1 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
2 * Copyright (c) 2016 Facebook
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of version 2 of the GNU General Public
6 * License as published by the Free Software Foundation.
8 * This program is distributed in the hope that it will be useful, but
9 * WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
13 #include <linux/kernel.h>
14 #include <linux/types.h>
15 #include <linux/slab.h>
16 #include <linux/bpf.h>
17 #include <linux/bpf_verifier.h>
18 #include <linux/filter.h>
19 #include <net/netlink.h>
20 #include <linux/file.h>
21 #include <linux/vmalloc.h>
22 #include <linux/stringify.h>
26 static const struct bpf_verifier_ops
* const bpf_verifier_ops
[] = {
27 #define BPF_PROG_TYPE(_id, _name) \
28 [_id] = & _name ## _verifier_ops,
29 #define BPF_MAP_TYPE(_id, _ops)
30 #include <linux/bpf_types.h>
35 /* bpf_check() is a static code analyzer that walks eBPF program
36 * instruction by instruction and updates register/stack state.
37 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
39 * The first pass is depth-first-search to check that the program is a DAG.
40 * It rejects the following programs:
41 * - larger than BPF_MAXINSNS insns
42 * - if loop is present (detected via back-edge)
43 * - unreachable insns exist (shouldn't be a forest. program = one function)
44 * - out of bounds or malformed jumps
45 * The second pass is all possible path descent from the 1st insn.
46 * Since it's analyzing all pathes through the program, the length of the
47 * analysis is limited to 64k insn, which may be hit even if total number of
48 * insn is less then 4K, but there are too many branches that change stack/regs.
49 * Number of 'branches to be analyzed' is limited to 1k
51 * On entry to each instruction, each register has a type, and the instruction
52 * changes the types of the registers depending on instruction semantics.
53 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
56 * All registers are 64-bit.
57 * R0 - return register
58 * R1-R5 argument passing registers
59 * R6-R9 callee saved registers
60 * R10 - frame pointer read-only
62 * At the start of BPF program the register R1 contains a pointer to bpf_context
63 * and has type PTR_TO_CTX.
65 * Verifier tracks arithmetic operations on pointers in case:
66 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
67 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
68 * 1st insn copies R10 (which has FRAME_PTR) type into R1
69 * and 2nd arithmetic instruction is pattern matched to recognize
70 * that it wants to construct a pointer to some element within stack.
71 * So after 2nd insn, the register R1 has type PTR_TO_STACK
72 * (and -20 constant is saved for further stack bounds checking).
73 * Meaning that this reg is a pointer to stack plus known immediate constant.
75 * Most of the time the registers have SCALAR_VALUE type, which
76 * means the register has some value, but it's not a valid pointer.
77 * (like pointer plus pointer becomes SCALAR_VALUE type)
79 * When verifier sees load or store instructions the type of base register
80 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK. These are three pointer
81 * types recognized by check_mem_access() function.
83 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
84 * and the range of [ptr, ptr + map's value_size) is accessible.
86 * registers used to pass values to function calls are checked against
87 * function argument constraints.
89 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
90 * It means that the register type passed to this function must be
91 * PTR_TO_STACK and it will be used inside the function as
92 * 'pointer to map element key'
94 * For example the argument constraints for bpf_map_lookup_elem():
95 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
96 * .arg1_type = ARG_CONST_MAP_PTR,
97 * .arg2_type = ARG_PTR_TO_MAP_KEY,
99 * ret_type says that this function returns 'pointer to map elem value or null'
100 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
101 * 2nd argument should be a pointer to stack, which will be used inside
102 * the helper function as a pointer to map element key.
104 * On the kernel side the helper function looks like:
105 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
107 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
108 * void *key = (void *) (unsigned long) r2;
111 * here kernel can access 'key' and 'map' pointers safely, knowing that
112 * [key, key + map->key_size) bytes are valid and were initialized on
113 * the stack of eBPF program.
116 * Corresponding eBPF program may look like:
117 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
118 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
119 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
120 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
121 * here verifier looks at prototype of map_lookup_elem() and sees:
122 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
123 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
125 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
126 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
127 * and were initialized prior to this call.
128 * If it's ok, then verifier allows this BPF_CALL insn and looks at
129 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
130 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
131 * returns ether pointer to map value or NULL.
133 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
134 * insn, the register holding that pointer in the true branch changes state to
135 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
136 * branch. See check_cond_jmp_op().
138 * After the call R0 is set to return type of the function and registers R1-R5
139 * are set to NOT_INIT to indicate that they are no longer readable.
142 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
143 struct bpf_verifier_stack_elem
{
144 /* verifer state is 'st'
145 * before processing instruction 'insn_idx'
146 * and after processing instruction 'prev_insn_idx'
148 struct bpf_verifier_state st
;
151 struct bpf_verifier_stack_elem
*next
;
154 #define BPF_COMPLEXITY_LIMIT_INSNS 131072
155 #define BPF_COMPLEXITY_LIMIT_STACK 1024
157 #define BPF_MAP_PTR_POISON ((void *)0xeB9F + POISON_POINTER_DELTA)
159 struct bpf_call_arg_meta
{
160 struct bpf_map
*map_ptr
;
167 static DEFINE_MUTEX(bpf_verifier_lock
);
169 /* log_level controls verbosity level of eBPF verifier.
170 * verbose() is used to dump the verification trace to the log, so the user
171 * can figure out what's wrong with the program
173 static __printf(2, 3) void verbose(struct bpf_verifier_env
*env
,
174 const char *fmt
, ...)
176 struct bpf_verifer_log
*log
= &env
->log
;
180 if (!log
->level
|| !log
->ubuf
|| bpf_verifier_log_full(log
))
184 n
= vscnprintf(log
->kbuf
, BPF_VERIFIER_TMP_LOG_SIZE
, fmt
, args
);
187 WARN_ONCE(n
>= BPF_VERIFIER_TMP_LOG_SIZE
- 1,
188 "verifier log line truncated - local buffer too short\n");
190 n
= min(log
->len_total
- log
->len_used
- 1, n
);
193 if (!copy_to_user(log
->ubuf
+ log
->len_used
, log
->kbuf
, n
+ 1))
199 static bool type_is_pkt_pointer(enum bpf_reg_type type
)
201 return type
== PTR_TO_PACKET
||
202 type
== PTR_TO_PACKET_META
;
205 /* string representation of 'enum bpf_reg_type' */
206 static const char * const reg_type_str
[] = {
208 [SCALAR_VALUE
] = "inv",
209 [PTR_TO_CTX
] = "ctx",
210 [CONST_PTR_TO_MAP
] = "map_ptr",
211 [PTR_TO_MAP_VALUE
] = "map_value",
212 [PTR_TO_MAP_VALUE_OR_NULL
] = "map_value_or_null",
213 [PTR_TO_STACK
] = "fp",
214 [PTR_TO_PACKET
] = "pkt",
215 [PTR_TO_PACKET_META
] = "pkt_meta",
216 [PTR_TO_PACKET_END
] = "pkt_end",
219 static void print_verifier_state(struct bpf_verifier_env
*env
,
220 struct bpf_verifier_state
*state
)
222 struct bpf_reg_state
*reg
;
226 for (i
= 0; i
< MAX_BPF_REG
; i
++) {
227 reg
= &state
->regs
[i
];
231 verbose(env
, " R%d=%s", i
, reg_type_str
[t
]);
232 if ((t
== SCALAR_VALUE
|| t
== PTR_TO_STACK
) &&
233 tnum_is_const(reg
->var_off
)) {
234 /* reg->off should be 0 for SCALAR_VALUE */
235 verbose(env
, "%lld", reg
->var_off
.value
+ reg
->off
);
237 verbose(env
, "(id=%d", reg
->id
);
238 if (t
!= SCALAR_VALUE
)
239 verbose(env
, ",off=%d", reg
->off
);
240 if (type_is_pkt_pointer(t
))
241 verbose(env
, ",r=%d", reg
->range
);
242 else if (t
== CONST_PTR_TO_MAP
||
243 t
== PTR_TO_MAP_VALUE
||
244 t
== PTR_TO_MAP_VALUE_OR_NULL
)
245 verbose(env
, ",ks=%d,vs=%d",
246 reg
->map_ptr
->key_size
,
247 reg
->map_ptr
->value_size
);
248 if (tnum_is_const(reg
->var_off
)) {
249 /* Typically an immediate SCALAR_VALUE, but
250 * could be a pointer whose offset is too big
253 verbose(env
, ",imm=%llx", reg
->var_off
.value
);
255 if (reg
->smin_value
!= reg
->umin_value
&&
256 reg
->smin_value
!= S64_MIN
)
257 verbose(env
, ",smin_value=%lld",
258 (long long)reg
->smin_value
);
259 if (reg
->smax_value
!= reg
->umax_value
&&
260 reg
->smax_value
!= S64_MAX
)
261 verbose(env
, ",smax_value=%lld",
262 (long long)reg
->smax_value
);
263 if (reg
->umin_value
!= 0)
264 verbose(env
, ",umin_value=%llu",
265 (unsigned long long)reg
->umin_value
);
266 if (reg
->umax_value
!= U64_MAX
)
267 verbose(env
, ",umax_value=%llu",
268 (unsigned long long)reg
->umax_value
);
269 if (!tnum_is_unknown(reg
->var_off
)) {
272 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
273 verbose(env
, ",var_off=%s", tn_buf
);
279 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
280 if (state
->stack
[i
].slot_type
[0] == STACK_SPILL
)
281 verbose(env
, " fp%d=%s",
282 -MAX_BPF_STACK
+ i
* BPF_REG_SIZE
,
283 reg_type_str
[state
->stack
[i
].spilled_ptr
.type
]);
288 static int copy_stack_state(struct bpf_verifier_state
*dst
,
289 const struct bpf_verifier_state
*src
)
293 if (WARN_ON_ONCE(dst
->allocated_stack
< src
->allocated_stack
)) {
294 /* internal bug, make state invalid to reject the program */
295 memset(dst
, 0, sizeof(*dst
));
298 memcpy(dst
->stack
, src
->stack
,
299 sizeof(*src
->stack
) * (src
->allocated_stack
/ BPF_REG_SIZE
));
303 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
304 * make it consume minimal amount of memory. check_stack_write() access from
305 * the program calls into realloc_verifier_state() to grow the stack size.
306 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
307 * which this function copies over. It points to previous bpf_verifier_state
308 * which is never reallocated
310 static int realloc_verifier_state(struct bpf_verifier_state
*state
, int size
,
313 u32 old_size
= state
->allocated_stack
;
314 struct bpf_stack_state
*new_stack
;
315 int slot
= size
/ BPF_REG_SIZE
;
317 if (size
<= old_size
|| !size
) {
320 state
->allocated_stack
= slot
* BPF_REG_SIZE
;
321 if (!size
&& old_size
) {
327 new_stack
= kmalloc_array(slot
, sizeof(struct bpf_stack_state
),
333 memcpy(new_stack
, state
->stack
,
334 sizeof(*new_stack
) * (old_size
/ BPF_REG_SIZE
));
335 memset(new_stack
+ old_size
/ BPF_REG_SIZE
, 0,
336 sizeof(*new_stack
) * (size
- old_size
) / BPF_REG_SIZE
);
338 state
->allocated_stack
= slot
* BPF_REG_SIZE
;
340 state
->stack
= new_stack
;
344 static void free_verifier_state(struct bpf_verifier_state
*state
,
352 /* copy verifier state from src to dst growing dst stack space
353 * when necessary to accommodate larger src stack
355 static int copy_verifier_state(struct bpf_verifier_state
*dst
,
356 const struct bpf_verifier_state
*src
)
360 err
= realloc_verifier_state(dst
, src
->allocated_stack
, false);
363 memcpy(dst
, src
, offsetof(struct bpf_verifier_state
, allocated_stack
));
364 return copy_stack_state(dst
, src
);
367 static int pop_stack(struct bpf_verifier_env
*env
, int *prev_insn_idx
,
370 struct bpf_verifier_state
*cur
= env
->cur_state
;
371 struct bpf_verifier_stack_elem
*elem
, *head
= env
->head
;
374 if (env
->head
== NULL
)
378 err
= copy_verifier_state(cur
, &head
->st
);
383 *insn_idx
= head
->insn_idx
;
385 *prev_insn_idx
= head
->prev_insn_idx
;
387 free_verifier_state(&head
->st
, false);
394 static struct bpf_verifier_state
*push_stack(struct bpf_verifier_env
*env
,
395 int insn_idx
, int prev_insn_idx
)
397 struct bpf_verifier_state
*cur
= env
->cur_state
;
398 struct bpf_verifier_stack_elem
*elem
;
401 elem
= kzalloc(sizeof(struct bpf_verifier_stack_elem
), GFP_KERNEL
);
405 elem
->insn_idx
= insn_idx
;
406 elem
->prev_insn_idx
= prev_insn_idx
;
407 elem
->next
= env
->head
;
410 err
= copy_verifier_state(&elem
->st
, cur
);
413 if (env
->stack_size
> BPF_COMPLEXITY_LIMIT_STACK
) {
414 verbose(env
, "BPF program is too complex\n");
419 /* pop all elements and return */
420 while (!pop_stack(env
, NULL
, NULL
));
424 #define CALLER_SAVED_REGS 6
425 static const int caller_saved
[CALLER_SAVED_REGS
] = {
426 BPF_REG_0
, BPF_REG_1
, BPF_REG_2
, BPF_REG_3
, BPF_REG_4
, BPF_REG_5
429 static void __mark_reg_not_init(struct bpf_reg_state
*reg
);
431 /* Mark the unknown part of a register (variable offset or scalar value) as
432 * known to have the value @imm.
434 static void __mark_reg_known(struct bpf_reg_state
*reg
, u64 imm
)
437 reg
->var_off
= tnum_const(imm
);
438 reg
->smin_value
= (s64
)imm
;
439 reg
->smax_value
= (s64
)imm
;
440 reg
->umin_value
= imm
;
441 reg
->umax_value
= imm
;
444 /* Mark the 'variable offset' part of a register as zero. This should be
445 * used only on registers holding a pointer type.
447 static void __mark_reg_known_zero(struct bpf_reg_state
*reg
)
449 __mark_reg_known(reg
, 0);
452 static void mark_reg_known_zero(struct bpf_verifier_env
*env
,
453 struct bpf_reg_state
*regs
, u32 regno
)
455 if (WARN_ON(regno
>= MAX_BPF_REG
)) {
456 verbose(env
, "mark_reg_known_zero(regs, %u)\n", regno
);
457 /* Something bad happened, let's kill all regs */
458 for (regno
= 0; regno
< MAX_BPF_REG
; regno
++)
459 __mark_reg_not_init(regs
+ regno
);
462 __mark_reg_known_zero(regs
+ regno
);
465 static bool reg_is_pkt_pointer(const struct bpf_reg_state
*reg
)
467 return type_is_pkt_pointer(reg
->type
);
470 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state
*reg
)
472 return reg_is_pkt_pointer(reg
) ||
473 reg
->type
== PTR_TO_PACKET_END
;
476 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
477 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state
*reg
,
478 enum bpf_reg_type which
)
480 /* The register can already have a range from prior markings.
481 * This is fine as long as it hasn't been advanced from its
484 return reg
->type
== which
&&
487 tnum_equals_const(reg
->var_off
, 0);
490 /* Attempts to improve min/max values based on var_off information */
491 static void __update_reg_bounds(struct bpf_reg_state
*reg
)
493 /* min signed is max(sign bit) | min(other bits) */
494 reg
->smin_value
= max_t(s64
, reg
->smin_value
,
495 reg
->var_off
.value
| (reg
->var_off
.mask
& S64_MIN
));
496 /* max signed is min(sign bit) | max(other bits) */
497 reg
->smax_value
= min_t(s64
, reg
->smax_value
,
498 reg
->var_off
.value
| (reg
->var_off
.mask
& S64_MAX
));
499 reg
->umin_value
= max(reg
->umin_value
, reg
->var_off
.value
);
500 reg
->umax_value
= min(reg
->umax_value
,
501 reg
->var_off
.value
| reg
->var_off
.mask
);
504 /* Uses signed min/max values to inform unsigned, and vice-versa */
505 static void __reg_deduce_bounds(struct bpf_reg_state
*reg
)
507 /* Learn sign from signed bounds.
508 * If we cannot cross the sign boundary, then signed and unsigned bounds
509 * are the same, so combine. This works even in the negative case, e.g.
510 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
512 if (reg
->smin_value
>= 0 || reg
->smax_value
< 0) {
513 reg
->smin_value
= reg
->umin_value
= max_t(u64
, reg
->smin_value
,
515 reg
->smax_value
= reg
->umax_value
= min_t(u64
, reg
->smax_value
,
519 /* Learn sign from unsigned bounds. Signed bounds cross the sign
520 * boundary, so we must be careful.
522 if ((s64
)reg
->umax_value
>= 0) {
523 /* Positive. We can't learn anything from the smin, but smax
524 * is positive, hence safe.
526 reg
->smin_value
= reg
->umin_value
;
527 reg
->smax_value
= reg
->umax_value
= min_t(u64
, reg
->smax_value
,
529 } else if ((s64
)reg
->umin_value
< 0) {
530 /* Negative. We can't learn anything from the smax, but smin
531 * is negative, hence safe.
