf_vmlinux
, t
, off
, size
, atype
, &btf_id
);
3397 if (value_regno
>= 0)
3398 mark_btf_ld_reg(env
, regs
, value_regno
, ret
, btf_vmlinux
, btf_id
);
3404 /* check whether memory at (regno + off) is accessible for t = (read | write)
3405 * if t==write, value_regno is a register which value is stored into memory
3406 * if t==read, value_regno is a register which will receive the value from memory
3407 * if t==write && value_regno==-1, some unknown value is stored into memory
3408 * if t==read && value_regno==-1, don't care what we read from memory
3410 static int check_mem_access(struct bpf_verifier_env
*env
, int insn_idx
, u32 regno
,
3411 int off
, int bpf_size
, enum bpf_access_type t
,
3412 int value_regno
, bool strict_alignment_once
)
3414 struct bpf_reg_state
*regs
= cur_regs(env
);
3415 struct bpf_reg_state
*reg
= regs
+ regno
;
3416 struct bpf_func_state
*state
;
3419 size
= bpf_size_to_bytes(bpf_size
);
3423 /* alignment checks will add in reg->off themselves */
3424 err
= check_ptr_alignment(env
, reg
, off
, size
, strict_alignment_once
);
3428 /* for access checks, reg->off is just part of off */
3431 if (reg
->type
== PTR_TO_MAP_VALUE
) {
3432 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
3433 is_pointer_value(env
, value_regno
)) {
3434 verbose(env
, "R%d leaks addr into map\n", value_regno
);
3437 err
= check_map_access_type(env
, regno
, off
, size
, t
);
3440 err
= check_map_access(env
, regno
, off
, size
, false);
3441 if (!err
&& t
== BPF_READ
&& value_regno
>= 0) {
3442 struct bpf_map
*map
= reg
->map_ptr
;
3444 /* if map is read-only, track its contents as scalars */
3445 if (tnum_is_const(reg
->var_off
) &&
3446 bpf_map_is_rdonly(map
) &&
3447 map
->ops
->map_direct_value_addr
) {
3448 int map_off
= off
+ reg
->var_off
.value
;
3451 err
= bpf_map_direct_read(map
, map_off
, size
,
3456 regs
[value_regno
].type
= SCALAR_VALUE
;
3457 __mark_reg_known(®s
[value_regno
], val
);
3459 mark_reg_unknown(env
, regs
, value_regno
);
3462 } else if (reg
->type
== PTR_TO_MEM
) {
3463 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
3464 is_pointer_value(env
, value_regno
)) {
3465 verbose(env
, "R%d leaks addr into mem\n", value_regno
);
3468 err
= check_mem_region_access(env
, regno
, off
, size
,
3469 reg
->mem_size
, false);
3470 if (!err
&& t
== BPF_READ
&& value_regno
>= 0)
3471 mark_reg_unknown(env
, regs
, value_regno
);
3472 } else if (reg
->type
== PTR_TO_CTX
) {
3473 enum bpf_reg_type reg_type
= SCALAR_VALUE
;
3474 struct btf
*btf
= NULL
;
3477 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
3478 is_pointer_value(env
, value_regno
)) {
3479 verbose(env
, "R%d leaks addr into ctx\n", value_regno
);
3483 err
= check_ctx_reg(env
, reg
, regno
);
3487 err
= check_ctx_access(env
, insn_idx
, off
, size
, t
, ®_type
, &btf
, &btf_id
);
3489 verbose_linfo(env
, insn_idx
, "; ");
3490 if (!err
&& t
== BPF_READ
&& value_regno
>= 0) {
3491 /* ctx access returns either a scalar, or a
3492 * PTR_TO_PACKET[_META,_END]. In the latter
3493 * case, we know the offset is zero.
3495 if (reg_type
== SCALAR_VALUE
) {
3496 mark_reg_unknown(env
, regs
, value_regno
);
3498 mark_reg_known_zero(env
, regs
,
3500 if (reg_type_may_be_null(reg_type
))
3501 regs
[value_regno
].id
= ++env
->id_gen
;
3502 /* A load of ctx field could have different
3503 * actual load size with the one encoded in the
3504 * insn. When the dst is PTR, it is for sure not
3507 regs
[value_regno
].subreg_def
= DEF_NOT_SUBREG
;
3508 if (reg_type
== PTR_TO_BTF_ID
||
3509 reg_type
== PTR_TO_BTF_ID_OR_NULL
) {
3510 regs
[value_regno
].btf
= btf
;
3511 regs
[value_regno
].btf_id
= btf_id
;
3514 regs
[value_regno
].type
= reg_type
;
3517 } else if (reg
->type
== PTR_TO_STACK
) {
3518 off
+= reg
->var_off
.value
;
3519 err
= check_stack_access(env
, reg
, off
, size
);
3523 state
= func(env
, reg
);
3524 err
= update_stack_depth(env
, state
, off
);
3529 err
= check_stack_write(env
, state
, off
, size
,
3530 value_regno
, insn_idx
);
3532 err
= check_stack_read(env
, state
, off
, size
,
3534 } else if (reg_is_pkt_pointer(reg
)) {
3535 if (t
== BPF_WRITE
&& !may_access_direct_pkt_data(env
, NULL
, t
)) {
3536 verbose(env
, "cannot write into packet\n");
3539 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
3540 is_pointer_value(env
, value_regno
)) {
3541 verbose(env
, "R%d leaks addr into packet\n",
3545 err
= check_packet_access(env
, regno
, off
, size
, false);
3546 if (!err
&& t
== BPF_READ
&& value_regno
>= 0)
3547 mark_reg_unknown(env
, regs
, value_regno
);
3548 } else if (reg
->type
== PTR_TO_FLOW_KEYS
) {
3549 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
3550 is_pointer_value(env
, value_regno
)) {
3551 verbose(env
, "R%d leaks addr into flow keys\n",
3556 err
= check_flow_keys_access(env
, off
, size
);
3557 if (!err
&& t
== BPF_READ
&& value_regno
>= 0)
3558 mark_reg_unknown(env
, regs
, value_regno
);
3559 } else if (type_is_sk_pointer(reg
->type
)) {
3560 if (t
== BPF_WRITE
) {
3561 verbose(env
, "R%d cannot write into %s\n",
3562 regno
, reg_type_str
[reg
->type
]);
3565 err
= check_sock_access(env
, insn_idx
, regno
, off
, size
, t
);
3566 if (!err
&& value_regno
>= 0)
3567 mark_reg_unknown(env
, regs
, value_regno
);
3568 } else if (reg
->type
== PTR_TO_TP_BUFFER
) {
3569 err
= check_tp_buffer_access(env
, reg
, regno
, off
, size
);
3570 if (!err
&& t
== BPF_READ
&& value_regno
>= 0)
3571 mark_reg_unknown(env
, regs
, value_regno
);
3572 } else if (reg
->type
== PTR_TO_BTF_ID
) {
3573 err
= check_ptr_to_btf_access(env
, regs
, regno
, off
, size
, t
,
3575 } else if (reg
->type
== CONST_PTR_TO_MAP
) {
3576 err
= check_ptr_to_map_access(env
, regs
, regno
, off
, size
, t
,
3578 } else if (reg
->type
== PTR_TO_RDONLY_BUF
) {
3579 if (t
== BPF_WRITE
) {
3580 verbose(env
, "R%d cannot write into %s\n",
3581 regno
, reg_type_str
[reg
->type
]);
3584 err
= check_buffer_access(env
, reg
, regno
, off
, size
, false,
3586 &env
->prog
->aux
->max_rdonly_access
);
3587 if (!err
&& value_regno
>= 0)
3588 mark_reg_unknown(env
, regs
, value_regno
);
3589 } else if (reg
->type
== PTR_TO_RDWR_BUF
) {
3590 err
= check_buffer_access(env
, reg
, regno
, off
, size
, false,
3592 &env
->prog
->aux
->max_rdwr_access
);
3593 if (!err
&& t
== BPF_READ
&& value_regno
>= 0)
3594 mark_reg_unknown(env
, regs
, value_regno
);
3596 verbose(env
, "R%d invalid mem access '%s'\n", regno
,
3597 reg_type_str
[reg
->type
]);
3601 if (!err
&& size
< BPF_REG_SIZE
&& value_regno
>= 0 && t
== BPF_READ
&&
3602 regs
[value_regno
].type
== SCALAR_VALUE
) {
3603 /* b/h/w load zero-extends, mark upper bits as known 0 */
3604 coerce_reg_to_size(®s
[value_regno
], size
);
3609 static int check_xadd(struct bpf_verifier_env
*env
, int insn_idx
, struct bpf_insn
*insn
)
3613 if ((BPF_SIZE(insn
->code
) != BPF_W
&& BPF_SIZE(insn
->code
) != BPF_DW
) ||
3615 verbose(env
, "BPF_XADD uses reserved fields\n");
3619 /* check src1 operand */
3620 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
3624 /* check src2 operand */
3625 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
3629 if (is_pointer_value(env
, insn
->src_reg
)) {
3630 verbose(env
, "R%d leaks addr into mem\n", insn
->src_reg
);
3634 if (is_ctx_reg(env
, insn
->dst_reg
) ||
3635 is_pkt_reg(env
, insn
->dst_reg
) ||
3636 is_flow_key_reg(env
, insn
->dst_reg
) ||
3637 is_sk_reg(env
, insn
->dst_reg
)) {
3638 verbose(env
, "BPF_XADD stores into R%d %s is not allowed\n",
3640 reg_type_str
[reg_state(env
, insn
->dst_reg
)->type
]);
3644 /* check whether atomic_add can read the memory */
3645 err
= check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
3646 BPF_SIZE(insn
->code
), BPF_READ
, -1, true);
3650 /* check whether atomic_add can write into the same memory */
3651 return check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
3652 BPF_SIZE(insn
->code
), BPF_WRITE
, -1, true);
3655 static int __check_stack_boundary(struct bpf_verifier_env
*env
, u32 regno
,
3656 int off
, int access_size
,
3657 bool zero_size_allowed
)
3659 struct bpf_reg_state
*reg
= reg_state(env
, regno
);
3661 if (off
>= 0 || off
< -MAX_BPF_STACK
|| off
+ access_size
> 0 ||
3662 access_size
< 0 || (access_size
== 0 && !zero_size_allowed
)) {
3663 if (tnum_is_const(reg
->var_off
)) {
3664 verbose(env
, "invalid stack type R%d off=%d access_size=%d\n",
3665 regno
, off
, access_size
);
3669 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
3670 verbose(env
, "invalid stack type R%d var_off=%s access_size=%d\n",
3671 regno
, tn_buf
, access_size
);
3678 /* when register 'regno' is passed into function that will read 'access_size'
3679 * bytes from that pointer, make sure that it's within stack boundary
3680 * and all elements of stack are initialized.
3681 * Unlike most pointer bounds-checking functions, this one doesn't take an
3682 * 'off' argument, so it has to add in reg->off itself.
3684 static int check_stack_boundary(struct bpf_verifier_env
*env
, int regno
,
3685 int access_size
, bool zero_size_allowed
,
3686 struct bpf_call_arg_meta
*meta
)
3688 struct bpf_reg_state
*reg
= reg_state(env
, regno
);
3689 struct bpf_func_state
*state
= func(env
, reg
);
3690 int err
, min_off
, max_off
, i
, j
, slot
, spi
;
3692 if (tnum_is_const(reg
->var_off
)) {
3693 min_off
= max_off
= reg
->var_off
.value
+ reg
->off
;
3694 err
= __check_stack_boundary(env
, regno
, min_off
, access_size
,
3699 /* Variable offset is prohibited for unprivileged mode for
3700 * simplicity since it requires corresponding support in
3701 * Spectre masking for stack ALU.
3702 * See also retrieve_ptr_limit().
3704 if (!env
->bypass_spec_v1
) {
3707 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
3708 verbose(env
, "R%d indirect variable offset stack access prohibited for !root, var_off=%s\n",
3712 /* Only initialized buffer on stack is allowed to be accessed
3713 * with variable offset. With uninitialized buffer it's hard to
3714 * guarantee that whole memory is marked as initialized on
3715 * helper return since specific bounds are unknown what may
3716 * cause uninitialized stack leaking.
3718 if (meta
&& meta
->raw_mode
)
3721 if (reg
->smax_value
>= BPF_MAX_VAR_OFF
||
3722 reg
->smax_value
<= -BPF_MAX_VAR_OFF
) {
3723 verbose(env
, "R%d unbounded indirect variable offset stack access\n",
3727 min_off
= reg
->smin_value
+ reg
->off
;
3728 max_off
= reg
->smax_value
+ reg
->off
;
3729 err
= __check_stack_boundary(env
, regno
, min_off
, access_size
,
3732 verbose(env
, "R%d min value is outside of stack bound\n",
3736 err
= __check_stack_boundary(env
, regno
, max_off
, access_size
,
3739 verbose(env
, "R%d max value is outside of stack bound\n",
3745 if (meta
&& meta
->raw_mode
) {
3746 meta
->access_size
= access_size
;
3747 meta
->regno
= regno
;
3751 for (i
= min_off
; i
< max_off
+ access_size
; i
++) {
3755 spi
= slot
/ BPF_REG_SIZE
;
3756 if (state
->allocated_stack
<= slot
)
3758 stype
= &state
->stack
[spi
].slot_type
[slot
% BPF_REG_SIZE
];
3759 if (*stype
== STACK_MISC
)
3761 if (*stype
== STACK_ZERO
) {
3762 /* helper can write anything into the stack */
3763 *stype
= STACK_MISC
;
3767 if (state
->stack
[spi
].slot_type
[0] == STACK_SPILL
&&
3768 state
->stack
[spi
].spilled_ptr
.type
== PTR_TO_BTF_ID
)
3771 if (state
->stack
[spi
].slot_type
[0] == STACK_SPILL
&&
3772 (state
->stack
[spi
].spilled_ptr
.type
== SCALAR_VALUE
||
3773 env
->allow_ptr_leaks
)) {
3774 __mark_reg_unknown(env
, &state
->stack
[spi
].spilled_ptr
);
3775 for (j
= 0; j
< BPF_REG_SIZE
; j
++)
3776 state
->stack
[spi
].slot_type
[j
] = STACK_MISC
;
3781 if (tnum_is_const(reg
->var_off
)) {
3782 verbose(env
, "invalid indirect read from stack off %d+%d size %d\n",
3783 min_off
, i
- min_off
, access_size
);
3787 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
3788 verbose(env
, "invalid indirect read from stack var_off %s+%d size %d\n",
3789 tn_buf
, i
- min_off
, access_size
);
3793 /* reading any byte out of 8-byte 'spill_slot' will cause
3794 * the whole slot to be marked as 'read'
3796 mark_reg_read(env
, &state
->stack
[spi
].spilled_ptr
,
3797 state
->stack
[spi
].spilled_ptr
.parent
,
3800 return update_stack_depth(env
, state
, min_off
);
3803 static int check_helper_mem_access(struct bpf_verifier_env
*env
, int regno
,
3804 int access_size
, bool zero_size_allowed
,
3805 struct bpf_call_arg_meta
*meta
)
3807 struct bpf_reg_state
*regs
= cur_regs(env
), *reg
= ®s
[regno
];
3809 switch (reg
->type
) {
3811 case PTR_TO_PACKET_META
:
3812 return check_packet_access(env
, regno
, reg
->off
, access_size
,
3814 case PTR_TO_MAP_VALUE
:
3815 if (check_map_access_type(env
, regno
, reg
->off
, access_size
,
3816 meta
&& meta
->raw_mode
? BPF_WRITE
:
3819 return check_map_access(env
, regno
, reg
->off
, access_size
,
3822 return check_mem_region_access(env
, regno
, reg
->off
,
3823 access_size
, reg
->mem_size
,
3825 case PTR_TO_RDONLY_BUF
:
3826 if (meta
&& meta
->raw_mode
)
3828 return check_buffer_access(env
, reg
, regno
, reg
->off
,
3829 access_size
, zero_size_allowed
,
3831 &env
->prog
->aux
->max_rdonly_access
);
3832 case PTR_TO_RDWR_BUF
:
3833 return check_buffer_access(env
, reg
, regno
, reg
->off
,
3834 access_size
, zero_size_allowed
,
3836 &env
->prog
->aux
->max_rdwr_access
);
3838 return check_stack_boundary(env
, regno
, access_size
,
3839 zero_size_allowed
, meta
);
3840 default: /* scalar_value or invalid ptr */
3841 /* Allow zero-byte read from NULL, regardless of pointer type */
3842 if (zero_size_allowed
&& access_size
== 0 &&
3843 register_is_null(reg
))
3846 verbose(env
, "R%d type=%s expected=%s\n", regno
,
3847 reg_type_str
[reg
->type
],
3848 reg_type_str
[PTR_TO_STACK
]);
3853 /* Implementation details:
3854 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
3855 * Two bpf_map_lookups (even with the same key) will have different reg->id.
3856 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
3857 * value_or_null->value transition, since the verifier only cares about
3858 * the range of access to valid map value pointer and doesn't care about actual
3859 * address of the map element.
3860 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
3861 * reg->id > 0 after value_or_null->value transition. By doing so
3862 * two bpf_map_lookups will be considered two different pointers that
3863 * point to different bpf_spin_locks.
3864 * The verifier allows taking only one bpf_spin_lock at a time to avoid
3866 * Since only one bpf_spin_lock is allowed the checks are simpler than
3867 * reg_is_refcounted() logic. The verifier needs to remember only
3868 * one spin_lock instead of array of acquired_refs.
3869 * cur_state->active_spin_lock remembers which map value element got locked
3870 * and clears it after bpf_spin_unlock.
3872 static int process_spin_lock(struct bpf_verifier_env
*env
, int regno
,
3875 struct bpf_reg_state
*regs
= cur_regs(env
), *reg
= ®s
[regno
];
3876 struct bpf_verifier_state
*cur
= env
->cur_state
;
3877 bool is_const
= tnum_is_const(reg
->var_off
);
3878 struct bpf_map
*map
= reg
->map_ptr
;
3879 u64 val
= reg
->var_off
.value
;
3883 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
3889 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
3893 if (!map_value_has_spin_lock(map
)) {
3894 if (map
->spin_lock_off
== -E2BIG
)
3896 "map '%s' has more than one 'struct bpf_spin_lock'\n",
3898 else if (map
->spin_lock_off
== -ENOENT
)
3900 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
3904 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
3908 if (map
->spin_lock_off
!= val
+ reg
->off
) {
3909 verbose(env
, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
3914 if (cur
->active_spin_lock
) {
3916 "Locking two bpf_spin_locks are not allowed\n");
3919 cur
->active_spin_lock
= reg
->id
;
3921 if (!cur
->active_spin_lock
) {
3922 verbose(env
, "bpf_spin_unlock without taking a lock\n");
3925 if (cur
->active_spin_lock
!= reg
->id
) {
3926 verbose(env
, "bpf_spin_unlock of different lock\n");
3929 cur
->active_spin_lock
= 0;
3934 static bool arg_type_is_mem_ptr(enum bpf_arg_type type
)
3936 return type
== ARG_PTR_TO_MEM
||
3937 type
== ARG_PTR_TO_MEM_OR_NULL
||
3938 type
== ARG_PTR_TO_UNINIT_MEM
;
3941 static bool arg_type_is_mem_size(enum bpf_arg_type type
)
3943 return type
== ARG_CONST_SIZE
||
3944 type
== ARG_CONST_SIZE_OR_ZERO
;
3947 static bool arg_type_is_alloc_size(enum bpf_arg_type type
)
3949 return type
== ARG_CONST_ALLOC_SIZE_OR_ZERO
;
3952 static bool arg_type_is_int_ptr(enum bpf_arg_type type
)
3954 return type
== ARG_PTR_TO_INT
||
3955 type
== ARG_PTR_TO_LONG
;
3958 static int int_ptr_type_to_size(enum bpf_arg_type type
)
3960 if (type
== ARG_PTR_TO_INT
)
3962 else if (type
== ARG_PTR_TO_LONG
)
3968 static int resolve_map_arg_type(struct bpf_verifier_env
*env
,
3969 const struct bpf_call_arg_meta
*meta
,
3970 enum bpf_arg_type
*arg_type
)
3972 if (!meta
->map_ptr
) {
3973 /* kernel subsystem misconfigured verifier */
3974 verbose(env
, "invalid map_ptr to access map->type\n");
3978 switch (meta
->map_ptr
->map_type
) {
3979 case BPF_MAP_TYPE_SOCKMAP
:
3980 case BPF_MAP_TYPE_SOCKHASH
:
3981 if (*arg_type
== ARG_PTR_TO_MAP_VALUE
) {
3982 *arg_type
= ARG_PTR_TO_BTF_ID_SOCK_COMMON
;
3984 verbose(env
, "invalid arg_type for sockmap/sockhash\n");
3995 struct bpf_reg_types
{
3996 const enum bpf_reg_type types
[10];
4000 static const struct bpf_reg_types map_key_value_types
= {
4009 static const struct bpf_reg_types sock_types
= {
4019 static const struct bpf_reg_types btf_id_sock_common_types
= {
4027 .btf_id
= &btf_sock_ids
[BTF_SOCK_TYPE_SOCK_COMMON
],
4031 static const struct bpf_reg_types mem_types
= {
4043 static const struct bpf_reg_types int_ptr_types
= {
4052 static const struct bpf_reg_types fullsock_types
= { .types
= { PTR_TO_SOCKET
} };
4053 static const struct bpf_reg_types scalar_types
= { .types
= { SCALAR_VALUE
} };
4054 static const struct bpf_reg_types context_types
= { .types
= { PTR_TO_CTX
} };
4055 static const struct bpf_reg_types alloc_mem_types
= { .types
= { PTR_TO_MEM
} };
4056 static const struct bpf_reg_types const_map_ptr_types
= { .types
= { CONST_PTR_TO_MAP
} };
4057 static const struct bpf_reg_types btf_ptr_types
= { .types
= { PTR_TO_BTF_ID
} };
4058 static const struct bpf_reg_types spin_lock_types
= { .types
= { PTR_TO_MAP_VALUE
} };
4059 static const struct bpf_reg_types percpu_btf_ptr_types
= { .types
= { PTR_TO_PERCPU_BTF_ID
} };
4061 static const struct bpf_reg_types
*compatible_reg_types
[__BPF_ARG_TYPE_MAX
] = {
4062 [ARG_PTR_TO_MAP_KEY
] = &map_key_value_types
,
4063 [ARG_PTR_TO_MAP_VALUE
] = &map_key_value_types
,
4064 [ARG_PTR_TO_UNINIT_MAP_VALUE
] = &map_key_value_types
,
4065 [ARG_PTR_TO_MAP_VALUE_OR_NULL
] = &map_key_value_types
,
4066 [ARG_CONST_SIZE
] = &scalar_types
,
4067 [ARG_CONST_SIZE_OR_ZERO
] = &scalar_types
,
4068 [ARG_CONST_ALLOC_SIZE_OR_ZERO
] = &scalar_types
,
4069 [ARG_CONST_MAP_PTR
] = &const_map_ptr_types
,
4070 [ARG_PTR_TO_CTX
] = &context_types
,
4071 [ARG_PTR_TO_CTX_OR_NULL
] = &context_types
,
4072 [ARG_PTR_TO_SOCK_COMMON
] = &sock_types
,
4074 [ARG_PTR_TO_BTF_ID_SOCK_COMMON
] = &btf_id_sock_common_types
,
4076 [ARG_PTR_TO_SOCKET
] = &fullsock_types
,
4077 [ARG_PTR_TO_SOCKET_OR_NULL
] = &fullsock_types
,
4078 [ARG_PTR_TO_BTF_ID
] = &btf_ptr_types
,
4079 [ARG_PTR_TO_SPIN_LOCK
] = &spin_lock_types
,
4080 [ARG_PTR_TO_MEM
] = &mem_types
,
4081 [ARG_PTR_TO_MEM_OR_NULL
] = &mem_types
,
4082 [ARG_PTR_TO_UNINIT_MEM
] = &mem_types
,
4083 [ARG_PTR_TO_ALLOC_MEM
] = &alloc_mem_types
,
4084 [ARG_PTR_TO_ALLOC_MEM_OR_NULL
] = &alloc_mem_types
,
4085 [ARG_PTR_TO_INT
] = &int_ptr_types
,
4086 [ARG_PTR_TO_LONG
] = &int_ptr_types
,
4087 [ARG_PTR_TO_PERCPU_BTF_ID
] = &percpu_btf_ptr_types
,
4090 static int check_reg_type(struct bpf_verifier_env
*env
, u32 regno
,
4091 enum bpf_arg_type arg_type
,
4092 const u32
*arg_btf_id
)
4094 struct bpf_reg_state
*regs
= cur_regs(env
), *reg
= ®s
[regno
];
4095 enum bpf_reg_type expected
, type
= reg
->type
;
4096 const struct bpf_reg_types
*compatible
;
4099 compatible
= compatible_reg_types
[arg_type
];
4101 verbose(env
, "verifier internal error: unsupported arg type %d\n", arg_type
);
4105 for (i
= 0; i
< ARRAY_SIZE(compatible
->types
); i
++) {
4106 expected
= compatible
->types
[i
];
4107 if (expected
== NOT_INIT
)
4110 if (type
== expected
)
4114 verbose(env
, "R%d type=%s expected=", regno
, reg_type_str
[type
]);
4115 for (j
= 0; j
+ 1 < i
; j
++)
4116 verbose(env
, "%s, ", reg_type_str
[compatible
->types
[j
]]);
4117 verbose(env
, "%s\n", reg_type_str
[compatible
->types
[j
]]);
4121 if (type
== PTR_TO_BTF_ID
) {
4123 if (!compatible
->btf_id
) {
4124 verbose(env
, "verifier internal error: missing arg compatible BTF ID\n");
4127 arg_btf_id
= compatible
->btf_id
;
4130 if (!btf_struct_ids_match(&env
->log
, reg
->btf
, reg
->btf_id
, reg
->off
,
4131 btf_vmlinux
, *arg_btf_id
)) {
4132 verbose(env
, "R%d is of type %s but %s is expected\n",
4133 regno
, kernel_type_name(reg
->btf
, reg
->btf_id
),
4134 kernel_type_name(btf_vmlinux
, *arg_btf_id
));
4138 if (!tnum_is_const(reg
->var_off
) || reg
->var_off
.value
) {
4139 verbose(env
, "R%d is a pointer to in-kernel struct with non-zero offset\n",
4148 static int check_func_arg(struct bpf_verifier_env
*env
, u32 arg
,
4149 struct bpf_call_arg_meta
*meta
,
4150 const struct bpf_func_proto
*fn
)
4152 u32 regno
= BPF_REG_1
+ arg
;
4153 struct bpf_reg_state
*regs
= cur_regs(env
), *reg
= ®s
[regno
];
4154 enum bpf_arg_type arg_type
= fn
->arg_type
[arg
];
4155 enum bpf_reg_type type
= reg
->type
;
4158 if (arg_type
== ARG_DONTCARE
)
4161 err
= check_reg_arg(env
, regno
, SRC_OP
);
4165 if (arg_type
== ARG_ANYTHING
) {
4166 if (is_pointer_value(env
, regno
)) {
4167 verbose(env
, "R%d leaks addr into helper function\n",
4174 if (type_is_pkt_pointer(type
) &&
4175 !may_access_direct_pkt_data(env
, meta
, BPF_READ
)) {
4176 verbose(env
, "helper access to the packet is not allowed\n");
4180 if (arg_type
== ARG_PTR_TO_MAP_VALUE
||
4181 arg_type
== ARG_PTR_TO_UNINIT_MAP_VALUE
||
4182 arg_type
== ARG_PTR_TO_MAP_VALUE_OR_NULL
) {
4183 err
= resolve_map_arg_type(env
, meta
, &arg_type
);
4188 if (register_is_null(reg
) && arg_type_may_be_null(arg_type
))
4189 /* A NULL register has a SCALAR_VALUE type, so skip
4192 goto skip_type_check
;
4194 err
= check_reg_type(env
, regno
, arg_type
, fn
->arg_btf_id
[arg
]);
4198 if (type
== PTR_TO_CTX
) {
4199 err
= check_ctx_reg(env
, reg
, regno
);
4205 if (reg
->ref_obj_id
) {
4206 if (meta
->ref_obj_id
) {
4207 verbose(env
, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
4208 regno
, reg
->ref_obj_id
,
4212 meta
->ref_obj_id
= reg
->ref_obj_id
;
4215 if (arg_type
== ARG_CONST_MAP_PTR
) {
4216 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
4217 meta
->map_ptr
= reg
->map_ptr
;
4218 } else if (arg_type
== ARG_PTR_TO_MAP_KEY
) {
4219 /* bpf_map_xxx(..., map_ptr, ..., key) call:
4220 * check that [key, key + map->key_size) are within
4221 * stack limits and initialized
4223 if (!meta
->map_ptr
) {
4224 /* in function declaration map_ptr must come before
4225 * map_key, so that it's verified and known before
4226 * we have to check map_key here. Otherwise it means
4227 * that kernel subsystem misconfigured verifier
4229 verbose(env
, "invalid map_ptr to access map->key\n");
4232 err
= check_helper_mem_access(env
, regno
,
4233 meta
->map_ptr
->key_size
, false,
4235 } else if (arg_type
== ARG_PTR_TO_MAP_VALUE
||
4236 (arg_type
== ARG_PTR_TO_MAP_VALUE_OR_NULL
&&
4237 !register_is_null(reg
)) ||
4238 arg_type
== ARG_PTR_TO_UNINIT_MAP_VALUE
) {
4239 /* bpf_map_xxx(..., map_ptr, ..., value) call:
4240 * check [value, value + map->value_size) validity
4242 if (!meta
->map_ptr
) {
4243 /* kernel subsystem misconfigured verifier */
4244 verbose(env
, "invalid map_ptr to access map->value\n");
4247 meta
->raw_mode
= (arg_type
== ARG_PTR_TO_UNINIT_MAP_VALUE
);
4248 err
= check_helper_mem_access(env
, regno
,
4249 meta
->map_ptr
->value_size
, false,
4251 } else if (arg_type
== ARG_PTR_TO_PERCPU_BTF_ID
) {
4253 verbose(env
, "Helper has invalid btf_id in R%d\n", regno
);
4256 meta
->ret_btf
= reg
->btf
;
4257 meta
->ret_btf_id
= reg
->btf_id
;
4258 } else if (arg_type
== ARG_PTR_TO_SPIN_LOCK
) {
4259 if (meta
->func_id
== BPF_FUNC_spin_lock
) {
4260 if (process_spin_lock(env
, regno
, true))
4262 } else if (meta
->func_id
== BPF_FUNC_spin_unlock
) {
4263 if (process_spin_lock(env
, regno
, false))
4266 verbose(env
, "verifier internal error\n");
4269 } else if (arg_type_is_mem_ptr(arg_type
)) {
4270 /* The access to this pointer is only checked when we hit the
4271 * next is_mem_size argument below.
4273 meta
->raw_mode
= (arg_type
== ARG_PTR_TO_UNINIT_MEM
);
4274 } else if (arg_type_is_mem_size(arg_type
)) {
4275 bool zero_size_allowed
= (arg_type
== ARG_CONST_SIZE_OR_ZERO
);
4277 /* This is used to refine r0 return value bounds for helpers
4278 * that enforce this value as an upper bound on return values.
4279 * See do_refine_retval_range() for helpers that can refine
4280 * the return value. C type of helper is u32 so we pull register
4281 * bound from umax_value however, if negative verifier errors
4282 * out. Only upper bounds can be learned because retval is an
4283 * int type and negative retvals are allowed.
4285 meta
->msize_max_value
= reg
->umax_value
;
4287 /* The register is SCALAR_VALUE; the access check
4288 * happens using its boundaries.
4290 if (!tnum_is_const(reg
->var_off
))
4291 /* For unprivileged variable accesses, disable raw
4292 * mode so that the program is required to
4293 * initialize all the memory that the helper could
4294 * just partially fill up.
4298 if (reg
->smin_value
< 0) {
4299 verbose(env
, "R%d min value is negative, either use unsigned or 'var &= const'\n",
4304 if (reg
->umin_value
== 0) {
4305 err
= check_helper_mem_access(env
, regno
- 1, 0,
4312 if (reg
->umax_value
>= BPF_MAX_VAR_SIZ
) {
4313 verbose(env
, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
4317 err
= check_helper_mem_access(env
, regno
- 1,
4319 zero_size_allowed
, meta
);
4321 err
= mark_chain_precision(env
, regno
);
4322 } else if (arg_type_is_alloc_size(arg_type
)) {
4323 if (!tnum_is_const(reg
->var_off
)) {
4324 verbose(env
, "R%d unbounded size, use 'var &= const' or 'if (var < const)'\n",
4328 meta
->mem_size
= reg
->var_off
.value
;
4329 } else if (arg_type_is_int_ptr(arg_type
)) {
4330 int size
= int_ptr_type_to_size(arg_type
);
4332 err
= check_helper_mem_access(env
, regno
, size
, false, meta
);
4335 err
= check_ptr_alignment(env
, reg
, 0, size
, true);
4341 static bool may_update_sockmap(struct bpf_verifier_env
*env
, int func_id
)
4343 enum bpf_attach_type eatype
= env
->prog
->expected_attach_type
;
4344 enum bpf_prog_type type
= resolve_prog_type(env
->prog
);
4346 if (func_id
!= BPF_FUNC_map_update_elem
)
4349 /* It's not possible to get access to a locked struct sock in these
4350 * contexts, so updating is safe.
