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}_fixed_off functions track spill/fill of registers,
2275 * stack boundary and alignment are checked in check_mem_access()
2277 static int check_stack_write_fixed_off(struct bpf_verifier_env
*env
,
2278 /* stack frame we're writing to */
2279 struct bpf_func_state
*state
,
2280 int off
, int size
, int value_regno
,
2283 struct bpf_func_state
*cur
; /* state of the current function */
2284 int i
, slot
= -off
- 1, spi
= slot
/ BPF_REG_SIZE
, err
;
2285 u32 dst_reg
= env
->prog
->insnsi
[insn_idx
].dst_reg
;
2286 struct bpf_reg_state
*reg
= NULL
;
2288 err
= realloc_func_state(state
, round_up(slot
+ 1, BPF_REG_SIZE
),
2289 state
->acquired_refs
, true);
2292 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
2293 * so it's aligned access and [off, off + size) are within stack limits
2295 if (!env
->allow_ptr_leaks
&&
2296 state
->stack
[spi
].slot_type
[0] == STACK_SPILL
&&
2297 size
!= BPF_REG_SIZE
) {
2298 verbose(env
, "attempt to corrupt spilled pointer on stack\n");
2302 cur
= env
->cur_state
->frame
[env
->cur_state
->curframe
];
2303 if (value_regno
>= 0)
2304 reg
= &cur
->regs
[value_regno
];
2306 if (reg
&& size
== BPF_REG_SIZE
&& register_is_bounded(reg
) &&
2307 !register_is_null(reg
) && env
->bpf_capable
) {
2308 if (dst_reg
!= BPF_REG_FP
) {
2309 /* The backtracking logic can only recognize explicit
2310 * stack slot address like [fp - 8]. Other spill of
2311 * scalar via different register has to be conervative.
2312 * Backtrack from here and mark all registers as precise
2313 * that contributed into 'reg' being a constant.
2315 err
= mark_chain_precision(env
, value_regno
);
2319 save_register_state(state
, spi
, reg
);
2320 } else if (reg
&& is_spillable_regtype(reg
->type
)) {
2321 /* register containing pointer is being spilled into stack */
2322 if (size
!= BPF_REG_SIZE
) {
2323 verbose_linfo(env
, insn_idx
, "; ");
2324 verbose(env
, "invalid size of register spill\n");
2328 if (state
!= cur
&& reg
->type
== PTR_TO_STACK
) {
2329 verbose(env
, "cannot spill pointers to stack into stack frame of the caller\n");
2333 if (!env
->bypass_spec_v4
) {
2334 bool sanitize
= false;
2336 if (state
->stack
[spi
].slot_type
[0] == STACK_SPILL
&&
2337 register_is_const(&state
->stack
[spi
].spilled_ptr
))
2339 for (i
= 0; i
< BPF_REG_SIZE
; i
++)
2340 if (state
->stack
[spi
].slot_type
[i
] == STACK_MISC
) {
2345 int *poff
= &env
->insn_aux_data
[insn_idx
].sanitize_stack_off
;
2346 int soff
= (-spi
- 1) * BPF_REG_SIZE
;
2348 /* detected reuse of integer stack slot with a pointer
2349 * which means either llvm is reusing stack slot or
2350 * an attacker is trying to exploit CVE-2018-3639
2351 * (speculative store bypass)
2352 * Have to sanitize that slot with preemptive
2355 if (*poff
&& *poff
!= soff
) {
2356 /* disallow programs where single insn stores
2357 * into two different stack slots, since verifier
2358 * cannot sanitize them
2361 "insn %d cannot access two stack slots fp%d and fp%d",
2362 insn_idx
, *poff
, soff
);
2368 save_register_state(state
, spi
, reg
);
2370 u8 type
= STACK_MISC
;
2372 /* regular write of data into stack destroys any spilled ptr */
2373 state
->stack
[spi
].spilled_ptr
.type
= NOT_INIT
;
2374 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
2375 if (state
->stack
[spi
].slot_type
[0] == STACK_SPILL
)
2376 for (i
= 0; i
< BPF_REG_SIZE
; i
++)
2377 state
->stack
[spi
].slot_type
[i
] = STACK_MISC
;
2379 /* only mark the slot as written if all 8 bytes were written
2380 * otherwise read propagation may incorrectly stop too soon
2381 * when stack slots are partially written.
2382 * This heuristic means that read propagation will be
2383 * conservative, since it will add reg_live_read marks
2384 * to stack slots all the way to first state when programs
2385 * writes+reads less than 8 bytes
2387 if (size
== BPF_REG_SIZE
)
2388 state
->stack
[spi
].spilled_ptr
.live
|= REG_LIVE_WRITTEN
;
2390 /* when we zero initialize stack slots mark them as such */
2391 if (reg
&& register_is_null(reg
)) {
2392 /* backtracking doesn't work for STACK_ZERO yet. */
2393 err
= mark_chain_precision(env
, value_regno
);
2399 /* Mark slots affected by this stack write. */
2400 for (i
= 0; i
< size
; i
++)
2401 state
->stack
[spi
].slot_type
[(slot
- i
) % BPF_REG_SIZE
] =
2407 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
2408 * known to contain a variable offset.
2409 * This function checks whether the write is permitted and conservatively
2410 * tracks the effects of the write, considering that each stack slot in the
2411 * dynamic range is potentially written to.
2413 * 'off' includes 'regno->off'.
2414 * 'value_regno' can be -1, meaning that an unknown value is being written to
2417 * Spilled pointers in range are not marked as written because we don't know
2418 * what's going to be actually written. This means that read propagation for
2419 * future reads cannot be terminated by this write.
2421 * For privileged programs, uninitialized stack slots are considered
2422 * initialized by this write (even though we don't know exactly what offsets
2423 * are going to be written to). The idea is that we don't want the verifier to
2424 * reject future reads that access slots written to through variable offsets.
2426 static int check_stack_write_var_off(struct bpf_verifier_env
*env
,
2427 /* func where register points to */
2428 struct bpf_func_state
*state
,
2429 int ptr_regno
, int off
, int size
,
2430 int value_regno
, int insn_idx
)
2432 struct bpf_func_state
*cur
; /* state of the current function */
2433 int min_off
, max_off
;
2435 struct bpf_reg_state
*ptr_reg
= NULL
, *value_reg
= NULL
;
2436 bool writing_zero
= false;
2437 /* set if the fact that we're writing a zero is used to let any
2438 * stack slots remain STACK_ZERO
2440 bool zero_used
= false;
2442 cur
= env
->cur_state
->frame
[env
->cur_state
->curframe
];
2443 ptr_reg
= &cur
->regs
[ptr_regno
];
2444 min_off
= ptr_reg
->smin_value
+ off
;
2445 max_off
= ptr_reg
->smax_value
+ off
+ size
;
2446 if (value_regno
>= 0)
2447 value_reg
= &cur
->regs
[value_regno
];
2448 if (value_reg
&& register_is_null(value_reg
))
2449 writing_zero
= true;
2451 err
= realloc_func_state(state
, round_up(-min_off
, BPF_REG_SIZE
),
2452 state
->acquired_refs
, true);
2457 /* Variable offset writes destroy any spilled pointers in range. */
2458 for (i
= min_off
; i
< max_off
; i
++) {
2459 u8 new_type
, *stype
;
2463 spi
= slot
/ BPF_REG_SIZE
;
2464 stype
= &state
->stack
[spi
].slot_type
[slot
% BPF_REG_SIZE
];
2466 if (!env
->allow_ptr_leaks
2467 && *stype
!= NOT_INIT
2468 && *stype
!= SCALAR_VALUE
) {
2469 /* Reject the write if there's are spilled pointers in
2470 * range. If we didn't reject here, the ptr status
2471 * would be erased below (even though not all slots are
2472 * actually overwritten), possibly opening the door to
2475 verbose(env
, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
2480 /* Erase all spilled pointers. */
2481 state
->stack
[spi
].spilled_ptr
.type
= NOT_INIT
;
2483 /* Update the slot type. */
2484 new_type
= STACK_MISC
;
2485 if (writing_zero
&& *stype
== STACK_ZERO
) {
2486 new_type
= STACK_ZERO
;
2489 /* If the slot is STACK_INVALID, we check whether it's OK to
2490 * pretend that it will be initialized by this write. The slot
2491 * might not actually be written to, and so if we mark it as
2492 * initialized future reads might leak uninitialized memory.
2493 * For privileged programs, we will accept such reads to slots
2494 * that may or may not be written because, if we're reject
2495 * them, the error would be too confusing.
2497 if (*stype
== STACK_INVALID
&& !env
->allow_uninit_stack
) {
2498 verbose(env
, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
2505 /* backtracking doesn't work for STACK_ZERO yet. */
2506 err
= mark_chain_precision(env
, value_regno
);
2513 /* When register 'dst_regno' is assigned some values from stack[min_off,
2514 * max_off), we set the register's type according to the types of the
2515 * respective stack slots. If all the stack values are known to be zeros, then
2516 * so is the destination reg. Otherwise, the register is considered to be
2517 * SCALAR. This function does not deal with register filling; the caller must
2518 * ensure that all spilled registers in the stack range have been marked as
2521 static void mark_reg_stack_read(struct bpf_verifier_env
*env
,
2522 /* func where src register points to */
2523 struct bpf_func_state
*ptr_state
,
2524 int min_off
, int max_off
, int dst_regno
)
2526 struct bpf_verifier_state
*vstate
= env
->cur_state
;
2527 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
2532 for (i
= min_off
; i
< max_off
; i
++) {
2534 spi
= slot
/ BPF_REG_SIZE
;
2535 stype
= ptr_state
->stack
[spi
].slot_type
;
2536 if (stype
[slot
% BPF_REG_SIZE
] != STACK_ZERO
)
2540 if (zeros
== max_off
- min_off
) {
2541 /* any access_size read into register is zero extended,
2542 * so the whole register == const_zero
2544 __mark_reg_const_zero(&state
->regs
[dst_regno
]);
2545 /* backtracking doesn't support STACK_ZERO yet,
2546 * so mark it precise here, so that later
2547 * backtracking can stop here.
2548 * Backtracking may not need this if this register
2549 * doesn't participate in pointer adjustment.
2550 * Forward propagation of precise flag is not
2551 * necessary either. This mark is only to stop
2552 * backtracking. Any register that contributed
2553 * to const 0 was marked precise before spill.
2555 state
->regs
[dst_regno
].precise
= true;
2557 /* have read misc data from the stack */
2558 mark_reg_unknown(env
, state
->regs
, dst_regno
);
2560 state
->regs
[dst_regno
].live
|= REG_LIVE_WRITTEN
;
2563 /* Read the stack at 'off' and put the results into the register indicated by
2564 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
2567 * 'dst_regno' can be -1, meaning that the read value is not going to a
2570 * The access is assumed to be within the current stack bounds.
2572 static int check_stack_read_fixed_off(struct bpf_verifier_env
*env
,
2573 /* func where src register points to */
2574 struct bpf_func_state
*reg_state
,
2575 int off
, int size
, int dst_regno
)
2577 struct bpf_verifier_state
*vstate
= env
->cur_state
;
2578 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
2579 int i
, slot
= -off
- 1, spi
= slot
/ BPF_REG_SIZE
;
2580 struct bpf_reg_state
*reg
;
2583 stype
= reg_state
->stack
[spi
].slot_type
;
2584 reg
= ®_state
->stack
[spi
].spilled_ptr
;
2586 if (stype
[0] == STACK_SPILL
) {
2587 if (size
!= BPF_REG_SIZE
) {
2588 if (reg
->type
!= SCALAR_VALUE
) {
2589 verbose_linfo(env
, env
->insn_idx
, "; ");
2590 verbose(env
, "invalid size of register fill\n");
2593 if (dst_regno
>= 0) {
2594 mark_reg_unknown(env
, state
->regs
, dst_regno
);
2595 state
->regs
[dst_regno
].live
|= REG_LIVE_WRITTEN
;
2597 mark_reg_read(env
, reg
, reg
->parent
, REG_LIVE_READ64
);
2600 for (i
= 1; i
< BPF_REG_SIZE
; i
++) {
2601 if (stype
[(slot
- i
) % BPF_REG_SIZE
] != STACK_SPILL
) {
2602 verbose(env
, "corrupted spill memory\n");
2607 if (dst_regno
>= 0) {
2608 /* restore register state from stack */
2609 state
->regs
[dst_regno
] = *reg
;
2610 /* mark reg as written since spilled pointer state likely
2611 * has its liveness marks cleared by is_state_visited()
2612 * which resets stack/reg liveness for state transitions
2614 state
->regs
[dst_regno
].live
|= REG_LIVE_WRITTEN
;
2615 } else if (__is_pointer_value(env
->allow_ptr_leaks
, reg
)) {
2616 /* If dst_regno==-1, the caller is asking us whether
2617 * it is acceptable to use this value as a SCALAR_VALUE
2619 * We must not allow unprivileged callers to do that
2620 * with spilled pointers.
2622 verbose(env
, "leaking pointer from stack off %d\n",
2626 mark_reg_read(env
, reg
, reg
->parent
, REG_LIVE_READ64
);
2630 for (i
= 0; i
< size
; i
++) {
2631 type
= stype
[(slot
- i
) % BPF_REG_SIZE
];
2632 if (type
== STACK_MISC
)
2634 if (type
== STACK_ZERO
)
2636 verbose(env
, "invalid read from stack off %d+%d size %d\n",
2640 mark_reg_read(env
, reg
, reg
->parent
, REG_LIVE_READ64
);
2642 mark_reg_stack_read(env
, reg_state
, off
, off
+ size
, dst_regno
);
2647 enum stack_access_src
{
2648 ACCESS_DIRECT
= 1, /* the access is performed by an instruction */
2649 ACCESS_HELPER
= 2, /* the access is performed by a helper */
2652 static int check_stack_range_initialized(struct bpf_verifier_env
*env
,
2653 int regno
, int off
, int access_size
,
2654 bool zero_size_allowed
,
2655 enum stack_access_src type
,
2656 struct bpf_call_arg_meta
*meta
);
2658 static struct bpf_reg_state
*reg_state(struct bpf_verifier_env
*env
, int regno
)
2660 return cur_regs(env
) + regno
;
2663 /* Read the stack at 'ptr_regno + off' and put the result into the register
2665 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
2666 * but not its variable offset.
2667 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
2669 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
2670 * filling registers (i.e. reads of spilled register cannot be detected when
2671 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
2672 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
2673 * offset; for a fixed offset check_stack_read_fixed_off should be used
2676 static int check_stack_read_var_off(struct bpf_verifier_env
*env
,
2677 int ptr_regno
, int off
, int size
, int dst_regno
)
2679 /* The state of the source register. */
2680 struct bpf_reg_state
*reg
= reg_state(env
, ptr_regno
);
2681 struct bpf_func_state
*ptr_state
= func(env
, reg
);
2683 int min_off
, max_off
;
2685 /* Note that we pass a NULL meta, so raw access will not be permitted.
2687 err
= check_stack_range_initialized(env
, ptr_regno
, off
, size
,
2688 false, ACCESS_DIRECT
, NULL
);
2692 min_off
= reg
->smin_value
+ off
;
2693 max_off
= reg
->smax_value
+ off
;
2694 mark_reg_stack_read(env
, ptr_state
, min_off
, max_off
+ size
, dst_regno
);
2698 /* check_stack_read dispatches to check_stack_read_fixed_off or
2699 * check_stack_read_var_off.
2701 * The caller must ensure that the offset falls within the allocated stack
2704 * 'dst_regno' is a register which will receive the value from the stack. It
2705 * can be -1, meaning that the read value is not going to a register.
2707 static int check_stack_read(struct bpf_verifier_env
*env
,
2708 int ptr_regno
, int off
, int size
,
2711 struct bpf_reg_state
*reg
= reg_state(env
, ptr_regno
);
2712 struct bpf_func_state
*state
= func(env
, reg
);
2714 /* Some accesses are only permitted with a static offset. */
2715 bool var_off
= !tnum_is_const(reg
->var_off
);
2717 /* The offset is required to be static when reads don't go to a
2718 * register, in order to not leak pointers (see
2719 * check_stack_read_fixed_off).
2721 if (dst_regno
< 0 && var_off
) {
2724 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
2725 verbose(env
, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
2729 /* Variable offset is prohibited for unprivileged mode for simplicity
2730 * since it requires corresponding support in Spectre masking for stack
2731 * ALU. See also retrieve_ptr_limit().
2733 if (!env
->bypass_spec_v1
&& var_off
) {
2736 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
2737 verbose(env
, "R%d variable offset stack access prohibited for !root, var_off=%s\n",
2743 off
+= reg
->var_off
.value
;
2744 err
= check_stack_read_fixed_off(env
, state
, off
, size
,
2747 /* Variable offset stack reads need more conservative handling
2748 * than fixed offset ones. Note that dst_regno >= 0 on this
2751 err
= check_stack_read_var_off(env
, ptr_regno
, off
, size
,
2758 /* check_stack_write dispatches to check_stack_write_fixed_off or
2759 * check_stack_write_var_off.
2761 * 'ptr_regno' is the register used as a pointer into the stack.
2762 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
2763 * 'value_regno' is the register whose value we're writing to the stack. It can
2764 * be -1, meaning that we're not writing from a register.
2766 * The caller must ensure that the offset falls within the maximum stack size.
2768 static int check_stack_write(struct bpf_verifier_env
*env
,
2769 int ptr_regno
, int off
, int size
,
2770 int value_regno
, int insn_idx
)
2772 struct bpf_reg_state
*reg
= reg_state(env
, ptr_regno
);
2773 struct bpf_func_state
*state
= func(env
, reg
);
2776 if (tnum_is_const(reg
->var_off
)) {
2777 off
+= reg
->var_off
.value
;
2778 err
= check_stack_write_fixed_off(env
, state
, off
, size
,
2779 value_regno
, insn_idx
);
2781 /* Variable offset stack reads need more conservative handling
2782 * than fixed offset ones.
2784 err
= check_stack_write_var_off(env
, state
,
2785 ptr_regno
, off
, size
,
2786 value_regno
, insn_idx
);
2791 static int check_map_access_type(struct bpf_verifier_env
*env
, u32 regno
,
2792 int off
, int size
, enum bpf_access_type type
)
2794 struct bpf_reg_state
*regs
= cur_regs(env
);
2795 struct bpf_map
*map
= regs
[regno
].map_ptr
;
2796 u32 cap
= bpf_map_flags_to_cap(map
);
2798 if (type
== BPF_WRITE
&& !(cap
& BPF_MAP_CAN_WRITE
)) {
2799 verbose(env
, "write into map forbidden, value_size=%d off=%d size=%d\n",
2800 map
->value_size
, off
, size
);
2804 if (type
== BPF_READ
&& !(cap
& BPF_MAP_CAN_READ
)) {
2805 verbose(env
, "read from map forbidden, value_size=%d off=%d size=%d\n",
2806 map
->value_size
, off
, size
);
2813 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
2814 static int __check_mem_access(struct bpf_verifier_env
*env
, int regno
,
2815 int off
, int size
, u32 mem_size
,
2816 bool zero_size_allowed
)
2818 bool size_ok
= size
> 0 || (size
== 0 && zero_size_allowed
);
2819 struct bpf_reg_state
*reg
;
2821 if (off
>= 0 && size_ok
&& (u64
)off
+ size
<= mem_size
)
2824 reg
= &cur_regs(env
)[regno
];
2825 switch (reg
->type
) {
2826 case PTR_TO_MAP_VALUE
:
2827 verbose(env
, "invalid access to map value, value_size=%d off=%d size=%d\n",
2828 mem_size
, off
, size
);
2831 case PTR_TO_PACKET_META
:
2832 case PTR_TO_PACKET_END
:
2833 verbose(env
, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
2834 off
, size
, regno
, reg
->id
, off
, mem_size
);
2838 verbose(env
, "invalid access to memory, mem_size=%u off=%d size=%d\n",
2839 mem_size
, off
, size
);
2845 /* check read/write into a memory region with possible variable offset */
2846 static int check_mem_region_access(struct bpf_verifier_env
*env
, u32 regno
,
2847 int off
, int size
, u32 mem_size
,
2848 bool zero_size_allowed
)
2850 struct bpf_verifier_state
*vstate
= env
->cur_state
;
2851 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
2852 struct bpf_reg_state
*reg
= &state
->regs
[regno
];
2855 /* We may have adjusted the register pointing to memory region, so we
2856 * need to try adding each of min_value and max_value to off
2857 * to make sure our theoretical access will be safe.
2859 if (env
->log
.level
& BPF_LOG_LEVEL
)
2860 print_verifier_state(env
, state
);
2862 /* The minimum value is only important with signed
2863 * comparisons where we can't assume the floor of a
2864 * value is 0. If we are using signed variables for our
2865 * index'es we need to make sure that whatever we use
2866 * will have a set floor within our range.
2868 if (reg
->smin_value
< 0 &&
2869 (reg
->smin_value
== S64_MIN
||
2870 (off
+ reg
->smin_value
!= (s64
)(s32
)(off
+ reg
->smin_value
)) ||
2871 reg
->smin_value
+ off
< 0)) {
2872 verbose(env
, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2876 err
= __check_mem_access(env
, regno
, reg
->smin_value
+ off
, size
,
2877 mem_size
, zero_size_allowed
);
2879 verbose(env
, "R%d min value is outside of the allowed memory range\n",
2884 /* If we haven't set a max value then we need to bail since we can't be
2885 * sure we won't do bad things.
2886 * If reg->umax_value + off could overflow, treat that as unbounded too.
2888 if (reg
->umax_value
>= BPF_MAX_VAR_OFF
) {
2889 verbose(env
, "R%d unbounded memory access, make sure to bounds check any such access\n",
2893 err
= __check_mem_access(env
, regno
, reg
->umax_value
+ off
, size
,
2894 mem_size
, zero_size_allowed
);
2896 verbose(env
, "R%d max value is outside of the allowed memory range\n",
2904 /* check read/write into a map element with possible variable offset */
2905 static int check_map_access(struct bpf_verifier_env
*env
, u32 regno
,
2906 int off
, int size
, bool zero_size_allowed
)
2908 struct bpf_verifier_state
*vstate
= env
->cur_state
;
2909 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
2910 struct bpf_reg_state
*reg
= &state
->regs
[regno
];
2911 struct bpf_map
*map
= reg
->map_ptr
;
2914 err
= check_mem_region_access(env
, regno
, off
, size
, map
->value_size
,
2919 if (map_value_has_spin_lock(map
)) {
2920 u32 lock
= map
->spin_lock_off
;
2922 /* if any part of struct bpf_spin_lock can be touched by
2923 * load/store reject this program.
2924 * To check that [x1, x2) overlaps with [y1, y2)
2925 * it is sufficient to check x1 < y2 && y1 < x2.
2927 if (reg
->smin_value
+ off
< lock
+ sizeof(struct bpf_spin_lock
) &&
2928 lock
< reg
->umax_value
+ off
+ size
) {
2929 verbose(env
, "bpf_spin_lock cannot be accessed directly by load/store\n");
2936 #define MAX_PACKET_OFF 0xffff
2938 static enum bpf_prog_type
resolve_prog_type(struct bpf_prog
*prog
)
2940 return prog
->aux
->dst_prog
? prog
->aux
->dst_prog
->type
: prog
->type
;
2943 static bool may_access_direct_pkt_data(struct bpf_verifier_env
*env
,
2944 const struct bpf_call_arg_meta
*meta
,
2945 enum bpf_access_type t
)
2947 enum bpf_prog_type prog_type
= resolve_prog_type(env
->prog
);
2949 switch (prog_type
) {
2950 /* Program types only with direct read access go here! */
2951 case BPF_PROG_TYPE_LWT_IN
:
2952 case BPF_PROG_TYPE_LWT_OUT
:
2953 case BPF_PROG_TYPE_LWT_SEG6LOCAL
:
2954 case BPF_PROG_TYPE_SK_REUSEPORT
:
2955 case BPF_PROG_TYPE_FLOW_DISSECTOR
:
2956 case BPF_PROG_TYPE_CGROUP_SKB
:
2961 /* Program types with direct read + write access go here! */
2962 case BPF_PROG_TYPE_SCHED_CLS
:
2963 case BPF_PROG_TYPE_SCHED_ACT
:
2964 case BPF_PROG_TYPE_XDP
:
2965 case BPF_PROG_TYPE_LWT_XMIT
:
2966 case BPF_PROG_TYPE_SK_SKB
:
2967 case BPF_PROG_TYPE_SK_MSG
:
2969 return meta
->pkt_access
;
2971 env
->seen_direct_write
= true;
2974 case BPF_PROG_TYPE_CGROUP_SOCKOPT
:
2976 env
->seen_direct_write
= true;
2985 static int check_packet_access(struct bpf_verifier_env
*env
, u32 regno
, int off
,
2986 int size
, bool zero_size_allowed
)
2988 struct bpf_reg_state
*regs
= cur_regs(env
);
2989 struct bpf_reg_state
*reg
= ®s
[regno
];
2992 /* We may have added a variable offset to the packet pointer; but any
2993 * reg->range we have comes after that. We are only checking the fixed
2997 /* We don't allow negative numbers, because we aren't tracking enough
2998 * detail to prove they're safe.
3000 if (reg
->smin_value
< 0) {
3001 verbose(env
, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3006 err
= reg
->range
< 0 ? -EINVAL
:
3007 __check_mem_access(env
, regno
, off
, size
, reg
->range
,
3010 verbose(env
, "R%d offset is outside of the packet\n", regno
);
3014 /* __check_mem_access has made sure "off + size - 1" is within u16.
3015 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
3016 * otherwise find_good_pkt_pointers would have refused to set range info
3017 * that __check_mem_access would have rejected this pkt access.
3018 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
3020 env
->prog
->aux
->max_pkt_offset
=
3021 max_t(u32
, env
->prog
->aux
->max_pkt_offset
,
3022 off
+ reg
->umax_value
+ size
- 1);
3027 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
3028 static int check_ctx_access(struct bpf_verifier_env
*env
, int insn_idx
, int off
, int size
,
3029 enum bpf_access_type t
, enum bpf_reg_type
*reg_type
,
3030 struct btf
**btf
, u32
*btf_id
)
3032 struct bpf_insn_access_aux info
= {
3033 .reg_type
= *reg_type
,
3037 if (env
->ops
->is_valid_access
&&
3038 env
->ops
->is_valid_access(off
, size
, t
, env
->prog
, &info
)) {
3039 /* A non zero info.ctx_field_size indicates that this field is a
3040 * candidate for later verifier transformation to load the whole
3041 * field and then apply a mask when accessed with a narrower
3042 * access than actual ctx access size. A zero info.ctx_field_size
3043 * will only allow for whole field access and rejects any other
3044 * type of narrower access.
3046 *reg_type
= info
.reg_type
;
3048 if (*reg_type
== PTR_TO_BTF_ID
|| *reg_type
== PTR_TO_BTF_ID_OR_NULL
) {
3050 *btf_id
= info
.btf_id
;
3052 env
->insn_aux_data
[insn_idx
].ctx_field_size
= info
.ctx_field_size
;
3054 /* remember the offset of last byte accessed in ctx */
3055 if (env
->prog
->aux
->max_ctx_offset
< off
+ size
)
3056 env
->prog
->aux
->max_ctx_offset
= off
+ size
;
3060 verbose(env
, "invalid bpf_context access off=%d size=%d\n", off
, size
);
3064 static int check_flow_keys_access(struct bpf_verifier_env
*env
, int off
,
3067 if (size
< 0 || off
< 0 ||
3068 (u64
)off
+ size
> sizeof(struct bpf_flow_keys
)) {
3069 verbose(env
, "invalid access to flow keys off=%d size=%d\n",
3076 static int check_sock_access(struct bpf_verifier_env
*env
, int insn_idx
,
3077 u32 regno
, int off
, int size
,
3078 enum bpf_access_type t
)
3080 struct bpf_reg_state
*regs
= cur_regs(env
);
3081 struct bpf_reg_state
*reg
= ®s
[regno
];
3082 struct bpf_insn_access_aux info
= {};
3085 if (reg
->smin_value
< 0) {
3086 verbose(env
, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3091 switch (reg
->type
) {
3092 case PTR_TO_SOCK_COMMON
:
3093 valid
= bpf_sock_common_is_valid_access(off
, size
, t
, &info
);
3096 valid
= bpf_sock_is_valid_access(off
, size
, t
, &info
);
3098 case PTR_TO_TCP_SOCK
:
3099 valid
= bpf_tcp_sock_is_valid_access(off
, size
, t
, &info
);
3101 case PTR_TO_XDP_SOCK
:
3102 valid
= bpf_xdp_sock_is_valid_access(off
, size
, t
, &info
);
3110 env
->insn_aux_data
[insn_idx
].ctx_field_size
=
3111 info
.ctx_field_size
;
3115 verbose(env
, "R%d invalid %s access off=%d size=%d\n",
3116 regno
, reg_type_str
[reg
->type
], off
, size
);
3121 static bool is_pointer_value(struct bpf_verifier_env
*env
, int regno
)
3123 return __is_pointer_value(env
->allow_ptr_leaks
, reg_state(env
, regno
));
3126 static bool is_ctx_reg(struct bpf_verifier_env
*env
, int regno
)
3128 const struct bpf_reg_state
*reg
= reg_state(env
, regno
);
3130 return reg
->type
== PTR_TO_CTX
;
3133 static bool is_sk_reg(struct bpf_verifier_env
*env
, int regno
)
3135 const struct bpf_reg_state
*reg
= reg_state(env
, regno
);
3137 return type_is_sk_pointer(reg
->type
);
3140 static bool is_pkt_reg(struct bpf_verifier_env
*env
, int regno
)
3142 const struct bpf_reg_state
*reg
= reg_state(env
, regno
);
3144 return type_is_pkt_pointer(reg
->type
);
3147 static bool is_flow_key_reg(struct bpf_verifier_env
*env
, int regno
)
3149 const struct bpf_reg_state
*reg
= reg_state(env
, regno
);
3151 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
3152 return reg
->type
== PTR_TO_FLOW_KEYS
;
3155 static int check_pkt_ptr_alignment(struct bpf_verifier_env
*env
,
3156 const struct bpf_reg_state
*reg
,
3157 int off
, int size
, bool strict
)
3159 struct tnum reg_off
;
3162 /* Byte size accesses are always allowed. */
3163 if (!strict
|| size
== 1)
3166 /* For platforms that do not have a Kconfig enabling
3167 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
3168 * NET_IP_ALIGN is universally set to '2'. And on platforms
3169 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
3170 * to this code only in strict mode where we want to emulate
3171 * the NET_IP_ALIGN==2 checking. Therefore use an
3172 * unconditional IP align value of '2'.
3176 reg_off
= tnum_add(reg
->var_off
, tnum_const(ip_align
+ reg
->off
+ off
));
3177 if (!tnum_is_aligned(reg_off
, size
)) {
3180 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
3182 "misaligned packet access off %d+%s+%d+%d size %d\n",
3183 ip_align
, tn_buf
, reg
->off
, off
, size
);
3190 static int check_generic_ptr_alignment(struct bpf_verifier_env
*env
,
3191 const struct bpf_reg_state
*reg
,
3192 const char *pointer_desc
,
3193 int off
, int size
, bool strict
)
3195 struct tnum reg_off
;
3197 /* Byte size accesses are always allowed. */
3198 if (!strict
|| size
== 1)
3201 reg_off
= tnum_add(reg
->var_off
, tnum_const(reg
->off
+ off
));
3202 if (!tnum_is_aligned(reg_off
, size
)) {
3205 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
3206 verbose(env
, "misaligned %saccess off %s+%d+%d size %d\n",
3207 pointer_desc
, tn_buf
, reg
->off
, off
, size
);
3214 static int check_ptr_alignment(struct bpf_verifier_env
*env
,
3215 const struct bpf_reg_state
*reg
, int off
,
3216 int size
, bool strict_alignment_once
)
3218 bool strict
= env
->strict_alignment
|| strict_alignment_once
;
3219 const char *pointer_desc
= "";
3221 switch (reg
->type
) {
3223 case PTR_TO_PACKET_META
:
3224 /* Special case, because of NET_IP_ALIGN. Given metadata sits
3225 * right in front, treat it the very same way.
3227 return check_pkt_ptr_alignment(env
, reg
, off
, size
, strict
);
3228 case PTR_TO_FLOW_KEYS
:
3229 pointer_desc
= "flow keys ";
3231 case PTR_TO_MAP_VALUE
:
3232 pointer_desc
= "value ";
3235 pointer_desc
= "context ";
3238 pointer_desc
= "stack ";
3239 /* The stack spill tracking logic in check_stack_write_fixed_off()
3240 * and check_stack_read_fixed_off() relies on stack accesses being
3246 pointer_desc
= "sock ";
3248 case PTR_TO_SOCK_COMMON
:
3249 pointer_desc
= "sock_common ";
3251 case PTR_TO_TCP_SOCK
:
3252 pointer_desc
= "tcp_sock ";
3254 case PTR_TO_XDP_SOCK
:
3255 pointer_desc
= "xdp_sock ";
3260 return check_generic_ptr_alignment(env
, reg
, pointer_desc
, off
, size
,
3264 static int update_stack_depth(struct bpf_verifier_env
*env
,
3265 const struct bpf_func_state
*func
,
3268 u16 stack
= env
->subprog_info
[func
->subprogno
].stack_depth
;
3273 /* update known max for given subprogram */
3274 env
->subprog_info
[func
->subprogno
].stack_depth
= -off
;
3278 /* starting from main bpf function walk all instructions of the function
3279 * and recursively walk all callees that given function can call.
3280 * Ignore jump and exit insns.
3281 * Since recursion is prevented by check_cfg() this algorithm
3282 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
3284 static int check_max_stack_depth(struct bpf_verifier_env
*env
)
3286 int depth
= 0, frame
= 0, idx
= 0, i
= 0, subprog_end
;
3287 struct bpf_subprog_info
*subprog
= env
->subprog_info
;
3288 struct bpf_insn
*insn
= env
->prog
->insnsi
;
3289 bool tail_call_reachable
= false;
3290 int ret_insn
[MAX_CALL_FRAMES
];
3291 int ret_prog
[MAX_CALL_FRAMES
];
3295 /* protect against potential stack overflow that might happen when
3296 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
3297 * depth for such case down to 256 so that the worst case scenario
3298 * would result in 8k stack size (32 which is tailcall limit * 256 =
3301 * To get the idea what might happen, see an example:
3302 * func1 -> sub rsp, 128
3303 * subfunc1 -> sub rsp, 256
3304 * tailcall1 -> add rsp, 256
3305 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
3306 * subfunc2 -> sub rsp, 64
3307 * subfunc22 -> sub rsp, 128
3308 * tailcall2 -> add rsp, 128
3309 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
3311 * tailcall will unwind the current stack frame but it will not get rid
3312 * of caller's stack as shown on the example above.
3314 if (idx
&& subprog
[idx
].has_tail_call
&& depth
>= 256) {
3316 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
3320 /* round up to 32-bytes, since this is granularity
3321 * of interpreter stack size
3323 depth
+= round_up(max_t(u32
, subprog
[idx
].stack_depth
, 1), 32);
3324 if (depth
> MAX_BPF_STACK
) {
3325 verbose(env
, "combined stack size of %d calls is %d. Too large\n",
3330 subprog_end
= subprog
[idx
+ 1].start
;
3331 for (; i
< subprog_end
; i
++) {
3332 if (insn
[i
].code
!= (BPF_JMP
| BPF_CALL
))
3334 if (insn
[i
].src_reg
!= BPF_PSEUDO_CALL
)
3336 /* remember insn and function to return to */
3337 ret_insn
[frame
] = i
+ 1;
3338 ret_prog
[frame
] = idx
;
3340 /* find the callee */
3341 i
= i
+ insn
[i
].imm
+ 1;
3342 idx
= find_subprog(env
, i
);
3344 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3349 if (subprog
[idx
].has_tail_call
)
3350 tail_call_reachable
= true;
3353 if (frame
>= MAX_CALL_FRAMES
) {
3354 verbose(env
, "the call stack of %d frames is too deep !\n",
3360 /* if tail call got detected across bpf2bpf calls then mark each of the
3361 * currently present subprog frames as tail call reachable subprogs;
3362 * this info will be utilized by JIT so that we will be preserving the
3363 * tail call counter throughout bpf2bpf calls combined with tailcalls
3365 if (tail_call_reachable
)
3366 for (j
= 0; j
< frame
; j
++)
3367 subprog
[ret_prog
[j
]].tail_call_reachable
= true;
3369 /* end of for() loop means the last insn of the 'subprog'
3370 * was reached. Doesn't matter whether it was JA or EXIT
3374 depth
-= round_up(max_t(u32
, subprog
[idx
].stack_depth
, 1), 32);
3376 i
= ret_insn
[frame
];
3377 idx
= ret_prog
[frame
];
3381 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
3382 static int get_callee_stack_depth(struct bpf_verifier_env
*env
,
3383 const struct bpf_insn
*insn
, int idx
)
3385 int start
= idx
+ insn
->imm
+ 1, subprog
;
3387 subprog
= find_subprog(env
, start
);
3389 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3393 return env
->subprog_info
[subprog
].stack_depth
;
3397 int check_ctx_reg(struct bpf_verifier_env
*env
,
3398 const struct bpf_reg_state
*reg
, int regno
)
3400 /* Access to ctx or passing it to a helper is only allowed in
3401 * its original, unmodified form.
3405 verbose(env
, "dereference of modified ctx ptr R%d off=%d disallowed\n",
3410 if (!tnum_is_const(reg
->var_off
) || reg
->var_off
.value
) {
3413 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
3414 verbose(env
, "variable ctx access var_off=%s disallowed\n", tn_buf
);
3421 static int __check_buffer_access(struct bpf_verifier_env
*env
,
3422 const char *buf_info
,
3423 const struct bpf_reg_state
*reg
,
3424 int regno
, int off
, int size
)
3428 "R%d invalid %s buffer access: off=%d, size=%d\n",
3429 regno
, buf_info
, off
, size
);
3432 if (!tnum_is_const(reg
->var_off
) || reg
->var_off
.value
) {
3435 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
3437 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
3438 regno
, off
, tn_buf
);
3445 static int check_tp_buffer_access(struct bpf_verifier_env
*env
,
3446 const struct bpf_reg_state
*reg
,
3447 int regno
, int off
, int size
)
3451 err
= __check_buffer_access(env
, "tracepoint", reg
, regno
, off
, size
);
3455 if (off
+ size
> env
->prog
->aux
->max_tp_access
)
3456 env
->prog
->aux
->max_tp_access
= off
+ size
;
3461 static int check_buffer_access(struct bpf_verifier_env
*env
,
3462 const struct bpf_reg_state
*reg
,
3463 int regno
, int off
, int size
,
3464 bool zero_size_allowed
,
3465 const char *buf_info
,
3470 err
= __check_buffer_access(env
, buf_info
, reg
, regno
, off
, size
);
3474 if (off
+ size
> *max_access
)
3475 *max_access
= off
+ size
;
3480 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
3481 static void zext_32_to_64(struct bpf_reg_state
*reg
)
3483 reg
->var_off
= tnum_subreg(reg
->var_off
);
3484 __reg_assign_32_into_64(reg
);
3487 /* truncate register to smaller size (in bytes)
3488 * must be called with size < BPF_REG_SIZE
3490 static void coerce_reg_to_size(struct bpf_reg_state
*reg
, int size
)
3494 /* clear high bits in bit representation */
3495 reg
->var_off
= tnum_cast(reg
->var_off
, size
);
3497 /* fix arithmetic bounds */
3498 mask
= ((u64
)1 << (size
* 8)) - 1;
3499 if ((reg
->umin_value
& ~mask
) == (reg
->umax_value
& ~mask
)) {
3500 reg
->umin_value
&= mask
;
3501 reg
->umax_value
&= mask
;
3503 reg
->umin_value
= 0;
3504 reg
->umax_value
= mask
;
3506 reg
->smin_value
= reg
->umin_value
;
3507 reg
->smax_value
= reg
->umax_value
;
3509 /* If size is smaller than 32bit register the 32bit register
3510 * values are also truncated so we push 64-bit bounds into
3511 * 32-bit bounds. Above were truncated < 32-bits already.
3515 __reg_combine_64_into_32(reg
);
3518 static bool bpf_map_is_rdonly(const struct bpf_map
*map
)
3520 return (map
->map_flags
& BPF_F_RDONLY_PROG
) && map
->frozen
;
3523 static int bpf_map_direct_read(struct bpf_map
*map
, int off
, int size
, u64
*val
)
3529 err
= map
->ops
->map_direct_value_addr(map
, &addr
, off
);
3532 ptr
= (void *)(long)addr
+ off
;
3536 *val
= (u64
)*(u8
*)ptr
;
3539 *val
= (u64
)*(u16
*)ptr
;
3542 *val
= (u64
)*(u32
*)ptr
;
3553 static int check_ptr_to_btf_access(struct bpf_verifier_env
*env
,
3554 struct bpf_reg_state
*regs
,
3555 int regno
, int off
, int size
,
3556 enum bpf_access_type atype
,
3559 struct bpf_reg_state
*reg
= regs
+ regno
;
3560 const struct btf_type
*t
= btf_type_by_id(reg
->btf
, reg
->btf_id
);
3561 const char *tname
= btf_name_by_offset(reg
->btf
, t
->name_off
);
3567 "R%d is ptr_%s invalid negative access: off=%d\n",
3571 if (!tnum_is_const(reg
->var_off
) || reg
->var_off
.value
) {
3574 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
3576 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
3577 regno
, tname
, off
, tn_buf
);
3581 if (env
->ops
->btf_struct_access
) {
3582 ret
= env
->ops
->btf_struct_access(&env
->log
, reg
->btf
, t
,
3583 off
, size
, atype
, &btf_id
);
3585 if (atype
!= BPF_READ
) {
3586 verbose(env
, "only read is supported\n");
3590 ret
= btf_struct_access(&env
->log
, reg
->btf
, t
, off
, size
,
3597 if (atype
== BPF_READ
&& value_regno
>= 0)
3598 mark_btf_ld_reg(env
, regs
, value_regno
, ret
, reg
->btf
, btf_id
);
3603 static int check_ptr_to_map_access(struct bpf_verifier_env
*env
,
3604 struct bpf_reg_state
*regs
,
3605 int regno
, int off
, int size
,
3606 enum bpf_access_type atype
,
3609 struct bpf_reg_state
*reg
= regs
+ regno
;
3610 struct bpf_map
*map
= reg
->map_ptr
;
3611 const struct btf_type
*t
;
3617 verbose(env
, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
3621 if (!map
->ops
->map_btf_id
|| !*map
->ops
->map_btf_id
) {
3622 verbose(env
, "map_ptr access not supported for map type %d\n",
3627 t
= btf_type_by_id(btf_vmlinux
, *map
->ops
->map_btf_id
);
3628 tname
= btf_name_by_offset(btf_vmlinux
, t
->name_off
);
3630 if (!env
->allow_ptr_to_map_access
) {
3632 "%s access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
3638 verbose(env
, "R%d is %s invalid negative access: off=%d\n",
3643 if (atype
!= BPF_READ
) {
3644 verbose(env
, "only read from %s is supported\n", tname
);
3648 ret
= btf_struct_access(&env
->log
, btf_vmlinux
, t
, off
, size
, atype
, &btf_id
);
3652 if (value_regno
>= 0)
3653 mark_btf_ld_reg(env
, regs
, value_regno
, ret
, btf_vmlinux
, btf_id
);
3658 /* Check that the stack access at the given offset is within bounds. The
3659 * maximum valid offset is -1.
3661 * The minimum valid offset is -MAX_BPF_STACK for writes, and
3662 * -state->allocated_stack for reads.
3664 static int check_stack_slot_within_bounds(int off
,
3665 struct bpf_func_state
*state
,
3666 enum bpf_access_type t
)
3671 min_valid_off
= -MAX_BPF_STACK
;
3673 min_valid_off
= -state
->allocated_stack
;
3675 if (off
< min_valid_off
|| off
> -1)
3680 /* Check that the stack access at 'regno + off' falls within the maximum stack
3683 * 'off' includes `regno->offset`, but not its dynamic part (if any).
