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 case BPF_PROG_TYPE_SK_REUSEPORT
:
1314 /* dst_input() and dst_output() can't write for now */
1318 case BPF_PROG_TYPE_SCHED_CLS
:
1319 case BPF_PROG_TYPE_SCHED_ACT
:
1320 case BPF_PROG_TYPE_XDP
:
1321 case BPF_PROG_TYPE_LWT_XMIT
:
1322 case BPF_PROG_TYPE_SK_SKB
:
1323 case BPF_PROG_TYPE_SK_MSG
:
1325 return meta
->pkt_access
;
1327 env
->seen_direct_write
= true;
1334 static int __check_packet_access(struct bpf_verifier_env
*env
, u32 regno
,
1335 int off
, int size
, bool zero_size_allowed
)
1337 struct bpf_reg_state
*regs
= cur_regs(env
);
1338 struct bpf_reg_state
*reg
= ®s
[regno
];
1340 if (off
< 0 || size
< 0 || (size
== 0 && !zero_size_allowed
) ||
1341 (u64
)off
+ size
> reg
->range
) {
1342 verbose(env
, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
1343 off
, size
, regno
, reg
->id
, reg
->off
, reg
->range
);
1349 static int check_packet_access(struct bpf_verifier_env
*env
, u32 regno
, int off
,
1350 int size
, bool zero_size_allowed
)
1352 struct bpf_reg_state
*regs
= cur_regs(env
);
1353 struct bpf_reg_state
*reg
= ®s
[regno
];
1356 /* We may have added a variable offset to the packet pointer; but any
1357 * reg->range we have comes after that. We are only checking the fixed
1361 /* We don't allow negative numbers, because we aren't tracking enough
1362 * detail to prove they're safe.
1364 if (reg
->smin_value
< 0) {
1365 verbose(env
, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1369 err
= __check_packet_access(env
, regno
, off
, size
, zero_size_allowed
);
1371 verbose(env
, "R%d offset is outside of the packet\n", regno
);
1377 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
1378 static int check_ctx_access(struct bpf_verifier_env
*env
, int insn_idx
, int off
, int size
,
1379 enum bpf_access_type t
, enum bpf_reg_type
*reg_type
)
1381 struct bpf_insn_access_aux info
= {
1382 .reg_type
= *reg_type
,
1385 if (env
->ops
->is_valid_access
&&
1386 env
->ops
->is_valid_access(off
, size
, t
, env
->prog
, &info
)) {
1387 /* A non zero info.ctx_field_size indicates that this field is a
1388 * candidate for later verifier transformation to load the whole
1389 * field and then apply a mask when accessed with a narrower
1390 * access than actual ctx access size. A zero info.ctx_field_size
1391 * will only allow for whole field access and rejects any other
1392 * type of narrower access.
1394 *reg_type
= info
.reg_type
;
1396 env
->insn_aux_data
[insn_idx
].ctx_field_size
= info
.ctx_field_size
;
1397 /* remember the offset of last byte accessed in ctx */
1398 if (env
->prog
->aux
->max_ctx_offset
< off
+ size
)
1399 env
->prog
->aux
->max_ctx_offset
= off
+ size
;
1403 verbose(env
, "invalid bpf_context access off=%d size=%d\n", off
, size
);
1407 static bool __is_pointer_value(bool allow_ptr_leaks
,
1408 const struct bpf_reg_state
*reg
)
1410 if (allow_ptr_leaks
)
1413 return reg
->type
!= SCALAR_VALUE
;
1416 static bool is_pointer_value(struct bpf_verifier_env
*env
, int regno
)
1418 return __is_pointer_value(env
->allow_ptr_leaks
, cur_regs(env
) + regno
);
1421 static bool is_ctx_reg(struct bpf_verifier_env
*env
, int regno
)
1423 const struct bpf_reg_state
*reg
= cur_regs(env
) + regno
;
1425 return reg
->type
== PTR_TO_CTX
;
1428 static bool is_pkt_reg(struct bpf_verifier_env
*env
, int regno
)
1430 const struct bpf_reg_state
*reg
= cur_regs(env
) + regno
;
1432 return type_is_pkt_pointer(reg
->type
);
1435 static int check_pkt_ptr_alignment(struct bpf_verifier_env
*env
,
1436 const struct bpf_reg_state
*reg
,
1437 int off
, int size
, bool strict
)
1439 struct tnum reg_off
;
1442 /* Byte size accesses are always allowed. */
1443 if (!strict
|| size
== 1)
1446 /* For platforms that do not have a Kconfig enabling
1447 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
1448 * NET_IP_ALIGN is universally set to '2'. And on platforms
1449 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
1450 * to this code only in strict mode where we want to emulate
1451 * the NET_IP_ALIGN==2 checking. Therefore use an
1452 * unconditional IP align value of '2'.
1456 reg_off
= tnum_add(reg
->var_off
, tnum_const(ip_align
+ reg
->off
+ off
));
1457 if (!tnum_is_aligned(reg_off
, size
)) {
1460 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1462 "misaligned packet access off %d+%s+%d+%d size %d\n",
1463 ip_align
, tn_buf
, reg
->off
, off
, size
);
1470 static int check_generic_ptr_alignment(struct bpf_verifier_env
*env
,
1471 const struct bpf_reg_state
*reg
,
1472 const char *pointer_desc
,
1473 int off
, int size
, bool strict
)
1475 struct tnum reg_off
;
1477 /* Byte size accesses are always allowed. */
1478 if (!strict
|| size
== 1)
1481 reg_off
= tnum_add(reg
->var_off
, tnum_const(reg
->off
+ off
));
1482 if (!tnum_is_aligned(reg_off
, size
)) {
1485 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1486 verbose(env
, "misaligned %saccess off %s+%d+%d size %d\n",
1487 pointer_desc
, tn_buf
, reg
->off
, off
, size
);
1494 static int check_ptr_alignment(struct bpf_verifier_env
*env
,
1495 const struct bpf_reg_state
*reg
, int off
,
1496 int size
, bool strict_alignment_once
)
1498 bool strict
= env
->strict_alignment
|| strict_alignment_once
;
1499 const char *pointer_desc
= "";
1501 switch (reg
->type
) {
1503 case PTR_TO_PACKET_META
:
1504 /* Special case, because of NET_IP_ALIGN. Given metadata sits
1505 * right in front, treat it the very same way.
1507 return check_pkt_ptr_alignment(env
, reg
, off
, size
, strict
);
1508 case PTR_TO_MAP_VALUE
:
1509 pointer_desc
= "value ";
1512 pointer_desc
= "context ";
1515 pointer_desc
= "stack ";
1516 /* The stack spill tracking logic in check_stack_write()
1517 * and check_stack_read() relies on stack accesses being
1525 return check_generic_ptr_alignment(env
, reg
, pointer_desc
, off
, size
,
1529 static int update_stack_depth(struct bpf_verifier_env
*env
,
1530 const struct bpf_func_state
*func
,
1533 u16 stack
= env
->subprog_info
[func
->subprogno
].stack_depth
;
1538 /* update known max for given subprogram */
1539 env
->subprog_info
[func
->subprogno
].stack_depth
= -off
;
1543 /* starting from main bpf function walk all instructions of the function
1544 * and recursively walk all callees that given function can call.
1545 * Ignore jump and exit insns.
1546 * Since recursion is prevented by check_cfg() this algorithm
1547 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
1549 static int check_max_stack_depth(struct bpf_verifier_env
*env
)
1551 int depth
= 0, frame
= 0, idx
= 0, i
= 0, subprog_end
;
1552 struct bpf_subprog_info
*subprog
= env
->subprog_info
;
1553 struct bpf_insn
*insn
= env
->prog
->insnsi
;
1554 int ret_insn
[MAX_CALL_FRAMES
];
1555 int ret_prog
[MAX_CALL_FRAMES
];
1558 /* round up to 32-bytes, since this is granularity
1559 * of interpreter stack size
1561 depth
+= round_up(max_t(u32
, subprog
[idx
].stack_depth
, 1), 32);
1562 if (depth
> MAX_BPF_STACK
) {
1563 verbose(env
, "combined stack size of %d calls is %d. Too large\n",
1568 subprog_end
= subprog
[idx
+ 1].start
;
1569 for (; i
< subprog_end
; i
++) {
1570 if (insn
[i
].code
!= (BPF_JMP
| BPF_CALL
))
1572 if (insn
[i
].src_reg
!= BPF_PSEUDO_CALL
)
1574 /* remember insn and function to return to */
1575 ret_insn
[frame
] = i
+ 1;
1576 ret_prog
[frame
] = idx
;
1578 /* find the callee */
1579 i
= i
+ insn
[i
].imm
+ 1;
1580 idx
= find_subprog(env
, i
);
1582 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1587 if (frame
>= MAX_CALL_FRAMES
) {
1588 WARN_ONCE(1, "verifier bug. Call stack is too deep\n");
1593 /* end of for() loop means the last insn of the 'subprog'
1594 * was reached. Doesn't matter whether it was JA or EXIT
1598 depth
-= round_up(max_t(u32
, subprog
[idx
].stack_depth
, 1), 32);
1600 i
= ret_insn
[frame
];
1601 idx
= ret_prog
[frame
];
1605 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1606 static int get_callee_stack_depth(struct bpf_verifier_env
*env
,
1607 const struct bpf_insn
*insn
, int idx
)
1609 int start
= idx
+ insn
->imm
+ 1, subprog
;
1611 subprog
= find_subprog(env
, start
);
1613 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1617 return env
->subprog_info
[subprog
].stack_depth
;
1621 static int check_ctx_reg(struct bpf_verifier_env
*env
,
1622 const struct bpf_reg_state
*reg
, int regno
)
1624 /* Access to ctx or passing it to a helper is only allowed in
1625 * its original, unmodified form.
1629 verbose(env
, "dereference of modified ctx ptr R%d off=%d disallowed\n",
1634 if (!tnum_is_const(reg
->var_off
) || reg
->var_off
.value
) {
1637 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1638 verbose(env
, "variable ctx access var_off=%s disallowed\n", tn_buf
);
1645 /* truncate register to smaller size (in bytes)
1646 * must be called with size < BPF_REG_SIZE
1648 static void coerce_reg_to_size(struct bpf_reg_state
*reg
, int size
)
1652 /* clear high bits in bit representation */
1653 reg
->var_off
= tnum_cast(reg
->var_off
, size
);
1655 /* fix arithmetic bounds */
1656 mask
= ((u64
)1 << (size
* 8)) - 1;
1657 if ((reg
->umin_value
& ~mask
) == (reg
->umax_value
& ~mask
)) {
1658 reg
->umin_value
&= mask
;
1659 reg
->umax_value
&= mask
;
1661 reg
->umin_value
= 0;
1662 reg
->umax_value
= mask
;
1664 reg
->smin_value
= reg
->umin_value
;
1665 reg
->smax_value
= reg
->umax_value
;
1668 /* check whether memory at (regno + off) is accessible for t = (read | write)
1669 * if t==write, value_regno is a register which value is stored into memory
1670 * if t==read, value_regno is a register which will receive the value from memory
1671 * if t==write && value_regno==-1, some unknown value is stored into memory
1672 * if t==read && value_regno==-1, don't care what we read from memory
1674 static int check_mem_access(struct bpf_verifier_env
*env
, int insn_idx
, u32 regno
,
1675 int off
, int bpf_size
, enum bpf_access_type t
,
1676 int value_regno
, bool strict_alignment_once
)
1678 struct bpf_reg_state
*regs
= cur_regs(env
);
1679 struct bpf_reg_state
*reg
= regs
+ regno
;
1680 struct bpf_func_state
*state
;
1683 size
= bpf_size_to_bytes(bpf_size
);
1687 /* alignment checks will add in reg->off themselves */
1688 err
= check_ptr_alignment(env
, reg
, off
, size
, strict_alignment_once
);
1692 /* for access checks, reg->off is just part of off */
1695 if (reg
->type
== PTR_TO_MAP_VALUE
) {
1696 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
1697 is_pointer_value(env
, value_regno
)) {
1698 verbose(env
, "R%d leaks addr into map\n", value_regno
);
1702 err
= check_map_access(env
, regno
, off
, size
, false);
1703 if (!err
&& t
== BPF_READ
&& value_regno
>= 0)
1704 mark_reg_unknown(env
, regs
, value_regno
);
1706 } else if (reg
->type
== PTR_TO_CTX
) {
1707 enum bpf_reg_type reg_type
= SCALAR_VALUE
;
1709 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
1710 is_pointer_value(env
, value_regno
)) {
1711 verbose(env
, "R%d leaks addr into ctx\n", value_regno
);
1715 err
= check_ctx_reg(env
, reg
, regno
);
1719 err
= check_ctx_access(env
, insn_idx
, off
, size
, t
, ®_type
);
1720 if (!err
&& t
== BPF_READ
&& value_regno
>= 0) {
1721 /* ctx access returns either a scalar, or a
1722 * PTR_TO_PACKET[_META,_END]. In the latter
1723 * case, we know the offset is zero.
1725 if (reg_type
== SCALAR_VALUE
)
1726 mark_reg_unknown(env
, regs
, value_regno
);
1728 mark_reg_known_zero(env
, regs
,
1730 regs
[value_regno
].id
= 0;
1731 regs
[value_regno
].off
= 0;
1732 regs
[value_regno
].range
= 0;
1733 regs
[value_regno
].type
= reg_type
;
1736 } else if (reg
->type
== PTR_TO_STACK
) {
1737 /* stack accesses must be at a fixed offset, so that we can
1738 * determine what type of data were returned.
1739 * See check_stack_read().
1741 if (!tnum_is_const(reg
->var_off
)) {
1744 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1745 verbose(env
, "variable stack access var_off=%s off=%d size=%d",
1749 off
+= reg
->var_off
.value
;
1750 if (off
>= 0 || off
< -MAX_BPF_STACK
) {
1751 verbose(env
, "invalid stack off=%d size=%d\n", off
,
1756 state
= func(env
, reg
);
1757 err
= update_stack_depth(env
, state
, off
);
1762 err
= check_stack_write(env
, state
, off
, size
,
1763 value_regno
, insn_idx
);
1765 err
= check_stack_read(env
, state
, off
, size
,
1767 } else if (reg_is_pkt_pointer(reg
)) {
1768 if (t
== BPF_WRITE
&& !may_access_direct_pkt_data(env
, NULL
, t
)) {
1769 verbose(env
, "cannot write into packet\n");
1772 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
1773 is_pointer_value(env
, value_regno
)) {
1774 verbose(env
, "R%d leaks addr into packet\n",
1778 err
= check_packet_access(env
, regno
, off
, size
, false);
1779 if (!err
&& t
== BPF_READ
&& value_regno
>= 0)
1780 mark_reg_unknown(env
, regs
, value_regno
);
1782 verbose(env
, "R%d invalid mem access '%s'\n", regno
,
1783 reg_type_str
[reg
->type
]);
1787 if (!err
&& size
< BPF_REG_SIZE
&& value_regno
>= 0 && t
== BPF_READ
&&
1788 regs
[value_regno
].type
== SCALAR_VALUE
) {
1789 /* b/h/w load zero-extends, mark upper bits as known 0 */
1790 coerce_reg_to_size(®s
[value_regno
], size
);
1795 static int check_xadd(struct bpf_verifier_env
*env
, int insn_idx
, struct bpf_insn
*insn
)
1799 if ((BPF_SIZE(insn
->code
) != BPF_W
&& BPF_SIZE(insn
->code
) != BPF_DW
) ||
1801 verbose(env
, "BPF_XADD uses reserved fields\n");
1805 /* check src1 operand */
1806 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
1810 /* check src2 operand */
1811 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
1815 if (is_pointer_value(env
, insn
->src_reg
)) {
1816 verbose(env
, "R%d leaks addr into mem\n", insn
->src_reg
);
1820 if (is_ctx_reg(env
, insn
->dst_reg
) ||
1821 is_pkt_reg(env
, insn
->dst_reg
)) {
1822 verbose(env
, "BPF_XADD stores into R%d %s is not allowed\n",
1823 insn
->dst_reg
, is_ctx_reg(env
, insn
->dst_reg
) ?
1824 "context" : "packet");
1828 /* check whether atomic_add can read the memory */
1829 err
= check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
1830 BPF_SIZE(insn
->code
), BPF_READ
, -1, true);
1834 /* check whether atomic_add can write into the same memory */
1835 return check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
1836 BPF_SIZE(insn
->code
), BPF_WRITE
, -1, true);
1839 /* when register 'regno' is passed into function that will read 'access_size'
1840 * bytes from that pointer, make sure that it's within stack boundary
1841 * and all elements of stack are initialized.
1842 * Unlike most pointer bounds-checking functions, this one doesn't take an
1843 * 'off' argument, so it has to add in reg->off itself.