4353 case BPF_PROG_TYPE_TRACING
:
4354 if (eatype
== BPF_TRACE_ITER
)
4357 case BPF_PROG_TYPE_SOCKET_FILTER
:
4358 case BPF_PROG_TYPE_SCHED_CLS
:
4359 case BPF_PROG_TYPE_SCHED_ACT
:
4360 case BPF_PROG_TYPE_XDP
:
4361 case BPF_PROG_TYPE_SK_REUSEPORT
:
4362 case BPF_PROG_TYPE_FLOW_DISSECTOR
:
4363 case BPF_PROG_TYPE_SK_LOOKUP
:
4369 verbose(env
, "cannot update sockmap in this context\n");
4373 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env
*env
)
4375 return env
->prog
->jit_requested
&& IS_ENABLED(CONFIG_X86_64
);
4378 static int check_map_func_compatibility(struct bpf_verifier_env
*env
,
4379 struct bpf_map
*map
, int func_id
)
4384 /* We need a two way check, first is from map perspective ... */
4385 switch (map
->map_type
) {
4386 case BPF_MAP_TYPE_PROG_ARRAY
:
4387 if (func_id
!= BPF_FUNC_tail_call
)
4390 case BPF_MAP_TYPE_PERF_EVENT_ARRAY
:
4391 if (func_id
!= BPF_FUNC_perf_event_read
&&
4392 func_id
!= BPF_FUNC_perf_event_output
&&
4393 func_id
!= BPF_FUNC_skb_output
&&
4394 func_id
!= BPF_FUNC_perf_event_read_value
&&
4395 func_id
!= BPF_FUNC_xdp_output
)
4398 case BPF_MAP_TYPE_RINGBUF
:
4399 if (func_id
!= BPF_FUNC_ringbuf_output
&&
4400 func_id
!= BPF_FUNC_ringbuf_reserve
&&
4401 func_id
!= BPF_FUNC_ringbuf_submit
&&
4402 func_id
!= BPF_FUNC_ringbuf_discard
&&
4403 func_id
!= BPF_FUNC_ringbuf_query
)
4406 case BPF_MAP_TYPE_STACK_TRACE
:
4407 if (func_id
!= BPF_FUNC_get_stackid
)
4410 case BPF_MAP_TYPE_CGROUP_ARRAY
:
4411 if (func_id
!= BPF_FUNC_skb_under_cgroup
&&
4412 func_id
!= BPF_FUNC_current_task_under_cgroup
)
4415 case BPF_MAP_TYPE_CGROUP_STORAGE
:
4416 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE
:
4417 if (func_id
!= BPF_FUNC_get_local_storage
)
4420 case BPF_MAP_TYPE_DEVMAP
:
4421 case BPF_MAP_TYPE_DEVMAP_HASH
:
4422 if (func_id
!= BPF_FUNC_redirect_map
&&
4423 func_id
!= BPF_FUNC_map_lookup_elem
)
4426 /* Restrict bpf side of cpumap and xskmap, open when use-cases
4429 case BPF_MAP_TYPE_CPUMAP
:
4430 if (func_id
!= BPF_FUNC_redirect_map
)
4433 case BPF_MAP_TYPE_XSKMAP
:
4434 if (func_id
!= BPF_FUNC_redirect_map
&&
4435 func_id
!= BPF_FUNC_map_lookup_elem
)
4438 case BPF_MAP_TYPE_ARRAY_OF_MAPS
:
4439 case BPF_MAP_TYPE_HASH_OF_MAPS
:
4440 if (func_id
!= BPF_FUNC_map_lookup_elem
)
4443 case BPF_MAP_TYPE_SOCKMAP
:
4444 if (func_id
!= BPF_FUNC_sk_redirect_map
&&
4445 func_id
!= BPF_FUNC_sock_map_update
&&
4446 func_id
!= BPF_FUNC_map_delete_elem
&&
4447 func_id
!= BPF_FUNC_msg_redirect_map
&&
4448 func_id
!= BPF_FUNC_sk_select_reuseport
&&
4449 func_id
!= BPF_FUNC_map_lookup_elem
&&
4450 !may_update_sockmap(env
, func_id
))
4453 case BPF_MAP_TYPE_SOCKHASH
:
4454 if (func_id
!= BPF_FUNC_sk_redirect_hash
&&
4455 func_id
!= BPF_FUNC_sock_hash_update
&&
4456 func_id
!= BPF_FUNC_map_delete_elem
&&
4457 func_id
!= BPF_FUNC_msg_redirect_hash
&&
4458 func_id
!= BPF_FUNC_sk_select_reuseport
&&
4459 func_id
!= BPF_FUNC_map_lookup_elem
&&
4460 !may_update_sockmap(env
, func_id
))
4463 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY
:
4464 if (func_id
!= BPF_FUNC_sk_select_reuseport
)
4467 case BPF_MAP_TYPE_QUEUE
:
4468 case BPF_MAP_TYPE_STACK
:
4469 if (func_id
!= BPF_FUNC_map_peek_elem
&&
4470 func_id
!= BPF_FUNC_map_pop_elem
&&
4471 func_id
!= BPF_FUNC_map_push_elem
)
4474 case BPF_MAP_TYPE_SK_STORAGE
:
4475 if (func_id
!= BPF_FUNC_sk_storage_get
&&
4476 func_id
!= BPF_FUNC_sk_storage_delete
)
4479 case BPF_MAP_TYPE_INODE_STORAGE
:
4480 if (func_id
!= BPF_FUNC_inode_storage_get
&&
4481 func_id
!= BPF_FUNC_inode_storage_delete
)
4484 case BPF_MAP_TYPE_TASK_STORAGE
:
4485 if (func_id
!= BPF_FUNC_task_storage_get
&&
4486 func_id
!= BPF_FUNC_task_storage_delete
)
4493 /* ... and second from the function itself. */
4495 case BPF_FUNC_tail_call
:
4496 if (map
->map_type
!= BPF_MAP_TYPE_PROG_ARRAY
)
4498 if (env
->subprog_cnt
> 1 && !allow_tail_call_in_subprogs(env
)) {
4499 verbose(env
, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
4503 case BPF_FUNC_perf_event_read
:
4504 case BPF_FUNC_perf_event_output
:
4505 case BPF_FUNC_perf_event_read_value
:
4506 case BPF_FUNC_skb_output
:
4507 case BPF_FUNC_xdp_output
:
4508 if (map
->map_type
!= BPF_MAP_TYPE_PERF_EVENT_ARRAY
)
4511 case BPF_FUNC_get_stackid
:
4512 if (map
->map_type
!= BPF_MAP_TYPE_STACK_TRACE
)
4515 case BPF_FUNC_current_task_under_cgroup
:
4516 case BPF_FUNC_skb_under_cgroup
:
4517 if (map
->map_type
!= BPF_MAP_TYPE_CGROUP_ARRAY
)
4520 case BPF_FUNC_redirect_map
:
4521 if (map
->map_type
!= BPF_MAP_TYPE_DEVMAP
&&
4522 map
->map_type
!= BPF_MAP_TYPE_DEVMAP_HASH
&&
4523 map
->map_type
!= BPF_MAP_TYPE_CPUMAP
&&
4524 map
->map_type
!= BPF_MAP_TYPE_XSKMAP
)
4527 case BPF_FUNC_sk_redirect_map
:
4528 case BPF_FUNC_msg_redirect_map
:
4529 case BPF_FUNC_sock_map_update
:
4530 if (map
->map_type
!= BPF_MAP_TYPE_SOCKMAP
)
4533 case BPF_FUNC_sk_redirect_hash
:
4534 case BPF_FUNC_msg_redirect_hash
:
4535 case BPF_FUNC_sock_hash_update
:
4536 if (map
->map_type
!= BPF_MAP_TYPE_SOCKHASH
)
4539 case BPF_FUNC_get_local_storage
:
4540 if (map
->map_type
!= BPF_MAP_TYPE_CGROUP_STORAGE
&&
4541 map
->map_type
!= BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE
)
4544 case BPF_FUNC_sk_select_reuseport
:
4545 if (map
->map_type
!= BPF_MAP_TYPE_REUSEPORT_SOCKARRAY
&&
4546 map
->map_type
!= BPF_MAP_TYPE_SOCKMAP
&&
4547 map
->map_type
!= BPF_MAP_TYPE_SOCKHASH
)
4550 case BPF_FUNC_map_peek_elem
:
4551 case BPF_FUNC_map_pop_elem
:
4552 case BPF_FUNC_map_push_elem
:
4553 if (map
->map_type
!= BPF_MAP_TYPE_QUEUE
&&
4554 map
->map_type
!= BPF_MAP_TYPE_STACK
)
4557 case BPF_FUNC_sk_storage_get
:
4558 case BPF_FUNC_sk_storage_delete
:
4559 if (map
->map_type
!= BPF_MAP_TYPE_SK_STORAGE
)
4562 case BPF_FUNC_inode_storage_get
:
4563 case BPF_FUNC_inode_storage_delete
:
4564 if (map
->map_type
!= BPF_MAP_TYPE_INODE_STORAGE
)
4567 case BPF_FUNC_task_storage_get
:
4568 case BPF_FUNC_task_storage_delete
:
4569 if (map
->map_type
!= BPF_MAP_TYPE_TASK_STORAGE
)
4578 verbose(env
, "cannot pass map_type %d into func %s#%d\n",
4579 map
->map_type
, func_id_name(func_id
), func_id
);
4583 static bool check_raw_mode_ok(const struct bpf_func_proto
*fn
)
4587 if (fn
->arg1_type
== ARG_PTR_TO_UNINIT_MEM
)
4589 if (fn
->arg2_type
== ARG_PTR_TO_UNINIT_MEM
)
4591 if (fn
->arg3_type
== ARG_PTR_TO_UNINIT_MEM
)
4593 if (fn
->arg4_type
== ARG_PTR_TO_UNINIT_MEM
)
4595 if (fn
->arg5_type
== ARG_PTR_TO_UNINIT_MEM
)
4598 /* We only support one arg being in raw mode at the moment,
4599 * which is sufficient for the helper functions we have
4605 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr
,
4606 enum bpf_arg_type arg_next
)
4608 return (arg_type_is_mem_ptr(arg_curr
) &&
4609 !arg_type_is_mem_size(arg_next
)) ||
4610 (!arg_type_is_mem_ptr(arg_curr
) &&
4611 arg_type_is_mem_size(arg_next
));
4614 static bool check_arg_pair_ok(const struct bpf_func_proto
*fn
)
4616 /* bpf_xxx(..., buf, len) call will access 'len'
4617 * bytes from memory 'buf'. Both arg types need
4618 * to be paired, so make sure there's no buggy
4619 * helper function specification.
4621 if (arg_type_is_mem_size(fn
->arg1_type
) ||
4622 arg_type_is_mem_ptr(fn
->arg5_type
) ||
4623 check_args_pair_invalid(fn
->arg1_type
, fn
->arg2_type
) ||
4624 check_args_pair_invalid(fn
->arg2_type
, fn
->arg3_type
) ||
4625 check_args_pair_invalid(fn
->arg3_type
, fn
->arg4_type
) ||
4626 check_args_pair_invalid(fn
->arg4_type
, fn
->arg5_type
))
4632 static bool check_refcount_ok(const struct bpf_func_proto
*fn
, int func_id
)
4636 if (arg_type_may_be_refcounted(fn
->arg1_type
))
4638 if (arg_type_may_be_refcounted(fn
->arg2_type
))
4640 if (arg_type_may_be_refcounted(fn
->arg3_type
))
4642 if (arg_type_may_be_refcounted(fn
->arg4_type
))
4644 if (arg_type_may_be_refcounted(fn
->arg5_type
))
4647 /* A reference acquiring function cannot acquire
4648 * another refcounted ptr.
4650 if (may_be_acquire_function(func_id
) && count
)
4653 /* We only support one arg being unreferenced at the moment,
4654 * which is sufficient for the helper functions we have right now.
4659 static bool check_btf_id_ok(const struct bpf_func_proto
*fn
)
4663 for (i
= 0; i
< ARRAY_SIZE(fn
->arg_type
); i
++) {
4664 if (fn
->arg_type
[i
] == ARG_PTR_TO_BTF_ID
&& !fn
->arg_btf_id
[i
])
4667 if (fn
->arg_type
[i
] != ARG_PTR_TO_BTF_ID
&& fn
->arg_btf_id
[i
])
4674 static int check_func_proto(const struct bpf_func_proto
*fn
, int func_id
)
4676 return check_raw_mode_ok(fn
) &&
4677 check_arg_pair_ok(fn
) &&
4678 check_btf_id_ok(fn
) &&
4679 check_refcount_ok(fn
, func_id
) ? 0 : -EINVAL
;
4682 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
4683 * are now invalid, so turn them into unknown SCALAR_VALUE.
4685 static void __clear_all_pkt_pointers(struct bpf_verifier_env
*env
,
4686 struct bpf_func_state
*state
)
4688 struct bpf_reg_state
*regs
= state
->regs
, *reg
;
4691 for (i
= 0; i
< MAX_BPF_REG
; i
++)
4692 if (reg_is_pkt_pointer_any(®s
[i
]))
4693 mark_reg_unknown(env
, regs
, i
);
4695 bpf_for_each_spilled_reg(i
, state
, reg
) {
4698 if (reg_is_pkt_pointer_any(reg
))
4699 __mark_reg_unknown(env
, reg
);
4703 static void clear_all_pkt_pointers(struct bpf_verifier_env
*env
)
4705 struct bpf_verifier_state
*vstate
= env
->cur_state
;
4708 for (i
= 0; i
<= vstate
->curframe
; i
++)
4709 __clear_all_pkt_pointers(env
, vstate
->frame
[i
]);
4714 BEYOND_PKT_END
= -2,
4717 static void mark_pkt_end(struct bpf_verifier_state
*vstate
, int regn
, bool range_open
)
4719 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
4720 struct bpf_reg_state
*reg
= &state
->regs
[regn
];
4722 if (reg
->type
!= PTR_TO_PACKET
)
4723 /* PTR_TO_PACKET_META is not supported yet */
4726 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
4727 * How far beyond pkt_end it goes is unknown.
4728 * if (!range_open) it's the case of pkt >= pkt_end
4729 * if (range_open) it's the case of pkt > pkt_end
4730 * hence this pointer is at least 1 byte bigger than pkt_end
4733 reg
->range
= BEYOND_PKT_END
;
4735 reg
->range
= AT_PKT_END
;
4738 static void release_reg_references(struct bpf_verifier_env
*env
,
4739 struct bpf_func_state
*state
,
4742 struct bpf_reg_state
*regs
= state
->regs
, *reg
;
4745 for (i
= 0; i
< MAX_BPF_REG
; i
++)
4746 if (regs
[i
].ref_obj_id
== ref_obj_id
)
4747 mark_reg_unknown(env
, regs
, i
);
4749 bpf_for_each_spilled_reg(i
, state
, reg
) {
4752 if (reg
->ref_obj_id
== ref_obj_id
)
4753 __mark_reg_unknown(env
, reg
);
4757 /* The pointer with the specified id has released its reference to kernel
4758 * resources. Identify all copies of the same pointer and clear the reference.
4760 static int release_reference(struct bpf_verifier_env
*env
,
4763 struct bpf_verifier_state
*vstate
= env
->cur_state
;
4767 err
= release_reference_state(cur_func(env
), ref_obj_id
);
4771 for (i
= 0; i
<= vstate
->curframe
; i
++)
4772 release_reg_references(env
, vstate
->frame
[i
], ref_obj_id
);
4777 static void clear_caller_saved_regs(struct bpf_verifier_env
*env
,
4778 struct bpf_reg_state
*regs
)
4782 /* after the call registers r0 - r5 were scratched */
4783 for (i
= 0; i
< CALLER_SAVED_REGS
; i
++) {
4784 mark_reg_not_init(env
, regs
, caller_saved
[i
]);
4785 check_reg_arg(env
, caller_saved
[i
], DST_OP_NO_MARK
);
4789 static int check_func_call(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
,
4792 struct bpf_verifier_state
*state
= env
->cur_state
;
4793 struct bpf_func_info_aux
*func_info_aux
;
4794 struct bpf_func_state
*caller
, *callee
;
4795 int i
, err
, subprog
, target_insn
;
4796 bool is_global
= false;
4798 if (state
->curframe
+ 1 >= MAX_CALL_FRAMES
) {
4799 verbose(env
, "the call stack of %d frames is too deep\n",
4800 state
->curframe
+ 2);
4804 target_insn
= *insn_idx
+ insn
->imm
;
4805 subprog
= find_subprog(env
, target_insn
+ 1);
4807 verbose(env
, "verifier bug. No program starts at insn %d\n",
4812 caller
= state
->frame
[state
->curframe
];
4813 if (state
->frame
[state
->curframe
+ 1]) {
4814 verbose(env
, "verifier bug. Frame %d already allocated\n",
4815 state
->curframe
+ 1);
4819 func_info_aux
= env
->prog
->aux
->func_info_aux
;
4821 is_global
= func_info_aux
[subprog
].linkage
== BTF_FUNC_GLOBAL
;
4822 err
= btf_check_func_arg_match(env
, subprog
, caller
->regs
);
4827 verbose(env
, "Caller passes invalid args into func#%d\n",
4831 if (env
->log
.level
& BPF_LOG_LEVEL
)
4833 "Func#%d is global and valid. Skipping.\n",
4835 clear_caller_saved_regs(env
, caller
->regs
);
4837 /* All global functions return a 64-bit SCALAR_VALUE */
4838 mark_reg_unknown(env
, caller
->regs
, BPF_REG_0
);
4839 caller
->regs
[BPF_REG_0
].subreg_def
= DEF_NOT_SUBREG
;
4841 /* continue with next insn after call */
4846 callee
= kzalloc(sizeof(*callee
), GFP_KERNEL
);
4849 state
->frame
[state
->curframe
+ 1] = callee
;
4851 /* callee cannot access r0, r6 - r9 for reading and has to write
4852 * into its own stack before reading from it.
4853 * callee can read/write into caller's stack
4855 init_func_state(env
, callee
,
4856 /* remember the callsite, it will be used by bpf_exit */
4857 *insn_idx
/* callsite */,
4858 state
->curframe
+ 1 /* frameno within this callchain */,
4859 subprog
/* subprog number within this prog */);
4861 /* Transfer references to the callee */
4862 err
= transfer_reference_state(callee
, caller
);
4866 /* copy r1 - r5 args that callee can access. The copy includes parent
4867 * pointers, which connects us up to the liveness chain
4869 for (i
= BPF_REG_1
; i
<= BPF_REG_5
; i
++)
4870 callee
->regs
[i
] = caller
->regs
[i
];
4872 clear_caller_saved_regs(env
, caller
->regs
);
4874 /* only increment it after check_reg_arg() finished */
4877 /* and go analyze first insn of the callee */
4878 *insn_idx
= target_insn
;
4880 if (env
->log
.level
& BPF_LOG_LEVEL
) {
4881 verbose(env
, "caller:\n");
4882 print_verifier_state(env
, caller
);
4883 verbose(env
, "callee:\n");
4884 print_verifier_state(env
, callee
);
4889 static int prepare_func_exit(struct bpf_verifier_env
*env
, int *insn_idx
)
4891 struct bpf_verifier_state
*state
= env
->cur_state
;
4892 struct bpf_func_state
*caller
, *callee
;
4893 struct bpf_reg_state
*r0
;
4896 callee
= state
->frame
[state
->curframe
];
4897 r0
= &callee
->regs
[BPF_REG_0
];
4898 if (r0
->type
== PTR_TO_STACK
) {
4899 /* technically it's ok to return caller's stack pointer
4900 * (or caller's caller's pointer) back to the caller,
4901 * since these pointers are valid. Only current stack
4902 * pointer will be invalid as soon as function exits,
4903 * but let's be conservative
4905 verbose(env
, "cannot return stack pointer to the caller\n");
4910 caller
= state
->frame
[state
->curframe
];
4911 /* return to the caller whatever r0 had in the callee */
4912 caller
->regs
[BPF_REG_0
] = *r0
;
4914 /* Transfer references to the caller */
4915 err
= transfer_reference_state(caller
, callee
);
4919 *insn_idx
= callee
->callsite
+ 1;
4920 if (env
->log
.level
& BPF_LOG_LEVEL
) {
4921 verbose(env
, "returning from callee:\n");
4922 print_verifier_state(env
, callee
);
4923 verbose(env
, "to caller at %d:\n", *insn_idx
);
4924 print_verifier_state(env
, caller
);
4926 /* clear everything in the callee */
4927 free_func_state(callee
);
4928 state
->frame
[state
->curframe
+ 1] = NULL
;
4932 static void do_refine_retval_range(struct bpf_reg_state
*regs
, int ret_type
,
4934 struct bpf_call_arg_meta
*meta
)
4936 struct bpf_reg_state
*ret_reg
= ®s
[BPF_REG_0
];
4938 if (ret_type
!= RET_INTEGER
||
4939 (func_id
!= BPF_FUNC_get_stack
&&
4940 func_id
!= BPF_FUNC_probe_read_str
&&
4941 func_id
!= BPF_FUNC_probe_read_kernel_str
&&
4942 func_id
!= BPF_FUNC_probe_read_user_str
))
4945 ret_reg
->smax_value
= meta
->msize_max_value
;
4946 ret_reg
->s32_max_value
= meta
->msize_max_value
;
4947 ret_reg
->smin_value
= -MAX_ERRNO
;
4948 ret_reg
->s32_min_value
= -MAX_ERRNO
;
4949 __reg_deduce_bounds(ret_reg
);
4950 __reg_bound_offset(ret_reg
);
4951 __update_reg_bounds(ret_reg
);
4955 record_func_map(struct bpf_verifier_env
*env
, struct bpf_call_arg_meta
*meta
,
4956 int func_id
, int insn_idx
)
4958 struct bpf_insn_aux_data
*aux
= &env
->insn_aux_data
[insn_idx
];
4959 struct bpf_map
*map
= meta
->map_ptr
;
4961 if (func_id
!= BPF_FUNC_tail_call
&&
4962 func_id
!= BPF_FUNC_map_lookup_elem
&&
4963 func_id
!= BPF_FUNC_map_update_elem
&&
4964 func_id
!= BPF_FUNC_map_delete_elem
&&
4965 func_id
!= BPF_FUNC_map_push_elem
&&
4966 func_id
!= BPF_FUNC_map_pop_elem
&&
4967 func_id
!= BPF_FUNC_map_peek_elem
)
4971 verbose(env
, "kernel subsystem misconfigured verifier\n");
4975 /* In case of read-only, some additional restrictions
4976 * need to be applied in order to prevent altering the
4977 * state of the map from program side.
4979 if ((map
->map_flags
& BPF_F_RDONLY_PROG
) &&
4980 (func_id
== BPF_FUNC_map_delete_elem
||
4981 func_id
== BPF_FUNC_map_update_elem
||
4982 func_id
== BPF_FUNC_map_push_elem
||
4983 func_id
== BPF_FUNC_map_pop_elem
)) {
4984 verbose(env
, "write into map forbidden\n");
4988 if (!BPF_MAP_PTR(aux
->map_ptr_state
))
4989 bpf_map_ptr_store(aux
, meta
->map_ptr
,
4990 !meta
->map_ptr
->bypass_spec_v1
);
4991 else if (BPF_MAP_PTR(aux
->map_ptr_state
) != meta
->map_ptr
)
4992 bpf_map_ptr_store(aux
, BPF_MAP_PTR_POISON
,
4993 !meta
->map_ptr
->bypass_spec_v1
);
4998 record_func_key(struct bpf_verifier_env
*env
, struct bpf_call_arg_meta
*meta
,
4999 int func_id
, int insn_idx
)
5001 struct bpf_insn_aux_data
*aux
= &env
->insn_aux_data
[insn_idx
];
5002 struct bpf_reg_state
*regs
= cur_regs(env
), *reg
;
5003 struct bpf_map
*map
= meta
->map_ptr
;
5008 if (func_id
!= BPF_FUNC_tail_call
)
5010 if (!map
|| map
->map_type
!= BPF_MAP_TYPE_PROG_ARRAY
) {
5011 verbose(env
, "kernel subsystem misconfigured verifier\n");
5015 range
= tnum_range(0, map
->max_entries
- 1);
5016 reg
= ®s
[BPF_REG_3
];
5018 if (!register_is_const(reg
) || !tnum_in(range
, reg
->var_off
)) {
5019 bpf_map_key_store(aux
, BPF_MAP_KEY_POISON
);
5023 err
= mark_chain_precision(env
, BPF_REG_3
);
5027 val
= reg
->var_off
.value
;
5028 if (bpf_map_key_unseen(aux
))
5029 bpf_map_key_store(aux
, val
);
5030 else if (!bpf_map_key_poisoned(aux
) &&
5031 bpf_map_key_immediate(aux
) != val
)
5032 bpf_map_key_store(aux
, BPF_MAP_KEY_POISON
);
5036 static int check_reference_leak(struct bpf_verifier_env
*env
)
5038 struct bpf_func_state
*state
= cur_func(env
);
5041 for (i
= 0; i
< state
->acquired_refs
; i
++) {
5042 verbose(env
, "Unreleased reference id=%d alloc_insn=%d\n",
5043 state
->refs
[i
].id
, state
->refs
[i
].insn_idx
);
5045 return state
->acquired_refs
? -EINVAL
: 0;
5048 static int check_helper_call(struct bpf_verifier_env
*env
, int func_id
, int insn_idx
)
5050 const struct bpf_func_proto
*fn
= NULL
;
5051 struct bpf_reg_state
*regs
;
5052 struct bpf_call_arg_meta meta
;
5056 /* find function prototype */
5057 if (func_id
< 0 || func_id
>= __BPF_FUNC_MAX_ID
) {
5058 verbose(env
, "invalid func %s#%d\n", func_id_name(func_id
),
5063 if (env
->ops
->get_func_proto
)
5064 fn
= env
->ops
->get_func_proto(func_id
, env
->prog
);
5066 verbose(env
, "unknown func %s#%d\n", func_id_name(func_id
),
5071 /* eBPF programs must be GPL compatible to use GPL-ed functions */
5072 if (!env
->prog
->gpl_compatible
&& fn
->gpl_only
) {
5073 verbose(env
, "cannot call GPL-restricted function from non-GPL compatible program\n");
5077 if (fn
->allowed
&& !fn
->allowed(env
->prog
)) {
5078 verbose(env
, "helper call is not allowed in probe\n");
5082 /* With LD_ABS/IND some JITs save/restore skb from r1. */
5083 changes_data
= bpf_helper_changes_pkt_data(fn
->func
);
5084 if (changes_data
&& fn
->arg1_type
!= ARG_PTR_TO_CTX
) {
5085 verbose(env
, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
5086 func_id_name(func_id
), func_id
);
5090 memset(&meta
, 0, sizeof(meta
));
5091 meta
.pkt_access
= fn
->pkt_access
;
5093 err
= check_func_proto(fn
, func_id
);
5095 verbose(env
, "kernel subsystem misconfigured func %s#%d\n",
5096 func_id_name(func_id
), func_id
);
5100 meta
.func_id
= func_id
;
5102 for (i
= 0; i
< 5; i
++) {
5103 err
= check_func_arg(env
, i
, &meta
, fn
);
5108 err
= record_func_map(env
, &meta
, func_id
, insn_idx
);
5112 err
= record_func_key(env
, &meta
, func_id
, insn_idx
);
5116 /* Mark slots with STACK_MISC in case of raw mode, stack offset
5117 * is inferred from register state.
5119 for (i
= 0; i
< meta
.access_size
; i
++) {
5120 err
= check_mem_access(env
, insn_idx
, meta
.regno
, i
, BPF_B
,
5121 BPF_WRITE
, -1, false);
5126 if (func_id
== BPF_FUNC_tail_call
) {
5127 err
= check_reference_leak(env
);
5129 verbose(env
, "tail_call would lead to reference leak\n");
5132 } else if (is_release_function(func_id
)) {
5133 err
= release_reference(env
, meta
.ref_obj_id
);
5135 verbose(env
, "func %s#%d reference has not been acquired before\n",
5136 func_id_name(func_id
), func_id
);
5141 regs
= cur_regs(env
);
5143 /* check that flags argument in get_local_storage(map, flags) is 0,
5144 * this is required because get_local_storage() can't return an error.
5146 if (func_id
== BPF_FUNC_get_local_storage
&&
5147 !register_is_null(®s
[BPF_REG_2
])) {
5148 verbose(env
, "get_local_storage() doesn't support non-zero flags\n");
5152 /* reset caller saved regs */
5153 for (i
= 0; i
< CALLER_SAVED_REGS
; i
++) {
5154 mark_reg_not_init(env
, regs
, caller_saved
[i
]);
5155 check_reg_arg(env
, caller_saved
[i
], DST_OP_NO_MARK
);
5158 /* helper call returns 64-bit value. */
5159 regs
[BPF_REG_0
].subreg_def
= DEF_NOT_SUBREG
;
5161 /* update return register (already marked as written above) */
5162 if (fn
->ret_type
== RET_INTEGER
) {
5163 /* sets type to SCALAR_VALUE */
5164 mark_reg_unknown(env
, regs
, BPF_REG_0
);
5165 } else if (fn
->ret_type
== RET_VOID
) {
5166 regs
[BPF_REG_0
].type
= NOT_INIT
;
5167 } else if (fn
->ret_type
== RET_PTR_TO_MAP_VALUE_OR_NULL
||
5168 fn
->ret_type
== RET_PTR_TO_MAP_VALUE
) {
5169 /* There is no offset yet applied, variable or fixed */
5170 mark_reg_known_zero(env
, regs
, BPF_REG_0
);
5171 /* remember map_ptr, so that check_map_access()
5172 * can check 'value_size' boundary of memory access
5173 * to map element returned from bpf_map_lookup_elem()
5175 if (meta
.map_ptr
== NULL
) {
5177 "kernel subsystem misconfigured verifier\n");
5180 regs
[BPF_REG_0
].map_ptr
= meta
.map_ptr
;
5181 if (fn
->ret_type
== RET_PTR_TO_MAP_VALUE
) {
5182 regs
[BPF_REG_0
].type
= PTR_TO_MAP_VALUE
;
5183 if (map_value_has_spin_lock(meta
.map_ptr
))
5184 regs
[BPF_REG_0
].id
= ++env
->id_gen
;
5186 regs
[BPF_REG_0
].type
= PTR_TO_MAP_VALUE_OR_NULL
;
5188 } else if (fn
->ret_type
== RET_PTR_TO_SOCKET_OR_NULL
) {
5189 mark_reg_known_zero(env
, regs
, BPF_REG_0
);
5190 regs
[BPF_REG_0
].type
= PTR_TO_SOCKET_OR_NULL
;
5191 } else if (fn
->ret_type
== RET_PTR_TO_SOCK_COMMON_OR_NULL
) {
5192 mark_reg_known_zero(env
, regs
, BPF_REG_0
);
5193 regs
[BPF_REG_0
].type
= PTR_TO_SOCK_COMMON_OR_NULL
;
5194 } else if (fn
->ret_type
== RET_PTR_TO_TCP_SOCK_OR_NULL
) {
5195 mark_reg_known_zero(env
, regs
, BPF_REG_0
);
5196 regs
[BPF_REG_0
].type
= PTR_TO_TCP_SOCK_OR_NULL
;
5197 } else if (fn
->ret_type
== RET_PTR_TO_ALLOC_MEM_OR_NULL
) {
5198 mark_reg_known_zero(env
, regs
, BPF_REG_0
);
5199 regs
[BPF_REG_0
].type
= PTR_TO_MEM_OR_NULL
;
5200 regs
[BPF_REG_0
].mem_size
= meta
.mem_size
;
5201 } else if (fn
->ret_type
== RET_PTR_TO_MEM_OR_BTF_ID_OR_NULL
||
5202 fn
->ret_type
== RET_PTR_TO_MEM_OR_BTF_ID
) {
5203 const struct btf_type
*t
;
5205 mark_reg_known_zero(env
, regs
, BPF_REG_0
);
5206 t
= btf_type_skip_modifiers(meta
.ret_btf
, meta
.ret_btf_id
, NULL
);
5207 if (!btf_type_is_struct(t
)) {
5209 const struct btf_type
*ret
;
5212 /* resolve the type size of ksym. */
5213 ret
= btf_resolve_size(meta
.ret_btf
, t
, &tsize
);
5215 tname
= btf_name_by_offset(meta
.ret_btf
, t
->name_off
);
5216 verbose(env
, "unable to resolve the size of type '%s': %ld\n",
5217 tname
, PTR_ERR(ret
));
5220 regs
[BPF_REG_0
].type
=
5221 fn
->ret_type
== RET_PTR_TO_MEM_OR_BTF_ID
?
5222 PTR_TO_MEM
: PTR_TO_MEM_OR_NULL
;
5223 regs
[BPF_REG_0
].mem_size
= tsize
;
5225 regs
[BPF_REG_0
].type
=
5226 fn
->ret_type
== RET_PTR_TO_MEM_OR_BTF_ID
?
5227 PTR_TO_BTF_ID
: PTR_TO_BTF_ID_OR_NULL
;
5228 regs
[BPF_REG_0
].btf
= meta
.ret_btf
;
5229 regs
[BPF_REG_0
].btf_id
= meta
.ret_btf_id
;
5231 } else if (fn
->ret_type
== RET_PTR_TO_BTF_ID_OR_NULL
||
5232 fn
->ret_type
== RET_PTR_TO_BTF_ID
) {
5235 mark_reg_known_zero(env
, regs
, BPF_REG_0
);
5236 regs
[BPF_REG_0
].type
= fn
->ret_type
== RET_PTR_TO_BTF_ID
?
5238 PTR_TO_BTF_ID_OR_NULL
;
5239 ret_btf_id
= *fn
->ret_btf_id
;
5240 if (ret_btf_id
== 0) {
5241 verbose(env
, "invalid return type %d of func %s#%d\n",
5242 fn
->ret_type
, func_id_name(func_id
), func_id
);
5245 /* current BPF helper definitions are only coming from
5246 * built-in code with type IDs from vmlinux BTF
5248 regs
[BPF_REG_0
].btf
= btf_vmlinux
;
5249 regs
[BPF_REG_0
].btf_id
= ret_btf_id
;
5251 verbose(env
, "unknown return type %d of func %s#%d\n",
5252 fn
->ret_type
, func_id_name(func_id
), func_id
);
5256 if (reg_type_may_be_null(regs
[BPF_REG_0
].type
))
5257 regs
[BPF_REG_0
].id
= ++env
->id_gen
;
5259 if (is_ptr_cast_function(func_id
)) {
5260 /* For release_reference() */
5261 regs
[BPF_REG_0
].ref_obj_id
= meta
.ref_obj_id
;
5262 } else if (is_acquire_function(func_id
, meta
.map_ptr
)) {
5263 int id
= acquire_reference_state(env
, insn_idx
);
5267 /* For mark_ptr_or_null_reg() */
5268 regs
[BPF_REG_0
].id
= id
;
5269 /* For release_reference() */
5270 regs
[BPF_REG_0
].ref_obj_id
= id
;
5273 do_refine_retval_range(regs
, fn
->ret_type
, func_id
, &meta
);
5275 err
= check_map_func_compatibility(env
, meta
.map_ptr
, func_id
);
5279 if ((func_id
== BPF_FUNC_get_stack
||
5280 func_id
== BPF_FUNC_get_task_stack
) &&
5281 !env
->prog
->has_callchain_buf
) {
5282 const char *err_str
;
5284 #ifdef CONFIG_PERF_EVENTS
5285 err
= get_callchain_buffers(sysctl_perf_event_max_stack
);
5286 err_str
= "cannot get callchain buffer for func %s#%d\n";
5289 err_str
= "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
5292 verbose(env
, err_str
, func_id_name(func_id
), func_id
);
5296 env
->prog
->has_callchain_buf
= true;
5299 if (func_id
== BPF_FUNC_get_stackid
|| func_id
== BPF_FUNC_get_stack
)
5300 env
->prog
->call_get_stack
= true;
5303 clear_all_pkt_pointers(env
);
5307 static bool signed_add_overflows(s64 a
, s64 b
)
5309 /* Do the add in u64, where overflow is well-defined */
5310 s64 res
= (s64
)((u64
)a
+ (u64
)b
);
5317 static bool signed_add32_overflows(s32 a
, s32 b
)
5319 /* Do the add in u32, where overflow is well-defined */
5320 s32 res
= (s32
)((u32
)a
+ (u32
)b
);
5327 static bool signed_sub_overflows(s64 a
, s64 b
)
5329 /* Do the sub in u64, where overflow is well-defined */
5330 s64 res
= (s64
)((u64
)a
- (u64
)b
);
5337 static bool signed_sub32_overflows(s32 a
, s32 b
)
5339 /* Do the sub in u32, where overflow is well-defined */
5340 s32 res
= (s32
)((u32
)a
- (u32
)b
);
5347 static bool check_reg_sane_offset(struct bpf_verifier_env
*env
,
5348 const struct bpf_reg_state
*reg
,
5349 enum bpf_reg_type type
)
5351 bool known
= tnum_is_const(reg
->var_off
);
5352 s64 val
= reg
->var_off
.value
;
5353 s64 smin
= reg
->smin_value
;
5355 if (known
&& (val
>= BPF_MAX_VAR_OFF
|| val
<= -BPF_MAX_VAR_OFF
)) {
5356 verbose(env
, "math between %s pointer and %lld is not allowed\n",
5357 reg_type_str
[type
], val
);
5361 if (reg
->off
>= BPF_MAX_VAR_OFF
|| reg
->off
<= -BPF_MAX_VAR_OFF
) {
5362 verbose(env
, "%s pointer offset %d is not allowed\n",
5363 reg_type_str
[type
], reg
->off
);
5367 if (smin
== S64_MIN
) {
5368 verbose(env
, "math between %s pointer and register with unbounded min value is not allowed\n",
5369 reg_type_str
[type
]);
5373 if (smin
>= BPF_MAX_VAR_OFF
|| smin
<= -BPF_MAX_VAR_OFF
) {
5374 verbose(env
, "value %lld makes %s pointer be out of bounds\n",
5375 smin
, reg_type_str
[type
]);
5382 static struct bpf_insn_aux_data
*cur_aux(struct bpf_verifier_env
*env
)
5384 return &env
->insn_aux_data
[env
->insn_idx
];
5387 static int retrieve_ptr_limit(const struct bpf_reg_state
*ptr_reg
,
5388 const struct bpf_reg_state
*off_reg
,
5389 u32
*alu_limit
, u8 opcode
)
5391 bool off_is_neg
= off_reg
->smin_value
< 0;
5392 bool mask_to_left
= (opcode
== BPF_ADD
&& off_is_neg
) ||
5393 (opcode
== BPF_SUB
&& !off_is_neg
);
5394 u32 off
, max
= 0, ptr_limit
= 0;
5396 if (!tnum_is_const(off_reg
->var_off
) &&
5397 (off_reg
->smin_value
< 0) != (off_reg
->smax_value
< 0))
5400 switch (ptr_reg
->type
) {
5402 /* Offset 0 is out-of-bounds, but acceptable start for the
5403 * left direction, see BPF_REG_FP.
5405 max
= MAX_BPF_STACK
+ mask_to_left
;
5406 /* Indirect variable offset stack access is prohibited in
5407 * unprivileged mode so it's not handled here.
5409 off
= ptr_reg
->off
+ ptr_reg
->var_off
.value
;
5411 ptr_limit
= MAX_BPF_STACK
+ off
;
5413 ptr_limit
= -off
- 1;
5415 case PTR_TO_MAP_VALUE
:
5416 max
= ptr_reg
->map_ptr
->value_size
;
5418 ptr_limit
= ptr_reg
->umax_value
+ ptr_reg
->off
;
5420 off
= ptr_reg
->smin_value
+ ptr_reg
->off
;
5421 ptr_limit
= ptr_reg
->map_ptr
->value_size
- off
- 1;
5428 if (ptr_limit
>= max
)
5430 *alu_limit
= ptr_limit
;
5434 static bool can_skip_alu_sanitation(const struct bpf_verifier_env
*env
,
5435 const struct bpf_insn
*insn
)
5437 return env
->bypass_spec_v1
|| BPF_SRC(insn
->code
) == BPF_K
;
5440 static int update_alu_sanitation_state(struct bpf_insn_aux_data
*aux
,
5441 u32 alu_state
, u32 alu_limit
)
5443 /* If we arrived here from different branches with different
5444 * state or limits to sanitize, then this won't work.
5446 if (aux
->alu_state
&&
5447 (aux
->alu_state
!= alu_state
||
5448 aux
->alu_limit
!= alu_limit
))
5451 /* Corresponding fixup done in fixup_bpf_calls(). */
5452 aux
->alu_state
= alu_state
;
5453 aux
->alu_limit
= alu_limit
;
5457 static int sanitize_val_alu(struct bpf_verifier_env
*env
,
5458 struct bpf_insn
*insn
)
5460 struct bpf_insn_aux_data
*aux
= cur_aux(env
);
5462 if (can_skip_alu_sanitation(env
, insn
))
5465 return update_alu_sanitation_state(aux
, BPF_ALU_NON_POINTER
, 0);
5468 static int sanitize_ptr_alu(struct bpf_verifier_env
*env
,
5469 struct bpf_insn
*insn
,
5470 const struct bpf_reg_state
*ptr_reg
,
5471 const struct bpf_reg_state
*off_reg
,
5472 struct bpf_reg_state
*dst_reg
)
5474 struct bpf_verifier_state
*vstate
= env
->cur_state
;
5475 struct bpf_insn_aux_data
*aux
= cur_aux(env
);
5476 bool off_is_neg
= off_reg
->smin_value
< 0;
5477 bool ptr_is_dst_reg
= ptr_reg
== dst_reg
;
5478 u8 opcode
= BPF_OP(insn
->code
);
5479 u32 alu_state
, alu_limit
;
5480 struct bpf_reg_state tmp
;
5484 if (can_skip_alu_sanitation(env
, insn
))
5487 /* We already marked aux for masking from non-speculative
5488 * paths, thus we got here in the first place. We only care
5489 * to explore bad access from here.
5491 if (vstate
->speculative
)
5494 alu_state
= off_is_neg
? BPF_ALU_NEG_VALUE
: 0;
5495 alu_state
|= ptr_is_dst_reg
?
