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>
23 #include <linux/bsearch.h>
24 #include <linux/sort.h>
25 #include <linux/perf_event.h>
29 static const struct bpf_verifier_ops
* const bpf_verifier_ops
[] = {
30 #define BPF_PROG_TYPE(_id, _name) \
31 [_id] = & _name ## _verifier_ops,
32 #define BPF_MAP_TYPE(_id, _ops)
33 #include <linux/bpf_types.h>
38 /* bpf_check() is a static code analyzer that walks eBPF program
39 * instruction by instruction and updates register/stack state.
40 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
42 * The first pass is depth-first-search to check that the program is a DAG.
43 * It rejects the following programs:
44 * - larger than BPF_MAXINSNS insns
45 * - if loop is present (detected via back-edge)
46 * - unreachable insns exist (shouldn't be a forest. program = one function)
47 * - out of bounds or malformed jumps
48 * The second pass is all possible path descent from the 1st insn.
49 * Since it's analyzing all pathes through the program, the length of the
50 * analysis is limited to 64k insn, which may be hit even if total number of
51 * insn is less then 4K, but there are too many branches that change stack/regs.
52 * Number of 'branches to be analyzed' is limited to 1k
54 * On entry to each instruction, each register has a type, and the instruction
55 * changes the types of the registers depending on instruction semantics.
56 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
59 * All registers are 64-bit.
60 * R0 - return register
61 * R1-R5 argument passing registers
62 * R6-R9 callee saved registers
63 * R10 - frame pointer read-only
65 * At the start of BPF program the register R1 contains a pointer to bpf_context
66 * and has type PTR_TO_CTX.
68 * Verifier tracks arithmetic operations on pointers in case:
69 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
70 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
71 * 1st insn copies R10 (which has FRAME_PTR) type into R1
72 * and 2nd arithmetic instruction is pattern matched to recognize
73 * that it wants to construct a pointer to some element within stack.
74 * So after 2nd insn, the register R1 has type PTR_TO_STACK
75 * (and -20 constant is saved for further stack bounds checking).
76 * Meaning that this reg is a pointer to stack plus known immediate constant.
78 * Most of the time the registers have SCALAR_VALUE type, which
79 * means the register has some value, but it's not a valid pointer.
80 * (like pointer plus pointer becomes SCALAR_VALUE type)
82 * When verifier sees load or store instructions the type of base register
83 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK. These are three pointer
84 * types recognized by check_mem_access() function.
86 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
87 * and the range of [ptr, ptr + map's value_size) is accessible.
89 * registers used to pass values to function calls are checked against
90 * function argument constraints.
92 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
93 * It means that the register type passed to this function must be
94 * PTR_TO_STACK and it will be used inside the function as
95 * 'pointer to map element key'
97 * For example the argument constraints for bpf_map_lookup_elem():
98 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
99 * .arg1_type = ARG_CONST_MAP_PTR,
100 * .arg2_type = ARG_PTR_TO_MAP_KEY,
102 * ret_type says that this function returns 'pointer to map elem value or null'
103 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
104 * 2nd argument should be a pointer to stack, which will be used inside
105 * the helper function as a pointer to map element key.
107 * On the kernel side the helper function looks like:
108 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
110 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
111 * void *key = (void *) (unsigned long) r2;
114 * here kernel can access 'key' and 'map' pointers safely, knowing that
115 * [key, key + map->key_size) bytes are valid and were initialized on
116 * the stack of eBPF program.
119 * Corresponding eBPF program may look like:
120 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
121 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
122 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
123 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
124 * here verifier looks at prototype of map_lookup_elem() and sees:
125 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
126 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
128 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
129 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
130 * and were initialized prior to this call.
131 * If it's ok, then verifier allows this BPF_CALL insn and looks at
132 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
133 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
134 * returns ether pointer to map value or NULL.
136 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
137 * insn, the register holding that pointer in the true branch changes state to
138 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
139 * branch. See check_cond_jmp_op().
141 * After the call R0 is set to return type of the function and registers R1-R5
142 * are set to NOT_INIT to indicate that they are no longer readable.
145 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
146 struct bpf_verifier_stack_elem
{
147 /* verifer state is 'st'
148 * before processing instruction 'insn_idx'
149 * and after processing instruction 'prev_insn_idx'
151 struct bpf_verifier_state st
;
154 struct bpf_verifier_stack_elem
*next
;
157 #define BPF_COMPLEXITY_LIMIT_INSNS 131072
158 #define BPF_COMPLEXITY_LIMIT_STACK 1024
160 #define BPF_MAP_PTR_UNPRIV 1UL
161 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
162 POISON_POINTER_DELTA))
163 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
165 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data
*aux
)
167 return BPF_MAP_PTR(aux
->map_state
) == BPF_MAP_PTR_POISON
;
170 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data
*aux
)
172 return aux
->map_state
& BPF_MAP_PTR_UNPRIV
;
175 static void bpf_map_ptr_store(struct bpf_insn_aux_data
*aux
,
176 const struct bpf_map
*map
, bool unpriv
)
178 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON
& BPF_MAP_PTR_UNPRIV
);
179 unpriv
|= bpf_map_ptr_unpriv(aux
);
180 aux
->map_state
= (unsigned long)map
|
181 (unpriv
? BPF_MAP_PTR_UNPRIV
: 0UL);
184 struct bpf_call_arg_meta
{
185 struct bpf_map
*map_ptr
;
190 s64 msize_smax_value
;
191 u64 msize_umax_value
;
194 static DEFINE_MUTEX(bpf_verifier_lock
);
196 void bpf_verifier_vlog(struct bpf_verifier_log
*log
, const char *fmt
,
201 n
= vscnprintf(log
->kbuf
, BPF_VERIFIER_TMP_LOG_SIZE
, fmt
, args
);
203 WARN_ONCE(n
>= BPF_VERIFIER_TMP_LOG_SIZE
- 1,
204 "verifier log line truncated - local buffer too short\n");
206 n
= min(log
->len_total
- log
->len_used
- 1, n
);
209 if (!copy_to_user(log
->ubuf
+ log
->len_used
, log
->kbuf
, n
+ 1))
215 /* log_level controls verbosity level of eBPF verifier.
216 * bpf_verifier_log_write() is used to dump the verification trace to the log,
217 * so the user can figure out what's wrong with the program
219 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env
*env
,
220 const char *fmt
, ...)
224 if (!bpf_verifier_log_needed(&env
->log
))
228 bpf_verifier_vlog(&env
->log
, fmt
, args
);
231 EXPORT_SYMBOL_GPL(bpf_verifier_log_write
);
233 __printf(2, 3) static void verbose(void *private_data
, const char *fmt
, ...)
235 struct bpf_verifier_env
*env
= private_data
;
238 if (!bpf_verifier_log_needed(&env
->log
))
242 bpf_verifier_vlog(&env
->log
, fmt
, args
);
246 static bool type_is_pkt_pointer(enum bpf_reg_type type
)
248 return type
== PTR_TO_PACKET
||
249 type
== PTR_TO_PACKET_META
;
252 /* string representation of 'enum bpf_reg_type' */
253 static const char * const reg_type_str
[] = {
255 [SCALAR_VALUE
] = "inv",
256 [PTR_TO_CTX
] = "ctx",
257 [CONST_PTR_TO_MAP
] = "map_ptr",
258 [PTR_TO_MAP_VALUE
] = "map_value",
259 [PTR_TO_MAP_VALUE_OR_NULL
] = "map_value_or_null",
260 [PTR_TO_STACK
] = "fp",
261 [PTR_TO_PACKET
] = "pkt",
262 [PTR_TO_PACKET_META
] = "pkt_meta",
263 [PTR_TO_PACKET_END
] = "pkt_end",
266 static void print_liveness(struct bpf_verifier_env
*env
,
267 enum bpf_reg_liveness live
)
269 if (live
& (REG_LIVE_READ
| REG_LIVE_WRITTEN
))
271 if (live
& REG_LIVE_READ
)
273 if (live
& REG_LIVE_WRITTEN
)
277 static struct bpf_func_state
*func(struct bpf_verifier_env
*env
,
278 const struct bpf_reg_state
*reg
)
280 struct bpf_verifier_state
*cur
= env
->cur_state
;
282 return cur
->frame
[reg
->frameno
];
285 static void print_verifier_state(struct bpf_verifier_env
*env
,
286 const struct bpf_func_state
*state
)
288 const struct bpf_reg_state
*reg
;
293 verbose(env
, " frame%d:", state
->frameno
);
294 for (i
= 0; i
< MAX_BPF_REG
; i
++) {
295 reg
= &state
->regs
[i
];
299 verbose(env
, " R%d", i
);
300 print_liveness(env
, reg
->live
);
301 verbose(env
, "=%s", reg_type_str
[t
]);
302 if ((t
== SCALAR_VALUE
|| t
== PTR_TO_STACK
) &&
303 tnum_is_const(reg
->var_off
)) {
304 /* reg->off should be 0 for SCALAR_VALUE */
305 verbose(env
, "%lld", reg
->var_off
.value
+ reg
->off
);
306 if (t
== PTR_TO_STACK
)
307 verbose(env
, ",call_%d", func(env
, reg
)->callsite
);
309 verbose(env
, "(id=%d", reg
->id
);
310 if (t
!= SCALAR_VALUE
)
311 verbose(env
, ",off=%d", reg
->off
);
312 if (type_is_pkt_pointer(t
))
313 verbose(env
, ",r=%d", reg
->range
);
314 else if (t
== CONST_PTR_TO_MAP
||
315 t
== PTR_TO_MAP_VALUE
||
316 t
== PTR_TO_MAP_VALUE_OR_NULL
)
317 verbose(env
, ",ks=%d,vs=%d",
318 reg
->map_ptr
->key_size
,
319 reg
->map_ptr
->value_size
);
320 if (tnum_is_const(reg
->var_off
)) {
321 /* Typically an immediate SCALAR_VALUE, but
322 * could be a pointer whose offset is too big
325 verbose(env
, ",imm=%llx", reg
->var_off
.value
);
327 if (reg
->smin_value
!= reg
->umin_value
&&
328 reg
->smin_value
!= S64_MIN
)
329 verbose(env
, ",smin_value=%lld",
330 (long long)reg
->smin_value
);
331 if (reg
->smax_value
!= reg
->umax_value
&&
332 reg
->smax_value
!= S64_MAX
)
333 verbose(env
, ",smax_value=%lld",
334 (long long)reg
->smax_value
);
335 if (reg
->umin_value
!= 0)
336 verbose(env
, ",umin_value=%llu",
337 (unsigned long long)reg
->umin_value
);
338 if (reg
->umax_value
!= U64_MAX
)
339 verbose(env
, ",umax_value=%llu",
340 (unsigned long long)reg
->umax_value
);
341 if (!tnum_is_unknown(reg
->var_off
)) {
344 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
345 verbose(env
, ",var_off=%s", tn_buf
);
351 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
352 if (state
->stack
[i
].slot_type
[0] == STACK_SPILL
) {
353 verbose(env
, " fp%d",
354 (-i
- 1) * BPF_REG_SIZE
);
355 print_liveness(env
, state
->stack
[i
].spilled_ptr
.live
);
357 reg_type_str
[state
->stack
[i
].spilled_ptr
.type
]);
359 if (state
->stack
[i
].slot_type
[0] == STACK_ZERO
)
360 verbose(env
, " fp%d=0", (-i
- 1) * BPF_REG_SIZE
);
365 static int copy_stack_state(struct bpf_func_state
*dst
,
366 const struct bpf_func_state
*src
)
370 if (WARN_ON_ONCE(dst
->allocated_stack
< src
->allocated_stack
)) {
371 /* internal bug, make state invalid to reject the program */
372 memset(dst
, 0, sizeof(*dst
));
375 memcpy(dst
->stack
, src
->stack
,
376 sizeof(*src
->stack
) * (src
->allocated_stack
/ BPF_REG_SIZE
));
380 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
381 * make it consume minimal amount of memory. check_stack_write() access from
382 * the program calls into realloc_func_state() to grow the stack size.
383 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
384 * which this function copies over. It points to previous bpf_verifier_state
385 * which is never reallocated
387 static int realloc_func_state(struct bpf_func_state
*state
, int size
,
390 u32 old_size
= state
->allocated_stack
;
391 struct bpf_stack_state
*new_stack
;
392 int slot
= size
/ BPF_REG_SIZE
;
394 if (size
<= old_size
|| !size
) {
397 state
->allocated_stack
= slot
* BPF_REG_SIZE
;
398 if (!size
&& old_size
) {
404 new_stack
= kmalloc_array(slot
, sizeof(struct bpf_stack_state
),
410 memcpy(new_stack
, state
->stack
,
411 sizeof(*new_stack
) * (old_size
/ BPF_REG_SIZE
));
412 memset(new_stack
+ old_size
/ BPF_REG_SIZE
, 0,
413 sizeof(*new_stack
) * (size
- old_size
) / BPF_REG_SIZE
);
415 state
->allocated_stack
= slot
* BPF_REG_SIZE
;
417 state
->stack
= new_stack
;
421 static void free_func_state(struct bpf_func_state
*state
)
429 static void free_verifier_state(struct bpf_verifier_state
*state
,
434 for (i
= 0; i
<= state
->curframe
; i
++) {
435 free_func_state(state
->frame
[i
]);
436 state
->frame
[i
] = NULL
;
442 /* copy verifier state from src to dst growing dst stack space
443 * when necessary to accommodate larger src stack
445 static int copy_func_state(struct bpf_func_state
*dst
,
446 const struct bpf_func_state
*src
)
450 err
= realloc_func_state(dst
, src
->allocated_stack
, false);
453 memcpy(dst
, src
, offsetof(struct bpf_func_state
, allocated_stack
));
454 return copy_stack_state(dst
, src
);
457 static int copy_verifier_state(struct bpf_verifier_state
*dst_state
,
458 const struct bpf_verifier_state
*src
)
460 struct bpf_func_state
*dst
;
463 /* if dst has more stack frames then src frame, free them */
464 for (i
= src
->curframe
+ 1; i
<= dst_state
->curframe
; i
++) {
465 free_func_state(dst_state
->frame
[i
]);
466 dst_state
->frame
[i
] = NULL
;
468 dst_state
->curframe
= src
->curframe
;
469 dst_state
->parent
= src
->parent
;
470 for (i
= 0; i
<= src
->curframe
; i
++) {
471 dst
= dst_state
->frame
[i
];
473 dst
= kzalloc(sizeof(*dst
), GFP_KERNEL
);
476 dst_state
->frame
[i
] = dst
;
478 err
= copy_func_state(dst
, src
->frame
[i
]);
485 static int pop_stack(struct bpf_verifier_env
*env
, int *prev_insn_idx
,
488 struct bpf_verifier_state
*cur
= env
->cur_state
;
489 struct bpf_verifier_stack_elem
*elem
, *head
= env
->head
;
492 if (env
->head
== NULL
)
496 err
= copy_verifier_state(cur
, &head
->st
);
501 *insn_idx
= head
->insn_idx
;
503 *prev_insn_idx
= head
->prev_insn_idx
;
505 free_verifier_state(&head
->st
, false);
512 static struct bpf_verifier_state
*push_stack(struct bpf_verifier_env
*env
,
513 int insn_idx
, int prev_insn_idx
)
515 struct bpf_verifier_state
*cur
= env
->cur_state
;
516 struct bpf_verifier_stack_elem
*elem
;
519 elem
= kzalloc(sizeof(struct bpf_verifier_stack_elem
), GFP_KERNEL
);
523 elem
->insn_idx
= insn_idx
;
524 elem
->prev_insn_idx
= prev_insn_idx
;
525 elem
->next
= env
->head
;
528 err
= copy_verifier_state(&elem
->st
, cur
);
531 if (env
->stack_size
> BPF_COMPLEXITY_LIMIT_STACK
) {
532 verbose(env
, "BPF program is too complex\n");
537 free_verifier_state(env
->cur_state
, true);
538 env
->cur_state
= NULL
;
539 /* pop all elements and return */
540 while (!pop_stack(env
, NULL
, NULL
));
544 #define CALLER_SAVED_REGS 6
545 static const int caller_saved
[CALLER_SAVED_REGS
] = {
546 BPF_REG_0
, BPF_REG_1
, BPF_REG_2
, BPF_REG_3
, BPF_REG_4
, BPF_REG_5
549 static void __mark_reg_not_init(struct bpf_reg_state
*reg
);
551 /* Mark the unknown part of a register (variable offset or scalar value) as
552 * known to have the value @imm.
554 static void __mark_reg_known(struct bpf_reg_state
*reg
, u64 imm
)
557 reg
->var_off
= tnum_const(imm
);
558 reg
->smin_value
= (s64
)imm
;
559 reg
->smax_value
= (s64
)imm
;
560 reg
->umin_value
= imm
;
561 reg
->umax_value
= imm
;
564 /* Mark the 'variable offset' part of a register as zero. This should be
565 * used only on registers holding a pointer type.
567 static void __mark_reg_known_zero(struct bpf_reg_state
*reg
)
569 __mark_reg_known(reg
, 0);
572 static void __mark_reg_const_zero(struct bpf_reg_state
*reg
)
574 __mark_reg_known(reg
, 0);
576 reg
->type
= SCALAR_VALUE
;
579 static void mark_reg_known_zero(struct bpf_verifier_env
*env
,
580 struct bpf_reg_state
*regs
, u32 regno
)
582 if (WARN_ON(regno
>= MAX_BPF_REG
)) {
583 verbose(env
, "mark_reg_known_zero(regs, %u)\n", regno
);
584 /* Something bad happened, let's kill all regs */
585 for (regno
= 0; regno
< MAX_BPF_REG
; regno
++)
586 __mark_reg_not_init(regs
+ regno
);
589 __mark_reg_known_zero(regs
+ regno
);
592 static bool reg_is_pkt_pointer(const struct bpf_reg_state
*reg
)
594 return type_is_pkt_pointer(reg
->type
);
597 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state
*reg
)
599 return reg_is_pkt_pointer(reg
) ||
600 reg
->type
== PTR_TO_PACKET_END
;
603 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
604 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state
*reg
,
605 enum bpf_reg_type which
)
607 /* The register can already have a range from prior markings.
608 * This is fine as long as it hasn't been advanced from its
611 return reg
->type
== which
&&
614 tnum_equals_const(reg
->var_off
, 0);
617 /* Attempts to improve min/max values based on var_off information */
618 static void __update_reg_bounds(struct bpf_reg_state
*reg
)
620 /* min signed is max(sign bit) | min(other bits) */
621 reg
->smin_value
= max_t(s64
, reg
->smin_value
,
622 reg
->var_off
.value
| (reg
->var_off
.mask
& S64_MIN
));
623 /* max signed is min(sign bit) | max(other bits) */
624 reg
->smax_value
= min_t(s64
, reg
->smax_value
,
625 reg
->var_off
.value
| (reg
->var_off
.mask
& S64_MAX
));
626 reg
->umin_value
= max(reg
->umin_value
, reg
->var_off
.value
);
627 reg
->umax_value
= min(reg
->umax_value
,
628 reg
->var_off
.value
| reg
->var_off
.mask
);
631 /* Uses signed min/max values to inform unsigned, and vice-versa */
632 static void __reg_deduce_bounds(struct bpf_reg_state
*reg
)
634 /* Learn sign from signed bounds.
635 * If we cannot cross the sign boundary, then signed and unsigned bounds
636 * are the same, so combine. This works even in the negative case, e.g.
637 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
639 if (reg
->smin_value
>= 0 || reg
->smax_value
< 0) {
640 reg
->smin_value
= reg
->umin_value
= max_t(u64
, reg
->smin_value
,
642 reg
->smax_value
= reg
->umax_value
= min_t(u64
, reg
->smax_value
,
646 /* Learn sign from unsigned bounds. Signed bounds cross the sign
647 * boundary, so we must be careful.
649 if ((s64
)reg
->umax_value
>= 0) {
650 /* Positive. We can't learn anything from the smin, but smax
651 * is positive, hence safe.
653 reg
->smin_value
= reg
->umin_value
;
654 reg
->smax_value
= reg
->umax_value
= min_t(u64
, reg
->smax_value
,
656 } else if ((s64
)reg
->umin_value
< 0) {
657 /* Negative. We can't learn anything from the smax, but smin
658 * is negative, hence safe.
660 reg
->smin_value
= reg
->umin_value
= max_t(u64
, reg
->smin_value
,
662 reg
->smax_value
= reg
->umax_value
;
666 /* Attempts to improve var_off based on unsigned min/max information */
667 static void __reg_bound_offset(struct bpf_reg_state
*reg
)
669 reg
->var_off
= tnum_intersect(reg
->var_off
,
670 tnum_range(reg
->umin_value
,
674 /* Reset the min/max bounds of a register */
675 static void __mark_reg_unbounded(struct bpf_reg_state
*reg
)
677 reg
->smin_value
= S64_MIN
;
678 reg
->smax_value
= S64_MAX
;
680 reg
->umax_value
= U64_MAX
;
683 /* Mark a register as having a completely unknown (scalar) value. */
684 static void __mark_reg_unknown(struct bpf_reg_state
*reg
)
686 reg
->type
= SCALAR_VALUE
;
689 reg
->var_off
= tnum_unknown
;
691 __mark_reg_unbounded(reg
);
694 static void mark_reg_unknown(struct bpf_verifier_env
*env
,
695 struct bpf_reg_state
*regs
, u32 regno
)
697 if (WARN_ON(regno
>= MAX_BPF_REG
)) {
698 verbose(env
, "mark_reg_unknown(regs, %u)\n", regno
);
699 /* Something bad happened, let's kill all regs except FP */
700 for (regno
= 0; regno
< BPF_REG_FP
; regno
++)
701 __mark_reg_not_init(regs
+ regno
);
704 __mark_reg_unknown(regs
+ regno
);
707 static void __mark_reg_not_init(struct bpf_reg_state
*reg
)
709 __mark_reg_unknown(reg
);
710 reg
->type
= NOT_INIT
;
713 static void mark_reg_not_init(struct bpf_verifier_env
*env
,
714 struct bpf_reg_state
*regs
, u32 regno
)
716 if (WARN_ON(regno
>= MAX_BPF_REG
)) {
717 verbose(env
, "mark_reg_not_init(regs, %u)\n", regno
);
718 /* Something bad happened, let's kill all regs except FP */
719 for (regno
= 0; regno
< BPF_REG_FP
; regno
++)
720 __mark_reg_not_init(regs
+ regno
);
723 __mark_reg_not_init(regs
+ regno
);
726 static void init_reg_state(struct bpf_verifier_env
*env
,
727 struct bpf_func_state
*state
)
729 struct bpf_reg_state
*regs
= state
->regs
;
732 for (i
= 0; i
< MAX_BPF_REG
; i
++) {
733 mark_reg_not_init(env
, regs
, i
);
734 regs
[i
].live
= REG_LIVE_NONE
;
738 regs
[BPF_REG_FP
].type
= PTR_TO_STACK
;
739 mark_reg_known_zero(env
, regs
, BPF_REG_FP
);
740 regs
[BPF_REG_FP
].frameno
= state
->frameno
;
742 /* 1st arg to a function */
743 regs
[BPF_REG_1
].type
= PTR_TO_CTX
;
744 mark_reg_known_zero(env
, regs
, BPF_REG_1
);
747 #define BPF_MAIN_FUNC (-1)
748 static void init_func_state(struct bpf_verifier_env
*env
,
749 struct bpf_func_state
*state
,
750 int callsite
, int frameno
, int subprogno
)
752 state
->callsite
= callsite
;
753 state
->frameno
= frameno
;
754 state
->subprogno
= subprogno
;
755 init_reg_state(env
, state
);
759 SRC_OP
, /* register is used as source operand */
760 DST_OP
, /* register is used as destination operand */
761 DST_OP_NO_MARK
/* same as above, check only, don't mark */
764 static int cmp_subprogs(const void *a
, const void *b
)
766 return ((struct bpf_subprog_info
*)a
)->start
-
767 ((struct bpf_subprog_info
*)b
)->start
;
770 static int find_subprog(struct bpf_verifier_env
*env
, int off
)
772 struct bpf_subprog_info
*p
;
774 p
= bsearch(&off
, env
->subprog_info
, env
->subprog_cnt
,
775 sizeof(env
->subprog_info
[0]), cmp_subprogs
);
778 return p
- env
->subprog_info
;
782 static int add_subprog(struct bpf_verifier_env
*env
, int off
)
784 int insn_cnt
= env
->prog
->len
;
787 if (off
>= insn_cnt
|| off
< 0) {
788 verbose(env
, "call to invalid destination\n");
791 ret
= find_subprog(env
, off
);
794 if (env
->subprog_cnt
>= BPF_MAX_SUBPROGS
) {
795 verbose(env
, "too many subprograms\n");
798 env
->subprog_info
[env
->subprog_cnt
++].start
= off
;
799 sort(env
->subprog_info
, env
->subprog_cnt
,
800 sizeof(env
->subprog_info
[0]), cmp_subprogs
, NULL
);
804 static int check_subprogs(struct bpf_verifier_env
*env
)
806 int i
, ret
, subprog_start
, subprog_end
, off
, cur_subprog
= 0;
807 struct bpf_subprog_info
*subprog
= env
->subprog_info
;
808 struct bpf_insn
*insn
= env
->prog
->insnsi
;
809 int insn_cnt
= env
->prog
->len
;
811 /* Add entry function. */
812 ret
= add_subprog(env
, 0);
816 /* determine subprog starts. The end is one before the next starts */
817 for (i
= 0; i
< insn_cnt
; i
++) {
818 if (insn
[i
].code
!= (BPF_JMP
| BPF_CALL
))
820 if (insn
[i
].src_reg
!= BPF_PSEUDO_CALL
)
822 if (!env
->allow_ptr_leaks
) {
823 verbose(env
, "function calls to other bpf functions are allowed for root only\n");
826 if (bpf_prog_is_dev_bound(env
->prog
->aux
)) {
827 verbose(env
, "function calls in offloaded programs are not supported yet\n");
830 ret
= add_subprog(env
, i
+ insn
[i
].imm
+ 1);
835 /* Add a fake 'exit' subprog which could simplify subprog iteration
836 * logic. 'subprog_cnt' should not be increased.