3685 static int check_stack_access_within_bounds(
3686 struct bpf_verifier_env
*env
,
3687 int regno
, int off
, int access_size
,
3688 enum stack_access_src src
, enum bpf_access_type type
)
3690 struct bpf_reg_state
*regs
= cur_regs(env
);
3691 struct bpf_reg_state
*reg
= regs
+ regno
;
3692 struct bpf_func_state
*state
= func(env
, reg
);
3693 int min_off
, max_off
;
3697 if (src
== ACCESS_HELPER
)
3698 /* We don't know if helpers are reading or writing (or both). */
3699 err_extra
= " indirect access to";
3700 else if (type
== BPF_READ
)
3701 err_extra
= " read from";
3703 err_extra
= " write to";
3705 if (tnum_is_const(reg
->var_off
)) {
3706 min_off
= reg
->var_off
.value
+ off
;
3707 if (access_size
> 0)
3708 max_off
= min_off
+ access_size
- 1;
3712 if (reg
->smax_value
>= BPF_MAX_VAR_OFF
||
3713 reg
->smin_value
<= -BPF_MAX_VAR_OFF
) {
3714 verbose(env
, "invalid unbounded variable-offset%s stack R%d\n",
3718 min_off
= reg
->smin_value
+ off
;
3719 if (access_size
> 0)
3720 max_off
= reg
->smax_value
+ off
+ access_size
- 1;
3725 err
= check_stack_slot_within_bounds(min_off
, state
, type
);
3727 err
= check_stack_slot_within_bounds(max_off
, state
, type
);
3730 if (tnum_is_const(reg
->var_off
)) {
3731 verbose(env
, "invalid%s stack R%d off=%d size=%d\n",
3732 err_extra
, regno
, off
, access_size
);
3736 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
3737 verbose(env
, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
3738 err_extra
, regno
, tn_buf
, access_size
);
3744 /* check whether memory at (regno + off) is accessible for t = (read | write)
3745 * if t==write, value_regno is a register which value is stored into memory
3746 * if t==read, value_regno is a register which will receive the value from memory
3747 * if t==write && value_regno==-1, some unknown value is stored into memory
3748 * if t==read && value_regno==-1, don't care what we read from memory
3750 static int check_mem_access(struct bpf_verifier_env
*env
, int insn_idx
, u32 regno
,
3751 int off
, int bpf_size
, enum bpf_access_type t
,
3752 int value_regno
, bool strict_alignment_once
)
3754 struct bpf_reg_state
*regs
= cur_regs(env
);
3755 struct bpf_reg_state
*reg
= regs
+ regno
;
3756 struct bpf_func_state
*state
;
3759 size
= bpf_size_to_bytes(bpf_size
);
3763 /* alignment checks will add in reg->off themselves */
3764 err
= check_ptr_alignment(env
, reg
, off
, size
, strict_alignment_once
);
3768 /* for access checks, reg->off is just part of off */
3771 if (reg
->type
== PTR_TO_MAP_VALUE
) {
3772 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
3773 is_pointer_value(env
, value_regno
)) {
3774 verbose(env
, "R%d leaks addr into map\n", value_regno
);
3777 err
= check_map_access_type(env
, regno
, off
, size
, t
);
3780 err
= check_map_access(env
, regno
, off
, size
, false);
3781 if (!err
&& t
== BPF_READ
&& value_regno
>= 0) {
3782 struct bpf_map
*map
= reg
->map_ptr
;
3784 /* if map is read-only, track its contents as scalars */
3785 if (tnum_is_const(reg
->var_off
) &&
3786 bpf_map_is_rdonly(map
) &&
3787 map
->ops
->map_direct_value_addr
) {
3788 int map_off
= off
+ reg
->var_off
.value
;
3791 err
= bpf_map_direct_read(map
, map_off
, size
,
3796 regs
[value_regno
].type
= SCALAR_VALUE
;
3797 __mark_reg_known(®s
[value_regno
], val
);
3799 mark_reg_unknown(env
, regs
, value_regno
);
3802 } else if (reg
->type
== PTR_TO_MEM
) {
3803 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
3804 is_pointer_value(env
, value_regno
)) {
3805 verbose(env
, "R%d leaks addr into mem\n", value_regno
);
3808 err
= check_mem_region_access(env
, regno
, off
, size
,
3809 reg
->mem_size
, false);
3810 if (!err
&& t
== BPF_READ
&& value_regno
>= 0)
3811 mark_reg_unknown(env
, regs
, value_regno
);
3812 } else if (reg
->type
== PTR_TO_CTX
) {
3813 enum bpf_reg_type reg_type
= SCALAR_VALUE
;
3814 struct btf
*btf
= NULL
;
3817 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
3818 is_pointer_value(env
, value_regno
)) {
3819 verbose(env
, "R%d leaks addr into ctx\n", value_regno
);
3823 err
= check_ctx_reg(env
, reg
, regno
);
3827 err
= check_ctx_access(env
, insn_idx
, off
, size
, t
, ®_type
, &btf
, &btf_id
);
3829 verbose_linfo(env
, insn_idx
, "; ");
3830 if (!err
&& t
== BPF_READ
&& value_regno
>= 0) {
3831 /* ctx access returns either a scalar, or a
3832 * PTR_TO_PACKET[_META,_END]. In the latter
3833 * case, we know the offset is zero.
3835 if (reg_type
== SCALAR_VALUE
) {
3836 mark_reg_unknown(env
, regs
, value_regno
);
3838 mark_reg_known_zero(env
, regs
,
3840 if (reg_type_may_be_null(reg_type
))
3841 regs
[value_regno
].id
= ++env
->id_gen
;
3842 /* A load of ctx field could have different
3843 * actual load size with the one encoded in the
3844 * insn. When the dst is PTR, it is for sure not
3847 regs
[value_regno
].subreg_def
= DEF_NOT_SUBREG
;
3848 if (reg_type
== PTR_TO_BTF_ID
||
3849 reg_type
== PTR_TO_BTF_ID_OR_NULL
) {
3850 regs
[value_regno
].btf
= btf
;
3851 regs
[value_regno
].btf_id
= btf_id
;
3854 regs
[value_regno
].type
= reg_type
;
3857 } else if (reg
->type
== PTR_TO_STACK
) {
3858 /* Basic bounds checks. */
3859 err
= check_stack_access_within_bounds(env
, regno
, off
, size
, ACCESS_DIRECT
, t
);
3863 state
= func(env
, reg
);
3864 err
= update_stack_depth(env
, state
, off
);
3869 err
= check_stack_read(env
, regno
, off
, size
,
3872 err
= check_stack_write(env
, regno
, off
, size
,
3873 value_regno
, insn_idx
);
3874 } else if (reg_is_pkt_pointer(reg
)) {
3875 if (t
== BPF_WRITE
&& !may_access_direct_pkt_data(env
, NULL
, t
)) {
3876 verbose(env
, "cannot write into packet\n");
3879 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
3880 is_pointer_value(env
, value_regno
)) {
3881 verbose(env
, "R%d leaks addr into packet\n",
3885 err
= check_packet_access(env
, regno
, off
, size
, false);
3886 if (!err
&& t
== BPF_READ
&& value_regno
>= 0)
3887 mark_reg_unknown(env
, regs
, value_regno
);
3888 } else if (reg
->type
== PTR_TO_FLOW_KEYS
) {
3889 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
3890 is_pointer_value(env
, value_regno
)) {
3891 verbose(env
, "R%d leaks addr into flow keys\n",
3896 err
= check_flow_keys_access(env
, off
, size
);
3897 if (!err
&& t
== BPF_READ
&& value_regno
>= 0)
3898 mark_reg_unknown(env
, regs
, value_regno
);
3899 } else if (type_is_sk_pointer(reg
->type
)) {
3900 if (t
== BPF_WRITE
) {
3901 verbose(env
, "R%d cannot write into %s\n",
3902 regno
, reg_type_str
[reg
->type
]);
3905 err
= check_sock_access(env
, insn_idx
, regno
, off
, size
, t
);
3906 if (!err
&& value_regno
>= 0)
3907 mark_reg_unknown(env
, regs
, value_regno
);
3908 } else if (reg
->type
== PTR_TO_TP_BUFFER
) {
3909 err
= check_tp_buffer_access(env
, reg
, regno
, off
, size
);
3910 if (!err
&& t
== BPF_READ
&& value_regno
>= 0)
3911 mark_reg_unknown(env
, regs
, value_regno
);
3912 } else if (reg
->type
== PTR_TO_BTF_ID
) {
3913 err
= check_ptr_to_btf_access(env
, regs
, regno
, off
, size
, t
,
3915 } else if (reg
->type
== CONST_PTR_TO_MAP
) {
3916 err
= check_ptr_to_map_access(env
, regs
, regno
, off
, size
, t
,
3918 } else if (reg
->type
== PTR_TO_RDONLY_BUF
) {
3919 if (t
== BPF_WRITE
) {
3920 verbose(env
, "R%d cannot write into %s\n",
3921 regno
, reg_type_str
[reg
->type
]);
3924 err
= check_buffer_access(env
, reg
, regno
, off
, size
, false,
3926 &env
->prog
->aux
->max_rdonly_access
);
3927 if (!err
&& value_regno
>= 0)
3928 mark_reg_unknown(env
, regs
, value_regno
);
3929 } else if (reg
->type
== PTR_TO_RDWR_BUF
) {
3930 err
= check_buffer_access(env
, reg
, regno
, off
, size
, false,
3932 &env
->prog
->aux
->max_rdwr_access
);
3933 if (!err
&& t
== BPF_READ
&& value_regno
>= 0)
3934 mark_reg_unknown(env
, regs
, value_regno
);
3936 verbose(env
, "R%d invalid mem access '%s'\n", regno
,
3937 reg_type_str
[reg
->type
]);
3941 if (!err
&& size
< BPF_REG_SIZE
&& value_regno
>= 0 && t
== BPF_READ
&&
3942 regs
[value_regno
].type
== SCALAR_VALUE
) {
3943 /* b/h/w load zero-extends, mark upper bits as known 0 */
3944 coerce_reg_to_size(®s
[value_regno
], size
);
3949 static int check_xadd(struct bpf_verifier_env
*env
, int insn_idx
, struct bpf_insn
*insn
)
3953 if ((BPF_SIZE(insn
->code
) != BPF_W
&& BPF_SIZE(insn
->code
) != BPF_DW
) ||
3955 verbose(env
, "BPF_XADD uses reserved fields\n");
3959 /* check src1 operand */
3960 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
3964 /* check src2 operand */
3965 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
3969 if (is_pointer_value(env
, insn
->src_reg
)) {
3970 verbose(env
, "R%d leaks addr into mem\n", insn
->src_reg
);
3974 if (is_ctx_reg(env
, insn
->dst_reg
) ||
3975 is_pkt_reg(env
, insn
->dst_reg
) ||
3976 is_flow_key_reg(env
, insn
->dst_reg
) ||
3977 is_sk_reg(env
, insn
->dst_reg
)) {
3978 verbose(env
, "BPF_XADD stores into R%d %s is not allowed\n",
3980 reg_type_str
[reg_state(env
, insn
->dst_reg
)->type
]);
3984 /* check whether atomic_add can read the memory */
3985 err
= check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
3986 BPF_SIZE(insn
->code
), BPF_READ
, -1, true);
3990 /* check whether atomic_add can write into the same memory */
3991 return check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
3992 BPF_SIZE(insn
->code
), BPF_WRITE
, -1, true);
3995 /* When register 'regno' is used to read the stack (either directly or through
3996 * a helper function) make sure that it's within stack boundary and, depending
3997 * on the access type, that all elements of the stack are initialized.
3999 * 'off' includes 'regno->off', but not its dynamic part (if any).
4001 * All registers that have been spilled on the stack in the slots within the
4002 * read offsets are marked as read.
4004 static int check_stack_range_initialized(
4005 struct bpf_verifier_env
*env
, int regno
, int off
,
4006 int access_size
, bool zero_size_allowed
,
4007 enum stack_access_src type
, struct bpf_call_arg_meta
*meta
)
4009 struct bpf_reg_state
*reg
= reg_state(env
, regno
);
4010 struct bpf_func_state
*state
= func(env
, reg
);
4011 int err
, min_off
, max_off
, i
, j
, slot
, spi
;
4012 char *err_extra
= type
== ACCESS_HELPER
? " indirect" : "";
4013 enum bpf_access_type bounds_check_type
;
4014 /* Some accesses can write anything into the stack, others are
4017 bool clobber
= false;
4019 if (access_size
== 0 && !zero_size_allowed
) {
4020 verbose(env
, "invalid zero-sized read\n");
4024 if (type
== ACCESS_HELPER
) {
4025 /* The bounds checks for writes are more permissive than for
4026 * reads. However, if raw_mode is not set, we'll do extra
4029 bounds_check_type
= BPF_WRITE
;
4032 bounds_check_type
= BPF_READ
;
4034 err
= check_stack_access_within_bounds(env
, regno
, off
, access_size
,
4035 type
, bounds_check_type
);
4040 if (tnum_is_const(reg
->var_off
)) {
4041 min_off
= max_off
= reg
->var_off
.value
+ off
;
4043 /* Variable offset is prohibited for unprivileged mode for
4044 * simplicity since it requires corresponding support in
4045 * Spectre masking for stack ALU.
4046 * See also retrieve_ptr_limit().
4048 if (!env
->bypass_spec_v1
) {
4051 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
4052 verbose(env
, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
4053 regno
, err_extra
, tn_buf
);
4056 /* Only initialized buffer on stack is allowed to be accessed
4057 * with variable offset. With uninitialized buffer it's hard to
4058 * guarantee that whole memory is marked as initialized on
4059 * helper return since specific bounds are unknown what may
4060 * cause uninitialized stack leaking.
4062 if (meta
&& meta
->raw_mode
)
4065 min_off
= reg
->smin_value
+ off
;
4066 max_off
= reg
->smax_value
+ off
;
4069 if (meta
&& meta
->raw_mode
) {
4070 meta
->access_size
= access_size
;
4071 meta
->regno
= regno
;
4075 for (i
= min_off
; i
< max_off
+ access_size
; i
++) {
4079 spi
= slot
/ BPF_REG_SIZE
;
4080 if (state
->allocated_stack
<= slot
)
4082 stype
= &state
->stack
[spi
].slot_type
[slot
% BPF_REG_SIZE
];
4083 if (*stype
== STACK_MISC
)
4085 if (*stype
== STACK_ZERO
) {
4087 /* helper can write anything into the stack */
4088 *stype
= STACK_MISC
;
4093 if (state
->stack
[spi
].slot_type
[0] == STACK_SPILL
&&
4094 state
->stack
[spi
].spilled_ptr
.type
== PTR_TO_BTF_ID
)
4097 if (state
->stack
[spi
].slot_type
[0] == STACK_SPILL
&&
4098 (state
->stack
[spi
].spilled_ptr
.type
== SCALAR_VALUE
||
4099 env
->allow_ptr_leaks
)) {
4101 __mark_reg_unknown(env
, &state
->stack
[spi
].spilled_ptr
);
4102 for (j
= 0; j
< BPF_REG_SIZE
; j
++)
4103 state
->stack
[spi
].slot_type
[j
] = STACK_MISC
;
4109 if (tnum_is_const(reg
->var_off
)) {
4110 verbose(env
, "invalid%s read from stack R%d off %d+%d size %d\n",
4111 err_extra
, regno
, min_off
, i
- min_off
, access_size
);
4115 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
4116 verbose(env
, "invalid%s read from stack R%d var_off %s+%d size %d\n",
4117 err_extra
, regno
, tn_buf
, i
- min_off
, access_size
);
4121 /* reading any byte out of 8-byte 'spill_slot' will cause
4122 * the whole slot to be marked as 'read'
4124 mark_reg_read(env
, &state
->stack
[spi
].spilled_ptr
,
4125 state
->stack
[spi
].spilled_ptr
.parent
,
4128 return update_stack_depth(env
, state
, min_off
);
4131 static int check_helper_mem_access(struct bpf_verifier_env
*env
, int regno
,
4132 int access_size
, bool zero_size_allowed
,
4133 struct bpf_call_arg_meta
*meta
)
4135 struct bpf_reg_state
*regs
= cur_regs(env
), *reg
= ®s
[regno
];
4137 switch (reg
->type
) {
4139 case PTR_TO_PACKET_META
:
4140 return check_packet_access(env
, regno
, reg
->off
, access_size
,
4142 case PTR_TO_MAP_VALUE
:
4143 if (check_map_access_type(env
, regno
, reg
->off
, access_size
,
4144 meta
&& meta
->raw_mode
? BPF_WRITE
:
4147 return check_map_access(env
, regno
, reg
->off
, access_size
,
4150 return check_mem_region_access(env
, regno
, reg
->off
,
4151 access_size
, reg
->mem_size
,
4153 case PTR_TO_RDONLY_BUF
:
4154 if (meta
&& meta
->raw_mode
)
4156 return check_buffer_access(env
, reg
, regno
, reg
->off
,
4157 access_size
, zero_size_allowed
,
4159 &env
->prog
->aux
->max_rdonly_access
);
4160 case PTR_TO_RDWR_BUF
:
4161 return check_buffer_access(env
, reg
, regno
, reg
->off
,
4162 access_size
, zero_size_allowed
,
4164 &env
->prog
->aux
->max_rdwr_access
);
4166 return check_stack_range_initialized(
4168 regno
, reg
->off
, access_size
,
4169 zero_size_allowed
, ACCESS_HELPER
, meta
);
4170 default: /* scalar_value or invalid ptr */
4171 /* Allow zero-byte read from NULL, regardless of pointer type */
4172 if (zero_size_allowed
&& access_size
== 0 &&
4173 register_is_null(reg
))
4176 verbose(env
, "R%d type=%s expected=%s\n", regno
,
4177 reg_type_str
[reg
->type
],
4178 reg_type_str
[PTR_TO_STACK
]);
4183 /* Implementation details:
4184 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
4185 * Two bpf_map_lookups (even with the same key) will have different reg->id.
4186 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
4187 * value_or_null->value transition, since the verifier only cares about
4188 * the range of access to valid map value pointer and doesn't care about actual
4189 * address of the map element.
4190 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
4191 * reg->id > 0 after value_or_null->value transition. By doing so
4192 * two bpf_map_lookups will be considered two different pointers that
4193 * point to different bpf_spin_locks.
4194 * The verifier allows taking only one bpf_spin_lock at a time to avoid
4196 * Since only one bpf_spin_lock is allowed the checks are simpler than
4197 * reg_is_refcounted() logic. The verifier needs to remember only
4198 * one spin_lock instead of array of acquired_refs.
4199 * cur_state->active_spin_lock remembers which map value element got locked
4200 * and clears it after bpf_spin_unlock.
4202 static int process_spin_lock(struct bpf_verifier_env
*env
, int regno
,
4205 struct bpf_reg_state
*regs
= cur_regs(env
), *reg
= ®s
[regno
];
4206 struct bpf_verifier_state
*cur
= env
->cur_state
;
4207 bool is_const
= tnum_is_const(reg
->var_off
);
4208 struct bpf_map
*map
= reg
->map_ptr
;
4209 u64 val
= reg
->var_off
.value
;
4213 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
4219 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
4223 if (!map_value_has_spin_lock(map
)) {
4224 if (map
->spin_lock_off
== -E2BIG
)
4226 "map '%s' has more than one 'struct bpf_spin_lock'\n",
4228 else if (map
->spin_lock_off
== -ENOENT
)
4230 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
4234 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
4238 if (map
->spin_lock_off
!= val
+ reg
->off
) {
4239 verbose(env
, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
4244 if (cur
->active_spin_lock
) {
4246 "Locking two bpf_spin_locks are not allowed\n");
4249 cur
->active_spin_lock
= reg
->id
;
4251 if (!cur
->active_spin_lock
) {
4252 verbose(env
, "bpf_spin_unlock without taking a lock\n");
4255 if (cur
->active_spin_lock
!= reg
->id
) {
4256 verbose(env
, "bpf_spin_unlock of different lock\n");
4259 cur
->active_spin_lock
= 0;
4264 static bool arg_type_is_mem_ptr(enum bpf_arg_type type
)
4266 return type
== ARG_PTR_TO_MEM
||
4267 type
== ARG_PTR_TO_MEM_OR_NULL
||
4268 type
== ARG_PTR_TO_UNINIT_MEM
;
4271 static bool arg_type_is_mem_size(enum bpf_arg_type type
)
4273 return type
== ARG_CONST_SIZE
||
4274 type
== ARG_CONST_SIZE_OR_ZERO
;
4277 static bool arg_type_is_alloc_size(enum bpf_arg_type type
)
4279 return type
== ARG_CONST_ALLOC_SIZE_OR_ZERO
;
4282 static bool arg_type_is_int_ptr(enum bpf_arg_type type
)
4284 return type
== ARG_PTR_TO_INT
||
4285 type
== ARG_PTR_TO_LONG
;
4288 static int int_ptr_type_to_size(enum bpf_arg_type type
)
4290 if (type
== ARG_PTR_TO_INT
)
4292 else if (type
== ARG_PTR_TO_LONG
)
4298 static int resolve_map_arg_type(struct bpf_verifier_env
*env
,
4299 const struct bpf_call_arg_meta
*meta
,
4300 enum bpf_arg_type
*arg_type
)
4302 if (!meta
->map_ptr
) {
4303 /* kernel subsystem misconfigured verifier */
4304 verbose(env
, "invalid map_ptr to access map->type\n");
4308 switch (meta
->map_ptr
->map_type
) {
4309 case BPF_MAP_TYPE_SOCKMAP
:
4310 case BPF_MAP_TYPE_SOCKHASH
:
4311 if (*arg_type
== ARG_PTR_TO_MAP_VALUE
) {
4312 *arg_type
= ARG_PTR_TO_BTF_ID_SOCK_COMMON
;
4314 verbose(env
, "invalid arg_type for sockmap/sockhash\n");
4325 struct bpf_reg_types
{
4326 const enum bpf_reg_type types
[10];
4330 static const struct bpf_reg_types map_key_value_types
= {
4339 static const struct bpf_reg_types sock_types
= {
4349 static const struct bpf_reg_types btf_id_sock_common_types
= {
4357 .btf_id
= &btf_sock_ids
[BTF_SOCK_TYPE_SOCK_COMMON
],
4361 static const struct bpf_reg_types mem_types
= {
4373 static const struct bpf_reg_types int_ptr_types
= {
4382 static const struct bpf_reg_types fullsock_types
= { .types
= { PTR_TO_SOCKET
} };
4383 static const struct bpf_reg_types scalar_types
= { .types
= { SCALAR_VALUE
} };
4384 static const struct bpf_reg_types context_types
= { .types
= { PTR_TO_CTX
} };
4385 static const struct bpf_reg_types alloc_mem_types
= { .types
= { PTR_TO_MEM
} };
4386 static const struct bpf_reg_types const_map_ptr_types
= { .types
= { CONST_PTR_TO_MAP
} };
4387 static const struct bpf_reg_types btf_ptr_types
= { .types
= { PTR_TO_BTF_ID
} };
4388 static const struct bpf_reg_types spin_lock_types
= { .types
= { PTR_TO_MAP_VALUE
} };
4389 static const struct bpf_reg_types percpu_btf_ptr_types
= { .types
= { PTR_TO_PERCPU_BTF_ID
} };
4391 static const struct bpf_reg_types
*compatible_reg_types
[__BPF_ARG_TYPE_MAX
] = {
4392 [ARG_PTR_TO_MAP_KEY
] = &map_key_value_types
,
4393 [ARG_PTR_TO_MAP_VALUE
] = &map_key_value_types
,
4394 [ARG_PTR_TO_UNINIT_MAP_VALUE
] = &map_key_value_types
,
4395 [ARG_PTR_TO_MAP_VALUE_OR_NULL
] = &map_key_value_types
,
4396 [ARG_CONST_SIZE
] = &scalar_types
,
4397 [ARG_CONST_SIZE_OR_ZERO
] = &scalar_types
,
4398 [ARG_CONST_ALLOC_SIZE_OR_ZERO
] = &scalar_types
,
4399 [ARG_CONST_MAP_PTR
] = &const_map_ptr_types
,
4400 [ARG_PTR_TO_CTX
] = &context_types
,
4401 [ARG_PTR_TO_CTX_OR_NULL
] = &context_types
,
4402 [ARG_PTR_TO_SOCK_COMMON
] = &sock_types
,
4404 [ARG_PTR_TO_BTF_ID_SOCK_COMMON
] = &btf_id_sock_common_types
,
4406 [ARG_PTR_TO_SOCKET
] = &fullsock_types
,
4407 [ARG_PTR_TO_SOCKET_OR_NULL
] = &fullsock_types
,
4408 [ARG_PTR_TO_BTF_ID
] = &btf_ptr_types
,
4409 [ARG_PTR_TO_SPIN_LOCK
] = &spin_lock_types
,
4410 [ARG_PTR_TO_MEM
] = &mem_types
,
4411 [ARG_PTR_TO_MEM_OR_NULL
] = &mem_types
,
4412 [ARG_PTR_TO_UNINIT_MEM
] = &mem_types
,
4413 [ARG_PTR_TO_ALLOC_MEM
] = &alloc_mem_types
,
4414 [ARG_PTR_TO_ALLOC_MEM_OR_NULL
] = &alloc_mem_types
,
4415 [ARG_PTR_TO_INT
] = &int_ptr_types
,
4416 [ARG_PTR_TO_LONG
] = &int_ptr_types
,
4417 [ARG_PTR_TO_PERCPU_BTF_ID
] = &percpu_btf_ptr_types
,
4420 static int check_reg_type(struct bpf_verifier_env
*env
, u32 regno
,
4421 enum bpf_arg_type arg_type
,
4422 const u32
*arg_btf_id
)
4424 struct bpf_reg_state
*regs
= cur_regs(env
), *reg
= ®s
[regno
];
4425 enum bpf_reg_type expected
, type
= reg
->type
;
4426 const struct bpf_reg_types
*compatible
;
4429 compatible
= compatible_reg_types
[arg_type
];
4431 verbose(env
, "verifier internal error: unsupported arg type %d\n", arg_type
);
4435 for (i
= 0; i
< ARRAY_SIZE(compatible
->types
); i
++) {
4436 expected
= compatible
->types
[i
];
4437 if (expected
== NOT_INIT
)
4440 if (type
== expected
)
4444 verbose(env
, "R%d type=%s expected=", regno
, reg_type_str
[type
]);
4445 for (j
= 0; j
+ 1 < i
; j
++)
4446 verbose(env
, "%s, ", reg_type_str
[compatible
->types
[j
]]);
4447 verbose(env
, "%s\n", reg_type_str
[compatible
->types
[j
]]);
4451 if (type
== PTR_TO_BTF_ID
) {
4453 if (!compatible
->btf_id
) {
4454 verbose(env
, "verifier internal error: missing arg compatible BTF ID\n");
4457 arg_btf_id
= compatible
->btf_id
;
4460 if (!btf_struct_ids_match(&env
->log
, reg
->btf
, reg
->btf_id
, reg
->off
,
4461 btf_vmlinux
, *arg_btf_id
)) {
4462 verbose(env
, "R%d is of type %s but %s is expected\n",
4463 regno
, kernel_type_name(reg
->btf
, reg
->btf_id
),
4464 kernel_type_name(btf_vmlinux
, *arg_btf_id
));
4468 if (!tnum_is_const(reg
->var_off
) || reg
->var_off
.value
) {
4469 verbose(env
, "R%d is a pointer to in-kernel struct with non-zero offset\n",
4478 static int check_func_arg(struct bpf_verifier_env
*env
, u32 arg
,
4479 struct bpf_call_arg_meta
*meta
,
4480 const struct bpf_func_proto
*fn
)
4482 u32 regno
= BPF_REG_1
+ arg
;
4483 struct bpf_reg_state
*regs
= cur_regs(env
), *reg
= ®s
[regno
];
4484 enum bpf_arg_type arg_type
= fn
->arg_type
[arg
];
4485 enum bpf_reg_type type
= reg
->type
;
4488 if (arg_type
== ARG_DONTCARE
)
4491 err
= check_reg_arg(env
, regno
, SRC_OP
);
4495 if (arg_type
== ARG_ANYTHING
) {
4496 if (is_pointer_value(env
, regno
)) {
4497 verbose(env
, "R%d leaks addr into helper function\n",
4504 if (type_is_pkt_pointer(type
) &&
4505 !may_access_direct_pkt_data(env
, meta
, BPF_READ
)) {
4506 verbose(env
, "helper access to the packet is not allowed\n");
4510 if (arg_type
== ARG_PTR_TO_MAP_VALUE
||
4511 arg_type
== ARG_PTR_TO_UNINIT_MAP_VALUE
||
4512 arg_type
== ARG_PTR_TO_MAP_VALUE_OR_NULL
) {
4513 err
= resolve_map_arg_type(env
, meta
, &arg_type
);
4518 if (register_is_null(reg
) && arg_type_may_be_null(arg_type
))
4519 /* A NULL register has a SCALAR_VALUE type, so skip
4522 goto skip_type_check
;
4524 err
= check_reg_type(env
, regno
, arg_type
, fn
->arg_btf_id
[arg
]);
4528 if (type
== PTR_TO_CTX
) {
4529 err
= check_ctx_reg(env
, reg
, regno
);
4535 if (reg
->ref_obj_id
) {
4536 if (meta
->ref_obj_id
) {
4537 verbose(env
, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
4538 regno
, reg
->ref_obj_id
,
4542 meta
->ref_obj_id
= reg
->ref_obj_id
;
4545 if (arg_type
== ARG_CONST_MAP_PTR
) {
4546 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
4547 meta
->map_ptr
= reg
->map_ptr
;
4548 } else if (arg_type
== ARG_PTR_TO_MAP_KEY
) {
4549 /* bpf_map_xxx(..., map_ptr, ..., key) call:
4550 * check that [key, key + map->key_size) are within
4551 * stack limits and initialized
4553 if (!meta
->map_ptr
) {
4554 /* in function declaration map_ptr must come before
4555 * map_key, so that it's verified and known before
4556 * we have to check map_key here. Otherwise it means
4557 * that kernel subsystem misconfigured verifier
4559 verbose(env
, "invalid map_ptr to access map->key\n");
4562 err
= check_helper_mem_access(env
, regno
,
4563 meta
->map_ptr
->key_size
, false,
4565 } else if (arg_type
== ARG_PTR_TO_MAP_VALUE
||
4566 (arg_type
== ARG_PTR_TO_MAP_VALUE_OR_NULL
&&
4567 !register_is_null(reg
)) ||
4568 arg_type
== ARG_PTR_TO_UNINIT_MAP_VALUE
) {
4569 /* bpf_map_xxx(..., map_ptr, ..., value) call:
4570 * check [value, value + map->value_size) validity
4572 if (!meta
->map_ptr
) {
4573 /* kernel subsystem misconfigured verifier */
4574 verbose(env
, "invalid map_ptr to access map->value\n");
4577 meta
->raw_mode
= (arg_type
== ARG_PTR_TO_UNINIT_MAP_VALUE
);
4578 err
= check_helper_mem_access(env
, regno
,
4579 meta
->map_ptr
->value_size
, false,
4581 } else if (arg_type
== ARG_PTR_TO_PERCPU_BTF_ID
) {
4583 verbose(env
, "Helper has invalid btf_id in R%d\n", regno
);
4586 meta
->ret_btf
= reg
->btf
;
4587 meta
->ret_btf_id
= reg
->btf_id
;
4588 } else if (arg_type
== ARG_PTR_TO_SPIN_LOCK
) {
4589 if (meta
->func_id
== BPF_FUNC_spin_lock
) {
4590 if (process_spin_lock(env
, regno
, true))
4592 } else if (meta
->func_id
== BPF_FUNC_spin_unlock
) {
4593 if (process_spin_lock(env
, regno
, false))
4596 verbose(env
, "verifier internal error\n");
4599 } else if (arg_type_is_mem_ptr(arg_type
)) {
4600 /* The access to this pointer is only checked when we hit the
4601 * next is_mem_size argument below.
4603 meta
->raw_mode
= (arg_type
== ARG_PTR_TO_UNINIT_MEM
);
4604 } else if (arg_type_is_mem_size(arg_type
)) {
4605 bool zero_size_allowed
= (arg_type
== ARG_CONST_SIZE_OR_ZERO
);
4607 /* This is used to refine r0 return value bounds for helpers
4608 * that enforce this value as an upper bound on return values.
4609 * See do_refine_retval_range() for helpers that can refine
4610 * the return value. C type of helper is u32 so we pull register
4611 * bound from umax_value however, if negative verifier errors
4612 * out. Only upper bounds can be learned because retval is an
4613 * int type and negative retvals are allowed.
4615 meta
->msize_max_value
= reg
->umax_value
;
4617 /* The register is SCALAR_VALUE; the access check
4618 * happens using its boundaries.
4620 if (!tnum_is_const(reg
->var_off
))
4621 /* For unprivileged variable accesses, disable raw
4622 * mode so that the program is required to
4623 * initialize all the memory that the helper could
4624 * just partially fill up.
4628 if (reg
->smin_value
< 0) {
4629 verbose(env
, "R%d min value is negative, either use unsigned or 'var &= const'\n",
4634 if (reg
->umin_value
== 0) {
4635 err
= check_helper_mem_access(env
, regno
- 1, 0,
4642 if (reg
->umax_value
>= BPF_MAX_VAR_SIZ
) {
4643 verbose(env
, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
4647 err
= check_helper_mem_access(env
, regno
- 1,
4649 zero_size_allowed
, meta
);
4651 err
= mark_chain_precision(env
, regno
);
4652 } else if (arg_type_is_alloc_size(arg_type
)) {
4653 if (!tnum_is_const(reg
->var_off
)) {
4654 verbose(env
, "R%d unbounded size, use 'var &= const' or 'if (var < const)'\n",
4658 meta
->mem_size
= reg
->var_off
.value
;
4659 } else if (arg_type_is_int_ptr(arg_type
)) {
4660 int size
= int_ptr_type_to_size(arg_type
);
4662 err
= check_helper_mem_access(env
, regno
, size
, false, meta
);
4665 err
= check_ptr_alignment(env
, reg
, 0, size
, true);
4671 static bool may_update_sockmap(struct bpf_verifier_env
*env
, int func_id
)
4673 enum bpf_attach_type eatype
= env
->prog
->expected_attach_type
;
4674 enum bpf_prog_type type
= resolve_prog_type(env
->prog
);
4676 if (func_id
!= BPF_FUNC_map_update_elem
)
4679 /* It's not possible to get access to a locked struct sock in these
4680 * contexts, so updating is safe.
4683 case BPF_PROG_TYPE_TRACING
:
4684 if (eatype
== BPF_TRACE_ITER
)
4687 case BPF_PROG_TYPE_SOCKET_FILTER
:
4688 case BPF_PROG_TYPE_SCHED_CLS
:
4689 case BPF_PROG_TYPE_SCHED_ACT
:
4690 case BPF_PROG_TYPE_XDP
:
4691 case BPF_PROG_TYPE_SK_REUSEPORT
:
4692 case BPF_PROG_TYPE_FLOW_DISSECTOR
:
4693 case BPF_PROG_TYPE_SK_LOOKUP
:
4699 verbose(env
, "cannot update sockmap in this context\n");
4703 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env
*env
)
4705 return env
->prog
->jit_requested
&& IS_ENABLED(CONFIG_X86_64
);
4708 static int check_map_func_compatibility(struct bpf_verifier_env
*env
,
4709 struct bpf_map
*map
, int func_id
)
4714 /* We need a two way check, first is from map perspective ... */
4715 switch (map
->map_type
) {
4716 case BPF_MAP_TYPE_PROG_ARRAY
:
4717 if (func_id
!= BPF_FUNC_tail_call
)
4720 case BPF_MAP_TYPE_PERF_EVENT_ARRAY
:
4721 if (func_id
!= BPF_FUNC_perf_event_read
&&
4722 func_id
!= BPF_FUNC_perf_event_output
&&
4723 func_id
!= BPF_FUNC_skb_output
&&
4724 func_id
!= BPF_FUNC_perf_event_read_value
&&
4725 func_id
!= BPF_FUNC_xdp_output
)
4728 case BPF_MAP_TYPE_RINGBUF
:
4729 if (func_id
!= BPF_FUNC_ringbuf_output
&&
4730 func_id
!= BPF_FUNC_ringbuf_reserve
&&
4731 func_id
!= BPF_FUNC_ringbuf_submit
&&
4732 func_id
!= BPF_FUNC_ringbuf_discard
&&
4733 func_id
!= BPF_FUNC_ringbuf_query
)
4736 case BPF_MAP_TYPE_STACK_TRACE
:
4737 if (func_id
!= BPF_FUNC_get_stackid
)
4740 case BPF_MAP_TYPE_CGROUP_ARRAY
:
4741 if (func_id
!= BPF_FUNC_skb_under_cgroup
&&
4742 func_id
!= BPF_FUNC_current_task_under_cgroup
)
4745 case BPF_MAP_TYPE_CGROUP_STORAGE
:
4746 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE
:
4747 if (func_id
!= BPF_FUNC_get_local_storage
)
4750 case BPF_MAP_TYPE_DEVMAP
:
4751 case BPF_MAP_TYPE_DEVMAP_HASH
:
4752 if (func_id
!= BPF_FUNC_redirect_map
&&
4753 func_id
!= BPF_FUNC_map_lookup_elem
)
4756 /* Restrict bpf side of cpumap and xskmap, open when use-cases
4759 case BPF_MAP_TYPE_CPUMAP
:
4760 if (func_id
!= BPF_FUNC_redirect_map
)
4763 case BPF_MAP_TYPE_XSKMAP
:
4764 if (func_id
!= BPF_FUNC_redirect_map
&&
4765 func_id
!= BPF_FUNC_map_lookup_elem
)
4768 case BPF_MAP_TYPE_ARRAY_OF_MAPS
:
4769 case BPF_MAP_TYPE_HASH_OF_MAPS
:
4770 if (func_id
!= BPF_FUNC_map_lookup_elem
)
4773 case BPF_MAP_TYPE_SOCKMAP
:
4774 if (func_id
!= BPF_FUNC_sk_redirect_map
&&
4775 func_id
!= BPF_FUNC_sock_map_update
&&
4776 func_id
!= BPF_FUNC_map_delete_elem
&&
4777 func_id
!= BPF_FUNC_msg_redirect_map
&&
4778 func_id
!= BPF_FUNC_sk_select_reuseport
&&
4779 func_id
!= BPF_FUNC_map_lookup_elem
&&
4780 !may_update_sockmap(env
, func_id
))
4783 case BPF_MAP_TYPE_SOCKHASH
:
4784 if (func_id
!= BPF_FUNC_sk_redirect_hash
&&
4785 func_id
!= BPF_FUNC_sock_hash_update
&&
4786 func_id
!= BPF_FUNC_map_delete_elem
&&
4787 func_id
!= BPF_FUNC_msg_redirect_hash
&&
4788 func_id
!= BPF_FUNC_sk_select_reuseport
&&
4789 func_id
!= BPF_FUNC_map_lookup_elem
&&
4790 !may_update_sockmap(env
, func_id
))
4793 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY
:
4794 if (func_id
!= BPF_FUNC_sk_select_reuseport
)
4797 case BPF_MAP_TYPE_QUEUE
:
4798 case BPF_MAP_TYPE_STACK
:
4799 if (func_id
!= BPF_FUNC_map_peek_elem
&&
4800 func_id
!= BPF_FUNC_map_pop_elem
&&
4801 func_id
!= BPF_FUNC_map_push_elem
)
4804 case BPF_MAP_TYPE_SK_STORAGE
:
4805 if (func_id
!= BPF_FUNC_sk_storage_get
&&
4806 func_id
!= BPF_FUNC_sk_storage_delete
)
4809 case BPF_MAP_TYPE_INODE_STORAGE
:
4810 if (func_id
!= BPF_FUNC_inode_storage_get
&&
4811 func_id
!= BPF_FUNC_inode_storage_delete
)
4814 case BPF_MAP_TYPE_TASK_STORAGE
:
4815 if (func_id
!= BPF_FUNC_task_storage_get
&&
4816 func_id
!= BPF_FUNC_task_storage_delete
)
4823 /* ... and second from the function itself. */
4825 case BPF_FUNC_tail_call
:
4826 if (map
->map_type
!= BPF_MAP_TYPE_PROG_ARRAY
)
4828 if (env
->subprog_cnt
> 1 && !allow_tail_call_in_subprogs(env
)) {
4829 verbose(env
, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
4833 case BPF_FUNC_perf_event_read
:
4834 case BPF_FUNC_perf_event_output
:
4835 case BPF_FUNC_perf_event_read_value
:
4836 case BPF_FUNC_skb_output
:
4837 case BPF_FUNC_xdp_output
:
4838 if (map
->map_type
!= BPF_MAP_TYPE_PERF_EVENT_ARRAY
)
4841 case BPF_FUNC_get_stackid
:
4842 if (map
->map_type
!= BPF_MAP_TYPE_STACK_TRACE
)
4845 case BPF_FUNC_current_task_under_cgroup
:
4846 case BPF_FUNC_skb_under_cgroup
:
4847 if (map
->map_type
!= BPF_MAP_TYPE_CGROUP_ARRAY
)
4850 case BPF_FUNC_redirect_map
:
4851 if (map
->map_type
!= BPF_MAP_TYPE_DEVMAP
&&
4852 map
->map_type
!= BPF_MAP_TYPE_DEVMAP_HASH
&&
4853 map
->map_type
!= BPF_MAP_TYPE_CPUMAP
&&
4854 map
->map_type
!= BPF_MAP_TYPE_XSKMAP
)
4857 case BPF_FUNC_sk_redirect_map
:
4858 case BPF_FUNC_msg_redirect_map
:
4859 case BPF_FUNC_sock_map_update
:
4860 if (map
->map_type
!= BPF_MAP_TYPE_SOCKMAP
)
4863 case BPF_FUNC_sk_redirect_hash
:
4864 case BPF_FUNC_msg_redirect_hash
:
4865 case BPF_FUNC_sock_hash_update
:
4866 if (map
->map_type
!= BPF_MAP_TYPE_SOCKHASH
)
4869 case BPF_FUNC_get_local_storage
:
4870 if (map
->map_type
!= BPF_MAP_TYPE_CGROUP_STORAGE
&&
4871 map
->map_type
!= BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE
)
4874 case BPF_FUNC_sk_select_reuseport
:
4875 if (map
->map_type
!= BPF_MAP_TYPE_REUSEPORT_SOCKARRAY
&&
4876 map
->map_type
!= BPF_MAP_TYPE_SOCKMAP
&&
4877 map
->map_type
!= BPF_MAP_TYPE_SOCKHASH
)
4880 case BPF_FUNC_map_peek_elem
:
4881 case BPF_FUNC_map_pop_elem
:
4882 case BPF_FUNC_map_push_elem
:
4883 if (map
->map_type
!= BPF_MAP_TYPE_QUEUE
&&
4884 map
->map_type
!= BPF_MAP_TYPE_STACK
)
4887 case BPF_FUNC_sk_storage_get
:
4888 case BPF_FUNC_sk_storage_delete
:
4889 if (map
->map_type
!= BPF_MAP_TYPE_SK_STORAGE
)
4892 case BPF_FUNC_inode_storage_get
:
4893 case BPF_FUNC_inode_storage_delete
:
4894 if (map
->map_type
!= BPF_MAP_TYPE_INODE_STORAGE
)
4897 case BPF_FUNC_task_storage_get
:
4898 case BPF_FUNC_task_storage_delete
:
4899 if (map
->map_type
!= BPF_MAP_TYPE_TASK_STORAGE
)
4908 verbose(env
, "cannot pass map_type %d into func %s#%d\n",
4909 map
->map_type
, func_id_name(func_id
), func_id
);
4913 static bool check_raw_mode_ok(const struct bpf_func_proto
*fn
)
4917 if (fn
->arg1_type
== ARG_PTR_TO_UNINIT_MEM
)
4919 if (fn
->arg2_type
== ARG_PTR_TO_UNINIT_MEM
)
4921 if (fn
->arg3_type
== ARG_PTR_TO_UNINIT_MEM
)
4923 if (fn
->arg4_type
== ARG_PTR_TO_UNINIT_MEM
)
4925 if (fn
->arg5_type
== ARG_PTR_TO_UNINIT_MEM
)
4928 /* We only support one arg being in raw mode at the moment,
4929 * which is sufficient for the helper functions we have
4935 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr
,
4936 enum bpf_arg_type arg_next
)
4938 return (arg_type_is_mem_ptr(arg_curr
) &&
4939 !arg_type_is_mem_size(arg_next
)) ||
4940 (!arg_type_is_mem_ptr(arg_curr
) &&
4941 arg_type_is_mem_size(arg_next
));
4944 static bool check_arg_pair_ok(const struct bpf_func_proto
*fn
)
4946 /* bpf_xxx(..., buf, len) call will access 'len'
4947 * bytes from memory 'buf'. Both arg types need
4948 * to be paired, so make sure there's no buggy
4949 * helper function specification.