5496 BPF_ALU_SANITIZE_SRC
: BPF_ALU_SANITIZE_DST
;
5498 err
= retrieve_ptr_limit(ptr_reg
, off_reg
, &alu_limit
, opcode
);
5502 err
= update_alu_sanitation_state(aux
, alu_state
, alu_limit
);
5506 /* Simulate and find potential out-of-bounds access under
5507 * speculative execution from truncation as a result of
5508 * masking when off was not within expected range. If off
5509 * sits in dst, then we temporarily need to move ptr there
5510 * to simulate dst (== 0) +/-= ptr. Needed, for example,
5511 * for cases where we use K-based arithmetic in one direction
5512 * and truncated reg-based in the other in order to explore
5515 if (!ptr_is_dst_reg
) {
5517 *dst_reg
= *ptr_reg
;
5519 ret
= push_stack(env
, env
->insn_idx
+ 1, env
->insn_idx
, true);
5520 if (!ptr_is_dst_reg
&& ret
)
5522 return !ret
? -EFAULT
: 0;
5525 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
5526 * Caller should also handle BPF_MOV case separately.
5527 * If we return -EACCES, caller may want to try again treating pointer as a
5528 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
5530 static int adjust_ptr_min_max_vals(struct bpf_verifier_env
*env
,
5531 struct bpf_insn
*insn
,
5532 const struct bpf_reg_state
*ptr_reg
,
5533 const struct bpf_reg_state
*off_reg
)
5535 struct bpf_verifier_state
*vstate
= env
->cur_state
;
5536 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
5537 struct bpf_reg_state
*regs
= state
->regs
, *dst_reg
;
5538 bool known
= tnum_is_const(off_reg
->var_off
);
5539 s64 smin_val
= off_reg
->smin_value
, smax_val
= off_reg
->smax_value
,
5540 smin_ptr
= ptr_reg
->smin_value
, smax_ptr
= ptr_reg
->smax_value
;
5541 u64 umin_val
= off_reg
->umin_value
, umax_val
= off_reg
->umax_value
,
5542 umin_ptr
= ptr_reg
->umin_value
, umax_ptr
= ptr_reg
->umax_value
;
5543 u8 opcode
= BPF_OP(insn
->code
);
5544 u32 dst
= insn
->dst_reg
;
5547 dst_reg
= ®s
[dst
];
5549 if ((known
&& (smin_val
!= smax_val
|| umin_val
!= umax_val
)) ||
5550 smin_val
> smax_val
|| umin_val
> umax_val
) {
5551 /* Taint dst register if offset had invalid bounds derived from
5552 * e.g. dead branches.
5554 __mark_reg_unknown(env
, dst_reg
);
5558 if (BPF_CLASS(insn
->code
) != BPF_ALU64
) {
5559 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
5560 if (opcode
== BPF_SUB
&& env
->allow_ptr_leaks
) {
5561 __mark_reg_unknown(env
, dst_reg
);
5566 "R%d 32-bit pointer arithmetic prohibited\n",
5571 switch (ptr_reg
->type
) {
5572 case PTR_TO_MAP_VALUE_OR_NULL
:
5573 verbose(env
, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
5574 dst
, reg_type_str
[ptr_reg
->type
]);
5576 case CONST_PTR_TO_MAP
:
5577 /* smin_val represents the known value */
5578 if (known
&& smin_val
== 0 && opcode
== BPF_ADD
)
5581 case PTR_TO_PACKET_END
:
5583 case PTR_TO_SOCKET_OR_NULL
:
5584 case PTR_TO_SOCK_COMMON
:
5585 case PTR_TO_SOCK_COMMON_OR_NULL
:
5586 case PTR_TO_TCP_SOCK
:
5587 case PTR_TO_TCP_SOCK_OR_NULL
:
5588 case PTR_TO_XDP_SOCK
:
5589 verbose(env
, "R%d pointer arithmetic on %s prohibited\n",
5590 dst
, reg_type_str
[ptr_reg
->type
]);
5596 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
5597 * The id may be overwritten later if we create a new variable offset.
5599 dst_reg
->type
= ptr_reg
->type
;
5600 dst_reg
->id
= ptr_reg
->id
;
5602 if (!check_reg_sane_offset(env
, off_reg
, ptr_reg
->type
) ||
5603 !check_reg_sane_offset(env
, ptr_reg
, ptr_reg
->type
))
5606 /* pointer types do not carry 32-bit bounds at the moment. */
5607 __mark_reg32_unbounded(dst_reg
);
5611 ret
= sanitize_ptr_alu(env
, insn
, ptr_reg
, off_reg
, dst_reg
);
5613 verbose(env
, "R%d tried to add from different maps, paths, or prohibited types\n", dst
);
5616 /* We can take a fixed offset as long as it doesn't overflow
5617 * the s32 'off' field
5619 if (known
&& (ptr_reg
->off
+ smin_val
==
5620 (s64
)(s32
)(ptr_reg
->off
+ smin_val
))) {
5621 /* pointer += K. Accumulate it into fixed offset */
5622 dst_reg
->smin_value
= smin_ptr
;
5623 dst_reg
->smax_value
= smax_ptr
;
5624 dst_reg
->umin_value
= umin_ptr
;
5625 dst_reg
->umax_value
= umax_ptr
;
5626 dst_reg
->var_off
= ptr_reg
->var_off
;
5627 dst_reg
->off
= ptr_reg
->off
+ smin_val
;
5628 dst_reg
->raw
= ptr_reg
->raw
;
5631 /* A new variable offset is created. Note that off_reg->off
5632 * == 0, since it's a scalar.
5633 * dst_reg gets the pointer type and since some positive
5634 * integer value was added to the pointer, give it a new 'id'
5635 * if it's a PTR_TO_PACKET.
5636 * this creates a new 'base' pointer, off_reg (variable) gets
5637 * added into the variable offset, and we copy the fixed offset
5640 if (signed_add_overflows(smin_ptr
, smin_val
) ||
5641 signed_add_overflows(smax_ptr
, smax_val
)) {
5642 dst_reg
->smin_value
= S64_MIN
;
5643 dst_reg
->smax_value
= S64_MAX
;
5645 dst_reg
->smin_value
= smin_ptr
+ smin_val
;
5646 dst_reg
->smax_value
= smax_ptr
+ smax_val
;
5648 if (umin_ptr
+ umin_val
< umin_ptr
||
5649 umax_ptr
+ umax_val
< umax_ptr
) {
5650 dst_reg
->umin_value
= 0;
5651 dst_reg
->umax_value
= U64_MAX
;
5653 dst_reg
->umin_value
= umin_ptr
+ umin_val
;
5654 dst_reg
->umax_value
= umax_ptr
+ umax_val
;
5656 dst_reg
->var_off
= tnum_add(ptr_reg
->var_off
, off_reg
->var_off
);
5657 dst_reg
->off
= ptr_reg
->off
;
5658 dst_reg
->raw
= ptr_reg
->raw
;
5659 if (reg_is_pkt_pointer(ptr_reg
)) {
5660 dst_reg
->id
= ++env
->id_gen
;
5661 /* something was added to pkt_ptr, set range to zero */
5662 memset(&dst_reg
->raw
, 0, sizeof(dst_reg
->raw
));
5666 ret
= sanitize_ptr_alu(env
, insn
, ptr_reg
, off_reg
, dst_reg
);
5668 verbose(env
, "R%d tried to sub from different maps, paths, or prohibited types\n", dst
);
5671 if (dst_reg
== off_reg
) {
5672 /* scalar -= pointer. Creates an unknown scalar */
5673 verbose(env
, "R%d tried to subtract pointer from scalar\n",
5677 /* We don't allow subtraction from FP, because (according to
5678 * test_verifier.c test "invalid fp arithmetic", JITs might not
5679 * be able to deal with it.
5681 if (ptr_reg
->type
== PTR_TO_STACK
) {
5682 verbose(env
, "R%d subtraction from stack pointer prohibited\n",
5686 if (known
&& (ptr_reg
->off
- smin_val
==
5687 (s64
)(s32
)(ptr_reg
->off
- smin_val
))) {
5688 /* pointer -= K. Subtract it from fixed offset */
5689 dst_reg
->smin_value
= smin_ptr
;
5690 dst_reg
->smax_value
= smax_ptr
;
5691 dst_reg
->umin_value
= umin_ptr
;
5692 dst_reg
->umax_value
= umax_ptr
;
5693 dst_reg
->var_off
= ptr_reg
->var_off
;
5694 dst_reg
->id
= ptr_reg
->id
;
5695 dst_reg
->off
= ptr_reg
->off
- smin_val
;
5696 dst_reg
->raw
= ptr_reg
->raw
;
5699 /* A new variable offset is created. If the subtrahend is known
5700 * nonnegative, then any reg->range we had before is still good.
5702 if (signed_sub_overflows(smin_ptr
, smax_val
) ||
5703 signed_sub_overflows(smax_ptr
, smin_val
)) {
5704 /* Overflow possible, we know nothing */
5705 dst_reg
->smin_value
= S64_MIN
;
5706 dst_reg
->smax_value
= S64_MAX
;
5708 dst_reg
->smin_value
= smin_ptr
- smax_val
;
5709 dst_reg
->smax_value
= smax_ptr
- smin_val
;
5711 if (umin_ptr
< umax_val
) {
5712 /* Overflow possible, we know nothing */
5713 dst_reg
->umin_value
= 0;
5714 dst_reg
->umax_value
= U64_MAX
;
5716 /* Cannot overflow (as long as bounds are consistent) */
5717 dst_reg
->umin_value
= umin_ptr
- umax_val
;
5718 dst_reg
->umax_value
= umax_ptr
- umin_val
;
5720 dst_reg
->var_off
= tnum_sub(ptr_reg
->var_off
, off_reg
->var_off
);
5721 dst_reg
->off
= ptr_reg
->off
;
5722 dst_reg
->raw
= ptr_reg
->raw
;
5723 if (reg_is_pkt_pointer(ptr_reg
)) {
5724 dst_reg
->id
= ++env
->id_gen
;
5725 /* something was added to pkt_ptr, set range to zero */
5727 memset(&dst_reg
->raw
, 0, sizeof(dst_reg
->raw
));
5733 /* bitwise ops on pointers are troublesome, prohibit. */
5734 verbose(env
, "R%d bitwise operator %s on pointer prohibited\n",
5735 dst
, bpf_alu_string
[opcode
>> 4]);
5738 /* other operators (e.g. MUL,LSH) produce non-pointer results */
5739 verbose(env
, "R%d pointer arithmetic with %s operator prohibited\n",
5740 dst
, bpf_alu_string
[opcode
>> 4]);
5744 if (!check_reg_sane_offset(env
, dst_reg
, ptr_reg
->type
))
5747 __update_reg_bounds(dst_reg
);
5748 __reg_deduce_bounds(dst_reg
);
5749 __reg_bound_offset(dst_reg
);
5751 /* For unprivileged we require that resulting offset must be in bounds
5752 * in order to be able to sanitize access later on.
5754 if (!env
->bypass_spec_v1
) {
5755 if (dst_reg
->type
== PTR_TO_MAP_VALUE
&&
5756 check_map_access(env
, dst
, dst_reg
->off
, 1, false)) {
5757 verbose(env
, "R%d pointer arithmetic of map value goes out of range, "
5758 "prohibited for !root\n", dst
);
5760 } else if (dst_reg
->type
== PTR_TO_STACK
&&
5761 check_stack_access(env
, dst_reg
, dst_reg
->off
+
5762 dst_reg
->var_off
.value
, 1)) {
5763 verbose(env
, "R%d stack pointer arithmetic goes out of range, "
5764 "prohibited for !root\n", dst
);
5772 static void scalar32_min_max_add(struct bpf_reg_state
*dst_reg
,
5773 struct bpf_reg_state
*src_reg
)
5775 s32 smin_val
= src_reg
->s32_min_value
;
5776 s32 smax_val
= src_reg
->s32_max_value
;
5777 u32 umin_val
= src_reg
->u32_min_value
;
5778 u32 umax_val
= src_reg
->u32_max_value
;
5780 if (signed_add32_overflows(dst_reg
->s32_min_value
, smin_val
) ||
5781 signed_add32_overflows(dst_reg
->s32_max_value
, smax_val
)) {
5782 dst_reg
->s32_min_value
= S32_MIN
;
5783 dst_reg
->s32_max_value
= S32_MAX
;
5785 dst_reg
->s32_min_value
+= smin_val
;
5786 dst_reg
->s32_max_value
+= smax_val
;
5788 if (dst_reg
->u32_min_value
+ umin_val
< umin_val
||
5789 dst_reg
->u32_max_value
+ umax_val
< umax_val
) {
5790 dst_reg
->u32_min_value
= 0;
5791 dst_reg
->u32_max_value
= U32_MAX
;
5793 dst_reg
->u32_min_value
+= umin_val
;
5794 dst_reg
->u32_max_value
+= umax_val
;
5798 static void scalar_min_max_add(struct bpf_reg_state
*dst_reg
,
5799 struct bpf_reg_state
*src_reg
)
5801 s64 smin_val
= src_reg
->smin_value
;
5802 s64 smax_val
= src_reg
->smax_value
;
5803 u64 umin_val
= src_reg
->umin_value
;
5804 u64 umax_val
= src_reg
->umax_value
;
5806 if (signed_add_overflows(dst_reg
->smin_value
, smin_val
) ||
5807 signed_add_overflows(dst_reg
->smax_value
, smax_val
)) {
5808 dst_reg
->smin_value
= S64_MIN
;
5809 dst_reg
->smax_value
= S64_MAX
;
5811 dst_reg
->smin_value
+= smin_val
;
5812 dst_reg
->smax_value
+= smax_val
;
5814 if (dst_reg
->umin_value
+ umin_val
< umin_val
||
5815 dst_reg
->umax_value
+ umax_val
< umax_val
) {
5816 dst_reg
->umin_value
= 0;
5817 dst_reg
->umax_value
= U64_MAX
;
5819 dst_reg
->umin_value
+= umin_val
;
5820 dst_reg
->umax_value
+= umax_val
;
5824 static void scalar32_min_max_sub(struct bpf_reg_state
*dst_reg
,
5825 struct bpf_reg_state
*src_reg
)
5827 s32 smin_val
= src_reg
->s32_min_value
;
5828 s32 smax_val
= src_reg
->s32_max_value
;
5829 u32 umin_val
= src_reg
->u32_min_value
;
5830 u32 umax_val
= src_reg
->u32_max_value
;
5832 if (signed_sub32_overflows(dst_reg
->s32_min_value
, smax_val
) ||
5833 signed_sub32_overflows(dst_reg
->s32_max_value
, smin_val
)) {
5834 /* Overflow possible, we know nothing */
5835 dst_reg
->s32_min_value
= S32_MIN
;
5836 dst_reg
->s32_max_value
= S32_MAX
;
5838 dst_reg
->s32_min_value
-= smax_val
;
5839 dst_reg
->s32_max_value
-= smin_val
;
5841 if (dst_reg
->u32_min_value
< umax_val
) {
5842 /* Overflow possible, we know nothing */
5843 dst_reg
->u32_min_value
= 0;
5844 dst_reg
->u32_max_value
= U32_MAX
;
5846 /* Cannot overflow (as long as bounds are consistent) */
5847 dst_reg
->u32_min_value
-= umax_val
;
5848 dst_reg
->u32_max_value
-= umin_val
;
5852 static void scalar_min_max_sub(struct bpf_reg_state
*dst_reg
,
5853 struct bpf_reg_state
*src_reg
)
5855 s64 smin_val
= src_reg
->smin_value
;
5856 s64 smax_val
= src_reg
->smax_value
;
5857 u64 umin_val
= src_reg
->umin_value
;
5858 u64 umax_val
= src_reg
->umax_value
;
5860 if (signed_sub_overflows(dst_reg
->smin_value
, smax_val
) ||
5861 signed_sub_overflows(dst_reg
->smax_value
, smin_val
)) {
5862 /* Overflow possible, we know nothing */
5863 dst_reg
->smin_value
= S64_MIN
;
5864 dst_reg
->smax_value
= S64_MAX
;
5866 dst_reg
->smin_value
-= smax_val
;
5867 dst_reg
->smax_value
-= smin_val
;
5869 if (dst_reg
->umin_value
< umax_val
) {
5870 /* Overflow possible, we know nothing */
5871 dst_reg
->umin_value
= 0;
5872 dst_reg
->umax_value
= U64_MAX
;
5874 /* Cannot overflow (as long as bounds are consistent) */
5875 dst_reg
->umin_value
-= umax_val
;
5876 dst_reg
->umax_value
-= umin_val
;
5880 static void scalar32_min_max_mul(struct bpf_reg_state
*dst_reg
,
5881 struct bpf_reg_state
*src_reg
)
5883 s32 smin_val
= src_reg
->s32_min_value
;
5884 u32 umin_val
= src_reg
->u32_min_value
;
5885 u32 umax_val
= src_reg
->u32_max_value
;
5887 if (smin_val
< 0 || dst_reg
->s32_min_value
< 0) {
5888 /* Ain't nobody got time to multiply that sign */
5889 __mark_reg32_unbounded(dst_reg
);
5892 /* Both values are positive, so we can work with unsigned and
5893 * copy the result to signed (unless it exceeds S32_MAX).
5895 if (umax_val
> U16_MAX
|| dst_reg
->u32_max_value
> U16_MAX
) {
5896 /* Potential overflow, we know nothing */
5897 __mark_reg32_unbounded(dst_reg
);
5900 dst_reg
->u32_min_value
*= umin_val
;
5901 dst_reg
->u32_max_value
*= umax_val
;
5902 if (dst_reg
->u32_max_value
> S32_MAX
) {
5903 /* Overflow possible, we know nothing */
5904 dst_reg
->s32_min_value
= S32_MIN
;
5905 dst_reg
->s32_max_value
= S32_MAX
;
5907 dst_reg
->s32_min_value
= dst_reg
->u32_min_value
;
5908 dst_reg
->s32_max_value
= dst_reg
->u32_max_value
;
5912 static void scalar_min_max_mul(struct bpf_reg_state
*dst_reg
,
5913 struct bpf_reg_state
*src_reg
)
5915 s64 smin_val
= src_reg
->smin_value
;
5916 u64 umin_val
= src_reg
->umin_value
;
5917 u64 umax_val
= src_reg
->umax_value
;
5919 if (smin_val
< 0 || dst_reg
->smin_value
< 0) {
5920 /* Ain't nobody got time to multiply that sign */
5921 __mark_reg64_unbounded(dst_reg
);
5924 /* Both values are positive, so we can work with unsigned and
5925 * copy the result to signed (unless it exceeds S64_MAX).
5927 if (umax_val
> U32_MAX
|| dst_reg
->umax_value
> U32_MAX
) {
5928 /* Potential overflow, we know nothing */
5929 __mark_reg64_unbounded(dst_reg
);
5932 dst_reg
->umin_value
*= umin_val
;
5933 dst_reg
->umax_value
*= umax_val
;
5934 if (dst_reg
->umax_value
> S64_MAX
) {
5935 /* Overflow possible, we know nothing */
5936 dst_reg
->smin_value
= S64_MIN
;
5937 dst_reg
->smax_value
= S64_MAX
;
5939 dst_reg
->smin_value
= dst_reg
->umin_value
;
5940 dst_reg
->smax_value
= dst_reg
->umax_value
;
5944 static void scalar32_min_max_and(struct bpf_reg_state
*dst_reg
,
5945 struct bpf_reg_state
*src_reg
)
5947 struct tnum var32_off
= tnum_subreg(dst_reg
->var_off
);
5948 s32 smin_val
= src_reg
->s32_min_value
;
5949 u32 umax_val
= src_reg
->u32_max_value
;
5951 /* We get our minimum from the var_off, since that's inherently
5952 * bitwise. Our maximum is the minimum of the operands' maxima.
5954 dst_reg
->u32_min_value
= var32_off
.value
;
5955 dst_reg
->u32_max_value
= min(dst_reg
->u32_max_value
, umax_val
);
5956 if (dst_reg
->s32_min_value
< 0 || smin_val
< 0) {
5957 /* Lose signed bounds when ANDing negative numbers,
5958 * ain't nobody got time for that.
5960 dst_reg
->s32_min_value
= S32_MIN
;
5961 dst_reg
->s32_max_value
= S32_MAX
;
5963 /* ANDing two positives gives a positive, so safe to
5964 * cast result into s64.
5966 dst_reg
->s32_min_value
= dst_reg
->u32_min_value
;
5967 dst_reg
->s32_max_value
= dst_reg
->u32_max_value
;
5972 static void scalar_min_max_and(struct bpf_reg_state
*dst_reg
,
5973 struct bpf_reg_state
*src_reg
)
5975 bool src_known
= tnum_is_const(src_reg
->var_off
);
5976 bool dst_known
= tnum_is_const(dst_reg
->var_off
);
5977 s64 smin_val
= src_reg
->smin_value
;
5978 u64 umax_val
= src_reg
->umax_value
;
5980 if (src_known
&& dst_known
) {
5981 __mark_reg_known(dst_reg
, dst_reg
->var_off
.value
);
5985 /* We get our minimum from the var_off, since that's inherently
5986 * bitwise. Our maximum is the minimum of the operands' maxima.
5988 dst_reg
->umin_value
= dst_reg
->var_off
.value
;
5989 dst_reg
->umax_value
= min(dst_reg
->umax_value
, umax_val
);
5990 if (dst_reg
->smin_value
< 0 || smin_val
< 0) {
5991 /* Lose signed bounds when ANDing negative numbers,
5992 * ain't nobody got time for that.
5994 dst_reg
->smin_value
= S64_MIN
;
5995 dst_reg
->smax_value
= S64_MAX
;
5997 /* ANDing two positives gives a positive, so safe to
5998 * cast result into s64.
6000 dst_reg
->smin_value
= dst_reg
->umin_value
;
6001 dst_reg
->smax_value
= dst_reg
->umax_value
;
6003 /* We may learn something more from the var_off */
6004 __update_reg_bounds(dst_reg
);
6007 static void scalar32_min_max_or(struct bpf_reg_state
*dst_reg
,
6008 struct bpf_reg_state
*src_reg
)
6010 struct tnum var32_off
= tnum_subreg(dst_reg
->var_off
);
6011 s32 smin_val
= src_reg
->s32_min_value
;
6012 u32 umin_val
= src_reg
->u32_min_value
;
6014 /* We get our maximum from the var_off, and our minimum is the
6015 * maximum of the operands' minima
6017 dst_reg
->u32_min_value
= max(dst_reg
->u32_min_value
, umin_val
);
6018 dst_reg
->u32_max_value
= var32_off
.value
| var32_off
.mask
;
6019 if (dst_reg
->s32_min_value
< 0 || smin_val
< 0) {
6020 /* Lose signed bounds when ORing negative numbers,
6021 * ain't nobody got time for that.
6023 dst_reg
->s32_min_value
= S32_MIN
;
6024 dst_reg
->s32_max_value
= S32_MAX
;
6026 /* ORing two positives gives a positive, so safe to
6027 * cast result into s64.
6029 dst_reg
->s32_min_value
= dst_reg
->u32_min_value
;
6030 dst_reg
->s32_max_value
= dst_reg
->u32_max_value
;
6034 static void scalar_min_max_or(struct bpf_reg_state
*dst_reg
,
6035 struct bpf_reg_state
*src_reg
)
6037 bool src_known
= tnum_is_const(src_reg
->var_off
);
6038 bool dst_known
= tnum_is_const(dst_reg
->var_off
);
6039 s64 smin_val
= src_reg
->smin_value
;
6040 u64 umin_val
= src_reg
->umin_value
;
6042 if (src_known
&& dst_known
) {
6043 __mark_reg_known(dst_reg
, dst_reg
->var_off
.value
);
6047 /* We get our maximum from the var_off, and our minimum is the
6048 * maximum of the operands' minima
6050 dst_reg
->umin_value
= max(dst_reg
->umin_value
, umin_val
);
6051 dst_reg
->umax_value
= dst_reg
->var_off
.value
| dst_reg
->var_off
.mask
;
6052 if (dst_reg
->smin_value
< 0 || smin_val
< 0) {
6053 /* Lose signed bounds when ORing negative numbers,
6054 * ain't nobody got time for that.
6056 dst_reg
->smin_value
= S64_MIN
;
6057 dst_reg
->smax_value
= S64_MAX
;
6059 /* ORing two positives gives a positive, so safe to
6060 * cast result into s64.
6062 dst_reg
->smin_value
= dst_reg
->umin_value
;
6063 dst_reg
->smax_value
= dst_reg
->umax_value
;
6065 /* We may learn something more from the var_off */
6066 __update_reg_bounds(dst_reg
);
6069 static void scalar32_min_max_xor(struct bpf_reg_state
*dst_reg
,
6070 struct bpf_reg_state
*src_reg
)
6072 struct tnum var32_off
= tnum_subreg(dst_reg
->var_off
);
6073 s32 smin_val
= src_reg
->s32_min_value
;
6075 /* We get both minimum and maximum from the var32_off. */
6076 dst_reg
->u32_min_value
= var32_off
.value
;
6077 dst_reg
->u32_max_value
= var32_off
.value
| var32_off
.mask
;
6079 if (dst_reg
->s32_min_value
>= 0 && smin_val
>= 0) {
6080 /* XORing two positive sign numbers gives a positive,
6081 * so safe to cast u32 result into s32.
6083 dst_reg
->s32_min_value
= dst_reg
->u32_min_value
;
6084 dst_reg
->s32_max_value
= dst_reg
->u32_max_value
;
6086 dst_reg
->s32_min_value
= S32_MIN
;
6087 dst_reg
->s32_max_value
= S32_MAX
;
6091 static void scalar_min_max_xor(struct bpf_reg_state
*dst_reg
,
6092 struct bpf_reg_state
*src_reg
)
6094 bool src_known
= tnum_is_const(src_reg
->var_off
);
6095 bool dst_known
= tnum_is_const(dst_reg
->var_off
);
6096 s64 smin_val
= src_reg
->smin_value
;
6098 if (src_known
&& dst_known
) {
6099 /* dst_reg->var_off.value has been updated earlier */
6100 __mark_reg_known(dst_reg
, dst_reg
->var_off
.value
);
6104 /* We get both minimum and maximum from the var_off. */
6105 dst_reg
->umin_value
= dst_reg
->var_off
.value
;
6106 dst_reg
->umax_value
= dst_reg
->var_off
.value
| dst_reg
->var_off
.mask
;
6108 if (dst_reg
->smin_value
>= 0 && smin_val
>= 0) {
6109 /* XORing two positive sign numbers gives a positive,
6110 * so safe to cast u64 result into s64.
6112 dst_reg
->smin_value
= dst_reg
->umin_value
;
6113 dst_reg
->smax_value
= dst_reg
->umax_value
;
6115 dst_reg
->smin_value
= S64_MIN
;
6116 dst_reg
->smax_value
= S64_MAX
;
6119 __update_reg_bounds(dst_reg
);
6122 static void __scalar32_min_max_lsh(struct bpf_reg_state
*dst_reg
,
6123 u64 umin_val
, u64 umax_val
)
6125 /* We lose all sign bit information (except what we can pick
6128 dst_reg
->s32_min_value
= S32_MIN
;
6129 dst_reg
->s32_max_value
= S32_MAX
;
6130 /* If we might shift our top bit out, then we know nothing */
6131 if (umax_val
> 31 || dst_reg
->u32_max_value
> 1ULL << (31 - umax_val
)) {
6132 dst_reg
->u32_min_value
= 0;
6133 dst_reg
->u32_max_value
= U32_MAX
;
6135 dst_reg
->u32_min_value
<<= umin_val
;
6136 dst_reg
->u32_max_value
<<= umax_val
;
6140 static void scalar32_min_max_lsh(struct bpf_reg_state
*dst_reg
,
6141 struct bpf_reg_state
*src_reg
)
6143 u32 umax_val
= src_reg
->u32_max_value
;
6144 u32 umin_val
= src_reg
->u32_min_value
;
6145 /* u32 alu operation will zext upper bits */
6146 struct tnum subreg
= tnum_subreg(dst_reg
->var_off
);
6148 __scalar32_min_max_lsh(dst_reg
, umin_val
, umax_val
);
6149 dst_reg
->var_off
= tnum_subreg(tnum_lshift(subreg
, umin_val
));
6150 /* Not required but being careful mark reg64 bounds as unknown so
6151 * that we are forced to pick them up from tnum and zext later and
6152 * if some path skips this step we are still safe.
6154 __mark_reg64_unbounded(dst_reg
);
6155 __update_reg32_bounds(dst_reg
);
6158 static void __scalar64_min_max_lsh(struct bpf_reg_state
*dst_reg
,
6159 u64 umin_val
, u64 umax_val
)
6161 /* Special case <<32 because it is a common compiler pattern to sign
6162 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
6163 * positive we know this shift will also be positive so we can track
6164 * bounds correctly. Otherwise we lose all sign bit information except
6165 * what we can pick up from var_off. Perhaps we can generalize this
6166 * later to shifts of any length.
6168 if (umin_val
== 32 && umax_val
== 32 && dst_reg
->s32_max_value
>= 0)
6169 dst_reg
->smax_value
= (s64
)dst_reg
->s32_max_value
<< 32;
6171 dst_reg
->smax_value
= S64_MAX
;
6173 if (umin_val
== 32 && umax_val
== 32 && dst_reg
->s32_min_value
>= 0)
6174 dst_reg
->smin_value
= (s64
)dst_reg
->s32_min_value
<< 32;
6176 dst_reg
->smin_value
= S64_MIN
;
6178 /* If we might shift our top bit out, then we know nothing */
6179 if (dst_reg
->umax_value
> 1ULL << (63 - umax_val
)) {
6180 dst_reg
->umin_value
= 0;
6181 dst_reg
->umax_value
= U64_MAX
;
6183 dst_reg
->umin_value
<<= umin_val
;
6184 dst_reg
->umax_value
<<= umax_val
;
6188 static void scalar_min_max_lsh(struct bpf_reg_state
*dst_reg
,
6189 struct bpf_reg_state
*src_reg
)
6191 u64 umax_val
= src_reg
->umax_value
;
6192 u64 umin_val
= src_reg
->umin_value
;
6194 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
6195 __scalar64_min_max_lsh(dst_reg
, umin_val
, umax_val
);
6196 __scalar32_min_max_lsh(dst_reg
, umin_val
, umax_val
);
6198 dst_reg
->var_off
= tnum_lshift(dst_reg
->var_off
, umin_val
);
6199 /* We may learn something more from the var_off */
6200 __update_reg_bounds(dst_reg
);
6203 static void scalar32_min_max_rsh(struct bpf_reg_state
*dst_reg
,
6204 struct bpf_reg_state
*src_reg
)
6206 struct tnum subreg
= tnum_subreg(dst_reg
->var_off
);
6207 u32 umax_val
= src_reg
->u32_max_value
;
6208 u32 umin_val
= src_reg
->u32_min_value
;
6210 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
6211 * be negative, then either:
6212 * 1) src_reg might be zero, so the sign bit of the result is
6213 * unknown, so we lose our signed bounds
6214 * 2) it's known negative, thus the unsigned bounds capture the
6216 * 3) the signed bounds cross zero, so they tell us nothing
6218 * If the value in dst_reg is known nonnegative, then again the
6219 * unsigned bounts capture the signed bounds.
6220 * Thus, in all cases it suffices to blow away our signed bounds
6221 * and rely on inferring new ones from the unsigned bounds and
6222 * var_off of the result.
6224 dst_reg
->s32_min_value
= S32_MIN
;
6225 dst_reg
->s32_max_value
= S32_MAX
;
6227 dst_reg
->var_off
= tnum_rshift(subreg
, umin_val
);
6228 dst_reg
->u32_min_value
>>= umax_val
;
6229 dst_reg
->u32_max_value
>>= umin_val
;
6231 __mark_reg64_unbounded(dst_reg
);
6232 __update_reg32_bounds(dst_reg
);
6235 static void scalar_min_max_rsh(struct bpf_reg_state
*dst_reg
,
6236 struct bpf_reg_state
*src_reg
)
6238 u64 umax_val
= src_reg
->umax_value
;
6239 u64 umin_val
= src_reg
->umin_value
;
6241 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
6242 * be negative, then either:
6243 * 1) src_reg might be zero, so the sign bit of the result is
6244 * unknown, so we lose our signed bounds
6245 * 2) it's known negative, thus the unsigned bounds capture the
6247 * 3) the signed bounds cross zero, so they tell us nothing
6249 * If the value in dst_reg is known nonnegative, then again the
6250 * unsigned bounts capture the signed bounds.
6251 * Thus, in all cases it suffices to blow away our signed bounds
6252 * and rely on inferring new ones from the unsigned bounds and
6253 * var_off of the result.
6255 dst_reg
->smin_value
= S64_MIN
;
6256 dst_reg
->smax_value
= S64_MAX
;
6257 dst_reg
->var_off
= tnum_rshift(dst_reg
->var_off
, umin_val
);
6258 dst_reg
->umin_value
>>= umax_val
;
6259 dst_reg
->umax_value
>>= umin_val
;
6261 /* Its not easy to operate on alu32 bounds here because it depends
6262 * on bits being shifted in. Take easy way out and mark unbounded
6263 * so we can recalculate later from tnum.
6265 __mark_reg32_unbounded(dst_reg
);
6266 __update_reg_bounds(dst_reg
);
6269 static void scalar32_min_max_arsh(struct bpf_reg_state
*dst_reg
,
6270 struct bpf_reg_state
*src_reg
)
6272 u64 umin_val
= src_reg
->u32_min_value
;
6274 /* Upon reaching here, src_known is true and
6275 * umax_val is equal to umin_val.
6277 dst_reg
->s32_min_value
= (u32
)(((s32
)dst_reg
->s32_min_value
) >> umin_val
);
6278 dst_reg
->s32_max_value
= (u32
)(((s32
)dst_reg
->s32_max_value
) >> umin_val
);
6280 dst_reg
->var_off
= tnum_arshift(tnum_subreg(dst_reg
->var_off
), umin_val
, 32);
6282 /* blow away the dst_reg umin_value/umax_value and rely on
6283 * dst_reg var_off to refine the result.
6285 dst_reg
->u32_min_value
= 0;
6286 dst_reg
->u32_max_value
= U32_MAX
;
6288 __mark_reg64_unbounded(dst_reg
);
6289 __update_reg32_bounds(dst_reg
);
6292 static void scalar_min_max_arsh(struct bpf_reg_state
*dst_reg
,
6293 struct bpf_reg_state
*src_reg
)
6295 u64 umin_val
= src_reg
->umin_value
;
6297 /* Upon reaching here, src_known is true and umax_val is equal
6300 dst_reg
->smin_value
>>= umin_val
;
6301 dst_reg
->smax_value
>>= umin_val
;
6303 dst_reg
->var_off
= tnum_arshift(dst_reg
->var_off
, umin_val
, 64);
6305 /* blow away the dst_reg umin_value/umax_value and rely on
6306 * dst_reg var_off to refine the result.
6308 dst_reg
->umin_value
= 0;
6309 dst_reg
->umax_value
= U64_MAX
;
6311 /* Its not easy to operate on alu32 bounds here because it depends
6312 * on bits being shifted in from upper 32-bits. Take easy way out
6313 * and mark unbounded so we can recalculate later from tnum.
6315 __mark_reg32_unbounded(dst_reg
);
6316 __update_reg_bounds(dst_reg
);
6319 /* WARNING: This function does calculations on 64-bit values, but the actual
6320 * execution may occur on 32-bit values. Therefore, things like bitshifts
6321 * need extra checks in the 32-bit case.
6323 static int adjust_scalar_min_max_vals(struct bpf_verifier_env
*env
,
6324 struct bpf_insn
*insn
,
6325 struct bpf_reg_state
*dst_reg
,
6326 struct bpf_reg_state src_reg
)
6328 struct bpf_reg_state
*regs
= cur_regs(env
);
6329 u8 opcode
= BPF_OP(insn
->code
);
6331 s64 smin_val
, smax_val
;
6332 u64 umin_val
, umax_val
;
6333 s32 s32_min_val
, s32_max_val
;
6334 u32 u32_min_val
, u32_max_val
;
6335 u64 insn_bitness
= (BPF_CLASS(insn
->code
) == BPF_ALU64
) ? 64 : 32;
6336 u32 dst
= insn
->dst_reg
;
6338 bool alu32
= (BPF_CLASS(insn
->code
) != BPF_ALU64
);
6340 smin_val
= src_reg
.smin_value
;
6341 smax_val
= src_reg
.smax_value
;
6342 umin_val
= src_reg
.umin_value
;
6343 umax_val
= src_reg
.umax_value
;
6345 s32_min_val
= src_reg
.s32_min_value
;
6346 s32_max_val
= src_reg
.s32_max_value
;
6347 u32_min_val
= src_reg
.u32_min_value
;
6348 u32_max_val
= src_reg
.u32_max_value
;
6351 src_known
= tnum_subreg_is_const(src_reg
.var_off
);
6353 (s32_min_val
!= s32_max_val
|| u32_min_val
!= u32_max_val
)) ||
6354 s32_min_val
> s32_max_val
|| u32_min_val
> u32_max_val
) {
6355 /* Taint dst register if offset had invalid bounds
6356 * derived from e.g. dead branches.
6358 __mark_reg_unknown(env
, dst_reg
);
6362 src_known
= tnum_is_const(src_reg
.var_off
);
6364 (smin_val
!= smax_val
|| umin_val
!= umax_val
)) ||
6365 smin_val
> smax_val
|| umin_val
> umax_val
) {
6366 /* Taint dst register if offset had invalid bounds
6367 * derived from e.g. dead branches.