838 subprog
[env
->subprog_cnt
].start
= insn_cnt
;
840 if (env
->log
.level
> 1)
841 for (i
= 0; i
< env
->subprog_cnt
; i
++)
842 verbose(env
, "func#%d @%d\n", i
, subprog
[i
].start
);
844 /* now check that all jumps are within the same subprog */
845 subprog_start
= subprog
[cur_subprog
].start
;
846 subprog_end
= subprog
[cur_subprog
+ 1].start
;
847 for (i
= 0; i
< insn_cnt
; i
++) {
848 u8 code
= insn
[i
].code
;
850 if (BPF_CLASS(code
) != BPF_JMP
)
852 if (BPF_OP(code
) == BPF_EXIT
|| BPF_OP(code
) == BPF_CALL
)
854 off
= i
+ insn
[i
].off
+ 1;
855 if (off
< subprog_start
|| off
>= subprog_end
) {
856 verbose(env
, "jump out of range from insn %d to %d\n", i
, off
);
860 if (i
== subprog_end
- 1) {
861 /* to avoid fall-through from one subprog into another
862 * the last insn of the subprog should be either exit
863 * or unconditional jump back
865 if (code
!= (BPF_JMP
| BPF_EXIT
) &&
866 code
!= (BPF_JMP
| BPF_JA
)) {
867 verbose(env
, "last insn is not an exit or jmp\n");
870 subprog_start
= subprog_end
;
872 if (cur_subprog
< env
->subprog_cnt
)
873 subprog_end
= subprog
[cur_subprog
+ 1].start
;
880 struct bpf_verifier_state
*skip_callee(struct bpf_verifier_env
*env
,
881 const struct bpf_verifier_state
*state
,
882 struct bpf_verifier_state
*parent
,
885 struct bpf_verifier_state
*tmp
= NULL
;
887 /* 'parent' could be a state of caller and
888 * 'state' could be a state of callee. In such case
889 * parent->curframe < state->curframe
890 * and it's ok for r1 - r5 registers
892 * 'parent' could be a callee's state after it bpf_exit-ed.
893 * In such case parent->curframe > state->curframe
894 * and it's ok for r0 only
896 if (parent
->curframe
== state
->curframe
||
897 (parent
->curframe
< state
->curframe
&&
898 regno
>= BPF_REG_1
&& regno
<= BPF_REG_5
) ||
899 (parent
->curframe
> state
->curframe
&&
903 if (parent
->curframe
> state
->curframe
&&
904 regno
>= BPF_REG_6
) {
905 /* for callee saved regs we have to skip the whole chain
906 * of states that belong to callee and mark as LIVE_READ
907 * the registers before the call
910 while (tmp
&& tmp
->curframe
!= state
->curframe
) {
921 verbose(env
, "verifier bug regno %d tmp %p\n", regno
, tmp
);
922 verbose(env
, "regno %d parent frame %d current frame %d\n",
923 regno
, parent
->curframe
, state
->curframe
);
927 static int mark_reg_read(struct bpf_verifier_env
*env
,
928 const struct bpf_verifier_state
*state
,
929 struct bpf_verifier_state
*parent
,
932 bool writes
= parent
== state
->parent
; /* Observe write marks */
934 if (regno
== BPF_REG_FP
)
935 /* We don't need to worry about FP liveness because it's read-only */
939 /* if read wasn't screened by an earlier write ... */
940 if (writes
&& state
->frame
[state
->curframe
]->regs
[regno
].live
& REG_LIVE_WRITTEN
)
942 parent
= skip_callee(env
, state
, parent
, regno
);
945 /* ... then we depend on parent's value */
946 parent
->frame
[parent
->curframe
]->regs
[regno
].live
|= REG_LIVE_READ
;
948 parent
= state
->parent
;
954 static int check_reg_arg(struct bpf_verifier_env
*env
, u32 regno
,
957 struct bpf_verifier_state
*vstate
= env
->cur_state
;
958 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
959 struct bpf_reg_state
*regs
= state
->regs
;
961 if (regno
>= MAX_BPF_REG
) {
962 verbose(env
, "R%d is invalid\n", regno
);
967 /* check whether register used as source operand can be read */
968 if (regs
[regno
].type
== NOT_INIT
) {
969 verbose(env
, "R%d !read_ok\n", regno
);
972 return mark_reg_read(env
, vstate
, vstate
->parent
, regno
);
974 /* check whether register used as dest operand can be written to */
975 if (regno
== BPF_REG_FP
) {
976 verbose(env
, "frame pointer is read only\n");
979 regs
[regno
].live
|= REG_LIVE_WRITTEN
;
981 mark_reg_unknown(env
, regs
, regno
);
986 static bool is_spillable_regtype(enum bpf_reg_type type
)
989 case PTR_TO_MAP_VALUE
:
990 case PTR_TO_MAP_VALUE_OR_NULL
:
994 case PTR_TO_PACKET_META
:
995 case PTR_TO_PACKET_END
:
996 case CONST_PTR_TO_MAP
:
1003 /* Does this register contain a constant zero? */
1004 static bool register_is_null(struct bpf_reg_state
*reg
)
1006 return reg
->type
== SCALAR_VALUE
&& tnum_equals_const(reg
->var_off
, 0);
1009 /* check_stack_read/write functions track spill/fill of registers,
1010 * stack boundary and alignment are checked in check_mem_access()
1012 static int check_stack_write(struct bpf_verifier_env
*env
,
1013 struct bpf_func_state
*state
, /* func where register points to */
1014 int off
, int size
, int value_regno
, int insn_idx
)
1016 struct bpf_func_state
*cur
; /* state of the current function */
1017 int i
, slot
= -off
- 1, spi
= slot
/ BPF_REG_SIZE
, err
;
1018 enum bpf_reg_type type
;
1020 err
= realloc_func_state(state
, round_up(slot
+ 1, BPF_REG_SIZE
),
1024 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
1025 * so it's aligned access and [off, off + size) are within stack limits
1027 if (!env
->allow_ptr_leaks
&&
1028 state
->stack
[spi
].slot_type
[0] == STACK_SPILL
&&
1029 size
!= BPF_REG_SIZE
) {
1030 verbose(env
, "attempt to corrupt spilled pointer on stack\n");
1034 cur
= env
->cur_state
->frame
[env
->cur_state
->curframe
];
1035 if (value_regno
>= 0 &&
1036 is_spillable_regtype((type
= cur
->regs
[value_regno
].type
))) {
1038 /* register containing pointer is being spilled into stack */
1039 if (size
!= BPF_REG_SIZE
) {
1040 verbose(env
, "invalid size of register spill\n");
1044 if (state
!= cur
&& type
== PTR_TO_STACK
) {
1045 verbose(env
, "cannot spill pointers to stack into stack frame of the caller\n");
1049 /* save register state */
1050 state
->stack
[spi
].spilled_ptr
= cur
->regs
[value_regno
];
1051 state
->stack
[spi
].spilled_ptr
.live
|= REG_LIVE_WRITTEN
;
1053 for (i
= 0; i
< BPF_REG_SIZE
; i
++) {
1054 if (state
->stack
[spi
].slot_type
[i
] == STACK_MISC
&&
1055 !env
->allow_ptr_leaks
) {
1056 int *poff
= &env
->insn_aux_data
[insn_idx
].sanitize_stack_off
;
1057 int soff
= (-spi
- 1) * BPF_REG_SIZE
;
1059 /* detected reuse of integer stack slot with a pointer
1060 * which means either llvm is reusing stack slot or
1061 * an attacker is trying to exploit CVE-2018-3639
1062 * (speculative store bypass)
1063 * Have to sanitize that slot with preemptive
1066 if (*poff
&& *poff
!= soff
) {
1067 /* disallow programs where single insn stores
1068 * into two different stack slots, since verifier
1069 * cannot sanitize them
1072 "insn %d cannot access two stack slots fp%d and fp%d",
1073 insn_idx
, *poff
, soff
);
1078 state
->stack
[spi
].slot_type
[i
] = STACK_SPILL
;
1081 u8 type
= STACK_MISC
;
1083 /* regular write of data into stack */
1084 state
->stack
[spi
].spilled_ptr
= (struct bpf_reg_state
) {};
1086 /* only mark the slot as written if all 8 bytes were written
1087 * otherwise read propagation may incorrectly stop too soon
1088 * when stack slots are partially written.
1089 * This heuristic means that read propagation will be
1090 * conservative, since it will add reg_live_read marks
1091 * to stack slots all the way to first state when programs
1092 * writes+reads less than 8 bytes
1094 if (size
== BPF_REG_SIZE
)
1095 state
->stack
[spi
].spilled_ptr
.live
|= REG_LIVE_WRITTEN
;
1097 /* when we zero initialize stack slots mark them as such */
1098 if (value_regno
>= 0 &&
1099 register_is_null(&cur
->regs
[value_regno
]))
1102 for (i
= 0; i
< size
; i
++)
1103 state
->stack
[spi
].slot_type
[(slot
- i
) % BPF_REG_SIZE
] =
1109 /* registers of every function are unique and mark_reg_read() propagates
1110 * the liveness in the following cases:
1111 * - from callee into caller for R1 - R5 that were used as arguments
1112 * - from caller into callee for R0 that used as result of the call
1113 * - from caller to the same caller skipping states of the callee for R6 - R9,
1114 * since R6 - R9 are callee saved by implicit function prologue and
1115 * caller's R6 != callee's R6, so when we propagate liveness up to
1116 * parent states we need to skip callee states for R6 - R9.
1118 * stack slot marking is different, since stacks of caller and callee are
1119 * accessible in both (since caller can pass a pointer to caller's stack to
1120 * callee which can pass it to another function), hence mark_stack_slot_read()
1121 * has to propagate the stack liveness to all parent states at given frame number.
1131 * First *ptr is reading from f1's stack and mark_stack_slot_read() has
1132 * to mark liveness at the f1's frame and not f2's frame.
1133 * Second *ptr is also reading from f1's stack and mark_stack_slot_read() has
1134 * to propagate liveness to f2 states at f1's frame level and further into
1135 * f1 states at f1's frame level until write into that stack slot
1137 static void mark_stack_slot_read(struct bpf_verifier_env
*env
,
1138 const struct bpf_verifier_state
*state
,
1139 struct bpf_verifier_state
*parent
,
1140 int slot
, int frameno
)
1142 bool writes
= parent
== state
->parent
; /* Observe write marks */
1145 if (parent
->frame
[frameno
]->allocated_stack
<= slot
* BPF_REG_SIZE
)
1146 /* since LIVE_WRITTEN mark is only done for full 8-byte
1147 * write the read marks are conservative and parent
1148 * state may not even have the stack allocated. In such case
1149 * end the propagation, since the loop reached beginning
1153 /* if read wasn't screened by an earlier write ... */
1154 if (writes
&& state
->frame
[frameno
]->stack
[slot
].spilled_ptr
.live
& REG_LIVE_WRITTEN
)
1156 /* ... then we depend on parent's value */
1157 parent
->frame
[frameno
]->stack
[slot
].spilled_ptr
.live
|= REG_LIVE_READ
;
1159 parent
= state
->parent
;
1164 static int check_stack_read(struct bpf_verifier_env
*env
,
1165 struct bpf_func_state
*reg_state
/* func where register points to */,
1166 int off
, int size
, int value_regno
)
1168 struct bpf_verifier_state
*vstate
= env
->cur_state
;
1169 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
1170 int i
, slot
= -off
- 1, spi
= slot
/ BPF_REG_SIZE
;
1173 if (reg_state
->allocated_stack
<= slot
) {
1174 verbose(env
, "invalid read from stack off %d+0 size %d\n",
1178 stype
= reg_state
->stack
[spi
].slot_type
;
1180 if (stype
[0] == STACK_SPILL
) {
1181 if (size
!= BPF_REG_SIZE
) {
1182 verbose(env
, "invalid size of register spill\n");
1185 for (i
= 1; i
< BPF_REG_SIZE
; i
++) {
1186 if (stype
[(slot
- i
) % BPF_REG_SIZE
] != STACK_SPILL
) {
1187 verbose(env
, "corrupted spill memory\n");
1192 if (value_regno
>= 0) {
1193 /* restore register state from stack */
1194 state
->regs
[value_regno
] = reg_state
->stack
[spi
].spilled_ptr
;
1195 /* mark reg as written since spilled pointer state likely
1196 * has its liveness marks cleared by is_state_visited()
1197 * which resets stack/reg liveness for state transitions
1199 state
->regs
[value_regno
].live
|= REG_LIVE_WRITTEN
;
1201 mark_stack_slot_read(env
, vstate
, vstate
->parent
, spi
,
1202 reg_state
->frameno
);
1207 for (i
= 0; i
< size
; i
++) {
1208 if (stype
[(slot
- i
) % BPF_REG_SIZE
] == STACK_MISC
)
1210 if (stype
[(slot
- i
) % BPF_REG_SIZE
] == STACK_ZERO
) {
1214 verbose(env
, "invalid read from stack off %d+%d size %d\n",
1218 mark_stack_slot_read(env
, vstate
, vstate
->parent
, spi
,
1219 reg_state
->frameno
);
1220 if (value_regno
>= 0) {
1221 if (zeros
== size
) {
1222 /* any size read into register is zero extended,
1223 * so the whole register == const_zero
1225 __mark_reg_const_zero(&state
->regs
[value_regno
]);
1227 /* have read misc data from the stack */
1228 mark_reg_unknown(env
, state
->regs
, value_regno
);
1230 state
->regs
[value_regno
].live
|= REG_LIVE_WRITTEN
;
1236 /* check read/write into map element returned by bpf_map_lookup_elem() */
1237 static int __check_map_access(struct bpf_verifier_env
*env
, u32 regno
, int off
,
1238 int size
, bool zero_size_allowed
)
1240 struct bpf_reg_state
*regs
= cur_regs(env
);
1241 struct bpf_map
*map
= regs
[regno
].map_ptr
;
1243 if (off
< 0 || size
< 0 || (size
== 0 && !zero_size_allowed
) ||
1244 off
+ size
> map
->value_size
) {
1245 verbose(env
, "invalid access to map value, value_size=%d off=%d size=%d\n",
1246 map
->value_size
, off
, size
);
1252 /* check read/write into a map element with possible variable offset */
1253 static int check_map_access(struct bpf_verifier_env
*env
, u32 regno
,
1254 int off
, int size
, bool zero_size_allowed
)
1256 struct bpf_verifier_state
*vstate
= env
->cur_state
;
1257 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
1258 struct bpf_reg_state
*reg
= &state
->regs
[regno
];
1261 /* We may have adjusted the register to this map value, so we
1262 * need to try adding each of min_value and max_value to off
1263 * to make sure our theoretical access will be safe.
1266 print_verifier_state(env
, state
);
1267 /* The minimum value is only important with signed
1268 * comparisons where we can't assume the floor of a
1269 * value is 0. If we are using signed variables for our
1270 * index'es we need to make sure that whatever we use
1271 * will have a set floor within our range.
1273 if (reg
->smin_value
< 0) {
1274 verbose(env
, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1278 err
= __check_map_access(env
, regno
, reg
->smin_value
+ off
, size
,
1281 verbose(env
, "R%d min value is outside of the array range\n",
1286 /* If we haven't set a max value then we need to bail since we can't be
1287 * sure we won't do bad things.
1288 * If reg->umax_value + off could overflow, treat that as unbounded too.
1290 if (reg
->umax_value
>= BPF_MAX_VAR_OFF
) {
1291 verbose(env
, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
1295 err
= __check_map_access(env
, regno
, reg
->umax_value
+ off
, size
,
1298 verbose(env
, "R%d max value is outside of the array range\n",
1303 #define MAX_PACKET_OFF 0xffff
1305 static bool may_access_direct_pkt_data(struct bpf_verifier_env
*env
,
1306 const struct bpf_call_arg_meta
*meta
,
1307 enum bpf_access_type t
)
1309 switch (env
->prog
->type
) {
1310 case BPF_PROG_TYPE_LWT_IN
:
1311 case BPF_PROG_TYPE_LWT_OUT
:
1312 case BPF_PROG_TYPE_LWT_SEG6LOCAL
:
1313 /* dst_input() and dst_output() can't write for now */
1317 case BPF_PROG_TYPE_SCHED_CLS
:
1318 case BPF_PROG_TYPE_SCHED_ACT
:
1319 case BPF_PROG_TYPE_XDP
:
1320 case BPF_PROG_TYPE_LWT_XMIT
:
1321 case BPF_PROG_TYPE_SK_SKB
:
1322 case BPF_PROG_TYPE_SK_MSG
:
1324 return meta
->pkt_access
;
1326 env
->seen_direct_write
= true;
1333 static int __check_packet_access(struct bpf_verifier_env
*env
, u32 regno
,
1334 int off
, int size
, bool zero_size_allowed
)
1336 struct bpf_reg_state
*regs
= cur_regs(env
);
1337 struct bpf_reg_state
*reg
= ®s
[regno
];
1339 if (off
< 0 || size
< 0 || (size
== 0 && !zero_size_allowed
) ||
1340 (u64
)off
+ size
> reg
->range
) {
1341 verbose(env
, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
1342 off
, size
, regno
, reg
->id
, reg
->off
, reg
->range
);
1348 static int check_packet_access(struct bpf_verifier_env
*env
, u32 regno
, int off
,
1349 int size
, bool zero_size_allowed
)
1351 struct bpf_reg_state
*regs
= cur_regs(env
);
1352 struct bpf_reg_state
*reg
= ®s
[regno
];
1355 /* We may have added a variable offset to the packet pointer; but any
1356 * reg->range we have comes after that. We are only checking the fixed
1360 /* We don't allow negative numbers, because we aren't tracking enough
1361 * detail to prove they're safe.
1363 if (reg
->smin_value
< 0) {
1364 verbose(env
, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1368 err
= __check_packet_access(env
, regno
, off
, size
, zero_size_allowed
);
1370 verbose(env
, "R%d offset is outside of the packet\n", regno
);
1376 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
1377 static int check_ctx_access(struct bpf_verifier_env
*env
, int insn_idx
, int off
, int size
,
1378 enum bpf_access_type t
, enum bpf_reg_type
*reg_type
)
1380 struct bpf_insn_access_aux info
= {
1381 .reg_type
= *reg_type
,
1384 if (env
->ops
->is_valid_access
&&
1385 env
->ops
->is_valid_access(off
, size
, t
, env
->prog
, &info
)) {
1386 /* A non zero info.ctx_field_size indicates that this field is a
1387 * candidate for later verifier transformation to load the whole
1388 * field and then apply a mask when accessed with a narrower
1389 * access than actual ctx access size. A zero info.ctx_field_size
1390 * will only allow for whole field access and rejects any other
1391 * type of narrower access.
1393 *reg_type
= info
.reg_type
;
1395 env
->insn_aux_data
[insn_idx
].ctx_field_size
= info
.ctx_field_size
;
1396 /* remember the offset of last byte accessed in ctx */
1397 if (env
->prog
->aux
->max_ctx_offset
< off
+ size
)
1398 env
->prog
->aux
->max_ctx_offset
= off
+ size
;
1402 verbose(env
, "invalid bpf_context access off=%d size=%d\n", off
, size
);
1406 static bool __is_pointer_value(bool allow_ptr_leaks
,
1407 const struct bpf_reg_state
*reg
)
1409 if (allow_ptr_leaks
)
1412 return reg
->type
!= SCALAR_VALUE
;
1415 static bool is_pointer_value(struct bpf_verifier_env
*env
, int regno
)
1417 return __is_pointer_value(env
->allow_ptr_leaks
, cur_regs(env
) + regno
);
1420 static bool is_ctx_reg(struct bpf_verifier_env
*env
, int regno
)
1422 const struct bpf_reg_state
*reg
= cur_regs(env
) + regno
;
1424 return reg
->type
== PTR_TO_CTX
;
1427 static bool is_pkt_reg(struct bpf_verifier_env
*env
, int regno
)
1429 const struct bpf_reg_state
*reg
= cur_regs(env
) + regno
;
1431 return type_is_pkt_pointer(reg
->type
);
1434 static int check_pkt_ptr_alignment(struct bpf_verifier_env
*env
,
1435 const struct bpf_reg_state
*reg
,
1436 int off
, int size
, bool strict
)
1438 struct tnum reg_off
;
1441 /* Byte size accesses are always allowed. */
1442 if (!strict
|| size
== 1)
1445 /* For platforms that do not have a Kconfig enabling
1446 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
1447 * NET_IP_ALIGN is universally set to '2'. And on platforms
1448 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
1449 * to this code only in strict mode where we want to emulate
1450 * the NET_IP_ALIGN==2 checking. Therefore use an
1451 * unconditional IP align value of '2'.
1455 reg_off
= tnum_add(reg
->var_off
, tnum_const(ip_align
+ reg
->off
+ off
));
1456 if (!tnum_is_aligned(reg_off
, size
)) {
1459 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1461 "misaligned packet access off %d+%s+%d+%d size %d\n",
1462 ip_align
, tn_buf
, reg
->off
, off
, size
);
1469 static int check_generic_ptr_alignment(struct bpf_verifier_env
*env
,
1470 const struct bpf_reg_state
*reg
,
1471 const char *pointer_desc
,
1472 int off
, int size
, bool strict
)
1474 struct tnum reg_off
;
1476 /* Byte size accesses are always allowed. */
1477 if (!strict
|| size
== 1)
1480 reg_off
= tnum_add(reg
->var_off
, tnum_const(reg
->off
+ off
));
1481 if (!tnum_is_aligned(reg_off
, size
)) {
1484 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1485 verbose(env
, "misaligned %saccess off %s+%d+%d size %d\n",
1486 pointer_desc
, tn_buf
, reg
->off
, off
, size
);
1493 static int check_ptr_alignment(struct bpf_verifier_env
*env
,
1494 const struct bpf_reg_state
*reg
, int off
,
1495 int size
, bool strict_alignment_once
)
1497 bool strict
= env
->strict_alignment
|| strict_alignment_once
;
1498 const char *pointer_desc
= "";
1500 switch (reg
->type
) {
1502 case PTR_TO_PACKET_META
:
1503 /* Special case, because of NET_IP_ALIGN. Given metadata sits
1504 * right in front, treat it the very same way.
1506 return check_pkt_ptr_alignment(env
, reg
, off
, size
, strict
);
1507 case PTR_TO_MAP_VALUE
:
1508 pointer_desc
= "value ";
1511 pointer_desc
= "context ";
1514 pointer_desc
= "stack ";
1515 /* The stack spill tracking logic in check_stack_write()
1516 * and check_stack_read() relies on stack accesses being
1524 return check_generic_ptr_alignment(env
, reg
, pointer_desc
, off
, size
,
1528 static int update_stack_depth(struct bpf_verifier_env
*env
,
1529 const struct bpf_func_state
*func
,
1532 u16 stack
= env
->subprog_info
[func
->subprogno
].stack_depth
;
1537 /* update known max for given subprogram */
1538 env
->subprog_info
[func
->subprogno
].stack_depth
= -off
;
1542 /* starting from main bpf function walk all instructions of the function
1543 * and recursively walk all callees that given function can call.
1544 * Ignore jump and exit insns.
