1 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
2 * Copyright (c) 2016 Facebook
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of version 2 of the GNU General Public
6 * License as published by the Free Software Foundation.
8 * This program is distributed in the hope that it will be useful, but
9 * WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
13 #include <linux/kernel.h>
14 #include <linux/types.h>
15 #include <linux/slab.h>
16 #include <linux/bpf.h>
17 #include <linux/bpf_verifier.h>
18 #include <linux/filter.h>
19 #include <net/netlink.h>
20 #include <linux/file.h>
21 #include <linux/vmalloc.h>
22 #include <linux/stringify.h>
26 /* bpf_check() is a static code analyzer that walks eBPF program
27 * instruction by instruction and updates register/stack state.
28 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
30 * The first pass is depth-first-search to check that the program is a DAG.
31 * It rejects the following programs:
32 * - larger than BPF_MAXINSNS insns
33 * - if loop is present (detected via back-edge)
34 * - unreachable insns exist (shouldn't be a forest. program = one function)
35 * - out of bounds or malformed jumps
36 * The second pass is all possible path descent from the 1st insn.
37 * Since it's analyzing all pathes through the program, the length of the
38 * analysis is limited to 64k insn, which may be hit even if total number of
39 * insn is less then 4K, but there are too many branches that change stack/regs.
40 * Number of 'branches to be analyzed' is limited to 1k
42 * On entry to each instruction, each register has a type, and the instruction
43 * changes the types of the registers depending on instruction semantics.
44 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
47 * All registers are 64-bit.
48 * R0 - return register
49 * R1-R5 argument passing registers
50 * R6-R9 callee saved registers
51 * R10 - frame pointer read-only
53 * At the start of BPF program the register R1 contains a pointer to bpf_context
54 * and has type PTR_TO_CTX.
56 * Verifier tracks arithmetic operations on pointers in case:
57 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
58 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
59 * 1st insn copies R10 (which has FRAME_PTR) type into R1
60 * and 2nd arithmetic instruction is pattern matched to recognize
61 * that it wants to construct a pointer to some element within stack.
62 * So after 2nd insn, the register R1 has type PTR_TO_STACK
63 * (and -20 constant is saved for further stack bounds checking).
64 * Meaning that this reg is a pointer to stack plus known immediate constant.
66 * Most of the time the registers have SCALAR_VALUE type, which
67 * means the register has some value, but it's not a valid pointer.
68 * (like pointer plus pointer becomes SCALAR_VALUE type)
70 * When verifier sees load or store instructions the type of base register
71 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK. These are three pointer
72 * types recognized by check_mem_access() function.
74 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
75 * and the range of [ptr, ptr + map's value_size) is accessible.
77 * registers used to pass values to function calls are checked against
78 * function argument constraints.
80 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
81 * It means that the register type passed to this function must be
82 * PTR_TO_STACK and it will be used inside the function as
83 * 'pointer to map element key'
85 * For example the argument constraints for bpf_map_lookup_elem():
86 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
87 * .arg1_type = ARG_CONST_MAP_PTR,
88 * .arg2_type = ARG_PTR_TO_MAP_KEY,
90 * ret_type says that this function returns 'pointer to map elem value or null'
91 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
92 * 2nd argument should be a pointer to stack, which will be used inside
93 * the helper function as a pointer to map element key.
95 * On the kernel side the helper function looks like:
96 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
98 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
99 * void *key = (void *) (unsigned long) r2;
102 * here kernel can access 'key' and 'map' pointers safely, knowing that
103 * [key, key + map->key_size) bytes are valid and were initialized on
104 * the stack of eBPF program.
107 * Corresponding eBPF program may look like:
108 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
109 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
110 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
111 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
112 * here verifier looks at prototype of map_lookup_elem() and sees:
113 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
114 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
116 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
117 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
118 * and were initialized prior to this call.
119 * If it's ok, then verifier allows this BPF_CALL insn and looks at
120 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
121 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
122 * returns ether pointer to map value or NULL.
124 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
125 * insn, the register holding that pointer in the true branch changes state to
126 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
127 * branch. See check_cond_jmp_op().
129 * After the call R0 is set to return type of the function and registers R1-R5
130 * are set to NOT_INIT to indicate that they are no longer readable.
133 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
134 struct bpf_verifier_stack_elem
{
135 /* verifer state is 'st'
136 * before processing instruction 'insn_idx'
137 * and after processing instruction 'prev_insn_idx'
139 struct bpf_verifier_state st
;
142 struct bpf_verifier_stack_elem
*next
;
145 #define BPF_COMPLEXITY_LIMIT_INSNS 131072
146 #define BPF_COMPLEXITY_LIMIT_STACK 1024
148 #define BPF_MAP_PTR_POISON ((void *)0xeB9F + POISON_POINTER_DELTA)
150 struct bpf_call_arg_meta
{
151 struct bpf_map
*map_ptr
;
158 static DEFINE_MUTEX(bpf_verifier_lock
);
160 /* log_level controls verbosity level of eBPF verifier.
161 * verbose() is used to dump the verification trace to the log, so the user
162 * can figure out what's wrong with the program
164 static __printf(2, 3) void verbose(struct bpf_verifier_env
*env
,
165 const char *fmt
, ...)
167 struct bpf_verifer_log
*log
= &env
->log
;
171 if (!log
->level
|| !log
->ubuf
|| bpf_verifier_log_full(log
))
175 n
= vscnprintf(log
->kbuf
, BPF_VERIFIER_TMP_LOG_SIZE
, fmt
, args
);
178 WARN_ONCE(n
>= BPF_VERIFIER_TMP_LOG_SIZE
- 1,
179 "verifier log line truncated - local buffer too short\n");
181 n
= min(log
->len_total
- log
->len_used
- 1, n
);
184 if (!copy_to_user(log
->ubuf
+ log
->len_used
, log
->kbuf
, n
+ 1))
190 static bool type_is_pkt_pointer(enum bpf_reg_type type
)
192 return type
== PTR_TO_PACKET
||
193 type
== PTR_TO_PACKET_META
;
196 /* string representation of 'enum bpf_reg_type' */
197 static const char * const reg_type_str
[] = {
199 [SCALAR_VALUE
] = "inv",
200 [PTR_TO_CTX
] = "ctx",
201 [CONST_PTR_TO_MAP
] = "map_ptr",
202 [PTR_TO_MAP_VALUE
] = "map_value",
203 [PTR_TO_MAP_VALUE_OR_NULL
] = "map_value_or_null",
204 [PTR_TO_STACK
] = "fp",
205 [PTR_TO_PACKET
] = "pkt",
206 [PTR_TO_PACKET_META
] = "pkt_meta",
207 [PTR_TO_PACKET_END
] = "pkt_end",
210 static void print_verifier_state(struct bpf_verifier_env
*env
,
211 struct bpf_verifier_state
*state
)
213 struct bpf_reg_state
*reg
;
217 for (i
= 0; i
< MAX_BPF_REG
; i
++) {
218 reg
= &state
->regs
[i
];
222 verbose(env
, " R%d=%s", i
, reg_type_str
[t
]);
223 if ((t
== SCALAR_VALUE
|| t
== PTR_TO_STACK
) &&
224 tnum_is_const(reg
->var_off
)) {
225 /* reg->off should be 0 for SCALAR_VALUE */
226 verbose(env
, "%lld", reg
->var_off
.value
+ reg
->off
);
228 verbose(env
, "(id=%d", reg
->id
);
229 if (t
!= SCALAR_VALUE
)
230 verbose(env
, ",off=%d", reg
->off
);
231 if (type_is_pkt_pointer(t
))
232 verbose(env
, ",r=%d", reg
->range
);
233 else if (t
== CONST_PTR_TO_MAP
||
234 t
== PTR_TO_MAP_VALUE
||
235 t
== PTR_TO_MAP_VALUE_OR_NULL
)
236 verbose(env
, ",ks=%d,vs=%d",
237 reg
->map_ptr
->key_size
,
238 reg
->map_ptr
->value_size
);
239 if (tnum_is_const(reg
->var_off
)) {
240 /* Typically an immediate SCALAR_VALUE, but
241 * could be a pointer whose offset is too big
244 verbose(env
, ",imm=%llx", reg
->var_off
.value
);
246 if (reg
->smin_value
!= reg
->umin_value
&&
247 reg
->smin_value
!= S64_MIN
)
248 verbose(env
, ",smin_value=%lld",
249 (long long)reg
->smin_value
);
250 if (reg
->smax_value
!= reg
->umax_value
&&
251 reg
->smax_value
!= S64_MAX
)
252 verbose(env
, ",smax_value=%lld",
253 (long long)reg
->smax_value
);
254 if (reg
->umin_value
!= 0)
255 verbose(env
, ",umin_value=%llu",
256 (unsigned long long)reg
->umin_value
);
257 if (reg
->umax_value
!= U64_MAX
)
258 verbose(env
, ",umax_value=%llu",
259 (unsigned long long)reg
->umax_value
);
260 if (!tnum_is_unknown(reg
->var_off
)) {
263 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
264 verbose(env
, ",var_off=%s", tn_buf
);
270 for (i
= 0; i
< MAX_BPF_STACK
; i
+= BPF_REG_SIZE
) {
271 if (state
->stack_slot_type
[i
] == STACK_SPILL
)
272 verbose(env
, " fp%d=%s", -MAX_BPF_STACK
+ i
,
273 reg_type_str
[state
->spilled_regs
[i
/ BPF_REG_SIZE
].type
]);
278 static int pop_stack(struct bpf_verifier_env
*env
, int *prev_insn_idx
)
280 struct bpf_verifier_stack_elem
*elem
;
283 if (env
->head
== NULL
)
286 memcpy(&env
->cur_state
, &env
->head
->st
, sizeof(env
->cur_state
));
287 insn_idx
= env
->head
->insn_idx
;
289 *prev_insn_idx
= env
->head
->prev_insn_idx
;
290 elem
= env
->head
->next
;
297 static struct bpf_verifier_state
*push_stack(struct bpf_verifier_env
*env
,
298 int insn_idx
, int prev_insn_idx
)
300 struct bpf_verifier_stack_elem
*elem
;
302 elem
= kmalloc(sizeof(struct bpf_verifier_stack_elem
), GFP_KERNEL
);
306 memcpy(&elem
->st
, &env
->cur_state
, sizeof(env
->cur_state
));
307 elem
->insn_idx
= insn_idx
;
308 elem
->prev_insn_idx
= prev_insn_idx
;
309 elem
->next
= env
->head
;
312 if (env
->stack_size
> BPF_COMPLEXITY_LIMIT_STACK
) {
313 verbose(env
, "BPF program is too complex\n");
318 /* pop all elements and return */
319 while (pop_stack(env
, NULL
) >= 0);
323 #define CALLER_SAVED_REGS 6
324 static const int caller_saved
[CALLER_SAVED_REGS
] = {
325 BPF_REG_0
, BPF_REG_1
, BPF_REG_2
, BPF_REG_3
, BPF_REG_4
, BPF_REG_5
328 static void __mark_reg_not_init(struct bpf_reg_state
*reg
);
330 /* Mark the unknown part of a register (variable offset or scalar value) as
331 * known to have the value @imm.
333 static void __mark_reg_known(struct bpf_reg_state
*reg
, u64 imm
)
336 reg
->var_off
= tnum_const(imm
);
337 reg
->smin_value
= (s64
)imm
;
338 reg
->smax_value
= (s64
)imm
;
339 reg
->umin_value
= imm
;
340 reg
->umax_value
= imm
;
343 /* Mark the 'variable offset' part of a register as zero. This should be
344 * used only on registers holding a pointer type.
346 static void __mark_reg_known_zero(struct bpf_reg_state
*reg
)
348 __mark_reg_known(reg
, 0);
351 static void mark_reg_known_zero(struct bpf_verifier_env
*env
,
352 struct bpf_reg_state
*regs
, u32 regno
)
354 if (WARN_ON(regno
>= MAX_BPF_REG
)) {
355 verbose(env
, "mark_reg_known_zero(regs, %u)\n", regno
);
356 /* Something bad happened, let's kill all regs */
357 for (regno
= 0; regno
< MAX_BPF_REG
; regno
++)
358 __mark_reg_not_init(regs
+ regno
);
361 __mark_reg_known_zero(regs
+ regno
);
364 static bool reg_is_pkt_pointer(const struct bpf_reg_state
*reg
)
366 return type_is_pkt_pointer(reg
->type
);
369 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state
*reg
)
371 return reg_is_pkt_pointer(reg
) ||
372 reg
->type
== PTR_TO_PACKET_END
;
375 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
376 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state
*reg
,
377 enum bpf_reg_type which
)
379 /* The register can already have a range from prior markings.
380 * This is fine as long as it hasn't been advanced from its
383 return reg
->type
== which
&&
386 tnum_equals_const(reg
->var_off
, 0);
389 /* Attempts to improve min/max values based on var_off information */
390 static void __update_reg_bounds(struct bpf_reg_state
*reg
)
392 /* min signed is max(sign bit) | min(other bits) */
393 reg
->smin_value
= max_t(s64
, reg
->smin_value
,
394 reg
->var_off
.value
| (reg
->var_off
.mask
& S64_MIN
));
395 /* max signed is min(sign bit) | max(other bits) */
396 reg
->smax_value
= min_t(s64
, reg
->smax_value
,
397 reg
->var_off
.value
| (reg
->var_off
.mask
& S64_MAX
));
398 reg
->umin_value
= max(reg
->umin_value
, reg
->var_off
.value
);
399 reg
->umax_value
= min(reg
->umax_value
,
400 reg
->var_off
.value
| reg
->var_off
.mask
);
403 /* Uses signed min/max values to inform unsigned, and vice-versa */
404 static void __reg_deduce_bounds(struct bpf_reg_state
*reg
)
406 /* Learn sign from signed bounds.
407 * If we cannot cross the sign boundary, then signed and unsigned bounds
408 * are the same, so combine. This works even in the negative case, e.g.
409 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
411 if (reg
->smin_value
>= 0 || reg
->smax_value
< 0) {
412 reg
->smin_value
= reg
->umin_value
= max_t(u64
, reg
->smin_value
,
414 reg
->smax_value
= reg
->umax_value
= min_t(u64
, reg
->smax_value
,
418 /* Learn sign from unsigned bounds. Signed bounds cross the sign
419 * boundary, so we must be careful.
421 if ((s64
)reg
->umax_value
>= 0) {
422 /* Positive. We can't learn anything from the smin, but smax
423 * is positive, hence safe.
425 reg
->smin_value
= reg
->umin_value
;
426 reg
->smax_value
= reg
->umax_value
= min_t(u64
, reg
->smax_value
,
428 } else if ((s64
)reg
->umin_value
< 0) {
429 /* Negative. We can't learn anything from the smax, but smin
430 * is negative, hence safe.
