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 (-i
- 1) * 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
, int insn_idx
)
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 if (state
->stack
[spi
].slot_type
[i
] == STACK_MISC
&&
727 !env
->allow_ptr_leaks
) {
728 int *poff
= &env
->insn_aux_data
[insn_idx
].sanitize_stack_off
;
729 int soff
= (-spi
- 1) * BPF_REG_SIZE
;
731 /* detected reuse of integer stack slot with a pointer
732 * which means either llvm is reusing stack slot or
733 * an attacker is trying to exploit CVE-2018-3639
734 * (speculative store bypass)
735 * Have to sanitize that slot with preemptive
738 if (*poff
&& *poff
!= soff
) {
739 /* disallow programs where single insn stores
740 * into two different stack slots, since verifier
741 * cannot sanitize them
744 "insn %d cannot access two stack slots fp%d and fp%d",
745 insn_idx
, *poff
, soff
);
750 state
->stack
[spi
].slot_type
[i
] = STACK_SPILL
;
753 /* regular write of data into stack */
754 state
->stack
[spi
].spilled_ptr
= (struct bpf_reg_state
) {};
756 for (i
= 0; i
< size
; i
++)
757 state
->stack
[spi
].slot_type
[(slot
- i
) % BPF_REG_SIZE
] =
763 static void mark_stack_slot_read(const struct bpf_verifier_state
*state
, int slot
)
765 struct bpf_verifier_state
*parent
= state
->parent
;
768 /* if read wasn't screened by an earlier write ... */
769 if (state
->stack
[slot
].spilled_ptr
.live
& REG_LIVE_WRITTEN
)
771 /* ... then we depend on parent's value */
772 parent
->stack
[slot
].spilled_ptr
.live
|= REG_LIVE_READ
;
774 parent
= state
->parent
;
778 static int check_stack_read(struct bpf_verifier_env
*env
,
779 struct bpf_verifier_state
*state
, int off
, int size
,
782 int i
, slot
= -off
- 1, spi
= slot
/ BPF_REG_SIZE
;
785 if (state
->allocated_stack
<= slot
) {
786 verbose(env
, "invalid read from stack off %d+0 size %d\n",
790 stype
= state
->stack
[spi
].slot_type
;
792 if (stype
[0] == STACK_SPILL
) {
793 if (size
!= BPF_REG_SIZE
) {
794 verbose(env
, "invalid size of register spill\n");
797 for (i
= 1; i
< BPF_REG_SIZE
; i
++) {
798 if (stype
[(slot
- i
) % BPF_REG_SIZE
] != STACK_SPILL
) {
799 verbose(env
, "corrupted spill memory\n");
804 if (value_regno
>= 0) {
805 /* restore register state from stack */
806 state
->regs
[value_regno
] = state
->stack
[spi
].spilled_ptr
;
807 mark_stack_slot_read(state
, spi
);
811 for (i
= 0; i
< size
; i
++) {
812 if (stype
[(slot
- i
) % BPF_REG_SIZE
] != STACK_MISC
) {
813 verbose(env
, "invalid read from stack off %d+%d size %d\n",
818 if (value_regno
>= 0)
819 /* have read misc data from the stack */
820 mark_reg_unknown(env
, state
->regs
, value_regno
);
825 /* check read/write into map element returned by bpf_map_lookup_elem() */
826 static int __check_map_access(struct bpf_verifier_env
*env
, u32 regno
, int off
,
827 int size
, bool zero_size_allowed
)
829 struct bpf_reg_state
*regs
= cur_regs(env
);
830 struct bpf_map
*map
= regs
[regno
].map_ptr
;
832 if (off
< 0 || size
< 0 || (size
== 0 && !zero_size_allowed
) ||
833 off
+ size
> map
->value_size
) {
834 verbose(env
, "invalid access to map value, value_size=%d off=%d size=%d\n",
835 map
->value_size
, off
, size
);
841 /* check read/write into a map element with possible variable offset */
842 static int check_map_access(struct bpf_verifier_env
*env
, u32 regno
,
843 int off
, int size
, bool zero_size_allowed
)
845 struct bpf_verifier_state
*state
= env
->cur_state
;
846 struct bpf_reg_state
*reg
= &state
->regs
[regno
];
849 /* We may have adjusted the register to this map value, so we
850 * need to try adding each of min_value and max_value to off
851 * to make sure our theoretical access will be safe.
854 print_verifier_state(env
, state
);
855 /* The minimum value is only important with signed
856 * comparisons where we can't assume the floor of a
857 * value is 0. If we are using signed variables for our
858 * index'es we need to make sure that whatever we use
859 * will have a set floor within our range.
861 if (reg
->smin_value
< 0) {
862 verbose(env
, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
866 err
= __check_map_access(env
, regno
, reg
->smin_value
+ off
, size
,
869 verbose(env
, "R%d min value is outside of the array range\n",
874 /* If we haven't set a max value then we need to bail since we can't be
875 * sure we won't do bad things.
876 * If reg->umax_value + off could overflow, treat that as unbounded too.
878 if (reg
->umax_value
>= BPF_MAX_VAR_OFF
) {
879 verbose(env
, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
883 err
= __check_map_access(env
, regno
, reg
->umax_value
+ off
, size
,
886 verbose(env
, "R%d max value is outside of the array range\n",
891 #define MAX_PACKET_OFF 0xffff
893 static bool may_access_direct_pkt_data(struct bpf_verifier_env
*env
,
894 const struct bpf_call_arg_meta
*meta
,
895 enum bpf_access_type t
)
897 switch (env
->prog
->type
) {
898 case BPF_PROG_TYPE_LWT_IN
:
899 case BPF_PROG_TYPE_LWT_OUT
:
900 /* dst_input() and dst_output() can't write for now */
904 case BPF_PROG_TYPE_SCHED_CLS
:
905 case BPF_PROG_TYPE_SCHED_ACT
:
906 case BPF_PROG_TYPE_XDP
:
907 case BPF_PROG_TYPE_LWT_XMIT
:
908 case BPF_PROG_TYPE_SK_SKB
:
910 return meta
->pkt_access
;
912 env
->seen_direct_write
= true;
919 static int __check_packet_access(struct bpf_verifier_env
*env
, u32 regno
,
920 int off
, int size
, bool zero_size_allowed
)
922 struct bpf_reg_state
*regs
= cur_regs(env
);
923 struct bpf_reg_state
*reg
= ®s
[regno
];
925 if (off
< 0 || size
< 0 || (size
== 0 && !zero_size_allowed
) ||
926 (u64
)off
+ size
> reg
->range
) {
927 verbose(env
, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
928 off
, size
, regno
, reg
->id
, reg
->off
, reg
->range
);
934 static int check_packet_access(struct bpf_verifier_env
*env
, u32 regno
, int off
,
935 int size
, bool zero_size_allowed
)
937 struct bpf_reg_state
*regs
= cur_regs(env
);
938 struct bpf_reg_state
*reg
= ®s
[regno
];
941 /* We may have added a variable offset to the packet pointer; but any
942 * reg->range we have comes after that. We are only checking the fixed
946 /* We don't allow negative numbers, because we aren't tracking enough
947 * detail to prove they're safe.
949 if (reg
->smin_value
< 0) {
950 verbose(env
, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
954 err
= __check_packet_access(env
, regno
, off
, size
, zero_size_allowed
);
956 verbose(env
, "R%d offset is outside of the packet\n", regno
);
962 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
963 static int check_ctx_access(struct bpf_verifier_env
*env
, int insn_idx
, int off
, int size
,
964 enum bpf_access_type t
, enum bpf_reg_type
*reg_type
)
966 struct bpf_insn_access_aux info
= {
967 .reg_type
= *reg_type
,
970 if (env
->ops
->is_valid_access
&&
971 env
->ops
->is_valid_access(off
, size
, t
, &info
)) {
972 /* A non zero info.ctx_field_size indicates that this field is a
973 * candidate for later verifier transformation to load the whole
974 * field and then apply a mask when accessed with a narrower
975 * access than actual ctx access size. A zero info.ctx_field_size
976 * will only allow for whole field access and rejects any other
977 * type of narrower access.
979 *reg_type
= info
.reg_type
;
981 env
->insn_aux_data
[insn_idx
].ctx_field_size
= info
.ctx_field_size
;
982 /* remember the offset of last byte accessed in ctx */
983 if (env
->prog
->aux
->max_ctx_offset
< off
+ size
)
984 env
->prog
->aux
->max_ctx_offset
= off
+ size
;
988 verbose(env
, "invalid bpf_context access off=%d size=%d\n", off
, size
);
992 static bool __is_pointer_value(bool allow_ptr_leaks
,
993 const struct bpf_reg_state
*reg
)
998 return reg
->type
!= SCALAR_VALUE
;
1001 static bool is_pointer_value(struct bpf_verifier_env
*env
, int regno
)
1003 return __is_pointer_value(env
->allow_ptr_leaks
, cur_regs(env
) + regno
);
1006 static bool is_ctx_reg(struct bpf_verifier_env
*env
, int regno
)
1008 const struct bpf_reg_state
*reg
= cur_regs(env
) + regno
;
1010 return reg
->type
== PTR_TO_CTX
;
1013 static bool is_pkt_reg(struct bpf_verifier_env
*env
, int regno
)
1015 const struct bpf_reg_state
*reg
= cur_regs(env
) + regno
;
1017 return type_is_pkt_pointer(reg
->type
);
1020 static int check_pkt_ptr_alignment(struct bpf_verifier_env
*env
,
1021 const struct bpf_reg_state
*reg
,
1022 int off
, int size
, bool strict
)
1024 struct tnum reg_off
;
1027 /* Byte size accesses are always allowed. */
1028 if (!strict
|| size
== 1)
1031 /* For platforms that do not have a Kconfig enabling
1032 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
1033 * NET_IP_ALIGN is universally set to '2'. And on platforms
1034 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
1035 * to this code only in strict mode where we want to emulate
1036 * the NET_IP_ALIGN==2 checking. Therefore use an
1037 * unconditional IP align value of '2'.
1041 reg_off
= tnum_add(reg
->var_off
, tnum_const(ip_align
+ reg
->off
+ off
));
1042 if (!tnum_is_aligned(reg_off
, size
)) {
1045 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1047 "misaligned packet access off %d+%s+%d+%d size %d\n",
1048 ip_align
, tn_buf
, reg
->off
, off
, size
);
1055 static int check_generic_ptr_alignment(struct bpf_verifier_env
*env
,
1056 const struct bpf_reg_state
*reg
,
1057 const char *pointer_desc
,
1058 int off
, int size
, bool strict
)
1060 struct tnum reg_off
;
1062 /* Byte size accesses are always allowed. */
1063 if (!strict
|| size
== 1)
1066 reg_off
= tnum_add(reg
->var_off
, tnum_const(reg
->off
+ off
));
1067 if (!tnum_is_aligned(reg_off
, size
)) {
1070 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1071 verbose(env
, "misaligned %saccess off %s+%d+%d size %d\n",
1072 pointer_desc
, tn_buf
, reg
->off
, off
, size
);
1079 static int check_ptr_alignment(struct bpf_verifier_env
*env
,
1080 const struct bpf_reg_state
*reg
, int off
,
1081 int size
, bool strict_alignment_once
)
1083 bool strict
= env
->strict_alignment
|| strict_alignment_once
;
1084 const char *pointer_desc
= "";
1086 switch (reg
->type
) {
1088 case PTR_TO_PACKET_META
:
1089 /* Special case, because of NET_IP_ALIGN. Given metadata sits
1090 * right in front, treat it the very same way.
1092 return check_pkt_ptr_alignment(env
, reg
, off
, size
, strict
);
1093 case PTR_TO_MAP_VALUE
:
1094 pointer_desc
= "value ";
1097 pointer_desc
= "context ";
1100 pointer_desc
= "stack ";
1101 /* The stack spill tracking logic in check_stack_write()
1102 * and check_stack_read() relies on stack accesses being
1110 return check_generic_ptr_alignment(env
, reg
, pointer_desc
, off
, size
,
1114 /* truncate register to smaller size (in bytes)
1115 * must be called with size < BPF_REG_SIZE
1117 static void coerce_reg_to_size(struct bpf_reg_state
*reg
, int size
)
1121 /* clear high bits in bit representation */
1122 reg
->var_off
= tnum_cast(reg
->var_off
, size
);
1124 /* fix arithmetic bounds */
1125 mask
= ((u64
)1 << (size
* 8)) - 1;
1126 if ((reg
->umin_value
& ~mask
) == (reg
->umax_value
& ~mask
)) {
1127 reg
->umin_value
&= mask
;
1128 reg
->umax_value
&= mask
;
1130 reg
->umin_value
= 0;
1131 reg
->umax_value
= mask
;
1133 reg
->smin_value
= reg
->umin_value
;
1134 reg
->smax_value
= reg
->umax_value
;
1137 /* check whether memory at (regno + off) is accessible for t = (read | write)
1138 * if t==write, value_regno is a register which value is stored into memory
1139 * if t==read, value_regno is a register which will receive the value from memory
1140 * if t==write && value_regno==-1, some unknown value is stored into memory
1141 * if t==read && value_regno==-1, don't care what we read from memory
1143 static int check_mem_access(struct bpf_verifier_env
*env
, int insn_idx
, u32 regno
,
1144 int off
, int bpf_size
, enum bpf_access_type t
,
1145 int value_regno
, bool strict_alignment_once
)
1147 struct bpf_verifier_state
*state
= env
->cur_state
;
1148 struct bpf_reg_state
*regs
= cur_regs(env
);
1149 struct bpf_reg_state
*reg
= regs
+ regno
;
1152 size
= bpf_size_to_bytes(bpf_size
);
1156 /* alignment checks will add in reg->off themselves */
1157 err
= check_ptr_alignment(env
, reg
, off
, size
, strict_alignment_once
);
1161 /* for access checks, reg->off is just part of off */
1164 if (reg
->type
== PTR_TO_MAP_VALUE
) {
1165 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
1166 is_pointer_value(env
, value_regno
)) {
1167 verbose(env
, "R%d leaks addr into map\n", value_regno
);
1171 err
= check_map_access(env
, regno
, off
, size
, false);
1172 if (!err
&& t
== BPF_READ
&& value_regno
>= 0)
1173 mark_reg_unknown(env
, regs
, value_regno
);
1175 } else if (reg
->type
== PTR_TO_CTX
) {
1176 enum bpf_reg_type reg_type
= SCALAR_VALUE
;
1178 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
1179 is_pointer_value(env
, value_regno
)) {
1180 verbose(env
, "R%d leaks addr into ctx\n", value_regno
);
1183 /* ctx accesses must be at a fixed offset, so that we can
1184 * determine what type of data were returned.
1188 "dereference of modified ctx ptr R%d off=%d+%d, ctx+const is allowed, ctx+const+const is not\n",
1189 regno
, reg
->off
, off
- reg
->off
);
1192 if (!tnum_is_const(reg
->var_off
) || reg
->var_off
.value
) {
1195 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1197 "variable ctx access var_off=%s off=%d size=%d",
1201 err
= check_ctx_access(env
, insn_idx
, off
, size
, t
, ®_type
);
1202 if (!err
&& t
== BPF_READ
&& value_regno
>= 0) {
1203 /* ctx access returns either a scalar, or a
1204 * PTR_TO_PACKET[_META,_END]. In the latter
1205 * case, we know the offset is zero.
1207 if (reg_type
== SCALAR_VALUE
)
1208 mark_reg_unknown(env
, regs
, value_regno
);
1210 mark_reg_known_zero(env
, regs
,
1212 regs
[value_regno
].id
= 0;
1213 regs
[value_regno
].off
= 0;
1214 regs
[value_regno
].range
= 0;
1215 regs
[value_regno
].type
= reg_type
;
1218 } else if (reg
->type
== PTR_TO_STACK
) {
1219 /* stack accesses must be at a fixed offset, so that we can
1220 * determine what type of data were returned.
1221 * See check_stack_read().