1845 static int check_stack_boundary(struct bpf_verifier_env
*env
, int regno
,
1846 int access_size
, bool zero_size_allowed
,
1847 struct bpf_call_arg_meta
*meta
)
1849 struct bpf_reg_state
*reg
= cur_regs(env
) + regno
;
1850 struct bpf_func_state
*state
= func(env
, reg
);
1851 int off
, i
, slot
, spi
;
1853 if (reg
->type
!= PTR_TO_STACK
) {
1854 /* Allow zero-byte read from NULL, regardless of pointer type */
1855 if (zero_size_allowed
&& access_size
== 0 &&
1856 register_is_null(reg
))
1859 verbose(env
, "R%d type=%s expected=%s\n", regno
,
1860 reg_type_str
[reg
->type
],
1861 reg_type_str
[PTR_TO_STACK
]);
1865 /* Only allow fixed-offset stack reads */
1866 if (!tnum_is_const(reg
->var_off
)) {
1869 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1870 verbose(env
, "invalid variable stack read R%d var_off=%s\n",
1874 off
= reg
->off
+ reg
->var_off
.value
;
1875 if (off
>= 0 || off
< -MAX_BPF_STACK
|| off
+ access_size
> 0 ||
1876 access_size
< 0 || (access_size
== 0 && !zero_size_allowed
)) {
1877 verbose(env
, "invalid stack type R%d off=%d access_size=%d\n",
1878 regno
, off
, access_size
);
1882 if (meta
&& meta
->raw_mode
) {
1883 meta
->access_size
= access_size
;
1884 meta
->regno
= regno
;
1888 for (i
= 0; i
< access_size
; i
++) {
1891 slot
= -(off
+ i
) - 1;
1892 spi
= slot
/ BPF_REG_SIZE
;
1893 if (state
->allocated_stack
<= slot
)
1895 stype
= &state
->stack
[spi
].slot_type
[slot
% BPF_REG_SIZE
];
1896 if (*stype
== STACK_MISC
)
1898 if (*stype
== STACK_ZERO
) {
1899 /* helper can write anything into the stack */
1900 *stype
= STACK_MISC
;
1904 verbose(env
, "invalid indirect read from stack off %d+%d size %d\n",
1905 off
, i
, access_size
);
1908 /* reading any byte out of 8-byte 'spill_slot' will cause
1909 * the whole slot to be marked as 'read'
1911 mark_stack_slot_read(env
, env
->cur_state
, env
->cur_state
->parent
,
1912 spi
, state
->frameno
);
1914 return update_stack_depth(env
, state
, off
);
1917 static int check_helper_mem_access(struct bpf_verifier_env
*env
, int regno
,
1918 int access_size
, bool zero_size_allowed
,
1919 struct bpf_call_arg_meta
*meta
)
1921 struct bpf_reg_state
*regs
= cur_regs(env
), *reg
= ®s
[regno
];
1923 switch (reg
->type
) {
1925 case PTR_TO_PACKET_META
:
1926 return check_packet_access(env
, regno
, reg
->off
, access_size
,
1928 case PTR_TO_MAP_VALUE
:
1929 return check_map_access(env
, regno
, reg
->off
, access_size
,
1931 default: /* scalar_value|ptr_to_stack or invalid ptr */
1932 return check_stack_boundary(env
, regno
, access_size
,
1933 zero_size_allowed
, meta
);
1937 static bool arg_type_is_mem_ptr(enum bpf_arg_type type
)
1939 return type
== ARG_PTR_TO_MEM
||
1940 type
== ARG_PTR_TO_MEM_OR_NULL
||
1941 type
== ARG_PTR_TO_UNINIT_MEM
;
1944 static bool arg_type_is_mem_size(enum bpf_arg_type type
)
1946 return type
== ARG_CONST_SIZE
||
1947 type
== ARG_CONST_SIZE_OR_ZERO
;
1950 static int check_func_arg(struct bpf_verifier_env
*env
, u32 regno
,
1951 enum bpf_arg_type arg_type
,
1952 struct bpf_call_arg_meta
*meta
)
1954 struct bpf_reg_state
*regs
= cur_regs(env
), *reg
= ®s
[regno
];
1955 enum bpf_reg_type expected_type
, type
= reg
->type
;
1958 if (arg_type
== ARG_DONTCARE
)
1961 err
= check_reg_arg(env
, regno
, SRC_OP
);
1965 if (arg_type
== ARG_ANYTHING
) {
1966 if (is_pointer_value(env
, regno
)) {
1967 verbose(env
, "R%d leaks addr into helper function\n",
1974 if (type_is_pkt_pointer(type
) &&
1975 !may_access_direct_pkt_data(env
, meta
, BPF_READ
)) {
1976 verbose(env
, "helper access to the packet is not allowed\n");
1980 if (arg_type
== ARG_PTR_TO_MAP_KEY
||
1981 arg_type
== ARG_PTR_TO_MAP_VALUE
) {
1982 expected_type
= PTR_TO_STACK
;
1983 if (!type_is_pkt_pointer(type
) && type
!= PTR_TO_MAP_VALUE
&&
1984 type
!= expected_type
)
1986 } else if (arg_type
== ARG_CONST_SIZE
||
1987 arg_type
== ARG_CONST_SIZE_OR_ZERO
) {
1988 expected_type
= SCALAR_VALUE
;
1989 if (type
!= expected_type
)
1991 } else if (arg_type
== ARG_CONST_MAP_PTR
) {
1992 expected_type
= CONST_PTR_TO_MAP
;
1993 if (type
!= expected_type
)
1995 } else if (arg_type
== ARG_PTR_TO_CTX
) {
1996 expected_type
= PTR_TO_CTX
;
1997 if (type
!= expected_type
)
1999 err
= check_ctx_reg(env
, reg
, regno
);
2002 } else if (arg_type_is_mem_ptr(arg_type
)) {
2003 expected_type
= PTR_TO_STACK
;
2004 /* One exception here. In case function allows for NULL to be
2005 * passed in as argument, it's a SCALAR_VALUE type. Final test
2006 * happens during stack boundary checking.
2008 if (register_is_null(reg
) &&
2009 arg_type
== ARG_PTR_TO_MEM_OR_NULL
)
2010 /* final test in check_stack_boundary() */;
2011 else if (!type_is_pkt_pointer(type
) &&
2012 type
!= PTR_TO_MAP_VALUE
&&
2013 type
!= expected_type
)
2015 meta
->raw_mode
= arg_type
== ARG_PTR_TO_UNINIT_MEM
;
2017 verbose(env
, "unsupported arg_type %d\n", arg_type
);
2021 if (arg_type
== ARG_CONST_MAP_PTR
) {
2022 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
2023 meta
->map_ptr
= reg
->map_ptr
;
2024 } else if (arg_type
== ARG_PTR_TO_MAP_KEY
) {
2025 /* bpf_map_xxx(..., map_ptr, ..., key) call:
2026 * check that [key, key + map->key_size) are within
2027 * stack limits and initialized
2029 if (!meta
->map_ptr
) {
2030 /* in function declaration map_ptr must come before
2031 * map_key, so that it's verified and known before
2032 * we have to check map_key here. Otherwise it means
2033 * that kernel subsystem misconfigured verifier
2035 verbose(env
, "invalid map_ptr to access map->key\n");
2038 err
= check_helper_mem_access(env
, regno
,
2039 meta
->map_ptr
->key_size
, false,
2041 } else if (arg_type
== ARG_PTR_TO_MAP_VALUE
) {
2042 /* bpf_map_xxx(..., map_ptr, ..., value) call:
2043 * check [value, value + map->value_size) validity
2045 if (!meta
->map_ptr
) {
2046 /* kernel subsystem misconfigured verifier */
2047 verbose(env
, "invalid map_ptr to access map->value\n");
2050 err
= check_helper_mem_access(env
, regno
,
2051 meta
->map_ptr
->value_size
, false,
2053 } else if (arg_type_is_mem_size(arg_type
)) {
2054 bool zero_size_allowed
= (arg_type
== ARG_CONST_SIZE_OR_ZERO
);
2056 /* remember the mem_size which may be used later
2057 * to refine return values.
2059 meta
->msize_smax_value
= reg
->smax_value
;
2060 meta
->msize_umax_value
= reg
->umax_value
;
2062 /* The register is SCALAR_VALUE; the access check
2063 * happens using its boundaries.
2065 if (!tnum_is_const(reg
->var_off
))
2066 /* For unprivileged variable accesses, disable raw
2067 * mode so that the program is required to
2068 * initialize all the memory that the helper could
2069 * just partially fill up.
2073 if (reg
->smin_value
< 0) {
2074 verbose(env
, "R%d min value is negative, either use unsigned or 'var &= const'\n",
2079 if (reg
->umin_value
== 0) {
2080 err
= check_helper_mem_access(env
, regno
- 1, 0,
2087 if (reg
->umax_value
>= BPF_MAX_VAR_SIZ
) {
2088 verbose(env
, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
2092 err
= check_helper_mem_access(env
, regno
- 1,
2094 zero_size_allowed
, meta
);
2099 verbose(env
, "R%d type=%s expected=%s\n", regno
,
2100 reg_type_str
[type
], reg_type_str
[expected_type
]);
2104 static int check_map_func_compatibility(struct bpf_verifier_env
*env
,
2105 struct bpf_map
*map
, int func_id
)
2110 /* We need a two way check, first is from map perspective ... */
2111 switch (map
->map_type
) {
2112 case BPF_MAP_TYPE_PROG_ARRAY
:
2113 if (func_id
!= BPF_FUNC_tail_call
)
2116 case BPF_MAP_TYPE_PERF_EVENT_ARRAY
:
2117 if (func_id
!= BPF_FUNC_perf_event_read
&&
2118 func_id
!= BPF_FUNC_perf_event_output
&&
2119 func_id
!= BPF_FUNC_perf_event_read_value
)
2122 case BPF_MAP_TYPE_STACK_TRACE
:
2123 if (func_id
!= BPF_FUNC_get_stackid
)
2126 case BPF_MAP_TYPE_CGROUP_ARRAY
:
2127 if (func_id
!= BPF_FUNC_skb_under_cgroup
&&
2128 func_id
!= BPF_FUNC_current_task_under_cgroup
)
2131 case BPF_MAP_TYPE_CGROUP_STORAGE
:
2132 if (func_id
!= BPF_FUNC_get_local_storage
)
2135 /* devmap returns a pointer to a live net_device ifindex that we cannot
2136 * allow to be modified from bpf side. So do not allow lookup elements
2139 case BPF_MAP_TYPE_DEVMAP
:
2140 if (func_id
!= BPF_FUNC_redirect_map
)
2143 /* Restrict bpf side of cpumap and xskmap, open when use-cases
2146 case BPF_MAP_TYPE_CPUMAP
:
2147 case BPF_MAP_TYPE_XSKMAP
:
2148 if (func_id
!= BPF_FUNC_redirect_map
)
2151 case BPF_MAP_TYPE_ARRAY_OF_MAPS
:
2152 case BPF_MAP_TYPE_HASH_OF_MAPS
:
2153 if (func_id
!= BPF_FUNC_map_lookup_elem
)
2156 case BPF_MAP_TYPE_SOCKMAP
:
2157 if (func_id
!= BPF_FUNC_sk_redirect_map
&&
2158 func_id
!= BPF_FUNC_sock_map_update
&&
2159 func_id
!= BPF_FUNC_map_delete_elem
&&
2160 func_id
!= BPF_FUNC_msg_redirect_map
)
2163 case BPF_MAP_TYPE_SOCKHASH
:
2164 if (func_id
!= BPF_FUNC_sk_redirect_hash
&&
2165 func_id
!= BPF_FUNC_sock_hash_update
&&
2166 func_id
!= BPF_FUNC_map_delete_elem
&&
2167 func_id
!= BPF_FUNC_msg_redirect_hash
)
2170 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY
:
2171 if (func_id
!= BPF_FUNC_sk_select_reuseport
)
2178 /* ... and second from the function itself. */
2180 case BPF_FUNC_tail_call
:
2181 if (map
->map_type
!= BPF_MAP_TYPE_PROG_ARRAY
)
2183 if (env
->subprog_cnt
> 1) {
2184 verbose(env
, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
2188 case BPF_FUNC_perf_event_read
:
2189 case BPF_FUNC_perf_event_output
:
2190 case BPF_FUNC_perf_event_read_value
:
2191 if (map
->map_type
!= BPF_MAP_TYPE_PERF_EVENT_ARRAY
)
2194 case BPF_FUNC_get_stackid
:
2195 if (map
->map_type
!= BPF_MAP_TYPE_STACK_TRACE
)
2198 case BPF_FUNC_current_task_under_cgroup
:
2199 case BPF_FUNC_skb_under_cgroup
:
2200 if (map
->map_type
!= BPF_MAP_TYPE_CGROUP_ARRAY
)
2203 case BPF_FUNC_redirect_map
:
2204 if (map
->map_type
!= BPF_MAP_TYPE_DEVMAP
&&
2205 map
->map_type
!= BPF_MAP_TYPE_CPUMAP
&&
2206 map
->map_type
!= BPF_MAP_TYPE_XSKMAP
)
2209 case BPF_FUNC_sk_redirect_map
:
2210 case BPF_FUNC_msg_redirect_map
:
2211 case BPF_FUNC_sock_map_update
:
2212 if (map
->map_type
!= BPF_MAP_TYPE_SOCKMAP
)
2215 case BPF_FUNC_sk_redirect_hash
:
2216 case BPF_FUNC_msg_redirect_hash
:
2217 case BPF_FUNC_sock_hash_update
:
2218 if (map
->map_type
!= BPF_MAP_TYPE_SOCKHASH
)
2221 case BPF_FUNC_get_local_storage
:
2222 if (map
->map_type
!= BPF_MAP_TYPE_CGROUP_STORAGE
)
2225 case BPF_FUNC_sk_select_reuseport
:
2226 if (map
->map_type
!= BPF_MAP_TYPE_REUSEPORT_SOCKARRAY
)
2235 verbose(env
, "cannot pass map_type %d into func %s#%d\n",
2236 map
->map_type
, func_id_name(func_id
), func_id
);
2240 static bool check_raw_mode_ok(const struct bpf_func_proto
*fn
)
2244 if (fn
->arg1_type
== ARG_PTR_TO_UNINIT_MEM
)
2246 if (fn
->arg2_type
== ARG_PTR_TO_UNINIT_MEM
)
2248 if (fn
->arg3_type
== ARG_PTR_TO_UNINIT_MEM
)
2250 if (fn
->arg4_type
== ARG_PTR_TO_UNINIT_MEM
)
2252 if (fn
->arg5_type
== ARG_PTR_TO_UNINIT_MEM
)
2255 /* We only support one arg being in raw mode at the moment,
2256 * which is sufficient for the helper functions we have
2262 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr
,
2263 enum bpf_arg_type arg_next
)
2265 return (arg_type_is_mem_ptr(arg_curr
) &&
2266 !arg_type_is_mem_size(arg_next
)) ||
2267 (!arg_type_is_mem_ptr(arg_curr
) &&
2268 arg_type_is_mem_size(arg_next
));
2271 static bool check_arg_pair_ok(const struct bpf_func_proto
*fn
)
2273 /* bpf_xxx(..., buf, len) call will access 'len'
2274 * bytes from memory 'buf'. Both arg types need
2275 * to be paired, so make sure there's no buggy
2276 * helper function specification.
2278 if (arg_type_is_mem_size(fn
->arg1_type
) ||
2279 arg_type_is_mem_ptr(fn
->arg5_type
) ||
2280 check_args_pair_invalid(fn
->arg1_type
, fn
->arg2_type
) ||
2281 check_args_pair_invalid(fn
->arg2_type
, fn
->arg3_type
) ||
2282 check_args_pair_invalid(fn
->arg3_type
, fn
->arg4_type
) ||
2283 check_args_pair_invalid(fn
->arg4_type
, fn
->arg5_type
))
2289 static int check_func_proto(const struct bpf_func_proto
*fn
)
2291 return check_raw_mode_ok(fn
) &&
2292 check_arg_pair_ok(fn
) ? 0 : -EINVAL
;
2295 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
2296 * are now invalid, so turn them into unknown SCALAR_VALUE.
2298 static void __clear_all_pkt_pointers(struct bpf_verifier_env
*env
,
2299 struct bpf_func_state
*state
)
2301 struct bpf_reg_state
*regs
= state
->regs
, *reg
;
2304 for (i
= 0; i
< MAX_BPF_REG
; i
++)
2305 if (reg_is_pkt_pointer_any(®s
[i
]))
2306 mark_reg_unknown(env
, regs
, i
);
2308 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
2309 if (state
->stack
[i
].slot_type
[0] != STACK_SPILL
)
2311 reg
= &state
->stack
[i
].spilled_ptr
;
2312 if (reg_is_pkt_pointer_any(reg
))
2313 __mark_reg_unknown(reg
);
2317 static void clear_all_pkt_pointers(struct bpf_verifier_env
*env
)
2319 struct bpf_verifier_state
*vstate
= env
->cur_state
;
2322 for (i
= 0; i
<= vstate
->curframe
; i
++)
2323 __clear_all_pkt_pointers(env
, vstate
->frame
[i
]);
2326 static int check_func_call(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
,
2329 struct bpf_verifier_state
*state
= env
->cur_state
;
2330 struct bpf_func_state
*caller
, *callee
;
2331 int i
, subprog
, target_insn
;
2333 if (state
->curframe
+ 1 >= MAX_CALL_FRAMES
) {
2334 verbose(env
, "the call stack of %d frames is too deep\n",
2335 state
->curframe
+ 2);
2339 target_insn
= *insn_idx
+ insn
->imm
;
2340 subprog
= find_subprog(env
, target_insn
+ 1);
2342 verbose(env
, "verifier bug. No program starts at insn %d\n",
2347 caller
= state
->frame
[state
->curframe
];
2348 if (state
->frame
[state
->curframe
+ 1]) {
2349 verbose(env
, "verifier bug. Frame %d already allocated\n",
2350 state
->curframe
+ 1);
2354 callee
= kzalloc(sizeof(*callee
), GFP_KERNEL
);
2357 state
->frame
[state
->curframe
+ 1] = callee
;
2359 /* callee cannot access r0, r6 - r9 for reading and has to write
2360 * into its own stack before reading from it.
2361 * callee can read/write into caller's stack
2363 init_func_state(env
, callee
,
2364 /* remember the callsite, it will be used by bpf_exit */
2365 *insn_idx
/* callsite */,
2366 state
->curframe
+ 1 /* frameno within this callchain */,
2367 subprog
/* subprog number within this prog */);
2369 /* copy r1 - r5 args that callee can access */
2370 for (i
= BPF_REG_1
; i
<= BPF_REG_5
; i
++)
2371 callee
->regs
[i
] = caller
->regs
[i
];
2373 /* after the call regsiters r0 - r5 were scratched */
2374 for (i
= 0; i
< CALLER_SAVED_REGS
; i
++) {
2375 mark_reg_not_init(env
, caller
->regs
, caller_saved
[i
]);
2376 check_reg_arg(env
, caller_saved
[i
], DST_OP_NO_MARK
);
2379 /* only increment it after check_reg_arg() finished */
2382 /* and go analyze first insn of the callee */
2383 *insn_idx
= target_insn
;
2385 if (env
->log
.level
) {
2386 verbose(env
, "caller:\n");
2387 print_verifier_state(env
, caller
);
2388 verbose(env
, "callee:\n");
2389 print_verifier_state(env
, callee
);
2394 static int prepare_func_exit(struct bpf_verifier_env
*env
, int *insn_idx
)
2396 struct bpf_verifier_state
*state
= env
->cur_state
;
2397 struct bpf_func_state
*caller
, *callee
;
2398 struct bpf_reg_state
*r0
;
2400 callee
= state
->frame
[state
->curframe
];
2401 r0
= &callee
->regs
[BPF_REG_0
];
2402 if (r0
->type
== PTR_TO_STACK
) {
2403 /* technically it's ok to return caller's stack pointer
2404 * (or caller's caller's pointer) back to the caller,
2405 * since these pointers are valid. Only current stack
2406 * pointer will be invalid as soon as function exits,
2407 * but let's be conservative
2409 verbose(env
, "cannot return stack pointer to the caller\n");
2414 caller
= state
->frame
[state
->curframe
];
2415 /* return to the caller whatever r0 had in the callee */
2416 caller
->regs
[BPF_REG_0
] = *r0
;
2418 *insn_idx
= callee
->callsite
+ 1;
2419 if (env
->log
.level
) {
2420 verbose(env
, "returning from callee:\n");
2421 print_verifier_state(env
, callee
);
2422 verbose(env
, "to caller at %d:\n", *insn_idx
);
2423 print_verifier_state(env
, caller
);
2425 /* clear everything in the callee */
2426 free_func_state(callee
);
2427 state
->frame
[state
->curframe
+ 1] = NULL
;
2431 static void do_refine_retval_range(struct bpf_reg_state
*regs
, int ret_type
,
2433 struct bpf_call_arg_meta
*meta
)
2435 struct bpf_reg_state
*ret_reg
= ®s
[BPF_REG_0
];
2437 if (ret_type
!= RET_INTEGER
||
2438 (func_id
!= BPF_FUNC_get_stack
&&
2439 func_id
!= BPF_FUNC_probe_read_str
))
2442 ret_reg
->smax_value
= meta
->msize_smax_value
;
2443 ret_reg
->umax_value
= meta
->msize_umax_value
;
2444 __reg_deduce_bounds(ret_reg
);
2445 __reg_bound_offset(ret_reg
);
2449 record_func_map(struct bpf_verifier_env
*env
, struct bpf_call_arg_meta
*meta
,
2450 int func_id
, int insn_idx
)
2452 struct bpf_insn_aux_data
*aux
= &env
->insn_aux_data
[insn_idx
];
2454 if (func_id
!= BPF_FUNC_tail_call
&&
2455 func_id
!= BPF_FUNC_map_lookup_elem
&&
2456 func_id
!= BPF_FUNC_map_update_elem
&&
2457 func_id
!= BPF_FUNC_map_delete_elem
)
2460 if (meta
->map_ptr
== NULL
) {
2461 verbose(env
, "kernel subsystem misconfigured verifier\n");
2465 if (!BPF_MAP_PTR(aux
->map_state
))
2466 bpf_map_ptr_store(aux
, meta
->map_ptr
,
2467 meta
->map_ptr
->unpriv_array
);
2468 else if (BPF_MAP_PTR(aux
->map_state
) != meta
->map_ptr
)
2469 bpf_map_ptr_store(aux
, BPF_MAP_PTR_POISON
,
2470 meta
->map_ptr
->unpriv_array
);
2474 static int check_helper_call(struct bpf_verifier_env
*env
, int func_id
, int insn_idx
)
2476 const struct bpf_func_proto
*fn
= NULL
;
2477 struct bpf_reg_state
*regs
;
2478 struct bpf_call_arg_meta meta
;
2482 /* find function prototype */
2483 if (func_id
< 0 || func_id
>= __BPF_FUNC_MAX_ID
) {
2484 verbose(env
, "invalid func %s#%d\n", func_id_name(func_id
),
2489 if (env
->ops
->get_func_proto
)
2490 fn
= env
->ops
->get_func_proto(func_id
, env
->prog
);
2492 verbose(env
, "unknown func %s#%d\n", func_id_name(func_id
),
2497 /* eBPF programs must be GPL compatible to use GPL-ed functions */
2498 if (!env
->prog
->gpl_compatible
&& fn
->gpl_only
) {
2499 verbose(env
, "cannot call GPL-restricted function from non-GPL compatible program\n");
2503 /* With LD_ABS/IND some JITs save/restore skb from r1. */
2504 changes_data
= bpf_helper_changes_pkt_data(fn
->func
);
2505 if (changes_data
&& fn
->arg1_type
!= ARG_PTR_TO_CTX
) {
2506 verbose(env
, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
2507 func_id_name(func_id
), func_id
);
2511 memset(&meta
, 0, sizeof(meta
));
2512 meta
.pkt_access
= fn
->pkt_access
;
2514 err
= check_func_proto(fn
);
2516 verbose(env
, "kernel subsystem misconfigured func %s#%d\n",
2517 func_id_name(func_id
), func_id
);
2522 err
= check_func_arg(env
, BPF_REG_1
, fn
->arg1_type
, &meta
);
2525 err
= check_func_arg(env
, BPF_REG_2
, fn
->arg2_type
, &meta
);
2528 err
= check_func_arg(env
, BPF_REG_3
, fn
->arg3_type
, &meta
);
2531 err
= check_func_arg(env
, BPF_REG_4
, fn
->arg4_type
, &meta
);
2534 err
= check_func_arg(env
, BPF_REG_5
, fn
->arg5_type
, &meta
);
2538 err
= record_func_map(env
, &meta
, func_id
, insn_idx
);
2542 /* Mark slots with STACK_MISC in case of raw mode, stack offset
2543 * is inferred from register state.