4951 if (arg_type_is_mem_size(fn
->arg1_type
) ||
4952 arg_type_is_mem_ptr(fn
->arg5_type
) ||
4953 check_args_pair_invalid(fn
->arg1_type
, fn
->arg2_type
) ||
4954 check_args_pair_invalid(fn
->arg2_type
, fn
->arg3_type
) ||
4955 check_args_pair_invalid(fn
->arg3_type
, fn
->arg4_type
) ||
4956 check_args_pair_invalid(fn
->arg4_type
, fn
->arg5_type
))
4962 static bool check_refcount_ok(const struct bpf_func_proto
*fn
, int func_id
)
4966 if (arg_type_may_be_refcounted(fn
->arg1_type
))
4968 if (arg_type_may_be_refcounted(fn
->arg2_type
))
4970 if (arg_type_may_be_refcounted(fn
->arg3_type
))
4972 if (arg_type_may_be_refcounted(fn
->arg4_type
))
4974 if (arg_type_may_be_refcounted(fn
->arg5_type
))
4977 /* A reference acquiring function cannot acquire
4978 * another refcounted ptr.
4980 if (may_be_acquire_function(func_id
) && count
)
4983 /* We only support one arg being unreferenced at the moment,
4984 * which is sufficient for the helper functions we have right now.
4989 static bool check_btf_id_ok(const struct bpf_func_proto
*fn
)
4993 for (i
= 0; i
< ARRAY_SIZE(fn
->arg_type
); i
++) {
4994 if (fn
->arg_type
[i
] == ARG_PTR_TO_BTF_ID
&& !fn
->arg_btf_id
[i
])
4997 if (fn
->arg_type
[i
] != ARG_PTR_TO_BTF_ID
&& fn
->arg_btf_id
[i
])
5004 static int check_func_proto(const struct bpf_func_proto
*fn
, int func_id
)
5006 return check_raw_mode_ok(fn
) &&
5007 check_arg_pair_ok(fn
) &&
5008 check_btf_id_ok(fn
) &&
5009 check_refcount_ok(fn
, func_id
) ? 0 : -EINVAL
;
5012 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
5013 * are now invalid, so turn them into unknown SCALAR_VALUE.
5015 static void __clear_all_pkt_pointers(struct bpf_verifier_env
*env
,
5016 struct bpf_func_state
*state
)
5018 struct bpf_reg_state
*regs
= state
->regs
, *reg
;
5021 for (i
= 0; i
< MAX_BPF_REG
; i
++)
5022 if (reg_is_pkt_pointer_any(®s
[i
]))
5023 mark_reg_unknown(env
, regs
, i
);
5025 bpf_for_each_spilled_reg(i
, state
, reg
) {
5028 if (reg_is_pkt_pointer_any(reg
))
5029 __mark_reg_unknown(env
, reg
);
5033 static void clear_all_pkt_pointers(struct bpf_verifier_env
*env
)
5035 struct bpf_verifier_state
*vstate
= env
->cur_state
;
5038 for (i
= 0; i
<= vstate
->curframe
; i
++)
5039 __clear_all_pkt_pointers(env
, vstate
->frame
[i
]);
5044 BEYOND_PKT_END
= -2,
5047 static void mark_pkt_end(struct bpf_verifier_state
*vstate
, int regn
, bool range_open
)
5049 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
5050 struct bpf_reg_state
*reg
= &state
->regs
[regn
];
5052 if (reg
->type
!= PTR_TO_PACKET
)
5053 /* PTR_TO_PACKET_META is not supported yet */
5056 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
5057 * How far beyond pkt_end it goes is unknown.
5058 * if (!range_open) it's the case of pkt >= pkt_end
5059 * if (range_open) it's the case of pkt > pkt_end
5060 * hence this pointer is at least 1 byte bigger than pkt_end
5063 reg
->range
= BEYOND_PKT_END
;
5065 reg
->range
= AT_PKT_END
;
5068 static void release_reg_references(struct bpf_verifier_env
*env
,
5069 struct bpf_func_state
*state
,
5072 struct bpf_reg_state
*regs
= state
->regs
, *reg
;
5075 for (i
= 0; i
< MAX_BPF_REG
; i
++)
5076 if (regs
[i
].ref_obj_id
== ref_obj_id
)
5077 mark_reg_unknown(env
, regs
, i
);
5079 bpf_for_each_spilled_reg(i
, state
, reg
) {
5082 if (reg
->ref_obj_id
== ref_obj_id
)
5083 __mark_reg_unknown(env
, reg
);
5087 /* The pointer with the specified id has released its reference to kernel
5088 * resources. Identify all copies of the same pointer and clear the reference.
5090 static int release_reference(struct bpf_verifier_env
*env
,
5093 struct bpf_verifier_state
*vstate
= env
->cur_state
;
5097 err
= release_reference_state(cur_func(env
), ref_obj_id
);
5101 for (i
= 0; i
<= vstate
->curframe
; i
++)
5102 release_reg_references(env
, vstate
->frame
[i
], ref_obj_id
);
5107 static void clear_caller_saved_regs(struct bpf_verifier_env
*env
,
5108 struct bpf_reg_state
*regs
)
5112 /* after the call registers r0 - r5 were scratched */
5113 for (i
= 0; i
< CALLER_SAVED_REGS
; i
++) {
5114 mark_reg_not_init(env
, regs
, caller_saved
[i
]);
5115 check_reg_arg(env
, caller_saved
[i
], DST_OP_NO_MARK
);
5119 static int check_func_call(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
,
5122 struct bpf_verifier_state
*state
= env
->cur_state
;
5123 struct bpf_func_info_aux
*func_info_aux
;
5124 struct bpf_func_state
*caller
, *callee
;
5125 int i
, err
, subprog
, target_insn
;
5126 bool is_global
= false;
5128 if (state
->curframe
+ 1 >= MAX_CALL_FRAMES
) {
5129 verbose(env
, "the call stack of %d frames is too deep\n",
5130 state
->curframe
+ 2);
5134 target_insn
= *insn_idx
+ insn
->imm
;
5135 subprog
= find_subprog(env
, target_insn
+ 1);
5137 verbose(env
, "verifier bug. No program starts at insn %d\n",
5142 caller
= state
->frame
[state
->curframe
];
5143 if (state
->frame
[state
->curframe
+ 1]) {
5144 verbose(env
, "verifier bug. Frame %d already allocated\n",
5145 state
->curframe
+ 1);
5149 func_info_aux
= env
->prog
->aux
->func_info_aux
;
5151 is_global
= func_info_aux
[subprog
].linkage
== BTF_FUNC_GLOBAL
;
5152 err
= btf_check_func_arg_match(env
, subprog
, caller
->regs
);
5157 verbose(env
, "Caller passes invalid args into func#%d\n",
5161 if (env
->log
.level
& BPF_LOG_LEVEL
)
5163 "Func#%d is global and valid. Skipping.\n",
5165 clear_caller_saved_regs(env
, caller
->regs
);
5167 /* All global functions return a 64-bit SCALAR_VALUE */
5168 mark_reg_unknown(env
, caller
->regs
, BPF_REG_0
);
5169 caller
->regs
[BPF_REG_0
].subreg_def
= DEF_NOT_SUBREG
;
5171 /* continue with next insn after call */
5176 callee
= kzalloc(sizeof(*callee
), GFP_KERNEL
);
5179 state
->frame
[state
->curframe
+ 1] = callee
;
5181 /* callee cannot access r0, r6 - r9 for reading and has to write
5182 * into its own stack before reading from it.
5183 * callee can read/write into caller's stack
5185 init_func_state(env
, callee
,
5186 /* remember the callsite, it will be used by bpf_exit */
5187 *insn_idx
/* callsite */,
5188 state
->curframe
+ 1 /* frameno within this callchain */,
5189 subprog
/* subprog number within this prog */);
5191 /* Transfer references to the callee */
5192 err
= transfer_reference_state(callee
, caller
);
5196 /* copy r1 - r5 args that callee can access. The copy includes parent
5197 * pointers, which connects us up to the liveness chain
5199 for (i
= BPF_REG_1
; i
<= BPF_REG_5
; i
++)
5200 callee
->regs
[i
] = caller
->regs
[i
];
5202 clear_caller_saved_regs(env
, caller
->regs
);
5204 /* only increment it after check_reg_arg() finished */
5207 /* and go analyze first insn of the callee */
5208 *insn_idx
= target_insn
;
5210 if (env
->log
.level
& BPF_LOG_LEVEL
) {
5211 verbose(env
, "caller:\n");
5212 print_verifier_state(env
, caller
);
5213 verbose(env
, "callee:\n");
5214 print_verifier_state(env
, callee
);
5219 static int prepare_func_exit(struct bpf_verifier_env
*env
, int *insn_idx
)
5221 struct bpf_verifier_state
*state
= env
->cur_state
;
5222 struct bpf_func_state
*caller
, *callee
;
5223 struct bpf_reg_state
*r0
;
5226 callee
= state
->frame
[state
->curframe
];
5227 r0
= &callee
->regs
[BPF_REG_0
];
5228 if (r0
->type
== PTR_TO_STACK
) {
5229 /* technically it's ok to return caller's stack pointer
5230 * (or caller's caller's pointer) back to the caller,
5231 * since these pointers are valid. Only current stack
5232 * pointer will be invalid as soon as function exits,
5233 * but let's be conservative
5235 verbose(env
, "cannot return stack pointer to the caller\n");
5240 caller
= state
->frame
[state
->curframe
];
5241 /* return to the caller whatever r0 had in the callee */
5242 caller
->regs
[BPF_REG_0
] = *r0
;
5244 /* Transfer references to the caller */
5245 err
= transfer_reference_state(caller
, callee
);
5249 *insn_idx
= callee
->callsite
+ 1;
5250 if (env
->log
.level
& BPF_LOG_LEVEL
) {
5251 verbose(env
, "returning from callee:\n");
5252 print_verifier_state(env
, callee
);
5253 verbose(env
, "to caller at %d:\n", *insn_idx
);
5254 print_verifier_state(env
, caller
);
5256 /* clear everything in the callee */
5257 free_func_state(callee
);
5258 state
->frame
[state
->curframe
+ 1] = NULL
;
5262 static void do_refine_retval_range(struct bpf_reg_state
*regs
, int ret_type
,
5264 struct bpf_call_arg_meta
*meta
)
5266 struct bpf_reg_state
*ret_reg
= ®s
[BPF_REG_0
];
5268 if (ret_type
!= RET_INTEGER
||
5269 (func_id
!= BPF_FUNC_get_stack
&&
5270 func_id
!= BPF_FUNC_probe_read_str
&&
5271 func_id
!= BPF_FUNC_probe_read_kernel_str
&&
5272 func_id
!= BPF_FUNC_probe_read_user_str
))
5275 ret_reg
->smax_value
= meta
->msize_max_value
;
5276 ret_reg
->s32_max_value
= meta
->msize_max_value
;
5277 ret_reg
->smin_value
= -MAX_ERRNO
;
5278 ret_reg
->s32_min_value
= -MAX_ERRNO
;
5279 __reg_deduce_bounds(ret_reg
);
5280 __reg_bound_offset(ret_reg
);
5281 __update_reg_bounds(ret_reg
);
5285 record_func_map(struct bpf_verifier_env
*env
, struct bpf_call_arg_meta
*meta
,
5286 int func_id
, int insn_idx
)
5288 struct bpf_insn_aux_data
*aux
= &env
->insn_aux_data
[insn_idx
];
5289 struct bpf_map
*map
= meta
->map_ptr
;
5291 if (func_id
!= BPF_FUNC_tail_call
&&
5292 func_id
!= BPF_FUNC_map_lookup_elem
&&
5293 func_id
!= BPF_FUNC_map_update_elem
&&
5294 func_id
!= BPF_FUNC_map_delete_elem
&&
5295 func_id
!= BPF_FUNC_map_push_elem
&&
5296 func_id
!= BPF_FUNC_map_pop_elem
&&
5297 func_id
!= BPF_FUNC_map_peek_elem
)
5301 verbose(env
, "kernel subsystem misconfigured verifier\n");
5305 /* In case of read-only, some additional restrictions
5306 * need to be applied in order to prevent altering the
5307 * state of the map from program side.
5309 if ((map
->map_flags
& BPF_F_RDONLY_PROG
) &&
5310 (func_id
== BPF_FUNC_map_delete_elem
||
5311 func_id
== BPF_FUNC_map_update_elem
||
5312 func_id
== BPF_FUNC_map_push_elem
||
5313 func_id
== BPF_FUNC_map_pop_elem
)) {
5314 verbose(env
, "write into map forbidden\n");
5318 if (!BPF_MAP_PTR(aux
->map_ptr_state
))
5319 bpf_map_ptr_store(aux
, meta
->map_ptr
,
5320 !meta
->map_ptr
->bypass_spec_v1
);
5321 else if (BPF_MAP_PTR(aux
->map_ptr_state
) != meta
->map_ptr
)
5322 bpf_map_ptr_store(aux
, BPF_MAP_PTR_POISON
,
5323 !meta
->map_ptr
->bypass_spec_v1
);
5328 record_func_key(struct bpf_verifier_env
*env
, struct bpf_call_arg_meta
*meta
,
5329 int func_id
, int insn_idx
)
5331 struct bpf_insn_aux_data
*aux
= &env
->insn_aux_data
[insn_idx
];
5332 struct bpf_reg_state
*regs
= cur_regs(env
), *reg
;
5333 struct bpf_map
*map
= meta
->map_ptr
;
5338 if (func_id
!= BPF_FUNC_tail_call
)
5340 if (!map
|| map
->map_type
!= BPF_MAP_TYPE_PROG_ARRAY
) {
5341 verbose(env
, "kernel subsystem misconfigured verifier\n");
5345 range
= tnum_range(0, map
->max_entries
- 1);
5346 reg
= ®s
[BPF_REG_3
];
5348 if (!register_is_const(reg
) || !tnum_in(range
, reg
->var_off
)) {
5349 bpf_map_key_store(aux
, BPF_MAP_KEY_POISON
);
5353 err
= mark_chain_precision(env
, BPF_REG_3
);
5357 val
= reg
->var_off
.value
;
5358 if (bpf_map_key_unseen(aux
))
5359 bpf_map_key_store(aux
, val
);
5360 else if (!bpf_map_key_poisoned(aux
) &&
5361 bpf_map_key_immediate(aux
) != val
)
5362 bpf_map_key_store(aux
, BPF_MAP_KEY_POISON
);
5366 static int check_reference_leak(struct bpf_verifier_env
*env
)
5368 struct bpf_func_state
*state
= cur_func(env
);
5371 for (i
= 0; i
< state
->acquired_refs
; i
++) {
5372 verbose(env
, "Unreleased reference id=%d alloc_insn=%d\n",
5373 state
->refs
[i
].id
, state
->refs
[i
].insn_idx
);
5375 return state
->acquired_refs
? -EINVAL
: 0;
5378 static int check_helper_call(struct bpf_verifier_env
*env
, int func_id
, int insn_idx
)
5380 const struct bpf_func_proto
*fn
= NULL
;
5381 struct bpf_reg_state
*regs
;
5382 struct bpf_call_arg_meta meta
;
5386 /* find function prototype */
5387 if (func_id
< 0 || func_id
>= __BPF_FUNC_MAX_ID
) {
5388 verbose(env
, "invalid func %s#%d\n", func_id_name(func_id
),
5393 if (env
->ops
->get_func_proto
)
5394 fn
= env
->ops
->get_func_proto(func_id
, env
->prog
);
5396 verbose(env
, "unknown func %s#%d\n", func_id_name(func_id
),
5401 /* eBPF programs must be GPL compatible to use GPL-ed functions */
5402 if (!env
->prog
->gpl_compatible
&& fn
->gpl_only
) {
5403 verbose(env
, "cannot call GPL-restricted function from non-GPL compatible program\n");
5407 if (fn
->allowed
&& !fn
->allowed(env
->prog
)) {
5408 verbose(env
, "helper call is not allowed in probe\n");
5412 /* With LD_ABS/IND some JITs save/restore skb from r1. */
5413 changes_data
= bpf_helper_changes_pkt_data(fn
->func
);
5414 if (changes_data
&& fn
->arg1_type
!= ARG_PTR_TO_CTX
) {
5415 verbose(env
, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
5416 func_id_name(func_id
), func_id
);
5420 memset(&meta
, 0, sizeof(meta
));
5421 meta
.pkt_access
= fn
->pkt_access
;
5423 err
= check_func_proto(fn
, func_id
);
5425 verbose(env
, "kernel subsystem misconfigured func %s#%d\n",
5426 func_id_name(func_id
), func_id
);
5430 meta
.func_id
= func_id
;
5432 for (i
= 0; i
< 5; i
++) {
5433 err
= check_func_arg(env
, i
, &meta
, fn
);
5438 err
= record_func_map(env
, &meta
, func_id
, insn_idx
);
5442 err
= record_func_key(env
, &meta
, func_id
, insn_idx
);
5446 /* Mark slots with STACK_MISC in case of raw mode, stack offset
5447 * is inferred from register state.
5449 for (i
= 0; i
< meta
.access_size
; i
++) {
5450 err
= check_mem_access(env
, insn_idx
, meta
.regno
, i
, BPF_B
,
5451 BPF_WRITE
, -1, false);
5456 if (func_id
== BPF_FUNC_tail_call
) {
5457 err
= check_reference_leak(env
);
5459 verbose(env
, "tail_call would lead to reference leak\n");
5462 } else if (is_release_function(func_id
)) {
5463 err
= release_reference(env
, meta
.ref_obj_id
);
5465 verbose(env
, "func %s#%d reference has not been acquired before\n",
5466 func_id_name(func_id
), func_id
);
5471 regs
= cur_regs(env
);
5473 /* check that flags argument in get_local_storage(map, flags) is 0,
5474 * this is required because get_local_storage() can't return an error.
5476 if (func_id
== BPF_FUNC_get_local_storage
&&
5477 !register_is_null(®s
[BPF_REG_2
])) {
5478 verbose(env
, "get_local_storage() doesn't support non-zero flags\n");
5482 /* reset caller saved regs */
5483 for (i
= 0; i
< CALLER_SAVED_REGS
; i
++) {
5484 mark_reg_not_init(env
, regs
, caller_saved
[i
]);
5485 check_reg_arg(env
, caller_saved
[i
], DST_OP_NO_MARK
);
5488 /* helper call returns 64-bit value. */
5489 regs
[BPF_REG_0
].subreg_def
= DEF_NOT_SUBREG
;
5491 /* update return register (already marked as written above) */
5492 if (fn
->ret_type
== RET_INTEGER
) {
5493 /* sets type to SCALAR_VALUE */
5494 mark_reg_unknown(env
, regs
, BPF_REG_0
);
5495 } else if (fn
->ret_type
== RET_VOID
) {
5496 regs
[BPF_REG_0
].type
= NOT_INIT
;
5497 } else if (fn
->ret_type
== RET_PTR_TO_MAP_VALUE_OR_NULL
||
5498 fn
->ret_type
== RET_PTR_TO_MAP_VALUE
) {
5499 /* There is no offset yet applied, variable or fixed */
5500 mark_reg_known_zero(env
, regs
, BPF_REG_0
);
5501 /* remember map_ptr, so that check_map_access()
5502 * can check 'value_size' boundary of memory access
5503 * to map element returned from bpf_map_lookup_elem()
5505 if (meta
.map_ptr
== NULL
) {
5507 "kernel subsystem misconfigured verifier\n");
5510 regs
[BPF_REG_0
].map_ptr
= meta
.map_ptr
;
5511 if (fn
->ret_type
== RET_PTR_TO_MAP_VALUE
) {
5512 regs
[BPF_REG_0
].type
= PTR_TO_MAP_VALUE
;
5513 if (map_value_has_spin_lock(meta
.map_ptr
))
5514 regs
[BPF_REG_0
].id
= ++env
->id_gen
;
5516 regs
[BPF_REG_0
].type
= PTR_TO_MAP_VALUE_OR_NULL
;
5518 } else if (fn
->ret_type
== RET_PTR_TO_SOCKET_OR_NULL
) {
5519 mark_reg_known_zero(env
, regs
, BPF_REG_0
);
5520 regs
[BPF_REG_0
].type
= PTR_TO_SOCKET_OR_NULL
;
5521 } else if (fn
->ret_type
== RET_PTR_TO_SOCK_COMMON_OR_NULL
) {
5522 mark_reg_known_zero(env
, regs
, BPF_REG_0
);
5523 regs
[BPF_REG_0
].type
= PTR_TO_SOCK_COMMON_OR_NULL
;
5524 } else if (fn
->ret_type
== RET_PTR_TO_TCP_SOCK_OR_NULL
) {
5525 mark_reg_known_zero(env
, regs
, BPF_REG_0
);
5526 regs
[BPF_REG_0
].type
= PTR_TO_TCP_SOCK_OR_NULL
;
5527 } else if (fn
->ret_type
== RET_PTR_TO_ALLOC_MEM_OR_NULL
) {
5528 mark_reg_known_zero(env
, regs
, BPF_REG_0
);
5529 regs
[BPF_REG_0
].type
= PTR_TO_MEM_OR_NULL
;
5530 regs
[BPF_REG_0
].mem_size
= meta
.mem_size
;
5531 } else if (fn
->ret_type
== RET_PTR_TO_MEM_OR_BTF_ID_OR_NULL
||
5532 fn
->ret_type
== RET_PTR_TO_MEM_OR_BTF_ID
) {
5533 const struct btf_type
*t
;
5535 mark_reg_known_zero(env
, regs
, BPF_REG_0
);
5536 t
= btf_type_skip_modifiers(meta
.ret_btf
, meta
.ret_btf_id
, NULL
);
5537 if (!btf_type_is_struct(t
)) {
5539 const struct btf_type
*ret
;
5542 /* resolve the type size of ksym. */
5543 ret
= btf_resolve_size(meta
.ret_btf
, t
, &tsize
);
5545 tname
= btf_name_by_offset(meta
.ret_btf
, t
->name_off
);
5546 verbose(env
, "unable to resolve the size of type '%s': %ld\n",
5547 tname
, PTR_ERR(ret
));
5550 regs
[BPF_REG_0
].type
=
5551 fn
->ret_type
== RET_PTR_TO_MEM_OR_BTF_ID
?
5552 PTR_TO_MEM
: PTR_TO_MEM_OR_NULL
;
5553 regs
[BPF_REG_0
].mem_size
= tsize
;
5555 regs
[BPF_REG_0
].type
=
5556 fn
->ret_type
== RET_PTR_TO_MEM_OR_BTF_ID
?
5557 PTR_TO_BTF_ID
: PTR_TO_BTF_ID_OR_NULL
;
5558 regs
[BPF_REG_0
].btf
= meta
.ret_btf
;
5559 regs
[BPF_REG_0
].btf_id
= meta
.ret_btf_id
;
5561 } else if (fn
->ret_type
== RET_PTR_TO_BTF_ID_OR_NULL
||
5562 fn
->ret_type
== RET_PTR_TO_BTF_ID
) {
5565 mark_reg_known_zero(env
, regs
, BPF_REG_0
);
5566 regs
[BPF_REG_0
].type
= fn
->ret_type
== RET_PTR_TO_BTF_ID
?
5568 PTR_TO_BTF_ID_OR_NULL
;
5569 ret_btf_id
= *fn
->ret_btf_id
;
5570 if (ret_btf_id
== 0) {
5571 verbose(env
, "invalid return type %d of func %s#%d\n",
5572 fn
->ret_type
, func_id_name(func_id
), func_id
);
5575 /* current BPF helper definitions are only coming from
5576 * built-in code with type IDs from vmlinux BTF
5578 regs
[BPF_REG_0
].btf
= btf_vmlinux
;
5579 regs
[BPF_REG_0
].btf_id
= ret_btf_id
;
5581 verbose(env
, "unknown return type %d of func %s#%d\n",
5582 fn
->ret_type
, func_id_name(func_id
), func_id
);
5586 if (reg_type_may_be_null(regs
[BPF_REG_0
].type
))
5587 regs
[BPF_REG_0
].id
= ++env
->id_gen
;
5589 if (is_ptr_cast_function(func_id
)) {
5590 /* For release_reference() */
5591 regs
[BPF_REG_0
].ref_obj_id
= meta
.ref_obj_id
;
5592 } else if (is_acquire_function(func_id
, meta
.map_ptr
)) {
5593 int id
= acquire_reference_state(env
, insn_idx
);
5597 /* For mark_ptr_or_null_reg() */
5598 regs
[BPF_REG_0
].id
= id
;
5599 /* For release_reference() */
5600 regs
[BPF_REG_0
].ref_obj_id
= id
;
5603 do_refine_retval_range(regs
, fn
->ret_type
, func_id
, &meta
);
5605 err
= check_map_func_compatibility(env
, meta
.map_ptr
, func_id
);
5609 if ((func_id
== BPF_FUNC_get_stack
||
5610 func_id
== BPF_FUNC_get_task_stack
) &&
5611 !env
->prog
->has_callchain_buf
) {
5612 const char *err_str
;
5614 #ifdef CONFIG_PERF_EVENTS
5615 err
= get_callchain_buffers(sysctl_perf_event_max_stack
);
5616 err_str
= "cannot get callchain buffer for func %s#%d\n";
5619 err_str
= "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
5622 verbose(env
, err_str
, func_id_name(func_id
), func_id
);
5626 env
->prog
->has_callchain_buf
= true;
5629 if (func_id
== BPF_FUNC_get_stackid
|| func_id
== BPF_FUNC_get_stack
)
5630 env
->prog
->call_get_stack
= true;
5633 clear_all_pkt_pointers(env
);
5637 static bool signed_add_overflows(s64 a
, s64 b
)
5639 /* Do the add in u64, where overflow is well-defined */
5640 s64 res
= (s64
)((u64
)a
+ (u64
)b
);
5647 static bool signed_add32_overflows(s32 a
, s32 b
)
5649 /* Do the add in u32, where overflow is well-defined */
5650 s32 res
= (s32
)((u32
)a
+ (u32
)b
);
5657 static bool signed_sub_overflows(s64 a
, s64 b
)
5659 /* Do the sub in u64, where overflow is well-defined */
5660 s64 res
= (s64
)((u64
)a
- (u64
)b
);
5667 static bool signed_sub32_overflows(s32 a
, s32 b
)
5669 /* Do the sub in u32, where overflow is well-defined */
5670 s32 res
= (s32
)((u32
)a
- (u32
)b
);
5677 static bool check_reg_sane_offset(struct bpf_verifier_env
*env
,
5678 const struct bpf_reg_state
*reg
,
5679 enum bpf_reg_type type
)
5681 bool known
= tnum_is_const(reg
->var_off
);
5682 s64 val
= reg
->var_off
.value
;
5683 s64 smin
= reg
->smin_value
;
5685 if (known
&& (val
>= BPF_MAX_VAR_OFF
|| val
<= -BPF_MAX_VAR_OFF
)) {
5686 verbose(env
, "math between %s pointer and %lld is not allowed\n",
5687 reg_type_str
[type
], val
);
5691 if (reg
->off
>= BPF_MAX_VAR_OFF
|| reg
->off
<= -BPF_MAX_VAR_OFF
) {
5692 verbose(env
, "%s pointer offset %d is not allowed\n",
5693 reg_type_str
[type
], reg
->off
);
5697 if (smin
== S64_MIN
) {
5698 verbose(env
, "math between %s pointer and register with unbounded min value is not allowed\n",
5699 reg_type_str
[type
]);
5703 if (smin
>= BPF_MAX_VAR_OFF
|| smin
<= -BPF_MAX_VAR_OFF
) {
5704 verbose(env
, "value %lld makes %s pointer be out of bounds\n",
5705 smin
, reg_type_str
[type
]);
5712 static struct bpf_insn_aux_data
*cur_aux(struct bpf_verifier_env
*env
)
5714 return &env
->insn_aux_data
[env
->insn_idx
];
5725 static int retrieve_ptr_limit(const struct bpf_reg_state
*ptr_reg
,
5726 const struct bpf_reg_state
*off_reg
,
5727 u32
*alu_limit
, u8 opcode
)
5729 bool off_is_neg
= off_reg
->smin_value
< 0;
5730 bool mask_to_left
= (opcode
== BPF_ADD
&& off_is_neg
) ||
5731 (opcode
== BPF_SUB
&& !off_is_neg
);
5732 u32 max
= 0, ptr_limit
= 0;
5734 if (!tnum_is_const(off_reg
->var_off
) &&
5735 (off_reg
->smin_value
< 0) != (off_reg
->smax_value
< 0))
5736 return REASON_BOUNDS
;
5738 switch (ptr_reg
->type
) {
5740 /* Offset 0 is out-of-bounds, but acceptable start for the
5741 * left direction, see BPF_REG_FP. Also, unknown scalar
5742 * offset where we would need to deal with min/max bounds is
5743 * currently prohibited for unprivileged.
5745 max
= MAX_BPF_STACK
+ mask_to_left
;
5746 ptr_limit
= -(ptr_reg
->var_off
.value
+ ptr_reg
->off
);
5748 case PTR_TO_MAP_VALUE
:
5749 max
= ptr_reg
->map_ptr
->value_size
;
5750 ptr_limit
= (mask_to_left
?
5751 ptr_reg
->smin_value
:
5752 ptr_reg
->umax_value
) + ptr_reg
->off
;
5758 if (ptr_limit
>= max
)
5759 return REASON_LIMIT
;
5760 *alu_limit
= ptr_limit
;
5764 static bool can_skip_alu_sanitation(const struct bpf_verifier_env
*env
,
5765 const struct bpf_insn
*insn
)
5767 return env
->bypass_spec_v1
|| BPF_SRC(insn
->code
) == BPF_K
;
5770 static int update_alu_sanitation_state(struct bpf_insn_aux_data
*aux
,
5771 u32 alu_state
, u32 alu_limit
)
5773 /* If we arrived here from different branches with different
5774 * state or limits to sanitize, then this won't work.
5776 if (aux
->alu_state
&&
5777 (aux
->alu_state
!= alu_state
||
5778 aux
->alu_limit
!= alu_limit
))
5779 return REASON_PATHS
;
5781 /* Corresponding fixup done in fixup_bpf_calls(). */
5782 aux
->alu_state
= alu_state
;
5783 aux
->alu_limit
= alu_limit
;
5787 static int sanitize_val_alu(struct bpf_verifier_env
*env
,
5788 struct bpf_insn
*insn
)
5790 struct bpf_insn_aux_data
*aux
= cur_aux(env
);
5792 if (can_skip_alu_sanitation(env
, insn
))
5795 return update_alu_sanitation_state(aux
, BPF_ALU_NON_POINTER
, 0);
5798 static bool sanitize_needed(u8 opcode
)
5800 return opcode
== BPF_ADD
|| opcode
== BPF_SUB
;
5803 static int sanitize_ptr_alu(struct bpf_verifier_env
*env
,
5804 struct bpf_insn
*insn
,
5805 const struct bpf_reg_state
*ptr_reg
,
5806 const struct bpf_reg_state
*off_reg
,
5807 struct bpf_reg_state
*dst_reg
,
5808 struct bpf_insn_aux_data
*tmp_aux
,
5809 const bool commit_window
)
5811 struct bpf_insn_aux_data
*aux
= commit_window
? cur_aux(env
) : tmp_aux
;
5812 struct bpf_verifier_state
*vstate
= env
->cur_state
;
5813 bool off_is_imm
= tnum_is_const(off_reg
->var_off
);
5814 bool off_is_neg
= off_reg
->smin_value
< 0;
5815 bool ptr_is_dst_reg
= ptr_reg
== dst_reg
;
5816 u8 opcode
= BPF_OP(insn
->code
);
5817 u32 alu_state
, alu_limit
;
5818 struct bpf_reg_state tmp
;
5822 if (can_skip_alu_sanitation(env
, insn
))
5825 /* We already marked aux for masking from non-speculative
5826 * paths, thus we got here in the first place. We only care
5827 * to explore bad access from here.
5829 if (vstate
->speculative
)
5832 err
= retrieve_ptr_limit(ptr_reg
, off_reg
, &alu_limit
, opcode
);
5836 if (commit_window
) {
5837 /* In commit phase we narrow the masking window based on
5838 * the observed pointer move after the simulated operation.
5840 alu_state
= tmp_aux
->alu_state
;
5841 alu_limit
= abs(tmp_aux
->alu_limit
- alu_limit
);
5843 alu_state
= off_is_neg
? BPF_ALU_NEG_VALUE
: 0;
5844 alu_state
|= off_is_imm
? BPF_ALU_IMMEDIATE
: 0;
5845 alu_state
|= ptr_is_dst_reg
?
5846 BPF_ALU_SANITIZE_SRC
: BPF_ALU_SANITIZE_DST
;
5849 err
= update_alu_sanitation_state(aux
, alu_state
, alu_limit
);
5853 /* If we're in commit phase, we're done here given we already
5854 * pushed the truncated dst_reg into the speculative verification
5860 /* Simulate and find potential out-of-bounds access under
5861 * speculative execution from truncation as a result of
5862 * masking when off was not within expected range. If off
5863 * sits in dst, then we temporarily need to move ptr there
5864 * to simulate dst (== 0) +/-= ptr. Needed, for example,
5865 * for cases where we use K-based arithmetic in one direction
5866 * and truncated reg-based in the other in order to explore
5869 if (!ptr_is_dst_reg
) {
5871 *dst_reg
= *ptr_reg
;
5873 ret
= push_stack(env
, env
->insn_idx
+ 1, env
->insn_idx
, true);
5874 if (!ptr_is_dst_reg
&& ret
)
5876 return !ret
? REASON_STACK
: 0;
5879 static int sanitize_err(struct bpf_verifier_env
*env
,
5880 const struct bpf_insn
*insn
, int reason
,
5881 const struct bpf_reg_state
*off_reg
,
5882 const struct bpf_reg_state
*dst_reg
)
5884 static const char *err
= "pointer arithmetic with it prohibited for !root";
5885 const char *op
= BPF_OP(insn
->code
) == BPF_ADD
? "add" : "sub";
5886 u32 dst
= insn
->dst_reg
, src
= insn
->src_reg
;
5890 verbose(env
, "R%d has unknown scalar with mixed signed bounds, %s\n",
5891 off_reg
== dst_reg
? dst
: src
, err
);
5894 verbose(env
, "R%d has pointer with unsupported alu operation, %s\n",
5895 off_reg
== dst_reg
? src
: dst
, err
);
5898 verbose(env
, "R%d tried to %s from different maps, paths or scalars, %s\n",
5902 verbose(env
, "R%d tried to %s beyond pointer bounds, %s\n",
5906 verbose(env
, "R%d could not be pushed for speculative verification, %s\n",
5910 verbose(env
, "verifier internal error: unknown reason (%d)\n",
5918 /* check that stack access falls within stack limits and that 'reg' doesn't
5919 * have a variable offset.
5921 * Variable offset is prohibited for unprivileged mode for simplicity since it
5922 * requires corresponding support in Spectre masking for stack ALU. See also
5923 * retrieve_ptr_limit().
5926 * 'off' includes 'reg->off'.
5928 static int check_stack_access_for_ptr_arithmetic(
5929 struct bpf_verifier_env
*env
,
5931 const struct bpf_reg_state
*reg
,
5934 if (!tnum_is_const(reg
->var_off
)) {
5937 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
5938 verbose(env
, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
5939 regno
, tn_buf
, off
);
5943 if (off
>= 0 || off
< -MAX_BPF_STACK
) {
5944 verbose(env
, "R%d stack pointer arithmetic goes out of range, "
5945 "prohibited for !root; off=%d\n", regno
, off
);
5952 static int sanitize_check_bounds(struct bpf_verifier_env
*env
,
5953 const struct bpf_insn
*insn
,
5954 const struct bpf_reg_state
*dst_reg
)
5956 u32 dst
= insn
->dst_reg
;
5958 /* For unprivileged we require that resulting offset must be in bounds
5959 * in order to be able to sanitize access later on.
5961 if (env
->bypass_spec_v1
)
5964 switch (dst_reg
->type
) {
5966 if (check_stack_access_for_ptr_arithmetic(env
, dst
, dst_reg
,
5967 dst_reg
->off
+ dst_reg
->var_off
.value
))
5970 case PTR_TO_MAP_VALUE
:
5971 if (check_map_access(env
, dst
, dst_reg
->off
, 1, false)) {
5972 verbose(env
, "R%d pointer arithmetic of map value goes out of range, "
5973 "prohibited for !root\n", dst
);
5984 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
5985 * Caller should also handle BPF_MOV case separately.
5986 * If we return -EACCES, caller may want to try again treating pointer as a
5987 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
5989 static int adjust_ptr_min_max_vals(struct bpf_verifier_env
*env
,
5990 struct bpf_insn
*insn
,
5991 const struct bpf_reg_state
*ptr_reg
,
5992 const struct bpf_reg_state
*off_reg
)
5994 struct bpf_verifier_state
*vstate
= env
->cur_state
;
5995 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
5996 struct bpf_reg_state
*regs
= state
->regs
, *dst_reg
;
5997 bool known
= tnum_is_const(off_reg
->var_off
);
5998 s64 smin_val
= off_reg
->smin_value
, smax_val
= off_reg
->smax_value
,
5999 smin_ptr
= ptr_reg
->smin_value
, smax_ptr
= ptr_reg
->smax_value
;
6000 u64 umin_val
= off_reg
->umin_value
, umax_val
= off_reg
->umax_value
,
6001 umin_ptr
= ptr_reg
->umin_value
, umax_ptr
= ptr_reg
->umax_value
;
6002 struct bpf_insn_aux_data tmp_aux
= {};
6003 u8 opcode
= BPF_OP(insn
->code
);
6004 u32 dst
= insn
->dst_reg
;
6007 dst_reg
= ®s
[dst
];
6009 if ((known
&& (smin_val
!= smax_val
|| umin_val
!= umax_val
)) ||
6010 smin_val
> smax_val
|| umin_val
> umax_val
) {
6011 /* Taint dst register if offset had invalid bounds derived from
6012 * e.g. dead branches.
6014 __mark_reg_unknown(env
, dst_reg
);
6018 if (BPF_CLASS(insn
->code
) != BPF_ALU64
) {
6019 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
6020 if (opcode
== BPF_SUB
&& env
->allow_ptr_leaks
) {
6021 __mark_reg_unknown(env
, dst_reg
);
6026 "R%d 32-bit pointer arithmetic prohibited\n",
6031 switch (ptr_reg
->type
) {
6032 case PTR_TO_MAP_VALUE_OR_NULL
:
6033 verbose(env
, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
6034 dst
, reg_type_str
[ptr_reg
->type
]);
6036 case CONST_PTR_TO_MAP
:
6037 /* smin_val represents the known value */
6038 if (known
&& smin_val
== 0 && opcode
== BPF_ADD
)
6041 case PTR_TO_PACKET_END
:
6043 case PTR_TO_SOCKET_OR_NULL
:
6044 case PTR_TO_SOCK_COMMON
:
6045 case PTR_TO_SOCK_COMMON_OR_NULL
:
6046 case PTR_TO_TCP_SOCK
:
6047 case PTR_TO_TCP_SOCK_OR_NULL
:
6048 case PTR_TO_XDP_SOCK
:
6049 verbose(env
, "R%d pointer arithmetic on %s prohibited\n",
6050 dst
, reg_type_str
[ptr_reg
->type
]);
6056 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
6057 * The id may be overwritten later if we create a new variable offset.
6059 dst_reg
->type
= ptr_reg
->type
;
6060 dst_reg
->id
= ptr_reg
->id
;
6062 if (!check_reg_sane_offset(env
, off_reg
, ptr_reg
->type
) ||
6063 !check_reg_sane_offset(env
, ptr_reg
, ptr_reg
->type
))
6066 /* pointer types do not carry 32-bit bounds at the moment. */
6067 __mark_reg32_unbounded(dst_reg
);
6069 if (sanitize_needed(opcode
)) {
6070 ret
= sanitize_ptr_alu(env
, insn
, ptr_reg
, off_reg
, dst_reg
,
6073 return sanitize_err(env
, insn
, ret
, off_reg
, dst_reg
);
6078 /* We can take a fixed offset as long as it doesn't overflow
6079 * the s32 'off' field
6081 if (known
&& (ptr_reg
->off
+ smin_val
==
6082 (s64
)(s32
)(ptr_reg
->off
+ smin_val
))) {
6083 /* pointer += K. Accumulate it into fixed offset */
6084 dst_reg
->smin_value
= smin_ptr
;
6085 dst_reg
->smax_value
= smax_ptr
;
6086 dst_reg
->umin_value
= umin_ptr
;
6087 dst_reg
->umax_value
= umax_ptr
;
6088 dst_reg
->var_off
= ptr_reg
->var_off
;
6089 dst_reg
->off
= ptr_reg
->off
+ smin_val
;
6090 dst_reg
->raw
= ptr_reg
->raw
;
6093 /* A new variable offset is created. Note that off_reg->off
6094 * == 0, since it's a scalar.
6095 * dst_reg gets the pointer type and since some positive
6096 * integer value was added to the pointer, give it a new 'id'
6097 * if it's a PTR_TO_PACKET.
6098 * this creates a new 'base' pointer, off_reg (variable) gets
6099 * added into the variable offset, and we copy the fixed offset
6102 if (signed_add_overflows(smin_ptr
, smin_val
) ||
6103 signed_add_overflows(smax_ptr
, smax_val
)) {
6104 dst_reg
->smin_value
= S64_MIN
;
6105 dst_reg
->smax_value
= S64_MAX
;
6107 dst_reg
->smin_value
= smin_ptr
+ smin_val
;
6108 dst_reg
->smax_value
= smax_ptr
+ smax_val
;
6110 if (umin_ptr
+ umin_val
< umin_ptr
||
6111 umax_ptr
+ umax_val
< umax_ptr
) {
6112 dst_reg
->umin_value
= 0;
6113 dst_reg
->umax_value
= U64_MAX
;
6115 dst_reg
->umin_value
= umin_ptr
+ umin_val
;
6116 dst_reg
->umax_value
= umax_ptr
+ umax_val
;
6118 dst_reg
->var_off
= tnum_add(ptr_reg
->var_off
, off_reg
->var_off
);
6119 dst_reg
->off
= ptr_reg
->off
;
6120 dst_reg
->raw
= ptr_reg
->raw
;
6121 if (reg_is_pkt_pointer(ptr_reg
)) {
6122 dst_reg
->id
= ++env
->id_gen
;
6123 /* something was added to pkt_ptr, set range to zero */
6124 memset(&dst_reg
->raw
, 0, sizeof(dst_reg
->raw
));
6128 if (dst_reg
== off_reg
) {
6129 /* scalar -= pointer. Creates an unknown scalar */
6130 verbose(env
, "R%d tried to subtract pointer from scalar\n",
6134 /* We don't allow subtraction from FP, because (according to
6135 * test_verifier.c test "invalid fp arithmetic", JITs might not
6136 * be able to deal with it.