533 reg
->smin_value
= reg
->umin_value
= max_t(u64
, reg
->smin_value
,
535 reg
->smax_value
= reg
->umax_value
;
539 /* Attempts to improve var_off based on unsigned min/max information */
540 static void __reg_bound_offset(struct bpf_reg_state
*reg
)
542 reg
->var_off
= tnum_intersect(reg
->var_off
,
543 tnum_range(reg
->umin_value
,
547 /* Reset the min/max bounds of a register */
548 static void __mark_reg_unbounded(struct bpf_reg_state
*reg
)
550 reg
->smin_value
= S64_MIN
;
551 reg
->smax_value
= S64_MAX
;
553 reg
->umax_value
= U64_MAX
;
556 /* Mark a register as having a completely unknown (scalar) value. */
557 static void __mark_reg_unknown(struct bpf_reg_state
*reg
)
559 reg
->type
= SCALAR_VALUE
;
562 reg
->var_off
= tnum_unknown
;
563 __mark_reg_unbounded(reg
);
566 static void mark_reg_unknown(struct bpf_verifier_env
*env
,
567 struct bpf_reg_state
*regs
, u32 regno
)
569 if (WARN_ON(regno
>= MAX_BPF_REG
)) {
570 verbose(env
, "mark_reg_unknown(regs, %u)\n", regno
);
571 /* Something bad happened, let's kill all regs */
572 for (regno
= 0; regno
< MAX_BPF_REG
; regno
++)
573 __mark_reg_not_init(regs
+ regno
);
576 __mark_reg_unknown(regs
+ regno
);
579 static void __mark_reg_not_init(struct bpf_reg_state
*reg
)
581 __mark_reg_unknown(reg
);
582 reg
->type
= NOT_INIT
;
585 static void mark_reg_not_init(struct bpf_verifier_env
*env
,
586 struct bpf_reg_state
*regs
, u32 regno
)
588 if (WARN_ON(regno
>= MAX_BPF_REG
)) {
589 verbose(env
, "mark_reg_not_init(regs, %u)\n", regno
);
590 /* Something bad happened, let's kill all regs */
591 for (regno
= 0; regno
< MAX_BPF_REG
; regno
++)
592 __mark_reg_not_init(regs
+ regno
);
595 __mark_reg_not_init(regs
+ regno
);
598 static void init_reg_state(struct bpf_verifier_env
*env
,
599 struct bpf_reg_state
*regs
)
603 for (i
= 0; i
< MAX_BPF_REG
; i
++) {
604 mark_reg_not_init(env
, regs
, i
);
605 regs
[i
].live
= REG_LIVE_NONE
;
609 regs
[BPF_REG_FP
].type
= PTR_TO_STACK
;
610 mark_reg_known_zero(env
, regs
, BPF_REG_FP
);
612 /* 1st arg to a function */
613 regs
[BPF_REG_1
].type
= PTR_TO_CTX
;
614 mark_reg_known_zero(env
, regs
, BPF_REG_1
);
618 SRC_OP
, /* register is used as source operand */
619 DST_OP
, /* register is used as destination operand */
620 DST_OP_NO_MARK
/* same as above, check only, don't mark */
623 static void mark_reg_read(const struct bpf_verifier_state
*state
, u32 regno
)
625 struct bpf_verifier_state
*parent
= state
->parent
;
627 if (regno
== BPF_REG_FP
)
628 /* We don't need to worry about FP liveness because it's read-only */
632 /* if read wasn't screened by an earlier write ... */
633 if (state
->regs
[regno
].live
& REG_LIVE_WRITTEN
)
635 /* ... then we depend on parent's value */
636 parent
->regs
[regno
].live
|= REG_LIVE_READ
;
638 parent
= state
->parent
;
642 static int check_reg_arg(struct bpf_verifier_env
*env
, u32 regno
,
645 struct bpf_reg_state
*regs
= env
->cur_state
->regs
;
647 if (regno
>= MAX_BPF_REG
) {
648 verbose(env
, "R%d is invalid\n", regno
);
653 /* check whether register used as source operand can be read */
654 if (regs
[regno
].type
== NOT_INIT
) {
655 verbose(env
, "R%d !read_ok\n", regno
);
658 mark_reg_read(env
->cur_state
, regno
);
660 /* check whether register used as dest operand can be written to */
661 if (regno
== BPF_REG_FP
) {
662 verbose(env
, "frame pointer is read only\n");
665 regs
[regno
].live
|= REG_LIVE_WRITTEN
;
667 mark_reg_unknown(env
, regs
, regno
);
672 static bool is_spillable_regtype(enum bpf_reg_type type
)
675 case PTR_TO_MAP_VALUE
:
676 case PTR_TO_MAP_VALUE_OR_NULL
:
680 case PTR_TO_PACKET_META
:
681 case PTR_TO_PACKET_END
:
682 case CONST_PTR_TO_MAP
:
689 /* check_stack_read/write functions track spill/fill of registers,
690 * stack boundary and alignment are checked in check_mem_access()
692 static int check_stack_write(struct bpf_verifier_env
*env
,
693 struct bpf_verifier_state
*state
, int off
,
694 int size
, int value_regno
)
696 int i
, slot
= -off
- 1, spi
= slot
/ BPF_REG_SIZE
, err
;
698 err
= realloc_verifier_state(state
, round_up(slot
+ 1, BPF_REG_SIZE
),
702 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
703 * so it's aligned access and [off, off + size) are within stack limits
705 if (!env
->allow_ptr_leaks
&&
706 state
->stack
[spi
].slot_type
[0] == STACK_SPILL
&&
707 size
!= BPF_REG_SIZE
) {
708 verbose(env
, "attempt to corrupt spilled pointer on stack\n");
712 if (value_regno
>= 0 &&
713 is_spillable_regtype(state
->regs
[value_regno
].type
)) {
715 /* register containing pointer is being spilled into stack */
716 if (size
!= BPF_REG_SIZE
) {
717 verbose(env
, "invalid size of register spill\n");
721 /* save register state */
722 state
->stack
[spi
].spilled_ptr
= state
->regs
[value_regno
];
723 state
->stack
[spi
].spilled_ptr
.live
|= REG_LIVE_WRITTEN
;
725 for (i
= 0; i
< BPF_REG_SIZE
; i
++)
726 state
->stack
[spi
].slot_type
[i
] = STACK_SPILL
;
728 /* regular write of data into stack */
729 state
->stack
[spi
].spilled_ptr
= (struct bpf_reg_state
) {};
731 for (i
= 0; i
< size
; i
++)
732 state
->stack
[spi
].slot_type
[(slot
- i
) % BPF_REG_SIZE
] =
738 static void mark_stack_slot_read(const struct bpf_verifier_state
*state
, int slot
)
740 struct bpf_verifier_state
*parent
= state
->parent
;
743 /* if read wasn't screened by an earlier write ... */
744 if (state
->stack
[slot
].spilled_ptr
.live
& REG_LIVE_WRITTEN
)
746 /* ... then we depend on parent's value */
747 parent
->stack
[slot
].spilled_ptr
.live
|= REG_LIVE_READ
;
749 parent
= state
->parent
;
753 static int check_stack_read(struct bpf_verifier_env
*env
,
754 struct bpf_verifier_state
*state
, int off
, int size
,
757 int i
, slot
= -off
- 1, spi
= slot
/ BPF_REG_SIZE
;
760 if (state
->allocated_stack
<= slot
) {
761 verbose(env
, "invalid read from stack off %d+0 size %d\n",
765 stype
= state
->stack
[spi
].slot_type
;
767 if (stype
[0] == STACK_SPILL
) {
768 if (size
!= BPF_REG_SIZE
) {
769 verbose(env
, "invalid size of register spill\n");
772 for (i
= 1; i
< BPF_REG_SIZE
; i
++) {
773 if (stype
[(slot
- i
) % BPF_REG_SIZE
] != STACK_SPILL
) {
774 verbose(env
, "corrupted spill memory\n");
779 if (value_regno
>= 0) {
780 /* restore register state from stack */
781 state
->regs
[value_regno
] = state
->stack
[spi
].spilled_ptr
;
782 mark_stack_slot_read(state
, spi
);
786 for (i
= 0; i
< size
; i
++) {
787 if (stype
[(slot
- i
) % BPF_REG_SIZE
] != STACK_MISC
) {
788 verbose(env
, "invalid read from stack off %d+%d size %d\n",
793 if (value_regno
>= 0)
794 /* have read misc data from the stack */
795 mark_reg_unknown(env
, state
->regs
, value_regno
);
800 /* check read/write into map element returned by bpf_map_lookup_elem() */
801 static int __check_map_access(struct bpf_verifier_env
*env
, u32 regno
, int off
,
802 int size
, bool zero_size_allowed
)
804 struct bpf_reg_state
*regs
= cur_regs(env
);
805 struct bpf_map
*map
= regs
[regno
].map_ptr
;
807 if (off
< 0 || size
< 0 || (size
== 0 && !zero_size_allowed
) ||
808 off
+ size
> map
->value_size
) {
809 verbose(env
, "invalid access to map value, value_size=%d off=%d size=%d\n",
810 map
->value_size
, off
, size
);
816 /* check read/write into a map element with possible variable offset */
817 static int check_map_access(struct bpf_verifier_env
*env
, u32 regno
,
818 int off
, int size
, bool zero_size_allowed
)
820 struct bpf_verifier_state
*state
= env
->cur_state
;
821 struct bpf_reg_state
*reg
= &state
->regs
[regno
];
824 /* We may have adjusted the register to this map value, so we
825 * need to try adding each of min_value and max_value to off
826 * to make sure our theoretical access will be safe.
829 print_verifier_state(env
, state
);
830 /* The minimum value is only important with signed
831 * comparisons where we can't assume the floor of a
832 * value is 0. If we are using signed variables for our
833 * index'es we need to make sure that whatever we use
834 * will have a set floor within our range.
836 if (reg
->smin_value
< 0) {
837 verbose(env
, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
841 err
= __check_map_access(env
, regno
, reg
->smin_value
+ off
, size
,
844 verbose(env
, "R%d min value is outside of the array range\n",
849 /* If we haven't set a max value then we need to bail since we can't be
850 * sure we won't do bad things.
851 * If reg->umax_value + off could overflow, treat that as unbounded too.
853 if (reg
->umax_value
>= BPF_MAX_VAR_OFF
) {
854 verbose(env
, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
858 err
= __check_map_access(env
, regno
, reg
->umax_value
+ off
, size
,
861 verbose(env
, "R%d max value is outside of the array range\n",
866 #define MAX_PACKET_OFF 0xffff
868 static bool may_access_direct_pkt_data(struct bpf_verifier_env
*env
,
869 const struct bpf_call_arg_meta
*meta
,
870 enum bpf_access_type t
)
872 switch (env
->prog
->type
) {
873 case BPF_PROG_TYPE_LWT_IN
:
874 case BPF_PROG_TYPE_LWT_OUT
:
875 /* dst_input() and dst_output() can't write for now */
879 case BPF_PROG_TYPE_SCHED_CLS
:
880 case BPF_PROG_TYPE_SCHED_ACT
:
881 case BPF_PROG_TYPE_XDP
:
882 case BPF_PROG_TYPE_LWT_XMIT
:
883 case BPF_PROG_TYPE_SK_SKB
:
885 return meta
->pkt_access
;
887 env
->seen_direct_write
= true;
894 static int __check_packet_access(struct bpf_verifier_env
*env
, u32 regno
,
895 int off
, int size
, bool zero_size_allowed
)
897 struct bpf_reg_state
*regs
= cur_regs(env
);
898 struct bpf_reg_state
*reg
= ®s
[regno
];
900 if (off
< 0 || size
< 0 || (size
== 0 && !zero_size_allowed
) ||
901 (u64
)off
+ size
> reg
->range
) {
902 verbose(env
, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
903 off
, size
, regno
, reg
->id
, reg
->off
, reg
->range
);
909 static int check_packet_access(struct bpf_verifier_env
*env
, u32 regno
, int off
,
910 int size
, bool zero_size_allowed
)
912 struct bpf_reg_state
*regs
= cur_regs(env
);
913 struct bpf_reg_state
*reg
= ®s
[regno
];
916 /* We may have added a variable offset to the packet pointer; but any
917 * reg->range we have comes after that. We are only checking the fixed
921 /* We don't allow negative numbers, because we aren't tracking enough
922 * detail to prove they're safe.
924 if (reg
->smin_value
< 0) {
925 verbose(env
, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
929 err
= __check_packet_access(env
, regno
, off
, size
, zero_size_allowed
);
931 verbose(env
, "R%d offset is outside of the packet\n", regno
);
937 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
938 static int check_ctx_access(struct bpf_verifier_env
*env
, int insn_idx
, int off
, int size
,
939 enum bpf_access_type t
, enum bpf_reg_type
*reg_type
)
941 struct bpf_insn_access_aux info
= {
942 .reg_type
= *reg_type
,
945 if (env
->ops
->is_valid_access
&&
946 env
->ops
->is_valid_access(off
, size
, t
, &info
)) {
947 /* A non zero info.ctx_field_size indicates that this field is a
948 * candidate for later verifier transformation to load the whole
949 * field and then apply a mask when accessed with a narrower
950 * access than actual ctx access size. A zero info.ctx_field_size
951 * will only allow for whole field access and rejects any other
952 * type of narrower access.
954 *reg_type
= info
.reg_type
;
956 env
->insn_aux_data
[insn_idx
].ctx_field_size
= info
.ctx_field_size
;
957 /* remember the offset of last byte accessed in ctx */
958 if (env
->prog
->aux
->max_ctx_offset
< off
+ size
)
959 env
->prog
->aux
->max_ctx_offset
= off
+ size
;
963 verbose(env
, "invalid bpf_context access off=%d size=%d\n", off
, size
);
967 static bool __is_pointer_value(bool allow_ptr_leaks
,
968 const struct bpf_reg_state
*reg
)
973 return reg
->type
!= SCALAR_VALUE
;
976 static bool is_pointer_value(struct bpf_verifier_env
*env
, int regno
)
978 return __is_pointer_value(env
->allow_ptr_leaks
, cur_regs(env
) + regno
);
981 static int check_pkt_ptr_alignment(struct bpf_verifier_env
*env
,
982 const struct bpf_reg_state
*reg
,
983 int off
, int size
, bool strict
)
988 /* Byte size accesses are always allowed. */
989 if (!strict
|| size
== 1)
992 /* For platforms that do not have a Kconfig enabling
993 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
994 * NET_IP_ALIGN is universally set to '2'. And on platforms
995 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
996 * to this code only in strict mode where we want to emulate
997 * the NET_IP_ALIGN==2 checking. Therefore use an
998 * unconditional IP align value of '2'.
1002 reg_off
= tnum_add(reg
->var_off
, tnum_const(ip_align
+ reg
->off
+ off
));
1003 if (!tnum_is_aligned(reg_off
, size
)) {
1006 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1008 "misaligned packet access off %d+%s+%d+%d size %d\n",
1009 ip_align
, tn_buf
, reg
->off
, off
, size
);
1016 static int check_generic_ptr_alignment(struct bpf_verifier_env
*env
,
1017 const struct bpf_reg_state
*reg
,
1018 const char *pointer_desc
,
1019 int off
, int size
, bool strict
)
1021 struct tnum reg_off
;
1023 /* Byte size accesses are always allowed. */
1024 if (!strict
|| size
== 1)
1027 reg_off
= tnum_add(reg
->var_off
, tnum_const(reg
->off
+ off
));
1028 if (!tnum_is_aligned(reg_off
, size
)) {
1031 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1032 verbose(env
, "misaligned %saccess off %s+%d+%d size %d\n",
1033 pointer_desc
, tn_buf
, reg
->off
, off
, size
);
1040 static int check_ptr_alignment(struct bpf_verifier_env
*env
,
1041 const struct bpf_reg_state
*reg
,
1044 bool strict
= env
->strict_alignment
;
1045 const char *pointer_desc
= "";
1047 switch (reg
->type
) {
1049 case PTR_TO_PACKET_META
:
1050 /* Special case, because of NET_IP_ALIGN. Given metadata sits
1051 * right in front, treat it the very same way.
1053 return check_pkt_ptr_alignment(env
, reg
, off
, size
, strict
);
1054 case PTR_TO_MAP_VALUE
:
1055 pointer_desc
= "value ";
1058 pointer_desc
= "context ";
1061 pointer_desc
= "stack ";
1066 return check_generic_ptr_alignment(env
, reg
, pointer_desc
, off
, size
,
1070 /* check whether memory at (regno + off) is accessible for t = (read | write)
1071 * if t==write, value_regno is a register which value is stored into memory
1072 * if t==read, value_regno is a register which will receive the value from memory
1073 * if t==write && value_regno==-1, some unknown value is stored into memory
1074 * if t==read && value_regno==-1, don't care what we read from memory
1076 static int check_mem_access(struct bpf_verifier_env
*env
, int insn_idx
, u32 regno
, int off
,
1077 int bpf_size
, enum bpf_access_type t
,
1080 struct bpf_verifier_state
*state
= env
->cur_state
;
1081 struct bpf_reg_state
*regs
= cur_regs(env
);
1082 struct bpf_reg_state
*reg
= regs
+ regno
;
1085 size
= bpf_size_to_bytes(bpf_size
);
1089 /* alignment checks will add in reg->off themselves */
1090 err
= check_ptr_alignment(env
, reg
, off
, size
);
1094 /* for access checks, reg->off is just part of off */
1097 if (reg
->type
== PTR_TO_MAP_VALUE
) {
1098 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
1099 is_pointer_value(env
, value_regno
)) {
1100 verbose(env
, "R%d leaks addr into map\n", value_regno
);
1104 err
= check_map_access(env
, regno
, off
, size
, false);
1105 if (!err
&& t
== BPF_READ
&& value_regno
>= 0)
1106 mark_reg_unknown(env
, regs
, value_regno
);
1108 } else if (reg
->type
== PTR_TO_CTX
) {
1109 enum bpf_reg_type reg_type
= SCALAR_VALUE
;
1111 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
1112 is_pointer_value(env
, value_regno
)) {
1113 verbose(env
, "R%d leaks addr into ctx\n", value_regno
);
1116 /* ctx accesses must be at a fixed offset, so that we can
1117 * determine what type of data were returned.
1121 "dereference of modified ctx ptr R%d off=%d+%d, ctx+const is allowed, ctx+const+const is not\n",
1122 regno
, reg
->off
, off
- reg
->off
);
1125 if (!tnum_is_const(reg
->var_off
) || reg
->var_off
.value
) {
1128 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1130 "variable ctx access var_off=%s off=%d size=%d",
1134 err
= check_ctx_access(env
, insn_idx
, off
, size
, t
, ®_type
);
1135 if (!err
&& t
== BPF_READ
&& value_regno
>= 0) {
1136 /* ctx access returns either a scalar, or a
1137 * PTR_TO_PACKET[_META,_END]. In the latter
1138 * case, we know the offset is zero.
1140 if (reg_type
== SCALAR_VALUE
)
1141 mark_reg_unknown(env
, regs
, value_regno
);
1143 mark_reg_known_zero(env
, regs
,
1145 regs
[value_regno
].id
= 0;
1146 regs
[value_regno
].off
= 0;
1147 regs
[value_regno
].range
= 0;
1148 regs
[value_regno
].type
= reg_type
;
1151 } else if (reg
->type
== PTR_TO_STACK
) {
1152 /* stack accesses must be at a fixed offset, so that we can
1153 * determine what type of data were returned.
1154 * See check_stack_read().