2545 for (i
= 0; i
< meta
.access_size
; i
++) {
2546 err
= check_mem_access(env
, insn_idx
, meta
.regno
, i
, BPF_B
,
2547 BPF_WRITE
, -1, false);
2552 regs
= cur_regs(env
);
2554 /* check that flags argument in get_local_storage(map, flags) is 0,
2555 * this is required because get_local_storage() can't return an error.
2557 if (func_id
== BPF_FUNC_get_local_storage
&&
2558 !register_is_null(®s
[BPF_REG_2
])) {
2559 verbose(env
, "get_local_storage() doesn't support non-zero flags\n");
2563 /* reset caller saved regs */
2564 for (i
= 0; i
< CALLER_SAVED_REGS
; i
++) {
2565 mark_reg_not_init(env
, regs
, caller_saved
[i
]);
2566 check_reg_arg(env
, caller_saved
[i
], DST_OP_NO_MARK
);
2569 /* update return register (already marked as written above) */
2570 if (fn
->ret_type
== RET_INTEGER
) {
2571 /* sets type to SCALAR_VALUE */
2572 mark_reg_unknown(env
, regs
, BPF_REG_0
);
2573 } else if (fn
->ret_type
== RET_VOID
) {
2574 regs
[BPF_REG_0
].type
= NOT_INIT
;
2575 } else if (fn
->ret_type
== RET_PTR_TO_MAP_VALUE_OR_NULL
||
2576 fn
->ret_type
== RET_PTR_TO_MAP_VALUE
) {
2577 if (fn
->ret_type
== RET_PTR_TO_MAP_VALUE
)
2578 regs
[BPF_REG_0
].type
= PTR_TO_MAP_VALUE
;
2580 regs
[BPF_REG_0
].type
= PTR_TO_MAP_VALUE_OR_NULL
;
2581 /* There is no offset yet applied, variable or fixed */
2582 mark_reg_known_zero(env
, regs
, BPF_REG_0
);
2583 regs
[BPF_REG_0
].off
= 0;
2584 /* remember map_ptr, so that check_map_access()
2585 * can check 'value_size' boundary of memory access
2586 * to map element returned from bpf_map_lookup_elem()
2588 if (meta
.map_ptr
== NULL
) {
2590 "kernel subsystem misconfigured verifier\n");
2593 regs
[BPF_REG_0
].map_ptr
= meta
.map_ptr
;
2594 regs
[BPF_REG_0
].id
= ++env
->id_gen
;
2596 verbose(env
, "unknown return type %d of func %s#%d\n",
2597 fn
->ret_type
, func_id_name(func_id
), func_id
);
2601 do_refine_retval_range(regs
, fn
->ret_type
, func_id
, &meta
);
2603 err
= check_map_func_compatibility(env
, meta
.map_ptr
, func_id
);
2607 if (func_id
== BPF_FUNC_get_stack
&& !env
->prog
->has_callchain_buf
) {
2608 const char *err_str
;
2610 #ifdef CONFIG_PERF_EVENTS
2611 err
= get_callchain_buffers(sysctl_perf_event_max_stack
);
2612 err_str
= "cannot get callchain buffer for func %s#%d\n";
2615 err_str
= "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
2618 verbose(env
, err_str
, func_id_name(func_id
), func_id
);
2622 env
->prog
->has_callchain_buf
= true;
2626 clear_all_pkt_pointers(env
);
2630 static bool signed_add_overflows(s64 a
, s64 b
)
2632 /* Do the add in u64, where overflow is well-defined */
2633 s64 res
= (s64
)((u64
)a
+ (u64
)b
);
2640 static bool signed_sub_overflows(s64 a
, s64 b
)
2642 /* Do the sub in u64, where overflow is well-defined */
2643 s64 res
= (s64
)((u64
)a
- (u64
)b
);
2650 static bool check_reg_sane_offset(struct bpf_verifier_env
*env
,
2651 const struct bpf_reg_state
*reg
,
2652 enum bpf_reg_type type
)
2654 bool known
= tnum_is_const(reg
->var_off
);
2655 s64 val
= reg
->var_off
.value
;
2656 s64 smin
= reg
->smin_value
;
2658 if (known
&& (val
>= BPF_MAX_VAR_OFF
|| val
<= -BPF_MAX_VAR_OFF
)) {
2659 verbose(env
, "math between %s pointer and %lld is not allowed\n",
2660 reg_type_str
[type
], val
);
2664 if (reg
->off
>= BPF_MAX_VAR_OFF
|| reg
->off
<= -BPF_MAX_VAR_OFF
) {
2665 verbose(env
, "%s pointer offset %d is not allowed\n",
2666 reg_type_str
[type
], reg
->off
);
2670 if (smin
== S64_MIN
) {
2671 verbose(env
, "math between %s pointer and register with unbounded min value is not allowed\n",
2672 reg_type_str
[type
]);
2676 if (smin
>= BPF_MAX_VAR_OFF
|| smin
<= -BPF_MAX_VAR_OFF
) {
2677 verbose(env
, "value %lld makes %s pointer be out of bounds\n",
2678 smin
, reg_type_str
[type
]);
2685 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
2686 * Caller should also handle BPF_MOV case separately.
2687 * If we return -EACCES, caller may want to try again treating pointer as a
2688 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
2690 static int adjust_ptr_min_max_vals(struct bpf_verifier_env
*env
,
2691 struct bpf_insn
*insn
,
2692 const struct bpf_reg_state
*ptr_reg
,
2693 const struct bpf_reg_state
*off_reg
)
2695 struct bpf_verifier_state
*vstate
= env
->cur_state
;
2696 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
2697 struct bpf_reg_state
*regs
= state
->regs
, *dst_reg
;
2698 bool known
= tnum_is_const(off_reg
->var_off
);
2699 s64 smin_val
= off_reg
->smin_value
, smax_val
= off_reg
->smax_value
,
2700 smin_ptr
= ptr_reg
->smin_value
, smax_ptr
= ptr_reg
->smax_value
;
2701 u64 umin_val
= off_reg
->umin_value
, umax_val
= off_reg
->umax_value
,
2702 umin_ptr
= ptr_reg
->umin_value
, umax_ptr
= ptr_reg
->umax_value
;
2703 u8 opcode
= BPF_OP(insn
->code
);
2704 u32 dst
= insn
->dst_reg
;
2706 dst_reg
= ®s
[dst
];
2708 if ((known
&& (smin_val
!= smax_val
|| umin_val
!= umax_val
)) ||
2709 smin_val
> smax_val
|| umin_val
> umax_val
) {
2710 /* Taint dst register if offset had invalid bounds derived from
2711 * e.g. dead branches.
2713 __mark_reg_unknown(dst_reg
);
2717 if (BPF_CLASS(insn
->code
) != BPF_ALU64
) {
2718 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
2720 "R%d 32-bit pointer arithmetic prohibited\n",
2725 if (ptr_reg
->type
== PTR_TO_MAP_VALUE_OR_NULL
) {
2726 verbose(env
, "R%d pointer arithmetic on PTR_TO_MAP_VALUE_OR_NULL prohibited, null-check it first\n",
2730 if (ptr_reg
->type
== CONST_PTR_TO_MAP
) {
2731 verbose(env
, "R%d pointer arithmetic on CONST_PTR_TO_MAP prohibited\n",
2735 if (ptr_reg
->type
== PTR_TO_PACKET_END
) {
2736 verbose(env
, "R%d pointer arithmetic on PTR_TO_PACKET_END prohibited\n",
2741 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
2742 * The id may be overwritten later if we create a new variable offset.
2744 dst_reg
->type
= ptr_reg
->type
;
2745 dst_reg
->id
= ptr_reg
->id
;
2747 if (!check_reg_sane_offset(env
, off_reg
, ptr_reg
->type
) ||
2748 !check_reg_sane_offset(env
, ptr_reg
, ptr_reg
->type
))
2753 /* We can take a fixed offset as long as it doesn't overflow
2754 * the s32 'off' field
2756 if (known
&& (ptr_reg
->off
+ smin_val
==
2757 (s64
)(s32
)(ptr_reg
->off
+ smin_val
))) {
2758 /* pointer += K. Accumulate it into fixed offset */
2759 dst_reg
->smin_value
= smin_ptr
;
2760 dst_reg
->smax_value
= smax_ptr
;
2761 dst_reg
->umin_value
= umin_ptr
;
2762 dst_reg
->umax_value
= umax_ptr
;
2763 dst_reg
->var_off
= ptr_reg
->var_off
;
2764 dst_reg
->off
= ptr_reg
->off
+ smin_val
;
2765 dst_reg
->range
= ptr_reg
->range
;
2768 /* A new variable offset is created. Note that off_reg->off
2769 * == 0, since it's a scalar.
2770 * dst_reg gets the pointer type and since some positive
2771 * integer value was added to the pointer, give it a new 'id'
2772 * if it's a PTR_TO_PACKET.
2773 * this creates a new 'base' pointer, off_reg (variable) gets
2774 * added into the variable offset, and we copy the fixed offset
2777 if (signed_add_overflows(smin_ptr
, smin_val
) ||
2778 signed_add_overflows(smax_ptr
, smax_val
)) {
2779 dst_reg
->smin_value
= S64_MIN
;
2780 dst_reg
->smax_value
= S64_MAX
;
2782 dst_reg
->smin_value
= smin_ptr
+ smin_val
;
2783 dst_reg
->smax_value
= smax_ptr
+ smax_val
;
2785 if (umin_ptr
+ umin_val
< umin_ptr
||
2786 umax_ptr
+ umax_val
< umax_ptr
) {
2787 dst_reg
->umin_value
= 0;
2788 dst_reg
->umax_value
= U64_MAX
;
2790 dst_reg
->umin_value
= umin_ptr
+ umin_val
;
2791 dst_reg
->umax_value
= umax_ptr
+ umax_val
;
2793 dst_reg
->var_off
= tnum_add(ptr_reg
->var_off
, off_reg
->var_off
);
2794 dst_reg
->off
= ptr_reg
->off
;
2795 if (reg_is_pkt_pointer(ptr_reg
)) {
2796 dst_reg
->id
= ++env
->id_gen
;
2797 /* something was added to pkt_ptr, set range to zero */
2802 if (dst_reg
== off_reg
) {
2803 /* scalar -= pointer. Creates an unknown scalar */
2804 verbose(env
, "R%d tried to subtract pointer from scalar\n",
2808 /* We don't allow subtraction from FP, because (according to
2809 * test_verifier.c test "invalid fp arithmetic", JITs might not
2810 * be able to deal with it.
2812 if (ptr_reg
->type
== PTR_TO_STACK
) {
2813 verbose(env
, "R%d subtraction from stack pointer prohibited\n",
2817 if (known
&& (ptr_reg
->off
- smin_val
==
2818 (s64
)(s32
)(ptr_reg
->off
- smin_val
))) {
2819 /* pointer -= K. Subtract it from fixed offset */
2820 dst_reg
->smin_value
= smin_ptr
;
2821 dst_reg
->smax_value
= smax_ptr
;
2822 dst_reg
->umin_value
= umin_ptr
;
2823 dst_reg
->umax_value
= umax_ptr
;
2824 dst_reg
->var_off
= ptr_reg
->var_off
;
2825 dst_reg
->id
= ptr_reg
->id
;
2826 dst_reg
->off
= ptr_reg
->off
- smin_val
;
2827 dst_reg
->range
= ptr_reg
->range
;
2830 /* A new variable offset is created. If the subtrahend is known
2831 * nonnegative, then any reg->range we had before is still good.
2833 if (signed_sub_overflows(smin_ptr
, smax_val
) ||
2834 signed_sub_overflows(smax_ptr
, smin_val
)) {
2835 /* Overflow possible, we know nothing */
2836 dst_reg
->smin_value
= S64_MIN
;
2837 dst_reg
->smax_value
= S64_MAX
;
2839 dst_reg
->smin_value
= smin_ptr
- smax_val
;
2840 dst_reg
->smax_value
= smax_ptr
- smin_val
;
2842 if (umin_ptr
< umax_val
) {
2843 /* Overflow possible, we know nothing */
2844 dst_reg
->umin_value
= 0;
2845 dst_reg
->umax_value
= U64_MAX
;
2847 /* Cannot overflow (as long as bounds are consistent) */
2848 dst_reg
->umin_value
= umin_ptr
- umax_val
;
2849 dst_reg
->umax_value
= umax_ptr
- umin_val
;
2851 dst_reg
->var_off
= tnum_sub(ptr_reg
->var_off
, off_reg
->var_off
);
2852 dst_reg
->off
= ptr_reg
->off
;
2853 if (reg_is_pkt_pointer(ptr_reg
)) {
2854 dst_reg
->id
= ++env
->id_gen
;
2855 /* something was added to pkt_ptr, set range to zero */
2863 /* bitwise ops on pointers are troublesome, prohibit. */
2864 verbose(env
, "R%d bitwise operator %s on pointer prohibited\n",
2865 dst
, bpf_alu_string
[opcode
>> 4]);
2868 /* other operators (e.g. MUL,LSH) produce non-pointer results */
2869 verbose(env
, "R%d pointer arithmetic with %s operator prohibited\n",
2870 dst
, bpf_alu_string
[opcode
>> 4]);
2874 if (!check_reg_sane_offset(env
, dst_reg
, ptr_reg
->type
))
2877 __update_reg_bounds(dst_reg
);
2878 __reg_deduce_bounds(dst_reg
);
2879 __reg_bound_offset(dst_reg
);
2883 /* WARNING: This function does calculations on 64-bit values, but the actual
2884 * execution may occur on 32-bit values. Therefore, things like bitshifts
2885 * need extra checks in the 32-bit case.
2887 static int adjust_scalar_min_max_vals(struct bpf_verifier_env
*env
,
2888 struct bpf_insn
*insn
,
2889 struct bpf_reg_state
*dst_reg
,
2890 struct bpf_reg_state src_reg
)
2892 struct bpf_reg_state
*regs
= cur_regs(env
);
2893 u8 opcode
= BPF_OP(insn
->code
);
2894 bool src_known
, dst_known
;
2895 s64 smin_val
, smax_val
;
2896 u64 umin_val
, umax_val
;
2897 u64 insn_bitness
= (BPF_CLASS(insn
->code
) == BPF_ALU64
) ? 64 : 32;
2899 if (insn_bitness
== 32) {
2900 /* Relevant for 32-bit RSH: Information can propagate towards
2901 * LSB, so it isn't sufficient to only truncate the output to
2904 coerce_reg_to_size(dst_reg
, 4);
2905 coerce_reg_to_size(&src_reg
, 4);
2908 smin_val
= src_reg
.smin_value
;
2909 smax_val
= src_reg
.smax_value
;
2910 umin_val
= src_reg
.umin_value
;
2911 umax_val
= src_reg
.umax_value
;
2912 src_known
= tnum_is_const(src_reg
.var_off
);
2913 dst_known
= tnum_is_const(dst_reg
->var_off
);
2915 if ((src_known
&& (smin_val
!= smax_val
|| umin_val
!= umax_val
)) ||
2916 smin_val
> smax_val
|| umin_val
> umax_val
) {
2917 /* Taint dst register if offset had invalid bounds derived from
2918 * e.g. dead branches.
2920 __mark_reg_unknown(dst_reg
);
2925 opcode
!= BPF_ADD
&& opcode
!= BPF_SUB
&& opcode
!= BPF_AND
) {
2926 __mark_reg_unknown(dst_reg
);
2932 if (signed_add_overflows(dst_reg
->smin_value
, smin_val
) ||
2933 signed_add_overflows(dst_reg
->smax_value
, smax_val
)) {
2934 dst_reg
->smin_value
= S64_MIN
;
2935 dst_reg
->smax_value
= S64_MAX
;
2937 dst_reg
->smin_value
+= smin_val
;
2938 dst_reg
->smax_value
+= smax_val
;
2940 if (dst_reg
->umin_value
+ umin_val
< umin_val
||
2941 dst_reg
->umax_value
+ umax_val
< umax_val
) {
2942 dst_reg
->umin_value
= 0;
2943 dst_reg
->umax_value
= U64_MAX
;
2945 dst_reg
->umin_value
+= umin_val
;
2946 dst_reg
->umax_value
+= umax_val
;
2948 dst_reg
->var_off
= tnum_add(dst_reg
->var_off
, src_reg
.var_off
);
2951 if (signed_sub_overflows(dst_reg
->smin_value
, smax_val
) ||
2952 signed_sub_overflows(dst_reg
->smax_value
, smin_val
)) {
2953 /* Overflow possible, we know nothing */
2954 dst_reg
->smin_value
= S64_MIN
;
2955 dst_reg
->smax_value
= S64_MAX
;
2957 dst_reg
->smin_value
-= smax_val
;
2958 dst_reg
->smax_value
-= smin_val
;
2960 if (dst_reg
->umin_value
< umax_val
) {
2961 /* Overflow possible, we know nothing */
2962 dst_reg
->umin_value
= 0;
2963 dst_reg
->umax_value
= U64_MAX
;
2965 /* Cannot overflow (as long as bounds are consistent) */
2966 dst_reg
->umin_value
-= umax_val
;
2967 dst_reg
->umax_value
-= umin_val
;
2969 dst_reg
->var_off
= tnum_sub(dst_reg
->var_off
, src_reg
.var_off
);
2972 dst_reg
->var_off
= tnum_mul(dst_reg
->var_off
, src_reg
.var_off
);
2973 if (smin_val
< 0 || dst_reg
->smin_value
< 0) {
2974 /* Ain't nobody got time to multiply that sign */
2975 __mark_reg_unbounded(dst_reg
);
2976 __update_reg_bounds(dst_reg
);
2979 /* Both values are positive, so we can work with unsigned and
2980 * copy the result to signed (unless it exceeds S64_MAX).