6369 __mark_reg_unknown(env
, dst_reg
);
6375 opcode
!= BPF_ADD
&& opcode
!= BPF_SUB
&& opcode
!= BPF_AND
) {
6376 __mark_reg_unknown(env
, dst_reg
);
6380 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
6381 * There are two classes of instructions: The first class we track both
6382 * alu32 and alu64 sign/unsigned bounds independently this provides the
6383 * greatest amount of precision when alu operations are mixed with jmp32
6384 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
6385 * and BPF_OR. This is possible because these ops have fairly easy to
6386 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
6387 * See alu32 verifier tests for examples. The second class of
6388 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
6389 * with regards to tracking sign/unsigned bounds because the bits may
6390 * cross subreg boundaries in the alu64 case. When this happens we mark
6391 * the reg unbounded in the subreg bound space and use the resulting
6392 * tnum to calculate an approximation of the sign/unsigned bounds.
6396 ret
= sanitize_val_alu(env
, insn
);
6398 verbose(env
, "R%d tried to add from different pointers or scalars\n", dst
);
6401 scalar32_min_max_add(dst_reg
, &src_reg
);
6402 scalar_min_max_add(dst_reg
, &src_reg
);
6403 dst_reg
->var_off
= tnum_add(dst_reg
->var_off
, src_reg
.var_off
);
6406 ret
= sanitize_val_alu(env
, insn
);
6408 verbose(env
, "R%d tried to sub from different pointers or scalars\n", dst
);
6411 scalar32_min_max_sub(dst_reg
, &src_reg
);
6412 scalar_min_max_sub(dst_reg
, &src_reg
);
6413 dst_reg
->var_off
= tnum_sub(dst_reg
->var_off
, src_reg
.var_off
);
6416 dst_reg
->var_off
= tnum_mul(dst_reg
->var_off
, src_reg
.var_off
);
6417 scalar32_min_max_mul(dst_reg
, &src_reg
);
6418 scalar_min_max_mul(dst_reg
, &src_reg
);
6421 dst_reg
->var_off
= tnum_and(dst_reg
->var_off
, src_reg
.var_off
);
6422 scalar32_min_max_and(dst_reg
, &src_reg
);
6423 scalar_min_max_and(dst_reg
, &src_reg
);
6426 dst_reg
->var_off
= tnum_or(dst_reg
->var_off
, src_reg
.var_off
);
6427 scalar32_min_max_or(dst_reg
, &src_reg
);
6428 scalar_min_max_or(dst_reg
, &src_reg
);
6431 dst_reg
->var_off
= tnum_xor(dst_reg
->var_off
, src_reg
.var_off
);
6432 scalar32_min_max_xor(dst_reg
, &src_reg
);
6433 scalar_min_max_xor(dst_reg
, &src_reg
);
6436 if (umax_val
>= insn_bitness
) {
6437 /* Shifts greater than 31 or 63 are undefined.
6438 * This includes shifts by a negative number.
6440 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
6444 scalar32_min_max_lsh(dst_reg
, &src_reg
);
6446 scalar_min_max_lsh(dst_reg
, &src_reg
);
6449 if (umax_val
>= insn_bitness
) {
6450 /* Shifts greater than 31 or 63 are undefined.
6451 * This includes shifts by a negative number.
6453 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
6457 scalar32_min_max_rsh(dst_reg
, &src_reg
);
6459 scalar_min_max_rsh(dst_reg
, &src_reg
);
6462 if (umax_val
>= insn_bitness
) {
6463 /* Shifts greater than 31 or 63 are undefined.
6464 * This includes shifts by a negative number.
6466 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
6470 scalar32_min_max_arsh(dst_reg
, &src_reg
);
6472 scalar_min_max_arsh(dst_reg
, &src_reg
);
6475 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
6479 /* ALU32 ops are zero extended into 64bit register */
6481 zext_32_to_64(dst_reg
);
6483 __update_reg_bounds(dst_reg
);
6484 __reg_deduce_bounds(dst_reg
);
6485 __reg_bound_offset(dst_reg
);
6489 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
6492 static int adjust_reg_min_max_vals(struct bpf_verifier_env
*env
,
6493 struct bpf_insn
*insn
)
6495 struct bpf_verifier_state
*vstate
= env
->cur_state
;
6496 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
6497 struct bpf_reg_state
*regs
= state
->regs
, *dst_reg
, *src_reg
;
6498 struct bpf_reg_state
*ptr_reg
= NULL
, off_reg
= {0};
6499 u8 opcode
= BPF_OP(insn
->code
);
6502 dst_reg
= ®s
[insn
->dst_reg
];
6504 if (dst_reg
->type
!= SCALAR_VALUE
)
6507 /* Make sure ID is cleared otherwise dst_reg min/max could be
6508 * incorrectly propagated into other registers by find_equal_scalars()
6511 if (BPF_SRC(insn
->code
) == BPF_X
) {
6512 src_reg
= ®s
[insn
->src_reg
];
6513 if (src_reg
->type
!= SCALAR_VALUE
) {
6514 if (dst_reg
->type
!= SCALAR_VALUE
) {
6515 /* Combining two pointers by any ALU op yields
6516 * an arbitrary scalar. Disallow all math except
6517 * pointer subtraction
6519 if (opcode
== BPF_SUB
&& env
->allow_ptr_leaks
) {
6520 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
6523 verbose(env
, "R%d pointer %s pointer prohibited\n",
6525 bpf_alu_string
[opcode
>> 4]);
6528 /* scalar += pointer
6529 * This is legal, but we have to reverse our
6530 * src/dest handling in computing the range
6532 err
= mark_chain_precision(env
, insn
->dst_reg
);
6535 return adjust_ptr_min_max_vals(env
, insn
,
6538 } else if (ptr_reg
) {
6539 /* pointer += scalar */
6540 err
= mark_chain_precision(env
, insn
->src_reg
);
6543 return adjust_ptr_min_max_vals(env
, insn
,
6547 /* Pretend the src is a reg with a known value, since we only
6548 * need to be able to read from this state.
6550 off_reg
.type
= SCALAR_VALUE
;
6551 __mark_reg_known(&off_reg
, insn
->imm
);
6553 if (ptr_reg
) /* pointer += K */
6554 return adjust_ptr_min_max_vals(env
, insn
,
6558 /* Got here implies adding two SCALAR_VALUEs */
6559 if (WARN_ON_ONCE(ptr_reg
)) {
6560 print_verifier_state(env
, state
);
6561 verbose(env
, "verifier internal error: unexpected ptr_reg\n");
6564 if (WARN_ON(!src_reg
)) {
6565 print_verifier_state(env
, state
);
6566 verbose(env
, "verifier internal error: no src_reg\n");
6569 return adjust_scalar_min_max_vals(env
, insn
, dst_reg
, *src_reg
);
6572 /* check validity of 32-bit and 64-bit arithmetic operations */
6573 static int check_alu_op(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
6575 struct bpf_reg_state
*regs
= cur_regs(env
);
6576 u8 opcode
= BPF_OP(insn
->code
);
6579 if (opcode
== BPF_END
|| opcode
== BPF_NEG
) {
6580 if (opcode
== BPF_NEG
) {
6581 if (BPF_SRC(insn
->code
) != 0 ||
6582 insn
->src_reg
!= BPF_REG_0
||
6583 insn
->off
!= 0 || insn
->imm
!= 0) {
6584 verbose(env
, "BPF_NEG uses reserved fields\n");
6588 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0 ||
6589 (insn
->imm
!= 16 && insn
->imm
!= 32 && insn
->imm
!= 64) ||
6590 BPF_CLASS(insn
->code
) == BPF_ALU64
) {
6591 verbose(env
, "BPF_END uses reserved fields\n");
6596 /* check src operand */
6597 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
6601 if (is_pointer_value(env
, insn
->dst_reg
)) {
6602 verbose(env
, "R%d pointer arithmetic prohibited\n",
6607 /* check dest operand */
6608 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP
);
6612 } else if (opcode
== BPF_MOV
) {
6614 if (BPF_SRC(insn
->code
) == BPF_X
) {
6615 if (insn
->imm
!= 0 || insn
->off
!= 0) {
6616 verbose(env
, "BPF_MOV uses reserved fields\n");
6620 /* check src operand */
6621 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
6625 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0) {
6626 verbose(env
, "BPF_MOV uses reserved fields\n");
6631 /* check dest operand, mark as required later */
6632 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP_NO_MARK
);
6636 if (BPF_SRC(insn
->code
) == BPF_X
) {
6637 struct bpf_reg_state
*src_reg
= regs
+ insn
->src_reg
;
6638 struct bpf_reg_state
*dst_reg
= regs
+ insn
->dst_reg
;
6640 if (BPF_CLASS(insn
->code
) == BPF_ALU64
) {
6642 * copy register state to dest reg
6644 if (src_reg
->type
== SCALAR_VALUE
&& !src_reg
->id
)
6645 /* Assign src and dst registers the same ID
6646 * that will be used by find_equal_scalars()
6647 * to propagate min/max range.
6649 src_reg
->id
= ++env
->id_gen
;
6650 *dst_reg
= *src_reg
;
6651 dst_reg
->live
|= REG_LIVE_WRITTEN
;
6652 dst_reg
->subreg_def
= DEF_NOT_SUBREG
;
6655 if (is_pointer_value(env
, insn
->src_reg
)) {
6657 "R%d partial copy of pointer\n",
6660 } else if (src_reg
->type
== SCALAR_VALUE
) {
6661 *dst_reg
= *src_reg
;
6662 /* Make sure ID is cleared otherwise
6663 * dst_reg min/max could be incorrectly
6664 * propagated into src_reg by find_equal_scalars()
6667 dst_reg
->live
|= REG_LIVE_WRITTEN
;
6668 dst_reg
->subreg_def
= env
->insn_idx
+ 1;
6670 mark_reg_unknown(env
, regs
,
6673 zext_32_to_64(dst_reg
);
6677 * remember the value we stored into this reg
6679 /* clear any state __mark_reg_known doesn't set */
6680 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
6681 regs
[insn
->dst_reg
].type
= SCALAR_VALUE
;
6682 if (BPF_CLASS(insn
->code
) == BPF_ALU64
) {
6683 __mark_reg_known(regs
+ insn
->dst_reg
,
6686 __mark_reg_known(regs
+ insn
->dst_reg
,
6691 } else if (opcode
> BPF_END
) {
6692 verbose(env
, "invalid BPF_ALU opcode %x\n", opcode
);
6695 } else { /* all other ALU ops: and, sub, xor, add, ... */
6697 if (BPF_SRC(insn
->code
) == BPF_X
) {
6698 if (insn
->imm
!= 0 || insn
->off
!= 0) {
6699 verbose(env
, "BPF_ALU uses reserved fields\n");
6702 /* check src1 operand */
6703 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
6707 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0) {
6708 verbose(env
, "BPF_ALU uses reserved fields\n");
6713 /* check src2 operand */
6714 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
6718 if ((opcode
== BPF_MOD
|| opcode
== BPF_DIV
) &&
6719 BPF_SRC(insn
->code
) == BPF_K
&& insn
->imm
== 0) {
6720 verbose(env
, "div by zero\n");
6724 if ((opcode
== BPF_LSH
|| opcode
== BPF_RSH
||
6725 opcode
== BPF_ARSH
) && BPF_SRC(insn
->code
) == BPF_K
) {
6726 int size
= BPF_CLASS(insn
->code
) == BPF_ALU64
? 64 : 32;
6728 if (insn
->imm
< 0 || insn
->imm
>= size
) {
6729 verbose(env
, "invalid shift %d\n", insn
->imm
);
6734 /* check dest operand */
6735 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP_NO_MARK
);
6739 return adjust_reg_min_max_vals(env
, insn
);
6745 static void __find_good_pkt_pointers(struct bpf_func_state
*state
,
6746 struct bpf_reg_state
*dst_reg
,
6747 enum bpf_reg_type type
, int new_range
)
6749 struct bpf_reg_state
*reg
;
6752 for (i
= 0; i
< MAX_BPF_REG
; i
++) {
6753 reg
= &state
->regs
[i
];
6754 if (reg
->type
== type
&& reg
->id
== dst_reg
->id
)
6755 /* keep the maximum range already checked */
6756 reg
->range
= max(reg
->range
, new_range
);
6759 bpf_for_each_spilled_reg(i
, state
, reg
) {
6762 if (reg
->type
== type
&& reg
->id
== dst_reg
->id
)
6763 reg
->range
= max(reg
->range
, new_range
);
6767 static void find_good_pkt_pointers(struct bpf_verifier_state
*vstate
,
6768 struct bpf_reg_state
*dst_reg
,
6769 enum bpf_reg_type type
,
6770 bool range_right_open
)
6774 if (dst_reg
->off
< 0 ||
6775 (dst_reg
->off
== 0 && range_right_open
))
6776 /* This doesn't give us any range */
6779 if (dst_reg
->umax_value
> MAX_PACKET_OFF
||
6780 dst_reg
->umax_value
+ dst_reg
->off
> MAX_PACKET_OFF
)
6781 /* Risk of overflow. For instance, ptr + (1<<63) may be less
6782 * than pkt_end, but that's because it's also less than pkt.
6786 new_range
= dst_reg
->off
;
6787 if (range_right_open
)
6790 /* Examples for register markings:
6792 * pkt_data in dst register:
6796 * if (r2 > pkt_end) goto <handle exception>
6801 * if (r2 < pkt_end) goto <access okay>
6802 * <handle exception>
6805 * r2 == dst_reg, pkt_end == src_reg
6806 * r2=pkt(id=n,off=8,r=0)
6807 * r3=pkt(id=n,off=0,r=0)
6809 * pkt_data in src register:
6813 * if (pkt_end >= r2) goto <access okay>
6814 * <handle exception>
6818 * if (pkt_end <= r2) goto <handle exception>
6822 * pkt_end == dst_reg, r2 == src_reg
6823 * r2=pkt(id=n,off=8,r=0)
6824 * r3=pkt(id=n,off=0,r=0)
6826 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
6827 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
6828 * and [r3, r3 + 8-1) respectively is safe to access depending on
6832 /* If our ids match, then we must have the same max_value. And we
6833 * don't care about the other reg's fixed offset, since if it's too big
6834 * the range won't allow anything.
6835 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
6837 for (i
= 0; i
<= vstate
->curframe
; i
++)
6838 __find_good_pkt_pointers(vstate
->frame
[i
], dst_reg
, type
,
6842 static int is_branch32_taken(struct bpf_reg_state
*reg
, u32 val
, u8 opcode
)
6844 struct tnum subreg
= tnum_subreg(reg
->var_off
);
6845 s32 sval
= (s32
)val
;
6849 if (tnum_is_const(subreg
))
6850 return !!tnum_equals_const(subreg
, val
);
6853 if (tnum_is_const(subreg
))
6854 return !tnum_equals_const(subreg
, val
);
6857 if ((~subreg
.mask
& subreg
.value
) & val
)
6859 if (!((subreg
.mask
| subreg
.value
) & val
))
6863 if (reg
->u32_min_value
> val
)
6865 else if (reg
->u32_max_value
<= val
)
6869 if (reg
->s32_min_value
> sval
)
6871 else if (reg
->s32_max_value
<= sval
)
6875 if (reg
->u32_max_value
< val
)
6877 else if (reg
->u32_min_value
>= val
)
6881 if (reg
->s32_max_value
< sval
)
6883 else if (reg
->s32_min_value
>= sval
)
6887 if (reg
->u32_min_value
>= val
)
6889 else if (reg
->u32_max_value
< val
)
6893 if (reg
->s32_min_value
>= sval
)
6895 else if (reg
->s32_max_value
< sval
)
6899 if (reg
->u32_max_value
<= val
)
6901 else if (reg
->u32_min_value
> val
)
6905 if (reg
->s32_max_value
<= sval
)
6907 else if (reg
->s32_min_value
> sval
)
6916 static int is_branch64_taken(struct bpf_reg_state
*reg
, u64 val
, u8 opcode
)
6918 s64 sval
= (s64
)val
;
6922 if (tnum_is_const(reg
->var_off
))
6923 return !!tnum_equals_const(reg
->var_off
, val
);
6926 if (tnum_is_const(reg
->var_off
))
6927 return !tnum_equals_const(reg
->var_off
, val
);
6930 if ((~reg
->var_off
.mask
& reg
->var_off
.value
) & val
)
6932 if (!((reg
->var_off
.mask
| reg
->var_off
.value
) & val
))
6936 if (reg
->umin_value
> val
)
6938 else if (reg
->umax_value
<= val
)
6942 if (reg
->smin_value
> sval
)
6944 else if (reg
->smax_value
<= sval
)
6948 if (reg
->umax_value
< val
)
6950 else if (reg
->umin_value
>= val
)
6954 if (reg
->smax_value
< sval
)
6956 else if (reg
->smin_value
>= sval
)
6960 if (reg
->umin_value
>= val
)
6962 else if (reg
->umax_value
< val
)
6966 if (reg
->smin_value
>= sval
)
6968 else if (reg
->smax_value
< sval
)
6972 if (reg
->umax_value
<= val
)
6974 else if (reg
->umin_value
> val
)
6978 if (reg
->smax_value
<= sval
)
6980 else if (reg
->smin_value
> sval
)
6988 /* compute branch direction of the expression "if (reg opcode val) goto target;"
6990 * 1 - branch will be taken and "goto target" will be executed
6991 * 0 - branch will not be taken and fall-through to next insn
6992 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
6995 static int is_branch_taken(struct bpf_reg_state
*reg
, u64 val
, u8 opcode
,
6998 if (__is_pointer_value(false, reg
)) {
6999 if (!reg_type_not_null(reg
->type
))
7002 /* If pointer is valid tests against zero will fail so we can
7003 * use this to direct branch taken.
7019 return is_branch32_taken(reg
, val
, opcode
);
7020 return is_branch64_taken(reg
, val
, opcode
);
7023 static int flip_opcode(u32 opcode
)
7025 /* How can we transform "a <op> b" into "b <op> a"? */
7026 static const u8 opcode_flip
[16] = {
7027 /* these stay the same */
7028 [BPF_JEQ
>> 4] = BPF_JEQ
,
7029 [BPF_JNE
>> 4] = BPF_JNE
,
7030 [BPF_JSET
>> 4] = BPF_JSET
,
7031 /* these swap "lesser" and "greater" (L and G in the opcodes) */
7032 [BPF_JGE
>> 4] = BPF_JLE
,
7033 [BPF_JGT
>> 4] = BPF_JLT
,
7034 [BPF_JLE
>> 4] = BPF_JGE
,
7035 [BPF_JLT
>> 4] = BPF_JGT
,
7036 [BPF_JSGE
>> 4] = BPF_JSLE
,
7037 [BPF_JSGT
>> 4] = BPF_JSLT
,
7038 [BPF_JSLE
>> 4] = BPF_JSGE
,
7039 [BPF_JSLT
>> 4] = BPF_JSGT
7041 return opcode_flip
[opcode
>> 4];
7044 static int is_pkt_ptr_branch_taken(struct bpf_reg_state
*dst_reg
,
7045 struct bpf_reg_state
*src_reg
,
7048 struct bpf_reg_state
*pkt
;
7050 if (src_reg
->type
== PTR_TO_PACKET_END
) {
7052 } else if (dst_reg
->type
== PTR_TO_PACKET_END
) {
7054 opcode
= flip_opcode(opcode
);
7059 if (pkt
->range
>= 0)
7064 /* pkt <= pkt_end */
7068 if (pkt
->range
== BEYOND_PKT_END
)
7069 /* pkt has at last one extra byte beyond pkt_end */
7070 return opcode
== BPF_JGT
;
7076 /* pkt >= pkt_end */
7077 if (pkt
->range
== BEYOND_PKT_END
|| pkt
->range
== AT_PKT_END
)
7078 return opcode
== BPF_JGE
;
7084 /* Adjusts the register min/max values in the case that the dst_reg is the
7085 * variable register that we are working on, and src_reg is a constant or we're
7086 * simply doing a BPF_K check.
7087 * In JEQ/JNE cases we also adjust the var_off values.
7089 static void reg_set_min_max(struct bpf_reg_state
*true_reg
,
7090 struct bpf_reg_state
*false_reg
,
7092 u8 opcode
, bool is_jmp32
)
7094 struct tnum false_32off
= tnum_subreg(false_reg
->var_off
);
7095 struct tnum false_64off
= false_reg
->var_off
;
7096 struct tnum true_32off
= tnum_subreg(true_reg
->var_off
);
7097 struct tnum true_64off
= true_reg
->var_off
;
7098 s64 sval
= (s64
)val
;
7099 s32 sval32
= (s32
)val32
;
7101 /* If the dst_reg is a pointer, we can't learn anything about its
7102 * variable offset from the compare (unless src_reg were a pointer into
7103 * the same object, but we don't bother with that.
7104 * Since false_reg and true_reg have the same type by construction, we
7105 * only need to check one of them for pointerness.
7107 if (__is_pointer_value(false, false_reg
))
7114 struct bpf_reg_state
*reg
=
7115 opcode
== BPF_JEQ
? true_reg
: false_reg
;
7117 /* JEQ/JNE comparison doesn't change the register equivalence.
7119 * if (r1 == 42) goto label;
7121 * label: // here both r1 and r2 are known to be 42.
7123 * Hence when marking register as known preserve it's ID.
7126 __mark_reg32_known(reg
, val32
);
7128 ___mark_reg_known(reg
, val
);
7133 false_32off
= tnum_and(false_32off
, tnum_const(~val32
));
7134 if (is_power_of_2(val32
))
7135 true_32off
= tnum_or(true_32off
,
7138 false_64off
= tnum_and(false_64off
, tnum_const(~val
));
7139 if (is_power_of_2(val
))
7140 true_64off
= tnum_or(true_64off
,
7148 u32 false_umax
= opcode
== BPF_JGT
? val32
: val32
- 1;
7149 u32 true_umin
= opcode
== BPF_JGT
? val32
+ 1 : val32
;
7151 false_reg
->u32_max_value
= min(false_reg
->u32_max_value
,
7153 true_reg
->u32_min_value
= max(true_reg
->u32_min_value
,
7156 u64 false_umax
= opcode
== BPF_JGT
? val
: val
- 1;
7157 u64 true_umin
= opcode
== BPF_JGT
? val
+ 1 : val
;
7159 false_reg
->umax_value
= min(false_reg
->umax_value
, false_umax
);
7160 true_reg
->umin_value
= max(true_reg
->umin_value
, true_umin
);
7168 s32 false_smax
= opcode
== BPF_JSGT
? sval32
: sval32
- 1;
7169 s32 true_smin
= opcode
== BPF_JSGT
? sval32
+ 1 : sval32
;
7171 false_reg
->s32_max_value
= min(false_reg
->s32_max_value
, false_smax
);
7172 true_reg
->s32_min_value
= max(true_reg
->s32_min_value
, true_smin
);
7174 s64 false_smax
= opcode
== BPF_JSGT
? sval
: sval
- 1;
7175 s64 true_smin
= opcode
== BPF_JSGT
? sval
+ 1 : sval
;
7177 false_reg
->smax_value
= min(false_reg
->smax_value
, false_smax
);
7178 true_reg
->smin_value
= max(true_reg
->smin_value
, true_smin
);
7186 u32 false_umin
= opcode
== BPF_JLT
? val32
: val32
+ 1;
7187 u32 true_umax
= opcode
== BPF_JLT
? val32
- 1 : val32
;
7189 false_reg
->u32_min_value
= max(false_reg
->u32_min_value
,
7191 true_reg
->u32_max_value
= min(true_reg
->u32_max_value
,
7194 u64 false_umin
= opcode
== BPF_JLT
? val
: val
+ 1;
7195 u64 true_umax
= opcode
== BPF_JLT
? val
- 1 : val
;
7197 false_reg
->umin_value
= max(false_reg
->umin_value
, false_umin
);
7198 true_reg
->umax_value
= min(true_reg
->umax_value
, true_umax
);
7206 s32 false_smin
= opcode
== BPF_JSLT
? sval32
: sval32
+ 1;
7207 s32 true_smax
= opcode
== BPF_JSLT
? sval32
- 1 : sval32
;
7209 false_reg
->s32_min_value
= max(false_reg
->s32_min_value
, false_smin
);
7210 true_reg
->s32_max_value
= min(true_reg
->s32_max_value
, true_smax
);
7212 s64 false_smin
= opcode
== BPF_JSLT
? sval
: sval
+ 1;
7213 s64 true_smax
= opcode
== BPF_JSLT
? sval
- 1 : sval
;
7215 false_reg
->smin_value
= max(false_reg
->smin_value
, false_smin
);
7216 true_reg
->smax_value
= min(true_reg
->smax_value
, true_smax
);
7225 false_reg
->var_off
= tnum_or(tnum_clear_subreg(false_64off
),
7226 tnum_subreg(false_32off
));
7227 true_reg
->var_off
= tnum_or(tnum_clear_subreg(true_64off
),
7228 tnum_subreg(true_32off
));
7229 __reg_combine_32_into_64(false_reg
);
7230 __reg_combine_32_into_64(true_reg
);
7232 false_reg
->var_off
= false_64off
;
7233 true_reg
->var_off
= true_64off
;
7234 __reg_combine_64_into_32(false_reg
);
7235 __reg_combine_64_into_32(true_reg
);
7239 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
7242 static void reg_set_min_max_inv(struct bpf_reg_state
*true_reg
,
7243 struct bpf_reg_state
*false_reg
,
7245 u8 opcode
, bool is_jmp32
)
7247 opcode
= flip_opcode(opcode
);
7248 /* This uses zero as "not present in table"; luckily the zero opcode,
7249 * BPF_JA, can't get here.
7252 reg_set_min_max(true_reg
, false_reg
, val
, val32
, opcode
, is_jmp32
);
7255 /* Regs are known to be equal, so intersect their min/max/var_off */
7256 static void __reg_combine_min_max(struct bpf_reg_state
*src_reg
,
7257 struct bpf_reg_state
*dst_reg
)
7259 src_reg
->umin_value
= dst_reg
->umin_value
= max(src_reg
->umin_value
,
7260 dst_reg
->umin_value
);
7261 src_reg
->umax_value
= dst_reg
->umax_value
= min(src_reg
->umax_value
,
7262 dst_reg
->umax_value
);
7263 src_reg
->smin_value
= dst_reg
->smin_value
= max(src_reg
->smin_value
,
7264 dst_reg
->smin_value
);
7265 src_reg
->smax_value
= dst_reg
->smax_value
= min(src_reg
->smax_value
,
7266 dst_reg
->smax_value
);
7267 src_reg
->var_off
= dst_reg
->var_off
= tnum_intersect(src_reg
->var_off
,
7269 /* We might have learned new bounds from the var_off. */
7270 __update_reg_bounds(src_reg
);
7271 __update_reg_bounds(dst_reg
);
7272 /* We might have learned something about the sign bit. */
7273 __reg_deduce_bounds(src_reg
);
7274 __reg_deduce_bounds(dst_reg
);
7275 /* We might have learned some bits from the bounds. */
7276 __reg_bound_offset(src_reg
);
7277 __reg_bound_offset(dst_reg
);
7278 /* Intersecting with the old var_off might have improved our bounds
7279 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
7280 * then new var_off is (0; 0x7f...fc) which improves our umax.
7282 __update_reg_bounds(src_reg
);
7283 __update_reg_bounds(dst_reg
);
7286 static void reg_combine_min_max(struct bpf_reg_state
*true_src
,
7287 struct bpf_reg_state
*true_dst
,
7288 struct bpf_reg_state
*false_src
,
7289 struct bpf_reg_state
*false_dst
,
7294 __reg_combine_min_max(true_src
, true_dst
);
7297 __reg_combine_min_max(false_src
, false_dst
);
7302 static void mark_ptr_or_null_reg(struct bpf_func_state
*state
,
7303 struct bpf_reg_state
*reg
, u32 id
,
7306 if (reg_type_may_be_null(reg
->type
) && reg
->id
== id
&&
7307 !WARN_ON_ONCE(!reg
->id
)) {
7308 /* Old offset (both fixed and variable parts) should
7309 * have been known-zero, because we don't allow pointer
7310 * arithmetic on pointers that might be NULL.
7312 if (WARN_ON_ONCE(reg
->smin_value
|| reg
->smax_value
||
7313 !tnum_equals_const(reg
->var_off
, 0) ||
7315 __mark_reg_known_zero(reg
);
7319 reg
->type
= SCALAR_VALUE
;
7320 } else if (reg
->type
== PTR_TO_MAP_VALUE_OR_NULL
) {
7321 const struct bpf_map
*map
= reg
->map_ptr
;
7323 if (map
->inner_map_meta
) {
7324 reg
->type
= CONST_PTR_TO_MAP
;
7325 reg
->map_ptr
= map
->inner_map_meta
;
7326 } else if (map
->map_type
== BPF_MAP_TYPE_XSKMAP
) {
7327 reg
->type
= PTR_TO_XDP_SOCK
;
7328 } else if (map
->map_type
== BPF_MAP_TYPE_SOCKMAP
||
7329 map
->map_type
== BPF_MAP_TYPE_SOCKHASH
) {
7330 reg
->type
= PTR_TO_SOCKET
;
7332 reg
->type
= PTR_TO_MAP_VALUE
;
7334 } else if (reg
->type
== PTR_TO_SOCKET_OR_NULL
) {
7335 reg
->type
= PTR_TO_SOCKET
;
7336 } else if (reg
->type
== PTR_TO_SOCK_COMMON_OR_NULL
) {
7337 reg
->type
= PTR_TO_SOCK_COMMON
;
7338 } else if (reg
->type
== PTR_TO_TCP_SOCK_OR_NULL
) {
7339 reg
->type
= PTR_TO_TCP_SOCK
;
7340 } else if (reg
->type
== PTR_TO_BTF_ID_OR_NULL
) {
7341 reg
->type
= PTR_TO_BTF_ID
;
7342 } else if (reg
->type
== PTR_TO_MEM_OR_NULL
) {
7343 reg
->type
= PTR_TO_MEM
;
7344 } else if (reg
->type
== PTR_TO_RDONLY_BUF_OR_NULL
) {
7345 reg
->type
= PTR_TO_RDONLY_BUF
;
7346 } else if (reg
->type
== PTR_TO_RDWR_BUF_OR_NULL
) {
7347 reg
->type
= PTR_TO_RDWR_BUF
;
7350 /* We don't need id and ref_obj_id from this point
7351 * onwards anymore, thus we should better reset it,
7352 * so that state pruning has chances to take effect.
7355 reg
->ref_obj_id
= 0;
7356 } else if (!reg_may_point_to_spin_lock(reg
)) {
7357 /* For not-NULL ptr, reg->ref_obj_id will be reset
7358 * in release_reg_references().
7360 * reg->id is still used by spin_lock ptr. Other
7361 * than spin_lock ptr type, reg->id can be reset.
7368 static void __mark_ptr_or_null_regs(struct bpf_func_state
*state
, u32 id
,
7371 struct bpf_reg_state
*reg
;
7374 for (i
= 0; i
< MAX_BPF_REG
; i
++)
7375 mark_ptr_or_null_reg(state
, &state
->regs
[i
], id
, is_null
);
7377 bpf_for_each_spilled_reg(i
, state
, reg
) {
7380 mark_ptr_or_null_reg(state
, reg
, id
, is_null
);
7384 /* The logic is similar to find_good_pkt_pointers(), both could eventually
7385 * be folded together at some point.
7387 static void mark_ptr_or_null_regs(struct bpf_verifier_state
*vstate
, u32 regno
,
7390 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
7391 struct bpf_reg_state
*regs
= state
->regs
;
7392 u32 ref_obj_id
= regs
[regno
].ref_obj_id
;
7393 u32 id
= regs
[regno
].id
;
7396 if (ref_obj_id
&& ref_obj_id
== id
&& is_null
)
7397 /* regs[regno] is in the " == NULL" branch.
7398 * No one could have freed the reference state before
7399 * doing the NULL check.
7401 WARN_ON_ONCE(release_reference_state(state
, id
));
7403 for (i
= 0; i
<= vstate
->curframe
; i
++)
7404 __mark_ptr_or_null_regs(vstate
->frame
[i
], id
, is_null
);
7407 static bool try_match_pkt_pointers(const struct bpf_insn
*insn
,
7408 struct bpf_reg_state
*dst_reg
,
7409 struct bpf_reg_state
*src_reg
,
7410 struct bpf_verifier_state
*this_branch
,
7411 struct bpf_verifier_state
*other_branch
)
7413 if (BPF_SRC(insn
->code
) != BPF_X
)
7416 /* Pointers are always 64-bit. */
7417 if (BPF_CLASS(insn
->code
) == BPF_JMP32
)
7420 switch (BPF_OP(insn
->code
)) {
7422 if ((dst_reg
->type
== PTR_TO_PACKET
&&
7423 src_reg
->type
== PTR_TO_PACKET_END
) ||
7424 (dst_reg
->type
== PTR_TO_PACKET_META
&&
7425 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
7426 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
7427 find_good_pkt_pointers(this_branch
, dst_reg
,
7428 dst_reg
->type
, false);
7429 mark_pkt_end(other_branch
, insn
->dst_reg
, true);
7430 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
7431 src_reg
->type
== PTR_TO_PACKET
) ||
7432 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
7433 src_reg
->type
== PTR_TO_PACKET_META
)) {
7434 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
7435 find_good_pkt_pointers(other_branch
, src_reg
,
7436 src_reg
->type
, true);
7437 mark_pkt_end(this_branch
, insn
->src_reg
, false);
7443 if ((dst_reg
->type
== PTR_TO_PACKET
&&
7444 src_reg
->type
== PTR_TO_PACKET_END
) ||
7445 (dst_reg
->type
== PTR_TO_PACKET_META
&&
7446 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
7447 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
7448 find_good_pkt_pointers(other_branch
, dst_reg
,
7449 dst_reg
->type
, true);
7450 mark_pkt_end(this_branch
, insn
->dst_reg
, false);
7451 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
7452 src_reg
->type
== PTR_TO_PACKET
) ||
7453 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
7454 src_reg
->type
== PTR_TO_PACKET_META
)) {
7455 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
7456 find_good_pkt_pointers(this_branch
, src_reg
,
7457 src_reg
->type
, false);
7458 mark_pkt_end(other_branch
, insn
->src_reg
, true);
7464 if ((dst_reg
->type
== PTR_TO_PACKET
&&
7465 src_reg
->type
== PTR_TO_PACKET_END
) ||
7466 (dst_reg
->type
== PTR_TO_PACKET_META
&&
7467 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
7468 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
7469 find_good_pkt_pointers(this_branch
, dst_reg
,
7470 dst_reg
->type
, true);
7471 mark_pkt_end(other_branch
, insn
->dst_reg
, false);
7472 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
7473 src_reg
->type
== PTR_TO_PACKET
) ||
7474 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
7475 src_reg
->type
== PTR_TO_PACKET_META
)) {
7476 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
7477 find_good_pkt_pointers(other_branch
, src_reg
,
7478 src_reg
->type
, false);
7479 mark_pkt_end(this_branch
, insn
->src_reg
, true);
7485 if ((dst_reg
->type
== PTR_TO_PACKET
&&
7486 src_reg
->type
== PTR_TO_PACKET_END
) ||
7487 (dst_reg
->type
== PTR_TO_PACKET_META
&&
7488 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
7489 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
7490 find_good_pkt_pointers(other_branch
, dst_reg
,
7491 dst_reg
->type
, false);
7492 mark_pkt_end(this_branch
, insn
->dst_reg
, true);
7493 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
7494 src_reg
->type
== PTR_TO_PACKET
) ||
7495 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
7496 src_reg
->type
== PTR_TO_PACKET_META
)) {
7497 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
7498 find_good_pkt_pointers(this_branch
, src_reg
,
7499 src_reg
->type
, true);
7500 mark_pkt_end(other_branch
, insn
->src_reg
, false);
7512 static void find_equal_scalars(struct bpf_verifier_state
*vstate
,
7513 struct bpf_reg_state
*known_reg
)
7515 struct bpf_func_state
*state
;
7516 struct bpf_reg_state
*reg
;
7519 for (i
= 0; i
<= vstate
->curframe
; i
++) {
7520 state
= vstate
->frame
[i
];
7521 for (j
= 0; j
< MAX_BPF_REG
; j
++) {
7522 reg
= &state
->regs
[j
];
7523 if (reg
->type
== SCALAR_VALUE
&& reg
->id
== known_reg
->id
)
7527 bpf_for_each_spilled_reg(j
, state
, reg
) {
7530 if (reg
->type
== SCALAR_VALUE
&& reg
->id
== known_reg
->id
)
7536 static int check_cond_jmp_op(struct bpf_verifier_env
*env
,
7537 struct bpf_insn
*insn
, int *insn_idx
)
7539 struct bpf_verifier_state
*this_branch
= env
->cur_state
;
7540 struct bpf_verifier_state
*other_branch
;
7541 struct bpf_reg_state
*regs
= this_branch
->frame
[this_branch
->curframe
]->regs
;
7542 struct bpf_reg_state
*dst_reg
, *other_branch_regs
, *src_reg
= NULL
;
7543 u8 opcode
= BPF_OP(insn
->code
);
7548 /* Only conditional jumps are expected to reach here. */
7549 if (opcode
== BPF_JA
|| opcode
> BPF_JSLE
) {
7550 verbose(env
, "invalid BPF_JMP/JMP32 opcode %x\n", opcode
);
7554 if (BPF_SRC(insn
->code
) == BPF_X
) {
7555 if (insn
->imm
!= 0) {
7556 verbose(env
, "BPF_JMP/JMP32 uses reserved fields\n");
7560 /* check src1 operand */
7561 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
7565 if (is_pointer_value(env
, insn
->src_reg
)) {
7566 verbose(env
, "R%d pointer comparison prohibited\n",
7570 src_reg
= ®s
[insn
->src_reg
];
7572 if (insn
->src_reg
!= BPF_REG_0
) {
7573 verbose(env
, "BPF_JMP/JMP32 uses reserved fields\n");
7578 /* check src2 operand */
7579 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
7583 dst_reg
= ®s
[insn
->dst_reg
];
7584 is_jmp32
= BPF_CLASS(insn
->code
) == BPF_JMP32
;
7586 if (BPF_SRC(insn
->code
) == BPF_K
) {
7587 pred
= is_branch_taken(dst_reg
, insn
->imm
, opcode
, is_jmp32
);
7588 } else if (src_reg
->type
== SCALAR_VALUE
&&
7589 is_jmp32
&& tnum_is_const(tnum_subreg(src_reg
->var_off
))) {
7590 pred
= is_branch_taken(dst_reg
,
7591 tnum_subreg(src_reg
->var_off
).value
,
7594 } else if (src_reg
->type
== SCALAR_VALUE
&&
7595 !is_jmp32
&& tnum_is_const(src_reg
->var_off
)) {
7596 pred
= is_branch_taken(dst_reg
,
7597 src_reg
->var_off
.value
,
7600 } else if (reg_is_pkt_pointer_any(dst_reg
) &&
7601 reg_is_pkt_pointer_any(src_reg
) &&
7603 pred
= is_pkt_ptr_branch_taken(dst_reg
, src_reg
, opcode
);
7607 /* If we get here with a dst_reg pointer type it is because
7608 * above is_branch_taken() special cased the 0 comparison.