1156 if (!tnum_is_const(reg
->var_off
)) {
1159 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1160 verbose(env
, "variable stack access var_off=%s off=%d size=%d",
1164 off
+= reg
->var_off
.value
;
1165 if (off
>= 0 || off
< -MAX_BPF_STACK
) {
1166 verbose(env
, "invalid stack off=%d size=%d\n", off
,
1171 if (env
->prog
->aux
->stack_depth
< -off
)
1172 env
->prog
->aux
->stack_depth
= -off
;
1175 err
= check_stack_write(env
, state
, off
, size
,
1178 err
= check_stack_read(env
, state
, off
, size
,
1180 } else if (reg_is_pkt_pointer(reg
)) {
1181 if (t
== BPF_WRITE
&& !may_access_direct_pkt_data(env
, NULL
, t
)) {
1182 verbose(env
, "cannot write into packet\n");
1185 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
1186 is_pointer_value(env
, value_regno
)) {
1187 verbose(env
, "R%d leaks addr into packet\n",
1191 err
= check_packet_access(env
, regno
, off
, size
, false);
1192 if (!err
&& t
== BPF_READ
&& value_regno
>= 0)
1193 mark_reg_unknown(env
, regs
, value_regno
);
1195 verbose(env
, "R%d invalid mem access '%s'\n", regno
,
1196 reg_type_str
[reg
->type
]);
1200 if (!err
&& size
< BPF_REG_SIZE
&& value_regno
>= 0 && t
== BPF_READ
&&
1201 regs
[value_regno
].type
== SCALAR_VALUE
) {
1202 /* b/h/w load zero-extends, mark upper bits as known 0 */
1203 regs
[value_regno
].var_off
=
1204 tnum_cast(regs
[value_regno
].var_off
, size
);
1205 __update_reg_bounds(®s
[value_regno
]);
1210 static int check_xadd(struct bpf_verifier_env
*env
, int insn_idx
, struct bpf_insn
*insn
)
1214 if ((BPF_SIZE(insn
->code
) != BPF_W
&& BPF_SIZE(insn
->code
) != BPF_DW
) ||
1216 verbose(env
, "BPF_XADD uses reserved fields\n");
1220 /* check src1 operand */
1221 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
1225 /* check src2 operand */
1226 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
1230 if (is_pointer_value(env
, insn
->src_reg
)) {
1231 verbose(env
, "R%d leaks addr into mem\n", insn
->src_reg
);
1235 /* check whether atomic_add can read the memory */
1236 err
= check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
1237 BPF_SIZE(insn
->code
), BPF_READ
, -1);
1241 /* check whether atomic_add can write into the same memory */
1242 return check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
1243 BPF_SIZE(insn
->code
), BPF_WRITE
, -1);
1246 /* Does this register contain a constant zero? */
1247 static bool register_is_null(struct bpf_reg_state reg
)
1249 return reg
.type
== SCALAR_VALUE
&& tnum_equals_const(reg
.var_off
, 0);
1252 /* when register 'regno' is passed into function that will read 'access_size'
1253 * bytes from that pointer, make sure that it's within stack boundary
1254 * and all elements of stack are initialized.
1255 * Unlike most pointer bounds-checking functions, this one doesn't take an
1256 * 'off' argument, so it has to add in reg->off itself.
1258 static int check_stack_boundary(struct bpf_verifier_env
*env
, int regno
,
1259 int access_size
, bool zero_size_allowed
,
1260 struct bpf_call_arg_meta
*meta
)
1262 struct bpf_verifier_state
*state
= env
->cur_state
;
1263 struct bpf_reg_state
*regs
= state
->regs
;
1264 int off
, i
, slot
, spi
;
1266 if (regs
[regno
].type
!= PTR_TO_STACK
) {
1267 /* Allow zero-byte read from NULL, regardless of pointer type */
1268 if (zero_size_allowed
&& access_size
== 0 &&
1269 register_is_null(regs
[regno
]))
1272 verbose(env
, "R%d type=%s expected=%s\n", regno
,
1273 reg_type_str
[regs
[regno
].type
],
1274 reg_type_str
[PTR_TO_STACK
]);
1278 /* Only allow fixed-offset stack reads */
1279 if (!tnum_is_const(regs
[regno
].var_off
)) {
1282 tnum_strn(tn_buf
, sizeof(tn_buf
), regs
[regno
].var_off
);
1283 verbose(env
, "invalid variable stack read R%d var_off=%s\n",
1286 off
= regs
[regno
].off
+ regs
[regno
].var_off
.value
;
1287 if (off
>= 0 || off
< -MAX_BPF_STACK
|| off
+ access_size
> 0 ||
1288 access_size
< 0 || (access_size
== 0 && !zero_size_allowed
)) {
1289 verbose(env
, "invalid stack type R%d off=%d access_size=%d\n",
1290 regno
, off
, access_size
);
1294 if (env
->prog
->aux
->stack_depth
< -off
)
1295 env
->prog
->aux
->stack_depth
= -off
;
1297 if (meta
&& meta
->raw_mode
) {
1298 meta
->access_size
= access_size
;
1299 meta
->regno
= regno
;
1303 for (i
= 0; i
< access_size
; i
++) {
1304 slot
= -(off
+ i
) - 1;
1305 spi
= slot
/ BPF_REG_SIZE
;
1306 if (state
->allocated_stack
<= slot
||
1307 state
->stack
[spi
].slot_type
[slot
% BPF_REG_SIZE
] !=
1309 verbose(env
, "invalid indirect read from stack off %d+%d size %d\n",
1310 off
, i
, access_size
);
1317 static int check_helper_mem_access(struct bpf_verifier_env
*env
, int regno
,
1318 int access_size
, bool zero_size_allowed
,
1319 struct bpf_call_arg_meta
*meta
)
1321 struct bpf_reg_state
*regs
= cur_regs(env
), *reg
= ®s
[regno
];
1323 switch (reg
->type
) {
1325 case PTR_TO_PACKET_META
:
1326 return check_packet_access(env
, regno
, reg
->off
, access_size
,
1328 case PTR_TO_MAP_VALUE
:
1329 return check_map_access(env
, regno
, reg
->off
, access_size
,
1331 default: /* scalar_value|ptr_to_stack or invalid ptr */
1332 return check_stack_boundary(env
, regno
, access_size
,
1333 zero_size_allowed
, meta
);
1337 static int check_func_arg(struct bpf_verifier_env
*env
, u32 regno
,
1338 enum bpf_arg_type arg_type
,
1339 struct bpf_call_arg_meta
*meta
)
1341 struct bpf_reg_state
*regs
= cur_regs(env
), *reg
= ®s
[regno
];
1342 enum bpf_reg_type expected_type
, type
= reg
->type
;
1345 if (arg_type
== ARG_DONTCARE
)
1348 err
= check_reg_arg(env
, regno
, SRC_OP
);
1352 if (arg_type
== ARG_ANYTHING
) {
1353 if (is_pointer_value(env
, regno
)) {
1354 verbose(env
, "R%d leaks addr into helper function\n",
1361 if (type_is_pkt_pointer(type
) &&
1362 !may_access_direct_pkt_data(env
, meta
, BPF_READ
)) {
1363 verbose(env
, "helper access to the packet is not allowed\n");
1367 if (arg_type
== ARG_PTR_TO_MAP_KEY
||
1368 arg_type
== ARG_PTR_TO_MAP_VALUE
) {
1369 expected_type
= PTR_TO_STACK
;
1370 if (!type_is_pkt_pointer(type
) &&
1371 type
!= expected_type
)
1373 } else if (arg_type
== ARG_CONST_SIZE
||
1374 arg_type
== ARG_CONST_SIZE_OR_ZERO
) {
1375 expected_type
= SCALAR_VALUE
;
1376 if (type
!= expected_type
)
1378 } else if (arg_type
== ARG_CONST_MAP_PTR
) {
1379 expected_type
= CONST_PTR_TO_MAP
;
1380 if (type
!= expected_type
)
1382 } else if (arg_type
== ARG_PTR_TO_CTX
) {
1383 expected_type
= PTR_TO_CTX
;
1384 if (type
!= expected_type
)
1386 } else if (arg_type
== ARG_PTR_TO_MEM
||
1387 arg_type
== ARG_PTR_TO_MEM_OR_NULL
||
1388 arg_type
== ARG_PTR_TO_UNINIT_MEM
) {
1389 expected_type
= PTR_TO_STACK
;
1390 /* One exception here. In case function allows for NULL to be
1391 * passed in as argument, it's a SCALAR_VALUE type. Final test
1392 * happens during stack boundary checking.
1394 if (register_is_null(*reg
) &&
1395 arg_type
== ARG_PTR_TO_MEM_OR_NULL
)
1396 /* final test in check_stack_boundary() */;
1397 else if (!type_is_pkt_pointer(type
) &&
1398 type
!= PTR_TO_MAP_VALUE
&&
1399 type
!= expected_type
)
1401 meta
->raw_mode
= arg_type
== ARG_PTR_TO_UNINIT_MEM
;
1403 verbose(env
, "unsupported arg_type %d\n", arg_type
);
1407 if (arg_type
== ARG_CONST_MAP_PTR
) {
1408 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
1409 meta
->map_ptr
= reg
->map_ptr
;
1410 } else if (arg_type
== ARG_PTR_TO_MAP_KEY
) {
1411 /* bpf_map_xxx(..., map_ptr, ..., key) call:
1412 * check that [key, key + map->key_size) are within
1413 * stack limits and initialized
1415 if (!meta
->map_ptr
) {
1416 /* in function declaration map_ptr must come before
1417 * map_key, so that it's verified and known before
1418 * we have to check map_key here. Otherwise it means
1419 * that kernel subsystem misconfigured verifier
1421 verbose(env
, "invalid map_ptr to access map->key\n");
1424 if (type_is_pkt_pointer(type
))
1425 err
= check_packet_access(env
, regno
, reg
->off
,
1426 meta
->map_ptr
->key_size
,
1429 err
= check_stack_boundary(env
, regno
,
1430 meta
->map_ptr
->key_size
,
1432 } else if (arg_type
== ARG_PTR_TO_MAP_VALUE
) {
1433 /* bpf_map_xxx(..., map_ptr, ..., value) call:
1434 * check [value, value + map->value_size) validity
1436 if (!meta
->map_ptr
) {
1437 /* kernel subsystem misconfigured verifier */
1438 verbose(env
, "invalid map_ptr to access map->value\n");
1441 if (type_is_pkt_pointer(type
))
1442 err
= check_packet_access(env
, regno
, reg
->off
,
1443 meta
->map_ptr
->value_size
,
1446 err
= check_stack_boundary(env
, regno
,
1447 meta
->map_ptr
->value_size
,
1449 } else if (arg_type
== ARG_CONST_SIZE
||
1450 arg_type
== ARG_CONST_SIZE_OR_ZERO
) {
1451 bool zero_size_allowed
= (arg_type
== ARG_CONST_SIZE_OR_ZERO
);
1453 /* bpf_xxx(..., buf, len) call will access 'len' bytes
1454 * from stack pointer 'buf'. Check it
1455 * note: regno == len, regno - 1 == buf
1458 /* kernel subsystem misconfigured verifier */
1460 "ARG_CONST_SIZE cannot be first argument\n");
1464 /* The register is SCALAR_VALUE; the access check
1465 * happens using its boundaries.
1468 if (!tnum_is_const(reg
->var_off
))
1469 /* For unprivileged variable accesses, disable raw
1470 * mode so that the program is required to
1471 * initialize all the memory that the helper could
1472 * just partially fill up.
1476 if (reg
->smin_value
< 0) {
1477 verbose(env
, "R%d min value is negative, either use unsigned or 'var &= const'\n",
1482 if (reg
->umin_value
== 0) {
1483 err
= check_helper_mem_access(env
, regno
- 1, 0,
1490 if (reg
->umax_value
>= BPF_MAX_VAR_SIZ
) {
1491 verbose(env
, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
1495 err
= check_helper_mem_access(env
, regno
- 1,
1497 zero_size_allowed
, meta
);
1502 verbose(env
, "R%d type=%s expected=%s\n", regno
,
1503 reg_type_str
[type
], reg_type_str
[expected_type
]);
1507 static int check_map_func_compatibility(struct bpf_verifier_env
*env
,
1508 struct bpf_map
*map
, int func_id
)
1513 /* We need a two way check, first is from map perspective ... */
1514 switch (map
->map_type
) {
1515 case BPF_MAP_TYPE_PROG_ARRAY
:
1516 if (func_id
!= BPF_FUNC_tail_call
)
1519 case BPF_MAP_TYPE_PERF_EVENT_ARRAY
:
1520 if (func_id
!= BPF_FUNC_perf_event_read
&&
1521 func_id
!= BPF_FUNC_perf_event_output
&&
1522 func_id
!= BPF_FUNC_perf_event_read_value
)
1525 case BPF_MAP_TYPE_STACK_TRACE
:
1526 if (func_id
!= BPF_FUNC_get_stackid
)
1529 case BPF_MAP_TYPE_CGROUP_ARRAY
:
1530 if (func_id
!= BPF_FUNC_skb_under_cgroup
&&
1531 func_id
!= BPF_FUNC_current_task_under_cgroup
)
1534 /* devmap returns a pointer to a live net_device ifindex that we cannot
1535 * allow to be modified from bpf side. So do not allow lookup elements
1538 case BPF_MAP_TYPE_DEVMAP
:
1539 if (func_id
!= BPF_FUNC_redirect_map
)
1542 /* Restrict bpf side of cpumap, open when use-cases appear */
1543 case BPF_MAP_TYPE_CPUMAP
:
1544 if (func_id
!= BPF_FUNC_redirect_map
)
1547 case BPF_MAP_TYPE_ARRAY_OF_MAPS
:
1548 case BPF_MAP_TYPE_HASH_OF_MAPS
:
1549 if (func_id
!= BPF_FUNC_map_lookup_elem
)
1552 case BPF_MAP_TYPE_SOCKMAP
:
1553 if (func_id
!= BPF_FUNC_sk_redirect_map
&&
1554 func_id
!= BPF_FUNC_sock_map_update
&&
1555 func_id
!= BPF_FUNC_map_delete_elem
)
1562 /* ... and second from the function itself. */
1564 case BPF_FUNC_tail_call
:
1565 if (map
->map_type
!= BPF_MAP_TYPE_PROG_ARRAY
)
1568 case BPF_FUNC_perf_event_read
:
1569 case BPF_FUNC_perf_event_output
:
1570 case BPF_FUNC_perf_event_read_value
:
1571 if (map
->map_type
!= BPF_MAP_TYPE_PERF_EVENT_ARRAY
)
1574 case BPF_FUNC_get_stackid
:
1575 if (map
->map_type
!= BPF_MAP_TYPE_STACK_TRACE
)
1578 case BPF_FUNC_current_task_under_cgroup
:
1579 case BPF_FUNC_skb_under_cgroup
:
1580 if (map
->map_type
!= BPF_MAP_TYPE_CGROUP_ARRAY
)
1583 case BPF_FUNC_redirect_map
:
1584 if (map
->map_type
!= BPF_MAP_TYPE_DEVMAP
&&
1585 map
->map_type
!= BPF_MAP_TYPE_CPUMAP
)
1588 case BPF_FUNC_sk_redirect_map
:
1589 if (map
->map_type
!= BPF_MAP_TYPE_SOCKMAP
)
1592 case BPF_FUNC_sock_map_update
:
1593 if (map
->map_type
!= BPF_MAP_TYPE_SOCKMAP
)
1602 verbose(env
, "cannot pass map_type %d into func %s#%d\n",
1603 map
->map_type
, func_id_name(func_id
), func_id
);
1607 static int check_raw_mode(const struct bpf_func_proto
*fn
)
1611 if (fn
->arg1_type
== ARG_PTR_TO_UNINIT_MEM
)
1613 if (fn
->arg2_type
== ARG_PTR_TO_UNINIT_MEM
)
1615 if (fn
->arg3_type
== ARG_PTR_TO_UNINIT_MEM
)
1617 if (fn
->arg4_type
== ARG_PTR_TO_UNINIT_MEM
)
1619 if (fn
->arg5_type
== ARG_PTR_TO_UNINIT_MEM
)
1622 return count
> 1 ? -EINVAL
: 0;
1625 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
1626 * are now invalid, so turn them into unknown SCALAR_VALUE.
1628 static void clear_all_pkt_pointers(struct bpf_verifier_env
*env
)
1630 struct bpf_verifier_state
*state
= env
->cur_state
;
1631 struct bpf_reg_state
*regs
= state
->regs
, *reg
;
1634 for (i
= 0; i
< MAX_BPF_REG
; i
++)
1635 if (reg_is_pkt_pointer_any(®s
[i
]))
1636 mark_reg_unknown(env
, regs
, i
);
1638 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
1639 if (state
->stack
[i
].slot_type
[0] != STACK_SPILL
)
1641 reg
= &state
->stack
[i
].spilled_ptr
;
1642 if (reg_is_pkt_pointer_any(reg
))
1643 __mark_reg_unknown(reg
);
1647 static int check_call(struct bpf_verifier_env
*env
, int func_id
, int insn_idx
)
1649 const struct bpf_func_proto
*fn
= NULL
;
1650 struct bpf_reg_state
*regs
;
1651 struct bpf_call_arg_meta meta
;
1655 /* find function prototype */
1656 if (func_id
< 0 || func_id
>= __BPF_FUNC_MAX_ID
) {
1657 verbose(env
, "invalid func %s#%d\n", func_id_name(func_id
),
1662 if (env
->ops
->get_func_proto
)
1663 fn
= env
->ops
->get_func_proto(func_id
);
1666 verbose(env
, "unknown func %s#%d\n", func_id_name(func_id
),
1671 /* eBPF programs must be GPL compatible to use GPL-ed functions */
1672 if (!env
->prog
->gpl_compatible
&& fn
->gpl_only
) {
1673 verbose(env
, "cannot call GPL only function from proprietary program\n");
1677 changes_data
= bpf_helper_changes_pkt_data(fn
->func
);
1679 memset(&meta
, 0, sizeof(meta
));
1680 meta
.pkt_access
= fn
->pkt_access
;
1682 /* We only support one arg being in raw mode at the moment, which
1683 * is sufficient for the helper functions we have right now.
1685 err
= check_raw_mode(fn
);
1687 verbose(env
, "kernel subsystem misconfigured func %s#%d\n",
1688 func_id_name(func_id
), func_id
);
1693 err
= check_func_arg(env
, BPF_REG_1
, fn
->arg1_type
, &meta
);
1696 err
= check_func_arg(env
, BPF_REG_2
, fn
->arg2_type
, &meta
);
1699 err
= check_func_arg(env
, BPF_REG_3
, fn
->arg3_type
, &meta
);
1702 err
= check_func_arg(env
, BPF_REG_4
, fn
->arg4_type
, &meta
);
1705 err
= check_func_arg(env
, BPF_REG_5
, fn
->arg5_type
, &meta
);
1709 /* Mark slots with STACK_MISC in case of raw mode, stack offset
1710 * is inferred from register state.