2982 if (umax_val
> U32_MAX
|| dst_reg
->umax_value
> U32_MAX
) {
2983 /* Potential overflow, we know nothing */
2984 __mark_reg_unbounded(dst_reg
);
2985 /* (except what we can learn from the var_off) */
2986 __update_reg_bounds(dst_reg
);
2989 dst_reg
->umin_value
*= umin_val
;
2990 dst_reg
->umax_value
*= umax_val
;
2991 if (dst_reg
->umax_value
> S64_MAX
) {
2992 /* Overflow possible, we know nothing */
2993 dst_reg
->smin_value
= S64_MIN
;
2994 dst_reg
->smax_value
= S64_MAX
;
2996 dst_reg
->smin_value
= dst_reg
->umin_value
;
2997 dst_reg
->smax_value
= dst_reg
->umax_value
;
3001 if (src_known
&& dst_known
) {
3002 __mark_reg_known(dst_reg
, dst_reg
->var_off
.value
&
3003 src_reg
.var_off
.value
);
3006 /* We get our minimum from the var_off, since that's inherently
3007 * bitwise. Our maximum is the minimum of the operands' maxima.
3009 dst_reg
->var_off
= tnum_and(dst_reg
->var_off
, src_reg
.var_off
);
3010 dst_reg
->umin_value
= dst_reg
->var_off
.value
;
3011 dst_reg
->umax_value
= min(dst_reg
->umax_value
, umax_val
);
3012 if (dst_reg
->smin_value
< 0 || smin_val
< 0) {
3013 /* Lose signed bounds when ANDing negative numbers,
3014 * ain't nobody got time for that.
3016 dst_reg
->smin_value
= S64_MIN
;
3017 dst_reg
->smax_value
= S64_MAX
;
3019 /* ANDing two positives gives a positive, so safe to
3020 * cast result into s64.
3022 dst_reg
->smin_value
= dst_reg
->umin_value
;
3023 dst_reg
->smax_value
= dst_reg
->umax_value
;
3025 /* We may learn something more from the var_off */
3026 __update_reg_bounds(dst_reg
);
3029 if (src_known
&& dst_known
) {
3030 __mark_reg_known(dst_reg
, dst_reg
->var_off
.value
|
3031 src_reg
.var_off
.value
);
3034 /* We get our maximum from the var_off, and our minimum is the
3035 * maximum of the operands' minima
3037 dst_reg
->var_off
= tnum_or(dst_reg
->var_off
, src_reg
.var_off
);
3038 dst_reg
->umin_value
= max(dst_reg
->umin_value
, umin_val
);
3039 dst_reg
->umax_value
= dst_reg
->var_off
.value
|
3040 dst_reg
->var_off
.mask
;
3041 if (dst_reg
->smin_value
< 0 || smin_val
< 0) {
3042 /* Lose signed bounds when ORing negative numbers,
3043 * ain't nobody got time for that.
3045 dst_reg
->smin_value
= S64_MIN
;
3046 dst_reg
->smax_value
= S64_MAX
;
3048 /* ORing two positives gives a positive, so safe to
3049 * cast result into s64.
3051 dst_reg
->smin_value
= dst_reg
->umin_value
;
3052 dst_reg
->smax_value
= dst_reg
->umax_value
;
3054 /* We may learn something more from the var_off */
3055 __update_reg_bounds(dst_reg
);
3058 if (umax_val
>= insn_bitness
) {
3059 /* Shifts greater than 31 or 63 are undefined.
3060 * This includes shifts by a negative number.
3062 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
3065 /* We lose all sign bit information (except what we can pick
3068 dst_reg
->smin_value
= S64_MIN
;
3069 dst_reg
->smax_value
= S64_MAX
;
3070 /* If we might shift our top bit out, then we know nothing */
3071 if (dst_reg
->umax_value
> 1ULL << (63 - umax_val
)) {
3072 dst_reg
->umin_value
= 0;
3073 dst_reg
->umax_value
= U64_MAX
;
3075 dst_reg
->umin_value
<<= umin_val
;
3076 dst_reg
->umax_value
<<= umax_val
;
3078 dst_reg
->var_off
= tnum_lshift(dst_reg
->var_off
, umin_val
);
3079 /* We may learn something more from the var_off */
3080 __update_reg_bounds(dst_reg
);
3083 if (umax_val
>= insn_bitness
) {
3084 /* Shifts greater than 31 or 63 are undefined.
3085 * This includes shifts by a negative number.
3087 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
3090 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
3091 * be negative, then either:
3092 * 1) src_reg might be zero, so the sign bit of the result is
3093 * unknown, so we lose our signed bounds
3094 * 2) it's known negative, thus the unsigned bounds capture the
3096 * 3) the signed bounds cross zero, so they tell us nothing
3098 * If the value in dst_reg is known nonnegative, then again the
3099 * unsigned bounts capture the signed bounds.
3100 * Thus, in all cases it suffices to blow away our signed bounds
3101 * and rely on inferring new ones from the unsigned bounds and
3102 * var_off of the result.
3104 dst_reg
->smin_value
= S64_MIN
;
3105 dst_reg
->smax_value
= S64_MAX
;
3106 dst_reg
->var_off
= tnum_rshift(dst_reg
->var_off
, umin_val
);
3107 dst_reg
->umin_value
>>= umax_val
;
3108 dst_reg
->umax_value
>>= umin_val
;
3109 /* We may learn something more from the var_off */
3110 __update_reg_bounds(dst_reg
);
3113 if (umax_val
>= insn_bitness
) {
3114 /* Shifts greater than 31 or 63 are undefined.
3115 * This includes shifts by a negative number.
3117 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
3121 /* Upon reaching here, src_known is true and
3122 * umax_val is equal to umin_val.
3124 dst_reg
->smin_value
>>= umin_val
;
3125 dst_reg
->smax_value
>>= umin_val
;
3126 dst_reg
->var_off
= tnum_arshift(dst_reg
->var_off
, umin_val
);
3128 /* blow away the dst_reg umin_value/umax_value and rely on
3129 * dst_reg var_off to refine the result.
3131 dst_reg
->umin_value
= 0;
3132 dst_reg
->umax_value
= U64_MAX
;
3133 __update_reg_bounds(dst_reg
);
3136 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
3140 if (BPF_CLASS(insn
->code
) != BPF_ALU64
) {
3141 /* 32-bit ALU ops are (32,32)->32 */
3142 coerce_reg_to_size(dst_reg
, 4);
3145 __reg_deduce_bounds(dst_reg
);
3146 __reg_bound_offset(dst_reg
);
3150 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
3153 static int adjust_reg_min_max_vals(struct bpf_verifier_env
*env
,
3154 struct bpf_insn
*insn
)
3156 struct bpf_verifier_state
*vstate
= env
->cur_state
;
3157 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
3158 struct bpf_reg_state
*regs
= state
->regs
, *dst_reg
, *src_reg
;
3159 struct bpf_reg_state
*ptr_reg
= NULL
, off_reg
= {0};
3160 u8 opcode
= BPF_OP(insn
->code
);
3162 dst_reg
= ®s
[insn
->dst_reg
];
3164 if (dst_reg
->type
!= SCALAR_VALUE
)
3166 if (BPF_SRC(insn
->code
) == BPF_X
) {
3167 src_reg
= ®s
[insn
->src_reg
];
3168 if (src_reg
->type
!= SCALAR_VALUE
) {
3169 if (dst_reg
->type
!= SCALAR_VALUE
) {
3170 /* Combining two pointers by any ALU op yields
3171 * an arbitrary scalar. Disallow all math except
3172 * pointer subtraction
3174 if (opcode
== BPF_SUB
&& env
->allow_ptr_leaks
) {
3175 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
3178 verbose(env
, "R%d pointer %s pointer prohibited\n",
3180 bpf_alu_string
[opcode
>> 4]);
3183 /* scalar += pointer
3184 * This is legal, but we have to reverse our
3185 * src/dest handling in computing the range
3187 return adjust_ptr_min_max_vals(env
, insn
,
3190 } else if (ptr_reg
) {
3191 /* pointer += scalar */
3192 return adjust_ptr_min_max_vals(env
, insn
,
3196 /* Pretend the src is a reg with a known value, since we only
3197 * need to be able to read from this state.
3199 off_reg
.type
= SCALAR_VALUE
;
3200 __mark_reg_known(&off_reg
, insn
->imm
);
3202 if (ptr_reg
) /* pointer += K */
3203 return adjust_ptr_min_max_vals(env
, insn
,
3207 /* Got here implies adding two SCALAR_VALUEs */
3208 if (WARN_ON_ONCE(ptr_reg
)) {
3209 print_verifier_state(env
, state
);
3210 verbose(env
, "verifier internal error: unexpected ptr_reg\n");
3213 if (WARN_ON(!src_reg
)) {
3214 print_verifier_state(env
, state
);
3215 verbose(env
, "verifier internal error: no src_reg\n");
3218 return adjust_scalar_min_max_vals(env
, insn
, dst_reg
, *src_reg
);
3221 /* check validity of 32-bit and 64-bit arithmetic operations */
3222 static int check_alu_op(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
3224 struct bpf_reg_state
*regs
= cur_regs(env
);
3225 u8 opcode
= BPF_OP(insn
->code
);
3228 if (opcode
== BPF_END
|| opcode
== BPF_NEG
) {
3229 if (opcode
== BPF_NEG
) {
3230 if (BPF_SRC(insn
->code
) != 0 ||
3231 insn
->src_reg
!= BPF_REG_0
||
3232 insn
->off
!= 0 || insn
->imm
!= 0) {
3233 verbose(env
, "BPF_NEG uses reserved fields\n");
3237 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0 ||
3238 (insn
->imm
!= 16 && insn
->imm
!= 32 && insn
->imm
!= 64) ||
3239 BPF_CLASS(insn
->code
) == BPF_ALU64
) {
3240 verbose(env
, "BPF_END uses reserved fields\n");
3245 /* check src operand */
3246 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
3250 if (is_pointer_value(env
, insn
->dst_reg
)) {
3251 verbose(env
, "R%d pointer arithmetic prohibited\n",
3256 /* check dest operand */
3257 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP
);
3261 } else if (opcode
== BPF_MOV
) {
3263 if (BPF_SRC(insn
->code
) == BPF_X
) {
3264 if (insn
->imm
!= 0 || insn
->off
!= 0) {
3265 verbose(env
, "BPF_MOV uses reserved fields\n");
3269 /* check src operand */
3270 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
3274 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0) {
3275 verbose(env
, "BPF_MOV uses reserved fields\n");
3280 /* check dest operand, mark as required later */
3281 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP_NO_MARK
);
3285 if (BPF_SRC(insn
->code
) == BPF_X
) {
3286 if (BPF_CLASS(insn
->code
) == BPF_ALU64
) {
3288 * copy register state to dest reg
3290 regs
[insn
->dst_reg
] = regs
[insn
->src_reg
];
3291 regs
[insn
->dst_reg
].live
|= REG_LIVE_WRITTEN
;
3294 if (is_pointer_value(env
, insn
->src_reg
)) {
3296 "R%d partial copy of pointer\n",
3300 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
3301 coerce_reg_to_size(®s
[insn
->dst_reg
], 4);
3305 * remember the value we stored into this reg
3307 /* clear any state __mark_reg_known doesn't set */
3308 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
3309 regs
[insn
->dst_reg
].type
= SCALAR_VALUE
;
3310 if (BPF_CLASS(insn
->code
) == BPF_ALU64
) {
3311 __mark_reg_known(regs
+ insn
->dst_reg
,
3314 __mark_reg_known(regs
+ insn
->dst_reg
,
3319 } else if (opcode
> BPF_END
) {
3320 verbose(env
, "invalid BPF_ALU opcode %x\n", opcode
);
3323 } else { /* all other ALU ops: and, sub, xor, add, ... */
3325 if (BPF_SRC(insn
->code
) == BPF_X
) {
3326 if (insn
->imm
!= 0 || insn
->off
!= 0) {
3327 verbose(env
, "BPF_ALU uses reserved fields\n");
3330 /* check src1 operand */
3331 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
3335 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0) {
3336 verbose(env
, "BPF_ALU uses reserved fields\n");
3341 /* check src2 operand */
3342 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
3346 if ((opcode
== BPF_MOD
|| opcode
== BPF_DIV
) &&
3347 BPF_SRC(insn
->code
) == BPF_K
&& insn
->imm
== 0) {
3348 verbose(env
, "div by zero\n");
3352 if (opcode
== BPF_ARSH
&& BPF_CLASS(insn
->code
) != BPF_ALU64
) {
3353 verbose(env
, "BPF_ARSH not supported for 32 bit ALU\n");
3357 if ((opcode
== BPF_LSH
|| opcode
== BPF_RSH
||
3358 opcode
== BPF_ARSH
) && BPF_SRC(insn
->code
) == BPF_K
) {
3359 int size
= BPF_CLASS(insn
->code
) == BPF_ALU64
? 64 : 32;
3361 if (insn
->imm
< 0 || insn
->imm
>= size
) {
3362 verbose(env
, "invalid shift %d\n", insn
->imm
);
3367 /* check dest operand */
3368 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP_NO_MARK
);
3372 return adjust_reg_min_max_vals(env
, insn
);
3378 static void find_good_pkt_pointers(struct bpf_verifier_state
*vstate
,
3379 struct bpf_reg_state
*dst_reg
,
3380 enum bpf_reg_type type
,
3381 bool range_right_open
)
3383 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
3384 struct bpf_reg_state
*regs
= state
->regs
, *reg
;
3388 if (dst_reg
->off
< 0 ||
3389 (dst_reg
->off
== 0 && range_right_open
))
3390 /* This doesn't give us any range */
3393 if (dst_reg
->umax_value
> MAX_PACKET_OFF
||
3394 dst_reg
->umax_value
+ dst_reg
->off
> MAX_PACKET_OFF
)
3395 /* Risk of overflow. For instance, ptr + (1<<63) may be less
3396 * than pkt_end, but that's because it's also less than pkt.
3400 new_range
= dst_reg
->off
;
3401 if (range_right_open
)
3404 /* Examples for register markings:
3406 * pkt_data in dst register:
3410 * if (r2 > pkt_end) goto <handle exception>
3415 * if (r2 < pkt_end) goto <access okay>
3416 * <handle exception>
3419 * r2 == dst_reg, pkt_end == src_reg
3420 * r2=pkt(id=n,off=8,r=0)
3421 * r3=pkt(id=n,off=0,r=0)
3423 * pkt_data in src register:
3427 * if (pkt_end >= r2) goto <access okay>
3428 * <handle exception>
3432 * if (pkt_end <= r2) goto <handle exception>
3436 * pkt_end == dst_reg, r2 == src_reg
3437 * r2=pkt(id=n,off=8,r=0)
3438 * r3=pkt(id=n,off=0,r=0)
3440 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
3441 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
3442 * and [r3, r3 + 8-1) respectively is safe to access depending on
3446 /* If our ids match, then we must have the same max_value. And we
3447 * don't care about the other reg's fixed offset, since if it's too big
3448 * the range won't allow anything.
3449 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
3451 for (i
= 0; i
< MAX_BPF_REG
; i
++)
3452 if (regs
[i
].type
== type
&& regs
[i
].id
== dst_reg
->id
)
3453 /* keep the maximum range already checked */
3454 regs
[i
].range
= max(regs
[i
].range
, new_range
);
3456 for (j
= 0; j
<= vstate
->curframe
; j
++) {
3457 state
= vstate
->frame
[j
];
3458 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
3459 if (state
->stack
[i
].slot_type
[0] != STACK_SPILL
)
3461 reg
= &state
->stack
[i
].spilled_ptr
;
3462 if (reg
->type
== type
&& reg
->id
== dst_reg
->id
)
3463 reg
->range
= max(reg
->range
, new_range
);
3468 /* Adjusts the register min/max values in the case that the dst_reg is the
3469 * variable register that we are working on, and src_reg is a constant or we're
3470 * simply doing a BPF_K check.
3471 * In JEQ/JNE cases we also adjust the var_off values.
3473 static void reg_set_min_max(struct bpf_reg_state
*true_reg
,
3474 struct bpf_reg_state
*false_reg
, u64 val
,
3477 /* If the dst_reg is a pointer, we can't learn anything about its
3478 * variable offset from the compare (unless src_reg were a pointer into
3479 * the same object, but we don't bother with that.
3480 * Since false_reg and true_reg have the same type by construction, we
3481 * only need to check one of them for pointerness.
3483 if (__is_pointer_value(false, false_reg
))
3488 /* If this is false then we know nothing Jon Snow, but if it is
3489 * true then we know for sure.
3491 __mark_reg_known(true_reg
, val
);
3494 /* If this is true we know nothing Jon Snow, but if it is false
3495 * we know the value for sure;
3497 __mark_reg_known(false_reg
, val
);
3500 false_reg
->umax_value
= min(false_reg
->umax_value
, val
);
3501 true_reg
->umin_value
= max(true_reg
->umin_value
, val
+ 1);
3504 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
);
3505 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
+ 1);
3508 false_reg
->umin_value
= max(false_reg
->umin_value
, val
);
3509 true_reg
->umax_value
= min(true_reg
->umax_value
, val
- 1);
3512 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
);
3513 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
- 1);
3516 false_reg
->umax_value
= min(false_reg
->umax_value
, val
- 1);
3517 true_reg
->umin_value
= max(true_reg
->umin_value
, val
);
3520 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
- 1);
3521 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
);
3524 false_reg
->umin_value
= max(false_reg
->umin_value
, val
+ 1);
3525 true_reg
->umax_value
= min(true_reg
->umax_value
, val
);
3528 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
+ 1);
3529 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
);
3535 __reg_deduce_bounds(false_reg
);
3536 __reg_deduce_bounds(true_reg
);
3537 /* We might have learned some bits from the bounds. */
3538 __reg_bound_offset(false_reg
);
3539 __reg_bound_offset(true_reg
);
3540 /* Intersecting with the old var_off might have improved our bounds
3541 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3542 * then new var_off is (0; 0x7f...fc) which improves our umax.
3544 __update_reg_bounds(false_reg
);
3545 __update_reg_bounds(true_reg
);
3548 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
3551 static void reg_set_min_max_inv(struct bpf_reg_state
*true_reg
,
3552 struct bpf_reg_state
*false_reg
, u64 val
,
3555 if (__is_pointer_value(false, false_reg
))
3560 /* If this is false then we know nothing Jon Snow, but if it is
3561 * true then we know for sure.
3563 __mark_reg_known(true_reg
, val
);
3566 /* If this is true we know nothing Jon Snow, but if it is false
3567 * we know the value for sure;
3569 __mark_reg_known(false_reg
, val
);
3572 true_reg
->umax_value
= min(true_reg
->umax_value
, val
- 1);
3573 false_reg
->umin_value
= max(false_reg
->umin_value
, val
);
3576 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
- 1);
3577 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
);
3580 true_reg
->umin_value
= max(true_reg
->umin_value
, val
+ 1);
3581 false_reg
->umax_value
= min(false_reg
->umax_value
, val
);
3584 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
+ 1);
3585 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
);
3588 true_reg
->umax_value
= min(true_reg
->umax_value
, val
);
3589 false_reg
->umin_value
= max(false_reg
->umin_value
, val
+ 1);
3592 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
);
3593 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
+ 1);
3596 true_reg
->umin_value
= max(true_reg
->umin_value
, val
);
3597 false_reg
->umax_value
= min(false_reg
->umax_value
, val
- 1);
3600 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
);
3601 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
- 1);
3607 __reg_deduce_bounds(false_reg
);
3608 __reg_deduce_bounds(true_reg
);
3609 /* We might have learned some bits from the bounds. */
3610 __reg_bound_offset(false_reg
);
3611 __reg_bound_offset(true_reg
);
3612 /* Intersecting with the old var_off might have improved our bounds
3613 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3614 * then new var_off is (0; 0x7f...fc) which improves our umax.