1545 * Since recursion is prevented by check_cfg() this algorithm
1546 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
1548 static int check_max_stack_depth(struct bpf_verifier_env
*env
)
1550 int depth
= 0, frame
= 0, idx
= 0, i
= 0, subprog_end
;
1551 struct bpf_subprog_info
*subprog
= env
->subprog_info
;
1552 struct bpf_insn
*insn
= env
->prog
->insnsi
;
1553 int ret_insn
[MAX_CALL_FRAMES
];
1554 int ret_prog
[MAX_CALL_FRAMES
];
1557 /* round up to 32-bytes, since this is granularity
1558 * of interpreter stack size
1560 depth
+= round_up(max_t(u32
, subprog
[idx
].stack_depth
, 1), 32);
1561 if (depth
> MAX_BPF_STACK
) {
1562 verbose(env
, "combined stack size of %d calls is %d. Too large\n",
1567 subprog_end
= subprog
[idx
+ 1].start
;
1568 for (; i
< subprog_end
; i
++) {
1569 if (insn
[i
].code
!= (BPF_JMP
| BPF_CALL
))
1571 if (insn
[i
].src_reg
!= BPF_PSEUDO_CALL
)
1573 /* remember insn and function to return to */
1574 ret_insn
[frame
] = i
+ 1;
1575 ret_prog
[frame
] = idx
;
1577 /* find the callee */
1578 i
= i
+ insn
[i
].imm
+ 1;
1579 idx
= find_subprog(env
, i
);
1581 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1586 if (frame
>= MAX_CALL_FRAMES
) {
1587 WARN_ONCE(1, "verifier bug. Call stack is too deep\n");
1592 /* end of for() loop means the last insn of the 'subprog'
1593 * was reached. Doesn't matter whether it was JA or EXIT
1597 depth
-= round_up(max_t(u32
, subprog
[idx
].stack_depth
, 1), 32);
1599 i
= ret_insn
[frame
];
1600 idx
= ret_prog
[frame
];
1604 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1605 static int get_callee_stack_depth(struct bpf_verifier_env
*env
,
1606 const struct bpf_insn
*insn
, int idx
)
1608 int start
= idx
+ insn
->imm
+ 1, subprog
;
1610 subprog
= find_subprog(env
, start
);
1612 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1616 return env
->subprog_info
[subprog
].stack_depth
;
1620 static int check_ctx_reg(struct bpf_verifier_env
*env
,
1621 const struct bpf_reg_state
*reg
, int regno
)
1623 /* Access to ctx or passing it to a helper is only allowed in
1624 * its original, unmodified form.
1628 verbose(env
, "dereference of modified ctx ptr R%d off=%d disallowed\n",
1633 if (!tnum_is_const(reg
->var_off
) || reg
->var_off
.value
) {
1636 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1637 verbose(env
, "variable ctx access var_off=%s disallowed\n", tn_buf
);
1644 /* truncate register to smaller size (in bytes)
1645 * must be called with size < BPF_REG_SIZE
1647 static void coerce_reg_to_size(struct bpf_reg_state
*reg
, int size
)
1651 /* clear high bits in bit representation */
1652 reg
->var_off
= tnum_cast(reg
->var_off
, size
);
1654 /* fix arithmetic bounds */
1655 mask
= ((u64
)1 << (size
* 8)) - 1;
1656 if ((reg
->umin_value
& ~mask
) == (reg
->umax_value
& ~mask
)) {
1657 reg
->umin_value
&= mask
;
1658 reg
->umax_value
&= mask
;
1660 reg
->umin_value
= 0;
1661 reg
->umax_value
= mask
;
1663 reg
->smin_value
= reg
->umin_value
;
1664 reg
->smax_value
= reg
->umax_value
;
1667 /* check whether memory at (regno + off) is accessible for t = (read | write)
1668 * if t==write, value_regno is a register which value is stored into memory
1669 * if t==read, value_regno is a register which will receive the value from memory
1670 * if t==write && value_regno==-1, some unknown value is stored into memory
1671 * if t==read && value_regno==-1, don't care what we read from memory
1673 static int check_mem_access(struct bpf_verifier_env
*env
, int insn_idx
, u32 regno
,
1674 int off
, int bpf_size
, enum bpf_access_type t
,
1675 int value_regno
, bool strict_alignment_once
)
1677 struct bpf_reg_state
*regs
= cur_regs(env
);
1678 struct bpf_reg_state
*reg
= regs
+ regno
;
1679 struct bpf_func_state
*state
;
1682 size
= bpf_size_to_bytes(bpf_size
);
1686 /* alignment checks will add in reg->off themselves */
1687 err
= check_ptr_alignment(env
, reg
, off
, size
, strict_alignment_once
);
1691 /* for access checks, reg->off is just part of off */
1694 if (reg
->type
== PTR_TO_MAP_VALUE
) {
1695 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
1696 is_pointer_value(env
, value_regno
)) {
1697 verbose(env
, "R%d leaks addr into map\n", value_regno
);
1701 err
= check_map_access(env
, regno
, off
, size
, false);
1702 if (!err
&& t
== BPF_READ
&& value_regno
>= 0)
1703 mark_reg_unknown(env
, regs
, value_regno
);
1705 } else if (reg
->type
== PTR_TO_CTX
) {
1706 enum bpf_reg_type reg_type
= SCALAR_VALUE
;
1708 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
1709 is_pointer_value(env
, value_regno
)) {
1710 verbose(env
, "R%d leaks addr into ctx\n", value_regno
);
1714 err
= check_ctx_reg(env
, reg
, regno
);
1718 err
= check_ctx_access(env
, insn_idx
, off
, size
, t
, ®_type
);
1719 if (!err
&& t
== BPF_READ
&& value_regno
>= 0) {
1720 /* ctx access returns either a scalar, or a
1721 * PTR_TO_PACKET[_META,_END]. In the latter
1722 * case, we know the offset is zero.
1724 if (reg_type
== SCALAR_VALUE
)
1725 mark_reg_unknown(env
, regs
, value_regno
);
1727 mark_reg_known_zero(env
, regs
,
1729 regs
[value_regno
].id
= 0;
1730 regs
[value_regno
].off
= 0;
1731 regs
[value_regno
].range
= 0;
1732 regs
[value_regno
].type
= reg_type
;
1735 } else if (reg
->type
== PTR_TO_STACK
) {
1736 /* stack accesses must be at a fixed offset, so that we can
1737 * determine what type of data were returned.
1738 * See check_stack_read().
1740 if (!tnum_is_const(reg
->var_off
)) {
1743 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1744 verbose(env
, "variable stack access var_off=%s off=%d size=%d",
1748 off
+= reg
->var_off
.value
;
1749 if (off
>= 0 || off
< -MAX_BPF_STACK
) {
1750 verbose(env
, "invalid stack off=%d size=%d\n", off
,
1755 state
= func(env
, reg
);
1756 err
= update_stack_depth(env
, state
, off
);
1761 err
= check_stack_write(env
, state
, off
, size
,
1762 value_regno
, insn_idx
);
1764 err
= check_stack_read(env
, state
, off
, size
,
1766 } else if (reg_is_pkt_pointer(reg
)) {
1767 if (t
== BPF_WRITE
&& !may_access_direct_pkt_data(env
, NULL
, t
)) {
1768 verbose(env
, "cannot write into packet\n");
1771 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
1772 is_pointer_value(env
, value_regno
)) {
1773 verbose(env
, "R%d leaks addr into packet\n",
1777 err
= check_packet_access(env
, regno
, off
, size
, false);
1778 if (!err
&& t
== BPF_READ
&& value_regno
>= 0)
1779 mark_reg_unknown(env
, regs
, value_regno
);
1781 verbose(env
, "R%d invalid mem access '%s'\n", regno
,
1782 reg_type_str
[reg
->type
]);
1786 if (!err
&& size
< BPF_REG_SIZE
&& value_regno
>= 0 && t
== BPF_READ
&&
1787 regs
[value_regno
].type
== SCALAR_VALUE
) {
1788 /* b/h/w load zero-extends, mark upper bits as known 0 */
1789 coerce_reg_to_size(®s
[value_regno
], size
);
1794 static int check_xadd(struct bpf_verifier_env
*env
, int insn_idx
, struct bpf_insn
*insn
)
1798 if ((BPF_SIZE(insn
->code
) != BPF_W
&& BPF_SIZE(insn
->code
) != BPF_DW
) ||
1800 verbose(env
, "BPF_XADD uses reserved fields\n");
1804 /* check src1 operand */
1805 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
1809 /* check src2 operand */
1810 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
1814 if (is_pointer_value(env
, insn
->src_reg
)) {
1815 verbose(env
, "R%d leaks addr into mem\n", insn
->src_reg
);
1819 if (is_ctx_reg(env
, insn
->dst_reg
) ||
1820 is_pkt_reg(env
, insn
->dst_reg
)) {
1821 verbose(env
, "BPF_XADD stores into R%d %s is not allowed\n",
1822 insn
->dst_reg
, is_ctx_reg(env
, insn
->dst_reg
) ?
1823 "context" : "packet");
1827 /* check whether atomic_add can read the memory */
1828 err
= check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
1829 BPF_SIZE(insn
->code
), BPF_READ
, -1, true);
1833 /* check whether atomic_add can write into the same memory */
1834 return check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
1835 BPF_SIZE(insn
->code
), BPF_WRITE
, -1, true);
1838 /* when register 'regno' is passed into function that will read 'access_size'
1839 * bytes from that pointer, make sure that it's within stack boundary
1840 * and all elements of stack are initialized.
1841 * Unlike most pointer bounds-checking functions, this one doesn't take an
1842 * 'off' argument, so it has to add in reg->off itself.
1844 static int check_stack_boundary(struct bpf_verifier_env
*env
, int regno
,
1845 int access_size
, bool zero_size_allowed
,
1846 struct bpf_call_arg_meta
*meta
)
1848 struct bpf_reg_state
*reg
= cur_regs(env
) + regno
;
1849 struct bpf_func_state
*state
= func(env
, reg
);
1850 int off
, i
, slot
, spi
;
1852 if (reg
->type
!= PTR_TO_STACK
) {
1853 /* Allow zero-byte read from NULL, regardless of pointer type */
1854 if (zero_size_allowed
&& access_size
== 0 &&
1855 register_is_null(reg
))
1858 verbose(env
, "R%d type=%s expected=%s\n", regno
,
1859 reg_type_str
[reg
->type
],
1860 reg_type_str
[PTR_TO_STACK
]);
1864 /* Only allow fixed-offset stack reads */
1865 if (!tnum_is_const(reg
->var_off
)) {
1868 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1869 verbose(env
, "invalid variable stack read R%d var_off=%s\n",
1873 off
= reg
->off
+ reg
->var_off
.value
;
1874 if (off
>= 0 || off
< -MAX_BPF_STACK
|| off
+ access_size
> 0 ||
1875 access_size
< 0 || (access_size
== 0 && !zero_size_allowed
)) {
1876 verbose(env
, "invalid stack type R%d off=%d access_size=%d\n",
1877 regno
, off
, access_size
);
1881 if (meta
&& meta
->raw_mode
) {
1882 meta
->access_size
= access_size
;
1883 meta
->regno
= regno
;
1887 for (i
= 0; i
< access_size
; i
++) {
1890 slot
= -(off
+ i
) - 1;
1891 spi
= slot
/ BPF_REG_SIZE
;
1892 if (state
->allocated_stack
<= slot
)
1894 stype
= &state
->stack
[spi
].slot_type
[slot
% BPF_REG_SIZE
];
1895 if (*stype
== STACK_MISC
)
1897 if (*stype
== STACK_ZERO
) {
1898 /* helper can write anything into the stack */
1899 *stype
= STACK_MISC
;
1903 verbose(env
, "invalid indirect read from stack off %d+%d size %d\n",
1904 off
, i
, access_size
);
1907 /* reading any byte out of 8-byte 'spill_slot' will cause
1908 * the whole slot to be marked as 'read'
1910 mark_stack_slot_read(env
, env
->cur_state
, env
->cur_state
->parent
,
1911 spi
, state
->frameno
);
1913 return update_stack_depth(env
, state
, off
);
1916 static int check_helper_mem_access(struct bpf_verifier_env
*env
, int regno
,
1917 int access_size
, bool zero_size_allowed
,
1918 struct bpf_call_arg_meta
*meta
)
1920 struct bpf_reg_state
*regs
= cur_regs(env
), *reg
= ®s
[regno
];
1922 switch (reg
->type
) {
1924 case PTR_TO_PACKET_META
:
1925 return check_packet_access(env
, regno
, reg
->off
, access_size
,
1927 case PTR_TO_MAP_VALUE
:
1928 return check_map_access(env
, regno
, reg
->off
, access_size
,
1930 default: /* scalar_value|ptr_to_stack or invalid ptr */
1931 return check_stack_boundary(env
, regno
, access_size
,
1932 zero_size_allowed
, meta
);
1936 static bool arg_type_is_mem_ptr(enum bpf_arg_type type
)
1938 return type
== ARG_PTR_TO_MEM
||
1939 type
== ARG_PTR_TO_MEM_OR_NULL
||
1940 type
== ARG_PTR_TO_UNINIT_MEM
;
1943 static bool arg_type_is_mem_size(enum bpf_arg_type type
)
1945 return type
== ARG_CONST_SIZE
||
1946 type
== ARG_CONST_SIZE_OR_ZERO
;
1949 static int check_func_arg(struct bpf_verifier_env
*env
, u32 regno
,
1950 enum bpf_arg_type arg_type
,
1951 struct bpf_call_arg_meta
*meta
)
1953 struct bpf_reg_state
*regs
= cur_regs(env
), *reg
= ®s
[regno
];
1954 enum bpf_reg_type expected_type
, type
= reg
->type
;
1957 if (arg_type
== ARG_DONTCARE
)
1960 err
= check_reg_arg(env
, regno
, SRC_OP
);
1964 if (arg_type
== ARG_ANYTHING
) {
1965 if (is_pointer_value(env
, regno
)) {
1966 verbose(env
, "R%d leaks addr into helper function\n",
1973 if (type_is_pkt_pointer(type
) &&
1974 !may_access_direct_pkt_data(env
, meta
, BPF_READ
)) {
1975 verbose(env
, "helper access to the packet is not allowed\n");
1979 if (arg_type
== ARG_PTR_TO_MAP_KEY
||
1980 arg_type
== ARG_PTR_TO_MAP_VALUE
) {
1981 expected_type
= PTR_TO_STACK
;
1982 if (!type_is_pkt_pointer(type
) && type
!= PTR_TO_MAP_VALUE
&&
1983 type
!= expected_type
)
1985 } else if (arg_type
== ARG_CONST_SIZE
||
1986 arg_type
== ARG_CONST_SIZE_OR_ZERO
) {
1987 expected_type
= SCALAR_VALUE
;
1988 if (type
!= expected_type
)
1990 } else if (arg_type
== ARG_CONST_MAP_PTR
) {
1991 expected_type
= CONST_PTR_TO_MAP
;
1992 if (type
!= expected_type
)
1994 } else if (arg_type
== ARG_PTR_TO_CTX
) {
1995 expected_type
= PTR_TO_CTX
;
1996 if (type
!= expected_type
)
1998 err
= check_ctx_reg(env
, reg
, regno
);
2001 } else if (arg_type_is_mem_ptr(arg_type
)) {
2002 expected_type
= PTR_TO_STACK
;
2003 /* One exception here. In case function allows for NULL to be
2004 * passed in as argument, it's a SCALAR_VALUE type. Final test
2005 * happens during stack boundary checking.
2007 if (register_is_null(reg
) &&
2008 arg_type
== ARG_PTR_TO_MEM_OR_NULL
)
2009 /* final test in check_stack_boundary() */;
2010 else if (!type_is_pkt_pointer(type
) &&
2011 type
!= PTR_TO_MAP_VALUE
&&
2012 type
!= expected_type
)
2014 meta
->raw_mode
= arg_type
== ARG_PTR_TO_UNINIT_MEM
;
2016 verbose(env
, "unsupported arg_type %d\n", arg_type
);
2020 if (arg_type
== ARG_CONST_MAP_PTR
) {
2021 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
2022 meta
->map_ptr
= reg
->map_ptr
;
2023 } else if (arg_type
== ARG_PTR_TO_MAP_KEY
) {
2024 /* bpf_map_xxx(..., map_ptr, ..., key) call:
2025 * check that [key, key + map->key_size) are within
2026 * stack limits and initialized
2028 if (!meta
->map_ptr
) {
2029 /* in function declaration map_ptr must come before
2030 * map_key, so that it's verified and known before
2031 * we have to check map_key here. Otherwise it means
2032 * that kernel subsystem misconfigured verifier
2034 verbose(env
, "invalid map_ptr to access map->key\n");
2037 err
= check_helper_mem_access(env
, regno
,
2038 meta
->map_ptr
->key_size
, false,
2040 } else if (arg_type
== ARG_PTR_TO_MAP_VALUE
) {
2041 /* bpf_map_xxx(..., map_ptr, ..., value) call:
2042 * check [value, value + map->value_size) validity
2044 if (!meta
->map_ptr
) {
2045 /* kernel subsystem misconfigured verifier */
2046 verbose(env
, "invalid map_ptr to access map->value\n");
2049 err
= check_helper_mem_access(env
, regno
,
2050 meta
->map_ptr
->value_size
, false,
2052 } else if (arg_type_is_mem_size(arg_type
)) {
2053 bool zero_size_allowed
= (arg_type
== ARG_CONST_SIZE_OR_ZERO
);
2055 /* remember the mem_size which may be used later
2056 * to refine return values.
2058 meta
->msize_smax_value
= reg
->smax_value
;
2059 meta
->msize_umax_value
= reg
->umax_value
;
2061 /* The register is SCALAR_VALUE; the access check
2062 * happens using its boundaries.
2064 if (!tnum_is_const(reg
->var_off
))
2065 /* For unprivileged variable accesses, disable raw
2066 * mode so that the program is required to
2067 * initialize all the memory that the helper could
2068 * just partially fill up.
2072 if (reg
->smin_value
< 0) {
2073 verbose(env
, "R%d min value is negative, either use unsigned or 'var &= const'\n",
2078 if (reg
->umin_value
== 0) {
2079 err
= check_helper_mem_access(env
, regno
- 1, 0,
2086 if (reg
->umax_value
>= BPF_MAX_VAR_SIZ
) {
2087 verbose(env
, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
2091 err
= check_helper_mem_access(env
, regno
- 1,
2093 zero_size_allowed
, meta
);
2098 verbose(env
, "R%d type=%s expected=%s\n", regno
,
2099 reg_type_str
[type
], reg_type_str
[expected_type
]);
2103 static int check_map_func_compatibility(struct bpf_verifier_env
*env
,
2104 struct bpf_map
*map
, int func_id
)
2109 /* We need a two way check, first is from map perspective ... */
2110 switch (map
->map_type
) {
2111 case BPF_MAP_TYPE_PROG_ARRAY
:
2112 if (func_id
!= BPF_FUNC_tail_call
)
2115 case BPF_MAP_TYPE_PERF_EVENT_ARRAY
:
2116 if (func_id
!= BPF_FUNC_perf_event_read
&&
2117 func_id
!= BPF_FUNC_perf_event_output
&&
2118 func_id
!= BPF_FUNC_perf_event_read_value
)
2121 case BPF_MAP_TYPE_STACK_TRACE
:
2122 if (func_id
!= BPF_FUNC_get_stackid
)
2125 case BPF_MAP_TYPE_CGROUP_ARRAY
:
2126 if (func_id
!= BPF_FUNC_skb_under_cgroup
&&
2127 func_id
!= BPF_FUNC_current_task_under_cgroup
)
2130 /* devmap returns a pointer to a live net_device ifindex that we cannot
2131 * allow to be modified from bpf side. So do not allow lookup elements
2134 case BPF_MAP_TYPE_DEVMAP
:
2135 if (func_id
!= BPF_FUNC_redirect_map
)
2138 /* Restrict bpf side of cpumap and xskmap, open when use-cases
2141 case BPF_MAP_TYPE_CPUMAP
:
2142 case BPF_MAP_TYPE_XSKMAP
:
2143 if (func_id
!= BPF_FUNC_redirect_map
)
2146 case BPF_MAP_TYPE_ARRAY_OF_MAPS
:
2147 case BPF_MAP_TYPE_HASH_OF_MAPS
:
2148 if (func_id
!= BPF_FUNC_map_lookup_elem
)
2151 case BPF_MAP_TYPE_SOCKMAP
:
2152 if (func_id
!= BPF_FUNC_sk_redirect_map
&&
2153 func_id
!= BPF_FUNC_sock_map_update
&&
2154 func_id
!= BPF_FUNC_map_delete_elem
&&
2155 func_id
!= BPF_FUNC_msg_redirect_map
)
2158 case BPF_MAP_TYPE_SOCKHASH
:
2159 if (func_id
!= BPF_FUNC_sk_redirect_hash
&&
2160 func_id
!= BPF_FUNC_sock_hash_update
&&
2161 func_id
!= BPF_FUNC_map_delete_elem
&&
2162 func_id
!= BPF_FUNC_msg_redirect_hash
)
2169 /* ... and second from the function itself. */
2171 case BPF_FUNC_tail_call
:
2172 if (map
->map_type
!= BPF_MAP_TYPE_PROG_ARRAY
)
2174 if (env
->subprog_cnt
> 1) {
2175 verbose(env
, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
2179 case BPF_FUNC_perf_event_read
:
2180 case BPF_FUNC_perf_event_output
:
2181 case BPF_FUNC_perf_event_read_value
:
2182 if (map
->map_type
!= BPF_MAP_TYPE_PERF_EVENT_ARRAY
)
2185 case BPF_FUNC_get_stackid
:
2186 if (map
->map_type
!= BPF_MAP_TYPE_STACK_TRACE
)
2189 case BPF_FUNC_current_task_under_cgroup
:
2190 case BPF_FUNC_skb_under_cgroup
:
2191 if (map
->map_type
!= BPF_MAP_TYPE_CGROUP_ARRAY
)
2194 case BPF_FUNC_redirect_map
:
2195 if (map
->map_type
!= BPF_MAP_TYPE_DEVMAP
&&
2196 map
->map_type
!= BPF_MAP_TYPE_CPUMAP
&&
2197 map
->map_type
!= BPF_MAP_TYPE_XSKMAP
)
2200 case BPF_FUNC_sk_redirect_map
:
2201 case BPF_FUNC_msg_redirect_map
:
2202 case BPF_FUNC_sock_map_update
:
2203 if (map
->map_type
!= BPF_MAP_TYPE_SOCKMAP
)
2206 case BPF_FUNC_sk_redirect_hash
:
2207 case BPF_FUNC_msg_redirect_hash
:
2208 case BPF_FUNC_sock_hash_update
:
2209 if (map
->map_type
!= BPF_MAP_TYPE_SOCKHASH
)
2218 verbose(env
, "cannot pass map_type %d into func %s#%d\n",
2219 map
->map_type
, func_id_name(func_id
), func_id
);
2223 static bool check_raw_mode_ok(const struct bpf_func_proto
*fn
)
2227 if (fn
->arg1_type
== ARG_PTR_TO_UNINIT_MEM
)
2229 if (fn
->arg2_type
== ARG_PTR_TO_UNINIT_MEM
)
2231 if (fn
->arg3_type
== ARG_PTR_TO_UNINIT_MEM
)
2233 if (fn
->arg4_type
== ARG_PTR_TO_UNINIT_MEM
)
2235 if (fn
->arg5_type
== ARG_PTR_TO_UNINIT_MEM
)
2238 /* We only support one arg being in raw mode at the moment,
2239 * which is sufficient for the helper functions we have
2245 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr
,
2246 enum bpf_arg_type arg_next
)
2248 return (arg_type_is_mem_ptr(arg_curr
) &&
2249 !arg_type_is_mem_size(arg_next
)) ||
2250 (!arg_type_is_mem_ptr(arg_curr
) &&
2251 arg_type_is_mem_size(arg_next
));
2254 static bool check_arg_pair_ok(const struct bpf_func_proto
*fn
)
2256 /* bpf_xxx(..., buf, len) call will access 'len'
2257 * bytes from memory 'buf'. Both arg types need
2258 * to be paired, so make sure there's no buggy
2259 * helper function specification.