7610 if (!__is_pointer_value(false, dst_reg
))
7611 err
= mark_chain_precision(env
, insn
->dst_reg
);
7612 if (BPF_SRC(insn
->code
) == BPF_X
&& !err
&&
7613 !__is_pointer_value(false, src_reg
))
7614 err
= mark_chain_precision(env
, insn
->src_reg
);
7619 /* only follow the goto, ignore fall-through */
7620 *insn_idx
+= insn
->off
;
7622 } else if (pred
== 0) {
7623 /* only follow fall-through branch, since
7624 * that's where the program will go
7629 other_branch
= push_stack(env
, *insn_idx
+ insn
->off
+ 1, *insn_idx
,
7633 other_branch_regs
= other_branch
->frame
[other_branch
->curframe
]->regs
;
7635 /* detect if we are comparing against a constant value so we can adjust
7636 * our min/max values for our dst register.
7637 * this is only legit if both are scalars (or pointers to the same
7638 * object, I suppose, but we don't support that right now), because
7639 * otherwise the different base pointers mean the offsets aren't
7642 if (BPF_SRC(insn
->code
) == BPF_X
) {
7643 struct bpf_reg_state
*src_reg
= ®s
[insn
->src_reg
];
7645 if (dst_reg
->type
== SCALAR_VALUE
&&
7646 src_reg
->type
== SCALAR_VALUE
) {
7647 if (tnum_is_const(src_reg
->var_off
) ||
7649 tnum_is_const(tnum_subreg(src_reg
->var_off
))))
7650 reg_set_min_max(&other_branch_regs
[insn
->dst_reg
],
7652 src_reg
->var_off
.value
,
7653 tnum_subreg(src_reg
->var_off
).value
,
7655 else if (tnum_is_const(dst_reg
->var_off
) ||
7657 tnum_is_const(tnum_subreg(dst_reg
->var_off
))))
7658 reg_set_min_max_inv(&other_branch_regs
[insn
->src_reg
],
7660 dst_reg
->var_off
.value
,
7661 tnum_subreg(dst_reg
->var_off
).value
,
7663 else if (!is_jmp32
&&
7664 (opcode
== BPF_JEQ
|| opcode
== BPF_JNE
))
7665 /* Comparing for equality, we can combine knowledge */
7666 reg_combine_min_max(&other_branch_regs
[insn
->src_reg
],
7667 &other_branch_regs
[insn
->dst_reg
],
7668 src_reg
, dst_reg
, opcode
);
7670 !WARN_ON_ONCE(src_reg
->id
!= other_branch_regs
[insn
->src_reg
].id
)) {
7671 find_equal_scalars(this_branch
, src_reg
);
7672 find_equal_scalars(other_branch
, &other_branch_regs
[insn
->src_reg
]);
7676 } else if (dst_reg
->type
== SCALAR_VALUE
) {
7677 reg_set_min_max(&other_branch_regs
[insn
->dst_reg
],
7678 dst_reg
, insn
->imm
, (u32
)insn
->imm
,
7682 if (dst_reg
->type
== SCALAR_VALUE
&& dst_reg
->id
&&
7683 !WARN_ON_ONCE(dst_reg
->id
!= other_branch_regs
[insn
->dst_reg
].id
)) {
7684 find_equal_scalars(this_branch
, dst_reg
);
7685 find_equal_scalars(other_branch
, &other_branch_regs
[insn
->dst_reg
]);
7688 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
7689 * NOTE: these optimizations below are related with pointer comparison
7690 * which will never be JMP32.
7692 if (!is_jmp32
&& BPF_SRC(insn
->code
) == BPF_K
&&
7693 insn
->imm
== 0 && (opcode
== BPF_JEQ
|| opcode
== BPF_JNE
) &&
7694 reg_type_may_be_null(dst_reg
->type
)) {
7695 /* Mark all identical registers in each branch as either
7696 * safe or unknown depending R == 0 or R != 0 conditional.
7698 mark_ptr_or_null_regs(this_branch
, insn
->dst_reg
,
7700 mark_ptr_or_null_regs(other_branch
, insn
->dst_reg
,
7702 } else if (!try_match_pkt_pointers(insn
, dst_reg
, ®s
[insn
->src_reg
],
7703 this_branch
, other_branch
) &&
7704 is_pointer_value(env
, insn
->dst_reg
)) {
7705 verbose(env
, "R%d pointer comparison prohibited\n",
7709 if (env
->log
.level
& BPF_LOG_LEVEL
)
7710 print_verifier_state(env
, this_branch
->frame
[this_branch
->curframe
]);
7714 /* verify BPF_LD_IMM64 instruction */
7715 static int check_ld_imm(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
7717 struct bpf_insn_aux_data
*aux
= cur_aux(env
);
7718 struct bpf_reg_state
*regs
= cur_regs(env
);
7719 struct bpf_reg_state
*dst_reg
;
7720 struct bpf_map
*map
;
7723 if (BPF_SIZE(insn
->code
) != BPF_DW
) {
7724 verbose(env
, "invalid BPF_LD_IMM insn\n");
7727 if (insn
->off
!= 0) {
7728 verbose(env
, "BPF_LD_IMM64 uses reserved fields\n");
7732 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP
);
7736 dst_reg
= ®s
[insn
->dst_reg
];
7737 if (insn
->src_reg
== 0) {
7738 u64 imm
= ((u64
)(insn
+ 1)->imm
<< 32) | (u32
)insn
->imm
;
7740 dst_reg
->type
= SCALAR_VALUE
;
7741 __mark_reg_known(®s
[insn
->dst_reg
], imm
);
7745 if (insn
->src_reg
== BPF_PSEUDO_BTF_ID
) {
7746 mark_reg_known_zero(env
, regs
, insn
->dst_reg
);
7748 dst_reg
->type
= aux
->btf_var
.reg_type
;
7749 switch (dst_reg
->type
) {
7751 dst_reg
->mem_size
= aux
->btf_var
.mem_size
;
7754 case PTR_TO_PERCPU_BTF_ID
:
7755 dst_reg
->btf
= aux
->btf_var
.btf
;
7756 dst_reg
->btf_id
= aux
->btf_var
.btf_id
;
7759 verbose(env
, "bpf verifier is misconfigured\n");
7765 map
= env
->used_maps
[aux
->map_index
];
7766 mark_reg_known_zero(env
, regs
, insn
->dst_reg
);
7767 dst_reg
->map_ptr
= map
;
7769 if (insn
->src_reg
== BPF_PSEUDO_MAP_VALUE
) {
7770 dst_reg
->type
= PTR_TO_MAP_VALUE
;
7771 dst_reg
->off
= aux
->map_off
;
7772 if (map_value_has_spin_lock(map
))
7773 dst_reg
->id
= ++env
->id_gen
;
7774 } else if (insn
->src_reg
== BPF_PSEUDO_MAP_FD
) {
7775 dst_reg
->type
= CONST_PTR_TO_MAP
;
7777 verbose(env
, "bpf verifier is misconfigured\n");
7784 static bool may_access_skb(enum bpf_prog_type type
)
7787 case BPF_PROG_TYPE_SOCKET_FILTER
:
7788 case BPF_PROG_TYPE_SCHED_CLS
:
7789 case BPF_PROG_TYPE_SCHED_ACT
:
7796 /* verify safety of LD_ABS|LD_IND instructions:
7797 * - they can only appear in the programs where ctx == skb
7798 * - since they are wrappers of function calls, they scratch R1-R5 registers,
7799 * preserve R6-R9, and store return value into R0
7802 * ctx == skb == R6 == CTX
7805 * SRC == any register
7806 * IMM == 32-bit immediate
7809 * R0 - 8/16/32-bit skb data converted to cpu endianness
7811 static int check_ld_abs(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
7813 struct bpf_reg_state
*regs
= cur_regs(env
);
7814 static const int ctx_reg
= BPF_REG_6
;
7815 u8 mode
= BPF_MODE(insn
->code
);
7818 if (!may_access_skb(resolve_prog_type(env
->prog
))) {
7819 verbose(env
, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
7823 if (!env
->ops
->gen_ld_abs
) {
7824 verbose(env
, "bpf verifier is misconfigured\n");
7828 if (insn
->dst_reg
!= BPF_REG_0
|| insn
->off
!= 0 ||
7829 BPF_SIZE(insn
->code
) == BPF_DW
||
7830 (mode
== BPF_ABS
&& insn
->src_reg
!= BPF_REG_0
)) {
7831 verbose(env
, "BPF_LD_[ABS|IND] uses reserved fields\n");
7835 /* check whether implicit source operand (register R6) is readable */
7836 err
= check_reg_arg(env
, ctx_reg
, SRC_OP
);
7840 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
7841 * gen_ld_abs() may terminate the program at runtime, leading to
7844 err
= check_reference_leak(env
);
7846 verbose(env
, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
7850 if (env
->cur_state
->active_spin_lock
) {
7851 verbose(env
, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
7855 if (regs
[ctx_reg
].type
!= PTR_TO_CTX
) {
7857 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
7861 if (mode
== BPF_IND
) {
7862 /* check explicit source operand */
7863 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
7868 err
= check_ctx_reg(env
, ®s
[ctx_reg
], ctx_reg
);
7872 /* reset caller saved regs to unreadable */
7873 for (i
= 0; i
< CALLER_SAVED_REGS
; i
++) {
7874 mark_reg_not_init(env
, regs
, caller_saved
[i
]);
7875 check_reg_arg(env
, caller_saved
[i
], DST_OP_NO_MARK
);
7878 /* mark destination R0 register as readable, since it contains
7879 * the value fetched from the packet.
7880 * Already marked as written above.
7882 mark_reg_unknown(env
, regs
, BPF_REG_0
);
7883 /* ld_abs load up to 32-bit skb data. */
7884 regs
[BPF_REG_0
].subreg_def
= env
->insn_idx
+ 1;
7888 static int check_return_code(struct bpf_verifier_env
*env
)
7890 struct tnum enforce_attach_type_range
= tnum_unknown
;
7891 const struct bpf_prog
*prog
= env
->prog
;
7892 struct bpf_reg_state
*reg
;
7893 struct tnum range
= tnum_range(0, 1);
7894 enum bpf_prog_type prog_type
= resolve_prog_type(env
->prog
);
7896 const bool is_subprog
= env
->cur_state
->frame
[0]->subprogno
;
7898 /* LSM and struct_ops func-ptr's return type could be "void" */
7900 (prog_type
== BPF_PROG_TYPE_STRUCT_OPS
||
7901 prog_type
== BPF_PROG_TYPE_LSM
) &&
7902 !prog
->aux
->attach_func_proto
->type
)
7905 /* eBPF calling convetion is such that R0 is used
7906 * to return the value from eBPF program.
7907 * Make sure that it's readable at this time
7908 * of bpf_exit, which means that program wrote
7909 * something into it earlier
7911 err
= check_reg_arg(env
, BPF_REG_0
, SRC_OP
);
7915 if (is_pointer_value(env
, BPF_REG_0
)) {
7916 verbose(env
, "R0 leaks addr as return value\n");
7920 reg
= cur_regs(env
) + BPF_REG_0
;
7922 if (reg
->type
!= SCALAR_VALUE
) {
7923 verbose(env
, "At subprogram exit the register R0 is not a scalar value (%s)\n",
7924 reg_type_str
[reg
->type
]);
7930 switch (prog_type
) {
7931 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR
:
7932 if (env
->prog
->expected_attach_type
== BPF_CGROUP_UDP4_RECVMSG
||
7933 env
->prog
->expected_attach_type
== BPF_CGROUP_UDP6_RECVMSG
||
7934 env
->prog
->expected_attach_type
== BPF_CGROUP_INET4_GETPEERNAME
||
7935 env
->prog
->expected_attach_type
== BPF_CGROUP_INET6_GETPEERNAME
||
7936 env
->prog
->expected_attach_type
== BPF_CGROUP_INET4_GETSOCKNAME
||
7937 env
->prog
->expected_attach_type
== BPF_CGROUP_INET6_GETSOCKNAME
)
7938 range
= tnum_range(1, 1);
7940 case BPF_PROG_TYPE_CGROUP_SKB
:
7941 if (env
->prog
->expected_attach_type
== BPF_CGROUP_INET_EGRESS
) {
7942 range
= tnum_range(0, 3);
7943 enforce_attach_type_range
= tnum_range(2, 3);
7946 case BPF_PROG_TYPE_CGROUP_SOCK
:
7947 case BPF_PROG_TYPE_SOCK_OPS
:
7948 case BPF_PROG_TYPE_CGROUP_DEVICE
:
7949 case BPF_PROG_TYPE_CGROUP_SYSCTL
:
7950 case BPF_PROG_TYPE_CGROUP_SOCKOPT
:
7952 case BPF_PROG_TYPE_RAW_TRACEPOINT
:
7953 if (!env
->prog
->aux
->attach_btf_id
)
7955 range
= tnum_const(0);
7957 case BPF_PROG_TYPE_TRACING
:
7958 switch (env
->prog
->expected_attach_type
) {
7959 case BPF_TRACE_FENTRY
:
7960 case BPF_TRACE_FEXIT
:
7961 range
= tnum_const(0);
7963 case BPF_TRACE_RAW_TP
:
7964 case BPF_MODIFY_RETURN
:
7966 case BPF_TRACE_ITER
:
7972 case BPF_PROG_TYPE_SK_LOOKUP
:
7973 range
= tnum_range(SK_DROP
, SK_PASS
);
7975 case BPF_PROG_TYPE_EXT
:
7976 /* freplace program can return anything as its return value
7977 * depends on the to-be-replaced kernel func or bpf program.
7983 if (reg
->type
!= SCALAR_VALUE
) {
7984 verbose(env
, "At program exit the register R0 is not a known value (%s)\n",
7985 reg_type_str
[reg
->type
]);
7989 if (!tnum_in(range
, reg
->var_off
)) {
7992 verbose(env
, "At program exit the register R0 ");
7993 if (!tnum_is_unknown(reg
->var_off
)) {
7994 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
7995 verbose(env
, "has value %s", tn_buf
);
7997 verbose(env
, "has unknown scalar value");
7999 tnum_strn(tn_buf
, sizeof(tn_buf
), range
);
8000 verbose(env
, " should have been in %s\n", tn_buf
);
8004 if (!tnum_is_unknown(enforce_attach_type_range
) &&
8005 tnum_in(enforce_attach_type_range
, reg
->var_off
))
8006 env
->prog
->enforce_expected_attach_type
= 1;
8010 /* non-recursive DFS pseudo code
8011 * 1 procedure DFS-iterative(G,v):
8012 * 2 label v as discovered
8013 * 3 let S be a stack
8015 * 5 while S is not empty
8017 * 7 if t is what we're looking for:
8019 * 9 for all edges e in G.adjacentEdges(t) do
8020 * 10 if edge e is already labelled
8021 * 11 continue with the next edge
8022 * 12 w <- G.adjacentVertex(t,e)
8023 * 13 if vertex w is not discovered and not explored
8024 * 14 label e as tree-edge
8025 * 15 label w as discovered
8028 * 18 else if vertex w is discovered
8029 * 19 label e as back-edge
8031 * 21 // vertex w is explored
8032 * 22 label e as forward- or cross-edge
8033 * 23 label t as explored
8038 * 0x11 - discovered and fall-through edge labelled
8039 * 0x12 - discovered and fall-through and branch edges labelled
8050 static u32
state_htab_size(struct bpf_verifier_env
*env
)
8052 return env
->prog
->len
;
8055 static struct bpf_verifier_state_list
**explored_state(
8056 struct bpf_verifier_env
*env
,
8059 struct bpf_verifier_state
*cur
= env
->cur_state
;
8060 struct bpf_func_state
*state
= cur
->frame
[cur
->curframe
];
8062 return &env
->explored_states
[(idx
^ state
->callsite
) % state_htab_size(env
)];
8065 static void init_explored_state(struct bpf_verifier_env
*env
, int idx
)
8067 env
->insn_aux_data
[idx
].prune_point
= true;
8075 /* t, w, e - match pseudo-code above:
8076 * t - index of current instruction
8077 * w - next instruction
8080 static int push_insn(int t
, int w
, int e
, struct bpf_verifier_env
*env
,
8083 int *insn_stack
= env
->cfg
.insn_stack
;
8084 int *insn_state
= env
->cfg
.insn_state
;
8086 if (e
== FALLTHROUGH
&& insn_state
[t
] >= (DISCOVERED
| FALLTHROUGH
))
8087 return DONE_EXPLORING
;
8089 if (e
== BRANCH
&& insn_state
[t
] >= (DISCOVERED
| BRANCH
))
8090 return DONE_EXPLORING
;
8092 if (w
< 0 || w
>= env
->prog
->len
) {
8093 verbose_linfo(env
, t
, "%d: ", t
);
8094 verbose(env
, "jump out of range from insn %d to %d\n", t
, w
);
8099 /* mark branch target for state pruning */
8100 init_explored_state(env
, w
);
8102 if (insn_state
[w
] == 0) {
8104 insn_state
[t
] = DISCOVERED
| e
;
8105 insn_state
[w
] = DISCOVERED
;
8106 if (env
->cfg
.cur_stack
>= env
->prog
->len
)
8108 insn_stack
[env
->cfg
.cur_stack
++] = w
;
8109 return KEEP_EXPLORING
;
8110 } else if ((insn_state
[w
] & 0xF0) == DISCOVERED
) {
8111 if (loop_ok
&& env
->bpf_capable
)
8112 return DONE_EXPLORING
;
8113 verbose_linfo(env
, t
, "%d: ", t
);
8114 verbose_linfo(env
, w
, "%d: ", w
);
8115 verbose(env
, "back-edge from insn %d to %d\n", t
, w
);
8117 } else if (insn_state
[w
] == EXPLORED
) {
8118 /* forward- or cross-edge */
8119 insn_state
[t
] = DISCOVERED
| e
;
8121 verbose(env
, "insn state internal bug\n");
8124 return DONE_EXPLORING
;
8127 /* Visits the instruction at index t and returns one of the following:
8128 * < 0 - an error occurred
8129 * DONE_EXPLORING - the instruction was fully explored
8130 * KEEP_EXPLORING - there is still work to be done before it is fully explored
8132 static int visit_insn(int t
, int insn_cnt
, struct bpf_verifier_env
*env
)
8134 struct bpf_insn
*insns
= env
->prog
->insnsi
;
8137 /* All non-branch instructions have a single fall-through edge. */
8138 if (BPF_CLASS(insns
[t
].code
) != BPF_JMP
&&
8139 BPF_CLASS(insns
[t
].code
) != BPF_JMP32
)
8140 return push_insn(t
, t
+ 1, FALLTHROUGH
, env
, false);
8142 switch (BPF_OP(insns
[t
].code
)) {
8144 return DONE_EXPLORING
;
8147 ret
= push_insn(t
, t
+ 1, FALLTHROUGH
, env
, false);
8151 if (t
+ 1 < insn_cnt
)
8152 init_explored_state(env
, t
+ 1);
8153 if (insns
[t
].src_reg
== BPF_PSEUDO_CALL
) {
8154 init_explored_state(env
, t
);
8155 ret
= push_insn(t
, t
+ insns
[t
].imm
+ 1, BRANCH
,
8161 if (BPF_SRC(insns
[t
].code
) != BPF_K
)
8164 /* unconditional jump with single edge */
8165 ret
= push_insn(t
, t
+ insns
[t
].off
+ 1, FALLTHROUGH
, env
,
8170 /* unconditional jmp is not a good pruning point,
8171 * but it's marked, since backtracking needs
8172 * to record jmp history in is_state_visited().
8174 init_explored_state(env
, t
+ insns
[t
].off
+ 1);
8175 /* tell verifier to check for equivalent states
8176 * after every call and jump
8178 if (t
+ 1 < insn_cnt
)
8179 init_explored_state(env
, t
+ 1);
8184 /* conditional jump with two edges */
8185 init_explored_state(env
, t
);
8186 ret
= push_insn(t
, t
+ 1, FALLTHROUGH
, env
, true);
8190 return push_insn(t
, t
+ insns
[t
].off
+ 1, BRANCH
, env
, true);
8194 /* non-recursive depth-first-search to detect loops in BPF program
8195 * loop == back-edge in directed graph
8197 static int check_cfg(struct bpf_verifier_env
*env
)
8199 int insn_cnt
= env
->prog
->len
;
8200 int *insn_stack
, *insn_state
;
8204 insn_state
= env
->cfg
.insn_state
= kvcalloc(insn_cnt
, sizeof(int), GFP_KERNEL
);
8208 insn_stack
= env
->cfg
.insn_stack
= kvcalloc(insn_cnt
, sizeof(int), GFP_KERNEL
);
8214 insn_state
[0] = DISCOVERED
; /* mark 1st insn as discovered */
8215 insn_stack
[0] = 0; /* 0 is the first instruction */
8216 env
->cfg
.cur_stack
= 1;
8218 while (env
->cfg
.cur_stack
> 0) {
8219 int t
= insn_stack
[env
->cfg
.cur_stack
- 1];
8221 ret
= visit_insn(t
, insn_cnt
, env
);
8223 case DONE_EXPLORING
:
8224 insn_state
[t
] = EXPLORED
;
8225 env
->cfg
.cur_stack
--;
8227 case KEEP_EXPLORING
:
8231 verbose(env
, "visit_insn internal bug\n");
8238 if (env
->cfg
.cur_stack
< 0) {
8239 verbose(env
, "pop stack internal bug\n");
8244 for (i
= 0; i
< insn_cnt
; i
++) {
8245 if (insn_state
[i
] != EXPLORED
) {
8246 verbose(env
, "unreachable insn %d\n", i
);
8251 ret
= 0; /* cfg looks good */
8256 env
->cfg
.insn_state
= env
->cfg
.insn_stack
= NULL
;
8260 static int check_abnormal_return(struct bpf_verifier_env
*env
)
8264 for (i
= 1; i
< env
->subprog_cnt
; i
++) {
8265 if (env
->subprog_info
[i
].has_ld_abs
) {
8266 verbose(env
, "LD_ABS is not allowed in subprogs without BTF\n");
8269 if (env
->subprog_info
[i
].has_tail_call
) {
8270 verbose(env
, "tail_call is not allowed in subprogs without BTF\n");
8277 /* The minimum supported BTF func info size */
8278 #define MIN_BPF_FUNCINFO_SIZE 8
8279 #define MAX_FUNCINFO_REC_SIZE 252
8281 static int check_btf_func(struct bpf_verifier_env
*env
,
8282 const union bpf_attr
*attr
,
8283 union bpf_attr __user
*uattr
)
8285 const struct btf_type
*type
, *func_proto
, *ret_type
;
8286 u32 i
, nfuncs
, urec_size
, min_size
;
8287 u32 krec_size
= sizeof(struct bpf_func_info
);
8288 struct bpf_func_info
*krecord
;
8289 struct bpf_func_info_aux
*info_aux
= NULL
;
8290 struct bpf_prog
*prog
;
8291 const struct btf
*btf
;
8292 void __user
*urecord
;
8293 u32 prev_offset
= 0;
8297 nfuncs
= attr
->func_info_cnt
;
8299 if (check_abnormal_return(env
))
8304 if (nfuncs
!= env
->subprog_cnt
) {
8305 verbose(env
, "number of funcs in func_info doesn't match number of subprogs\n");
8309 urec_size
= attr
->func_info_rec_size
;
8310 if (urec_size
< MIN_BPF_FUNCINFO_SIZE
||
8311 urec_size
> MAX_FUNCINFO_REC_SIZE
||
8312 urec_size
% sizeof(u32
)) {
8313 verbose(env
, "invalid func info rec size %u\n", urec_size
);
8318 btf
= prog
->aux
->btf
;
8320 urecord
= u64_to_user_ptr(attr
->func_info
);
8321 min_size
= min_t(u32
, krec_size
, urec_size
);
8323 krecord
= kvcalloc(nfuncs
, krec_size
, GFP_KERNEL
| __GFP_NOWARN
);
8326 info_aux
= kcalloc(nfuncs
, sizeof(*info_aux
), GFP_KERNEL
| __GFP_NOWARN
);
8330 for (i
= 0; i
< nfuncs
; i
++) {
8331 ret
= bpf_check_uarg_tail_zero(urecord
, krec_size
, urec_size
);
8333 if (ret
== -E2BIG
) {
8334 verbose(env
, "nonzero tailing record in func info");
8335 /* set the size kernel expects so loader can zero
8336 * out the rest of the record.
8338 if (put_user(min_size
, &uattr
->func_info_rec_size
))
8344 if (copy_from_user(&krecord
[i
], urecord
, min_size
)) {
8349 /* check insn_off */
8352 if (krecord
[i
].insn_off
) {
8354 "nonzero insn_off %u for the first func info record",
8355 krecord
[i
].insn_off
);
8358 } else if (krecord
[i
].insn_off
<= prev_offset
) {
8360 "same or smaller insn offset (%u) than previous func info record (%u)",
8361 krecord
[i
].insn_off
, prev_offset
);
8365 if (env
->subprog_info
[i
].start
!= krecord
[i
].insn_off
) {
8366 verbose(env
, "func_info BTF section doesn't match subprog layout in BPF program\n");
8371 type
= btf_type_by_id(btf
, krecord
[i
].type_id
);
8372 if (!type
|| !btf_type_is_func(type
)) {
8373 verbose(env
, "invalid type id %d in func info",
8374 krecord
[i
].type_id
);
8377 info_aux
[i
].linkage
= BTF_INFO_VLEN(type
->info
);
8379 func_proto
= btf_type_by_id(btf
, type
->type
);
8380 if (unlikely(!func_proto
|| !btf_type_is_func_proto(func_proto
)))
8381 /* btf_func_check() already verified it during BTF load */
8383 ret_type
= btf_type_skip_modifiers(btf
, func_proto
->type
, NULL
);
8385 btf_type_is_small_int(ret_type
) || btf_type_is_enum(ret_type
);
8386 if (i
&& !scalar_return
&& env
->subprog_info
[i
].has_ld_abs
) {
8387 verbose(env
, "LD_ABS is only allowed in functions that return 'int'.\n");
8390 if (i
&& !scalar_return
&& env
->subprog_info
[i
].has_tail_call
) {
8391 verbose(env
, "tail_call is only allowed in functions that return 'int'.\n");
8395 prev_offset
= krecord
[i
].insn_off
;
8396 urecord
+= urec_size
;
8399 prog
->aux
->func_info
= krecord
;
8400 prog
->aux
->func_info_cnt
= nfuncs
;
8401 prog
->aux
->func_info_aux
= info_aux
;
8410 static void adjust_btf_func(struct bpf_verifier_env
*env
)
8412 struct bpf_prog_aux
*aux
= env
->prog
->aux
;
8415 if (!aux
->func_info
)
8418 for (i
= 0; i
< env
->subprog_cnt
; i
++)
8419 aux
->func_info
[i
].insn_off
= env
->subprog_info
[i
].start
;
8422 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
8423 sizeof(((struct bpf_line_info *)(0))->line_col))
8424 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
8426 static int check_btf_line(struct bpf_verifier_env
*env
,
8427 const union bpf_attr
*attr
,
8428 union bpf_attr __user
*uattr
)
8430 u32 i
, s
, nr_linfo
, ncopy
, expected_size
, rec_size
, prev_offset
= 0;
8431 struct bpf_subprog_info
*sub
;
8432 struct bpf_line_info
*linfo
;
8433 struct bpf_prog
*prog
;
8434 const struct btf
*btf
;
8435 void __user
*ulinfo
;
8438 nr_linfo
= attr
->line_info_cnt
;
8442 rec_size
= attr
->line_info_rec_size
;
8443 if (rec_size
< MIN_BPF_LINEINFO_SIZE
||
8444 rec_size
> MAX_LINEINFO_REC_SIZE
||
8445 rec_size
& (sizeof(u32
) - 1))
8448 /* Need to zero it in case the userspace may
8449 * pass in a smaller bpf_line_info object.
8451 linfo
= kvcalloc(nr_linfo
, sizeof(struct bpf_line_info
),
8452 GFP_KERNEL
| __GFP_NOWARN
);
8457 btf
= prog
->aux
->btf
;
8460 sub
= env
->subprog_info
;
8461 ulinfo
= u64_to_user_ptr(attr
->line_info
);
8462 expected_size
= sizeof(struct bpf_line_info
);
8463 ncopy
= min_t(u32
, expected_size
, rec_size
);
8464 for (i
= 0; i
< nr_linfo
; i
++) {
8465 err
= bpf_check_uarg_tail_zero(ulinfo
, expected_size
, rec_size
);
8467 if (err
== -E2BIG
) {
8468 verbose(env
, "nonzero tailing record in line_info");
8469 if (put_user(expected_size
,
8470 &uattr
->line_info_rec_size
))
8476 if (copy_from_user(&linfo
[i
], ulinfo
, ncopy
)) {
8482 * Check insn_off to ensure
8483 * 1) strictly increasing AND
8484 * 2) bounded by prog->len
8486 * The linfo[0].insn_off == 0 check logically falls into
8487 * the later "missing bpf_line_info for func..." case
8488 * because the first linfo[0].insn_off must be the
8489 * first sub also and the first sub must have
8490 * subprog_info[0].start == 0.
8492 if ((i
&& linfo
[i
].insn_off
<= prev_offset
) ||
8493 linfo
[i
].insn_off
>= prog
->len
) {
8494 verbose(env
, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
8495 i
, linfo
[i
].insn_off
, prev_offset
,
8501 if (!prog
->insnsi
[linfo
[i
].insn_off
].code
) {
8503 "Invalid insn code at line_info[%u].insn_off\n",
8509 if (!btf_name_by_offset(btf
, linfo
[i
].line_off
) ||
8510 !btf_name_by_offset(btf
, linfo
[i
].file_name_off
)) {
8511 verbose(env
, "Invalid line_info[%u].line_off or .file_name_off\n", i
);
8516 if (s
!= env
->subprog_cnt
) {
8517 if (linfo
[i
].insn_off
== sub
[s
].start
) {
8518 sub
[s
].linfo_idx
= i
;
8520 } else if (sub
[s
].start
< linfo
[i
].insn_off
) {
8521 verbose(env
, "missing bpf_line_info for func#%u\n", s
);
8527 prev_offset
= linfo
[i
].insn_off
;
8531 if (s
!= env
->subprog_cnt
) {
8532 verbose(env
, "missing bpf_line_info for %u funcs starting from func#%u\n",
8533 env
->subprog_cnt
- s
, s
);
8538 prog
->aux
->linfo
= linfo
;
8539 prog
->aux
->nr_linfo
= nr_linfo
;
8548 static int check_btf_info(struct bpf_verifier_env
*env
,
8549 const union bpf_attr
*attr
,
8550 union bpf_attr __user
*uattr
)
8555 if (!attr
->func_info_cnt
&& !attr
->line_info_cnt
) {
8556 if (check_abnormal_return(env
))
8561 btf
= btf_get_by_fd(attr
->prog_btf_fd
);
8563 return PTR_ERR(btf
);
8564 if (btf_is_kernel(btf
)) {
8568 env
->prog
->aux
->btf
= btf
;
8570 err
= check_btf_func(env
, attr
, uattr
);
8574 err
= check_btf_line(env
, attr
, uattr
);
8581 /* check %cur's range satisfies %old's */
8582 static bool range_within(struct bpf_reg_state
*old
,
8583 struct bpf_reg_state
*cur
)
8585 return old
->umin_value
<= cur
->umin_value
&&
8586 old
->umax_value
>= cur
->umax_value
&&
8587 old
->smin_value
<= cur
->smin_value
&&
8588 old
->smax_value
>= cur
->smax_value
&&
8589 old
->u32_min_value
<= cur
->u32_min_value
&&
8590 old
->u32_max_value
>= cur
->u32_max_value
&&
8591 old
->s32_min_value
<= cur
->s32_min_value
&&
8592 old
->s32_max_value
>= cur
->s32_max_value
;
8595 /* Maximum number of register states that can exist at once */
8596 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
8602 /* If in the old state two registers had the same id, then they need to have
8603 * the same id in the new state as well. But that id could be different from
8604 * the old state, so we need to track the mapping from old to new ids.
8605 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
8606 * regs with old id 5 must also have new id 9 for the new state to be safe. But
8607 * regs with a different old id could still have new id 9, we don't care about
8609 * So we look through our idmap to see if this old id has been seen before. If
8610 * so, we require the new id to match; otherwise, we add the id pair to the map.
8612 static bool check_ids(u32 old_id
, u32 cur_id
, struct idpair
*idmap
)
8616 for (i
= 0; i
< ID_MAP_SIZE
; i
++) {
8617 if (!idmap
[i
].old
) {
8618 /* Reached an empty slot; haven't seen this id before */
8619 idmap
[i
].old
= old_id
;
8620 idmap
[i
].cur
= cur_id
;
8623 if (idmap
[i
].old
== old_id
)
8624 return idmap
[i
].cur
== cur_id
;
8626 /* We ran out of idmap slots, which should be impossible */
8631 static void clean_func_state(struct bpf_verifier_env
*env
,
8632 struct bpf_func_state
*st
)
8634 enum bpf_reg_liveness live
;
8637 for (i
= 0; i
< BPF_REG_FP
; i
++) {
8638 live
= st
->regs
[i
].live
;
8639 /* liveness must not touch this register anymore */
8640 st
->regs
[i
].live
|= REG_LIVE_DONE
;
8641 if (!(live
& REG_LIVE_READ
))
8642 /* since the register is unused, clear its state
8643 * to make further comparison simpler
8645 __mark_reg_not_init(env
, &st
->regs
[i
]);
8648 for (i
= 0; i
< st
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
8649 live
= st
->stack
[i
].spilled_ptr
.live
;
8650 /* liveness must not touch this stack slot anymore */
8651 st
->stack
[i
].spilled_ptr
.live
|= REG_LIVE_DONE
;
8652 if (!(live
& REG_LIVE_READ
)) {
8653 __mark_reg_not_init(env
, &st
->stack
[i
].spilled_ptr
);
8654 for (j
= 0; j
< BPF_REG_SIZE
; j
++)
8655 st
->stack
[i
].slot_type
[j
] = STACK_INVALID
;
8660 static void clean_verifier_state(struct bpf_verifier_env
*env
,
8661 struct bpf_verifier_state
*st
)
8665 if (st
->frame
[0]->regs
[0].live
& REG_LIVE_DONE
)
8666 /* all regs in this state in all frames were already marked */
8669 for (i
= 0; i
<= st
->curframe
; i
++)
8670 clean_func_state(env
, st
->frame
[i
]);
8673 /* the parentage chains form a tree.
8674 * the verifier states are added to state lists at given insn and
8675 * pushed into state stack for future exploration.
8676 * when the verifier reaches bpf_exit insn some of the verifer states
8677 * stored in the state lists have their final liveness state already,
8678 * but a lot of states will get revised from liveness point of view when
8679 * the verifier explores other branches.
8682 * 2: if r1 == 100 goto pc+1
8685 * when the verifier reaches exit insn the register r0 in the state list of
8686 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
8687 * of insn 2 and goes exploring further. At the insn 4 it will walk the
8688 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
8690 * Since the verifier pushes the branch states as it sees them while exploring
8691 * the program the condition of walking the branch instruction for the second
8692 * time means that all states below this branch were already explored and
8693 * their final liveness markes are already propagated.
8694 * Hence when the verifier completes the search of state list in is_state_visited()
8695 * we can call this clean_live_states() function to mark all liveness states
8696 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
8698 * This function also clears the registers and stack for states that !READ
8699 * to simplify state merging.
8701 * Important note here that walking the same branch instruction in the callee
8702 * doesn't meant that the states are DONE. The verifier has to compare
8705 static void clean_live_states(struct bpf_verifier_env
*env
, int insn
,
8706 struct bpf_verifier_state
*cur
)
8708 struct bpf_verifier_state_list
*sl
;
8711 sl
= *explored_state(env
, insn
);
8713 if (sl
->state
.branches
)
8715 if (sl
->state
.insn_idx
!= insn
||
8716 sl
->state
.curframe
!= cur
->curframe
)
8718 for (i
= 0; i
<= cur
->curframe
; i
++)
8719 if (sl
->state
.frame
[i
]->callsite
!= cur
->frame
[i
]->callsite
)
8721 clean_verifier_state(env
, &sl
->state
);
8727 /* Returns true if (rold safe implies rcur safe) */
8728 static bool regsafe(struct bpf_reg_state
*rold
, struct bpf_reg_state
*rcur
,
8729 struct idpair
*idmap
)
8733 if (!(rold
->live
& REG_LIVE_READ
))
8734 /* explored state didn't use this */
8737 equal
= memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, parent
)) == 0;
8739 if (rold
->type
== PTR_TO_STACK
)
8740 /* two stack pointers are equal only if they're pointing to
8741 * the same stack frame, since fp-8 in foo != fp-8 in bar
8743 return equal
&& rold
->frameno
== rcur
->frameno
;
8748 if (rold
->type
== NOT_INIT
)
8749 /* explored state can't have used this */
8751 if (rcur
->type
== NOT_INIT
)
8753 switch (rold
->type
) {
8755 if (rcur
->type
== SCALAR_VALUE
) {
8756 if (!rold
->precise
&& !rcur
->precise
)
8758 /* new val must satisfy old val knowledge */
8759 return range_within(rold
, rcur
) &&
8760 tnum_in(rold
->var_off
, rcur
->var_off
);
8762 /* We're trying to use a pointer in place of a scalar.