432 reg
->smin_value
= reg
->umin_value
= max_t(u64
, reg
->smin_value
,
434 reg
->smax_value
= reg
->umax_value
;
438 /* Attempts to improve var_off based on unsigned min/max information */
439 static void __reg_bound_offset(struct bpf_reg_state
*reg
)
441 reg
->var_off
= tnum_intersect(reg
->var_off
,
442 tnum_range(reg
->umin_value
,
446 /* Reset the min/max bounds of a register */
447 static void __mark_reg_unbounded(struct bpf_reg_state
*reg
)
449 reg
->smin_value
= S64_MIN
;
450 reg
->smax_value
= S64_MAX
;
452 reg
->umax_value
= U64_MAX
;
455 /* Mark a register as having a completely unknown (scalar) value. */
456 static void __mark_reg_unknown(struct bpf_reg_state
*reg
)
458 reg
->type
= SCALAR_VALUE
;
461 reg
->var_off
= tnum_unknown
;
462 __mark_reg_unbounded(reg
);
465 static void mark_reg_unknown(struct bpf_verifier_env
*env
,
466 struct bpf_reg_state
*regs
, u32 regno
)
468 if (WARN_ON(regno
>= MAX_BPF_REG
)) {
469 verbose(env
, "mark_reg_unknown(regs, %u)\n", regno
);
470 /* Something bad happened, let's kill all regs */
471 for (regno
= 0; regno
< MAX_BPF_REG
; regno
++)
472 __mark_reg_not_init(regs
+ regno
);
475 __mark_reg_unknown(regs
+ regno
);
478 static void __mark_reg_not_init(struct bpf_reg_state
*reg
)
480 __mark_reg_unknown(reg
);
481 reg
->type
= NOT_INIT
;
484 static void mark_reg_not_init(struct bpf_verifier_env
*env
,
485 struct bpf_reg_state
*regs
, u32 regno
)
487 if (WARN_ON(regno
>= MAX_BPF_REG
)) {
488 verbose(env
, "mark_reg_not_init(regs, %u)\n", regno
);
489 /* Something bad happened, let's kill all regs */
490 for (regno
= 0; regno
< MAX_BPF_REG
; regno
++)
491 __mark_reg_not_init(regs
+ regno
);
494 __mark_reg_not_init(regs
+ regno
);
497 static void init_reg_state(struct bpf_verifier_env
*env
,
498 struct bpf_reg_state
*regs
)
502 for (i
= 0; i
< MAX_BPF_REG
; i
++) {
503 mark_reg_not_init(env
, regs
, i
);
504 regs
[i
].live
= REG_LIVE_NONE
;
508 regs
[BPF_REG_FP
].type
= PTR_TO_STACK
;
509 mark_reg_known_zero(env
, regs
, BPF_REG_FP
);
511 /* 1st arg to a function */
512 regs
[BPF_REG_1
].type
= PTR_TO_CTX
;
513 mark_reg_known_zero(env
, regs
, BPF_REG_1
);
517 SRC_OP
, /* register is used as source operand */
518 DST_OP
, /* register is used as destination operand */
519 DST_OP_NO_MARK
/* same as above, check only, don't mark */
522 static void mark_reg_read(const struct bpf_verifier_state
*state
, u32 regno
)
524 struct bpf_verifier_state
*parent
= state
->parent
;
526 if (regno
== BPF_REG_FP
)
527 /* We don't need to worry about FP liveness because it's read-only */
531 /* if read wasn't screened by an earlier write ... */
532 if (state
->regs
[regno
].live
& REG_LIVE_WRITTEN
)
534 /* ... then we depend on parent's value */
535 parent
->regs
[regno
].live
|= REG_LIVE_READ
;
537 parent
= state
->parent
;
541 static int check_reg_arg(struct bpf_verifier_env
*env
, u32 regno
,
544 struct bpf_reg_state
*regs
= env
->cur_state
.regs
;
546 if (regno
>= MAX_BPF_REG
) {
547 verbose(env
, "R%d is invalid\n", regno
);
552 /* check whether register used as source operand can be read */
553 if (regs
[regno
].type
== NOT_INIT
) {
554 verbose(env
, "R%d !read_ok\n", regno
);
557 mark_reg_read(&env
->cur_state
, regno
);
559 /* check whether register used as dest operand can be written to */
560 if (regno
== BPF_REG_FP
) {
561 verbose(env
, "frame pointer is read only\n");
564 regs
[regno
].live
|= REG_LIVE_WRITTEN
;
566 mark_reg_unknown(env
, regs
, regno
);
571 static bool is_spillable_regtype(enum bpf_reg_type type
)
574 case PTR_TO_MAP_VALUE
:
575 case PTR_TO_MAP_VALUE_OR_NULL
:
579 case PTR_TO_PACKET_META
:
580 case PTR_TO_PACKET_END
:
581 case CONST_PTR_TO_MAP
:
588 /* check_stack_read/write functions track spill/fill of registers,
589 * stack boundary and alignment are checked in check_mem_access()
591 static int check_stack_write(struct bpf_verifier_env
*env
,
592 struct bpf_verifier_state
*state
, int off
,
593 int size
, int value_regno
)
595 int i
, spi
= (MAX_BPF_STACK
+ off
) / BPF_REG_SIZE
;
596 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
597 * so it's aligned access and [off, off + size) are within stack limits
600 if (value_regno
>= 0 &&
601 is_spillable_regtype(state
->regs
[value_regno
].type
)) {
603 /* register containing pointer is being spilled into stack */
604 if (size
!= BPF_REG_SIZE
) {
605 verbose(env
, "invalid size of register spill\n");
609 /* save register state */
610 state
->spilled_regs
[spi
] = state
->regs
[value_regno
];
611 state
->spilled_regs
[spi
].live
|= REG_LIVE_WRITTEN
;
613 for (i
= 0; i
< BPF_REG_SIZE
; i
++)
614 state
->stack_slot_type
[MAX_BPF_STACK
+ off
+ i
] = STACK_SPILL
;
616 /* regular write of data into stack */
617 state
->spilled_regs
[spi
] = (struct bpf_reg_state
) {};
619 for (i
= 0; i
< size
; i
++)
620 state
->stack_slot_type
[MAX_BPF_STACK
+ off
+ i
] = STACK_MISC
;
625 static void mark_stack_slot_read(const struct bpf_verifier_state
*state
, int slot
)
627 struct bpf_verifier_state
*parent
= state
->parent
;
630 /* if read wasn't screened by an earlier write ... */
631 if (state
->spilled_regs
[slot
].live
& REG_LIVE_WRITTEN
)
633 /* ... then we depend on parent's value */
634 parent
->spilled_regs
[slot
].live
|= REG_LIVE_READ
;
636 parent
= state
->parent
;
640 static int check_stack_read(struct bpf_verifier_env
*env
,
641 struct bpf_verifier_state
*state
, int off
, int size
,
647 slot_type
= &state
->stack_slot_type
[MAX_BPF_STACK
+ off
];
649 if (slot_type
[0] == STACK_SPILL
) {
650 if (size
!= BPF_REG_SIZE
) {
651 verbose(env
, "invalid size of register spill\n");
654 for (i
= 1; i
< BPF_REG_SIZE
; i
++) {
655 if (slot_type
[i
] != STACK_SPILL
) {
656 verbose(env
, "corrupted spill memory\n");
661 spi
= (MAX_BPF_STACK
+ off
) / BPF_REG_SIZE
;
663 if (value_regno
>= 0) {
664 /* restore register state from stack */
665 state
->regs
[value_regno
] = state
->spilled_regs
[spi
];
666 mark_stack_slot_read(state
, spi
);
670 for (i
= 0; i
< size
; i
++) {
671 if (slot_type
[i
] != STACK_MISC
) {
672 verbose(env
, "invalid read from stack off %d+%d size %d\n",
677 if (value_regno
>= 0)
678 /* have read misc data from the stack */
679 mark_reg_unknown(env
, state
->regs
, value_regno
);
684 /* check read/write into map element returned by bpf_map_lookup_elem() */
685 static int __check_map_access(struct bpf_verifier_env
*env
, u32 regno
, int off
,
688 struct bpf_map
*map
= env
->cur_state
.regs
[regno
].map_ptr
;
690 if (off
< 0 || size
<= 0 || off
+ size
> map
->value_size
) {
691 verbose(env
, "invalid access to map value, value_size=%d off=%d size=%d\n",
692 map
->value_size
, off
, size
);
698 /* check read/write into a map element with possible variable offset */
699 static int check_map_access(struct bpf_verifier_env
*env
, u32 regno
,
702 struct bpf_verifier_state
*state
= &env
->cur_state
;
703 struct bpf_reg_state
*reg
= &state
->regs
[regno
];
706 /* We may have adjusted the register to this map value, so we
707 * need to try adding each of min_value and max_value to off
708 * to make sure our theoretical access will be safe.
711 print_verifier_state(env
, state
);
712 /* The minimum value is only important with signed
713 * comparisons where we can't assume the floor of a
714 * value is 0. If we are using signed variables for our
715 * index'es we need to make sure that whatever we use
716 * will have a set floor within our range.
718 if (reg
->smin_value
< 0) {
719 verbose(env
, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
723 err
= __check_map_access(env
, regno
, reg
->smin_value
+ off
, size
);
725 verbose(env
, "R%d min value is outside of the array range\n",
730 /* If we haven't set a max value then we need to bail since we can't be
731 * sure we won't do bad things.
732 * If reg->umax_value + off could overflow, treat that as unbounded too.
734 if (reg
->umax_value
>= BPF_MAX_VAR_OFF
) {
735 verbose(env
, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
739 err
= __check_map_access(env
, regno
, reg
->umax_value
+ off
, size
);
741 verbose(env
, "R%d max value is outside of the array range\n",
746 #define MAX_PACKET_OFF 0xffff
748 static bool may_access_direct_pkt_data(struct bpf_verifier_env
*env
,
749 const struct bpf_call_arg_meta
*meta
,
750 enum bpf_access_type t
)
752 switch (env
->prog
->type
) {
753 case BPF_PROG_TYPE_LWT_IN
:
754 case BPF_PROG_TYPE_LWT_OUT
:
755 /* dst_input() and dst_output() can't write for now */
759 case BPF_PROG_TYPE_SCHED_CLS
:
760 case BPF_PROG_TYPE_SCHED_ACT
:
761 case BPF_PROG_TYPE_XDP
:
762 case BPF_PROG_TYPE_LWT_XMIT
:
763 case BPF_PROG_TYPE_SK_SKB
:
765 return meta
->pkt_access
;
767 env
->seen_direct_write
= true;
774 static int __check_packet_access(struct bpf_verifier_env
*env
, u32 regno
,
777 struct bpf_reg_state
*regs
= env
->cur_state
.regs
;
778 struct bpf_reg_state
*reg
= ®s
[regno
];
780 if (off
< 0 || size
<= 0 || (u64
)off
+ size
> reg
->range
) {
781 verbose(env
, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
782 off
, size
, regno
, reg
->id
, reg
->off
, reg
->range
);
788 static int check_packet_access(struct bpf_verifier_env
*env
, u32 regno
, int off
,
791 struct bpf_reg_state
*regs
= env
->cur_state
.regs
;
792 struct bpf_reg_state
*reg
= ®s
[regno
];
795 /* We may have added a variable offset to the packet pointer; but any
796 * reg->range we have comes after that. We are only checking the fixed
800 /* We don't allow negative numbers, because we aren't tracking enough
801 * detail to prove they're safe.
803 if (reg
->smin_value
< 0) {
804 verbose(env
, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
808 err
= __check_packet_access(env
, regno
, off
, size
);
810 verbose(env
, "R%d offset is outside of the packet\n", regno
);
816 static bool analyzer_is_valid_access(struct bpf_verifier_env
*env
, int off
,
817 struct bpf_insn_access_aux
*info
)
819 switch (env
->prog
->type
) {
820 case BPF_PROG_TYPE_XDP
:
822 case offsetof(struct xdp_buff
, data
):
823 info
->reg_type
= PTR_TO_PACKET
;
825 case offsetof(struct xdp_buff
, data_end
):
826 info
->reg_type
= PTR_TO_PACKET_END
;
830 case BPF_PROG_TYPE_SCHED_CLS
:
832 case offsetof(struct sk_buff
, data
):
833 info
->reg_type
= PTR_TO_PACKET
;
835 case offsetof(struct sk_buff
, cb
) +
836 offsetof(struct bpf_skb_data_end
, data_end
):
837 info
->reg_type
= PTR_TO_PACKET_END
;
846 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
847 static int check_ctx_access(struct bpf_verifier_env
*env
, int insn_idx
, int off
, int size
,
848 enum bpf_access_type t
, enum bpf_reg_type
*reg_type
)
850 struct bpf_insn_access_aux info
= {
851 .reg_type
= *reg_type
,
854 if (env
->analyzer_ops
) {
855 if (analyzer_is_valid_access(env
, off
, &info
)) {
856 *reg_type
= info
.reg_type
;
859 } else if (env
->prog
->aux
->ops
->is_valid_access
&&
860 env
->prog
->aux
->ops
->is_valid_access(off
, size
, t
, &info
)) {
861 /* A non zero info.ctx_field_size indicates that this field is a
862 * candidate for later verifier transformation to load the whole
863 * field and then apply a mask when accessed with a narrower
864 * access than actual ctx access size. A zero info.ctx_field_size
865 * will only allow for whole field access and rejects any other
866 * type of narrower access.
868 env
->insn_aux_data
[insn_idx
].ctx_field_size
= info
.ctx_field_size
;
869 *reg_type
= info
.reg_type
;
871 /* remember the offset of last byte accessed in ctx */
872 if (env
->prog
->aux
->max_ctx_offset
< off
+ size
)
873 env
->prog
->aux
->max_ctx_offset
= off
+ size
;
877 verbose(env
, "invalid bpf_context access off=%d size=%d\n", off
, size
);
881 static bool __is_pointer_value(bool allow_ptr_leaks
,
882 const struct bpf_reg_state
*reg
)
887 return reg
->type
!= SCALAR_VALUE
;
890 static bool is_pointer_value(struct bpf_verifier_env
*env
, int regno
)
892 return __is_pointer_value(env
->allow_ptr_leaks
, &env
->cur_state
.regs
[regno
]);
895 static int check_pkt_ptr_alignment(struct bpf_verifier_env
*env
,
896 const struct bpf_reg_state
*reg
,
897 int off
, int size
, bool strict
)
902 /* Byte size accesses are always allowed. */
903 if (!strict
|| size
== 1)
906 /* For platforms that do not have a Kconfig enabling
907 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
908 * NET_IP_ALIGN is universally set to '2'. And on platforms
909 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
910 * to this code only in strict mode where we want to emulate
911 * the NET_IP_ALIGN==2 checking. Therefore use an
912 * unconditional IP align value of '2'.
916 reg_off
= tnum_add(reg
->var_off
, tnum_const(ip_align
+ reg
->off
+ off
));
917 if (!tnum_is_aligned(reg_off
, size
)) {
920 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
922 "misaligned packet access off %d+%s+%d+%d size %d\n",
923 ip_align
, tn_buf
, reg
->off
, off
, size
);
930 static int check_generic_ptr_alignment(struct bpf_verifier_env
*env
,
931 const struct bpf_reg_state
*reg
,
932 const char *pointer_desc
,
933 int off
, int size
, bool strict
)
937 /* Byte size accesses are always allowed. */
938 if (!strict
|| size
== 1)
941 reg_off
= tnum_add(reg
->var_off
, tnum_const(reg
->off
+ off
));
942 if (!tnum_is_aligned(reg_off
, size
)) {
945 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
946 verbose(env
, "misaligned %saccess off %s+%d+%d size %d\n",
947 pointer_desc
, tn_buf
, reg
->off
, off
, size
);
954 static int check_ptr_alignment(struct bpf_verifier_env
*env
,
955 const struct bpf_reg_state
*reg
,
958 bool strict
= env
->strict_alignment
;
959 const char *pointer_desc
= "";
963 case PTR_TO_PACKET_META
:
964 /* Special case, because of NET_IP_ALIGN. Given metadata sits
965 * right in front, treat it the very same way.
967 return check_pkt_ptr_alignment(env
, reg
, off
, size
, strict
);
968 case PTR_TO_MAP_VALUE
:
969 pointer_desc
= "value ";
972 pointer_desc
= "context ";
975 pointer_desc
= "stack ";
980 return check_generic_ptr_alignment(env
, reg
, pointer_desc
, off
, size
,
984 /* check whether memory at (regno + off) is accessible for t = (read | write)
985 * if t==write, value_regno is a register which value is stored into memory
986 * if t==read, value_regno is a register which will receive the value from memory
987 * if t==write && value_regno==-1, some unknown value is stored into memory
988 * if t==read && value_regno==-1, don't care what we read from memory
990 static int check_mem_access(struct bpf_verifier_env
*env
, int insn_idx
, u32 regno
, int off
,
991 int bpf_size
, enum bpf_access_type t
,
994 struct bpf_verifier_state
*state
= &env
->cur_state
;
995 struct bpf_reg_state
*reg
= &state
->regs
[regno
];
998 size
= bpf_size_to_bytes(bpf_size
);
1002 /* alignment checks will add in reg->off themselves */
1003 err
= check_ptr_alignment(env
, reg
, off
, size
);
1007 /* for access checks, reg->off is just part of off */
1010 if (reg
->type
== PTR_TO_MAP_VALUE
) {
1011 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
1012 is_pointer_value(env
, value_regno
)) {
1013 verbose(env
, "R%d leaks addr into map\n", value_regno
);
1017 err
= check_map_access(env
, regno
, off
, size
);
1018 if (!err
&& t
== BPF_READ
&& value_regno
>= 0)
1019 mark_reg_unknown(env
, state
->regs
, value_regno
);
1021 } else if (reg
->type
== PTR_TO_CTX
) {
1022 enum bpf_reg_type reg_type
= SCALAR_VALUE
;
1024 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
1025 is_pointer_value(env
, value_regno
)) {
1026 verbose(env
, "R%d leaks addr into ctx\n", value_regno
);
1029 /* ctx accesses must be at a fixed offset, so that we can
1030 * determine what type of data were returned.
1032 if (!tnum_is_const(reg
->var_off
)) {
1035 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1037 "variable ctx access var_off=%s off=%d size=%d",
1041 off
+= reg
->var_off
.value
;
1042 err
= check_ctx_access(env
, insn_idx
, off
, size
, t
, ®_type
);
1043 if (!err
&& t
== BPF_READ
&& value_regno
>= 0) {
1044 /* ctx access returns either a scalar, or a
1045 * PTR_TO_PACKET[_META,_END]. In the latter
1046 * case, we know the offset is zero.
1048 if (reg_type
== SCALAR_VALUE
)
1049 mark_reg_unknown(env
, state
->regs
, value_regno
);
1051 mark_reg_known_zero(env
, state
->regs
,
1053 state
->regs
[value_regno
].id
= 0;
1054 state
->regs
[value_regno
].off
= 0;
1055 state
->regs
[value_regno
].range
= 0;
1056 state
->regs
[value_regno
].type
= reg_type
;
1059 } else if (reg
->type
== PTR_TO_STACK
) {
1060 /* stack accesses must be at a fixed offset, so that we can
1061 * determine what type of data were returned.
1062 * See check_stack_read().
1064 if (!tnum_is_const(reg
->var_off
)) {
1067 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1068 verbose(env
, "variable stack access var_off=%s off=%d size=%d",
1072 off
+= reg
->var_off
.value
;
1073 if (off
>= 0 || off
< -MAX_BPF_STACK
) {
1074 verbose(env
, "invalid stack off=%d size=%d\n", off
,
1079 if (env
->prog
->aux
->stack_depth
< -off
)
1080 env
->prog
->aux
->stack_depth
= -off
;
1082 if (t
== BPF_WRITE
) {
1083 if (!env
->allow_ptr_leaks
&&
1084 state
->stack_slot_type
[MAX_BPF_STACK
+ off
] == STACK_SPILL
&&
1085 size
!= BPF_REG_SIZE
) {
1086 verbose(env
, "attempt to corrupt spilled pointer on stack\n");
1089 err
= check_stack_write(env
, state
, off
, size
,
1092 err
= check_stack_read(env
, state
, off
, size
,
1095 } else if (reg_is_pkt_pointer(reg
)) {
1096 if (t
== BPF_WRITE
&& !may_access_direct_pkt_data(env
, NULL
, t
)) {
1097 verbose(env
, "cannot write into packet\n");
1100 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
1101 is_pointer_value(env
, value_regno
)) {
1102 verbose(env
, "R%d leaks addr into packet\n",
1106 err
= check_packet_access(env
, regno
, off
, size
);
1107 if (!err
&& t
== BPF_READ
&& value_regno
>= 0)
1108 mark_reg_unknown(env
, state
->regs
, value_regno
);
1110 verbose(env
, "R%d invalid mem access '%s'\n", regno
,
1111 reg_type_str
[reg
->type
]);
1115 if (!err
&& size
< BPF_REG_SIZE
&& value_regno
>= 0 && t
== BPF_READ
&&
1116 state
->regs
[value_regno
].type
== SCALAR_VALUE
) {
1117 /* b/h/w load zero-extends, mark upper bits as known 0 */
1118 state
->regs
[value_regno
].var_off
= tnum_cast(
1119 state
->regs
[value_regno
].var_off
, size
);
1120 __update_reg_bounds(&state
->regs
[value_regno
]);
1125 static int check_xadd(struct bpf_verifier_env
*env
, int insn_idx
, struct bpf_insn
*insn
)
1129 if ((BPF_SIZE(insn
->code
) != BPF_W
&& BPF_SIZE(insn
->code
) != BPF_DW
) ||
1131 verbose(env
, "BPF_XADD uses reserved fields\n");
1135 /* check src1 operand */
1136 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
1140 /* check src2 operand */
1141 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
1145 if (is_pointer_value(env
, insn
->src_reg
)) {
1146 verbose(env
, "R%d leaks addr into mem\n", insn
->src_reg
);
1150 /* check whether atomic_add can read the memory */
1151 err
= check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
1152 BPF_SIZE(insn
->code
), BPF_READ
, -1);
1156 /* check whether atomic_add can write into the same memory */
1157 return check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
1158 BPF_SIZE(insn
->code
), BPF_WRITE
, -1);
1161 /* Does this register contain a constant zero? */
1162 static bool register_is_null(struct bpf_reg_state reg
)
1164 return reg
.type
== SCALAR_VALUE
&& tnum_equals_const(reg
.var_off
, 0);
1167 /* when register 'regno' is passed into function that will read 'access_size'
1168 * bytes from that pointer, make sure that it's within stack boundary
1169 * and all elements of stack are initialized.
