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>
24 /* bpf_check() is a static code analyzer that walks eBPF program
25 * instruction by instruction and updates register/stack state.
26 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
28 * The first pass is depth-first-search to check that the program is a DAG.
29 * It rejects the following programs:
30 * - larger than BPF_MAXINSNS insns
31 * - if loop is present (detected via back-edge)
32 * - unreachable insns exist (shouldn't be a forest. program = one function)
33 * - out of bounds or malformed jumps
34 * The second pass is all possible path descent from the 1st insn.
35 * Since it's analyzing all pathes through the program, the length of the
36 * analysis is limited to 64k insn, which may be hit even if total number of
37 * insn is less then 4K, but there are too many branches that change stack/regs.
38 * Number of 'branches to be analyzed' is limited to 1k
40 * On entry to each instruction, each register has a type, and the instruction
41 * changes the types of the registers depending on instruction semantics.
42 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
45 * All registers are 64-bit.
46 * R0 - return register
47 * R1-R5 argument passing registers
48 * R6-R9 callee saved registers
49 * R10 - frame pointer read-only
51 * At the start of BPF program the register R1 contains a pointer to bpf_context
52 * and has type PTR_TO_CTX.
54 * Verifier tracks arithmetic operations on pointers in case:
55 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
56 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
57 * 1st insn copies R10 (which has FRAME_PTR) type into R1
58 * and 2nd arithmetic instruction is pattern matched to recognize
59 * that it wants to construct a pointer to some element within stack.
60 * So after 2nd insn, the register R1 has type PTR_TO_STACK
61 * (and -20 constant is saved for further stack bounds checking).
62 * Meaning that this reg is a pointer to stack plus known immediate constant.
64 * Most of the time the registers have SCALAR_VALUE type, which
65 * means the register has some value, but it's not a valid pointer.
66 * (like pointer plus pointer becomes SCALAR_VALUE type)
68 * When verifier sees load or store instructions the type of base register
69 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK. These are three pointer
70 * types recognized by check_mem_access() function.
72 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
73 * and the range of [ptr, ptr + map's value_size) is accessible.
75 * registers used to pass values to function calls are checked against
76 * function argument constraints.
78 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
79 * It means that the register type passed to this function must be
80 * PTR_TO_STACK and it will be used inside the function as
81 * 'pointer to map element key'
83 * For example the argument constraints for bpf_map_lookup_elem():
84 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
85 * .arg1_type = ARG_CONST_MAP_PTR,
86 * .arg2_type = ARG_PTR_TO_MAP_KEY,
88 * ret_type says that this function returns 'pointer to map elem value or null'
89 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
90 * 2nd argument should be a pointer to stack, which will be used inside
91 * the helper function as a pointer to map element key.
93 * On the kernel side the helper function looks like:
94 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
96 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
97 * void *key = (void *) (unsigned long) r2;
100 * here kernel can access 'key' and 'map' pointers safely, knowing that
101 * [key, key + map->key_size) bytes are valid and were initialized on
102 * the stack of eBPF program.
105 * Corresponding eBPF program may look like:
106 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
107 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
108 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
109 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
110 * here verifier looks at prototype of map_lookup_elem() and sees:
111 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
112 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
114 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
115 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
116 * and were initialized prior to this call.
117 * If it's ok, then verifier allows this BPF_CALL insn and looks at
118 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
119 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
120 * returns ether pointer to map value or NULL.
122 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
123 * insn, the register holding that pointer in the true branch changes state to
124 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
125 * branch. See check_cond_jmp_op().
127 * After the call R0 is set to return type of the function and registers R1-R5
128 * are set to NOT_INIT to indicate that they are no longer readable.
131 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
132 struct bpf_verifier_stack_elem
{
133 /* verifer state is 'st'
134 * before processing instruction 'insn_idx'
135 * and after processing instruction 'prev_insn_idx'
137 struct bpf_verifier_state st
;
140 struct bpf_verifier_stack_elem
*next
;
143 #define BPF_COMPLEXITY_LIMIT_INSNS 131072
144 #define BPF_COMPLEXITY_LIMIT_STACK 1024
146 #define BPF_MAP_PTR_POISON ((void *)0xeB9F + POISON_POINTER_DELTA)
148 struct bpf_call_arg_meta
{
149 struct bpf_map
*map_ptr
;
156 /* verbose verifier prints what it's seeing
157 * bpf_check() is called under lock, so no race to access these global vars
159 static u32 log_level
, log_size
, log_len
;
160 static char *log_buf
;
162 static DEFINE_MUTEX(bpf_verifier_lock
);
164 /* log_level controls verbosity level of eBPF verifier.
165 * verbose() is used to dump the verification trace to the log, so the user
166 * can figure out what's wrong with the program
168 static __printf(1, 2) void verbose(const char *fmt
, ...)
172 if (log_level
== 0 || log_len
>= log_size
- 1)
176 log_len
+= vscnprintf(log_buf
+ log_len
, log_size
- log_len
, fmt
, args
);
180 static bool type_is_pkt_pointer(enum bpf_reg_type type
)
182 return type
== PTR_TO_PACKET
||
183 type
== PTR_TO_PACKET_META
;
186 /* string representation of 'enum bpf_reg_type' */
187 static const char * const reg_type_str
[] = {
189 [SCALAR_VALUE
] = "inv",
190 [PTR_TO_CTX
] = "ctx",
191 [CONST_PTR_TO_MAP
] = "map_ptr",
192 [PTR_TO_MAP_VALUE
] = "map_value",
193 [PTR_TO_MAP_VALUE_OR_NULL
] = "map_value_or_null",
194 [PTR_TO_STACK
] = "fp",
195 [PTR_TO_PACKET
] = "pkt",
196 [PTR_TO_PACKET_META
] = "pkt_meta",
197 [PTR_TO_PACKET_END
] = "pkt_end",
200 #define __BPF_FUNC_STR_FN(x) [BPF_FUNC_ ## x] = __stringify(bpf_ ## x)
201 static const char * const func_id_str
[] = {
202 __BPF_FUNC_MAPPER(__BPF_FUNC_STR_FN
)
204 #undef __BPF_FUNC_STR_FN
206 static const char *func_id_name(int id
)
208 BUILD_BUG_ON(ARRAY_SIZE(func_id_str
) != __BPF_FUNC_MAX_ID
);
210 if (id
>= 0 && id
< __BPF_FUNC_MAX_ID
&& func_id_str
[id
])
211 return func_id_str
[id
];
216 static void print_verifier_state(struct bpf_verifier_state
*state
)
218 struct bpf_reg_state
*reg
;
222 for (i
= 0; i
< MAX_BPF_REG
; i
++) {
223 reg
= &state
->regs
[i
];
227 verbose(" R%d=%s", i
, reg_type_str
[t
]);
228 if ((t
== SCALAR_VALUE
|| t
== PTR_TO_STACK
) &&
229 tnum_is_const(reg
->var_off
)) {
230 /* reg->off should be 0 for SCALAR_VALUE */
231 verbose("%lld", reg
->var_off
.value
+ reg
->off
);
233 verbose("(id=%d", reg
->id
);
234 if (t
!= SCALAR_VALUE
)
235 verbose(",off=%d", reg
->off
);
236 if (type_is_pkt_pointer(t
))
237 verbose(",r=%d", reg
->range
);
238 else if (t
== CONST_PTR_TO_MAP
||
239 t
== PTR_TO_MAP_VALUE
||
240 t
== PTR_TO_MAP_VALUE_OR_NULL
)
241 verbose(",ks=%d,vs=%d",
242 reg
->map_ptr
->key_size
,
243 reg
->map_ptr
->value_size
);
244 if (tnum_is_const(reg
->var_off
)) {
245 /* Typically an immediate SCALAR_VALUE, but
246 * could be a pointer whose offset is too big
249 verbose(",imm=%llx", reg
->var_off
.value
);
251 if (reg
->smin_value
!= reg
->umin_value
&&
252 reg
->smin_value
!= S64_MIN
)
253 verbose(",smin_value=%lld",
254 (long long)reg
->smin_value
);
255 if (reg
->smax_value
!= reg
->umax_value
&&
256 reg
->smax_value
!= S64_MAX
)
257 verbose(",smax_value=%lld",
258 (long long)reg
->smax_value
);
259 if (reg
->umin_value
!= 0)
260 verbose(",umin_value=%llu",
261 (unsigned long long)reg
->umin_value
);
262 if (reg
->umax_value
!= U64_MAX
)
263 verbose(",umax_value=%llu",
264 (unsigned long long)reg
->umax_value
);
265 if (!tnum_is_unknown(reg
->var_off
)) {
268 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
269 verbose(",var_off=%s", tn_buf
);
275 for (i
= 0; i
< MAX_BPF_STACK
; i
+= BPF_REG_SIZE
) {
276 if (state
->stack_slot_type
[i
] == STACK_SPILL
)
277 verbose(" fp%d=%s", -MAX_BPF_STACK
+ i
,
278 reg_type_str
[state
->spilled_regs
[i
/ BPF_REG_SIZE
].type
]);
283 static const char *const bpf_class_string
[] = {
291 [BPF_ALU64
] = "alu64",
294 static const char *const bpf_alu_string
[16] = {
295 [BPF_ADD
>> 4] = "+=",
296 [BPF_SUB
>> 4] = "-=",
297 [BPF_MUL
>> 4] = "*=",
298 [BPF_DIV
>> 4] = "/=",
299 [BPF_OR
>> 4] = "|=",
300 [BPF_AND
>> 4] = "&=",
301 [BPF_LSH
>> 4] = "<<=",
302 [BPF_RSH
>> 4] = ">>=",
303 [BPF_NEG
>> 4] = "neg",
304 [BPF_MOD
>> 4] = "%=",
305 [BPF_XOR
>> 4] = "^=",
306 [BPF_MOV
>> 4] = "=",
307 [BPF_ARSH
>> 4] = "s>>=",
308 [BPF_END
>> 4] = "endian",
311 static const char *const bpf_ldst_string
[] = {
312 [BPF_W
>> 3] = "u32",
313 [BPF_H
>> 3] = "u16",
315 [BPF_DW
>> 3] = "u64",
318 static const char *const bpf_jmp_string
[16] = {
319 [BPF_JA
>> 4] = "jmp",
320 [BPF_JEQ
>> 4] = "==",
321 [BPF_JGT
>> 4] = ">",
322 [BPF_JLT
>> 4] = "<",
323 [BPF_JGE
>> 4] = ">=",
324 [BPF_JLE
>> 4] = "<=",
325 [BPF_JSET
>> 4] = "&",
326 [BPF_JNE
>> 4] = "!=",
327 [BPF_JSGT
>> 4] = "s>",
328 [BPF_JSLT
>> 4] = "s<",
329 [BPF_JSGE
>> 4] = "s>=",
330 [BPF_JSLE
>> 4] = "s<=",
331 [BPF_CALL
>> 4] = "call",
332 [BPF_EXIT
>> 4] = "exit",
335 static void print_bpf_end_insn(const struct bpf_verifier_env
*env
,
336 const struct bpf_insn
*insn
)
338 verbose("(%02x) r%d = %s%d r%d\n", insn
->code
, insn
->dst_reg
,
339 BPF_SRC(insn
->code
) == BPF_TO_BE
? "be" : "le",
340 insn
->imm
, insn
->dst_reg
);
343 static void print_bpf_insn(const struct bpf_verifier_env
*env
,
344 const struct bpf_insn
*insn
)
346 u8
class = BPF_CLASS(insn
->code
);
348 if (class == BPF_ALU
|| class == BPF_ALU64
) {
349 if (BPF_OP(insn
->code
) == BPF_END
) {
350 if (class == BPF_ALU64
)
351 verbose("BUG_alu64_%02x\n", insn
->code
);
353 print_bpf_end_insn(env
, insn
);
354 } else if (BPF_OP(insn
->code
) == BPF_NEG
) {
355 verbose("(%02x) r%d = %s-r%d\n",
356 insn
->code
, insn
->dst_reg
,
357 class == BPF_ALU
? "(u32) " : "",
359 } else if (BPF_SRC(insn
->code
) == BPF_X
) {
360 verbose("(%02x) %sr%d %s %sr%d\n",
361 insn
->code
, class == BPF_ALU
? "(u32) " : "",
363 bpf_alu_string
[BPF_OP(insn
->code
) >> 4],
364 class == BPF_ALU
? "(u32) " : "",
367 verbose("(%02x) %sr%d %s %s%d\n",
368 insn
->code
, class == BPF_ALU
? "(u32) " : "",
370 bpf_alu_string
[BPF_OP(insn
->code
) >> 4],
371 class == BPF_ALU
? "(u32) " : "",
374 } else if (class == BPF_STX
) {
375 if (BPF_MODE(insn
->code
) == BPF_MEM
)
376 verbose("(%02x) *(%s *)(r%d %+d) = r%d\n",
378 bpf_ldst_string
[BPF_SIZE(insn
->code
) >> 3],
380 insn
->off
, insn
->src_reg
);
381 else if (BPF_MODE(insn
->code
) == BPF_XADD
)
382 verbose("(%02x) lock *(%s *)(r%d %+d) += r%d\n",
384 bpf_ldst_string
[BPF_SIZE(insn
->code
) >> 3],
385 insn
->dst_reg
, insn
->off
,
388 verbose("BUG_%02x\n", insn
->code
);
389 } else if (class == BPF_ST
) {
390 if (BPF_MODE(insn
->code
) != BPF_MEM
) {
391 verbose("BUG_st_%02x\n", insn
->code
);
394 verbose("(%02x) *(%s *)(r%d %+d) = %d\n",
396 bpf_ldst_string
[BPF_SIZE(insn
->code
) >> 3],
398 insn
->off
, insn
->imm
);
399 } else if (class == BPF_LDX
) {
400 if (BPF_MODE(insn
->code
) != BPF_MEM
) {
401 verbose("BUG_ldx_%02x\n", insn
->code
);
404 verbose("(%02x) r%d = *(%s *)(r%d %+d)\n",
405 insn
->code
, insn
->dst_reg
,
406 bpf_ldst_string
[BPF_SIZE(insn
->code
) >> 3],
407 insn
->src_reg
, insn
->off
);
408 } else if (class == BPF_LD
) {
409 if (BPF_MODE(insn
->code
) == BPF_ABS
) {
410 verbose("(%02x) r0 = *(%s *)skb[%d]\n",
412 bpf_ldst_string
[BPF_SIZE(insn
->code
) >> 3],
414 } else if (BPF_MODE(insn
->code
) == BPF_IND
) {
415 verbose("(%02x) r0 = *(%s *)skb[r%d + %d]\n",
417 bpf_ldst_string
[BPF_SIZE(insn
->code
) >> 3],
418 insn
->src_reg
, insn
->imm
);
419 } else if (BPF_MODE(insn
->code
) == BPF_IMM
&&
420 BPF_SIZE(insn
->code
) == BPF_DW
) {
421 /* At this point, we already made sure that the second
422 * part of the ldimm64 insn is accessible.