8763 * Even if the scalar was unbounded, this could lead to
8764 * pointer leaks because scalars are allowed to leak
8765 * while pointers are not. We could make this safe in
8766 * special cases if root is calling us, but it's
8767 * probably not worth the hassle.
8771 case PTR_TO_MAP_VALUE
:
8772 /* If the new min/max/var_off satisfy the old ones and
8773 * everything else matches, we are OK.
8774 * 'id' is not compared, since it's only used for maps with
8775 * bpf_spin_lock inside map element and in such cases if
8776 * the rest of the prog is valid for one map element then
8777 * it's valid for all map elements regardless of the key
8778 * used in bpf_map_lookup()
8780 return memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, id
)) == 0 &&
8781 range_within(rold
, rcur
) &&
8782 tnum_in(rold
->var_off
, rcur
->var_off
);
8783 case PTR_TO_MAP_VALUE_OR_NULL
:
8784 /* a PTR_TO_MAP_VALUE could be safe to use as a
8785 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
8786 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
8787 * checked, doing so could have affected others with the same
8788 * id, and we can't check for that because we lost the id when
8789 * we converted to a PTR_TO_MAP_VALUE.
8791 if (rcur
->type
!= PTR_TO_MAP_VALUE_OR_NULL
)
8793 if (memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, id
)))
8795 /* Check our ids match any regs they're supposed to */
8796 return check_ids(rold
->id
, rcur
->id
, idmap
);
8797 case PTR_TO_PACKET_META
:
8799 if (rcur
->type
!= rold
->type
)
8801 /* We must have at least as much range as the old ptr
8802 * did, so that any accesses which were safe before are
8803 * still safe. This is true even if old range < old off,
8804 * since someone could have accessed through (ptr - k), or
8805 * even done ptr -= k in a register, to get a safe access.
8807 if (rold
->range
> rcur
->range
)
8809 /* If the offsets don't match, we can't trust our alignment;
8810 * nor can we be sure that we won't fall out of range.
8812 if (rold
->off
!= rcur
->off
)
8814 /* id relations must be preserved */
8815 if (rold
->id
&& !check_ids(rold
->id
, rcur
->id
, idmap
))
8817 /* new val must satisfy old val knowledge */
8818 return range_within(rold
, rcur
) &&
8819 tnum_in(rold
->var_off
, rcur
->var_off
);
8821 case CONST_PTR_TO_MAP
:
8822 case PTR_TO_PACKET_END
:
8823 case PTR_TO_FLOW_KEYS
:
8825 case PTR_TO_SOCKET_OR_NULL
:
8826 case PTR_TO_SOCK_COMMON
:
8827 case PTR_TO_SOCK_COMMON_OR_NULL
:
8828 case PTR_TO_TCP_SOCK
:
8829 case PTR_TO_TCP_SOCK_OR_NULL
:
8830 case PTR_TO_XDP_SOCK
:
8831 /* Only valid matches are exact, which memcmp() above
8832 * would have accepted
8835 /* Don't know what's going on, just say it's not safe */
8839 /* Shouldn't get here; if we do, say it's not safe */
8844 static bool stacksafe(struct bpf_func_state
*old
,
8845 struct bpf_func_state
*cur
,
8846 struct idpair
*idmap
)
8850 /* walk slots of the explored stack and ignore any additional
8851 * slots in the current stack, since explored(safe) state
8854 for (i
= 0; i
< old
->allocated_stack
; i
++) {
8855 spi
= i
/ BPF_REG_SIZE
;
8857 if (!(old
->stack
[spi
].spilled_ptr
.live
& REG_LIVE_READ
)) {
8858 i
+= BPF_REG_SIZE
- 1;
8859 /* explored state didn't use this */
8863 if (old
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
] == STACK_INVALID
)
8866 /* explored stack has more populated slots than current stack
8867 * and these slots were used
8869 if (i
>= cur
->allocated_stack
)
8872 /* if old state was safe with misc data in the stack
8873 * it will be safe with zero-initialized stack.
8874 * The opposite is not true
8876 if (old
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
] == STACK_MISC
&&
8877 cur
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
] == STACK_ZERO
)
8879 if (old
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
] !=
8880 cur
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
])
8881 /* Ex: old explored (safe) state has STACK_SPILL in
8882 * this stack slot, but current has STACK_MISC ->
8883 * this verifier states are not equivalent,
8884 * return false to continue verification of this path
8887 if (i
% BPF_REG_SIZE
)
8889 if (old
->stack
[spi
].slot_type
[0] != STACK_SPILL
)
8891 if (!regsafe(&old
->stack
[spi
].spilled_ptr
,
8892 &cur
->stack
[spi
].spilled_ptr
,
8894 /* when explored and current stack slot are both storing
8895 * spilled registers, check that stored pointers types
8896 * are the same as well.
8897 * Ex: explored safe path could have stored
8898 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
8899 * but current path has stored:
8900 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
8901 * such verifier states are not equivalent.
8902 * return false to continue verification of this path
8909 static bool refsafe(struct bpf_func_state
*old
, struct bpf_func_state
*cur
)
8911 if (old
->acquired_refs
!= cur
->acquired_refs
)
8913 return !memcmp(old
->refs
, cur
->refs
,
8914 sizeof(*old
->refs
) * old
->acquired_refs
);
8917 /* compare two verifier states
8919 * all states stored in state_list are known to be valid, since
8920 * verifier reached 'bpf_exit' instruction through them
8922 * this function is called when verifier exploring different branches of
8923 * execution popped from the state stack. If it sees an old state that has
8924 * more strict register state and more strict stack state then this execution
8925 * branch doesn't need to be explored further, since verifier already
8926 * concluded that more strict state leads to valid finish.
8928 * Therefore two states are equivalent if register state is more conservative
8929 * and explored stack state is more conservative than the current one.
8932 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
8933 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
8935 * In other words if current stack state (one being explored) has more
8936 * valid slots than old one that already passed validation, it means
8937 * the verifier can stop exploring and conclude that current state is valid too
8939 * Similarly with registers. If explored state has register type as invalid
8940 * whereas register type in current state is meaningful, it means that
8941 * the current state will reach 'bpf_exit' instruction safely
8943 static bool func_states_equal(struct bpf_func_state
*old
,
8944 struct bpf_func_state
*cur
)
8946 struct idpair
*idmap
;
8950 idmap
= kcalloc(ID_MAP_SIZE
, sizeof(struct idpair
), GFP_KERNEL
);
8951 /* If we failed to allocate the idmap, just say it's not safe */
8955 for (i
= 0; i
< MAX_BPF_REG
; i
++) {
8956 if (!regsafe(&old
->regs
[i
], &cur
->regs
[i
], idmap
))
8960 if (!stacksafe(old
, cur
, idmap
))
8963 if (!refsafe(old
, cur
))
8971 static bool states_equal(struct bpf_verifier_env
*env
,
8972 struct bpf_verifier_state
*old
,
8973 struct bpf_verifier_state
*cur
)
8977 if (old
->curframe
!= cur
->curframe
)
8980 /* Verification state from speculative execution simulation
8981 * must never prune a non-speculative execution one.
8983 if (old
->speculative
&& !cur
->speculative
)
8986 if (old
->active_spin_lock
!= cur
->active_spin_lock
)
8989 /* for states to be equal callsites have to be the same
8990 * and all frame states need to be equivalent
8992 for (i
= 0; i
<= old
->curframe
; i
++) {
8993 if (old
->frame
[i
]->callsite
!= cur
->frame
[i
]->callsite
)
8995 if (!func_states_equal(old
->frame
[i
], cur
->frame
[i
]))
9001 /* Return 0 if no propagation happened. Return negative error code if error
9002 * happened. Otherwise, return the propagated bit.
9004 static int propagate_liveness_reg(struct bpf_verifier_env
*env
,
9005 struct bpf_reg_state
*reg
,
9006 struct bpf_reg_state
*parent_reg
)
9008 u8 parent_flag
= parent_reg
->live
& REG_LIVE_READ
;
9009 u8 flag
= reg
->live
& REG_LIVE_READ
;
9012 /* When comes here, read flags of PARENT_REG or REG could be any of
9013 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
9014 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
9016 if (parent_flag
== REG_LIVE_READ64
||
9017 /* Or if there is no read flag from REG. */
9019 /* Or if the read flag from REG is the same as PARENT_REG. */
9020 parent_flag
== flag
)
9023 err
= mark_reg_read(env
, reg
, parent_reg
, flag
);
9030 /* A write screens off any subsequent reads; but write marks come from the
9031 * straight-line code between a state and its parent. When we arrive at an
9032 * equivalent state (jump target or such) we didn't arrive by the straight-line
9033 * code, so read marks in the state must propagate to the parent regardless
9034 * of the state's write marks. That's what 'parent == state->parent' comparison
9035 * in mark_reg_read() is for.
9037 static int propagate_liveness(struct bpf_verifier_env
*env
,
9038 const struct bpf_verifier_state
*vstate
,
9039 struct bpf_verifier_state
*vparent
)
9041 struct bpf_reg_state
*state_reg
, *parent_reg
;
9042 struct bpf_func_state
*state
, *parent
;
9043 int i
, frame
, err
= 0;
9045 if (vparent
->curframe
!= vstate
->curframe
) {
9046 WARN(1, "propagate_live: parent frame %d current frame %d\n",
9047 vparent
->curframe
, vstate
->curframe
);
9050 /* Propagate read liveness of registers... */
9051 BUILD_BUG_ON(BPF_REG_FP
+ 1 != MAX_BPF_REG
);
9052 for (frame
= 0; frame
<= vstate
->curframe
; frame
++) {
9053 parent
= vparent
->frame
[frame
];
9054 state
= vstate
->frame
[frame
];
9055 parent_reg
= parent
->regs
;
9056 state_reg
= state
->regs
;
9057 /* We don't need to worry about FP liveness, it's read-only */
9058 for (i
= frame
< vstate
->curframe
? BPF_REG_6
: 0; i
< BPF_REG_FP
; i
++) {
9059 err
= propagate_liveness_reg(env
, &state_reg
[i
],
9063 if (err
== REG_LIVE_READ64
)
9064 mark_insn_zext(env
, &parent_reg
[i
]);
9067 /* Propagate stack slots. */
9068 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
&&
9069 i
< parent
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
9070 parent_reg
= &parent
->stack
[i
].spilled_ptr
;
9071 state_reg
= &state
->stack
[i
].spilled_ptr
;
9072 err
= propagate_liveness_reg(env
, state_reg
,
9081 /* find precise scalars in the previous equivalent state and
9082 * propagate them into the current state
9084 static int propagate_precision(struct bpf_verifier_env
*env
,
9085 const struct bpf_verifier_state
*old
)
9087 struct bpf_reg_state
*state_reg
;
9088 struct bpf_func_state
*state
;
9091 state
= old
->frame
[old
->curframe
];
9092 state_reg
= state
->regs
;
9093 for (i
= 0; i
< BPF_REG_FP
; i
++, state_reg
++) {
9094 if (state_reg
->type
!= SCALAR_VALUE
||
9095 !state_reg
->precise
)
9097 if (env
->log
.level
& BPF_LOG_LEVEL2
)
9098 verbose(env
, "propagating r%d\n", i
);
9099 err
= mark_chain_precision(env
, i
);
9104 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
9105 if (state
->stack
[i
].slot_type
[0] != STACK_SPILL
)
9107 state_reg
= &state
->stack
[i
].spilled_ptr
;
9108 if (state_reg
->type
!= SCALAR_VALUE
||
9109 !state_reg
->precise
)
9111 if (env
->log
.level
& BPF_LOG_LEVEL2
)
9112 verbose(env
, "propagating fp%d\n",
9113 (-i
- 1) * BPF_REG_SIZE
);
9114 err
= mark_chain_precision_stack(env
, i
);
9121 static bool states_maybe_looping(struct bpf_verifier_state
*old
,
9122 struct bpf_verifier_state
*cur
)
9124 struct bpf_func_state
*fold
, *fcur
;
9125 int i
, fr
= cur
->curframe
;
9127 if (old
->curframe
!= fr
)
9130 fold
= old
->frame
[fr
];
9131 fcur
= cur
->frame
[fr
];
9132 for (i
= 0; i
< MAX_BPF_REG
; i
++)
9133 if (memcmp(&fold
->regs
[i
], &fcur
->regs
[i
],
9134 offsetof(struct bpf_reg_state
, parent
)))
9140 static int is_state_visited(struct bpf_verifier_env
*env
, int insn_idx
)
9142 struct bpf_verifier_state_list
*new_sl
;
9143 struct bpf_verifier_state_list
*sl
, **pprev
;
9144 struct bpf_verifier_state
*cur
= env
->cur_state
, *new;
9145 int i
, j
, err
, states_cnt
= 0;
9146 bool add_new_state
= env
->test_state_freq
? true : false;
9148 cur
->last_insn_idx
= env
->prev_insn_idx
;
9149 if (!env
->insn_aux_data
[insn_idx
].prune_point
)
9150 /* this 'insn_idx' instruction wasn't marked, so we will not
9151 * be doing state search here
9155 /* bpf progs typically have pruning point every 4 instructions
9156 * http://vger.kernel.org/bpfconf2019.html#session-1
9157 * Do not add new state for future pruning if the verifier hasn't seen
9158 * at least 2 jumps and at least 8 instructions.
9159 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
9160 * In tests that amounts to up to 50% reduction into total verifier
9161 * memory consumption and 20% verifier time speedup.
9163 if (env
->jmps_processed
- env
->prev_jmps_processed
>= 2 &&
9164 env
->insn_processed
- env
->prev_insn_processed
>= 8)
9165 add_new_state
= true;
9167 pprev
= explored_state(env
, insn_idx
);
9170 clean_live_states(env
, insn_idx
, cur
);
9174 if (sl
->state
.insn_idx
!= insn_idx
)
9176 if (sl
->state
.branches
) {
9177 if (states_maybe_looping(&sl
->state
, cur
) &&
9178 states_equal(env
, &sl
->state
, cur
)) {
9179 verbose_linfo(env
, insn_idx
, "; ");
9180 verbose(env
, "infinite loop detected at insn %d\n", insn_idx
);
9183 /* if the verifier is processing a loop, avoid adding new state
9184 * too often, since different loop iterations have distinct
9185 * states and may not help future pruning.
9186 * This threshold shouldn't be too low to make sure that
9187 * a loop with large bound will be rejected quickly.
9188 * The most abusive loop will be:
9190 * if r1 < 1000000 goto pc-2
9191 * 1M insn_procssed limit / 100 == 10k peak states.
9192 * This threshold shouldn't be too high either, since states
9193 * at the end of the loop are likely to be useful in pruning.
9195 if (env
->jmps_processed
- env
->prev_jmps_processed
< 20 &&
9196 env
->insn_processed
- env
->prev_insn_processed
< 100)
9197 add_new_state
= false;
9200 if (states_equal(env
, &sl
->state
, cur
)) {
9202 /* reached equivalent register/stack state,
9204 * Registers read by the continuation are read by us.
9205 * If we have any write marks in env->cur_state, they
9206 * will prevent corresponding reads in the continuation
9207 * from reaching our parent (an explored_state). Our
9208 * own state will get the read marks recorded, but
9209 * they'll be immediately forgotten as we're pruning
9210 * this state and will pop a new one.
9212 err
= propagate_liveness(env
, &sl
->state
, cur
);
9214 /* if previous state reached the exit with precision and
9215 * current state is equivalent to it (except precsion marks)
9216 * the precision needs to be propagated back in
9217 * the current state.
9219 err
= err
? : push_jmp_history(env
, cur
);
9220 err
= err
? : propagate_precision(env
, &sl
->state
);
9226 /* when new state is not going to be added do not increase miss count.
9227 * Otherwise several loop iterations will remove the state
9228 * recorded earlier. The goal of these heuristics is to have
9229 * states from some iterations of the loop (some in the beginning
9230 * and some at the end) to help pruning.
9234 /* heuristic to determine whether this state is beneficial
9235 * to keep checking from state equivalence point of view.
9236 * Higher numbers increase max_states_per_insn and verification time,
9237 * but do not meaningfully decrease insn_processed.
9239 if (sl
->miss_cnt
> sl
->hit_cnt
* 3 + 3) {
9240 /* the state is unlikely to be useful. Remove it to
9241 * speed up verification
9244 if (sl
->state
.frame
[0]->regs
[0].live
& REG_LIVE_DONE
) {
9245 u32 br
= sl
->state
.branches
;
9248 "BUG live_done but branches_to_explore %d\n",
9250 free_verifier_state(&sl
->state
, false);
9254 /* cannot free this state, since parentage chain may
9255 * walk it later. Add it for free_list instead to
9256 * be freed at the end of verification
9258 sl
->next
= env
->free_list
;
9259 env
->free_list
= sl
;
9269 if (env
->max_states_per_insn
< states_cnt
)
9270 env
->max_states_per_insn
= states_cnt
;
9272 if (!env
->bpf_capable
&& states_cnt
> BPF_COMPLEXITY_LIMIT_STATES
)
9273 return push_jmp_history(env
, cur
);
9276 return push_jmp_history(env
, cur
);
9278 /* There were no equivalent states, remember the current one.
9279 * Technically the current state is not proven to be safe yet,
9280 * but it will either reach outer most bpf_exit (which means it's safe)
9281 * or it will be rejected. When there are no loops the verifier won't be
9282 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
9283 * again on the way to bpf_exit.
9284 * When looping the sl->state.branches will be > 0 and this state
9285 * will not be considered for equivalence until branches == 0.
9287 new_sl
= kzalloc(sizeof(struct bpf_verifier_state_list
), GFP_KERNEL
);
9290 env
->total_states
++;
9292 env
->prev_jmps_processed
= env
->jmps_processed
;
9293 env
->prev_insn_processed
= env
->insn_processed
;
9295 /* add new state to the head of linked list */
9296 new = &new_sl
->state
;
9297 err
= copy_verifier_state(new, cur
);
9299 free_verifier_state(new, false);
9303 new->insn_idx
= insn_idx
;
9304 WARN_ONCE(new->branches
!= 1,
9305 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches
, insn_idx
);
9308 cur
->first_insn_idx
= insn_idx
;
9309 clear_jmp_history(cur
);
9310 new_sl
->next
= *explored_state(env
, insn_idx
);
9311 *explored_state(env
, insn_idx
) = new_sl
;
9312 /* connect new state to parentage chain. Current frame needs all
9313 * registers connected. Only r6 - r9 of the callers are alive (pushed
9314 * to the stack implicitly by JITs) so in callers' frames connect just
9315 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
9316 * the state of the call instruction (with WRITTEN set), and r0 comes
9317 * from callee with its full parentage chain, anyway.
9319 /* clear write marks in current state: the writes we did are not writes
9320 * our child did, so they don't screen off its reads from us.
9321 * (There are no read marks in current state, because reads always mark
9322 * their parent and current state never has children yet. Only
9323 * explored_states can get read marks.)
9325 for (j
= 0; j
<= cur
->curframe
; j
++) {
9326 for (i
= j
< cur
->curframe
? BPF_REG_6
: 0; i
< BPF_REG_FP
; i
++)
9327 cur
->frame
[j
]->regs
[i
].parent
= &new->frame
[j
]->regs
[i
];
9328 for (i
= 0; i
< BPF_REG_FP
; i
++)
9329 cur
->frame
[j
]->regs
[i
].live
= REG_LIVE_NONE
;
9332 /* all stack frames are accessible from callee, clear them all */
9333 for (j
= 0; j
<= cur
->curframe
; j
++) {
9334 struct bpf_func_state
*frame
= cur
->frame
[j
];
9335 struct bpf_func_state
*newframe
= new->frame
[j
];
9337 for (i
= 0; i
< frame
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
9338 frame
->stack
[i
].spilled_ptr
.live
= REG_LIVE_NONE
;
9339 frame
->stack
[i
].spilled_ptr
.parent
=
9340 &newframe
->stack
[i
].spilled_ptr
;
9346 /* Return true if it's OK to have the same insn return a different type. */
9347 static bool reg_type_mismatch_ok(enum bpf_reg_type type
)
9352 case PTR_TO_SOCKET_OR_NULL
:
9353 case PTR_TO_SOCK_COMMON
:
9354 case PTR_TO_SOCK_COMMON_OR_NULL
:
9355 case PTR_TO_TCP_SOCK
:
9356 case PTR_TO_TCP_SOCK_OR_NULL
:
9357 case PTR_TO_XDP_SOCK
:
9359 case PTR_TO_BTF_ID_OR_NULL
:
9366 /* If an instruction was previously used with particular pointer types, then we
9367 * need to be careful to avoid cases such as the below, where it may be ok
9368 * for one branch accessing the pointer, but not ok for the other branch:
9373 * R1 = some_other_valid_ptr;
9376 * R2 = *(u32 *)(R1 + 0);
9378 static bool reg_type_mismatch(enum bpf_reg_type src
, enum bpf_reg_type prev
)
9380 return src
!= prev
&& (!reg_type_mismatch_ok(src
) ||
9381 !reg_type_mismatch_ok(prev
));
9384 static int do_check(struct bpf_verifier_env
*env
)
9386 bool pop_log
= !(env
->log
.level
& BPF_LOG_LEVEL2
);
9387 struct bpf_verifier_state
*state
= env
->cur_state
;
9388 struct bpf_insn
*insns
= env
->prog
->insnsi
;
9389 struct bpf_reg_state
*regs
;
9390 int insn_cnt
= env
->prog
->len
;
9391 bool do_print_state
= false;
9392 int prev_insn_idx
= -1;
9395 struct bpf_insn
*insn
;
9399 env
->prev_insn_idx
= prev_insn_idx
;
9400 if (env
->insn_idx
>= insn_cnt
) {
9401 verbose(env
, "invalid insn idx %d insn_cnt %d\n",
9402 env
->insn_idx
, insn_cnt
);
9406 insn
= &insns
[env
->insn_idx
];
9407 class = BPF_CLASS(insn
->code
);
9409 if (++env
->insn_processed
> BPF_COMPLEXITY_LIMIT_INSNS
) {
9411 "BPF program is too large. Processed %d insn\n",
9412 env
->insn_processed
);
9416 err
= is_state_visited(env
, env
->insn_idx
);
9420 /* found equivalent state, can prune the search */
9421 if (env
->log
.level
& BPF_LOG_LEVEL
) {
9423 verbose(env
, "\nfrom %d to %d%s: safe\n",
9424 env
->prev_insn_idx
, env
->insn_idx
,
9425 env
->cur_state
->speculative
?
9426 " (speculative execution)" : "");
9428 verbose(env
, "%d: safe\n", env
->insn_idx
);
9430 goto process_bpf_exit
;
9433 if (signal_pending(current
))
9439 if (env
->log
.level
& BPF_LOG_LEVEL2
||
9440 (env
->log
.level
& BPF_LOG_LEVEL
&& do_print_state
)) {
9441 if (env
->log
.level
& BPF_LOG_LEVEL2
)
9442 verbose(env
, "%d:", env
->insn_idx
);
9444 verbose(env
, "\nfrom %d to %d%s:",
9445 env
->prev_insn_idx
, env
->insn_idx
,
9446 env
->cur_state
->speculative
?
9447 " (speculative execution)" : "");
9448 print_verifier_state(env
, state
->frame
[state
->curframe
]);
9449 do_print_state
= false;
9452 if (env
->log
.level
& BPF_LOG_LEVEL
) {
9453 const struct bpf_insn_cbs cbs
= {
9454 .cb_print
= verbose
,
9455 .private_data
= env
,
9458 verbose_linfo(env
, env
->insn_idx
, "; ");
9459 verbose(env
, "%d: ", env
->insn_idx
);
9460 print_bpf_insn(&cbs
, insn
, env
->allow_ptr_leaks
);
9463 if (bpf_prog_is_dev_bound(env
->prog
->aux
)) {
9464 err
= bpf_prog_offload_verify_insn(env
, env
->insn_idx
,
9465 env
->prev_insn_idx
);
9470 regs
= cur_regs(env
);
9471 env
->insn_aux_data
[env
->insn_idx
].seen
= env
->pass_cnt
;
9472 prev_insn_idx
= env
->insn_idx
;
9474 if (class == BPF_ALU
|| class == BPF_ALU64
) {
9475 err
= check_alu_op(env
, insn
);
9479 } else if (class == BPF_LDX
) {
9480 enum bpf_reg_type
*prev_src_type
, src_reg_type
;
9482 /* check for reserved fields is already done */
9484 /* check src operand */
9485 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
9489 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP_NO_MARK
);
9493 src_reg_type
= regs
[insn
->src_reg
].type
;
9495 /* check that memory (src_reg + off) is readable,
9496 * the state of dst_reg will be updated by this func
9498 err
= check_mem_access(env
, env
->insn_idx
, insn
->src_reg
,
9499 insn
->off
, BPF_SIZE(insn
->code
),
9500 BPF_READ
, insn
->dst_reg
, false);
9504 prev_src_type
= &env
->insn_aux_data
[env
->insn_idx
].ptr_type
;
9506 if (*prev_src_type
== NOT_INIT
) {
9508 * dst_reg = *(u32 *)(src_reg + off)
9509 * save type to validate intersecting paths
9511 *prev_src_type
= src_reg_type
;
9513 } else if (reg_type_mismatch(src_reg_type
, *prev_src_type
)) {
9514 /* ABuser program is trying to use the same insn
9515 * dst_reg = *(u32*) (src_reg + off)
9516 * with different pointer types:
9517 * src_reg == ctx in one branch and
9518 * src_reg == stack|map in some other branch.
9521 verbose(env
, "same insn cannot be used with different pointers\n");
9525 } else if (class == BPF_STX
) {
9526 enum bpf_reg_type
*prev_dst_type
, dst_reg_type
;
9528 if (BPF_MODE(insn
->code
) == BPF_XADD
) {
9529 err
= check_xadd(env
, env
->insn_idx
, insn
);
9536 /* check src1 operand */
9537 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
9540 /* check src2 operand */
9541 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
9545 dst_reg_type
= regs
[insn
->dst_reg
].type
;
9547 /* check that memory (dst_reg + off) is writeable */
9548 err
= check_mem_access(env
, env
->insn_idx
, insn
->dst_reg
,
9549 insn
->off
, BPF_SIZE(insn
->code
),
9550 BPF_WRITE
, insn
->src_reg
, false);
9554 prev_dst_type
= &env
->insn_aux_data
[env
->insn_idx
].ptr_type
;
9556 if (*prev_dst_type
== NOT_INIT
) {
9557 *prev_dst_type
= dst_reg_type
;
9558 } else if (reg_type_mismatch(dst_reg_type
, *prev_dst_type
)) {
9559 verbose(env
, "same insn cannot be used with different pointers\n");
9563 } else if (class == BPF_ST
) {
9564 if (BPF_MODE(insn
->code
) != BPF_MEM
||
9565 insn
->src_reg
!= BPF_REG_0
) {
9566 verbose(env
, "BPF_ST uses reserved fields\n");
9569 /* check src operand */
9570 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
9574 if (is_ctx_reg(env
, insn
->dst_reg
)) {
9575 verbose(env
, "BPF_ST stores into R%d %s is not allowed\n",
9577 reg_type_str
[reg_state(env
, insn
->dst_reg
)->type
]);
9581 /* check that memory (dst_reg + off) is writeable */
9582 err
= check_mem_access(env
, env
->insn_idx
, insn
->dst_reg
,
9583 insn
->off
, BPF_SIZE(insn
->code
),
9584 BPF_WRITE
, -1, false);
9588 } else if (class == BPF_JMP
|| class == BPF_JMP32
) {
9589 u8 opcode
= BPF_OP(insn
->code
);
9591 env
->jmps_processed
++;
9592 if (opcode
== BPF_CALL
) {
9593 if (BPF_SRC(insn
->code
) != BPF_K
||
9595 (insn
->src_reg
!= BPF_REG_0
&&
9596 insn
->src_reg
!= BPF_PSEUDO_CALL
) ||
9597 insn
->dst_reg
!= BPF_REG_0
||
9598 class == BPF_JMP32
) {
9599 verbose(env
, "BPF_CALL uses reserved fields\n");
9603 if (env
->cur_state
->active_spin_lock
&&
9604 (insn
->src_reg
== BPF_PSEUDO_CALL
||
9605 insn
->imm
!= BPF_FUNC_spin_unlock
)) {
9606 verbose(env
, "function calls are not allowed while holding a lock\n");
9609 if (insn
->src_reg
== BPF_PSEUDO_CALL
)
9610 err
= check_func_call(env
, insn
, &env
->insn_idx
);
9612 err
= check_helper_call(env
, insn
->imm
, env
->insn_idx
);
9616 } else if (opcode
== BPF_JA
) {
9617 if (BPF_SRC(insn
->code
) != BPF_K
||
9619 insn
->src_reg
!= BPF_REG_0
||
9620 insn
->dst_reg
!= BPF_REG_0
||
9621 class == BPF_JMP32
) {
9622 verbose(env
, "BPF_JA uses reserved fields\n");
9626 env
->insn_idx
+= insn
->off
+ 1;
9629 } else if (opcode
== BPF_EXIT
) {
9630 if (BPF_SRC(insn
->code
) != BPF_K
||
9632 insn
->src_reg
!= BPF_REG_0
||
9633 insn
->dst_reg
!= BPF_REG_0
||
9634 class == BPF_JMP32
) {
9635 verbose(env
, "BPF_EXIT uses reserved fields\n");
9639 if (env
->cur_state
->active_spin_lock
) {
9640 verbose(env
, "bpf_spin_unlock is missing\n");
9644 if (state
->curframe
) {
9645 /* exit from nested function */
9646 err
= prepare_func_exit(env
, &env
->insn_idx
);
9649 do_print_state
= true;
9653 err
= check_reference_leak(env
);
9657 err
= check_return_code(env
);
9661 update_branch_counts(env
, env
->cur_state
);
9662 err
= pop_stack(env
, &prev_insn_idx
,
9663 &env
->insn_idx
, pop_log
);
9669 do_print_state
= true;
9673 err
= check_cond_jmp_op(env
, insn
, &env
->insn_idx
);
9677 } else if (class == BPF_LD
) {
9678 u8 mode
= BPF_MODE(insn
->code
);
9680 if (mode
== BPF_ABS
|| mode
== BPF_IND
) {
9681 err
= check_ld_abs(env
, insn
);
9685 } else if (mode
== BPF_IMM
) {
9686 err
= check_ld_imm(env
, insn
);
9691 env
->insn_aux_data
[env
->insn_idx
].seen
= env
->pass_cnt
;
9693 verbose(env
, "invalid BPF_LD mode\n");
9697 verbose(env
, "unknown insn class %d\n", class);
9707 /* replace pseudo btf_id with kernel symbol address */
9708 static int check_pseudo_btf_id(struct bpf_verifier_env
*env
,
9709 struct bpf_insn
*insn
,
9710 struct bpf_insn_aux_data
*aux
)
9712 const struct btf_var_secinfo
*vsi
;
9713 const struct btf_type
*datasec
;
9714 const struct btf_type
*t
;
9715 const char *sym_name
;
9716 bool percpu
= false;
9717 u32 type
, id
= insn
->imm
;
9723 verbose(env
, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
9727 if (insn
[1].imm
!= 0) {
9728 verbose(env
, "reserved field (insn[1].imm) is used in pseudo_btf_id ldimm64 insn.\n");
9732 t
= btf_type_by_id(btf_vmlinux
, id
);
9734 verbose(env
, "ldimm64 insn specifies invalid btf_id %d.\n", id
);
9738 if (!btf_type_is_var(t
)) {
9739 verbose(env
, "pseudo btf_id %d in ldimm64 isn't KIND_VAR.\n",
9744 sym_name
= btf_name_by_offset(btf_vmlinux
, t
->name_off
);
9745 addr
= kallsyms_lookup_name(sym_name
);
9747 verbose(env
, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
9752 datasec_id
= btf_find_by_name_kind(btf_vmlinux
, ".data..percpu",
9754 if (datasec_id
> 0) {
9755 datasec
= btf_type_by_id(btf_vmlinux
, datasec_id
);
9756 for_each_vsi(i
, datasec
, vsi
) {
9757 if (vsi
->type
== id
) {
9764 insn
[0].imm
= (u32
)addr
;
9765 insn
[1].imm
= addr
>> 32;
9768 t
= btf_type_skip_modifiers(btf_vmlinux
, type
, NULL
);
9770 aux
->btf_var
.reg_type
= PTR_TO_PERCPU_BTF_ID
;
9771 aux
->btf_var
.btf
= btf_vmlinux
;
9772 aux
->btf_var
.btf_id
= type
;
9773 } else if (!btf_type_is_struct(t
)) {
9774 const struct btf_type
*ret
;
9778 /* resolve the type size of ksym. */
9779 ret
= btf_resolve_size(btf_vmlinux
, t
, &tsize
);
9781 tname
= btf_name_by_offset(btf_vmlinux
, t
->name_off
);
9782 verbose(env
, "ldimm64 unable to resolve the size of type '%s': %ld\n",
9783 tname
, PTR_ERR(ret
));
9786 aux
->btf_var
.reg_type
= PTR_TO_MEM
;
9787 aux
->btf_var
.mem_size
= tsize
;
9789 aux
->btf_var
.reg_type
= PTR_TO_BTF_ID
;
9790 aux
->btf_var
.btf
= btf_vmlinux
;
9791 aux
->btf_var
.btf_id
= type
;
9796 static int check_map_prealloc(struct bpf_map
*map
)
9798 return (map
->map_type
!= BPF_MAP_TYPE_HASH
&&
9799 map
->map_type
!= BPF_MAP_TYPE_PERCPU_HASH
&&
9800 map
->map_type
!= BPF_MAP_TYPE_HASH_OF_MAPS
) ||
9801 !(map
->map_flags
& BPF_F_NO_PREALLOC
);
9804 static bool is_tracing_prog_type(enum bpf_prog_type type
)
9807 case BPF_PROG_TYPE_KPROBE
:
9808 case BPF_PROG_TYPE_TRACEPOINT
:
9809 case BPF_PROG_TYPE_PERF_EVENT
:
9810 case BPF_PROG_TYPE_RAW_TRACEPOINT
:
9817 static bool is_preallocated_map(struct bpf_map
*map
)
9819 if (!check_map_prealloc(map
))
9821 if (map
->inner_map_meta
&& !check_map_prealloc(map
->inner_map_meta
))
9826 static int check_map_prog_compatibility(struct bpf_verifier_env
*env
,
9827 struct bpf_map
*map
,
9828 struct bpf_prog
*prog
)
9831 enum bpf_prog_type prog_type
= resolve_prog_type(prog
);
9833 * Validate that trace type programs use preallocated hash maps.
9835 * For programs attached to PERF events this is mandatory as the
9836 * perf NMI can hit any arbitrary code sequence.
9838 * All other trace types using preallocated hash maps are unsafe as
9839 * well because tracepoint or kprobes can be inside locked regions
9840 * of the memory allocator or at a place where a recursion into the
9841 * memory allocator would see inconsistent state.
9843 * On RT enabled kernels run-time allocation of all trace type
9844 * programs is strictly prohibited due to lock type constraints. On
9845 * !RT kernels it is allowed for backwards compatibility reasons for
9846 * now, but warnings are emitted so developers are made aware of
9847 * the unsafety and can fix their programs before this is enforced.
9849 if (is_tracing_prog_type(prog_type
) && !is_preallocated_map(map
)) {
9850 if (prog_type
== BPF_PROG_TYPE_PERF_EVENT
) {
9851 verbose(env
, "perf_event programs can only use preallocated hash map\n");
9854 if (IS_ENABLED(CONFIG_PREEMPT_RT
)) {
9855 verbose(env
, "trace type programs can only use preallocated hash map\n");
9858 WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
9859 verbose(env
, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
9862 if (map_value_has_spin_lock(map
)) {
9863 if (prog_type
== BPF_PROG_TYPE_SOCKET_FILTER
) {
9864 verbose(env
, "socket filter progs cannot use bpf_spin_lock yet\n");
9868 if (is_tracing_prog_type(prog_type
)) {
9869 verbose(env
, "tracing progs cannot use bpf_spin_lock yet\n");
9873 if (prog
->aux
->sleepable
) {
9874 verbose(env
, "sleepable progs cannot use bpf_spin_lock yet\n");
9879 if ((bpf_prog_is_dev_bound(prog
->aux
) || bpf_map_is_dev_bound(map
)) &&
9880 !bpf_offload_prog_map_match(prog
, map
)) {
9881 verbose(env
, "offload device mismatch between prog and map\n");
9885 if (map
->map_type
== BPF_MAP_TYPE_STRUCT_OPS
) {
9886 verbose(env
, "bpf_struct_ops map cannot be used in prog\n");
9890 if (prog
->aux
->sleepable
)
9891 switch (map
->map_type
) {
9892 case BPF_MAP_TYPE_HASH
:
9893 case BPF_MAP_TYPE_LRU_HASH
:
9894 case BPF_MAP_TYPE_ARRAY
:
9895 if (!is_preallocated_map(map
)) {
9897 "Sleepable programs can only use preallocated hash maps\n");
9903 "Sleepable programs can only use array and hash maps\n");
9910 static bool bpf_map_is_cgroup_storage(struct bpf_map
*map
)
9912 return (map
->map_type
== BPF_MAP_TYPE_CGROUP_STORAGE
||
9913 map
->map_type
== BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE
);
9916 /* find and rewrite pseudo imm in ld_imm64 instructions:
9918 * 1. if it accesses map FD, replace it with actual map pointer.