3616 __update_reg_bounds(false_reg
);
3617 __update_reg_bounds(true_reg
);
3620 /* Regs are known to be equal, so intersect their min/max/var_off */
3621 static void __reg_combine_min_max(struct bpf_reg_state
*src_reg
,
3622 struct bpf_reg_state
*dst_reg
)
3624 src_reg
->umin_value
= dst_reg
->umin_value
= max(src_reg
->umin_value
,
3625 dst_reg
->umin_value
);
3626 src_reg
->umax_value
= dst_reg
->umax_value
= min(src_reg
->umax_value
,
3627 dst_reg
->umax_value
);
3628 src_reg
->smin_value
= dst_reg
->smin_value
= max(src_reg
->smin_value
,
3629 dst_reg
->smin_value
);
3630 src_reg
->smax_value
= dst_reg
->smax_value
= min(src_reg
->smax_value
,
3631 dst_reg
->smax_value
);
3632 src_reg
->var_off
= dst_reg
->var_off
= tnum_intersect(src_reg
->var_off
,
3634 /* We might have learned new bounds from the var_off. */
3635 __update_reg_bounds(src_reg
);
3636 __update_reg_bounds(dst_reg
);
3637 /* We might have learned something about the sign bit. */
3638 __reg_deduce_bounds(src_reg
);
3639 __reg_deduce_bounds(dst_reg
);
3640 /* We might have learned some bits from the bounds. */
3641 __reg_bound_offset(src_reg
);
3642 __reg_bound_offset(dst_reg
);
3643 /* Intersecting with the old var_off might have improved our bounds
3644 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3645 * then new var_off is (0; 0x7f...fc) which improves our umax.
3647 __update_reg_bounds(src_reg
);
3648 __update_reg_bounds(dst_reg
);
3651 static void reg_combine_min_max(struct bpf_reg_state
*true_src
,
3652 struct bpf_reg_state
*true_dst
,
3653 struct bpf_reg_state
*false_src
,
3654 struct bpf_reg_state
*false_dst
,
3659 __reg_combine_min_max(true_src
, true_dst
);
3662 __reg_combine_min_max(false_src
, false_dst
);
3667 static void mark_map_reg(struct bpf_reg_state
*regs
, u32 regno
, u32 id
,
3670 struct bpf_reg_state
*reg
= ®s
[regno
];
3672 if (reg
->type
== PTR_TO_MAP_VALUE_OR_NULL
&& reg
->id
== id
) {
3673 /* Old offset (both fixed and variable parts) should
3674 * have been known-zero, because we don't allow pointer
3675 * arithmetic on pointers that might be NULL.
3677 if (WARN_ON_ONCE(reg
->smin_value
|| reg
->smax_value
||
3678 !tnum_equals_const(reg
->var_off
, 0) ||
3680 __mark_reg_known_zero(reg
);
3684 reg
->type
= SCALAR_VALUE
;
3685 } else if (reg
->map_ptr
->inner_map_meta
) {
3686 reg
->type
= CONST_PTR_TO_MAP
;
3687 reg
->map_ptr
= reg
->map_ptr
->inner_map_meta
;
3689 reg
->type
= PTR_TO_MAP_VALUE
;
3691 /* We don't need id from this point onwards anymore, thus we
3692 * should better reset it, so that state pruning has chances
3699 /* The logic is similar to find_good_pkt_pointers(), both could eventually
3700 * be folded together at some point.
3702 static void mark_map_regs(struct bpf_verifier_state
*vstate
, u32 regno
,
3705 struct bpf_func_state
*state
= vstate
->frame
[vstate
->curframe
];
3706 struct bpf_reg_state
*regs
= state
->regs
;
3707 u32 id
= regs
[regno
].id
;
3710 for (i
= 0; i
< MAX_BPF_REG
; i
++)
3711 mark_map_reg(regs
, i
, id
, is_null
);
3713 for (j
= 0; j
<= vstate
->curframe
; j
++) {
3714 state
= vstate
->frame
[j
];
3715 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
3716 if (state
->stack
[i
].slot_type
[0] != STACK_SPILL
)
3718 mark_map_reg(&state
->stack
[i
].spilled_ptr
, 0, id
, is_null
);
3723 static bool try_match_pkt_pointers(const struct bpf_insn
*insn
,
3724 struct bpf_reg_state
*dst_reg
,
3725 struct bpf_reg_state
*src_reg
,
3726 struct bpf_verifier_state
*this_branch
,
3727 struct bpf_verifier_state
*other_branch
)
3729 if (BPF_SRC(insn
->code
) != BPF_X
)
3732 switch (BPF_OP(insn
->code
)) {
3734 if ((dst_reg
->type
== PTR_TO_PACKET
&&
3735 src_reg
->type
== PTR_TO_PACKET_END
) ||
3736 (dst_reg
->type
== PTR_TO_PACKET_META
&&
3737 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
3738 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
3739 find_good_pkt_pointers(this_branch
, dst_reg
,
3740 dst_reg
->type
, false);
3741 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
3742 src_reg
->type
== PTR_TO_PACKET
) ||
3743 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
3744 src_reg
->type
== PTR_TO_PACKET_META
)) {
3745 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
3746 find_good_pkt_pointers(other_branch
, src_reg
,
3747 src_reg
->type
, true);
3753 if ((dst_reg
->type
== PTR_TO_PACKET
&&
3754 src_reg
->type
== PTR_TO_PACKET_END
) ||
3755 (dst_reg
->type
== PTR_TO_PACKET_META
&&
3756 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
3757 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
3758 find_good_pkt_pointers(other_branch
, dst_reg
,
3759 dst_reg
->type
, true);
3760 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
3761 src_reg
->type
== PTR_TO_PACKET
) ||
3762 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
3763 src_reg
->type
== PTR_TO_PACKET_META
)) {
3764 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
3765 find_good_pkt_pointers(this_branch
, src_reg
,
3766 src_reg
->type
, false);
3772 if ((dst_reg
->type
== PTR_TO_PACKET
&&
3773 src_reg
->type
== PTR_TO_PACKET_END
) ||
3774 (dst_reg
->type
== PTR_TO_PACKET_META
&&
3775 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
3776 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
3777 find_good_pkt_pointers(this_branch
, dst_reg
,
3778 dst_reg
->type
, true);
3779 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
3780 src_reg
->type
== PTR_TO_PACKET
) ||
3781 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
3782 src_reg
->type
== PTR_TO_PACKET_META
)) {
3783 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
3784 find_good_pkt_pointers(other_branch
, src_reg
,
3785 src_reg
->type
, false);
3791 if ((dst_reg
->type
== PTR_TO_PACKET
&&
3792 src_reg
->type
== PTR_TO_PACKET_END
) ||
3793 (dst_reg
->type
== PTR_TO_PACKET_META
&&
3794 reg_is_init_pkt_pointer(src_reg
, PTR_TO_PACKET
))) {
3795 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
3796 find_good_pkt_pointers(other_branch
, dst_reg
,
3797 dst_reg
->type
, false);
3798 } else if ((dst_reg
->type
== PTR_TO_PACKET_END
&&
3799 src_reg
->type
== PTR_TO_PACKET
) ||
3800 (reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
3801 src_reg
->type
== PTR_TO_PACKET_META
)) {
3802 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
3803 find_good_pkt_pointers(this_branch
, src_reg
,
3804 src_reg
->type
, true);
3816 static int check_cond_jmp_op(struct bpf_verifier_env
*env
,
3817 struct bpf_insn
*insn
, int *insn_idx
)
3819 struct bpf_verifier_state
*this_branch
= env
->cur_state
;
3820 struct bpf_verifier_state
*other_branch
;
3821 struct bpf_reg_state
*regs
= this_branch
->frame
[this_branch
->curframe
]->regs
;
3822 struct bpf_reg_state
*dst_reg
, *other_branch_regs
;
3823 u8 opcode
= BPF_OP(insn
->code
);
3826 if (opcode
> BPF_JSLE
) {
3827 verbose(env
, "invalid BPF_JMP opcode %x\n", opcode
);
3831 if (BPF_SRC(insn
->code
) == BPF_X
) {
3832 if (insn
->imm
!= 0) {
3833 verbose(env
, "BPF_JMP uses reserved fields\n");
3837 /* check src1 operand */
3838 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
3842 if (is_pointer_value(env
, insn
->src_reg
)) {
3843 verbose(env
, "R%d pointer comparison prohibited\n",
3848 if (insn
->src_reg
!= BPF_REG_0
) {
3849 verbose(env
, "BPF_JMP uses reserved fields\n");
3854 /* check src2 operand */
3855 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
3859 dst_reg
= ®s
[insn
->dst_reg
];
3861 /* detect if R == 0 where R was initialized to zero earlier */
3862 if (BPF_SRC(insn
->code
) == BPF_K
&&
3863 (opcode
== BPF_JEQ
|| opcode
== BPF_JNE
) &&
3864 dst_reg
->type
== SCALAR_VALUE
&&
3865 tnum_is_const(dst_reg
->var_off
)) {
3866 if ((opcode
== BPF_JEQ
&& dst_reg
->var_off
.value
== insn
->imm
) ||
3867 (opcode
== BPF_JNE
&& dst_reg
->var_off
.value
!= insn
->imm
)) {
3868 /* if (imm == imm) goto pc+off;
3869 * only follow the goto, ignore fall-through
3871 *insn_idx
+= insn
->off
;
3874 /* if (imm != imm) goto pc+off;
3875 * only follow fall-through branch, since
3876 * that's where the program will go
3882 other_branch
= push_stack(env
, *insn_idx
+ insn
->off
+ 1, *insn_idx
);
3885 other_branch_regs
= other_branch
->frame
[other_branch
->curframe
]->regs
;
3887 /* detect if we are comparing against a constant value so we can adjust
3888 * our min/max values for our dst register.
3889 * this is only legit if both are scalars (or pointers to the same
3890 * object, I suppose, but we don't support that right now), because
3891 * otherwise the different base pointers mean the offsets aren't
3894 if (BPF_SRC(insn
->code
) == BPF_X
) {
3895 if (dst_reg
->type
== SCALAR_VALUE
&&
3896 regs
[insn
->src_reg
].type
== SCALAR_VALUE
) {
3897 if (tnum_is_const(regs
[insn
->src_reg
].var_off
))
3898 reg_set_min_max(&other_branch_regs
[insn
->dst_reg
],
3899 dst_reg
, regs
[insn
->src_reg
].var_off
.value
,
3901 else if (tnum_is_const(dst_reg
->var_off
))
3902 reg_set_min_max_inv(&other_branch_regs
[insn
->src_reg
],
3903 ®s
[insn
->src_reg
],
3904 dst_reg
->var_off
.value
, opcode
);
3905 else if (opcode
== BPF_JEQ
|| opcode
== BPF_JNE
)
3906 /* Comparing for equality, we can combine knowledge */
3907 reg_combine_min_max(&other_branch_regs
[insn
->src_reg
],
3908 &other_branch_regs
[insn
->dst_reg
],
3909 ®s
[insn
->src_reg
],
3910 ®s
[insn
->dst_reg
], opcode
);
3912 } else if (dst_reg
->type
== SCALAR_VALUE
) {
3913 reg_set_min_max(&other_branch_regs
[insn
->dst_reg
],
3914 dst_reg
, insn
->imm
, opcode
);
3917 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
3918 if (BPF_SRC(insn
->code
) == BPF_K
&&
3919 insn
->imm
== 0 && (opcode
== BPF_JEQ
|| opcode
== BPF_JNE
) &&
3920 dst_reg
->type
== PTR_TO_MAP_VALUE_OR_NULL
) {
3921 /* Mark all identical map registers in each branch as either
3922 * safe or unknown depending R == 0 or R != 0 conditional.
3924 mark_map_regs(this_branch
, insn
->dst_reg
, opcode
== BPF_JNE
);
3925 mark_map_regs(other_branch
, insn
->dst_reg
, opcode
== BPF_JEQ
);
3926 } else if (!try_match_pkt_pointers(insn
, dst_reg
, ®s
[insn
->src_reg
],
3927 this_branch
, other_branch
) &&
3928 is_pointer_value(env
, insn
->dst_reg
)) {
3929 verbose(env
, "R%d pointer comparison prohibited\n",
3934 print_verifier_state(env
, this_branch
->frame
[this_branch
->curframe
]);
3938 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
3939 static struct bpf_map
*ld_imm64_to_map_ptr(struct bpf_insn
*insn
)
3941 u64 imm64
= ((u64
) (u32
) insn
[0].imm
) | ((u64
) (u32
) insn
[1].imm
) << 32;
3943 return (struct bpf_map
*) (unsigned long) imm64
;
3946 /* verify BPF_LD_IMM64 instruction */
3947 static int check_ld_imm(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
3949 struct bpf_reg_state
*regs
= cur_regs(env
);
3952 if (BPF_SIZE(insn
->code
) != BPF_DW
) {
3953 verbose(env
, "invalid BPF_LD_IMM insn\n");
3956 if (insn
->off
!= 0) {
3957 verbose(env
, "BPF_LD_IMM64 uses reserved fields\n");
3961 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP
);
3965 if (insn
->src_reg
== 0) {
3966 u64 imm
= ((u64
)(insn
+ 1)->imm
<< 32) | (u32
)insn
->imm
;
3968 regs
[insn
->dst_reg
].type
= SCALAR_VALUE
;
3969 __mark_reg_known(®s
[insn
->dst_reg
], imm
);
3973 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
3974 BUG_ON(insn
->src_reg
!= BPF_PSEUDO_MAP_FD
);
3976 regs
[insn
->dst_reg
].type
= CONST_PTR_TO_MAP
;
3977 regs
[insn
->dst_reg
].map_ptr
= ld_imm64_to_map_ptr(insn
);
3981 static bool may_access_skb(enum bpf_prog_type type
)
3984 case BPF_PROG_TYPE_SOCKET_FILTER
:
3985 case BPF_PROG_TYPE_SCHED_CLS
:
3986 case BPF_PROG_TYPE_SCHED_ACT
:
3993 /* verify safety of LD_ABS|LD_IND instructions:
3994 * - they can only appear in the programs where ctx == skb
3995 * - since they are wrappers of function calls, they scratch R1-R5 registers,
3996 * preserve R6-R9, and store return value into R0
3999 * ctx == skb == R6 == CTX
4002 * SRC == any register
4003 * IMM == 32-bit immediate
4006 * R0 - 8/16/32-bit skb data converted to cpu endianness
4008 static int check_ld_abs(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
4010 struct bpf_reg_state
*regs
= cur_regs(env
);
4011 u8 mode
= BPF_MODE(insn
->code
);
4014 if (!may_access_skb(env
->prog
->type
)) {
4015 verbose(env
, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
4019 if (!env
->ops
->gen_ld_abs
) {
4020 verbose(env
, "bpf verifier is misconfigured\n");
4024 if (env
->subprog_cnt
> 1) {
4025 /* when program has LD_ABS insn JITs and interpreter assume
4026 * that r1 == ctx == skb which is not the case for callees
4027 * that can have arbitrary arguments. It's problematic
4028 * for main prog as well since JITs would need to analyze
4029 * all functions in order to make proper register save/restore
4030 * decisions in the main prog. Hence disallow LD_ABS with calls
4032 verbose(env
, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
4036 if (insn
->dst_reg
!= BPF_REG_0
|| insn
->off
!= 0 ||
4037 BPF_SIZE(insn
->code
) == BPF_DW
||
4038 (mode
== BPF_ABS
&& insn
->src_reg
!= BPF_REG_0
)) {
4039 verbose(env
, "BPF_LD_[ABS|IND] uses reserved fields\n");
4043 /* check whether implicit source operand (register R6) is readable */
4044 err
= check_reg_arg(env
, BPF_REG_6
, SRC_OP
);
4048 if (regs
[BPF_REG_6
].type
!= PTR_TO_CTX
) {
4050 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
4054 if (mode
== BPF_IND
) {
4055 /* check explicit source operand */
4056 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
4061 /* reset caller saved regs to unreadable */
4062 for (i
= 0; i
< CALLER_SAVED_REGS
; i
++) {
4063 mark_reg_not_init(env
, regs
, caller_saved
[i
]);
4064 check_reg_arg(env
, caller_saved
[i
], DST_OP_NO_MARK
);
4067 /* mark destination R0 register as readable, since it contains
4068 * the value fetched from the packet.
4069 * Already marked as written above.