1170 * Unlike most pointer bounds-checking functions, this one doesn't take an
1171 * 'off' argument, so it has to add in reg->off itself.
1173 static int check_stack_boundary(struct bpf_verifier_env
*env
, int regno
,
1174 int access_size
, bool zero_size_allowed
,
1175 struct bpf_call_arg_meta
*meta
)
1177 struct bpf_verifier_state
*state
= &env
->cur_state
;
1178 struct bpf_reg_state
*regs
= state
->regs
;
1181 if (regs
[regno
].type
!= PTR_TO_STACK
) {
1182 /* Allow zero-byte read from NULL, regardless of pointer type */
1183 if (zero_size_allowed
&& access_size
== 0 &&
1184 register_is_null(regs
[regno
]))
1187 verbose(env
, "R%d type=%s expected=%s\n", regno
,
1188 reg_type_str
[regs
[regno
].type
],
1189 reg_type_str
[PTR_TO_STACK
]);
1193 /* Only allow fixed-offset stack reads */
1194 if (!tnum_is_const(regs
[regno
].var_off
)) {
1197 tnum_strn(tn_buf
, sizeof(tn_buf
), regs
[regno
].var_off
);
1198 verbose(env
, "invalid variable stack read R%d var_off=%s\n",
1201 off
= regs
[regno
].off
+ regs
[regno
].var_off
.value
;
1202 if (off
>= 0 || off
< -MAX_BPF_STACK
|| off
+ access_size
> 0 ||
1204 verbose(env
, "invalid stack type R%d off=%d access_size=%d\n",
1205 regno
, off
, access_size
);
1209 if (env
->prog
->aux
->stack_depth
< -off
)
1210 env
->prog
->aux
->stack_depth
= -off
;
1212 if (meta
&& meta
->raw_mode
) {
1213 meta
->access_size
= access_size
;
1214 meta
->regno
= regno
;
1218 for (i
= 0; i
< access_size
; i
++) {
1219 if (state
->stack_slot_type
[MAX_BPF_STACK
+ off
+ i
] != STACK_MISC
) {
1220 verbose(env
, "invalid indirect read from stack off %d+%d size %d\n",
1221 off
, i
, access_size
);
1228 static int check_helper_mem_access(struct bpf_verifier_env
*env
, int regno
,
1229 int access_size
, bool zero_size_allowed
,
1230 struct bpf_call_arg_meta
*meta
)
1232 struct bpf_reg_state
*regs
= env
->cur_state
.regs
, *reg
= ®s
[regno
];
1234 switch (reg
->type
) {
1236 case PTR_TO_PACKET_META
:
1237 return check_packet_access(env
, regno
, reg
->off
, access_size
);
1238 case PTR_TO_MAP_VALUE
:
1239 return check_map_access(env
, regno
, reg
->off
, access_size
);
1240 default: /* scalar_value|ptr_to_stack or invalid ptr */
1241 return check_stack_boundary(env
, regno
, access_size
,
1242 zero_size_allowed
, meta
);
1246 static int check_func_arg(struct bpf_verifier_env
*env
, u32 regno
,
1247 enum bpf_arg_type arg_type
,
1248 struct bpf_call_arg_meta
*meta
)
1250 struct bpf_reg_state
*regs
= env
->cur_state
.regs
, *reg
= ®s
[regno
];
1251 enum bpf_reg_type expected_type
, type
= reg
->type
;
1254 if (arg_type
== ARG_DONTCARE
)
1257 err
= check_reg_arg(env
, regno
, SRC_OP
);
1261 if (arg_type
== ARG_ANYTHING
) {
1262 if (is_pointer_value(env
, regno
)) {
1263 verbose(env
, "R%d leaks addr into helper function\n",
1270 if (type_is_pkt_pointer(type
) &&
1271 !may_access_direct_pkt_data(env
, meta
, BPF_READ
)) {
1272 verbose(env
, "helper access to the packet is not allowed\n");
1276 if (arg_type
== ARG_PTR_TO_MAP_KEY
||
1277 arg_type
== ARG_PTR_TO_MAP_VALUE
) {
1278 expected_type
= PTR_TO_STACK
;
1279 if (!type_is_pkt_pointer(type
) &&
1280 type
!= expected_type
)
1282 } else if (arg_type
== ARG_CONST_SIZE
||
1283 arg_type
== ARG_CONST_SIZE_OR_ZERO
) {
1284 expected_type
= SCALAR_VALUE
;
1285 if (type
!= expected_type
)
1287 } else if (arg_type
== ARG_CONST_MAP_PTR
) {
1288 expected_type
= CONST_PTR_TO_MAP
;
1289 if (type
!= expected_type
)
1291 } else if (arg_type
== ARG_PTR_TO_CTX
) {
1292 expected_type
= PTR_TO_CTX
;
1293 if (type
!= expected_type
)
1295 } else if (arg_type
== ARG_PTR_TO_MEM
||
1296 arg_type
== ARG_PTR_TO_UNINIT_MEM
) {
1297 expected_type
= PTR_TO_STACK
;
1298 /* One exception here. In case function allows for NULL to be
1299 * passed in as argument, it's a SCALAR_VALUE type. Final test
1300 * happens during stack boundary checking.
1302 if (register_is_null(*reg
))
1303 /* final test in check_stack_boundary() */;
1304 else if (!type_is_pkt_pointer(type
) &&
1305 type
!= PTR_TO_MAP_VALUE
&&
1306 type
!= expected_type
)
1308 meta
->raw_mode
= arg_type
== ARG_PTR_TO_UNINIT_MEM
;
1310 verbose(env
, "unsupported arg_type %d\n", arg_type
);
1314 if (arg_type
== ARG_CONST_MAP_PTR
) {
1315 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
1316 meta
->map_ptr
= reg
->map_ptr
;
1317 } else if (arg_type
== ARG_PTR_TO_MAP_KEY
) {
1318 /* bpf_map_xxx(..., map_ptr, ..., key) call:
1319 * check that [key, key + map->key_size) are within
1320 * stack limits and initialized
1322 if (!meta
->map_ptr
) {
1323 /* in function declaration map_ptr must come before
1324 * map_key, so that it's verified and known before
1325 * we have to check map_key here. Otherwise it means
1326 * that kernel subsystem misconfigured verifier
1328 verbose(env
, "invalid map_ptr to access map->key\n");
1331 if (type_is_pkt_pointer(type
))
1332 err
= check_packet_access(env
, regno
, reg
->off
,
1333 meta
->map_ptr
->key_size
);
1335 err
= check_stack_boundary(env
, regno
,
1336 meta
->map_ptr
->key_size
,
1338 } else if (arg_type
== ARG_PTR_TO_MAP_VALUE
) {
1339 /* bpf_map_xxx(..., map_ptr, ..., value) call:
1340 * check [value, value + map->value_size) validity
1342 if (!meta
->map_ptr
) {
1343 /* kernel subsystem misconfigured verifier */
1344 verbose(env
, "invalid map_ptr to access map->value\n");
1347 if (type_is_pkt_pointer(type
))
1348 err
= check_packet_access(env
, regno
, reg
->off
,
1349 meta
->map_ptr
->value_size
);
1351 err
= check_stack_boundary(env
, regno
,
1352 meta
->map_ptr
->value_size
,
1354 } else if (arg_type
== ARG_CONST_SIZE
||
1355 arg_type
== ARG_CONST_SIZE_OR_ZERO
) {
1356 bool zero_size_allowed
= (arg_type
== ARG_CONST_SIZE_OR_ZERO
);
1358 /* bpf_xxx(..., buf, len) call will access 'len' bytes
1359 * from stack pointer 'buf'. Check it
1360 * note: regno == len, regno - 1 == buf
1363 /* kernel subsystem misconfigured verifier */
1365 "ARG_CONST_SIZE cannot be first argument\n");
1369 /* The register is SCALAR_VALUE; the access check
1370 * happens using its boundaries.
1373 if (!tnum_is_const(reg
->var_off
))
1374 /* For unprivileged variable accesses, disable raw
1375 * mode so that the program is required to
1376 * initialize all the memory that the helper could
1377 * just partially fill up.
1381 if (reg
->smin_value
< 0) {
1382 verbose(env
, "R%d min value is negative, either use unsigned or 'var &= const'\n",
1387 if (reg
->umin_value
== 0) {
1388 err
= check_helper_mem_access(env
, regno
- 1, 0,
1395 if (reg
->umax_value
>= BPF_MAX_VAR_SIZ
) {
1396 verbose(env
, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
1400 err
= check_helper_mem_access(env
, regno
- 1,
1402 zero_size_allowed
, meta
);
1407 verbose(env
, "R%d type=%s expected=%s\n", regno
,
1408 reg_type_str
[type
], reg_type_str
[expected_type
]);
1412 static int check_map_func_compatibility(struct bpf_verifier_env
*env
,
1413 struct bpf_map
*map
, int func_id
)
1418 /* We need a two way check, first is from map perspective ... */
1419 switch (map
->map_type
) {
1420 case BPF_MAP_TYPE_PROG_ARRAY
:
1421 if (func_id
!= BPF_FUNC_tail_call
)
1424 case BPF_MAP_TYPE_PERF_EVENT_ARRAY
:
1425 if (func_id
!= BPF_FUNC_perf_event_read
&&
1426 func_id
!= BPF_FUNC_perf_event_output
&&
1427 func_id
!= BPF_FUNC_perf_event_read_value
)
1430 case BPF_MAP_TYPE_STACK_TRACE
:
1431 if (func_id
!= BPF_FUNC_get_stackid
)
1434 case BPF_MAP_TYPE_CGROUP_ARRAY
:
1435 if (func_id
!= BPF_FUNC_skb_under_cgroup
&&
1436 func_id
!= BPF_FUNC_current_task_under_cgroup
)
1439 /* devmap returns a pointer to a live net_device ifindex that we cannot
1440 * allow to be modified from bpf side. So do not allow lookup elements
1443 case BPF_MAP_TYPE_DEVMAP
:
1444 if (func_id
!= BPF_FUNC_redirect_map
)
1447 case BPF_MAP_TYPE_ARRAY_OF_MAPS
:
1448 case BPF_MAP_TYPE_HASH_OF_MAPS
:
1449 if (func_id
!= BPF_FUNC_map_lookup_elem
)
1452 case BPF_MAP_TYPE_SOCKMAP
:
1453 if (func_id
!= BPF_FUNC_sk_redirect_map
&&
1454 func_id
!= BPF_FUNC_sock_map_update
&&
1455 func_id
!= BPF_FUNC_map_delete_elem
)
1462 /* ... and second from the function itself. */
1464 case BPF_FUNC_tail_call
:
1465 if (map
->map_type
!= BPF_MAP_TYPE_PROG_ARRAY
)
1468 case BPF_FUNC_perf_event_read
:
1469 case BPF_FUNC_perf_event_output
:
1470 case BPF_FUNC_perf_event_read_value
:
1471 if (map
->map_type
!= BPF_MAP_TYPE_PERF_EVENT_ARRAY
)
1474 case BPF_FUNC_get_stackid
:
1475 if (map
->map_type
!= BPF_MAP_TYPE_STACK_TRACE
)
1478 case BPF_FUNC_current_task_under_cgroup
:
1479 case BPF_FUNC_skb_under_cgroup
:
1480 if (map
->map_type
!= BPF_MAP_TYPE_CGROUP_ARRAY
)
1483 case BPF_FUNC_redirect_map
:
1484 if (map
->map_type
!= BPF_MAP_TYPE_DEVMAP
)
1487 case BPF_FUNC_sk_redirect_map
:
1488 if (map
->map_type
!= BPF_MAP_TYPE_SOCKMAP
)
1491 case BPF_FUNC_sock_map_update
:
1492 if (map
->map_type
!= BPF_MAP_TYPE_SOCKMAP
)
1501 verbose(env
, "cannot pass map_type %d into func %s#%d\n",
1502 map
->map_type
, func_id_name(func_id
), func_id
);
1506 static int check_raw_mode(const struct bpf_func_proto
*fn
)
1510 if (fn
->arg1_type
== ARG_PTR_TO_UNINIT_MEM
)
1512 if (fn
->arg2_type
== ARG_PTR_TO_UNINIT_MEM
)
1514 if (fn
->arg3_type
== ARG_PTR_TO_UNINIT_MEM
)
1516 if (fn
->arg4_type
== ARG_PTR_TO_UNINIT_MEM
)
1518 if (fn
->arg5_type
== ARG_PTR_TO_UNINIT_MEM
)
1521 return count
> 1 ? -EINVAL
: 0;
1524 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
1525 * are now invalid, so turn them into unknown SCALAR_VALUE.
1527 static void clear_all_pkt_pointers(struct bpf_verifier_env
*env
)
1529 struct bpf_verifier_state
*state
= &env
->cur_state
;
1530 struct bpf_reg_state
*regs
= state
->regs
, *reg
;
1533 for (i
= 0; i
< MAX_BPF_REG
; i
++)
1534 if (reg_is_pkt_pointer_any(®s
[i
]))
1535 mark_reg_unknown(env
, regs
, i
);
1537 for (i
= 0; i
< MAX_BPF_STACK
; i
+= BPF_REG_SIZE
) {
1538 if (state
->stack_slot_type
[i
] != STACK_SPILL
)
1540 reg
= &state
->spilled_regs
[i
/ BPF_REG_SIZE
];
1541 if (reg_is_pkt_pointer_any(reg
))
1542 __mark_reg_unknown(reg
);
1546 static int check_call(struct bpf_verifier_env
*env
, int func_id
, int insn_idx
)
1548 struct bpf_verifier_state
*state
= &env
->cur_state
;
1549 const struct bpf_func_proto
*fn
= NULL
;
1550 struct bpf_reg_state
*regs
= state
->regs
;
1551 struct bpf_call_arg_meta meta
;
1555 /* find function prototype */
1556 if (func_id
< 0 || func_id
>= __BPF_FUNC_MAX_ID
) {
1557 verbose(env
, "invalid func %s#%d\n", func_id_name(func_id
),
1562 if (env
->prog
->aux
->ops
->get_func_proto
)
1563 fn
= env
->prog
->aux
->ops
->get_func_proto(func_id
);
1566 verbose(env
, "unknown func %s#%d\n", func_id_name(func_id
),
1571 /* eBPF programs must be GPL compatible to use GPL-ed functions */
1572 if (!env
->prog
->gpl_compatible
&& fn
->gpl_only
) {
1573 verbose(env
, "cannot call GPL only function from proprietary program\n");
1577 changes_data
= bpf_helper_changes_pkt_data(fn
->func
);
1579 memset(&meta
, 0, sizeof(meta
));
1580 meta
.pkt_access
= fn
->pkt_access
;
1582 /* We only support one arg being in raw mode at the moment, which
1583 * is sufficient for the helper functions we have right now.
1585 err
= check_raw_mode(fn
);
1587 verbose(env
, "kernel subsystem misconfigured func %s#%d\n",
1588 func_id_name(func_id
), func_id
);
1593 err
= check_func_arg(env
, BPF_REG_1
, fn
->arg1_type
, &meta
);
1596 err
= check_func_arg(env
, BPF_REG_2
, fn
->arg2_type
, &meta
);
1599 err
= check_func_arg(env
, BPF_REG_3
, fn
->arg3_type
, &meta
);
1602 err
= check_func_arg(env
, BPF_REG_4
, fn
->arg4_type
, &meta
);
1605 err
= check_func_arg(env
, BPF_REG_5
, fn
->arg5_type
, &meta
);
1609 /* Mark slots with STACK_MISC in case of raw mode, stack offset
1610 * is inferred from register state.
1612 for (i
= 0; i
< meta
.access_size
; i
++) {
1613 err
= check_mem_access(env
, insn_idx
, meta
.regno
, i
, BPF_B
, BPF_WRITE
, -1);
1618 /* reset caller saved regs */
1619 for (i
= 0; i
< CALLER_SAVED_REGS
; i
++) {
1620 mark_reg_not_init(env
, regs
, caller_saved
[i
]);
1621 check_reg_arg(env
, caller_saved
[i
], DST_OP_NO_MARK
);
1624 /* update return register (already marked as written above) */
1625 if (fn
->ret_type
== RET_INTEGER
) {
1626 /* sets type to SCALAR_VALUE */
1627 mark_reg_unknown(env
, regs
, BPF_REG_0
);
1628 } else if (fn
->ret_type
== RET_VOID
) {
1629 regs
[BPF_REG_0
].type
= NOT_INIT
;
1630 } else if (fn
->ret_type
== RET_PTR_TO_MAP_VALUE_OR_NULL
) {
1631 struct bpf_insn_aux_data
*insn_aux
;
1633 regs
[BPF_REG_0
].type
= PTR_TO_MAP_VALUE_OR_NULL
;
1634 /* There is no offset yet applied, variable or fixed */
1635 mark_reg_known_zero(env
, regs
, BPF_REG_0
);
1636 regs
[BPF_REG_0
].off
= 0;
1637 /* remember map_ptr, so that check_map_access()
1638 * can check 'value_size' boundary of memory access
1639 * to map element returned from bpf_map_lookup_elem()
1641 if (meta
.map_ptr
== NULL
) {
1643 "kernel subsystem misconfigured verifier\n");
1646 regs
[BPF_REG_0
].map_ptr
= meta
.map_ptr
;
1647 regs
[BPF_REG_0
].id
= ++env
->id_gen
;
1648 insn_aux
= &env
->insn_aux_data
[insn_idx
];
1649 if (!insn_aux
->map_ptr
)
1650 insn_aux
->map_ptr
= meta
.map_ptr
;
1651 else if (insn_aux
->map_ptr
!= meta
.map_ptr
)
1652 insn_aux
->map_ptr
= BPF_MAP_PTR_POISON
;
1654 verbose(env
, "unknown return type %d of func %s#%d\n",
1655 fn
->ret_type
, func_id_name(func_id
), func_id
);
1659 err
= check_map_func_compatibility(env
, meta
.map_ptr
, func_id
);
1664 clear_all_pkt_pointers(env
);
1668 static void coerce_reg_to_32(struct bpf_reg_state
*reg
)
1670 /* clear high 32 bits */
1671 reg
->var_off
= tnum_cast(reg
->var_off
, 4);
1673 __update_reg_bounds(reg
);
1676 static bool signed_add_overflows(s64 a
, s64 b
)
1678 /* Do the add in u64, where overflow is well-defined */
1679 s64 res
= (s64
)((u64
)a
+ (u64
)b
);
1686 static bool signed_sub_overflows(s64 a
, s64 b
)
1688 /* Do the sub in u64, where overflow is well-defined */
1689 s64 res
= (s64
)((u64
)a
- (u64
)b
);
1696 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
1697 * Caller should also handle BPF_MOV case separately.