1712 for (i
= 0; i
< meta
.access_size
; i
++) {
1713 err
= check_mem_access(env
, insn_idx
, meta
.regno
, i
, BPF_B
, BPF_WRITE
, -1);
1718 regs
= cur_regs(env
);
1719 /* reset caller saved regs */
1720 for (i
= 0; i
< CALLER_SAVED_REGS
; i
++) {
1721 mark_reg_not_init(env
, regs
, caller_saved
[i
]);
1722 check_reg_arg(env
, caller_saved
[i
], DST_OP_NO_MARK
);
1725 /* update return register (already marked as written above) */
1726 if (fn
->ret_type
== RET_INTEGER
) {
1727 /* sets type to SCALAR_VALUE */
1728 mark_reg_unknown(env
, regs
, BPF_REG_0
);
1729 } else if (fn
->ret_type
== RET_VOID
) {
1730 regs
[BPF_REG_0
].type
= NOT_INIT
;
1731 } else if (fn
->ret_type
== RET_PTR_TO_MAP_VALUE_OR_NULL
) {
1732 struct bpf_insn_aux_data
*insn_aux
;
1734 regs
[BPF_REG_0
].type
= PTR_TO_MAP_VALUE_OR_NULL
;
1735 /* There is no offset yet applied, variable or fixed */
1736 mark_reg_known_zero(env
, regs
, BPF_REG_0
);
1737 regs
[BPF_REG_0
].off
= 0;
1738 /* remember map_ptr, so that check_map_access()
1739 * can check 'value_size' boundary of memory access
1740 * to map element returned from bpf_map_lookup_elem()
1742 if (meta
.map_ptr
== NULL
) {
1744 "kernel subsystem misconfigured verifier\n");
1747 regs
[BPF_REG_0
].map_ptr
= meta
.map_ptr
;
1748 regs
[BPF_REG_0
].id
= ++env
->id_gen
;
1749 insn_aux
= &env
->insn_aux_data
[insn_idx
];
1750 if (!insn_aux
->map_ptr
)
1751 insn_aux
->map_ptr
= meta
.map_ptr
;
1752 else if (insn_aux
->map_ptr
!= meta
.map_ptr
)
1753 insn_aux
->map_ptr
= BPF_MAP_PTR_POISON
;
1755 verbose(env
, "unknown return type %d of func %s#%d\n",
1756 fn
->ret_type
, func_id_name(func_id
), func_id
);
1760 err
= check_map_func_compatibility(env
, meta
.map_ptr
, func_id
);
1765 clear_all_pkt_pointers(env
);
1769 static void coerce_reg_to_32(struct bpf_reg_state
*reg
)
1771 /* clear high 32 bits */
1772 reg
->var_off
= tnum_cast(reg
->var_off
, 4);
1774 __update_reg_bounds(reg
);
1777 static bool signed_add_overflows(s64 a
, s64 b
)
1779 /* Do the add in u64, where overflow is well-defined */
1780 s64 res
= (s64
)((u64
)a
+ (u64
)b
);
1787 static bool signed_sub_overflows(s64 a
, s64 b
)
1789 /* Do the sub in u64, where overflow is well-defined */
1790 s64 res
= (s64
)((u64
)a
- (u64
)b
);
1797 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
1798 * Caller should also handle BPF_MOV case separately.
1799 * If we return -EACCES, caller may want to try again treating pointer as a
1800 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
1802 static int adjust_ptr_min_max_vals(struct bpf_verifier_env
*env
,
1803 struct bpf_insn
*insn
,
1804 const struct bpf_reg_state
*ptr_reg
,
1805 const struct bpf_reg_state
*off_reg
)
1807 struct bpf_reg_state
*regs
= cur_regs(env
), *dst_reg
;
1808 bool known
= tnum_is_const(off_reg
->var_off
);
1809 s64 smin_val
= off_reg
->smin_value
, smax_val
= off_reg
->smax_value
,
1810 smin_ptr
= ptr_reg
->smin_value
, smax_ptr
= ptr_reg
->smax_value
;
1811 u64 umin_val
= off_reg
->umin_value
, umax_val
= off_reg
->umax_value
,
1812 umin_ptr
= ptr_reg
->umin_value
, umax_ptr
= ptr_reg
->umax_value
;
1813 u8 opcode
= BPF_OP(insn
->code
);
1814 u32 dst
= insn
->dst_reg
;
1816 dst_reg
= ®s
[dst
];
1818 if (WARN_ON_ONCE(known
&& (smin_val
!= smax_val
))) {
1819 print_verifier_state(env
, env
->cur_state
);
1821 "verifier internal error: known but bad sbounds\n");
1824 if (WARN_ON_ONCE(known
&& (umin_val
!= umax_val
))) {
1825 print_verifier_state(env
, env
->cur_state
);
1827 "verifier internal error: known but bad ubounds\n");
1831 if (BPF_CLASS(insn
->code
) != BPF_ALU64
) {
1832 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
1833 if (!env
->allow_ptr_leaks
)
1835 "R%d 32-bit pointer arithmetic prohibited\n",
1840 if (ptr_reg
->type
== PTR_TO_MAP_VALUE_OR_NULL
) {
1841 if (!env
->allow_ptr_leaks
)
1842 verbose(env
, "R%d pointer arithmetic on PTR_TO_MAP_VALUE_OR_NULL prohibited, null-check it first\n",
1846 if (ptr_reg
->type
== CONST_PTR_TO_MAP
) {
1847 if (!env
->allow_ptr_leaks
)
1848 verbose(env
, "R%d pointer arithmetic on CONST_PTR_TO_MAP prohibited\n",
1852 if (ptr_reg
->type
== PTR_TO_PACKET_END
) {
1853 if (!env
->allow_ptr_leaks
)
1854 verbose(env
, "R%d pointer arithmetic on PTR_TO_PACKET_END prohibited\n",
1859 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
1860 * The id may be overwritten later if we create a new variable offset.
1862 dst_reg
->type
= ptr_reg
->type
;
1863 dst_reg
->id
= ptr_reg
->id
;
1867 /* We can take a fixed offset as long as it doesn't overflow
1868 * the s32 'off' field
1870 if (known
&& (ptr_reg
->off
+ smin_val
==
1871 (s64
)(s32
)(ptr_reg
->off
+ smin_val
))) {
1872 /* pointer += K. Accumulate it into fixed offset */
1873 dst_reg
->smin_value
= smin_ptr
;
1874 dst_reg
->smax_value
= smax_ptr
;
1875 dst_reg
->umin_value
= umin_ptr
;
1876 dst_reg
->umax_value
= umax_ptr
;
1877 dst_reg
->var_off
= ptr_reg
->var_off
;
1878 dst_reg
->off
= ptr_reg
->off
+ smin_val
;
1879 dst_reg
->range
= ptr_reg
->range
;
1882 /* A new variable offset is created. Note that off_reg->off
1883 * == 0, since it's a scalar.
1884 * dst_reg gets the pointer type and since some positive
1885 * integer value was added to the pointer, give it a new 'id'
1886 * if it's a PTR_TO_PACKET.
1887 * this creates a new 'base' pointer, off_reg (variable) gets
1888 * added into the variable offset, and we copy the fixed offset
1891 if (signed_add_overflows(smin_ptr
, smin_val
) ||
1892 signed_add_overflows(smax_ptr
, smax_val
)) {
1893 dst_reg
->smin_value
= S64_MIN
;
1894 dst_reg
->smax_value
= S64_MAX
;
1896 dst_reg
->smin_value
= smin_ptr
+ smin_val
;
1897 dst_reg
->smax_value
= smax_ptr
+ smax_val
;
1899 if (umin_ptr
+ umin_val
< umin_ptr
||
1900 umax_ptr
+ umax_val
< umax_ptr
) {
1901 dst_reg
->umin_value
= 0;
1902 dst_reg
->umax_value
= U64_MAX
;
1904 dst_reg
->umin_value
= umin_ptr
+ umin_val
;
1905 dst_reg
->umax_value
= umax_ptr
+ umax_val
;
1907 dst_reg
->var_off
= tnum_add(ptr_reg
->var_off
, off_reg
->var_off
);
1908 dst_reg
->off
= ptr_reg
->off
;
1909 if (reg_is_pkt_pointer(ptr_reg
)) {
1910 dst_reg
->id
= ++env
->id_gen
;
1911 /* something was added to pkt_ptr, set range to zero */
1916 if (dst_reg
== off_reg
) {
1917 /* scalar -= pointer. Creates an unknown scalar */
1918 if (!env
->allow_ptr_leaks
)
1919 verbose(env
, "R%d tried to subtract pointer from scalar\n",
1923 /* We don't allow subtraction from FP, because (according to
1924 * test_verifier.c test "invalid fp arithmetic", JITs might not
1925 * be able to deal with it.
1927 if (ptr_reg
->type
== PTR_TO_STACK
) {
1928 if (!env
->allow_ptr_leaks
)
1929 verbose(env
, "R%d subtraction from stack pointer prohibited\n",
1933 if (known
&& (ptr_reg
->off
- smin_val
==
1934 (s64
)(s32
)(ptr_reg
->off
- smin_val
))) {
1935 /* pointer -= K. Subtract it from fixed offset */
1936 dst_reg
->smin_value
= smin_ptr
;
1937 dst_reg
->smax_value
= smax_ptr
;
1938 dst_reg
->umin_value
= umin_ptr
;
1939 dst_reg
->umax_value
= umax_ptr
;
1940 dst_reg
->var_off
= ptr_reg
->var_off
;
1941 dst_reg
->id
= ptr_reg
->id
;
1942 dst_reg
->off
= ptr_reg
->off
- smin_val
;
1943 dst_reg
->range
= ptr_reg
->range
;
1946 /* A new variable offset is created. If the subtrahend is known
1947 * nonnegative, then any reg->range we had before is still good.
1949 if (signed_sub_overflows(smin_ptr
, smax_val
) ||
1950 signed_sub_overflows(smax_ptr
, smin_val
)) {
1951 /* Overflow possible, we know nothing */
1952 dst_reg
->smin_value
= S64_MIN
;
1953 dst_reg
->smax_value
= S64_MAX
;
1955 dst_reg
->smin_value
= smin_ptr
- smax_val
;
1956 dst_reg
->smax_value
= smax_ptr
- smin_val
;
1958 if (umin_ptr
< umax_val
) {
1959 /* Overflow possible, we know nothing */
1960 dst_reg
->umin_value
= 0;
1961 dst_reg
->umax_value
= U64_MAX
;
1963 /* Cannot overflow (as long as bounds are consistent) */
1964 dst_reg
->umin_value
= umin_ptr
- umax_val
;
1965 dst_reg
->umax_value
= umax_ptr
- umin_val
;
1967 dst_reg
->var_off
= tnum_sub(ptr_reg
->var_off
, off_reg
->var_off
);
1968 dst_reg
->off
= ptr_reg
->off
;
1969 if (reg_is_pkt_pointer(ptr_reg
)) {
1970 dst_reg
->id
= ++env
->id_gen
;
1971 /* something was added to pkt_ptr, set range to zero */
1979 /* bitwise ops on pointers are troublesome, prohibit for now.
1980 * (However, in principle we could allow some cases, e.g.
1981 * ptr &= ~3 which would reduce min_value by 3.)
1983 if (!env
->allow_ptr_leaks
)
1984 verbose(env
, "R%d bitwise operator %s on pointer prohibited\n",
1985 dst
, bpf_alu_string
[opcode
>> 4]);
1988 /* other operators (e.g. MUL,LSH) produce non-pointer results */
1989 if (!env
->allow_ptr_leaks
)
1990 verbose(env
, "R%d pointer arithmetic with %s operator prohibited\n",
1991 dst
, bpf_alu_string
[opcode
>> 4]);
1995 __update_reg_bounds(dst_reg
);
1996 __reg_deduce_bounds(dst_reg
);
1997 __reg_bound_offset(dst_reg
);
2001 static int adjust_scalar_min_max_vals(struct bpf_verifier_env
*env
,
2002 struct bpf_insn
*insn
,
2003 struct bpf_reg_state
*dst_reg
,
2004 struct bpf_reg_state src_reg
)
2006 struct bpf_reg_state
*regs
= cur_regs(env
);
2007 u8 opcode
= BPF_OP(insn
->code
);
2008 bool src_known
, dst_known
;
2009 s64 smin_val
, smax_val
;
2010 u64 umin_val
, umax_val
;
2012 if (BPF_CLASS(insn
->code
) != BPF_ALU64
) {
2013 /* 32-bit ALU ops are (32,32)->64 */
2014 coerce_reg_to_32(dst_reg
);
2015 coerce_reg_to_32(&src_reg
);
2017 smin_val
= src_reg
.smin_value
;
2018 smax_val
= src_reg
.smax_value
;
2019 umin_val
= src_reg
.umin_value
;
2020 umax_val
= src_reg
.umax_value
;
2021 src_known
= tnum_is_const(src_reg
.var_off
);
2022 dst_known
= tnum_is_const(dst_reg
->var_off
);
2026 if (signed_add_overflows(dst_reg
->smin_value
, smin_val
) ||
2027 signed_add_overflows(dst_reg
->smax_value
, smax_val
)) {
2028 dst_reg
->smin_value
= S64_MIN
;
2029 dst_reg
->smax_value
= S64_MAX
;
2031 dst_reg
->smin_value
+= smin_val
;
2032 dst_reg
->smax_value
+= smax_val
;
2034 if (dst_reg
->umin_value
+ umin_val
< umin_val
||
2035 dst_reg
->umax_value
+ umax_val
< umax_val
) {
2036 dst_reg
->umin_value
= 0;
2037 dst_reg
->umax_value
= U64_MAX
;
2039 dst_reg
->umin_value
+= umin_val
;
2040 dst_reg
->umax_value
+= umax_val
;
2042 dst_reg
->var_off
= tnum_add(dst_reg
->var_off
, src_reg
.var_off
);
2045 if (signed_sub_overflows(dst_reg
->smin_value
, smax_val
) ||
2046 signed_sub_overflows(dst_reg
->smax_value
, smin_val
)) {
2047 /* Overflow possible, we know nothing */
2048 dst_reg
->smin_value
= S64_MIN
;
2049 dst_reg
->smax_value
= S64_MAX
;
2051 dst_reg
->smin_value
-= smax_val
;
2052 dst_reg
->smax_value
-= smin_val
;
2054 if (dst_reg
->umin_value
< umax_val
) {
2055 /* Overflow possible, we know nothing */
2056 dst_reg
->umin_value
= 0;
2057 dst_reg
->umax_value
= U64_MAX
;
2059 /* Cannot overflow (as long as bounds are consistent) */
2060 dst_reg
->umin_value
-= umax_val
;
2061 dst_reg
->umax_value
-= umin_val
;
2063 dst_reg
->var_off
= tnum_sub(dst_reg
->var_off
, src_reg
.var_off
);
2066 dst_reg
->var_off
= tnum_mul(dst_reg
->var_off
, src_reg
.var_off
);
2067 if (smin_val
< 0 || dst_reg
->smin_value
< 0) {
2068 /* Ain't nobody got time to multiply that sign */
2069 __mark_reg_unbounded(dst_reg
);
2070 __update_reg_bounds(dst_reg
);
2073 /* Both values are positive, so we can work with unsigned and
2074 * copy the result to signed (unless it exceeds S64_MAX).
2076 if (umax_val
> U32_MAX
|| dst_reg
->umax_value
> U32_MAX
) {
2077 /* Potential overflow, we know nothing */
2078 __mark_reg_unbounded(dst_reg
);
2079 /* (except what we can learn from the var_off) */
2080 __update_reg_bounds(dst_reg
);
2083 dst_reg
->umin_value
*= umin_val
;
2084 dst_reg
->umax_value
*= umax_val
;
2085 if (dst_reg
->umax_value
> S64_MAX
) {
2086 /* Overflow possible, we know nothing */
2087 dst_reg
->smin_value
= S64_MIN
;
2088 dst_reg
->smax_value
= S64_MAX
;
2090 dst_reg
->smin_value
= dst_reg
->umin_value
;
2091 dst_reg
->smax_value
= dst_reg
->umax_value
;
2095 if (src_known
&& dst_known
) {
2096 __mark_reg_known(dst_reg
, dst_reg
->var_off
.value
&
2097 src_reg
.var_off
.value
);
2100 /* We get our minimum from the var_off, since that's inherently
2101 * bitwise. Our maximum is the minimum of the operands' maxima.
2103 dst_reg
->var_off
= tnum_and(dst_reg
->var_off
, src_reg
.var_off
);
2104 dst_reg
->umin_value
= dst_reg
->var_off
.value
;
2105 dst_reg
->umax_value
= min(dst_reg
->umax_value
, umax_val
);
2106 if (dst_reg
->smin_value
< 0 || smin_val
< 0) {
2107 /* Lose signed bounds when ANDing negative numbers,
2108 * ain't nobody got time for that.
2110 dst_reg
->smin_value
= S64_MIN
;
2111 dst_reg
->smax_value
= S64_MAX
;
2113 /* ANDing two positives gives a positive, so safe to
2114 * cast result into s64.
2116 dst_reg
->smin_value
= dst_reg
->umin_value
;
2117 dst_reg
->smax_value
= dst_reg
->umax_value
;
2119 /* We may learn something more from the var_off */
2120 __update_reg_bounds(dst_reg
);
2123 if (src_known
&& dst_known
) {
2124 __mark_reg_known(dst_reg
, dst_reg
->var_off
.value
|
2125 src_reg
.var_off
.value
);
2128 /* We get our maximum from the var_off, and our minimum is the
2129 * maximum of the operands' minima
2131 dst_reg
->var_off
= tnum_or(dst_reg
->var_off
, src_reg
.var_off
);
2132 dst_reg
->umin_value
= max(dst_reg
->umin_value
, umin_val
);
2133 dst_reg
->umax_value
= dst_reg
->var_off
.value
|
2134 dst_reg
->var_off
.mask
;
2135 if (dst_reg
->smin_value
< 0 || smin_val
< 0) {
2136 /* Lose signed bounds when ORing negative numbers,
2137 * ain't nobody got time for that.
2139 dst_reg
->smin_value
= S64_MIN
;
2140 dst_reg
->smax_value
= S64_MAX
;
2142 /* ORing two positives gives a positive, so safe to
2143 * cast result into s64.
2145 dst_reg
->smin_value
= dst_reg
->umin_value
;
2146 dst_reg
->smax_value
= dst_reg
->umax_value
;
2148 /* We may learn something more from the var_off */
2149 __update_reg_bounds(dst_reg
);
2152 if (umax_val
> 63) {
2153 /* Shifts greater than 63 are undefined. This includes
2154 * shifts by a negative number.
2156 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
2159 /* We lose all sign bit information (except what we can pick
2162 dst_reg
->smin_value
= S64_MIN
;
2163 dst_reg
->smax_value
= S64_MAX
;
2164 /* If we might shift our top bit out, then we know nothing */
2165 if (dst_reg
->umax_value
> 1ULL << (63 - umax_val
)) {
2166 dst_reg
->umin_value
= 0;
2167 dst_reg
->umax_value
= U64_MAX
;
2169 dst_reg
->umin_value
<<= umin_val
;
2170 dst_reg
->umax_value
<<= umax_val
;
2173 dst_reg
->var_off
= tnum_lshift(dst_reg
->var_off
, umin_val
);
2175 dst_reg
->var_off
= tnum_lshift(tnum_unknown
, umin_val
);
2176 /* We may learn something more from the var_off */
2177 __update_reg_bounds(dst_reg
);
2180 if (umax_val
> 63) {
2181 /* Shifts greater than 63 are undefined. This includes
2182 * shifts by a negative number.