9919 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
9921 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
9923 static int resolve_pseudo_ldimm64(struct bpf_verifier_env
*env
)
9925 struct bpf_insn
*insn
= env
->prog
->insnsi
;
9926 int insn_cnt
= env
->prog
->len
;
9929 err
= bpf_prog_calc_tag(env
->prog
);
9933 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
9934 if (BPF_CLASS(insn
->code
) == BPF_LDX
&&
9935 (BPF_MODE(insn
->code
) != BPF_MEM
|| insn
->imm
!= 0)) {
9936 verbose(env
, "BPF_LDX uses reserved fields\n");
9940 if (BPF_CLASS(insn
->code
) == BPF_STX
&&
9941 ((BPF_MODE(insn
->code
) != BPF_MEM
&&
9942 BPF_MODE(insn
->code
) != BPF_XADD
) || insn
->imm
!= 0)) {
9943 verbose(env
, "BPF_STX uses reserved fields\n");
9947 if (insn
[0].code
== (BPF_LD
| BPF_IMM
| BPF_DW
)) {
9948 struct bpf_insn_aux_data
*aux
;
9949 struct bpf_map
*map
;
9953 if (i
== insn_cnt
- 1 || insn
[1].code
!= 0 ||
9954 insn
[1].dst_reg
!= 0 || insn
[1].src_reg
!= 0 ||
9956 verbose(env
, "invalid bpf_ld_imm64 insn\n");
9960 if (insn
[0].src_reg
== 0)
9961 /* valid generic load 64-bit imm */
9964 if (insn
[0].src_reg
== BPF_PSEUDO_BTF_ID
) {
9965 aux
= &env
->insn_aux_data
[i
];
9966 err
= check_pseudo_btf_id(env
, insn
, aux
);
9972 /* In final convert_pseudo_ld_imm64() step, this is
9973 * converted into regular 64-bit imm load insn.
9975 if ((insn
[0].src_reg
!= BPF_PSEUDO_MAP_FD
&&
9976 insn
[0].src_reg
!= BPF_PSEUDO_MAP_VALUE
) ||
9977 (insn
[0].src_reg
== BPF_PSEUDO_MAP_FD
&&
9978 insn
[1].imm
!= 0)) {
9980 "unrecognized bpf_ld_imm64 insn\n");
9984 f
= fdget(insn
[0].imm
);
9985 map
= __bpf_map_get(f
);
9987 verbose(env
, "fd %d is not pointing to valid bpf_map\n",
9989 return PTR_ERR(map
);
9992 err
= check_map_prog_compatibility(env
, map
, env
->prog
);
9998 aux
= &env
->insn_aux_data
[i
];
9999 if (insn
->src_reg
== BPF_PSEUDO_MAP_FD
) {
10000 addr
= (unsigned long)map
;
10002 u32 off
= insn
[1].imm
;
10004 if (off
>= BPF_MAX_VAR_OFF
) {
10005 verbose(env
, "direct value offset of %u is not allowed\n", off
);
10010 if (!map
->ops
->map_direct_value_addr
) {
10011 verbose(env
, "no direct value access support for this map type\n");
10016 err
= map
->ops
->map_direct_value_addr(map
, &addr
, off
);
10018 verbose(env
, "invalid access to map value pointer, value_size=%u off=%u\n",
10019 map
->value_size
, off
);
10024 aux
->map_off
= off
;
10028 insn
[0].imm
= (u32
)addr
;
10029 insn
[1].imm
= addr
>> 32;
10031 /* check whether we recorded this map already */
10032 for (j
= 0; j
< env
->used_map_cnt
; j
++) {
10033 if (env
->used_maps
[j
] == map
) {
10034 aux
->map_index
= j
;
10040 if (env
->used_map_cnt
>= MAX_USED_MAPS
) {
10045 /* hold the map. If the program is rejected by verifier,
10046 * the map will be released by release_maps() or it
10047 * will be used by the valid program until it's unloaded
10048 * and all maps are released in free_used_maps()
10052 aux
->map_index
= env
->used_map_cnt
;
10053 env
->used_maps
[env
->used_map_cnt
++] = map
;
10055 if (bpf_map_is_cgroup_storage(map
) &&
10056 bpf_cgroup_storage_assign(env
->prog
->aux
, map
)) {
10057 verbose(env
, "only one cgroup storage of each type is allowed\n");
10069 /* Basic sanity check before we invest more work here. */
10070 if (!bpf_opcode_in_insntable(insn
->code
)) {
10071 verbose(env
, "unknown opcode %02x\n", insn
->code
);
10076 /* now all pseudo BPF_LD_IMM64 instructions load valid
10077 * 'struct bpf_map *' into a register instead of user map_fd.
10078 * These pointers will be used later by verifier to validate map access.
10083 /* drop refcnt of maps used by the rejected program */
10084 static void release_maps(struct bpf_verifier_env
*env
)
10086 __bpf_free_used_maps(env
->prog
->aux
, env
->used_maps
,
10087 env
->used_map_cnt
);
10090 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
10091 static void convert_pseudo_ld_imm64(struct bpf_verifier_env
*env
)
10093 struct bpf_insn
*insn
= env
->prog
->insnsi
;
10094 int insn_cnt
= env
->prog
->len
;
10097 for (i
= 0; i
< insn_cnt
; i
++, insn
++)
10098 if (insn
->code
== (BPF_LD
| BPF_IMM
| BPF_DW
))
10102 /* single env->prog->insni[off] instruction was replaced with the range
10103 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
10104 * [0, off) and [off, end) to new locations, so the patched range stays zero
10106 static int adjust_insn_aux_data(struct bpf_verifier_env
*env
,
10107 struct bpf_prog
*new_prog
, u32 off
, u32 cnt
)
10109 struct bpf_insn_aux_data
*new_data
, *old_data
= env
->insn_aux_data
;
10110 struct bpf_insn
*insn
= new_prog
->insnsi
;
10114 /* aux info at OFF always needs adjustment, no matter fast path
10115 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
10116 * original insn at old prog.
10118 old_data
[off
].zext_dst
= insn_has_def32(env
, insn
+ off
+ cnt
- 1);
10122 prog_len
= new_prog
->len
;
10123 new_data
= vzalloc(array_size(prog_len
,
10124 sizeof(struct bpf_insn_aux_data
)));
10127 memcpy(new_data
, old_data
, sizeof(struct bpf_insn_aux_data
) * off
);
10128 memcpy(new_data
+ off
+ cnt
- 1, old_data
+ off
,
10129 sizeof(struct bpf_insn_aux_data
) * (prog_len
- off
- cnt
+ 1));
10130 for (i
= off
; i
< off
+ cnt
- 1; i
++) {
10131 new_data
[i
].seen
= env
->pass_cnt
;
10132 new_data
[i
].zext_dst
= insn_has_def32(env
, insn
+ i
);
10134 env
->insn_aux_data
= new_data
;
10139 static void adjust_subprog_starts(struct bpf_verifier_env
*env
, u32 off
, u32 len
)
10145 /* NOTE: fake 'exit' subprog should be updated as well. */
10146 for (i
= 0; i
<= env
->subprog_cnt
; i
++) {
10147 if (env
->subprog_info
[i
].start
<= off
)
10149 env
->subprog_info
[i
].start
+= len
- 1;
10153 static void adjust_poke_descs(struct bpf_prog
*prog
, u32 len
)
10155 struct bpf_jit_poke_descriptor
*tab
= prog
->aux
->poke_tab
;
10156 int i
, sz
= prog
->aux
->size_poke_tab
;
10157 struct bpf_jit_poke_descriptor
*desc
;
10159 for (i
= 0; i
< sz
; i
++) {
10161 desc
->insn_idx
+= len
- 1;
10165 static struct bpf_prog
*bpf_patch_insn_data(struct bpf_verifier_env
*env
, u32 off
,
10166 const struct bpf_insn
*patch
, u32 len
)
10168 struct bpf_prog
*new_prog
;
10170 new_prog
= bpf_patch_insn_single(env
->prog
, off
, patch
, len
);
10171 if (IS_ERR(new_prog
)) {
10172 if (PTR_ERR(new_prog
) == -ERANGE
)
10174 "insn %d cannot be patched due to 16-bit range\n",
10175 env
->insn_aux_data
[off
].orig_idx
);
10178 if (adjust_insn_aux_data(env
, new_prog
, off
, len
))
10180 adjust_subprog_starts(env
, off
, len
);
10181 adjust_poke_descs(new_prog
, len
);
10185 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env
*env
,
10190 /* find first prog starting at or after off (first to remove) */
10191 for (i
= 0; i
< env
->subprog_cnt
; i
++)
10192 if (env
->subprog_info
[i
].start
>= off
)
10194 /* find first prog starting at or after off + cnt (first to stay) */
10195 for (j
= i
; j
< env
->subprog_cnt
; j
++)
10196 if (env
->subprog_info
[j
].start
>= off
+ cnt
)
10198 /* if j doesn't start exactly at off + cnt, we are just removing
10199 * the front of previous prog
10201 if (env
->subprog_info
[j
].start
!= off
+ cnt
)
10205 struct bpf_prog_aux
*aux
= env
->prog
->aux
;
10208 /* move fake 'exit' subprog as well */
10209 move
= env
->subprog_cnt
+ 1 - j
;
10211 memmove(env
->subprog_info
+ i
,
10212 env
->subprog_info
+ j
,
10213 sizeof(*env
->subprog_info
) * move
);
10214 env
->subprog_cnt
-= j
- i
;
10216 /* remove func_info */
10217 if (aux
->func_info
) {
10218 move
= aux
->func_info_cnt
- j
;
10220 memmove(aux
->func_info
+ i
,
10221 aux
->func_info
+ j
,
10222 sizeof(*aux
->func_info
) * move
);
10223 aux
->func_info_cnt
-= j
- i
;
10224 /* func_info->insn_off is set after all code rewrites,
10225 * in adjust_btf_func() - no need to adjust
10229 /* convert i from "first prog to remove" to "first to adjust" */
10230 if (env
->subprog_info
[i
].start
== off
)
10234 /* update fake 'exit' subprog as well */
10235 for (; i
<= env
->subprog_cnt
; i
++)
10236 env
->subprog_info
[i
].start
-= cnt
;
10241 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env
*env
, u32 off
,
10244 struct bpf_prog
*prog
= env
->prog
;
10245 u32 i
, l_off
, l_cnt
, nr_linfo
;
10246 struct bpf_line_info
*linfo
;
10248 nr_linfo
= prog
->aux
->nr_linfo
;
10252 linfo
= prog
->aux
->linfo
;
10254 /* find first line info to remove, count lines to be removed */
10255 for (i
= 0; i
< nr_linfo
; i
++)
10256 if (linfo
[i
].insn_off
>= off
)
10261 for (; i
< nr_linfo
; i
++)
10262 if (linfo
[i
].insn_off
< off
+ cnt
)
10267 /* First live insn doesn't match first live linfo, it needs to "inherit"
10268 * last removed linfo. prog is already modified, so prog->len == off
10269 * means no live instructions after (tail of the program was removed).
10271 if (prog
->len
!= off
&& l_cnt
&&
10272 (i
== nr_linfo
|| linfo
[i
].insn_off
!= off
+ cnt
)) {
10274 linfo
[--i
].insn_off
= off
+ cnt
;
10277 /* remove the line info which refer to the removed instructions */
10279 memmove(linfo
+ l_off
, linfo
+ i
,
10280 sizeof(*linfo
) * (nr_linfo
- i
));
10282 prog
->aux
->nr_linfo
-= l_cnt
;
10283 nr_linfo
= prog
->aux
->nr_linfo
;
10286 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
10287 for (i
= l_off
; i
< nr_linfo
; i
++)
10288 linfo
[i
].insn_off
-= cnt
;
10290 /* fix up all subprogs (incl. 'exit') which start >= off */
10291 for (i
= 0; i
<= env
->subprog_cnt
; i
++)
10292 if (env
->subprog_info
[i
].linfo_idx
> l_off
) {
10293 /* program may have started in the removed region but
10294 * may not be fully removed
10296 if (env
->subprog_info
[i
].linfo_idx
>= l_off
+ l_cnt
)
10297 env
->subprog_info
[i
].linfo_idx
-= l_cnt
;
10299 env
->subprog_info
[i
].linfo_idx
= l_off
;
10305 static int verifier_remove_insns(struct bpf_verifier_env
*env
, u32 off
, u32 cnt
)
10307 struct bpf_insn_aux_data
*aux_data
= env
->insn_aux_data
;
10308 unsigned int orig_prog_len
= env
->prog
->len
;
10311 if (bpf_prog_is_dev_bound(env
->prog
->aux
))
10312 bpf_prog_offload_remove_insns(env
, off
, cnt
);
10314 err
= bpf_remove_insns(env
->prog
, off
, cnt
);
10318 err
= adjust_subprog_starts_after_remove(env
, off
, cnt
);
10322 err
= bpf_adj_linfo_after_remove(env
, off
, cnt
);
10326 memmove(aux_data
+ off
, aux_data
+ off
+ cnt
,
10327 sizeof(*aux_data
) * (orig_prog_len
- off
- cnt
));
10332 /* The verifier does more data flow analysis than llvm and will not
10333 * explore branches that are dead at run time. Malicious programs can
10334 * have dead code too. Therefore replace all dead at-run-time code
10337 * Just nops are not optimal, e.g. if they would sit at the end of the
10338 * program and through another bug we would manage to jump there, then
10339 * we'd execute beyond program memory otherwise. Returning exception
10340 * code also wouldn't work since we can have subprogs where the dead
10341 * code could be located.
10343 static void sanitize_dead_code(struct bpf_verifier_env
*env
)
10345 struct bpf_insn_aux_data
*aux_data
= env
->insn_aux_data
;
10346 struct bpf_insn trap
= BPF_JMP_IMM(BPF_JA
, 0, 0, -1);
10347 struct bpf_insn
*insn
= env
->prog
->insnsi
;
10348 const int insn_cnt
= env
->prog
->len
;
10351 for (i
= 0; i
< insn_cnt
; i
++) {
10352 if (aux_data
[i
].seen
)
10354 memcpy(insn
+ i
, &trap
, sizeof(trap
));
10358 static bool insn_is_cond_jump(u8 code
)
10362 if (BPF_CLASS(code
) == BPF_JMP32
)
10365 if (BPF_CLASS(code
) != BPF_JMP
)
10369 return op
!= BPF_JA
&& op
!= BPF_EXIT
&& op
!= BPF_CALL
;
10372 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env
*env
)
10374 struct bpf_insn_aux_data
*aux_data
= env
->insn_aux_data
;
10375 struct bpf_insn ja
= BPF_JMP_IMM(BPF_JA
, 0, 0, 0);
10376 struct bpf_insn
*insn
= env
->prog
->insnsi
;
10377 const int insn_cnt
= env
->prog
->len
;
10380 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
10381 if (!insn_is_cond_jump(insn
->code
))
10384 if (!aux_data
[i
+ 1].seen
)
10385 ja
.off
= insn
->off
;
10386 else if (!aux_data
[i
+ 1 + insn
->off
].seen
)
10391 if (bpf_prog_is_dev_bound(env
->prog
->aux
))
10392 bpf_prog_offload_replace_insn(env
, i
, &ja
);
10394 memcpy(insn
, &ja
, sizeof(ja
));
10398 static int opt_remove_dead_code(struct bpf_verifier_env
*env
)
10400 struct bpf_insn_aux_data
*aux_data
= env
->insn_aux_data
;
10401 int insn_cnt
= env
->prog
->len
;
10404 for (i
= 0; i
< insn_cnt
; i
++) {
10408 while (i
+ j
< insn_cnt
&& !aux_data
[i
+ j
].seen
)
10413 err
= verifier_remove_insns(env
, i
, j
);
10416 insn_cnt
= env
->prog
->len
;
10422 static int opt_remove_nops(struct bpf_verifier_env
*env
)
10424 const struct bpf_insn ja
= BPF_JMP_IMM(BPF_JA
, 0, 0, 0);
10425 struct bpf_insn
*insn
= env
->prog
->insnsi
;
10426 int insn_cnt
= env
->prog
->len
;
10429 for (i
= 0; i
< insn_cnt
; i
++) {
10430 if (memcmp(&insn
[i
], &ja
, sizeof(ja
)))
10433 err
= verifier_remove_insns(env
, i
, 1);
10443 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env
*env
,
10444 const union bpf_attr
*attr
)
10446 struct bpf_insn
*patch
, zext_patch
[2], rnd_hi32_patch
[4];
10447 struct bpf_insn_aux_data
*aux
= env
->insn_aux_data
;
10448 int i
, patch_len
, delta
= 0, len
= env
->prog
->len
;
10449 struct bpf_insn
*insns
= env
->prog
->insnsi
;
10450 struct bpf_prog
*new_prog
;
10453 rnd_hi32
= attr
->prog_flags
& BPF_F_TEST_RND_HI32
;
10454 zext_patch
[1] = BPF_ZEXT_REG(0);
10455 rnd_hi32_patch
[1] = BPF_ALU64_IMM(BPF_MOV
, BPF_REG_AX
, 0);
10456 rnd_hi32_patch
[2] = BPF_ALU64_IMM(BPF_LSH
, BPF_REG_AX
, 32);
10457 rnd_hi32_patch
[3] = BPF_ALU64_REG(BPF_OR
, 0, BPF_REG_AX
);
10458 for (i
= 0; i
< len
; i
++) {
10459 int adj_idx
= i
+ delta
;
10460 struct bpf_insn insn
;
10462 insn
= insns
[adj_idx
];
10463 if (!aux
[adj_idx
].zext_dst
) {
10471 class = BPF_CLASS(code
);
10472 if (insn_no_def(&insn
))
10475 /* NOTE: arg "reg" (the fourth one) is only used for
10476 * BPF_STX which has been ruled out in above
10477 * check, it is safe to pass NULL here.
10479 if (is_reg64(env
, &insn
, insn
.dst_reg
, NULL
, DST_OP
)) {
10480 if (class == BPF_LD
&&
10481 BPF_MODE(code
) == BPF_IMM
)
10486 /* ctx load could be transformed into wider load. */
10487 if (class == BPF_LDX
&&
10488 aux
[adj_idx
].ptr_type
== PTR_TO_CTX
)
10491 imm_rnd
= get_random_int();
10492 rnd_hi32_patch
[0] = insn
;
10493 rnd_hi32_patch
[1].imm
= imm_rnd
;
10494 rnd_hi32_patch
[3].dst_reg
= insn
.dst_reg
;
10495 patch
= rnd_hi32_patch
;
10497 goto apply_patch_buffer
;
10500 if (!bpf_jit_needs_zext())
10503 zext_patch
[0] = insn
;
10504 zext_patch
[1].dst_reg
= insn
.dst_reg
;
10505 zext_patch
[1].src_reg
= insn
.dst_reg
;
10506 patch
= zext_patch
;
10508 apply_patch_buffer
:
10509 new_prog
= bpf_patch_insn_data(env
, adj_idx
, patch
, patch_len
);
10512 env
->prog
= new_prog
;
10513 insns
= new_prog
->insnsi
;
10514 aux
= env
->insn_aux_data
;
10515 delta
+= patch_len
- 1;
10521 /* convert load instructions that access fields of a context type into a
10522 * sequence of instructions that access fields of the underlying structure:
10523 * struct __sk_buff -> struct sk_buff
10524 * struct bpf_sock_ops -> struct sock
10526 static int convert_ctx_accesses(struct bpf_verifier_env
*env
)
10528 const struct bpf_verifier_ops
*ops
= env
->ops
;
10529 int i
, cnt
, size
, ctx_field_size
, delta
= 0;
10530 const int insn_cnt
= env
->prog
->len
;
10531 struct bpf_insn insn_buf
[16], *insn
;
10532 u32 target_size
, size_default
, off
;
10533 struct bpf_prog
*new_prog
;
10534 enum bpf_access_type type
;
10535 bool is_narrower_load
;
10537 if (ops
->gen_prologue
|| env
->seen_direct_write
) {
10538 if (!ops
->gen_prologue
) {
10539 verbose(env
, "bpf verifier is misconfigured\n");
10542 cnt
= ops
->gen_prologue(insn_buf
, env
->seen_direct_write
,
10544 if (cnt
>= ARRAY_SIZE(insn_buf
)) {
10545 verbose(env
, "bpf verifier is misconfigured\n");
10548 new_prog
= bpf_patch_insn_data(env
, 0, insn_buf
, cnt
);
10552 env
->prog
= new_prog
;
10557 if (bpf_prog_is_dev_bound(env
->prog
->aux
))
10560 insn
= env
->prog
->insnsi
+ delta
;
10562 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
10563 bpf_convert_ctx_access_t convert_ctx_access
;
10565 if (insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_B
) ||
10566 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_H
) ||
10567 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_W
) ||
10568 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_DW
))
10570 else if (insn
->code
== (BPF_STX
| BPF_MEM
| BPF_B
) ||
10571 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_H
) ||
10572 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_W
) ||
10573 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_DW
))
10578 if (type
== BPF_WRITE
&&
10579 env
->insn_aux_data
[i
+ delta
].sanitize_stack_off
) {
10580 struct bpf_insn patch
[] = {
10581 /* Sanitize suspicious stack slot with zero.
10582 * There are no memory dependencies for this store,
10583 * since it's only using frame pointer and immediate
10586 BPF_ST_MEM(BPF_DW
, BPF_REG_FP
,
10587 env
->insn_aux_data
[i
+ delta
].sanitize_stack_off
,
10589 /* the original STX instruction will immediately
10590 * overwrite the same stack slot with appropriate value
10595 cnt
= ARRAY_SIZE(patch
);
10596 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, patch
, cnt
);
10601 env
->prog
= new_prog
;
10602 insn
= new_prog
->insnsi
+ i
+ delta
;
10606 switch (env
->insn_aux_data
[i
+ delta
].ptr_type
) {
10608 if (!ops
->convert_ctx_access
)
10610 convert_ctx_access
= ops
->convert_ctx_access
;
10612 case PTR_TO_SOCKET
:
10613 case PTR_TO_SOCK_COMMON
:
10614 convert_ctx_access
= bpf_sock_convert_ctx_access
;
10616 case PTR_TO_TCP_SOCK
:
10617 convert_ctx_access
= bpf_tcp_sock_convert_ctx_access
;
10619 case PTR_TO_XDP_SOCK
:
10620 convert_ctx_access
= bpf_xdp_sock_convert_ctx_access
;
10622 case PTR_TO_BTF_ID
:
10623 if (type
== BPF_READ
) {
10624 insn
->code
= BPF_LDX
| BPF_PROBE_MEM
|
10625 BPF_SIZE((insn
)->code
);
10626 env
->prog
->aux
->num_exentries
++;
10627 } else if (resolve_prog_type(env
->prog
) != BPF_PROG_TYPE_STRUCT_OPS
) {
10628 verbose(env
, "Writes through BTF pointers are not allowed\n");
10636 ctx_field_size
= env
->insn_aux_data
[i
+ delta
].ctx_field_size
;
10637 size
= BPF_LDST_BYTES(insn
);
10639 /* If the read access is a narrower load of the field,
10640 * convert to a 4/8-byte load, to minimum program type specific
10641 * convert_ctx_access changes. If conversion is successful,
10642 * we will apply proper mask to the result.
10644 is_narrower_load
= size
< ctx_field_size
;
10645 size_default
= bpf_ctx_off_adjust_machine(ctx_field_size
);
10647 if (is_narrower_load
) {
10650 if (type
== BPF_WRITE
) {
10651 verbose(env
, "bpf verifier narrow ctx access misconfigured\n");
10656 if (ctx_field_size
== 4)
10658 else if (ctx_field_size
== 8)
10659 size_code
= BPF_DW
;
10661 insn
->off
= off
& ~(size_default
- 1);
10662 insn
->code
= BPF_LDX
| BPF_MEM
| size_code
;
10666 cnt
= convert_ctx_access(type
, insn
, insn_buf
, env
->prog
,
10668 if (cnt
== 0 || cnt
>= ARRAY_SIZE(insn_buf
) ||
10669 (ctx_field_size
&& !target_size
)) {
10670 verbose(env
, "bpf verifier is misconfigured\n");
10674 if (is_narrower_load
&& size
< target_size
) {
10675 u8 shift
= bpf_ctx_narrow_access_offset(
10676 off
, size
, size_default
) * 8;
10677 if (ctx_field_size
<= 4) {
10679 insn_buf
[cnt
++] = BPF_ALU32_IMM(BPF_RSH
,
10682 insn_buf
[cnt
++] = BPF_ALU32_IMM(BPF_AND
, insn
->dst_reg
,
10683 (1 << size
* 8) - 1);
10686 insn_buf
[cnt
++] = BPF_ALU64_IMM(BPF_RSH
,
10689 insn_buf
[cnt
++] = BPF_ALU64_IMM(BPF_AND
, insn
->dst_reg
,
10690 (1ULL << size
* 8) - 1);
10694 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
, cnt
);
10700 /* keep walking new program and skip insns we just inserted */
10701 env
->prog
= new_prog
;
10702 insn
= new_prog
->insnsi
+ i
+ delta
;
10708 static int jit_subprogs(struct bpf_verifier_env
*env
)
10710 struct bpf_prog
*prog
= env
->prog
, **func
, *tmp
;
10711 int i
, j
, subprog_start
, subprog_end
= 0, len
, subprog
;
10712 struct bpf_map
*map_ptr
;
10713 struct bpf_insn
*insn
;
10714 void *old_bpf_func
;
10715 int err
, num_exentries
;
10717 if (env
->subprog_cnt
<= 1)
10720 for (i
= 0, insn
= prog
->insnsi
; i
< prog
->len
; i
++, insn
++) {
10721 if (insn
->code
!= (BPF_JMP
| BPF_CALL
) ||
10722 insn
->src_reg
!= BPF_PSEUDO_CALL
)
10724 /* Upon error here we cannot fall back to interpreter but
10725 * need a hard reject of the program. Thus -EFAULT is
10726 * propagated in any case.
10728 subprog
= find_subprog(env
, i
+ insn
->imm
+ 1);
10730 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
10731 i
+ insn
->imm
+ 1);
10734 /* temporarily remember subprog id inside insn instead of
10735 * aux_data, since next loop will split up all insns into funcs
10737 insn
->off
= subprog
;
10738 /* remember original imm in case JIT fails and fallback
10739 * to interpreter will be needed
10741 env
->insn_aux_data
[i
].call_imm
= insn
->imm
;
10742 /* point imm to __bpf_call_base+1 from JITs point of view */
10746 err
= bpf_prog_alloc_jited_linfo(prog
);
10748 goto out_undo_insn
;
10751 func
= kcalloc(env
->subprog_cnt
, sizeof(prog
), GFP_KERNEL
);
10753 goto out_undo_insn
;
10755 for (i
= 0; i
< env
->subprog_cnt
; i
++) {
10756 subprog_start
= subprog_end
;
10757 subprog_end
= env
->subprog_info
[i
+ 1].start
;
10759 len
= subprog_end
- subprog_start
;
10760 /* BPF_PROG_RUN doesn't call subprogs directly,
10761 * hence main prog stats include the runtime of subprogs.
10762 * subprogs don't have IDs and not reachable via prog_get_next_id
10763 * func[i]->aux->stats will never be accessed and stays NULL
10765 func
[i
] = bpf_prog_alloc_no_stats(bpf_prog_size(len
), GFP_USER
);
10768 memcpy(func
[i
]->insnsi
, &prog
->insnsi
[subprog_start
],
10769 len
* sizeof(struct bpf_insn
));
10770 func
[i
]->type
= prog
->type
;
10771 func
[i
]->len
= len
;
10772 if (bpf_prog_calc_tag(func
[i
]))
10774 func
[i
]->is_func
= 1;
10775 func
[i
]->aux
->func_idx
= i
;
10776 /* the btf and func_info will be freed only at prog->aux */
10777 func
[i
]->aux
->btf
= prog
->aux
->btf
;
10778 func
[i
]->aux
->func_info
= prog
->aux
->func_info
;
10780 for (j
= 0; j
< prog
->aux
->size_poke_tab
; j
++) {
10781 u32 insn_idx
= prog
->aux
->poke_tab
[j
].insn_idx
;
10784 if (!(insn_idx
>= subprog_start
&&
10785 insn_idx
<= subprog_end
))
10788 ret
= bpf_jit_add_poke_descriptor(func
[i
],
10789 &prog
->aux
->poke_tab
[j
]);
10791 verbose(env
, "adding tail call poke descriptor failed\n");
10795 func
[i
]->insnsi
[insn_idx
- subprog_start
].imm
= ret
+ 1;
10797 map_ptr
= func
[i
]->aux
->poke_tab
[ret
].tail_call
.map
;
10798 ret
= map_ptr
->ops
->map_poke_track(map_ptr
, func
[i
]->aux
);
10800 verbose(env
, "tracking tail call prog failed\n");
10805 /* Use bpf_prog_F_tag to indicate functions in stack traces.
10806 * Long term would need debug info to populate names
10808 func
[i
]->aux
->name
[0] = 'F';
10809 func
[i
]->aux
->stack_depth
= env
->subprog_info
[i
].stack_depth
;
10810 func
[i
]->jit_requested
= 1;
10811 func
[i
]->aux
->linfo
= prog
->aux
->linfo
;
10812 func
[i
]->aux
->nr_linfo
= prog
->aux
->nr_linfo
;
10813 func
[i
]->aux
->jited_linfo
= prog
->aux
->jited_linfo
;
10814 func
[i
]->aux
->linfo_idx
= env
->subprog_info
[i
].linfo_idx
;
10816 insn
= func
[i
]->insnsi
;
10817 for (j
= 0; j
< func
[i
]->len
; j
++, insn
++) {
10818 if (BPF_CLASS(insn
->code
) == BPF_LDX
&&
10819 BPF_MODE(insn
->code
) == BPF_PROBE_MEM
)
10822 func
[i
]->aux
->num_exentries
= num_exentries
;
10823 func
[i
]->aux
->tail_call_reachable
= env
->subprog_info
[i
].tail_call_reachable
;
10824 func
[i
] = bpf_int_jit_compile(func
[i
]);
10825 if (!func
[i
]->jited
) {
10832 /* Untrack main program's aux structs so that during map_poke_run()
10833 * we will not stumble upon the unfilled poke descriptors; each
10834 * of the main program's poke descs got distributed across subprogs
10835 * and got tracked onto map, so we are sure that none of them will
10836 * be missed after the operation below
10838 for (i
= 0; i
< prog
->aux
->size_poke_tab
; i
++) {
10839 map_ptr
= prog
->aux
->poke_tab
[i
].tail_call
.map
;
10841 map_ptr
->ops
->map_poke_untrack(map_ptr
, prog
->aux
);
10844 /* at this point all bpf functions were successfully JITed
10845 * now populate all bpf_calls with correct addresses and
10846 * run last pass of JIT
10848 for (i
= 0; i
< env
->subprog_cnt
; i
++) {
10849 insn
= func
[i
]->insnsi
;
10850 for (j
= 0; j
< func
[i
]->len
; j
++, insn
++) {
10851 if (insn
->code
!= (BPF_JMP
| BPF_CALL
) ||
10852 insn
->src_reg
!= BPF_PSEUDO_CALL
)
10854 subprog
= insn
->off
;
10855 insn
->imm
= BPF_CAST_CALL(func
[subprog
]->bpf_func
) -
10859 /* we use the aux data to keep a list of the start addresses
10860 * of the JITed images for each function in the program
10862 * for some architectures, such as powerpc64, the imm field
10863 * might not be large enough to hold the offset of the start
10864 * address of the callee's JITed image from __bpf_call_base
10866 * in such cases, we can lookup the start address of a callee
10867 * by using its subprog id, available from the off field of
10868 * the call instruction, as an index for this list
10870 func
[i
]->aux
->func
= func
;
10871 func
[i
]->aux
->func_cnt
= env
->subprog_cnt
;
10873 for (i
= 0; i
< env
->subprog_cnt
; i
++) {
10874 old_bpf_func
= func
[i
]->bpf_func
;
10875 tmp
= bpf_int_jit_compile(func
[i
]);
10876 if (tmp
!= func
[i
] || func
[i
]->bpf_func
!= old_bpf_func
) {
10877 verbose(env
, "JIT doesn't support bpf-to-bpf calls\n");
10884 /* finally lock prog and jit images for all functions and
10885 * populate kallsysm
10887 for (i
= 0; i
< env
->subprog_cnt
; i
++) {
10888 bpf_prog_lock_ro(func
[i
]);
10889 bpf_prog_kallsyms_add(func
[i
]);
10892 /* Last step: make now unused interpreter insns from main
10893 * prog consistent for later dump requests, so they can
10894 * later look the same as if they were interpreted only.
10896 for (i
= 0, insn
= prog
->insnsi
; i
< prog
->len
; i
++, insn
++) {
10897 if (insn
->code
!= (BPF_JMP
| BPF_CALL
) ||
10898 insn
->src_reg
!= BPF_PSEUDO_CALL
)
10900 insn
->off
= env
->insn_aux_data
[i
].call_imm
;
10901 subprog
= find_subprog(env
, i
+ insn
->off
+ 1);
10902 insn
->imm
= subprog
;
10906 prog
->bpf_func
= func
[0]->bpf_func
;
10907 prog
->aux
->func
= func
;
10908 prog
->aux
->func_cnt
= env
->subprog_cnt
;
10909 bpf_prog_free_unused_jited_linfo(prog
);
10912 for (i
= 0; i
< env
->subprog_cnt
; i
++) {
10916 for (j
= 0; j
< func
[i
]->aux
->size_poke_tab
; j
++) {
10917 map_ptr
= func
[i
]->aux
->poke_tab
[j
].tail_call
.map
;
10918 map_ptr
->ops
->map_poke_untrack(map_ptr
, func
[i
]->aux
);
10920 bpf_jit_free(func
[i
]);
10924 /* cleanup main prog to be interpreted */
10925 prog
->jit_requested
= 0;
10926 for (i
= 0, insn
= prog
->insnsi
; i
< prog
->len
; i
++, insn
++) {
10927 if (insn
->code
!= (BPF_JMP
| BPF_CALL
) ||
10928 insn
->src_reg
!= BPF_PSEUDO_CALL
)
10931 insn
->imm
= env
->insn_aux_data
[i
].call_imm
;
10933 bpf_prog_free_jited_linfo(prog
);
10937 static int fixup_call_args(struct bpf_verifier_env
*env
)
10939 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
10940 struct bpf_prog
*prog
= env
->prog
;
10941 struct bpf_insn
*insn
= prog
->insnsi
;
10946 if (env
->prog
->jit_requested
&&
10947 !bpf_prog_is_dev_bound(env
->prog
->aux
)) {
10948 err
= jit_subprogs(env
);
10951 if (err
== -EFAULT
)
10954 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
10955 if (env
->subprog_cnt
> 1 && env
->prog
->aux
->tail_call_reachable
) {
10956 /* When JIT fails the progs with bpf2bpf calls and tail_calls
10957 * have to be rejected, since interpreter doesn't support them yet.