1698 * If we return -EACCES, caller may want to try again treating pointer as a
1699 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
1701 static int adjust_ptr_min_max_vals(struct bpf_verifier_env
*env
,
1702 struct bpf_insn
*insn
,
1703 const struct bpf_reg_state
*ptr_reg
,
1704 const struct bpf_reg_state
*off_reg
)
1706 struct bpf_reg_state
*regs
= env
->cur_state
.regs
, *dst_reg
;
1707 bool known
= tnum_is_const(off_reg
->var_off
);
1708 s64 smin_val
= off_reg
->smin_value
, smax_val
= off_reg
->smax_value
,
1709 smin_ptr
= ptr_reg
->smin_value
, smax_ptr
= ptr_reg
->smax_value
;
1710 u64 umin_val
= off_reg
->umin_value
, umax_val
= off_reg
->umax_value
,
1711 umin_ptr
= ptr_reg
->umin_value
, umax_ptr
= ptr_reg
->umax_value
;
1712 u8 opcode
= BPF_OP(insn
->code
);
1713 u32 dst
= insn
->dst_reg
;
1715 dst_reg
= ®s
[dst
];
1717 if (WARN_ON_ONCE(known
&& (smin_val
!= smax_val
))) {
1718 print_verifier_state(env
, &env
->cur_state
);
1720 "verifier internal error: known but bad sbounds\n");
1723 if (WARN_ON_ONCE(known
&& (umin_val
!= umax_val
))) {
1724 print_verifier_state(env
, &env
->cur_state
);
1726 "verifier internal error: known but bad ubounds\n");
1730 if (BPF_CLASS(insn
->code
) != BPF_ALU64
) {
1731 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
1732 if (!env
->allow_ptr_leaks
)
1734 "R%d 32-bit pointer arithmetic prohibited\n",
1739 if (ptr_reg
->type
== PTR_TO_MAP_VALUE_OR_NULL
) {
1740 if (!env
->allow_ptr_leaks
)
1741 verbose(env
, "R%d pointer arithmetic on PTR_TO_MAP_VALUE_OR_NULL prohibited, null-check it first\n",
1745 if (ptr_reg
->type
== CONST_PTR_TO_MAP
) {
1746 if (!env
->allow_ptr_leaks
)
1747 verbose(env
, "R%d pointer arithmetic on CONST_PTR_TO_MAP prohibited\n",
1751 if (ptr_reg
->type
== PTR_TO_PACKET_END
) {
1752 if (!env
->allow_ptr_leaks
)
1753 verbose(env
, "R%d pointer arithmetic on PTR_TO_PACKET_END prohibited\n",
1758 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
1759 * The id may be overwritten later if we create a new variable offset.
1761 dst_reg
->type
= ptr_reg
->type
;
1762 dst_reg
->id
= ptr_reg
->id
;
1766 /* We can take a fixed offset as long as it doesn't overflow
1767 * the s32 'off' field
1769 if (known
&& (ptr_reg
->off
+ smin_val
==
1770 (s64
)(s32
)(ptr_reg
->off
+ smin_val
))) {
1771 /* pointer += K. Accumulate it into fixed offset */
1772 dst_reg
->smin_value
= smin_ptr
;
1773 dst_reg
->smax_value
= smax_ptr
;
1774 dst_reg
->umin_value
= umin_ptr
;
1775 dst_reg
->umax_value
= umax_ptr
;
1776 dst_reg
->var_off
= ptr_reg
->var_off
;
1777 dst_reg
->off
= ptr_reg
->off
+ smin_val
;
1778 dst_reg
->range
= ptr_reg
->range
;
1781 /* A new variable offset is created. Note that off_reg->off
1782 * == 0, since it's a scalar.
1783 * dst_reg gets the pointer type and since some positive
1784 * integer value was added to the pointer, give it a new 'id'
1785 * if it's a PTR_TO_PACKET.
1786 * this creates a new 'base' pointer, off_reg (variable) gets
1787 * added into the variable offset, and we copy the fixed offset
1790 if (signed_add_overflows(smin_ptr
, smin_val
) ||
1791 signed_add_overflows(smax_ptr
, smax_val
)) {
1792 dst_reg
->smin_value
= S64_MIN
;
1793 dst_reg
->smax_value
= S64_MAX
;
1795 dst_reg
->smin_value
= smin_ptr
+ smin_val
;
1796 dst_reg
->smax_value
= smax_ptr
+ smax_val
;
1798 if (umin_ptr
+ umin_val
< umin_ptr
||
1799 umax_ptr
+ umax_val
< umax_ptr
) {
1800 dst_reg
->umin_value
= 0;
1801 dst_reg
->umax_value
= U64_MAX
;
1803 dst_reg
->umin_value
= umin_ptr
+ umin_val
;
1804 dst_reg
->umax_value
= umax_ptr
+ umax_val
;
1806 dst_reg
->var_off
= tnum_add(ptr_reg
->var_off
, off_reg
->var_off
);
1807 dst_reg
->off
= ptr_reg
->off
;
1808 if (reg_is_pkt_pointer(ptr_reg
)) {
1809 dst_reg
->id
= ++env
->id_gen
;
1810 /* something was added to pkt_ptr, set range to zero */
1815 if (dst_reg
== off_reg
) {
1816 /* scalar -= pointer. Creates an unknown scalar */
1817 if (!env
->allow_ptr_leaks
)
1818 verbose(env
, "R%d tried to subtract pointer from scalar\n",
1822 /* We don't allow subtraction from FP, because (according to
1823 * test_verifier.c test "invalid fp arithmetic", JITs might not
1824 * be able to deal with it.
1826 if (ptr_reg
->type
== PTR_TO_STACK
) {
1827 if (!env
->allow_ptr_leaks
)
1828 verbose(env
, "R%d subtraction from stack pointer prohibited\n",
1832 if (known
&& (ptr_reg
->off
- smin_val
==
1833 (s64
)(s32
)(ptr_reg
->off
- smin_val
))) {
1834 /* pointer -= K. Subtract it from fixed offset */
1835 dst_reg
->smin_value
= smin_ptr
;
1836 dst_reg
->smax_value
= smax_ptr
;
1837 dst_reg
->umin_value
= umin_ptr
;
1838 dst_reg
->umax_value
= umax_ptr
;
1839 dst_reg
->var_off
= ptr_reg
->var_off
;
1840 dst_reg
->id
= ptr_reg
->id
;
1841 dst_reg
->off
= ptr_reg
->off
- smin_val
;
1842 dst_reg
->range
= ptr_reg
->range
;
1845 /* A new variable offset is created. If the subtrahend is known
1846 * nonnegative, then any reg->range we had before is still good.
1848 if (signed_sub_overflows(smin_ptr
, smax_val
) ||
1849 signed_sub_overflows(smax_ptr
, smin_val
)) {
1850 /* Overflow possible, we know nothing */
1851 dst_reg
->smin_value
= S64_MIN
;
1852 dst_reg
->smax_value
= S64_MAX
;
1854 dst_reg
->smin_value
= smin_ptr
- smax_val
;
1855 dst_reg
->smax_value
= smax_ptr
- smin_val
;
1857 if (umin_ptr
< umax_val
) {
1858 /* Overflow possible, we know nothing */
1859 dst_reg
->umin_value
= 0;
1860 dst_reg
->umax_value
= U64_MAX
;
1862 /* Cannot overflow (as long as bounds are consistent) */
1863 dst_reg
->umin_value
= umin_ptr
- umax_val
;
1864 dst_reg
->umax_value
= umax_ptr
- umin_val
;
1866 dst_reg
->var_off
= tnum_sub(ptr_reg
->var_off
, off_reg
->var_off
);
1867 dst_reg
->off
= ptr_reg
->off
;
1868 if (reg_is_pkt_pointer(ptr_reg
)) {
1869 dst_reg
->id
= ++env
->id_gen
;
1870 /* something was added to pkt_ptr, set range to zero */
1878 /* bitwise ops on pointers are troublesome, prohibit for now.
1879 * (However, in principle we could allow some cases, e.g.
1880 * ptr &= ~3 which would reduce min_value by 3.)
1882 if (!env
->allow_ptr_leaks
)
1883 verbose(env
, "R%d bitwise operator %s on pointer prohibited\n",
1884 dst
, bpf_alu_string
[opcode
>> 4]);
1887 /* other operators (e.g. MUL,LSH) produce non-pointer results */
1888 if (!env
->allow_ptr_leaks
)
1889 verbose(env
, "R%d pointer arithmetic with %s operator prohibited\n",
1890 dst
, bpf_alu_string
[opcode
>> 4]);
1894 __update_reg_bounds(dst_reg
);
1895 __reg_deduce_bounds(dst_reg
);
1896 __reg_bound_offset(dst_reg
);
1900 static int adjust_scalar_min_max_vals(struct bpf_verifier_env
*env
,
1901 struct bpf_insn
*insn
,
1902 struct bpf_reg_state
*dst_reg
,
1903 struct bpf_reg_state src_reg
)
1905 struct bpf_reg_state
*regs
= env
->cur_state
.regs
;
1906 u8 opcode
= BPF_OP(insn
->code
);
1907 bool src_known
, dst_known
;
1908 s64 smin_val
, smax_val
;
1909 u64 umin_val
, umax_val
;
1911 if (BPF_CLASS(insn
->code
) != BPF_ALU64
) {
1912 /* 32-bit ALU ops are (32,32)->64 */
1913 coerce_reg_to_32(dst_reg
);
1914 coerce_reg_to_32(&src_reg
);
1916 smin_val
= src_reg
.smin_value
;
1917 smax_val
= src_reg
.smax_value
;
1918 umin_val
= src_reg
.umin_value
;
1919 umax_val
= src_reg
.umax_value
;
1920 src_known
= tnum_is_const(src_reg
.var_off
);
1921 dst_known
= tnum_is_const(dst_reg
->var_off
);
1925 if (signed_add_overflows(dst_reg
->smin_value
, smin_val
) ||
1926 signed_add_overflows(dst_reg
->smax_value
, smax_val
)) {
1927 dst_reg
->smin_value
= S64_MIN
;
1928 dst_reg
->smax_value
= S64_MAX
;
1930 dst_reg
->smin_value
+= smin_val
;
1931 dst_reg
->smax_value
+= smax_val
;
1933 if (dst_reg
->umin_value
+ umin_val
< umin_val
||
1934 dst_reg
->umax_value
+ umax_val
< umax_val
) {
1935 dst_reg
->umin_value
= 0;
1936 dst_reg
->umax_value
= U64_MAX
;
1938 dst_reg
->umin_value
+= umin_val
;
1939 dst_reg
->umax_value
+= umax_val
;
1941 dst_reg
->var_off
= tnum_add(dst_reg
->var_off
, src_reg
.var_off
);
1944 if (signed_sub_overflows(dst_reg
->smin_value
, smax_val
) ||
1945 signed_sub_overflows(dst_reg
->smax_value
, smin_val
)) {
1946 /* Overflow possible, we know nothing */
1947 dst_reg
->smin_value
= S64_MIN
;
1948 dst_reg
->smax_value
= S64_MAX
;
1950 dst_reg
->smin_value
-= smax_val
;
1951 dst_reg
->smax_value
-= smin_val
;
1953 if (dst_reg
->umin_value
< umax_val
) {
1954 /* Overflow possible, we know nothing */
1955 dst_reg
->umin_value
= 0;
1956 dst_reg
->umax_value
= U64_MAX
;
1958 /* Cannot overflow (as long as bounds are consistent) */
1959 dst_reg
->umin_value
-= umax_val
;
1960 dst_reg
->umax_value
-= umin_val
;
1962 dst_reg
->var_off
= tnum_sub(dst_reg
->var_off
, src_reg
.var_off
);
1965 dst_reg
->var_off
= tnum_mul(dst_reg
->var_off
, src_reg
.var_off
);
1966 if (smin_val
< 0 || dst_reg
->smin_value
< 0) {
1967 /* Ain't nobody got time to multiply that sign */
1968 __mark_reg_unbounded(dst_reg
);
1969 __update_reg_bounds(dst_reg
);
1972 /* Both values are positive, so we can work with unsigned and
1973 * copy the result to signed (unless it exceeds S64_MAX).
1975 if (umax_val
> U32_MAX
|| dst_reg
->umax_value
> U32_MAX
) {
1976 /* Potential overflow, we know nothing */
1977 __mark_reg_unbounded(dst_reg
);
1978 /* (except what we can learn from the var_off) */
1979 __update_reg_bounds(dst_reg
);
1982 dst_reg
->umin_value
*= umin_val
;
1983 dst_reg
->umax_value
*= umax_val
;
1984 if (dst_reg
->umax_value
> S64_MAX
) {
1985 /* Overflow possible, we know nothing */
1986 dst_reg
->smin_value
= S64_MIN
;
1987 dst_reg
->smax_value
= S64_MAX
;
1989 dst_reg
->smin_value
= dst_reg
->umin_value
;
1990 dst_reg
->smax_value
= dst_reg
->umax_value
;
1994 if (src_known
&& dst_known
) {
1995 __mark_reg_known(dst_reg
, dst_reg
->var_off
.value
&
1996 src_reg
.var_off
.value
);
1999 /* We get our minimum from the var_off, since that's inherently
2000 * bitwise. Our maximum is the minimum of the operands' maxima.
2002 dst_reg
->var_off
= tnum_and(dst_reg
->var_off
, src_reg
.var_off
);
2003 dst_reg
->umin_value
= dst_reg
->var_off
.value
;
2004 dst_reg
->umax_value
= min(dst_reg
->umax_value
, umax_val
);
2005 if (dst_reg
->smin_value
< 0 || smin_val
< 0) {
2006 /* Lose signed bounds when ANDing negative numbers,
2007 * ain't nobody got time for that.
2009 dst_reg
->smin_value
= S64_MIN
;
2010 dst_reg
->smax_value
= S64_MAX
;
2012 /* ANDing two positives gives a positive, so safe to
2013 * cast result into s64.
2015 dst_reg
->smin_value
= dst_reg
->umin_value
;
2016 dst_reg
->smax_value
= dst_reg
->umax_value
;
2018 /* We may learn something more from the var_off */
2019 __update_reg_bounds(dst_reg
);
2022 if (src_known
&& dst_known
) {
2023 __mark_reg_known(dst_reg
, dst_reg
->var_off
.value
|
2024 src_reg
.var_off
.value
);
2027 /* We get our maximum from the var_off, and our minimum is the
2028 * maximum of the operands' minima
2030 dst_reg
->var_off
= tnum_or(dst_reg
->var_off
, src_reg
.var_off
);
2031 dst_reg
->umin_value
= max(dst_reg
->umin_value
, umin_val
);
2032 dst_reg
->umax_value
= dst_reg
->var_off
.value
|
2033 dst_reg
->var_off
.mask
;
2034 if (dst_reg
->smin_value
< 0 || smin_val
< 0) {
2035 /* Lose signed bounds when ORing negative numbers,
2036 * ain't nobody got time for that.
2038 dst_reg
->smin_value
= S64_MIN
;
2039 dst_reg
->smax_value
= S64_MAX
;
2041 /* ORing two positives gives a positive, so safe to
2042 * cast result into s64.
2044 dst_reg
->smin_value
= dst_reg
->umin_value
;
2045 dst_reg
->smax_value
= dst_reg
->umax_value
;
2047 /* We may learn something more from the var_off */
2048 __update_reg_bounds(dst_reg
);
2051 if (umax_val
> 63) {
2052 /* Shifts greater than 63 are undefined. This includes
2053 * shifts by a negative number.
2055 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
2058 /* We lose all sign bit information (except what we can pick
2061 dst_reg
->smin_value
= S64_MIN
;
2062 dst_reg
->smax_value
= S64_MAX
;
2063 /* If we might shift our top bit out, then we know nothing */
2064 if (dst_reg
->umax_value
> 1ULL << (63 - umax_val
)) {
2065 dst_reg
->umin_value
= 0;
2066 dst_reg
->umax_value
= U64_MAX
;
2068 dst_reg
->umin_value
<<= umin_val
;
2069 dst_reg
->umax_value
<<= umax_val
;
2072 dst_reg
->var_off
= tnum_lshift(dst_reg
->var_off
, umin_val
);
2074 dst_reg
->var_off
= tnum_lshift(tnum_unknown
, umin_val
);
2075 /* We may learn something more from the var_off */
2076 __update_reg_bounds(dst_reg
);
2079 if (umax_val
> 63) {
2080 /* Shifts greater than 63 are undefined. This includes
2081 * shifts by a negative number.
2083 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
2086 /* BPF_RSH is an unsigned shift, so make the appropriate casts */
2087 if (dst_reg
->smin_value
< 0) {
2089 /* Sign bit will be cleared */
2090 dst_reg
->smin_value
= 0;
2092 /* Lost sign bit information */
2093 dst_reg
->smin_value
= S64_MIN
;
2094 dst_reg
->smax_value
= S64_MAX
;
2097 dst_reg
->smin_value
=
2098 (u64
)(dst_reg
->smin_value
) >> umax_val
;
2101 dst_reg
->var_off
= tnum_rshift(dst_reg
->var_off
,
2104 dst_reg
->var_off
= tnum_rshift(tnum_unknown
, umin_val
);
2105 dst_reg
->umin_value
>>= umax_val
;
2106 dst_reg
->umax_value
>>= umin_val
;
2107 /* We may learn something more from the var_off */
2108 __update_reg_bounds(dst_reg
);
2111 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
2115 __reg_deduce_bounds(dst_reg
);
2116 __reg_bound_offset(dst_reg
);
2120 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
2123 static int adjust_reg_min_max_vals(struct bpf_verifier_env
*env
,
2124 struct bpf_insn
*insn
)
2126 struct bpf_reg_state
*regs
= env
->cur_state
.regs
, *dst_reg
, *src_reg
;
2127 struct bpf_reg_state
*ptr_reg
= NULL
, off_reg
= {0};
2128 u8 opcode
= BPF_OP(insn
->code
);
2131 dst_reg
= ®s
[insn
->dst_reg
];
2133 if (dst_reg
->type
!= SCALAR_VALUE
)
2135 if (BPF_SRC(insn
->code
) == BPF_X
) {
2136 src_reg
= ®s
[insn
->src_reg
];
2137 if (src_reg
->type
!= SCALAR_VALUE
) {
2138 if (dst_reg
->type
!= SCALAR_VALUE
) {
2139 /* Combining two pointers by any ALU op yields
2140 * an arbitrary scalar.