2184 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
2187 /* BPF_RSH is an unsigned shift, so make the appropriate casts */
2188 if (dst_reg
->smin_value
< 0) {
2190 /* Sign bit will be cleared */
2191 dst_reg
->smin_value
= 0;
2193 /* Lost sign bit information */
2194 dst_reg
->smin_value
= S64_MIN
;
2195 dst_reg
->smax_value
= S64_MAX
;
2198 dst_reg
->smin_value
=
2199 (u64
)(dst_reg
->smin_value
) >> umax_val
;
2202 dst_reg
->var_off
= tnum_rshift(dst_reg
->var_off
,
2205 dst_reg
->var_off
= tnum_rshift(tnum_unknown
, umin_val
);
2206 dst_reg
->umin_value
>>= umax_val
;
2207 dst_reg
->umax_value
>>= umin_val
;
2208 /* We may learn something more from the var_off */
2209 __update_reg_bounds(dst_reg
);
2212 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
2216 __reg_deduce_bounds(dst_reg
);
2217 __reg_bound_offset(dst_reg
);
2221 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
2224 static int adjust_reg_min_max_vals(struct bpf_verifier_env
*env
,
2225 struct bpf_insn
*insn
)
2227 struct bpf_reg_state
*regs
= cur_regs(env
), *dst_reg
, *src_reg
;
2228 struct bpf_reg_state
*ptr_reg
= NULL
, off_reg
= {0};
2229 u8 opcode
= BPF_OP(insn
->code
);
2232 dst_reg
= ®s
[insn
->dst_reg
];
2234 if (dst_reg
->type
!= SCALAR_VALUE
)
2236 if (BPF_SRC(insn
->code
) == BPF_X
) {
2237 src_reg
= ®s
[insn
->src_reg
];
2238 if (src_reg
->type
!= SCALAR_VALUE
) {
2239 if (dst_reg
->type
!= SCALAR_VALUE
) {
2240 /* Combining two pointers by any ALU op yields
2241 * an arbitrary scalar.
2243 if (!env
->allow_ptr_leaks
) {
2244 verbose(env
, "R%d pointer %s pointer prohibited\n",
2246 bpf_alu_string
[opcode
>> 4]);
2249 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
2252 /* scalar += pointer
2253 * This is legal, but we have to reverse our
2254 * src/dest handling in computing the range
2256 rc
= adjust_ptr_min_max_vals(env
, insn
,
2258 if (rc
== -EACCES
&& env
->allow_ptr_leaks
) {
2259 /* scalar += unknown scalar */
2260 __mark_reg_unknown(&off_reg
);
2261 return adjust_scalar_min_max_vals(
2267 } else if (ptr_reg
) {
2268 /* pointer += scalar */
2269 rc
= adjust_ptr_min_max_vals(env
, insn
,
2271 if (rc
== -EACCES
&& env
->allow_ptr_leaks
) {
2272 /* unknown scalar += scalar */
2273 __mark_reg_unknown(dst_reg
);
2274 return adjust_scalar_min_max_vals(
2275 env
, insn
, dst_reg
, *src_reg
);
2280 /* Pretend the src is a reg with a known value, since we only
2281 * need to be able to read from this state.
2283 off_reg
.type
= SCALAR_VALUE
;
2284 __mark_reg_known(&off_reg
, insn
->imm
);
2286 if (ptr_reg
) { /* pointer += K */
2287 rc
= adjust_ptr_min_max_vals(env
, insn
,
2289 if (rc
== -EACCES
&& env
->allow_ptr_leaks
) {
2290 /* unknown scalar += K */
2291 __mark_reg_unknown(dst_reg
);
2292 return adjust_scalar_min_max_vals(
2293 env
, insn
, dst_reg
, off_reg
);
2299 /* Got here implies adding two SCALAR_VALUEs */
2300 if (WARN_ON_ONCE(ptr_reg
)) {
2301 print_verifier_state(env
, env
->cur_state
);
2302 verbose(env
, "verifier internal error: unexpected ptr_reg\n");
2305 if (WARN_ON(!src_reg
)) {
2306 print_verifier_state(env
, env
->cur_state
);
2307 verbose(env
, "verifier internal error: no src_reg\n");
2310 return adjust_scalar_min_max_vals(env
, insn
, dst_reg
, *src_reg
);
2313 /* check validity of 32-bit and 64-bit arithmetic operations */
2314 static int check_alu_op(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
2316 struct bpf_reg_state
*regs
= cur_regs(env
);
2317 u8 opcode
= BPF_OP(insn
->code
);
2320 if (opcode
== BPF_END
|| opcode
== BPF_NEG
) {
2321 if (opcode
== BPF_NEG
) {
2322 if (BPF_SRC(insn
->code
) != 0 ||
2323 insn
->src_reg
!= BPF_REG_0
||
2324 insn
->off
!= 0 || insn
->imm
!= 0) {
2325 verbose(env
, "BPF_NEG uses reserved fields\n");
2329 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0 ||
2330 (insn
->imm
!= 16 && insn
->imm
!= 32 && insn
->imm
!= 64) ||
2331 BPF_CLASS(insn
->code
) == BPF_ALU64
) {
2332 verbose(env
, "BPF_END uses reserved fields\n");
2337 /* check src operand */
2338 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
2342 if (is_pointer_value(env
, insn
->dst_reg
)) {
2343 verbose(env
, "R%d pointer arithmetic prohibited\n",
2348 /* check dest operand */
2349 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP
);
2353 } else if (opcode
== BPF_MOV
) {
2355 if (BPF_SRC(insn
->code
) == BPF_X
) {
2356 if (insn
->imm
!= 0 || insn
->off
!= 0) {
2357 verbose(env
, "BPF_MOV uses reserved fields\n");
2361 /* check src operand */
2362 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
2366 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0) {
2367 verbose(env
, "BPF_MOV uses reserved fields\n");
2372 /* check dest operand */
2373 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP
);
2377 if (BPF_SRC(insn
->code
) == BPF_X
) {
2378 if (BPF_CLASS(insn
->code
) == BPF_ALU64
) {
2380 * copy register state to dest reg
2382 regs
[insn
->dst_reg
] = regs
[insn
->src_reg
];
2383 regs
[insn
->dst_reg
].live
|= REG_LIVE_WRITTEN
;
2386 if (is_pointer_value(env
, insn
->src_reg
)) {
2388 "R%d partial copy of pointer\n",
2392 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
2393 /* high 32 bits are known zero. */
2394 regs
[insn
->dst_reg
].var_off
= tnum_cast(
2395 regs
[insn
->dst_reg
].var_off
, 4);
2396 __update_reg_bounds(®s
[insn
->dst_reg
]);
2400 * remember the value we stored into this reg
2402 regs
[insn
->dst_reg
].type
= SCALAR_VALUE
;
2403 __mark_reg_known(regs
+ insn
->dst_reg
, insn
->imm
);
2406 } else if (opcode
> BPF_END
) {
2407 verbose(env
, "invalid BPF_ALU opcode %x\n", opcode
);
2410 } else { /* all other ALU ops: and, sub, xor, add, ... */
2412 if (BPF_SRC(insn
->code
) == BPF_X
) {
2413 if (insn
->imm
!= 0 || insn
->off
!= 0) {
2414 verbose(env
, "BPF_ALU uses reserved fields\n");
2417 /* check src1 operand */
2418 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
2422 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0) {
2423 verbose(env
, "BPF_ALU uses reserved fields\n");
2428 /* check src2 operand */
2429 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
2433 if ((opcode
== BPF_MOD
|| opcode
== BPF_DIV
) &&
2434 BPF_SRC(insn
->code
) == BPF_K
&& insn
->imm
== 0) {
2435 verbose(env
, "div by zero\n");
2439 if ((opcode
== BPF_LSH
|| opcode
== BPF_RSH
||
2440 opcode
== BPF_ARSH
) && BPF_SRC(insn
->code
) == BPF_K
) {
2441 int size
= BPF_CLASS(insn
->code
) == BPF_ALU64
? 64 : 32;
2443 if (insn
->imm
< 0 || insn
->imm
>= size
) {
2444 verbose(env
, "invalid shift %d\n", insn
->imm
);
2449 /* check dest operand */
2450 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP_NO_MARK
);
2454 return adjust_reg_min_max_vals(env
, insn
);
2460 static void find_good_pkt_pointers(struct bpf_verifier_state
*state
,
2461 struct bpf_reg_state
*dst_reg
,
2462 enum bpf_reg_type type
,
2463 bool range_right_open
)
2465 struct bpf_reg_state
*regs
= state
->regs
, *reg
;
2469 if (dst_reg
->off
< 0 ||
2470 (dst_reg
->off
== 0 && range_right_open
))
2471 /* This doesn't give us any range */
2474 if (dst_reg
->umax_value
> MAX_PACKET_OFF
||
2475 dst_reg
->umax_value
+ dst_reg
->off
> MAX_PACKET_OFF
)
2476 /* Risk of overflow. For instance, ptr + (1<<63) may be less
2477 * than pkt_end, but that's because it's also less than pkt.
2481 new_range
= dst_reg
->off
;
2482 if (range_right_open
)
2485 /* Examples for register markings:
2487 * pkt_data in dst register:
2491 * if (r2 > pkt_end) goto <handle exception>
2496 * if (r2 < pkt_end) goto <access okay>
2497 * <handle exception>
2500 * r2 == dst_reg, pkt_end == src_reg
2501 * r2=pkt(id=n,off=8,r=0)
2502 * r3=pkt(id=n,off=0,r=0)
2504 * pkt_data in src register:
2508 * if (pkt_end >= r2) goto <access okay>
2509 * <handle exception>
2513 * if (pkt_end <= r2) goto <handle exception>
2517 * pkt_end == dst_reg, r2 == src_reg
2518 * r2=pkt(id=n,off=8,r=0)
2519 * r3=pkt(id=n,off=0,r=0)
2521 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
2522 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
2523 * and [r3, r3 + 8-1) respectively is safe to access depending on
2527 /* If our ids match, then we must have the same max_value. And we
2528 * don't care about the other reg's fixed offset, since if it's too big
2529 * the range won't allow anything.
2530 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
2532 for (i
= 0; i
< MAX_BPF_REG
; i
++)
2533 if (regs
[i
].type
== type
&& regs
[i
].id
== dst_reg
->id
)
2534 /* keep the maximum range already checked */
2535 regs
[i
].range
= max(regs
[i
].range
, new_range
);
2537 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
2538 if (state
->stack
[i
].slot_type
[0] != STACK_SPILL
)
2540 reg
= &state
->stack
[i
].spilled_ptr
;
2541 if (reg
->type
== type
&& reg
->id
== dst_reg
->id
)
2542 reg
->range
= max(reg
->range
, new_range
);
2546 /* Adjusts the register min/max values in the case that the dst_reg is the
2547 * variable register that we are working on, and src_reg is a constant or we're
2548 * simply doing a BPF_K check.
2549 * In JEQ/JNE cases we also adjust the var_off values.
2551 static void reg_set_min_max(struct bpf_reg_state
*true_reg
,
2552 struct bpf_reg_state
*false_reg
, u64 val
,
2555 /* If the dst_reg is a pointer, we can't learn anything about its
2556 * variable offset from the compare (unless src_reg were a pointer into
2557 * the same object, but we don't bother with that.
2558 * Since false_reg and true_reg have the same type by construction, we
2559 * only need to check one of them for pointerness.
2561 if (__is_pointer_value(false, false_reg
))
2566 /* If this is false then we know nothing Jon Snow, but if it is
2567 * true then we know for sure.
2569 __mark_reg_known(true_reg
, val
);
2572 /* If this is true we know nothing Jon Snow, but if it is false
2573 * we know the value for sure;
2575 __mark_reg_known(false_reg
, val
);
2578 false_reg
->umax_value
= min(false_reg
->umax_value
, val
);
2579 true_reg
->umin_value
= max(true_reg
->umin_value
, val
+ 1);
2582 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
);
2583 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
+ 1);
2586 false_reg
->umin_value
= max(false_reg
->umin_value
, val
);
2587 true_reg
->umax_value
= min(true_reg
->umax_value
, val
- 1);
2590 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
);
2591 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
- 1);
2594 false_reg
->umax_value
= min(false_reg
->umax_value
, val
- 1);
2595 true_reg
->umin_value
= max(true_reg
->umin_value
, val
);
2598 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
- 1);
2599 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
);
2602 false_reg
->umin_value
= max(false_reg
->umin_value
, val
+ 1);
2603 true_reg
->umax_value
= min(true_reg
->umax_value
, val
);
2606 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
+ 1);
2607 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
);
2613 __reg_deduce_bounds(false_reg
);
2614 __reg_deduce_bounds(true_reg
);
2615 /* We might have learned some bits from the bounds. */
2616 __reg_bound_offset(false_reg
);
2617 __reg_bound_offset(true_reg
);
2618 /* Intersecting with the old var_off might have improved our bounds
2619 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2620 * then new var_off is (0; 0x7f...fc) which improves our umax.
2622 __update_reg_bounds(false_reg
);
2623 __update_reg_bounds(true_reg
);
2626 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
2629 static void reg_set_min_max_inv(struct bpf_reg_state
*true_reg
,
2630 struct bpf_reg_state
*false_reg
, u64 val
,
2633 if (__is_pointer_value(false, false_reg
))
2638 /* If this is false then we know nothing Jon Snow, but if it is
2639 * true then we know for sure.
2641 __mark_reg_known(true_reg
, val
);
2644 /* If this is true we know nothing Jon Snow, but if it is false
2645 * we know the value for sure;
2647 __mark_reg_known(false_reg
, val
);
2650 true_reg
->umax_value
= min(true_reg
->umax_value
, val
- 1);
2651 false_reg
->umin_value
= max(false_reg
->umin_value
, val
);
2654 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
- 1);
2655 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
);
2658 true_reg
->umin_value
= max(true_reg
->umin_value
, val
+ 1);
2659 false_reg
->umax_value
= min(false_reg
->umax_value
, val
);
2662 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
+ 1);
2663 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
);
2666 true_reg
->umax_value
= min(true_reg
->umax_value
, val
);
2667 false_reg
->umin_value
= max(false_reg
->umin_value
, val
+ 1);
2670 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
);
2671 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
+ 1);
2674 true_reg
->umin_value
= max(true_reg
->umin_value
, val
);
2675 false_reg
->umax_value
= min(false_reg
->umax_value
, val
- 1);
2678 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
);
2679 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
- 1);
2685 __reg_deduce_bounds(false_reg
);
2686 __reg_deduce_bounds(true_reg
);
2687 /* We might have learned some bits from the bounds. */
2688 __reg_bound_offset(false_reg
);
2689 __reg_bound_offset(true_reg
);
2690 /* Intersecting with the old var_off might have improved our bounds
2691 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2692 * then new var_off is (0; 0x7f...fc) which improves our umax.
2694 __update_reg_bounds(false_reg
);
2695 __update_reg_bounds(true_reg
);
2698 /* Regs are known to be equal, so intersect their min/max/var_off */
2699 static void __reg_combine_min_max(struct bpf_reg_state
*src_reg
,
2700 struct bpf_reg_state
*dst_reg
)
2702 src_reg
->umin_value
= dst_reg
->umin_value
= max(src_reg
->umin_value
,
2703 dst_reg
->umin_value
);
2704 src_reg
->umax_value
= dst_reg
->umax_value
= min(src_reg
->umax_value
,
2705 dst_reg
->umax_value
);
2706 src_reg
->smin_value
= dst_reg
->smin_value
= max(src_reg
->smin_value
,
2707 dst_reg
->smin_value
);
2708 src_reg
->smax_value
= dst_reg
->smax_value
= min(src_reg
->smax_value
,
2709 dst_reg
->smax_value
);
2710 src_reg
->var_off
= dst_reg
->var_off
= tnum_intersect(src_reg
->var_off
,
2712 /* We might have learned new bounds from the var_off. */
2713 __update_reg_bounds(src_reg
);
2714 __update_reg_bounds(dst_reg
);
2715 /* We might have learned something about the sign bit. */
2716 __reg_deduce_bounds(src_reg
);
2717 __reg_deduce_bounds(dst_reg
);
2718 /* We might have learned some bits from the bounds. */
2719 __reg_bound_offset(src_reg
);
2720 __reg_bound_offset(dst_reg
);
2721 /* Intersecting with the old var_off might have improved our bounds
2722 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2723 * then new var_off is (0; 0x7f...fc) which improves our umax.
2725 __update_reg_bounds(src_reg
);
2726 __update_reg_bounds(dst_reg
);
2729 static void reg_combine_min_max(struct bpf_reg_state
*true_src
,
2730 struct bpf_reg_state
*true_dst
,
2731 struct bpf_reg_state
*false_src
,
2732 struct bpf_reg_state
*false_dst
,
2737 __reg_combine_min_max(true_src
, true_dst
);
2740 __reg_combine_min_max(false_src
, false_dst
);
2745 static void mark_map_reg(struct bpf_reg_state
*regs
, u32 regno
, u32 id
,
2748 struct bpf_reg_state
*reg
= ®s
[regno
];
2750 if (reg
->type
== PTR_TO_MAP_VALUE_OR_NULL
&& reg
->id
== id
) {
2751 /* Old offset (both fixed and variable parts) should
2752 * have been known-zero, because we don't allow pointer
2753 * arithmetic on pointers that might be NULL.
2755 if (WARN_ON_ONCE(reg
->smin_value
|| reg
->smax_value
||
2756 !tnum_equals_const(reg
->var_off
, 0) ||
2758 __mark_reg_known_zero(reg
);
2762 reg
->type
= SCALAR_VALUE
;
2763 } else if (reg
->map_ptr
->inner_map_meta
) {
2764 reg
->type
= CONST_PTR_TO_MAP
;
2765 reg
->map_ptr
= reg
->map_ptr
->inner_map_meta
;
2767 reg
->type
= PTR_TO_MAP_VALUE
;
2769 /* We don't need id from this point onwards anymore, thus we
2770 * should better reset it, so that state pruning has chances
2777 /* The logic is similar to find_good_pkt_pointers(), both could eventually
2778 * be folded together at some point.