1223 if (!tnum_is_const(reg
->var_off
)) {
1226 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1227 verbose(env
, "variable stack access var_off=%s off=%d size=%d",
1231 off
+= reg
->var_off
.value
;
1232 if (off
>= 0 || off
< -MAX_BPF_STACK
) {
1233 verbose(env
, "invalid stack off=%d size=%d\n", off
,
1238 if (env
->prog
->aux
->stack_depth
< -off
)
1239 env
->prog
->aux
->stack_depth
= -off
;
1242 err
= check_stack_write(env
, state
, off
, size
,
1243 value_regno
, insn_idx
);
1245 err
= check_stack_read(env
, state
, off
, size
,
1247 } else if (reg_is_pkt_pointer(reg
)) {
1248 if (t
== BPF_WRITE
&& !may_access_direct_pkt_data(env
, NULL
, t
)) {
1249 verbose(env
, "cannot write into packet\n");
1252 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
1253 is_pointer_value(env
, value_regno
)) {
1254 verbose(env
, "R%d leaks addr into packet\n",
1258 err
= check_packet_access(env
, regno
, off
, size
, false);
1259 if (!err
&& t
== BPF_READ
&& value_regno
>= 0)
1260 mark_reg_unknown(env
, regs
, value_regno
);
1262 verbose(env
, "R%d invalid mem access '%s'\n", regno
,
1263 reg_type_str
[reg
->type
]);
1267 if (!err
&& size
< BPF_REG_SIZE
&& value_regno
>= 0 && t
== BPF_READ
&&
1268 regs
[value_regno
].type
== SCALAR_VALUE
) {
1269 /* b/h/w load zero-extends, mark upper bits as known 0 */
1270 coerce_reg_to_size(®s
[value_regno
], size
);
1275 static int check_xadd(struct bpf_verifier_env
*env
, int insn_idx
, struct bpf_insn
*insn
)
1279 if ((BPF_SIZE(insn
->code
) != BPF_W
&& BPF_SIZE(insn
->code
) != BPF_DW
) ||
1281 verbose(env
, "BPF_XADD uses reserved fields\n");
1285 /* check src1 operand */
1286 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
1290 /* check src2 operand */
1291 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
1295 if (is_pointer_value(env
, insn
->src_reg
)) {
1296 verbose(env
, "R%d leaks addr into mem\n", insn
->src_reg
);
1300 if (is_ctx_reg(env
, insn
->dst_reg
) ||
1301 is_pkt_reg(env
, insn
->dst_reg
)) {
1302 verbose(env
, "BPF_XADD stores into R%d %s is not allowed\n",
1303 insn
->dst_reg
, is_ctx_reg(env
, insn
->dst_reg
) ?
1304 "context" : "packet");
1308 /* check whether atomic_add can read the memory */
1309 err
= check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
1310 BPF_SIZE(insn
->code
), BPF_READ
, -1, true);
1314 /* check whether atomic_add can write into the same memory */
1315 return check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
1316 BPF_SIZE(insn
->code
), BPF_WRITE
, -1, true);
1319 /* Does this register contain a constant zero? */
1320 static bool register_is_null(struct bpf_reg_state reg
)
1322 return reg
.type
== SCALAR_VALUE
&& tnum_equals_const(reg
.var_off
, 0);
1325 /* when register 'regno' is passed into function that will read 'access_size'
1326 * bytes from that pointer, make sure that it's within stack boundary
1327 * and all elements of stack are initialized.
1328 * Unlike most pointer bounds-checking functions, this one doesn't take an
1329 * 'off' argument, so it has to add in reg->off itself.
1331 static int check_stack_boundary(struct bpf_verifier_env
*env
, int regno
,
1332 int access_size
, bool zero_size_allowed
,
1333 struct bpf_call_arg_meta
*meta
)
1335 struct bpf_verifier_state
*state
= env
->cur_state
;
1336 struct bpf_reg_state
*regs
= state
->regs
;
1337 int off
, i
, slot
, spi
;
1339 if (regs
[regno
].type
!= PTR_TO_STACK
) {
1340 /* Allow zero-byte read from NULL, regardless of pointer type */
1341 if (zero_size_allowed
&& access_size
== 0 &&
1342 register_is_null(regs
[regno
]))
1345 verbose(env
, "R%d type=%s expected=%s\n", regno
,
1346 reg_type_str
[regs
[regno
].type
],
1347 reg_type_str
[PTR_TO_STACK
]);
1351 /* Only allow fixed-offset stack reads */
1352 if (!tnum_is_const(regs
[regno
].var_off
)) {
1355 tnum_strn(tn_buf
, sizeof(tn_buf
), regs
[regno
].var_off
);
1356 verbose(env
, "invalid variable stack read R%d var_off=%s\n",
1360 off
= regs
[regno
].off
+ regs
[regno
].var_off
.value
;
1361 if (off
>= 0 || off
< -MAX_BPF_STACK
|| off
+ access_size
> 0 ||
1362 access_size
< 0 || (access_size
== 0 && !zero_size_allowed
)) {
1363 verbose(env
, "invalid stack type R%d off=%d access_size=%d\n",
1364 regno
, off
, access_size
);
1368 if (env
->prog
->aux
->stack_depth
< -off
)
1369 env
->prog
->aux
->stack_depth
= -off
;
1371 if (meta
&& meta
->raw_mode
) {
1372 meta
->access_size
= access_size
;
1373 meta
->regno
= regno
;
1377 for (i
= 0; i
< access_size
; i
++) {
1378 slot
= -(off
+ i
) - 1;
1379 spi
= slot
/ BPF_REG_SIZE
;
1380 if (state
->allocated_stack
<= slot
||
1381 state
->stack
[spi
].slot_type
[slot
% BPF_REG_SIZE
] !=
1383 verbose(env
, "invalid indirect read from stack off %d+%d size %d\n",
1384 off
, i
, access_size
);
1391 static int check_helper_mem_access(struct bpf_verifier_env
*env
, int regno
,
1392 int access_size
, bool zero_size_allowed
,
1393 struct bpf_call_arg_meta
*meta
)
1395 struct bpf_reg_state
*regs
= cur_regs(env
), *reg
= ®s
[regno
];
1397 switch (reg
->type
) {
1399 case PTR_TO_PACKET_META
:
1400 return check_packet_access(env
, regno
, reg
->off
, access_size
,
1402 case PTR_TO_MAP_VALUE
:
1403 return check_map_access(env
, regno
, reg
->off
, access_size
,
1405 default: /* scalar_value|ptr_to_stack or invalid ptr */
1406 return check_stack_boundary(env
, regno
, access_size
,
1407 zero_size_allowed
, meta
);
1411 static int check_func_arg(struct bpf_verifier_env
*env
, u32 regno
,
1412 enum bpf_arg_type arg_type
,
1413 struct bpf_call_arg_meta
*meta
)
1415 struct bpf_reg_state
*regs
= cur_regs(env
), *reg
= ®s
[regno
];
1416 enum bpf_reg_type expected_type
, type
= reg
->type
;
1419 if (arg_type
== ARG_DONTCARE
)
1422 err
= check_reg_arg(env
, regno
, SRC_OP
);
1426 if (arg_type
== ARG_ANYTHING
) {
1427 if (is_pointer_value(env
, regno
)) {
1428 verbose(env
, "R%d leaks addr into helper function\n",
1435 if (type_is_pkt_pointer(type
) &&
1436 !may_access_direct_pkt_data(env
, meta
, BPF_READ
)) {
1437 verbose(env
, "helper access to the packet is not allowed\n");
1441 if (arg_type
== ARG_PTR_TO_MAP_KEY
||
1442 arg_type
== ARG_PTR_TO_MAP_VALUE
) {
1443 expected_type
= PTR_TO_STACK
;
1444 if (!type_is_pkt_pointer(type
) &&
1445 type
!= expected_type
)
1447 } else if (arg_type
== ARG_CONST_SIZE
||
1448 arg_type
== ARG_CONST_SIZE_OR_ZERO
) {
1449 expected_type
= SCALAR_VALUE
;
1450 if (type
!= expected_type
)
1452 } else if (arg_type
== ARG_CONST_MAP_PTR
) {
1453 expected_type
= CONST_PTR_TO_MAP
;
1454 if (type
!= expected_type
)
1456 } else if (arg_type
== ARG_PTR_TO_CTX
) {
1457 expected_type
= PTR_TO_CTX
;
1458 if (type
!= expected_type
)
1460 } else if (arg_type
== ARG_PTR_TO_MEM
||
1461 arg_type
== ARG_PTR_TO_MEM_OR_NULL
||
1462 arg_type
== ARG_PTR_TO_UNINIT_MEM
) {
1463 expected_type
= PTR_TO_STACK
;
1464 /* One exception here. In case function allows for NULL to be
1465 * passed in as argument, it's a SCALAR_VALUE type. Final test
1466 * happens during stack boundary checking.
1468 if (register_is_null(*reg
) &&
1469 arg_type
== ARG_PTR_TO_MEM_OR_NULL
)
1470 /* final test in check_stack_boundary() */;
1471 else if (!type_is_pkt_pointer(type
) &&
1472 type
!= PTR_TO_MAP_VALUE
&&
1473 type
!= expected_type
)
1475 meta
->raw_mode
= arg_type
== ARG_PTR_TO_UNINIT_MEM
;
1477 verbose(env
, "unsupported arg_type %d\n", arg_type
);
1481 if (arg_type
== ARG_CONST_MAP_PTR
) {
1482 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
1483 meta
->map_ptr
= reg
->map_ptr
;
1484 } else if (arg_type
== ARG_PTR_TO_MAP_KEY
) {
1485 /* bpf_map_xxx(..., map_ptr, ..., key) call:
1486 * check that [key, key + map->key_size) are within
1487 * stack limits and initialized
1489 if (!meta
->map_ptr
) {
1490 /* in function declaration map_ptr must come before
1491 * map_key, so that it's verified and known before
1492 * we have to check map_key here. Otherwise it means
1493 * that kernel subsystem misconfigured verifier
1495 verbose(env
, "invalid map_ptr to access map->key\n");
1498 if (type_is_pkt_pointer(type
))
1499 err
= check_packet_access(env
, regno
, reg
->off
,
1500 meta
->map_ptr
->key_size
,
1503 err
= check_stack_boundary(env
, regno
,
1504 meta
->map_ptr
->key_size
,
1506 } else if (arg_type
== ARG_PTR_TO_MAP_VALUE
) {
1507 /* bpf_map_xxx(..., map_ptr, ..., value) call:
1508 * check [value, value + map->value_size) validity
1510 if (!meta
->map_ptr
) {
1511 /* kernel subsystem misconfigured verifier */
1512 verbose(env
, "invalid map_ptr to access map->value\n");
1515 if (type_is_pkt_pointer(type
))
1516 err
= check_packet_access(env
, regno
, reg
->off
,
1517 meta
->map_ptr
->value_size
,
1520 err
= check_stack_boundary(env
, regno
,
1521 meta
->map_ptr
->value_size
,
1523 } else if (arg_type
== ARG_CONST_SIZE
||
1524 arg_type
== ARG_CONST_SIZE_OR_ZERO
) {
1525 bool zero_size_allowed
= (arg_type
== ARG_CONST_SIZE_OR_ZERO
);
1527 /* bpf_xxx(..., buf, len) call will access 'len' bytes
1528 * from stack pointer 'buf'. Check it
1529 * note: regno == len, regno - 1 == buf
1532 /* kernel subsystem misconfigured verifier */
1534 "ARG_CONST_SIZE cannot be first argument\n");
1538 /* The register is SCALAR_VALUE; the access check
1539 * happens using its boundaries.
1542 if (!tnum_is_const(reg
->var_off
))
1543 /* For unprivileged variable accesses, disable raw
1544 * mode so that the program is required to
1545 * initialize all the memory that the helper could
1546 * just partially fill up.
1550 if (reg
->smin_value
< 0) {
1551 verbose(env
, "R%d min value is negative, either use unsigned or 'var &= const'\n",
1556 if (reg
->umin_value
== 0) {
1557 err
= check_helper_mem_access(env
, regno
- 1, 0,
1564 if (reg
->umax_value
>= BPF_MAX_VAR_SIZ
) {
1565 verbose(env
, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
1569 err
= check_helper_mem_access(env
, regno
- 1,
1571 zero_size_allowed
, meta
);
1576 verbose(env
, "R%d type=%s expected=%s\n", regno
,
1577 reg_type_str
[type
], reg_type_str
[expected_type
]);
1581 static int check_map_func_compatibility(struct bpf_verifier_env
*env
,
1582 struct bpf_map
*map
, int func_id
)
1587 /* We need a two way check, first is from map perspective ... */
1588 switch (map
->map_type
) {
1589 case BPF_MAP_TYPE_PROG_ARRAY
:
1590 if (func_id
!= BPF_FUNC_tail_call
)
1593 case BPF_MAP_TYPE_PERF_EVENT_ARRAY
:
1594 if (func_id
!= BPF_FUNC_perf_event_read
&&
1595 func_id
!= BPF_FUNC_perf_event_output
&&
1596 func_id
!= BPF_FUNC_perf_event_read_value
)
1599 case BPF_MAP_TYPE_STACK_TRACE
:
1600 if (func_id
!= BPF_FUNC_get_stackid
)
1603 case BPF_MAP_TYPE_CGROUP_ARRAY
:
1604 if (func_id
!= BPF_FUNC_skb_under_cgroup
&&
1605 func_id
!= BPF_FUNC_current_task_under_cgroup
)
1608 /* devmap returns a pointer to a live net_device ifindex that we cannot
1609 * allow to be modified from bpf side. So do not allow lookup elements
1612 case BPF_MAP_TYPE_DEVMAP
:
1613 if (func_id
!= BPF_FUNC_redirect_map
)
1616 /* Restrict bpf side of cpumap, open when use-cases appear */
1617 case BPF_MAP_TYPE_CPUMAP
:
1618 if (func_id
!= BPF_FUNC_redirect_map
)
1621 case BPF_MAP_TYPE_ARRAY_OF_MAPS
:
1622 case BPF_MAP_TYPE_HASH_OF_MAPS
:
1623 if (func_id
!= BPF_FUNC_map_lookup_elem
)
1626 case BPF_MAP_TYPE_SOCKMAP
:
1627 if (func_id
!= BPF_FUNC_sk_redirect_map
&&
1628 func_id
!= BPF_FUNC_sock_map_update
&&
1629 func_id
!= BPF_FUNC_map_delete_elem
)
1636 /* ... and second from the function itself. */
1638 case BPF_FUNC_tail_call
:
1639 if (map
->map_type
!= BPF_MAP_TYPE_PROG_ARRAY
)
1642 case BPF_FUNC_perf_event_read
:
1643 case BPF_FUNC_perf_event_output
:
1644 case BPF_FUNC_perf_event_read_value
:
1645 if (map
->map_type
!= BPF_MAP_TYPE_PERF_EVENT_ARRAY
)
1648 case BPF_FUNC_get_stackid
:
1649 if (map
->map_type
!= BPF_MAP_TYPE_STACK_TRACE
)
1652 case BPF_FUNC_current_task_under_cgroup
:
1653 case BPF_FUNC_skb_under_cgroup
:
1654 if (map
->map_type
!= BPF_MAP_TYPE_CGROUP_ARRAY
)
1657 case BPF_FUNC_redirect_map
:
1658 if (map
->map_type
!= BPF_MAP_TYPE_DEVMAP
&&
1659 map
->map_type
!= BPF_MAP_TYPE_CPUMAP
)
1662 case BPF_FUNC_sk_redirect_map
:
1663 if (map
->map_type
!= BPF_MAP_TYPE_SOCKMAP
)
1666 case BPF_FUNC_sock_map_update
:
1667 if (map
->map_type
!= BPF_MAP_TYPE_SOCKMAP
)
1676 verbose(env
, "cannot pass map_type %d into func %s#%d\n",
1677 map
->map_type
, func_id_name(func_id
), func_id
);
1681 static int check_raw_mode(const struct bpf_func_proto
*fn
)
1685 if (fn
->arg1_type
== ARG_PTR_TO_UNINIT_MEM
)
1687 if (fn
->arg2_type
== ARG_PTR_TO_UNINIT_MEM
)
1689 if (fn
->arg3_type
== ARG_PTR_TO_UNINIT_MEM
)
1691 if (fn
->arg4_type
== ARG_PTR_TO_UNINIT_MEM
)
1693 if (fn
->arg5_type
== ARG_PTR_TO_UNINIT_MEM
)
1696 return count
> 1 ? -EINVAL
: 0;
1699 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
1700 * are now invalid, so turn them into unknown SCALAR_VALUE.
1702 static void clear_all_pkt_pointers(struct bpf_verifier_env
*env
)
1704 struct bpf_verifier_state
*state
= env
->cur_state
;
1705 struct bpf_reg_state
*regs
= state
->regs
, *reg
;
1708 for (i
= 0; i
< MAX_BPF_REG
; i
++)
1709 if (reg_is_pkt_pointer_any(®s
[i
]))
1710 mark_reg_unknown(env
, regs
, i
);
1712 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
1713 if (state
->stack
[i
].slot_type
[0] != STACK_SPILL
)
1715 reg
= &state
->stack
[i
].spilled_ptr
;
1716 if (reg_is_pkt_pointer_any(reg
))
1717 __mark_reg_unknown(reg
);
1721 static int check_call(struct bpf_verifier_env
*env
, int func_id
, int insn_idx
)
1723 const struct bpf_func_proto
*fn
= NULL
;
1724 struct bpf_reg_state
*regs
;
1725 struct bpf_call_arg_meta meta
;
1729 /* find function prototype */
1730 if (func_id
< 0 || func_id
>= __BPF_FUNC_MAX_ID
) {
1731 verbose(env
, "invalid func %s#%d\n", func_id_name(func_id
),
1736 if (env
->ops
->get_func_proto
)
1737 fn
= env
->ops
->get_func_proto(func_id
);
1740 verbose(env
, "unknown func %s#%d\n", func_id_name(func_id
),
1745 /* eBPF programs must be GPL compatible to use GPL-ed functions */
1746 if (!env
->prog
->gpl_compatible
&& fn
->gpl_only
) {
1747 verbose(env
, "cannot call GPL only function from proprietary program\n");
1751 /* With LD_ABS/IND some JITs save/restore skb from r1. */
1752 changes_data
= bpf_helper_changes_pkt_data(fn
->func
);
1753 if (changes_data
&& fn
->arg1_type
!= ARG_PTR_TO_CTX
) {
1754 verbose(env
, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
1755 func_id_name(func_id
), func_id
);
1759 memset(&meta
, 0, sizeof(meta
));
1760 meta
.pkt_access
= fn
->pkt_access
;
1762 /* We only support one arg being in raw mode at the moment, which
1763 * is sufficient for the helper functions we have right now.