424 u64 imm
= ((u64
)(insn
+ 1)->imm
<< 32) | (u32
)insn
->imm
;
425 bool map_ptr
= insn
->src_reg
== BPF_PSEUDO_MAP_FD
;
427 if (map_ptr
&& !env
->allow_ptr_leaks
)
430 verbose("(%02x) r%d = 0x%llx\n", insn
->code
,
431 insn
->dst_reg
, (unsigned long long)imm
);
433 verbose("BUG_ld_%02x\n", insn
->code
);
436 } else if (class == BPF_JMP
) {
437 u8 opcode
= BPF_OP(insn
->code
);
439 if (opcode
== BPF_CALL
) {
440 verbose("(%02x) call %s#%d\n", insn
->code
,
441 func_id_name(insn
->imm
), insn
->imm
);
442 } else if (insn
->code
== (BPF_JMP
| BPF_JA
)) {
443 verbose("(%02x) goto pc%+d\n",
444 insn
->code
, insn
->off
);
445 } else if (insn
->code
== (BPF_JMP
| BPF_EXIT
)) {
446 verbose("(%02x) exit\n", insn
->code
);
447 } else if (BPF_SRC(insn
->code
) == BPF_X
) {
448 verbose("(%02x) if r%d %s r%d goto pc%+d\n",
449 insn
->code
, insn
->dst_reg
,
450 bpf_jmp_string
[BPF_OP(insn
->code
) >> 4],
451 insn
->src_reg
, insn
->off
);
453 verbose("(%02x) if r%d %s 0x%x goto pc%+d\n",
454 insn
->code
, insn
->dst_reg
,
455 bpf_jmp_string
[BPF_OP(insn
->code
) >> 4],
456 insn
->imm
, insn
->off
);
459 verbose("(%02x) %s\n", insn
->code
, bpf_class_string
[class]);
463 static int pop_stack(struct bpf_verifier_env
*env
, int *prev_insn_idx
)
465 struct bpf_verifier_stack_elem
*elem
;
468 if (env
->head
== NULL
)
471 memcpy(&env
->cur_state
, &env
->head
->st
, sizeof(env
->cur_state
));
472 insn_idx
= env
->head
->insn_idx
;
474 *prev_insn_idx
= env
->head
->prev_insn_idx
;
475 elem
= env
->head
->next
;
482 static struct bpf_verifier_state
*push_stack(struct bpf_verifier_env
*env
,
483 int insn_idx
, int prev_insn_idx
)
485 struct bpf_verifier_stack_elem
*elem
;
487 elem
= kmalloc(sizeof(struct bpf_verifier_stack_elem
), GFP_KERNEL
);
491 memcpy(&elem
->st
, &env
->cur_state
, sizeof(env
->cur_state
));
492 elem
->insn_idx
= insn_idx
;
493 elem
->prev_insn_idx
= prev_insn_idx
;
494 elem
->next
= env
->head
;
497 if (env
->stack_size
> BPF_COMPLEXITY_LIMIT_STACK
) {
498 verbose("BPF program is too complex\n");
503 /* pop all elements and return */
504 while (pop_stack(env
, NULL
) >= 0);
508 #define CALLER_SAVED_REGS 6
509 static const int caller_saved
[CALLER_SAVED_REGS
] = {
510 BPF_REG_0
, BPF_REG_1
, BPF_REG_2
, BPF_REG_3
, BPF_REG_4
, BPF_REG_5
513 static void __mark_reg_not_init(struct bpf_reg_state
*reg
);
515 /* Mark the unknown part of a register (variable offset or scalar value) as
516 * known to have the value @imm.
518 static void __mark_reg_known(struct bpf_reg_state
*reg
, u64 imm
)
521 reg
->var_off
= tnum_const(imm
);
522 reg
->smin_value
= (s64
)imm
;
523 reg
->smax_value
= (s64
)imm
;
524 reg
->umin_value
= imm
;
525 reg
->umax_value
= imm
;
528 /* Mark the 'variable offset' part of a register as zero. This should be
529 * used only on registers holding a pointer type.
531 static void __mark_reg_known_zero(struct bpf_reg_state
*reg
)
533 __mark_reg_known(reg
, 0);
536 static void mark_reg_known_zero(struct bpf_reg_state
*regs
, u32 regno
)
538 if (WARN_ON(regno
>= MAX_BPF_REG
)) {
539 verbose("mark_reg_known_zero(regs, %u)\n", regno
);
540 /* Something bad happened, let's kill all regs */
541 for (regno
= 0; regno
< MAX_BPF_REG
; regno
++)
542 __mark_reg_not_init(regs
+ regno
);
545 __mark_reg_known_zero(regs
+ regno
);
548 static bool reg_is_pkt_pointer(const struct bpf_reg_state
*reg
)
550 return type_is_pkt_pointer(reg
->type
);
553 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state
*reg
)
555 return reg_is_pkt_pointer(reg
) ||
556 reg
->type
== PTR_TO_PACKET_END
;
559 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
560 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state
*reg
,
561 enum bpf_reg_type which
)
563 /* The register can already have a range from prior markings.
564 * This is fine as long as it hasn't been advanced from its
567 return reg
->type
== which
&&
570 tnum_equals_const(reg
->var_off
, 0);
573 /* Attempts to improve min/max values based on var_off information */
574 static void __update_reg_bounds(struct bpf_reg_state
*reg
)
576 /* min signed is max(sign bit) | min(other bits) */
577 reg
->smin_value
= max_t(s64
, reg
->smin_value
,
578 reg
->var_off
.value
| (reg
->var_off
.mask
& S64_MIN
));
579 /* max signed is min(sign bit) | max(other bits) */
580 reg
->smax_value
= min_t(s64
, reg
->smax_value
,
581 reg
->var_off
.value
| (reg
->var_off
.mask
& S64_MAX
));
582 reg
->umin_value
= max(reg
->umin_value
, reg
->var_off
.value
);
583 reg
->umax_value
= min(reg
->umax_value
,
584 reg
->var_off
.value
| reg
->var_off
.mask
);
587 /* Uses signed min/max values to inform unsigned, and vice-versa */
588 static void __reg_deduce_bounds(struct bpf_reg_state
*reg
)
590 /* Learn sign from signed bounds.
591 * If we cannot cross the sign boundary, then signed and unsigned bounds
592 * are the same, so combine. This works even in the negative case, e.g.
593 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
595 if (reg
->smin_value
>= 0 || reg
->smax_value
< 0) {
596 reg
->smin_value
= reg
->umin_value
= max_t(u64
, reg
->smin_value
,
598 reg
->smax_value
= reg
->umax_value
= min_t(u64
, reg
->smax_value
,
602 /* Learn sign from unsigned bounds. Signed bounds cross the sign
603 * boundary, so we must be careful.
605 if ((s64
)reg
->umax_value
>= 0) {
606 /* Positive. We can't learn anything from the smin, but smax
607 * is positive, hence safe.
609 reg
->smin_value
= reg
->umin_value
;
610 reg
->smax_value
= reg
->umax_value
= min_t(u64
, reg
->smax_value
,
612 } else if ((s64
)reg
->umin_value
< 0) {
613 /* Negative. We can't learn anything from the smax, but smin
614 * is negative, hence safe.
616 reg
->smin_value
= reg
->umin_value
= max_t(u64
, reg
->smin_value
,
618 reg
->smax_value
= reg
->umax_value
;
622 /* Attempts to improve var_off based on unsigned min/max information */
623 static void __reg_bound_offset(struct bpf_reg_state
*reg
)
625 reg
->var_off
= tnum_intersect(reg
->var_off
,
626 tnum_range(reg
->umin_value
,
630 /* Reset the min/max bounds of a register */
631 static void __mark_reg_unbounded(struct bpf_reg_state
*reg
)
633 reg
->smin_value
= S64_MIN
;
634 reg
->smax_value
= S64_MAX
;
636 reg
->umax_value
= U64_MAX
;
639 /* Mark a register as having a completely unknown (scalar) value. */
640 static void __mark_reg_unknown(struct bpf_reg_state
*reg
)
642 reg
->type
= SCALAR_VALUE
;
645 reg
->var_off
= tnum_unknown
;
646 __mark_reg_unbounded(reg
);
649 static void mark_reg_unknown(struct bpf_reg_state
*regs
, u32 regno
)
651 if (WARN_ON(regno
>= MAX_BPF_REG
)) {
652 verbose("mark_reg_unknown(regs, %u)\n", regno
);
653 /* Something bad happened, let's kill all regs */
654 for (regno
= 0; regno
< MAX_BPF_REG
; regno
++)
655 __mark_reg_not_init(regs
+ regno
);
658 __mark_reg_unknown(regs
+ regno
);
661 static void __mark_reg_not_init(struct bpf_reg_state
*reg
)
663 __mark_reg_unknown(reg
);
664 reg
->type
= NOT_INIT
;
667 static void mark_reg_not_init(struct bpf_reg_state
*regs
, u32 regno
)
669 if (WARN_ON(regno
>= MAX_BPF_REG
)) {
670 verbose("mark_reg_not_init(regs, %u)\n", regno
);
671 /* Something bad happened, let's kill all regs */
672 for (regno
= 0; regno
< MAX_BPF_REG
; regno
++)
673 __mark_reg_not_init(regs
+ regno
);
676 __mark_reg_not_init(regs
+ regno
);
679 static void init_reg_state(struct bpf_reg_state
*regs
)
683 for (i
= 0; i
< MAX_BPF_REG
; i
++) {
684 mark_reg_not_init(regs
, i
);
685 regs
[i
].live
= REG_LIVE_NONE
;
689 regs
[BPF_REG_FP
].type
= PTR_TO_STACK
;
690 mark_reg_known_zero(regs
, BPF_REG_FP
);
692 /* 1st arg to a function */
693 regs
[BPF_REG_1
].type
= PTR_TO_CTX
;
694 mark_reg_known_zero(regs
, BPF_REG_1
);
698 SRC_OP
, /* register is used as source operand */
699 DST_OP
, /* register is used as destination operand */
700 DST_OP_NO_MARK
/* same as above, check only, don't mark */
703 static void mark_reg_read(const struct bpf_verifier_state
*state
, u32 regno
)
705 struct bpf_verifier_state
*parent
= state
->parent
;
708 /* if read wasn't screened by an earlier write ... */
709 if (state
->regs
[regno
].live
& REG_LIVE_WRITTEN
)
711 /* ... then we depend on parent's value */
712 parent
->regs
[regno
].live
|= REG_LIVE_READ
;
714 parent
= state
->parent
;
718 static int check_reg_arg(struct bpf_verifier_env
*env
, u32 regno
,
721 struct bpf_reg_state
*regs
= env
->cur_state
.regs
;
723 if (regno
>= MAX_BPF_REG
) {
724 verbose("R%d is invalid\n", regno
);
729 /* check whether register used as source operand can be read */
730 if (regs
[regno
].type
== NOT_INIT
) {
731 verbose("R%d !read_ok\n", regno
);
734 mark_reg_read(&env
->cur_state
, regno
);
736 /* check whether register used as dest operand can be written to */
737 if (regno
== BPF_REG_FP
) {
738 verbose("frame pointer is read only\n");
741 regs
[regno
].live
|= REG_LIVE_WRITTEN
;
743 mark_reg_unknown(regs
, regno
);
748 static bool is_spillable_regtype(enum bpf_reg_type type
)
751 case PTR_TO_MAP_VALUE
:
752 case PTR_TO_MAP_VALUE_OR_NULL
:
756 case PTR_TO_PACKET_META
:
757 case PTR_TO_PACKET_END
:
758 case CONST_PTR_TO_MAP
:
765 /* check_stack_read/write functions track spill/fill of registers,
766 * stack boundary and alignment are checked in check_mem_access()
768 static int check_stack_write(struct bpf_verifier_state
*state
, int off
,
769 int size
, int value_regno
)
771 int i
, spi
= (MAX_BPF_STACK
+ off
) / BPF_REG_SIZE
;
772 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
773 * so it's aligned access and [off, off + size) are within stack limits
776 if (value_regno
>= 0 &&
777 is_spillable_regtype(state
->regs
[value_regno
].type
)) {
779 /* register containing pointer is being spilled into stack */
780 if (size
!= BPF_REG_SIZE
) {
781 verbose("invalid size of register spill\n");
785 /* save register state */
786 state
->spilled_regs
[spi
] = state
->regs
[value_regno
];
787 state
->spilled_regs
[spi
].live
|= REG_LIVE_WRITTEN
;
789 for (i
= 0; i
< BPF_REG_SIZE
; i
++)
790 state
->stack_slot_type
[MAX_BPF_STACK
+ off
+ i
] = STACK_SPILL
;
792 /* regular write of data into stack */
793 state
->spilled_regs
[spi
] = (struct bpf_reg_state
) {};
795 for (i
= 0; i
< size
; i
++)
796 state
->stack_slot_type
[MAX_BPF_STACK
+ off
+ i
] = STACK_MISC
;
801 static void mark_stack_slot_read(const struct bpf_verifier_state
*state
, int slot
)
803 struct bpf_verifier_state
*parent
= state
->parent
;
806 /* if read wasn't screened by an earlier write ... */
807 if (state
->spilled_regs
[slot
].live
& REG_LIVE_WRITTEN
)
809 /* ... then we depend on parent's value */
810 parent
->spilled_regs
[slot
].live
|= REG_LIVE_READ
;
812 parent
= state
->parent
;
816 static int check_stack_read(struct bpf_verifier_state
*state
, int off
, int size
,
822 slot_type
= &state
->stack_slot_type
[MAX_BPF_STACK
+ off
];
824 if (slot_type
[0] == STACK_SPILL
) {
825 if (size
!= BPF_REG_SIZE
) {
826 verbose("invalid size of register spill\n");
829 for (i
= 1; i
< BPF_REG_SIZE
; i
++) {
830 if (slot_type
[i
] != STACK_SPILL
) {
831 verbose("corrupted spill memory\n");
836 spi
= (MAX_BPF_STACK
+ off
) / BPF_REG_SIZE
;
838 if (value_regno
>= 0) {
839 /* restore register state from stack */
840 state
->regs
[value_regno
] = state
->spilled_regs
[spi
];
841 mark_stack_slot_read(state
, spi
);
845 for (i
= 0; i
< size
; i
++) {
846 if (slot_type
[i
] != STACK_MISC
) {
847 verbose("invalid read from stack off %d+%d size %d\n",
852 if (value_regno
>= 0)
853 /* have read misc data from the stack */
854 mark_reg_unknown(state
->regs
, value_regno
);
859 /* check read/write into map element returned by bpf_map_lookup_elem() */
860 static int __check_map_access(struct bpf_verifier_env
*env
, u32 regno
, int off
,
863 struct bpf_map
*map
= env
->cur_state
.regs
[regno
].map_ptr
;
865 if (off
< 0 || size
<= 0 || off
+ size
> map
->value_size
) {
866 verbose("invalid access to map value, value_size=%d off=%d size=%d\n",
867 map
->value_size
, off
, size
);
873 /* check read/write into a map element with possible variable offset */
874 static int check_map_access(struct bpf_verifier_env
*env
, u32 regno
,
877 struct bpf_verifier_state
*state
= &env
->cur_state
;
878 struct bpf_reg_state
*reg
= &state
->regs
[regno
];
881 /* We may have adjusted the register to this map value, so we
882 * need to try adding each of min_value and max_value to off
883 * to make sure our theoretical access will be safe.