4071 mark_reg_unknown(env
, regs
, BPF_REG_0
);
4075 static int check_return_code(struct bpf_verifier_env
*env
)
4077 struct bpf_reg_state
*reg
;
4078 struct tnum range
= tnum_range(0, 1);
4080 switch (env
->prog
->type
) {
4081 case BPF_PROG_TYPE_CGROUP_SKB
:
4082 case BPF_PROG_TYPE_CGROUP_SOCK
:
4083 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR
:
4084 case BPF_PROG_TYPE_SOCK_OPS
:
4085 case BPF_PROG_TYPE_CGROUP_DEVICE
:
4091 reg
= cur_regs(env
) + BPF_REG_0
;
4092 if (reg
->type
!= SCALAR_VALUE
) {
4093 verbose(env
, "At program exit the register R0 is not a known value (%s)\n",
4094 reg_type_str
[reg
->type
]);
4098 if (!tnum_in(range
, reg
->var_off
)) {
4099 verbose(env
, "At program exit the register R0 ");
4100 if (!tnum_is_unknown(reg
->var_off
)) {
4103 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
4104 verbose(env
, "has value %s", tn_buf
);
4106 verbose(env
, "has unknown scalar value");
4108 verbose(env
, " should have been 0 or 1\n");
4114 /* non-recursive DFS pseudo code
4115 * 1 procedure DFS-iterative(G,v):
4116 * 2 label v as discovered
4117 * 3 let S be a stack
4119 * 5 while S is not empty
4121 * 7 if t is what we're looking for:
4123 * 9 for all edges e in G.adjacentEdges(t) do
4124 * 10 if edge e is already labelled
4125 * 11 continue with the next edge
4126 * 12 w <- G.adjacentVertex(t,e)
4127 * 13 if vertex w is not discovered and not explored
4128 * 14 label e as tree-edge
4129 * 15 label w as discovered
4132 * 18 else if vertex w is discovered
4133 * 19 label e as back-edge
4135 * 21 // vertex w is explored
4136 * 22 label e as forward- or cross-edge
4137 * 23 label t as explored
4142 * 0x11 - discovered and fall-through edge labelled
4143 * 0x12 - discovered and fall-through and branch edges labelled
4154 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
4156 static int *insn_stack
; /* stack of insns to process */
4157 static int cur_stack
; /* current stack index */
4158 static int *insn_state
;
4160 /* t, w, e - match pseudo-code above:
4161 * t - index of current instruction
4162 * w - next instruction
4165 static int push_insn(int t
, int w
, int e
, struct bpf_verifier_env
*env
)
4167 if (e
== FALLTHROUGH
&& insn_state
[t
] >= (DISCOVERED
| FALLTHROUGH
))
4170 if (e
== BRANCH
&& insn_state
[t
] >= (DISCOVERED
| BRANCH
))
4173 if (w
< 0 || w
>= env
->prog
->len
) {
4174 verbose(env
, "jump out of range from insn %d to %d\n", t
, w
);
4179 /* mark branch target for state pruning */
4180 env
->explored_states
[w
] = STATE_LIST_MARK
;
4182 if (insn_state
[w
] == 0) {
4184 insn_state
[t
] = DISCOVERED
| e
;
4185 insn_state
[w
] = DISCOVERED
;
4186 if (cur_stack
>= env
->prog
->len
)
4188 insn_stack
[cur_stack
++] = w
;
4190 } else if ((insn_state
[w
] & 0xF0) == DISCOVERED
) {
4191 verbose(env
, "back-edge from insn %d to %d\n", t
, w
);
4193 } else if (insn_state
[w
] == EXPLORED
) {
4194 /* forward- or cross-edge */
4195 insn_state
[t
] = DISCOVERED
| e
;
4197 verbose(env
, "insn state internal bug\n");
4203 /* non-recursive depth-first-search to detect loops in BPF program
4204 * loop == back-edge in directed graph
4206 static int check_cfg(struct bpf_verifier_env
*env
)
4208 struct bpf_insn
*insns
= env
->prog
->insnsi
;
4209 int insn_cnt
= env
->prog
->len
;
4213 ret
= check_subprogs(env
);
4217 insn_state
= kcalloc(insn_cnt
, sizeof(int), GFP_KERNEL
);
4221 insn_stack
= kcalloc(insn_cnt
, sizeof(int), GFP_KERNEL
);
4227 insn_state
[0] = DISCOVERED
; /* mark 1st insn as discovered */
4228 insn_stack
[0] = 0; /* 0 is the first instruction */
4234 t
= insn_stack
[cur_stack
- 1];
4236 if (BPF_CLASS(insns
[t
].code
) == BPF_JMP
) {
4237 u8 opcode
= BPF_OP(insns
[t
].code
);
4239 if (opcode
== BPF_EXIT
) {
4241 } else if (opcode
== BPF_CALL
) {
4242 ret
= push_insn(t
, t
+ 1, FALLTHROUGH
, env
);
4247 if (t
+ 1 < insn_cnt
)
4248 env
->explored_states
[t
+ 1] = STATE_LIST_MARK
;
4249 if (insns
[t
].src_reg
== BPF_PSEUDO_CALL
) {
4250 env
->explored_states
[t
] = STATE_LIST_MARK
;
4251 ret
= push_insn(t
, t
+ insns
[t
].imm
+ 1, BRANCH
, env
);
4257 } else if (opcode
== BPF_JA
) {
4258 if (BPF_SRC(insns
[t
].code
) != BPF_K
) {
4262 /* unconditional jump with single edge */
4263 ret
= push_insn(t
, t
+ insns
[t
].off
+ 1,
4269 /* tell verifier to check for equivalent states
4270 * after every call and jump
4272 if (t
+ 1 < insn_cnt
)
4273 env
->explored_states
[t
+ 1] = STATE_LIST_MARK
;
4275 /* conditional jump with two edges */
4276 env
->explored_states
[t
] = STATE_LIST_MARK
;
4277 ret
= push_insn(t
, t
+ 1, FALLTHROUGH
, env
);
4283 ret
= push_insn(t
, t
+ insns
[t
].off
+ 1, BRANCH
, env
);
4290 /* all other non-branch instructions with single
4293 ret
= push_insn(t
, t
+ 1, FALLTHROUGH
, env
);
4301 insn_state
[t
] = EXPLORED
;
4302 if (cur_stack
-- <= 0) {
4303 verbose(env
, "pop stack internal bug\n");
4310 for (i
= 0; i
< insn_cnt
; i
++) {
4311 if (insn_state
[i
] != EXPLORED
) {
4312 verbose(env
, "unreachable insn %d\n", i
);
4317 ret
= 0; /* cfg looks good */
4325 /* check %cur's range satisfies %old's */
4326 static bool range_within(struct bpf_reg_state
*old
,
4327 struct bpf_reg_state
*cur
)
4329 return old
->umin_value
<= cur
->umin_value
&&
4330 old
->umax_value
>= cur
->umax_value
&&
4331 old
->smin_value
<= cur
->smin_value
&&
4332 old
->smax_value
>= cur
->smax_value
;
4335 /* Maximum number of register states that can exist at once */
4336 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
4342 /* If in the old state two registers had the same id, then they need to have
4343 * the same id in the new state as well. But that id could be different from
4344 * the old state, so we need to track the mapping from old to new ids.
4345 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
4346 * regs with old id 5 must also have new id 9 for the new state to be safe. But
4347 * regs with a different old id could still have new id 9, we don't care about
4349 * So we look through our idmap to see if this old id has been seen before. If
4350 * so, we require the new id to match; otherwise, we add the id pair to the map.
4352 static bool check_ids(u32 old_id
, u32 cur_id
, struct idpair
*idmap
)
4356 for (i
= 0; i
< ID_MAP_SIZE
; i
++) {
4357 if (!idmap
[i
].old
) {
4358 /* Reached an empty slot; haven't seen this id before */
4359 idmap
[i
].old
= old_id
;
4360 idmap
[i
].cur
= cur_id
;
4363 if (idmap
[i
].old
== old_id
)
4364 return idmap
[i
].cur
== cur_id
;
4366 /* We ran out of idmap slots, which should be impossible */
4371 /* Returns true if (rold safe implies rcur safe) */
4372 static bool regsafe(struct bpf_reg_state
*rold
, struct bpf_reg_state
*rcur
,
4373 struct idpair
*idmap
)
4377 if (!(rold
->live
& REG_LIVE_READ
))
4378 /* explored state didn't use this */
4381 equal
= memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, frameno
)) == 0;
4383 if (rold
->type
== PTR_TO_STACK
)
4384 /* two stack pointers are equal only if they're pointing to
4385 * the same stack frame, since fp-8 in foo != fp-8 in bar
4387 return equal
&& rold
->frameno
== rcur
->frameno
;
4392 if (rold
->type
== NOT_INIT
)
4393 /* explored state can't have used this */
4395 if (rcur
->type
== NOT_INIT
)
4397 switch (rold
->type
) {
4399 if (rcur
->type
== SCALAR_VALUE
) {
4400 /* new val must satisfy old val knowledge */
4401 return range_within(rold
, rcur
) &&
4402 tnum_in(rold
->var_off
, rcur
->var_off
);
4404 /* We're trying to use a pointer in place of a scalar.
4405 * Even if the scalar was unbounded, this could lead to
4406 * pointer leaks because scalars are allowed to leak
4407 * while pointers are not. We could make this safe in
4408 * special cases if root is calling us, but it's
4409 * probably not worth the hassle.
4413 case PTR_TO_MAP_VALUE
:
4414 /* If the new min/max/var_off satisfy the old ones and
4415 * everything else matches, we are OK.
4416 * We don't care about the 'id' value, because nothing
4417 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
4419 return memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, id
)) == 0 &&
4420 range_within(rold
, rcur
) &&
4421 tnum_in(rold
->var_off
, rcur
->var_off
);
4422 case PTR_TO_MAP_VALUE_OR_NULL
:
4423 /* a PTR_TO_MAP_VALUE could be safe to use as a
4424 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
4425 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
4426 * checked, doing so could have affected others with the same
4427 * id, and we can't check for that because we lost the id when
4428 * we converted to a PTR_TO_MAP_VALUE.
4430 if (rcur
->type
!= PTR_TO_MAP_VALUE_OR_NULL
)
4432 if (memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, id
)))
4434 /* Check our ids match any regs they're supposed to */
4435 return check_ids(rold
->id
, rcur
->id
, idmap
);
4436 case PTR_TO_PACKET_META
:
4438 if (rcur
->type
!= rold
->type
)
4440 /* We must have at least as much range as the old ptr
4441 * did, so that any accesses which were safe before are
4442 * still safe. This is true even if old range < old off,
4443 * since someone could have accessed through (ptr - k), or
4444 * even done ptr -= k in a register, to get a safe access.
4446 if (rold
->range
> rcur
->range
)
4448 /* If the offsets don't match, we can't trust our alignment;
4449 * nor can we be sure that we won't fall out of range.
4451 if (rold
->off
!= rcur
->off
)
4453 /* id relations must be preserved */
4454 if (rold
->id
&& !check_ids(rold
->id
, rcur
->id
, idmap
))
4456 /* new val must satisfy old val knowledge */
4457 return range_within(rold
, rcur
) &&
4458 tnum_in(rold
->var_off
, rcur
->var_off
);
4460 case CONST_PTR_TO_MAP
:
4461 case PTR_TO_PACKET_END
:
4462 /* Only valid matches are exact, which memcmp() above
4463 * would have accepted
4466 /* Don't know what's going on, just say it's not safe */
4470 /* Shouldn't get here; if we do, say it's not safe */
4475 static bool stacksafe(struct bpf_func_state
*old
,
4476 struct bpf_func_state
*cur
,
4477 struct idpair
*idmap
)
4481 /* if explored stack has more populated slots than current stack
4482 * such stacks are not equivalent
4484 if (old
->allocated_stack
> cur
->allocated_stack
)
4487 /* walk slots of the explored stack and ignore any additional
4488 * slots in the current stack, since explored(safe) state
4491 for (i
= 0; i
< old
->allocated_stack
; i
++) {
4492 spi
= i
/ BPF_REG_SIZE
;
4494 if (!(old
->stack
[spi
].spilled_ptr
.live
& REG_LIVE_READ
))
4495 /* explored state didn't use this */
4498 if (old
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
] == STACK_INVALID
)
4500 /* if old state was safe with misc data in the stack
4501 * it will be safe with zero-initialized stack.
4502 * The opposite is not true
4504 if (old
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
] == STACK_MISC
&&
4505 cur
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
] == STACK_ZERO
)
4507 if (old
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
] !=
4508 cur
->stack
[spi
].slot_type
[i
% BPF_REG_SIZE
])
4509 /* Ex: old explored (safe) state has STACK_SPILL in
4510 * this stack slot, but current has has STACK_MISC ->
4511 * this verifier states are not equivalent,
4512 * return false to continue verification of this path
4515 if (i
% BPF_REG_SIZE
)
4517 if (old
->stack
[spi
].slot_type
[0] != STACK_SPILL
)
4519 if (!regsafe(&old
->stack
[spi
].spilled_ptr
,
4520 &cur
->stack
[spi
].spilled_ptr
,
4522 /* when explored and current stack slot are both storing
4523 * spilled registers, check that stored pointers types
4524 * are the same as well.
4525 * Ex: explored safe path could have stored
4526 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
4527 * but current path has stored:
4528 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
4529 * such verifier states are not equivalent.
4530 * return false to continue verification of this path
4537 /* compare two verifier states
4539 * all states stored in state_list are known to be valid, since
4540 * verifier reached 'bpf_exit' instruction through them
4542 * this function is called when verifier exploring different branches of
4543 * execution popped from the state stack. If it sees an old state that has
4544 * more strict register state and more strict stack state then this execution
4545 * branch doesn't need to be explored further, since verifier already
4546 * concluded that more strict state leads to valid finish.
4548 * Therefore two states are equivalent if register state is more conservative
4549 * and explored stack state is more conservative than the current one.
4552 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
4553 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
4555 * In other words if current stack state (one being explored) has more
4556 * valid slots than old one that already passed validation, it means
4557 * the verifier can stop exploring and conclude that current state is valid too
4559 * Similarly with registers. If explored state has register type as invalid
4560 * whereas register type in current state is meaningful, it means that
4561 * the current state will reach 'bpf_exit' instruction safely
4563 static bool func_states_equal(struct bpf_func_state
*old
,
4564 struct bpf_func_state
*cur
)
4566 struct idpair
*idmap
;
4570 idmap
= kcalloc(ID_MAP_SIZE
, sizeof(struct idpair
), GFP_KERNEL
);
4571 /* If we failed to allocate the idmap, just say it's not safe */
4575 for (i
= 0; i
< MAX_BPF_REG
; i
++) {
4576 if (!regsafe(&old
->regs
[i
], &cur
->regs
[i
], idmap
))
4580 if (!stacksafe(old
, cur
, idmap
))
4588 static bool states_equal(struct bpf_verifier_env
*env
,
4589 struct bpf_verifier_state
*old
,
4590 struct bpf_verifier_state
*cur
)
4594 if (old
->curframe
!= cur
->curframe
)
4597 /* for states to be equal callsites have to be the same
4598 * and all frame states need to be equivalent
4600 for (i
= 0; i
<= old
->curframe
; i
++) {
4601 if (old
->frame
[i
]->callsite
!= cur
->frame
[i
]->callsite
)
4603 if (!func_states_equal(old
->frame
[i
], cur
->frame
[i
]))
4609 /* A write screens off any subsequent reads; but write marks come from the
4610 * straight-line code between a state and its parent. When we arrive at an
4611 * equivalent state (jump target or such) we didn't arrive by the straight-line
4612 * code, so read marks in the state must propagate to the parent regardless
4613 * of the state's write marks. That's what 'parent == state->parent' comparison
4614 * in mark_reg_read() and mark_stack_slot_read() is for.
4616 static int propagate_liveness(struct bpf_verifier_env
*env
,
4617 const struct bpf_verifier_state
*vstate
,
4618 struct bpf_verifier_state
*vparent
)
4620 int i
, frame
, err
= 0;
4621 struct bpf_func_state
*state
, *parent
;
4623 if (vparent
->curframe
!= vstate
->curframe
) {
4624 WARN(1, "propagate_live: parent frame %d current frame %d\n",
4625 vparent
->curframe
, vstate
->curframe
);
4628 /* Propagate read liveness of registers... */
4629 BUILD_BUG_ON(BPF_REG_FP
+ 1 != MAX_BPF_REG
);
4630 /* We don't need to worry about FP liveness because it's read-only */
4631 for (i
= 0; i
< BPF_REG_FP
; i
++) {
4632 if (vparent
->frame
[vparent
->curframe
]->regs
[i
].live
& REG_LIVE_READ
)
4634 if (vstate
->frame
[vstate
->curframe
]->regs
[i
].live
& REG_LIVE_READ
) {
4635 err
= mark_reg_read(env
, vstate
, vparent
, i
);
4641 /* ... and stack slots */
4642 for (frame
= 0; frame
<= vstate
->curframe
; frame
++) {
4643 state
= vstate
->frame
[frame
];
4644 parent
= vparent
->frame
[frame
];
4645 for (i
= 0; i
< state
->allocated_stack
/ BPF_REG_SIZE
&&
4646 i
< parent
->allocated_stack
/ BPF_REG_SIZE
; i
++) {
4647 if (parent
->stack
[i
].spilled_ptr
.live
& REG_LIVE_READ
)
4649 if (state
->stack
[i
].spilled_ptr
.live
& REG_LIVE_READ
)
4650 mark_stack_slot_read(env
, vstate
, vparent
, i
, frame
);
4656 static int is_state_visited(struct bpf_verifier_env
*env
, int insn_idx
)
4658 struct bpf_verifier_state_list
*new_sl
;
4659 struct bpf_verifier_state_list
*sl
;
4660 struct bpf_verifier_state
*cur
= env
->cur_state
;
4663 sl
= env
->explored_states
[insn_idx
];
4665 /* this 'insn_idx' instruction wasn't marked, so we will not
4666 * be doing state search here
4670 while (sl
!= STATE_LIST_MARK
) {
4671 if (states_equal(env
, &sl
->state
, cur
)) {
4672 /* reached equivalent register/stack state,
4674 * Registers read by the continuation are read by us.
4675 * If we have any write marks in env->cur_state, they
4676 * will prevent corresponding reads in the continuation
4677 * from reaching our parent (an explored_state). Our
4678 * own state will get the read marks recorded, but
4679 * they'll be immediately forgotten as we're pruning
4680 * this state and will pop a new one.
4682 err
= propagate_liveness(env
, &sl
->state
, cur
);
4690 /* there were no equivalent states, remember current one.
4691 * technically the current state is not proven to be safe yet,
4692 * but it will either reach outer most bpf_exit (which means it's safe)
4693 * or it will be rejected. Since there are no loops, we won't be
4694 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
4695 * again on the way to bpf_exit
4697 new_sl
= kzalloc(sizeof(struct bpf_verifier_state_list
), GFP_KERNEL
);
4701 /* add new state to the head of linked list */
4702 err
= copy_verifier_state(&new_sl
->state
, cur
);
4704 free_verifier_state(&new_sl
->state
, false);
4708 new_sl
->next
= env
->explored_states
[insn_idx
];
4709 env
->explored_states
[insn_idx
] = new_sl
;
4710 /* connect new state to parentage chain */
4711 cur
->parent
= &new_sl
->state
;
4712 /* clear write marks in current state: the writes we did are not writes
4713 * our child did, so they don't screen off its reads from us.
4714 * (There are no read marks in current state, because reads always mark
4715 * their parent and current state never has children yet. Only
4716 * explored_states can get read marks.)
4718 for (i
= 0; i
< BPF_REG_FP
; i
++)
4719 cur
->frame
[cur
->curframe
]->regs
[i
].live
= REG_LIVE_NONE
;
4721 /* all stack frames are accessible from callee, clear them all */
4722 for (j
= 0; j
<= cur
->curframe
; j
++) {
4723 struct bpf_func_state
*frame
= cur
->frame
[j
];
4725 for (i
= 0; i
< frame
->allocated_stack
/ BPF_REG_SIZE
; i
++)
4726 frame
->stack
[i
].spilled_ptr
.live
= REG_LIVE_NONE
;
4731 static int do_check(struct bpf_verifier_env
*env
)
4733 struct bpf_verifier_state
*state
;
4734 struct bpf_insn
*insns
= env
->prog
->insnsi
;
4735 struct bpf_reg_state
*regs
;
4736 int insn_cnt
= env
->prog
->len
, i
;
4737 int insn_idx
, prev_insn_idx
= 0;
4738 int insn_processed
= 0;
4739 bool do_print_state
= false;
4741 state
= kzalloc(sizeof(struct bpf_verifier_state
), GFP_KERNEL
);
4744 state
->curframe
= 0;
4745 state
->parent
= NULL
;
4746 state
->frame
[0] = kzalloc(sizeof(struct bpf_func_state
), GFP_KERNEL
);
4747 if (!state
->frame
[0]) {
4751 env
->cur_state
= state
;
4752 init_func_state(env
, state
->frame
[0],
4753 BPF_MAIN_FUNC
/* callsite */,
4755 0 /* subprogno, zero == main subprog */);
4758 struct bpf_insn
*insn
;
4762 if (insn_idx
>= insn_cnt
) {
4763 verbose(env
, "invalid insn idx %d insn_cnt %d\n",
4764 insn_idx
, insn_cnt
);
4768 insn
= &insns
[insn_idx
];
4769 class = BPF_CLASS(insn
->code
);
4771 if (++insn_processed
> BPF_COMPLEXITY_LIMIT_INSNS
) {
4773 "BPF program is too large. Processed %d insn\n",
4778 err
= is_state_visited(env
, insn_idx
);
4782 /* found equivalent state, can prune the search */
4783 if (env
->log
.level
) {
4785 verbose(env
, "\nfrom %d to %d: safe\n",
4786 prev_insn_idx
, insn_idx
);
4788 verbose(env
, "%d: safe\n", insn_idx
);
4790 goto process_bpf_exit
;
4796 if (env
->log
.level
> 1 || (env
->log
.level
&& do_print_state
)) {
4797 if (env
->log
.level
> 1)
4798 verbose(env
, "%d:", insn_idx
);
4800 verbose(env
, "\nfrom %d to %d:",
4801 prev_insn_idx
, insn_idx
);
4802 print_verifier_state(env
, state
->frame
[state
->curframe
]);
4803 do_print_state
= false;
4806 if (env
->log
.level
) {
4807 const struct bpf_insn_cbs cbs
= {
4808 .cb_print
= verbose
,
4809 .private_data
= env
,
4812 verbose(env
, "%d: ", insn_idx
);
4813 print_bpf_insn(&cbs
, insn
, env
->allow_ptr_leaks
);
4816 if (bpf_prog_is_dev_bound(env
->prog
->aux
)) {
4817 err
= bpf_prog_offload_verify_insn(env
, insn_idx
,
4823 regs
= cur_regs(env
);
4824 env
->insn_aux_data
[insn_idx
].seen
= true;
4825 if (class == BPF_ALU
|| class == BPF_ALU64
) {
4826 err
= check_alu_op(env
, insn
);
4830 } else if (class == BPF_LDX
) {
4831 enum bpf_reg_type
*prev_src_type
, src_reg_type
;
4833 /* check for reserved fields is already done */
4835 /* check src operand */
4836 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
4840 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP_NO_MARK
);
4844 src_reg_type
= regs
[insn
->src_reg
].type
;
4846 /* check that memory (src_reg + off) is readable,
4847 * the state of dst_reg will be updated by this func
4849 err
= check_mem_access(env
, insn_idx
, insn
->src_reg
, insn
->off
,
4850 BPF_SIZE(insn
->code
), BPF_READ
,
4851 insn
->dst_reg
, false);
4855 prev_src_type
= &env
->insn_aux_data
[insn_idx
].ptr_type
;
4857 if (*prev_src_type
== NOT_INIT
) {
4859 * dst_reg = *(u32 *)(src_reg + off)
4860 * save type to validate intersecting paths
4862 *prev_src_type
= src_reg_type
;
4864 } else if (src_reg_type
!= *prev_src_type
&&
4865 (src_reg_type
== PTR_TO_CTX
||
4866 *prev_src_type
== PTR_TO_CTX
)) {
4867 /* ABuser program is trying to use the same insn
4868 * dst_reg = *(u32*) (src_reg + off)
4869 * with different pointer types:
4870 * src_reg == ctx in one branch and
4871 * src_reg == stack|map in some other branch.