2261 if (arg_type_is_mem_size(fn
->arg1_type
) ||
2262 arg_type_is_mem_ptr(fn
->arg5_type
) ||
2263 check_args_pair_invalid(fn
->arg1_type
, fn
->arg2_type
) ||
2264 check_args_pair_invalid(fn
->arg2_type
, fn
->arg3_type
) ||
2265 check_args_pair_invalid(fn
->arg3_type
, fn
->arg4_type
) ||
2266 check_args_pair_invalid(fn
->arg4_type
, fn
->arg5_type
))
2272 static int check_func_proto(const struct bpf_func_proto
*fn
)
2274 return check_raw_mode_ok(fn
) &&
2275 check_arg_pair_ok(fn
) ? 0 : -EINVAL
;
2278 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
2279 * are now invalid, so turn them into unknown SCALAR_VALUE.
2281 static void __clear_all_pkt_pointers(struct bpf_verifier_env
*env
,
2282 struct bpf_func_state
*state
)
2284 struct bpf_reg_state
*regs
= state
->regs
, *reg
;
2287 for (i
= 0; i
< MAX_BPF_REG
; i
++)
2288 if (reg_is_pkt_pointer_any(®s
[i
]))
2289 mark_reg_unknown(env
, regs
, i
);
2291 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
2292 if (state
->stack
[i
].slot_type
[0] != STACK_SPILL
)
2294 reg
= &state
->stack
[i
].spilled_ptr
;
2295 if (reg_is_pkt_pointer_any(reg
))
2296 __mark_reg_unknown(reg
);
2300 static void clear_all_pkt_pointers(struct bpf_verifier_env
*env
)
2302 struct bpf_verifier_state
*vstate
= env
->cur_state
;
2305 for (i
= 0; i
<= vstate
->curframe
; i
++)
2306 __clear_all_pkt_pointers(env
, vstate
->frame
[i
]);
2309 static int check_func_call(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
,
2312 struct bpf_verifier_state
*state
= env
->cur_state
;
2313 struct bpf_func_state
*caller
, *callee
;
2314 int i
, subprog
, target_insn
;
2316 if (state
->curframe
+ 1 >= MAX_CALL_FRAMES
) {
2317 verbose(env
, "the call stack of %d frames is too deep\n",
2318 state
->curframe
+ 2);
2322 target_insn
= *insn_idx
+ insn
->imm
;
2323 subprog
= find_subprog(env
, target_insn
+ 1);
2325 verbose(env
, "verifier bug. No program starts at insn %d\n",
2330 caller
= state
->frame
[state
->curframe
];
2331 if (state
->frame
[state
->curframe
+ 1]) {
2332 verbose(env
, "verifier bug. Frame %d already allocated\n",
2333 state
->curframe
+ 1);
2337 callee
= kzalloc(sizeof(*callee
), GFP_KERNEL
);
2340 state
->frame
[state
->curframe
+ 1] = callee
;
2342 /* callee cannot access r0, r6 - r9 for reading and has to write
2343 * into its own stack before reading from it.
2344 * callee can read/write into caller's stack
2346 init_func_state(env
, callee
,
2347 /* remember the callsite, it will be used by bpf_exit */
2348 *insn_idx
/* callsite */,
2349 state
->curframe
+ 1 /* frameno within this callchain */,
2350 subprog
/* subprog number within this prog */);
2352 /* copy r1 - r5 args that callee can access */
2353 for (i
= BPF_REG_1
; i
<= BPF_REG_5
; i
++)
2354 callee
->regs
[i
] = caller
->regs
[i
];
2356 /* after the call regsiters r0 - r5 were scratched */
2357 for (i
= 0; i
< CALLER_SAVED_REGS
; i
++) {
2358 mark_reg_not_init(env
, caller
->regs
, caller_saved
[i
]);
2359 check_reg_arg(env
, caller_saved
[i
], DST_OP_NO_MARK
);
2362 /* only increment it after check_reg_arg() finished */
2365 /* and go analyze first insn of the callee */
2366 *insn_idx
= target_insn
;
2368 if (env
->log
.level
) {
2369 verbose(env
, "caller:\n");
2370 print_verifier_state(env
, caller
);
2371 verbose(env
, "callee:\n");
2372 print_verifier_state(env
, callee
);
2377 static int prepare_func_exit(struct bpf_verifier_env
*env
, int *insn_idx
)
2379 struct bpf_verifier_state
*state
= env
->cur_state
;
2380 struct bpf_func_state
*caller
, *callee
;
2381 struct bpf_reg_state
*r0
;
2383 callee
= state
->frame
[state
->curframe
];
2384 r0
= &callee
->regs
[BPF_REG_0
];
2385 if (r0
->type
== PTR_TO_STACK
) {
2386 /* technically it's ok to return caller's stack pointer
2387 * (or caller's caller's pointer) back to the caller,
2388 * since these pointers are valid. Only current stack
2389 * pointer will be invalid as soon as function exits,
2390 * but let's be conservative
2392 verbose(env
, "cannot return stack pointer to the caller\n");
2397 caller
= state
->frame
[state
->curframe
];
2398 /* return to the caller whatever r0 had in the callee */
2399 caller
->regs
[BPF_REG_0
] = *r0
;
2401 *insn_idx
= callee
->callsite
+ 1;
2402 if (env
->log
.level
) {
2403 verbose(env
, "returning from callee:\n");
2404 print_verifier_state(env
, callee
);
2405 verbose(env
, "to caller at %d:\n", *insn_idx
);
2406 print_verifier_state(env
, caller
);
2408 /* clear everything in the callee */
2409 free_func_state(callee
);
2410 state
->frame
[state
->curframe
+ 1] = NULL
;
2414 static void do_refine_retval_range(struct bpf_reg_state
*regs
, int ret_type
,
2416 struct bpf_call_arg_meta
*meta
)
2418 struct bpf_reg_state
*ret_reg
= ®s
[BPF_REG_0
];
2420 if (ret_type
!= RET_INTEGER
||
2421 (func_id
!= BPF_FUNC_get_stack
&&
2422 func_id
!= BPF_FUNC_probe_read_str
))
2425 ret_reg
->smax_value
= meta
->msize_smax_value
;
2426 ret_reg
->umax_value
= meta
->msize_umax_value
;
2427 __reg_deduce_bounds(ret_reg
);
2428 __reg_bound_offset(ret_reg
);
2432 record_func_map(struct bpf_verifier_env
*env
, struct bpf_call_arg_meta
*meta
,
2433 int func_id
, int insn_idx
)
2435 struct bpf_insn_aux_data
*aux
= &env
->insn_aux_data
[insn_idx
];
2437 if (func_id
!= BPF_FUNC_tail_call
&&
2438 func_id
!= BPF_FUNC_map_lookup_elem
&&
2439 func_id
!= BPF_FUNC_map_update_elem
&&
2440 func_id
!= BPF_FUNC_map_delete_elem
)
2443 if (meta
->map_ptr
== NULL
) {
2444 verbose(env
, "kernel subsystem misconfigured verifier\n");
2448 if (!BPF_MAP_PTR(aux
->map_state
))
2449 bpf_map_ptr_store(aux
, meta
->map_ptr
,
2450 meta
->map_ptr
->unpriv_array
);
2451 else if (BPF_MAP_PTR(aux
->map_state
) != meta
->map_ptr
)
2452 bpf_map_ptr_store(aux
, BPF_MAP_PTR_POISON
,
2453 meta
->map_ptr
->unpriv_array
);
2457 static int check_helper_call(struct bpf_verifier_env
*env
, int func_id
, int insn_idx
)
2459 const struct bpf_func_proto
*fn
= NULL
;
2460 struct bpf_reg_state
*regs
;
2461 struct bpf_call_arg_meta meta
;
2465 /* find function prototype */
2466 if (func_id
< 0 || func_id
>= __BPF_FUNC_MAX_ID
) {
2467 verbose(env
, "invalid func %s#%d\n", func_id_name(func_id
),
2472 if (env
->ops
->get_func_proto
)
2473 fn
= env
->ops
->get_func_proto(func_id
, env
->prog
);
2475 verbose(env
, "unknown func %s#%d\n", func_id_name(func_id
),
2480 /* eBPF programs must be GPL compatible to use GPL-ed functions */
2481 if (!env
->prog
->gpl_compatible
&& fn
->gpl_only
) {
2482 verbose(env
, "cannot call GPL-restricted function from non-GPL compatible program\n");
2486 /* With LD_ABS/IND some JITs save/restore skb from r1. */
2487 changes_data
= bpf_helper_changes_pkt_data(fn
->func
);
2488 if (changes_data
&& fn
->arg1_type
!= ARG_PTR_TO_CTX
) {
2489 verbose(env
, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
2490 func_id_name(func_id
), func_id
);
2494 memset(&meta
, 0, sizeof(meta
));
2495 meta
.pkt_access
= fn
->pkt_access
;
2497 err
= check_func_proto(fn
);
2499 verbose(env
, "kernel subsystem misconfigured func %s#%d\n",
2500 func_id_name(func_id
), func_id
);
2505 err
= check_func_arg(env
, BPF_REG_1
, fn
->arg1_type
, &meta
);
2508 err
= check_func_arg(env
, BPF_REG_2
, fn
->arg2_type
, &meta
);
2511 err
= check_func_arg(env
, BPF_REG_3
, fn
->arg3_type
, &meta
);
2514 err
= check_func_arg(env
, BPF_REG_4
, fn
->arg4_type
, &meta
);
2517 err
= check_func_arg(env
, BPF_REG_5
, fn
->arg5_type
, &meta
);
2521 err
= record_func_map(env
, &meta
, func_id
, insn_idx
);
2525 /* Mark slots with STACK_MISC in case of raw mode, stack offset
2526 * is inferred from register state.
2528 for (i
= 0; i
< meta
.access_size
; i
++) {
2529 err
= check_mem_access(env
, insn_idx
, meta
.regno
, i
, BPF_B
,
2530 BPF_WRITE
, -1, false);
2535 regs
= cur_regs(env
);
2536 /* reset caller saved regs */
2537 for (i
= 0; i
< CALLER_SAVED_REGS
; i
++) {
2538 mark_reg_not_init(env
, regs
, caller_saved
[i
]);
2539 check_reg_arg(env
, caller_saved
[i
], DST_OP_NO_MARK
);
2542 /* update return register (already marked as written above) */
2543 if (fn
->ret_type
== RET_INTEGER
) {
2544 /* sets type to SCALAR_VALUE */
2545 mark_reg_unknown(env
, regs
, BPF_REG_0
);
2546 } else if (fn
->ret_type
== RET_VOID
) {
2547 regs
[BPF_REG_0
].type
= NOT_INIT
;
2548 } else if (fn
->ret_type
== RET_PTR_TO_MAP_VALUE_OR_NULL
) {
2549 regs
[BPF_REG_0
].type
= PTR_TO_MAP_VALUE_OR_NULL
;
2550 /* There is no offset yet applied, variable or fixed */
2551 mark_reg_known_zero(env
, regs
, BPF_REG_0
);
2552 regs
[BPF_REG_0
].off
= 0;
2553 /* remember map_ptr, so that check_map_access()
2554 * can check 'value_size' boundary of memory access
2555 * to map element returned from bpf_map_lookup_elem()
2557 if (meta
.map_ptr
== NULL
) {
2559 "kernel subsystem misconfigured verifier\n");
2562 regs
[BPF_REG_0
].map_ptr
= meta
.map_ptr
;
2563 regs
[BPF_REG_0
].id
= ++env
->id_gen
;
2565 verbose(env
, "unknown return type %d of func %s#%d\n",
2566 fn
->ret_type
, func_id_name(func_id
), func_id
);
2570 do_refine_retval_range(regs
, fn
->ret_type
, func_id
, &meta
);
2572 err
= check_map_func_compatibility(env
, meta
.map_ptr
, func_id
);
2576 if (func_id
== BPF_FUNC_get_stack
&& !env
->prog
->has_callchain_buf
) {
2577 const char *err_str
;
2579 #ifdef CONFIG_PERF_EVENTS
2580 err
= get_callchain_buffers(sysctl_perf_event_max_stack
);
2581 err_str
= "cannot get callchain buffer for func %s#%d\n";
2584 err_str
= "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
2587 verbose(env
, err_str
, func_id_name(func_id
), func_id
);
2591 env
->prog
->has_callchain_buf
= true;
2595 clear_all_pkt_pointers(env
);
2599 static bool signed_add_overflows(s64 a
, s64 b
)
2601 /* Do the add in u64, where overflow is well-defined */
2602 s64 res
= (s64
)((u64
)a
+ (u64
)b
);
2609 static bool signed_sub_overflows(s64 a
, s64 b
)
2611 /* Do the sub in u64, where overflow is well-defined */
2612 s64 res
= (s64
)((u64
)a
- (u64
)b
);
2619 static bool check_reg_sane_offset(struct bpf_verifier_env
*env
,
2620 const struct bpf_reg_state
*reg
,
2621 enum bpf_reg_type type
)
2623 bool known
= tnum_is_const(reg
->var_off
);
2624 s64 val
= reg
->var_off
.value
;
2625 s64 smin
= reg
->smin_value
;
2627 if (known
&& (val
>= BPF_MAX_VAR_OFF
|| val
<= -BPF_MAX_VAR_OFF
)) {
2628 verbose(env
, "math between %s pointer and %lld is not allowed\n",
2629 reg_type_str
[type
], val
);
2633 if (reg
->off
>= BPF_MAX_VAR_OFF
|| reg
->off
<= -BPF_MAX_VAR_OFF
) {
2634 verbose(env
, "%s pointer offset %d is not allowed\n",
2635 reg_type_str
[type
], reg
->off
);
2639 if (smin
== S64_MIN
) {
2640 verbose(env
, "math between %s pointer and register with unbounded min value is not allowed\n",
2641 reg_type_str
[type
]);
2645 if (smin
>= BPF_MAX_VAR_OFF
|| smin
<= -BPF_MAX_VAR_OFF
) {
2646 verbose(env
, "value %lld makes %s pointer be out of bounds\n",
2647 smin
, reg_type_str
[type
]);
2654 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
2655 * Caller should also handle BPF_MOV case separately.
2656 * If we return -EACCES, caller may want to try again treating pointer as a
2657 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
2659 static int adjust_ptr_min_max_vals(struct bpf_verifier_env
*env
,
2660 struct bpf_insn
*insn
,
2661 const struct bpf_reg_state
*ptr_reg
,
2662 const struct bpf_reg_state
*off_reg
)
2664 struct bpf_verifier_state
*vstate
= env
->cur_state
;
2665 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
2666 struct bpf_reg_state
*regs
= state
->regs
, *dst_reg
;
2667 bool known
= tnum_is_const(off_reg
->var_off
);
2668 s64 smin_val
= off_reg
->smin_value
, smax_val
= off_reg
->smax_value
,
2669 smin_ptr
= ptr_reg
->smin_value
, smax_ptr
= ptr_reg
->smax_value
;
2670 u64 umin_val
= off_reg
->umin_value
, umax_val
= off_reg
->umax_value
,
2671 umin_ptr
= ptr_reg
->umin_value
, umax_ptr
= ptr_reg
->umax_value
;
2672 u8 opcode
= BPF_OP(insn
->code
);
2673 u32 dst
= insn
->dst_reg
;
2675 dst_reg
= ®s
[dst
];
2677 if ((known
&& (smin_val
!= smax_val
|| umin_val
!= umax_val
)) ||
2678 smin_val
> smax_val
|| umin_val
> umax_val
) {
2679 /* Taint dst register if offset had invalid bounds derived from
2680 * e.g. dead branches.
2682 __mark_reg_unknown(dst_reg
);
2686 if (BPF_CLASS(insn
->code
) != BPF_ALU64
) {
2687 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
2689 "R%d 32-bit pointer arithmetic prohibited\n",
2694 if (ptr_reg
->type
== PTR_TO_MAP_VALUE_OR_NULL
) {
2695 verbose(env
, "R%d pointer arithmetic on PTR_TO_MAP_VALUE_OR_NULL prohibited, null-check it first\n",
2699 if (ptr_reg
->type
== CONST_PTR_TO_MAP
) {
2700 verbose(env
, "R%d pointer arithmetic on CONST_PTR_TO_MAP prohibited\n",
2704 if (ptr_reg
->type
== PTR_TO_PACKET_END
) {
2705 verbose(env
, "R%d pointer arithmetic on PTR_TO_PACKET_END prohibited\n",
2710 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
2711 * The id may be overwritten later if we create a new variable offset.
2713 dst_reg
->type
= ptr_reg
->type
;
2714 dst_reg
->id
= ptr_reg
->id
;
2716 if (!check_reg_sane_offset(env
, off_reg
, ptr_reg
->type
) ||
2717 !check_reg_sane_offset(env
, ptr_reg
, ptr_reg
->type
))
2722 /* We can take a fixed offset as long as it doesn't overflow
2723 * the s32 'off' field
2725 if (known
&& (ptr_reg
->off
+ smin_val
==
2726 (s64
)(s32
)(ptr_reg
->off
+ smin_val
))) {
2727 /* pointer += K. Accumulate it into fixed offset */
2728 dst_reg
->smin_value
= smin_ptr
;
2729 dst_reg
->smax_value
= smax_ptr
;
2730 dst_reg
->umin_value
= umin_ptr
;
2731 dst_reg
->umax_value
= umax_ptr
;
2732 dst_reg
->var_off
= ptr_reg
->var_off
;
2733 dst_reg
->off
= ptr_reg
->off
+ smin_val
;
2734 dst_reg
->range
= ptr_reg
->range
;
2737 /* A new variable offset is created. Note that off_reg->off
2738 * == 0, since it's a scalar.
2739 * dst_reg gets the pointer type and since some positive
2740 * integer value was added to the pointer, give it a new 'id'
2741 * if it's a PTR_TO_PACKET.
2742 * this creates a new 'base' pointer, off_reg (variable) gets
2743 * added into the variable offset, and we copy the fixed offset
2746 if (signed_add_overflows(smin_ptr
, smin_val
) ||
2747 signed_add_overflows(smax_ptr
, smax_val
)) {
2748 dst_reg
->smin_value
= S64_MIN
;
2749 dst_reg
->smax_value
= S64_MAX
;
2751 dst_reg
->smin_value
= smin_ptr
+ smin_val
;
2752 dst_reg
->smax_value
= smax_ptr
+ smax_val
;
2754 if (umin_ptr
+ umin_val
< umin_ptr
||
2755 umax_ptr
+ umax_val
< umax_ptr
) {
2756 dst_reg
->umin_value
= 0;
2757 dst_reg
->umax_value
= U64_MAX
;
2759 dst_reg
->umin_value
= umin_ptr
+ umin_val
;
2760 dst_reg
->umax_value
= umax_ptr
+ umax_val
;
2762 dst_reg
->var_off
= tnum_add(ptr_reg
->var_off
, off_reg
->var_off
);
2763 dst_reg
->off
= ptr_reg
->off
;
2764 if (reg_is_pkt_pointer(ptr_reg
)) {
2765 dst_reg
->id
= ++env
->id_gen
;
2766 /* something was added to pkt_ptr, set range to zero */
2771 if (dst_reg
== off_reg
) {
2772 /* scalar -= pointer. Creates an unknown scalar */
2773 verbose(env
, "R%d tried to subtract pointer from scalar\n",
2777 /* We don't allow subtraction from FP, because (according to
2778 * test_verifier.c test "invalid fp arithmetic", JITs might not
2779 * be able to deal with it.
2781 if (ptr_reg
->type
== PTR_TO_STACK
) {
2782 verbose(env
, "R%d subtraction from stack pointer prohibited\n",
2786 if (known
&& (ptr_reg
->off
- smin_val
==
2787 (s64
)(s32
)(ptr_reg
->off
- smin_val
))) {
2788 /* pointer -= K. Subtract it from fixed offset */
2789 dst_reg
->smin_value
= smin_ptr
;
2790 dst_reg
->smax_value
= smax_ptr
;
2791 dst_reg
->umin_value
= umin_ptr
;
2792 dst_reg
->umax_value
= umax_ptr
;
2793 dst_reg
->var_off
= ptr_reg
->var_off
;
2794 dst_reg
->id
= ptr_reg
->id
;
2795 dst_reg
->off
= ptr_reg
->off
- smin_val
;
2796 dst_reg
->range
= ptr_reg
->range
;
2799 /* A new variable offset is created. If the subtrahend is known
2800 * nonnegative, then any reg->range we had before is still good.
2802 if (signed_sub_overflows(smin_ptr
, smax_val
) ||
2803 signed_sub_overflows(smax_ptr
, smin_val
)) {
2804 /* Overflow possible, we know nothing */
2805 dst_reg
->smin_value
= S64_MIN
;
2806 dst_reg
->smax_value
= S64_MAX
;
2808 dst_reg
->smin_value
= smin_ptr
- smax_val
;
2809 dst_reg
->smax_value
= smax_ptr
- smin_val
;
2811 if (umin_ptr
< umax_val
) {
2812 /* Overflow possible, we know nothing */
2813 dst_reg
->umin_value
= 0;
2814 dst_reg
->umax_value
= U64_MAX
;
2816 /* Cannot overflow (as long as bounds are consistent) */
2817 dst_reg
->umin_value
= umin_ptr
- umax_val
;
2818 dst_reg
->umax_value
= umax_ptr
- umin_val
;
2820 dst_reg
->var_off
= tnum_sub(ptr_reg
->var_off
, off_reg
->var_off
);
2821 dst_reg
->off
= ptr_reg
->off
;
2822 if (reg_is_pkt_pointer(ptr_reg
)) {
2823 dst_reg
->id
= ++env
->id_gen
;
2824 /* something was added to pkt_ptr, set range to zero */
2832 /* bitwise ops on pointers are troublesome, prohibit. */
2833 verbose(env
, "R%d bitwise operator %s on pointer prohibited\n",
2834 dst
, bpf_alu_string
[opcode
>> 4]);
2837 /* other operators (e.g. MUL,LSH) produce non-pointer results */
2838 verbose(env
, "R%d pointer arithmetic with %s operator prohibited\n",
2839 dst
, bpf_alu_string
[opcode
>> 4]);
2843 if (!check_reg_sane_offset(env
, dst_reg
, ptr_reg
->type
))
2846 __update_reg_bounds(dst_reg
);
2847 __reg_deduce_bounds(dst_reg
);
2848 __reg_bound_offset(dst_reg
);
2852 /* WARNING: This function does calculations on 64-bit values, but the actual
2853 * execution may occur on 32-bit values. Therefore, things like bitshifts
2854 * need extra checks in the 32-bit case.
2856 static int adjust_scalar_min_max_vals(struct bpf_verifier_env
*env
,
2857 struct bpf_insn
*insn
,
2858 struct bpf_reg_state
*dst_reg
,
2859 struct bpf_reg_state src_reg
)
2861 struct bpf_reg_state
*regs
= cur_regs(env
);
2862 u8 opcode
= BPF_OP(insn
->code
);
2863 bool src_known
, dst_known
;
2864 s64 smin_val
, smax_val
;
2865 u64 umin_val
, umax_val
;
2866 u64 insn_bitness
= (BPF_CLASS(insn
->code
) == BPF_ALU64
) ? 64 : 32;
2868 smin_val
= src_reg
.smin_value
;
2869 smax_val
= src_reg
.smax_value
;
2870 umin_val
= src_reg
.umin_value
;
2871 umax_val
= src_reg
.umax_value
;
2872 src_known
= tnum_is_const(src_reg
.var_off
);
2873 dst_known
= tnum_is_const(dst_reg
->var_off
);
2875 if ((src_known
&& (smin_val
!= smax_val
|| umin_val
!= umax_val
)) ||
2876 smin_val
> smax_val
|| umin_val
> umax_val
) {
2877 /* Taint dst register if offset had invalid bounds derived from
2878 * e.g. dead branches.