886 print_verifier_state(state
);
887 /* The minimum value is only important with signed
888 * comparisons where we can't assume the floor of a
889 * value is 0. If we are using signed variables for our
890 * index'es we need to make sure that whatever we use
891 * will have a set floor within our range.
893 if (reg
->smin_value
< 0) {
894 verbose("R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
898 err
= __check_map_access(env
, regno
, reg
->smin_value
+ off
, size
);
900 verbose("R%d min value is outside of the array range\n", regno
);
904 /* If we haven't set a max value then we need to bail since we can't be
905 * sure we won't do bad things.
906 * If reg->umax_value + off could overflow, treat that as unbounded too.
908 if (reg
->umax_value
>= BPF_MAX_VAR_OFF
) {
909 verbose("R%d unbounded memory access, make sure to bounds check any array access into a map\n",
913 err
= __check_map_access(env
, regno
, reg
->umax_value
+ off
, size
);
915 verbose("R%d max value is outside of the array range\n", regno
);
919 #define MAX_PACKET_OFF 0xffff
921 static bool may_access_direct_pkt_data(struct bpf_verifier_env
*env
,
922 const struct bpf_call_arg_meta
*meta
,
923 enum bpf_access_type t
)
925 switch (env
->prog
->type
) {
926 case BPF_PROG_TYPE_LWT_IN
:
927 case BPF_PROG_TYPE_LWT_OUT
:
928 /* dst_input() and dst_output() can't write for now */
932 case BPF_PROG_TYPE_SCHED_CLS
:
933 case BPF_PROG_TYPE_SCHED_ACT
:
934 case BPF_PROG_TYPE_XDP
:
935 case BPF_PROG_TYPE_LWT_XMIT
:
936 case BPF_PROG_TYPE_SK_SKB
:
938 return meta
->pkt_access
;
940 env
->seen_direct_write
= true;
947 static int __check_packet_access(struct bpf_verifier_env
*env
, u32 regno
,
950 struct bpf_reg_state
*regs
= env
->cur_state
.regs
;
951 struct bpf_reg_state
*reg
= ®s
[regno
];
953 if (off
< 0 || size
<= 0 || (u64
)off
+ size
> reg
->range
) {
954 verbose("invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
955 off
, size
, regno
, reg
->id
, reg
->off
, reg
->range
);
961 static int check_packet_access(struct bpf_verifier_env
*env
, u32 regno
, int off
,
964 struct bpf_reg_state
*regs
= env
->cur_state
.regs
;
965 struct bpf_reg_state
*reg
= ®s
[regno
];
968 /* We may have added a variable offset to the packet pointer; but any
969 * reg->range we have comes after that. We are only checking the fixed
973 /* We don't allow negative numbers, because we aren't tracking enough
974 * detail to prove they're safe.
976 if (reg
->smin_value
< 0) {
977 verbose("R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
981 err
= __check_packet_access(env
, regno
, off
, size
);
983 verbose("R%d offset is outside of the packet\n", regno
);
989 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
990 static int check_ctx_access(struct bpf_verifier_env
*env
, int insn_idx
, int off
, int size
,
991 enum bpf_access_type t
, enum bpf_reg_type
*reg_type
)
993 struct bpf_insn_access_aux info
= {
994 .reg_type
= *reg_type
,
997 /* for analyzer ctx accesses are already validated and converted */
998 if (env
->analyzer_ops
)
1001 if (env
->prog
->aux
->ops
->is_valid_access
&&
1002 env
->prog
->aux
->ops
->is_valid_access(off
, size
, t
, &info
)) {
1003 /* A non zero info.ctx_field_size indicates that this field is a
1004 * candidate for later verifier transformation to load the whole
1005 * field and then apply a mask when accessed with a narrower
1006 * access than actual ctx access size. A zero info.ctx_field_size
1007 * will only allow for whole field access and rejects any other
1008 * type of narrower access.
1010 env
->insn_aux_data
[insn_idx
].ctx_field_size
= info
.ctx_field_size
;
1011 *reg_type
= info
.reg_type
;
1013 /* remember the offset of last byte accessed in ctx */
1014 if (env
->prog
->aux
->max_ctx_offset
< off
+ size
)
1015 env
->prog
->aux
->max_ctx_offset
= off
+ size
;
1019 verbose("invalid bpf_context access off=%d size=%d\n", off
, size
);
1023 static bool __is_pointer_value(bool allow_ptr_leaks
,
1024 const struct bpf_reg_state
*reg
)
1026 if (allow_ptr_leaks
)
1029 return reg
->type
!= SCALAR_VALUE
;
1032 static bool is_pointer_value(struct bpf_verifier_env
*env
, int regno
)
1034 return __is_pointer_value(env
->allow_ptr_leaks
, &env
->cur_state
.regs
[regno
]);
1037 static int check_pkt_ptr_alignment(const struct bpf_reg_state
*reg
,
1038 int off
, int size
, bool strict
)
1040 struct tnum reg_off
;
1043 /* Byte size accesses are always allowed. */
1044 if (!strict
|| size
== 1)
1047 /* For platforms that do not have a Kconfig enabling
1048 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
1049 * NET_IP_ALIGN is universally set to '2'. And on platforms
1050 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
1051 * to this code only in strict mode where we want to emulate
1052 * the NET_IP_ALIGN==2 checking. Therefore use an
1053 * unconditional IP align value of '2'.
1057 reg_off
= tnum_add(reg
->var_off
, tnum_const(ip_align
+ reg
->off
+ off
));
1058 if (!tnum_is_aligned(reg_off
, size
)) {
1061 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1062 verbose("misaligned packet access off %d+%s+%d+%d size %d\n",
1063 ip_align
, tn_buf
, reg
->off
, off
, size
);
1070 static int check_generic_ptr_alignment(const struct bpf_reg_state
*reg
,
1071 const char *pointer_desc
,
1072 int off
, int size
, bool strict
)
1074 struct tnum reg_off
;
1076 /* Byte size accesses are always allowed. */
1077 if (!strict
|| size
== 1)
1080 reg_off
= tnum_add(reg
->var_off
, tnum_const(reg
->off
+ off
));
1081 if (!tnum_is_aligned(reg_off
, size
)) {
1084 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1085 verbose("misaligned %saccess off %s+%d+%d size %d\n",
1086 pointer_desc
, tn_buf
, reg
->off
, off
, size
);
1093 static int check_ptr_alignment(struct bpf_verifier_env
*env
,
1094 const struct bpf_reg_state
*reg
,
1097 bool strict
= env
->strict_alignment
;
1098 const char *pointer_desc
= "";
1100 switch (reg
->type
) {
1102 case PTR_TO_PACKET_META
:
1103 /* Special case, because of NET_IP_ALIGN. Given metadata sits
1104 * right in front, treat it the very same way.
1106 return check_pkt_ptr_alignment(reg
, off
, size
, strict
);
1107 case PTR_TO_MAP_VALUE
:
1108 pointer_desc
= "value ";
1111 pointer_desc
= "context ";
1114 pointer_desc
= "stack ";
1119 return check_generic_ptr_alignment(reg
, pointer_desc
, off
, size
, strict
);
1122 /* check whether memory at (regno + off) is accessible for t = (read | write)
1123 * if t==write, value_regno is a register which value is stored into memory
1124 * if t==read, value_regno is a register which will receive the value from memory
1125 * if t==write && value_regno==-1, some unknown value is stored into memory
1126 * if t==read && value_regno==-1, don't care what we read from memory
1128 static int check_mem_access(struct bpf_verifier_env
*env
, int insn_idx
, u32 regno
, int off
,
1129 int bpf_size
, enum bpf_access_type t
,
1132 struct bpf_verifier_state
*state
= &env
->cur_state
;
1133 struct bpf_reg_state
*reg
= &state
->regs
[regno
];
1136 size
= bpf_size_to_bytes(bpf_size
);
1140 /* alignment checks will add in reg->off themselves */
1141 err
= check_ptr_alignment(env
, reg
, off
, size
);
1145 /* for access checks, reg->off is just part of off */
1148 if (reg
->type
== PTR_TO_MAP_VALUE
) {
1149 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
1150 is_pointer_value(env
, value_regno
)) {
1151 verbose("R%d leaks addr into map\n", value_regno
);
1155 err
= check_map_access(env
, regno
, off
, size
);
1156 if (!err
&& t
== BPF_READ
&& value_regno
>= 0)
1157 mark_reg_unknown(state
->regs
, value_regno
);
1159 } else if (reg
->type
== PTR_TO_CTX
) {
1160 enum bpf_reg_type reg_type
= SCALAR_VALUE
;
1162 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
1163 is_pointer_value(env
, value_regno
)) {
1164 verbose("R%d leaks addr into ctx\n", value_regno
);
1167 /* ctx accesses must be at a fixed offset, so that we can
1168 * determine what type of data were returned.
1170 if (!tnum_is_const(reg
->var_off
)) {
1173 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1174 verbose("variable ctx access var_off=%s off=%d size=%d",
1178 off
+= reg
->var_off
.value
;
1179 err
= check_ctx_access(env
, insn_idx
, off
, size
, t
, ®_type
);
1180 if (!err
&& t
== BPF_READ
&& value_regno
>= 0) {
1181 /* ctx access returns either a scalar, or a
1182 * PTR_TO_PACKET[_META,_END]. In the latter
1183 * case, we know the offset is zero.