10959 verbose(env
, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
10962 for (i
= 0; i
< prog
->len
; i
++, insn
++) {
10963 if (insn
->code
!= (BPF_JMP
| BPF_CALL
) ||
10964 insn
->src_reg
!= BPF_PSEUDO_CALL
)
10966 depth
= get_callee_stack_depth(env
, insn
, i
);
10969 bpf_patch_call_args(insn
, depth
);
10976 /* fixup insn->imm field of bpf_call instructions
10977 * and inline eligible helpers as explicit sequence of BPF instructions
10979 * this function is called after eBPF program passed verification
10981 static int fixup_bpf_calls(struct bpf_verifier_env
*env
)
10983 struct bpf_prog
*prog
= env
->prog
;
10984 bool expect_blinding
= bpf_jit_blinding_enabled(prog
);
10985 struct bpf_insn
*insn
= prog
->insnsi
;
10986 const struct bpf_func_proto
*fn
;
10987 const int insn_cnt
= prog
->len
;
10988 const struct bpf_map_ops
*ops
;
10989 struct bpf_insn_aux_data
*aux
;
10990 struct bpf_insn insn_buf
[16];
10991 struct bpf_prog
*new_prog
;
10992 struct bpf_map
*map_ptr
;
10993 int i
, ret
, cnt
, delta
= 0;
10995 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
10996 if (insn
->code
== (BPF_ALU64
| BPF_MOD
| BPF_X
) ||
10997 insn
->code
== (BPF_ALU64
| BPF_DIV
| BPF_X
) ||
10998 insn
->code
== (BPF_ALU
| BPF_MOD
| BPF_X
) ||
10999 insn
->code
== (BPF_ALU
| BPF_DIV
| BPF_X
)) {
11000 bool is64
= BPF_CLASS(insn
->code
) == BPF_ALU64
;
11001 bool isdiv
= BPF_OP(insn
->code
) == BPF_DIV
;
11002 struct bpf_insn
*patchlet
;
11003 struct bpf_insn chk_and_div
[] = {
11004 /* [R,W]x div 0 -> 0 */
11005 BPF_RAW_INSN((is64
? BPF_JMP
: BPF_JMP32
) |
11006 BPF_JNE
| BPF_K
, insn
->src_reg
,
11008 BPF_ALU32_REG(BPF_XOR
, insn
->dst_reg
, insn
->dst_reg
),
11009 BPF_JMP_IMM(BPF_JA
, 0, 0, 1),
11012 struct bpf_insn chk_and_mod
[] = {
11013 /* [R,W]x mod 0 -> [R,W]x */
11014 BPF_RAW_INSN((is64
? BPF_JMP
: BPF_JMP32
) |
11015 BPF_JEQ
| BPF_K
, insn
->src_reg
,
11016 0, 1 + (is64
? 0 : 1), 0),
11018 BPF_JMP_IMM(BPF_JA
, 0, 0, 1),
11019 BPF_MOV32_REG(insn
->dst_reg
, insn
->dst_reg
),
11022 patchlet
= isdiv
? chk_and_div
: chk_and_mod
;
11023 cnt
= isdiv
? ARRAY_SIZE(chk_and_div
) :
11024 ARRAY_SIZE(chk_and_mod
) - (is64
? 2 : 0);
11026 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, patchlet
, cnt
);
11031 env
->prog
= prog
= new_prog
;
11032 insn
= new_prog
->insnsi
+ i
+ delta
;
11036 if (BPF_CLASS(insn
->code
) == BPF_LD
&&
11037 (BPF_MODE(insn
->code
) == BPF_ABS
||
11038 BPF_MODE(insn
->code
) == BPF_IND
)) {
11039 cnt
= env
->ops
->gen_ld_abs(insn
, insn_buf
);
11040 if (cnt
== 0 || cnt
>= ARRAY_SIZE(insn_buf
)) {
11041 verbose(env
, "bpf verifier is misconfigured\n");
11045 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
, cnt
);
11050 env
->prog
= prog
= new_prog
;
11051 insn
= new_prog
->insnsi
+ i
+ delta
;
11055 if (insn
->code
== (BPF_ALU64
| BPF_ADD
| BPF_X
) ||
11056 insn
->code
== (BPF_ALU64
| BPF_SUB
| BPF_X
)) {
11057 const u8 code_add
= BPF_ALU64
| BPF_ADD
| BPF_X
;
11058 const u8 code_sub
= BPF_ALU64
| BPF_SUB
| BPF_X
;
11059 struct bpf_insn insn_buf
[16];
11060 struct bpf_insn
*patch
= &insn_buf
[0];
11064 aux
= &env
->insn_aux_data
[i
+ delta
];
11065 if (!aux
->alu_state
||
11066 aux
->alu_state
== BPF_ALU_NON_POINTER
)
11069 isneg
= aux
->alu_state
& BPF_ALU_NEG_VALUE
;
11070 issrc
= (aux
->alu_state
& BPF_ALU_SANITIZE
) ==
11071 BPF_ALU_SANITIZE_SRC
;
11073 off_reg
= issrc
? insn
->src_reg
: insn
->dst_reg
;
11075 *patch
++ = BPF_ALU64_IMM(BPF_MUL
, off_reg
, -1);
11076 *patch
++ = BPF_MOV32_IMM(BPF_REG_AX
, aux
->alu_limit
);
11077 *patch
++ = BPF_ALU64_REG(BPF_SUB
, BPF_REG_AX
, off_reg
);
11078 *patch
++ = BPF_ALU64_REG(BPF_OR
, BPF_REG_AX
, off_reg
);
11079 *patch
++ = BPF_ALU64_IMM(BPF_NEG
, BPF_REG_AX
, 0);
11080 *patch
++ = BPF_ALU64_IMM(BPF_ARSH
, BPF_REG_AX
, 63);
11082 *patch
++ = BPF_ALU64_REG(BPF_AND
, BPF_REG_AX
,
11084 insn
->src_reg
= BPF_REG_AX
;
11086 *patch
++ = BPF_ALU64_REG(BPF_AND
, off_reg
,
11090 insn
->code
= insn
->code
== code_add
?
11091 code_sub
: code_add
;
11093 if (issrc
&& isneg
)
11094 *patch
++ = BPF_ALU64_IMM(BPF_MUL
, off_reg
, -1);
11095 cnt
= patch
- insn_buf
;
11097 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
, cnt
);
11102 env
->prog
= prog
= new_prog
;
11103 insn
= new_prog
->insnsi
+ i
+ delta
;
11107 if (insn
->code
!= (BPF_JMP
| BPF_CALL
))
11109 if (insn
->src_reg
== BPF_PSEUDO_CALL
)
11112 if (insn
->imm
== BPF_FUNC_get_route_realm
)
11113 prog
->dst_needed
= 1;
11114 if (insn
->imm
== BPF_FUNC_get_prandom_u32
)
11115 bpf_user_rnd_init_once();
11116 if (insn
->imm
== BPF_FUNC_override_return
)
11117 prog
->kprobe_override
= 1;
11118 if (insn
->imm
== BPF_FUNC_tail_call
) {
11119 /* If we tail call into other programs, we
11120 * cannot make any assumptions since they can
11121 * be replaced dynamically during runtime in
11122 * the program array.
11124 prog
->cb_access
= 1;
11125 if (!allow_tail_call_in_subprogs(env
))
11126 prog
->aux
->stack_depth
= MAX_BPF_STACK
;
11127 prog
->aux
->max_pkt_offset
= MAX_PACKET_OFF
;
11129 /* mark bpf_tail_call as different opcode to avoid
11130 * conditional branch in the interpeter for every normal
11131 * call and to prevent accidental JITing by JIT compiler
11132 * that doesn't support bpf_tail_call yet
11135 insn
->code
= BPF_JMP
| BPF_TAIL_CALL
;
11137 aux
= &env
->insn_aux_data
[i
+ delta
];
11138 if (env
->bpf_capable
&& !expect_blinding
&&
11139 prog
->jit_requested
&&
11140 !bpf_map_key_poisoned(aux
) &&
11141 !bpf_map_ptr_poisoned(aux
) &&
11142 !bpf_map_ptr_unpriv(aux
)) {
11143 struct bpf_jit_poke_descriptor desc
= {
11144 .reason
= BPF_POKE_REASON_TAIL_CALL
,
11145 .tail_call
.map
= BPF_MAP_PTR(aux
->map_ptr_state
),
11146 .tail_call
.key
= bpf_map_key_immediate(aux
),
11147 .insn_idx
= i
+ delta
,
11150 ret
= bpf_jit_add_poke_descriptor(prog
, &desc
);
11152 verbose(env
, "adding tail call poke descriptor failed\n");
11156 insn
->imm
= ret
+ 1;
11160 if (!bpf_map_ptr_unpriv(aux
))
11163 /* instead of changing every JIT dealing with tail_call
11164 * emit two extra insns:
11165 * if (index >= max_entries) goto out;
11166 * index &= array->index_mask;
11167 * to avoid out-of-bounds cpu speculation
11169 if (bpf_map_ptr_poisoned(aux
)) {
11170 verbose(env
, "tail_call abusing map_ptr\n");
11174 map_ptr
= BPF_MAP_PTR(aux
->map_ptr_state
);
11175 insn_buf
[0] = BPF_JMP_IMM(BPF_JGE
, BPF_REG_3
,
11176 map_ptr
->max_entries
, 2);
11177 insn_buf
[1] = BPF_ALU32_IMM(BPF_AND
, BPF_REG_3
,
11178 container_of(map_ptr
,
11181 insn_buf
[2] = *insn
;
11183 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
, cnt
);
11188 env
->prog
= prog
= new_prog
;
11189 insn
= new_prog
->insnsi
+ i
+ delta
;
11193 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
11194 * and other inlining handlers are currently limited to 64 bit
11197 if (prog
->jit_requested
&& BITS_PER_LONG
== 64 &&
11198 (insn
->imm
== BPF_FUNC_map_lookup_elem
||
11199 insn
->imm
== BPF_FUNC_map_update_elem
||
11200 insn
->imm
== BPF_FUNC_map_delete_elem
||
11201 insn
->imm
== BPF_FUNC_map_push_elem
||
11202 insn
->imm
== BPF_FUNC_map_pop_elem
||
11203 insn
->imm
== BPF_FUNC_map_peek_elem
)) {
11204 aux
= &env
->insn_aux_data
[i
+ delta
];
11205 if (bpf_map_ptr_poisoned(aux
))
11206 goto patch_call_imm
;
11208 map_ptr
= BPF_MAP_PTR(aux
->map_ptr_state
);
11209 ops
= map_ptr
->ops
;
11210 if (insn
->imm
== BPF_FUNC_map_lookup_elem
&&
11211 ops
->map_gen_lookup
) {
11212 cnt
= ops
->map_gen_lookup(map_ptr
, insn_buf
);
11213 if (cnt
== -EOPNOTSUPP
)
11214 goto patch_map_ops_generic
;
11215 if (cnt
<= 0 || cnt
>= ARRAY_SIZE(insn_buf
)) {
11216 verbose(env
, "bpf verifier is misconfigured\n");
11220 new_prog
= bpf_patch_insn_data(env
, i
+ delta
,
11226 env
->prog
= prog
= new_prog
;
11227 insn
= new_prog
->insnsi
+ i
+ delta
;
11231 BUILD_BUG_ON(!__same_type(ops
->map_lookup_elem
,
11232 (void *(*)(struct bpf_map
*map
, void *key
))NULL
));
11233 BUILD_BUG_ON(!__same_type(ops
->map_delete_elem
,
11234 (int (*)(struct bpf_map
*map
, void *key
))NULL
));
11235 BUILD_BUG_ON(!__same_type(ops
->map_update_elem
,
11236 (int (*)(struct bpf_map
*map
, void *key
, void *value
,
11238 BUILD_BUG_ON(!__same_type(ops
->map_push_elem
,
11239 (int (*)(struct bpf_map
*map
, void *value
,
11241 BUILD_BUG_ON(!__same_type(ops
->map_pop_elem
,
11242 (int (*)(struct bpf_map
*map
, void *value
))NULL
));
11243 BUILD_BUG_ON(!__same_type(ops
->map_peek_elem
,
11244 (int (*)(struct bpf_map
*map
, void *value
))NULL
));
11245 patch_map_ops_generic
:
11246 switch (insn
->imm
) {
11247 case BPF_FUNC_map_lookup_elem
:
11248 insn
->imm
= BPF_CAST_CALL(ops
->map_lookup_elem
) -
11251 case BPF_FUNC_map_update_elem
:
11252 insn
->imm
= BPF_CAST_CALL(ops
->map_update_elem
) -
11255 case BPF_FUNC_map_delete_elem
:
11256 insn
->imm
= BPF_CAST_CALL(ops
->map_delete_elem
) -
11259 case BPF_FUNC_map_push_elem
:
11260 insn
->imm
= BPF_CAST_CALL(ops
->map_push_elem
) -
11263 case BPF_FUNC_map_pop_elem
:
11264 insn
->imm
= BPF_CAST_CALL(ops
->map_pop_elem
) -
11267 case BPF_FUNC_map_peek_elem
:
11268 insn
->imm
= BPF_CAST_CALL(ops
->map_peek_elem
) -
11273 goto patch_call_imm
;
11276 if (prog
->jit_requested
&& BITS_PER_LONG
== 64 &&
11277 insn
->imm
== BPF_FUNC_jiffies64
) {
11278 struct bpf_insn ld_jiffies_addr
[2] = {
11279 BPF_LD_IMM64(BPF_REG_0
,
11280 (unsigned long)&jiffies
),
11283 insn_buf
[0] = ld_jiffies_addr
[0];
11284 insn_buf
[1] = ld_jiffies_addr
[1];
11285 insn_buf
[2] = BPF_LDX_MEM(BPF_DW
, BPF_REG_0
,
11289 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
,
11295 env
->prog
= prog
= new_prog
;
11296 insn
= new_prog
->insnsi
+ i
+ delta
;
11301 fn
= env
->ops
->get_func_proto(insn
->imm
, env
->prog
);
11302 /* all functions that have prototype and verifier allowed
11303 * programs to call them, must be real in-kernel functions
11307 "kernel subsystem misconfigured func %s#%d\n",
11308 func_id_name(insn
->imm
), insn
->imm
);
11311 insn
->imm
= fn
->func
- __bpf_call_base
;
11314 /* Since poke tab is now finalized, publish aux to tracker. */
11315 for (i
= 0; i
< prog
->aux
->size_poke_tab
; i
++) {
11316 map_ptr
= prog
->aux
->poke_tab
[i
].tail_call
.map
;
11317 if (!map_ptr
->ops
->map_poke_track
||
11318 !map_ptr
->ops
->map_poke_untrack
||
11319 !map_ptr
->ops
->map_poke_run
) {
11320 verbose(env
, "bpf verifier is misconfigured\n");
11324 ret
= map_ptr
->ops
->map_poke_track(map_ptr
, prog
->aux
);
11326 verbose(env
, "tracking tail call prog failed\n");
11334 static void free_states(struct bpf_verifier_env
*env
)
11336 struct bpf_verifier_state_list
*sl
, *sln
;
11339 sl
= env
->free_list
;
11342 free_verifier_state(&sl
->state
, false);
11346 env
->free_list
= NULL
;
11348 if (!env
->explored_states
)
11351 for (i
= 0; i
< state_htab_size(env
); i
++) {
11352 sl
= env
->explored_states
[i
];
11356 free_verifier_state(&sl
->state
, false);
11360 env
->explored_states
[i
] = NULL
;
11364 /* The verifier is using insn_aux_data[] to store temporary data during
11365 * verification and to store information for passes that run after the
11366 * verification like dead code sanitization. do_check_common() for subprogram N
11367 * may analyze many other subprograms. sanitize_insn_aux_data() clears all
11368 * temporary data after do_check_common() finds that subprogram N cannot be
11369 * verified independently. pass_cnt counts the number of times
11370 * do_check_common() was run and insn->aux->seen tells the pass number
11371 * insn_aux_data was touched. These variables are compared to clear temporary
11372 * data from failed pass. For testing and experiments do_check_common() can be
11373 * run multiple times even when prior attempt to verify is unsuccessful.
11375 static void sanitize_insn_aux_data(struct bpf_verifier_env
*env
)
11377 struct bpf_insn
*insn
= env
->prog
->insnsi
;
11378 struct bpf_insn_aux_data
*aux
;
11381 for (i
= 0; i
< env
->prog
->len
; i
++) {
11382 class = BPF_CLASS(insn
[i
].code
);
11383 if (class != BPF_LDX
&& class != BPF_STX
)
11385 aux
= &env
->insn_aux_data
[i
];
11386 if (aux
->seen
!= env
->pass_cnt
)
11388 memset(aux
, 0, offsetof(typeof(*aux
), orig_idx
));
11392 static int do_check_common(struct bpf_verifier_env
*env
, int subprog
)
11394 bool pop_log
= !(env
->log
.level
& BPF_LOG_LEVEL2
);
11395 struct bpf_verifier_state
*state
;
11396 struct bpf_reg_state
*regs
;
11399 env
->prev_linfo
= NULL
;
11402 state
= kzalloc(sizeof(struct bpf_verifier_state
), GFP_KERNEL
);
11405 state
->curframe
= 0;
11406 state
->speculative
= false;
11407 state
->branches
= 1;
11408 state
->frame
[0] = kzalloc(sizeof(struct bpf_func_state
), GFP_KERNEL
);
11409 if (!state
->frame
[0]) {
11413 env
->cur_state
= state
;
11414 init_func_state(env
, state
->frame
[0],
11415 BPF_MAIN_FUNC
/* callsite */,
11419 regs
= state
->frame
[state
->curframe
]->regs
;
11420 if (subprog
|| env
->prog
->type
== BPF_PROG_TYPE_EXT
) {
11421 ret
= btf_prepare_func_args(env
, subprog
, regs
);
11424 for (i
= BPF_REG_1
; i
<= BPF_REG_5
; i
++) {
11425 if (regs
[i
].type
== PTR_TO_CTX
)
11426 mark_reg_known_zero(env
, regs
, i
);
11427 else if (regs
[i
].type
== SCALAR_VALUE
)
11428 mark_reg_unknown(env
, regs
, i
);
11431 /* 1st arg to a function */
11432 regs
[BPF_REG_1
].type
= PTR_TO_CTX
;
11433 mark_reg_known_zero(env
, regs
, BPF_REG_1
);
11434 ret
= btf_check_func_arg_match(env
, subprog
, regs
);
11435 if (ret
== -EFAULT
)
11436 /* unlikely verifier bug. abort.
11437 * ret == 0 and ret < 0 are sadly acceptable for
11438 * main() function due to backward compatibility.
11439 * Like socket filter program may be written as:
11440 * int bpf_prog(struct pt_regs *ctx)
11441 * and never dereference that ctx in the program.
11442 * 'struct pt_regs' is a type mismatch for socket
11443 * filter that should be using 'struct __sk_buff'.
11448 ret
= do_check(env
);
11450 /* check for NULL is necessary, since cur_state can be freed inside
11451 * do_check() under memory pressure.
11453 if (env
->cur_state
) {
11454 free_verifier_state(env
->cur_state
, true);
11455 env
->cur_state
= NULL
;
11457 while (!pop_stack(env
, NULL
, NULL
, false));
11458 if (!ret
&& pop_log
)
11459 bpf_vlog_reset(&env
->log
, 0);
11462 /* clean aux data in case subprog was rejected */
11463 sanitize_insn_aux_data(env
);
11467 /* Verify all global functions in a BPF program one by one based on their BTF.
11468 * All global functions must pass verification. Otherwise the whole program is rejected.
11479 * foo() will be verified first for R1=any_scalar_value. During verification it
11480 * will be assumed that bar() already verified successfully and call to bar()
11481 * from foo() will be checked for type match only. Later bar() will be verified
11482 * independently to check that it's safe for R1=any_scalar_value.
11484 static int do_check_subprogs(struct bpf_verifier_env
*env
)
11486 struct bpf_prog_aux
*aux
= env
->prog
->aux
;
11489 if (!aux
->func_info
)
11492 for (i
= 1; i
< env
->subprog_cnt
; i
++) {
11493 if (aux
->func_info_aux
[i
].linkage
!= BTF_FUNC_GLOBAL
)
11495 env
->insn_idx
= env
->subprog_info
[i
].start
;
11496 WARN_ON_ONCE(env
->insn_idx
== 0);
11497 ret
= do_check_common(env
, i
);
11500 } else if (env
->log
.level
& BPF_LOG_LEVEL
) {
11502 "Func#%d is safe for any args that match its prototype\n",
11509 static int do_check_main(struct bpf_verifier_env
*env
)
11514 ret
= do_check_common(env
, 0);
11516 env
->prog
->aux
->stack_depth
= env
->subprog_info
[0].stack_depth
;
11521 static void print_verification_stats(struct bpf_verifier_env
*env
)
11525 if (env
->log
.level
& BPF_LOG_STATS
) {
11526 verbose(env
, "verification time %lld usec\n",
11527 div_u64(env
->verification_time
, 1000));
11528 verbose(env
, "stack depth ");
11529 for (i
= 0; i
< env
->subprog_cnt
; i
++) {
11530 u32 depth
= env
->subprog_info
[i
].stack_depth
;
11532 verbose(env
, "%d", depth
);
11533 if (i
+ 1 < env
->subprog_cnt
)
11536 verbose(env
, "\n");
11538 verbose(env
, "processed %d insns (limit %d) max_states_per_insn %d "
11539 "total_states %d peak_states %d mark_read %d\n",
11540 env
->insn_processed
, BPF_COMPLEXITY_LIMIT_INSNS
,
11541 env
->max_states_per_insn
, env
->total_states
,
11542 env
->peak_states
, env
->longest_mark_read_walk
);
11545 static int check_struct_ops_btf_id(struct bpf_verifier_env
*env
)
11547 const struct btf_type
*t
, *func_proto
;
11548 const struct bpf_struct_ops
*st_ops
;
11549 const struct btf_member
*member
;
11550 struct bpf_prog
*prog
= env
->prog
;
11551 u32 btf_id
, member_idx
;
11554 if (!prog
->gpl_compatible
) {
11555 verbose(env
, "struct ops programs must have a GPL compatible license\n");
11559 btf_id
= prog
->aux
->attach_btf_id
;
11560 st_ops
= bpf_struct_ops_find(btf_id
);
11562 verbose(env
, "attach_btf_id %u is not a supported struct\n",
11568 member_idx
= prog
->expected_attach_type
;
11569 if (member_idx
>= btf_type_vlen(t
)) {
11570 verbose(env
, "attach to invalid member idx %u of struct %s\n",
11571 member_idx
, st_ops
->name
);
11575 member
= &btf_type_member(t
)[member_idx
];
11576 mname
= btf_name_by_offset(btf_vmlinux
, member
->name_off
);
11577 func_proto
= btf_type_resolve_func_ptr(btf_vmlinux
, member
->type
,
11580 verbose(env
, "attach to invalid member %s(@idx %u) of struct %s\n",
11581 mname
, member_idx
, st_ops
->name
);
11585 if (st_ops
->check_member
) {
11586 int err
= st_ops
->check_member(t
, member
);
11589 verbose(env
, "attach to unsupported member %s of struct %s\n",
11590 mname
, st_ops
->name
);
11595 prog
->aux
->attach_func_proto
= func_proto
;
11596 prog
->aux
->attach_func_name
= mname
;
11597 env
->ops
= st_ops
->verifier_ops
;
11601 #define SECURITY_PREFIX "security_"
11603 static int check_attach_modify_return(unsigned long addr
, const char *func_name
)
11605 if (within_error_injection_list(addr
) ||
11606 !strncmp(SECURITY_PREFIX
, func_name
, sizeof(SECURITY_PREFIX
) - 1))
11612 /* list of non-sleepable functions that are otherwise on
11613 * ALLOW_ERROR_INJECTION list
11615 BTF_SET_START(btf_non_sleepable_error_inject
)
11616 /* Three functions below can be called from sleepable and non-sleepable context.
11617 * Assume non-sleepable from bpf safety point of view.
11619 BTF_ID(func
, __add_to_page_cache_locked
)
11620 BTF_ID(func
, should_fail_alloc_page
)
11621 BTF_ID(func
, should_failslab
)
11622 BTF_SET_END(btf_non_sleepable_error_inject
)
11624 static int check_non_sleepable_error_inject(u32 btf_id
)
11626 return btf_id_set_contains(&btf_non_sleepable_error_inject
, btf_id
);
11629 int bpf_check_attach_target(struct bpf_verifier_log
*log
,
11630 const struct bpf_prog
*prog
,
11631 const struct bpf_prog
*tgt_prog
,
11633 struct bpf_attach_target_info
*tgt_info
)
11635 bool prog_extension
= prog
->type
== BPF_PROG_TYPE_EXT
;
11636 const char prefix
[] = "btf_trace_";
11637 int ret
= 0, subprog
= -1, i
;
11638 const struct btf_type
*t
;
11639 bool conservative
= true;
11645 bpf_log(log
, "Tracing programs must provide btf_id\n");
11648 btf
= tgt_prog
? tgt_prog
->aux
->btf
: prog
->aux
->attach_btf
;
11651 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
11654 t
= btf_type_by_id(btf
, btf_id
);
11656 bpf_log(log
, "attach_btf_id %u is invalid\n", btf_id
);
11659 tname
= btf_name_by_offset(btf
, t
->name_off
);
11661 bpf_log(log
, "attach_btf_id %u doesn't have a name\n", btf_id
);
11665 struct bpf_prog_aux
*aux
= tgt_prog
->aux
;
11667 for (i
= 0; i
< aux
->func_info_cnt
; i
++)
11668 if (aux
->func_info
[i
].type_id
== btf_id
) {
11672 if (subprog
== -1) {
11673 bpf_log(log
, "Subprog %s doesn't exist\n", tname
);
11676 conservative
= aux
->func_info_aux
[subprog
].unreliable
;
11677 if (prog_extension
) {
11678 if (conservative
) {
11680 "Cannot replace static functions\n");
11683 if (!prog
->jit_requested
) {
11685 "Extension programs should be JITed\n");
11689 if (!tgt_prog
->jited
) {
11690 bpf_log(log
, "Can attach to only JITed progs\n");
11693 if (tgt_prog
->type
== prog
->type
) {
11694 /* Cannot fentry/fexit another fentry/fexit program.
11695 * Cannot attach program extension to another extension.
11696 * It's ok to attach fentry/fexit to extension program.
11698 bpf_log(log
, "Cannot recursively attach\n");
11701 if (tgt_prog
->type
== BPF_PROG_TYPE_TRACING
&&
11703 (tgt_prog
->expected_attach_type
== BPF_TRACE_FENTRY
||
11704 tgt_prog
->expected_attach_type
== BPF_TRACE_FEXIT
)) {
11705 /* Program extensions can extend all program types
11706 * except fentry/fexit. The reason is the following.
11707 * The fentry/fexit programs are used for performance
11708 * analysis, stats and can be attached to any program
11709 * type except themselves. When extension program is
11710 * replacing XDP function it is necessary to allow
11711 * performance analysis of all functions. Both original
11712 * XDP program and its program extension. Hence
11713 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
11714 * allowed. If extending of fentry/fexit was allowed it
11715 * would be possible to create long call chain
11716 * fentry->extension->fentry->extension beyond
11717 * reasonable stack size. Hence extending fentry is not
11720 bpf_log(log
, "Cannot extend fentry/fexit\n");
11724 if (prog_extension
) {
11725 bpf_log(log
, "Cannot replace kernel functions\n");
11730 switch (prog
->expected_attach_type
) {
11731 case BPF_TRACE_RAW_TP
:
11734 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
11737 if (!btf_type_is_typedef(t
)) {
11738 bpf_log(log
, "attach_btf_id %u is not a typedef\n",
11742 if (strncmp(prefix
, tname
, sizeof(prefix
) - 1)) {
11743 bpf_log(log
, "attach_btf_id %u points to wrong type name %s\n",
11747 tname
+= sizeof(prefix
) - 1;
11748 t
= btf_type_by_id(btf
, t
->type
);
11749 if (!btf_type_is_ptr(t
))
11750 /* should never happen in valid vmlinux build */
11752 t
= btf_type_by_id(btf
, t
->type
);
11753 if (!btf_type_is_func_proto(t
))
11754 /* should never happen in valid vmlinux build */
11758 case BPF_TRACE_ITER
:
11759 if (!btf_type_is_func(t
)) {
11760 bpf_log(log
, "attach_btf_id %u is not a function\n",
11764 t
= btf_type_by_id(btf
, t
->type
);
11765 if (!btf_type_is_func_proto(t
))
11767 ret
= btf_distill_func_proto(log
, btf
, t
, tname
, &tgt_info
->fmodel
);
11772 if (!prog_extension
)
11775 case BPF_MODIFY_RETURN
:
11777 case BPF_TRACE_FENTRY
:
11778 case BPF_TRACE_FEXIT
:
11779 if (!btf_type_is_func(t
)) {
11780 bpf_log(log
, "attach_btf_id %u is not a function\n",
11784 if (prog_extension
&&
11785 btf_check_type_match(log
, prog
, btf
, t
))
11787 t
= btf_type_by_id(btf
, t
->type
);
11788 if (!btf_type_is_func_proto(t
))
11791 if ((prog
->aux
->saved_dst_prog_type
|| prog
->aux
->saved_dst_attach_type
) &&
11792 (!tgt_prog
|| prog
->aux
->saved_dst_prog_type
!= tgt_prog
->type
||
11793 prog
->aux
->saved_dst_attach_type
!= tgt_prog
->expected_attach_type
))
11796 if (tgt_prog
&& conservative
)
11799 ret
= btf_distill_func_proto(log
, btf
, t
, tname
, &tgt_info
->fmodel
);
11805 addr
= (long) tgt_prog
->bpf_func
;
11807 addr
= (long) tgt_prog
->aux
->func
[subprog
]->bpf_func
;
11809 addr
= kallsyms_lookup_name(tname
);
11812 "The address of function %s cannot be found\n",
11818 if (prog
->aux
->sleepable
) {
11820 switch (prog
->type
) {
11821 case BPF_PROG_TYPE_TRACING
:
11822 /* fentry/fexit/fmod_ret progs can be sleepable only if they are
11823 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
11825 if (!check_non_sleepable_error_inject(btf_id
) &&
11826 within_error_injection_list(addr
))
11829 case BPF_PROG_TYPE_LSM
:
11830 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
11831 * Only some of them are sleepable.
11833 if (bpf_lsm_is_sleepable_hook(btf_id
))
11840 bpf_log(log
, "%s is not sleepable\n", tname
);
11843 } else if (prog
->expected_attach_type
== BPF_MODIFY_RETURN
) {
11845 bpf_log(log
, "can't modify return codes of BPF programs\n");
11848 ret
= check_attach_modify_return(addr
, tname
);
11850 bpf_log(log
, "%s() is not modifiable\n", tname
);
11857 tgt_info
->tgt_addr
= addr
;
11858 tgt_info
->tgt_name
= tname
;
11859 tgt_info
->tgt_type
= t
;
11863 static int check_attach_btf_id(struct bpf_verifier_env
*env
)
11865 struct bpf_prog
*prog
= env
->prog
;
11866 struct bpf_prog
*tgt_prog
= prog
->aux
->dst_prog
;
11867 struct bpf_attach_target_info tgt_info
= {};
11868 u32 btf_id
= prog
->aux
->attach_btf_id
;
11869 struct bpf_trampoline
*tr
;
11873 if (prog
->aux
->sleepable
&& prog
->type
!= BPF_PROG_TYPE_TRACING
&&
11874 prog
->type
!= BPF_PROG_TYPE_LSM
) {
11875 verbose(env
, "Only fentry/fexit/fmod_ret and lsm programs can be sleepable\n");
11879 if (prog
->type
== BPF_PROG_TYPE_STRUCT_OPS
)
11880 return check_struct_ops_btf_id(env
);
11882 if (prog
->type
!= BPF_PROG_TYPE_TRACING
&&
11883 prog
->type
!= BPF_PROG_TYPE_LSM
&&
11884 prog
->type
!= BPF_PROG_TYPE_EXT
)
11887 ret
= bpf_check_attach_target(&env
->log
, prog
, tgt_prog
, btf_id
, &tgt_info
);
11891 if (tgt_prog
&& prog
->type
== BPF_PROG_TYPE_EXT
) {
11892 /* to make freplace equivalent to their targets, they need to
11893 * inherit env->ops and expected_attach_type for the rest of the
11896 env
->ops
= bpf_verifier_ops
[tgt_prog
->type
];
11897 prog
->expected_attach_type
= tgt_prog
->expected_attach_type
;
11900 /* store info about the attachment target that will be used later */
11901 prog
->aux
->attach_func_proto
= tgt_info
.tgt_type
;
11902 prog
->aux
->attach_func_name
= tgt_info
.tgt_name
;
11905 prog
->aux
->saved_dst_prog_type
= tgt_prog
->type
;
11906 prog
->aux
->saved_dst_attach_type
= tgt_prog
->expected_attach_type
;
11909 if (prog
->expected_attach_type
== BPF_TRACE_RAW_TP
) {
11910 prog
->aux
->attach_btf_trace
= true;
11912 } else if (prog
->expected_attach_type
== BPF_TRACE_ITER
) {
11913 if (!bpf_iter_prog_supported(prog
))
11918 if (prog
->type
== BPF_PROG_TYPE_LSM
) {
11919 ret
= bpf_lsm_verify_prog(&env
->log
, prog
);
11924 key
= bpf_trampoline_compute_key(tgt_prog
, prog
->aux
->attach_btf
, btf_id
);
11925 tr
= bpf_trampoline_get(key
, &tgt_info
);
11929 prog
->aux
->dst_trampoline
= tr
;
11933 struct btf
*bpf_get_btf_vmlinux(void)
11935 if (!btf_vmlinux
&& IS_ENABLED(CONFIG_DEBUG_INFO_BTF
)) {
11936 mutex_lock(&bpf_verifier_lock
);
11938 btf_vmlinux
= btf_parse_vmlinux();
11939 mutex_unlock(&bpf_verifier_lock
);
11941 return btf_vmlinux
;
11944 int bpf_check(struct bpf_prog
**prog
, union bpf_attr
*attr
,
11945 union bpf_attr __user
*uattr
)
11947 u64 start_time
= ktime_get_ns();
11948 struct bpf_verifier_env
*env
;
11949 struct bpf_verifier_log
*log
;
11950 int i
, len
, ret
= -EINVAL
;
11953 /* no program is valid */
11954 if (ARRAY_SIZE(bpf_verifier_ops
) == 0)
11957 /* 'struct bpf_verifier_env' can be global, but since it's not small,
11958 * allocate/free it every time bpf_check() is called
11960 env
= kzalloc(sizeof(struct bpf_verifier_env
), GFP_KERNEL
);
11965 len
= (*prog
)->len
;
11966 env
->insn_aux_data
=
11967 vzalloc(array_size(sizeof(struct bpf_insn_aux_data
), len
));
11969 if (!env
->insn_aux_data
)
11971 for (i
= 0; i
< len
; i
++)
11972 env
->insn_aux_data
[i
].orig_idx
= i
;
11974 env
->ops
= bpf_verifier_ops
[env
->prog
->type
];
11975 is_priv
= bpf_capable();
11977 bpf_get_btf_vmlinux();
11979 /* grab the mutex to protect few globals used by verifier */
11981 mutex_lock(&bpf_verifier_lock
);
11983 if (attr
->log_level
|| attr
->log_buf
|| attr
->log_size
) {
11984 /* user requested verbose verifier output
11985 * and supplied buffer to store the verification trace
11987 log
->level
= attr
->log_level
;
11988 log
->ubuf
= (char __user
*) (unsigned long) attr
->log_buf
;
11989 log
->len_total
= attr
->log_size
;
11992 /* log attributes have to be sane */
11993 if (log
->len_total
< 128 || log
->len_total
> UINT_MAX
>> 2 ||
11994 !log
->level
|| !log
->ubuf
|| log
->level
& ~BPF_LOG_MASK
)
11998 if (IS_ERR(btf_vmlinux
)) {
11999 /* Either gcc or pahole or kernel are broken. */
12000 verbose(env
, "in-kernel BTF is malformed\n");
12001 ret
= PTR_ERR(btf_vmlinux
);
12002 goto skip_full_check
;
12005 env
->strict_alignment
= !!(attr
->prog_flags
& BPF_F_STRICT_ALIGNMENT
);
12006 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
))
12007 env
->strict_alignment
= true;
12008 if (attr
->prog_flags
& BPF_F_ANY_ALIGNMENT
)
12009 env
->strict_alignment
= false;
12011 env
->allow_ptr_leaks
= bpf_allow_ptr_leaks();
12012 env
->allow_ptr_to_map_access
= bpf_allow_ptr_to_map_access();
12013 env
->bypass_spec_v1
= bpf_bypass_spec_v1();
12014 env
->bypass_spec_v4
= bpf_bypass_spec_v4();
12015 env
->bpf_capable
= bpf_capable();
12018 env
->test_state_freq
= attr
->prog_flags
& BPF_F_TEST_STATE_FREQ
;
12020 if (bpf_prog_is_dev_bound(env
->prog
->aux
)) {
12021 ret
= bpf_prog_offload_verifier_prep(env
->prog
);
12023 goto skip_full_check
;
12026 env
->explored_states
= kvcalloc(state_htab_size(env
),
12027 sizeof(struct bpf_verifier_state_list
*),
12030 if (!env
->explored_states
)
12031 goto skip_full_check
;
12033 ret
= check_subprogs(env
);
12035 goto skip_full_check
;
12037 ret
= check_btf_info(env
, attr
, uattr
);
12039 goto skip_full_check
;
12041 ret
= check_attach_btf_id(env
);
12043 goto skip_full_check
;
12045 ret
= resolve_pseudo_ldimm64(env
);
12047 goto skip_full_check
;
12049 ret
= check_cfg(env
);
12051 goto skip_full_check
;
12053 ret
= do_check_subprogs(env
);
12054 ret
= ret
?: do_check_main(env
);
12056 if (ret
== 0 && bpf_prog_is_dev_bound(env
->prog
->aux
))
12057 ret
= bpf_prog_offload_finalize(env
);
12060 kvfree(env
->explored_states
);
12063 ret
= check_max_stack_depth(env
);
12065 /* instruction rewrites happen after this point */
12068 opt_hard_wire_dead_code_branches(env
);
12070 ret
= opt_remove_dead_code(env
);
12072 ret
= opt_remove_nops(env
);
12075 sanitize_dead_code(env
);
12079 /* program is valid, convert *(u32*)(ctx + off) accesses */
12080 ret
= convert_ctx_accesses(env
);
12083 ret
= fixup_bpf_calls(env
);
12085 /* do 32-bit optimization after insn patching has done so those patched
12086 * insns could be handled correctly.
12088 if (ret
== 0 && !bpf_prog_is_dev_bound(env
->prog
->aux
)) {
12089 ret
= opt_subreg_zext_lo32_rnd_hi32(env
, attr
);
12090 env
->prog
->aux
->verifier_zext
= bpf_jit_needs_zext() ? !ret
12095 ret
= fixup_call_args(env
);
12097 env
->verification_time
= ktime_get_ns() - start_time
;
12098 print_verification_stats(env
);
12100 if (log
->level
&& bpf_verifier_log_full(log
))
12102 if (log
->level
&& !log
->ubuf
) {
12104 goto err_release_maps
;
12107 if (ret
== 0 && env
->used_map_cnt
) {
12108 /* if program passed verifier, update used_maps in bpf_prog_info */
12109 env
->prog
->aux
->used_maps
= kmalloc_array(env
->used_map_cnt
,
12110 sizeof(env
->used_maps
[0]),
12113 if (!env
->prog
->aux
->used_maps
) {
12115 goto err_release_maps
;
12118 memcpy(env
->prog
->aux
->used_maps
, env
->used_maps
,
12119 sizeof(env
->used_maps
[0]) * env
->used_map_cnt
);
12120 env
->prog
->aux
->used_map_cnt
= env
->used_map_cnt
;
12122 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
12123 * bpf_ld_imm64 instructions
12125 convert_pseudo_ld_imm64(env
);
12129 adjust_btf_func(env
);
12132 if (!env
->prog
->aux
->used_maps
)
12133 /* if we didn't copy map pointers into bpf_prog_info, release
12134 * them now. Otherwise free_used_maps() will release them.
12138 /* extension progs temporarily inherit the attach_type of their targets
12139 for verification purposes, so set it back to zero before returning
12141 if (env
->prog
->type
== BPF_PROG_TYPE_EXT
)
12142 env
->prog
->expected_attach_type
= 0;
12147 mutex_unlock(&bpf_verifier_lock
);
12148 vfree(env
->insn_aux_data
);