2780 static void mark_map_regs(struct bpf_verifier_state
*state
, u32 regno
,
2783 struct bpf_reg_state
*regs
= state
->regs
;
2784 u32 id
= regs
[regno
].id
;
2787 for (i
= 0; i
< MAX_BPF_REG
; i
++)
2788 mark_map_reg(regs
, i
, id
, is_null
);
2790 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
2791 if (state
->stack
[i
].slot_type
[0] != STACK_SPILL
)
2793 mark_map_reg(&state
->stack
[i
].spilled_ptr
, 0, id
, is_null
);
2797 static bool try_match_pkt_pointers(const struct bpf_insn
*insn
,
2798 struct bpf_reg_state
*dst_reg
,
2799 struct bpf_reg_state
*src_reg
,
2800 struct bpf_verifier_state
*this_branch
,
2801 struct bpf_verifier_state
*other_branch
)
2803 if (BPF_SRC(insn
->code
) != BPF_X
)
2806 switch (BPF_OP(insn
->code
)) {
2808 if ((dst_reg
->type
== PTR_TO_PACKET
&&
2809 src_reg
->type
== PTR_TO_PACKET_END
) ||
2810 (dst_reg
->type
== PTR_TO_PACKET_META
&&
2811 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
2812 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
2813 find_good_pkt_pointers(this_branch
, dst_reg
,
2814 dst_reg
->type
, false);
2815 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
2816 src_reg
->type
== PTR_TO_PACKET
) ||
2817 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
2818 src_reg
->type
== PTR_TO_PACKET_META
)) {
2819 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
2820 find_good_pkt_pointers(other_branch
, src_reg
,
2821 src_reg
->type
, true);
2827 if ((dst_reg
->type
== PTR_TO_PACKET
&&
2828 src_reg
->type
== PTR_TO_PACKET_END
) ||
2829 (dst_reg
->type
== PTR_TO_PACKET_META
&&
2830 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
2831 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
2832 find_good_pkt_pointers(other_branch
, dst_reg
,
2833 dst_reg
->type
, true);
2834 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
2835 src_reg
->type
== PTR_TO_PACKET
) ||
2836 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
2837 src_reg
->type
== PTR_TO_PACKET_META
)) {
2838 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
2839 find_good_pkt_pointers(this_branch
, src_reg
,
2840 src_reg
->type
, false);
2846 if ((dst_reg
->type
== PTR_TO_PACKET
&&
2847 src_reg
->type
== PTR_TO_PACKET_END
) ||
2848 (dst_reg
->type
== PTR_TO_PACKET_META
&&
2849 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
2850 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
2851 find_good_pkt_pointers(this_branch
, dst_reg
,
2852 dst_reg
->type
, true);
2853 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
2854 src_reg
->type
== PTR_TO_PACKET
) ||
2855 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
2856 src_reg
->type
== PTR_TO_PACKET_META
)) {
2857 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
2858 find_good_pkt_pointers(other_branch
, src_reg
,
2859 src_reg
->type
, false);
2865 if ((dst_reg
->type
== PTR_TO_PACKET
&&
2866 src_reg
->type
== PTR_TO_PACKET_END
) ||
2867 (dst_reg
->type
== PTR_TO_PACKET_META
&&
2868 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
2869 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
2870 find_good_pkt_pointers(other_branch
, dst_reg
,
2871 dst_reg
->type
, false);
2872 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
2873 src_reg
->type
== PTR_TO_PACKET
) ||
2874 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
2875 src_reg
->type
== PTR_TO_PACKET_META
)) {
2876 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
2877 find_good_pkt_pointers(this_branch
, src_reg
,
2878 src_reg
->type
, true);
2890 static int check_cond_jmp_op(struct bpf_verifier_env
*env
,
2891 struct bpf_insn
*insn
, int *insn_idx
)
2893 struct bpf_verifier_state
*other_branch
, *this_branch
= env
->cur_state
;
2894 struct bpf_reg_state
*regs
= this_branch
->regs
, *dst_reg
;
2895 u8 opcode
= BPF_OP(insn
->code
);
2898 if (opcode
> BPF_JSLE
) {
2899 verbose(env
, "invalid BPF_JMP opcode %x\n", opcode
);
2903 if (BPF_SRC(insn
->code
) == BPF_X
) {
2904 if (insn
->imm
!= 0) {
2905 verbose(env
, "BPF_JMP uses reserved fields\n");
2909 /* check src1 operand */
2910 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
2914 if (is_pointer_value(env
, insn
->src_reg
)) {
2915 verbose(env
, "R%d pointer comparison prohibited\n",
2920 if (insn
->src_reg
!= BPF_REG_0
) {
2921 verbose(env
, "BPF_JMP uses reserved fields\n");
2926 /* check src2 operand */
2927 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
2931 dst_reg
= ®s
[insn
->dst_reg
];
2933 /* detect if R == 0 where R was initialized to zero earlier */
2934 if (BPF_SRC(insn
->code
) == BPF_K
&&
2935 (opcode
== BPF_JEQ
|| opcode
== BPF_JNE
) &&
2936 dst_reg
->type
== SCALAR_VALUE
&&
2937 tnum_equals_const(dst_reg
->var_off
, insn
->imm
)) {
2938 if (opcode
== BPF_JEQ
) {
2939 /* if (imm == imm) goto pc+off;
2940 * only follow the goto, ignore fall-through
2942 *insn_idx
+= insn
->off
;
2945 /* if (imm != imm) goto pc+off;
2946 * only follow fall-through branch, since
2947 * that's where the program will go
2953 other_branch
= push_stack(env
, *insn_idx
+ insn
->off
+ 1, *insn_idx
);
2957 /* detect if we are comparing against a constant value so we can adjust
2958 * our min/max values for our dst register.
2959 * this is only legit if both are scalars (or pointers to the same
2960 * object, I suppose, but we don't support that right now), because
2961 * otherwise the different base pointers mean the offsets aren't
2964 if (BPF_SRC(insn
->code
) == BPF_X
) {
2965 if (dst_reg
->type
== SCALAR_VALUE
&&
2966 regs
[insn
->src_reg
].type
== SCALAR_VALUE
) {
2967 if (tnum_is_const(regs
[insn
->src_reg
].var_off
))
2968 reg_set_min_max(&other_branch
->regs
[insn
->dst_reg
],
2969 dst_reg
, regs
[insn
->src_reg
].var_off
.value
,
2971 else if (tnum_is_const(dst_reg
->var_off
))
2972 reg_set_min_max_inv(&other_branch
->regs
[insn
->src_reg
],
2973 ®s
[insn
->src_reg
],
2974 dst_reg
->var_off
.value
, opcode
);
2975 else if (opcode
== BPF_JEQ
|| opcode
== BPF_JNE
)
2976 /* Comparing for equality, we can combine knowledge */
2977 reg_combine_min_max(&other_branch
->regs
[insn
->src_reg
],
2978 &other_branch
->regs
[insn
->dst_reg
],
2979 ®s
[insn
->src_reg
],
2980 ®s
[insn
->dst_reg
], opcode
);
2982 } else if (dst_reg
->type
== SCALAR_VALUE
) {
2983 reg_set_min_max(&other_branch
->regs
[insn
->dst_reg
],
2984 dst_reg
, insn
->imm
, opcode
);
2987 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
2988 if (BPF_SRC(insn
->code
) == BPF_K
&&
2989 insn
->imm
== 0 && (opcode
== BPF_JEQ
|| opcode
== BPF_JNE
) &&
2990 dst_reg
->type
== PTR_TO_MAP_VALUE_OR_NULL
) {
2991 /* Mark all identical map registers in each branch as either
2992 * safe or unknown depending R == 0 or R != 0 conditional.
2994 mark_map_regs(this_branch
, insn
->dst_reg
, opcode
== BPF_JNE
);
2995 mark_map_regs(other_branch
, insn
->dst_reg
, opcode
== BPF_JEQ
);
2996 } else if (!try_match_pkt_pointers(insn
, dst_reg
, ®s
[insn
->src_reg
],
2997 this_branch
, other_branch
) &&
2998 is_pointer_value(env
, insn
->dst_reg
)) {
2999 verbose(env
, "R%d pointer comparison prohibited\n",
3004 print_verifier_state(env
, this_branch
);
3008 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
3009 static struct bpf_map
*ld_imm64_to_map_ptr(struct bpf_insn
*insn
)
3011 u64 imm64
= ((u64
) (u32
) insn
[0].imm
) | ((u64
) (u32
) insn
[1].imm
) << 32;
3013 return (struct bpf_map
*) (unsigned long) imm64
;
3016 /* verify BPF_LD_IMM64 instruction */
3017 static int check_ld_imm(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
3019 struct bpf_reg_state
*regs
= cur_regs(env
);
3022 if (BPF_SIZE(insn
->code
) != BPF_DW
) {
3023 verbose(env
, "invalid BPF_LD_IMM insn\n");
3026 if (insn
->off
!= 0) {
3027 verbose(env
, "BPF_LD_IMM64 uses reserved fields\n");
3031 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP
);
3035 if (insn
->src_reg
== 0) {
3036 u64 imm
= ((u64
)(insn
+ 1)->imm
<< 32) | (u32
)insn
->imm
;
3038 regs
[insn
->dst_reg
].type
= SCALAR_VALUE
;
3039 __mark_reg_known(®s
[insn
->dst_reg
], imm
);
3043 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
3044 BUG_ON(insn
->src_reg
!= BPF_PSEUDO_MAP_FD
);
3046 regs
[insn
->dst_reg
].type
= CONST_PTR_TO_MAP
;
3047 regs
[insn
->dst_reg
].map_ptr
= ld_imm64_to_map_ptr(insn
);
3051 static bool may_access_skb(enum bpf_prog_type type
)
3054 case BPF_PROG_TYPE_SOCKET_FILTER
:
3055 case BPF_PROG_TYPE_SCHED_CLS
:
3056 case BPF_PROG_TYPE_SCHED_ACT
:
3063 /* verify safety of LD_ABS|LD_IND instructions:
3064 * - they can only appear in the programs where ctx == skb
3065 * - since they are wrappers of function calls, they scratch R1-R5 registers,
3066 * preserve R6-R9, and store return value into R0
3069 * ctx == skb == R6 == CTX
3072 * SRC == any register
3073 * IMM == 32-bit immediate
3076 * R0 - 8/16/32-bit skb data converted to cpu endianness
3078 static int check_ld_abs(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
3080 struct bpf_reg_state
*regs
= cur_regs(env
);
3081 u8 mode
= BPF_MODE(insn
->code
);
3084 if (!may_access_skb(env
->prog
->type
)) {
3085 verbose(env
, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
3089 if (insn
->dst_reg
!= BPF_REG_0
|| insn
->off
!= 0 ||
3090 BPF_SIZE(insn
->code
) == BPF_DW
||
3091 (mode
== BPF_ABS
&& insn
->src_reg
!= BPF_REG_0
)) {
3092 verbose(env
, "BPF_LD_[ABS|IND] uses reserved fields\n");
3096 /* check whether implicit source operand (register R6) is readable */
3097 err
= check_reg_arg(env
, BPF_REG_6
, SRC_OP
);
3101 if (regs
[BPF_REG_6
].type
!= PTR_TO_CTX
) {
3103 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
3107 if (mode
== BPF_IND
) {
3108 /* check explicit source operand */
3109 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
3114 /* reset caller saved regs to unreadable */
3115 for (i
= 0; i
< CALLER_SAVED_REGS
; i
++) {
3116 mark_reg_not_init(env
, regs
, caller_saved
[i
]);
3117 check_reg_arg(env
, caller_saved
[i
], DST_OP_NO_MARK
);
3120 /* mark destination R0 register as readable, since it contains
3121 * the value fetched from the packet.
3122 * Already marked as written above.
3124 mark_reg_unknown(env
, regs
, BPF_REG_0
);
3128 static int check_return_code(struct bpf_verifier_env
*env
)
3130 struct bpf_reg_state
*reg
;
3131 struct tnum range
= tnum_range(0, 1);
3133 switch (env
->prog
->type
) {
3134 case BPF_PROG_TYPE_CGROUP_SKB
:
3135 case BPF_PROG_TYPE_CGROUP_SOCK
:
3136 case BPF_PROG_TYPE_SOCK_OPS
:
3137 case BPF_PROG_TYPE_CGROUP_DEVICE
:
3143 reg
= cur_regs(env
) + BPF_REG_0
;
3144 if (reg
->type
!= SCALAR_VALUE
) {
3145 verbose(env
, "At program exit the register R0 is not a known value (%s)\n",
3146 reg_type_str
[reg
->type
]);
3150 if (!tnum_in(range
, reg
->var_off
)) {
3151 verbose(env
, "At program exit the register R0 ");
3152 if (!tnum_is_unknown(reg
->var_off
)) {
3155 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
3156 verbose(env
, "has value %s", tn_buf
);
3158 verbose(env
, "has unknown scalar value");
3160 verbose(env
, " should have been 0 or 1\n");
3166 /* non-recursive DFS pseudo code
3167 * 1 procedure DFS-iterative(G,v):
3168 * 2 label v as discovered
3169 * 3 let S be a stack
3171 * 5 while S is not empty
3173 * 7 if t is what we're looking for:
3175 * 9 for all edges e in G.adjacentEdges(t) do
3176 * 10 if edge e is already labelled
3177 * 11 continue with the next edge
3178 * 12 w <- G.adjacentVertex(t,e)
3179 * 13 if vertex w is not discovered and not explored
3180 * 14 label e as tree-edge
3181 * 15 label w as discovered
3184 * 18 else if vertex w is discovered
3185 * 19 label e as back-edge
3187 * 21 // vertex w is explored
3188 * 22 label e as forward- or cross-edge
3189 * 23 label t as explored
3194 * 0x11 - discovered and fall-through edge labelled
3195 * 0x12 - discovered and fall-through and branch edges labelled
3206 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
3208 static int *insn_stack
; /* stack of insns to process */
3209 static int cur_stack
; /* current stack index */
3210 static int *insn_state
;
3212 /* t, w, e - match pseudo-code above:
3213 * t - index of current instruction
3214 * w - next instruction
3217 static int push_insn(int t
, int w
, int e
, struct bpf_verifier_env
*env
)
3219 if (e
== FALLTHROUGH
&& insn_state
[t
] >= (DISCOVERED
| FALLTHROUGH
))
3222 if (e
== BRANCH
&& insn_state
[t
] >= (DISCOVERED
| BRANCH
))
3225 if (w
< 0 || w
>= env
->prog
->len
) {
3226 verbose(env
, "jump out of range from insn %d to %d\n", t
, w
);
3231 /* mark branch target for state pruning */
3232 env
->explored_states
[w
] = STATE_LIST_MARK
;
3234 if (insn_state
[w
] == 0) {
3236 insn_state
[t
] = DISCOVERED
| e
;
3237 insn_state
[w
] = DISCOVERED
;
3238 if (cur_stack
>= env
->prog
->len
)
3240 insn_stack
[cur_stack
++] = w
;
3242 } else if ((insn_state
[w
] & 0xF0) == DISCOVERED
) {
3243 verbose(env
, "back-edge from insn %d to %d\n", t
, w
);
3245 } else if (insn_state
[w
] == EXPLORED
) {
3246 /* forward- or cross-edge */
3247 insn_state
[t
] = DISCOVERED
| e
;
3249 verbose(env
, "insn state internal bug\n");
3255 /* non-recursive depth-first-search to detect loops in BPF program
3256 * loop == back-edge in directed graph
3258 static int check_cfg(struct bpf_verifier_env
*env
)
3260 struct bpf_insn
*insns
= env
->prog
->insnsi
;
3261 int insn_cnt
= env
->prog
->len
;
3265 insn_state
= kcalloc(insn_cnt
, sizeof(int), GFP_KERNEL
);
3269 insn_stack
= kcalloc(insn_cnt
, sizeof(int), GFP_KERNEL
);
3275 insn_state
[0] = DISCOVERED
; /* mark 1st insn as discovered */
3276 insn_stack
[0] = 0; /* 0 is the first instruction */
3282 t
= insn_stack
[cur_stack
- 1];
3284 if (BPF_CLASS(insns
[t
].code
) == BPF_JMP
) {
3285 u8 opcode
= BPF_OP(insns
[t
].code
);
3287 if (opcode
== BPF_EXIT
) {
3289 } else if (opcode
== BPF_CALL
) {
3290 ret
= push_insn(t
, t
+ 1, FALLTHROUGH
, env
);
3295 if (t
+ 1 < insn_cnt
)
3296 env
->explored_states
[t
+ 1] = STATE_LIST_MARK
;
3297 } else if (opcode
== BPF_JA
) {
3298 if (BPF_SRC(insns
[t
].code
) != BPF_K
) {
3302 /* unconditional jump with single edge */
3303 ret
= push_insn(t
, t
+ insns
[t
].off
+ 1,
3309 /* tell verifier to check for equivalent states
3310 * after every call and jump
3312 if (t
+ 1 < insn_cnt
)
3313 env
->explored_states
[t
+ 1] = STATE_LIST_MARK
;
3315 /* conditional jump with two edges */
3316 env
->explored_states
[t
] = STATE_LIST_MARK
;
3317 ret
= push_insn(t
, t
+ 1, FALLTHROUGH
, env
);
3323 ret
= push_insn(t
, t
+ insns
[t
].off
+ 1, BRANCH
, env
);
3330 /* all other non-branch instructions with single
3333 ret
= push_insn(t
, t
+ 1, FALLTHROUGH
, env
);
3341 insn_state
[t
] = EXPLORED
;
3342 if (cur_stack
-- <= 0) {
3343 verbose(env
, "pop stack internal bug\n");
3350 for (i
= 0; i
< insn_cnt
; i
++) {
3351 if (insn_state
[i
] != EXPLORED
) {
3352 verbose(env
, "unreachable insn %d\n", i
);
3357 ret
= 0; /* cfg looks good */
3365 /* check %cur's range satisfies %old's */
3366 static bool range_within(struct bpf_reg_state
*old
,
3367 struct bpf_reg_state
*cur
)
3369 return old
->umin_value
<= cur
->umin_value
&&
3370 old
->umax_value
>= cur
->umax_value
&&
3371 old
->smin_value
<= cur
->smin_value
&&
3372 old
->smax_value
>= cur
->smax_value
;
3375 /* Maximum number of register states that can exist at once */
3376 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
3382 /* If in the old state two registers had the same id, then they need to have
3383 * the same id in the new state as well. But that id could be different from
3384 * the old state, so we need to track the mapping from old to new ids.
3385 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
3386 * regs with old id 5 must also have new id 9 for the new state to be safe. But
3387 * regs with a different old id could still have new id 9, we don't care about
3389 * So we look through our idmap to see if this old id has been seen before. If
3390 * so, we require the new id to match; otherwise, we add the id pair to the map.
3392 static bool check_ids(u32 old_id
, u32 cur_id
, struct idpair
*idmap
)
3396 for (i
= 0; i
< ID_MAP_SIZE
; i
++) {
3397 if (!idmap
[i
].old
) {
3398 /* Reached an empty slot; haven't seen this id before */
3399 idmap
[i
].old
= old_id
;
3400 idmap
[i
].cur
= cur_id
;
3403 if (idmap
[i
].old
== old_id
)
3404 return idmap
[i
].cur
== cur_id
;
3406 /* We ran out of idmap slots, which should be impossible */
3411 /* Returns true if (rold safe implies rcur safe) */
3412 static bool regsafe(struct bpf_reg_state
*rold
, struct bpf_reg_state
*rcur
,
3413 struct idpair
*idmap
)
3415 if (!(rold
->live
& REG_LIVE_READ
))
3416 /* explored state didn't use this */
3419 if (memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, live
)) == 0)
3422 if (rold
->type
== NOT_INIT
)
3423 /* explored state can't have used this */
3425 if (rcur
->type
== NOT_INIT
)
3427 switch (rold
->type
) {
3429 if (rcur
->type
== SCALAR_VALUE
) {
3430 /* new val must satisfy old val knowledge */
3431 return range_within(rold
, rcur
) &&
3432 tnum_in(rold
->var_off
, rcur
->var_off
);
3434 /* if we knew anything about the old value, we're not
3435 * equal, because we can't know anything about the
3436 * scalar value of the pointer in the new value.