1765 err
= check_raw_mode(fn
);
1767 verbose(env
, "kernel subsystem misconfigured func %s#%d\n",
1768 func_id_name(func_id
), func_id
);
1773 err
= check_func_arg(env
, BPF_REG_1
, fn
->arg1_type
, &meta
);
1776 err
= check_func_arg(env
, BPF_REG_2
, fn
->arg2_type
, &meta
);
1779 if (func_id
== BPF_FUNC_tail_call
) {
1780 if (meta
.map_ptr
== NULL
) {
1781 verbose(env
, "verifier bug\n");
1784 env
->insn_aux_data
[insn_idx
].map_ptr
= meta
.map_ptr
;
1786 err
= check_func_arg(env
, BPF_REG_3
, fn
->arg3_type
, &meta
);
1789 err
= check_func_arg(env
, BPF_REG_4
, fn
->arg4_type
, &meta
);
1792 err
= check_func_arg(env
, BPF_REG_5
, fn
->arg5_type
, &meta
);
1796 /* Mark slots with STACK_MISC in case of raw mode, stack offset
1797 * is inferred from register state.
1799 for (i
= 0; i
< meta
.access_size
; i
++) {
1800 err
= check_mem_access(env
, insn_idx
, meta
.regno
, i
, BPF_B
,
1801 BPF_WRITE
, -1, false);
1806 regs
= cur_regs(env
);
1807 /* reset caller saved regs */
1808 for (i
= 0; i
< CALLER_SAVED_REGS
; i
++) {
1809 mark_reg_not_init(env
, regs
, caller_saved
[i
]);
1810 check_reg_arg(env
, caller_saved
[i
], DST_OP_NO_MARK
);
1813 /* update return register (already marked as written above) */
1814 if (fn
->ret_type
== RET_INTEGER
) {
1815 /* sets type to SCALAR_VALUE */
1816 mark_reg_unknown(env
, regs
, BPF_REG_0
);
1817 } else if (fn
->ret_type
== RET_VOID
) {
1818 regs
[BPF_REG_0
].type
= NOT_INIT
;
1819 } else if (fn
->ret_type
== RET_PTR_TO_MAP_VALUE_OR_NULL
) {
1820 struct bpf_insn_aux_data
*insn_aux
;
1822 regs
[BPF_REG_0
].type
= PTR_TO_MAP_VALUE_OR_NULL
;
1823 /* There is no offset yet applied, variable or fixed */
1824 mark_reg_known_zero(env
, regs
, BPF_REG_0
);
1825 regs
[BPF_REG_0
].off
= 0;
1826 /* remember map_ptr, so that check_map_access()
1827 * can check 'value_size' boundary of memory access
1828 * to map element returned from bpf_map_lookup_elem()
1830 if (meta
.map_ptr
== NULL
) {
1832 "kernel subsystem misconfigured verifier\n");
1835 regs
[BPF_REG_0
].map_ptr
= meta
.map_ptr
;
1836 regs
[BPF_REG_0
].id
= ++env
->id_gen
;
1837 insn_aux
= &env
->insn_aux_data
[insn_idx
];
1838 if (!insn_aux
->map_ptr
)
1839 insn_aux
->map_ptr
= meta
.map_ptr
;
1840 else if (insn_aux
->map_ptr
!= meta
.map_ptr
)
1841 insn_aux
->map_ptr
= BPF_MAP_PTR_POISON
;
1843 verbose(env
, "unknown return type %d of func %s#%d\n",
1844 fn
->ret_type
, func_id_name(func_id
), func_id
);
1848 err
= check_map_func_compatibility(env
, meta
.map_ptr
, func_id
);
1853 clear_all_pkt_pointers(env
);
1857 static bool signed_add_overflows(s64 a
, s64 b
)
1859 /* Do the add in u64, where overflow is well-defined */
1860 s64 res
= (s64
)((u64
)a
+ (u64
)b
);
1867 static bool signed_sub_overflows(s64 a
, s64 b
)
1869 /* Do the sub in u64, where overflow is well-defined */
1870 s64 res
= (s64
)((u64
)a
- (u64
)b
);
1877 static bool check_reg_sane_offset(struct bpf_verifier_env
*env
,
1878 const struct bpf_reg_state
*reg
,
1879 enum bpf_reg_type type
)
1881 bool known
= tnum_is_const(reg
->var_off
);
1882 s64 val
= reg
->var_off
.value
;
1883 s64 smin
= reg
->smin_value
;
1885 if (known
&& (val
>= BPF_MAX_VAR_OFF
|| val
<= -BPF_MAX_VAR_OFF
)) {
1886 verbose(env
, "math between %s pointer and %lld is not allowed\n",
1887 reg_type_str
[type
], val
);
1891 if (reg
->off
>= BPF_MAX_VAR_OFF
|| reg
->off
<= -BPF_MAX_VAR_OFF
) {
1892 verbose(env
, "%s pointer offset %d is not allowed\n",
1893 reg_type_str
[type
], reg
->off
);
1897 if (smin
== S64_MIN
) {
1898 verbose(env
, "math between %s pointer and register with unbounded min value is not allowed\n",
1899 reg_type_str
[type
]);
1903 if (smin
>= BPF_MAX_VAR_OFF
|| smin
<= -BPF_MAX_VAR_OFF
) {
1904 verbose(env
, "value %lld makes %s pointer be out of bounds\n",
1905 smin
, reg_type_str
[type
]);
1912 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
1913 * Caller should also handle BPF_MOV case separately.
1914 * If we return -EACCES, caller may want to try again treating pointer as a
1915 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
1917 static int adjust_ptr_min_max_vals(struct bpf_verifier_env
*env
,
1918 struct bpf_insn
*insn
,
1919 const struct bpf_reg_state
*ptr_reg
,
1920 const struct bpf_reg_state
*off_reg
)
1922 struct bpf_reg_state
*regs
= cur_regs(env
), *dst_reg
;
1923 bool known
= tnum_is_const(off_reg
->var_off
);
1924 s64 smin_val
= off_reg
->smin_value
, smax_val
= off_reg
->smax_value
,
1925 smin_ptr
= ptr_reg
->smin_value
, smax_ptr
= ptr_reg
->smax_value
;
1926 u64 umin_val
= off_reg
->umin_value
, umax_val
= off_reg
->umax_value
,
1927 umin_ptr
= ptr_reg
->umin_value
, umax_ptr
= ptr_reg
->umax_value
;
1928 u8 opcode
= BPF_OP(insn
->code
);
1929 u32 dst
= insn
->dst_reg
;
1931 dst_reg
= ®s
[dst
];
1933 if ((known
&& (smin_val
!= smax_val
|| umin_val
!= umax_val
)) ||
1934 smin_val
> smax_val
|| umin_val
> umax_val
) {
1935 /* Taint dst register if offset had invalid bounds derived from
1936 * e.g. dead branches.
1938 __mark_reg_unknown(dst_reg
);
1942 if (BPF_CLASS(insn
->code
) != BPF_ALU64
) {
1943 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
1945 "R%d 32-bit pointer arithmetic prohibited\n",
1950 if (ptr_reg
->type
== PTR_TO_MAP_VALUE_OR_NULL
) {
1951 verbose(env
, "R%d pointer arithmetic on PTR_TO_MAP_VALUE_OR_NULL prohibited, null-check it first\n",
1955 if (ptr_reg
->type
== CONST_PTR_TO_MAP
) {
1956 verbose(env
, "R%d pointer arithmetic on CONST_PTR_TO_MAP prohibited\n",
1960 if (ptr_reg
->type
== PTR_TO_PACKET_END
) {
1961 verbose(env
, "R%d pointer arithmetic on PTR_TO_PACKET_END prohibited\n",
1966 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
1967 * The id may be overwritten later if we create a new variable offset.
1969 dst_reg
->type
= ptr_reg
->type
;
1970 dst_reg
->id
= ptr_reg
->id
;
1972 if (!check_reg_sane_offset(env
, off_reg
, ptr_reg
->type
) ||
1973 !check_reg_sane_offset(env
, ptr_reg
, ptr_reg
->type
))
1978 /* We can take a fixed offset as long as it doesn't overflow
1979 * the s32 'off' field
1981 if (known
&& (ptr_reg
->off
+ smin_val
==
1982 (s64
)(s32
)(ptr_reg
->off
+ smin_val
))) {
1983 /* pointer += K. Accumulate it into fixed offset */
1984 dst_reg
->smin_value
= smin_ptr
;
1985 dst_reg
->smax_value
= smax_ptr
;
1986 dst_reg
->umin_value
= umin_ptr
;
1987 dst_reg
->umax_value
= umax_ptr
;
1988 dst_reg
->var_off
= ptr_reg
->var_off
;
1989 dst_reg
->off
= ptr_reg
->off
+ smin_val
;
1990 dst_reg
->range
= ptr_reg
->range
;
1993 /* A new variable offset is created. Note that off_reg->off
1994 * == 0, since it's a scalar.
1995 * dst_reg gets the pointer type and since some positive
1996 * integer value was added to the pointer, give it a new 'id'
1997 * if it's a PTR_TO_PACKET.
1998 * this creates a new 'base' pointer, off_reg (variable) gets
1999 * added into the variable offset, and we copy the fixed offset
2002 if (signed_add_overflows(smin_ptr
, smin_val
) ||
2003 signed_add_overflows(smax_ptr
, smax_val
)) {
2004 dst_reg
->smin_value
= S64_MIN
;
2005 dst_reg
->smax_value
= S64_MAX
;
2007 dst_reg
->smin_value
= smin_ptr
+ smin_val
;
2008 dst_reg
->smax_value
= smax_ptr
+ smax_val
;
2010 if (umin_ptr
+ umin_val
< umin_ptr
||
2011 umax_ptr
+ umax_val
< umax_ptr
) {
2012 dst_reg
->umin_value
= 0;
2013 dst_reg
->umax_value
= U64_MAX
;
2015 dst_reg
->umin_value
= umin_ptr
+ umin_val
;
2016 dst_reg
->umax_value
= umax_ptr
+ umax_val
;
2018 dst_reg
->var_off
= tnum_add(ptr_reg
->var_off
, off_reg
->var_off
);
2019 dst_reg
->off
= ptr_reg
->off
;
2020 if (reg_is_pkt_pointer(ptr_reg
)) {
2021 dst_reg
->id
= ++env
->id_gen
;
2022 /* something was added to pkt_ptr, set range to zero */
2027 if (dst_reg
== off_reg
) {
2028 /* scalar -= pointer. Creates an unknown scalar */
2029 verbose(env
, "R%d tried to subtract pointer from scalar\n",
2033 /* We don't allow subtraction from FP, because (according to
2034 * test_verifier.c test "invalid fp arithmetic", JITs might not
2035 * be able to deal with it.
2037 if (ptr_reg
->type
== PTR_TO_STACK
) {
2038 verbose(env
, "R%d subtraction from stack pointer prohibited\n",
2042 if (known
&& (ptr_reg
->off
- smin_val
==
2043 (s64
)(s32
)(ptr_reg
->off
- smin_val
))) {
2044 /* pointer -= K. Subtract it from fixed offset */
2045 dst_reg
->smin_value
= smin_ptr
;
2046 dst_reg
->smax_value
= smax_ptr
;
2047 dst_reg
->umin_value
= umin_ptr
;
2048 dst_reg
->umax_value
= umax_ptr
;
2049 dst_reg
->var_off
= ptr_reg
->var_off
;
2050 dst_reg
->id
= ptr_reg
->id
;
2051 dst_reg
->off
= ptr_reg
->off
- smin_val
;
2052 dst_reg
->range
= ptr_reg
->range
;
2055 /* A new variable offset is created. If the subtrahend is known
2056 * nonnegative, then any reg->range we had before is still good.
2058 if (signed_sub_overflows(smin_ptr
, smax_val
) ||
2059 signed_sub_overflows(smax_ptr
, smin_val
)) {
2060 /* Overflow possible, we know nothing */
2061 dst_reg
->smin_value
= S64_MIN
;
2062 dst_reg
->smax_value
= S64_MAX
;
2064 dst_reg
->smin_value
= smin_ptr
- smax_val
;
2065 dst_reg
->smax_value
= smax_ptr
- smin_val
;
2067 if (umin_ptr
< umax_val
) {
2068 /* Overflow possible, we know nothing */
2069 dst_reg
->umin_value
= 0;
2070 dst_reg
->umax_value
= U64_MAX
;
2072 /* Cannot overflow (as long as bounds are consistent) */
2073 dst_reg
->umin_value
= umin_ptr
- umax_val
;
2074 dst_reg
->umax_value
= umax_ptr
- umin_val
;
2076 dst_reg
->var_off
= tnum_sub(ptr_reg
->var_off
, off_reg
->var_off
);
2077 dst_reg
->off
= ptr_reg
->off
;
2078 if (reg_is_pkt_pointer(ptr_reg
)) {
2079 dst_reg
->id
= ++env
->id_gen
;
2080 /* something was added to pkt_ptr, set range to zero */
2088 /* bitwise ops on pointers are troublesome, prohibit. */
2089 verbose(env
, "R%d bitwise operator %s on pointer prohibited\n",
2090 dst
, bpf_alu_string
[opcode
>> 4]);
2093 /* other operators (e.g. MUL,LSH) produce non-pointer results */
2094 verbose(env
, "R%d pointer arithmetic with %s operator prohibited\n",
2095 dst
, bpf_alu_string
[opcode
>> 4]);
2099 if (!check_reg_sane_offset(env
, dst_reg
, ptr_reg
->type
))
2102 __update_reg_bounds(dst_reg
);
2103 __reg_deduce_bounds(dst_reg
);
2104 __reg_bound_offset(dst_reg
);
2108 /* WARNING: This function does calculations on 64-bit values, but the actual
2109 * execution may occur on 32-bit values. Therefore, things like bitshifts
2110 * need extra checks in the 32-bit case.
2112 static int adjust_scalar_min_max_vals(struct bpf_verifier_env
*env
,
2113 struct bpf_insn
*insn
,
2114 struct bpf_reg_state
*dst_reg
,
2115 struct bpf_reg_state src_reg
)
2117 struct bpf_reg_state
*regs
= cur_regs(env
);
2118 u8 opcode
= BPF_OP(insn
->code
);
2119 bool src_known
, dst_known
;
2120 s64 smin_val
, smax_val
;
2121 u64 umin_val
, umax_val
;
2122 u64 insn_bitness
= (BPF_CLASS(insn
->code
) == BPF_ALU64
) ? 64 : 32;
2124 smin_val
= src_reg
.smin_value
;
2125 smax_val
= src_reg
.smax_value
;
2126 umin_val
= src_reg
.umin_value
;
2127 umax_val
= src_reg
.umax_value
;
2128 src_known
= tnum_is_const(src_reg
.var_off
);
2129 dst_known
= tnum_is_const(dst_reg
->var_off
);
2131 if ((src_known
&& (smin_val
!= smax_val
|| umin_val
!= umax_val
)) ||
2132 smin_val
> smax_val
|| umin_val
> umax_val
) {
2133 /* Taint dst register if offset had invalid bounds derived from
2134 * e.g. dead branches.