1185 if (reg_type
== SCALAR_VALUE
)
1186 mark_reg_unknown(state
->regs
, value_regno
);
1188 mark_reg_known_zero(state
->regs
, value_regno
);
1189 state
->regs
[value_regno
].id
= 0;
1190 state
->regs
[value_regno
].off
= 0;
1191 state
->regs
[value_regno
].range
= 0;
1192 state
->regs
[value_regno
].type
= reg_type
;
1195 } else if (reg
->type
== PTR_TO_STACK
) {
1196 /* stack accesses must be at a fixed offset, so that we can
1197 * determine what type of data were returned.
1198 * See check_stack_read().
1200 if (!tnum_is_const(reg
->var_off
)) {
1203 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
1204 verbose("variable stack access var_off=%s off=%d size=%d",
1208 off
+= reg
->var_off
.value
;
1209 if (off
>= 0 || off
< -MAX_BPF_STACK
) {
1210 verbose("invalid stack off=%d size=%d\n", off
, size
);
1214 if (env
->prog
->aux
->stack_depth
< -off
)
1215 env
->prog
->aux
->stack_depth
= -off
;
1217 if (t
== BPF_WRITE
) {
1218 if (!env
->allow_ptr_leaks
&&
1219 state
->stack_slot_type
[MAX_BPF_STACK
+ off
] == STACK_SPILL
&&
1220 size
!= BPF_REG_SIZE
) {
1221 verbose("attempt to corrupt spilled pointer on stack\n");
1224 err
= check_stack_write(state
, off
, size
, value_regno
);
1226 err
= check_stack_read(state
, off
, size
, value_regno
);
1228 } else if (reg_is_pkt_pointer(reg
)) {
1229 if (t
== BPF_WRITE
&& !may_access_direct_pkt_data(env
, NULL
, t
)) {
1230 verbose("cannot write into packet\n");
1233 if (t
== BPF_WRITE
&& value_regno
>= 0 &&
1234 is_pointer_value(env
, value_regno
)) {
1235 verbose("R%d leaks addr into packet\n", value_regno
);
1238 err
= check_packet_access(env
, regno
, off
, size
);
1239 if (!err
&& t
== BPF_READ
&& value_regno
>= 0)
1240 mark_reg_unknown(state
->regs
, value_regno
);
1242 verbose("R%d invalid mem access '%s'\n",
1243 regno
, reg_type_str
[reg
->type
]);
1247 if (!err
&& size
< BPF_REG_SIZE
&& value_regno
>= 0 && t
== BPF_READ
&&
1248 state
->regs
[value_regno
].type
== SCALAR_VALUE
) {
1249 /* b/h/w load zero-extends, mark upper bits as known 0 */
1250 state
->regs
[value_regno
].var_off
= tnum_cast(
1251 state
->regs
[value_regno
].var_off
, size
);
1252 __update_reg_bounds(&state
->regs
[value_regno
]);
1257 static int check_xadd(struct bpf_verifier_env
*env
, int insn_idx
, struct bpf_insn
*insn
)
1261 if ((BPF_SIZE(insn
->code
) != BPF_W
&& BPF_SIZE(insn
->code
) != BPF_DW
) ||
1263 verbose("BPF_XADD uses reserved fields\n");
1267 /* check src1 operand */
1268 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
1272 /* check src2 operand */
1273 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
1277 if (is_pointer_value(env
, insn
->src_reg
)) {
1278 verbose("R%d leaks addr into mem\n", insn
->src_reg
);
1282 /* check whether atomic_add can read the memory */
1283 err
= check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
1284 BPF_SIZE(insn
->code
), BPF_READ
, -1);
1288 /* check whether atomic_add can write into the same memory */
1289 return check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
1290 BPF_SIZE(insn
->code
), BPF_WRITE
, -1);
1293 /* Does this register contain a constant zero? */
1294 static bool register_is_null(struct bpf_reg_state reg
)
1296 return reg
.type
== SCALAR_VALUE
&& tnum_equals_const(reg
.var_off
, 0);
1299 /* when register 'regno' is passed into function that will read 'access_size'
1300 * bytes from that pointer, make sure that it's within stack boundary
1301 * and all elements of stack are initialized.
1302 * Unlike most pointer bounds-checking functions, this one doesn't take an
1303 * 'off' argument, so it has to add in reg->off itself.
1305 static int check_stack_boundary(struct bpf_verifier_env
*env
, int regno
,
1306 int access_size
, bool zero_size_allowed
,
1307 struct bpf_call_arg_meta
*meta
)
1309 struct bpf_verifier_state
*state
= &env
->cur_state
;
1310 struct bpf_reg_state
*regs
= state
->regs
;
1313 if (regs
[regno
].type
!= PTR_TO_STACK
) {
1314 /* Allow zero-byte read from NULL, regardless of pointer type */
1315 if (zero_size_allowed
&& access_size
== 0 &&
1316 register_is_null(regs
[regno
]))
1319 verbose("R%d type=%s expected=%s\n", regno
,
1320 reg_type_str
[regs
[regno
].type
],
1321 reg_type_str
[PTR_TO_STACK
]);
1325 /* Only allow fixed-offset stack reads */
1326 if (!tnum_is_const(regs
[regno
].var_off
)) {
1329 tnum_strn(tn_buf
, sizeof(tn_buf
), regs
[regno
].var_off
);
1330 verbose("invalid variable stack read R%d var_off=%s\n",
1333 off
= regs
[regno
].off
+ regs
[regno
].var_off
.value
;
1334 if (off
>= 0 || off
< -MAX_BPF_STACK
|| off
+ access_size
> 0 ||
1336 verbose("invalid stack type R%d off=%d access_size=%d\n",
1337 regno
, off
, access_size
);
1341 if (env
->prog
->aux
->stack_depth
< -off
)
1342 env
->prog
->aux
->stack_depth
= -off
;
1344 if (meta
&& meta
->raw_mode
) {
1345 meta
->access_size
= access_size
;
1346 meta
->regno
= regno
;
1350 for (i
= 0; i
< access_size
; i
++) {
1351 if (state
->stack_slot_type
[MAX_BPF_STACK
+ off
+ i
] != STACK_MISC
) {
1352 verbose("invalid indirect read from stack off %d+%d size %d\n",
1353 off
, i
, access_size
);
1360 static int check_helper_mem_access(struct bpf_verifier_env
*env
, int regno
,
1361 int access_size
, bool zero_size_allowed
,
1362 struct bpf_call_arg_meta
*meta
)
1364 struct bpf_reg_state
*regs
= env
->cur_state
.regs
, *reg
= ®s
[regno
];
1366 switch (reg
->type
) {
1368 case PTR_TO_PACKET_META
:
1369 return check_packet_access(env
, regno
, reg
->off
, access_size
);
1370 case PTR_TO_MAP_VALUE
:
1371 return check_map_access(env
, regno
, reg
->off
, access_size
);
1372 default: /* scalar_value|ptr_to_stack or invalid ptr */
1373 return check_stack_boundary(env
, regno
, access_size
,
1374 zero_size_allowed
, meta
);
1378 static int check_func_arg(struct bpf_verifier_env
*env
, u32 regno
,
1379 enum bpf_arg_type arg_type
,
1380 struct bpf_call_arg_meta
*meta
)
1382 struct bpf_reg_state
*regs
= env
->cur_state
.regs
, *reg
= ®s
[regno
];
1383 enum bpf_reg_type expected_type
, type
= reg
->type
;
1386 if (arg_type
== ARG_DONTCARE
)
1389 err
= check_reg_arg(env
, regno
, SRC_OP
);
1393 if (arg_type
== ARG_ANYTHING
) {
1394 if (is_pointer_value(env
, regno
)) {
1395 verbose("R%d leaks addr into helper function\n", regno
);
1401 if (type_is_pkt_pointer(type
) &&
1402 !may_access_direct_pkt_data(env
, meta
, BPF_READ
)) {
1403 verbose("helper access to the packet is not allowed\n");
1407 if (arg_type
== ARG_PTR_TO_MAP_KEY
||
1408 arg_type
== ARG_PTR_TO_MAP_VALUE
) {
1409 expected_type
= PTR_TO_STACK
;
1410 if (!type_is_pkt_pointer(type
) &&
1411 type
!= expected_type
)
1413 } else if (arg_type
== ARG_CONST_SIZE
||
1414 arg_type
== ARG_CONST_SIZE_OR_ZERO
) {
1415 expected_type
= SCALAR_VALUE
;
1416 if (type
!= expected_type
)
1418 } else if (arg_type
== ARG_CONST_MAP_PTR
) {
1419 expected_type
= CONST_PTR_TO_MAP
;
1420 if (type
!= expected_type
)
1422 } else if (arg_type
== ARG_PTR_TO_CTX
) {
1423 expected_type
= PTR_TO_CTX
;
1424 if (type
!= expected_type
)
1426 } else if (arg_type
== ARG_PTR_TO_MEM
||
1427 arg_type
== ARG_PTR_TO_UNINIT_MEM
) {
1428 expected_type
= PTR_TO_STACK
;
1429 /* One exception here. In case function allows for NULL to be
1430 * passed in as argument, it's a SCALAR_VALUE type. Final test
1431 * happens during stack boundary checking.
1433 if (register_is_null(*reg
))
1434 /* final test in check_stack_boundary() */;
1435 else if (!type_is_pkt_pointer(type
) &&
1436 type
!= PTR_TO_MAP_VALUE
&&
1437 type
!= expected_type
)
1439 meta
->raw_mode
= arg_type
== ARG_PTR_TO_UNINIT_MEM
;
1441 verbose("unsupported arg_type %d\n", arg_type
);
1445 if (arg_type
== ARG_CONST_MAP_PTR
) {
1446 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
1447 meta
->map_ptr
= reg
->map_ptr
;
1448 } else if (arg_type
== ARG_PTR_TO_MAP_KEY
) {
1449 /* bpf_map_xxx(..., map_ptr, ..., key) call:
1450 * check that [key, key + map->key_size) are within
1451 * stack limits and initialized
1453 if (!meta
->map_ptr
) {
1454 /* in function declaration map_ptr must come before
1455 * map_key, so that it's verified and known before
1456 * we have to check map_key here. Otherwise it means
1457 * that kernel subsystem misconfigured verifier
1459 verbose("invalid map_ptr to access map->key\n");
1462 if (type_is_pkt_pointer(type
))
1463 err
= check_packet_access(env
, regno
, reg
->off
,
1464 meta
->map_ptr
->key_size
);
1466 err
= check_stack_boundary(env
, regno
,
1467 meta
->map_ptr
->key_size
,
1469 } else if (arg_type
== ARG_PTR_TO_MAP_VALUE
) {
1470 /* bpf_map_xxx(..., map_ptr, ..., value) call:
1471 * check [value, value + map->value_size) validity
1473 if (!meta
->map_ptr
) {
1474 /* kernel subsystem misconfigured verifier */
1475 verbose("invalid map_ptr to access map->value\n");
1478 if (type_is_pkt_pointer(type
))
1479 err
= check_packet_access(env
, regno
, reg
->off
,
1480 meta
->map_ptr
->value_size
);
1482 err
= check_stack_boundary(env
, regno
,
1483 meta
->map_ptr
->value_size
,
1485 } else if (arg_type
== ARG_CONST_SIZE
||
1486 arg_type
== ARG_CONST_SIZE_OR_ZERO
) {
1487 bool zero_size_allowed
= (arg_type
== ARG_CONST_SIZE_OR_ZERO
);
1489 /* bpf_xxx(..., buf, len) call will access 'len' bytes
1490 * from stack pointer 'buf'. Check it
1491 * note: regno == len, regno - 1 == buf
1494 /* kernel subsystem misconfigured verifier */
1495 verbose("ARG_CONST_SIZE cannot be first argument\n");
1499 /* The register is SCALAR_VALUE; the access check
1500 * happens using its boundaries.
1503 if (!tnum_is_const(reg
->var_off
))
1504 /* For unprivileged variable accesses, disable raw
1505 * mode so that the program is required to
1506 * initialize all the memory that the helper could
1507 * just partially fill up.