4874 verbose(env
, "same insn cannot be used with different pointers\n");
4878 } else if (class == BPF_STX
) {
4879 enum bpf_reg_type
*prev_dst_type
, dst_reg_type
;
4881 if (BPF_MODE(insn
->code
) == BPF_XADD
) {
4882 err
= check_xadd(env
, insn_idx
, insn
);
4889 /* check src1 operand */
4890 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
4893 /* check src2 operand */
4894 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
4898 dst_reg_type
= regs
[insn
->dst_reg
].type
;
4900 /* check that memory (dst_reg + off) is writeable */
4901 err
= check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
4902 BPF_SIZE(insn
->code
), BPF_WRITE
,
4903 insn
->src_reg
, false);
4907 prev_dst_type
= &env
->insn_aux_data
[insn_idx
].ptr_type
;
4909 if (*prev_dst_type
== NOT_INIT
) {
4910 *prev_dst_type
= dst_reg_type
;
4911 } else if (dst_reg_type
!= *prev_dst_type
&&
4912 (dst_reg_type
== PTR_TO_CTX
||
4913 *prev_dst_type
== PTR_TO_CTX
)) {
4914 verbose(env
, "same insn cannot be used with different pointers\n");
4918 } else if (class == BPF_ST
) {
4919 if (BPF_MODE(insn
->code
) != BPF_MEM
||
4920 insn
->src_reg
!= BPF_REG_0
) {
4921 verbose(env
, "BPF_ST uses reserved fields\n");
4924 /* check src operand */
4925 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
4929 if (is_ctx_reg(env
, insn
->dst_reg
)) {
4930 verbose(env
, "BPF_ST stores into R%d context is not allowed\n",
4935 /* check that memory (dst_reg + off) is writeable */
4936 err
= check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
4937 BPF_SIZE(insn
->code
), BPF_WRITE
,
4942 } else if (class == BPF_JMP
) {
4943 u8 opcode
= BPF_OP(insn
->code
);
4945 if (opcode
== BPF_CALL
) {
4946 if (BPF_SRC(insn
->code
) != BPF_K
||
4948 (insn
->src_reg
!= BPF_REG_0
&&
4949 insn
->src_reg
!= BPF_PSEUDO_CALL
) ||
4950 insn
->dst_reg
!= BPF_REG_0
) {
4951 verbose(env
, "BPF_CALL uses reserved fields\n");
4955 if (insn
->src_reg
== BPF_PSEUDO_CALL
)
4956 err
= check_func_call(env
, insn
, &insn_idx
);
4958 err
= check_helper_call(env
, insn
->imm
, insn_idx
);
4962 } else if (opcode
== BPF_JA
) {
4963 if (BPF_SRC(insn
->code
) != BPF_K
||
4965 insn
->src_reg
!= BPF_REG_0
||
4966 insn
->dst_reg
!= BPF_REG_0
) {
4967 verbose(env
, "BPF_JA uses reserved fields\n");
4971 insn_idx
+= insn
->off
+ 1;
4974 } else if (opcode
== BPF_EXIT
) {
4975 if (BPF_SRC(insn
->code
) != BPF_K
||
4977 insn
->src_reg
!= BPF_REG_0
||
4978 insn
->dst_reg
!= BPF_REG_0
) {
4979 verbose(env
, "BPF_EXIT uses reserved fields\n");
4983 if (state
->curframe
) {
4984 /* exit from nested function */
4985 prev_insn_idx
= insn_idx
;
4986 err
= prepare_func_exit(env
, &insn_idx
);
4989 do_print_state
= true;
4993 /* eBPF calling convetion is such that R0 is used
4994 * to return the value from eBPF program.
4995 * Make sure that it's readable at this time
4996 * of bpf_exit, which means that program wrote
4997 * something into it earlier
4999 err
= check_reg_arg(env
, BPF_REG_0
, SRC_OP
);
5003 if (is_pointer_value(env
, BPF_REG_0
)) {
5004 verbose(env
, "R0 leaks addr as return value\n");
5008 err
= check_return_code(env
);
5012 err
= pop_stack(env
, &prev_insn_idx
, &insn_idx
);
5018 do_print_state
= true;
5022 err
= check_cond_jmp_op(env
, insn
, &insn_idx
);
5026 } else if (class == BPF_LD
) {
5027 u8 mode
= BPF_MODE(insn
->code
);
5029 if (mode
== BPF_ABS
|| mode
== BPF_IND
) {
5030 err
= check_ld_abs(env
, insn
);
5034 } else if (mode
== BPF_IMM
) {
5035 err
= check_ld_imm(env
, insn
);
5040 env
->insn_aux_data
[insn_idx
].seen
= true;
5042 verbose(env
, "invalid BPF_LD mode\n");
5046 verbose(env
, "unknown insn class %d\n", class);
5053 verbose(env
, "processed %d insns (limit %d), stack depth ",
5054 insn_processed
, BPF_COMPLEXITY_LIMIT_INSNS
);
5055 for (i
= 0; i
< env
->subprog_cnt
; i
++) {
5056 u32 depth
= env
->subprog_info
[i
].stack_depth
;
5058 verbose(env
, "%d", depth
);
5059 if (i
+ 1 < env
->subprog_cnt
)
5063 env
->prog
->aux
->stack_depth
= env
->subprog_info
[0].stack_depth
;
5067 static int check_map_prealloc(struct bpf_map
*map
)
5069 return (map
->map_type
!= BPF_MAP_TYPE_HASH
&&
5070 map
->map_type
!= BPF_MAP_TYPE_PERCPU_HASH
&&
5071 map
->map_type
!= BPF_MAP_TYPE_HASH_OF_MAPS
) ||
5072 !(map
->map_flags
& BPF_F_NO_PREALLOC
);
5075 static int check_map_prog_compatibility(struct bpf_verifier_env
*env
,
5076 struct bpf_map
*map
,
5077 struct bpf_prog
*prog
)
5080 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
5081 * preallocated hash maps, since doing memory allocation
5082 * in overflow_handler can crash depending on where nmi got
5085 if (prog
->type
== BPF_PROG_TYPE_PERF_EVENT
) {
5086 if (!check_map_prealloc(map
)) {
5087 verbose(env
, "perf_event programs can only use preallocated hash map\n");
5090 if (map
->inner_map_meta
&&
5091 !check_map_prealloc(map
->inner_map_meta
)) {
5092 verbose(env
, "perf_event programs can only use preallocated inner hash map\n");
5097 if ((bpf_prog_is_dev_bound(prog
->aux
) || bpf_map_is_dev_bound(map
)) &&
5098 !bpf_offload_prog_map_match(prog
, map
)) {
5099 verbose(env
, "offload device mismatch between prog and map\n");
5106 /* look for pseudo eBPF instructions that access map FDs and
5107 * replace them with actual map pointers
5109 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env
*env
)
5111 struct bpf_insn
*insn
= env
->prog
->insnsi
;
5112 int insn_cnt
= env
->prog
->len
;
5115 err
= bpf_prog_calc_tag(env
->prog
);
5119 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
5120 if (BPF_CLASS(insn
->code
) == BPF_LDX
&&
5121 (BPF_MODE(insn
->code
) != BPF_MEM
|| insn
->imm
!= 0)) {
5122 verbose(env
, "BPF_LDX uses reserved fields\n");
5126 if (BPF_CLASS(insn
->code
) == BPF_STX
&&
5127 ((BPF_MODE(insn
->code
) != BPF_MEM
&&
5128 BPF_MODE(insn
->code
) != BPF_XADD
) || insn
->imm
!= 0)) {
5129 verbose(env
, "BPF_STX uses reserved fields\n");
5133 if (insn
[0].code
== (BPF_LD
| BPF_IMM
| BPF_DW
)) {
5134 struct bpf_map
*map
;
5137 if (i
== insn_cnt
- 1 || insn
[1].code
!= 0 ||
5138 insn
[1].dst_reg
!= 0 || insn
[1].src_reg
!= 0 ||
5140 verbose(env
, "invalid bpf_ld_imm64 insn\n");
5144 if (insn
->src_reg
== 0)
5145 /* valid generic load 64-bit imm */
5148 if (insn
->src_reg
!= BPF_PSEUDO_MAP_FD
) {
5150 "unrecognized bpf_ld_imm64 insn\n");
5154 f
= fdget(insn
->imm
);
5155 map
= __bpf_map_get(f
);
5157 verbose(env
, "fd %d is not pointing to valid bpf_map\n",
5159 return PTR_ERR(map
);
5162 err
= check_map_prog_compatibility(env
, map
, env
->prog
);
5168 /* store map pointer inside BPF_LD_IMM64 instruction */
5169 insn
[0].imm
= (u32
) (unsigned long) map
;
5170 insn
[1].imm
= ((u64
) (unsigned long) map
) >> 32;
5172 /* check whether we recorded this map already */
5173 for (j
= 0; j
< env
->used_map_cnt
; j
++)
5174 if (env
->used_maps
[j
] == map
) {
5179 if (env
->used_map_cnt
>= MAX_USED_MAPS
) {
5184 /* hold the map. If the program is rejected by verifier,
5185 * the map will be released by release_maps() or it
5186 * will be used by the valid program until it's unloaded
5187 * and all maps are released in free_used_maps()
5189 map
= bpf_map_inc(map
, false);
5192 return PTR_ERR(map
);
5194 env
->used_maps
[env
->used_map_cnt
++] = map
;
5196 if (map
->map_type
== BPF_MAP_TYPE_CGROUP_STORAGE
&&
5197 bpf_cgroup_storage_assign(env
->prog
, map
)) {
5199 "only one cgroup storage is allowed\n");
5211 /* Basic sanity check before we invest more work here. */
5212 if (!bpf_opcode_in_insntable(insn
->code
)) {
5213 verbose(env
, "unknown opcode %02x\n", insn
->code
);
5218 /* now all pseudo BPF_LD_IMM64 instructions load valid
5219 * 'struct bpf_map *' into a register instead of user map_fd.
5220 * These pointers will be used later by verifier to validate map access.
5225 /* drop refcnt of maps used by the rejected program */
5226 static void release_maps(struct bpf_verifier_env
*env
)
5230 if (env
->prog
->aux
->cgroup_storage
)
5231 bpf_cgroup_storage_release(env
->prog
,
5232 env
->prog
->aux
->cgroup_storage
);
5234 for (i
= 0; i
< env
->used_map_cnt
; i
++)
5235 bpf_map_put(env
->used_maps
[i
]);
5238 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
5239 static void convert_pseudo_ld_imm64(struct bpf_verifier_env
*env
)
5241 struct bpf_insn
*insn
= env
->prog
->insnsi
;
5242 int insn_cnt
= env
->prog
->len
;
5245 for (i
= 0; i
< insn_cnt
; i
++, insn
++)
5246 if (insn
->code
== (BPF_LD
| BPF_IMM
| BPF_DW
))
5250 /* single env->prog->insni[off] instruction was replaced with the range
5251 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
5252 * [0, off) and [off, end) to new locations, so the patched range stays zero
5254 static int adjust_insn_aux_data(struct bpf_verifier_env
*env
, u32 prog_len
,
5257 struct bpf_insn_aux_data
*new_data
, *old_data
= env
->insn_aux_data
;
5262 new_data
= vzalloc(array_size(prog_len
,
5263 sizeof(struct bpf_insn_aux_data
)));
5266 memcpy(new_data
, old_data
, sizeof(struct bpf_insn_aux_data
) * off
);
5267 memcpy(new_data
+ off
+ cnt
- 1, old_data
+ off
,
5268 sizeof(struct bpf_insn_aux_data
) * (prog_len
- off
- cnt
+ 1));
5269 for (i
= off
; i
< off
+ cnt
- 1; i
++)
5270 new_data
[i
].seen
= true;
5271 env
->insn_aux_data
= new_data
;
5276 static void adjust_subprog_starts(struct bpf_verifier_env
*env
, u32 off
, u32 len
)
5282 /* NOTE: fake 'exit' subprog should be updated as well. */
5283 for (i
= 0; i
<= env
->subprog_cnt
; i
++) {
5284 if (env
->subprog_info
[i
].start
< off
)
5286 env
->subprog_info
[i
].start
+= len
- 1;
5290 static struct bpf_prog
*bpf_patch_insn_data(struct bpf_verifier_env
*env
, u32 off
,
5291 const struct bpf_insn
*patch
, u32 len
)
5293 struct bpf_prog
*new_prog
;
5295 new_prog
= bpf_patch_insn_single(env
->prog
, off
, patch
, len
);
5298 if (adjust_insn_aux_data(env
, new_prog
->len
, off
, len
))
5300 adjust_subprog_starts(env
, off
, len
);
5304 /* The verifier does more data flow analysis than llvm and will not
5305 * explore branches that are dead at run time. Malicious programs can
5306 * have dead code too. Therefore replace all dead at-run-time code
5309 * Just nops are not optimal, e.g. if they would sit at the end of the
5310 * program and through another bug we would manage to jump there, then
5311 * we'd execute beyond program memory otherwise. Returning exception
5312 * code also wouldn't work since we can have subprogs where the dead
5313 * code could be located.
5315 static void sanitize_dead_code(struct bpf_verifier_env
*env
)
5317 struct bpf_insn_aux_data
*aux_data
= env
->insn_aux_data
;
5318 struct bpf_insn trap
= BPF_JMP_IMM(BPF_JA
, 0, 0, -1);
5319 struct bpf_insn
*insn
= env
->prog
->insnsi
;
5320 const int insn_cnt
= env
->prog
->len
;
5323 for (i
= 0; i
< insn_cnt
; i
++) {
5324 if (aux_data
[i
].seen
)
5326 memcpy(insn
+ i
, &trap
, sizeof(trap
));
5330 /* convert load instructions that access fields of 'struct __sk_buff'
5331 * into sequence of instructions that access fields of 'struct sk_buff'
5333 static int convert_ctx_accesses(struct bpf_verifier_env
*env
)
5335 const struct bpf_verifier_ops
*ops
= env
->ops
;
5336 int i
, cnt
, size
, ctx_field_size
, delta
= 0;
5337 const int insn_cnt
= env
->prog
->len
;
5338 struct bpf_insn insn_buf
[16], *insn
;
5339 struct bpf_prog
*new_prog
;
5340 enum bpf_access_type type
;
5341 bool is_narrower_load
;
5344 if (ops
->gen_prologue
) {
5345 cnt
= ops
->gen_prologue(insn_buf
, env
->seen_direct_write
,
5347 if (cnt
>= ARRAY_SIZE(insn_buf
)) {
5348 verbose(env
, "bpf verifier is misconfigured\n");
5351 new_prog
= bpf_patch_insn_data(env
, 0, insn_buf
, cnt
);
5355 env
->prog
= new_prog
;
5360 if (!ops
->convert_ctx_access
|| bpf_prog_is_dev_bound(env
->prog
->aux
))
5363 insn
= env
->prog
->insnsi
+ delta
;
5365 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
5366 if (insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_B
) ||
5367 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_H
) ||
5368 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_W
) ||
5369 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_DW
))
5371 else if (insn
->code
== (BPF_STX
| BPF_MEM
| BPF_B
) ||
5372 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_H
) ||
5373 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_W
) ||
5374 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_DW
))
5379 if (type
== BPF_WRITE
&&
5380 env
->insn_aux_data
[i
+ delta
].sanitize_stack_off
) {
5381 struct bpf_insn patch
[] = {
5382 /* Sanitize suspicious stack slot with zero.
5383 * There are no memory dependencies for this store,
5384 * since it's only using frame pointer and immediate
5387 BPF_ST_MEM(BPF_DW
, BPF_REG_FP
,
5388 env
->insn_aux_data
[i
+ delta
].sanitize_stack_off
,
5390 /* the original STX instruction will immediately
5391 * overwrite the same stack slot with appropriate value
5396 cnt
= ARRAY_SIZE(patch
);
5397 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, patch
, cnt
);
5402 env
->prog
= new_prog
;
5403 insn
= new_prog
->insnsi
+ i
+ delta
;
5407 if (env
->insn_aux_data
[i
+ delta
].ptr_type
!= PTR_TO_CTX
)
5410 ctx_field_size
= env
->insn_aux_data
[i
+ delta
].ctx_field_size
;
5411 size
= BPF_LDST_BYTES(insn
);
5413 /* If the read access is a narrower load of the field,
5414 * convert to a 4/8-byte load, to minimum program type specific
5415 * convert_ctx_access changes. If conversion is successful,
5416 * we will apply proper mask to the result.