2880 __mark_reg_unknown(dst_reg
);
2885 opcode
!= BPF_ADD
&& opcode
!= BPF_SUB
&& opcode
!= BPF_AND
) {
2886 __mark_reg_unknown(dst_reg
);
2892 if (signed_add_overflows(dst_reg
->smin_value
, smin_val
) ||
2893 signed_add_overflows(dst_reg
->smax_value
, smax_val
)) {
2894 dst_reg
->smin_value
= S64_MIN
;
2895 dst_reg
->smax_value
= S64_MAX
;
2897 dst_reg
->smin_value
+= smin_val
;
2898 dst_reg
->smax_value
+= smax_val
;
2900 if (dst_reg
->umin_value
+ umin_val
< umin_val
||
2901 dst_reg
->umax_value
+ umax_val
< umax_val
) {
2902 dst_reg
->umin_value
= 0;
2903 dst_reg
->umax_value
= U64_MAX
;
2905 dst_reg
->umin_value
+= umin_val
;
2906 dst_reg
->umax_value
+= umax_val
;
2908 dst_reg
->var_off
= tnum_add(dst_reg
->var_off
, src_reg
.var_off
);
2911 if (signed_sub_overflows(dst_reg
->smin_value
, smax_val
) ||
2912 signed_sub_overflows(dst_reg
->smax_value
, smin_val
)) {
2913 /* Overflow possible, we know nothing */
2914 dst_reg
->smin_value
= S64_MIN
;
2915 dst_reg
->smax_value
= S64_MAX
;
2917 dst_reg
->smin_value
-= smax_val
;
2918 dst_reg
->smax_value
-= smin_val
;
2920 if (dst_reg
->umin_value
< umax_val
) {
2921 /* Overflow possible, we know nothing */
2922 dst_reg
->umin_value
= 0;
2923 dst_reg
->umax_value
= U64_MAX
;
2925 /* Cannot overflow (as long as bounds are consistent) */
2926 dst_reg
->umin_value
-= umax_val
;
2927 dst_reg
->umax_value
-= umin_val
;
2929 dst_reg
->var_off
= tnum_sub(dst_reg
->var_off
, src_reg
.var_off
);
2932 dst_reg
->var_off
= tnum_mul(dst_reg
->var_off
, src_reg
.var_off
);
2933 if (smin_val
< 0 || dst_reg
->smin_value
< 0) {
2934 /* Ain't nobody got time to multiply that sign */
2935 __mark_reg_unbounded(dst_reg
);
2936 __update_reg_bounds(dst_reg
);
2939 /* Both values are positive, so we can work with unsigned and
2940 * copy the result to signed (unless it exceeds S64_MAX).
2942 if (umax_val
> U32_MAX
|| dst_reg
->umax_value
> U32_MAX
) {
2943 /* Potential overflow, we know nothing */
2944 __mark_reg_unbounded(dst_reg
);
2945 /* (except what we can learn from the var_off) */
2946 __update_reg_bounds(dst_reg
);
2949 dst_reg
->umin_value
*= umin_val
;
2950 dst_reg
->umax_value
*= umax_val
;
2951 if (dst_reg
->umax_value
> S64_MAX
) {
2952 /* Overflow possible, we know nothing */
2953 dst_reg
->smin_value
= S64_MIN
;
2954 dst_reg
->smax_value
= S64_MAX
;
2956 dst_reg
->smin_value
= dst_reg
->umin_value
;
2957 dst_reg
->smax_value
= dst_reg
->umax_value
;
2961 if (src_known
&& dst_known
) {
2962 __mark_reg_known(dst_reg
, dst_reg
->var_off
.value
&
2963 src_reg
.var_off
.value
);
2966 /* We get our minimum from the var_off, since that's inherently
2967 * bitwise. Our maximum is the minimum of the operands' maxima.
2969 dst_reg
->var_off
= tnum_and(dst_reg
->var_off
, src_reg
.var_off
);
2970 dst_reg
->umin_value
= dst_reg
->var_off
.value
;
2971 dst_reg
->umax_value
= min(dst_reg
->umax_value
, umax_val
);
2972 if (dst_reg
->smin_value
< 0 || smin_val
< 0) {
2973 /* Lose signed bounds when ANDing negative numbers,
2974 * ain't nobody got time for that.
2976 dst_reg
->smin_value
= S64_MIN
;
2977 dst_reg
->smax_value
= S64_MAX
;
2979 /* ANDing two positives gives a positive, so safe to
2980 * cast result into s64.
2982 dst_reg
->smin_value
= dst_reg
->umin_value
;
2983 dst_reg
->smax_value
= dst_reg
->umax_value
;
2985 /* We may learn something more from the var_off */
2986 __update_reg_bounds(dst_reg
);
2989 if (src_known
&& dst_known
) {
2990 __mark_reg_known(dst_reg
, dst_reg
->var_off
.value
|
2991 src_reg
.var_off
.value
);
2994 /* We get our maximum from the var_off, and our minimum is the
2995 * maximum of the operands' minima
2997 dst_reg
->var_off
= tnum_or(dst_reg
->var_off
, src_reg
.var_off
);
2998 dst_reg
->umin_value
= max(dst_reg
->umin_value
, umin_val
);
2999 dst_reg
->umax_value
= dst_reg
->var_off
.value
|
3000 dst_reg
->var_off
.mask
;
3001 if (dst_reg
->smin_value
< 0 || smin_val
< 0) {
3002 /* Lose signed bounds when ORing negative numbers,
3003 * ain't nobody got time for that.
3005 dst_reg
->smin_value
= S64_MIN
;
3006 dst_reg
->smax_value
= S64_MAX
;
3008 /* ORing two positives gives a positive, so safe to
3009 * cast result into s64.
3011 dst_reg
->smin_value
= dst_reg
->umin_value
;
3012 dst_reg
->smax_value
= dst_reg
->umax_value
;
3014 /* We may learn something more from the var_off */
3015 __update_reg_bounds(dst_reg
);
3018 if (umax_val
>= insn_bitness
) {
3019 /* Shifts greater than 31 or 63 are undefined.
3020 * This includes shifts by a negative number.
3022 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
3025 /* We lose all sign bit information (except what we can pick
3028 dst_reg
->smin_value
= S64_MIN
;
3029 dst_reg
->smax_value
= S64_MAX
;
3030 /* If we might shift our top bit out, then we know nothing */
3031 if (dst_reg
->umax_value
> 1ULL << (63 - umax_val
)) {
3032 dst_reg
->umin_value
= 0;
3033 dst_reg
->umax_value
= U64_MAX
;
3035 dst_reg
->umin_value
<<= umin_val
;
3036 dst_reg
->umax_value
<<= umax_val
;
3038 dst_reg
->var_off
= tnum_lshift(dst_reg
->var_off
, umin_val
);
3039 /* We may learn something more from the var_off */
3040 __update_reg_bounds(dst_reg
);
3043 if (umax_val
>= insn_bitness
) {
3044 /* Shifts greater than 31 or 63 are undefined.
3045 * This includes shifts by a negative number.
3047 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
3050 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
3051 * be negative, then either:
3052 * 1) src_reg might be zero, so the sign bit of the result is
3053 * unknown, so we lose our signed bounds
3054 * 2) it's known negative, thus the unsigned bounds capture the
3056 * 3) the signed bounds cross zero, so they tell us nothing
3058 * If the value in dst_reg is known nonnegative, then again the
3059 * unsigned bounts capture the signed bounds.
3060 * Thus, in all cases it suffices to blow away our signed bounds
3061 * and rely on inferring new ones from the unsigned bounds and
3062 * var_off of the result.
3064 dst_reg
->smin_value
= S64_MIN
;
3065 dst_reg
->smax_value
= S64_MAX
;
3066 dst_reg
->var_off
= tnum_rshift(dst_reg
->var_off
, umin_val
);
3067 dst_reg
->umin_value
>>= umax_val
;
3068 dst_reg
->umax_value
>>= umin_val
;
3069 /* We may learn something more from the var_off */
3070 __update_reg_bounds(dst_reg
);
3073 if (umax_val
>= insn_bitness
) {
3074 /* Shifts greater than 31 or 63 are undefined.
3075 * This includes shifts by a negative number.
3077 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
3081 /* Upon reaching here, src_known is true and
3082 * umax_val is equal to umin_val.
3084 dst_reg
->smin_value
>>= umin_val
;
3085 dst_reg
->smax_value
>>= umin_val
;
3086 dst_reg
->var_off
= tnum_arshift(dst_reg
->var_off
, umin_val
);
3088 /* blow away the dst_reg umin_value/umax_value and rely on
3089 * dst_reg var_off to refine the result.
3091 dst_reg
->umin_value
= 0;
3092 dst_reg
->umax_value
= U64_MAX
;
3093 __update_reg_bounds(dst_reg
);
3096 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
3100 if (BPF_CLASS(insn
->code
) != BPF_ALU64
) {
3101 /* 32-bit ALU ops are (32,32)->32 */
3102 coerce_reg_to_size(dst_reg
, 4);
3103 coerce_reg_to_size(&src_reg
, 4);
3106 __reg_deduce_bounds(dst_reg
);
3107 __reg_bound_offset(dst_reg
);
3111 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
3114 static int adjust_reg_min_max_vals(struct bpf_verifier_env
*env
,
3115 struct bpf_insn
*insn
)
3117 struct bpf_verifier_state
*vstate
= env
->cur_state
;
3118 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
3119 struct bpf_reg_state
*regs
= state
->regs
, *dst_reg
, *src_reg
;
3120 struct bpf_reg_state
*ptr_reg
= NULL
, off_reg
= {0};
3121 u8 opcode
= BPF_OP(insn
->code
);
3123 dst_reg
= ®s
[insn
->dst_reg
];
3125 if (dst_reg
->type
!= SCALAR_VALUE
)
3127 if (BPF_SRC(insn
->code
) == BPF_X
) {
3128 src_reg
= ®s
[insn
->src_reg
];
3129 if (src_reg
->type
!= SCALAR_VALUE
) {
3130 if (dst_reg
->type
!= SCALAR_VALUE
) {
3131 /* Combining two pointers by any ALU op yields
3132 * an arbitrary scalar. Disallow all math except
3133 * pointer subtraction
3135 if (opcode
== BPF_SUB
){
3136 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
3139 verbose(env
, "R%d pointer %s pointer prohibited\n",
3141 bpf_alu_string
[opcode
>> 4]);
3144 /* scalar += pointer
3145 * This is legal, but we have to reverse our
3146 * src/dest handling in computing the range
3148 return adjust_ptr_min_max_vals(env
, insn
,
3151 } else if (ptr_reg
) {
3152 /* pointer += scalar */
3153 return adjust_ptr_min_max_vals(env
, insn
,
3157 /* Pretend the src is a reg with a known value, since we only
3158 * need to be able to read from this state.
3160 off_reg
.type
= SCALAR_VALUE
;
3161 __mark_reg_known(&off_reg
, insn
->imm
);
3163 if (ptr_reg
) /* pointer += K */
3164 return adjust_ptr_min_max_vals(env
, insn
,
3168 /* Got here implies adding two SCALAR_VALUEs */
3169 if (WARN_ON_ONCE(ptr_reg
)) {
3170 print_verifier_state(env
, state
);
3171 verbose(env
, "verifier internal error: unexpected ptr_reg\n");
3174 if (WARN_ON(!src_reg
)) {
3175 print_verifier_state(env
, state
);
3176 verbose(env
, "verifier internal error: no src_reg\n");
3179 return adjust_scalar_min_max_vals(env
, insn
, dst_reg
, *src_reg
);
3182 /* check validity of 32-bit and 64-bit arithmetic operations */
3183 static int check_alu_op(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
3185 struct bpf_reg_state
*regs
= cur_regs(env
);
3186 u8 opcode
= BPF_OP(insn
->code
);
3189 if (opcode
== BPF_END
|| opcode
== BPF_NEG
) {
3190 if (opcode
== BPF_NEG
) {
3191 if (BPF_SRC(insn
->code
) != 0 ||
3192 insn
->src_reg
!= BPF_REG_0
||
3193 insn
->off
!= 0 || insn
->imm
!= 0) {
3194 verbose(env
, "BPF_NEG uses reserved fields\n");
3198 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0 ||
3199 (insn
->imm
!= 16 && insn
->imm
!= 32 && insn
->imm
!= 64) ||
3200 BPF_CLASS(insn
->code
) == BPF_ALU64
) {
3201 verbose(env
, "BPF_END uses reserved fields\n");
3206 /* check src operand */
3207 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
3211 if (is_pointer_value(env
, insn
->dst_reg
)) {
3212 verbose(env
, "R%d pointer arithmetic prohibited\n",
3217 /* check dest operand */
3218 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP
);
3222 } else if (opcode
== BPF_MOV
) {
3224 if (BPF_SRC(insn
->code
) == BPF_X
) {
3225 if (insn
->imm
!= 0 || insn
->off
!= 0) {
3226 verbose(env
, "BPF_MOV uses reserved fields\n");
3230 /* check src operand */
3231 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
3235 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0) {
3236 verbose(env
, "BPF_MOV uses reserved fields\n");
3241 /* check dest operand */
3242 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP
);
3246 if (BPF_SRC(insn
->code
) == BPF_X
) {
3247 if (BPF_CLASS(insn
->code
) == BPF_ALU64
) {
3249 * copy register state to dest reg
3251 regs
[insn
->dst_reg
] = regs
[insn
->src_reg
];
3252 regs
[insn
->dst_reg
].live
|= REG_LIVE_WRITTEN
;
3255 if (is_pointer_value(env
, insn
->src_reg
)) {
3257 "R%d partial copy of pointer\n",
3261 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
3262 coerce_reg_to_size(®s
[insn
->dst_reg
], 4);
3266 * remember the value we stored into this reg
3268 regs
[insn
->dst_reg
].type
= SCALAR_VALUE
;
3269 if (BPF_CLASS(insn
->code
) == BPF_ALU64
) {
3270 __mark_reg_known(regs
+ insn
->dst_reg
,
3273 __mark_reg_known(regs
+ insn
->dst_reg
,
3278 } else if (opcode
> BPF_END
) {
3279 verbose(env
, "invalid BPF_ALU opcode %x\n", opcode
);
3282 } else { /* all other ALU ops: and, sub, xor, add, ... */
3284 if (BPF_SRC(insn
->code
) == BPF_X
) {
3285 if (insn
->imm
!= 0 || insn
->off
!= 0) {
3286 verbose(env
, "BPF_ALU uses reserved fields\n");
3289 /* check src1 operand */
3290 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
3294 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0) {
3295 verbose(env
, "BPF_ALU uses reserved fields\n");
3300 /* check src2 operand */
3301 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
3305 if ((opcode
== BPF_MOD
|| opcode
== BPF_DIV
) &&
3306 BPF_SRC(insn
->code
) == BPF_K
&& insn
->imm
== 0) {
3307 verbose(env
, "div by zero\n");
3311 if (opcode
== BPF_ARSH
&& BPF_CLASS(insn
->code
) != BPF_ALU64
) {
3312 verbose(env
, "BPF_ARSH not supported for 32 bit ALU\n");
3316 if ((opcode
== BPF_LSH
|| opcode
== BPF_RSH
||
3317 opcode
== BPF_ARSH
) && BPF_SRC(insn
->code
) == BPF_K
) {
3318 int size
= BPF_CLASS(insn
->code
) == BPF_ALU64
? 64 : 32;
3320 if (insn
->imm
< 0 || insn
->imm
>= size
) {
3321 verbose(env
, "invalid shift %d\n", insn
->imm
);
3326 /* check dest operand */
3327 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP_NO_MARK
);
3331 return adjust_reg_min_max_vals(env
, insn
);
3337 static void find_good_pkt_pointers(struct bpf_verifier_state
*vstate
,
3338 struct bpf_reg_state
*dst_reg
,
3339 enum bpf_reg_type type
,
3340 bool range_right_open
)
3342 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
3343 struct bpf_reg_state
*regs
= state
->regs
, *reg
;
3347 if (dst_reg
->off
< 0 ||
3348 (dst_reg
->off
== 0 && range_right_open
))
3349 /* This doesn't give us any range */
3352 if (dst_reg
->umax_value
> MAX_PACKET_OFF
||
3353 dst_reg
->umax_value
+ dst_reg
->off
> MAX_PACKET_OFF
)
3354 /* Risk of overflow. For instance, ptr + (1<<63) may be less
3355 * than pkt_end, but that's because it's also less than pkt.
3359 new_range
= dst_reg
->off
;
3360 if (range_right_open
)
3363 /* Examples for register markings:
3365 * pkt_data in dst register:
3369 * if (r2 > pkt_end) goto <handle exception>
3374 * if (r2 < pkt_end) goto <access okay>
3375 * <handle exception>
3378 * r2 == dst_reg, pkt_end == src_reg
3379 * r2=pkt(id=n,off=8,r=0)
3380 * r3=pkt(id=n,off=0,r=0)
3382 * pkt_data in src register:
3386 * if (pkt_end >= r2) goto <access okay>
3387 * <handle exception>
3391 * if (pkt_end <= r2) goto <handle exception>
3395 * pkt_end == dst_reg, r2 == src_reg
3396 * r2=pkt(id=n,off=8,r=0)
3397 * r3=pkt(id=n,off=0,r=0)
3399 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
3400 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
3401 * and [r3, r3 + 8-1) respectively is safe to access depending on
3405 /* If our ids match, then we must have the same max_value. And we
3406 * don't care about the other reg's fixed offset, since if it's too big
3407 * the range won't allow anything.
3408 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
3410 for (i
= 0; i
< MAX_BPF_REG
; i
++)
3411 if (regs
[i
].type
== type
&& regs
[i
].id
== dst_reg
->id
)
3412 /* keep the maximum range already checked */
3413 regs
[i
].range
= max(regs
[i
].range
, new_range
);
3415 for (j
= 0; j
<= vstate
->curframe
; j
++) {
3416 state
= vstate
->frame
[j
];
3417 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
3418 if (state
->stack
[i
].slot_type
[0] != STACK_SPILL
)
3420 reg
= &state
->stack
[i
].spilled_ptr
;
3421 if (reg
->type
== type
&& reg
->id
== dst_reg
->id
)
3422 reg
->range
= max(reg
->range
, new_range
);
3427 /* Adjusts the register min/max values in the case that the dst_reg is the
3428 * variable register that we are working on, and src_reg is a constant or we're
3429 * simply doing a BPF_K check.
3430 * In JEQ/JNE cases we also adjust the var_off values.
3432 static void reg_set_min_max(struct bpf_reg_state
*true_reg
,
3433 struct bpf_reg_state
*false_reg
, u64 val
,
3436 /* If the dst_reg is a pointer, we can't learn anything about its
3437 * variable offset from the compare (unless src_reg were a pointer into
3438 * the same object, but we don't bother with that.
3439 * Since false_reg and true_reg have the same type by construction, we
3440 * only need to check one of them for pointerness.
3442 if (__is_pointer_value(false, false_reg
))
3447 /* If this is false then we know nothing Jon Snow, but if it is
3448 * true then we know for sure.
3450 __mark_reg_known(true_reg
, val
);
3453 /* If this is true we know nothing Jon Snow, but if it is false
3454 * we know the value for sure;
3456 __mark_reg_known(false_reg
, val
);
3459 false_reg
->umax_value
= min(false_reg
->umax_value
, val
);
3460 true_reg
->umin_value
= max(true_reg
->umin_value
, val
+ 1);
3463 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
);
3464 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
+ 1);
3467 false_reg
->umin_value
= max(false_reg
->umin_value
, val
);
3468 true_reg
->umax_value
= min(true_reg
->umax_value
, val
- 1);
3471 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
);
3472 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
- 1);
3475 false_reg
->umax_value
= min(false_reg
->umax_value
, val
- 1);
3476 true_reg
->umin_value
= max(true_reg
->umin_value
, val
);
3479 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
- 1);
3480 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
);
3483 false_reg
->umin_value
= max(false_reg
->umin_value
, val
+ 1);
3484 true_reg
->umax_value
= min(true_reg
->umax_value
, val
);
3487 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
+ 1);
3488 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
);
3494 __reg_deduce_bounds(false_reg
);
3495 __reg_deduce_bounds(true_reg
);
3496 /* We might have learned some bits from the bounds. */
3497 __reg_bound_offset(false_reg
);
3498 __reg_bound_offset(true_reg
);
3499 /* Intersecting with the old var_off might have improved our bounds
3500 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3501 * then new var_off is (0; 0x7f...fc) which improves our umax.
3503 __update_reg_bounds(false_reg
);
3504 __update_reg_bounds(true_reg
);
3507 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
3510 static void reg_set_min_max_inv(struct bpf_reg_state
*true_reg
,
3511 struct bpf_reg_state
*false_reg
, u64 val
,
3514 if (__is_pointer_value(false, false_reg
))
3519 /* If this is false then we know nothing Jon Snow, but if it is
3520 * true then we know for sure.
3522 __mark_reg_known(true_reg
, val
);
3525 /* If this is true we know nothing Jon Snow, but if it is false
3526 * we know the value for sure;
3528 __mark_reg_known(false_reg
, val
);
3531 true_reg
->umax_value
= min(true_reg
->umax_value
, val
- 1);
3532 false_reg
->umin_value
= max(false_reg
->umin_value
, val
);
3535 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
- 1);
3536 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
);
3539 true_reg
->umin_value
= max(true_reg
->umin_value
, val
+ 1);
3540 false_reg
->umax_value
= min(false_reg
->umax_value
, val
);
3543 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
+ 1);
3544 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
);
3547 true_reg
->umax_value
= min(true_reg
->umax_value
, val
);
3548 false_reg
->umin_value
= max(false_reg
->umin_value
, val
+ 1);
3551 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
);
3552 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
+ 1);
3555 true_reg
->umin_value
= max(true_reg
->umin_value
, val
);
3556 false_reg
->umax_value
= min(false_reg
->umax_value
, val
- 1);
3559 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
);
3560 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
- 1);
3566 __reg_deduce_bounds(false_reg
);
3567 __reg_deduce_bounds(true_reg
);
3568 /* We might have learned some bits from the bounds. */
3569 __reg_bound_offset(false_reg
);
3570 __reg_bound_offset(true_reg
);
3571 /* Intersecting with the old var_off might have improved our bounds
3572 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3573 * then new var_off is (0; 0x7f...fc) which improves our umax.
3575 __update_reg_bounds(false_reg
);
3576 __update_reg_bounds(true_reg
);
3579 /* Regs are known to be equal, so intersect their min/max/var_off */
3580 static void __reg_combine_min_max(struct bpf_reg_state
*src_reg
,
3581 struct bpf_reg_state
*dst_reg
)
3583 src_reg
->umin_value
= dst_reg
->umin_value
= max(src_reg
->umin_value
,
3584 dst_reg
->umin_value
);
3585 src_reg
->umax_value
= dst_reg
->umax_value
= min(src_reg
->umax_value
,
3586 dst_reg
->umax_value
);
3587 src_reg
->smin_value
= dst_reg
->smin_value
= max(src_reg
->smin_value
,
3588 dst_reg
->smin_value
);
3589 src_reg
->smax_value
= dst_reg
->smax_value
= min(src_reg
->smax_value
,
3590 dst_reg
->smax_value
);
3591 src_reg
->var_off
= dst_reg
->var_off
= tnum_intersect(src_reg
->var_off
,
3593 /* We might have learned new bounds from the var_off. */
3594 __update_reg_bounds(src_reg
);
3595 __update_reg_bounds(dst_reg
);
3596 /* We might have learned something about the sign bit. */
3597 __reg_deduce_bounds(src_reg
);
3598 __reg_deduce_bounds(dst_reg
);
3599 /* We might have learned some bits from the bounds. */
3600 __reg_bound_offset(src_reg
);
3601 __reg_bound_offset(dst_reg
);
3602 /* Intersecting with the old var_off might have improved our bounds
3603 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3604 * then new var_off is (0; 0x7f...fc) which improves our umax.
3606 __update_reg_bounds(src_reg
);
3607 __update_reg_bounds(dst_reg
);
3610 static void reg_combine_min_max(struct bpf_reg_state
*true_src
,
3611 struct bpf_reg_state
*true_dst
,
3612 struct bpf_reg_state
*false_src
,
3613 struct bpf_reg_state
*false_dst
,
3618 __reg_combine_min_max(true_src
, true_dst
);
3621 __reg_combine_min_max(false_src
, false_dst
);
3626 static void mark_map_reg(struct bpf_reg_state
*regs
, u32 regno
, u32 id
,
3629 struct bpf_reg_state
*reg
= ®s
[regno
];
3631 if (reg
->type
== PTR_TO_MAP_VALUE_OR_NULL
&& reg
->id
== id
) {
3632 /* Old offset (both fixed and variable parts) should
3633 * have been known-zero, because we don't allow pointer
3634 * arithmetic on pointers that might be NULL.
3636 if (WARN_ON_ONCE(reg
->smin_value
|| reg
->smax_value
||
3637 !tnum_equals_const(reg
->var_off
, 0) ||
3639 __mark_reg_known_zero(reg
);
3643 reg
->type
= SCALAR_VALUE
;
3644 } else if (reg
->map_ptr
->inner_map_meta
) {
3645 reg
->type
= CONST_PTR_TO_MAP
;
3646 reg
->map_ptr
= reg
->map_ptr
->inner_map_meta
;
3648 reg
->type
= PTR_TO_MAP_VALUE
;
3650 /* We don't need id from this point onwards anymore, thus we
3651 * should better reset it, so that state pruning has chances
3658 /* The logic is similar to find_good_pkt_pointers(), both could eventually
3659 * be folded together at some point.