2136 __mark_reg_unknown(dst_reg
);
2141 opcode
!= BPF_ADD
&& opcode
!= BPF_SUB
&& opcode
!= BPF_AND
) {
2142 __mark_reg_unknown(dst_reg
);
2148 if (signed_add_overflows(dst_reg
->smin_value
, smin_val
) ||
2149 signed_add_overflows(dst_reg
->smax_value
, smax_val
)) {
2150 dst_reg
->smin_value
= S64_MIN
;
2151 dst_reg
->smax_value
= S64_MAX
;
2153 dst_reg
->smin_value
+= smin_val
;
2154 dst_reg
->smax_value
+= smax_val
;
2156 if (dst_reg
->umin_value
+ umin_val
< umin_val
||
2157 dst_reg
->umax_value
+ umax_val
< umax_val
) {
2158 dst_reg
->umin_value
= 0;
2159 dst_reg
->umax_value
= U64_MAX
;
2161 dst_reg
->umin_value
+= umin_val
;
2162 dst_reg
->umax_value
+= umax_val
;
2164 dst_reg
->var_off
= tnum_add(dst_reg
->var_off
, src_reg
.var_off
);
2167 if (signed_sub_overflows(dst_reg
->smin_value
, smax_val
) ||
2168 signed_sub_overflows(dst_reg
->smax_value
, smin_val
)) {
2169 /* Overflow possible, we know nothing */
2170 dst_reg
->smin_value
= S64_MIN
;
2171 dst_reg
->smax_value
= S64_MAX
;
2173 dst_reg
->smin_value
-= smax_val
;
2174 dst_reg
->smax_value
-= smin_val
;
2176 if (dst_reg
->umin_value
< umax_val
) {
2177 /* Overflow possible, we know nothing */
2178 dst_reg
->umin_value
= 0;
2179 dst_reg
->umax_value
= U64_MAX
;
2181 /* Cannot overflow (as long as bounds are consistent) */
2182 dst_reg
->umin_value
-= umax_val
;
2183 dst_reg
->umax_value
-= umin_val
;
2185 dst_reg
->var_off
= tnum_sub(dst_reg
->var_off
, src_reg
.var_off
);
2188 dst_reg
->var_off
= tnum_mul(dst_reg
->var_off
, src_reg
.var_off
);
2189 if (smin_val
< 0 || dst_reg
->smin_value
< 0) {
2190 /* Ain't nobody got time to multiply that sign */
2191 __mark_reg_unbounded(dst_reg
);
2192 __update_reg_bounds(dst_reg
);
2195 /* Both values are positive, so we can work with unsigned and
2196 * copy the result to signed (unless it exceeds S64_MAX).
2198 if (umax_val
> U32_MAX
|| dst_reg
->umax_value
> U32_MAX
) {
2199 /* Potential overflow, we know nothing */
2200 __mark_reg_unbounded(dst_reg
);
2201 /* (except what we can learn from the var_off) */
2202 __update_reg_bounds(dst_reg
);
2205 dst_reg
->umin_value
*= umin_val
;
2206 dst_reg
->umax_value
*= umax_val
;
2207 if (dst_reg
->umax_value
> S64_MAX
) {
2208 /* Overflow possible, we know nothing */
2209 dst_reg
->smin_value
= S64_MIN
;
2210 dst_reg
->smax_value
= S64_MAX
;
2212 dst_reg
->smin_value
= dst_reg
->umin_value
;
2213 dst_reg
->smax_value
= dst_reg
->umax_value
;
2217 if (src_known
&& dst_known
) {
2218 __mark_reg_known(dst_reg
, dst_reg
->var_off
.value
&
2219 src_reg
.var_off
.value
);
2222 /* We get our minimum from the var_off, since that's inherently
2223 * bitwise. Our maximum is the minimum of the operands' maxima.
2225 dst_reg
->var_off
= tnum_and(dst_reg
->var_off
, src_reg
.var_off
);
2226 dst_reg
->umin_value
= dst_reg
->var_off
.value
;
2227 dst_reg
->umax_value
= min(dst_reg
->umax_value
, umax_val
);
2228 if (dst_reg
->smin_value
< 0 || smin_val
< 0) {
2229 /* Lose signed bounds when ANDing negative numbers,
2230 * ain't nobody got time for that.
2232 dst_reg
->smin_value
= S64_MIN
;
2233 dst_reg
->smax_value
= S64_MAX
;
2235 /* ANDing two positives gives a positive, so safe to
2236 * cast result into s64.
2238 dst_reg
->smin_value
= dst_reg
->umin_value
;
2239 dst_reg
->smax_value
= dst_reg
->umax_value
;
2241 /* We may learn something more from the var_off */
2242 __update_reg_bounds(dst_reg
);
2245 if (src_known
&& dst_known
) {
2246 __mark_reg_known(dst_reg
, dst_reg
->var_off
.value
|
2247 src_reg
.var_off
.value
);
2250 /* We get our maximum from the var_off, and our minimum is the
2251 * maximum of the operands' minima
2253 dst_reg
->var_off
= tnum_or(dst_reg
->var_off
, src_reg
.var_off
);
2254 dst_reg
->umin_value
= max(dst_reg
->umin_value
, umin_val
);
2255 dst_reg
->umax_value
= dst_reg
->var_off
.value
|
2256 dst_reg
->var_off
.mask
;
2257 if (dst_reg
->smin_value
< 0 || smin_val
< 0) {
2258 /* Lose signed bounds when ORing negative numbers,
2259 * ain't nobody got time for that.
2261 dst_reg
->smin_value
= S64_MIN
;
2262 dst_reg
->smax_value
= S64_MAX
;
2264 /* ORing two positives gives a positive, so safe to
2265 * cast result into s64.
2267 dst_reg
->smin_value
= dst_reg
->umin_value
;
2268 dst_reg
->smax_value
= dst_reg
->umax_value
;
2270 /* We may learn something more from the var_off */
2271 __update_reg_bounds(dst_reg
);
2274 if (umax_val
>= insn_bitness
) {
2275 /* Shifts greater than 31 or 63 are undefined.
2276 * This includes shifts by a negative number.
2278 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
2281 /* We lose all sign bit information (except what we can pick
2284 dst_reg
->smin_value
= S64_MIN
;
2285 dst_reg
->smax_value
= S64_MAX
;
2286 /* If we might shift our top bit out, then we know nothing */
2287 if (dst_reg
->umax_value
> 1ULL << (63 - umax_val
)) {
2288 dst_reg
->umin_value
= 0;
2289 dst_reg
->umax_value
= U64_MAX
;
2291 dst_reg
->umin_value
<<= umin_val
;
2292 dst_reg
->umax_value
<<= umax_val
;
2295 dst_reg
->var_off
= tnum_lshift(dst_reg
->var_off
, umin_val
);
2297 dst_reg
->var_off
= tnum_lshift(tnum_unknown
, umin_val
);
2298 /* We may learn something more from the var_off */
2299 __update_reg_bounds(dst_reg
);
2302 if (umax_val
>= insn_bitness
) {
2303 /* Shifts greater than 31 or 63 are undefined.
2304 * This includes shifts by a negative number.
2306 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
2309 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
2310 * be negative, then either:
2311 * 1) src_reg might be zero, so the sign bit of the result is
2312 * unknown, so we lose our signed bounds
2313 * 2) it's known negative, thus the unsigned bounds capture the
2315 * 3) the signed bounds cross zero, so they tell us nothing
2317 * If the value in dst_reg is known nonnegative, then again the
2318 * unsigned bounts capture the signed bounds.
2319 * Thus, in all cases it suffices to blow away our signed bounds
2320 * and rely on inferring new ones from the unsigned bounds and
2321 * var_off of the result.
2323 dst_reg
->smin_value
= S64_MIN
;
2324 dst_reg
->smax_value
= S64_MAX
;
2326 dst_reg
->var_off
= tnum_rshift(dst_reg
->var_off
,
2329 dst_reg
->var_off
= tnum_rshift(tnum_unknown
, umin_val
);
2330 dst_reg
->umin_value
>>= umax_val
;
2331 dst_reg
->umax_value
>>= umin_val
;
2332 /* We may learn something more from the var_off */
2333 __update_reg_bounds(dst_reg
);
2336 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
2340 if (BPF_CLASS(insn
->code
) != BPF_ALU64
) {
2341 /* 32-bit ALU ops are (32,32)->32 */
2342 coerce_reg_to_size(dst_reg
, 4);
2343 coerce_reg_to_size(&src_reg
, 4);
2346 __reg_deduce_bounds(dst_reg
);
2347 __reg_bound_offset(dst_reg
);
2351 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
2354 static int adjust_reg_min_max_vals(struct bpf_verifier_env
*env
,
2355 struct bpf_insn
*insn
)
2357 struct bpf_reg_state
*regs
= cur_regs(env
), *dst_reg
, *src_reg
;
2358 struct bpf_reg_state
*ptr_reg
= NULL
, off_reg
= {0};
2359 u8 opcode
= BPF_OP(insn
->code
);
2361 dst_reg
= ®s
[insn
->dst_reg
];
2363 if (dst_reg
->type
!= SCALAR_VALUE
)
2365 if (BPF_SRC(insn
->code
) == BPF_X
) {
2366 src_reg
= ®s
[insn
->src_reg
];
2367 if (src_reg
->type
!= SCALAR_VALUE
) {
2368 if (dst_reg
->type
!= SCALAR_VALUE
) {
2369 /* Combining two pointers by any ALU op yields
2370 * an arbitrary scalar. Disallow all math except
2371 * pointer subtraction
2373 if (opcode
== BPF_SUB
){
2374 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
2377 verbose(env
, "R%d pointer %s pointer prohibited\n",
2379 bpf_alu_string
[opcode
>> 4]);
2382 /* scalar += pointer
2383 * This is legal, but we have to reverse our
2384 * src/dest handling in computing the range
2386 return adjust_ptr_min_max_vals(env
, insn
,
2389 } else if (ptr_reg
) {
2390 /* pointer += scalar */
2391 return adjust_ptr_min_max_vals(env
, insn
,
2395 /* Pretend the src is a reg with a known value, since we only
2396 * need to be able to read from this state.
2398 off_reg
.type
= SCALAR_VALUE
;
2399 __mark_reg_known(&off_reg
, insn
->imm
);
2401 if (ptr_reg
) /* pointer += K */
2402 return adjust_ptr_min_max_vals(env
, insn
,
2406 /* Got here implies adding two SCALAR_VALUEs */
2407 if (WARN_ON_ONCE(ptr_reg
)) {
2408 print_verifier_state(env
, env
->cur_state
);
2409 verbose(env
, "verifier internal error: unexpected ptr_reg\n");
2412 if (WARN_ON(!src_reg
)) {
2413 print_verifier_state(env
, env
->cur_state
);
2414 verbose(env
, "verifier internal error: no src_reg\n");
2417 return adjust_scalar_min_max_vals(env
, insn
, dst_reg
, *src_reg
);
2420 /* check validity of 32-bit and 64-bit arithmetic operations */
2421 static int check_alu_op(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
2423 struct bpf_reg_state
*regs
= cur_regs(env
);
2424 u8 opcode
= BPF_OP(insn
->code
);
2427 if (opcode
== BPF_END
|| opcode
== BPF_NEG
) {
2428 if (opcode
== BPF_NEG
) {
2429 if (BPF_SRC(insn
->code
) != 0 ||
2430 insn
->src_reg
!= BPF_REG_0
||
2431 insn
->off
!= 0 || insn
->imm
!= 0) {
2432 verbose(env
, "BPF_NEG uses reserved fields\n");
2436 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0 ||
2437 (insn
->imm
!= 16 && insn
->imm
!= 32 && insn
->imm
!= 64) ||
2438 BPF_CLASS(insn
->code
) == BPF_ALU64
) {
2439 verbose(env
, "BPF_END uses reserved fields\n");
2444 /* check src operand */
2445 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
2449 if (is_pointer_value(env
, insn
->dst_reg
)) {
2450 verbose(env
, "R%d pointer arithmetic prohibited\n",
2455 /* check dest operand */
2456 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP
);
2460 } else if (opcode
== BPF_MOV
) {
2462 if (BPF_SRC(insn
->code
) == BPF_X
) {
2463 if (insn
->imm
!= 0 || insn
->off
!= 0) {
2464 verbose(env
, "BPF_MOV uses reserved fields\n");
2468 /* check src operand */
2469 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
2473 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0) {
2474 verbose(env
, "BPF_MOV uses reserved fields\n");
2479 /* check dest operand */
2480 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP
);
2484 if (BPF_SRC(insn
->code
) == BPF_X
) {
2485 if (BPF_CLASS(insn
->code
) == BPF_ALU64
) {
2487 * copy register state to dest reg
2489 regs
[insn
->dst_reg
] = regs
[insn
->src_reg
];
2490 regs
[insn
->dst_reg
].live
|= REG_LIVE_WRITTEN
;
2493 if (is_pointer_value(env
, insn
->src_reg
)) {
2495 "R%d partial copy of pointer\n",
2499 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
2500 coerce_reg_to_size(®s
[insn
->dst_reg
], 4);
2504 * remember the value we stored into this reg
2506 regs
[insn
->dst_reg
].type
= SCALAR_VALUE
;
2507 if (BPF_CLASS(insn
->code
) == BPF_ALU64
) {
2508 __mark_reg_known(regs
+ insn
->dst_reg
,
2511 __mark_reg_known(regs
+ insn
->dst_reg
,
2516 } else if (opcode
> BPF_END
) {
2517 verbose(env
, "invalid BPF_ALU opcode %x\n", opcode
);
2520 } else { /* all other ALU ops: and, sub, xor, add, ... */
2522 if (BPF_SRC(insn
->code
) == BPF_X
) {
2523 if (insn
->imm
!= 0 || insn
->off
!= 0) {
2524 verbose(env
, "BPF_ALU uses reserved fields\n");
2527 /* check src1 operand */
2528 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
2532 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0) {
2533 verbose(env
, "BPF_ALU uses reserved fields\n");
2538 /* check src2 operand */
2539 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
2543 if ((opcode
== BPF_MOD
|| opcode
== BPF_DIV
) &&
2544 BPF_SRC(insn
->code
) == BPF_K
&& insn
->imm
== 0) {
2545 verbose(env
, "div by zero\n");
2549 if (opcode
== BPF_ARSH
&& BPF_CLASS(insn
->code
) != BPF_ALU64
) {
2550 verbose(env
, "BPF_ARSH not supported for 32 bit ALU\n");
2554 if ((opcode
== BPF_LSH
|| opcode
== BPF_RSH
||
2555 opcode
== BPF_ARSH
) && BPF_SRC(insn
->code
) == BPF_K
) {
2556 int size
= BPF_CLASS(insn
->code
) == BPF_ALU64
? 64 : 32;
2558 if (insn
->imm
< 0 || insn
->imm
>= size
) {
2559 verbose(env
, "invalid shift %d\n", insn
->imm
);
2564 /* check dest operand */
2565 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP_NO_MARK
);
2569 return adjust_reg_min_max_vals(env
, insn
);
2575 static void find_good_pkt_pointers(struct bpf_verifier_state
*state
,
2576 struct bpf_reg_state
*dst_reg
,
2577 enum bpf_reg_type type
,
2578 bool range_right_open
)
2580 struct bpf_reg_state
*regs
= state
->regs
, *reg
;
2584 if (dst_reg
->off
< 0 ||
2585 (dst_reg
->off
== 0 && range_right_open
))
2586 /* This doesn't give us any range */
2589 if (dst_reg
->umax_value
> MAX_PACKET_OFF
||
2590 dst_reg
->umax_value
+ dst_reg
->off
> MAX_PACKET_OFF
)
2591 /* Risk of overflow. For instance, ptr + (1<<63) may be less
2592 * than pkt_end, but that's because it's also less than pkt.
2596 new_range
= dst_reg
->off
;
2597 if (range_right_open
)
2600 /* Examples for register markings:
2602 * pkt_data in dst register:
2606 * if (r2 > pkt_end) goto <handle exception>
2611 * if (r2 < pkt_end) goto <access okay>
2612 * <handle exception>
2615 * r2 == dst_reg, pkt_end == src_reg
2616 * r2=pkt(id=n,off=8,r=0)
2617 * r3=pkt(id=n,off=0,r=0)
2619 * pkt_data in src register:
2623 * if (pkt_end >= r2) goto <access okay>
2624 * <handle exception>
2628 * if (pkt_end <= r2) goto <handle exception>
2632 * pkt_end == dst_reg, r2 == src_reg
2633 * r2=pkt(id=n,off=8,r=0)
2634 * r3=pkt(id=n,off=0,r=0)
2636 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
2637 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
2638 * and [r3, r3 + 8-1) respectively is safe to access depending on
2642 /* If our ids match, then we must have the same max_value. And we
2643 * don't care about the other reg's fixed offset, since if it's too big
2644 * the range won't allow anything.
2645 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
2647 for (i
= 0; i
< MAX_BPF_REG
; i
++)
2648 if (regs
[i
].type
== type
&& regs
[i
].id
== dst_reg
->id
)
2649 /* keep the maximum range already checked */
2650 regs
[i
].range
= max(regs
[i
].range
, new_range
);
2652 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
2653 if (state
->stack
[i
].slot_type
[0] != STACK_SPILL
)
2655 reg
= &state
->stack
[i
].spilled_ptr
;
2656 if (reg
->type
== type
&& reg
->id
== dst_reg
->id
)
2657 reg
->range
= max(reg
->range
, new_range
);
2661 /* Adjusts the register min/max values in the case that the dst_reg is the
2662 * variable register that we are working on, and src_reg is a constant or we're
2663 * simply doing a BPF_K check.