6138 if (ptr_reg
->type
== PTR_TO_STACK
) {
6139 verbose(env
, "R%d subtraction from stack pointer prohibited\n",
6143 if (known
&& (ptr_reg
->off
- smin_val
==
6144 (s64
)(s32
)(ptr_reg
->off
- smin_val
))) {
6145 /* pointer -= K. Subtract it from fixed offset */
6146 dst_reg
->smin_value
= smin_ptr
;
6147 dst_reg
->smax_value
= smax_ptr
;
6148 dst_reg
->umin_value
= umin_ptr
;
6149 dst_reg
->umax_value
= umax_ptr
;
6150 dst_reg
->var_off
= ptr_reg
->var_off
;
6151 dst_reg
->id
= ptr_reg
->id
;
6152 dst_reg
->off
= ptr_reg
->off
- smin_val
;
6153 dst_reg
->raw
= ptr_reg
->raw
;
6156 /* A new variable offset is created. If the subtrahend is known
6157 * nonnegative, then any reg->range we had before is still good.
6159 if (signed_sub_overflows(smin_ptr
, smax_val
) ||
6160 signed_sub_overflows(smax_ptr
, smin_val
)) {
6161 /* Overflow possible, we know nothing */
6162 dst_reg
->smin_value
= S64_MIN
;
6163 dst_reg
->smax_value
= S64_MAX
;
6165 dst_reg
->smin_value
= smin_ptr
- smax_val
;
6166 dst_reg
->smax_value
= smax_ptr
- smin_val
;
6168 if (umin_ptr
< umax_val
) {
6169 /* Overflow possible, we know nothing */
6170 dst_reg
->umin_value
= 0;
6171 dst_reg
->umax_value
= U64_MAX
;
6173 /* Cannot overflow (as long as bounds are consistent) */
6174 dst_reg
->umin_value
= umin_ptr
- umax_val
;
6175 dst_reg
->umax_value
= umax_ptr
- umin_val
;
6177 dst_reg
->var_off
= tnum_sub(ptr_reg
->var_off
, off_reg
->var_off
);
6178 dst_reg
->off
= ptr_reg
->off
;
6179 dst_reg
->raw
= ptr_reg
->raw
;
6180 if (reg_is_pkt_pointer(ptr_reg
)) {
6181 dst_reg
->id
= ++env
->id_gen
;
6182 /* something was added to pkt_ptr, set range to zero */
6184 memset(&dst_reg
->raw
, 0, sizeof(dst_reg
->raw
));
6190 /* bitwise ops on pointers are troublesome, prohibit. */
6191 verbose(env
, "R%d bitwise operator %s on pointer prohibited\n",
6192 dst
, bpf_alu_string
[opcode
>> 4]);
6195 /* other operators (e.g. MUL,LSH) produce non-pointer results */
6196 verbose(env
, "R%d pointer arithmetic with %s operator prohibited\n",
6197 dst
, bpf_alu_string
[opcode
>> 4]);
6201 if (!check_reg_sane_offset(env
, dst_reg
, ptr_reg
->type
))
6204 __update_reg_bounds(dst_reg
);
6205 __reg_deduce_bounds(dst_reg
);
6206 __reg_bound_offset(dst_reg
);
6208 if (sanitize_check_bounds(env
, insn
, dst_reg
) < 0)
6210 if (sanitize_needed(opcode
)) {
6211 ret
= sanitize_ptr_alu(env
, insn
, dst_reg
, off_reg
, dst_reg
,
6214 return sanitize_err(env
, insn
, ret
, off_reg
, dst_reg
);
6220 static void scalar32_min_max_add(struct bpf_reg_state
*dst_reg
,
6221 struct bpf_reg_state
*src_reg
)
6223 s32 smin_val
= src_reg
->s32_min_value
;
6224 s32 smax_val
= src_reg
->s32_max_value
;
6225 u32 umin_val
= src_reg
->u32_min_value
;
6226 u32 umax_val
= src_reg
->u32_max_value
;
6228 if (signed_add32_overflows(dst_reg
->s32_min_value
, smin_val
) ||
6229 signed_add32_overflows(dst_reg
->s32_max_value
, smax_val
)) {
6230 dst_reg
->s32_min_value
= S32_MIN
;
6231 dst_reg
->s32_max_value
= S32_MAX
;
6233 dst_reg
->s32_min_value
+= smin_val
;
6234 dst_reg
->s32_max_value
+= smax_val
;
6236 if (dst_reg
->u32_min_value
+ umin_val
< umin_val
||
6237 dst_reg
->u32_max_value
+ umax_val
< umax_val
) {
6238 dst_reg
->u32_min_value
= 0;
6239 dst_reg
->u32_max_value
= U32_MAX
;
6241 dst_reg
->u32_min_value
+= umin_val
;
6242 dst_reg
->u32_max_value
+= umax_val
;
6246 static void scalar_min_max_add(struct bpf_reg_state
*dst_reg
,
6247 struct bpf_reg_state
*src_reg
)
6249 s64 smin_val
= src_reg
->smin_value
;
6250 s64 smax_val
= src_reg
->smax_value
;
6251 u64 umin_val
= src_reg
->umin_value
;
6252 u64 umax_val
= src_reg
->umax_value
;
6254 if (signed_add_overflows(dst_reg
->smin_value
, smin_val
) ||
6255 signed_add_overflows(dst_reg
->smax_value
, smax_val
)) {
6256 dst_reg
->smin_value
= S64_MIN
;
6257 dst_reg
->smax_value
= S64_MAX
;
6259 dst_reg
->smin_value
+= smin_val
;
6260 dst_reg
->smax_value
+= smax_val
;
6262 if (dst_reg
->umin_value
+ umin_val
< umin_val
||
6263 dst_reg
->umax_value
+ umax_val
< umax_val
) {
6264 dst_reg
->umin_value
= 0;
6265 dst_reg
->umax_value
= U64_MAX
;
6267 dst_reg
->umin_value
+= umin_val
;
6268 dst_reg
->umax_value
+= umax_val
;
6272 static void scalar32_min_max_sub(struct bpf_reg_state
*dst_reg
,
6273 struct bpf_reg_state
*src_reg
)
6275 s32 smin_val
= src_reg
->s32_min_value
;
6276 s32 smax_val
= src_reg
->s32_max_value
;
6277 u32 umin_val
= src_reg
->u32_min_value
;
6278 u32 umax_val
= src_reg
->u32_max_value
;
6280 if (signed_sub32_overflows(dst_reg
->s32_min_value
, smax_val
) ||
6281 signed_sub32_overflows(dst_reg
->s32_max_value
, smin_val
)) {
6282 /* Overflow possible, we know nothing */
6283 dst_reg
->s32_min_value
= S32_MIN
;
6284 dst_reg
->s32_max_value
= S32_MAX
;
6286 dst_reg
->s32_min_value
-= smax_val
;
6287 dst_reg
->s32_max_value
-= smin_val
;
6289 if (dst_reg
->u32_min_value
< umax_val
) {
6290 /* Overflow possible, we know nothing */
6291 dst_reg
->u32_min_value
= 0;
6292 dst_reg
->u32_max_value
= U32_MAX
;
6294 /* Cannot overflow (as long as bounds are consistent) */
6295 dst_reg
->u32_min_value
-= umax_val
;
6296 dst_reg
->u32_max_value
-= umin_val
;
6300 static void scalar_min_max_sub(struct bpf_reg_state
*dst_reg
,
6301 struct bpf_reg_state
*src_reg
)
6303 s64 smin_val
= src_reg
->smin_value
;
6304 s64 smax_val
= src_reg
->smax_value
;
6305 u64 umin_val
= src_reg
->umin_value
;
6306 u64 umax_val
= src_reg
->umax_value
;
6308 if (signed_sub_overflows(dst_reg
->smin_value
, smax_val
) ||
6309 signed_sub_overflows(dst_reg
->smax_value
, smin_val
)) {
6310 /* Overflow possible, we know nothing */
6311 dst_reg
->smin_value
= S64_MIN
;
6312 dst_reg
->smax_value
= S64_MAX
;
6314 dst_reg
->smin_value
-= smax_val
;
6315 dst_reg
->smax_value
-= smin_val
;
6317 if (dst_reg
->umin_value
< umax_val
) {
6318 /* Overflow possible, we know nothing */
6319 dst_reg
->umin_value
= 0;
6320 dst_reg
->umax_value
= U64_MAX
;
6322 /* Cannot overflow (as long as bounds are consistent) */
6323 dst_reg
->umin_value
-= umax_val
;
6324 dst_reg
->umax_value
-= umin_val
;
6328 static void scalar32_min_max_mul(struct bpf_reg_state
*dst_reg
,
6329 struct bpf_reg_state
*src_reg
)
6331 s32 smin_val
= src_reg
->s32_min_value
;
6332 u32 umin_val
= src_reg
->u32_min_value
;
6333 u32 umax_val
= src_reg
->u32_max_value
;
6335 if (smin_val
< 0 || dst_reg
->s32_min_value
< 0) {
6336 /* Ain't nobody got time to multiply that sign */
6337 __mark_reg32_unbounded(dst_reg
);
6340 /* Both values are positive, so we can work with unsigned and
6341 * copy the result to signed (unless it exceeds S32_MAX).
6343 if (umax_val
> U16_MAX
|| dst_reg
->u32_max_value
> U16_MAX
) {
6344 /* Potential overflow, we know nothing */
6345 __mark_reg32_unbounded(dst_reg
);
6348 dst_reg
->u32_min_value
*= umin_val
;
6349 dst_reg
->u32_max_value
*= umax_val
;
6350 if (dst_reg
->u32_max_value
> S32_MAX
) {
6351 /* Overflow possible, we know nothing */
6352 dst_reg
->s32_min_value
= S32_MIN
;
6353 dst_reg
->s32_max_value
= S32_MAX
;
6355 dst_reg
->s32_min_value
= dst_reg
->u32_min_value
;
6356 dst_reg
->s32_max_value
= dst_reg
->u32_max_value
;
6360 static void scalar_min_max_mul(struct bpf_reg_state
*dst_reg
,
6361 struct bpf_reg_state
*src_reg
)
6363 s64 smin_val
= src_reg
->smin_value
;
6364 u64 umin_val
= src_reg
->umin_value
;
6365 u64 umax_val
= src_reg
->umax_value
;
6367 if (smin_val
< 0 || dst_reg
->smin_value
< 0) {
6368 /* Ain't nobody got time to multiply that sign */
6369 __mark_reg64_unbounded(dst_reg
);
6372 /* Both values are positive, so we can work with unsigned and
6373 * copy the result to signed (unless it exceeds S64_MAX).
6375 if (umax_val
> U32_MAX
|| dst_reg
->umax_value
> U32_MAX
) {
6376 /* Potential overflow, we know nothing */
6377 __mark_reg64_unbounded(dst_reg
);
6380 dst_reg
->umin_value
*= umin_val
;
6381 dst_reg
->umax_value
*= umax_val
;
6382 if (dst_reg
->umax_value
> S64_MAX
) {
6383 /* Overflow possible, we know nothing */
6384 dst_reg
->smin_value
= S64_MIN
;
6385 dst_reg
->smax_value
= S64_MAX
;
6387 dst_reg
->smin_value
= dst_reg
->umin_value
;
6388 dst_reg
->smax_value
= dst_reg
->umax_value
;
6392 static void scalar32_min_max_and(struct bpf_reg_state
*dst_reg
,
6393 struct bpf_reg_state
*src_reg
)
6395 struct tnum var32_off
= tnum_subreg(dst_reg
->var_off
);
6396 s32 smin_val
= src_reg
->s32_min_value
;
6397 u32 umax_val
= src_reg
->u32_max_value
;
6399 /* We get our minimum from the var_off, since that's inherently
6400 * bitwise. Our maximum is the minimum of the operands' maxima.
6402 dst_reg
->u32_min_value
= var32_off
.value
;
6403 dst_reg
->u32_max_value
= min(dst_reg
->u32_max_value
, umax_val
);
6404 if (dst_reg
->s32_min_value
< 0 || smin_val
< 0) {
6405 /* Lose signed bounds when ANDing negative numbers,
6406 * ain't nobody got time for that.
6408 dst_reg
->s32_min_value
= S32_MIN
;
6409 dst_reg
->s32_max_value
= S32_MAX
;
6411 /* ANDing two positives gives a positive, so safe to
6412 * cast result into s64.
6414 dst_reg
->s32_min_value
= dst_reg
->u32_min_value
;
6415 dst_reg
->s32_max_value
= dst_reg
->u32_max_value
;
6420 static void scalar_min_max_and(struct bpf_reg_state
*dst_reg
,
6421 struct bpf_reg_state
*src_reg
)
6423 bool src_known
= tnum_is_const(src_reg
->var_off
);
6424 bool dst_known
= tnum_is_const(dst_reg
->var_off
);
6425 s64 smin_val
= src_reg
->smin_value
;
6426 u64 umax_val
= src_reg
->umax_value
;
6428 if (src_known
&& dst_known
) {
6429 __mark_reg_known(dst_reg
, dst_reg
->var_off
.value
);
6433 /* We get our minimum from the var_off, since that's inherently
6434 * bitwise. Our maximum is the minimum of the operands' maxima.
6436 dst_reg
->umin_value
= dst_reg
->var_off
.value
;
6437 dst_reg
->umax_value
= min(dst_reg
->umax_value
, umax_val
);
6438 if (dst_reg
->smin_value
< 0 || smin_val
< 0) {
6439 /* Lose signed bounds when ANDing negative numbers,
6440 * ain't nobody got time for that.
6442 dst_reg
->smin_value
= S64_MIN
;
6443 dst_reg
->smax_value
= S64_MAX
;
6445 /* ANDing two positives gives a positive, so safe to
6446 * cast result into s64.
6448 dst_reg
->smin_value
= dst_reg
->umin_value
;
6449 dst_reg
->smax_value
= dst_reg
->umax_value
;
6451 /* We may learn something more from the var_off */
6452 __update_reg_bounds(dst_reg
);
6455 static void scalar32_min_max_or(struct bpf_reg_state
*dst_reg
,
6456 struct bpf_reg_state
*src_reg
)
6458 struct tnum var32_off
= tnum_subreg(dst_reg
->var_off
);
6459 s32 smin_val
= src_reg
->s32_min_value
;
6460 u32 umin_val
= src_reg
->u32_min_value
;
6462 /* We get our maximum from the var_off, and our minimum is the
6463 * maximum of the operands' minima
6465 dst_reg
->u32_min_value
= max(dst_reg
->u32_min_value
, umin_val
);
6466 dst_reg
->u32_max_value
= var32_off
.value
| var32_off
.mask
;
6467 if (dst_reg
->s32_min_value
< 0 || smin_val
< 0) {
6468 /* Lose signed bounds when ORing negative numbers,
6469 * ain't nobody got time for that.
6471 dst_reg
->s32_min_value
= S32_MIN
;
6472 dst_reg
->s32_max_value
= S32_MAX
;
6474 /* ORing two positives gives a positive, so safe to
6475 * cast result into s64.
6477 dst_reg
->s32_min_value
= dst_reg
->u32_min_value
;
6478 dst_reg
->s32_max_value
= dst_reg
->u32_max_value
;
6482 static void scalar_min_max_or(struct bpf_reg_state
*dst_reg
,
6483 struct bpf_reg_state
*src_reg
)
6485 bool src_known
= tnum_is_const(src_reg
->var_off
);
6486 bool dst_known
= tnum_is_const(dst_reg
->var_off
);
6487 s64 smin_val
= src_reg
->smin_value
;
6488 u64 umin_val
= src_reg
->umin_value
;
6490 if (src_known
&& dst_known
) {
6491 __mark_reg_known(dst_reg
, dst_reg
->var_off
.value
);
6495 /* We get our maximum from the var_off, and our minimum is the
6496 * maximum of the operands' minima
6498 dst_reg
->umin_value
= max(dst_reg
->umin_value
, umin_val
);
6499 dst_reg
->umax_value
= dst_reg
->var_off
.value
| dst_reg
->var_off
.mask
;
6500 if (dst_reg
->smin_value
< 0 || smin_val
< 0) {
6501 /* Lose signed bounds when ORing negative numbers,
6502 * ain't nobody got time for that.
6504 dst_reg
->smin_value
= S64_MIN
;
6505 dst_reg
->smax_value
= S64_MAX
;
6507 /* ORing two positives gives a positive, so safe to
6508 * cast result into s64.
6510 dst_reg
->smin_value
= dst_reg
->umin_value
;
6511 dst_reg
->smax_value
= dst_reg
->umax_value
;
6513 /* We may learn something more from the var_off */
6514 __update_reg_bounds(dst_reg
);
6517 static void scalar32_min_max_xor(struct bpf_reg_state
*dst_reg
,
6518 struct bpf_reg_state
*src_reg
)
6520 struct tnum var32_off
= tnum_subreg(dst_reg
->var_off
);
6521 s32 smin_val
= src_reg
->s32_min_value
;
6523 /* We get both minimum and maximum from the var32_off. */
6524 dst_reg
->u32_min_value
= var32_off
.value
;
6525 dst_reg
->u32_max_value
= var32_off
.value
| var32_off
.mask
;
6527 if (dst_reg
->s32_min_value
>= 0 && smin_val
>= 0) {
6528 /* XORing two positive sign numbers gives a positive,
6529 * so safe to cast u32 result into s32.
6531 dst_reg
->s32_min_value
= dst_reg
->u32_min_value
;
6532 dst_reg
->s32_max_value
= dst_reg
->u32_max_value
;
6534 dst_reg
->s32_min_value
= S32_MIN
;
6535 dst_reg
->s32_max_value
= S32_MAX
;
6539 static void scalar_min_max_xor(struct bpf_reg_state
*dst_reg
,
6540 struct bpf_reg_state
*src_reg
)
6542 bool src_known
= tnum_is_const(src_reg
->var_off
);
6543 bool dst_known
= tnum_is_const(dst_reg
->var_off
);
6544 s64 smin_val
= src_reg
->smin_value
;
6546 if (src_known
&& dst_known
) {
6547 /* dst_reg->var_off.value has been updated earlier */
6548 __mark_reg_known(dst_reg
, dst_reg
->var_off
.value
);
6552 /* We get both minimum and maximum from the var_off. */
6553 dst_reg
->umin_value
= dst_reg
->var_off
.value
;
6554 dst_reg
->umax_value
= dst_reg
->var_off
.value
| dst_reg
->var_off
.mask
;
6556 if (dst_reg
->smin_value
>= 0 && smin_val
>= 0) {
6557 /* XORing two positive sign numbers gives a positive,
6558 * so safe to cast u64 result into s64.
6560 dst_reg
->smin_value
= dst_reg
->umin_value
;
6561 dst_reg
->smax_value
= dst_reg
->umax_value
;
6563 dst_reg
->smin_value
= S64_MIN
;
6564 dst_reg
->smax_value
= S64_MAX
;
6567 __update_reg_bounds(dst_reg
);
6570 static void __scalar32_min_max_lsh(struct bpf_reg_state
*dst_reg
,
6571 u64 umin_val
, u64 umax_val
)
6573 /* We lose all sign bit information (except what we can pick
6576 dst_reg
->s32_min_value
= S32_MIN
;
6577 dst_reg
->s32_max_value
= S32_MAX
;
6578 /* If we might shift our top bit out, then we know nothing */
6579 if (umax_val
> 31 || dst_reg
->u32_max_value
> 1ULL << (31 - umax_val
)) {
6580 dst_reg
->u32_min_value
= 0;
6581 dst_reg
->u32_max_value
= U32_MAX
;
6583 dst_reg
->u32_min_value
<<= umin_val
;
6584 dst_reg
->u32_max_value
<<= umax_val
;
6588 static void scalar32_min_max_lsh(struct bpf_reg_state
*dst_reg
,
6589 struct bpf_reg_state
*src_reg
)
6591 u32 umax_val
= src_reg
->u32_max_value
;
6592 u32 umin_val
= src_reg
->u32_min_value
;
6593 /* u32 alu operation will zext upper bits */
6594 struct tnum subreg
= tnum_subreg(dst_reg
->var_off
);
6596 __scalar32_min_max_lsh(dst_reg
, umin_val
, umax_val
);
6597 dst_reg
->var_off
= tnum_subreg(tnum_lshift(subreg
, umin_val
));
6598 /* Not required but being careful mark reg64 bounds as unknown so
6599 * that we are forced to pick them up from tnum and zext later and
6600 * if some path skips this step we are still safe.
6602 __mark_reg64_unbounded(dst_reg
);
6603 __update_reg32_bounds(dst_reg
);
6606 static void __scalar64_min_max_lsh(struct bpf_reg_state
*dst_reg
,
6607 u64 umin_val
, u64 umax_val
)
6609 /* Special case <<32 because it is a common compiler pattern to sign
6610 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
6611 * positive we know this shift will also be positive so we can track
6612 * bounds correctly. Otherwise we lose all sign bit information except
6613 * what we can pick up from var_off. Perhaps we can generalize this
6614 * later to shifts of any length.
6616 if (umin_val
== 32 && umax_val
== 32 && dst_reg
->s32_max_value
>= 0)
6617 dst_reg
->smax_value
= (s64
)dst_reg
->s32_max_value
<< 32;
6619 dst_reg
->smax_value
= S64_MAX
;
6621 if (umin_val
== 32 && umax_val
== 32 && dst_reg
->s32_min_value
>= 0)
6622 dst_reg
->smin_value
= (s64
)dst_reg
->s32_min_value
<< 32;
6624 dst_reg
->smin_value
= S64_MIN
;
6626 /* If we might shift our top bit out, then we know nothing */
6627 if (dst_reg
->umax_value
> 1ULL << (63 - umax_val
)) {
6628 dst_reg
->umin_value
= 0;
6629 dst_reg
->umax_value
= U64_MAX
;
6631 dst_reg
->umin_value
<<= umin_val
;
6632 dst_reg
->umax_value
<<= umax_val
;
6636 static void scalar_min_max_lsh(struct bpf_reg_state
*dst_reg
,
6637 struct bpf_reg_state
*src_reg
)
6639 u64 umax_val
= src_reg
->umax_value
;
6640 u64 umin_val
= src_reg
->umin_value
;
6642 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
6643 __scalar64_min_max_lsh(dst_reg
, umin_val
, umax_val
);
6644 __scalar32_min_max_lsh(dst_reg
, umin_val
, umax_val
);
6646 dst_reg
->var_off
= tnum_lshift(dst_reg
->var_off
, umin_val
);
6647 /* We may learn something more from the var_off */
6648 __update_reg_bounds(dst_reg
);
6651 static void scalar32_min_max_rsh(struct bpf_reg_state
*dst_reg
,
6652 struct bpf_reg_state
*src_reg
)
6654 struct tnum subreg
= tnum_subreg(dst_reg
->var_off
);
6655 u32 umax_val
= src_reg
->u32_max_value
;
6656 u32 umin_val
= src_reg
->u32_min_value
;
6658 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
6659 * be negative, then either:
6660 * 1) src_reg might be zero, so the sign bit of the result is
6661 * unknown, so we lose our signed bounds
6662 * 2) it's known negative, thus the unsigned bounds capture the
6664 * 3) the signed bounds cross zero, so they tell us nothing
6666 * If the value in dst_reg is known nonnegative, then again the
6667 * unsigned bounts capture the signed bounds.
6668 * Thus, in all cases it suffices to blow away our signed bounds
6669 * and rely on inferring new ones from the unsigned bounds and
6670 * var_off of the result.
6672 dst_reg
->s32_min_value
= S32_MIN
;
6673 dst_reg
->s32_max_value
= S32_MAX
;
6675 dst_reg
->var_off
= tnum_rshift(subreg
, umin_val
);
6676 dst_reg
->u32_min_value
>>= umax_val
;
6677 dst_reg
->u32_max_value
>>= umin_val
;
6679 __mark_reg64_unbounded(dst_reg
);
6680 __update_reg32_bounds(dst_reg
);
6683 static void scalar_min_max_rsh(struct bpf_reg_state
*dst_reg
,
6684 struct bpf_reg_state
*src_reg
)
6686 u64 umax_val
= src_reg
->umax_value
;
6687 u64 umin_val
= src_reg
->umin_value
;
6689 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
6690 * be negative, then either:
6691 * 1) src_reg might be zero, so the sign bit of the result is
6692 * unknown, so we lose our signed bounds
6693 * 2) it's known negative, thus the unsigned bounds capture the
6695 * 3) the signed bounds cross zero, so they tell us nothing
6697 * If the value in dst_reg is known nonnegative, then again the
6698 * unsigned bounts capture the signed bounds.
6699 * Thus, in all cases it suffices to blow away our signed bounds
6700 * and rely on inferring new ones from the unsigned bounds and
6701 * var_off of the result.
6703 dst_reg
->smin_value
= S64_MIN
;
6704 dst_reg
->smax_value
= S64_MAX
;
6705 dst_reg
->var_off
= tnum_rshift(dst_reg
->var_off
, umin_val
);
6706 dst_reg
->umin_value
>>= umax_val
;
6707 dst_reg
->umax_value
>>= umin_val
;
6709 /* Its not easy to operate on alu32 bounds here because it depends
6710 * on bits being shifted in. Take easy way out and mark unbounded
6711 * so we can recalculate later from tnum.
6713 __mark_reg32_unbounded(dst_reg
);
6714 __update_reg_bounds(dst_reg
);
6717 static void scalar32_min_max_arsh(struct bpf_reg_state
*dst_reg
,
6718 struct bpf_reg_state
*src_reg
)
6720 u64 umin_val
= src_reg
->u32_min_value
;
6722 /* Upon reaching here, src_known is true and
6723 * umax_val is equal to umin_val.
6725 dst_reg
->s32_min_value
= (u32
)(((s32
)dst_reg
->s32_min_value
) >> umin_val
);
6726 dst_reg
->s32_max_value
= (u32
)(((s32
)dst_reg
->s32_max_value
) >> umin_val
);
6728 dst_reg
->var_off
= tnum_arshift(tnum_subreg(dst_reg
->var_off
), umin_val
, 32);
6730 /* blow away the dst_reg umin_value/umax_value and rely on
6731 * dst_reg var_off to refine the result.
6733 dst_reg
->u32_min_value
= 0;
6734 dst_reg
->u32_max_value
= U32_MAX
;
6736 __mark_reg64_unbounded(dst_reg
);
6737 __update_reg32_bounds(dst_reg
);
6740 static void scalar_min_max_arsh(struct bpf_reg_state
*dst_reg
,
6741 struct bpf_reg_state
*src_reg
)
6743 u64 umin_val
= src_reg
->umin_value
;
6745 /* Upon reaching here, src_known is true and umax_val is equal
6748 dst_reg
->smin_value
>>= umin_val
;
6749 dst_reg
->smax_value
>>= umin_val
;
6751 dst_reg
->var_off
= tnum_arshift(dst_reg
->var_off
, umin_val
, 64);
6753 /* blow away the dst_reg umin_value/umax_value and rely on
6754 * dst_reg var_off to refine the result.
6756 dst_reg
->umin_value
= 0;
6757 dst_reg
->umax_value
= U64_MAX
;
6759 /* Its not easy to operate on alu32 bounds here because it depends
6760 * on bits being shifted in from upper 32-bits. Take easy way out
6761 * and mark unbounded so we can recalculate later from tnum.
6763 __mark_reg32_unbounded(dst_reg
);
6764 __update_reg_bounds(dst_reg
);
6767 /* WARNING: This function does calculations on 64-bit values, but the actual
6768 * execution may occur on 32-bit values. Therefore, things like bitshifts
6769 * need extra checks in the 32-bit case.
6771 static int adjust_scalar_min_max_vals(struct bpf_verifier_env
*env
,
6772 struct bpf_insn
*insn
,
6773 struct bpf_reg_state
*dst_reg
,
6774 struct bpf_reg_state src_reg
)
6776 struct bpf_reg_state
*regs
= cur_regs(env
);
6777 u8 opcode
= BPF_OP(insn
->code
);
6779 s64 smin_val
, smax_val
;
6780 u64 umin_val
, umax_val
;
6781 s32 s32_min_val
, s32_max_val
;
6782 u32 u32_min_val
, u32_max_val
;
6783 u64 insn_bitness
= (BPF_CLASS(insn
->code
) == BPF_ALU64
) ? 64 : 32;
6784 bool alu32
= (BPF_CLASS(insn
->code
) != BPF_ALU64
);
6787 smin_val
= src_reg
.smin_value
;
6788 smax_val
= src_reg
.smax_value
;
6789 umin_val
= src_reg
.umin_value
;
6790 umax_val
= src_reg
.umax_value
;
6792 s32_min_val
= src_reg
.s32_min_value
;
6793 s32_max_val
= src_reg
.s32_max_value
;
6794 u32_min_val
= src_reg
.u32_min_value
;
6795 u32_max_val
= src_reg
.u32_max_value
;
6798 src_known
= tnum_subreg_is_const(src_reg
.var_off
);
6800 (s32_min_val
!= s32_max_val
|| u32_min_val
!= u32_max_val
)) ||
6801 s32_min_val
> s32_max_val
|| u32_min_val
> u32_max_val
) {
6802 /* Taint dst register if offset had invalid bounds
6803 * derived from e.g. dead branches.
6805 __mark_reg_unknown(env
, dst_reg
);
6809 src_known
= tnum_is_const(src_reg
.var_off
);
6811 (smin_val
!= smax_val
|| umin_val
!= umax_val
)) ||
6812 smin_val
> smax_val
|| umin_val
> umax_val
) {
6813 /* Taint dst register if offset had invalid bounds
6814 * derived from e.g. dead branches.
6816 __mark_reg_unknown(env
, dst_reg
);
6822 opcode
!= BPF_ADD
&& opcode
!= BPF_SUB
&& opcode
!= BPF_AND
) {
6823 __mark_reg_unknown(env
, dst_reg
);
6827 if (sanitize_needed(opcode
)) {
6828 ret
= sanitize_val_alu(env
, insn
);
6830 return sanitize_err(env
, insn
, ret
, NULL
, NULL
);
6833 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
6834 * There are two classes of instructions: The first class we track both
6835 * alu32 and alu64 sign/unsigned bounds independently this provides the
6836 * greatest amount of precision when alu operations are mixed with jmp32
6837 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
6838 * and BPF_OR. This is possible because these ops have fairly easy to
6839 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
6840 * See alu32 verifier tests for examples. The second class of
6841 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
6842 * with regards to tracking sign/unsigned bounds because the bits may
6843 * cross subreg boundaries in the alu64 case. When this happens we mark
6844 * the reg unbounded in the subreg bound space and use the resulting
6845 * tnum to calculate an approximation of the sign/unsigned bounds.
6849 scalar32_min_max_add(dst_reg
, &src_reg
);
6850 scalar_min_max_add(dst_reg
, &src_reg
);
6851 dst_reg
->var_off
= tnum_add(dst_reg
->var_off
, src_reg
.var_off
);
6854 scalar32_min_max_sub(dst_reg
, &src_reg
);
6855 scalar_min_max_sub(dst_reg
, &src_reg
);
6856 dst_reg
->var_off
= tnum_sub(dst_reg
->var_off
, src_reg
.var_off
);
6859 dst_reg
->var_off
= tnum_mul(dst_reg
->var_off
, src_reg
.var_off
);
6860 scalar32_min_max_mul(dst_reg
, &src_reg
);
6861 scalar_min_max_mul(dst_reg
, &src_reg
);
6864 dst_reg
->var_off
= tnum_and(dst_reg
->var_off
, src_reg
.var_off
);
6865 scalar32_min_max_and(dst_reg
, &src_reg
);
6866 scalar_min_max_and(dst_reg
, &src_reg
);
6869 dst_reg
->var_off
= tnum_or(dst_reg
->var_off
, src_reg
.var_off
);
6870 scalar32_min_max_or(dst_reg
, &src_reg
);
6871 scalar_min_max_or(dst_reg
, &src_reg
);
6874 dst_reg
->var_off
= tnum_xor(dst_reg
->var_off
, src_reg
.var_off
);
6875 scalar32_min_max_xor(dst_reg
, &src_reg
);
6876 scalar_min_max_xor(dst_reg
, &src_reg
);
6879 if (umax_val
>= insn_bitness
) {
6880 /* Shifts greater than 31 or 63 are undefined.
6881 * This includes shifts by a negative number.
6883 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
6887 scalar32_min_max_lsh(dst_reg
, &src_reg
);
6889 scalar_min_max_lsh(dst_reg
, &src_reg
);
6892 if (umax_val
>= insn_bitness
) {
6893 /* Shifts greater than 31 or 63 are undefined.
6894 * This includes shifts by a negative number.
6896 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
6900 scalar32_min_max_rsh(dst_reg
, &src_reg
);
6902 scalar_min_max_rsh(dst_reg
, &src_reg
);
6905 if (umax_val
>= insn_bitness
) {
6906 /* Shifts greater than 31 or 63 are undefined.
6907 * This includes shifts by a negative number.
6909 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
6913 scalar32_min_max_arsh(dst_reg
, &src_reg
);
6915 scalar_min_max_arsh(dst_reg
, &src_reg
);
6918 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
6922 /* ALU32 ops are zero extended into 64bit register */
6924 zext_32_to_64(dst_reg
);
6926 __update_reg_bounds(dst_reg
);
6927 __reg_deduce_bounds(dst_reg
);
6928 __reg_bound_offset(dst_reg
);
6932 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
6935 static int adjust_reg_min_max_vals(struct bpf_verifier_env
*env
,
6936 struct bpf_insn
*insn
)
6938 struct bpf_verifier_state
*vstate
= env
->cur_state
;
6939 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
6940 struct bpf_reg_state
*regs
= state
->regs
, *dst_reg
, *src_reg
;
6941 struct bpf_reg_state
*ptr_reg
= NULL
, off_reg
= {0};
6942 u8 opcode
= BPF_OP(insn
->code
);
6945 dst_reg
= ®s
[insn
->dst_reg
];
6947 if (dst_reg
->type
!= SCALAR_VALUE
)
6950 /* Make sure ID is cleared otherwise dst_reg min/max could be
6951 * incorrectly propagated into other registers by find_equal_scalars()
6954 if (BPF_SRC(insn
->code
) == BPF_X
) {
6955 src_reg
= ®s
[insn
->src_reg
];
6956 if (src_reg
->type
!= SCALAR_VALUE
) {
6957 if (dst_reg
->type
!= SCALAR_VALUE
) {
6958 /* Combining two pointers by any ALU op yields
6959 * an arbitrary scalar. Disallow all math except
6960 * pointer subtraction
6962 if (opcode
== BPF_SUB
&& env
->allow_ptr_leaks
) {
6963 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
6966 verbose(env
, "R%d pointer %s pointer prohibited\n",
6968 bpf_alu_string
[opcode
>> 4]);
6971 /* scalar += pointer
6972 * This is legal, but we have to reverse our
6973 * src/dest handling in computing the range
6975 err
= mark_chain_precision(env
, insn
->dst_reg
);
6978 return adjust_ptr_min_max_vals(env
, insn
,
6981 } else if (ptr_reg
) {
6982 /* pointer += scalar */
6983 err
= mark_chain_precision(env
, insn
->src_reg
);
6986 return adjust_ptr_min_max_vals(env
, insn
,
6990 /* Pretend the src is a reg with a known value, since we only
6991 * need to be able to read from this state.
6993 off_reg
.type
= SCALAR_VALUE
;
6994 __mark_reg_known(&off_reg
, insn
->imm
);
6996 if (ptr_reg
) /* pointer += K */
6997 return adjust_ptr_min_max_vals(env
, insn
,
7001 /* Got here implies adding two SCALAR_VALUEs */
7002 if (WARN_ON_ONCE(ptr_reg
)) {
7003 print_verifier_state(env
, state
);
7004 verbose(env
, "verifier internal error: unexpected ptr_reg\n");
7007 if (WARN_ON(!src_reg
)) {
7008 print_verifier_state(env
, state
);
7009 verbose(env
, "verifier internal error: no src_reg\n");
7012 return adjust_scalar_min_max_vals(env
, insn
, dst_reg
, *src_reg
);
7015 /* check validity of 32-bit and 64-bit arithmetic operations */
7016 static int check_alu_op(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
7018 struct bpf_reg_state
*regs
= cur_regs(env
);
7019 u8 opcode
= BPF_OP(insn
->code
);
7022 if (opcode
== BPF_END
|| opcode
== BPF_NEG
) {
7023 if (opcode
== BPF_NEG
) {
7024 if (BPF_SRC(insn
->code
) != 0 ||
7025 insn
->src_reg
!= BPF_REG_0
||
7026 insn
->off
!= 0 || insn
->imm
!= 0) {
7027 verbose(env
, "BPF_NEG uses reserved fields\n");
7031 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0 ||
7032 (insn
->imm
!= 16 && insn
->imm
!= 32 && insn
->imm
!= 64) ||
7033 BPF_CLASS(insn
->code
) == BPF_ALU64
) {
7034 verbose(env
, "BPF_END uses reserved fields\n");
7039 /* check src operand */
7040 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
7044 if (is_pointer_value(env
, insn
->dst_reg
)) {
7045 verbose(env
, "R%d pointer arithmetic prohibited\n",
7050 /* check dest operand */
7051 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP
);
7055 } else if (opcode
== BPF_MOV
) {
7057 if (BPF_SRC(insn
->code
) == BPF_X
) {
7058 if (insn
->imm
!= 0 || insn
->off
!= 0) {
7059 verbose(env
, "BPF_MOV uses reserved fields\n");
7063 /* check src operand */
7064 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
7068 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0) {
7069 verbose(env
, "BPF_MOV uses reserved fields\n");
7074 /* check dest operand, mark as required later */
7075 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP_NO_MARK
);
7079 if (BPF_SRC(insn
->code
) == BPF_X
) {
7080 struct bpf_reg_state
*src_reg
= regs
+ insn
->src_reg
;
7081 struct bpf_reg_state
*dst_reg
= regs
+ insn
->dst_reg
;
7083 if (BPF_CLASS(insn
->code
) == BPF_ALU64
) {
7085 * copy register state to dest reg
7087 if (src_reg
->type
== SCALAR_VALUE
&& !src_reg
->id
)
7088 /* Assign src and dst registers the same ID
7089 * that will be used by find_equal_scalars()
7090 * to propagate min/max range.
7092 src_reg
->id
= ++env
->id_gen
;
7093 *dst_reg
= *src_reg
;
7094 dst_reg
->live
|= REG_LIVE_WRITTEN
;
7095 dst_reg
->subreg_def
= DEF_NOT_SUBREG
;
7098 if (is_pointer_value(env
, insn
->src_reg
)) {
7100 "R%d partial copy of pointer\n",
7103 } else if (src_reg
->type
== SCALAR_VALUE
) {
7104 *dst_reg
= *src_reg
;
7105 /* Make sure ID is cleared otherwise
7106 * dst_reg min/max could be incorrectly
7107 * propagated into src_reg by find_equal_scalars()
7110 dst_reg
->live
|= REG_LIVE_WRITTEN
;
7111 dst_reg
->subreg_def
= env
->insn_idx
+ 1;
7113 mark_reg_unknown(env
, regs
,
7116 zext_32_to_64(dst_reg
);
7120 * remember the value we stored into this reg
7122 /* clear any state __mark_reg_known doesn't set */
7123 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
7124 regs
[insn
->dst_reg
].type
= SCALAR_VALUE
;
7125 if (BPF_CLASS(insn
->code
) == BPF_ALU64
) {
7126 __mark_reg_known(regs
+ insn
->dst_reg
,
7129 __mark_reg_known(regs
+ insn
->dst_reg
,
7134 } else if (opcode
> BPF_END
) {
7135 verbose(env
, "invalid BPF_ALU opcode %x\n", opcode
);
7138 } else { /* all other ALU ops: and, sub, xor, add, ... */
7140 if (BPF_SRC(insn
->code
) == BPF_X
) {
7141 if (insn
->imm
!= 0 || insn
->off
!= 0) {
7142 verbose(env
, "BPF_ALU uses reserved fields\n");
7145 /* check src1 operand */
7146 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
7150 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0) {
7151 verbose(env
, "BPF_ALU uses reserved fields\n");
7156 /* check src2 operand */
7157 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
7161 if ((opcode
== BPF_MOD
|| opcode
== BPF_DIV
) &&
7162 BPF_SRC(insn
->code
) == BPF_K
&& insn
->imm
== 0) {
7163 verbose(env
, "div by zero\n");
7167 if ((opcode
== BPF_LSH
|| opcode
== BPF_RSH
||
7168 opcode
== BPF_ARSH
) && BPF_SRC(insn
->code
) == BPF_K
) {
7169 int size
= BPF_CLASS(insn
->code
) == BPF_ALU64
? 64 : 32;
7171 if (insn
->imm
< 0 || insn
->imm
>= size
) {
7172 verbose(env
, "invalid shift %d\n", insn
->imm
);
7177 /* check dest operand */
7178 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP_NO_MARK
);
7182 return adjust_reg_min_max_vals(env
, insn
);
7188 static void __find_good_pkt_pointers(struct bpf_func_state
*state
,
7189 struct bpf_reg_state
*dst_reg
,
7190 enum bpf_reg_type type
, int new_range
)
7192 struct bpf_reg_state
*reg
;
7195 for (i
= 0; i
< MAX_BPF_REG
; i
++) {
7196 reg
= &state
->regs
[i
];
7197 if (reg
->type
== type
&& reg
->id
== dst_reg
->id
)
7198 /* keep the maximum range already checked */
7199 reg
->range
= max(reg
->range
, new_range
);
7202 bpf_for_each_spilled_reg(i
, state
, reg
) {
7205 if (reg
->type
== type
&& reg
->id
== dst_reg
->id
)
7206 reg
->range
= max(reg
->range
, new_range
);
7210 static void find_good_pkt_pointers(struct bpf_verifier_state
*vstate
,
7211 struct bpf_reg_state
*dst_reg
,
7212 enum bpf_reg_type type
,
7213 bool range_right_open
)
7217 if (dst_reg
->off
< 0 ||
7218 (dst_reg
->off
== 0 && range_right_open
))
7219 /* This doesn't give us any range */
7222 if (dst_reg
->umax_value
> MAX_PACKET_OFF
||
7223 dst_reg
->umax_value
+ dst_reg
->off
> MAX_PACKET_OFF
)
7224 /* Risk of overflow. For instance, ptr + (1<<63) may be less
7225 * than pkt_end, but that's because it's also less than pkt.
7229 new_range
= dst_reg
->off
;
7230 if (range_right_open
)
7233 /* Examples for register markings:
7235 * pkt_data in dst register:
7239 * if (r2 > pkt_end) goto <handle exception>
7244 * if (r2 < pkt_end) goto <access okay>
7245 * <handle exception>
7248 * r2 == dst_reg, pkt_end == src_reg
7249 * r2=pkt(id=n,off=8,r=0)
7250 * r3=pkt(id=n,off=0,r=0)
7252 * pkt_data in src register:
7256 * if (pkt_end >= r2) goto <access okay>
7257 * <handle exception>
7261 * if (pkt_end <= r2) goto <handle exception>
7265 * pkt_end == dst_reg, r2 == src_reg
7266 * r2=pkt(id=n,off=8,r=0)
7267 * r3=pkt(id=n,off=0,r=0)
7269 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
7270 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
7271 * and [r3, r3 + 8-1) respectively is safe to access depending on
7275 /* If our ids match, then we must have the same max_value. And we
7276 * don't care about the other reg's fixed offset, since if it's too big
7277 * the range won't allow anything.