2142 if (!env
->allow_ptr_leaks
) {
2143 verbose(env
, "R%d pointer %s pointer prohibited\n",
2145 bpf_alu_string
[opcode
>> 4]);
2148 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
2151 /* scalar += pointer
2152 * This is legal, but we have to reverse our
2153 * src/dest handling in computing the range
2155 rc
= adjust_ptr_min_max_vals(env
, insn
,
2157 if (rc
== -EACCES
&& env
->allow_ptr_leaks
) {
2158 /* scalar += unknown scalar */
2159 __mark_reg_unknown(&off_reg
);
2160 return adjust_scalar_min_max_vals(
2166 } else if (ptr_reg
) {
2167 /* pointer += scalar */
2168 rc
= adjust_ptr_min_max_vals(env
, insn
,
2170 if (rc
== -EACCES
&& env
->allow_ptr_leaks
) {
2171 /* unknown scalar += scalar */
2172 __mark_reg_unknown(dst_reg
);
2173 return adjust_scalar_min_max_vals(
2174 env
, insn
, dst_reg
, *src_reg
);
2179 /* Pretend the src is a reg with a known value, since we only
2180 * need to be able to read from this state.
2182 off_reg
.type
= SCALAR_VALUE
;
2183 __mark_reg_known(&off_reg
, insn
->imm
);
2185 if (ptr_reg
) { /* pointer += K */
2186 rc
= adjust_ptr_min_max_vals(env
, insn
,
2188 if (rc
== -EACCES
&& env
->allow_ptr_leaks
) {
2189 /* unknown scalar += K */
2190 __mark_reg_unknown(dst_reg
);
2191 return adjust_scalar_min_max_vals(
2192 env
, insn
, dst_reg
, off_reg
);
2198 /* Got here implies adding two SCALAR_VALUEs */
2199 if (WARN_ON_ONCE(ptr_reg
)) {
2200 print_verifier_state(env
, &env
->cur_state
);
2201 verbose(env
, "verifier internal error: unexpected ptr_reg\n");
2204 if (WARN_ON(!src_reg
)) {
2205 print_verifier_state(env
, &env
->cur_state
);
2206 verbose(env
, "verifier internal error: no src_reg\n");
2209 return adjust_scalar_min_max_vals(env
, insn
, dst_reg
, *src_reg
);
2212 /* check validity of 32-bit and 64-bit arithmetic operations */
2213 static int check_alu_op(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
2215 struct bpf_reg_state
*regs
= env
->cur_state
.regs
;
2216 u8 opcode
= BPF_OP(insn
->code
);
2219 if (opcode
== BPF_END
|| opcode
== BPF_NEG
) {
2220 if (opcode
== BPF_NEG
) {
2221 if (BPF_SRC(insn
->code
) != 0 ||
2222 insn
->src_reg
!= BPF_REG_0
||
2223 insn
->off
!= 0 || insn
->imm
!= 0) {
2224 verbose(env
, "BPF_NEG uses reserved fields\n");
2228 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0 ||
2229 (insn
->imm
!= 16 && insn
->imm
!= 32 && insn
->imm
!= 64) ||
2230 BPF_CLASS(insn
->code
) == BPF_ALU64
) {
2231 verbose(env
, "BPF_END uses reserved fields\n");
2236 /* check src operand */
2237 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
2241 if (is_pointer_value(env
, insn
->dst_reg
)) {
2242 verbose(env
, "R%d pointer arithmetic prohibited\n",
2247 /* check dest operand */
2248 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP
);
2252 } else if (opcode
== BPF_MOV
) {
2254 if (BPF_SRC(insn
->code
) == BPF_X
) {
2255 if (insn
->imm
!= 0 || insn
->off
!= 0) {
2256 verbose(env
, "BPF_MOV uses reserved fields\n");
2260 /* check src operand */
2261 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
2265 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0) {
2266 verbose(env
, "BPF_MOV uses reserved fields\n");
2271 /* check dest operand */
2272 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP
);
2276 if (BPF_SRC(insn
->code
) == BPF_X
) {
2277 if (BPF_CLASS(insn
->code
) == BPF_ALU64
) {
2279 * copy register state to dest reg
2281 regs
[insn
->dst_reg
] = regs
[insn
->src_reg
];
2282 regs
[insn
->dst_reg
].live
|= REG_LIVE_WRITTEN
;
2285 if (is_pointer_value(env
, insn
->src_reg
)) {
2287 "R%d partial copy of pointer\n",
2291 mark_reg_unknown(env
, regs
, insn
->dst_reg
);
2292 /* high 32 bits are known zero. */
2293 regs
[insn
->dst_reg
].var_off
= tnum_cast(
2294 regs
[insn
->dst_reg
].var_off
, 4);
2295 __update_reg_bounds(®s
[insn
->dst_reg
]);
2299 * remember the value we stored into this reg
2301 regs
[insn
->dst_reg
].type
= SCALAR_VALUE
;
2302 __mark_reg_known(regs
+ insn
->dst_reg
, insn
->imm
);
2305 } else if (opcode
> BPF_END
) {
2306 verbose(env
, "invalid BPF_ALU opcode %x\n", opcode
);
2309 } else { /* all other ALU ops: and, sub, xor, add, ... */
2311 if (BPF_SRC(insn
->code
) == BPF_X
) {
2312 if (insn
->imm
!= 0 || insn
->off
!= 0) {
2313 verbose(env
, "BPF_ALU uses reserved fields\n");
2316 /* check src1 operand */
2317 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
2321 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0) {
2322 verbose(env
, "BPF_ALU uses reserved fields\n");
2327 /* check src2 operand */
2328 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
2332 if ((opcode
== BPF_MOD
|| opcode
== BPF_DIV
) &&
2333 BPF_SRC(insn
->code
) == BPF_K
&& insn
->imm
== 0) {
2334 verbose(env
, "div by zero\n");
2338 if ((opcode
== BPF_LSH
|| opcode
== BPF_RSH
||
2339 opcode
== BPF_ARSH
) && BPF_SRC(insn
->code
) == BPF_K
) {
2340 int size
= BPF_CLASS(insn
->code
) == BPF_ALU64
? 64 : 32;
2342 if (insn
->imm
< 0 || insn
->imm
>= size
) {
2343 verbose(env
, "invalid shift %d\n", insn
->imm
);
2348 /* check dest operand */
2349 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP_NO_MARK
);
2353 return adjust_reg_min_max_vals(env
, insn
);
2359 static void find_good_pkt_pointers(struct bpf_verifier_state
*state
,
2360 struct bpf_reg_state
*dst_reg
,
2361 enum bpf_reg_type type
)
2363 struct bpf_reg_state
*regs
= state
->regs
, *reg
;
2366 if (dst_reg
->off
< 0)
2367 /* This doesn't give us any range */
2370 if (dst_reg
->umax_value
> MAX_PACKET_OFF
||
2371 dst_reg
->umax_value
+ dst_reg
->off
> MAX_PACKET_OFF
)
2372 /* Risk of overflow. For instance, ptr + (1<<63) may be less
2373 * than pkt_end, but that's because it's also less than pkt.
2377 /* LLVM can generate four kind of checks:
2383 * if (r2 > pkt_end) goto <handle exception>
2388 * if (r2 < pkt_end) goto <access okay>
2389 * <handle exception>
2392 * r2 == dst_reg, pkt_end == src_reg
2393 * r2=pkt(id=n,off=8,r=0)
2394 * r3=pkt(id=n,off=0,r=0)
2400 * if (pkt_end >= r2) goto <access okay>
2401 * <handle exception>
2405 * if (pkt_end <= r2) goto <handle exception>
2409 * pkt_end == dst_reg, r2 == src_reg
2410 * r2=pkt(id=n,off=8,r=0)
2411 * r3=pkt(id=n,off=0,r=0)
2413 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
2414 * so that range of bytes [r3, r3 + 8) is safe to access.
2417 /* If our ids match, then we must have the same max_value. And we
2418 * don't care about the other reg's fixed offset, since if it's too big
2419 * the range won't allow anything.
2420 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
2422 for (i
= 0; i
< MAX_BPF_REG
; i
++)
2423 if (regs
[i
].type
== type
&& regs
[i
].id
== dst_reg
->id
)
2424 /* keep the maximum range already checked */
2425 regs
[i
].range
= max_t(u16
, regs
[i
].range
, dst_reg
->off
);
2427 for (i
= 0; i
< MAX_BPF_STACK
; i
+= BPF_REG_SIZE
) {
2428 if (state
->stack_slot_type
[i
] != STACK_SPILL
)
2430 reg
= &state
->spilled_regs
[i
/ BPF_REG_SIZE
];
2431 if (reg
->type
== type
&& reg
->id
== dst_reg
->id
)
2432 reg
->range
= max_t(u16
, reg
->range
, dst_reg
->off
);
2436 /* Adjusts the register min/max values in the case that the dst_reg is the
2437 * variable register that we are working on, and src_reg is a constant or we're
2438 * simply doing a BPF_K check.
2439 * In JEQ/JNE cases we also adjust the var_off values.
2441 static void reg_set_min_max(struct bpf_reg_state
*true_reg
,
2442 struct bpf_reg_state
*false_reg
, u64 val
,
2445 /* If the dst_reg is a pointer, we can't learn anything about its
2446 * variable offset from the compare (unless src_reg were a pointer into
2447 * the same object, but we don't bother with that.
2448 * Since false_reg and true_reg have the same type by construction, we
2449 * only need to check one of them for pointerness.
2451 if (__is_pointer_value(false, false_reg
))
2456 /* If this is false then we know nothing Jon Snow, but if it is
2457 * true then we know for sure.
2459 __mark_reg_known(true_reg
, val
);
2462 /* If this is true we know nothing Jon Snow, but if it is false
2463 * we know the value for sure;
2465 __mark_reg_known(false_reg
, val
);
2468 false_reg
->umax_value
= min(false_reg
->umax_value
, val
);
2469 true_reg
->umin_value
= max(true_reg
->umin_value
, val
+ 1);
2472 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
);
2473 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
+ 1);
2476 false_reg
->umin_value
= max(false_reg
->umin_value
, val
);
2477 true_reg
->umax_value
= min(true_reg
->umax_value
, val
- 1);
2480 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
);
2481 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
- 1);
2484 false_reg
->umax_value
= min(false_reg
->umax_value
, val
- 1);
2485 true_reg
->umin_value
= max(true_reg
->umin_value
, val
);
2488 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
- 1);
2489 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
);
2492 false_reg
->umin_value
= max(false_reg
->umin_value
, val
+ 1);
2493 true_reg
->umax_value
= min(true_reg
->umax_value
, val
);
2496 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
+ 1);
2497 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
);
2503 __reg_deduce_bounds(false_reg
);
2504 __reg_deduce_bounds(true_reg
);
2505 /* We might have learned some bits from the bounds. */
2506 __reg_bound_offset(false_reg
);
2507 __reg_bound_offset(true_reg
);
2508 /* Intersecting with the old var_off might have improved our bounds
2509 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2510 * then new var_off is (0; 0x7f...fc) which improves our umax.
2512 __update_reg_bounds(false_reg
);
2513 __update_reg_bounds(true_reg
);
2516 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
2519 static void reg_set_min_max_inv(struct bpf_reg_state
*true_reg
,
2520 struct bpf_reg_state
*false_reg
, u64 val
,
2523 if (__is_pointer_value(false, false_reg
))
2528 /* If this is false then we know nothing Jon Snow, but if it is
2529 * true then we know for sure.
2531 __mark_reg_known(true_reg
, val
);
2534 /* If this is true we know nothing Jon Snow, but if it is false
2535 * we know the value for sure;
2537 __mark_reg_known(false_reg
, val
);
2540 true_reg
->umax_value
= min(true_reg
->umax_value
, val
- 1);
2541 false_reg
->umin_value
= max(false_reg
->umin_value
, val
);
2544 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
- 1);
2545 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
);
2548 true_reg
->umin_value
= max(true_reg
->umin_value
, val
+ 1);
2549 false_reg
->umax_value
= min(false_reg
->umax_value
, val
);
2552 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
+ 1);
2553 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
);
2556 true_reg
->umax_value
= min(true_reg
->umax_value
, val
);
2557 false_reg
->umin_value
= max(false_reg
->umin_value
, val
+ 1);
2560 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
);
2561 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
+ 1);
2564 true_reg
->umin_value
= max(true_reg
->umin_value
, val
);
2565 false_reg
->umax_value
= min(false_reg
->umax_value
, val
- 1);
2568 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
);
2569 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
- 1);
2575 __reg_deduce_bounds(false_reg
);
2576 __reg_deduce_bounds(true_reg
);
2577 /* We might have learned some bits from the bounds. */
2578 __reg_bound_offset(false_reg
);
2579 __reg_bound_offset(true_reg
);
2580 /* Intersecting with the old var_off might have improved our bounds
2581 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2582 * then new var_off is (0; 0x7f...fc) which improves our umax.
2584 __update_reg_bounds(false_reg
);
2585 __update_reg_bounds(true_reg
);
2588 /* Regs are known to be equal, so intersect their min/max/var_off */
2589 static void __reg_combine_min_max(struct bpf_reg_state
*src_reg
,
2590 struct bpf_reg_state
*dst_reg
)
2592 src_reg
->umin_value
= dst_reg
->umin_value
= max(src_reg
->umin_value
,
2593 dst_reg
->umin_value
);
2594 src_reg
->umax_value
= dst_reg
->umax_value
= min(src_reg
->umax_value
,
2595 dst_reg
->umax_value
);
2596 src_reg
->smin_value
= dst_reg
->smin_value
= max(src_reg
->smin_value
,
2597 dst_reg
->smin_value
);
2598 src_reg
->smax_value
= dst_reg
->smax_value
= min(src_reg
->smax_value
,
2599 dst_reg
->smax_value
);
2600 src_reg
->var_off
= dst_reg
->var_off
= tnum_intersect(src_reg
->var_off
,
2602 /* We might have learned new bounds from the var_off. */
2603 __update_reg_bounds(src_reg
);
2604 __update_reg_bounds(dst_reg
);
2605 /* We might have learned something about the sign bit. */
2606 __reg_deduce_bounds(src_reg
);
2607 __reg_deduce_bounds(dst_reg
);
2608 /* We might have learned some bits from the bounds. */
2609 __reg_bound_offset(src_reg
);
2610 __reg_bound_offset(dst_reg
);
2611 /* Intersecting with the old var_off might have improved our bounds
2612 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2613 * then new var_off is (0; 0x7f...fc) which improves our umax.
2615 __update_reg_bounds(src_reg
);
2616 __update_reg_bounds(dst_reg
);
2619 static void reg_combine_min_max(struct bpf_reg_state
*true_src
,
2620 struct bpf_reg_state
*true_dst
,
2621 struct bpf_reg_state
*false_src
,
2622 struct bpf_reg_state
*false_dst
,
2627 __reg_combine_min_max(true_src
, true_dst
);
2630 __reg_combine_min_max(false_src
, false_dst
);
2635 static void mark_map_reg(struct bpf_reg_state
*regs
, u32 regno
, u32 id
,
2638 struct bpf_reg_state
*reg
= ®s
[regno
];
2640 if (reg
->type
== PTR_TO_MAP_VALUE_OR_NULL
&& reg
->id
== id
) {
2641 /* Old offset (both fixed and variable parts) should
2642 * have been known-zero, because we don't allow pointer
2643 * arithmetic on pointers that might be NULL.
2645 if (WARN_ON_ONCE(reg
->smin_value
|| reg
->smax_value
||
2646 !tnum_equals_const(reg
->var_off
, 0) ||
2648 __mark_reg_known_zero(reg
);
2652 reg
->type
= SCALAR_VALUE
;
2653 } else if (reg
->map_ptr
->inner_map_meta
) {
2654 reg
->type
= CONST_PTR_TO_MAP
;
2655 reg
->map_ptr
= reg
->map_ptr
->inner_map_meta
;
2657 reg
->type
= PTR_TO_MAP_VALUE
;
2659 /* We don't need id from this point onwards anymore, thus we
2660 * should better reset it, so that state pruning has chances
2667 /* The logic is similar to find_good_pkt_pointers(), both could eventually
2668 * be folded together at some point.