2664 * In JEQ/JNE cases we also adjust the var_off values.
2666 static void reg_set_min_max(struct bpf_reg_state
*true_reg
,
2667 struct bpf_reg_state
*false_reg
, u64 val
,
2670 /* If the dst_reg is a pointer, we can't learn anything about its
2671 * variable offset from the compare (unless src_reg were a pointer into
2672 * the same object, but we don't bother with that.
2673 * Since false_reg and true_reg have the same type by construction, we
2674 * only need to check one of them for pointerness.
2676 if (__is_pointer_value(false, false_reg
))
2681 /* If this is false then we know nothing Jon Snow, but if it is
2682 * true then we know for sure.
2684 __mark_reg_known(true_reg
, val
);
2687 /* If this is true we know nothing Jon Snow, but if it is false
2688 * we know the value for sure;
2690 __mark_reg_known(false_reg
, val
);
2693 false_reg
->umax_value
= min(false_reg
->umax_value
, val
);
2694 true_reg
->umin_value
= max(true_reg
->umin_value
, val
+ 1);
2697 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
);
2698 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
+ 1);
2701 false_reg
->umin_value
= max(false_reg
->umin_value
, val
);
2702 true_reg
->umax_value
= min(true_reg
->umax_value
, val
- 1);
2705 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
);
2706 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
- 1);
2709 false_reg
->umax_value
= min(false_reg
->umax_value
, val
- 1);
2710 true_reg
->umin_value
= max(true_reg
->umin_value
, val
);
2713 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
- 1);
2714 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
);
2717 false_reg
->umin_value
= max(false_reg
->umin_value
, val
+ 1);
2718 true_reg
->umax_value
= min(true_reg
->umax_value
, val
);
2721 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
+ 1);
2722 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
);
2728 __reg_deduce_bounds(false_reg
);
2729 __reg_deduce_bounds(true_reg
);
2730 /* We might have learned some bits from the bounds. */
2731 __reg_bound_offset(false_reg
);
2732 __reg_bound_offset(true_reg
);
2733 /* Intersecting with the old var_off might have improved our bounds
2734 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2735 * then new var_off is (0; 0x7f...fc) which improves our umax.
2737 __update_reg_bounds(false_reg
);
2738 __update_reg_bounds(true_reg
);
2741 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
2744 static void reg_set_min_max_inv(struct bpf_reg_state
*true_reg
,
2745 struct bpf_reg_state
*false_reg
, u64 val
,
2748 if (__is_pointer_value(false, false_reg
))
2753 /* If this is false then we know nothing Jon Snow, but if it is
2754 * true then we know for sure.
2756 __mark_reg_known(true_reg
, val
);
2759 /* If this is true we know nothing Jon Snow, but if it is false
2760 * we know the value for sure;
2762 __mark_reg_known(false_reg
, val
);
2765 true_reg
->umax_value
= min(true_reg
->umax_value
, val
- 1);
2766 false_reg
->umin_value
= max(false_reg
->umin_value
, val
);
2769 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
- 1);
2770 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
);
2773 true_reg
->umin_value
= max(true_reg
->umin_value
, val
+ 1);
2774 false_reg
->umax_value
= min(false_reg
->umax_value
, val
);
2777 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
+ 1);
2778 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
);
2781 true_reg
->umax_value
= min(true_reg
->umax_value
, val
);
2782 false_reg
->umin_value
= max(false_reg
->umin_value
, val
+ 1);
2785 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
);
2786 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
+ 1);
2789 true_reg
->umin_value
= max(true_reg
->umin_value
, val
);
2790 false_reg
->umax_value
= min(false_reg
->umax_value
, val
- 1);
2793 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
);
2794 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
- 1);
2800 __reg_deduce_bounds(false_reg
);
2801 __reg_deduce_bounds(true_reg
);
2802 /* We might have learned some bits from the bounds. */
2803 __reg_bound_offset(false_reg
);
2804 __reg_bound_offset(true_reg
);
2805 /* Intersecting with the old var_off might have improved our bounds
2806 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2807 * then new var_off is (0; 0x7f...fc) which improves our umax.
2809 __update_reg_bounds(false_reg
);
2810 __update_reg_bounds(true_reg
);
2813 /* Regs are known to be equal, so intersect their min/max/var_off */
2814 static void __reg_combine_min_max(struct bpf_reg_state
*src_reg
,
2815 struct bpf_reg_state
*dst_reg
)
2817 src_reg
->umin_value
= dst_reg
->umin_value
= max(src_reg
->umin_value
,
2818 dst_reg
->umin_value
);
2819 src_reg
->umax_value
= dst_reg
->umax_value
= min(src_reg
->umax_value
,
2820 dst_reg
->umax_value
);
2821 src_reg
->smin_value
= dst_reg
->smin_value
= max(src_reg
->smin_value
,
2822 dst_reg
->smin_value
);
2823 src_reg
->smax_value
= dst_reg
->smax_value
= min(src_reg
->smax_value
,
2824 dst_reg
->smax_value
);
2825 src_reg
->var_off
= dst_reg
->var_off
= tnum_intersect(src_reg
->var_off
,
2827 /* We might have learned new bounds from the var_off. */
2828 __update_reg_bounds(src_reg
);
2829 __update_reg_bounds(dst_reg
);
2830 /* We might have learned something about the sign bit. */
2831 __reg_deduce_bounds(src_reg
);
2832 __reg_deduce_bounds(dst_reg
);
2833 /* We might have learned some bits from the bounds. */
2834 __reg_bound_offset(src_reg
);
2835 __reg_bound_offset(dst_reg
);
2836 /* Intersecting with the old var_off might have improved our bounds
2837 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2838 * then new var_off is (0; 0x7f...fc) which improves our umax.
2840 __update_reg_bounds(src_reg
);
2841 __update_reg_bounds(dst_reg
);
2844 static void reg_combine_min_max(struct bpf_reg_state
*true_src
,
2845 struct bpf_reg_state
*true_dst
,
2846 struct bpf_reg_state
*false_src
,
2847 struct bpf_reg_state
*false_dst
,
2852 __reg_combine_min_max(true_src
, true_dst
);
2855 __reg_combine_min_max(false_src
, false_dst
);
2860 static void mark_map_reg(struct bpf_reg_state
*regs
, u32 regno
, u32 id
,
2863 struct bpf_reg_state
*reg
= ®s
[regno
];
2865 if (reg
->type
== PTR_TO_MAP_VALUE_OR_NULL
&& reg
->id
== id
) {
2866 /* Old offset (both fixed and variable parts) should
2867 * have been known-zero, because we don't allow pointer
2868 * arithmetic on pointers that might be NULL.
2870 if (WARN_ON_ONCE(reg
->smin_value
|| reg
->smax_value
||
2871 !tnum_equals_const(reg
->var_off
, 0) ||
2873 __mark_reg_known_zero(reg
);
2877 reg
->type
= SCALAR_VALUE
;
2878 } else if (reg
->map_ptr
->inner_map_meta
) {
2879 reg
->type
= CONST_PTR_TO_MAP
;
2880 reg
->map_ptr
= reg
->map_ptr
->inner_map_meta
;
2882 reg
->type
= PTR_TO_MAP_VALUE
;
2884 /* We don't need id from this point onwards anymore, thus we
2885 * should better reset it, so that state pruning has chances
2892 /* The logic is similar to find_good_pkt_pointers(), both could eventually
2893 * be folded together at some point.
2895 static void mark_map_regs(struct bpf_verifier_state
*state
, u32 regno
,
2898 struct bpf_reg_state
*regs
= state
->regs
;
2899 u32 id
= regs
[regno
].id
;
2902 for (i
= 0; i
< MAX_BPF_REG
; i
++)
2903 mark_map_reg(regs
, i
, id
, is_null
);
2905 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
2906 if (state
->stack
[i
].slot_type
[0] != STACK_SPILL
)
2908 mark_map_reg(&state
->stack
[i
].spilled_ptr
, 0, id
, is_null
);
2912 static bool try_match_pkt_pointers(const struct bpf_insn
*insn
,
2913 struct bpf_reg_state
*dst_reg
,
2914 struct bpf_reg_state
*src_reg
,
2915 struct bpf_verifier_state
*this_branch
,
2916 struct bpf_verifier_state
*other_branch
)
2918 if (BPF_SRC(insn
->code
) != BPF_X
)
2921 switch (BPF_OP(insn
->code
)) {
2923 if ((dst_reg
->type
== PTR_TO_PACKET
&&
2924 src_reg
->type
== PTR_TO_PACKET_END
) ||
2925 (dst_reg
->type
== PTR_TO_PACKET_META
&&
2926 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
2927 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
2928 find_good_pkt_pointers(this_branch
, dst_reg
,
2929 dst_reg
->type
, false);
2930 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
2931 src_reg
->type
== PTR_TO_PACKET
) ||
2932 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
2933 src_reg
->type
== PTR_TO_PACKET_META
)) {
2934 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
2935 find_good_pkt_pointers(other_branch
, src_reg
,
2936 src_reg
->type
, true);
2942 if ((dst_reg
->type
== PTR_TO_PACKET
&&
2943 src_reg
->type
== PTR_TO_PACKET_END
) ||
2944 (dst_reg
->type
== PTR_TO_PACKET_META
&&
2945 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
2946 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
2947 find_good_pkt_pointers(other_branch
, dst_reg
,
2948 dst_reg
->type
, true);
2949 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
2950 src_reg
->type
== PTR_TO_PACKET
) ||
2951 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
2952 src_reg
->type
== PTR_TO_PACKET_META
)) {
2953 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
2954 find_good_pkt_pointers(this_branch
, src_reg
,
2955 src_reg
->type
, false);
2961 if ((dst_reg
->type
== PTR_TO_PACKET
&&
2962 src_reg
->type
== PTR_TO_PACKET_END
) ||
2963 (dst_reg
->type
== PTR_TO_PACKET_META
&&
2964 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
2965 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
2966 find_good_pkt_pointers(this_branch
, dst_reg
,
2967 dst_reg
->type
, true);
2968 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
2969 src_reg
->type
== PTR_TO_PACKET
) ||
2970 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
2971 src_reg
->type
== PTR_TO_PACKET_META
)) {
2972 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
2973 find_good_pkt_pointers(other_branch
, src_reg
,
2974 src_reg
->type
, false);
2980 if ((dst_reg
->type
== PTR_TO_PACKET
&&
2981 src_reg
->type
== PTR_TO_PACKET_END
) ||
2982 (dst_reg
->type
== PTR_TO_PACKET_META
&&
2983 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
2984 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
2985 find_good_pkt_pointers(other_branch
, dst_reg
,
2986 dst_reg
->type
, false);
2987 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
2988 src_reg
->type
== PTR_TO_PACKET
) ||
2989 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
2990 src_reg
->type
== PTR_TO_PACKET_META
)) {
2991 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
2992 find_good_pkt_pointers(this_branch
, src_reg
,
2993 src_reg
->type
, true);
3005 static int check_cond_jmp_op(struct bpf_verifier_env
*env
,
3006 struct bpf_insn
*insn
, int *insn_idx
)
3008 struct bpf_verifier_state
*other_branch
, *this_branch
= env
->cur_state
;
3009 struct bpf_reg_state
*regs
= this_branch
->regs
, *dst_reg
;
3010 u8 opcode
= BPF_OP(insn
->code
);
3013 if (opcode
> BPF_JSLE
) {
3014 verbose(env
, "invalid BPF_JMP opcode %x\n", opcode
);
3018 if (BPF_SRC(insn
->code
) == BPF_X
) {
3019 if (insn
->imm
!= 0) {
3020 verbose(env
, "BPF_JMP uses reserved fields\n");
3024 /* check src1 operand */
3025 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
3029 if (is_pointer_value(env
, insn
->src_reg
)) {
3030 verbose(env
, "R%d pointer comparison prohibited\n",
3035 if (insn
->src_reg
!= BPF_REG_0
) {
3036 verbose(env
, "BPF_JMP uses reserved fields\n");
3041 /* check src2 operand */
3042 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
3046 dst_reg
= ®s
[insn
->dst_reg
];
3048 /* detect if R == 0 where R was initialized to zero earlier */
3049 if (BPF_SRC(insn
->code
) == BPF_K
&&
3050 (opcode
== BPF_JEQ
|| opcode
== BPF_JNE
) &&
3051 dst_reg
->type
== SCALAR_VALUE
&&
3052 tnum_equals_const(dst_reg
->var_off
, insn
->imm
)) {
3053 if (opcode
== BPF_JEQ
) {
3054 /* if (imm == imm) goto pc+off;
3055 * only follow the goto, ignore fall-through
3057 *insn_idx
+= insn
->off
;
3060 /* if (imm != imm) goto pc+off;
3061 * only follow fall-through branch, since
3062 * that's where the program will go
3068 other_branch
= push_stack(env
, *insn_idx
+ insn
->off
+ 1, *insn_idx
);
3072 /* detect if we are comparing against a constant value so we can adjust
3073 * our min/max values for our dst register.
3074 * this is only legit if both are scalars (or pointers to the same
3075 * object, I suppose, but we don't support that right now), because
3076 * otherwise the different base pointers mean the offsets aren't
3079 if (BPF_SRC(insn
->code
) == BPF_X
) {
3080 if (dst_reg
->type
== SCALAR_VALUE
&&
3081 regs
[insn
->src_reg
].type
== SCALAR_VALUE
) {
3082 if (tnum_is_const(regs
[insn
->src_reg
].var_off
))
3083 reg_set_min_max(&other_branch
->regs
[insn
->dst_reg
],
3084 dst_reg
, regs
[insn
->src_reg
].var_off
.value
,
3086 else if (tnum_is_const(dst_reg
->var_off
))
3087 reg_set_min_max_inv(&other_branch
->regs
[insn
->src_reg
],
3088 ®s
[insn
->src_reg
],
3089 dst_reg
->var_off
.value
, opcode
);
3090 else if (opcode
== BPF_JEQ
|| opcode
== BPF_JNE
)
3091 /* Comparing for equality, we can combine knowledge */
3092 reg_combine_min_max(&other_branch
->regs
[insn
->src_reg
],
3093 &other_branch
->regs
[insn
->dst_reg
],
3094 ®s
[insn
->src_reg
],
3095 ®s
[insn
->dst_reg
], opcode
);
3097 } else if (dst_reg
->type
== SCALAR_VALUE
) {
3098 reg_set_min_max(&other_branch
->regs
[insn
->dst_reg
],
3099 dst_reg
, insn
->imm
, opcode
);
3102 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
3103 if (BPF_SRC(insn
->code
) == BPF_K
&&
3104 insn
->imm
== 0 && (opcode
== BPF_JEQ
|| opcode
== BPF_JNE
) &&
3105 dst_reg
->type
== PTR_TO_MAP_VALUE_OR_NULL
) {
3106 /* Mark all identical map registers in each branch as either
3107 * safe or unknown depending R == 0 or R != 0 conditional.