7278 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
7280 for (i
= 0; i
<= vstate
->curframe
; i
++)
7281 __find_good_pkt_pointers(vstate
->frame
[i
], dst_reg
, type
,
7285 static int is_branch32_taken(struct bpf_reg_state
*reg
, u32 val
, u8 opcode
)
7287 struct tnum subreg
= tnum_subreg(reg
->var_off
);
7288 s32 sval
= (s32
)val
;
7292 if (tnum_is_const(subreg
))
7293 return !!tnum_equals_const(subreg
, val
);
7296 if (tnum_is_const(subreg
))
7297 return !tnum_equals_const(subreg
, val
);
7300 if ((~subreg
.mask
& subreg
.value
) & val
)
7302 if (!((subreg
.mask
| subreg
.value
) & val
))
7306 if (reg
->u32_min_value
> val
)
7308 else if (reg
->u32_max_value
<= val
)
7312 if (reg
->s32_min_value
> sval
)
7314 else if (reg
->s32_max_value
<= sval
)
7318 if (reg
->u32_max_value
< val
)
7320 else if (reg
->u32_min_value
>= val
)
7324 if (reg
->s32_max_value
< sval
)
7326 else if (reg
->s32_min_value
>= sval
)
7330 if (reg
->u32_min_value
>= val
)
7332 else if (reg
->u32_max_value
< val
)
7336 if (reg
->s32_min_value
>= sval
)
7338 else if (reg
->s32_max_value
< sval
)
7342 if (reg
->u32_max_value
<= val
)
7344 else if (reg
->u32_min_value
> val
)
7348 if (reg
->s32_max_value
<= sval
)
7350 else if (reg
->s32_min_value
> sval
)
7359 static int is_branch64_taken(struct bpf_reg_state
*reg
, u64 val
, u8 opcode
)
7361 s64 sval
= (s64
)val
;
7365 if (tnum_is_const(reg
->var_off
))
7366 return !!tnum_equals_const(reg
->var_off
, val
);
7369 if (tnum_is_const(reg
->var_off
))
7370 return !tnum_equals_const(reg
->var_off
, val
);
7373 if ((~reg
->var_off
.mask
& reg
->var_off
.value
) & val
)
7375 if (!((reg
->var_off
.mask
| reg
->var_off
.value
) & val
))
7379 if (reg
->umin_value
> val
)
7381 else if (reg
->umax_value
<= val
)
7385 if (reg
->smin_value
> sval
)
7387 else if (reg
->smax_value
<= sval
)
7391 if (reg
->umax_value
< val
)
7393 else if (reg
->umin_value
>= val
)
7397 if (reg
->smax_value
< sval
)
7399 else if (reg
->smin_value
>= sval
)
7403 if (reg
->umin_value
>= val
)
7405 else if (reg
->umax_value
< val
)
7409 if (reg
->smin_value
>= sval
)
7411 else if (reg
->smax_value
< sval
)
7415 if (reg
->umax_value
<= val
)
7417 else if (reg
->umin_value
> val
)
7421 if (reg
->smax_value
<= sval
)
7423 else if (reg
->smin_value
> sval
)
7431 /* compute branch direction of the expression "if (reg opcode val) goto target;"
7433 * 1 - branch will be taken and "goto target" will be executed
7434 * 0 - branch will not be taken and fall-through to next insn
7435 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
7438 static int is_branch_taken(struct bpf_reg_state
*reg
, u64 val
, u8 opcode
,
7441 if (__is_pointer_value(false, reg
)) {
7442 if (!reg_type_not_null(reg
->type
))
7445 /* If pointer is valid tests against zero will fail so we can
7446 * use this to direct branch taken.
7462 return is_branch32_taken(reg
, val
, opcode
);
7463 return is_branch64_taken(reg
, val
, opcode
);
7466 static int flip_opcode(u32 opcode
)
7468 /* How can we transform "a <op> b" into "b <op> a"? */
7469 static const u8 opcode_flip
[16] = {
7470 /* these stay the same */
7471 [BPF_JEQ
>> 4] = BPF_JEQ
,
7472 [BPF_JNE
>> 4] = BPF_JNE
,
7473 [BPF_JSET
>> 4] = BPF_JSET
,
7474 /* these swap "lesser" and "greater" (L and G in the opcodes) */
7475 [BPF_JGE
>> 4] = BPF_JLE
,
7476 [BPF_JGT
>> 4] = BPF_JLT
,
7477 [BPF_JLE
>> 4] = BPF_JGE
,
7478 [BPF_JLT
>> 4] = BPF_JGT
,
7479 [BPF_JSGE
>> 4] = BPF_JSLE
,
7480 [BPF_JSGT
>> 4] = BPF_JSLT
,
7481 [BPF_JSLE
>> 4] = BPF_JSGE
,
7482 [BPF_JSLT
>> 4] = BPF_JSGT
7484 return opcode_flip
[opcode
>> 4];
7487 static int is_pkt_ptr_branch_taken(struct bpf_reg_state
*dst_reg
,
7488 struct bpf_reg_state
*src_reg
,
7491 struct bpf_reg_state
*pkt
;
7493 if (src_reg
->type
== PTR_TO_PACKET_END
) {
7495 } else if (dst_reg
->type
== PTR_TO_PACKET_END
) {
7497 opcode
= flip_opcode(opcode
);
7502 if (pkt
->range
>= 0)
7507 /* pkt <= pkt_end */
7511 if (pkt
->range
== BEYOND_PKT_END
)
7512 /* pkt has at last one extra byte beyond pkt_end */
7513 return opcode
== BPF_JGT
;
7519 /* pkt >= pkt_end */
7520 if (pkt
->range
== BEYOND_PKT_END
|| pkt
->range
== AT_PKT_END
)
7521 return opcode
== BPF_JGE
;
7527 /* Adjusts the register min/max values in the case that the dst_reg is the
7528 * variable register that we are working on, and src_reg is a constant or we're
7529 * simply doing a BPF_K check.
7530 * In JEQ/JNE cases we also adjust the var_off values.
7532 static void reg_set_min_max(struct bpf_reg_state
*true_reg
,
7533 struct bpf_reg_state
*false_reg
,
7535 u8 opcode
, bool is_jmp32
)
7537 struct tnum false_32off
= tnum_subreg(false_reg
->var_off
);
7538 struct tnum false_64off
= false_reg
->var_off
;
7539 struct tnum true_32off
= tnum_subreg(true_reg
->var_off
);
7540 struct tnum true_64off
= true_reg
->var_off
;
7541 s64 sval
= (s64
)val
;
7542 s32 sval32
= (s32
)val32
;
7544 /* If the dst_reg is a pointer, we can't learn anything about its
7545 * variable offset from the compare (unless src_reg were a pointer into
7546 * the same object, but we don't bother with that.
7547 * Since false_reg and true_reg have the same type by construction, we
7548 * only need to check one of them for pointerness.
7550 if (__is_pointer_value(false, false_reg
))
7557 struct bpf_reg_state
*reg
=
7558 opcode
== BPF_JEQ
? true_reg
: false_reg
;
7560 /* JEQ/JNE comparison doesn't change the register equivalence.
7562 * if (r1 == 42) goto label;
7564 * label: // here both r1 and r2 are known to be 42.
7566 * Hence when marking register as known preserve it's ID.
7569 __mark_reg32_known(reg
, val32
);
7571 ___mark_reg_known(reg
, val
);
7576 false_32off
= tnum_and(false_32off
, tnum_const(~val32
));
7577 if (is_power_of_2(val32
))
7578 true_32off
= tnum_or(true_32off
,
7581 false_64off
= tnum_and(false_64off
, tnum_const(~val
));
7582 if (is_power_of_2(val
))
7583 true_64off
= tnum_or(true_64off
,
7591 u32 false_umax
= opcode
== BPF_JGT
? val32
: val32
- 1;
7592 u32 true_umin
= opcode
== BPF_JGT
? val32
+ 1 : val32
;
7594 false_reg
->u32_max_value
= min(false_reg
->u32_max_value
,
7596 true_reg
->u32_min_value
= max(true_reg
->u32_min_value
,
7599 u64 false_umax
= opcode
== BPF_JGT
? val
: val
- 1;
7600 u64 true_umin
= opcode
== BPF_JGT
? val
+ 1 : val
;
7602 false_reg
->umax_value
= min(false_reg
->umax_value
, false_umax
);
7603 true_reg
->umin_value
= max(true_reg
->umin_value
, true_umin
);
7611 s32 false_smax
= opcode
== BPF_JSGT
? sval32
: sval32
- 1;
7612 s32 true_smin
= opcode
== BPF_JSGT
? sval32
+ 1 : sval32
;
7614 false_reg
->s32_max_value
= min(false_reg
->s32_max_value
, false_smax
);
7615 true_reg
->s32_min_value
= max(true_reg
->s32_min_value
, true_smin
);
7617 s64 false_smax
= opcode
== BPF_JSGT
? sval
: sval
- 1;
7618 s64 true_smin
= opcode
== BPF_JSGT
? sval
+ 1 : sval
;
7620 false_reg
->smax_value
= min(false_reg
->smax_value
, false_smax
);
7621 true_reg
->smin_value
= max(true_reg
->smin_value
, true_smin
);
7629 u32 false_umin
= opcode
== BPF_JLT
? val32
: val32
+ 1;
7630 u32 true_umax
= opcode
== BPF_JLT
? val32
- 1 : val32
;
7632 false_reg
->u32_min_value
= max(false_reg
->u32_min_value
,
7634 true_reg
->u32_max_value
= min(true_reg
->u32_max_value
,
7637 u64 false_umin
= opcode
== BPF_JLT
? val
: val
+ 1;
7638 u64 true_umax
= opcode
== BPF_JLT
? val
- 1 : val
;
7640 false_reg
->umin_value
= max(false_reg
->umin_value
, false_umin
);
7641 true_reg
->umax_value
= min(true_reg
->umax_value
, true_umax
);
7649 s32 false_smin
= opcode
== BPF_JSLT
? sval32
: sval32
+ 1;
7650 s32 true_smax
= opcode
== BPF_JSLT
? sval32
- 1 : sval32
;
7652 false_reg
->s32_min_value
= max(false_reg
->s32_min_value
, false_smin
);
7653 true_reg
->s32_max_value
= min(true_reg
->s32_max_value
, true_smax
);
7655 s64 false_smin
= opcode
== BPF_JSLT
? sval
: sval
+ 1;
7656 s64 true_smax
= opcode
== BPF_JSLT
? sval
- 1 : sval
;
7658 false_reg
->smin_value
= max(false_reg
->smin_value
, false_smin
);
7659 true_reg
->smax_value
= min(true_reg
->smax_value
, true_smax
);
7668 false_reg
->var_off
= tnum_or(tnum_clear_subreg(false_64off
),
7669 tnum_subreg(false_32off
));
7670 true_reg
->var_off
= tnum_or(tnum_clear_subreg(true_64off
),
7671 tnum_subreg(true_32off
));
7672 __reg_combine_32_into_64(false_reg
);
7673 __reg_combine_32_into_64(true_reg
);
7675 false_reg
->var_off
= false_64off
;
7676 true_reg
->var_off
= true_64off
;
7677 __reg_combine_64_into_32(false_reg
);
7678 __reg_combine_64_into_32(true_reg
);
7682 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
7685 static void reg_set_min_max_inv(struct bpf_reg_state
*true_reg
,
7686 struct bpf_reg_state
*false_reg
,
7688 u8 opcode
, bool is_jmp32
)
7690 opcode
= flip_opcode(opcode
);
7691 /* This uses zero as "not present in table"; luckily the zero opcode,
7692 * BPF_JA, can't get here.
7695 reg_set_min_max(true_reg
, false_reg
, val
, val32
, opcode
, is_jmp32
);
7698 /* Regs are known to be equal, so intersect their min/max/var_off */
7699 static void __reg_combine_min_max(struct bpf_reg_state
*src_reg
,
7700 struct bpf_reg_state
*dst_reg
)
7702 src_reg
->umin_value
= dst_reg
->umin_value
= max(src_reg
->umin_value
,
7703 dst_reg
->umin_value
);
7704 src_reg
->umax_value
= dst_reg
->umax_value
= min(src_reg
->umax_value
,
7705 dst_reg
->umax_value
);
7706 src_reg
->smin_value
= dst_reg
->smin_value
= max(src_reg
->smin_value
,
7707 dst_reg
->smin_value
);
7708 src_reg
->smax_value
= dst_reg
->smax_value
= min(src_reg
->smax_value
,
7709 dst_reg
->smax_value
);
7710 src_reg
->var_off
= dst_reg
->var_off
= tnum_intersect(src_reg
->var_off
,
7712 /* We might have learned new bounds from the var_off. */
7713 __update_reg_bounds(src_reg
);
7714 __update_reg_bounds(dst_reg
);
7715 /* We might have learned something about the sign bit. */
7716 __reg_deduce_bounds(src_reg
);
7717 __reg_deduce_bounds(dst_reg
);
7718 /* We might have learned some bits from the bounds. */
7719 __reg_bound_offset(src_reg
);
7720 __reg_bound_offset(dst_reg
);
7721 /* Intersecting with the old var_off might have improved our bounds
7722 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
7723 * then new var_off is (0; 0x7f...fc) which improves our umax.
7725 __update_reg_bounds(src_reg
);
7726 __update_reg_bounds(dst_reg
);
7729 static void reg_combine_min_max(struct bpf_reg_state
*true_src
,
7730 struct bpf_reg_state
*true_dst
,
7731 struct bpf_reg_state
*false_src
,
7732 struct bpf_reg_state
*false_dst
,
7737 __reg_combine_min_max(true_src
, true_dst
);
7740 __reg_combine_min_max(false_src
, false_dst
);
7745 static void mark_ptr_or_null_reg(struct bpf_func_state
*state
,
7746 struct bpf_reg_state
*reg
, u32 id
,
7749 if (reg_type_may_be_null(reg
->type
) && reg
->id
== id
&&
7750 !WARN_ON_ONCE(!reg
->id
)) {
7751 /* Old offset (both fixed and variable parts) should
7752 * have been known-zero, because we don't allow pointer
7753 * arithmetic on pointers that might be NULL.
7755 if (WARN_ON_ONCE(reg
->smin_value
|| reg
->smax_value
||
7756 !tnum_equals_const(reg
->var_off
, 0) ||
7758 __mark_reg_known_zero(reg
);
7762 reg
->type
= SCALAR_VALUE
;
7763 } else if (reg
->type
== PTR_TO_MAP_VALUE_OR_NULL
) {
7764 const struct bpf_map
*map
= reg
->map_ptr
;
7766 if (map
->inner_map_meta
) {
7767 reg
->type
= CONST_PTR_TO_MAP
;
7768 reg
->map_ptr
= map
->inner_map_meta
;
7769 } else if (map
->map_type
== BPF_MAP_TYPE_XSKMAP
) {
7770 reg
->type
= PTR_TO_XDP_SOCK
;
7771 } else if (map
->map_type
== BPF_MAP_TYPE_SOCKMAP
||
7772 map
->map_type
== BPF_MAP_TYPE_SOCKHASH
) {
7773 reg
->type
= PTR_TO_SOCKET
;
7775 reg
->type
= PTR_TO_MAP_VALUE
;
7777 } else if (reg
->type
== PTR_TO_SOCKET_OR_NULL
) {
7778 reg
->type
= PTR_TO_SOCKET
;
7779 } else if (reg
->type
== PTR_TO_SOCK_COMMON_OR_NULL
) {
7780 reg
->type
= PTR_TO_SOCK_COMMON
;
7781 } else if (reg
->type
== PTR_TO_TCP_SOCK_OR_NULL
) {
7782 reg
->type
= PTR_TO_TCP_SOCK
;
7783 } else if (reg
->type
== PTR_TO_BTF_ID_OR_NULL
) {
7784 reg
->type
= PTR_TO_BTF_ID
;
7785 } else if (reg
->type
== PTR_TO_MEM_OR_NULL
) {
7786 reg
->type
= PTR_TO_MEM
;
7787 } else if (reg
->type
== PTR_TO_RDONLY_BUF_OR_NULL
) {
7788 reg
->type
= PTR_TO_RDONLY_BUF
;
7789 } else if (reg
->type
== PTR_TO_RDWR_BUF_OR_NULL
) {
7790 reg
->type
= PTR_TO_RDWR_BUF
;
7793 /* We don't need id and ref_obj_id from this point
7794 * onwards anymore, thus we should better reset it,
7795 * so that state pruning has chances to take effect.
7798 reg
->ref_obj_id
= 0;
7799 } else if (!reg_may_point_to_spin_lock(reg
)) {
7800 /* For not-NULL ptr, reg->ref_obj_id will be reset
7801 * in release_reg_references().
7803 * reg->id is still used by spin_lock ptr. Other
7804 * than spin_lock ptr type, reg->id can be reset.
7811 static void __mark_ptr_or_null_regs(struct bpf_func_state
*state
, u32 id
,
7814 struct bpf_reg_state
*reg
;
7817 for (i
= 0; i
< MAX_BPF_REG
; i
++)
7818 mark_ptr_or_null_reg(state
, &state
->regs
[i
], id
, is_null
);
7820 bpf_for_each_spilled_reg(i
, state
, reg
) {
7823 mark_ptr_or_null_reg(state
, reg
, id
, is_null
);
7827 /* The logic is similar to find_good_pkt_pointers(), both could eventually
7828 * be folded together at some point.
7830 static void mark_ptr_or_null_regs(struct bpf_verifier_state
*vstate
, u32 regno
,
7833 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
7834 struct bpf_reg_state
*regs
= state
->regs
;
7835 u32 ref_obj_id
= regs
[regno
].ref_obj_id
;
7836 u32 id
= regs
[regno
].id
;
7839 if (ref_obj_id
&& ref_obj_id
== id
&& is_null
)
7840 /* regs[regno] is in the " == NULL" branch.
7841 * No one could have freed the reference state before
7842 * doing the NULL check.
7844 WARN_ON_ONCE(release_reference_state(state
, id
));
7846 for (i
= 0; i
<= vstate
->curframe
; i
++)
7847 __mark_ptr_or_null_regs(vstate
->frame
[i
], id
, is_null
);
7850 static bool try_match_pkt_pointers(const struct bpf_insn
*insn
,
7851 struct bpf_reg_state
*dst_reg
,
7852 struct bpf_reg_state
*src_reg
,
7853 struct bpf_verifier_state
*this_branch
,
7854 struct bpf_verifier_state
*other_branch
)
7856 if (BPF_SRC(insn
->code
) != BPF_X
)
7859 /* Pointers are always 64-bit. */
7860 if (BPF_CLASS(insn
->code
) == BPF_JMP32
)
7863 switch (BPF_OP(insn
->code
)) {
7865 if ((dst_reg
->type
== PTR_TO_PACKET
&&
7866 src_reg
->type
== PTR_TO_PACKET_END
) ||
7867 (dst_reg
->type
== PTR_TO_PACKET_META
&&
7868 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
7869 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
7870 find_good_pkt_pointers(this_branch
, dst_reg
,
7871 dst_reg
->type
, false);
7872 mark_pkt_end(other_branch
, insn
->dst_reg
, true);
7873 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
7874 src_reg
->type
== PTR_TO_PACKET
) ||
7875 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
7876 src_reg
->type
== PTR_TO_PACKET_META
)) {
7877 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
7878 find_good_pkt_pointers(other_branch
, src_reg
,
7879 src_reg
->type
, true);
7880 mark_pkt_end(this_branch
, insn
->src_reg
, false);
7886 if ((dst_reg
->type
== PTR_TO_PACKET
&&
7887 src_reg
->type
== PTR_TO_PACKET_END
) ||
7888 (dst_reg
->type
== PTR_TO_PACKET_META
&&
7889 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
7890 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
7891 find_good_pkt_pointers(other_branch
, dst_reg
,
7892 dst_reg
->type
, true);
7893 mark_pkt_end(this_branch
, insn
->dst_reg
, false);
7894 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
7895 src_reg
->type
== PTR_TO_PACKET
) ||
7896 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
7897 src_reg
->type
== PTR_TO_PACKET_META
)) {
7898 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
7899 find_good_pkt_pointers(this_branch
, src_reg
,
7900 src_reg
->type
, false);
7901 mark_pkt_end(other_branch
, insn
->src_reg
, true);
7907 if ((dst_reg
->type
== PTR_TO_PACKET
&&
7908 src_reg
->type
== PTR_TO_PACKET_END
) ||
7909 (dst_reg
->type
== PTR_TO_PACKET_META
&&
7910 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
7911 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
7912 find_good_pkt_pointers(this_branch
, dst_reg
,
7913 dst_reg
->type
, true);
7914 mark_pkt_end(other_branch
, insn
->dst_reg
, false);
7915 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
7916 src_reg
->type
== PTR_TO_PACKET
) ||
7917 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
7918 src_reg
->type
== PTR_TO_PACKET_META
)) {
7919 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
7920 find_good_pkt_pointers(other_branch
, src_reg
,
7921 src_reg
->type
, false);
7922 mark_pkt_end(this_branch
, insn
->src_reg
, true);
7928 if ((dst_reg
->type
== PTR_TO_PACKET
&&
7929 src_reg
->type
== PTR_TO_PACKET_END
) ||
7930 (dst_reg
->type
== PTR_TO_PACKET_META
&&
7931 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
7932 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
7933 find_good_pkt_pointers(other_branch
, dst_reg
,
7934 dst_reg
->type
, false);
7935 mark_pkt_end(this_branch
, insn
->dst_reg
, true);
7936 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
7937 src_reg
->type
== PTR_TO_PACKET
) ||
7938 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
7939 src_reg
->type
== PTR_TO_PACKET_META
)) {
7940 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
7941 find_good_pkt_pointers(this_branch
, src_reg
,
7942 src_reg
->type
, true);
7943 mark_pkt_end(other_branch
, insn
->src_reg
, false);
7955 static void find_equal_scalars(struct bpf_verifier_state
*vstate
,
7956 struct bpf_reg_state
*known_reg
)
7958 struct bpf_func_state
*state
;
7959 struct bpf_reg_state
*reg
;
7962 for (i
= 0; i
<= vstate
->curframe
; i
++) {
7963 state
= vstate
->frame
[i
];
7964 for (j
= 0; j
< MAX_BPF_REG
; j
++) {
7965 reg
= &state
->regs
[j
];
7966 if (reg
->type
== SCALAR_VALUE
&& reg
->id
== known_reg
->id
)
7970 bpf_for_each_spilled_reg(j
, state
, reg
) {
7973 if (reg
->type
== SCALAR_VALUE
&& reg
->id
== known_reg
->id
)
7979 static int check_cond_jmp_op(struct bpf_verifier_env
*env
,
7980 struct bpf_insn
*insn
, int *insn_idx
)
7982 struct bpf_verifier_state
*this_branch
= env
->cur_state
;
7983 struct bpf_verifier_state
*other_branch
;
7984 struct bpf_reg_state
*regs
= this_branch
->frame
[this_branch
->curframe
]->regs
;
7985 struct bpf_reg_state
*dst_reg
, *other_branch_regs
, *src_reg
= NULL
;
7986 u8 opcode
= BPF_OP(insn
->code
);
7991 /* Only conditional jumps are expected to reach here. */
7992 if (opcode
== BPF_JA
|| opcode
> BPF_JSLE
) {
7993 verbose(env
, "invalid BPF_JMP/JMP32 opcode %x\n", opcode
);
7997 if (BPF_SRC(insn
->code
) == BPF_X
) {
7998 if (insn
->imm
!= 0) {
7999 verbose(env
, "BPF_JMP/JMP32 uses reserved fields\n");
8003 /* check src1 operand */
8004 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
8008 if (is_pointer_value(env
, insn
->src_reg
)) {
8009 verbose(env
, "R%d pointer comparison prohibited\n",
8013 src_reg
= ®s
[insn
->src_reg
];
8015 if (insn
->src_reg
!= BPF_REG_0
) {
8016 verbose(env
, "BPF_JMP/JMP32 uses reserved fields\n");
8021 /* check src2 operand */
8022 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
8026 dst_reg
= ®s
[insn
->dst_reg
];
8027 is_jmp32
= BPF_CLASS(insn
->code
) == BPF_JMP32
;
8029 if (BPF_SRC(insn
->code
) == BPF_K
) {
8030 pred
= is_branch_taken(dst_reg
, insn
->imm
, opcode
, is_jmp32
);
8031 } else if (src_reg
->type
== SCALAR_VALUE
&&
8032 is_jmp32
&& tnum_is_const(tnum_subreg(src_reg
->var_off
))) {
8033 pred
= is_branch_taken(dst_reg
,
8034 tnum_subreg(src_reg
->var_off
).value
,
8037 } else if (src_reg
->type
== SCALAR_VALUE
&&
8038 !is_jmp32
&& tnum_is_const(src_reg
->var_off
)) {
8039 pred
= is_branch_taken(dst_reg
,
8040 src_reg
->var_off
.value
,
8043 } else if (reg_is_pkt_pointer_any(dst_reg
) &&
8044 reg_is_pkt_pointer_any(src_reg
) &&
8046 pred
= is_pkt_ptr_branch_taken(dst_reg
, src_reg
, opcode
);
8050 /* If we get here with a dst_reg pointer type it is because
8051 * above is_branch_taken() special cased the 0 comparison.
8053 if (!__is_pointer_value(false, dst_reg
))
8054 err
= mark_chain_precision(env
, insn
->dst_reg
);
8055 if (BPF_SRC(insn
->code
) == BPF_X
&& !err
&&
8056 !__is_pointer_value(false, src_reg
))
8057 err
= mark_chain_precision(env
, insn
->src_reg
);
8062 /* only follow the goto, ignore fall-through */
8063 *insn_idx
+= insn
->off
;
8065 } else if (pred
== 0) {
8066 /* only follow fall-through branch, since
8067 * that's where the program will go
8072 other_branch
= push_stack(env
, *insn_idx
+ insn
->off
+ 1, *insn_idx
,
8076 other_branch_regs
= other_branch
->frame
[other_branch
->curframe
]->regs
;
8078 /* detect if we are comparing against a constant value so we can adjust
8079 * our min/max values for our dst register.
8080 * this is only legit if both are scalars (or pointers to the same
8081 * object, I suppose, but we don't support that right now), because
8082 * otherwise the different base pointers mean the offsets aren't
8085 if (BPF_SRC(insn
->code
) == BPF_X
) {
8086 struct bpf_reg_state
*src_reg
= ®s
[insn
->src_reg
];
8088 if (dst_reg
->type
== SCALAR_VALUE
&&
8089 src_reg
->type
== SCALAR_VALUE
) {
8090 if (tnum_is_const(src_reg
->var_off
) ||
8092 tnum_is_const(tnum_subreg(src_reg
->var_off
))))
8093 reg_set_min_max(&other_branch_regs
[insn
->dst_reg
],
8095 src_reg
->var_off
.value
,
8096 tnum_subreg(src_reg
->var_off
).value
,
8098 else if (tnum_is_const(dst_reg
->var_off
) ||
8100 tnum_is_const(tnum_subreg(dst_reg
->var_off
))))
8101 reg_set_min_max_inv(&other_branch_regs
[insn
->src_reg
],
8103 dst_reg
->var_off
.value
,
8104 tnum_subreg(dst_reg
->var_off
).value
,
8106 else if (!is_jmp32
&&
8107 (opcode
== BPF_JEQ
|| opcode
== BPF_JNE
))
8108 /* Comparing for equality, we can combine knowledge */
8109 reg_combine_min_max(&other_branch_regs
[insn
->src_reg
],
8110 &other_branch_regs
[insn
->dst_reg
],
8111 src_reg
, dst_reg
, opcode
);
8113 !WARN_ON_ONCE(src_reg
->id
!= other_branch_regs
[insn
->src_reg
].id
)) {
8114 find_equal_scalars(this_branch
, src_reg
);
8115 find_equal_scalars(other_branch
, &other_branch_regs
[insn
->src_reg
]);
8119 } else if (dst_reg
->type
== SCALAR_VALUE
) {
8120 reg_set_min_max(&other_branch_regs
[insn
->dst_reg
],
8121 dst_reg
, insn
->imm
, (u32
)insn
->imm
,
8125 if (dst_reg
->type
== SCALAR_VALUE
&& dst_reg
->id
&&
8126 !WARN_ON_ONCE(dst_reg
->id
!= other_branch_regs
[insn
->dst_reg
].id
)) {
8127 find_equal_scalars(this_branch
, dst_reg
);
8128 find_equal_scalars(other_branch
, &other_branch_regs
[insn
->dst_reg
]);
8131 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
8132 * NOTE: these optimizations below are related with pointer comparison
8133 * which will never be JMP32.
8135 if (!is_jmp32
&& BPF_SRC(insn
->code
) == BPF_K
&&
8136 insn
->imm
== 0 && (opcode
== BPF_JEQ
|| opcode
== BPF_JNE
) &&
8137 reg_type_may_be_null(dst_reg
->type
)) {
8138 /* Mark all identical registers in each branch as either
8139 * safe or unknown depending R == 0 or R != 0 conditional.
8141 mark_ptr_or_null_regs(this_branch
, insn
->dst_reg
,
8143 mark_ptr_or_null_regs(other_branch
, insn
->dst_reg
,
8145 } else if (!try_match_pkt_pointers(insn
, dst_reg
, ®s
[insn
->src_reg
],
8146 this_branch
, other_branch
) &&
8147 is_pointer_value(env
, insn
->dst_reg
)) {
8148 verbose(env
, "R%d pointer comparison prohibited\n",
8152 if (env
->log
.level
& BPF_LOG_LEVEL
)
8153 print_verifier_state(env
, this_branch
->frame
[this_branch
->curframe
]);
8157 /* verify BPF_LD_IMM64 instruction */
8158 static int check_ld_imm(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
8160 struct bpf_insn_aux_data
*aux
= cur_aux(env
);
8161 struct bpf_reg_state
*regs
= cur_regs(env
);
8162 struct bpf_reg_state
*dst_reg
;
8163 struct bpf_map
*map
;
8166 if (BPF_SIZE(insn
->code
) != BPF_DW
) {
8167 verbose(env
, "invalid BPF_LD_IMM insn\n");
8170 if (insn
->off
!= 0) {
8171 verbose(env
, "BPF_LD_IMM64 uses reserved fields\n");
8175 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP
);
8179 dst_reg
= ®s
[insn
->dst_reg
];
8180 if (insn
->src_reg
== 0) {
8181 u64 imm
= ((u64
)(insn
+ 1)->imm
<< 32) | (u32
)insn
->imm
;
8183 dst_reg
->type
= SCALAR_VALUE
;
8184 __mark_reg_known(®s
[insn
->dst_reg
], imm
);
8188 if (insn
->src_reg
== BPF_PSEUDO_BTF_ID
) {
8189 mark_reg_known_zero(env
, regs
, insn
->dst_reg
);
8191 dst_reg
->type
= aux
->btf_var
.reg_type
;
8192 switch (dst_reg
->type
) {
8194 dst_reg
->mem_size
= aux
->btf_var
.mem_size
;
8197 case PTR_TO_PERCPU_BTF_ID
:
8198 dst_reg
->btf
= aux
->btf_var
.btf
;
8199 dst_reg
->btf_id
= aux
->btf_var
.btf_id
;
8202 verbose(env
, "bpf verifier is misconfigured\n");
8208 map
= env
->used_maps
[aux
->map_index
];
8209 mark_reg_known_zero(env
, regs
, insn
->dst_reg
);
8210 dst_reg
->map_ptr
= map
;
8212 if (insn
->src_reg
== BPF_PSEUDO_MAP_VALUE
) {
8213 dst_reg
->type
= PTR_TO_MAP_VALUE
;
8214 dst_reg
->off
= aux
->map_off
;
8215 if (map_value_has_spin_lock(map
))
8216 dst_reg
->id
= ++env
->id_gen
;
8217 } else if (insn
->src_reg
== BPF_PSEUDO_MAP_FD
) {
8218 dst_reg
->type
= CONST_PTR_TO_MAP
;
8220 verbose(env
, "bpf verifier is misconfigured\n");
8227 static bool may_access_skb(enum bpf_prog_type type
)
8230 case BPF_PROG_TYPE_SOCKET_FILTER
:
8231 case BPF_PROG_TYPE_SCHED_CLS
:
8232 case BPF_PROG_TYPE_SCHED_ACT
:
8239 /* verify safety of LD_ABS|LD_IND instructions:
8240 * - they can only appear in the programs where ctx == skb
8241 * - since they are wrappers of function calls, they scratch R1-R5 registers,
8242 * preserve R6-R9, and store return value into R0
8245 * ctx == skb == R6 == CTX
8248 * SRC == any register
8249 * IMM == 32-bit immediate
8252 * R0 - 8/16/32-bit skb data converted to cpu endianness
8254 static int check_ld_abs(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
8256 struct bpf_reg_state
*regs
= cur_regs(env
);
8257 static const int ctx_reg
= BPF_REG_6
;
8258 u8 mode
= BPF_MODE(insn
->code
);
8261 if (!may_access_skb(resolve_prog_type(env
->prog
))) {
8262 verbose(env
, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
8266 if (!env
->ops
->gen_ld_abs
) {
8267 verbose(env
, "bpf verifier is misconfigured\n");
8271 if (insn
->dst_reg
!= BPF_REG_0
|| insn
->off
!= 0 ||
8272 BPF_SIZE(insn
->code
) == BPF_DW
||
8273 (mode
== BPF_ABS
&& insn
->src_reg
!= BPF_REG_0
)) {
8274 verbose(env
, "BPF_LD_[ABS|IND] uses reserved fields\n");
8278 /* check whether implicit source operand (register R6) is readable */
8279 err
= check_reg_arg(env
, ctx_reg
, SRC_OP
);
8283 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
8284 * gen_ld_abs() may terminate the program at runtime, leading to
8287 err
= check_reference_leak(env
);
8289 verbose(env
, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
8293 if (env
->cur_state
->active_spin_lock
) {
8294 verbose(env
, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
8298 if (regs
[ctx_reg
].type
!= PTR_TO_CTX
) {
8300 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
8304 if (mode
== BPF_IND
) {
8305 /* check explicit source operand */
8306 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
8311 err
= check_ctx_reg(env
, ®s
[ctx_reg
], ctx_reg
);
8315 /* reset caller saved regs to unreadable */
8316 for (i
= 0; i
< CALLER_SAVED_REGS
; i
++) {
8317 mark_reg_not_init(env
, regs
, caller_saved
[i
]);
8318 check_reg_arg(env
, caller_saved
[i
], DST_OP_NO_MARK
);
8321 /* mark destination R0 register as readable, since it contains
8322 * the value fetched from the packet.
8323 * Already marked as written above.
8325 mark_reg_unknown(env
, regs
, BPF_REG_0
);
8326 /* ld_abs load up to 32-bit skb data. */
8327 regs
[BPF_REG_0
].subreg_def
= env
->insn_idx
+ 1;
8331 static int check_return_code(struct bpf_verifier_env
*env
)
8333 struct tnum enforce_attach_type_range
= tnum_unknown
;
8334 const struct bpf_prog
*prog
= env
->prog
;
8335 struct bpf_reg_state
*reg
;
8336 struct tnum range
= tnum_range(0, 1);
8337 enum bpf_prog_type prog_type
= resolve_prog_type(env
->prog
);
8339 const bool is_subprog
= env
->cur_state
->frame
[0]->subprogno
;
8341 /* LSM and struct_ops func-ptr's return type could be "void" */
8343 (prog_type
== BPF_PROG_TYPE_STRUCT_OPS
||
8344 prog_type
== BPF_PROG_TYPE_LSM
) &&
8345 !prog
->aux
->attach_func_proto
->type
)
8348 /* eBPF calling convetion is such that R0 is used
8349 * to return the value from eBPF program.
8350 * Make sure that it's readable at this time
8351 * of bpf_exit, which means that program wrote
8352 * something into it earlier
8354 err
= check_reg_arg(env
, BPF_REG_0
, SRC_OP
);
8358 if (is_pointer_value(env
, BPF_REG_0
)) {
8359 verbose(env
, "R0 leaks addr as return value\n");
8363 reg
= cur_regs(env
) + BPF_REG_0
;
8365 if (reg
->type
!= SCALAR_VALUE
) {
8366 verbose(env
, "At subprogram exit the register R0 is not a scalar value (%s)\n",
8367 reg_type_str
[reg
->type
]);
8373 switch (prog_type
) {
8374 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR
:
8375 if (env
->prog
->expected_attach_type
== BPF_CGROUP_UDP4_RECVMSG
||
8376 env
->prog
->expected_attach_type
== BPF_CGROUP_UDP6_RECVMSG
||
8377 env
->prog
->expected_attach_type
== BPF_CGROUP_INET4_GETPEERNAME
||
8378 env
->prog
->expected_attach_type
== BPF_CGROUP_INET6_GETPEERNAME
||
8379 env
->prog
->expected_attach_type
== BPF_CGROUP_INET4_GETSOCKNAME
||
8380 env
->prog
->expected_attach_type
== BPF_CGROUP_INET6_GETSOCKNAME
)
8381 range
= tnum_range(1, 1);
8383 case BPF_PROG_TYPE_CGROUP_SKB
:
8384 if (env
->prog
->expected_attach_type
== BPF_CGROUP_INET_EGRESS
) {
8385 range
= tnum_range(0, 3);
8386 enforce_attach_type_range
= tnum_range(2, 3);
8389 case BPF_PROG_TYPE_CGROUP_SOCK
:
8390 case BPF_PROG_TYPE_SOCK_OPS
:
8391 case BPF_PROG_TYPE_CGROUP_DEVICE
:
8392 case BPF_PROG_TYPE_CGROUP_SYSCTL
:
8393 case BPF_PROG_TYPE_CGROUP_SOCKOPT
:
8395 case BPF_PROG_TYPE_RAW_TRACEPOINT
:
8396 if (!env
->prog
->aux
->attach_btf_id
)
8398 range
= tnum_const(0);
8400 case BPF_PROG_TYPE_TRACING
:
8401 switch (env
->prog
->expected_attach_type
) {
8402 case BPF_TRACE_FENTRY
:
8403 case BPF_TRACE_FEXIT
:
8404 range
= tnum_const(0);
8406 case BPF_TRACE_RAW_TP
:
8407 case BPF_MODIFY_RETURN
:
8409 case BPF_TRACE_ITER
:
8415 case BPF_PROG_TYPE_SK_LOOKUP
:
8416 range
= tnum_range(SK_DROP
, SK_PASS
);
8418 case BPF_PROG_TYPE_EXT
:
8419 /* freplace program can return anything as its return value
8420 * depends on the to-be-replaced kernel func or bpf program.
8426 if (reg
->type
!= SCALAR_VALUE
) {
8427 verbose(env
, "At program exit the register R0 is not a known value (%s)\n",
8428 reg_type_str
[reg
->type
]);
8432 if (!tnum_in(range
, reg
->var_off
)) {
8435 verbose(env
, "At program exit the register R0 ");
8436 if (!tnum_is_unknown(reg
->var_off
)) {
8437 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
8438 verbose(env
, "has value %s", tn_buf
);
8440 verbose(env
, "has unknown scalar value");
8442 tnum_strn(tn_buf
, sizeof(tn_buf
), range
);
8443 verbose(env
, " should have been in %s\n", tn_buf
);
8447 if (!tnum_is_unknown(enforce_attach_type_range
) &&
8448 tnum_in(enforce_attach_type_range
, reg
->var_off
))
8449 env
->prog
->enforce_expected_attach_type
= 1;
8453 /* non-recursive DFS pseudo code
8454 * 1 procedure DFS-iterative(G,v):
8455 * 2 label v as discovered
8456 * 3 let S be a stack
8458 * 5 while S is not empty
8460 * 7 if t is what we're looking for:
8462 * 9 for all edges e in G.adjacentEdges(t) do
8463 * 10 if edge e is already labelled
8464 * 11 continue with the next edge
8465 * 12 w <- G.adjacentVertex(t,e)
8466 * 13 if vertex w is not discovered and not explored
8467 * 14 label e as tree-edge
8468 * 15 label w as discovered
8471 * 18 else if vertex w is discovered
8472 * 19 label e as back-edge
8474 * 21 // vertex w is explored
8475 * 22 label e as forward- or cross-edge
8476 * 23 label t as explored
8481 * 0x11 - discovered and fall-through edge labelled
8482 * 0x12 - discovered and fall-through and branch edges labelled
8493 static u32
state_htab_size(struct bpf_verifier_env
*env
)
8495 return env
->prog
->len
;
8498 static struct bpf_verifier_state_list
**explored_state(
8499 struct bpf_verifier_env
*env
,
8502 struct bpf_verifier_state
*cur
= env
->cur_state
;
8503 struct bpf_func_state
*state
= cur
->frame
[cur
->curframe
];
8505 return &env
->explored_states
[(idx
^ state
->callsite
) % state_htab_size(env
)];
8508 static void init_explored_state(struct bpf_verifier_env
*env
, int idx
)
8510 env
->insn_aux_data
[idx
].prune_point
= true;
8518 /* t, w, e - match pseudo-code above:
8519 * t - index of current instruction
8520 * w - next instruction
8523 static int push_insn(int t
, int w
, int e
, struct bpf_verifier_env
*env
,
8526 int *insn_stack
= env
->cfg
.insn_stack
;
8527 int *insn_state
= env
->cfg
.insn_state
;
8529 if (e
== FALLTHROUGH
&& insn_state
[t
] >= (DISCOVERED
| FALLTHROUGH
))
8530 return DONE_EXPLORING
;
8532 if (e
== BRANCH
&& insn_state
[t
] >= (DISCOVERED
| BRANCH
))
8533 return DONE_EXPLORING
;
8535 if (w
< 0 || w
>= env
->prog
->len
) {
8536 verbose_linfo(env
, t
, "%d: ", t
);
8537 verbose(env
, "jump out of range from insn %d to %d\n", t
, w
);
8542 /* mark branch target for state pruning */
8543 init_explored_state(env
, w
);
8545 if (insn_state
[w
] == 0) {
8547 insn_state
[t
] = DISCOVERED
| e
;
8548 insn_state
[w
] = DISCOVERED
;
8549 if (env
->cfg
.cur_stack
>= env
->prog
->len
)
8551 insn_stack
[env
->cfg
.cur_stack
++] = w
;
8552 return KEEP_EXPLORING
;
8553 } else if ((insn_state
[w
] & 0xF0) == DISCOVERED
) {
8554 if (loop_ok
&& env
->bpf_capable
)
8555 return DONE_EXPLORING
;
8556 verbose_linfo(env
, t
, "%d: ", t
);
8557 verbose_linfo(env
, w
, "%d: ", w
);
8558 verbose(env
, "back-edge from insn %d to %d\n", t
, w
);
8560 } else if (insn_state
[w
] == EXPLORED
) {
8561 /* forward- or cross-edge */
8562 insn_state
[t
] = DISCOVERED
| e
;
8564 verbose(env
, "insn state internal bug\n");
8567 return DONE_EXPLORING
;
8570 /* Visits the instruction at index t and returns one of the following:
8571 * < 0 - an error occurred
8572 * DONE_EXPLORING - the instruction was fully explored
8573 * KEEP_EXPLORING - there is still work to be done before it is fully explored
8575 static int visit_insn(int t
, int insn_cnt
, struct bpf_verifier_env
*env
)
8577 struct bpf_insn
*insns
= env
->prog
->insnsi
;
8580 /* All non-branch instructions have a single fall-through edge. */
8581 if (BPF_CLASS(insns
[t
].code
) != BPF_JMP
&&
8582 BPF_CLASS(insns
[t
].code
) != BPF_JMP32
)
8583 return push_insn(t
, t
+ 1, FALLTHROUGH
, env
, false);
8585 switch (BPF_OP(insns
[t
].code
)) {
8587 return DONE_EXPLORING
;
8590 ret
= push_insn(t
, t
+ 1, FALLTHROUGH
, env
, false);
8594 if (t
+ 1 < insn_cnt
)
8595 init_explored_state(env
, t
+ 1);
8596 if (insns
[t
].src_reg
== BPF_PSEUDO_CALL
) {
8597 init_explored_state(env
, t
);
8598 ret
= push_insn(t
, t
+ insns
[t
].imm
+ 1, BRANCH
,
8604 if (BPF_SRC(insns
[t
].code
) != BPF_K
)
8607 /* unconditional jump with single edge */
8608 ret
= push_insn(t
, t
+ insns
[t
].off
+ 1, FALLTHROUGH
, env
,
8613 /* unconditional jmp is not a good pruning point,
8614 * but it's marked, since backtracking needs
8615 * to record jmp history in is_state_visited().