2670 static void mark_map_regs(struct bpf_verifier_state
*state
, u32 regno
,
2673 struct bpf_reg_state
*regs
= state
->regs
;
2674 u32 id
= regs
[regno
].id
;
2677 for (i
= 0; i
< MAX_BPF_REG
; i
++)
2678 mark_map_reg(regs
, i
, id
, is_null
);
2680 for (i
= 0; i
< MAX_BPF_STACK
; i
+= BPF_REG_SIZE
) {
2681 if (state
->stack_slot_type
[i
] != STACK_SPILL
)
2683 mark_map_reg(state
->spilled_regs
, i
/ BPF_REG_SIZE
, id
, is_null
);
2687 static int check_cond_jmp_op(struct bpf_verifier_env
*env
,
2688 struct bpf_insn
*insn
, int *insn_idx
)
2690 struct bpf_verifier_state
*other_branch
, *this_branch
= &env
->cur_state
;
2691 struct bpf_reg_state
*regs
= this_branch
->regs
, *dst_reg
;
2692 u8 opcode
= BPF_OP(insn
->code
);
2695 if (opcode
> BPF_JSLE
) {
2696 verbose(env
, "invalid BPF_JMP opcode %x\n", opcode
);
2700 if (BPF_SRC(insn
->code
) == BPF_X
) {
2701 if (insn
->imm
!= 0) {
2702 verbose(env
, "BPF_JMP uses reserved fields\n");
2706 /* check src1 operand */
2707 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
2711 if (is_pointer_value(env
, insn
->src_reg
)) {
2712 verbose(env
, "R%d pointer comparison prohibited\n",
2717 if (insn
->src_reg
!= BPF_REG_0
) {
2718 verbose(env
, "BPF_JMP uses reserved fields\n");
2723 /* check src2 operand */
2724 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
2728 dst_reg
= ®s
[insn
->dst_reg
];
2730 /* detect if R == 0 where R was initialized to zero earlier */
2731 if (BPF_SRC(insn
->code
) == BPF_K
&&
2732 (opcode
== BPF_JEQ
|| opcode
== BPF_JNE
) &&
2733 dst_reg
->type
== SCALAR_VALUE
&&
2734 tnum_equals_const(dst_reg
->var_off
, insn
->imm
)) {
2735 if (opcode
== BPF_JEQ
) {
2736 /* if (imm == imm) goto pc+off;
2737 * only follow the goto, ignore fall-through
2739 *insn_idx
+= insn
->off
;
2742 /* if (imm != imm) goto pc+off;
2743 * only follow fall-through branch, since
2744 * that's where the program will go
2750 other_branch
= push_stack(env
, *insn_idx
+ insn
->off
+ 1, *insn_idx
);
2754 /* detect if we are comparing against a constant value so we can adjust
2755 * our min/max values for our dst register.
2756 * this is only legit if both are scalars (or pointers to the same
2757 * object, I suppose, but we don't support that right now), because
2758 * otherwise the different base pointers mean the offsets aren't
2761 if (BPF_SRC(insn
->code
) == BPF_X
) {
2762 if (dst_reg
->type
== SCALAR_VALUE
&&
2763 regs
[insn
->src_reg
].type
== SCALAR_VALUE
) {
2764 if (tnum_is_const(regs
[insn
->src_reg
].var_off
))
2765 reg_set_min_max(&other_branch
->regs
[insn
->dst_reg
],
2766 dst_reg
, regs
[insn
->src_reg
].var_off
.value
,
2768 else if (tnum_is_const(dst_reg
->var_off
))
2769 reg_set_min_max_inv(&other_branch
->regs
[insn
->src_reg
],
2770 ®s
[insn
->src_reg
],
2771 dst_reg
->var_off
.value
, opcode
);
2772 else if (opcode
== BPF_JEQ
|| opcode
== BPF_JNE
)
2773 /* Comparing for equality, we can combine knowledge */
2774 reg_combine_min_max(&other_branch
->regs
[insn
->src_reg
],
2775 &other_branch
->regs
[insn
->dst_reg
],
2776 ®s
[insn
->src_reg
],
2777 ®s
[insn
->dst_reg
], opcode
);
2779 } else if (dst_reg
->type
== SCALAR_VALUE
) {
2780 reg_set_min_max(&other_branch
->regs
[insn
->dst_reg
],
2781 dst_reg
, insn
->imm
, opcode
);
2784 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
2785 if (BPF_SRC(insn
->code
) == BPF_K
&&
2786 insn
->imm
== 0 && (opcode
== BPF_JEQ
|| opcode
== BPF_JNE
) &&
2787 dst_reg
->type
== PTR_TO_MAP_VALUE_OR_NULL
) {
2788 /* Mark all identical map registers in each branch as either
2789 * safe or unknown depending R == 0 or R != 0 conditional.
2791 mark_map_regs(this_branch
, insn
->dst_reg
, opcode
== BPF_JNE
);
2792 mark_map_regs(other_branch
, insn
->dst_reg
, opcode
== BPF_JEQ
);
2793 } else if (BPF_SRC(insn
->code
) == BPF_X
&& opcode
== BPF_JGT
&&
2794 dst_reg
->type
== PTR_TO_PACKET
&&
2795 regs
[insn
->src_reg
].type
== PTR_TO_PACKET_END
) {
2796 find_good_pkt_pointers(this_branch
, dst_reg
, PTR_TO_PACKET
);
2797 } else if (BPF_SRC(insn
->code
) == BPF_X
&& opcode
== BPF_JLT
&&
2798 dst_reg
->type
== PTR_TO_PACKET
&&
2799 regs
[insn
->src_reg
].type
== PTR_TO_PACKET_END
) {
2800 find_good_pkt_pointers(other_branch
, dst_reg
, PTR_TO_PACKET
);
2801 } else if (BPF_SRC(insn
->code
) == BPF_X
&& opcode
== BPF_JGE
&&
2802 dst_reg
->type
== PTR_TO_PACKET_END
&&
2803 regs
[insn
->src_reg
].type
== PTR_TO_PACKET
) {
2804 find_good_pkt_pointers(other_branch
, ®s
[insn
->src_reg
],
2806 } else if (BPF_SRC(insn
->code
) == BPF_X
&& opcode
== BPF_JLE
&&
2807 dst_reg
->type
== PTR_TO_PACKET_END
&&
2808 regs
[insn
->src_reg
].type
== PTR_TO_PACKET
) {
2809 find_good_pkt_pointers(this_branch
, ®s
[insn
->src_reg
],
2811 } else if (BPF_SRC(insn
->code
) == BPF_X
&& opcode
== BPF_JGT
&&
2812 dst_reg
->type
== PTR_TO_PACKET_META
&&
2813 reg_is_init_pkt_pointer(®s
[insn
->src_reg
], PTR_TO_PACKET
)) {
2814 find_good_pkt_pointers(this_branch
, dst_reg
, PTR_TO_PACKET_META
);
2815 } else if (BPF_SRC(insn
->code
) == BPF_X
&& opcode
== BPF_JLT
&&
2816 dst_reg
->type
== PTR_TO_PACKET_META
&&
2817 reg_is_init_pkt_pointer(®s
[insn
->src_reg
], PTR_TO_PACKET
)) {
2818 find_good_pkt_pointers(other_branch
, dst_reg
, PTR_TO_PACKET_META
);
2819 } else if (BPF_SRC(insn
->code
) == BPF_X
&& opcode
== BPF_JGE
&&
2820 reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
2821 regs
[insn
->src_reg
].type
== PTR_TO_PACKET_META
) {
2822 find_good_pkt_pointers(other_branch
, ®s
[insn
->src_reg
],
2823 PTR_TO_PACKET_META
);
2824 } else if (BPF_SRC(insn
->code
) == BPF_X
&& opcode
== BPF_JLE
&&
2825 reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
2826 regs
[insn
->src_reg
].type
== PTR_TO_PACKET_META
) {
2827 find_good_pkt_pointers(this_branch
, ®s
[insn
->src_reg
],
2828 PTR_TO_PACKET_META
);
2829 } else if (is_pointer_value(env
, insn
->dst_reg
)) {
2830 verbose(env
, "R%d pointer comparison prohibited\n",
2835 print_verifier_state(env
, this_branch
);
2839 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
2840 static struct bpf_map
*ld_imm64_to_map_ptr(struct bpf_insn
*insn
)
2842 u64 imm64
= ((u64
) (u32
) insn
[0].imm
) | ((u64
) (u32
) insn
[1].imm
) << 32;
2844 return (struct bpf_map
*) (unsigned long) imm64
;
2847 /* verify BPF_LD_IMM64 instruction */
2848 static int check_ld_imm(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
2850 struct bpf_reg_state
*regs
= env
->cur_state
.regs
;
2853 if (BPF_SIZE(insn
->code
) != BPF_DW
) {
2854 verbose(env
, "invalid BPF_LD_IMM insn\n");
2857 if (insn
->off
!= 0) {
2858 verbose(env
, "BPF_LD_IMM64 uses reserved fields\n");
2862 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP
);
2866 if (insn
->src_reg
== 0) {
2867 u64 imm
= ((u64
)(insn
+ 1)->imm
<< 32) | (u32
)insn
->imm
;
2869 regs
[insn
->dst_reg
].type
= SCALAR_VALUE
;
2870 __mark_reg_known(®s
[insn
->dst_reg
], imm
);
2874 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
2875 BUG_ON(insn
->src_reg
!= BPF_PSEUDO_MAP_FD
);
2877 regs
[insn
->dst_reg
].type
= CONST_PTR_TO_MAP
;
2878 regs
[insn
->dst_reg
].map_ptr
= ld_imm64_to_map_ptr(insn
);
2882 static bool may_access_skb(enum bpf_prog_type type
)
2885 case BPF_PROG_TYPE_SOCKET_FILTER
:
2886 case BPF_PROG_TYPE_SCHED_CLS
:
2887 case BPF_PROG_TYPE_SCHED_ACT
:
2894 /* verify safety of LD_ABS|LD_IND instructions:
2895 * - they can only appear in the programs where ctx == skb
2896 * - since they are wrappers of function calls, they scratch R1-R5 registers,
2897 * preserve R6-R9, and store return value into R0
2900 * ctx == skb == R6 == CTX
2903 * SRC == any register
2904 * IMM == 32-bit immediate
2907 * R0 - 8/16/32-bit skb data converted to cpu endianness
2909 static int check_ld_abs(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
2911 struct bpf_reg_state
*regs
= env
->cur_state
.regs
;
2912 u8 mode
= BPF_MODE(insn
->code
);
2915 if (!may_access_skb(env
->prog
->type
)) {
2916 verbose(env
, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
2920 if (insn
->dst_reg
!= BPF_REG_0
|| insn
->off
!= 0 ||
2921 BPF_SIZE(insn
->code
) == BPF_DW
||
2922 (mode
== BPF_ABS
&& insn
->src_reg
!= BPF_REG_0
)) {
2923 verbose(env
, "BPF_LD_[ABS|IND] uses reserved fields\n");
2927 /* check whether implicit source operand (register R6) is readable */
2928 err
= check_reg_arg(env
, BPF_REG_6
, SRC_OP
);
2932 if (regs
[BPF_REG_6
].type
!= PTR_TO_CTX
) {
2934 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
2938 if (mode
== BPF_IND
) {
2939 /* check explicit source operand */
2940 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
2945 /* reset caller saved regs to unreadable */
2946 for (i
= 0; i
< CALLER_SAVED_REGS
; i
++) {
2947 mark_reg_not_init(env
, regs
, caller_saved
[i
]);
2948 check_reg_arg(env
, caller_saved
[i
], DST_OP_NO_MARK
);
2951 /* mark destination R0 register as readable, since it contains
2952 * the value fetched from the packet.
2953 * Already marked as written above.
2955 mark_reg_unknown(env
, regs
, BPF_REG_0
);
2959 static int check_return_code(struct bpf_verifier_env
*env
)
2961 struct bpf_reg_state
*reg
;
2962 struct tnum range
= tnum_range(0, 1);
2964 switch (env
->prog
->type
) {
2965 case BPF_PROG_TYPE_CGROUP_SKB
:
2966 case BPF_PROG_TYPE_CGROUP_SOCK
:
2967 case BPF_PROG_TYPE_SOCK_OPS
:
2973 reg
= &env
->cur_state
.regs
[BPF_REG_0
];
2974 if (reg
->type
!= SCALAR_VALUE
) {
2975 verbose(env
, "At program exit the register R0 is not a known value (%s)\n",
2976 reg_type_str
[reg
->type
]);
2980 if (!tnum_in(range
, reg
->var_off
)) {
2981 verbose(env
, "At program exit the register R0 ");
2982 if (!tnum_is_unknown(reg
->var_off
)) {
2985 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
2986 verbose(env
, "has value %s", tn_buf
);
2988 verbose(env
, "has unknown scalar value");
2990 verbose(env
, " should have been 0 or 1\n");
2996 /* non-recursive DFS pseudo code
2997 * 1 procedure DFS-iterative(G,v):
2998 * 2 label v as discovered
2999 * 3 let S be a stack
3001 * 5 while S is not empty
3003 * 7 if t is what we're looking for:
3005 * 9 for all edges e in G.adjacentEdges(t) do
3006 * 10 if edge e is already labelled
3007 * 11 continue with the next edge
3008 * 12 w <- G.adjacentVertex(t,e)
3009 * 13 if vertex w is not discovered and not explored
3010 * 14 label e as tree-edge
3011 * 15 label w as discovered
3014 * 18 else if vertex w is discovered
3015 * 19 label e as back-edge
3017 * 21 // vertex w is explored
3018 * 22 label e as forward- or cross-edge
3019 * 23 label t as explored
3024 * 0x11 - discovered and fall-through edge labelled
3025 * 0x12 - discovered and fall-through and branch edges labelled
3036 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
3038 static int *insn_stack
; /* stack of insns to process */
3039 static int cur_stack
; /* current stack index */
3040 static int *insn_state
;
3042 /* t, w, e - match pseudo-code above:
3043 * t - index of current instruction
3044 * w - next instruction
3047 static int push_insn(int t
, int w
, int e
, struct bpf_verifier_env
*env
)
3049 if (e
== FALLTHROUGH
&& insn_state
[t
] >= (DISCOVERED
| FALLTHROUGH
))
3052 if (e
== BRANCH
&& insn_state
[t
] >= (DISCOVERED
| BRANCH
))
3055 if (w
< 0 || w
>= env
->prog
->len
) {
3056 verbose(env
, "jump out of range from insn %d to %d\n", t
, w
);
3061 /* mark branch target for state pruning */
3062 env
->explored_states
[w
] = STATE_LIST_MARK
;
3064 if (insn_state
[w
] == 0) {
3066 insn_state
[t
] = DISCOVERED
| e
;
3067 insn_state
[w
] = DISCOVERED
;
3068 if (cur_stack
>= env
->prog
->len
)
3070 insn_stack
[cur_stack
++] = w
;
3072 } else if ((insn_state
[w
] & 0xF0) == DISCOVERED
) {
3073 verbose(env
, "back-edge from insn %d to %d\n", t
, w
);
3075 } else if (insn_state
[w
] == EXPLORED
) {
3076 /* forward- or cross-edge */
3077 insn_state
[t
] = DISCOVERED
| e
;
3079 verbose(env
, "insn state internal bug\n");
3085 /* non-recursive depth-first-search to detect loops in BPF program
3086 * loop == back-edge in directed graph
3088 static int check_cfg(struct bpf_verifier_env
*env
)
3090 struct bpf_insn
*insns
= env
->prog
->insnsi
;
3091 int insn_cnt
= env
->prog
->len
;
3095 insn_state
= kcalloc(insn_cnt
, sizeof(int), GFP_KERNEL
);
3099 insn_stack
= kcalloc(insn_cnt
, sizeof(int), GFP_KERNEL
);
3105 insn_state
[0] = DISCOVERED
; /* mark 1st insn as discovered */
3106 insn_stack
[0] = 0; /* 0 is the first instruction */
3112 t
= insn_stack
[cur_stack
- 1];
3114 if (BPF_CLASS(insns
[t
].code
) == BPF_JMP
) {
3115 u8 opcode
= BPF_OP(insns
[t
].code
);
3117 if (opcode
== BPF_EXIT
) {
3119 } else if (opcode
== BPF_CALL
) {
3120 ret
= push_insn(t
, t
+ 1, FALLTHROUGH
, env
);
3125 if (t
+ 1 < insn_cnt
)
3126 env
->explored_states
[t
+ 1] = STATE_LIST_MARK
;
3127 } else if (opcode
== BPF_JA
) {
3128 if (BPF_SRC(insns
[t
].code
) != BPF_K
) {
3132 /* unconditional jump with single edge */
3133 ret
= push_insn(t
, t
+ insns
[t
].off
+ 1,
3139 /* tell verifier to check for equivalent states
3140 * after every call and jump
3142 if (t
+ 1 < insn_cnt
)
3143 env
->explored_states
[t
+ 1] = STATE_LIST_MARK
;
3145 /* conditional jump with two edges */
3146 env
->explored_states
[t
] = STATE_LIST_MARK
;
3147 ret
= push_insn(t
, t
+ 1, FALLTHROUGH
, env
);
3153 ret
= push_insn(t
, t
+ insns
[t
].off
+ 1, BRANCH
, env
);
3160 /* all other non-branch instructions with single
3163 ret
= push_insn(t
, t
+ 1, FALLTHROUGH
, env
);
3171 insn_state
[t
] = EXPLORED
;
3172 if (cur_stack
-- <= 0) {
3173 verbose(env
, "pop stack internal bug\n");
3180 for (i
= 0; i
< insn_cnt
; i
++) {
3181 if (insn_state
[i
] != EXPLORED
) {
3182 verbose(env
, "unreachable insn %d\n", i
);
3187 ret
= 0; /* cfg looks good */
3195 /* check %cur's range satisfies %old's */
3196 static bool range_within(struct bpf_reg_state
*old
,
3197 struct bpf_reg_state
*cur
)
3199 return old
->umin_value
<= cur
->umin_value
&&
3200 old
->umax_value
>= cur
->umax_value
&&
3201 old
->smin_value
<= cur
->smin_value
&&
3202 old
->smax_value
>= cur
->smax_value
;
3205 /* Maximum number of register states that can exist at once */
3206 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
3212 /* If in the old state two registers had the same id, then they need to have
3213 * the same id in the new state as well. But that id could be different from
3214 * the old state, so we need to track the mapping from old to new ids.
3215 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
3216 * regs with old id 5 must also have new id 9 for the new state to be safe. But
3217 * regs with a different old id could still have new id 9, we don't care about
3219 * So we look through our idmap to see if this old id has been seen before. If
3220 * so, we require the new id to match; otherwise, we add the id pair to the map.
3222 static bool check_ids(u32 old_id
, u32 cur_id
, struct idpair
*idmap
)
3226 for (i
= 0; i
< ID_MAP_SIZE
; i
++) {
3227 if (!idmap
[i
].old
) {
3228 /* Reached an empty slot; haven't seen this id before */
3229 idmap
[i
].old
= old_id
;
3230 idmap
[i
].cur
= cur_id
;
3233 if (idmap
[i
].old
== old_id
)
3234 return idmap
[i
].cur
== cur_id
;
3236 /* We ran out of idmap slots, which should be impossible */
3241 /* Returns true if (rold safe implies rcur safe) */
3242 static bool regsafe(struct bpf_reg_state
*rold
, struct bpf_reg_state
*rcur
,
3243 struct idpair
*idmap
)
3245 if (!(rold
->live
& REG_LIVE_READ
))
3246 /* explored state didn't use this */
3249 if (memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, live
)) == 0)
3252 if (rold
->type
== NOT_INIT
)
3253 /* explored state can't have used this */
3255 if (rcur
->type
== NOT_INIT
)
3257 switch (rold
->type
) {
3259 if (rcur
->type
== SCALAR_VALUE
) {
3260 /* new val must satisfy old val knowledge */
3261 return range_within(rold
, rcur
) &&
3262 tnum_in(rold
->var_off
, rcur
->var_off
);
3264 /* if we knew anything about the old value, we're not
3265 * equal, because we can't know anything about the
3266 * scalar value of the pointer in the new value.