3109 mark_map_regs(this_branch
, insn
->dst_reg
, opcode
== BPF_JNE
);
3110 mark_map_regs(other_branch
, insn
->dst_reg
, opcode
== BPF_JEQ
);
3111 } else if (!try_match_pkt_pointers(insn
, dst_reg
, ®s
[insn
->src_reg
],
3112 this_branch
, other_branch
) &&
3113 is_pointer_value(env
, insn
->dst_reg
)) {
3114 verbose(env
, "R%d pointer comparison prohibited\n",
3119 print_verifier_state(env
, this_branch
);
3123 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
3124 static struct bpf_map
*ld_imm64_to_map_ptr(struct bpf_insn
*insn
)
3126 u64 imm64
= ((u64
) (u32
) insn
[0].imm
) | ((u64
) (u32
) insn
[1].imm
) << 32;
3128 return (struct bpf_map
*) (unsigned long) imm64
;
3131 /* verify BPF_LD_IMM64 instruction */
3132 static int check_ld_imm(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
3134 struct bpf_reg_state
*regs
= cur_regs(env
);
3137 if (BPF_SIZE(insn
->code
) != BPF_DW
) {
3138 verbose(env
, "invalid BPF_LD_IMM insn\n");
3141 if (insn
->off
!= 0) {
3142 verbose(env
, "BPF_LD_IMM64 uses reserved fields\n");
3146 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP
);
3150 if (insn
->src_reg
== 0) {
3151 u64 imm
= ((u64
)(insn
+ 1)->imm
<< 32) | (u32
)insn
->imm
;
3153 regs
[insn
->dst_reg
].type
= SCALAR_VALUE
;
3154 __mark_reg_known(®s
[insn
->dst_reg
], imm
);
3158 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
3159 BUG_ON(insn
->src_reg
!= BPF_PSEUDO_MAP_FD
);
3161 regs
[insn
->dst_reg
].type
= CONST_PTR_TO_MAP
;
3162 regs
[insn
->dst_reg
].map_ptr
= ld_imm64_to_map_ptr(insn
);
3166 static bool may_access_skb(enum bpf_prog_type type
)
3169 case BPF_PROG_TYPE_SOCKET_FILTER
:
3170 case BPF_PROG_TYPE_SCHED_CLS
:
3171 case BPF_PROG_TYPE_SCHED_ACT
:
3178 /* verify safety of LD_ABS|LD_IND instructions:
3179 * - they can only appear in the programs where ctx == skb
3180 * - since they are wrappers of function calls, they scratch R1-R5 registers,
3181 * preserve R6-R9, and store return value into R0
3184 * ctx == skb == R6 == CTX
3187 * SRC == any register
3188 * IMM == 32-bit immediate
3191 * R0 - 8/16/32-bit skb data converted to cpu endianness
3193 static int check_ld_abs(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
3195 struct bpf_reg_state
*regs
= cur_regs(env
);
3196 u8 mode
= BPF_MODE(insn
->code
);
3199 if (!may_access_skb(env
->prog
->type
)) {
3200 verbose(env
, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
3204 if (insn
->dst_reg
!= BPF_REG_0
|| insn
->off
!= 0 ||
3205 BPF_SIZE(insn
->code
) == BPF_DW
||
3206 (mode
== BPF_ABS
&& insn
->src_reg
!= BPF_REG_0
)) {
3207 verbose(env
, "BPF_LD_[ABS|IND] uses reserved fields\n");
3211 /* check whether implicit source operand (register R6) is readable */
3212 err
= check_reg_arg(env
, BPF_REG_6
, SRC_OP
);
3216 if (regs
[BPF_REG_6
].type
!= PTR_TO_CTX
) {
3218 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
3222 if (mode
== BPF_IND
) {
3223 /* check explicit source operand */
3224 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
3229 /* reset caller saved regs to unreadable */
3230 for (i
= 0; i
< CALLER_SAVED_REGS
; i
++) {
3231 mark_reg_not_init(env
, regs
, caller_saved
[i
]);
3232 check_reg_arg(env
, caller_saved
[i
], DST_OP_NO_MARK
);
3235 /* mark destination R0 register as readable, since it contains
3236 * the value fetched from the packet.
3237 * Already marked as written above.
3239 mark_reg_unknown(env
, regs
, BPF_REG_0
);
3243 static int check_return_code(struct bpf_verifier_env
*env
)
3245 struct bpf_reg_state
*reg
;
3246 struct tnum range
= tnum_range(0, 1);
3248 switch (env
->prog
->type
) {
3249 case BPF_PROG_TYPE_CGROUP_SKB
:
3250 case BPF_PROG_TYPE_CGROUP_SOCK
:
3251 case BPF_PROG_TYPE_SOCK_OPS
:
3252 case BPF_PROG_TYPE_CGROUP_DEVICE
:
3258 reg
= cur_regs(env
) + BPF_REG_0
;
3259 if (reg
->type
!= SCALAR_VALUE
) {
3260 verbose(env
, "At program exit the register R0 is not a known value (%s)\n",
3261 reg_type_str
[reg
->type
]);
3265 if (!tnum_in(range
, reg
->var_off
)) {
3266 verbose(env
, "At program exit the register R0 ");
3267 if (!tnum_is_unknown(reg
->var_off
)) {
3270 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
3271 verbose(env
, "has value %s", tn_buf
);
3273 verbose(env
, "has unknown scalar value");
3275 verbose(env
, " should have been 0 or 1\n");
3281 /* non-recursive DFS pseudo code
3282 * 1 procedure DFS-iterative(G,v):
3283 * 2 label v as discovered
3284 * 3 let S be a stack
3286 * 5 while S is not empty
3288 * 7 if t is what we're looking for:
3290 * 9 for all edges e in G.adjacentEdges(t) do
3291 * 10 if edge e is already labelled
3292 * 11 continue with the next edge
3293 * 12 w <- G.adjacentVertex(t,e)
3294 * 13 if vertex w is not discovered and not explored
3295 * 14 label e as tree-edge
3296 * 15 label w as discovered
3299 * 18 else if vertex w is discovered
3300 * 19 label e as back-edge
3302 * 21 // vertex w is explored
3303 * 22 label e as forward- or cross-edge
3304 * 23 label t as explored
3309 * 0x11 - discovered and fall-through edge labelled
3310 * 0x12 - discovered and fall-through and branch edges labelled
3321 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
3323 static int *insn_stack
; /* stack of insns to process */
3324 static int cur_stack
; /* current stack index */
3325 static int *insn_state
;
3327 /* t, w, e - match pseudo-code above:
3328 * t - index of current instruction
3329 * w - next instruction
3332 static int push_insn(int t
, int w
, int e
, struct bpf_verifier_env
*env
)
3334 if (e
== FALLTHROUGH
&& insn_state
[t
] >= (DISCOVERED
| FALLTHROUGH
))
3337 if (e
== BRANCH
&& insn_state
[t
] >= (DISCOVERED
| BRANCH
))
3340 if (w
< 0 || w
>= env
->prog
->len
) {
3341 verbose(env
, "jump out of range from insn %d to %d\n", t
, w
);
3346 /* mark branch target for state pruning */
3347 env
->explored_states
[w
] = STATE_LIST_MARK
;
3349 if (insn_state
[w
] == 0) {
3351 insn_state
[t
] = DISCOVERED
| e
;
3352 insn_state
[w
] = DISCOVERED
;
3353 if (cur_stack
>= env
->prog
->len
)
3355 insn_stack
[cur_stack
++] = w
;
3357 } else if ((insn_state
[w
] & 0xF0) == DISCOVERED
) {
3358 verbose(env
, "back-edge from insn %d to %d\n", t
, w
);
3360 } else if (insn_state
[w
] == EXPLORED
) {
3361 /* forward- or cross-edge */
3362 insn_state
[t
] = DISCOVERED
| e
;
3364 verbose(env
, "insn state internal bug\n");
3370 /* non-recursive depth-first-search to detect loops in BPF program
3371 * loop == back-edge in directed graph
3373 static int check_cfg(struct bpf_verifier_env
*env
)
3375 struct bpf_insn
*insns
= env
->prog
->insnsi
;
3376 int insn_cnt
= env
->prog
->len
;
3380 insn_state
= kcalloc(insn_cnt
, sizeof(int), GFP_KERNEL
);
3384 insn_stack
= kcalloc(insn_cnt
, sizeof(int), GFP_KERNEL
);
3390 insn_state
[0] = DISCOVERED
; /* mark 1st insn as discovered */
3391 insn_stack
[0] = 0; /* 0 is the first instruction */
3397 t
= insn_stack
[cur_stack
- 1];
3399 if (BPF_CLASS(insns
[t
].code
) == BPF_JMP
) {
3400 u8 opcode
= BPF_OP(insns
[t
].code
);
3402 if (opcode
== BPF_EXIT
) {
3404 } else if (opcode
== BPF_CALL
) {
3405 ret
= push_insn(t
, t
+ 1, FALLTHROUGH
, env
);
3410 if (t
+ 1 < insn_cnt
)
3411 env
->explored_states
[t
+ 1] = STATE_LIST_MARK
;
3412 } else if (opcode
== BPF_JA
) {
3413 if (BPF_SRC(insns
[t
].code
) != BPF_K
) {
3417 /* unconditional jump with single edge */
3418 ret
= push_insn(t
, t
+ insns
[t
].off
+ 1,
3424 /* tell verifier to check for equivalent states
3425 * after every call and jump
3427 if (t
+ 1 < insn_cnt
)
3428 env
->explored_states
[t
+ 1] = STATE_LIST_MARK
;
3430 /* conditional jump with two edges */
3431 env
->explored_states
[t
] = STATE_LIST_MARK
;
3432 ret
= push_insn(t
, t
+ 1, FALLTHROUGH
, env
);
3438 ret
= push_insn(t
, t
+ insns
[t
].off
+ 1, BRANCH
, env
);
3445 /* all other non-branch instructions with single
3448 ret
= push_insn(t
, t
+ 1, FALLTHROUGH
, env
);
3456 insn_state
[t
] = EXPLORED
;
3457 if (cur_stack
-- <= 0) {
3458 verbose(env
, "pop stack internal bug\n");
3465 for (i
= 0; i
< insn_cnt
; i
++) {
3466 if (insn_state
[i
] != EXPLORED
) {
3467 verbose(env
, "unreachable insn %d\n", i
);
3472 ret
= 0; /* cfg looks good */
3480 /* check %cur's range satisfies %old's */
3481 static bool range_within(struct bpf_reg_state
*old
,
3482 struct bpf_reg_state
*cur
)
3484 return old
->umin_value
<= cur
->umin_value
&&
3485 old
->umax_value
>= cur
->umax_value
&&
3486 old
->smin_value
<= cur
->smin_value
&&
3487 old
->smax_value
>= cur
->smax_value
;
3490 /* Maximum number of register states that can exist at once */
3491 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
3497 /* If in the old state two registers had the same id, then they need to have
3498 * the same id in the new state as well. But that id could be different from
3499 * the old state, so we need to track the mapping from old to new ids.
3500 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
3501 * regs with old id 5 must also have new id 9 for the new state to be safe. But
3502 * regs with a different old id could still have new id 9, we don't care about
3504 * So we look through our idmap to see if this old id has been seen before. If
3505 * so, we require the new id to match; otherwise, we add the id pair to the map.
3507 static bool check_ids(u32 old_id
, u32 cur_id
, struct idpair
*idmap
)
3511 for (i
= 0; i
< ID_MAP_SIZE
; i
++) {
3512 if (!idmap
[i
].old
) {
3513 /* Reached an empty slot; haven't seen this id before */
3514 idmap
[i
].old
= old_id
;
3515 idmap
[i
].cur
= cur_id
;
3518 if (idmap
[i
].old
== old_id
)
3519 return idmap
[i
].cur
== cur_id
;
3521 /* We ran out of idmap slots, which should be impossible */
3526 /* Returns true if (rold safe implies rcur safe) */
3527 static bool regsafe(struct bpf_reg_state
*rold
, struct bpf_reg_state
*rcur
,
3528 struct idpair
*idmap
)
3530 if (!(rold
->live
& REG_LIVE_READ
))
3531 /* explored state didn't use this */
3534 if (memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, live
)) == 0)
3537 if (rold
->type
== NOT_INIT
)
3538 /* explored state can't have used this */
3540 if (rcur
->type
== NOT_INIT
)
3542 switch (rold
->type
) {
3544 if (rcur
->type
== SCALAR_VALUE
) {
3545 /* new val must satisfy old val knowledge */
3546 return range_within(rold
, rcur
) &&
3547 tnum_in(rold
->var_off
, rcur
->var_off
);
3549 /* We're trying to use a pointer in place of a scalar.
3550 * Even if the scalar was unbounded, this could lead to
3551 * pointer leaks because scalars are allowed to leak
3552 * while pointers are not. We could make this safe in
3553 * special cases if root is calling us, but it's
3554 * probably not worth the hassle.
3558 case PTR_TO_MAP_VALUE
:
3559 /* If the new min/max/var_off satisfy the old ones and
3560 * everything else matches, we are OK.
3561 * We don't care about the 'id' value, because nothing
3562 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
3564 return memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, id
)) == 0 &&
3565 range_within(rold
, rcur
) &&
3566 tnum_in(rold
->var_off
, rcur
->var_off
);
3567 case PTR_TO_MAP_VALUE_OR_NULL
:
3568 /* a PTR_TO_MAP_VALUE could be safe to use as a
3569 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
3570 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
3571 * checked, doing so could have affected others with the same
3572 * id, and we can't check for that because we lost the id when
3573 * we converted to a PTR_TO_MAP_VALUE.
3575 if (rcur
->type
!= PTR_TO_MAP_VALUE_OR_NULL
)
3577 if (memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, id
)))
3579 /* Check our ids match any regs they're supposed to */
3580 return check_ids(rold
->id
, rcur
->id
, idmap
);
3581 case PTR_TO_PACKET_META
:
3583 if (rcur
->type
!= rold
->type
)
3585 /* We must have at least as much range as the old ptr
3586 * did, so that any accesses which were safe before are
3587 * still safe. This is true even if old range < old off,
3588 * since someone could have accessed through (ptr - k), or
3589 * even done ptr -= k in a register, to get a safe access.
3591 if (rold
->range
> rcur
->range
)
3593 /* If the offsets don't match, we can't trust our alignment;
3594 * nor can we be sure that we won't fall out of range.
3596 if (rold
->off
!= rcur
->off
)
3598 /* id relations must be preserved */
3599 if (rold
->id
&& !check_ids(rold
->id
, rcur
->id
, idmap
))
3601 /* new val must satisfy old val knowledge */
3602 return range_within(rold
, rcur
) &&
3603 tnum_in(rold
->var_off
, rcur
->var_off
);
3605 case CONST_PTR_TO_MAP
:
3607 case PTR_TO_PACKET_END
:
3608 /* Only valid matches are exact, which memcmp() above
3609 * would have accepted
3612 /* Don't know what's going on, just say it's not safe */
3616 /* Shouldn't get here; if we do, say it's not safe */
3621 static bool stacksafe(struct bpf_verifier_state
*old
,
3622 struct bpf_verifier_state
*cur
,
3623 struct idpair
*idmap
)
3627 /* if explored stack has more populated slots than current stack
3628 * such stacks are not equivalent
3630 if (old
->allocated_stack
> cur
->allocated_stack
)
3633 /* walk slots of the explored stack and ignore any additional
3634 * slots in the current stack, since explored(safe) state
3637 for (i
= 0; i
< old
->allocated_stack
; i
++) {
3638 spi
= i
/ BPF_REG_SIZE
;
3640 if (old
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
] == STACK_INVALID
)
3642 if (old
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
] !=
3643 cur
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
])
3644 /* Ex: old explored (safe) state has STACK_SPILL in
3645 * this stack slot, but current has has STACK_MISC ->
3646 * this verifier states are not equivalent,
3647 * return false to continue verification of this path
3650 if (i
% BPF_REG_SIZE
)
3652 if (old
->stack
[spi
].slot_type
[0] != STACK_SPILL
)
3654 if (!regsafe(&old
->stack
[spi
].spilled_ptr
,
3655 &cur
->stack
[spi
].spilled_ptr
,
3657 /* when explored and current stack slot are both storing
3658 * spilled registers, check that stored pointers types
3659 * are the same as well.
3660 * Ex: explored safe path could have stored
3661 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
3662 * but current path has stored:
3663 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
3664 * such verifier states are not equivalent.
3665 * return false to continue verification of this path
3672 /* compare two verifier states
3674 * all states stored in state_list are known to be valid, since
3675 * verifier reached 'bpf_exit' instruction through them
3677 * this function is called when verifier exploring different branches of
3678 * execution popped from the state stack. If it sees an old state that has
3679 * more strict register state and more strict stack state then this execution
3680 * branch doesn't need to be explored further, since verifier already
3681 * concluded that more strict state leads to valid finish.
3683 * Therefore two states are equivalent if register state is more conservative
3684 * and explored stack state is more conservative than the current one.
3687 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
3688 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
3690 * In other words if current stack state (one being explored) has more
3691 * valid slots than old one that already passed validation, it means
3692 * the verifier can stop exploring and conclude that current state is valid too
3694 * Similarly with registers. If explored state has register type as invalid
3695 * whereas register type in current state is meaningful, it means that
3696 * the current state will reach 'bpf_exit' instruction safely
3698 static bool states_equal(struct bpf_verifier_env
*env
,
3699 struct bpf_verifier_state
*old
,
3700 struct bpf_verifier_state
*cur
)
3702 struct idpair
*idmap
;
3706 idmap
= kcalloc(ID_MAP_SIZE
, sizeof(struct idpair
), GFP_KERNEL
);
3707 /* If we failed to allocate the idmap, just say it's not safe */
3711 for (i
= 0; i
< MAX_BPF_REG
; i
++) {
3712 if (!regsafe(&old
->regs
[i
], &cur
->regs
[i
], idmap
))
3716 if (!stacksafe(old
, cur
, idmap
))
3724 /* A write screens off any subsequent reads; but write marks come from the
3725 * straight-line code between a state and its parent. When we arrive at a
3726 * jump target (in the first iteration of the propagate_liveness() loop),
3727 * we didn't arrive by the straight-line code, so read marks in state must
3728 * propagate to parent regardless of state's write marks.