5418 is_narrower_load
= size
< ctx_field_size
;
5419 if (is_narrower_load
) {
5420 u32 size_default
= bpf_ctx_off_adjust_machine(ctx_field_size
);
5421 u32 off
= insn
->off
;
5424 if (type
== BPF_WRITE
) {
5425 verbose(env
, "bpf verifier narrow ctx access misconfigured\n");
5430 if (ctx_field_size
== 4)
5432 else if (ctx_field_size
== 8)
5435 insn
->off
= off
& ~(size_default
- 1);
5436 insn
->code
= BPF_LDX
| BPF_MEM
| size_code
;
5440 cnt
= ops
->convert_ctx_access(type
, insn
, insn_buf
, env
->prog
,
5442 if (cnt
== 0 || cnt
>= ARRAY_SIZE(insn_buf
) ||
5443 (ctx_field_size
&& !target_size
)) {
5444 verbose(env
, "bpf verifier is misconfigured\n");
5448 if (is_narrower_load
&& size
< target_size
) {
5449 if (ctx_field_size
<= 4)
5450 insn_buf
[cnt
++] = BPF_ALU32_IMM(BPF_AND
, insn
->dst_reg
,
5451 (1 << size
* 8) - 1);
5453 insn_buf
[cnt
++] = BPF_ALU64_IMM(BPF_AND
, insn
->dst_reg
,
5454 (1 << size
* 8) - 1);
5457 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
, cnt
);
5463 /* keep walking new program and skip insns we just inserted */
5464 env
->prog
= new_prog
;
5465 insn
= new_prog
->insnsi
+ i
+ delta
;
5471 static int jit_subprogs(struct bpf_verifier_env
*env
)
5473 struct bpf_prog
*prog
= env
->prog
, **func
, *tmp
;
5474 int i
, j
, subprog_start
, subprog_end
= 0, len
, subprog
;
5475 struct bpf_insn
*insn
;
5479 if (env
->subprog_cnt
<= 1)
5482 for (i
= 0, insn
= prog
->insnsi
; i
< prog
->len
; i
++, insn
++) {
5483 if (insn
->code
!= (BPF_JMP
| BPF_CALL
) ||
5484 insn
->src_reg
!= BPF_PSEUDO_CALL
)
5486 /* Upon error here we cannot fall back to interpreter but
5487 * need a hard reject of the program. Thus -EFAULT is
5488 * propagated in any case.
5490 subprog
= find_subprog(env
, i
+ insn
->imm
+ 1);
5492 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5496 /* temporarily remember subprog id inside insn instead of
5497 * aux_data, since next loop will split up all insns into funcs
5499 insn
->off
= subprog
;
5500 /* remember original imm in case JIT fails and fallback
5501 * to interpreter will be needed
5503 env
->insn_aux_data
[i
].call_imm
= insn
->imm
;
5504 /* point imm to __bpf_call_base+1 from JITs point of view */
5508 func
= kcalloc(env
->subprog_cnt
, sizeof(prog
), GFP_KERNEL
);
5512 for (i
= 0; i
< env
->subprog_cnt
; i
++) {
5513 subprog_start
= subprog_end
;
5514 subprog_end
= env
->subprog_info
[i
+ 1].start
;
5516 len
= subprog_end
- subprog_start
;
5517 func
[i
] = bpf_prog_alloc(bpf_prog_size(len
), GFP_USER
);
5520 memcpy(func
[i
]->insnsi
, &prog
->insnsi
[subprog_start
],
5521 len
* sizeof(struct bpf_insn
));
5522 func
[i
]->type
= prog
->type
;
5524 if (bpf_prog_calc_tag(func
[i
]))
5526 func
[i
]->is_func
= 1;
5527 /* Use bpf_prog_F_tag to indicate functions in stack traces.
5528 * Long term would need debug info to populate names
5530 func
[i
]->aux
->name
[0] = 'F';
5531 func
[i
]->aux
->stack_depth
= env
->subprog_info
[i
].stack_depth
;
5532 func
[i
]->jit_requested
= 1;
5533 func
[i
] = bpf_int_jit_compile(func
[i
]);
5534 if (!func
[i
]->jited
) {
5540 /* at this point all bpf functions were successfully JITed
5541 * now populate all bpf_calls with correct addresses and
5542 * run last pass of JIT
5544 for (i
= 0; i
< env
->subprog_cnt
; i
++) {
5545 insn
= func
[i
]->insnsi
;
5546 for (j
= 0; j
< func
[i
]->len
; j
++, insn
++) {
5547 if (insn
->code
!= (BPF_JMP
| BPF_CALL
) ||
5548 insn
->src_reg
!= BPF_PSEUDO_CALL
)
5550 subprog
= insn
->off
;
5551 insn
->imm
= (u64 (*)(u64
, u64
, u64
, u64
, u64
))
5552 func
[subprog
]->bpf_func
-
5556 /* we use the aux data to keep a list of the start addresses
5557 * of the JITed images for each function in the program
5559 * for some architectures, such as powerpc64, the imm field
5560 * might not be large enough to hold the offset of the start
5561 * address of the callee's JITed image from __bpf_call_base
5563 * in such cases, we can lookup the start address of a callee
5564 * by using its subprog id, available from the off field of
5565 * the call instruction, as an index for this list
5567 func
[i
]->aux
->func
= func
;
5568 func
[i
]->aux
->func_cnt
= env
->subprog_cnt
;
5570 for (i
= 0; i
< env
->subprog_cnt
; i
++) {
5571 old_bpf_func
= func
[i
]->bpf_func
;
5572 tmp
= bpf_int_jit_compile(func
[i
]);
5573 if (tmp
!= func
[i
] || func
[i
]->bpf_func
!= old_bpf_func
) {
5574 verbose(env
, "JIT doesn't support bpf-to-bpf calls\n");
5581 /* finally lock prog and jit images for all functions and
5584 for (i
= 0; i
< env
->subprog_cnt
; i
++) {
5585 bpf_prog_lock_ro(func
[i
]);
5586 bpf_prog_kallsyms_add(func
[i
]);
5589 /* Last step: make now unused interpreter insns from main
5590 * prog consistent for later dump requests, so they can
5591 * later look the same as if they were interpreted only.
5593 for (i
= 0, insn
= prog
->insnsi
; i
< prog
->len
; i
++, insn
++) {
5594 if (insn
->code
!= (BPF_JMP
| BPF_CALL
) ||
5595 insn
->src_reg
!= BPF_PSEUDO_CALL
)
5597 insn
->off
= env
->insn_aux_data
[i
].call_imm
;
5598 subprog
= find_subprog(env
, i
+ insn
->off
+ 1);
5599 insn
->imm
= subprog
;
5603 prog
->bpf_func
= func
[0]->bpf_func
;
5604 prog
->aux
->func
= func
;
5605 prog
->aux
->func_cnt
= env
->subprog_cnt
;
5608 for (i
= 0; i
< env
->subprog_cnt
; i
++)
5610 bpf_jit_free(func
[i
]);
5613 /* cleanup main prog to be interpreted */
5614 prog
->jit_requested
= 0;
5615 for (i
= 0, insn
= prog
->insnsi
; i
< prog
->len
; i
++, insn
++) {
5616 if (insn
->code
!= (BPF_JMP
| BPF_CALL
) ||
5617 insn
->src_reg
!= BPF_PSEUDO_CALL
)
5620 insn
->imm
= env
->insn_aux_data
[i
].call_imm
;
5625 static int fixup_call_args(struct bpf_verifier_env
*env
)
5627 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
5628 struct bpf_prog
*prog
= env
->prog
;
5629 struct bpf_insn
*insn
= prog
->insnsi
;
5635 if (env
->prog
->jit_requested
) {
5636 err
= jit_subprogs(env
);
5642 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
5643 for (i
= 0; i
< prog
->len
; i
++, insn
++) {
5644 if (insn
->code
!= (BPF_JMP
| BPF_CALL
) ||
5645 insn
->src_reg
!= BPF_PSEUDO_CALL
)
5647 depth
= get_callee_stack_depth(env
, insn
, i
);
5650 bpf_patch_call_args(insn
, depth
);
5657 /* fixup insn->imm field of bpf_call instructions
5658 * and inline eligible helpers as explicit sequence of BPF instructions
5660 * this function is called after eBPF program passed verification
5662 static int fixup_bpf_calls(struct bpf_verifier_env
*env
)
5664 struct bpf_prog
*prog
= env
->prog
;
5665 struct bpf_insn
*insn
= prog
->insnsi
;
5666 const struct bpf_func_proto
*fn
;
5667 const int insn_cnt
= prog
->len
;
5668 const struct bpf_map_ops
*ops
;
5669 struct bpf_insn_aux_data
*aux
;
5670 struct bpf_insn insn_buf
[16];
5671 struct bpf_prog
*new_prog
;
5672 struct bpf_map
*map_ptr
;
5673 int i
, cnt
, delta
= 0;
5675 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
5676 if (insn
->code
== (BPF_ALU64
| BPF_MOD
| BPF_X
) ||
5677 insn
->code
== (BPF_ALU64
| BPF_DIV
| BPF_X
) ||
5678 insn
->code
== (BPF_ALU
| BPF_MOD
| BPF_X
) ||
5679 insn
->code
== (BPF_ALU
| BPF_DIV
| BPF_X
)) {
5680 bool is64
= BPF_CLASS(insn
->code
) == BPF_ALU64
;
5681 struct bpf_insn mask_and_div
[] = {
5682 BPF_MOV32_REG(insn
->src_reg
, insn
->src_reg
),
5684 BPF_JMP_IMM(BPF_JNE
, insn
->src_reg
, 0, 2),
5685 BPF_ALU32_REG(BPF_XOR
, insn
->dst_reg
, insn
->dst_reg
),
5686 BPF_JMP_IMM(BPF_JA
, 0, 0, 1),
5689 struct bpf_insn mask_and_mod
[] = {
5690 BPF_MOV32_REG(insn
->src_reg
, insn
->src_reg
),
5691 /* Rx mod 0 -> Rx */
5692 BPF_JMP_IMM(BPF_JEQ
, insn
->src_reg
, 0, 1),
5695 struct bpf_insn
*patchlet
;
5697 if (insn
->code
== (BPF_ALU64
| BPF_DIV
| BPF_X
) ||
5698 insn
->code
== (BPF_ALU
| BPF_DIV
| BPF_X
)) {
5699 patchlet
= mask_and_div
+ (is64
? 1 : 0);
5700 cnt
= ARRAY_SIZE(mask_and_div
) - (is64
? 1 : 0);
5702 patchlet
= mask_and_mod
+ (is64
? 1 : 0);
5703 cnt
= ARRAY_SIZE(mask_and_mod
) - (is64
? 1 : 0);
5706 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, patchlet
, cnt
);
5711 env
->prog
= prog
= new_prog
;
5712 insn
= new_prog
->insnsi
+ i
+ delta
;
5716 if (BPF_CLASS(insn
->code
) == BPF_LD
&&
5717 (BPF_MODE(insn
->code
) == BPF_ABS
||
5718 BPF_MODE(insn
->code
) == BPF_IND
)) {
5719 cnt
= env
->ops
->gen_ld_abs(insn
, insn_buf
);
5720 if (cnt
== 0 || cnt
>= ARRAY_SIZE(insn_buf
)) {
5721 verbose(env
, "bpf verifier is misconfigured\n");
5725 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
, cnt
);
5730 env
->prog
= prog
= new_prog
;
5731 insn
= new_prog
->insnsi
+ i
+ delta
;
5735 if (insn
->code
!= (BPF_JMP
| BPF_CALL
))
5737 if (insn
->src_reg
== BPF_PSEUDO_CALL
)
5740 if (insn
->imm
== BPF_FUNC_get_route_realm
)
5741 prog
->dst_needed
= 1;
5742 if (insn
->imm
== BPF_FUNC_get_prandom_u32
)
5743 bpf_user_rnd_init_once();
5744 if (insn
->imm
== BPF_FUNC_override_return
)
5745 prog
->kprobe_override
= 1;
5746 if (insn
->imm
== BPF_FUNC_tail_call
) {
5747 /* If we tail call into other programs, we
5748 * cannot make any assumptions since they can
5749 * be replaced dynamically during runtime in
5750 * the program array.
5752 prog
->cb_access
= 1;
5753 env
->prog
->aux
->stack_depth
= MAX_BPF_STACK
;
5755 /* mark bpf_tail_call as different opcode to avoid
5756 * conditional branch in the interpeter for every normal
5757 * call and to prevent accidental JITing by JIT compiler
5758 * that doesn't support bpf_tail_call yet
5761 insn
->code
= BPF_JMP
| BPF_TAIL_CALL
;
5763 aux
= &env
->insn_aux_data
[i
+ delta
];
5764 if (!bpf_map_ptr_unpriv(aux
))
5767 /* instead of changing every JIT dealing with tail_call
5768 * emit two extra insns:
5769 * if (index >= max_entries) goto out;
5770 * index &= array->index_mask;
5771 * to avoid out-of-bounds cpu speculation
5773 if (bpf_map_ptr_poisoned(aux
)) {
5774 verbose(env
, "tail_call abusing map_ptr\n");
5778 map_ptr
= BPF_MAP_PTR(aux
->map_state
);
5779 insn_buf
[0] = BPF_JMP_IMM(BPF_JGE
, BPF_REG_3
,
5780 map_ptr
->max_entries
, 2);
5781 insn_buf
[1] = BPF_ALU32_IMM(BPF_AND
, BPF_REG_3
,
5782 container_of(map_ptr
,
5785 insn_buf
[2] = *insn
;
5787 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
, cnt
);
5792 env
->prog
= prog
= new_prog
;
5793 insn
= new_prog
->insnsi
+ i
+ delta
;
5797 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
5798 * and other inlining handlers are currently limited to 64 bit
5801 if (prog
->jit_requested
&& BITS_PER_LONG
== 64 &&
5802 (insn
->imm
== BPF_FUNC_map_lookup_elem
||
5803 insn
->imm
== BPF_FUNC_map_update_elem
||
5804 insn
->imm
== BPF_FUNC_map_delete_elem
)) {
5805 aux
= &env
->insn_aux_data
[i
+ delta
];
5806 if (bpf_map_ptr_poisoned(aux
))
5807 goto patch_call_imm
;
5809 map_ptr
= BPF_MAP_PTR(aux
->map_state
);
5811 if (insn
->imm
== BPF_FUNC_map_lookup_elem
&&
5812 ops
->map_gen_lookup
) {
5813 cnt
= ops
->map_gen_lookup(map_ptr
, insn_buf
);
5814 if (cnt
== 0 || cnt
>= ARRAY_SIZE(insn_buf
)) {
5815 verbose(env
, "bpf verifier is misconfigured\n");
5819 new_prog
= bpf_patch_insn_data(env
, i
+ delta
,
5825 env
->prog
= prog
= new_prog
;
5826 insn
= new_prog
->insnsi
+ i
+ delta
;
5830 BUILD_BUG_ON(!__same_type(ops
->map_lookup_elem
,
5831 (void *(*)(struct bpf_map
*map
, void *key
))NULL
));
5832 BUILD_BUG_ON(!__same_type(ops
->map_delete_elem
,
5833 (int (*)(struct bpf_map
*map
, void *key
))NULL
));
5834 BUILD_BUG_ON(!__same_type(ops
->map_update_elem
,
5835 (int (*)(struct bpf_map
*map
, void *key
, void *value
,
5837 switch (insn
->imm
) {
5838 case BPF_FUNC_map_lookup_elem
:
5839 insn
->imm
= BPF_CAST_CALL(ops
->map_lookup_elem
) -
5842 case BPF_FUNC_map_update_elem
:
5843 insn
->imm
= BPF_CAST_CALL(ops
->map_update_elem
) -
5846 case BPF_FUNC_map_delete_elem
:
5847 insn
->imm
= BPF_CAST_CALL(ops
->map_delete_elem
) -
5852 goto patch_call_imm
;
5856 fn
= env
->ops
->get_func_proto(insn
->imm
, env
->prog
);
5857 /* all functions that have prototype and verifier allowed
5858 * programs to call them, must be real in-kernel functions
5862 "kernel subsystem misconfigured func %s#%d\n",
5863 func_id_name(insn
->imm
), insn
->imm
);
5866 insn
->imm
= fn
->func
- __bpf_call_base
;
5872 static void free_states(struct bpf_verifier_env
*env
)
5874 struct bpf_verifier_state_list
*sl
, *sln
;
5877 if (!env
->explored_states
)
5880 for (i
= 0; i
< env
->prog
->len
; i
++) {
5881 sl
= env
->explored_states
[i
];
5884 while (sl
!= STATE_LIST_MARK
) {
5886 free_verifier_state(&sl
->state
, false);
5892 kfree(env
->explored_states
);
5895 int bpf_check(struct bpf_prog
**prog
, union bpf_attr
*attr
)
5897 struct bpf_verifier_env
*env
;
5898 struct bpf_verifier_log
*log
;
5901 /* no program is valid */
5902 if (ARRAY_SIZE(bpf_verifier_ops
) == 0)
5905 /* 'struct bpf_verifier_env' can be global, but since it's not small,
5906 * allocate/free it every time bpf_check() is called
5908 env
= kzalloc(sizeof(struct bpf_verifier_env
), GFP_KERNEL
);
5913 env
->insn_aux_data
=
5914 vzalloc(array_size(sizeof(struct bpf_insn_aux_data
),
5917 if (!env
->insn_aux_data
)
5920 env
->ops
= bpf_verifier_ops
[env
->prog
->type
];
5922 /* grab the mutex to protect few globals used by verifier */
5923 mutex_lock(&bpf_verifier_lock
);
5925 if (attr
->log_level
|| attr
->log_buf
|| attr
->log_size
) {
5926 /* user requested verbose verifier output
5927 * and supplied buffer to store the verification trace
5929 log
->level
= attr
->log_level
;
5930 log
->ubuf
= (char __user
*) (unsigned long) attr
->log_buf
;
5931 log
->len_total
= attr
->log_size
;
5934 /* log attributes have to be sane */
5935 if (log
->len_total
< 128 || log
->len_total
> UINT_MAX
>> 8 ||
5936 !log
->level
|| !log
->ubuf
)
5940 env
->strict_alignment
= !!(attr
->prog_flags
& BPF_F_STRICT_ALIGNMENT
);
5941 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
))
5942 env
->strict_alignment
= true;
5944 ret
= replace_map_fd_with_map_ptr(env
);
5946 goto skip_full_check
;
5948 if (bpf_prog_is_dev_bound(env
->prog
->aux
)) {
5949 ret
= bpf_prog_offload_verifier_prep(env
);
5951 goto skip_full_check
;
5954 env
->explored_states
= kcalloc(env
->prog
->len
,
5955 sizeof(struct bpf_verifier_state_list
*),
5958 if (!env
->explored_states
)
5959 goto skip_full_check
;
5961 env
->allow_ptr_leaks
= capable(CAP_SYS_ADMIN
);
5963 ret
= check_cfg(env
);
5965 goto skip_full_check
;
5967 ret
= do_check(env
);
5968 if (env
->cur_state
) {
5969 free_verifier_state(env
->cur_state
, true);
5970 env
->cur_state
= NULL
;
5974 while (!pop_stack(env
, NULL
, NULL
));
5978 sanitize_dead_code(env
);
5981 ret
= check_max_stack_depth(env
);
5984 /* program is valid, convert *(u32*)(ctx + off) accesses */
5985 ret
= convert_ctx_accesses(env
);
5988 ret
= fixup_bpf_calls(env
);
5991 ret
= fixup_call_args(env
);
5993 if (log
->level
&& bpf_verifier_log_full(log
))
5995 if (log
->level
&& !log
->ubuf
) {
5997 goto err_release_maps
;
6000 if (ret
== 0 && env
->used_map_cnt
) {
6001 /* if program passed verifier, update used_maps in bpf_prog_info */
6002 env
->prog
->aux
->used_maps
= kmalloc_array(env
->used_map_cnt
,
6003 sizeof(env
->used_maps
[0]),
6006 if (!env
->prog
->aux
->used_maps
) {
6008 goto err_release_maps
;
6011 memcpy(env
->prog
->aux
->used_maps
, env
->used_maps
,
6012 sizeof(env
->used_maps
[0]) * env
->used_map_cnt
);
6013 env
->prog
->aux
->used_map_cnt
= env
->used_map_cnt
;
6015 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
6016 * bpf_ld_imm64 instructions
6018 convert_pseudo_ld_imm64(env
);
6022 if (!env
->prog
->aux
->used_maps
)
6023 /* if we didn't copy map pointers into bpf_prog_info, release
6024 * them now. Otherwise free_used_maps() will release them.
6029 mutex_unlock(&bpf_verifier_lock
);
6030 vfree(env
->insn_aux_data
);