8617 init_explored_state(env
, t
+ insns
[t
].off
+ 1);
8618 /* tell verifier to check for equivalent states
8619 * after every call and jump
8621 if (t
+ 1 < insn_cnt
)
8622 init_explored_state(env
, t
+ 1);
8627 /* conditional jump with two edges */
8628 init_explored_state(env
, t
);
8629 ret
= push_insn(t
, t
+ 1, FALLTHROUGH
, env
, true);
8633 return push_insn(t
, t
+ insns
[t
].off
+ 1, BRANCH
, env
, true);
8637 /* non-recursive depth-first-search to detect loops in BPF program
8638 * loop == back-edge in directed graph
8640 static int check_cfg(struct bpf_verifier_env
*env
)
8642 int insn_cnt
= env
->prog
->len
;
8643 int *insn_stack
, *insn_state
;
8647 insn_state
= env
->cfg
.insn_state
= kvcalloc(insn_cnt
, sizeof(int), GFP_KERNEL
);
8651 insn_stack
= env
->cfg
.insn_stack
= kvcalloc(insn_cnt
, sizeof(int), GFP_KERNEL
);
8657 insn_state
[0] = DISCOVERED
; /* mark 1st insn as discovered */
8658 insn_stack
[0] = 0; /* 0 is the first instruction */
8659 env
->cfg
.cur_stack
= 1;
8661 while (env
->cfg
.cur_stack
> 0) {
8662 int t
= insn_stack
[env
->cfg
.cur_stack
- 1];
8664 ret
= visit_insn(t
, insn_cnt
, env
);
8666 case DONE_EXPLORING
:
8667 insn_state
[t
] = EXPLORED
;
8668 env
->cfg
.cur_stack
--;
8670 case KEEP_EXPLORING
:
8674 verbose(env
, "visit_insn internal bug\n");
8681 if (env
->cfg
.cur_stack
< 0) {
8682 verbose(env
, "pop stack internal bug\n");
8687 for (i
= 0; i
< insn_cnt
; i
++) {
8688 if (insn_state
[i
] != EXPLORED
) {
8689 verbose(env
, "unreachable insn %d\n", i
);
8694 ret
= 0; /* cfg looks good */
8699 env
->cfg
.insn_state
= env
->cfg
.insn_stack
= NULL
;
8703 static int check_abnormal_return(struct bpf_verifier_env
*env
)
8707 for (i
= 1; i
< env
->subprog_cnt
; i
++) {
8708 if (env
->subprog_info
[i
].has_ld_abs
) {
8709 verbose(env
, "LD_ABS is not allowed in subprogs without BTF\n");
8712 if (env
->subprog_info
[i
].has_tail_call
) {
8713 verbose(env
, "tail_call is not allowed in subprogs without BTF\n");
8720 /* The minimum supported BTF func info size */
8721 #define MIN_BPF_FUNCINFO_SIZE 8
8722 #define MAX_FUNCINFO_REC_SIZE 252
8724 static int check_btf_func(struct bpf_verifier_env
*env
,
8725 const union bpf_attr
*attr
,
8726 union bpf_attr __user
*uattr
)
8728 const struct btf_type
*type
, *func_proto
, *ret_type
;
8729 u32 i
, nfuncs
, urec_size
, min_size
;
8730 u32 krec_size
= sizeof(struct bpf_func_info
);
8731 struct bpf_func_info
*krecord
;
8732 struct bpf_func_info_aux
*info_aux
= NULL
;
8733 struct bpf_prog
*prog
;
8734 const struct btf
*btf
;
8735 void __user
*urecord
;
8736 u32 prev_offset
= 0;
8740 nfuncs
= attr
->func_info_cnt
;
8742 if (check_abnormal_return(env
))
8747 if (nfuncs
!= env
->subprog_cnt
) {
8748 verbose(env
, "number of funcs in func_info doesn't match number of subprogs\n");
8752 urec_size
= attr
->func_info_rec_size
;
8753 if (urec_size
< MIN_BPF_FUNCINFO_SIZE
||
8754 urec_size
> MAX_FUNCINFO_REC_SIZE
||
8755 urec_size
% sizeof(u32
)) {
8756 verbose(env
, "invalid func info rec size %u\n", urec_size
);
8761 btf
= prog
->aux
->btf
;
8763 urecord
= u64_to_user_ptr(attr
->func_info
);
8764 min_size
= min_t(u32
, krec_size
, urec_size
);
8766 krecord
= kvcalloc(nfuncs
, krec_size
, GFP_KERNEL
| __GFP_NOWARN
);
8769 info_aux
= kcalloc(nfuncs
, sizeof(*info_aux
), GFP_KERNEL
| __GFP_NOWARN
);
8773 for (i
= 0; i
< nfuncs
; i
++) {
8774 ret
= bpf_check_uarg_tail_zero(urecord
, krec_size
, urec_size
);
8776 if (ret
== -E2BIG
) {
8777 verbose(env
, "nonzero tailing record in func info");
8778 /* set the size kernel expects so loader can zero
8779 * out the rest of the record.
8781 if (put_user(min_size
, &uattr
->func_info_rec_size
))
8787 if (copy_from_user(&krecord
[i
], urecord
, min_size
)) {
8792 /* check insn_off */
8795 if (krecord
[i
].insn_off
) {
8797 "nonzero insn_off %u for the first func info record",
8798 krecord
[i
].insn_off
);
8801 } else if (krecord
[i
].insn_off
<= prev_offset
) {
8803 "same or smaller insn offset (%u) than previous func info record (%u)",
8804 krecord
[i
].insn_off
, prev_offset
);
8808 if (env
->subprog_info
[i
].start
!= krecord
[i
].insn_off
) {
8809 verbose(env
, "func_info BTF section doesn't match subprog layout in BPF program\n");
8814 type
= btf_type_by_id(btf
, krecord
[i
].type_id
);
8815 if (!type
|| !btf_type_is_func(type
)) {
8816 verbose(env
, "invalid type id %d in func info",
8817 krecord
[i
].type_id
);
8820 info_aux
[i
].linkage
= BTF_INFO_VLEN(type
->info
);
8822 func_proto
= btf_type_by_id(btf
, type
->type
);
8823 if (unlikely(!func_proto
|| !btf_type_is_func_proto(func_proto
)))
8824 /* btf_func_check() already verified it during BTF load */
8826 ret_type
= btf_type_skip_modifiers(btf
, func_proto
->type
, NULL
);
8828 btf_type_is_small_int(ret_type
) || btf_type_is_enum(ret_type
);
8829 if (i
&& !scalar_return
&& env
->subprog_info
[i
].has_ld_abs
) {
8830 verbose(env
, "LD_ABS is only allowed in functions that return 'int'.\n");
8833 if (i
&& !scalar_return
&& env
->subprog_info
[i
].has_tail_call
) {
8834 verbose(env
, "tail_call is only allowed in functions that return 'int'.\n");
8838 prev_offset
= krecord
[i
].insn_off
;
8839 urecord
+= urec_size
;
8842 prog
->aux
->func_info
= krecord
;
8843 prog
->aux
->func_info_cnt
= nfuncs
;
8844 prog
->aux
->func_info_aux
= info_aux
;
8853 static void adjust_btf_func(struct bpf_verifier_env
*env
)
8855 struct bpf_prog_aux
*aux
= env
->prog
->aux
;
8858 if (!aux
->func_info
)
8861 for (i
= 0; i
< env
->subprog_cnt
; i
++)
8862 aux
->func_info
[i
].insn_off
= env
->subprog_info
[i
].start
;
8865 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
8866 sizeof(((struct bpf_line_info *)(0))->line_col))
8867 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
8869 static int check_btf_line(struct bpf_verifier_env
*env
,
8870 const union bpf_attr
*attr
,
8871 union bpf_attr __user
*uattr
)
8873 u32 i
, s
, nr_linfo
, ncopy
, expected_size
, rec_size
, prev_offset
= 0;
8874 struct bpf_subprog_info
*sub
;
8875 struct bpf_line_info
*linfo
;
8876 struct bpf_prog
*prog
;
8877 const struct btf
*btf
;
8878 void __user
*ulinfo
;
8881 nr_linfo
= attr
->line_info_cnt
;
8885 rec_size
= attr
->line_info_rec_size
;
8886 if (rec_size
< MIN_BPF_LINEINFO_SIZE
||
8887 rec_size
> MAX_LINEINFO_REC_SIZE
||
8888 rec_size
& (sizeof(u32
) - 1))
8891 /* Need to zero it in case the userspace may
8892 * pass in a smaller bpf_line_info object.
8894 linfo
= kvcalloc(nr_linfo
, sizeof(struct bpf_line_info
),
8895 GFP_KERNEL
| __GFP_NOWARN
);
8900 btf
= prog
->aux
->btf
;
8903 sub
= env
->subprog_info
;
8904 ulinfo
= u64_to_user_ptr(attr
->line_info
);
8905 expected_size
= sizeof(struct bpf_line_info
);
8906 ncopy
= min_t(u32
, expected_size
, rec_size
);
8907 for (i
= 0; i
< nr_linfo
; i
++) {
8908 err
= bpf_check_uarg_tail_zero(ulinfo
, expected_size
, rec_size
);
8910 if (err
== -E2BIG
) {
8911 verbose(env
, "nonzero tailing record in line_info");
8912 if (put_user(expected_size
,
8913 &uattr
->line_info_rec_size
))
8919 if (copy_from_user(&linfo
[i
], ulinfo
, ncopy
)) {
8925 * Check insn_off to ensure
8926 * 1) strictly increasing AND
8927 * 2) bounded by prog->len
8929 * The linfo[0].insn_off == 0 check logically falls into
8930 * the later "missing bpf_line_info for func..." case
8931 * because the first linfo[0].insn_off must be the
8932 * first sub also and the first sub must have
8933 * subprog_info[0].start == 0.
8935 if ((i
&& linfo
[i
].insn_off
<= prev_offset
) ||
8936 linfo
[i
].insn_off
>= prog
->len
) {
8937 verbose(env
, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
8938 i
, linfo
[i
].insn_off
, prev_offset
,
8944 if (!prog
->insnsi
[linfo
[i
].insn_off
].code
) {
8946 "Invalid insn code at line_info[%u].insn_off\n",
8952 if (!btf_name_by_offset(btf
, linfo
[i
].line_off
) ||
8953 !btf_name_by_offset(btf
, linfo
[i
].file_name_off
)) {
8954 verbose(env
, "Invalid line_info[%u].line_off or .file_name_off\n", i
);
8959 if (s
!= env
->subprog_cnt
) {
8960 if (linfo
[i
].insn_off
== sub
[s
].start
) {
8961 sub
[s
].linfo_idx
= i
;
8963 } else if (sub
[s
].start
< linfo
[i
].insn_off
) {
8964 verbose(env
, "missing bpf_line_info for func#%u\n", s
);
8970 prev_offset
= linfo
[i
].insn_off
;
8974 if (s
!= env
->subprog_cnt
) {
8975 verbose(env
, "missing bpf_line_info for %u funcs starting from func#%u\n",
8976 env
->subprog_cnt
- s
, s
);
8981 prog
->aux
->linfo
= linfo
;
8982 prog
->aux
->nr_linfo
= nr_linfo
;
8991 static int check_btf_info(struct bpf_verifier_env
*env
,
8992 const union bpf_attr
*attr
,
8993 union bpf_attr __user
*uattr
)
8998 if (!attr
->func_info_cnt
&& !attr
->line_info_cnt
) {
8999 if (check_abnormal_return(env
))
9004 btf
= btf_get_by_fd(attr
->prog_btf_fd
);
9006 return PTR_ERR(btf
);
9007 if (btf_is_kernel(btf
)) {
9011 env
->prog
->aux
->btf
= btf
;
9013 err
= check_btf_func(env
, attr
, uattr
);
9017 err
= check_btf_line(env
, attr
, uattr
);
9024 /* check %cur's range satisfies %old's */
9025 static bool range_within(struct bpf_reg_state
*old
,
9026 struct bpf_reg_state
*cur
)
9028 return old
->umin_value
<= cur
->umin_value
&&
9029 old
->umax_value
>= cur
->umax_value
&&
9030 old
->smin_value
<= cur
->smin_value
&&
9031 old
->smax_value
>= cur
->smax_value
&&
9032 old
->u32_min_value
<= cur
->u32_min_value
&&
9033 old
->u32_max_value
>= cur
->u32_max_value
&&
9034 old
->s32_min_value
<= cur
->s32_min_value
&&
9035 old
->s32_max_value
>= cur
->s32_max_value
;
9038 /* Maximum number of register states that can exist at once */
9039 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
9045 /* If in the old state two registers had the same id, then they need to have
9046 * the same id in the new state as well. But that id could be different from
9047 * the old state, so we need to track the mapping from old to new ids.
9048 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
9049 * regs with old id 5 must also have new id 9 for the new state to be safe. But
9050 * regs with a different old id could still have new id 9, we don't care about
9052 * So we look through our idmap to see if this old id has been seen before. If
9053 * so, we require the new id to match; otherwise, we add the id pair to the map.
9055 static bool check_ids(u32 old_id
, u32 cur_id
, struct idpair
*idmap
)
9059 for (i
= 0; i
< ID_MAP_SIZE
; i
++) {
9060 if (!idmap
[i
].old
) {
9061 /* Reached an empty slot; haven't seen this id before */
9062 idmap
[i
].old
= old_id
;
9063 idmap
[i
].cur
= cur_id
;
9066 if (idmap
[i
].old
== old_id
)
9067 return idmap
[i
].cur
== cur_id
;
9069 /* We ran out of idmap slots, which should be impossible */
9074 static void clean_func_state(struct bpf_verifier_env
*env
,
9075 struct bpf_func_state
*st
)
9077 enum bpf_reg_liveness live
;
9080 for (i
= 0; i
< BPF_REG_FP
; i
++) {
9081 live
= st
->regs
[i
].live
;
9082 /* liveness must not touch this register anymore */
9083 st
->regs
[i
].live
|= REG_LIVE_DONE
;
9084 if (!(live
& REG_LIVE_READ
))
9085 /* since the register is unused, clear its state
9086 * to make further comparison simpler
9088 __mark_reg_not_init(env
, &st
->regs
[i
]);
9091 for (i
= 0; i
< st
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
9092 live
= st
->stack
[i
].spilled_ptr
.live
;
9093 /* liveness must not touch this stack slot anymore */
9094 st
->stack
[i
].spilled_ptr
.live
|= REG_LIVE_DONE
;
9095 if (!(live
& REG_LIVE_READ
)) {
9096 __mark_reg_not_init(env
, &st
->stack
[i
].spilled_ptr
);
9097 for (j
= 0; j
< BPF_REG_SIZE
; j
++)
9098 st
->stack
[i
].slot_type
[j
] = STACK_INVALID
;
9103 static void clean_verifier_state(struct bpf_verifier_env
*env
,
9104 struct bpf_verifier_state
*st
)
9108 if (st
->frame
[0]->regs
[0].live
& REG_LIVE_DONE
)
9109 /* all regs in this state in all frames were already marked */
9112 for (i
= 0; i
<= st
->curframe
; i
++)
9113 clean_func_state(env
, st
->frame
[i
]);
9116 /* the parentage chains form a tree.
9117 * the verifier states are added to state lists at given insn and
9118 * pushed into state stack for future exploration.
9119 * when the verifier reaches bpf_exit insn some of the verifer states
9120 * stored in the state lists have their final liveness state already,
9121 * but a lot of states will get revised from liveness point of view when
9122 * the verifier explores other branches.
9125 * 2: if r1 == 100 goto pc+1
9128 * when the verifier reaches exit insn the register r0 in the state list of
9129 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
9130 * of insn 2 and goes exploring further. At the insn 4 it will walk the
9131 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
9133 * Since the verifier pushes the branch states as it sees them while exploring
9134 * the program the condition of walking the branch instruction for the second
9135 * time means that all states below this branch were already explored and
9136 * their final liveness markes are already propagated.
9137 * Hence when the verifier completes the search of state list in is_state_visited()
9138 * we can call this clean_live_states() function to mark all liveness states
9139 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
9141 * This function also clears the registers and stack for states that !READ
9142 * to simplify state merging.
9144 * Important note here that walking the same branch instruction in the callee
9145 * doesn't meant that the states are DONE. The verifier has to compare
9148 static void clean_live_states(struct bpf_verifier_env
*env
, int insn
,
9149 struct bpf_verifier_state
*cur
)
9151 struct bpf_verifier_state_list
*sl
;
9154 sl
= *explored_state(env
, insn
);
9156 if (sl
->state
.branches
)
9158 if (sl
->state
.insn_idx
!= insn
||
9159 sl
->state
.curframe
!= cur
->curframe
)
9161 for (i
= 0; i
<= cur
->curframe
; i
++)
9162 if (sl
->state
.frame
[i
]->callsite
!= cur
->frame
[i
]->callsite
)
9164 clean_verifier_state(env
, &sl
->state
);
9170 /* Returns true if (rold safe implies rcur safe) */
9171 static bool regsafe(struct bpf_reg_state
*rold
, struct bpf_reg_state
*rcur
,
9172 struct idpair
*idmap
)
9176 if (!(rold
->live
& REG_LIVE_READ
))
9177 /* explored state didn't use this */
9180 equal
= memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, parent
)) == 0;
9182 if (rold
->type
== PTR_TO_STACK
)
9183 /* two stack pointers are equal only if they're pointing to
9184 * the same stack frame, since fp-8 in foo != fp-8 in bar
9186 return equal
&& rold
->frameno
== rcur
->frameno
;
9191 if (rold
->type
== NOT_INIT
)
9192 /* explored state can't have used this */
9194 if (rcur
->type
== NOT_INIT
)
9196 switch (rold
->type
) {
9198 if (rcur
->type
== SCALAR_VALUE
) {
9199 if (!rold
->precise
&& !rcur
->precise
)
9201 /* new val must satisfy old val knowledge */
9202 return range_within(rold
, rcur
) &&
9203 tnum_in(rold
->var_off
, rcur
->var_off
);
9205 /* We're trying to use a pointer in place of a scalar.
9206 * Even if the scalar was unbounded, this could lead to
9207 * pointer leaks because scalars are allowed to leak
9208 * while pointers are not. We could make this safe in
9209 * special cases if root is calling us, but it's
9210 * probably not worth the hassle.
9214 case PTR_TO_MAP_VALUE
:
9215 /* If the new min/max/var_off satisfy the old ones and
9216 * everything else matches, we are OK.
9217 * 'id' is not compared, since it's only used for maps with
9218 * bpf_spin_lock inside map element and in such cases if
9219 * the rest of the prog is valid for one map element then
9220 * it's valid for all map elements regardless of the key
9221 * used in bpf_map_lookup()
9223 return memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, id
)) == 0 &&
9224 range_within(rold
, rcur
) &&
9225 tnum_in(rold
->var_off
, rcur
->var_off
);
9226 case PTR_TO_MAP_VALUE_OR_NULL
:
9227 /* a PTR_TO_MAP_VALUE could be safe to use as a
9228 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
9229 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
9230 * checked, doing so could have affected others with the same
9231 * id, and we can't check for that because we lost the id when
9232 * we converted to a PTR_TO_MAP_VALUE.
9234 if (rcur
->type
!= PTR_TO_MAP_VALUE_OR_NULL
)
9236 if (memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, id
)))
9238 /* Check our ids match any regs they're supposed to */
9239 return check_ids(rold
->id
, rcur
->id
, idmap
);
9240 case PTR_TO_PACKET_META
:
9242 if (rcur
->type
!= rold
->type
)
9244 /* We must have at least as much range as the old ptr
9245 * did, so that any accesses which were safe before are
9246 * still safe. This is true even if old range < old off,
9247 * since someone could have accessed through (ptr - k), or
9248 * even done ptr -= k in a register, to get a safe access.
9250 if (rold
->range
> rcur
->range
)
9252 /* If the offsets don't match, we can't trust our alignment;
9253 * nor can we be sure that we won't fall out of range.
9255 if (rold
->off
!= rcur
->off
)
9257 /* id relations must be preserved */
9258 if (rold
->id
&& !check_ids(rold
->id
, rcur
->id
, idmap
))
9260 /* new val must satisfy old val knowledge */
9261 return range_within(rold
, rcur
) &&
9262 tnum_in(rold
->var_off
, rcur
->var_off
);
9264 case CONST_PTR_TO_MAP
:
9265 case PTR_TO_PACKET_END
:
9266 case PTR_TO_FLOW_KEYS
:
9268 case PTR_TO_SOCKET_OR_NULL
:
9269 case PTR_TO_SOCK_COMMON
:
9270 case PTR_TO_SOCK_COMMON_OR_NULL
:
9271 case PTR_TO_TCP_SOCK
:
9272 case PTR_TO_TCP_SOCK_OR_NULL
:
9273 case PTR_TO_XDP_SOCK
:
9274 /* Only valid matches are exact, which memcmp() above
9275 * would have accepted
9278 /* Don't know what's going on, just say it's not safe */
9282 /* Shouldn't get here; if we do, say it's not safe */
9287 static bool stacksafe(struct bpf_func_state
*old
,
9288 struct bpf_func_state
*cur
,
9289 struct idpair
*idmap
)
9293 /* walk slots of the explored stack and ignore any additional
9294 * slots in the current stack, since explored(safe) state
9297 for (i
= 0; i
< old
->allocated_stack
; i
++) {
9298 spi
= i
/ BPF_REG_SIZE
;
9300 if (!(old
->stack
[spi
].spilled_ptr
.live
& REG_LIVE_READ
)) {
9301 i
+= BPF_REG_SIZE
- 1;
9302 /* explored state didn't use this */
9306 if (old
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
] == STACK_INVALID
)
9309 /* explored stack has more populated slots than current stack
9310 * and these slots were used
9312 if (i
>= cur
->allocated_stack
)
9315 /* if old state was safe with misc data in the stack
9316 * it will be safe with zero-initialized stack.
9317 * The opposite is not true
9319 if (old
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
] == STACK_MISC
&&
9320 cur
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
] == STACK_ZERO
)
9322 if (old
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
] !=
9323 cur
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
])
9324 /* Ex: old explored (safe) state has STACK_SPILL in
9325 * this stack slot, but current has STACK_MISC ->
9326 * this verifier states are not equivalent,
9327 * return false to continue verification of this path
9330 if (i
% BPF_REG_SIZE
)
9332 if (old
->stack
[spi
].slot_type
[0] != STACK_SPILL
)
9334 if (!regsafe(&old
->stack
[spi
].spilled_ptr
,
9335 &cur
->stack
[spi
].spilled_ptr
,
9337 /* when explored and current stack slot are both storing
9338 * spilled registers, check that stored pointers types
9339 * are the same as well.
9340 * Ex: explored safe path could have stored
9341 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
9342 * but current path has stored:
9343 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
9344 * such verifier states are not equivalent.
9345 * return false to continue verification of this path
9352 static bool refsafe(struct bpf_func_state
*old
, struct bpf_func_state
*cur
)
9354 if (old
->acquired_refs
!= cur
->acquired_refs
)
9356 return !memcmp(old
->refs
, cur
->refs
,
9357 sizeof(*old
->refs
) * old
->acquired_refs
);
9360 /* compare two verifier states
9362 * all states stored in state_list are known to be valid, since
9363 * verifier reached 'bpf_exit' instruction through them
9365 * this function is called when verifier exploring different branches of
9366 * execution popped from the state stack. If it sees an old state that has
9367 * more strict register state and more strict stack state then this execution
9368 * branch doesn't need to be explored further, since verifier already
9369 * concluded that more strict state leads to valid finish.
9371 * Therefore two states are equivalent if register state is more conservative
9372 * and explored stack state is more conservative than the current one.
9375 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
9376 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
9378 * In other words if current stack state (one being explored) has more
9379 * valid slots than old one that already passed validation, it means
9380 * the verifier can stop exploring and conclude that current state is valid too
9382 * Similarly with registers. If explored state has register type as invalid
9383 * whereas register type in current state is meaningful, it means that
9384 * the current state will reach 'bpf_exit' instruction safely
9386 static bool func_states_equal(struct bpf_func_state
*old
,
9387 struct bpf_func_state
*cur
)
9389 struct idpair
*idmap
;
9393 idmap
= kcalloc(ID_MAP_SIZE
, sizeof(struct idpair
), GFP_KERNEL
);
9394 /* If we failed to allocate the idmap, just say it's not safe */
9398 for (i
= 0; i
< MAX_BPF_REG
; i
++) {
9399 if (!regsafe(&old
->regs
[i
], &cur
->regs
[i
], idmap
))
9403 if (!stacksafe(old
, cur
, idmap
))
9406 if (!refsafe(old
, cur
))
9414 static bool states_equal(struct bpf_verifier_env
*env
,
9415 struct bpf_verifier_state
*old
,
9416 struct bpf_verifier_state
*cur
)
9420 if (old
->curframe
!= cur
->curframe
)
9423 /* Verification state from speculative execution simulation
9424 * must never prune a non-speculative execution one.
9426 if (old
->speculative
&& !cur
->speculative
)
9429 if (old
->active_spin_lock
!= cur
->active_spin_lock
)
9432 /* for states to be equal callsites have to be the same
9433 * and all frame states need to be equivalent
9435 for (i
= 0; i
<= old
->curframe
; i
++) {
9436 if (old
->frame
[i
]->callsite
!= cur
->frame
[i
]->callsite
)
9438 if (!func_states_equal(old
->frame
[i
], cur
->frame
[i
]))
9444 /* Return 0 if no propagation happened. Return negative error code if error
9445 * happened. Otherwise, return the propagated bit.
9447 static int propagate_liveness_reg(struct bpf_verifier_env
*env
,
9448 struct bpf_reg_state
*reg
,
9449 struct bpf_reg_state
*parent_reg
)
9451 u8 parent_flag
= parent_reg
->live
& REG_LIVE_READ
;
9452 u8 flag
= reg
->live
& REG_LIVE_READ
;
9455 /* When comes here, read flags of PARENT_REG or REG could be any of
9456 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
9457 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
9459 if (parent_flag
== REG_LIVE_READ64
||
9460 /* Or if there is no read flag from REG. */
9462 /* Or if the read flag from REG is the same as PARENT_REG. */
9463 parent_flag
== flag
)
9466 err
= mark_reg_read(env
, reg
, parent_reg
, flag
);
9473 /* A write screens off any subsequent reads; but write marks come from the
9474 * straight-line code between a state and its parent. When we arrive at an
9475 * equivalent state (jump target or such) we didn't arrive by the straight-line
9476 * code, so read marks in the state must propagate to the parent regardless
9477 * of the state's write marks. That's what 'parent == state->parent' comparison
9478 * in mark_reg_read() is for.
9480 static int propagate_liveness(struct bpf_verifier_env
*env
,
9481 const struct bpf_verifier_state
*vstate
,
9482 struct bpf_verifier_state
*vparent
)
9484 struct bpf_reg_state
*state_reg
, *parent_reg
;
9485 struct bpf_func_state
*state
, *parent
;
9486 int i
, frame
, err
= 0;
9488 if (vparent
->curframe
!= vstate
->curframe
) {
9489 WARN(1, "propagate_live: parent frame %d current frame %d\n",
9490 vparent
->curframe
, vstate
->curframe
);
9493 /* Propagate read liveness of registers... */
9494 BUILD_BUG_ON(BPF_REG_FP
+ 1 != MAX_BPF_REG
);
9495 for (frame
= 0; frame
<= vstate
->curframe
; frame
++) {
9496 parent
= vparent
->frame
[frame
];
9497 state
= vstate
->frame
[frame
];
9498 parent_reg
= parent
->regs
;
9499 state_reg
= state
->regs
;
9500 /* We don't need to worry about FP liveness, it's read-only */
9501 for (i
= frame
< vstate
->curframe
? BPF_REG_6
: 0; i
< BPF_REG_FP
; i
++) {
9502 err
= propagate_liveness_reg(env
, &state_reg
[i
],
9506 if (err
== REG_LIVE_READ64
)
9507 mark_insn_zext(env
, &parent_reg
[i
]);
9510 /* Propagate stack slots. */
9511 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
&&
9512 i
< parent
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
9513 parent_reg
= &parent
->stack
[i
].spilled_ptr
;
9514 state_reg
= &state
->stack
[i
].spilled_ptr
;
9515 err
= propagate_liveness_reg(env
, state_reg
,
9524 /* find precise scalars in the previous equivalent state and
9525 * propagate them into the current state
9527 static int propagate_precision(struct bpf_verifier_env
*env
,
9528 const struct bpf_verifier_state
*old
)
9530 struct bpf_reg_state
*state_reg
;
9531 struct bpf_func_state
*state
;
9534 state
= old
->frame
[old
->curframe
];
9535 state_reg
= state
->regs
;
9536 for (i
= 0; i
< BPF_REG_FP
; i
++, state_reg
++) {
9537 if (state_reg
->type
!= SCALAR_VALUE
||
9538 !state_reg
->precise
)
9540 if (env
->log
.level
& BPF_LOG_LEVEL2
)
9541 verbose(env
, "propagating r%d\n", i
);
9542 err
= mark_chain_precision(env
, i
);
9547 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
9548 if (state
->stack
[i
].slot_type
[0] != STACK_SPILL
)
9550 state_reg
= &state
->stack
[i
].spilled_ptr
;
9551 if (state_reg
->type
!= SCALAR_VALUE
||
9552 !state_reg
->precise
)
9554 if (env
->log
.level
& BPF_LOG_LEVEL2
)
9555 verbose(env
, "propagating fp%d\n",
9556 (-i
- 1) * BPF_REG_SIZE
);
9557 err
= mark_chain_precision_stack(env
, i
);
9564 static bool states_maybe_looping(struct bpf_verifier_state
*old
,
9565 struct bpf_verifier_state
*cur
)
9567 struct bpf_func_state
*fold
, *fcur
;
9568 int i
, fr
= cur
->curframe
;
9570 if (old
->curframe
!= fr
)
9573 fold
= old
->frame
[fr
];
9574 fcur
= cur
->frame
[fr
];
9575 for (i
= 0; i
< MAX_BPF_REG
; i
++)
9576 if (memcmp(&fold
->regs
[i
], &fcur
->regs
[i
],
9577 offsetof(struct bpf_reg_state
, parent
)))
9583 static int is_state_visited(struct bpf_verifier_env
*env
, int insn_idx
)
9585 struct bpf_verifier_state_list
*new_sl
;
9586 struct bpf_verifier_state_list
*sl
, **pprev
;
9587 struct bpf_verifier_state
*cur
= env
->cur_state
, *new;
9588 int i
, j
, err
, states_cnt
= 0;
9589 bool add_new_state
= env
->test_state_freq
? true : false;
9591 cur
->last_insn_idx
= env
->prev_insn_idx
;
9592 if (!env
->insn_aux_data
[insn_idx
].prune_point
)
9593 /* this 'insn_idx' instruction wasn't marked, so we will not
9594 * be doing state search here
9598 /* bpf progs typically have pruning point every 4 instructions
9599 * http://vger.kernel.org/bpfconf2019.html#session-1
9600 * Do not add new state for future pruning if the verifier hasn't seen
9601 * at least 2 jumps and at least 8 instructions.
9602 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
9603 * In tests that amounts to up to 50% reduction into total verifier
9604 * memory consumption and 20% verifier time speedup.
9606 if (env
->jmps_processed
- env
->prev_jmps_processed
>= 2 &&
9607 env
->insn_processed
- env
->prev_insn_processed
>= 8)
9608 add_new_state
= true;
9610 pprev
= explored_state(env
, insn_idx
);
9613 clean_live_states(env
, insn_idx
, cur
);
9617 if (sl
->state
.insn_idx
!= insn_idx
)
9619 if (sl
->state
.branches
) {
9620 if (states_maybe_looping(&sl
->state
, cur
) &&
9621 states_equal(env
, &sl
->state
, cur
)) {
9622 verbose_linfo(env
, insn_idx
, "; ");
9623 verbose(env
, "infinite loop detected at insn %d\n", insn_idx
);
9626 /* if the verifier is processing a loop, avoid adding new state
9627 * too often, since different loop iterations have distinct
9628 * states and may not help future pruning.
9629 * This threshold shouldn't be too low to make sure that
9630 * a loop with large bound will be rejected quickly.
9631 * The most abusive loop will be:
9633 * if r1 < 1000000 goto pc-2
9634 * 1M insn_procssed limit / 100 == 10k peak states.
9635 * This threshold shouldn't be too high either, since states
9636 * at the end of the loop are likely to be useful in pruning.
9638 if (env
->jmps_processed
- env
->prev_jmps_processed
< 20 &&
9639 env
->insn_processed
- env
->prev_insn_processed
< 100)
9640 add_new_state
= false;
9643 if (states_equal(env
, &sl
->state
, cur
)) {
9645 /* reached equivalent register/stack state,
9647 * Registers read by the continuation are read by us.
9648 * If we have any write marks in env->cur_state, they
9649 * will prevent corresponding reads in the continuation
9650 * from reaching our parent (an explored_state). Our
9651 * own state will get the read marks recorded, but
9652 * they'll be immediately forgotten as we're pruning
9653 * this state and will pop a new one.
9655 err
= propagate_liveness(env
, &sl
->state
, cur
);
9657 /* if previous state reached the exit with precision and
9658 * current state is equivalent to it (except precsion marks)
9659 * the precision needs to be propagated back in
9660 * the current state.
9662 err
= err
? : push_jmp_history(env
, cur
);
9663 err
= err
? : propagate_precision(env
, &sl
->state
);
9669 /* when new state is not going to be added do not increase miss count.
9670 * Otherwise several loop iterations will remove the state
9671 * recorded earlier. The goal of these heuristics is to have
9672 * states from some iterations of the loop (some in the beginning
9673 * and some at the end) to help pruning.
9677 /* heuristic to determine whether this state is beneficial
9678 * to keep checking from state equivalence point of view.
9679 * Higher numbers increase max_states_per_insn and verification time,
9680 * but do not meaningfully decrease insn_processed.
9682 if (sl
->miss_cnt
> sl
->hit_cnt
* 3 + 3) {
9683 /* the state is unlikely to be useful. Remove it to
9684 * speed up verification
9687 if (sl
->state
.frame
[0]->regs
[0].live
& REG_LIVE_DONE
) {
9688 u32 br
= sl
->state
.branches
;
9691 "BUG live_done but branches_to_explore %d\n",
9693 free_verifier_state(&sl
->state
, false);
9697 /* cannot free this state, since parentage chain may
9698 * walk it later. Add it for free_list instead to
9699 * be freed at the end of verification
9701 sl
->next
= env
->free_list
;
9702 env
->free_list
= sl
;
9712 if (env
->max_states_per_insn
< states_cnt
)
9713 env
->max_states_per_insn
= states_cnt
;
9715 if (!env
->bpf_capable
&& states_cnt
> BPF_COMPLEXITY_LIMIT_STATES
)
9716 return push_jmp_history(env
, cur
);
9719 return push_jmp_history(env
, cur
);
9721 /* There were no equivalent states, remember the current one.
9722 * Technically the current state is not proven to be safe yet,
9723 * but it will either reach outer most bpf_exit (which means it's safe)
9724 * or it will be rejected. When there are no loops the verifier won't be
9725 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
9726 * again on the way to bpf_exit.
9727 * When looping the sl->state.branches will be > 0 and this state
9728 * will not be considered for equivalence until branches == 0.
9730 new_sl
= kzalloc(sizeof(struct bpf_verifier_state_list
), GFP_KERNEL
);
9733 env
->total_states
++;
9735 env
->prev_jmps_processed
= env
->jmps_processed
;
9736 env
->prev_insn_processed
= env
->insn_processed
;
9738 /* add new state to the head of linked list */
9739 new = &new_sl
->state
;
9740 err
= copy_verifier_state(new, cur
);
9742 free_verifier_state(new, false);
9746 new->insn_idx
= insn_idx
;
9747 WARN_ONCE(new->branches
!= 1,
9748 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches
, insn_idx
);
9751 cur
->first_insn_idx
= insn_idx
;
9752 clear_jmp_history(cur
);
9753 new_sl
->next
= *explored_state(env
, insn_idx
);
9754 *explored_state(env
, insn_idx
) = new_sl
;
9755 /* connect new state to parentage chain. Current frame needs all
9756 * registers connected. Only r6 - r9 of the callers are alive (pushed
9757 * to the stack implicitly by JITs) so in callers' frames connect just
9758 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
9759 * the state of the call instruction (with WRITTEN set), and r0 comes
9760 * from callee with its full parentage chain, anyway.
9762 /* clear write marks in current state: the writes we did are not writes
9763 * our child did, so they don't screen off its reads from us.
9764 * (There are no read marks in current state, because reads always mark
9765 * their parent and current state never has children yet. Only
9766 * explored_states can get read marks.)
9768 for (j
= 0; j
<= cur
->curframe
; j
++) {
9769 for (i
= j
< cur
->curframe
? BPF_REG_6
: 0; i
< BPF_REG_FP
; i
++)
9770 cur
->frame
[j
]->regs
[i
].parent
= &new->frame
[j
]->regs
[i
];
9771 for (i
= 0; i
< BPF_REG_FP
; i
++)
9772 cur
->frame
[j
]->regs
[i
].live
= REG_LIVE_NONE
;
9775 /* all stack frames are accessible from callee, clear them all */
9776 for (j
= 0; j
<= cur
->curframe
; j
++) {
9777 struct bpf_func_state
*frame
= cur
->frame
[j
];
9778 struct bpf_func_state
*newframe
= new->frame
[j
];
9780 for (i
= 0; i
< frame
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
9781 frame
->stack
[i
].spilled_ptr
.live
= REG_LIVE_NONE
;
9782 frame
->stack
[i
].spilled_ptr
.parent
=
9783 &newframe
->stack
[i
].spilled_ptr
;
9789 /* Return true if it's OK to have the same insn return a different type. */
9790 static bool reg_type_mismatch_ok(enum bpf_reg_type type
)
9795 case PTR_TO_SOCKET_OR_NULL
:
9796 case PTR_TO_SOCK_COMMON
:
9797 case PTR_TO_SOCK_COMMON_OR_NULL
:
9798 case PTR_TO_TCP_SOCK
:
9799 case PTR_TO_TCP_SOCK_OR_NULL
:
9800 case PTR_TO_XDP_SOCK
:
9802 case PTR_TO_BTF_ID_OR_NULL
:
9809 /* If an instruction was previously used with particular pointer types, then we
9810 * need to be careful to avoid cases such as the below, where it may be ok
9811 * for one branch accessing the pointer, but not ok for the other branch:
9816 * R1 = some_other_valid_ptr;
9819 * R2 = *(u32 *)(R1 + 0);
9821 static bool reg_type_mismatch(enum bpf_reg_type src
, enum bpf_reg_type prev
)
9823 return src
!= prev
&& (!reg_type_mismatch_ok(src
) ||
9824 !reg_type_mismatch_ok(prev
));
9827 static int do_check(struct bpf_verifier_env
*env
)
9829 bool pop_log
= !(env
->log
.level
& BPF_LOG_LEVEL2
);
9830 struct bpf_verifier_state
*state
= env
->cur_state
;
9831 struct bpf_insn
*insns
= env
->prog
->insnsi
;
9832 struct bpf_reg_state
*regs
;
9833 int insn_cnt
= env
->prog
->len
;
9834 bool do_print_state
= false;
9835 int prev_insn_idx
= -1;
9838 struct bpf_insn
*insn
;
9842 env
->prev_insn_idx
= prev_insn_idx
;
9843 if (env
->insn_idx
>= insn_cnt
) {
9844 verbose(env
, "invalid insn idx %d insn_cnt %d\n",
9845 env
->insn_idx
, insn_cnt
);
9849 insn
= &insns
[env
->insn_idx
];
9850 class = BPF_CLASS(insn
->code
);
9852 if (++env
->insn_processed
> BPF_COMPLEXITY_LIMIT_INSNS
) {
9854 "BPF program is too large. Processed %d insn\n",
9855 env
->insn_processed
);
9859 err
= is_state_visited(env
, env
->insn_idx
);
9863 /* found equivalent state, can prune the search */
9864 if (env
->log
.level
& BPF_LOG_LEVEL
) {
9866 verbose(env
, "\nfrom %d to %d%s: safe\n",
9867 env
->prev_insn_idx
, env
->insn_idx
,
9868 env
->cur_state
->speculative
?
9869 " (speculative execution)" : "");
9871 verbose(env
, "%d: safe\n", env
->insn_idx
);
9873 goto process_bpf_exit
;
9876 if (signal_pending(current
))
9882 if (env
->log
.level
& BPF_LOG_LEVEL2
||
9883 (env
->log
.level
& BPF_LOG_LEVEL
&& do_print_state
)) {
9884 if (env
->log
.level
& BPF_LOG_LEVEL2
)
9885 verbose(env
, "%d:", env
->insn_idx
);
9887 verbose(env
, "\nfrom %d to %d%s:",
9888 env
->prev_insn_idx
, env
->insn_idx
,
9889 env
->cur_state
->speculative
?
9890 " (speculative execution)" : "");
9891 print_verifier_state(env
, state
->frame
[state
->curframe
]);
9892 do_print_state
= false;
9895 if (env
->log
.level
& BPF_LOG_LEVEL
) {
9896 const struct bpf_insn_cbs cbs
= {
9897 .cb_print
= verbose
,
9898 .private_data
= env
,
9901 verbose_linfo(env
, env
->insn_idx
, "; ");
9902 verbose(env
, "%d: ", env
->insn_idx
);
9903 print_bpf_insn(&cbs
, insn
, env
->allow_ptr_leaks
);
9906 if (bpf_prog_is_dev_bound(env
->prog
->aux
)) {
9907 err
= bpf_prog_offload_verify_insn(env
, env
->insn_idx
,
9908 env
->prev_insn_idx
);
9913 regs
= cur_regs(env
);
9914 env
->insn_aux_data
[env
->insn_idx
].seen
= env
->pass_cnt
;
9915 prev_insn_idx
= env
->insn_idx
;
9917 if (class == BPF_ALU
|| class == BPF_ALU64
) {
9918 err
= check_alu_op(env
, insn
);
9922 } else if (class == BPF_LDX
) {
9923 enum bpf_reg_type
*prev_src_type
, src_reg_type
;
9925 /* check for reserved fields is already done */
9927 /* check src operand */
9928 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
9932 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP_NO_MARK
);
9936 src_reg_type
= regs
[insn
->src_reg
].type
;
9938 /* check that memory (src_reg + off) is readable,
9939 * the state of dst_reg will be updated by this func
9941 err
= check_mem_access(env
, env
->insn_idx
, insn
->src_reg
,
9942 insn
->off
, BPF_SIZE(insn
->code
),
9943 BPF_READ
, insn
->dst_reg
, false);
9947 prev_src_type
= &env
->insn_aux_data
[env
->insn_idx
].ptr_type
;
9949 if (*prev_src_type
== NOT_INIT
) {
9951 * dst_reg = *(u32 *)(src_reg + off)
9952 * save type to validate intersecting paths
9954 *prev_src_type
= src_reg_type
;
9956 } else if (reg_type_mismatch(src_reg_type
, *prev_src_type
)) {
9957 /* ABuser program is trying to use the same insn
9958 * dst_reg = *(u32*) (src_reg + off)
9959 * with different pointer types:
9960 * src_reg == ctx in one branch and
9961 * src_reg == stack|map in some other branch.