3730 static bool do_propagate_liveness(const struct bpf_verifier_state
*state
,
3731 struct bpf_verifier_state
*parent
)
3733 bool writes
= parent
== state
->parent
; /* Observe write marks */
3734 bool touched
= false; /* any changes made? */
3739 /* Propagate read liveness of registers... */
3740 BUILD_BUG_ON(BPF_REG_FP
+ 1 != MAX_BPF_REG
);
3741 /* We don't need to worry about FP liveness because it's read-only */
3742 for (i
= 0; i
< BPF_REG_FP
; i
++) {
3743 if (parent
->regs
[i
].live
& REG_LIVE_READ
)
3745 if (writes
&& (state
->regs
[i
].live
& REG_LIVE_WRITTEN
))
3747 if (state
->regs
[i
].live
& REG_LIVE_READ
) {
3748 parent
->regs
[i
].live
|= REG_LIVE_READ
;
3752 /* ... and stack slots */
3753 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
&&
3754 i
< parent
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
3755 if (parent
->stack
[i
].slot_type
[0] != STACK_SPILL
)
3757 if (state
->stack
[i
].slot_type
[0] != STACK_SPILL
)
3759 if (parent
->stack
[i
].spilled_ptr
.live
& REG_LIVE_READ
)
3762 (state
->stack
[i
].spilled_ptr
.live
& REG_LIVE_WRITTEN
))
3764 if (state
->stack
[i
].spilled_ptr
.live
& REG_LIVE_READ
) {
3765 parent
->stack
[i
].spilled_ptr
.live
|= REG_LIVE_READ
;
3772 /* "parent" is "a state from which we reach the current state", but initially
3773 * it is not the state->parent (i.e. "the state whose straight-line code leads
3774 * to the current state"), instead it is the state that happened to arrive at
3775 * a (prunable) equivalent of the current state. See comment above
3776 * do_propagate_liveness() for consequences of this.
3777 * This function is just a more efficient way of calling mark_reg_read() or
3778 * mark_stack_slot_read() on each reg in "parent" that is read in "state",
3779 * though it requires that parent != state->parent in the call arguments.
3781 static void propagate_liveness(const struct bpf_verifier_state
*state
,
3782 struct bpf_verifier_state
*parent
)
3784 while (do_propagate_liveness(state
, parent
)) {
3785 /* Something changed, so we need to feed those changes onward */
3787 parent
= state
->parent
;
3791 static int is_state_visited(struct bpf_verifier_env
*env
, int insn_idx
)
3793 struct bpf_verifier_state_list
*new_sl
;
3794 struct bpf_verifier_state_list
*sl
;
3795 struct bpf_verifier_state
*cur
= env
->cur_state
;
3798 sl
= env
->explored_states
[insn_idx
];
3800 /* this 'insn_idx' instruction wasn't marked, so we will not
3801 * be doing state search here
3805 while (sl
!= STATE_LIST_MARK
) {
3806 if (states_equal(env
, &sl
->state
, cur
)) {
3807 /* reached equivalent register/stack state,
3809 * Registers read by the continuation are read by us.
3810 * If we have any write marks in env->cur_state, they
3811 * will prevent corresponding reads in the continuation
3812 * from reaching our parent (an explored_state). Our
3813 * own state will get the read marks recorded, but
3814 * they'll be immediately forgotten as we're pruning
3815 * this state and will pop a new one.
3817 propagate_liveness(&sl
->state
, cur
);
3823 /* there were no equivalent states, remember current one.
3824 * technically the current state is not proven to be safe yet,
3825 * but it will either reach bpf_exit (which means it's safe) or
3826 * it will be rejected. Since there are no loops, we won't be
3827 * seeing this 'insn_idx' instruction again on the way to bpf_exit
3829 new_sl
= kzalloc(sizeof(struct bpf_verifier_state_list
), GFP_KERNEL
);
3833 /* add new state to the head of linked list */
3834 err
= copy_verifier_state(&new_sl
->state
, cur
);
3836 free_verifier_state(&new_sl
->state
, false);
3840 new_sl
->next
= env
->explored_states
[insn_idx
];
3841 env
->explored_states
[insn_idx
] = new_sl
;
3842 /* connect new state to parentage chain */
3843 cur
->parent
= &new_sl
->state
;
3844 /* clear write marks in current state: the writes we did are not writes
3845 * our child did, so they don't screen off its reads from us.
3846 * (There are no read marks in current state, because reads always mark
3847 * their parent and current state never has children yet. Only
3848 * explored_states can get read marks.)
3850 for (i
= 0; i
< BPF_REG_FP
; i
++)
3851 cur
->regs
[i
].live
= REG_LIVE_NONE
;
3852 for (i
= 0; i
< cur
->allocated_stack
/ BPF_REG_SIZE
; i
++)
3853 if (cur
->stack
[i
].slot_type
[0] == STACK_SPILL
)
3854 cur
->stack
[i
].spilled_ptr
.live
= REG_LIVE_NONE
;
3858 static int ext_analyzer_insn_hook(struct bpf_verifier_env
*env
,
3859 int insn_idx
, int prev_insn_idx
)
3861 if (env
->dev_ops
&& env
->dev_ops
->insn_hook
)
3862 return env
->dev_ops
->insn_hook(env
, insn_idx
, prev_insn_idx
);
3867 static int do_check(struct bpf_verifier_env
*env
)
3869 struct bpf_verifier_state
*state
;
3870 struct bpf_insn
*insns
= env
->prog
->insnsi
;
3871 struct bpf_reg_state
*regs
;
3872 int insn_cnt
= env
->prog
->len
;
3873 int insn_idx
, prev_insn_idx
= 0;
3874 int insn_processed
= 0;
3875 bool do_print_state
= false;
3877 state
= kzalloc(sizeof(struct bpf_verifier_state
), GFP_KERNEL
);
3880 env
->cur_state
= state
;
3881 init_reg_state(env
, state
->regs
);
3882 state
->parent
= NULL
;
3885 struct bpf_insn
*insn
;
3889 if (insn_idx
>= insn_cnt
) {
3890 verbose(env
, "invalid insn idx %d insn_cnt %d\n",
3891 insn_idx
, insn_cnt
);
3895 insn
= &insns
[insn_idx
];
3896 class = BPF_CLASS(insn
->code
);
3898 if (++insn_processed
> BPF_COMPLEXITY_LIMIT_INSNS
) {
3900 "BPF program is too large. Processed %d insn\n",
3905 err
= is_state_visited(env
, insn_idx
);
3909 /* found equivalent state, can prune the search */
3910 if (env
->log
.level
) {
3912 verbose(env
, "\nfrom %d to %d: safe\n",
3913 prev_insn_idx
, insn_idx
);
3915 verbose(env
, "%d: safe\n", insn_idx
);
3917 goto process_bpf_exit
;
3923 if (env
->log
.level
> 1 || (env
->log
.level
&& do_print_state
)) {
3924 if (env
->log
.level
> 1)
3925 verbose(env
, "%d:", insn_idx
);
3927 verbose(env
, "\nfrom %d to %d:",
3928 prev_insn_idx
, insn_idx
);
3929 print_verifier_state(env
, state
);
3930 do_print_state
= false;
3933 if (env
->log
.level
) {
3934 verbose(env
, "%d: ", insn_idx
);
3935 print_bpf_insn(verbose
, env
, insn
,
3936 env
->allow_ptr_leaks
);
3939 err
= ext_analyzer_insn_hook(env
, insn_idx
, prev_insn_idx
);
3943 regs
= cur_regs(env
);
3944 env
->insn_aux_data
[insn_idx
].seen
= true;
3945 if (class == BPF_ALU
|| class == BPF_ALU64
) {
3946 err
= check_alu_op(env
, insn
);
3950 } else if (class == BPF_LDX
) {
3951 enum bpf_reg_type
*prev_src_type
, src_reg_type
;
3953 /* check for reserved fields is already done */
3955 /* check src operand */
3956 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
3960 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP_NO_MARK
);
3964 src_reg_type
= regs
[insn
->src_reg
].type
;
3966 /* check that memory (src_reg + off) is readable,
3967 * the state of dst_reg will be updated by this func
3969 err
= check_mem_access(env
, insn_idx
, insn
->src_reg
, insn
->off
,
3970 BPF_SIZE(insn
->code
), BPF_READ
,
3971 insn
->dst_reg
, false);
3975 prev_src_type
= &env
->insn_aux_data
[insn_idx
].ptr_type
;
3977 if (*prev_src_type
== NOT_INIT
) {
3979 * dst_reg = *(u32 *)(src_reg + off)
3980 * save type to validate intersecting paths
3982 *prev_src_type
= src_reg_type
;
3984 } else if (src_reg_type
!= *prev_src_type
&&
3985 (src_reg_type
== PTR_TO_CTX
||
3986 *prev_src_type
== PTR_TO_CTX
)) {
3987 /* ABuser program is trying to use the same insn
3988 * dst_reg = *(u32*) (src_reg + off)
3989 * with different pointer types:
3990 * src_reg == ctx in one branch and
3991 * src_reg == stack|map in some other branch.
3994 verbose(env
, "same insn cannot be used with different pointers\n");
3998 } else if (class == BPF_STX
) {
3999 enum bpf_reg_type
*prev_dst_type
, dst_reg_type
;
4001 if (BPF_MODE(insn
->code
) == BPF_XADD
) {
4002 err
= check_xadd(env
, insn_idx
, insn
);
4009 /* check src1 operand */
4010 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
4013 /* check src2 operand */
4014 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
4018 dst_reg_type
= regs
[insn
->dst_reg
].type
;
4020 /* check that memory (dst_reg + off) is writeable */
4021 err
= check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
4022 BPF_SIZE(insn
->code
), BPF_WRITE
,
4023 insn
->src_reg
, false);
4027 prev_dst_type
= &env
->insn_aux_data
[insn_idx
].ptr_type
;
4029 if (*prev_dst_type
== NOT_INIT
) {
4030 *prev_dst_type
= dst_reg_type
;
4031 } else if (dst_reg_type
!= *prev_dst_type
&&
4032 (dst_reg_type
== PTR_TO_CTX
||
4033 *prev_dst_type
== PTR_TO_CTX
)) {
4034 verbose(env
, "same insn cannot be used with different pointers\n");
4038 } else if (class == BPF_ST
) {
4039 if (BPF_MODE(insn
->code
) != BPF_MEM
||
4040 insn
->src_reg
!= BPF_REG_0
) {
4041 verbose(env
, "BPF_ST uses reserved fields\n");
4044 /* check src operand */
4045 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
4049 if (is_ctx_reg(env
, insn
->dst_reg
)) {
4050 verbose(env
, "BPF_ST stores into R%d context is not allowed\n",
4055 /* check that memory (dst_reg + off) is writeable */
4056 err
= check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
4057 BPF_SIZE(insn
->code
), BPF_WRITE
,
4062 } else if (class == BPF_JMP
) {
4063 u8 opcode
= BPF_OP(insn
->code
);
4065 if (opcode
== BPF_CALL
) {
4066 if (BPF_SRC(insn
->code
) != BPF_K
||
4068 insn
->src_reg
!= BPF_REG_0
||
4069 insn
->dst_reg
!= BPF_REG_0
) {
4070 verbose(env
, "BPF_CALL uses reserved fields\n");
4074 err
= check_call(env
, insn
->imm
, insn_idx
);
4078 } else if (opcode
== BPF_JA
) {
4079 if (BPF_SRC(insn
->code
) != BPF_K
||
4081 insn
->src_reg
!= BPF_REG_0
||
4082 insn
->dst_reg
!= BPF_REG_0
) {
4083 verbose(env
, "BPF_JA uses reserved fields\n");
4087 insn_idx
+= insn
->off
+ 1;
4090 } else if (opcode
== BPF_EXIT
) {
4091 if (BPF_SRC(insn
->code
) != BPF_K
||
4093 insn
->src_reg
!= BPF_REG_0
||
4094 insn
->dst_reg
!= BPF_REG_0
) {
4095 verbose(env
, "BPF_EXIT uses reserved fields\n");
4099 /* eBPF calling convetion is such that R0 is used
4100 * to return the value from eBPF program.
4101 * Make sure that it's readable at this time
4102 * of bpf_exit, which means that program wrote
4103 * something into it earlier
4105 err
= check_reg_arg(env
, BPF_REG_0
, SRC_OP
);
4109 if (is_pointer_value(env
, BPF_REG_0
)) {
4110 verbose(env
, "R0 leaks addr as return value\n");
4114 err
= check_return_code(env
);
4118 err
= pop_stack(env
, &prev_insn_idx
, &insn_idx
);
4124 do_print_state
= true;
4128 err
= check_cond_jmp_op(env
, insn
, &insn_idx
);
4132 } else if (class == BPF_LD
) {
4133 u8 mode
= BPF_MODE(insn
->code
);
4135 if (mode
== BPF_ABS
|| mode
== BPF_IND
) {
4136 err
= check_ld_abs(env
, insn
);
4140 } else if (mode
== BPF_IMM
) {
4141 err
= check_ld_imm(env
, insn
);
4146 env
->insn_aux_data
[insn_idx
].seen
= true;
4148 verbose(env
, "invalid BPF_LD mode\n");
4152 verbose(env
, "unknown insn class %d\n", class);
4159 verbose(env
, "processed %d insns, stack depth %d\n", insn_processed
,
4160 env
->prog
->aux
->stack_depth
);
4164 static int check_map_prealloc(struct bpf_map
*map
)
4166 return (map
->map_type
!= BPF_MAP_TYPE_HASH
&&
4167 map
->map_type
!= BPF_MAP_TYPE_PERCPU_HASH
&&
4168 map
->map_type
!= BPF_MAP_TYPE_HASH_OF_MAPS
) ||
4169 !(map
->map_flags
& BPF_F_NO_PREALLOC
);
4172 static int check_map_prog_compatibility(struct bpf_verifier_env
*env
,
4173 struct bpf_map
*map
,
4174 struct bpf_prog
*prog
)
4177 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
4178 * preallocated hash maps, since doing memory allocation
4179 * in overflow_handler can crash depending on where nmi got
4182 if (prog
->type
== BPF_PROG_TYPE_PERF_EVENT
) {
4183 if (!check_map_prealloc(map
)) {
4184 verbose(env
, "perf_event programs can only use preallocated hash map\n");
4187 if (map
->inner_map_meta
&&
4188 !check_map_prealloc(map
->inner_map_meta
)) {
4189 verbose(env
, "perf_event programs can only use preallocated inner hash map\n");
4196 /* look for pseudo eBPF instructions that access map FDs and
4197 * replace them with actual map pointers
4199 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env
*env
)
4201 struct bpf_insn
*insn
= env
->prog
->insnsi
;
4202 int insn_cnt
= env
->prog
->len
;
4205 err
= bpf_prog_calc_tag(env
->prog
);
4209 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
4210 if (BPF_CLASS(insn
->code
) == BPF_LDX
&&
4211 (BPF_MODE(insn
->code
) != BPF_MEM
|| insn
->imm
!= 0)) {
4212 verbose(env
, "BPF_LDX uses reserved fields\n");
4216 if (BPF_CLASS(insn
->code
) == BPF_STX
&&
4217 ((BPF_MODE(insn
->code
) != BPF_MEM
&&
4218 BPF_MODE(insn
->code
) != BPF_XADD
) || insn
->imm
!= 0)) {
4219 verbose(env
, "BPF_STX uses reserved fields\n");
4223 if (insn
[0].code
== (BPF_LD
| BPF_IMM
| BPF_DW
)) {
4224 struct bpf_map
*map
;
4227 if (i
== insn_cnt
- 1 || insn
[1].code
!= 0 ||
4228 insn
[1].dst_reg
!= 0 || insn
[1].src_reg
!= 0 ||
4230 verbose(env
, "invalid bpf_ld_imm64 insn\n");
4234 if (insn
->src_reg
== 0)
4235 /* valid generic load 64-bit imm */
4238 if (insn
->src_reg
!= BPF_PSEUDO_MAP_FD
) {
4240 "unrecognized bpf_ld_imm64 insn\n");
4244 f
= fdget(insn
->imm
);
4245 map
= __bpf_map_get(f
);
4247 verbose(env
, "fd %d is not pointing to valid bpf_map\n",
4249 return PTR_ERR(map
);
4252 err
= check_map_prog_compatibility(env
, map
, env
->prog
);
4258 /* store map pointer inside BPF_LD_IMM64 instruction */
4259 insn
[0].imm
= (u32
) (unsigned long) map
;
4260 insn
[1].imm
= ((u64
) (unsigned long) map
) >> 32;
4262 /* check whether we recorded this map already */
4263 for (j
= 0; j
< env
->used_map_cnt
; j
++)
4264 if (env
->used_maps
[j
] == map
) {
4269 if (env
->used_map_cnt
>= MAX_USED_MAPS
) {
4274 /* hold the map. If the program is rejected by verifier,
4275 * the map will be released by release_maps() or it
4276 * will be used by the valid program until it's unloaded
4277 * and all maps are released in free_bpf_prog_info()
4279 map
= bpf_map_inc(map
, false);
4282 return PTR_ERR(map
);
4284 env
->used_maps
[env
->used_map_cnt
++] = map
;
4293 /* now all pseudo BPF_LD_IMM64 instructions load valid
4294 * 'struct bpf_map *' into a register instead of user map_fd.