3661 static void mark_map_regs(struct bpf_verifier_state
*vstate
, u32 regno
,
3664 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
3665 struct bpf_reg_state
*regs
= state
->regs
;
3666 u32 id
= regs
[regno
].id
;
3669 for (i
= 0; i
< MAX_BPF_REG
; i
++)
3670 mark_map_reg(regs
, i
, id
, is_null
);
3672 for (j
= 0; j
<= vstate
->curframe
; j
++) {
3673 state
= vstate
->frame
[j
];
3674 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
3675 if (state
->stack
[i
].slot_type
[0] != STACK_SPILL
)
3677 mark_map_reg(&state
->stack
[i
].spilled_ptr
, 0, id
, is_null
);
3682 static bool try_match_pkt_pointers(const struct bpf_insn
*insn
,
3683 struct bpf_reg_state
*dst_reg
,
3684 struct bpf_reg_state
*src_reg
,
3685 struct bpf_verifier_state
*this_branch
,
3686 struct bpf_verifier_state
*other_branch
)
3688 if (BPF_SRC(insn
->code
) != BPF_X
)
3691 switch (BPF_OP(insn
->code
)) {
3693 if ((dst_reg
->type
== PTR_TO_PACKET
&&
3694 src_reg
->type
== PTR_TO_PACKET_END
) ||
3695 (dst_reg
->type
== PTR_TO_PACKET_META
&&
3696 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
3697 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
3698 find_good_pkt_pointers(this_branch
, dst_reg
,
3699 dst_reg
->type
, false);
3700 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
3701 src_reg
->type
== PTR_TO_PACKET
) ||
3702 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
3703 src_reg
->type
== PTR_TO_PACKET_META
)) {
3704 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
3705 find_good_pkt_pointers(other_branch
, src_reg
,
3706 src_reg
->type
, true);
3712 if ((dst_reg
->type
== PTR_TO_PACKET
&&
3713 src_reg
->type
== PTR_TO_PACKET_END
) ||
3714 (dst_reg
->type
== PTR_TO_PACKET_META
&&
3715 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
3716 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
3717 find_good_pkt_pointers(other_branch
, dst_reg
,
3718 dst_reg
->type
, true);
3719 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
3720 src_reg
->type
== PTR_TO_PACKET
) ||
3721 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
3722 src_reg
->type
== PTR_TO_PACKET_META
)) {
3723 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
3724 find_good_pkt_pointers(this_branch
, src_reg
,
3725 src_reg
->type
, false);
3731 if ((dst_reg
->type
== PTR_TO_PACKET
&&
3732 src_reg
->type
== PTR_TO_PACKET_END
) ||
3733 (dst_reg
->type
== PTR_TO_PACKET_META
&&
3734 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
3735 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
3736 find_good_pkt_pointers(this_branch
, dst_reg
,
3737 dst_reg
->type
, true);
3738 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
3739 src_reg
->type
== PTR_TO_PACKET
) ||
3740 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
3741 src_reg
->type
== PTR_TO_PACKET_META
)) {
3742 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
3743 find_good_pkt_pointers(other_branch
, src_reg
,
3744 src_reg
->type
, false);
3750 if ((dst_reg
->type
== PTR_TO_PACKET
&&
3751 src_reg
->type
== PTR_TO_PACKET_END
) ||
3752 (dst_reg
->type
== PTR_TO_PACKET_META
&&
3753 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
3754 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
3755 find_good_pkt_pointers(other_branch
, dst_reg
,
3756 dst_reg
->type
, false);
3757 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
3758 src_reg
->type
== PTR_TO_PACKET
) ||
3759 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
3760 src_reg
->type
== PTR_TO_PACKET_META
)) {
3761 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
3762 find_good_pkt_pointers(this_branch
, src_reg
,
3763 src_reg
->type
, true);
3775 static int check_cond_jmp_op(struct bpf_verifier_env
*env
,
3776 struct bpf_insn
*insn
, int *insn_idx
)
3778 struct bpf_verifier_state
*this_branch
= env
->cur_state
;
3779 struct bpf_verifier_state
*other_branch
;
3780 struct bpf_reg_state
*regs
= this_branch
->frame
[this_branch
->curframe
]->regs
;
3781 struct bpf_reg_state
*dst_reg
, *other_branch_regs
;
3782 u8 opcode
= BPF_OP(insn
->code
);
3785 if (opcode
> BPF_JSLE
) {
3786 verbose(env
, "invalid BPF_JMP opcode %x\n", opcode
);
3790 if (BPF_SRC(insn
->code
) == BPF_X
) {
3791 if (insn
->imm
!= 0) {
3792 verbose(env
, "BPF_JMP uses reserved fields\n");
3796 /* check src1 operand */
3797 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
3801 if (is_pointer_value(env
, insn
->src_reg
)) {
3802 verbose(env
, "R%d pointer comparison prohibited\n",
3807 if (insn
->src_reg
!= BPF_REG_0
) {
3808 verbose(env
, "BPF_JMP uses reserved fields\n");
3813 /* check src2 operand */
3814 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
3818 dst_reg
= ®s
[insn
->dst_reg
];
3820 /* detect if R == 0 where R was initialized to zero earlier */
3821 if (BPF_SRC(insn
->code
) == BPF_K
&&
3822 (opcode
== BPF_JEQ
|| opcode
== BPF_JNE
) &&
3823 dst_reg
->type
== SCALAR_VALUE
&&
3824 tnum_is_const(dst_reg
->var_off
)) {
3825 if ((opcode
== BPF_JEQ
&& dst_reg
->var_off
.value
== insn
->imm
) ||
3826 (opcode
== BPF_JNE
&& dst_reg
->var_off
.value
!= insn
->imm
)) {
3827 /* if (imm == imm) goto pc+off;
3828 * only follow the goto, ignore fall-through
3830 *insn_idx
+= insn
->off
;
3833 /* if (imm != imm) goto pc+off;
3834 * only follow fall-through branch, since
3835 * that's where the program will go
3841 other_branch
= push_stack(env
, *insn_idx
+ insn
->off
+ 1, *insn_idx
);
3844 other_branch_regs
= other_branch
->frame
[other_branch
->curframe
]->regs
;
3846 /* detect if we are comparing against a constant value so we can adjust
3847 * our min/max values for our dst register.
3848 * this is only legit if both are scalars (or pointers to the same
3849 * object, I suppose, but we don't support that right now), because
3850 * otherwise the different base pointers mean the offsets aren't
3853 if (BPF_SRC(insn
->code
) == BPF_X
) {
3854 if (dst_reg
->type
== SCALAR_VALUE
&&
3855 regs
[insn
->src_reg
].type
== SCALAR_VALUE
) {
3856 if (tnum_is_const(regs
[insn
->src_reg
].var_off
))
3857 reg_set_min_max(&other_branch_regs
[insn
->dst_reg
],
3858 dst_reg
, regs
[insn
->src_reg
].var_off
.value
,
3860 else if (tnum_is_const(dst_reg
->var_off
))
3861 reg_set_min_max_inv(&other_branch_regs
[insn
->src_reg
],
3862 ®s
[insn
->src_reg
],
3863 dst_reg
->var_off
.value
, opcode
);
3864 else if (opcode
== BPF_JEQ
|| opcode
== BPF_JNE
)
3865 /* Comparing for equality, we can combine knowledge */
3866 reg_combine_min_max(&other_branch_regs
[insn
->src_reg
],
3867 &other_branch_regs
[insn
->dst_reg
],
3868 ®s
[insn
->src_reg
],
3869 ®s
[insn
->dst_reg
], opcode
);
3871 } else if (dst_reg
->type
== SCALAR_VALUE
) {
3872 reg_set_min_max(&other_branch_regs
[insn
->dst_reg
],
3873 dst_reg
, insn
->imm
, opcode
);
3876 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
3877 if (BPF_SRC(insn
->code
) == BPF_K
&&
3878 insn
->imm
== 0 && (opcode
== BPF_JEQ
|| opcode
== BPF_JNE
) &&
3879 dst_reg
->type
== PTR_TO_MAP_VALUE_OR_NULL
) {
3880 /* Mark all identical map registers in each branch as either
3881 * safe or unknown depending R == 0 or R != 0 conditional.
3883 mark_map_regs(this_branch
, insn
->dst_reg
, opcode
== BPF_JNE
);
3884 mark_map_regs(other_branch
, insn
->dst_reg
, opcode
== BPF_JEQ
);
3885 } else if (!try_match_pkt_pointers(insn
, dst_reg
, ®s
[insn
->src_reg
],
3886 this_branch
, other_branch
) &&
3887 is_pointer_value(env
, insn
->dst_reg
)) {
3888 verbose(env
, "R%d pointer comparison prohibited\n",
3893 print_verifier_state(env
, this_branch
->frame
[this_branch
->curframe
]);
3897 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
3898 static struct bpf_map
*ld_imm64_to_map_ptr(struct bpf_insn
*insn
)
3900 u64 imm64
= ((u64
) (u32
) insn
[0].imm
) | ((u64
) (u32
) insn
[1].imm
) << 32;
3902 return (struct bpf_map
*) (unsigned long) imm64
;
3905 /* verify BPF_LD_IMM64 instruction */
3906 static int check_ld_imm(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
3908 struct bpf_reg_state
*regs
= cur_regs(env
);
3911 if (BPF_SIZE(insn
->code
) != BPF_DW
) {
3912 verbose(env
, "invalid BPF_LD_IMM insn\n");
3915 if (insn
->off
!= 0) {
3916 verbose(env
, "BPF_LD_IMM64 uses reserved fields\n");
3920 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP
);
3924 if (insn
->src_reg
== 0) {
3925 u64 imm
= ((u64
)(insn
+ 1)->imm
<< 32) | (u32
)insn
->imm
;
3927 regs
[insn
->dst_reg
].type
= SCALAR_VALUE
;
3928 __mark_reg_known(®s
[insn
->dst_reg
], imm
);
3932 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
3933 BUG_ON(insn
->src_reg
!= BPF_PSEUDO_MAP_FD
);
3935 regs
[insn
->dst_reg
].type
= CONST_PTR_TO_MAP
;
3936 regs
[insn
->dst_reg
].map_ptr
= ld_imm64_to_map_ptr(insn
);
3940 static bool may_access_skb(enum bpf_prog_type type
)
3943 case BPF_PROG_TYPE_SOCKET_FILTER
:
3944 case BPF_PROG_TYPE_SCHED_CLS
:
3945 case BPF_PROG_TYPE_SCHED_ACT
:
3952 /* verify safety of LD_ABS|LD_IND instructions:
3953 * - they can only appear in the programs where ctx == skb
3954 * - since they are wrappers of function calls, they scratch R1-R5 registers,
3955 * preserve R6-R9, and store return value into R0
3958 * ctx == skb == R6 == CTX
3961 * SRC == any register
3962 * IMM == 32-bit immediate
3965 * R0 - 8/16/32-bit skb data converted to cpu endianness
3967 static int check_ld_abs(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
3969 struct bpf_reg_state
*regs
= cur_regs(env
);
3970 u8 mode
= BPF_MODE(insn
->code
);
3973 if (!may_access_skb(env
->prog
->type
)) {
3974 verbose(env
, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
3978 if (!env
->ops
->gen_ld_abs
) {
3979 verbose(env
, "bpf verifier is misconfigured\n");
3983 if (env
->subprog_cnt
> 1) {
3984 /* when program has LD_ABS insn JITs and interpreter assume
3985 * that r1 == ctx == skb which is not the case for callees
3986 * that can have arbitrary arguments. It's problematic
3987 * for main prog as well since JITs would need to analyze
3988 * all functions in order to make proper register save/restore
3989 * decisions in the main prog. Hence disallow LD_ABS with calls
3991 verbose(env
, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
3995 if (insn
->dst_reg
!= BPF_REG_0
|| insn
->off
!= 0 ||
3996 BPF_SIZE(insn
->code
) == BPF_DW
||
3997 (mode
== BPF_ABS
&& insn
->src_reg
!= BPF_REG_0
)) {
3998 verbose(env
, "BPF_LD_[ABS|IND] uses reserved fields\n");
4002 /* check whether implicit source operand (register R6) is readable */
4003 err
= check_reg_arg(env
, BPF_REG_6
, SRC_OP
);
4007 if (regs
[BPF_REG_6
].type
!= PTR_TO_CTX
) {
4009 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
4013 if (mode
== BPF_IND
) {
4014 /* check explicit source operand */
4015 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
4020 /* reset caller saved regs to unreadable */
4021 for (i
= 0; i
< CALLER_SAVED_REGS
; i
++) {
4022 mark_reg_not_init(env
, regs
, caller_saved
[i
]);
4023 check_reg_arg(env
, caller_saved
[i
], DST_OP_NO_MARK
);
4026 /* mark destination R0 register as readable, since it contains
4027 * the value fetched from the packet.
4028 * Already marked as written above.
4030 mark_reg_unknown(env
, regs
, BPF_REG_0
);
4034 static int check_return_code(struct bpf_verifier_env
*env
)
4036 struct bpf_reg_state
*reg
;
4037 struct tnum range
= tnum_range(0, 1);
4039 switch (env
->prog
->type
) {
4040 case BPF_PROG_TYPE_CGROUP_SKB
:
4041 case BPF_PROG_TYPE_CGROUP_SOCK
:
4042 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR
:
4043 case BPF_PROG_TYPE_SOCK_OPS
:
4044 case BPF_PROG_TYPE_CGROUP_DEVICE
:
4050 reg
= cur_regs(env
) + BPF_REG_0
;
4051 if (reg
->type
!= SCALAR_VALUE
) {
4052 verbose(env
, "At program exit the register R0 is not a known value (%s)\n",
4053 reg_type_str
[reg
->type
]);
4057 if (!tnum_in(range
, reg
->var_off
)) {
4058 verbose(env
, "At program exit the register R0 ");
4059 if (!tnum_is_unknown(reg
->var_off
)) {
4062 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
4063 verbose(env
, "has value %s", tn_buf
);
4065 verbose(env
, "has unknown scalar value");
4067 verbose(env
, " should have been 0 or 1\n");
4073 /* non-recursive DFS pseudo code
4074 * 1 procedure DFS-iterative(G,v):
4075 * 2 label v as discovered
4076 * 3 let S be a stack
4078 * 5 while S is not empty
4080 * 7 if t is what we're looking for:
4082 * 9 for all edges e in G.adjacentEdges(t) do
4083 * 10 if edge e is already labelled
4084 * 11 continue with the next edge
4085 * 12 w <- G.adjacentVertex(t,e)
4086 * 13 if vertex w is not discovered and not explored
4087 * 14 label e as tree-edge
4088 * 15 label w as discovered
4091 * 18 else if vertex w is discovered
4092 * 19 label e as back-edge
4094 * 21 // vertex w is explored
4095 * 22 label e as forward- or cross-edge
4096 * 23 label t as explored
4101 * 0x11 - discovered and fall-through edge labelled
4102 * 0x12 - discovered and fall-through and branch edges labelled
4113 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
4115 static int *insn_stack
; /* stack of insns to process */
4116 static int cur_stack
; /* current stack index */
4117 static int *insn_state
;
4119 /* t, w, e - match pseudo-code above:
4120 * t - index of current instruction
4121 * w - next instruction
4124 static int push_insn(int t
, int w
, int e
, struct bpf_verifier_env
*env
)
4126 if (e
== FALLTHROUGH
&& insn_state
[t
] >= (DISCOVERED
| FALLTHROUGH
))
4129 if (e
== BRANCH
&& insn_state
[t
] >= (DISCOVERED
| BRANCH
))
4132 if (w
< 0 || w
>= env
->prog
->len
) {
4133 verbose(env
, "jump out of range from insn %d to %d\n", t
, w
);
4138 /* mark branch target for state pruning */
4139 env
->explored_states
[w
] = STATE_LIST_MARK
;
4141 if (insn_state
[w
] == 0) {
4143 insn_state
[t
] = DISCOVERED
| e
;
4144 insn_state
[w
] = DISCOVERED
;
4145 if (cur_stack
>= env
->prog
->len
)
4147 insn_stack
[cur_stack
++] = w
;
4149 } else if ((insn_state
[w
] & 0xF0) == DISCOVERED
) {
4150 verbose(env
, "back-edge from insn %d to %d\n", t
, w
);
4152 } else if (insn_state
[w
] == EXPLORED
) {
4153 /* forward- or cross-edge */
4154 insn_state
[t
] = DISCOVERED
| e
;
4156 verbose(env
, "insn state internal bug\n");
4162 /* non-recursive depth-first-search to detect loops in BPF program
4163 * loop == back-edge in directed graph
4165 static int check_cfg(struct bpf_verifier_env
*env
)
4167 struct bpf_insn
*insns
= env
->prog
->insnsi
;
4168 int insn_cnt
= env
->prog
->len
;
4172 ret
= check_subprogs(env
);
4176 insn_state
= kcalloc(insn_cnt
, sizeof(int), GFP_KERNEL
);
4180 insn_stack
= kcalloc(insn_cnt
, sizeof(int), GFP_KERNEL
);
4186 insn_state
[0] = DISCOVERED
; /* mark 1st insn as discovered */
4187 insn_stack
[0] = 0; /* 0 is the first instruction */
4193 t
= insn_stack
[cur_stack
- 1];
4195 if (BPF_CLASS(insns
[t
].code
) == BPF_JMP
) {
4196 u8 opcode
= BPF_OP(insns
[t
].code
);
4198 if (opcode
== BPF_EXIT
) {
4200 } else if (opcode
== BPF_CALL
) {
4201 ret
= push_insn(t
, t
+ 1, FALLTHROUGH
, env
);
4206 if (t
+ 1 < insn_cnt
)
4207 env
->explored_states
[t
+ 1] = STATE_LIST_MARK
;
4208 if (insns
[t
].src_reg
== BPF_PSEUDO_CALL
) {
4209 env
->explored_states
[t
] = STATE_LIST_MARK
;
4210 ret
= push_insn(t
, t
+ insns
[t
].imm
+ 1, BRANCH
, env
);
4216 } else if (opcode
== BPF_JA
) {
4217 if (BPF_SRC(insns
[t
].code
) != BPF_K
) {
4221 /* unconditional jump with single edge */
4222 ret
= push_insn(t
, t
+ insns
[t
].off
+ 1,
4228 /* tell verifier to check for equivalent states
4229 * after every call and jump
4231 if (t
+ 1 < insn_cnt
)
4232 env
->explored_states
[t
+ 1] = STATE_LIST_MARK
;
4234 /* conditional jump with two edges */
4235 env
->explored_states
[t
] = STATE_LIST_MARK
;
4236 ret
= push_insn(t
, t
+ 1, FALLTHROUGH
, env
);
4242 ret
= push_insn(t
, t
+ insns
[t
].off
+ 1, BRANCH
, env
);
4249 /* all other non-branch instructions with single
4252 ret
= push_insn(t
, t
+ 1, FALLTHROUGH
, env
);
4260 insn_state
[t
] = EXPLORED
;
4261 if (cur_stack
-- <= 0) {
4262 verbose(env
, "pop stack internal bug\n");
4269 for (i
= 0; i
< insn_cnt
; i
++) {
4270 if (insn_state
[i
] != EXPLORED
) {
4271 verbose(env
, "unreachable insn %d\n", i
);
4276 ret
= 0; /* cfg looks good */
4284 /* check %cur's range satisfies %old's */
4285 static bool range_within(struct bpf_reg_state
*old
,
4286 struct bpf_reg_state
*cur
)
4288 return old
->umin_value
<= cur
->umin_value
&&
4289 old
->umax_value
>= cur
->umax_value
&&
4290 old
->smin_value
<= cur
->smin_value
&&
4291 old
->smax_value
>= cur
->smax_value
;
4294 /* Maximum number of register states that can exist at once */
4295 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
4301 /* If in the old state two registers had the same id, then they need to have
4302 * the same id in the new state as well. But that id could be different from
4303 * the old state, so we need to track the mapping from old to new ids.
4304 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
4305 * regs with old id 5 must also have new id 9 for the new state to be safe. But
4306 * regs with a different old id could still have new id 9, we don't care about
4308 * So we look through our idmap to see if this old id has been seen before. If
4309 * so, we require the new id to match; otherwise, we add the id pair to the map.
4311 static bool check_ids(u32 old_id
, u32 cur_id
, struct idpair
*idmap
)
4315 for (i
= 0; i
< ID_MAP_SIZE
; i
++) {
4316 if (!idmap
[i
].old
) {
4317 /* Reached an empty slot; haven't seen this id before */
4318 idmap
[i
].old
= old_id
;
4319 idmap
[i
].cur
= cur_id
;
4322 if (idmap
[i
].old
== old_id
)
4323 return idmap
[i
].cur
== cur_id
;
4325 /* We ran out of idmap slots, which should be impossible */
4330 /* Returns true if (rold safe implies rcur safe) */
4331 static bool regsafe(struct bpf_reg_state
*rold
, struct bpf_reg_state
*rcur
,
4332 struct idpair
*idmap
)
4336 if (!(rold
->live
& REG_LIVE_READ
))
4337 /* explored state didn't use this */
4340 equal
= memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, frameno
)) == 0;
4342 if (rold
->type
== PTR_TO_STACK
)
4343 /* two stack pointers are equal only if they're pointing to
4344 * the same stack frame, since fp-8 in foo != fp-8 in bar
4346 return equal
&& rold
->frameno
== rcur
->frameno
;
4351 if (rold
->type
== NOT_INIT
)
4352 /* explored state can't have used this */
4354 if (rcur
->type
== NOT_INIT
)
4356 switch (rold
->type
) {
4358 if (rcur
->type
== SCALAR_VALUE
) {
4359 /* new val must satisfy old val knowledge */
4360 return range_within(rold
, rcur
) &&
4361 tnum_in(rold
->var_off
, rcur
->var_off
);
4363 /* We're trying to use a pointer in place of a scalar.
4364 * Even if the scalar was unbounded, this could lead to
4365 * pointer leaks because scalars are allowed to leak
4366 * while pointers are not. We could make this safe in
4367 * special cases if root is calling us, but it's
4368 * probably not worth the hassle.
4372 case PTR_TO_MAP_VALUE
:
4373 /* If the new min/max/var_off satisfy the old ones and
4374 * everything else matches, we are OK.
4375 * We don't care about the 'id' value, because nothing
4376 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
4378 return memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, id
)) == 0 &&
4379 range_within(rold
, rcur
) &&
4380 tnum_in(rold
->var_off
, rcur
->var_off
);
4381 case PTR_TO_MAP_VALUE_OR_NULL
:
4382 /* a PTR_TO_MAP_VALUE could be safe to use as a
4383 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
4384 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
4385 * checked, doing so could have affected others with the same
4386 * id, and we can't check for that because we lost the id when
4387 * we converted to a PTR_TO_MAP_VALUE.
4389 if (rcur
->type
!= PTR_TO_MAP_VALUE_OR_NULL
)
4391 if (memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, id
)))
4393 /* Check our ids match any regs they're supposed to */
4394 return check_ids(rold
->id
, rcur
->id
, idmap
);
4395 case PTR_TO_PACKET_META
:
4397 if (rcur
->type
!= rold
->type
)
4399 /* We must have at least as much range as the old ptr
4400 * did, so that any accesses which were safe before are
4401 * still safe. This is true even if old range < old off,
4402 * since someone could have accessed through (ptr - k), or
4403 * even done ptr -= k in a register, to get a safe access.
4405 if (rold
->range
> rcur
->range
)
4407 /* If the offsets don't match, we can't trust our alignment;
4408 * nor can we be sure that we won't fall out of range.
4410 if (rold
->off
!= rcur
->off
)
4412 /* id relations must be preserved */
4413 if (rold
->id
&& !check_ids(rold
->id
, rcur
->id
, idmap
))
4415 /* new val must satisfy old val knowledge */
4416 return range_within(rold
, rcur
) &&
4417 tnum_in(rold
->var_off
, rcur
->var_off
);
4419 case CONST_PTR_TO_MAP
:
4420 case PTR_TO_PACKET_END
:
4421 /* Only valid matches are exact, which memcmp() above
4422 * would have accepted
4425 /* Don't know what's going on, just say it's not safe */
4429 /* Shouldn't get here; if we do, say it's not safe */
4434 static bool stacksafe(struct bpf_func_state
*old
,
4435 struct bpf_func_state
*cur
,
4436 struct idpair
*idmap
)
4440 /* if explored stack has more populated slots than current stack
4441 * such stacks are not equivalent
4443 if (old
->allocated_stack
> cur
->allocated_stack
)
4446 /* walk slots of the explored stack and ignore any additional
4447 * slots in the current stack, since explored(safe) state
4450 for (i
= 0; i
< old
->allocated_stack
; i
++) {
4451 spi
= i
/ BPF_REG_SIZE
;
4453 if (!(old
->stack
[spi
].spilled_ptr
.live
& REG_LIVE_READ
))
4454 /* explored state didn't use this */
4457 if (old
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
] == STACK_INVALID
)
4459 /* if old state was safe with misc data in the stack
4460 * it will be safe with zero-initialized stack.