3268 return rold
->umin_value
== 0 &&
3269 rold
->umax_value
== U64_MAX
&&
3270 rold
->smin_value
== S64_MIN
&&
3271 rold
->smax_value
== S64_MAX
&&
3272 tnum_is_unknown(rold
->var_off
);
3274 case PTR_TO_MAP_VALUE
:
3275 /* If the new min/max/var_off satisfy the old ones and
3276 * everything else matches, we are OK.
3277 * We don't care about the 'id' value, because nothing
3278 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
3280 return memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, id
)) == 0 &&
3281 range_within(rold
, rcur
) &&
3282 tnum_in(rold
->var_off
, rcur
->var_off
);
3283 case PTR_TO_MAP_VALUE_OR_NULL
:
3284 /* a PTR_TO_MAP_VALUE could be safe to use as a
3285 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
3286 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
3287 * checked, doing so could have affected others with the same
3288 * id, and we can't check for that because we lost the id when
3289 * we converted to a PTR_TO_MAP_VALUE.
3291 if (rcur
->type
!= PTR_TO_MAP_VALUE_OR_NULL
)
3293 if (memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, id
)))
3295 /* Check our ids match any regs they're supposed to */
3296 return check_ids(rold
->id
, rcur
->id
, idmap
);
3297 case PTR_TO_PACKET_META
:
3299 if (rcur
->type
!= rold
->type
)
3301 /* We must have at least as much range as the old ptr
3302 * did, so that any accesses which were safe before are
3303 * still safe. This is true even if old range < old off,
3304 * since someone could have accessed through (ptr - k), or
3305 * even done ptr -= k in a register, to get a safe access.
3307 if (rold
->range
> rcur
->range
)
3309 /* If the offsets don't match, we can't trust our alignment;
3310 * nor can we be sure that we won't fall out of range.
3312 if (rold
->off
!= rcur
->off
)
3314 /* id relations must be preserved */
3315 if (rold
->id
&& !check_ids(rold
->id
, rcur
->id
, idmap
))
3317 /* new val must satisfy old val knowledge */
3318 return range_within(rold
, rcur
) &&
3319 tnum_in(rold
->var_off
, rcur
->var_off
);
3321 case CONST_PTR_TO_MAP
:
3323 case PTR_TO_PACKET_END
:
3324 /* Only valid matches are exact, which memcmp() above
3325 * would have accepted
3328 /* Don't know what's going on, just say it's not safe */
3332 /* Shouldn't get here; if we do, say it's not safe */
3337 /* compare two verifier states
3339 * all states stored in state_list are known to be valid, since
3340 * verifier reached 'bpf_exit' instruction through them
3342 * this function is called when verifier exploring different branches of
3343 * execution popped from the state stack. If it sees an old state that has
3344 * more strict register state and more strict stack state then this execution
3345 * branch doesn't need to be explored further, since verifier already
3346 * concluded that more strict state leads to valid finish.
3348 * Therefore two states are equivalent if register state is more conservative
3349 * and explored stack state is more conservative than the current one.
3352 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
3353 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
3355 * In other words if current stack state (one being explored) has more
3356 * valid slots than old one that already passed validation, it means
3357 * the verifier can stop exploring and conclude that current state is valid too
3359 * Similarly with registers. If explored state has register type as invalid
3360 * whereas register type in current state is meaningful, it means that
3361 * the current state will reach 'bpf_exit' instruction safely
3363 static bool states_equal(struct bpf_verifier_env
*env
,
3364 struct bpf_verifier_state
*old
,
3365 struct bpf_verifier_state
*cur
)
3367 struct idpair
*idmap
;
3371 idmap
= kcalloc(ID_MAP_SIZE
, sizeof(struct idpair
), GFP_KERNEL
);
3372 /* If we failed to allocate the idmap, just say it's not safe */
3376 for (i
= 0; i
< MAX_BPF_REG
; i
++) {
3377 if (!regsafe(&old
->regs
[i
], &cur
->regs
[i
], idmap
))
3381 for (i
= 0; i
< MAX_BPF_STACK
; i
++) {
3382 if (old
->stack_slot_type
[i
] == STACK_INVALID
)
3384 if (old
->stack_slot_type
[i
] != cur
->stack_slot_type
[i
])
3385 /* Ex: old explored (safe) state has STACK_SPILL in
3386 * this stack slot, but current has has STACK_MISC ->
3387 * this verifier states are not equivalent,
3388 * return false to continue verification of this path
3391 if (i
% BPF_REG_SIZE
)
3393 if (old
->stack_slot_type
[i
] != STACK_SPILL
)
3395 if (!regsafe(&old
->spilled_regs
[i
/ BPF_REG_SIZE
],
3396 &cur
->spilled_regs
[i
/ BPF_REG_SIZE
],
3398 /* when explored and current stack slot are both storing
3399 * spilled registers, check that stored pointers types
3400 * are the same as well.
3401 * Ex: explored safe path could have stored
3402 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
3403 * but current path has stored:
3404 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
3405 * such verifier states are not equivalent.
3406 * return false to continue verification of this path
3418 /* A write screens off any subsequent reads; but write marks come from the
3419 * straight-line code between a state and its parent. When we arrive at a
3420 * jump target (in the first iteration of the propagate_liveness() loop),
3421 * we didn't arrive by the straight-line code, so read marks in state must
3422 * propagate to parent regardless of state's write marks.
3424 static bool do_propagate_liveness(const struct bpf_verifier_state
*state
,
3425 struct bpf_verifier_state
*parent
)
3427 bool writes
= parent
== state
->parent
; /* Observe write marks */
3428 bool touched
= false; /* any changes made? */
3433 /* Propagate read liveness of registers... */
3434 BUILD_BUG_ON(BPF_REG_FP
+ 1 != MAX_BPF_REG
);
3435 /* We don't need to worry about FP liveness because it's read-only */
3436 for (i
= 0; i
< BPF_REG_FP
; i
++) {
3437 if (parent
->regs
[i
].live
& REG_LIVE_READ
)
3439 if (writes
&& (state
->regs
[i
].live
& REG_LIVE_WRITTEN
))
3441 if (state
->regs
[i
].live
& REG_LIVE_READ
) {
3442 parent
->regs
[i
].live
|= REG_LIVE_READ
;
3446 /* ... and stack slots */
3447 for (i
= 0; i
< MAX_BPF_STACK
/ BPF_REG_SIZE
; i
++) {
3448 if (parent
->stack_slot_type
[i
* BPF_REG_SIZE
] != STACK_SPILL
)
3450 if (state
->stack_slot_type
[i
* BPF_REG_SIZE
] != STACK_SPILL
)
3452 if (parent
->spilled_regs
[i
].live
& REG_LIVE_READ
)
3454 if (writes
&& (state
->spilled_regs
[i
].live
& REG_LIVE_WRITTEN
))
3456 if (state
->spilled_regs
[i
].live
& REG_LIVE_READ
) {
3457 parent
->spilled_regs
[i
].live
|= REG_LIVE_READ
;
3464 /* "parent" is "a state from which we reach the current state", but initially
3465 * it is not the state->parent (i.e. "the state whose straight-line code leads
3466 * to the current state"), instead it is the state that happened to arrive at
3467 * a (prunable) equivalent of the current state. See comment above
3468 * do_propagate_liveness() for consequences of this.
3469 * This function is just a more efficient way of calling mark_reg_read() or
3470 * mark_stack_slot_read() on each reg in "parent" that is read in "state",
3471 * though it requires that parent != state->parent in the call arguments.
3473 static void propagate_liveness(const struct bpf_verifier_state
*state
,
3474 struct bpf_verifier_state
*parent
)
3476 while (do_propagate_liveness(state
, parent
)) {
3477 /* Something changed, so we need to feed those changes onward */
3479 parent
= state
->parent
;
3483 static int is_state_visited(struct bpf_verifier_env
*env
, int insn_idx
)
3485 struct bpf_verifier_state_list
*new_sl
;
3486 struct bpf_verifier_state_list
*sl
;
3489 sl
= env
->explored_states
[insn_idx
];
3491 /* this 'insn_idx' instruction wasn't marked, so we will not
3492 * be doing state search here
3496 while (sl
!= STATE_LIST_MARK
) {
3497 if (states_equal(env
, &sl
->state
, &env
->cur_state
)) {
3498 /* reached equivalent register/stack state,
3500 * Registers read by the continuation are read by us.
3501 * If we have any write marks in env->cur_state, they
3502 * will prevent corresponding reads in the continuation
3503 * from reaching our parent (an explored_state). Our
3504 * own state will get the read marks recorded, but
3505 * they'll be immediately forgotten as we're pruning
3506 * this state and will pop a new one.
3508 propagate_liveness(&sl
->state
, &env
->cur_state
);
3514 /* there were no equivalent states, remember current one.
3515 * technically the current state is not proven to be safe yet,
3516 * but it will either reach bpf_exit (which means it's safe) or
3517 * it will be rejected. Since there are no loops, we won't be
3518 * seeing this 'insn_idx' instruction again on the way to bpf_exit
3520 new_sl
= kmalloc(sizeof(struct bpf_verifier_state_list
), GFP_USER
);
3524 /* add new state to the head of linked list */
3525 memcpy(&new_sl
->state
, &env
->cur_state
, sizeof(env
->cur_state
));
3526 new_sl
->next
= env
->explored_states
[insn_idx
];
3527 env
->explored_states
[insn_idx
] = new_sl
;
3528 /* connect new state to parentage chain */
3529 env
->cur_state
.parent
= &new_sl
->state
;
3530 /* clear write marks in current state: the writes we did are not writes
3531 * our child did, so they don't screen off its reads from us.
3532 * (There are no read marks in current state, because reads always mark
3533 * their parent and current state never has children yet. Only
3534 * explored_states can get read marks.)
3536 for (i
= 0; i
< BPF_REG_FP
; i
++)
3537 env
->cur_state
.regs
[i
].live
= REG_LIVE_NONE
;
3538 for (i
= 0; i
< MAX_BPF_STACK
/ BPF_REG_SIZE
; i
++)
3539 if (env
->cur_state
.stack_slot_type
[i
* BPF_REG_SIZE
] == STACK_SPILL
)
3540 env
->cur_state
.spilled_regs
[i
].live
= REG_LIVE_NONE
;
3544 static int ext_analyzer_insn_hook(struct bpf_verifier_env
*env
,
3545 int insn_idx
, int prev_insn_idx
)
3547 if (!env
->analyzer_ops
|| !env
->analyzer_ops
->insn_hook
)
3550 return env
->analyzer_ops
->insn_hook(env
, insn_idx
, prev_insn_idx
);
3553 static int do_check(struct bpf_verifier_env
*env
)
3555 struct bpf_verifier_state
*state
= &env
->cur_state
;
3556 struct bpf_insn
*insns
= env
->prog
->insnsi
;
3557 struct bpf_reg_state
*regs
= state
->regs
;
3558 int insn_cnt
= env
->prog
->len
;
3559 int insn_idx
, prev_insn_idx
= 0;
3560 int insn_processed
= 0;
3561 bool do_print_state
= false;
3563 init_reg_state(env
, regs
);
3564 state
->parent
= NULL
;
3567 struct bpf_insn
*insn
;
3571 if (insn_idx
>= insn_cnt
) {
3572 verbose(env
, "invalid insn idx %d insn_cnt %d\n",
3573 insn_idx
, insn_cnt
);
3577 insn
= &insns
[insn_idx
];
3578 class = BPF_CLASS(insn
->code
);
3580 if (++insn_processed
> BPF_COMPLEXITY_LIMIT_INSNS
) {
3582 "BPF program is too large. Processed %d insn\n",
3587 err
= is_state_visited(env
, insn_idx
);
3591 /* found equivalent state, can prune the search */
3592 if (env
->log
.level
) {
3594 verbose(env
, "\nfrom %d to %d: safe\n",
3595 prev_insn_idx
, insn_idx
);
3597 verbose(env
, "%d: safe\n", insn_idx
);
3599 goto process_bpf_exit
;
3605 if (env
->log
.level
> 1 || (env
->log
.level
&& do_print_state
)) {
3606 if (env
->log
.level
> 1)
3607 verbose(env
, "%d:", insn_idx
);
3609 verbose(env
, "\nfrom %d to %d:",
3610 prev_insn_idx
, insn_idx
);
3611 print_verifier_state(env
, &env
->cur_state
);
3612 do_print_state
= false;
3615 if (env
->log
.level
) {
3616 verbose(env
, "%d: ", insn_idx
);
3617 print_bpf_insn(verbose
, env
, insn
,
3618 env
->allow_ptr_leaks
);
3621 err
= ext_analyzer_insn_hook(env
, insn_idx
, prev_insn_idx
);
3625 if (class == BPF_ALU
|| class == BPF_ALU64
) {
3626 err
= check_alu_op(env
, insn
);
3630 } else if (class == BPF_LDX
) {
3631 enum bpf_reg_type
*prev_src_type
, src_reg_type
;
3633 /* check for reserved fields is already done */
3635 /* check src operand */
3636 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
3640 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP_NO_MARK
);
3644 src_reg_type
= regs
[insn
->src_reg
].type
;
3646 /* check that memory (src_reg + off) is readable,
3647 * the state of dst_reg will be updated by this func
3649 err
= check_mem_access(env
, insn_idx
, insn
->src_reg
, insn
->off
,
3650 BPF_SIZE(insn
->code
), BPF_READ
,
3655 prev_src_type
= &env
->insn_aux_data
[insn_idx
].ptr_type
;
3657 if (*prev_src_type
== NOT_INIT
) {
3659 * dst_reg = *(u32 *)(src_reg + off)
3660 * save type to validate intersecting paths
3662 *prev_src_type
= src_reg_type
;
3664 } else if (src_reg_type
!= *prev_src_type
&&
3665 (src_reg_type
== PTR_TO_CTX
||
3666 *prev_src_type
== PTR_TO_CTX
)) {
3667 /* ABuser program is trying to use the same insn
3668 * dst_reg = *(u32*) (src_reg + off)
3669 * with different pointer types:
3670 * src_reg == ctx in one branch and
3671 * src_reg == stack|map in some other branch.
3674 verbose(env
, "same insn cannot be used with different pointers\n");
3678 } else if (class == BPF_STX
) {
3679 enum bpf_reg_type
*prev_dst_type
, dst_reg_type
;
3681 if (BPF_MODE(insn
->code
) == BPF_XADD
) {
3682 err
= check_xadd(env
, insn_idx
, insn
);
3689 /* check src1 operand */
3690 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
3693 /* check src2 operand */
3694 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
3698 dst_reg_type
= regs
[insn
->dst_reg
].type
;
3700 /* check that memory (dst_reg + off) is writeable */
3701 err
= check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
3702 BPF_SIZE(insn
->code
), BPF_WRITE
,
3707 prev_dst_type
= &env
->insn_aux_data
[insn_idx
].ptr_type
;
3709 if (*prev_dst_type
== NOT_INIT
) {
3710 *prev_dst_type
= dst_reg_type
;
3711 } else if (dst_reg_type
!= *prev_dst_type
&&
3712 (dst_reg_type
== PTR_TO_CTX
||
3713 *prev_dst_type
== PTR_TO_CTX
)) {
3714 verbose(env
, "same insn cannot be used with different pointers\n");
3718 } else if (class == BPF_ST
) {
3719 if (BPF_MODE(insn
->code
) != BPF_MEM
||
3720 insn
->src_reg
!= BPF_REG_0
) {
3721 verbose(env
, "BPF_ST uses reserved fields\n");
3724 /* check src operand */
3725 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
3729 /* check that memory (dst_reg + off) is writeable */
3730 err
= check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
3731 BPF_SIZE(insn
->code
), BPF_WRITE
,
3736 } else if (class == BPF_JMP
) {
3737 u8 opcode
= BPF_OP(insn
->code
);
3739 if (opcode
== BPF_CALL
) {
3740 if (BPF_SRC(insn
->code
) != BPF_K
||
3742 insn
->src_reg
!= BPF_REG_0
||
3743 insn
->dst_reg
!= BPF_REG_0
) {
3744 verbose(env
, "BPF_CALL uses reserved fields\n");
3748 err
= check_call(env
, insn
->imm
, insn_idx
);
3752 } else if (opcode
== BPF_JA
) {
3753 if (BPF_SRC(insn
->code
) != BPF_K
||
3755 insn
->src_reg
!= BPF_REG_0
||
3756 insn
->dst_reg
!= BPF_REG_0
) {
3757 verbose(env
, "BPF_JA uses reserved fields\n");
3761 insn_idx
+= insn
->off
+ 1;
3764 } else if (opcode
== BPF_EXIT
) {
3765 if (BPF_SRC(insn
->code
) != BPF_K
||
3767 insn
->src_reg
!= BPF_REG_0
||
3768 insn
->dst_reg
!= BPF_REG_0
) {
3769 verbose(env
, "BPF_EXIT uses reserved fields\n");
3773 /* eBPF calling convetion is such that R0 is used
3774 * to return the value from eBPF program.