9964 verbose(env
, "same insn cannot be used with different pointers\n");
9968 } else if (class == BPF_STX
) {
9969 enum bpf_reg_type
*prev_dst_type
, dst_reg_type
;
9971 if (BPF_MODE(insn
->code
) == BPF_XADD
) {
9972 err
= check_xadd(env
, env
->insn_idx
, insn
);
9979 /* check src1 operand */
9980 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
9983 /* check src2 operand */
9984 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
9988 dst_reg_type
= regs
[insn
->dst_reg
].type
;
9990 /* check that memory (dst_reg + off) is writeable */
9991 err
= check_mem_access(env
, env
->insn_idx
, insn
->dst_reg
,
9992 insn
->off
, BPF_SIZE(insn
->code
),
9993 BPF_WRITE
, insn
->src_reg
, false);
9997 prev_dst_type
= &env
->insn_aux_data
[env
->insn_idx
].ptr_type
;
9999 if (*prev_dst_type
== NOT_INIT
) {
10000 *prev_dst_type
= dst_reg_type
;
10001 } else if (reg_type_mismatch(dst_reg_type
, *prev_dst_type
)) {
10002 verbose(env
, "same insn cannot be used with different pointers\n");
10006 } else if (class == BPF_ST
) {
10007 if (BPF_MODE(insn
->code
) != BPF_MEM
||
10008 insn
->src_reg
!= BPF_REG_0
) {
10009 verbose(env
, "BPF_ST uses reserved fields\n");
10012 /* check src operand */
10013 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
10017 if (is_ctx_reg(env
, insn
->dst_reg
)) {
10018 verbose(env
, "BPF_ST stores into R%d %s is not allowed\n",
10020 reg_type_str
[reg_state(env
, insn
->dst_reg
)->type
]);
10024 /* check that memory (dst_reg + off) is writeable */
10025 err
= check_mem_access(env
, env
->insn_idx
, insn
->dst_reg
,
10026 insn
->off
, BPF_SIZE(insn
->code
),
10027 BPF_WRITE
, -1, false);
10031 } else if (class == BPF_JMP
|| class == BPF_JMP32
) {
10032 u8 opcode
= BPF_OP(insn
->code
);
10034 env
->jmps_processed
++;
10035 if (opcode
== BPF_CALL
) {
10036 if (BPF_SRC(insn
->code
) != BPF_K
||
10038 (insn
->src_reg
!= BPF_REG_0
&&
10039 insn
->src_reg
!= BPF_PSEUDO_CALL
) ||
10040 insn
->dst_reg
!= BPF_REG_0
||
10041 class == BPF_JMP32
) {
10042 verbose(env
, "BPF_CALL uses reserved fields\n");
10046 if (env
->cur_state
->active_spin_lock
&&
10047 (insn
->src_reg
== BPF_PSEUDO_CALL
||
10048 insn
->imm
!= BPF_FUNC_spin_unlock
)) {
10049 verbose(env
, "function calls are not allowed while holding a lock\n");
10052 if (insn
->src_reg
== BPF_PSEUDO_CALL
)
10053 err
= check_func_call(env
, insn
, &env
->insn_idx
);
10055 err
= check_helper_call(env
, insn
->imm
, env
->insn_idx
);
10059 } else if (opcode
== BPF_JA
) {
10060 if (BPF_SRC(insn
->code
) != BPF_K
||
10062 insn
->src_reg
!= BPF_REG_0
||
10063 insn
->dst_reg
!= BPF_REG_0
||
10064 class == BPF_JMP32
) {
10065 verbose(env
, "BPF_JA uses reserved fields\n");
10069 env
->insn_idx
+= insn
->off
+ 1;
10072 } else if (opcode
== BPF_EXIT
) {
10073 if (BPF_SRC(insn
->code
) != BPF_K
||
10075 insn
->src_reg
!= BPF_REG_0
||
10076 insn
->dst_reg
!= BPF_REG_0
||
10077 class == BPF_JMP32
) {
10078 verbose(env
, "BPF_EXIT uses reserved fields\n");
10082 if (env
->cur_state
->active_spin_lock
) {
10083 verbose(env
, "bpf_spin_unlock is missing\n");
10087 if (state
->curframe
) {
10088 /* exit from nested function */
10089 err
= prepare_func_exit(env
, &env
->insn_idx
);
10092 do_print_state
= true;
10096 err
= check_reference_leak(env
);
10100 err
= check_return_code(env
);
10104 update_branch_counts(env
, env
->cur_state
);
10105 err
= pop_stack(env
, &prev_insn_idx
,
10106 &env
->insn_idx
, pop_log
);
10108 if (err
!= -ENOENT
)
10112 do_print_state
= true;
10116 err
= check_cond_jmp_op(env
, insn
, &env
->insn_idx
);
10120 } else if (class == BPF_LD
) {
10121 u8 mode
= BPF_MODE(insn
->code
);
10123 if (mode
== BPF_ABS
|| mode
== BPF_IND
) {
10124 err
= check_ld_abs(env
, insn
);
10128 } else if (mode
== BPF_IMM
) {
10129 err
= check_ld_imm(env
, insn
);
10134 env
->insn_aux_data
[env
->insn_idx
].seen
= env
->pass_cnt
;
10136 verbose(env
, "invalid BPF_LD mode\n");
10140 verbose(env
, "unknown insn class %d\n", class);
10150 /* replace pseudo btf_id with kernel symbol address */
10151 static int check_pseudo_btf_id(struct bpf_verifier_env
*env
,
10152 struct bpf_insn
*insn
,
10153 struct bpf_insn_aux_data
*aux
)
10155 const struct btf_var_secinfo
*vsi
;
10156 const struct btf_type
*datasec
;
10157 const struct btf_type
*t
;
10158 const char *sym_name
;
10159 bool percpu
= false;
10160 u32 type
, id
= insn
->imm
;
10165 if (!btf_vmlinux
) {
10166 verbose(env
, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
10170 if (insn
[1].imm
!= 0) {
10171 verbose(env
, "reserved field (insn[1].imm) is used in pseudo_btf_id ldimm64 insn.\n");
10175 t
= btf_type_by_id(btf_vmlinux
, id
);
10177 verbose(env
, "ldimm64 insn specifies invalid btf_id %d.\n", id
);
10181 if (!btf_type_is_var(t
)) {
10182 verbose(env
, "pseudo btf_id %d in ldimm64 isn't KIND_VAR.\n",
10187 sym_name
= btf_name_by_offset(btf_vmlinux
, t
->name_off
);
10188 addr
= kallsyms_lookup_name(sym_name
);
10190 verbose(env
, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
10195 datasec_id
= btf_find_by_name_kind(btf_vmlinux
, ".data..percpu",
10197 if (datasec_id
> 0) {
10198 datasec
= btf_type_by_id(btf_vmlinux
, datasec_id
);
10199 for_each_vsi(i
, datasec
, vsi
) {
10200 if (vsi
->type
== id
) {
10207 insn
[0].imm
= (u32
)addr
;
10208 insn
[1].imm
= addr
>> 32;
10211 t
= btf_type_skip_modifiers(btf_vmlinux
, type
, NULL
);
10213 aux
->btf_var
.reg_type
= PTR_TO_PERCPU_BTF_ID
;
10214 aux
->btf_var
.btf
= btf_vmlinux
;
10215 aux
->btf_var
.btf_id
= type
;
10216 } else if (!btf_type_is_struct(t
)) {
10217 const struct btf_type
*ret
;
10221 /* resolve the type size of ksym. */
10222 ret
= btf_resolve_size(btf_vmlinux
, t
, &tsize
);
10224 tname
= btf_name_by_offset(btf_vmlinux
, t
->name_off
);
10225 verbose(env
, "ldimm64 unable to resolve the size of type '%s': %ld\n",
10226 tname
, PTR_ERR(ret
));
10229 aux
->btf_var
.reg_type
= PTR_TO_MEM
;
10230 aux
->btf_var
.mem_size
= tsize
;
10232 aux
->btf_var
.reg_type
= PTR_TO_BTF_ID
;
10233 aux
->btf_var
.btf
= btf_vmlinux
;
10234 aux
->btf_var
.btf_id
= type
;
10239 static int check_map_prealloc(struct bpf_map
*map
)
10241 return (map
->map_type
!= BPF_MAP_TYPE_HASH
&&
10242 map
->map_type
!= BPF_MAP_TYPE_PERCPU_HASH
&&
10243 map
->map_type
!= BPF_MAP_TYPE_HASH_OF_MAPS
) ||
10244 !(map
->map_flags
& BPF_F_NO_PREALLOC
);
10247 static bool is_tracing_prog_type(enum bpf_prog_type type
)
10250 case BPF_PROG_TYPE_KPROBE
:
10251 case BPF_PROG_TYPE_TRACEPOINT
:
10252 case BPF_PROG_TYPE_PERF_EVENT
:
10253 case BPF_PROG_TYPE_RAW_TRACEPOINT
:
10260 static bool is_preallocated_map(struct bpf_map
*map
)
10262 if (!check_map_prealloc(map
))
10264 if (map
->inner_map_meta
&& !check_map_prealloc(map
->inner_map_meta
))
10269 static int check_map_prog_compatibility(struct bpf_verifier_env
*env
,
10270 struct bpf_map
*map
,
10271 struct bpf_prog
*prog
)
10274 enum bpf_prog_type prog_type
= resolve_prog_type(prog
);
10276 * Validate that trace type programs use preallocated hash maps.
10278 * For programs attached to PERF events this is mandatory as the
10279 * perf NMI can hit any arbitrary code sequence.
10281 * All other trace types using preallocated hash maps are unsafe as
10282 * well because tracepoint or kprobes can be inside locked regions
10283 * of the memory allocator or at a place where a recursion into the
10284 * memory allocator would see inconsistent state.
10286 * On RT enabled kernels run-time allocation of all trace type
10287 * programs is strictly prohibited due to lock type constraints. On
10288 * !RT kernels it is allowed for backwards compatibility reasons for
10289 * now, but warnings are emitted so developers are made aware of
10290 * the unsafety and can fix their programs before this is enforced.
10292 if (is_tracing_prog_type(prog_type
) && !is_preallocated_map(map
)) {
10293 if (prog_type
== BPF_PROG_TYPE_PERF_EVENT
) {
10294 verbose(env
, "perf_event programs can only use preallocated hash map\n");
10297 if (IS_ENABLED(CONFIG_PREEMPT_RT
)) {
10298 verbose(env
, "trace type programs can only use preallocated hash map\n");
10301 WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
10302 verbose(env
, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
10305 if (map_value_has_spin_lock(map
)) {
10306 if (prog_type
== BPF_PROG_TYPE_SOCKET_FILTER
) {
10307 verbose(env
, "socket filter progs cannot use bpf_spin_lock yet\n");
10311 if (is_tracing_prog_type(prog_type
)) {
10312 verbose(env
, "tracing progs cannot use bpf_spin_lock yet\n");
10316 if (prog
->aux
->sleepable
) {
10317 verbose(env
, "sleepable progs cannot use bpf_spin_lock yet\n");
10322 if ((bpf_prog_is_dev_bound(prog
->aux
) || bpf_map_is_dev_bound(map
)) &&
10323 !bpf_offload_prog_map_match(prog
, map
)) {
10324 verbose(env
, "offload device mismatch between prog and map\n");
10328 if (map
->map_type
== BPF_MAP_TYPE_STRUCT_OPS
) {
10329 verbose(env
, "bpf_struct_ops map cannot be used in prog\n");
10333 if (prog
->aux
->sleepable
)
10334 switch (map
->map_type
) {
10335 case BPF_MAP_TYPE_HASH
:
10336 case BPF_MAP_TYPE_LRU_HASH
:
10337 case BPF_MAP_TYPE_ARRAY
:
10338 if (!is_preallocated_map(map
)) {
10340 "Sleepable programs can only use preallocated hash maps\n");
10346 "Sleepable programs can only use array and hash maps\n");
10353 static bool bpf_map_is_cgroup_storage(struct bpf_map
*map
)
10355 return (map
->map_type
== BPF_MAP_TYPE_CGROUP_STORAGE
||
10356 map
->map_type
== BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE
);
10359 /* find and rewrite pseudo imm in ld_imm64 instructions:
10361 * 1. if it accesses map FD, replace it with actual map pointer.
10362 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
10364 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
10366 static int resolve_pseudo_ldimm64(struct bpf_verifier_env
*env
)
10368 struct bpf_insn
*insn
= env
->prog
->insnsi
;
10369 int insn_cnt
= env
->prog
->len
;
10372 err
= bpf_prog_calc_tag(env
->prog
);
10376 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
10377 if (BPF_CLASS(insn
->code
) == BPF_LDX
&&
10378 (BPF_MODE(insn
->code
) != BPF_MEM
|| insn
->imm
!= 0)) {
10379 verbose(env
, "BPF_LDX uses reserved fields\n");
10383 if (BPF_CLASS(insn
->code
) == BPF_STX
&&
10384 ((BPF_MODE(insn
->code
) != BPF_MEM
&&
10385 BPF_MODE(insn
->code
) != BPF_XADD
) || insn
->imm
!= 0)) {
10386 verbose(env
, "BPF_STX uses reserved fields\n");
10390 if (insn
[0].code
== (BPF_LD
| BPF_IMM
| BPF_DW
)) {
10391 struct bpf_insn_aux_data
*aux
;
10392 struct bpf_map
*map
;
10396 if (i
== insn_cnt
- 1 || insn
[1].code
!= 0 ||
10397 insn
[1].dst_reg
!= 0 || insn
[1].src_reg
!= 0 ||
10398 insn
[1].off
!= 0) {
10399 verbose(env
, "invalid bpf_ld_imm64 insn\n");
10403 if (insn
[0].src_reg
== 0)
10404 /* valid generic load 64-bit imm */
10407 if (insn
[0].src_reg
== BPF_PSEUDO_BTF_ID
) {
10408 aux
= &env
->insn_aux_data
[i
];
10409 err
= check_pseudo_btf_id(env
, insn
, aux
);
10415 /* In final convert_pseudo_ld_imm64() step, this is
10416 * converted into regular 64-bit imm load insn.
10418 if ((insn
[0].src_reg
!= BPF_PSEUDO_MAP_FD
&&
10419 insn
[0].src_reg
!= BPF_PSEUDO_MAP_VALUE
) ||
10420 (insn
[0].src_reg
== BPF_PSEUDO_MAP_FD
&&
10421 insn
[1].imm
!= 0)) {
10423 "unrecognized bpf_ld_imm64 insn\n");
10427 f
= fdget(insn
[0].imm
);
10428 map
= __bpf_map_get(f
);
10430 verbose(env
, "fd %d is not pointing to valid bpf_map\n",
10432 return PTR_ERR(map
);
10435 err
= check_map_prog_compatibility(env
, map
, env
->prog
);
10441 aux
= &env
->insn_aux_data
[i
];
10442 if (insn
->src_reg
== BPF_PSEUDO_MAP_FD
) {
10443 addr
= (unsigned long)map
;
10445 u32 off
= insn
[1].imm
;
10447 if (off
>= BPF_MAX_VAR_OFF
) {
10448 verbose(env
, "direct value offset of %u is not allowed\n", off
);
10453 if (!map
->ops
->map_direct_value_addr
) {
10454 verbose(env
, "no direct value access support for this map type\n");
10459 err
= map
->ops
->map_direct_value_addr(map
, &addr
, off
);
10461 verbose(env
, "invalid access to map value pointer, value_size=%u off=%u\n",
10462 map
->value_size
, off
);
10467 aux
->map_off
= off
;
10471 insn
[0].imm
= (u32
)addr
;
10472 insn
[1].imm
= addr
>> 32;
10474 /* check whether we recorded this map already */
10475 for (j
= 0; j
< env
->used_map_cnt
; j
++) {
10476 if (env
->used_maps
[j
] == map
) {
10477 aux
->map_index
= j
;
10483 if (env
->used_map_cnt
>= MAX_USED_MAPS
) {
10488 /* hold the map. If the program is rejected by verifier,
10489 * the map will be released by release_maps() or it
10490 * will be used by the valid program until it's unloaded
10491 * and all maps are released in free_used_maps()
10495 aux
->map_index
= env
->used_map_cnt
;
10496 env
->used_maps
[env
->used_map_cnt
++] = map
;
10498 if (bpf_map_is_cgroup_storage(map
) &&
10499 bpf_cgroup_storage_assign(env
->prog
->aux
, map
)) {
10500 verbose(env
, "only one cgroup storage of each type is allowed\n");
10512 /* Basic sanity check before we invest more work here. */
10513 if (!bpf_opcode_in_insntable(insn
->code
)) {
10514 verbose(env
, "unknown opcode %02x\n", insn
->code
);
10519 /* now all pseudo BPF_LD_IMM64 instructions load valid
10520 * 'struct bpf_map *' into a register instead of user map_fd.
10521 * These pointers will be used later by verifier to validate map access.
10526 /* drop refcnt of maps used by the rejected program */
10527 static void release_maps(struct bpf_verifier_env
*env
)
10529 __bpf_free_used_maps(env
->prog
->aux
, env
->used_maps
,
10530 env
->used_map_cnt
);
10533 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
10534 static void convert_pseudo_ld_imm64(struct bpf_verifier_env
*env
)
10536 struct bpf_insn
*insn
= env
->prog
->insnsi
;
10537 int insn_cnt
= env
->prog
->len
;
10540 for (i
= 0; i
< insn_cnt
; i
++, insn
++)
10541 if (insn
->code
== (BPF_LD
| BPF_IMM
| BPF_DW
))
10545 /* single env->prog->insni[off] instruction was replaced with the range
10546 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
10547 * [0, off) and [off, end) to new locations, so the patched range stays zero
10549 static int adjust_insn_aux_data(struct bpf_verifier_env
*env
,
10550 struct bpf_prog
*new_prog
, u32 off
, u32 cnt
)
10552 struct bpf_insn_aux_data
*new_data
, *old_data
= env
->insn_aux_data
;
10553 struct bpf_insn
*insn
= new_prog
->insnsi
;
10557 /* aux info at OFF always needs adjustment, no matter fast path
10558 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
10559 * original insn at old prog.
10561 old_data
[off
].zext_dst
= insn_has_def32(env
, insn
+ off
+ cnt
- 1);
10565 prog_len
= new_prog
->len
;
10566 new_data
= vzalloc(array_size(prog_len
,
10567 sizeof(struct bpf_insn_aux_data
)));
10570 memcpy(new_data
, old_data
, sizeof(struct bpf_insn_aux_data
) * off
);
10571 memcpy(new_data
+ off
+ cnt
- 1, old_data
+ off
,
10572 sizeof(struct bpf_insn_aux_data
) * (prog_len
- off
- cnt
+ 1));
10573 for (i
= off
; i
< off
+ cnt
- 1; i
++) {
10574 new_data
[i
].seen
= env
->pass_cnt
;
10575 new_data
[i
].zext_dst
= insn_has_def32(env
, insn
+ i
);
10577 env
->insn_aux_data
= new_data
;
10582 static void adjust_subprog_starts(struct bpf_verifier_env
*env
, u32 off
, u32 len
)
10588 /* NOTE: fake 'exit' subprog should be updated as well. */
10589 for (i
= 0; i
<= env
->subprog_cnt
; i
++) {
10590 if (env
->subprog_info
[i
].start
<= off
)
10592 env
->subprog_info
[i
].start
+= len
- 1;
10596 static void adjust_poke_descs(struct bpf_prog
*prog
, u32 len
)
10598 struct bpf_jit_poke_descriptor
*tab
= prog
->aux
->poke_tab
;
10599 int i
, sz
= prog
->aux
->size_poke_tab
;
10600 struct bpf_jit_poke_descriptor
*desc
;
10602 for (i
= 0; i
< sz
; i
++) {
10604 desc
->insn_idx
+= len
- 1;
10608 static struct bpf_prog
*bpf_patch_insn_data(struct bpf_verifier_env
*env
, u32 off
,
10609 const struct bpf_insn
*patch
, u32 len
)
10611 struct bpf_prog
*new_prog
;
10613 new_prog
= bpf_patch_insn_single(env
->prog
, off
, patch
, len
);
10614 if (IS_ERR(new_prog
)) {
10615 if (PTR_ERR(new_prog
) == -ERANGE
)
10617 "insn %d cannot be patched due to 16-bit range\n",
10618 env
->insn_aux_data
[off
].orig_idx
);
10621 if (adjust_insn_aux_data(env
, new_prog
, off
, len
))
10623 adjust_subprog_starts(env
, off
, len
);
10624 adjust_poke_descs(new_prog
, len
);
10628 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env
*env
,
10633 /* find first prog starting at or after off (first to remove) */
10634 for (i
= 0; i
< env
->subprog_cnt
; i
++)
10635 if (env
->subprog_info
[i
].start
>= off
)
10637 /* find first prog starting at or after off + cnt (first to stay) */
10638 for (j
= i
; j
< env
->subprog_cnt
; j
++)
10639 if (env
->subprog_info
[j
].start
>= off
+ cnt
)
10641 /* if j doesn't start exactly at off + cnt, we are just removing
10642 * the front of previous prog
10644 if (env
->subprog_info
[j
].start
!= off
+ cnt
)
10648 struct bpf_prog_aux
*aux
= env
->prog
->aux
;
10651 /* move fake 'exit' subprog as well */
10652 move
= env
->subprog_cnt
+ 1 - j
;
10654 memmove(env
->subprog_info
+ i
,
10655 env
->subprog_info
+ j
,
10656 sizeof(*env
->subprog_info
) * move
);
10657 env
->subprog_cnt
-= j
- i
;
10659 /* remove func_info */
10660 if (aux
->func_info
) {
10661 move
= aux
->func_info_cnt
- j
;
10663 memmove(aux
->func_info
+ i
,
10664 aux
->func_info
+ j
,
10665 sizeof(*aux
->func_info
) * move
);
10666 aux
->func_info_cnt
-= j
- i
;
10667 /* func_info->insn_off is set after all code rewrites,
10668 * in adjust_btf_func() - no need to adjust
10672 /* convert i from "first prog to remove" to "first to adjust" */
10673 if (env
->subprog_info
[i
].start
== off
)
10677 /* update fake 'exit' subprog as well */
10678 for (; i
<= env
->subprog_cnt
; i
++)
10679 env
->subprog_info
[i
].start
-= cnt
;
10684 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env
*env
, u32 off
,
10687 struct bpf_prog
*prog
= env
->prog
;
10688 u32 i
, l_off
, l_cnt
, nr_linfo
;
10689 struct bpf_line_info
*linfo
;
10691 nr_linfo
= prog
->aux
->nr_linfo
;
10695 linfo
= prog
->aux
->linfo
;
10697 /* find first line info to remove, count lines to be removed */
10698 for (i
= 0; i
< nr_linfo
; i
++)
10699 if (linfo
[i
].insn_off
>= off
)
10704 for (; i
< nr_linfo
; i
++)
10705 if (linfo
[i
].insn_off
< off
+ cnt
)
10710 /* First live insn doesn't match first live linfo, it needs to "inherit"
10711 * last removed linfo. prog is already modified, so prog->len == off
10712 * means no live instructions after (tail of the program was removed).
10714 if (prog
->len
!= off
&& l_cnt
&&
10715 (i
== nr_linfo
|| linfo
[i
].insn_off
!= off
+ cnt
)) {
10717 linfo
[--i
].insn_off
= off
+ cnt
;
10720 /* remove the line info which refer to the removed instructions */
10722 memmove(linfo
+ l_off
, linfo
+ i
,
10723 sizeof(*linfo
) * (nr_linfo
- i
));
10725 prog
->aux
->nr_linfo
-= l_cnt
;
10726 nr_linfo
= prog
->aux
->nr_linfo
;
10729 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
10730 for (i
= l_off
; i
< nr_linfo
; i
++)
10731 linfo
[i
].insn_off
-= cnt
;
10733 /* fix up all subprogs (incl. 'exit') which start >= off */
10734 for (i
= 0; i
<= env
->subprog_cnt
; i
++)
10735 if (env
->subprog_info
[i
].linfo_idx
> l_off
) {
10736 /* program may have started in the removed region but
10737 * may not be fully removed
10739 if (env
->subprog_info
[i
].linfo_idx
>= l_off
+ l_cnt
)
10740 env
->subprog_info
[i
].linfo_idx
-= l_cnt
;
10742 env
->subprog_info
[i
].linfo_idx
= l_off
;
10748 static int verifier_remove_insns(struct bpf_verifier_env
*env
, u32 off
, u32 cnt
)
10750 struct bpf_insn_aux_data
*aux_data
= env
->insn_aux_data
;
10751 unsigned int orig_prog_len
= env
->prog
->len
;
10754 if (bpf_prog_is_dev_bound(env
->prog
->aux
))
10755 bpf_prog_offload_remove_insns(env
, off
, cnt
);
10757 err
= bpf_remove_insns(env
->prog
, off
, cnt
);
10761 err
= adjust_subprog_starts_after_remove(env
, off
, cnt
);
10765 err
= bpf_adj_linfo_after_remove(env
, off
, cnt
);
10769 memmove(aux_data
+ off
, aux_data
+ off
+ cnt
,
10770 sizeof(*aux_data
) * (orig_prog_len
- off
- cnt
));
10775 /* The verifier does more data flow analysis than llvm and will not
10776 * explore branches that are dead at run time. Malicious programs can
10777 * have dead code too. Therefore replace all dead at-run-time code
10780 * Just nops are not optimal, e.g. if they would sit at the end of the
10781 * program and through another bug we would manage to jump there, then
10782 * we'd execute beyond program memory otherwise. Returning exception
10783 * code also wouldn't work since we can have subprogs where the dead
10784 * code could be located.
10786 static void sanitize_dead_code(struct bpf_verifier_env
*env
)
10788 struct bpf_insn_aux_data
*aux_data
= env
->insn_aux_data
;
10789 struct bpf_insn trap
= BPF_JMP_IMM(BPF_JA
, 0, 0, -1);
10790 struct bpf_insn
*insn
= env
->prog
->insnsi
;
10791 const int insn_cnt
= env
->prog
->len
;
10794 for (i
= 0; i
< insn_cnt
; i
++) {
10795 if (aux_data
[i
].seen
)
10797 memcpy(insn
+ i
, &trap
, sizeof(trap
));
10801 static bool insn_is_cond_jump(u8 code
)
10805 if (BPF_CLASS(code
) == BPF_JMP32
)
10808 if (BPF_CLASS(code
) != BPF_JMP
)
10812 return op
!= BPF_JA
&& op
!= BPF_EXIT
&& op
!= BPF_CALL
;
10815 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env
*env
)
10817 struct bpf_insn_aux_data
*aux_data
= env
->insn_aux_data
;
10818 struct bpf_insn ja
= BPF_JMP_IMM(BPF_JA
, 0, 0, 0);
10819 struct bpf_insn
*insn
= env
->prog
->insnsi
;
10820 const int insn_cnt
= env
->prog
->len
;
10823 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
10824 if (!insn_is_cond_jump(insn
->code
))
10827 if (!aux_data
[i
+ 1].seen
)
10828 ja
.off
= insn
->off
;
10829 else if (!aux_data
[i
+ 1 + insn
->off
].seen
)
10834 if (bpf_prog_is_dev_bound(env
->prog
->aux
))
10835 bpf_prog_offload_replace_insn(env
, i
, &ja
);
10837 memcpy(insn
, &ja
, sizeof(ja
));
10841 static int opt_remove_dead_code(struct bpf_verifier_env
*env
)
10843 struct bpf_insn_aux_data
*aux_data
= env
->insn_aux_data
;
10844 int insn_cnt
= env
->prog
->len
;
10847 for (i
= 0; i
< insn_cnt
; i
++) {
10851 while (i
+ j
< insn_cnt
&& !aux_data
[i
+ j
].seen
)
10856 err
= verifier_remove_insns(env
, i
, j
);
10859 insn_cnt
= env
->prog
->len
;
10865 static int opt_remove_nops(struct bpf_verifier_env
*env
)
10867 const struct bpf_insn ja
= BPF_JMP_IMM(BPF_JA
, 0, 0, 0);
10868 struct bpf_insn
*insn
= env
->prog
->insnsi
;
10869 int insn_cnt
= env
->prog
->len
;
10872 for (i
= 0; i
< insn_cnt
; i
++) {
10873 if (memcmp(&insn
[i
], &ja
, sizeof(ja
)))
10876 err
= verifier_remove_insns(env
, i
, 1);
10886 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env
*env
,
10887 const union bpf_attr
*attr
)
10889 struct bpf_insn
*patch
, zext_patch
[2], rnd_hi32_patch
[4];
10890 struct bpf_insn_aux_data
*aux
= env
->insn_aux_data
;
10891 int i
, patch_len
, delta
= 0, len
= env
->prog
->len
;
10892 struct bpf_insn
*insns
= env
->prog
->insnsi
;
10893 struct bpf_prog
*new_prog
;
10896 rnd_hi32
= attr
->prog_flags
& BPF_F_TEST_RND_HI32
;
10897 zext_patch
[1] = BPF_ZEXT_REG(0);
10898 rnd_hi32_patch
[1] = BPF_ALU64_IMM(BPF_MOV
, BPF_REG_AX
, 0);
10899 rnd_hi32_patch
[2] = BPF_ALU64_IMM(BPF_LSH
, BPF_REG_AX
, 32);
10900 rnd_hi32_patch
[3] = BPF_ALU64_REG(BPF_OR
, 0, BPF_REG_AX
);
10901 for (i
= 0; i
< len
; i
++) {
10902 int adj_idx
= i
+ delta
;
10903 struct bpf_insn insn
;
10905 insn
= insns
[adj_idx
];
10906 if (!aux
[adj_idx
].zext_dst
) {
10914 class = BPF_CLASS(code
);
10915 if (insn_no_def(&insn
))
10918 /* NOTE: arg "reg" (the fourth one) is only used for
10919 * BPF_STX which has been ruled out in above
10920 * check, it is safe to pass NULL here.
10922 if (is_reg64(env
, &insn
, insn
.dst_reg
, NULL
, DST_OP
)) {
10923 if (class == BPF_LD
&&
10924 BPF_MODE(code
) == BPF_IMM
)
10929 /* ctx load could be transformed into wider load. */
10930 if (class == BPF_LDX
&&
10931 aux
[adj_idx
].ptr_type
== PTR_TO_CTX
)
10934 imm_rnd
= get_random_int();
10935 rnd_hi32_patch
[0] = insn
;
10936 rnd_hi32_patch
[1].imm
= imm_rnd
;
10937 rnd_hi32_patch
[3].dst_reg
= insn
.dst_reg
;
10938 patch
= rnd_hi32_patch
;
10940 goto apply_patch_buffer
;
10943 if (!bpf_jit_needs_zext())
10946 zext_patch
[0] = insn
;
10947 zext_patch
[1].dst_reg
= insn
.dst_reg
;
10948 zext_patch
[1].src_reg
= insn
.dst_reg
;
10949 patch
= zext_patch
;
10951 apply_patch_buffer
:
10952 new_prog
= bpf_patch_insn_data(env
, adj_idx
, patch
, patch_len
);
10955 env
->prog
= new_prog
;
10956 insns
= new_prog
->insnsi
;
10957 aux
= env
->insn_aux_data
;
10958 delta
+= patch_len
- 1;
10964 /* convert load instructions that access fields of a context type into a
10965 * sequence of instructions that access fields of the underlying structure:
10966 * struct __sk_buff -> struct sk_buff
10967 * struct bpf_sock_ops -> struct sock
10969 static int convert_ctx_accesses(struct bpf_verifier_env
*env
)
10971 const struct bpf_verifier_ops
*ops
= env
->ops
;
10972 int i
, cnt
, size
, ctx_field_size
, delta
= 0;
10973 const int insn_cnt
= env
->prog
->len
;
10974 struct bpf_insn insn_buf
[16], *insn
;
10975 u32 target_size
, size_default
, off
;
10976 struct bpf_prog
*new_prog
;
10977 enum bpf_access_type type
;
10978 bool is_narrower_load
;
10980 if (ops
->gen_prologue
|| env
->seen_direct_write
) {
10981 if (!ops
->gen_prologue
) {
10982 verbose(env
, "bpf verifier is misconfigured\n");
10985 cnt
= ops
->gen_prologue(insn_buf
, env
->seen_direct_write
,
10987 if (cnt
>= ARRAY_SIZE(insn_buf
)) {
10988 verbose(env
, "bpf verifier is misconfigured\n");
10991 new_prog
= bpf_patch_insn_data(env
, 0, insn_buf
, cnt
);
10995 env
->prog
= new_prog
;
11000 if (bpf_prog_is_dev_bound(env
->prog
->aux
))
11003 insn
= env
->prog
->insnsi
+ delta
;
11005 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
11006 bpf_convert_ctx_access_t convert_ctx_access
;
11008 if (insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_B
) ||
11009 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_H
) ||
11010 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_W
) ||
11011 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_DW
))
11013 else if (insn
->code
== (BPF_STX
| BPF_MEM
| BPF_B
) ||
11014 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_H
) ||
11015 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_W
) ||
11016 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_DW
))
11021 if (type
== BPF_WRITE
&&
11022 env
->insn_aux_data
[i
+ delta
].sanitize_stack_off
) {
11023 struct bpf_insn patch
[] = {
11024 /* Sanitize suspicious stack slot with zero.
11025 * There are no memory dependencies for this store,
11026 * since it's only using frame pointer and immediate
11029 BPF_ST_MEM(BPF_DW
, BPF_REG_FP
,
11030 env
->insn_aux_data
[i
+ delta
].sanitize_stack_off
,
11032 /* the original STX instruction will immediately
11033 * overwrite the same stack slot with appropriate value
11038 cnt
= ARRAY_SIZE(patch
);
11039 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, patch
, cnt
);
11044 env
->prog
= new_prog
;
11045 insn
= new_prog
->insnsi
+ i
+ delta
;
11049 switch (env
->insn_aux_data
[i
+ delta
].ptr_type
) {
11051 if (!ops
->convert_ctx_access
)
11053 convert_ctx_access
= ops
->convert_ctx_access
;
11055 case PTR_TO_SOCKET
:
11056 case PTR_TO_SOCK_COMMON
:
11057 convert_ctx_access
= bpf_sock_convert_ctx_access
;
11059 case PTR_TO_TCP_SOCK
:
11060 convert_ctx_access
= bpf_tcp_sock_convert_ctx_access
;
11062 case PTR_TO_XDP_SOCK
:
11063 convert_ctx_access
= bpf_xdp_sock_convert_ctx_access
;
11065 case PTR_TO_BTF_ID
:
11066 if (type
== BPF_READ
) {
11067 insn
->code
= BPF_LDX
| BPF_PROBE_MEM
|
11068 BPF_SIZE((insn
)->code
);
11069 env
->prog
->aux
->num_exentries
++;
11070 } else if (resolve_prog_type(env
->prog
) != BPF_PROG_TYPE_STRUCT_OPS
) {
11071 verbose(env
, "Writes through BTF pointers are not allowed\n");
11079 ctx_field_size
= env
->insn_aux_data
[i
+ delta
].ctx_field_size
;
11080 size
= BPF_LDST_BYTES(insn
);
11082 /* If the read access is a narrower load of the field,
11083 * convert to a 4/8-byte load, to minimum program type specific
11084 * convert_ctx_access changes. If conversion is successful,
11085 * we will apply proper mask to the result.
11087 is_narrower_load
= size
< ctx_field_size
;
11088 size_default
= bpf_ctx_off_adjust_machine(ctx_field_size
);
11090 if (is_narrower_load
) {
11093 if (type
== BPF_WRITE
) {
11094 verbose(env
, "bpf verifier narrow ctx access misconfigured\n");
11099 if (ctx_field_size
== 4)
11101 else if (ctx_field_size
== 8)
11102 size_code
= BPF_DW
;
11104 insn
->off
= off
& ~(size_default
- 1);
11105 insn
->code
= BPF_LDX
| BPF_MEM
| size_code
;
11109 cnt
= convert_ctx_access(type
, insn
, insn_buf
, env
->prog
,
11111 if (cnt
== 0 || cnt
>= ARRAY_SIZE(insn_buf
) ||
11112 (ctx_field_size
&& !target_size
)) {
11113 verbose(env
, "bpf verifier is misconfigured\n");
11117 if (is_narrower_load
&& size
< target_size
) {
11118 u8 shift
= bpf_ctx_narrow_access_offset(
11119 off
, size
, size_default
) * 8;
11120 if (ctx_field_size
<= 4) {
11122 insn_buf
[cnt
++] = BPF_ALU32_IMM(BPF_RSH
,
11125 insn_buf
[cnt
++] = BPF_ALU32_IMM(BPF_AND
, insn
->dst_reg
,
11126 (1 << size
* 8) - 1);
11129 insn_buf
[cnt
++] = BPF_ALU64_IMM(BPF_RSH
,
11132 insn_buf
[cnt
++] = BPF_ALU64_IMM(BPF_AND
, insn
->dst_reg
,
11133 (1ULL << size
* 8) - 1);
11137 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
, cnt
);
11143 /* keep walking new program and skip insns we just inserted */
11144 env
->prog
= new_prog
;
11145 insn
= new_prog
->insnsi
+ i
+ delta
;
11151 static int jit_subprogs(struct bpf_verifier_env
*env
)
11153 struct bpf_prog
*prog
= env
->prog
, **func
, *tmp
;
11154 int i
, j
, subprog_start
, subprog_end
= 0, len
, subprog
;
11155 struct bpf_map
*map_ptr
;
11156 struct bpf_insn
*insn
;
11157 void *old_bpf_func
;
11158 int err
, num_exentries
;
11160 if (env
->subprog_cnt
<= 1)
11163 for (i
= 0, insn
= prog
->insnsi
; i
< prog
->len
; i
++, insn
++) {
11164 if (insn
->code
!= (BPF_JMP
| BPF_CALL
) ||
11165 insn
->src_reg
!= BPF_PSEUDO_CALL
)
11167 /* Upon error here we cannot fall back to interpreter but
11168 * need a hard reject of the program. Thus -EFAULT is
11169 * propagated in any case.
11171 subprog
= find_subprog(env
, i
+ insn
->imm
+ 1);
11173 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
11174 i
+ insn
->imm
+ 1);
11177 /* temporarily remember subprog id inside insn instead of
11178 * aux_data, since next loop will split up all insns into funcs
11180 insn
->off
= subprog
;
11181 /* remember original imm in case JIT fails and fallback
11182 * to interpreter will be needed
11184 env
->insn_aux_data
[i
].call_imm
= insn
->imm
;
11185 /* point imm to __bpf_call_base+1 from JITs point of view */
11189 err
= bpf_prog_alloc_jited_linfo(prog
);
11191 goto out_undo_insn
;
11194 func
= kcalloc(env
->subprog_cnt
, sizeof(prog
), GFP_KERNEL
);
11196 goto out_undo_insn
;
11198 for (i
= 0; i
< env
->subprog_cnt
; i
++) {
11199 subprog_start
= subprog_end
;
11200 subprog_end
= env
->subprog_info
[i
+ 1].start
;
11202 len
= subprog_end
- subprog_start
;
11203 /* BPF_PROG_RUN doesn't call subprogs directly,
11204 * hence main prog stats include the runtime of subprogs.
11205 * subprogs don't have IDs and not reachable via prog_get_next_id
11206 * func[i]->aux->stats will never be accessed and stays NULL
11208 func
[i
] = bpf_prog_alloc_no_stats(bpf_prog_size(len
), GFP_USER
);
11211 memcpy(func
[i
]->insnsi
, &prog
->insnsi
[subprog_start
],
11212 len
* sizeof(struct bpf_insn
));
11213 func
[i
]->type
= prog
->type
;
11214 func
[i
]->len
= len
;
11215 if (bpf_prog_calc_tag(func
[i
]))
11217 func
[i
]->is_func
= 1;
11218 func
[i
]->aux
->func_idx
= i
;
11219 /* the btf and func_info will be freed only at prog->aux */
11220 func
[i
]->aux
->btf
= prog
->aux
->btf
;
11221 func
[i
]->aux
->func_info
= prog
->aux
->func_info
;
11223 for (j
= 0; j
< prog
->aux
->size_poke_tab
; j
++) {
11224 u32 insn_idx
= prog
->aux
->poke_tab
[j
].insn_idx
;
11227 if (!(insn_idx
>= subprog_start
&&
11228 insn_idx
<= subprog_end
))
11231 ret
= bpf_jit_add_poke_descriptor(func
[i
],
11232 &prog
->aux
->poke_tab
[j
]);
11234 verbose(env
, "adding tail call poke descriptor failed\n");
11238 func
[i
]->insnsi
[insn_idx
- subprog_start
].imm
= ret
+ 1;
11240 map_ptr
= func
[i
]->aux
->poke_tab
[ret
].tail_call
.map
;
11241 ret
= map_ptr
->ops
->map_poke_track(map_ptr
, func
[i
]->aux
);
11243 verbose(env
, "tracking tail call prog failed\n");
11248 /* Use bpf_prog_F_tag to indicate functions in stack traces.
11249 * Long term would need debug info to populate names
11251 func
[i
]->aux
->name
[0] = 'F';
11252 func
[i
]->aux
->stack_depth
= env
->subprog_info
[i
].stack_depth
;
11253 func
[i
]->jit_requested
= 1;
11254 func
[i
]->aux
->linfo
= prog
->aux
->linfo
;
11255 func
[i
]->aux
->nr_linfo
= prog
->aux
->nr_linfo
;
11256 func
[i
]->aux
->jited_linfo
= prog
->aux
->jited_linfo
;
11257 func
[i
]->aux
->linfo_idx
= env
->subprog_info
[i
].linfo_idx
;
11259 insn
= func
[i
]->insnsi
;
11260 for (j
= 0; j
< func
[i
]->len
; j
++, insn
++) {
11261 if (BPF_CLASS(insn
->code
) == BPF_LDX
&&
11262 BPF_MODE(insn
->code
) == BPF_PROBE_MEM
)
11265 func
[i
]->aux
->num_exentries
= num_exentries
;
11266 func
[i
]->aux
->tail_call_reachable
= env
->subprog_info
[i
].tail_call_reachable
;
11267 func
[i
] = bpf_int_jit_compile(func
[i
]);
11268 if (!func
[i
]->jited
) {
11275 /* Untrack main program's aux structs so that during map_poke_run()
11276 * we will not stumble upon the unfilled poke descriptors; each
11277 * of the main program's poke descs got distributed across subprogs
11278 * and got tracked onto map, so we are sure that none of them will
11279 * be missed after the operation below
11281 for (i
= 0; i
< prog
->aux
->size_poke_tab
; i
++) {
11282 map_ptr
= prog
->aux
->poke_tab
[i
].tail_call
.map
;
11284 map_ptr
->ops
->map_poke_untrack(map_ptr
, prog
->aux
);
11287 /* at this point all bpf functions were successfully JITed
11288 * now populate all bpf_calls with correct addresses and
11289 * run last pass of JIT
11291 for (i
= 0; i
< env
->subprog_cnt
; i
++) {
11292 insn
= func
[i
]->insnsi
;
11293 for (j
= 0; j
< func
[i
]->len
; j
++, insn
++) {
11294 if (insn
->code
!= (BPF_JMP
| BPF_CALL
) ||
11295 insn
->src_reg
!= BPF_PSEUDO_CALL
)
11297 subprog
= insn
->off
;
11298 insn
->imm
= BPF_CAST_CALL(func
[subprog
]->bpf_func
) -
11302 /* we use the aux data to keep a list of the start addresses
11303 * of the JITed images for each function in the program
11305 * for some architectures, such as powerpc64, the imm field
11306 * might not be large enough to hold the offset of the start
11307 * address of the callee's JITed image from __bpf_call_base
11309 * in such cases, we can lookup the start address of a callee
11310 * by using its subprog id, available from the off field of
11311 * the call instruction, as an index for this list
11313 func
[i
]->aux
->func
= func
;
11314 func
[i
]->aux
->func_cnt
= env
->subprog_cnt
;
11316 for (i
= 0; i
< env
->subprog_cnt
; i
++) {
11317 old_bpf_func
= func
[i
]->bpf_func
;
11318 tmp
= bpf_int_jit_compile(func
[i
]);
11319 if (tmp
!= func
[i
] || func
[i
]->bpf_func
!= old_bpf_func
) {
11320 verbose(env
, "JIT doesn't support bpf-to-bpf calls\n");
11327 /* finally lock prog and jit images for all functions and
11328 * populate kallsysm
11330 for (i
= 0; i
< env
->subprog_cnt
; i
++) {
11331 bpf_prog_lock_ro(func
[i
]);
11332 bpf_prog_kallsyms_add(func
[i
]);
11335 /* Last step: make now unused interpreter insns from main
11336 * prog consistent for later dump requests, so they can
11337 * later look the same as if they were interpreted only.