1511 if (reg
->smin_value
< 0) {
1512 verbose("R%d min value is negative, either use unsigned or 'var &= const'\n",
1517 if (reg
->umin_value
== 0) {
1518 err
= check_helper_mem_access(env
, regno
- 1, 0,
1525 if (reg
->umax_value
>= BPF_MAX_VAR_SIZ
) {
1526 verbose("R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
1530 err
= check_helper_mem_access(env
, regno
- 1,
1532 zero_size_allowed
, meta
);
1537 verbose("R%d type=%s expected=%s\n", regno
,
1538 reg_type_str
[type
], reg_type_str
[expected_type
]);
1542 static int check_map_func_compatibility(struct bpf_map
*map
, int func_id
)
1547 /* We need a two way check, first is from map perspective ... */
1548 switch (map
->map_type
) {
1549 case BPF_MAP_TYPE_PROG_ARRAY
:
1550 if (func_id
!= BPF_FUNC_tail_call
)
1553 case BPF_MAP_TYPE_PERF_EVENT_ARRAY
:
1554 if (func_id
!= BPF_FUNC_perf_event_read
&&
1555 func_id
!= BPF_FUNC_perf_event_output
)
1558 case BPF_MAP_TYPE_STACK_TRACE
:
1559 if (func_id
!= BPF_FUNC_get_stackid
)
1562 case BPF_MAP_TYPE_CGROUP_ARRAY
:
1563 if (func_id
!= BPF_FUNC_skb_under_cgroup
&&
1564 func_id
!= BPF_FUNC_current_task_under_cgroup
)
1567 /* devmap returns a pointer to a live net_device ifindex that we cannot
1568 * allow to be modified from bpf side. So do not allow lookup elements
1571 case BPF_MAP_TYPE_DEVMAP
:
1572 if (func_id
!= BPF_FUNC_redirect_map
)
1575 case BPF_MAP_TYPE_ARRAY_OF_MAPS
:
1576 case BPF_MAP_TYPE_HASH_OF_MAPS
:
1577 if (func_id
!= BPF_FUNC_map_lookup_elem
)
1580 case BPF_MAP_TYPE_SOCKMAP
:
1581 if (func_id
!= BPF_FUNC_sk_redirect_map
&&
1582 func_id
!= BPF_FUNC_sock_map_update
&&
1583 func_id
!= BPF_FUNC_map_delete_elem
)
1590 /* ... and second from the function itself. */
1592 case BPF_FUNC_tail_call
:
1593 if (map
->map_type
!= BPF_MAP_TYPE_PROG_ARRAY
)
1596 case BPF_FUNC_perf_event_read
:
1597 case BPF_FUNC_perf_event_output
:
1598 if (map
->map_type
!= BPF_MAP_TYPE_PERF_EVENT_ARRAY
)
1601 case BPF_FUNC_get_stackid
:
1602 if (map
->map_type
!= BPF_MAP_TYPE_STACK_TRACE
)
1605 case BPF_FUNC_current_task_under_cgroup
:
1606 case BPF_FUNC_skb_under_cgroup
:
1607 if (map
->map_type
!= BPF_MAP_TYPE_CGROUP_ARRAY
)
1610 case BPF_FUNC_redirect_map
:
1611 if (map
->map_type
!= BPF_MAP_TYPE_DEVMAP
)
1614 case BPF_FUNC_sk_redirect_map
:
1615 if (map
->map_type
!= BPF_MAP_TYPE_SOCKMAP
)
1618 case BPF_FUNC_sock_map_update
:
1619 if (map
->map_type
!= BPF_MAP_TYPE_SOCKMAP
)
1628 verbose("cannot pass map_type %d into func %s#%d\n",
1629 map
->map_type
, func_id_name(func_id
), func_id
);
1633 static int check_raw_mode(const struct bpf_func_proto
*fn
)
1637 if (fn
->arg1_type
== ARG_PTR_TO_UNINIT_MEM
)
1639 if (fn
->arg2_type
== ARG_PTR_TO_UNINIT_MEM
)
1641 if (fn
->arg3_type
== ARG_PTR_TO_UNINIT_MEM
)
1643 if (fn
->arg4_type
== ARG_PTR_TO_UNINIT_MEM
)
1645 if (fn
->arg5_type
== ARG_PTR_TO_UNINIT_MEM
)
1648 return count
> 1 ? -EINVAL
: 0;
1651 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
1652 * are now invalid, so turn them into unknown SCALAR_VALUE.
1654 static void clear_all_pkt_pointers(struct bpf_verifier_env
*env
)
1656 struct bpf_verifier_state
*state
= &env
->cur_state
;
1657 struct bpf_reg_state
*regs
= state
->regs
, *reg
;
1660 for (i
= 0; i
< MAX_BPF_REG
; i
++)
1661 if (reg_is_pkt_pointer_any(®s
[i
]))
1662 mark_reg_unknown(regs
, i
);
1664 for (i
= 0; i
< MAX_BPF_STACK
; i
+= BPF_REG_SIZE
) {
1665 if (state
->stack_slot_type
[i
] != STACK_SPILL
)
1667 reg
= &state
->spilled_regs
[i
/ BPF_REG_SIZE
];
1668 if (reg_is_pkt_pointer_any(reg
))
1669 __mark_reg_unknown(reg
);
1673 static int check_call(struct bpf_verifier_env
*env
, int func_id
, int insn_idx
)
1675 struct bpf_verifier_state
*state
= &env
->cur_state
;
1676 const struct bpf_func_proto
*fn
= NULL
;
1677 struct bpf_reg_state
*regs
= state
->regs
;
1678 struct bpf_call_arg_meta meta
;
1682 /* find function prototype */
1683 if (func_id
< 0 || func_id
>= __BPF_FUNC_MAX_ID
) {
1684 verbose("invalid func %s#%d\n", func_id_name(func_id
), func_id
);
1688 if (env
->prog
->aux
->ops
->get_func_proto
)
1689 fn
= env
->prog
->aux
->ops
->get_func_proto(func_id
);
1692 verbose("unknown func %s#%d\n", func_id_name(func_id
), func_id
);
1696 /* eBPF programs must be GPL compatible to use GPL-ed functions */
1697 if (!env
->prog
->gpl_compatible
&& fn
->gpl_only
) {
1698 verbose("cannot call GPL only function from proprietary program\n");
1702 changes_data
= bpf_helper_changes_pkt_data(fn
->func
);
1704 memset(&meta
, 0, sizeof(meta
));
1705 meta
.pkt_access
= fn
->pkt_access
;
1707 /* We only support one arg being in raw mode at the moment, which
1708 * is sufficient for the helper functions we have right now.
1710 err
= check_raw_mode(fn
);
1712 verbose("kernel subsystem misconfigured func %s#%d\n",
1713 func_id_name(func_id
), func_id
);
1718 err
= check_func_arg(env
, BPF_REG_1
, fn
->arg1_type
, &meta
);
1721 err
= check_func_arg(env
, BPF_REG_2
, fn
->arg2_type
, &meta
);
1724 err
= check_func_arg(env
, BPF_REG_3
, fn
->arg3_type
, &meta
);
1727 err
= check_func_arg(env
, BPF_REG_4
, fn
->arg4_type
, &meta
);
1730 err
= check_func_arg(env
, BPF_REG_5
, fn
->arg5_type
, &meta
);
1734 /* Mark slots with STACK_MISC in case of raw mode, stack offset
1735 * is inferred from register state.
1737 for (i
= 0; i
< meta
.access_size
; i
++) {
1738 err
= check_mem_access(env
, insn_idx
, meta
.regno
, i
, BPF_B
, BPF_WRITE
, -1);
1743 /* reset caller saved regs */
1744 for (i
= 0; i
< CALLER_SAVED_REGS
; i
++) {
1745 mark_reg_not_init(regs
, caller_saved
[i
]);
1746 check_reg_arg(env
, caller_saved
[i
], DST_OP_NO_MARK
);
1749 /* update return register (already marked as written above) */
1750 if (fn
->ret_type
== RET_INTEGER
) {
1751 /* sets type to SCALAR_VALUE */
1752 mark_reg_unknown(regs
, BPF_REG_0
);
1753 } else if (fn
->ret_type
== RET_VOID
) {
1754 regs
[BPF_REG_0
].type
= NOT_INIT
;
1755 } else if (fn
->ret_type
== RET_PTR_TO_MAP_VALUE_OR_NULL
) {
1756 struct bpf_insn_aux_data
*insn_aux
;
1758 regs
[BPF_REG_0
].type
= PTR_TO_MAP_VALUE_OR_NULL
;
1759 /* There is no offset yet applied, variable or fixed */
1760 mark_reg_known_zero(regs
, BPF_REG_0
);
1761 regs
[BPF_REG_0
].off
= 0;
1762 /* remember map_ptr, so that check_map_access()
1763 * can check 'value_size' boundary of memory access
1764 * to map element returned from bpf_map_lookup_elem()
1766 if (meta
.map_ptr
== NULL
) {
1767 verbose("kernel subsystem misconfigured verifier\n");
1770 regs
[BPF_REG_0
].map_ptr
= meta
.map_ptr
;
1771 regs
[BPF_REG_0
].id
= ++env
->id_gen
;
1772 insn_aux
= &env
->insn_aux_data
[insn_idx
];
1773 if (!insn_aux
->map_ptr
)
1774 insn_aux
->map_ptr
= meta
.map_ptr
;
1775 else if (insn_aux
->map_ptr
!= meta
.map_ptr
)
1776 insn_aux
->map_ptr
= BPF_MAP_PTR_POISON
;
1778 verbose("unknown return type %d of func %s#%d\n",
1779 fn
->ret_type
, func_id_name(func_id
), func_id
);
1783 err
= check_map_func_compatibility(meta
.map_ptr
, func_id
);
1788 clear_all_pkt_pointers(env
);
1792 static void coerce_reg_to_32(struct bpf_reg_state
*reg
)
1794 /* clear high 32 bits */
1795 reg
->var_off
= tnum_cast(reg
->var_off
, 4);
1797 __update_reg_bounds(reg
);
1800 static bool signed_add_overflows(s64 a
, s64 b
)
1802 /* Do the add in u64, where overflow is well-defined */
1803 s64 res
= (s64
)((u64
)a
+ (u64
)b
);
1810 static bool signed_sub_overflows(s64 a
, s64 b
)
1812 /* Do the sub in u64, where overflow is well-defined */
1813 s64 res
= (s64
)((u64
)a
- (u64
)b
);
1820 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
1821 * Caller should also handle BPF_MOV case separately.
1822 * If we return -EACCES, caller may want to try again treating pointer as a
1823 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
1825 static int adjust_ptr_min_max_vals(struct bpf_verifier_env
*env
,
1826 struct bpf_insn
*insn
,
1827 const struct bpf_reg_state
*ptr_reg
,
1828 const struct bpf_reg_state
*off_reg
)
1830 struct bpf_reg_state
*regs
= env
->cur_state
.regs
, *dst_reg
;
1831 bool known
= tnum_is_const(off_reg
->var_off
);
1832 s64 smin_val
= off_reg
->smin_value
, smax_val
= off_reg
->smax_value
,
1833 smin_ptr
= ptr_reg
->smin_value
, smax_ptr
= ptr_reg
->smax_value
;
1834 u64 umin_val
= off_reg
->umin_value
, umax_val
= off_reg
->umax_value
,
1835 umin_ptr
= ptr_reg
->umin_value
, umax_ptr
= ptr_reg
->umax_value
;
1836 u8 opcode
= BPF_OP(insn
->code
);
1837 u32 dst
= insn
->dst_reg
;
1839 dst_reg
= ®s
[dst
];
1841 if (WARN_ON_ONCE(known
&& (smin_val
!= smax_val
))) {
1842 print_verifier_state(&env
->cur_state
);
1843 verbose("verifier internal error: known but bad sbounds\n");
1846 if (WARN_ON_ONCE(known
&& (umin_val
!= umax_val
))) {
1847 print_verifier_state(&env
->cur_state
);
1848 verbose("verifier internal error: known but bad ubounds\n");
1852 if (BPF_CLASS(insn
->code
) != BPF_ALU64
) {
1853 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
1854 if (!env
->allow_ptr_leaks
)
1855 verbose("R%d 32-bit pointer arithmetic prohibited\n",
1860 if (ptr_reg
->type
== PTR_TO_MAP_VALUE_OR_NULL
) {
1861 if (!env
->allow_ptr_leaks
)
1862 verbose("R%d pointer arithmetic on PTR_TO_MAP_VALUE_OR_NULL prohibited, null-check it first\n",
1866 if (ptr_reg
->type
== CONST_PTR_TO_MAP
) {
1867 if (!env
->allow_ptr_leaks
)
1868 verbose("R%d pointer arithmetic on CONST_PTR_TO_MAP prohibited\n",
1872 if (ptr_reg
->type
== PTR_TO_PACKET_END
) {
1873 if (!env
->allow_ptr_leaks
)
1874 verbose("R%d pointer arithmetic on PTR_TO_PACKET_END prohibited\n",
1879 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
1880 * The id may be overwritten later if we create a new variable offset.
1882 dst_reg
->type
= ptr_reg
->type
;
1883 dst_reg
->id
= ptr_reg
->id
;
1887 /* We can take a fixed offset as long as it doesn't overflow
1888 * the s32 'off' field
1890 if (known
&& (ptr_reg
->off
+ smin_val
==
1891 (s64
)(s32
)(ptr_reg
->off
+ smin_val
))) {
1892 /* pointer += K. Accumulate it into fixed offset */
1893 dst_reg
->smin_value
= smin_ptr
;
1894 dst_reg
->smax_value
= smax_ptr
;
1895 dst_reg
->umin_value
= umin_ptr
;
1896 dst_reg
->umax_value
= umax_ptr
;
1897 dst_reg
->var_off
= ptr_reg
->var_off
;
1898 dst_reg
->off
= ptr_reg
->off
+ smin_val
;
1899 dst_reg
->range
= ptr_reg
->range
;
1902 /* A new variable offset is created. Note that off_reg->off
1903 * == 0, since it's a scalar.
1904 * dst_reg gets the pointer type and since some positive
1905 * integer value was added to the pointer, give it a new 'id'
1906 * if it's a PTR_TO_PACKET.