4295 * These pointers will be used later by verifier to validate map access.
4300 /* drop refcnt of maps used by the rejected program */
4301 static void release_maps(struct bpf_verifier_env
*env
)
4305 for (i
= 0; i
< env
->used_map_cnt
; i
++)
4306 bpf_map_put(env
->used_maps
[i
]);
4309 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
4310 static void convert_pseudo_ld_imm64(struct bpf_verifier_env
*env
)
4312 struct bpf_insn
*insn
= env
->prog
->insnsi
;
4313 int insn_cnt
= env
->prog
->len
;
4316 for (i
= 0; i
< insn_cnt
; i
++, insn
++)
4317 if (insn
->code
== (BPF_LD
| BPF_IMM
| BPF_DW
))
4321 /* single env->prog->insni[off] instruction was replaced with the range
4322 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
4323 * [0, off) and [off, end) to new locations, so the patched range stays zero
4325 static int adjust_insn_aux_data(struct bpf_verifier_env
*env
, u32 prog_len
,
4328 struct bpf_insn_aux_data
*new_data
, *old_data
= env
->insn_aux_data
;
4333 new_data
= vzalloc(sizeof(struct bpf_insn_aux_data
) * prog_len
);
4336 memcpy(new_data
, old_data
, sizeof(struct bpf_insn_aux_data
) * off
);
4337 memcpy(new_data
+ off
+ cnt
- 1, old_data
+ off
,
4338 sizeof(struct bpf_insn_aux_data
) * (prog_len
- off
- cnt
+ 1));
4339 for (i
= off
; i
< off
+ cnt
- 1; i
++)
4340 new_data
[i
].seen
= true;
4341 env
->insn_aux_data
= new_data
;
4346 static struct bpf_prog
*bpf_patch_insn_data(struct bpf_verifier_env
*env
, u32 off
,
4347 const struct bpf_insn
*patch
, u32 len
)
4349 struct bpf_prog
*new_prog
;
4351 new_prog
= bpf_patch_insn_single(env
->prog
, off
, patch
, len
);
4354 if (adjust_insn_aux_data(env
, new_prog
->len
, off
, len
))
4359 /* The verifier does more data flow analysis than llvm and will not explore
4360 * branches that are dead at run time. Malicious programs can have dead code
4361 * too. Therefore replace all dead at-run-time code with nops.
4363 static void sanitize_dead_code(struct bpf_verifier_env
*env
)
4365 struct bpf_insn_aux_data
*aux_data
= env
->insn_aux_data
;
4366 struct bpf_insn nop
= BPF_MOV64_REG(BPF_REG_0
, BPF_REG_0
);
4367 struct bpf_insn
*insn
= env
->prog
->insnsi
;
4368 const int insn_cnt
= env
->prog
->len
;
4371 for (i
= 0; i
< insn_cnt
; i
++) {
4372 if (aux_data
[i
].seen
)
4374 memcpy(insn
+ i
, &nop
, sizeof(nop
));
4378 /* convert load instructions that access fields of 'struct __sk_buff'
4379 * into sequence of instructions that access fields of 'struct sk_buff'
4381 static int convert_ctx_accesses(struct bpf_verifier_env
*env
)
4383 const struct bpf_verifier_ops
*ops
= env
->ops
;
4384 int i
, cnt
, size
, ctx_field_size
, delta
= 0;
4385 const int insn_cnt
= env
->prog
->len
;
4386 struct bpf_insn insn_buf
[16], *insn
;
4387 struct bpf_prog
*new_prog
;
4388 enum bpf_access_type type
;
4389 bool is_narrower_load
;
4392 if (ops
->gen_prologue
) {
4393 cnt
= ops
->gen_prologue(insn_buf
, env
->seen_direct_write
,
4395 if (cnt
>= ARRAY_SIZE(insn_buf
)) {
4396 verbose(env
, "bpf verifier is misconfigured\n");
4399 new_prog
= bpf_patch_insn_data(env
, 0, insn_buf
, cnt
);
4403 env
->prog
= new_prog
;
4408 if (!ops
->convert_ctx_access
)
4411 insn
= env
->prog
->insnsi
+ delta
;
4413 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
4414 if (insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_B
) ||
4415 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_H
) ||
4416 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_W
) ||
4417 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_DW
))
4419 else if (insn
->code
== (BPF_STX
| BPF_MEM
| BPF_B
) ||
4420 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_H
) ||
4421 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_W
) ||
4422 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_DW
))
4427 if (type
== BPF_WRITE
&&
4428 env
->insn_aux_data
[i
+ delta
].sanitize_stack_off
) {
4429 struct bpf_insn patch
[] = {
4430 /* Sanitize suspicious stack slot with zero.
4431 * There are no memory dependencies for this store,
4432 * since it's only using frame pointer and immediate
4435 BPF_ST_MEM(BPF_DW
, BPF_REG_FP
,
4436 env
->insn_aux_data
[i
+ delta
].sanitize_stack_off
,
4438 /* the original STX instruction will immediately
4439 * overwrite the same stack slot with appropriate value
4444 cnt
= ARRAY_SIZE(patch
);
4445 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, patch
, cnt
);
4450 env
->prog
= new_prog
;
4451 insn
= new_prog
->insnsi
+ i
+ delta
;
4455 if (env
->insn_aux_data
[i
+ delta
].ptr_type
!= PTR_TO_CTX
)
4458 ctx_field_size
= env
->insn_aux_data
[i
+ delta
].ctx_field_size
;
4459 size
= BPF_LDST_BYTES(insn
);
4461 /* If the read access is a narrower load of the field,
4462 * convert to a 4/8-byte load, to minimum program type specific
4463 * convert_ctx_access changes. If conversion is successful,
4464 * we will apply proper mask to the result.
4466 is_narrower_load
= size
< ctx_field_size
;
4467 if (is_narrower_load
) {
4468 u32 off
= insn
->off
;
4471 if (type
== BPF_WRITE
) {
4472 verbose(env
, "bpf verifier narrow ctx access misconfigured\n");
4477 if (ctx_field_size
== 4)
4479 else if (ctx_field_size
== 8)
4482 insn
->off
= off
& ~(ctx_field_size
- 1);
4483 insn
->code
= BPF_LDX
| BPF_MEM
| size_code
;
4487 cnt
= ops
->convert_ctx_access(type
, insn
, insn_buf
, env
->prog
,
4489 if (cnt
== 0 || cnt
>= ARRAY_SIZE(insn_buf
) ||
4490 (ctx_field_size
&& !target_size
)) {
4491 verbose(env
, "bpf verifier is misconfigured\n");
4495 if (is_narrower_load
&& size
< target_size
) {
4496 if (ctx_field_size
<= 4)
4497 insn_buf
[cnt
++] = BPF_ALU32_IMM(BPF_AND
, insn
->dst_reg
,
4498 (1 << size
* 8) - 1);
4500 insn_buf
[cnt
++] = BPF_ALU64_IMM(BPF_AND
, insn
->dst_reg
,
4501 (1 << size
* 8) - 1);
4504 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
, cnt
);
4510 /* keep walking new program and skip insns we just inserted */
4511 env
->prog
= new_prog
;
4512 insn
= new_prog
->insnsi
+ i
+ delta
;
4518 /* fixup insn->imm field of bpf_call instructions
4519 * and inline eligible helpers as explicit sequence of BPF instructions
4521 * this function is called after eBPF program passed verification
4523 static int fixup_bpf_calls(struct bpf_verifier_env
*env
)
4525 struct bpf_prog
*prog
= env
->prog
;
4526 struct bpf_insn
*insn
= prog
->insnsi
;
4527 const struct bpf_func_proto
*fn
;
4528 const int insn_cnt
= prog
->len
;
4529 struct bpf_insn insn_buf
[16];
4530 struct bpf_prog
*new_prog
;
4531 struct bpf_map
*map_ptr
;
4532 int i
, cnt
, delta
= 0;
4534 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
4535 if (insn
->code
== (BPF_ALU
| BPF_MOD
| BPF_X
) ||
4536 insn
->code
== (BPF_ALU
| BPF_DIV
| BPF_X
)) {
4537 /* due to JIT bugs clear upper 32-bits of src register
4538 * before div/mod operation
4540 insn_buf
[0] = BPF_MOV32_REG(insn
->src_reg
, insn
->src_reg
);
4541 insn_buf
[1] = *insn
;
4543 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
, cnt
);
4548 env
->prog
= prog
= new_prog
;
4549 insn
= new_prog
->insnsi
+ i
+ delta
;
4553 if (insn
->code
!= (BPF_JMP
| BPF_CALL
))
4556 if (insn
->imm
== BPF_FUNC_get_route_realm
)
4557 prog
->dst_needed
= 1;
4558 if (insn
->imm
== BPF_FUNC_get_prandom_u32
)
4559 bpf_user_rnd_init_once();
4560 if (insn
->imm
== BPF_FUNC_tail_call
) {
4561 /* If we tail call into other programs, we
4562 * cannot make any assumptions since they can
4563 * be replaced dynamically during runtime in
4564 * the program array.
4566 prog
->cb_access
= 1;
4567 env
->prog
->aux
->stack_depth
= MAX_BPF_STACK
;
4569 /* mark bpf_tail_call as different opcode to avoid
4570 * conditional branch in the interpeter for every normal
4571 * call and to prevent accidental JITing by JIT compiler
4572 * that doesn't support bpf_tail_call yet
4575 insn
->code
= BPF_JMP
| BPF_TAIL_CALL
;
4577 /* instead of changing every JIT dealing with tail_call
4578 * emit two extra insns:
4579 * if (index >= max_entries) goto out;
4580 * index &= array->index_mask;
4581 * to avoid out-of-bounds cpu speculation
4583 map_ptr
= env
->insn_aux_data
[i
+ delta
].map_ptr
;
4584 if (map_ptr
== BPF_MAP_PTR_POISON
) {
4585 verbose(env
, "tail_call abusing map_ptr\n");
4588 if (!map_ptr
->unpriv_array
)
4590 insn_buf
[0] = BPF_JMP_IMM(BPF_JGE
, BPF_REG_3
,
4591 map_ptr
->max_entries
, 2);
4592 insn_buf
[1] = BPF_ALU32_IMM(BPF_AND
, BPF_REG_3
,
4593 container_of(map_ptr
,
4596 insn_buf
[2] = *insn
;
4598 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
, cnt
);
4603 env
->prog
= prog
= new_prog
;
4604 insn
= new_prog
->insnsi
+ i
+ delta
;
4608 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
4609 * handlers are currently limited to 64 bit only.
4611 if (ebpf_jit_enabled() && BITS_PER_LONG
== 64 &&
4612 insn
->imm
== BPF_FUNC_map_lookup_elem
) {
4613 map_ptr
= env
->insn_aux_data
[i
+ delta
].map_ptr
;
4614 if (map_ptr
== BPF_MAP_PTR_POISON
||
4615 !map_ptr
->ops
->map_gen_lookup
)
4616 goto patch_call_imm
;
4618 cnt
= map_ptr
->ops
->map_gen_lookup(map_ptr
, insn_buf
);
4619 if (cnt
== 0 || cnt
>= ARRAY_SIZE(insn_buf
)) {
4620 verbose(env
, "bpf verifier is misconfigured\n");
4624 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
,
4631 /* keep walking new program and skip insns we just inserted */
4632 env
->prog
= prog
= new_prog
;
4633 insn
= new_prog
->insnsi
+ i
+ delta
;
4637 if (insn
->imm
== BPF_FUNC_redirect_map
) {
4638 /* Note, we cannot use prog directly as imm as subsequent
4639 * rewrites would still change the prog pointer. The only
4640 * stable address we can use is aux, which also works with
4641 * prog clones during blinding.
4643 u64 addr
= (unsigned long)prog
->aux
;
4644 struct bpf_insn r4_ld
[] = {
4645 BPF_LD_IMM64(BPF_REG_4
, addr
),
4648 cnt
= ARRAY_SIZE(r4_ld
);
4650 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, r4_ld
, cnt
);
4655 env
->prog
= prog
= new_prog
;
4656 insn
= new_prog
->insnsi
+ i
+ delta
;
4659 fn
= env
->ops
->get_func_proto(insn
->imm
);
4660 /* all functions that have prototype and verifier allowed
4661 * programs to call them, must be real in-kernel functions
4665 "kernel subsystem misconfigured func %s#%d\n",
4666 func_id_name(insn
->imm
), insn
->imm
);
4669 insn
->imm
= fn
->func
- __bpf_call_base
;
4675 static void free_states(struct bpf_verifier_env
*env
)
4677 struct bpf_verifier_state_list
*sl
, *sln
;
4680 if (!env
->explored_states
)
4683 for (i
= 0; i
< env
->prog
->len
; i
++) {
4684 sl
= env
->explored_states
[i
];
4687 while (sl
!= STATE_LIST_MARK
) {
4689 free_verifier_state(&sl
->state
, false);
4695 kfree(env
->explored_states
);
4698 int bpf_check(struct bpf_prog
**prog
, union bpf_attr
*attr
)
4700 struct bpf_verifier_env
*env
;
4701 struct bpf_verifer_log
*log
;
4704 /* no program is valid */
4705 if (ARRAY_SIZE(bpf_verifier_ops
) == 0)
4708 /* 'struct bpf_verifier_env' can be global, but since it's not small,
4709 * allocate/free it every time bpf_check() is called
4711 env
= kzalloc(sizeof(struct bpf_verifier_env
), GFP_KERNEL
);
4716 env
->insn_aux_data
= vzalloc(sizeof(struct bpf_insn_aux_data
) *
4719 if (!env
->insn_aux_data
)
4722 env
->ops
= bpf_verifier_ops
[env
->prog
->type
];
4724 /* grab the mutex to protect few globals used by verifier */
4725 mutex_lock(&bpf_verifier_lock
);
4727 if (attr
->log_level
|| attr
->log_buf
|| attr
->log_size
) {
4728 /* user requested verbose verifier output
4729 * and supplied buffer to store the verification trace
4731 log
->level
= attr
->log_level
;
4732 log
->ubuf
= (char __user
*) (unsigned long) attr
->log_buf
;
4733 log
->len_total
= attr
->log_size
;
4736 /* log attributes have to be sane */
4737 if (log
->len_total
< 128 || log
->len_total
> UINT_MAX
>> 8 ||
4738 !log
->level
|| !log
->ubuf
)
4742 env
->strict_alignment
= !!(attr
->prog_flags
& BPF_F_STRICT_ALIGNMENT
);
4743 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
))
4744 env
->strict_alignment
= true;
4746 if (env
->prog
->aux
->offload
) {
4747 ret
= bpf_prog_offload_verifier_prep(env
);
4752 ret
= replace_map_fd_with_map_ptr(env
);
4754 goto skip_full_check
;
4756 env
->explored_states
= kcalloc(env
->prog
->len
,
4757 sizeof(struct bpf_verifier_state_list
*),
4760 if (!env
->explored_states
)
4761 goto skip_full_check
;
4763 ret
= check_cfg(env
);
4765 goto skip_full_check
;
4767 env
->allow_ptr_leaks
= capable(CAP_SYS_ADMIN
);
4769 ret
= do_check(env
);
4770 if (env
->cur_state
) {
4771 free_verifier_state(env
->cur_state
, true);
4772 env
->cur_state
= NULL
;
4776 while (!pop_stack(env
, NULL
, NULL
));
4780 sanitize_dead_code(env
);
4783 /* program is valid, convert *(u32*)(ctx + off) accesses */
4784 ret
= convert_ctx_accesses(env
);
4787 ret
= fixup_bpf_calls(env
);
4789 if (log
->level
&& bpf_verifier_log_full(log
))
4791 if (log
->level
&& !log
->ubuf
) {
4793 goto err_release_maps
;
4796 if (ret
== 0 && env
->used_map_cnt
) {
4797 /* if program passed verifier, update used_maps in bpf_prog_info */
4798 env
->prog
->aux
->used_maps
= kmalloc_array(env
->used_map_cnt
,
4799 sizeof(env
->used_maps
[0]),
4802 if (!env
->prog
->aux
->used_maps
) {
4804 goto err_release_maps
;
4807 memcpy(env
->prog
->aux
->used_maps
, env
->used_maps
,
4808 sizeof(env
->used_maps
[0]) * env
->used_map_cnt
);
4809 env
->prog
->aux
->used_map_cnt
= env
->used_map_cnt
;
4811 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
4812 * bpf_ld_imm64 instructions
4814 convert_pseudo_ld_imm64(env
);
4818 if (!env
->prog
->aux
->used_maps
)
4819 /* if we didn't copy map pointers into bpf_prog_info, release
4820 * them now. Otherwise free_bpf_prog_info() will release them.
4825 mutex_unlock(&bpf_verifier_lock
);
4826 vfree(env
->insn_aux_data
);