4461 * The opposite is not true
4463 if (old
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
] == STACK_MISC
&&
4464 cur
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
] == STACK_ZERO
)
4466 if (old
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
] !=
4467 cur
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
])
4468 /* Ex: old explored (safe) state has STACK_SPILL in
4469 * this stack slot, but current has has STACK_MISC ->
4470 * this verifier states are not equivalent,
4471 * return false to continue verification of this path
4474 if (i
% BPF_REG_SIZE
)
4476 if (old
->stack
[spi
].slot_type
[0] != STACK_SPILL
)
4478 if (!regsafe(&old
->stack
[spi
].spilled_ptr
,
4479 &cur
->stack
[spi
].spilled_ptr
,
4481 /* when explored and current stack slot are both storing
4482 * spilled registers, check that stored pointers types
4483 * are the same as well.
4484 * Ex: explored safe path could have stored
4485 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
4486 * but current path has stored:
4487 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
4488 * such verifier states are not equivalent.
4489 * return false to continue verification of this path
4496 /* compare two verifier states
4498 * all states stored in state_list are known to be valid, since
4499 * verifier reached 'bpf_exit' instruction through them
4501 * this function is called when verifier exploring different branches of
4502 * execution popped from the state stack. If it sees an old state that has
4503 * more strict register state and more strict stack state then this execution
4504 * branch doesn't need to be explored further, since verifier already
4505 * concluded that more strict state leads to valid finish.
4507 * Therefore two states are equivalent if register state is more conservative
4508 * and explored stack state is more conservative than the current one.
4511 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
4512 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
4514 * In other words if current stack state (one being explored) has more
4515 * valid slots than old one that already passed validation, it means
4516 * the verifier can stop exploring and conclude that current state is valid too
4518 * Similarly with registers. If explored state has register type as invalid
4519 * whereas register type in current state is meaningful, it means that
4520 * the current state will reach 'bpf_exit' instruction safely
4522 static bool func_states_equal(struct bpf_func_state
*old
,
4523 struct bpf_func_state
*cur
)
4525 struct idpair
*idmap
;
4529 idmap
= kcalloc(ID_MAP_SIZE
, sizeof(struct idpair
), GFP_KERNEL
);
4530 /* If we failed to allocate the idmap, just say it's not safe */
4534 for (i
= 0; i
< MAX_BPF_REG
; i
++) {
4535 if (!regsafe(&old
->regs
[i
], &cur
->regs
[i
], idmap
))
4539 if (!stacksafe(old
, cur
, idmap
))
4547 static bool states_equal(struct bpf_verifier_env
*env
,
4548 struct bpf_verifier_state
*old
,
4549 struct bpf_verifier_state
*cur
)
4553 if (old
->curframe
!= cur
->curframe
)
4556 /* for states to be equal callsites have to be the same
4557 * and all frame states need to be equivalent
4559 for (i
= 0; i
<= old
->curframe
; i
++) {
4560 if (old
->frame
[i
]->callsite
!= cur
->frame
[i
]->callsite
)
4562 if (!func_states_equal(old
->frame
[i
], cur
->frame
[i
]))
4568 /* A write screens off any subsequent reads; but write marks come from the
4569 * straight-line code between a state and its parent. When we arrive at an
4570 * equivalent state (jump target or such) we didn't arrive by the straight-line
4571 * code, so read marks in the state must propagate to the parent regardless
4572 * of the state's write marks. That's what 'parent == state->parent' comparison
4573 * in mark_reg_read() and mark_stack_slot_read() is for.
4575 static int propagate_liveness(struct bpf_verifier_env
*env
,
4576 const struct bpf_verifier_state
*vstate
,
4577 struct bpf_verifier_state
*vparent
)
4579 int i
, frame
, err
= 0;
4580 struct bpf_func_state
*state
, *parent
;
4582 if (vparent
->curframe
!= vstate
->curframe
) {
4583 WARN(1, "propagate_live: parent frame %d current frame %d\n",
4584 vparent
->curframe
, vstate
->curframe
);
4587 /* Propagate read liveness of registers... */
4588 BUILD_BUG_ON(BPF_REG_FP
+ 1 != MAX_BPF_REG
);
4589 /* We don't need to worry about FP liveness because it's read-only */
4590 for (i
= 0; i
< BPF_REG_FP
; i
++) {
4591 if (vparent
->frame
[vparent
->curframe
]->regs
[i
].live
& REG_LIVE_READ
)
4593 if (vstate
->frame
[vstate
->curframe
]->regs
[i
].live
& REG_LIVE_READ
) {
4594 err
= mark_reg_read(env
, vstate
, vparent
, i
);
4600 /* ... and stack slots */
4601 for (frame
= 0; frame
<= vstate
->curframe
; frame
++) {
4602 state
= vstate
->frame
[frame
];
4603 parent
= vparent
->frame
[frame
];
4604 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
&&
4605 i
< parent
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
4606 if (parent
->stack
[i
].spilled_ptr
.live
& REG_LIVE_READ
)
4608 if (state
->stack
[i
].spilled_ptr
.live
& REG_LIVE_READ
)
4609 mark_stack_slot_read(env
, vstate
, vparent
, i
, frame
);
4615 static int is_state_visited(struct bpf_verifier_env
*env
, int insn_idx
)
4617 struct bpf_verifier_state_list
*new_sl
;
4618 struct bpf_verifier_state_list
*sl
;
4619 struct bpf_verifier_state
*cur
= env
->cur_state
;
4622 sl
= env
->explored_states
[insn_idx
];
4624 /* this 'insn_idx' instruction wasn't marked, so we will not
4625 * be doing state search here
4629 while (sl
!= STATE_LIST_MARK
) {
4630 if (states_equal(env
, &sl
->state
, cur
)) {
4631 /* reached equivalent register/stack state,
4633 * Registers read by the continuation are read by us.
4634 * If we have any write marks in env->cur_state, they
4635 * will prevent corresponding reads in the continuation
4636 * from reaching our parent (an explored_state). Our
4637 * own state will get the read marks recorded, but
4638 * they'll be immediately forgotten as we're pruning
4639 * this state and will pop a new one.
4641 err
= propagate_liveness(env
, &sl
->state
, cur
);
4649 /* there were no equivalent states, remember current one.
4650 * technically the current state is not proven to be safe yet,
4651 * but it will either reach outer most bpf_exit (which means it's safe)
4652 * or it will be rejected. Since there are no loops, we won't be
4653 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
4654 * again on the way to bpf_exit
4656 new_sl
= kzalloc(sizeof(struct bpf_verifier_state_list
), GFP_KERNEL
);
4660 /* add new state to the head of linked list */
4661 err
= copy_verifier_state(&new_sl
->state
, cur
);
4663 free_verifier_state(&new_sl
->state
, false);
4667 new_sl
->next
= env
->explored_states
[insn_idx
];
4668 env
->explored_states
[insn_idx
] = new_sl
;
4669 /* connect new state to parentage chain */
4670 cur
->parent
= &new_sl
->state
;
4671 /* clear write marks in current state: the writes we did are not writes
4672 * our child did, so they don't screen off its reads from us.
4673 * (There are no read marks in current state, because reads always mark
4674 * their parent and current state never has children yet. Only
4675 * explored_states can get read marks.)
4677 for (i
= 0; i
< BPF_REG_FP
; i
++)
4678 cur
->frame
[cur
->curframe
]->regs
[i
].live
= REG_LIVE_NONE
;
4680 /* all stack frames are accessible from callee, clear them all */
4681 for (j
= 0; j
<= cur
->curframe
; j
++) {
4682 struct bpf_func_state
*frame
= cur
->frame
[j
];
4684 for (i
= 0; i
< frame
->allocated_stack
/ BPF_REG_SIZE
; i
++)
4685 frame
->stack
[i
].spilled_ptr
.live
= REG_LIVE_NONE
;
4690 static int do_check(struct bpf_verifier_env
*env
)
4692 struct bpf_verifier_state
*state
;
4693 struct bpf_insn
*insns
= env
->prog
->insnsi
;
4694 struct bpf_reg_state
*regs
;
4695 int insn_cnt
= env
->prog
->len
, i
;
4696 int insn_idx
, prev_insn_idx
= 0;
4697 int insn_processed
= 0;
4698 bool do_print_state
= false;
4700 state
= kzalloc(sizeof(struct bpf_verifier_state
), GFP_KERNEL
);
4703 state
->curframe
= 0;
4704 state
->parent
= NULL
;
4705 state
->frame
[0] = kzalloc(sizeof(struct bpf_func_state
), GFP_KERNEL
);
4706 if (!state
->frame
[0]) {
4710 env
->cur_state
= state
;
4711 init_func_state(env
, state
->frame
[0],
4712 BPF_MAIN_FUNC
/* callsite */,
4714 0 /* subprogno, zero == main subprog */);
4717 struct bpf_insn
*insn
;
4721 if (insn_idx
>= insn_cnt
) {
4722 verbose(env
, "invalid insn idx %d insn_cnt %d\n",
4723 insn_idx
, insn_cnt
);
4727 insn
= &insns
[insn_idx
];
4728 class = BPF_CLASS(insn
->code
);
4730 if (++insn_processed
> BPF_COMPLEXITY_LIMIT_INSNS
) {
4732 "BPF program is too large. Processed %d insn\n",
4737 err
= is_state_visited(env
, insn_idx
);
4741 /* found equivalent state, can prune the search */
4742 if (env
->log
.level
) {
4744 verbose(env
, "\nfrom %d to %d: safe\n",
4745 prev_insn_idx
, insn_idx
);
4747 verbose(env
, "%d: safe\n", insn_idx
);
4749 goto process_bpf_exit
;
4755 if (env
->log
.level
> 1 || (env
->log
.level
&& do_print_state
)) {
4756 if (env
->log
.level
> 1)
4757 verbose(env
, "%d:", insn_idx
);
4759 verbose(env
, "\nfrom %d to %d:",
4760 prev_insn_idx
, insn_idx
);
4761 print_verifier_state(env
, state
->frame
[state
->curframe
]);
4762 do_print_state
= false;
4765 if (env
->log
.level
) {
4766 const struct bpf_insn_cbs cbs
= {
4767 .cb_print
= verbose
,
4768 .private_data
= env
,
4771 verbose(env
, "%d: ", insn_idx
);
4772 print_bpf_insn(&cbs
, insn
, env
->allow_ptr_leaks
);
4775 if (bpf_prog_is_dev_bound(env
->prog
->aux
)) {
4776 err
= bpf_prog_offload_verify_insn(env
, insn_idx
,
4782 regs
= cur_regs(env
);
4783 env
->insn_aux_data
[insn_idx
].seen
= true;
4784 if (class == BPF_ALU
|| class == BPF_ALU64
) {
4785 err
= check_alu_op(env
, insn
);
4789 } else if (class == BPF_LDX
) {
4790 enum bpf_reg_type
*prev_src_type
, src_reg_type
;
4792 /* check for reserved fields is already done */
4794 /* check src operand */
4795 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
4799 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP_NO_MARK
);
4803 src_reg_type
= regs
[insn
->src_reg
].type
;
4805 /* check that memory (src_reg + off) is readable,
4806 * the state of dst_reg will be updated by this func
4808 err
= check_mem_access(env
, insn_idx
, insn
->src_reg
, insn
->off
,
4809 BPF_SIZE(insn
->code
), BPF_READ
,
4810 insn
->dst_reg
, false);
4814 prev_src_type
= &env
->insn_aux_data
[insn_idx
].ptr_type
;
4816 if (*prev_src_type
== NOT_INIT
) {
4818 * dst_reg = *(u32 *)(src_reg + off)
4819 * save type to validate intersecting paths
4821 *prev_src_type
= src_reg_type
;
4823 } else if (src_reg_type
!= *prev_src_type
&&
4824 (src_reg_type
== PTR_TO_CTX
||
4825 *prev_src_type
== PTR_TO_CTX
)) {
4826 /* ABuser program is trying to use the same insn
4827 * dst_reg = *(u32*) (src_reg + off)
4828 * with different pointer types:
4829 * src_reg == ctx in one branch and
4830 * src_reg == stack|map in some other branch.
4833 verbose(env
, "same insn cannot be used with different pointers\n");
4837 } else if (class == BPF_STX
) {
4838 enum bpf_reg_type
*prev_dst_type
, dst_reg_type
;
4840 if (BPF_MODE(insn
->code
) == BPF_XADD
) {
4841 err
= check_xadd(env
, insn_idx
, insn
);
4848 /* check src1 operand */
4849 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
4852 /* check src2 operand */
4853 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
4857 dst_reg_type
= regs
[insn
->dst_reg
].type
;
4859 /* check that memory (dst_reg + off) is writeable */
4860 err
= check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
4861 BPF_SIZE(insn
->code
), BPF_WRITE
,
4862 insn
->src_reg
, false);
4866 prev_dst_type
= &env
->insn_aux_data
[insn_idx
].ptr_type
;
4868 if (*prev_dst_type
== NOT_INIT
) {
4869 *prev_dst_type
= dst_reg_type
;
4870 } else if (dst_reg_type
!= *prev_dst_type
&&
4871 (dst_reg_type
== PTR_TO_CTX
||
4872 *prev_dst_type
== PTR_TO_CTX
)) {
4873 verbose(env
, "same insn cannot be used with different pointers\n");
4877 } else if (class == BPF_ST
) {
4878 if (BPF_MODE(insn
->code
) != BPF_MEM
||
4879 insn
->src_reg
!= BPF_REG_0
) {
4880 verbose(env
, "BPF_ST uses reserved fields\n");
4883 /* check src operand */
4884 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
4888 if (is_ctx_reg(env
, insn
->dst_reg
)) {
4889 verbose(env
, "BPF_ST stores into R%d context is not allowed\n",
4894 /* check that memory (dst_reg + off) is writeable */
4895 err
= check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
4896 BPF_SIZE(insn
->code
), BPF_WRITE
,
4901 } else if (class == BPF_JMP
) {
4902 u8 opcode
= BPF_OP(insn
->code
);
4904 if (opcode
== BPF_CALL
) {
4905 if (BPF_SRC(insn
->code
) != BPF_K
||
4907 (insn
->src_reg
!= BPF_REG_0
&&
4908 insn
->src_reg
!= BPF_PSEUDO_CALL
) ||
4909 insn
->dst_reg
!= BPF_REG_0
) {
4910 verbose(env
, "BPF_CALL uses reserved fields\n");
4914 if (insn
->src_reg
== BPF_PSEUDO_CALL
)
4915 err
= check_func_call(env
, insn
, &insn_idx
);
4917 err
= check_helper_call(env
, insn
->imm
, insn_idx
);
4921 } else if (opcode
== BPF_JA
) {
4922 if (BPF_SRC(insn
->code
) != BPF_K
||
4924 insn
->src_reg
!= BPF_REG_0
||
4925 insn
->dst_reg
!= BPF_REG_0
) {
4926 verbose(env
, "BPF_JA uses reserved fields\n");
4930 insn_idx
+= insn
->off
+ 1;
4933 } else if (opcode
== BPF_EXIT
) {
4934 if (BPF_SRC(insn
->code
) != BPF_K
||
4936 insn
->src_reg
!= BPF_REG_0
||
4937 insn
->dst_reg
!= BPF_REG_0
) {
4938 verbose(env
, "BPF_EXIT uses reserved fields\n");
4942 if (state
->curframe
) {
4943 /* exit from nested function */
4944 prev_insn_idx
= insn_idx
;
4945 err
= prepare_func_exit(env
, &insn_idx
);
4948 do_print_state
= true;
4952 /* eBPF calling convetion is such that R0 is used
4953 * to return the value from eBPF program.
4954 * Make sure that it's readable at this time
4955 * of bpf_exit, which means that program wrote
4956 * something into it earlier
4958 err
= check_reg_arg(env
, BPF_REG_0
, SRC_OP
);
4962 if (is_pointer_value(env
, BPF_REG_0
)) {
4963 verbose(env
, "R0 leaks addr as return value\n");
4967 err
= check_return_code(env
);
4971 err
= pop_stack(env
, &prev_insn_idx
, &insn_idx
);
4977 do_print_state
= true;
4981 err
= check_cond_jmp_op(env
, insn
, &insn_idx
);
4985 } else if (class == BPF_LD
) {
4986 u8 mode
= BPF_MODE(insn
->code
);
4988 if (mode
== BPF_ABS
|| mode
== BPF_IND
) {
4989 err
= check_ld_abs(env
, insn
);
4993 } else if (mode
== BPF_IMM
) {
4994 err
= check_ld_imm(env
, insn
);
4999 env
->insn_aux_data
[insn_idx
].seen
= true;
5001 verbose(env
, "invalid BPF_LD mode\n");
5005 verbose(env
, "unknown insn class %d\n", class);
5012 verbose(env
, "processed %d insns (limit %d), stack depth ",
5013 insn_processed
, BPF_COMPLEXITY_LIMIT_INSNS
);
5014 for (i
= 0; i
< env
->subprog_cnt
; i
++) {
5015 u32 depth
= env
->subprog_info
[i
].stack_depth
;
5017 verbose(env
, "%d", depth
);
5018 if (i
+ 1 < env
->subprog_cnt
)
5022 env
->prog
->aux
->stack_depth
= env
->subprog_info
[0].stack_depth
;
5026 static int check_map_prealloc(struct bpf_map
*map
)
5028 return (map
->map_type
!= BPF_MAP_TYPE_HASH
&&
5029 map
->map_type
!= BPF_MAP_TYPE_PERCPU_HASH
&&
5030 map
->map_type
!= BPF_MAP_TYPE_HASH_OF_MAPS
) ||
5031 !(map
->map_flags
& BPF_F_NO_PREALLOC
);
5034 static int check_map_prog_compatibility(struct bpf_verifier_env
*env
,
5035 struct bpf_map
*map
,
5036 struct bpf_prog
*prog
)
5039 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
5040 * preallocated hash maps, since doing memory allocation
5041 * in overflow_handler can crash depending on where nmi got
5044 if (prog
->type
== BPF_PROG_TYPE_PERF_EVENT
) {
5045 if (!check_map_prealloc(map
)) {
5046 verbose(env
, "perf_event programs can only use preallocated hash map\n");
5049 if (map
->inner_map_meta
&&
5050 !check_map_prealloc(map
->inner_map_meta
)) {
5051 verbose(env
, "perf_event programs can only use preallocated inner hash map\n");
5056 if ((bpf_prog_is_dev_bound(prog
->aux
) || bpf_map_is_dev_bound(map
)) &&
5057 !bpf_offload_dev_match(prog
, map
)) {
5058 verbose(env
, "offload device mismatch between prog and map\n");
5065 /* look for pseudo eBPF instructions that access map FDs and
5066 * replace them with actual map pointers
5068 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env
*env
)
5070 struct bpf_insn
*insn
= env
->prog
->insnsi
;
5071 int insn_cnt
= env
->prog
->len
;
5074 err
= bpf_prog_calc_tag(env
->prog
);
5078 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
5079 if (BPF_CLASS(insn
->code
) == BPF_LDX
&&
5080 (BPF_MODE(insn
->code
) != BPF_MEM
|| insn
->imm
!= 0)) {
5081 verbose(env
, "BPF_LDX uses reserved fields\n");
5085 if (BPF_CLASS(insn
->code
) == BPF_STX
&&
5086 ((BPF_MODE(insn
->code
) != BPF_MEM
&&
5087 BPF_MODE(insn
->code
) != BPF_XADD
) || insn
->imm
!= 0)) {
5088 verbose(env
, "BPF_STX uses reserved fields\n");
5092 if (insn
[0].code
== (BPF_LD
| BPF_IMM
| BPF_DW
)) {
5093 struct bpf_map
*map
;
5096 if (i
== insn_cnt
- 1 || insn
[1].code
!= 0 ||
5097 insn
[1].dst_reg
!= 0 || insn
[1].src_reg
!= 0 ||
5099 verbose(env
, "invalid bpf_ld_imm64 insn\n");
5103 if (insn
->src_reg
== 0)
5104 /* valid generic load 64-bit imm */
5107 if (insn
->src_reg
!= BPF_PSEUDO_MAP_FD
) {
5109 "unrecognized bpf_ld_imm64 insn\n");
5113 f
= fdget(insn
->imm
);
5114 map
= __bpf_map_get(f
);
5116 verbose(env
, "fd %d is not pointing to valid bpf_map\n",
5118 return PTR_ERR(map
);
5121 err
= check_map_prog_compatibility(env
, map
, env
->prog
);
5127 /* store map pointer inside BPF_LD_IMM64 instruction */
5128 insn
[0].imm
= (u32
) (unsigned long) map
;
5129 insn
[1].imm
= ((u64
) (unsigned long) map
) >> 32;
5131 /* check whether we recorded this map already */
5132 for (j
= 0; j
< env
->used_map_cnt
; j
++)
5133 if (env
->used_maps
[j
] == map
) {
5138 if (env
->used_map_cnt
>= MAX_USED_MAPS
) {
5143 /* hold the map. If the program is rejected by verifier,
5144 * the map will be released by release_maps() or it
5145 * will be used by the valid program until it's unloaded
5146 * and all maps are released in free_used_maps()
5148 map
= bpf_map_inc(map
, false);
5151 return PTR_ERR(map
);
5153 env
->used_maps
[env
->used_map_cnt
++] = map
;
5162 /* Basic sanity check before we invest more work here. */
5163 if (!bpf_opcode_in_insntable(insn
->code
)) {
5164 verbose(env
, "unknown opcode %02x\n", insn
->code
);
5169 /* now all pseudo BPF_LD_IMM64 instructions load valid
5170 * 'struct bpf_map *' into a register instead of user map_fd.
5171 * These pointers will be used later by verifier to validate map access.
5176 /* drop refcnt of maps used by the rejected program */
5177 static void release_maps(struct bpf_verifier_env
*env
)
5181 for (i
= 0; i
< env
->used_map_cnt
; i
++)
5182 bpf_map_put(env
->used_maps
[i
]);
5185 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
5186 static void convert_pseudo_ld_imm64(struct bpf_verifier_env
*env
)
5188 struct bpf_insn
*insn
= env
->prog
->insnsi
;
5189 int insn_cnt
= env
->prog
->len
;
5192 for (i
= 0; i
< insn_cnt
; i
++, insn
++)
5193 if (insn
->code
== (BPF_LD
| BPF_IMM
| BPF_DW
))
5197 /* single env->prog->insni[off] instruction was replaced with the range
5198 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
5199 * [0, off) and [off, end) to new locations, so the patched range stays zero
5201 static int adjust_insn_aux_data(struct bpf_verifier_env
*env
, u32 prog_len
,
5204 struct bpf_insn_aux_data
*new_data
, *old_data
= env
->insn_aux_data
;
5209 new_data
= vzalloc(sizeof(struct bpf_insn_aux_data
) * prog_len
);
5212 memcpy(new_data
, old_data
, sizeof(struct bpf_insn_aux_data
) * off
);
5213 memcpy(new_data
+ off
+ cnt
- 1, old_data
+ off
,
5214 sizeof(struct bpf_insn_aux_data
) * (prog_len
- off
- cnt
+ 1));
5215 for (i
= off
; i
< off
+ cnt
- 1; i
++)
5216 new_data
[i
].seen
= true;
5217 env
->insn_aux_data
= new_data
;
5222 static void adjust_subprog_starts(struct bpf_verifier_env
*env
, u32 off
, u32 len
)
5228 /* NOTE: fake 'exit' subprog should be updated as well. */
5229 for (i
= 0; i
<= env
->subprog_cnt
; i
++) {
5230 if (env
->subprog_info
[i
].start
< off
)
5232 env
->subprog_info
[i
].start
+= len
- 1;
5236 static struct bpf_prog
*bpf_patch_insn_data(struct bpf_verifier_env
*env
, u32 off
,
5237 const struct bpf_insn
*patch
, u32 len
)
5239 struct bpf_prog
*new_prog
;
5241 new_prog
= bpf_patch_insn_single(env
->prog
, off
, patch
, len
);
5244 if (adjust_insn_aux_data(env
, new_prog
->len
, off
, len
))
5246 adjust_subprog_starts(env
, off
, len
);
5250 /* The verifier does more data flow analysis than llvm and will not
5251 * explore branches that are dead at run time. Malicious programs can
5252 * have dead code too. Therefore replace all dead at-run-time code
5255 * Just nops are not optimal, e.g. if they would sit at the end of the
5256 * program and through another bug we would manage to jump there, then
5257 * we'd execute beyond program memory otherwise. Returning exception
5258 * code also wouldn't work since we can have subprogs where the dead
5259 * code could be located.