1907 * this creates a new 'base' pointer, off_reg (variable) gets
1908 * added into the variable offset, and we copy the fixed offset
1911 if (signed_add_overflows(smin_ptr
, smin_val
) ||
1912 signed_add_overflows(smax_ptr
, smax_val
)) {
1913 dst_reg
->smin_value
= S64_MIN
;
1914 dst_reg
->smax_value
= S64_MAX
;
1916 dst_reg
->smin_value
= smin_ptr
+ smin_val
;
1917 dst_reg
->smax_value
= smax_ptr
+ smax_val
;
1919 if (umin_ptr
+ umin_val
< umin_ptr
||
1920 umax_ptr
+ umax_val
< umax_ptr
) {
1921 dst_reg
->umin_value
= 0;
1922 dst_reg
->umax_value
= U64_MAX
;
1924 dst_reg
->umin_value
= umin_ptr
+ umin_val
;
1925 dst_reg
->umax_value
= umax_ptr
+ umax_val
;
1927 dst_reg
->var_off
= tnum_add(ptr_reg
->var_off
, off_reg
->var_off
);
1928 dst_reg
->off
= ptr_reg
->off
;
1929 if (reg_is_pkt_pointer(ptr_reg
)) {
1930 dst_reg
->id
= ++env
->id_gen
;
1931 /* something was added to pkt_ptr, set range to zero */
1936 if (dst_reg
== off_reg
) {
1937 /* scalar -= pointer. Creates an unknown scalar */
1938 if (!env
->allow_ptr_leaks
)
1939 verbose("R%d tried to subtract pointer from scalar\n",
1943 /* We don't allow subtraction from FP, because (according to
1944 * test_verifier.c test "invalid fp arithmetic", JITs might not
1945 * be able to deal with it.
1947 if (ptr_reg
->type
== PTR_TO_STACK
) {
1948 if (!env
->allow_ptr_leaks
)
1949 verbose("R%d subtraction from stack pointer prohibited\n",
1953 if (known
&& (ptr_reg
->off
- smin_val
==
1954 (s64
)(s32
)(ptr_reg
->off
- smin_val
))) {
1955 /* pointer -= K. Subtract it from fixed offset */
1956 dst_reg
->smin_value
= smin_ptr
;
1957 dst_reg
->smax_value
= smax_ptr
;
1958 dst_reg
->umin_value
= umin_ptr
;
1959 dst_reg
->umax_value
= umax_ptr
;
1960 dst_reg
->var_off
= ptr_reg
->var_off
;
1961 dst_reg
->id
= ptr_reg
->id
;
1962 dst_reg
->off
= ptr_reg
->off
- smin_val
;
1963 dst_reg
->range
= ptr_reg
->range
;
1966 /* A new variable offset is created. If the subtrahend is known
1967 * nonnegative, then any reg->range we had before is still good.
1969 if (signed_sub_overflows(smin_ptr
, smax_val
) ||
1970 signed_sub_overflows(smax_ptr
, smin_val
)) {
1971 /* Overflow possible, we know nothing */
1972 dst_reg
->smin_value
= S64_MIN
;
1973 dst_reg
->smax_value
= S64_MAX
;
1975 dst_reg
->smin_value
= smin_ptr
- smax_val
;
1976 dst_reg
->smax_value
= smax_ptr
- smin_val
;
1978 if (umin_ptr
< umax_val
) {
1979 /* Overflow possible, we know nothing */
1980 dst_reg
->umin_value
= 0;
1981 dst_reg
->umax_value
= U64_MAX
;
1983 /* Cannot overflow (as long as bounds are consistent) */
1984 dst_reg
->umin_value
= umin_ptr
- umax_val
;
1985 dst_reg
->umax_value
= umax_ptr
- umin_val
;
1987 dst_reg
->var_off
= tnum_sub(ptr_reg
->var_off
, off_reg
->var_off
);
1988 dst_reg
->off
= ptr_reg
->off
;
1989 if (reg_is_pkt_pointer(ptr_reg
)) {
1990 dst_reg
->id
= ++env
->id_gen
;
1991 /* something was added to pkt_ptr, set range to zero */
1999 /* bitwise ops on pointers are troublesome, prohibit for now.
2000 * (However, in principle we could allow some cases, e.g.
2001 * ptr &= ~3 which would reduce min_value by 3.)
2003 if (!env
->allow_ptr_leaks
)
2004 verbose("R%d bitwise operator %s on pointer prohibited\n",
2005 dst
, bpf_alu_string
[opcode
>> 4]);
2008 /* other operators (e.g. MUL,LSH) produce non-pointer results */
2009 if (!env
->allow_ptr_leaks
)
2010 verbose("R%d pointer arithmetic with %s operator prohibited\n",
2011 dst
, bpf_alu_string
[opcode
>> 4]);
2015 __update_reg_bounds(dst_reg
);
2016 __reg_deduce_bounds(dst_reg
);
2017 __reg_bound_offset(dst_reg
);
2021 static int adjust_scalar_min_max_vals(struct bpf_verifier_env
*env
,
2022 struct bpf_insn
*insn
,
2023 struct bpf_reg_state
*dst_reg
,
2024 struct bpf_reg_state src_reg
)
2026 struct bpf_reg_state
*regs
= env
->cur_state
.regs
;
2027 u8 opcode
= BPF_OP(insn
->code
);
2028 bool src_known
, dst_known
;
2029 s64 smin_val
, smax_val
;
2030 u64 umin_val
, umax_val
;
2032 if (BPF_CLASS(insn
->code
) != BPF_ALU64
) {
2033 /* 32-bit ALU ops are (32,32)->64 */
2034 coerce_reg_to_32(dst_reg
);
2035 coerce_reg_to_32(&src_reg
);
2037 smin_val
= src_reg
.smin_value
;
2038 smax_val
= src_reg
.smax_value
;
2039 umin_val
= src_reg
.umin_value
;
2040 umax_val
= src_reg
.umax_value
;
2041 src_known
= tnum_is_const(src_reg
.var_off
);
2042 dst_known
= tnum_is_const(dst_reg
->var_off
);
2046 if (signed_add_overflows(dst_reg
->smin_value
, smin_val
) ||
2047 signed_add_overflows(dst_reg
->smax_value
, smax_val
)) {
2048 dst_reg
->smin_value
= S64_MIN
;
2049 dst_reg
->smax_value
= S64_MAX
;
2051 dst_reg
->smin_value
+= smin_val
;
2052 dst_reg
->smax_value
+= smax_val
;
2054 if (dst_reg
->umin_value
+ umin_val
< umin_val
||
2055 dst_reg
->umax_value
+ umax_val
< umax_val
) {
2056 dst_reg
->umin_value
= 0;
2057 dst_reg
->umax_value
= U64_MAX
;
2059 dst_reg
->umin_value
+= umin_val
;
2060 dst_reg
->umax_value
+= umax_val
;
2062 dst_reg
->var_off
= tnum_add(dst_reg
->var_off
, src_reg
.var_off
);
2065 if (signed_sub_overflows(dst_reg
->smin_value
, smax_val
) ||
2066 signed_sub_overflows(dst_reg
->smax_value
, smin_val
)) {
2067 /* Overflow possible, we know nothing */
2068 dst_reg
->smin_value
= S64_MIN
;
2069 dst_reg
->smax_value
= S64_MAX
;
2071 dst_reg
->smin_value
-= smax_val
;
2072 dst_reg
->smax_value
-= smin_val
;
2074 if (dst_reg
->umin_value
< umax_val
) {
2075 /* Overflow possible, we know nothing */
2076 dst_reg
->umin_value
= 0;
2077 dst_reg
->umax_value
= U64_MAX
;
2079 /* Cannot overflow (as long as bounds are consistent) */
2080 dst_reg
->umin_value
-= umax_val
;
2081 dst_reg
->umax_value
-= umin_val
;
2083 dst_reg
->var_off
= tnum_sub(dst_reg
->var_off
, src_reg
.var_off
);
2086 dst_reg
->var_off
= tnum_mul(dst_reg
->var_off
, src_reg
.var_off
);
2087 if (smin_val
< 0 || dst_reg
->smin_value
< 0) {
2088 /* Ain't nobody got time to multiply that sign */
2089 __mark_reg_unbounded(dst_reg
);
2090 __update_reg_bounds(dst_reg
);
2093 /* Both values are positive, so we can work with unsigned and
2094 * copy the result to signed (unless it exceeds S64_MAX).
2096 if (umax_val
> U32_MAX
|| dst_reg
->umax_value
> U32_MAX
) {
2097 /* Potential overflow, we know nothing */
2098 __mark_reg_unbounded(dst_reg
);
2099 /* (except what we can learn from the var_off) */
2100 __update_reg_bounds(dst_reg
);
2103 dst_reg
->umin_value
*= umin_val
;
2104 dst_reg
->umax_value
*= umax_val
;
2105 if (dst_reg
->umax_value
> S64_MAX
) {
2106 /* Overflow possible, we know nothing */
2107 dst_reg
->smin_value
= S64_MIN
;
2108 dst_reg
->smax_value
= S64_MAX
;
2110 dst_reg
->smin_value
= dst_reg
->umin_value
;
2111 dst_reg
->smax_value
= dst_reg
->umax_value
;
2115 if (src_known
&& dst_known
) {
2116 __mark_reg_known(dst_reg
, dst_reg
->var_off
.value
&
2117 src_reg
.var_off
.value
);
2120 /* We get our minimum from the var_off, since that's inherently
2121 * bitwise. Our maximum is the minimum of the operands' maxima.
2123 dst_reg
->var_off
= tnum_and(dst_reg
->var_off
, src_reg
.var_off
);
2124 dst_reg
->umin_value
= dst_reg
->var_off
.value
;
2125 dst_reg
->umax_value
= min(dst_reg
->umax_value
, umax_val
);
2126 if (dst_reg
->smin_value
< 0 || smin_val
< 0) {
2127 /* Lose signed bounds when ANDing negative numbers,
2128 * ain't nobody got time for that.
2130 dst_reg
->smin_value
= S64_MIN
;
2131 dst_reg
->smax_value
= S64_MAX
;
2133 /* ANDing two positives gives a positive, so safe to
2134 * cast result into s64.
2136 dst_reg
->smin_value
= dst_reg
->umin_value
;
2137 dst_reg
->smax_value
= dst_reg
->umax_value
;
2139 /* We may learn something more from the var_off */
2140 __update_reg_bounds(dst_reg
);
2143 if (src_known
&& dst_known
) {
2144 __mark_reg_known(dst_reg
, dst_reg
->var_off
.value
|
2145 src_reg
.var_off
.value
);
2148 /* We get our maximum from the var_off, and our minimum is the
2149 * maximum of the operands' minima
2151 dst_reg
->var_off
= tnum_or(dst_reg
->var_off
, src_reg
.var_off
);
2152 dst_reg
->umin_value
= max(dst_reg
->umin_value
, umin_val
);
2153 dst_reg
->umax_value
= dst_reg
->var_off
.value
|
2154 dst_reg
->var_off
.mask
;
2155 if (dst_reg
->smin_value
< 0 || smin_val
< 0) {
2156 /* Lose signed bounds when ORing negative numbers,
2157 * ain't nobody got time for that.
2159 dst_reg
->smin_value
= S64_MIN
;
2160 dst_reg
->smax_value
= S64_MAX
;
2162 /* ORing two positives gives a positive, so safe to
2163 * cast result into s64.
2165 dst_reg
->smin_value
= dst_reg
->umin_value
;
2166 dst_reg
->smax_value
= dst_reg
->umax_value
;
2168 /* We may learn something more from the var_off */
2169 __update_reg_bounds(dst_reg
);
2172 if (umax_val
> 63) {
2173 /* Shifts greater than 63 are undefined. This includes
2174 * shifts by a negative number.
2176 mark_reg_unknown(regs
, insn
->dst_reg
);
2179 /* We lose all sign bit information (except what we can pick
2182 dst_reg
->smin_value
= S64_MIN
;
2183 dst_reg
->smax_value
= S64_MAX
;
2184 /* If we might shift our top bit out, then we know nothing */
2185 if (dst_reg
->umax_value
> 1ULL << (63 - umax_val
)) {
2186 dst_reg
->umin_value
= 0;
2187 dst_reg
->umax_value
= U64_MAX
;
2189 dst_reg
->umin_value
<<= umin_val
;
2190 dst_reg
->umax_value
<<= umax_val
;
2193 dst_reg
->var_off
= tnum_lshift(dst_reg
->var_off
, umin_val
);
2195 dst_reg
->var_off
= tnum_lshift(tnum_unknown
, umin_val
);
2196 /* We may learn something more from the var_off */
2197 __update_reg_bounds(dst_reg
);
2200 if (umax_val
> 63) {
2201 /* Shifts greater than 63 are undefined. This includes
2202 * shifts by a negative number.
2204 mark_reg_unknown(regs
, insn
->dst_reg
);
2207 /* BPF_RSH is an unsigned shift, so make the appropriate casts */
2208 if (dst_reg
->smin_value
< 0) {
2210 /* Sign bit will be cleared */
2211 dst_reg
->smin_value
= 0;
2213 /* Lost sign bit information */
2214 dst_reg
->smin_value
= S64_MIN
;
2215 dst_reg
->smax_value
= S64_MAX
;
2218 dst_reg
->smin_value
=
2219 (u64
)(dst_reg
->smin_value
) >> umax_val
;
2222 dst_reg
->var_off
= tnum_rshift(dst_reg
->var_off
,
2225 dst_reg
->var_off
= tnum_rshift(tnum_unknown
, umin_val
);
2226 dst_reg
->umin_value
>>= umax_val
;
2227 dst_reg
->umax_value
>>= umin_val
;
2228 /* We may learn something more from the var_off */
2229 __update_reg_bounds(dst_reg
);
2232 mark_reg_unknown(regs
, insn
->dst_reg
);
2236 __reg_deduce_bounds(dst_reg
);
2237 __reg_bound_offset(dst_reg
);
2241 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
2244 static int adjust_reg_min_max_vals(struct bpf_verifier_env
*env
,
2245 struct bpf_insn
*insn
)
2247 struct bpf_reg_state
*regs
= env
->cur_state
.regs
, *dst_reg
, *src_reg
;
2248 struct bpf_reg_state
*ptr_reg
= NULL
, off_reg
= {0};
2249 u8 opcode
= BPF_OP(insn
->code
);
2252 dst_reg
= ®s
[insn
->dst_reg
];
2254 if (dst_reg
->type
!= SCALAR_VALUE
)
2256 if (BPF_SRC(insn
->code
) == BPF_X
) {
2257 src_reg
= ®s
[insn
->src_reg
];
2258 if (src_reg
->type
!= SCALAR_VALUE
) {
2259 if (dst_reg
->type
!= SCALAR_VALUE
) {
2260 /* Combining two pointers by any ALU op yields
2261 * an arbitrary scalar.