11339 for (i
= 0, insn
= prog
->insnsi
; i
< prog
->len
; i
++, insn
++) {
11340 if (insn
->code
!= (BPF_JMP
| BPF_CALL
) ||
11341 insn
->src_reg
!= BPF_PSEUDO_CALL
)
11343 insn
->off
= env
->insn_aux_data
[i
].call_imm
;
11344 subprog
= find_subprog(env
, i
+ insn
->off
+ 1);
11345 insn
->imm
= subprog
;
11349 prog
->bpf_func
= func
[0]->bpf_func
;
11350 prog
->aux
->func
= func
;
11351 prog
->aux
->func_cnt
= env
->subprog_cnt
;
11352 bpf_prog_free_unused_jited_linfo(prog
);
11355 for (i
= 0; i
< env
->subprog_cnt
; i
++) {
11359 for (j
= 0; j
< func
[i
]->aux
->size_poke_tab
; j
++) {
11360 map_ptr
= func
[i
]->aux
->poke_tab
[j
].tail_call
.map
;
11361 map_ptr
->ops
->map_poke_untrack(map_ptr
, func
[i
]->aux
);
11363 bpf_jit_free(func
[i
]);
11367 /* cleanup main prog to be interpreted */
11368 prog
->jit_requested
= 0;
11369 for (i
= 0, insn
= prog
->insnsi
; i
< prog
->len
; i
++, insn
++) {
11370 if (insn
->code
!= (BPF_JMP
| BPF_CALL
) ||
11371 insn
->src_reg
!= BPF_PSEUDO_CALL
)
11374 insn
->imm
= env
->insn_aux_data
[i
].call_imm
;
11376 bpf_prog_free_jited_linfo(prog
);
11380 static int fixup_call_args(struct bpf_verifier_env
*env
)
11382 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
11383 struct bpf_prog
*prog
= env
->prog
;
11384 struct bpf_insn
*insn
= prog
->insnsi
;
11389 if (env
->prog
->jit_requested
&&
11390 !bpf_prog_is_dev_bound(env
->prog
->aux
)) {
11391 err
= jit_subprogs(env
);
11394 if (err
== -EFAULT
)
11397 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
11398 if (env
->subprog_cnt
> 1 && env
->prog
->aux
->tail_call_reachable
) {
11399 /* When JIT fails the progs with bpf2bpf calls and tail_calls
11400 * have to be rejected, since interpreter doesn't support them yet.
11402 verbose(env
, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
11405 for (i
= 0; i
< prog
->len
; i
++, insn
++) {
11406 if (insn
->code
!= (BPF_JMP
| BPF_CALL
) ||
11407 insn
->src_reg
!= BPF_PSEUDO_CALL
)
11409 depth
= get_callee_stack_depth(env
, insn
, i
);
11412 bpf_patch_call_args(insn
, depth
);
11419 /* fixup insn->imm field of bpf_call instructions
11420 * and inline eligible helpers as explicit sequence of BPF instructions
11422 * this function is called after eBPF program passed verification
11424 static int fixup_bpf_calls(struct bpf_verifier_env
*env
)
11426 struct bpf_prog
*prog
= env
->prog
;
11427 bool expect_blinding
= bpf_jit_blinding_enabled(prog
);
11428 struct bpf_insn
*insn
= prog
->insnsi
;
11429 const struct bpf_func_proto
*fn
;
11430 const int insn_cnt
= prog
->len
;
11431 const struct bpf_map_ops
*ops
;
11432 struct bpf_insn_aux_data
*aux
;
11433 struct bpf_insn insn_buf
[16];
11434 struct bpf_prog
*new_prog
;
11435 struct bpf_map
*map_ptr
;
11436 int i
, ret
, cnt
, delta
= 0;
11438 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
11439 if (insn
->code
== (BPF_ALU64
| BPF_MOD
| BPF_X
) ||
11440 insn
->code
== (BPF_ALU64
| BPF_DIV
| BPF_X
) ||
11441 insn
->code
== (BPF_ALU
| BPF_MOD
| BPF_X
) ||
11442 insn
->code
== (BPF_ALU
| BPF_DIV
| BPF_X
)) {
11443 bool is64
= BPF_CLASS(insn
->code
) == BPF_ALU64
;
11444 bool isdiv
= BPF_OP(insn
->code
) == BPF_DIV
;
11445 struct bpf_insn
*patchlet
;
11446 struct bpf_insn chk_and_div
[] = {
11447 /* [R,W]x div 0 -> 0 */
11448 BPF_RAW_INSN((is64
? BPF_JMP
: BPF_JMP32
) |
11449 BPF_JNE
| BPF_K
, insn
->src_reg
,
11451 BPF_ALU32_REG(BPF_XOR
, insn
->dst_reg
, insn
->dst_reg
),
11452 BPF_JMP_IMM(BPF_JA
, 0, 0, 1),
11455 struct bpf_insn chk_and_mod
[] = {
11456 /* [R,W]x mod 0 -> [R,W]x */
11457 BPF_RAW_INSN((is64
? BPF_JMP
: BPF_JMP32
) |
11458 BPF_JEQ
| BPF_K
, insn
->src_reg
,
11459 0, 1 + (is64
? 0 : 1), 0),
11461 BPF_JMP_IMM(BPF_JA
, 0, 0, 1),
11462 BPF_MOV32_REG(insn
->dst_reg
, insn
->dst_reg
),
11465 patchlet
= isdiv
? chk_and_div
: chk_and_mod
;
11466 cnt
= isdiv
? ARRAY_SIZE(chk_and_div
) :
11467 ARRAY_SIZE(chk_and_mod
) - (is64
? 2 : 0);
11469 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, patchlet
, cnt
);
11474 env
->prog
= prog
= new_prog
;
11475 insn
= new_prog
->insnsi
+ i
+ delta
;
11479 if (BPF_CLASS(insn
->code
) == BPF_LD
&&
11480 (BPF_MODE(insn
->code
) == BPF_ABS
||
11481 BPF_MODE(insn
->code
) == BPF_IND
)) {
11482 cnt
= env
->ops
->gen_ld_abs(insn
, insn_buf
);
11483 if (cnt
== 0 || cnt
>= ARRAY_SIZE(insn_buf
)) {
11484 verbose(env
, "bpf verifier is misconfigured\n");
11488 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
, cnt
);
11493 env
->prog
= prog
= new_prog
;
11494 insn
= new_prog
->insnsi
+ i
+ delta
;
11498 if (insn
->code
== (BPF_ALU64
| BPF_ADD
| BPF_X
) ||
11499 insn
->code
== (BPF_ALU64
| BPF_SUB
| BPF_X
)) {
11500 const u8 code_add
= BPF_ALU64
| BPF_ADD
| BPF_X
;
11501 const u8 code_sub
= BPF_ALU64
| BPF_SUB
| BPF_X
;
11502 struct bpf_insn insn_buf
[16];
11503 struct bpf_insn
*patch
= &insn_buf
[0];
11504 bool issrc
, isneg
, isimm
;
11507 aux
= &env
->insn_aux_data
[i
+ delta
];
11508 if (!aux
->alu_state
||
11509 aux
->alu_state
== BPF_ALU_NON_POINTER
)
11512 isneg
= aux
->alu_state
& BPF_ALU_NEG_VALUE
;
11513 issrc
= (aux
->alu_state
& BPF_ALU_SANITIZE
) ==
11514 BPF_ALU_SANITIZE_SRC
;
11515 isimm
= aux
->alu_state
& BPF_ALU_IMMEDIATE
;
11517 off_reg
= issrc
? insn
->src_reg
: insn
->dst_reg
;
11519 *patch
++ = BPF_MOV32_IMM(BPF_REG_AX
, aux
->alu_limit
);
11522 *patch
++ = BPF_ALU64_IMM(BPF_MUL
, off_reg
, -1);
11523 *patch
++ = BPF_MOV32_IMM(BPF_REG_AX
, aux
->alu_limit
);
11524 *patch
++ = BPF_ALU64_REG(BPF_SUB
, BPF_REG_AX
, off_reg
);
11525 *patch
++ = BPF_ALU64_REG(BPF_OR
, BPF_REG_AX
, off_reg
);
11526 *patch
++ = BPF_ALU64_IMM(BPF_NEG
, BPF_REG_AX
, 0);
11527 *patch
++ = BPF_ALU64_IMM(BPF_ARSH
, BPF_REG_AX
, 63);
11528 *patch
++ = BPF_ALU64_REG(BPF_AND
, BPF_REG_AX
, off_reg
);
11531 *patch
++ = BPF_MOV64_REG(insn
->dst_reg
, insn
->src_reg
);
11532 insn
->src_reg
= BPF_REG_AX
;
11534 insn
->code
= insn
->code
== code_add
?
11535 code_sub
: code_add
;
11537 if (issrc
&& isneg
&& !isimm
)
11538 *patch
++ = BPF_ALU64_IMM(BPF_MUL
, off_reg
, -1);
11539 cnt
= patch
- insn_buf
;
11541 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
, cnt
);
11546 env
->prog
= prog
= new_prog
;
11547 insn
= new_prog
->insnsi
+ i
+ delta
;
11551 if (insn
->code
!= (BPF_JMP
| BPF_CALL
))
11553 if (insn
->src_reg
== BPF_PSEUDO_CALL
)
11556 if (insn
->imm
== BPF_FUNC_get_route_realm
)
11557 prog
->dst_needed
= 1;
11558 if (insn
->imm
== BPF_FUNC_get_prandom_u32
)
11559 bpf_user_rnd_init_once();
11560 if (insn
->imm
== BPF_FUNC_override_return
)
11561 prog
->kprobe_override
= 1;
11562 if (insn
->imm
== BPF_FUNC_tail_call
) {
11563 /* If we tail call into other programs, we
11564 * cannot make any assumptions since they can
11565 * be replaced dynamically during runtime in
11566 * the program array.
11568 prog
->cb_access
= 1;
11569 if (!allow_tail_call_in_subprogs(env
))
11570 prog
->aux
->stack_depth
= MAX_BPF_STACK
;
11571 prog
->aux
->max_pkt_offset
= MAX_PACKET_OFF
;
11573 /* mark bpf_tail_call as different opcode to avoid
11574 * conditional branch in the interpeter for every normal
11575 * call and to prevent accidental JITing by JIT compiler
11576 * that doesn't support bpf_tail_call yet
11579 insn
->code
= BPF_JMP
| BPF_TAIL_CALL
;
11581 aux
= &env
->insn_aux_data
[i
+ delta
];
11582 if (env
->bpf_capable
&& !expect_blinding
&&
11583 prog
->jit_requested
&&
11584 !bpf_map_key_poisoned(aux
) &&
11585 !bpf_map_ptr_poisoned(aux
) &&
11586 !bpf_map_ptr_unpriv(aux
)) {
11587 struct bpf_jit_poke_descriptor desc
= {
11588 .reason
= BPF_POKE_REASON_TAIL_CALL
,
11589 .tail_call
.map
= BPF_MAP_PTR(aux
->map_ptr_state
),
11590 .tail_call
.key
= bpf_map_key_immediate(aux
),
11591 .insn_idx
= i
+ delta
,
11594 ret
= bpf_jit_add_poke_descriptor(prog
, &desc
);
11596 verbose(env
, "adding tail call poke descriptor failed\n");
11600 insn
->imm
= ret
+ 1;
11604 if (!bpf_map_ptr_unpriv(aux
))
11607 /* instead of changing every JIT dealing with tail_call
11608 * emit two extra insns:
11609 * if (index >= max_entries) goto out;
11610 * index &= array->index_mask;
11611 * to avoid out-of-bounds cpu speculation
11613 if (bpf_map_ptr_poisoned(aux
)) {
11614 verbose(env
, "tail_call abusing map_ptr\n");
11618 map_ptr
= BPF_MAP_PTR(aux
->map_ptr_state
);
11619 insn_buf
[0] = BPF_JMP_IMM(BPF_JGE
, BPF_REG_3
,
11620 map_ptr
->max_entries
, 2);
11621 insn_buf
[1] = BPF_ALU32_IMM(BPF_AND
, BPF_REG_3
,
11622 container_of(map_ptr
,
11625 insn_buf
[2] = *insn
;
11627 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
, cnt
);
11632 env
->prog
= prog
= new_prog
;
11633 insn
= new_prog
->insnsi
+ i
+ delta
;
11637 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
11638 * and other inlining handlers are currently limited to 64 bit
11641 if (prog
->jit_requested
&& BITS_PER_LONG
== 64 &&
11642 (insn
->imm
== BPF_FUNC_map_lookup_elem
||
11643 insn
->imm
== BPF_FUNC_map_update_elem
||
11644 insn
->imm
== BPF_FUNC_map_delete_elem
||
11645 insn
->imm
== BPF_FUNC_map_push_elem
||
11646 insn
->imm
== BPF_FUNC_map_pop_elem
||
11647 insn
->imm
== BPF_FUNC_map_peek_elem
)) {
11648 aux
= &env
->insn_aux_data
[i
+ delta
];
11649 if (bpf_map_ptr_poisoned(aux
))
11650 goto patch_call_imm
;
11652 map_ptr
= BPF_MAP_PTR(aux
->map_ptr_state
);
11653 ops
= map_ptr
->ops
;
11654 if (insn
->imm
== BPF_FUNC_map_lookup_elem
&&
11655 ops
->map_gen_lookup
) {
11656 cnt
= ops
->map_gen_lookup(map_ptr
, insn_buf
);
11657 if (cnt
== -EOPNOTSUPP
)
11658 goto patch_map_ops_generic
;
11659 if (cnt
<= 0 || cnt
>= ARRAY_SIZE(insn_buf
)) {
11660 verbose(env
, "bpf verifier is misconfigured\n");
11664 new_prog
= bpf_patch_insn_data(env
, i
+ delta
,
11670 env
->prog
= prog
= new_prog
;
11671 insn
= new_prog
->insnsi
+ i
+ delta
;
11675 BUILD_BUG_ON(!__same_type(ops
->map_lookup_elem
,
11676 (void *(*)(struct bpf_map
*map
, void *key
))NULL
));
11677 BUILD_BUG_ON(!__same_type(ops
->map_delete_elem
,
11678 (int (*)(struct bpf_map
*map
, void *key
))NULL
));
11679 BUILD_BUG_ON(!__same_type(ops
->map_update_elem
,
11680 (int (*)(struct bpf_map
*map
, void *key
, void *value
,
11682 BUILD_BUG_ON(!__same_type(ops
->map_push_elem
,
11683 (int (*)(struct bpf_map
*map
, void *value
,
11685 BUILD_BUG_ON(!__same_type(ops
->map_pop_elem
,
11686 (int (*)(struct bpf_map
*map
, void *value
))NULL
));
11687 BUILD_BUG_ON(!__same_type(ops
->map_peek_elem
,
11688 (int (*)(struct bpf_map
*map
, void *value
))NULL
));
11689 patch_map_ops_generic
:
11690 switch (insn
->imm
) {
11691 case BPF_FUNC_map_lookup_elem
:
11692 insn
->imm
= BPF_CAST_CALL(ops
->map_lookup_elem
) -
11695 case BPF_FUNC_map_update_elem
:
11696 insn
->imm
= BPF_CAST_CALL(ops
->map_update_elem
) -
11699 case BPF_FUNC_map_delete_elem
:
11700 insn
->imm
= BPF_CAST_CALL(ops
->map_delete_elem
) -
11703 case BPF_FUNC_map_push_elem
:
11704 insn
->imm
= BPF_CAST_CALL(ops
->map_push_elem
) -
11707 case BPF_FUNC_map_pop_elem
:
11708 insn
->imm
= BPF_CAST_CALL(ops
->map_pop_elem
) -
11711 case BPF_FUNC_map_peek_elem
:
11712 insn
->imm
= BPF_CAST_CALL(ops
->map_peek_elem
) -
11717 goto patch_call_imm
;
11720 if (prog
->jit_requested
&& BITS_PER_LONG
== 64 &&
11721 insn
->imm
== BPF_FUNC_jiffies64
) {
11722 struct bpf_insn ld_jiffies_addr
[2] = {
11723 BPF_LD_IMM64(BPF_REG_0
,
11724 (unsigned long)&jiffies
),
11727 insn_buf
[0] = ld_jiffies_addr
[0];
11728 insn_buf
[1] = ld_jiffies_addr
[1];
11729 insn_buf
[2] = BPF_LDX_MEM(BPF_DW
, BPF_REG_0
,
11733 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
,
11739 env
->prog
= prog
= new_prog
;
11740 insn
= new_prog
->insnsi
+ i
+ delta
;
11745 fn
= env
->ops
->get_func_proto(insn
->imm
, env
->prog
);
11746 /* all functions that have prototype and verifier allowed
11747 * programs to call them, must be real in-kernel functions
11751 "kernel subsystem misconfigured func %s#%d\n",
11752 func_id_name(insn
->imm
), insn
->imm
);
11755 insn
->imm
= fn
->func
- __bpf_call_base
;
11758 /* Since poke tab is now finalized, publish aux to tracker. */
11759 for (i
= 0; i
< prog
->aux
->size_poke_tab
; i
++) {
11760 map_ptr
= prog
->aux
->poke_tab
[i
].tail_call
.map
;
11761 if (!map_ptr
->ops
->map_poke_track
||
11762 !map_ptr
->ops
->map_poke_untrack
||
11763 !map_ptr
->ops
->map_poke_run
) {
11764 verbose(env
, "bpf verifier is misconfigured\n");
11768 ret
= map_ptr
->ops
->map_poke_track(map_ptr
, prog
->aux
);
11770 verbose(env
, "tracking tail call prog failed\n");
11778 static void free_states(struct bpf_verifier_env
*env
)
11780 struct bpf_verifier_state_list
*sl
, *sln
;
11783 sl
= env
->free_list
;
11786 free_verifier_state(&sl
->state
, false);
11790 env
->free_list
= NULL
;
11792 if (!env
->explored_states
)
11795 for (i
= 0; i
< state_htab_size(env
); i
++) {
11796 sl
= env
->explored_states
[i
];
11800 free_verifier_state(&sl
->state
, false);
11804 env
->explored_states
[i
] = NULL
;
11808 /* The verifier is using insn_aux_data[] to store temporary data during
11809 * verification and to store information for passes that run after the
11810 * verification like dead code sanitization. do_check_common() for subprogram N
11811 * may analyze many other subprograms. sanitize_insn_aux_data() clears all
11812 * temporary data after do_check_common() finds that subprogram N cannot be
11813 * verified independently. pass_cnt counts the number of times
11814 * do_check_common() was run and insn->aux->seen tells the pass number
11815 * insn_aux_data was touched. These variables are compared to clear temporary
11816 * data from failed pass. For testing and experiments do_check_common() can be
11817 * run multiple times even when prior attempt to verify is unsuccessful.
11819 static void sanitize_insn_aux_data(struct bpf_verifier_env
*env
)
11821 struct bpf_insn
*insn
= env
->prog
->insnsi
;
11822 struct bpf_insn_aux_data
*aux
;
11825 for (i
= 0; i
< env
->prog
->len
; i
++) {
11826 class = BPF_CLASS(insn
[i
].code
);
11827 if (class != BPF_LDX
&& class != BPF_STX
)
11829 aux
= &env
->insn_aux_data
[i
];
11830 if (aux
->seen
!= env
->pass_cnt
)
11832 memset(aux
, 0, offsetof(typeof(*aux
), orig_idx
));
11836 static int do_check_common(struct bpf_verifier_env
*env
, int subprog
)
11838 bool pop_log
= !(env
->log
.level
& BPF_LOG_LEVEL2
);
11839 struct bpf_verifier_state
*state
;
11840 struct bpf_reg_state
*regs
;
11843 env
->prev_linfo
= NULL
;
11846 state
= kzalloc(sizeof(struct bpf_verifier_state
), GFP_KERNEL
);
11849 state
->curframe
= 0;
11850 state
->speculative
= false;
11851 state
->branches
= 1;
11852 state
->frame
[0] = kzalloc(sizeof(struct bpf_func_state
), GFP_KERNEL
);
11853 if (!state
->frame
[0]) {
11857 env
->cur_state
= state
;
11858 init_func_state(env
, state
->frame
[0],
11859 BPF_MAIN_FUNC
/* callsite */,
11863 regs
= state
->frame
[state
->curframe
]->regs
;
11864 if (subprog
|| env
->prog
->type
== BPF_PROG_TYPE_EXT
) {
11865 ret
= btf_prepare_func_args(env
, subprog
, regs
);
11868 for (i
= BPF_REG_1
; i
<= BPF_REG_5
; i
++) {
11869 if (regs
[i
].type
== PTR_TO_CTX
)
11870 mark_reg_known_zero(env
, regs
, i
);
11871 else if (regs
[i
].type
== SCALAR_VALUE
)
11872 mark_reg_unknown(env
, regs
, i
);
11875 /* 1st arg to a function */
11876 regs
[BPF_REG_1
].type
= PTR_TO_CTX
;
11877 mark_reg_known_zero(env
, regs
, BPF_REG_1
);
11878 ret
= btf_check_func_arg_match(env
, subprog
, regs
);
11879 if (ret
== -EFAULT
)
11880 /* unlikely verifier bug. abort.
11881 * ret == 0 and ret < 0 are sadly acceptable for
11882 * main() function due to backward compatibility.
11883 * Like socket filter program may be written as:
11884 * int bpf_prog(struct pt_regs *ctx)
11885 * and never dereference that ctx in the program.
11886 * 'struct pt_regs' is a type mismatch for socket
11887 * filter that should be using 'struct __sk_buff'.
11892 ret
= do_check(env
);
11894 /* check for NULL is necessary, since cur_state can be freed inside
11895 * do_check() under memory pressure.
11897 if (env
->cur_state
) {
11898 free_verifier_state(env
->cur_state
, true);
11899 env
->cur_state
= NULL
;
11901 while (!pop_stack(env
, NULL
, NULL
, false));
11902 if (!ret
&& pop_log
)
11903 bpf_vlog_reset(&env
->log
, 0);
11906 /* clean aux data in case subprog was rejected */
11907 sanitize_insn_aux_data(env
);
11911 /* Verify all global functions in a BPF program one by one based on their BTF.
11912 * All global functions must pass verification. Otherwise the whole program is rejected.
11923 * foo() will be verified first for R1=any_scalar_value. During verification it
11924 * will be assumed that bar() already verified successfully and call to bar()
11925 * from foo() will be checked for type match only. Later bar() will be verified
11926 * independently to check that it's safe for R1=any_scalar_value.
11928 static int do_check_subprogs(struct bpf_verifier_env
*env
)
11930 struct bpf_prog_aux
*aux
= env
->prog
->aux
;
11933 if (!aux
->func_info
)
11936 for (i
= 1; i
< env
->subprog_cnt
; i
++) {
11937 if (aux
->func_info_aux
[i
].linkage
!= BTF_FUNC_GLOBAL
)
11939 env
->insn_idx
= env
->subprog_info
[i
].start
;
11940 WARN_ON_ONCE(env
->insn_idx
== 0);
11941 ret
= do_check_common(env
, i
);
11944 } else if (env
->log
.level
& BPF_LOG_LEVEL
) {
11946 "Func#%d is safe for any args that match its prototype\n",
11953 static int do_check_main(struct bpf_verifier_env
*env
)
11958 ret
= do_check_common(env
, 0);
11960 env
->prog
->aux
->stack_depth
= env
->subprog_info
[0].stack_depth
;
11965 static void print_verification_stats(struct bpf_verifier_env
*env
)
11969 if (env
->log
.level
& BPF_LOG_STATS
) {
11970 verbose(env
, "verification time %lld usec\n",
11971 div_u64(env
->verification_time
, 1000));
11972 verbose(env
, "stack depth ");
11973 for (i
= 0; i
< env
->subprog_cnt
; i
++) {
11974 u32 depth
= env
->subprog_info
[i
].stack_depth
;
11976 verbose(env
, "%d", depth
);
11977 if (i
+ 1 < env
->subprog_cnt
)
11980 verbose(env
, "\n");
11982 verbose(env
, "processed %d insns (limit %d) max_states_per_insn %d "
11983 "total_states %d peak_states %d mark_read %d\n",
11984 env
->insn_processed
, BPF_COMPLEXITY_LIMIT_INSNS
,
11985 env
->max_states_per_insn
, env
->total_states
,
11986 env
->peak_states
, env
->longest_mark_read_walk
);
11989 static int check_struct_ops_btf_id(struct bpf_verifier_env
*env
)
11991 const struct btf_type
*t
, *func_proto
;
11992 const struct bpf_struct_ops
*st_ops
;
11993 const struct btf_member
*member
;
11994 struct bpf_prog
*prog
= env
->prog
;
11995 u32 btf_id
, member_idx
;
11998 if (!prog
->gpl_compatible
) {
11999 verbose(env
, "struct ops programs must have a GPL compatible license\n");
12003 btf_id
= prog
->aux
->attach_btf_id
;
12004 st_ops
= bpf_struct_ops_find(btf_id
);
12006 verbose(env
, "attach_btf_id %u is not a supported struct\n",
12012 member_idx
= prog
->expected_attach_type
;
12013 if (member_idx
>= btf_type_vlen(t
)) {
12014 verbose(env
, "attach to invalid member idx %u of struct %s\n",
12015 member_idx
, st_ops
->name
);
12019 member
= &btf_type_member(t
)[member_idx
];
12020 mname
= btf_name_by_offset(btf_vmlinux
, member
->name_off
);
12021 func_proto
= btf_type_resolve_func_ptr(btf_vmlinux
, member
->type
,
12024 verbose(env
, "attach to invalid member %s(@idx %u) of struct %s\n",
12025 mname
, member_idx
, st_ops
->name
);
12029 if (st_ops
->check_member
) {
12030 int err
= st_ops
->check_member(t
, member
);
12033 verbose(env
, "attach to unsupported member %s of struct %s\n",
12034 mname
, st_ops
->name
);
12039 prog
->aux
->attach_func_proto
= func_proto
;
12040 prog
->aux
->attach_func_name
= mname
;
12041 env
->ops
= st_ops
->verifier_ops
;
12045 #define SECURITY_PREFIX "security_"
12047 static int check_attach_modify_return(unsigned long addr
, const char *func_name
)
12049 if (within_error_injection_list(addr
) ||
12050 !strncmp(SECURITY_PREFIX
, func_name
, sizeof(SECURITY_PREFIX
) - 1))
12056 /* list of non-sleepable functions that are otherwise on
12057 * ALLOW_ERROR_INJECTION list
12059 BTF_SET_START(btf_non_sleepable_error_inject
)
12060 /* Three functions below can be called from sleepable and non-sleepable context.
12061 * Assume non-sleepable from bpf safety point of view.
12063 BTF_ID(func
, __add_to_page_cache_locked
)
12064 BTF_ID(func
, should_fail_alloc_page
)
12065 BTF_ID(func
, should_failslab
)
12066 BTF_SET_END(btf_non_sleepable_error_inject
)
12068 static int check_non_sleepable_error_inject(u32 btf_id
)
12070 return btf_id_set_contains(&btf_non_sleepable_error_inject
, btf_id
);
12073 int bpf_check_attach_target(struct bpf_verifier_log
*log
,
12074 const struct bpf_prog
*prog
,
12075 const struct bpf_prog
*tgt_prog
,
12077 struct bpf_attach_target_info
*tgt_info
)
12079 bool prog_extension
= prog
->type
== BPF_PROG_TYPE_EXT
;
12080 const char prefix
[] = "btf_trace_";
12081 int ret
= 0, subprog
= -1, i
;
12082 const struct btf_type
*t
;
12083 bool conservative
= true;
12089 bpf_log(log
, "Tracing programs must provide btf_id\n");
12092 btf
= tgt_prog
? tgt_prog
->aux
->btf
: prog
->aux
->attach_btf
;
12095 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
12098 t
= btf_type_by_id(btf
, btf_id
);
12100 bpf_log(log
, "attach_btf_id %u is invalid\n", btf_id
);
12103 tname
= btf_name_by_offset(btf
, t
->name_off
);
12105 bpf_log(log
, "attach_btf_id %u doesn't have a name\n", btf_id
);
12109 struct bpf_prog_aux
*aux
= tgt_prog
->aux
;
12111 for (i
= 0; i
< aux
->func_info_cnt
; i
++)
12112 if (aux
->func_info
[i
].type_id
== btf_id
) {
12116 if (subprog
== -1) {
12117 bpf_log(log
, "Subprog %s doesn't exist\n", tname
);
12120 conservative
= aux
->func_info_aux
[subprog
].unreliable
;
12121 if (prog_extension
) {
12122 if (conservative
) {
12124 "Cannot replace static functions\n");
12127 if (!prog
->jit_requested
) {
12129 "Extension programs should be JITed\n");
12133 if (!tgt_prog
->jited
) {
12134 bpf_log(log
, "Can attach to only JITed progs\n");
12137 if (tgt_prog
->type
== prog
->type
) {
12138 /* Cannot fentry/fexit another fentry/fexit program.
12139 * Cannot attach program extension to another extension.
12140 * It's ok to attach fentry/fexit to extension program.
12142 bpf_log(log
, "Cannot recursively attach\n");
12145 if (tgt_prog
->type
== BPF_PROG_TYPE_TRACING
&&
12147 (tgt_prog
->expected_attach_type
== BPF_TRACE_FENTRY
||
12148 tgt_prog
->expected_attach_type
== BPF_TRACE_FEXIT
)) {
12149 /* Program extensions can extend all program types
12150 * except fentry/fexit. The reason is the following.
12151 * The fentry/fexit programs are used for performance
12152 * analysis, stats and can be attached to any program
12153 * type except themselves. When extension program is
12154 * replacing XDP function it is necessary to allow
12155 * performance analysis of all functions. Both original
12156 * XDP program and its program extension. Hence
12157 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
12158 * allowed. If extending of fentry/fexit was allowed it
12159 * would be possible to create long call chain
12160 * fentry->extension->fentry->extension beyond
12161 * reasonable stack size. Hence extending fentry is not
12164 bpf_log(log
, "Cannot extend fentry/fexit\n");
12168 if (prog_extension
) {
12169 bpf_log(log
, "Cannot replace kernel functions\n");
12174 switch (prog
->expected_attach_type
) {
12175 case BPF_TRACE_RAW_TP
:
12178 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
12181 if (!btf_type_is_typedef(t
)) {
12182 bpf_log(log
, "attach_btf_id %u is not a typedef\n",
12186 if (strncmp(prefix
, tname
, sizeof(prefix
) - 1)) {
12187 bpf_log(log
, "attach_btf_id %u points to wrong type name %s\n",
12191 tname
+= sizeof(prefix
) - 1;
12192 t
= btf_type_by_id(btf
, t
->type
);
12193 if (!btf_type_is_ptr(t
))
12194 /* should never happen in valid vmlinux build */
12196 t
= btf_type_by_id(btf
, t
->type
);
12197 if (!btf_type_is_func_proto(t
))
12198 /* should never happen in valid vmlinux build */
12202 case BPF_TRACE_ITER
:
12203 if (!btf_type_is_func(t
)) {
12204 bpf_log(log
, "attach_btf_id %u is not a function\n",
12208 t
= btf_type_by_id(btf
, t
->type
);
12209 if (!btf_type_is_func_proto(t
))
12211 ret
= btf_distill_func_proto(log
, btf
, t
, tname
, &tgt_info
->fmodel
);
12216 if (!prog_extension
)
12219 case BPF_MODIFY_RETURN
:
12221 case BPF_TRACE_FENTRY
:
12222 case BPF_TRACE_FEXIT
:
12223 if (!btf_type_is_func(t
)) {
12224 bpf_log(log
, "attach_btf_id %u is not a function\n",
12228 if (prog_extension
&&
12229 btf_check_type_match(log
, prog
, btf
, t
))
12231 t
= btf_type_by_id(btf
, t
->type
);
12232 if (!btf_type_is_func_proto(t
))
12235 if ((prog
->aux
->saved_dst_prog_type
|| prog
->aux
->saved_dst_attach_type
) &&
12236 (!tgt_prog
|| prog
->aux
->saved_dst_prog_type
!= tgt_prog
->type
||
12237 prog
->aux
->saved_dst_attach_type
!= tgt_prog
->expected_attach_type
))
12240 if (tgt_prog
&& conservative
)
12243 ret
= btf_distill_func_proto(log
, btf
, t
, tname
, &tgt_info
->fmodel
);
12249 addr
= (long) tgt_prog
->bpf_func
;
12251 addr
= (long) tgt_prog
->aux
->func
[subprog
]->bpf_func
;
12253 addr
= kallsyms_lookup_name(tname
);
12256 "The address of function %s cannot be found\n",
12262 if (prog
->aux
->sleepable
) {
12264 switch (prog
->type
) {
12265 case BPF_PROG_TYPE_TRACING
:
12266 /* fentry/fexit/fmod_ret progs can be sleepable only if they are
12267 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
12269 if (!check_non_sleepable_error_inject(btf_id
) &&
12270 within_error_injection_list(addr
))
12273 case BPF_PROG_TYPE_LSM
:
12274 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
12275 * Only some of them are sleepable.
12277 if (bpf_lsm_is_sleepable_hook(btf_id
))
12284 bpf_log(log
, "%s is not sleepable\n", tname
);
12287 } else if (prog
->expected_attach_type
== BPF_MODIFY_RETURN
) {
12289 bpf_log(log
, "can't modify return codes of BPF programs\n");
12292 ret
= check_attach_modify_return(addr
, tname
);
12294 bpf_log(log
, "%s() is not modifiable\n", tname
);
12301 tgt_info
->tgt_addr
= addr
;
12302 tgt_info
->tgt_name
= tname
;
12303 tgt_info
->tgt_type
= t
;
12307 static int check_attach_btf_id(struct bpf_verifier_env
*env
)
12309 struct bpf_prog
*prog
= env
->prog
;
12310 struct bpf_prog
*tgt_prog
= prog
->aux
->dst_prog
;
12311 struct bpf_attach_target_info tgt_info
= {};
12312 u32 btf_id
= prog
->aux
->attach_btf_id
;
12313 struct bpf_trampoline
*tr
;
12317 if (prog
->aux
->sleepable
&& prog
->type
!= BPF_PROG_TYPE_TRACING
&&
12318 prog
->type
!= BPF_PROG_TYPE_LSM
) {
12319 verbose(env
, "Only fentry/fexit/fmod_ret and lsm programs can be sleepable\n");
12323 if (prog
->type
== BPF_PROG_TYPE_STRUCT_OPS
)
12324 return check_struct_ops_btf_id(env
);
12326 if (prog
->type
!= BPF_PROG_TYPE_TRACING
&&
12327 prog
->type
!= BPF_PROG_TYPE_LSM
&&
12328 prog
->type
!= BPF_PROG_TYPE_EXT
)
12331 ret
= bpf_check_attach_target(&env
->log
, prog
, tgt_prog
, btf_id
, &tgt_info
);
12335 if (tgt_prog
&& prog
->type
== BPF_PROG_TYPE_EXT
) {
12336 /* to make freplace equivalent to their targets, they need to
12337 * inherit env->ops and expected_attach_type for the rest of the
12340 env
->ops
= bpf_verifier_ops
[tgt_prog
->type
];
12341 prog
->expected_attach_type
= tgt_prog
->expected_attach_type
;
12344 /* store info about the attachment target that will be used later */
12345 prog
->aux
->attach_func_proto
= tgt_info
.tgt_type
;
12346 prog
->aux
->attach_func_name
= tgt_info
.tgt_name
;
12349 prog
->aux
->saved_dst_prog_type
= tgt_prog
->type
;
12350 prog
->aux
->saved_dst_attach_type
= tgt_prog
->expected_attach_type
;
12353 if (prog
->expected_attach_type
== BPF_TRACE_RAW_TP
) {
12354 prog
->aux
->attach_btf_trace
= true;
12356 } else if (prog
->expected_attach_type
== BPF_TRACE_ITER
) {
12357 if (!bpf_iter_prog_supported(prog
))
12362 if (prog
->type
== BPF_PROG_TYPE_LSM
) {
12363 ret
= bpf_lsm_verify_prog(&env
->log
, prog
);
12368 key
= bpf_trampoline_compute_key(tgt_prog
, prog
->aux
->attach_btf
, btf_id
);
12369 tr
= bpf_trampoline_get(key
, &tgt_info
);
12373 prog
->aux
->dst_trampoline
= tr
;
12377 struct btf
*bpf_get_btf_vmlinux(void)
12379 if (!btf_vmlinux
&& IS_ENABLED(CONFIG_DEBUG_INFO_BTF
)) {
12380 mutex_lock(&bpf_verifier_lock
);
12382 btf_vmlinux
= btf_parse_vmlinux();
12383 mutex_unlock(&bpf_verifier_lock
);
12385 return btf_vmlinux
;
12388 int bpf_check(struct bpf_prog
**prog
, union bpf_attr
*attr
,
12389 union bpf_attr __user
*uattr
)
12391 u64 start_time
= ktime_get_ns();
12392 struct bpf_verifier_env
*env
;
12393 struct bpf_verifier_log
*log
;
12394 int i
, len
, ret
= -EINVAL
;
12397 /* no program is valid */
12398 if (ARRAY_SIZE(bpf_verifier_ops
) == 0)
12401 /* 'struct bpf_verifier_env' can be global, but since it's not small,
12402 * allocate/free it every time bpf_check() is called
12404 env
= kzalloc(sizeof(struct bpf_verifier_env
), GFP_KERNEL
);
12409 len
= (*prog
)->len
;
12410 env
->insn_aux_data
=
12411 vzalloc(array_size(sizeof(struct bpf_insn_aux_data
), len
));
12413 if (!env
->insn_aux_data
)
12415 for (i
= 0; i
< len
; i
++)
12416 env
->insn_aux_data
[i
].orig_idx
= i
;
12418 env
->ops
= bpf_verifier_ops
[env
->prog
->type
];
12419 is_priv
= bpf_capable();
12421 bpf_get_btf_vmlinux();
12423 /* grab the mutex to protect few globals used by verifier */
12425 mutex_lock(&bpf_verifier_lock
);
12427 if (attr
->log_level
|| attr
->log_buf
|| attr
->log_size
) {
12428 /* user requested verbose verifier output
12429 * and supplied buffer to store the verification trace
12431 log
->level
= attr
->log_level
;
12432 log
->ubuf
= (char __user
*) (unsigned long) attr
->log_buf
;
12433 log
->len_total
= attr
->log_size
;
12436 /* log attributes have to be sane */
12437 if (log
->len_total
< 128 || log
->len_total
> UINT_MAX
>> 2 ||
12438 !log
->level
|| !log
->ubuf
|| log
->level
& ~BPF_LOG_MASK
)
12442 if (IS_ERR(btf_vmlinux
)) {
12443 /* Either gcc or pahole or kernel are broken. */
12444 verbose(env
, "in-kernel BTF is malformed\n");
12445 ret
= PTR_ERR(btf_vmlinux
);
12446 goto skip_full_check
;
12449 env
->strict_alignment
= !!(attr
->prog_flags
& BPF_F_STRICT_ALIGNMENT
);
12450 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
))
12451 env
->strict_alignment
= true;
12452 if (attr
->prog_flags
& BPF_F_ANY_ALIGNMENT
)
12453 env
->strict_alignment
= false;
12455 env
->allow_ptr_leaks
= bpf_allow_ptr_leaks();
12456 env
->allow_uninit_stack
= bpf_allow_uninit_stack();
12457 env
->allow_ptr_to_map_access
= bpf_allow_ptr_to_map_access();
12458 env
->bypass_spec_v1
= bpf_bypass_spec_v1();
12459 env
->bypass_spec_v4
= bpf_bypass_spec_v4();
12460 env
->bpf_capable
= bpf_capable();
12463 env
->test_state_freq
= attr
->prog_flags
& BPF_F_TEST_STATE_FREQ
;
12465 if (bpf_prog_is_dev_bound(env
->prog
->aux
)) {
12466 ret
= bpf_prog_offload_verifier_prep(env
->prog
);
12468 goto skip_full_check
;
12471 env
->explored_states
= kvcalloc(state_htab_size(env
),
12472 sizeof(struct bpf_verifier_state_list
*),
12475 if (!env
->explored_states
)
12476 goto skip_full_check
;
12478 ret
= check_subprogs(env
);
12480 goto skip_full_check
;
12482 ret
= check_btf_info(env
, attr
, uattr
);
12484 goto skip_full_check
;
12486 ret
= check_attach_btf_id(env
);
12488 goto skip_full_check
;
12490 ret
= resolve_pseudo_ldimm64(env
);
12492 goto skip_full_check
;
12494 ret
= check_cfg(env
);
12496 goto skip_full_check
;
12498 ret
= do_check_subprogs(env
);
12499 ret
= ret
?: do_check_main(env
);
12501 if (ret
== 0 && bpf_prog_is_dev_bound(env
->prog
->aux
))
12502 ret
= bpf_prog_offload_finalize(env
);
12505 kvfree(env
->explored_states
);
12508 ret
= check_max_stack_depth(env
);
12510 /* instruction rewrites happen after this point */
12513 opt_hard_wire_dead_code_branches(env
);
12515 ret
= opt_remove_dead_code(env
);
12517 ret
= opt_remove_nops(env
);
12520 sanitize_dead_code(env
);
12524 /* program is valid, convert *(u32*)(ctx + off) accesses */
12525 ret
= convert_ctx_accesses(env
);
12528 ret
= fixup_bpf_calls(env
);
12530 /* do 32-bit optimization after insn patching has done so those patched
12531 * insns could be handled correctly.
12533 if (ret
== 0 && !bpf_prog_is_dev_bound(env
->prog
->aux
)) {
12534 ret
= opt_subreg_zext_lo32_rnd_hi32(env
, attr
);
12535 env
->prog
->aux
->verifier_zext
= bpf_jit_needs_zext() ? !ret
12540 ret
= fixup_call_args(env
);
12542 env
->verification_time
= ktime_get_ns() - start_time
;
12543 print_verification_stats(env
);
12545 if (log
->level
&& bpf_verifier_log_full(log
))
12547 if (log
->level
&& !log
->ubuf
) {
12549 goto err_release_maps
;
12552 if (ret
== 0 && env
->used_map_cnt
) {
12553 /* if program passed verifier, update used_maps in bpf_prog_info */
12554 env
->prog
->aux
->used_maps
= kmalloc_array(env
->used_map_cnt
,
12555 sizeof(env
->used_maps
[0]),
12558 if (!env
->prog
->aux
->used_maps
) {
12560 goto err_release_maps
;
12563 memcpy(env
->prog
->aux
->used_maps
, env
->used_maps
,
12564 sizeof(env
->used_maps
[0]) * env
->used_map_cnt
);
12565 env
->prog
->aux
->used_map_cnt
= env
->used_map_cnt
;
12567 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
12568 * bpf_ld_imm64 instructions
12570 convert_pseudo_ld_imm64(env
);
12574 adjust_btf_func(env
);
12577 if (!env
->prog
->aux
->used_maps
)
12578 /* if we didn't copy map pointers into bpf_prog_info, release
12579 * them now. Otherwise free_used_maps() will release them.
12583 /* extension progs temporarily inherit the attach_type of their targets
12584 for verification purposes, so set it back to zero before returning
12586 if (env
->prog
->type
== BPF_PROG_TYPE_EXT
)
12587 env
->prog
->expected_attach_type
= 0;
12592 mutex_unlock(&bpf_verifier_lock
);
12593 vfree(env
->insn_aux_data
);