3775 * Make sure that it's readable at this time
3776 * of bpf_exit, which means that program wrote
3777 * something into it earlier
3779 err
= check_reg_arg(env
, BPF_REG_0
, SRC_OP
);
3783 if (is_pointer_value(env
, BPF_REG_0
)) {
3784 verbose(env
, "R0 leaks addr as return value\n");
3788 err
= check_return_code(env
);
3792 insn_idx
= pop_stack(env
, &prev_insn_idx
);
3796 do_print_state
= true;
3800 err
= check_cond_jmp_op(env
, insn
, &insn_idx
);
3804 } else if (class == BPF_LD
) {
3805 u8 mode
= BPF_MODE(insn
->code
);
3807 if (mode
== BPF_ABS
|| mode
== BPF_IND
) {
3808 err
= check_ld_abs(env
, insn
);
3812 } else if (mode
== BPF_IMM
) {
3813 err
= check_ld_imm(env
, insn
);
3819 verbose(env
, "invalid BPF_LD mode\n");
3823 verbose(env
, "unknown insn class %d\n", class);
3830 verbose(env
, "processed %d insns, stack depth %d\n", insn_processed
,
3831 env
->prog
->aux
->stack_depth
);
3835 static int check_map_prealloc(struct bpf_map
*map
)
3837 return (map
->map_type
!= BPF_MAP_TYPE_HASH
&&
3838 map
->map_type
!= BPF_MAP_TYPE_PERCPU_HASH
&&
3839 map
->map_type
!= BPF_MAP_TYPE_HASH_OF_MAPS
) ||
3840 !(map
->map_flags
& BPF_F_NO_PREALLOC
);
3843 static int check_map_prog_compatibility(struct bpf_verifier_env
*env
,
3844 struct bpf_map
*map
,
3845 struct bpf_prog
*prog
)
3848 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
3849 * preallocated hash maps, since doing memory allocation
3850 * in overflow_handler can crash depending on where nmi got
3853 if (prog
->type
== BPF_PROG_TYPE_PERF_EVENT
) {
3854 if (!check_map_prealloc(map
)) {
3855 verbose(env
, "perf_event programs can only use preallocated hash map\n");
3858 if (map
->inner_map_meta
&&
3859 !check_map_prealloc(map
->inner_map_meta
)) {
3860 verbose(env
, "perf_event programs can only use preallocated inner hash map\n");
3867 /* look for pseudo eBPF instructions that access map FDs and
3868 * replace them with actual map pointers
3870 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env
*env
)
3872 struct bpf_insn
*insn
= env
->prog
->insnsi
;
3873 int insn_cnt
= env
->prog
->len
;
3876 err
= bpf_prog_calc_tag(env
->prog
);
3880 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
3881 if (BPF_CLASS(insn
->code
) == BPF_LDX
&&
3882 (BPF_MODE(insn
->code
) != BPF_MEM
|| insn
->imm
!= 0)) {
3883 verbose(env
, "BPF_LDX uses reserved fields\n");
3887 if (BPF_CLASS(insn
->code
) == BPF_STX
&&
3888 ((BPF_MODE(insn
->code
) != BPF_MEM
&&
3889 BPF_MODE(insn
->code
) != BPF_XADD
) || insn
->imm
!= 0)) {
3890 verbose(env
, "BPF_STX uses reserved fields\n");
3894 if (insn
[0].code
== (BPF_LD
| BPF_IMM
| BPF_DW
)) {
3895 struct bpf_map
*map
;
3898 if (i
== insn_cnt
- 1 || insn
[1].code
!= 0 ||
3899 insn
[1].dst_reg
!= 0 || insn
[1].src_reg
!= 0 ||
3901 verbose(env
, "invalid bpf_ld_imm64 insn\n");
3905 if (insn
->src_reg
== 0)
3906 /* valid generic load 64-bit imm */
3909 if (insn
->src_reg
!= BPF_PSEUDO_MAP_FD
) {
3911 "unrecognized bpf_ld_imm64 insn\n");
3915 f
= fdget(insn
->imm
);
3916 map
= __bpf_map_get(f
);
3918 verbose(env
, "fd %d is not pointing to valid bpf_map\n",
3920 return PTR_ERR(map
);
3923 err
= check_map_prog_compatibility(env
, map
, env
->prog
);
3929 /* store map pointer inside BPF_LD_IMM64 instruction */
3930 insn
[0].imm
= (u32
) (unsigned long) map
;
3931 insn
[1].imm
= ((u64
) (unsigned long) map
) >> 32;
3933 /* check whether we recorded this map already */
3934 for (j
= 0; j
< env
->used_map_cnt
; j
++)
3935 if (env
->used_maps
[j
] == map
) {
3940 if (env
->used_map_cnt
>= MAX_USED_MAPS
) {
3945 /* hold the map. If the program is rejected by verifier,
3946 * the map will be released by release_maps() or it
3947 * will be used by the valid program until it's unloaded
3948 * and all maps are released in free_bpf_prog_info()
3950 map
= bpf_map_inc(map
, false);
3953 return PTR_ERR(map
);
3955 env
->used_maps
[env
->used_map_cnt
++] = map
;
3964 /* now all pseudo BPF_LD_IMM64 instructions load valid
3965 * 'struct bpf_map *' into a register instead of user map_fd.
3966 * These pointers will be used later by verifier to validate map access.
3971 /* drop refcnt of maps used by the rejected program */
3972 static void release_maps(struct bpf_verifier_env
*env
)
3976 for (i
= 0; i
< env
->used_map_cnt
; i
++)
3977 bpf_map_put(env
->used_maps
[i
]);
3980 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
3981 static void convert_pseudo_ld_imm64(struct bpf_verifier_env
*env
)
3983 struct bpf_insn
*insn
= env
->prog
->insnsi
;
3984 int insn_cnt
= env
->prog
->len
;
3987 for (i
= 0; i
< insn_cnt
; i
++, insn
++)
3988 if (insn
->code
== (BPF_LD
| BPF_IMM
| BPF_DW
))
3992 /* single env->prog->insni[off] instruction was replaced with the range
3993 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
3994 * [0, off) and [off, end) to new locations, so the patched range stays zero
3996 static int adjust_insn_aux_data(struct bpf_verifier_env
*env
, u32 prog_len
,
3999 struct bpf_insn_aux_data
*new_data
, *old_data
= env
->insn_aux_data
;
4003 new_data
= vzalloc(sizeof(struct bpf_insn_aux_data
) * prog_len
);
4006 memcpy(new_data
, old_data
, sizeof(struct bpf_insn_aux_data
) * off
);
4007 memcpy(new_data
+ off
+ cnt
- 1, old_data
+ off
,
4008 sizeof(struct bpf_insn_aux_data
) * (prog_len
- off
- cnt
+ 1));
4009 env
->insn_aux_data
= new_data
;
4014 static struct bpf_prog
*bpf_patch_insn_data(struct bpf_verifier_env
*env
, u32 off
,
4015 const struct bpf_insn
*patch
, u32 len
)
4017 struct bpf_prog
*new_prog
;
4019 new_prog
= bpf_patch_insn_single(env
->prog
, off
, patch
, len
);
4022 if (adjust_insn_aux_data(env
, new_prog
->len
, off
, len
))
4027 /* convert load instructions that access fields of 'struct __sk_buff'
4028 * into sequence of instructions that access fields of 'struct sk_buff'
4030 static int convert_ctx_accesses(struct bpf_verifier_env
*env
)
4032 const struct bpf_verifier_ops
*ops
= env
->prog
->aux
->ops
;
4033 int i
, cnt
, size
, ctx_field_size
, delta
= 0;
4034 const int insn_cnt
= env
->prog
->len
;
4035 struct bpf_insn insn_buf
[16], *insn
;
4036 struct bpf_prog
*new_prog
;
4037 enum bpf_access_type type
;
4038 bool is_narrower_load
;
4041 if (ops
->gen_prologue
) {
4042 cnt
= ops
->gen_prologue(insn_buf
, env
->seen_direct_write
,
4044 if (cnt
>= ARRAY_SIZE(insn_buf
)) {
4045 verbose(env
, "bpf verifier is misconfigured\n");
4048 new_prog
= bpf_patch_insn_data(env
, 0, insn_buf
, cnt
);
4052 env
->prog
= new_prog
;
4057 if (!ops
->convert_ctx_access
)
4060 insn
= env
->prog
->insnsi
+ delta
;
4062 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
4063 if (insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_B
) ||
4064 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_H
) ||
4065 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_W
) ||
4066 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_DW
))
4068 else if (insn
->code
== (BPF_STX
| BPF_MEM
| BPF_B
) ||
4069 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_H
) ||
4070 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_W
) ||
4071 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_DW
))
4076 if (env
->insn_aux_data
[i
+ delta
].ptr_type
!= PTR_TO_CTX
)
4079 ctx_field_size
= env
->insn_aux_data
[i
+ delta
].ctx_field_size
;
4080 size
= BPF_LDST_BYTES(insn
);
4082 /* If the read access is a narrower load of the field,
4083 * convert to a 4/8-byte load, to minimum program type specific
4084 * convert_ctx_access changes. If conversion is successful,
4085 * we will apply proper mask to the result.
4087 is_narrower_load
= size
< ctx_field_size
;
4088 if (is_narrower_load
) {
4089 u32 off
= insn
->off
;
4092 if (type
== BPF_WRITE
) {
4093 verbose(env
, "bpf verifier narrow ctx access misconfigured\n");
4098 if (ctx_field_size
== 4)
4100 else if (ctx_field_size
== 8)
4103 insn
->off
= off
& ~(ctx_field_size
- 1);
4104 insn
->code
= BPF_LDX
| BPF_MEM
| size_code
;
4108 cnt
= ops
->convert_ctx_access(type
, insn
, insn_buf
, env
->prog
,
4110 if (cnt
== 0 || cnt
>= ARRAY_SIZE(insn_buf
) ||
4111 (ctx_field_size
&& !target_size
)) {
4112 verbose(env
, "bpf verifier is misconfigured\n");
4116 if (is_narrower_load
&& size
< target_size
) {
4117 if (ctx_field_size
<= 4)
4118 insn_buf
[cnt
++] = BPF_ALU32_IMM(BPF_AND
, insn
->dst_reg
,
4119 (1 << size
* 8) - 1);
4121 insn_buf
[cnt
++] = BPF_ALU64_IMM(BPF_AND
, insn
->dst_reg
,
4122 (1 << size
* 8) - 1);
4125 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
, cnt
);
4131 /* keep walking new program and skip insns we just inserted */
4132 env
->prog
= new_prog
;
4133 insn
= new_prog
->insnsi
+ i
+ delta
;
4139 /* fixup insn->imm field of bpf_call instructions
4140 * and inline eligible helpers as explicit sequence of BPF instructions
4142 * this function is called after eBPF program passed verification
4144 static int fixup_bpf_calls(struct bpf_verifier_env
*env
)
4146 struct bpf_prog
*prog
= env
->prog
;
4147 struct bpf_insn
*insn
= prog
->insnsi
;
4148 const struct bpf_func_proto
*fn
;
4149 const int insn_cnt
= prog
->len
;
4150 struct bpf_insn insn_buf
[16];
4151 struct bpf_prog
*new_prog
;
4152 struct bpf_map
*map_ptr
;
4153 int i
, cnt
, delta
= 0;
4155 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
4156 if (insn
->code
!= (BPF_JMP
| BPF_CALL
))
4159 if (insn
->imm
== BPF_FUNC_get_route_realm
)
4160 prog
->dst_needed
= 1;
4161 if (insn
->imm
== BPF_FUNC_get_prandom_u32
)
4162 bpf_user_rnd_init_once();
4163 if (insn
->imm
== BPF_FUNC_tail_call
) {
4164 /* If we tail call into other programs, we
4165 * cannot make any assumptions since they can
4166 * be replaced dynamically during runtime in
4167 * the program array.
4169 prog
->cb_access
= 1;
4170 env
->prog
->aux
->stack_depth
= MAX_BPF_STACK
;
4172 /* mark bpf_tail_call as different opcode to avoid
4173 * conditional branch in the interpeter for every normal
4174 * call and to prevent accidental JITing by JIT compiler
4175 * that doesn't support bpf_tail_call yet
4178 insn
->code
= BPF_JMP
| BPF_TAIL_CALL
;
4182 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
4183 * handlers are currently limited to 64 bit only.
4185 if (ebpf_jit_enabled() && BITS_PER_LONG
== 64 &&
4186 insn
->imm
== BPF_FUNC_map_lookup_elem
) {
4187 map_ptr
= env
->insn_aux_data
[i
+ delta
].map_ptr
;
4188 if (map_ptr
== BPF_MAP_PTR_POISON
||
4189 !map_ptr
->ops
->map_gen_lookup
)
4190 goto patch_call_imm
;
4192 cnt
= map_ptr
->ops
->map_gen_lookup(map_ptr
, insn_buf
);
4193 if (cnt
== 0 || cnt
>= ARRAY_SIZE(insn_buf
)) {
4194 verbose(env
, "bpf verifier is misconfigured\n");
4198 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
,
4205 /* keep walking new program and skip insns we just inserted */
4206 env
->prog
= prog
= new_prog
;
4207 insn
= new_prog
->insnsi
+ i
+ delta
;
4211 if (insn
->imm
== BPF_FUNC_redirect_map
) {
4212 /* Note, we cannot use prog directly as imm as subsequent
4213 * rewrites would still change the prog pointer. The only
4214 * stable address we can use is aux, which also works with
4215 * prog clones during blinding.
4217 u64 addr
= (unsigned long)prog
->aux
;
4218 struct bpf_insn r4_ld
[] = {
4219 BPF_LD_IMM64(BPF_REG_4
, addr
),
4222 cnt
= ARRAY_SIZE(r4_ld
);
4224 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, r4_ld
, cnt
);
4229 env
->prog
= prog
= new_prog
;
4230 insn
= new_prog
->insnsi
+ i
+ delta
;
4233 fn
= prog
->aux
->ops
->get_func_proto(insn
->imm
);
4234 /* all functions that have prototype and verifier allowed
4235 * programs to call them, must be real in-kernel functions
4239 "kernel subsystem misconfigured func %s#%d\n",
4240 func_id_name(insn
->imm
), insn
->imm
);
4243 insn
->imm
= fn
->func
- __bpf_call_base
;
4249 static void free_states(struct bpf_verifier_env
*env
)
4251 struct bpf_verifier_state_list
*sl
, *sln
;
4254 if (!env
->explored_states
)
4257 for (i
= 0; i
< env
->prog
->len
; i
++) {
4258 sl
= env
->explored_states
[i
];
4261 while (sl
!= STATE_LIST_MARK
) {
4268 kfree(env
->explored_states
);
4271 int bpf_check(struct bpf_prog
**prog
, union bpf_attr
*attr
)
4273 struct bpf_verifier_env
*env
;
4274 struct bpf_verifer_log
*log
;
4277 /* 'struct bpf_verifier_env' can be global, but since it's not small,
4278 * allocate/free it every time bpf_check() is called
4280 env
= kzalloc(sizeof(struct bpf_verifier_env
), GFP_KERNEL
);
4285 env
->insn_aux_data
= vzalloc(sizeof(struct bpf_insn_aux_data
) *
4288 if (!env
->insn_aux_data
)
4292 /* grab the mutex to protect few globals used by verifier */
4293 mutex_lock(&bpf_verifier_lock
);
4295 if (attr
->log_level
|| attr
->log_buf
|| attr
->log_size
) {
4296 /* user requested verbose verifier output
4297 * and supplied buffer to store the verification trace
4299 log
->level
= attr
->log_level
;
4300 log
->ubuf
= (char __user
*) (unsigned long) attr
->log_buf
;
4301 log
->len_total
= attr
->log_size
;
4304 /* log attributes have to be sane */
4305 if (log
->len_total
< 128 || log
->len_total
> UINT_MAX
>> 8 ||
4306 !log
->level
|| !log
->ubuf
)
4310 env
->strict_alignment
= !!(attr
->prog_flags
& BPF_F_STRICT_ALIGNMENT
);
4311 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
))
4312 env
->strict_alignment
= true;
4314 ret
= replace_map_fd_with_map_ptr(env
);
4316 goto skip_full_check
;
4318 env
->explored_states
= kcalloc(env
->prog
->len
,
4319 sizeof(struct bpf_verifier_state_list
*),
4322 if (!env
->explored_states
)
4323 goto skip_full_check
;
4325 ret
= check_cfg(env
);
4327 goto skip_full_check
;
4329 env
->allow_ptr_leaks
= capable(CAP_SYS_ADMIN
);
4331 ret
= do_check(env
);
4334 while (pop_stack(env
, NULL
) >= 0);
4338 /* program is valid, convert *(u32*)(ctx + off) accesses */
4339 ret
= convert_ctx_accesses(env
);
4342 ret
= fixup_bpf_calls(env
);
4344 if (log
->level
&& bpf_verifier_log_full(log
))
4346 if (log
->level
&& !log
->ubuf
) {
4348 goto err_release_maps
;
4351 if (ret
== 0 && env
->used_map_cnt
) {
4352 /* if program passed verifier, update used_maps in bpf_prog_info */
4353 env
->prog
->aux
->used_maps
= kmalloc_array(env
->used_map_cnt
,
4354 sizeof(env
->used_maps
[0]),
4357 if (!env
->prog
->aux
->used_maps
) {
4359 goto err_release_maps
;
4362 memcpy(env
->prog
->aux
->used_maps
, env
->used_maps
,
4363 sizeof(env
->used_maps
[0]) * env
->used_map_cnt
);
4364 env
->prog
->aux
->used_map_cnt
= env
->used_map_cnt
;
4366 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
4367 * bpf_ld_imm64 instructions
4369 convert_pseudo_ld_imm64(env
);
4373 if (!env
->prog
->aux
->used_maps
)
4374 /* if we didn't copy map pointers into bpf_prog_info, release
4375 * them now. Otherwise free_bpf_prog_info() will release them.
4380 mutex_unlock(&bpf_verifier_lock
);
4381 vfree(env
->insn_aux_data
);
4387 int bpf_analyzer(struct bpf_prog
*prog
, const struct bpf_ext_analyzer_ops
*ops
,
4390 struct bpf_verifier_env
*env
;
4393 env
= kzalloc(sizeof(struct bpf_verifier_env
), GFP_KERNEL
);
4397 env
->insn_aux_data
= vzalloc(sizeof(struct bpf_insn_aux_data
) *
4400 if (!env
->insn_aux_data
)
4403 env
->analyzer_ops
= ops
;
4404 env
->analyzer_priv
= priv
;
4406 /* grab the mutex to protect few globals used by verifier */
4407 mutex_lock(&bpf_verifier_lock
);
4409 env
->strict_alignment
= false;
4410 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
))
4411 env
->strict_alignment
= true;
4413 env
->explored_states
= kcalloc(env
->prog
->len
,
4414 sizeof(struct bpf_verifier_state_list
*),
4417 if (!env
->explored_states
)
4418 goto skip_full_check
;
4420 ret
= check_cfg(env
);
4422 goto skip_full_check
;
4424 env
->allow_ptr_leaks
= capable(CAP_SYS_ADMIN
);
4426 ret
= do_check(env
);
4429 while (pop_stack(env
, NULL
) >= 0);
4432 mutex_unlock(&bpf_verifier_lock
);
4433 vfree(env
->insn_aux_data
);
4438 EXPORT_SYMBOL_GPL(bpf_analyzer
);