2263 if (!env
->allow_ptr_leaks
) {
2264 verbose("R%d pointer %s pointer prohibited\n",
2266 bpf_alu_string
[opcode
>> 4]);
2269 mark_reg_unknown(regs
, insn
->dst_reg
);
2272 /* scalar += pointer
2273 * This is legal, but we have to reverse our
2274 * src/dest handling in computing the range
2276 rc
= adjust_ptr_min_max_vals(env
, insn
,
2278 if (rc
== -EACCES
&& env
->allow_ptr_leaks
) {
2279 /* scalar += unknown scalar */
2280 __mark_reg_unknown(&off_reg
);
2281 return adjust_scalar_min_max_vals(
2287 } else if (ptr_reg
) {
2288 /* pointer += scalar */
2289 rc
= adjust_ptr_min_max_vals(env
, insn
,
2291 if (rc
== -EACCES
&& env
->allow_ptr_leaks
) {
2292 /* unknown scalar += scalar */
2293 __mark_reg_unknown(dst_reg
);
2294 return adjust_scalar_min_max_vals(
2295 env
, insn
, dst_reg
, *src_reg
);
2300 /* Pretend the src is a reg with a known value, since we only
2301 * need to be able to read from this state.
2303 off_reg
.type
= SCALAR_VALUE
;
2304 __mark_reg_known(&off_reg
, insn
->imm
);
2306 if (ptr_reg
) { /* pointer += K */
2307 rc
= adjust_ptr_min_max_vals(env
, insn
,
2309 if (rc
== -EACCES
&& env
->allow_ptr_leaks
) {
2310 /* unknown scalar += K */
2311 __mark_reg_unknown(dst_reg
);
2312 return adjust_scalar_min_max_vals(
2313 env
, insn
, dst_reg
, off_reg
);
2319 /* Got here implies adding two SCALAR_VALUEs */
2320 if (WARN_ON_ONCE(ptr_reg
)) {
2321 print_verifier_state(&env
->cur_state
);
2322 verbose("verifier internal error: unexpected ptr_reg\n");
2325 if (WARN_ON(!src_reg
)) {
2326 print_verifier_state(&env
->cur_state
);
2327 verbose("verifier internal error: no src_reg\n");
2330 return adjust_scalar_min_max_vals(env
, insn
, dst_reg
, *src_reg
);
2333 /* check validity of 32-bit and 64-bit arithmetic operations */
2334 static int check_alu_op(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
2336 struct bpf_reg_state
*regs
= env
->cur_state
.regs
;
2337 u8 opcode
= BPF_OP(insn
->code
);
2340 if (opcode
== BPF_END
|| opcode
== BPF_NEG
) {
2341 if (opcode
== BPF_NEG
) {
2342 if (BPF_SRC(insn
->code
) != 0 ||
2343 insn
->src_reg
!= BPF_REG_0
||
2344 insn
->off
!= 0 || insn
->imm
!= 0) {
2345 verbose("BPF_NEG uses reserved fields\n");
2349 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0 ||
2350 (insn
->imm
!= 16 && insn
->imm
!= 32 && insn
->imm
!= 64) ||
2351 BPF_CLASS(insn
->code
) == BPF_ALU64
) {
2352 verbose("BPF_END uses reserved fields\n");
2357 /* check src operand */
2358 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
2362 if (is_pointer_value(env
, insn
->dst_reg
)) {
2363 verbose("R%d pointer arithmetic prohibited\n",
2368 /* check dest operand */
2369 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP
);
2373 } else if (opcode
== BPF_MOV
) {
2375 if (BPF_SRC(insn
->code
) == BPF_X
) {
2376 if (insn
->imm
!= 0 || insn
->off
!= 0) {
2377 verbose("BPF_MOV uses reserved fields\n");
2381 /* check src operand */
2382 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
2386 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0) {
2387 verbose("BPF_MOV uses reserved fields\n");
2392 /* check dest operand */
2393 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP
);
2397 if (BPF_SRC(insn
->code
) == BPF_X
) {
2398 if (BPF_CLASS(insn
->code
) == BPF_ALU64
) {
2400 * copy register state to dest reg
2402 regs
[insn
->dst_reg
] = regs
[insn
->src_reg
];
2405 if (is_pointer_value(env
, insn
->src_reg
)) {
2406 verbose("R%d partial copy of pointer\n",
2410 mark_reg_unknown(regs
, insn
->dst_reg
);
2411 /* high 32 bits are known zero. */
2412 regs
[insn
->dst_reg
].var_off
= tnum_cast(
2413 regs
[insn
->dst_reg
].var_off
, 4);
2414 __update_reg_bounds(®s
[insn
->dst_reg
]);
2418 * remember the value we stored into this reg
2420 regs
[insn
->dst_reg
].type
= SCALAR_VALUE
;
2421 __mark_reg_known(regs
+ insn
->dst_reg
, insn
->imm
);
2424 } else if (opcode
> BPF_END
) {
2425 verbose("invalid BPF_ALU opcode %x\n", opcode
);
2428 } else { /* all other ALU ops: and, sub, xor, add, ... */
2430 if (BPF_SRC(insn
->code
) == BPF_X
) {
2431 if (insn
->imm
!= 0 || insn
->off
!= 0) {
2432 verbose("BPF_ALU uses reserved fields\n");
2435 /* check src1 operand */
2436 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
2440 if (insn
->src_reg
!= BPF_REG_0
|| insn
->off
!= 0) {
2441 verbose("BPF_ALU uses reserved fields\n");
2446 /* check src2 operand */
2447 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
2451 if ((opcode
== BPF_MOD
|| opcode
== BPF_DIV
) &&
2452 BPF_SRC(insn
->code
) == BPF_K
&& insn
->imm
== 0) {
2453 verbose("div by zero\n");
2457 if ((opcode
== BPF_LSH
|| opcode
== BPF_RSH
||
2458 opcode
== BPF_ARSH
) && BPF_SRC(insn
->code
) == BPF_K
) {
2459 int size
= BPF_CLASS(insn
->code
) == BPF_ALU64
? 64 : 32;
2461 if (insn
->imm
< 0 || insn
->imm
>= size
) {
2462 verbose("invalid shift %d\n", insn
->imm
);
2467 /* check dest operand */
2468 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP_NO_MARK
);
2472 return adjust_reg_min_max_vals(env
, insn
);
2478 static void find_good_pkt_pointers(struct bpf_verifier_state
*state
,
2479 struct bpf_reg_state
*dst_reg
,
2480 enum bpf_reg_type type
)
2482 struct bpf_reg_state
*regs
= state
->regs
, *reg
;
2485 if (dst_reg
->off
< 0)
2486 /* This doesn't give us any range */
2489 if (dst_reg
->umax_value
> MAX_PACKET_OFF
||
2490 dst_reg
->umax_value
+ dst_reg
->off
> MAX_PACKET_OFF
)
2491 /* Risk of overflow. For instance, ptr + (1<<63) may be less
2492 * than pkt_end, but that's because it's also less than pkt.
2496 /* LLVM can generate four kind of checks:
2502 * if (r2 > pkt_end) goto <handle exception>
2507 * if (r2 < pkt_end) goto <access okay>
2508 * <handle exception>
2511 * r2 == dst_reg, pkt_end == src_reg
2512 * r2=pkt(id=n,off=8,r=0)
2513 * r3=pkt(id=n,off=0,r=0)
2519 * if (pkt_end >= r2) goto <access okay>
2520 * <handle exception>
2524 * if (pkt_end <= r2) goto <handle exception>
2528 * pkt_end == dst_reg, r2 == src_reg
2529 * r2=pkt(id=n,off=8,r=0)
2530 * r3=pkt(id=n,off=0,r=0)
2532 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
2533 * so that range of bytes [r3, r3 + 8) is safe to access.
2536 /* If our ids match, then we must have the same max_value. And we
2537 * don't care about the other reg's fixed offset, since if it's too big
2538 * the range won't allow anything.
2539 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
2541 for (i
= 0; i
< MAX_BPF_REG
; i
++)
2542 if (regs
[i
].type
== type
&& regs
[i
].id
== dst_reg
->id
)
2543 /* keep the maximum range already checked */
2544 regs
[i
].range
= max_t(u16
, regs
[i
].range
, dst_reg
->off
);
2546 for (i
= 0; i
< MAX_BPF_STACK
; i
+= BPF_REG_SIZE
) {
2547 if (state
->stack_slot_type
[i
] != STACK_SPILL
)
2549 reg
= &state
->spilled_regs
[i
/ BPF_REG_SIZE
];
2550 if (reg
->type
== type
&& reg
->id
== dst_reg
->id
)
2551 reg
->range
= max_t(u16
, reg
->range
, dst_reg
->off
);
2555 /* Adjusts the register min/max values in the case that the dst_reg is the
2556 * variable register that we are working on, and src_reg is a constant or we're
2557 * simply doing a BPF_K check.
2558 * In JEQ/JNE cases we also adjust the var_off values.
2560 static void reg_set_min_max(struct bpf_reg_state
*true_reg
,
2561 struct bpf_reg_state
*false_reg
, u64 val
,
2564 /* If the dst_reg is a pointer, we can't learn anything about its
2565 * variable offset from the compare (unless src_reg were a pointer into
2566 * the same object, but we don't bother with that.
2567 * Since false_reg and true_reg have the same type by construction, we
2568 * only need to check one of them for pointerness.
2570 if (__is_pointer_value(false, false_reg
))
2575 /* If this is false then we know nothing Jon Snow, but if it is
2576 * true then we know for sure.
2578 __mark_reg_known(true_reg
, val
);
2581 /* If this is true we know nothing Jon Snow, but if it is false
2582 * we know the value for sure;
2584 __mark_reg_known(false_reg
, val
);
2587 false_reg
->umax_value
= min(false_reg
->umax_value
, val
);
2588 true_reg
->umin_value
= max(true_reg
->umin_value
, val
+ 1);
2591 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
);
2592 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
+ 1);
2595 false_reg
->umin_value
= max(false_reg
->umin_value
, val
);
2596 true_reg
->umax_value
= min(true_reg
->umax_value
, val
- 1);
2599 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
);
2600 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
- 1);
2603 false_reg
->umax_value
= min(false_reg
->umax_value
, val
- 1);
2604 true_reg
->umin_value
= max(true_reg
->umin_value
, val
);
2607 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
- 1);
2608 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
);
2611 false_reg
->umin_value
= max(false_reg
->umin_value
, val
+ 1);
2612 true_reg
->umax_value
= min(true_reg
->umax_value
, val
);
2615 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
+ 1);
2616 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
);
2622 __reg_deduce_bounds(false_reg
);
2623 __reg_deduce_bounds(true_reg
);
2624 /* We might have learned some bits from the bounds. */
2625 __reg_bound_offset(false_reg
);
2626 __reg_bound_offset(true_reg
);
2627 /* Intersecting with the old var_off might have improved our bounds
2628 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2629 * then new var_off is (0; 0x7f...fc) which improves our umax.
2631 __update_reg_bounds(false_reg
);
2632 __update_reg_bounds(true_reg
);
2635 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
2638 static void reg_set_min_max_inv(struct bpf_reg_state
*true_reg
,
2639 struct bpf_reg_state
*false_reg
, u64 val
,
2642 if (__is_pointer_value(false, false_reg
))
2647 /* If this is false then we know nothing Jon Snow, but if it is
2648 * true then we know for sure.
2650 __mark_reg_known(true_reg
, val
);
2653 /* If this is true we know nothing Jon Snow, but if it is false
2654 * we know the value for sure;
2656 __mark_reg_known(false_reg
, val
);
2659 true_reg
->umax_value
= min(true_reg
->umax_value
, val
- 1);
2660 false_reg
->umin_value
= max(false_reg
->umin_value
, val
);
2663 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
- 1);
2664 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
);
2667 true_reg
->umin_value
= max(true_reg
->umin_value
, val
+ 1);
2668 false_reg
->umax_value
= min(false_reg
->umax_value
, val
);
2671 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
+ 1);
2672 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
);
2675 true_reg
->umax_value
= min(true_reg
->umax_value
, val
);
2676 false_reg
->umin_value
= max(false_reg
->umin_value
, val
+ 1);
2679 true_reg
->smax_value
= min_t(s64
, true_reg
->smax_value
, val
);
2680 false_reg
->smin_value
= max_t(s64
, false_reg
->smin_value
, val
+ 1);
2683 true_reg
->umin_value
= max(true_reg
->umin_value
, val
);
2684 false_reg
->umax_value
= min(false_reg
->umax_value
, val
- 1);
2687 true_reg
->smin_value
= max_t(s64
, true_reg
->smin_value
, val
);
2688 false_reg
->smax_value
= min_t(s64
, false_reg
->smax_value
, val
- 1);
2694 __reg_deduce_bounds(false_reg
);
2695 __reg_deduce_bounds(true_reg
);
2696 /* We might have learned some bits from the bounds. */
2697 __reg_bound_offset(false_reg
);
2698 __reg_bound_offset(true_reg
);
2699 /* Intersecting with the old var_off might have improved our bounds
2700 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2701 * then new var_off is (0; 0x7f...fc) which improves our umax.