5261 static void sanitize_dead_code(struct bpf_verifier_env
*env
)
5263 struct bpf_insn_aux_data
*aux_data
= env
->insn_aux_data
;
5264 struct bpf_insn trap
= BPF_JMP_IMM(BPF_JA
, 0, 0, -1);
5265 struct bpf_insn
*insn
= env
->prog
->insnsi
;
5266 const int insn_cnt
= env
->prog
->len
;
5269 for (i
= 0; i
< insn_cnt
; i
++) {
5270 if (aux_data
[i
].seen
)
5272 memcpy(insn
+ i
, &trap
, sizeof(trap
));
5276 /* convert load instructions that access fields of 'struct __sk_buff'
5277 * into sequence of instructions that access fields of 'struct sk_buff'
5279 static int convert_ctx_accesses(struct bpf_verifier_env
*env
)
5281 const struct bpf_verifier_ops
*ops
= env
->ops
;
5282 int i
, cnt
, size
, ctx_field_size
, delta
= 0;
5283 const int insn_cnt
= env
->prog
->len
;
5284 struct bpf_insn insn_buf
[16], *insn
;
5285 struct bpf_prog
*new_prog
;
5286 enum bpf_access_type type
;
5287 bool is_narrower_load
;
5290 if (ops
->gen_prologue
) {
5291 cnt
= ops
->gen_prologue(insn_buf
, env
->seen_direct_write
,
5293 if (cnt
>= ARRAY_SIZE(insn_buf
)) {
5294 verbose(env
, "bpf verifier is misconfigured\n");
5297 new_prog
= bpf_patch_insn_data(env
, 0, insn_buf
, cnt
);
5301 env
->prog
= new_prog
;
5306 if (!ops
->convert_ctx_access
|| bpf_prog_is_dev_bound(env
->prog
->aux
))
5309 insn
= env
->prog
->insnsi
+ delta
;
5311 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
5312 if (insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_B
) ||
5313 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_H
) ||
5314 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_W
) ||
5315 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_DW
))
5317 else if (insn
->code
== (BPF_STX
| BPF_MEM
| BPF_B
) ||
5318 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_H
) ||
5319 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_W
) ||
5320 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_DW
))
5325 if (type
== BPF_WRITE
&&
5326 env
->insn_aux_data
[i
+ delta
].sanitize_stack_off
) {
5327 struct bpf_insn patch
[] = {
5328 /* Sanitize suspicious stack slot with zero.
5329 * There are no memory dependencies for this store,
5330 * since it's only using frame pointer and immediate
5333 BPF_ST_MEM(BPF_DW
, BPF_REG_FP
,
5334 env
->insn_aux_data
[i
+ delta
].sanitize_stack_off
,
5336 /* the original STX instruction will immediately
5337 * overwrite the same stack slot with appropriate value
5342 cnt
= ARRAY_SIZE(patch
);
5343 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, patch
, cnt
);
5348 env
->prog
= new_prog
;
5349 insn
= new_prog
->insnsi
+ i
+ delta
;
5353 if (env
->insn_aux_data
[i
+ delta
].ptr_type
!= PTR_TO_CTX
)
5356 ctx_field_size
= env
->insn_aux_data
[i
+ delta
].ctx_field_size
;
5357 size
= BPF_LDST_BYTES(insn
);
5359 /* If the read access is a narrower load of the field,
5360 * convert to a 4/8-byte load, to minimum program type specific
5361 * convert_ctx_access changes. If conversion is successful,
5362 * we will apply proper mask to the result.
5364 is_narrower_load
= size
< ctx_field_size
;
5365 if (is_narrower_load
) {
5366 u32 size_default
= bpf_ctx_off_adjust_machine(ctx_field_size
);
5367 u32 off
= insn
->off
;
5370 if (type
== BPF_WRITE
) {
5371 verbose(env
, "bpf verifier narrow ctx access misconfigured\n");
5376 if (ctx_field_size
== 4)
5378 else if (ctx_field_size
== 8)
5381 insn
->off
= off
& ~(size_default
- 1);
5382 insn
->code
= BPF_LDX
| BPF_MEM
| size_code
;
5386 cnt
= ops
->convert_ctx_access(type
, insn
, insn_buf
, env
->prog
,
5388 if (cnt
== 0 || cnt
>= ARRAY_SIZE(insn_buf
) ||
5389 (ctx_field_size
&& !target_size
)) {
5390 verbose(env
, "bpf verifier is misconfigured\n");
5394 if (is_narrower_load
&& size
< target_size
) {
5395 if (ctx_field_size
<= 4)
5396 insn_buf
[cnt
++] = BPF_ALU32_IMM(BPF_AND
, insn
->dst_reg
,
5397 (1 << size
* 8) - 1);
5399 insn_buf
[cnt
++] = BPF_ALU64_IMM(BPF_AND
, insn
->dst_reg
,
5400 (1 << size
* 8) - 1);
5403 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
, cnt
);
5409 /* keep walking new program and skip insns we just inserted */
5410 env
->prog
= new_prog
;
5411 insn
= new_prog
->insnsi
+ i
+ delta
;
5417 static int jit_subprogs(struct bpf_verifier_env
*env
)
5419 struct bpf_prog
*prog
= env
->prog
, **func
, *tmp
;
5420 int i
, j
, subprog_start
, subprog_end
= 0, len
, subprog
;
5421 struct bpf_insn
*insn
;
5425 if (env
->subprog_cnt
<= 1)
5428 for (i
= 0, insn
= prog
->insnsi
; i
< prog
->len
; i
++, insn
++) {
5429 if (insn
->code
!= (BPF_JMP
| BPF_CALL
) ||
5430 insn
->src_reg
!= BPF_PSEUDO_CALL
)
5432 subprog
= find_subprog(env
, i
+ insn
->imm
+ 1);
5434 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5438 /* temporarily remember subprog id inside insn instead of
5439 * aux_data, since next loop will split up all insns into funcs
5441 insn
->off
= subprog
;
5442 /* remember original imm in case JIT fails and fallback
5443 * to interpreter will be needed
5445 env
->insn_aux_data
[i
].call_imm
= insn
->imm
;
5446 /* point imm to __bpf_call_base+1 from JITs point of view */
5450 func
= kzalloc(sizeof(prog
) * env
->subprog_cnt
, GFP_KERNEL
);
5454 for (i
= 0; i
< env
->subprog_cnt
; i
++) {
5455 subprog_start
= subprog_end
;
5456 subprog_end
= env
->subprog_info
[i
+ 1].start
;
5458 len
= subprog_end
- subprog_start
;
5459 func
[i
] = bpf_prog_alloc(bpf_prog_size(len
), GFP_USER
);
5462 memcpy(func
[i
]->insnsi
, &prog
->insnsi
[subprog_start
],
5463 len
* sizeof(struct bpf_insn
));
5464 func
[i
]->type
= prog
->type
;
5466 if (bpf_prog_calc_tag(func
[i
]))
5468 func
[i
]->is_func
= 1;
5469 /* Use bpf_prog_F_tag to indicate functions in stack traces.
5470 * Long term would need debug info to populate names
5472 func
[i
]->aux
->name
[0] = 'F';
5473 func
[i
]->aux
->stack_depth
= env
->subprog_info
[i
].stack_depth
;
5474 func
[i
]->jit_requested
= 1;
5475 func
[i
] = bpf_int_jit_compile(func
[i
]);
5476 if (!func
[i
]->jited
) {
5482 /* at this point all bpf functions were successfully JITed
5483 * now populate all bpf_calls with correct addresses and
5484 * run last pass of JIT
5486 for (i
= 0; i
< env
->subprog_cnt
; i
++) {
5487 insn
= func
[i
]->insnsi
;
5488 for (j
= 0; j
< func
[i
]->len
; j
++, insn
++) {
5489 if (insn
->code
!= (BPF_JMP
| BPF_CALL
) ||
5490 insn
->src_reg
!= BPF_PSEUDO_CALL
)
5492 subprog
= insn
->off
;
5493 insn
->imm
= (u64 (*)(u64
, u64
, u64
, u64
, u64
))
5494 func
[subprog
]->bpf_func
-
5498 /* we use the aux data to keep a list of the start addresses
5499 * of the JITed images for each function in the program
5501 * for some architectures, such as powerpc64, the imm field
5502 * might not be large enough to hold the offset of the start
5503 * address of the callee's JITed image from __bpf_call_base
5505 * in such cases, we can lookup the start address of a callee
5506 * by using its subprog id, available from the off field of
5507 * the call instruction, as an index for this list
5509 func
[i
]->aux
->func
= func
;
5510 func
[i
]->aux
->func_cnt
= env
->subprog_cnt
;
5512 for (i
= 0; i
< env
->subprog_cnt
; i
++) {
5513 old_bpf_func
= func
[i
]->bpf_func
;
5514 tmp
= bpf_int_jit_compile(func
[i
]);
5515 if (tmp
!= func
[i
] || func
[i
]->bpf_func
!= old_bpf_func
) {
5516 verbose(env
, "JIT doesn't support bpf-to-bpf calls\n");
5523 /* finally lock prog and jit images for all functions and
5526 for (i
= 0; i
< env
->subprog_cnt
; i
++) {
5527 bpf_prog_lock_ro(func
[i
]);
5528 bpf_prog_kallsyms_add(func
[i
]);
5531 /* Last step: make now unused interpreter insns from main
5532 * prog consistent for later dump requests, so they can
5533 * later look the same as if they were interpreted only.
5535 for (i
= 0, insn
= prog
->insnsi
; i
< prog
->len
; i
++, insn
++) {
5536 if (insn
->code
!= (BPF_JMP
| BPF_CALL
) ||
5537 insn
->src_reg
!= BPF_PSEUDO_CALL
)
5539 insn
->off
= env
->insn_aux_data
[i
].call_imm
;
5540 subprog
= find_subprog(env
, i
+ insn
->off
+ 1);
5541 insn
->imm
= subprog
;
5545 prog
->bpf_func
= func
[0]->bpf_func
;
5546 prog
->aux
->func
= func
;
5547 prog
->aux
->func_cnt
= env
->subprog_cnt
;
5550 for (i
= 0; i
< env
->subprog_cnt
; i
++)
5552 bpf_jit_free(func
[i
]);
5554 /* cleanup main prog to be interpreted */
5555 prog
->jit_requested
= 0;
5556 for (i
= 0, insn
= prog
->insnsi
; i
< prog
->len
; i
++, insn
++) {
5557 if (insn
->code
!= (BPF_JMP
| BPF_CALL
) ||
5558 insn
->src_reg
!= BPF_PSEUDO_CALL
)
5561 insn
->imm
= env
->insn_aux_data
[i
].call_imm
;
5566 static int fixup_call_args(struct bpf_verifier_env
*env
)
5568 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
5569 struct bpf_prog
*prog
= env
->prog
;
5570 struct bpf_insn
*insn
= prog
->insnsi
;
5576 if (env
->prog
->jit_requested
) {
5577 err
= jit_subprogs(env
);
5581 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
5582 for (i
= 0; i
< prog
->len
; i
++, insn
++) {
5583 if (insn
->code
!= (BPF_JMP
| BPF_CALL
) ||
5584 insn
->src_reg
!= BPF_PSEUDO_CALL
)
5586 depth
= get_callee_stack_depth(env
, insn
, i
);
5589 bpf_patch_call_args(insn
, depth
);
5596 /* fixup insn->imm field of bpf_call instructions
5597 * and inline eligible helpers as explicit sequence of BPF instructions
5599 * this function is called after eBPF program passed verification
5601 static int fixup_bpf_calls(struct bpf_verifier_env
*env
)
5603 struct bpf_prog
*prog
= env
->prog
;
5604 struct bpf_insn
*insn
= prog
->insnsi
;
5605 const struct bpf_func_proto
*fn
;
5606 const int insn_cnt
= prog
->len
;
5607 const struct bpf_map_ops
*ops
;
5608 struct bpf_insn_aux_data
*aux
;
5609 struct bpf_insn insn_buf
[16];
5610 struct bpf_prog
*new_prog
;
5611 struct bpf_map
*map_ptr
;
5612 int i
, cnt
, delta
= 0;
5614 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
5615 if (insn
->code
== (BPF_ALU64
| BPF_MOD
| BPF_X
) ||
5616 insn
->code
== (BPF_ALU64
| BPF_DIV
| BPF_X
) ||
5617 insn
->code
== (BPF_ALU
| BPF_MOD
| BPF_X
) ||
5618 insn
->code
== (BPF_ALU
| BPF_DIV
| BPF_X
)) {
5619 bool is64
= BPF_CLASS(insn
->code
) == BPF_ALU64
;
5620 struct bpf_insn mask_and_div
[] = {
5621 BPF_MOV32_REG(insn
->src_reg
, insn
->src_reg
),
5623 BPF_JMP_IMM(BPF_JNE
, insn
->src_reg
, 0, 2),
5624 BPF_ALU32_REG(BPF_XOR
, insn
->dst_reg
, insn
->dst_reg
),
5625 BPF_JMP_IMM(BPF_JA
, 0, 0, 1),
5628 struct bpf_insn mask_and_mod
[] = {
5629 BPF_MOV32_REG(insn
->src_reg
, insn
->src_reg
),
5630 /* Rx mod 0 -> Rx */
5631 BPF_JMP_IMM(BPF_JEQ
, insn
->src_reg
, 0, 1),
5634 struct bpf_insn
*patchlet
;
5636 if (insn
->code
== (BPF_ALU64
| BPF_DIV
| BPF_X
) ||
5637 insn
->code
== (BPF_ALU
| BPF_DIV
| BPF_X
)) {
5638 patchlet
= mask_and_div
+ (is64
? 1 : 0);
5639 cnt
= ARRAY_SIZE(mask_and_div
) - (is64
? 1 : 0);
5641 patchlet
= mask_and_mod
+ (is64
? 1 : 0);
5642 cnt
= ARRAY_SIZE(mask_and_mod
) - (is64
? 1 : 0);
5645 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, patchlet
, cnt
);
5650 env
->prog
= prog
= new_prog
;
5651 insn
= new_prog
->insnsi
+ i
+ delta
;
5655 if (BPF_CLASS(insn
->code
) == BPF_LD
&&
5656 (BPF_MODE(insn
->code
) == BPF_ABS
||
5657 BPF_MODE(insn
->code
) == BPF_IND
)) {
5658 cnt
= env
->ops
->gen_ld_abs(insn
, insn_buf
);
5659 if (cnt
== 0 || cnt
>= ARRAY_SIZE(insn_buf
)) {
5660 verbose(env
, "bpf verifier is misconfigured\n");
5664 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
, cnt
);
5669 env
->prog
= prog
= new_prog
;
5670 insn
= new_prog
->insnsi
+ i
+ delta
;
5674 if (insn
->code
!= (BPF_JMP
| BPF_CALL
))
5676 if (insn
->src_reg
== BPF_PSEUDO_CALL
)
5679 if (insn
->imm
== BPF_FUNC_get_route_realm
)
5680 prog
->dst_needed
= 1;
5681 if (insn
->imm
== BPF_FUNC_get_prandom_u32
)
5682 bpf_user_rnd_init_once();
5683 if (insn
->imm
== BPF_FUNC_override_return
)
5684 prog
->kprobe_override
= 1;
5685 if (insn
->imm
== BPF_FUNC_tail_call
) {
5686 /* If we tail call into other programs, we
5687 * cannot make any assumptions since they can
5688 * be replaced dynamically during runtime in
5689 * the program array.
5691 prog
->cb_access
= 1;
5692 env
->prog
->aux
->stack_depth
= MAX_BPF_STACK
;
5694 /* mark bpf_tail_call as different opcode to avoid
5695 * conditional branch in the interpeter for every normal
5696 * call and to prevent accidental JITing by JIT compiler
5697 * that doesn't support bpf_tail_call yet
5700 insn
->code
= BPF_JMP
| BPF_TAIL_CALL
;
5702 aux
= &env
->insn_aux_data
[i
+ delta
];
5703 if (!bpf_map_ptr_unpriv(aux
))
5706 /* instead of changing every JIT dealing with tail_call
5707 * emit two extra insns:
5708 * if (index >= max_entries) goto out;
5709 * index &= array->index_mask;
5710 * to avoid out-of-bounds cpu speculation
5712 if (bpf_map_ptr_poisoned(aux
)) {
5713 verbose(env
, "tail_call abusing map_ptr\n");
5717 map_ptr
= BPF_MAP_PTR(aux
->map_state
);
5718 insn_buf
[0] = BPF_JMP_IMM(BPF_JGE
, BPF_REG_3
,
5719 map_ptr
->max_entries
, 2);
5720 insn_buf
[1] = BPF_ALU32_IMM(BPF_AND
, BPF_REG_3
,
5721 container_of(map_ptr
,
5724 insn_buf
[2] = *insn
;
5726 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
, cnt
);
5731 env
->prog
= prog
= new_prog
;
5732 insn
= new_prog
->insnsi
+ i
+ delta
;
5736 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
5737 * and other inlining handlers are currently limited to 64 bit
5740 if (prog
->jit_requested
&& BITS_PER_LONG
== 64 &&
5741 (insn
->imm
== BPF_FUNC_map_lookup_elem
||
5742 insn
->imm
== BPF_FUNC_map_update_elem
||
5743 insn
->imm
== BPF_FUNC_map_delete_elem
)) {
5744 aux
= &env
->insn_aux_data
[i
+ delta
];
5745 if (bpf_map_ptr_poisoned(aux
))
5746 goto patch_call_imm
;
5748 map_ptr
= BPF_MAP_PTR(aux
->map_state
);
5750 if (insn
->imm
== BPF_FUNC_map_lookup_elem
&&
5751 ops
->map_gen_lookup
) {
5752 cnt
= ops
->map_gen_lookup(map_ptr
, insn_buf
);
5753 if (cnt
== 0 || cnt
>= ARRAY_SIZE(insn_buf
)) {
5754 verbose(env
, "bpf verifier is misconfigured\n");
5758 new_prog
= bpf_patch_insn_data(env
, i
+ delta
,
5764 env
->prog
= prog
= new_prog
;
5765 insn
= new_prog
->insnsi
+ i
+ delta
;
5769 BUILD_BUG_ON(!__same_type(ops
->map_lookup_elem
,
5770 (void *(*)(struct bpf_map
*map
, void *key
))NULL
));
5771 BUILD_BUG_ON(!__same_type(ops
->map_delete_elem
,
5772 (int (*)(struct bpf_map
*map
, void *key
))NULL
));
5773 BUILD_BUG_ON(!__same_type(ops
->map_update_elem
,
5774 (int (*)(struct bpf_map
*map
, void *key
, void *value
,
5776 switch (insn
->imm
) {
5777 case BPF_FUNC_map_lookup_elem
:
5778 insn
->imm
= BPF_CAST_CALL(ops
->map_lookup_elem
) -
5781 case BPF_FUNC_map_update_elem
:
5782 insn
->imm
= BPF_CAST_CALL(ops
->map_update_elem
) -
5785 case BPF_FUNC_map_delete_elem
:
5786 insn
->imm
= BPF_CAST_CALL(ops
->map_delete_elem
) -
5791 goto patch_call_imm
;
5794 if (insn
->imm
== BPF_FUNC_redirect_map
) {
5795 /* Note, we cannot use prog directly as imm as subsequent
5796 * rewrites would still change the prog pointer. The only
5797 * stable address we can use is aux, which also works with
5798 * prog clones during blinding.
5800 u64 addr
= (unsigned long)prog
->aux
;
5801 struct bpf_insn r4_ld
[] = {
5802 BPF_LD_IMM64(BPF_REG_4
, addr
),
5805 cnt
= ARRAY_SIZE(r4_ld
);
5807 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, r4_ld
, cnt
);
5812 env
->prog
= prog
= new_prog
;
5813 insn
= new_prog
->insnsi
+ i
+ delta
;
5816 fn
= env
->ops
->get_func_proto(insn
->imm
, env
->prog
);
5817 /* all functions that have prototype and verifier allowed
5818 * programs to call them, must be real in-kernel functions
5822 "kernel subsystem misconfigured func %s#%d\n",
5823 func_id_name(insn
->imm
), insn
->imm
);
5826 insn
->imm
= fn
->func
- __bpf_call_base
;
5832 static void free_states(struct bpf_verifier_env
*env
)
5834 struct bpf_verifier_state_list
*sl
, *sln
;
5837 if (!env
->explored_states
)
5840 for (i
= 0; i
< env
->prog
->len
; i
++) {
5841 sl
= env
->explored_states
[i
];
5844 while (sl
!= STATE_LIST_MARK
) {
5846 free_verifier_state(&sl
->state
, false);
5852 kfree(env
->explored_states
);
5855 int bpf_check(struct bpf_prog
**prog
, union bpf_attr
*attr
)
5857 struct bpf_verifier_env
*env
;
5858 struct bpf_verifier_log
*log
;
5861 /* no program is valid */
5862 if (ARRAY_SIZE(bpf_verifier_ops
) == 0)
5865 /* 'struct bpf_verifier_env' can be global, but since it's not small,
5866 * allocate/free it every time bpf_check() is called
5868 env
= kzalloc(sizeof(struct bpf_verifier_env
), GFP_KERNEL
);
5873 env
->insn_aux_data
= vzalloc(sizeof(struct bpf_insn_aux_data
) *
5876 if (!env
->insn_aux_data
)
5879 env
->ops
= bpf_verifier_ops
[env
->prog
->type
];
5881 /* grab the mutex to protect few globals used by verifier */
5882 mutex_lock(&bpf_verifier_lock
);
5884 if (attr
->log_level
|| attr
->log_buf
|| attr
->log_size
) {
5885 /* user requested verbose verifier output
5886 * and supplied buffer to store the verification trace
5888 log
->level
= attr
->log_level
;
5889 log
->ubuf
= (char __user
*) (unsigned long) attr
->log_buf
;
5890 log
->len_total
= attr
->log_size
;
5893 /* log attributes have to be sane */
5894 if (log
->len_total
< 128 || log
->len_total
> UINT_MAX
>> 8 ||
5895 !log
->level
|| !log
->ubuf
)
5899 env
->strict_alignment
= !!(attr
->prog_flags
& BPF_F_STRICT_ALIGNMENT
);
5900 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
))
5901 env
->strict_alignment
= true;
5903 ret
= replace_map_fd_with_map_ptr(env
);
5905 goto skip_full_check
;
5907 if (bpf_prog_is_dev_bound(env
->prog
->aux
)) {
5908 ret
= bpf_prog_offload_verifier_prep(env
);
5910 goto skip_full_check
;
5913 env
->explored_states
= kcalloc(env
->prog
->len
,
5914 sizeof(struct bpf_verifier_state_list
*),
5917 if (!env
->explored_states
)
5918 goto skip_full_check
;
5920 env
->allow_ptr_leaks
= capable(CAP_SYS_ADMIN
);
5922 ret
= check_cfg(env
);
5924 goto skip_full_check
;
5926 ret
= do_check(env
);
5927 if (env
->cur_state
) {
5928 free_verifier_state(env
->cur_state
, true);
5929 env
->cur_state
= NULL
;
5933 while (!pop_stack(env
, NULL
, NULL
));
5937 sanitize_dead_code(env
);
5940 ret
= check_max_stack_depth(env
);
5943 /* program is valid, convert *(u32*)(ctx + off) accesses */
5944 ret
= convert_ctx_accesses(env
);
5947 ret
= fixup_bpf_calls(env
);
5950 ret
= fixup_call_args(env
);
5952 if (log
->level
&& bpf_verifier_log_full(log
))
5954 if (log
->level
&& !log
->ubuf
) {
5956 goto err_release_maps
;
5959 if (ret
== 0 && env
->used_map_cnt
) {
5960 /* if program passed verifier, update used_maps in bpf_prog_info */
5961 env
->prog
->aux
->used_maps
= kmalloc_array(env
->used_map_cnt
,
5962 sizeof(env
->used_maps
[0]),
5965 if (!env
->prog
->aux
->used_maps
) {
5967 goto err_release_maps
;
5970 memcpy(env
->prog
->aux
->used_maps
, env
->used_maps
,
5971 sizeof(env
->used_maps
[0]) * env
->used_map_cnt
);
5972 env
->prog
->aux
->used_map_cnt
= env
->used_map_cnt
;
5974 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
5975 * bpf_ld_imm64 instructions
5977 convert_pseudo_ld_imm64(env
);
5981 if (!env
->prog
->aux
->used_maps
)
5982 /* if we didn't copy map pointers into bpf_prog_info, release
5983 * them now. Otherwise free_used_maps() will release them.
5988 mutex_unlock(&bpf_verifier_lock
);
5989 vfree(env
->insn_aux_data
);