3438 return rold
->umin_value
== 0 &&
3439 rold
->umax_value
== U64_MAX
&&
3440 rold
->smin_value
== S64_MIN
&&
3441 rold
->smax_value
== S64_MAX
&&
3442 tnum_is_unknown(rold
->var_off
);
3444 case PTR_TO_MAP_VALUE
:
3445 /* If the new min/max/var_off satisfy the old ones and
3446 * everything else matches, we are OK.
3447 * We don't care about the 'id' value, because nothing
3448 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
3450 return memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, id
)) == 0 &&
3451 range_within(rold
, rcur
) &&
3452 tnum_in(rold
->var_off
, rcur
->var_off
);
3453 case PTR_TO_MAP_VALUE_OR_NULL
:
3454 /* a PTR_TO_MAP_VALUE could be safe to use as a
3455 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
3456 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
3457 * checked, doing so could have affected others with the same
3458 * id, and we can't check for that because we lost the id when
3459 * we converted to a PTR_TO_MAP_VALUE.
3461 if (rcur
->type
!= PTR_TO_MAP_VALUE_OR_NULL
)
3463 if (memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, id
)))
3465 /* Check our ids match any regs they're supposed to */
3466 return check_ids(rold
->id
, rcur
->id
, idmap
);
3467 case PTR_TO_PACKET_META
:
3469 if (rcur
->type
!= rold
->type
)
3471 /* We must have at least as much range as the old ptr
3472 * did, so that any accesses which were safe before are
3473 * still safe. This is true even if old range < old off,
3474 * since someone could have accessed through (ptr - k), or
3475 * even done ptr -= k in a register, to get a safe access.
3477 if (rold
->range
> rcur
->range
)
3479 /* If the offsets don't match, we can't trust our alignment;
3480 * nor can we be sure that we won't fall out of range.
3482 if (rold
->off
!= rcur
->off
)
3484 /* id relations must be preserved */
3485 if (rold
->id
&& !check_ids(rold
->id
, rcur
->id
, idmap
))
3487 /* new val must satisfy old val knowledge */
3488 return range_within(rold
, rcur
) &&
3489 tnum_in(rold
->var_off
, rcur
->var_off
);
3491 case CONST_PTR_TO_MAP
:
3493 case PTR_TO_PACKET_END
:
3494 /* Only valid matches are exact, which memcmp() above
3495 * would have accepted
3498 /* Don't know what's going on, just say it's not safe */
3502 /* Shouldn't get here; if we do, say it's not safe */
3507 static bool stacksafe(struct bpf_verifier_state
*old
,
3508 struct bpf_verifier_state
*cur
,
3509 struct idpair
*idmap
)
3513 /* if explored stack has more populated slots than current stack
3514 * such stacks are not equivalent
3516 if (old
->allocated_stack
> cur
->allocated_stack
)
3519 /* walk slots of the explored stack and ignore any additional
3520 * slots in the current stack, since explored(safe) state
3523 for (i
= 0; i
< old
->allocated_stack
; i
++) {
3524 spi
= i
/ BPF_REG_SIZE
;
3526 if (old
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
] == STACK_INVALID
)
3528 if (old
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
] !=
3529 cur
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
])
3530 /* Ex: old explored (safe) state has STACK_SPILL in
3531 * this stack slot, but current has has STACK_MISC ->
3532 * this verifier states are not equivalent,
3533 * return false to continue verification of this path
3536 if (i
% BPF_REG_SIZE
)
3538 if (old
->stack
[spi
].slot_type
[0] != STACK_SPILL
)
3540 if (!regsafe(&old
->stack
[spi
].spilled_ptr
,
3541 &cur
->stack
[spi
].spilled_ptr
,
3543 /* when explored and current stack slot are both storing
3544 * spilled registers, check that stored pointers types
3545 * are the same as well.
3546 * Ex: explored safe path could have stored
3547 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
3548 * but current path has stored:
3549 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
3550 * such verifier states are not equivalent.
3551 * return false to continue verification of this path
3558 /* compare two verifier states
3560 * all states stored in state_list are known to be valid, since
3561 * verifier reached 'bpf_exit' instruction through them
3563 * this function is called when verifier exploring different branches of
3564 * execution popped from the state stack. If it sees an old state that has
3565 * more strict register state and more strict stack state then this execution
3566 * branch doesn't need to be explored further, since verifier already
3567 * concluded that more strict state leads to valid finish.
3569 * Therefore two states are equivalent if register state is more conservative
3570 * and explored stack state is more conservative than the current one.
3573 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
3574 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
3576 * In other words if current stack state (one being explored) has more
3577 * valid slots than old one that already passed validation, it means
3578 * the verifier can stop exploring and conclude that current state is valid too
3580 * Similarly with registers. If explored state has register type as invalid
3581 * whereas register type in current state is meaningful, it means that
3582 * the current state will reach 'bpf_exit' instruction safely
3584 static bool states_equal(struct bpf_verifier_env
*env
,
3585 struct bpf_verifier_state
*old
,
3586 struct bpf_verifier_state
*cur
)
3588 struct idpair
*idmap
;
3592 idmap
= kcalloc(ID_MAP_SIZE
, sizeof(struct idpair
), GFP_KERNEL
);
3593 /* If we failed to allocate the idmap, just say it's not safe */
3597 for (i
= 0; i
< MAX_BPF_REG
; i
++) {
3598 if (!regsafe(&old
->regs
[i
], &cur
->regs
[i
], idmap
))
3602 if (!stacksafe(old
, cur
, idmap
))
3610 /* A write screens off any subsequent reads; but write marks come from the
3611 * straight-line code between a state and its parent. When we arrive at a
3612 * jump target (in the first iteration of the propagate_liveness() loop),
3613 * we didn't arrive by the straight-line code, so read marks in state must
3614 * propagate to parent regardless of state's write marks.
3616 static bool do_propagate_liveness(const struct bpf_verifier_state
*state
,
3617 struct bpf_verifier_state
*parent
)
3619 bool writes
= parent
== state
->parent
; /* Observe write marks */
3620 bool touched
= false; /* any changes made? */
3625 /* Propagate read liveness of registers... */
3626 BUILD_BUG_ON(BPF_REG_FP
+ 1 != MAX_BPF_REG
);
3627 /* We don't need to worry about FP liveness because it's read-only */
3628 for (i
= 0; i
< BPF_REG_FP
; i
++) {
3629 if (parent
->regs
[i
].live
& REG_LIVE_READ
)
3631 if (writes
&& (state
->regs
[i
].live
& REG_LIVE_WRITTEN
))
3633 if (state
->regs
[i
].live
& REG_LIVE_READ
) {
3634 parent
->regs
[i
].live
|= REG_LIVE_READ
;
3638 /* ... and stack slots */
3639 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
&&
3640 i
< parent
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
3641 if (parent
->stack
[i
].slot_type
[0] != STACK_SPILL
)
3643 if (state
->stack
[i
].slot_type
[0] != STACK_SPILL
)
3645 if (parent
->stack
[i
].spilled_ptr
.live
& REG_LIVE_READ
)
3648 (state
->stack
[i
].spilled_ptr
.live
& REG_LIVE_WRITTEN
))
3650 if (state
->stack
[i
].spilled_ptr
.live
& REG_LIVE_READ
) {
3651 parent
->stack
[i
].spilled_ptr
.live
|= REG_LIVE_READ
;
3658 /* "parent" is "a state from which we reach the current state", but initially
3659 * it is not the state->parent (i.e. "the state whose straight-line code leads
3660 * to the current state"), instead it is the state that happened to arrive at
3661 * a (prunable) equivalent of the current state. See comment above
3662 * do_propagate_liveness() for consequences of this.
3663 * This function is just a more efficient way of calling mark_reg_read() or
3664 * mark_stack_slot_read() on each reg in "parent" that is read in "state",
3665 * though it requires that parent != state->parent in the call arguments.
3667 static void propagate_liveness(const struct bpf_verifier_state
*state
,
3668 struct bpf_verifier_state
*parent
)
3670 while (do_propagate_liveness(state
, parent
)) {
3671 /* Something changed, so we need to feed those changes onward */
3673 parent
= state
->parent
;
3677 static int is_state_visited(struct bpf_verifier_env
*env
, int insn_idx
)
3679 struct bpf_verifier_state_list
*new_sl
;
3680 struct bpf_verifier_state_list
*sl
;
3681 struct bpf_verifier_state
*cur
= env
->cur_state
;
3684 sl
= env
->explored_states
[insn_idx
];
3686 /* this 'insn_idx' instruction wasn't marked, so we will not
3687 * be doing state search here
3691 while (sl
!= STATE_LIST_MARK
) {
3692 if (states_equal(env
, &sl
->state
, cur
)) {
3693 /* reached equivalent register/stack state,
3695 * Registers read by the continuation are read by us.
3696 * If we have any write marks in env->cur_state, they
3697 * will prevent corresponding reads in the continuation
3698 * from reaching our parent (an explored_state). Our
3699 * own state will get the read marks recorded, but
3700 * they'll be immediately forgotten as we're pruning
3701 * this state and will pop a new one.
3703 propagate_liveness(&sl
->state
, cur
);
3709 /* there were no equivalent states, remember current one.
3710 * technically the current state is not proven to be safe yet,
3711 * but it will either reach bpf_exit (which means it's safe) or
3712 * it will be rejected. Since there are no loops, we won't be
3713 * seeing this 'insn_idx' instruction again on the way to bpf_exit
3715 new_sl
= kzalloc(sizeof(struct bpf_verifier_state_list
), GFP_KERNEL
);
3719 /* add new state to the head of linked list */
3720 err
= copy_verifier_state(&new_sl
->state
, cur
);
3722 free_verifier_state(&new_sl
->state
, false);
3726 new_sl
->next
= env
->explored_states
[insn_idx
];
3727 env
->explored_states
[insn_idx
] = new_sl
;
3728 /* connect new state to parentage chain */
3729 cur
->parent
= &new_sl
->state
;
3730 /* clear write marks in current state: the writes we did are not writes
3731 * our child did, so they don't screen off its reads from us.
3732 * (There are no read marks in current state, because reads always mark
3733 * their parent and current state never has children yet. Only
3734 * explored_states can get read marks.)
3736 for (i
= 0; i
< BPF_REG_FP
; i
++)
3737 cur
->regs
[i
].live
= REG_LIVE_NONE
;
3738 for (i
= 0; i
< cur
->allocated_stack
/ BPF_REG_SIZE
; i
++)
3739 if (cur
->stack
[i
].slot_type
[0] == STACK_SPILL
)
3740 cur
->stack
[i
].spilled_ptr
.live
= REG_LIVE_NONE
;
3744 static int ext_analyzer_insn_hook(struct bpf_verifier_env
*env
,
3745 int insn_idx
, int prev_insn_idx
)
3747 if (env
->dev_ops
&& env
->dev_ops
->insn_hook
)
3748 return env
->dev_ops
->insn_hook(env
, insn_idx
, prev_insn_idx
);
3753 static int do_check(struct bpf_verifier_env
*env
)
3755 struct bpf_verifier_state
*state
;
3756 struct bpf_insn
*insns
= env
->prog
->insnsi
;
3757 struct bpf_reg_state
*regs
;
3758 int insn_cnt
= env
->prog
->len
;
3759 int insn_idx
, prev_insn_idx
= 0;
3760 int insn_processed
= 0;
3761 bool do_print_state
= false;
3763 state
= kzalloc(sizeof(struct bpf_verifier_state
), GFP_KERNEL
);
3766 env
->cur_state
= state
;
3767 init_reg_state(env
, state
->regs
);
3768 state
->parent
= NULL
;
3771 struct bpf_insn
*insn
;
3775 if (insn_idx
>= insn_cnt
) {
3776 verbose(env
, "invalid insn idx %d insn_cnt %d\n",
3777 insn_idx
, insn_cnt
);
3781 insn
= &insns
[insn_idx
];
3782 class = BPF_CLASS(insn
->code
);
3784 if (++insn_processed
> BPF_COMPLEXITY_LIMIT_INSNS
) {
3786 "BPF program is too large. Processed %d insn\n",
3791 err
= is_state_visited(env
, insn_idx
);
3795 /* found equivalent state, can prune the search */
3796 if (env
->log
.level
) {
3798 verbose(env
, "\nfrom %d to %d: safe\n",
3799 prev_insn_idx
, insn_idx
);
3801 verbose(env
, "%d: safe\n", insn_idx
);
3803 goto process_bpf_exit
;
3809 if (env
->log
.level
> 1 || (env
->log
.level
&& do_print_state
)) {
3810 if (env
->log
.level
> 1)
3811 verbose(env
, "%d:", insn_idx
);
3813 verbose(env
, "\nfrom %d to %d:",
3814 prev_insn_idx
, insn_idx
);
3815 print_verifier_state(env
, state
);
3816 do_print_state
= false;
3819 if (env
->log
.level
) {
3820 verbose(env
, "%d: ", insn_idx
);
3821 print_bpf_insn(verbose
, env
, insn
,
3822 env
->allow_ptr_leaks
);
3825 err
= ext_analyzer_insn_hook(env
, insn_idx
, prev_insn_idx
);
3829 regs
= cur_regs(env
);
3830 env
->insn_aux_data
[insn_idx
].seen
= true;
3831 if (class == BPF_ALU
|| class == BPF_ALU64
) {
3832 err
= check_alu_op(env
, insn
);
3836 } else if (class == BPF_LDX
) {
3837 enum bpf_reg_type
*prev_src_type
, src_reg_type
;
3839 /* check for reserved fields is already done */
3841 /* check src operand */
3842 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
3846 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP_NO_MARK
);
3850 src_reg_type
= regs
[insn
->src_reg
].type
;
3852 /* check that memory (src_reg + off) is readable,
3853 * the state of dst_reg will be updated by this func
3855 err
= check_mem_access(env
, insn_idx
, insn
->src_reg
, insn
->off
,
3856 BPF_SIZE(insn
->code
), BPF_READ
,
3861 prev_src_type
= &env
->insn_aux_data
[insn_idx
].ptr_type
;
3863 if (*prev_src_type
== NOT_INIT
) {
3865 * dst_reg = *(u32 *)(src_reg + off)
3866 * save type to validate intersecting paths
3868 *prev_src_type
= src_reg_type
;
3870 } else if (src_reg_type
!= *prev_src_type
&&
3871 (src_reg_type
== PTR_TO_CTX
||
3872 *prev_src_type
== PTR_TO_CTX
)) {
3873 /* ABuser program is trying to use the same insn
3874 * dst_reg = *(u32*) (src_reg + off)
3875 * with different pointer types:
3876 * src_reg == ctx in one branch and
3877 * src_reg == stack|map in some other branch.
3880 verbose(env
, "same insn cannot be used with different pointers\n");
3884 } else if (class == BPF_STX
) {
3885 enum bpf_reg_type
*prev_dst_type
, dst_reg_type
;
3887 if (BPF_MODE(insn
->code
) == BPF_XADD
) {
3888 err
= check_xadd(env
, insn_idx
, insn
);
3895 /* check src1 operand */
3896 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
3899 /* check src2 operand */
3900 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
3904 dst_reg_type
= regs
[insn
->dst_reg
].type
;
3906 /* check that memory (dst_reg + off) is writeable */
3907 err
= check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
3908 BPF_SIZE(insn
->code
), BPF_WRITE
,
3913 prev_dst_type
= &env
->insn_aux_data
[insn_idx
].ptr_type
;
3915 if (*prev_dst_type
== NOT_INIT
) {
3916 *prev_dst_type
= dst_reg_type
;
3917 } else if (dst_reg_type
!= *prev_dst_type
&&
3918 (dst_reg_type
== PTR_TO_CTX
||
3919 *prev_dst_type
== PTR_TO_CTX
)) {
3920 verbose(env
, "same insn cannot be used with different pointers\n");
3924 } else if (class == BPF_ST
) {
3925 if (BPF_MODE(insn
->code
) != BPF_MEM
||
3926 insn
->src_reg
!= BPF_REG_0
) {
3927 verbose(env
, "BPF_ST uses reserved fields\n");
3930 /* check src operand */
3931 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
3935 /* check that memory (dst_reg + off) is writeable */
3936 err
= check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
3937 BPF_SIZE(insn
->code
), BPF_WRITE
,
3942 } else if (class == BPF_JMP
) {
3943 u8 opcode
= BPF_OP(insn
->code
);
3945 if (opcode
== BPF_CALL
) {
3946 if (BPF_SRC(insn
->code
) != BPF_K
||
3948 insn
->src_reg
!= BPF_REG_0
||
3949 insn
->dst_reg
!= BPF_REG_0
) {
3950 verbose(env
, "BPF_CALL uses reserved fields\n");
3954 err
= check_call(env
, insn
->imm
, insn_idx
);
3958 } else if (opcode
== BPF_JA
) {
3959 if (BPF_SRC(insn
->code
) != BPF_K
||
3961 insn
->src_reg
!= BPF_REG_0
||
3962 insn
->dst_reg
!= BPF_REG_0
) {
3963 verbose(env
, "BPF_JA uses reserved fields\n");
3967 insn_idx
+= insn
->off
+ 1;
3970 } else if (opcode
== BPF_EXIT
) {
3971 if (BPF_SRC(insn
->code
) != BPF_K
||
3973 insn
->src_reg
!= BPF_REG_0
||
3974 insn
->dst_reg
!= BPF_REG_0
) {
3975 verbose(env
, "BPF_EXIT uses reserved fields\n");
3979 /* eBPF calling convetion is such that R0 is used
3980 * to return the value from eBPF program.
3981 * Make sure that it's readable at this time
3982 * of bpf_exit, which means that program wrote
3983 * something into it earlier
3985 err
= check_reg_arg(env
, BPF_REG_0
, SRC_OP
);
3989 if (is_pointer_value(env
, BPF_REG_0
)) {
3990 verbose(env
, "R0 leaks addr as return value\n");
3994 err
= check_return_code(env
);
3998 err
= pop_stack(env
, &prev_insn_idx
, &insn_idx
);
4004 do_print_state
= true;
4008 err
= check_cond_jmp_op(env
, insn
, &insn_idx
);
4012 } else if (class == BPF_LD
) {
4013 u8 mode
= BPF_MODE(insn
->code
);
4015 if (mode
== BPF_ABS
|| mode
== BPF_IND
) {
4016 err
= check_ld_abs(env
, insn
);
4020 } else if (mode
== BPF_IMM
) {
4021 err
= check_ld_imm(env
, insn
);
4026 env
->insn_aux_data
[insn_idx
].seen
= true;
4028 verbose(env
, "invalid BPF_LD mode\n");
4032 verbose(env
, "unknown insn class %d\n", class);
4039 verbose(env
, "processed %d insns, stack depth %d\n", insn_processed
,
4040 env
->prog
->aux
->stack_depth
);
4044 static int check_map_prealloc(struct bpf_map
*map
)
4046 return (map
->map_type
!= BPF_MAP_TYPE_HASH
&&
4047 map
->map_type
!= BPF_MAP_TYPE_PERCPU_HASH
&&
4048 map
->map_type
!= BPF_MAP_TYPE_HASH_OF_MAPS
) ||
4049 !(map
->map_flags
& BPF_F_NO_PREALLOC
);
4052 static int check_map_prog_compatibility(struct bpf_verifier_env
*env
,
4053 struct bpf_map
*map
,
4054 struct bpf_prog
*prog
)
4057 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
4058 * preallocated hash maps, since doing memory allocation
4059 * in overflow_handler can crash depending on where nmi got
4062 if (prog
->type
== BPF_PROG_TYPE_PERF_EVENT
) {
4063 if (!check_map_prealloc(map
)) {
4064 verbose(env
, "perf_event programs can only use preallocated hash map\n");
4067 if (map
->inner_map_meta
&&
4068 !check_map_prealloc(map
->inner_map_meta
)) {
4069 verbose(env
, "perf_event programs can only use preallocated inner hash map\n");
4076 /* look for pseudo eBPF instructions that access map FDs and
4077 * replace them with actual map pointers
4079 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env
*env
)
4081 struct bpf_insn
*insn
= env
->prog
->insnsi
;
4082 int insn_cnt
= env
->prog
->len
;
4085 err
= bpf_prog_calc_tag(env
->prog
);
4089 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
4090 if (BPF_CLASS(insn
->code
) == BPF_LDX
&&
4091 (BPF_MODE(insn
->code
) != BPF_MEM
|| insn
->imm
!= 0)) {
4092 verbose(env
, "BPF_LDX uses reserved fields\n");
4096 if (BPF_CLASS(insn
->code
) == BPF_STX
&&
4097 ((BPF_MODE(insn
->code
) != BPF_MEM
&&
4098 BPF_MODE(insn
->code
) != BPF_XADD
) || insn
->imm
!= 0)) {
4099 verbose(env
, "BPF_STX uses reserved fields\n");
4103 if (insn
[0].code
== (BPF_LD
| BPF_IMM
| BPF_DW
)) {
4104 struct bpf_map
*map
;
4107 if (i
== insn_cnt
- 1 || insn
[1].code
!= 0 ||
4108 insn
[1].dst_reg
!= 0 || insn
[1].src_reg
!= 0 ||
4110 verbose(env
, "invalid bpf_ld_imm64 insn\n");
4114 if (insn
->src_reg
== 0)
4115 /* valid generic load 64-bit imm */
4118 if (insn
->src_reg
!= BPF_PSEUDO_MAP_FD
) {
4120 "unrecognized bpf_ld_imm64 insn\n");
4124 f
= fdget(insn
->imm
);
4125 map
= __bpf_map_get(f
);
4127 verbose(env
, "fd %d is not pointing to valid bpf_map\n",
4129 return PTR_ERR(map
);
4132 err
= check_map_prog_compatibility(env
, map
, env
->prog
);
4138 /* store map pointer inside BPF_LD_IMM64 instruction */
4139 insn
[0].imm
= (u32
) (unsigned long) map
;
4140 insn
[1].imm
= ((u64
) (unsigned long) map
) >> 32;
4142 /* check whether we recorded this map already */
4143 for (j
= 0; j
< env
->used_map_cnt
; j
++)
4144 if (env
->used_maps
[j
] == map
) {
4149 if (env
->used_map_cnt
>= MAX_USED_MAPS
) {
4154 /* hold the map. If the program is rejected by verifier,
4155 * the map will be released by release_maps() or it
4156 * will be used by the valid program until it's unloaded
4157 * and all maps are released in free_bpf_prog_info()
4159 map
= bpf_map_inc(map
, false);
4162 return PTR_ERR(map
);
4164 env
->used_maps
[env
->used_map_cnt
++] = map
;
4173 /* now all pseudo BPF_LD_IMM64 instructions load valid
4174 * 'struct bpf_map *' into a register instead of user map_fd.
4175 * These pointers will be used later by verifier to validate map access.
4180 /* drop refcnt of maps used by the rejected program */
4181 static void release_maps(struct bpf_verifier_env
*env
)
4185 for (i
= 0; i
< env
->used_map_cnt
; i
++)
4186 bpf_map_put(env
->used_maps
[i
]);
4189 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
4190 static void convert_pseudo_ld_imm64(struct bpf_verifier_env
*env
)
4192 struct bpf_insn
*insn
= env
->prog
->insnsi
;
4193 int insn_cnt
= env
->prog
->len
;
4196 for (i
= 0; i
< insn_cnt
; i
++, insn
++)
4197 if (insn
->code
== (BPF_LD
| BPF_IMM
| BPF_DW
))
4201 /* single env->prog->insni[off] instruction was replaced with the range
4202 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
4203 * [0, off) and [off, end) to new locations, so the patched range stays zero
4205 static int adjust_insn_aux_data(struct bpf_verifier_env
*env
, u32 prog_len
,
4208 struct bpf_insn_aux_data
*new_data
, *old_data
= env
->insn_aux_data
;
4213 new_data
= vzalloc(sizeof(struct bpf_insn_aux_data
) * prog_len
);
4216 memcpy(new_data
, old_data
, sizeof(struct bpf_insn_aux_data
) * off
);
4217 memcpy(new_data
+ off
+ cnt
- 1, old_data
+ off
,
4218 sizeof(struct bpf_insn_aux_data
) * (prog_len
- off
- cnt
+ 1));
4219 for (i
= off
; i
< off
+ cnt
- 1; i
++)
4220 new_data
[i
].seen
= true;
4221 env
->insn_aux_data
= new_data
;
4226 static struct bpf_prog
*bpf_patch_insn_data(struct bpf_verifier_env
*env
, u32 off
,
4227 const struct bpf_insn
*patch
, u32 len
)
4229 struct bpf_prog
*new_prog
;
4231 new_prog
= bpf_patch_insn_single(env
->prog
, off
, patch
, len
);
4234 if (adjust_insn_aux_data(env
, new_prog
->len
, off
, len
))
4239 /* The verifier does more data flow analysis than llvm and will not explore
4240 * branches that are dead at run time. Malicious programs can have dead code
4241 * too. Therefore replace all dead at-run-time code with nops.
4243 static void sanitize_dead_code(struct bpf_verifier_env
*env
)
4245 struct bpf_insn_aux_data
*aux_data
= env
->insn_aux_data
;
4246 struct bpf_insn nop
= BPF_MOV64_REG(BPF_REG_0
, BPF_REG_0
);
4247 struct bpf_insn
*insn
= env
->prog
->insnsi
;
4248 const int insn_cnt
= env
->prog
->len
;
4251 for (i
= 0; i
< insn_cnt
; i
++) {
4252 if (aux_data
[i
].seen
)
4254 memcpy(insn
+ i
, &nop
, sizeof(nop
));
4258 /* convert load instructions that access fields of 'struct __sk_buff'
4259 * into sequence of instructions that access fields of 'struct sk_buff'
4261 static int convert_ctx_accesses(struct bpf_verifier_env
*env
)
4263 const struct bpf_verifier_ops
*ops
= env
->ops
;
4264 int i
, cnt
, size
, ctx_field_size
, delta
= 0;
4265 const int insn_cnt
= env
->prog
->len
;
4266 struct bpf_insn insn_buf
[16], *insn
;
4267 struct bpf_prog
*new_prog
;
4268 enum bpf_access_type type
;
4269 bool is_narrower_load
;
4272 if (ops
->gen_prologue
) {
4273 cnt
= ops
->gen_prologue(insn_buf
, env
->seen_direct_write
,
4275 if (cnt
>= ARRAY_SIZE(insn_buf
)) {
4276 verbose(env
, "bpf verifier is misconfigured\n");
4279 new_prog
= bpf_patch_insn_data(env
, 0, insn_buf
, cnt
);
4283 env
->prog
= new_prog
;
4288 if (!ops
->convert_ctx_access
)
4291 insn
= env
->prog
->insnsi
+ delta
;
4293 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
4294 if (insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_B
) ||
4295 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_H
) ||
4296 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_W
) ||
4297 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_DW
))
4299 else if (insn
->code
== (BPF_STX
| BPF_MEM
| BPF_B
) ||
4300 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_H
) ||
4301 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_W
) ||
4302 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_DW
))
4307 if (env
->insn_aux_data
[i
+ delta
].ptr_type
!= PTR_TO_CTX
)
4310 ctx_field_size
= env
->insn_aux_data
[i
+ delta
].ctx_field_size
;
4311 size
= BPF_LDST_BYTES(insn
);
4313 /* If the read access is a narrower load of the field,
4314 * convert to a 4/8-byte load, to minimum program type specific
4315 * convert_ctx_access changes. If conversion is successful,
4316 * we will apply proper mask to the result.
4318 is_narrower_load
= size
< ctx_field_size
;
4319 if (is_narrower_load
) {
4320 u32 off
= insn
->off
;
4323 if (type
== BPF_WRITE
) {
4324 verbose(env
, "bpf verifier narrow ctx access misconfigured\n");
4329 if (ctx_field_size
== 4)
4331 else if (ctx_field_size
== 8)
4334 insn
->off
= off
& ~(ctx_field_size
- 1);
4335 insn
->code
= BPF_LDX
| BPF_MEM
| size_code
;
4339 cnt
= ops
->convert_ctx_access(type
, insn
, insn_buf
, env
->prog
,
4341 if (cnt
== 0 || cnt
>= ARRAY_SIZE(insn_buf
) ||
4342 (ctx_field_size
&& !target_size
)) {
4343 verbose(env
, "bpf verifier is misconfigured\n");
4347 if (is_narrower_load
&& size
< target_size
) {
4348 if (ctx_field_size
<= 4)
4349 insn_buf
[cnt
++] = BPF_ALU32_IMM(BPF_AND
, insn
->dst_reg
,
4350 (1 << size
* 8) - 1);
4352 insn_buf
[cnt
++] = BPF_ALU64_IMM(BPF_AND
, insn
->dst_reg
,
4353 (1 << size
* 8) - 1);
4356 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
, cnt
);
4362 /* keep walking new program and skip insns we just inserted */
4363 env
->prog
= new_prog
;
4364 insn
= new_prog
->insnsi
+ i
+ delta
;
4370 /* fixup insn->imm field of bpf_call instructions
4371 * and inline eligible helpers as explicit sequence of BPF instructions
4373 * this function is called after eBPF program passed verification
4375 static int fixup_bpf_calls(struct bpf_verifier_env
*env
)
4377 struct bpf_prog
*prog
= env
->prog
;
4378 struct bpf_insn
*insn
= prog
->insnsi
;
4379 const struct bpf_func_proto
*fn
;
4380 const int insn_cnt
= prog
->len
;
4381 struct bpf_insn insn_buf
[16];
4382 struct bpf_prog
*new_prog
;
4383 struct bpf_map
*map_ptr
;
4384 int i
, cnt
, delta
= 0;
4386 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
4387 if (insn
->code
!= (BPF_JMP
| BPF_CALL
))
4390 if (insn
->imm
== BPF_FUNC_get_route_realm
)
4391 prog
->dst_needed
= 1;
4392 if (insn
->imm
== BPF_FUNC_get_prandom_u32
)
4393 bpf_user_rnd_init_once();
4394 if (insn
->imm
== BPF_FUNC_tail_call
) {
4395 /* If we tail call into other programs, we
4396 * cannot make any assumptions since they can
4397 * be replaced dynamically during runtime in
4398 * the program array.
4400 prog
->cb_access
= 1;
4401 env
->prog
->aux
->stack_depth
= MAX_BPF_STACK
;
4403 /* mark bpf_tail_call as different opcode to avoid
4404 * conditional branch in the interpeter for every normal
4405 * call and to prevent accidental JITing by JIT compiler
4406 * that doesn't support bpf_tail_call yet
4409 insn
->code
= BPF_JMP
| BPF_TAIL_CALL
;
4413 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
4414 * handlers are currently limited to 64 bit only.
4416 if (ebpf_jit_enabled() && BITS_PER_LONG
== 64 &&
4417 insn
->imm
== BPF_FUNC_map_lookup_elem
) {
4418 map_ptr
= env
->insn_aux_data
[i
+ delta
].map_ptr
;
4419 if (map_ptr
== BPF_MAP_PTR_POISON
||
4420 !map_ptr
->ops
->map_gen_lookup
)
4421 goto patch_call_imm
;
4423 cnt
= map_ptr
->ops
->map_gen_lookup(map_ptr
, insn_buf
);
4424 if (cnt
== 0 || cnt
>= ARRAY_SIZE(insn_buf
)) {
4425 verbose(env
, "bpf verifier is misconfigured\n");
4429 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
,
4436 /* keep walking new program and skip insns we just inserted */
4437 env
->prog
= prog
= new_prog
;
4438 insn
= new_prog
->insnsi
+ i
+ delta
;
4442 if (insn
->imm
== BPF_FUNC_redirect_map
) {
4443 /* Note, we cannot use prog directly as imm as subsequent
4444 * rewrites would still change the prog pointer. The only
4445 * stable address we can use is aux, which also works with
4446 * prog clones during blinding.
4448 u64 addr
= (unsigned long)prog
->aux
;
4449 struct bpf_insn r4_ld
[] = {
4450 BPF_LD_IMM64(BPF_REG_4
, addr
),
4453 cnt
= ARRAY_SIZE(r4_ld
);
4455 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, r4_ld
, cnt
);
4460 env
->prog
= prog
= new_prog
;
4461 insn
= new_prog
->insnsi
+ i
+ delta
;
4464 fn
= env
->ops
->get_func_proto(insn
->imm
);
4465 /* all functions that have prototype and verifier allowed
4466 * programs to call them, must be real in-kernel functions
4470 "kernel subsystem misconfigured func %s#%d\n",
4471 func_id_name(insn
->imm
), insn
->imm
);
4474 insn
->imm
= fn
->func
- __bpf_call_base
;
4480 static void free_states(struct bpf_verifier_env
*env
)
4482 struct bpf_verifier_state_list
*sl
, *sln
;
4485 if (!env
->explored_states
)
4488 for (i
= 0; i
< env
->prog
->len
; i
++) {
4489 sl
= env
->explored_states
[i
];
4492 while (sl
!= STATE_LIST_MARK
) {
4494 free_verifier_state(&sl
->state
, false);
4500 kfree(env
->explored_states
);
4503 int bpf_check(struct bpf_prog
**prog
, union bpf_attr
*attr
)
4505 struct bpf_verifier_env
*env
;
4506 struct bpf_verifer_log
*log
;
4509 /* no program is valid */
4510 if (ARRAY_SIZE(bpf_verifier_ops
) == 0)
4513 /* 'struct bpf_verifier_env' can be global, but since it's not small,
4514 * allocate/free it every time bpf_check() is called
4516 env
= kzalloc(sizeof(struct bpf_verifier_env
), GFP_KERNEL
);
4521 env
->insn_aux_data
= vzalloc(sizeof(struct bpf_insn_aux_data
) *
4524 if (!env
->insn_aux_data
)
4527 env
->ops
= bpf_verifier_ops
[env
->prog
->type
];
4529 /* grab the mutex to protect few globals used by verifier */
4530 mutex_lock(&bpf_verifier_lock
);
4532 if (attr
->log_level
|| attr
->log_buf
|| attr
->log_size
) {
4533 /* user requested verbose verifier output
4534 * and supplied buffer to store the verification trace
4536 log
->level
= attr
->log_level
;
4537 log
->ubuf
= (char __user
*) (unsigned long) attr
->log_buf
;
4538 log
->len_total
= attr
->log_size
;
4541 /* log attributes have to be sane */
4542 if (log
->len_total
< 128 || log
->len_total
> UINT_MAX
>> 8 ||
4543 !log
->level
|| !log
->ubuf
)
4547 env
->strict_alignment
= !!(attr
->prog_flags
& BPF_F_STRICT_ALIGNMENT
);
4548 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
))
4549 env
->strict_alignment
= true;
4551 if (env
->prog
->aux
->offload
) {
4552 ret
= bpf_prog_offload_verifier_prep(env
);
4557 ret
= replace_map_fd_with_map_ptr(env
);
4559 goto skip_full_check
;
4561 env
->explored_states
= kcalloc(env
->prog
->len
,
4562 sizeof(struct bpf_verifier_state_list
*),
4565 if (!env
->explored_states
)
4566 goto skip_full_check
;
4568 ret
= check_cfg(env
);
4570 goto skip_full_check
;
4572 env
->allow_ptr_leaks
= capable(CAP_SYS_ADMIN
);
4574 ret
= do_check(env
);
4575 if (env
->cur_state
) {
4576 free_verifier_state(env
->cur_state
, true);
4577 env
->cur_state
= NULL
;
4581 while (!pop_stack(env
, NULL
, NULL
));
4585 sanitize_dead_code(env
);
4588 /* program is valid, convert *(u32*)(ctx + off) accesses */
4589 ret
= convert_ctx_accesses(env
);
4592 ret
= fixup_bpf_calls(env
);
4594 if (log
->level
&& bpf_verifier_log_full(log
))
4596 if (log
->level
&& !log
->ubuf
) {
4598 goto err_release_maps
;
4601 if (ret
== 0 && env
->used_map_cnt
) {
4602 /* if program passed verifier, update used_maps in bpf_prog_info */
4603 env
->prog
->aux
->used_maps
= kmalloc_array(env
->used_map_cnt
,
4604 sizeof(env
->used_maps
[0]),
4607 if (!env
->prog
->aux
->used_maps
) {
4609 goto err_release_maps
;
4612 memcpy(env
->prog
->aux
->used_maps
, env
->used_maps
,
4613 sizeof(env
->used_maps
[0]) * env
->used_map_cnt
);
4614 env
->prog
->aux
->used_map_cnt
= env
->used_map_cnt
;
4616 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
4617 * bpf_ld_imm64 instructions
4619 convert_pseudo_ld_imm64(env
);
4623 if (!env
->prog
->aux
->used_maps
)
4624 /* if we didn't copy map pointers into bpf_prog_info, release
4625 * them now. Otherwise free_bpf_prog_info() will release them.
4630 mutex_unlock(&bpf_verifier_lock
);
4631 vfree(env
->insn_aux_data
);