2703 __update_reg_bounds(false_reg
);
2704 __update_reg_bounds(true_reg
);
2707 /* Regs are known to be equal, so intersect their min/max/var_off */
2708 static void __reg_combine_min_max(struct bpf_reg_state
*src_reg
,
2709 struct bpf_reg_state
*dst_reg
)
2711 src_reg
->umin_value
= dst_reg
->umin_value
= max(src_reg
->umin_value
,
2712 dst_reg
->umin_value
);
2713 src_reg
->umax_value
= dst_reg
->umax_value
= min(src_reg
->umax_value
,
2714 dst_reg
->umax_value
);
2715 src_reg
->smin_value
= dst_reg
->smin_value
= max(src_reg
->smin_value
,
2716 dst_reg
->smin_value
);
2717 src_reg
->smax_value
= dst_reg
->smax_value
= min(src_reg
->smax_value
,
2718 dst_reg
->smax_value
);
2719 src_reg
->var_off
= dst_reg
->var_off
= tnum_intersect(src_reg
->var_off
,
2721 /* We might have learned new bounds from the var_off. */
2722 __update_reg_bounds(src_reg
);
2723 __update_reg_bounds(dst_reg
);
2724 /* We might have learned something about the sign bit. */
2725 __reg_deduce_bounds(src_reg
);
2726 __reg_deduce_bounds(dst_reg
);
2727 /* We might have learned some bits from the bounds. */
2728 __reg_bound_offset(src_reg
);
2729 __reg_bound_offset(dst_reg
);
2730 /* Intersecting with the old var_off might have improved our bounds
2731 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2732 * then new var_off is (0; 0x7f...fc) which improves our umax.
2734 __update_reg_bounds(src_reg
);
2735 __update_reg_bounds(dst_reg
);
2738 static void reg_combine_min_max(struct bpf_reg_state
*true_src
,
2739 struct bpf_reg_state
*true_dst
,
2740 struct bpf_reg_state
*false_src
,
2741 struct bpf_reg_state
*false_dst
,
2746 __reg_combine_min_max(true_src
, true_dst
);
2749 __reg_combine_min_max(false_src
, false_dst
);
2754 static void mark_map_reg(struct bpf_reg_state
*regs
, u32 regno
, u32 id
,
2757 struct bpf_reg_state
*reg
= ®s
[regno
];
2759 if (reg
->type
== PTR_TO_MAP_VALUE_OR_NULL
&& reg
->id
== id
) {
2760 /* Old offset (both fixed and variable parts) should
2761 * have been known-zero, because we don't allow pointer
2762 * arithmetic on pointers that might be NULL.
2764 if (WARN_ON_ONCE(reg
->smin_value
|| reg
->smax_value
||
2765 !tnum_equals_const(reg
->var_off
, 0) ||
2767 __mark_reg_known_zero(reg
);
2771 reg
->type
= SCALAR_VALUE
;
2772 } else if (reg
->map_ptr
->inner_map_meta
) {
2773 reg
->type
= CONST_PTR_TO_MAP
;
2774 reg
->map_ptr
= reg
->map_ptr
->inner_map_meta
;
2776 reg
->type
= PTR_TO_MAP_VALUE
;
2778 /* We don't need id from this point onwards anymore, thus we
2779 * should better reset it, so that state pruning has chances
2786 /* The logic is similar to find_good_pkt_pointers(), both could eventually
2787 * be folded together at some point.
2789 static void mark_map_regs(struct bpf_verifier_state
*state
, u32 regno
,
2792 struct bpf_reg_state
*regs
= state
->regs
;
2793 u32 id
= regs
[regno
].id
;
2796 for (i
= 0; i
< MAX_BPF_REG
; i
++)
2797 mark_map_reg(regs
, i
, id
, is_null
);
2799 for (i
= 0; i
< MAX_BPF_STACK
; i
+= BPF_REG_SIZE
) {
2800 if (state
->stack_slot_type
[i
] != STACK_SPILL
)
2802 mark_map_reg(state
->spilled_regs
, i
/ BPF_REG_SIZE
, id
, is_null
);
2806 static int check_cond_jmp_op(struct bpf_verifier_env
*env
,
2807 struct bpf_insn
*insn
, int *insn_idx
)
2809 struct bpf_verifier_state
*other_branch
, *this_branch
= &env
->cur_state
;
2810 struct bpf_reg_state
*regs
= this_branch
->regs
, *dst_reg
;
2811 u8 opcode
= BPF_OP(insn
->code
);
2814 if (opcode
> BPF_JSLE
) {
2815 verbose("invalid BPF_JMP opcode %x\n", opcode
);
2819 if (BPF_SRC(insn
->code
) == BPF_X
) {
2820 if (insn
->imm
!= 0) {
2821 verbose("BPF_JMP uses reserved fields\n");
2825 /* check src1 operand */
2826 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
2830 if (is_pointer_value(env
, insn
->src_reg
)) {
2831 verbose("R%d pointer comparison prohibited\n",
2836 if (insn
->src_reg
!= BPF_REG_0
) {
2837 verbose("BPF_JMP uses reserved fields\n");
2842 /* check src2 operand */
2843 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
2847 dst_reg
= ®s
[insn
->dst_reg
];
2849 /* detect if R == 0 where R was initialized to zero earlier */
2850 if (BPF_SRC(insn
->code
) == BPF_K
&&
2851 (opcode
== BPF_JEQ
|| opcode
== BPF_JNE
) &&
2852 dst_reg
->type
== SCALAR_VALUE
&&
2853 tnum_equals_const(dst_reg
->var_off
, insn
->imm
)) {
2854 if (opcode
== BPF_JEQ
) {
2855 /* if (imm == imm) goto pc+off;
2856 * only follow the goto, ignore fall-through
2858 *insn_idx
+= insn
->off
;
2861 /* if (imm != imm) goto pc+off;
2862 * only follow fall-through branch, since
2863 * that's where the program will go
2869 other_branch
= push_stack(env
, *insn_idx
+ insn
->off
+ 1, *insn_idx
);
2873 /* detect if we are comparing against a constant value so we can adjust
2874 * our min/max values for our dst register.
2875 * this is only legit if both are scalars (or pointers to the same
2876 * object, I suppose, but we don't support that right now), because
2877 * otherwise the different base pointers mean the offsets aren't
2880 if (BPF_SRC(insn
->code
) == BPF_X
) {
2881 if (dst_reg
->type
== SCALAR_VALUE
&&
2882 regs
[insn
->src_reg
].type
== SCALAR_VALUE
) {
2883 if (tnum_is_const(regs
[insn
->src_reg
].var_off
))
2884 reg_set_min_max(&other_branch
->regs
[insn
->dst_reg
],
2885 dst_reg
, regs
[insn
->src_reg
].var_off
.value
,
2887 else if (tnum_is_const(dst_reg
->var_off
))
2888 reg_set_min_max_inv(&other_branch
->regs
[insn
->src_reg
],
2889 ®s
[insn
->src_reg
],
2890 dst_reg
->var_off
.value
, opcode
);
2891 else if (opcode
== BPF_JEQ
|| opcode
== BPF_JNE
)
2892 /* Comparing for equality, we can combine knowledge */
2893 reg_combine_min_max(&other_branch
->regs
[insn
->src_reg
],
2894 &other_branch
->regs
[insn
->dst_reg
],
2895 ®s
[insn
->src_reg
],
2896 ®s
[insn
->dst_reg
], opcode
);
2898 } else if (dst_reg
->type
== SCALAR_VALUE
) {
2899 reg_set_min_max(&other_branch
->regs
[insn
->dst_reg
],
2900 dst_reg
, insn
->imm
, opcode
);
2903 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
2904 if (BPF_SRC(insn
->code
) == BPF_K
&&
2905 insn
->imm
== 0 && (opcode
== BPF_JEQ
|| opcode
== BPF_JNE
) &&
2906 dst_reg
->type
== PTR_TO_MAP_VALUE_OR_NULL
) {
2907 /* Mark all identical map registers in each branch as either
2908 * safe or unknown depending R == 0 or R != 0 conditional.
2910 mark_map_regs(this_branch
, insn
->dst_reg
, opcode
== BPF_JNE
);
2911 mark_map_regs(other_branch
, insn
->dst_reg
, opcode
== BPF_JEQ
);
2912 } else if (BPF_SRC(insn
->code
) == BPF_X
&& opcode
== BPF_JGT
&&
2913 dst_reg
->type
== PTR_TO_PACKET
&&
2914 regs
[insn
->src_reg
].type
== PTR_TO_PACKET_END
) {
2915 find_good_pkt_pointers(this_branch
, dst_reg
, PTR_TO_PACKET
);
2916 } else if (BPF_SRC(insn
->code
) == BPF_X
&& opcode
== BPF_JLT
&&
2917 dst_reg
->type
== PTR_TO_PACKET
&&
2918 regs
[insn
->src_reg
].type
== PTR_TO_PACKET_END
) {
2919 find_good_pkt_pointers(other_branch
, dst_reg
, PTR_TO_PACKET
);
2920 } else if (BPF_SRC(insn
->code
) == BPF_X
&& opcode
== BPF_JGE
&&
2921 dst_reg
->type
== PTR_TO_PACKET_END
&&
2922 regs
[insn
->src_reg
].type
== PTR_TO_PACKET
) {
2923 find_good_pkt_pointers(other_branch
, ®s
[insn
->src_reg
],
2925 } else if (BPF_SRC(insn
->code
) == BPF_X
&& opcode
== BPF_JLE
&&
2926 dst_reg
->type
== PTR_TO_PACKET_END
&&
2927 regs
[insn
->src_reg
].type
== PTR_TO_PACKET
) {
2928 find_good_pkt_pointers(this_branch
, ®s
[insn
->src_reg
],
2930 } else if (BPF_SRC(insn
->code
) == BPF_X
&& opcode
== BPF_JGT
&&
2931 dst_reg
->type
== PTR_TO_PACKET_META
&&
2932 reg_is_init_pkt_pointer(®s
[insn
->src_reg
], PTR_TO_PACKET
)) {
2933 find_good_pkt_pointers(this_branch
, dst_reg
, PTR_TO_PACKET_META
);
2934 } else if (BPF_SRC(insn
->code
) == BPF_X
&& opcode
== BPF_JLT
&&
2935 dst_reg
->type
== PTR_TO_PACKET_META
&&
2936 reg_is_init_pkt_pointer(®s
[insn
->src_reg
], PTR_TO_PACKET
)) {
2937 find_good_pkt_pointers(other_branch
, dst_reg
, PTR_TO_PACKET_META
);
2938 } else if (BPF_SRC(insn
->code
) == BPF_X
&& opcode
== BPF_JGE
&&
2939 reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
2940 regs
[insn
->src_reg
].type
== PTR_TO_PACKET_META
) {
2941 find_good_pkt_pointers(other_branch
, ®s
[insn
->src_reg
],
2942 PTR_TO_PACKET_META
);
2943 } else if (BPF_SRC(insn
->code
) == BPF_X
&& opcode
== BPF_JLE
&&
2944 reg_is_init_pkt_pointer(dst_reg
, PTR_TO_PACKET
) &&
2945 regs
[insn
->src_reg
].type
== PTR_TO_PACKET_META
) {
2946 find_good_pkt_pointers(this_branch
, ®s
[insn
->src_reg
],
2947 PTR_TO_PACKET_META
);
2948 } else if (is_pointer_value(env
, insn
->dst_reg
)) {
2949 verbose("R%d pointer comparison prohibited\n", insn
->dst_reg
);
2953 print_verifier_state(this_branch
);
2957 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
2958 static struct bpf_map
*ld_imm64_to_map_ptr(struct bpf_insn
*insn
)
2960 u64 imm64
= ((u64
) (u32
) insn
[0].imm
) | ((u64
) (u32
) insn
[1].imm
) << 32;
2962 return (struct bpf_map
*) (unsigned long) imm64
;
2965 /* verify BPF_LD_IMM64 instruction */
2966 static int check_ld_imm(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
2968 struct bpf_reg_state
*regs
= env
->cur_state
.regs
;
2971 if (BPF_SIZE(insn
->code
) != BPF_DW
) {
2972 verbose("invalid BPF_LD_IMM insn\n");
2975 if (insn
->off
!= 0) {
2976 verbose("BPF_LD_IMM64 uses reserved fields\n");
2980 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP
);
2984 if (insn
->src_reg
== 0) {
2985 u64 imm
= ((u64
)(insn
+ 1)->imm
<< 32) | (u32
)insn
->imm
;
2987 regs
[insn
->dst_reg
].type
= SCALAR_VALUE
;
2988 __mark_reg_known(®s
[insn
->dst_reg
], imm
);
2992 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
2993 BUG_ON(insn
->src_reg
!= BPF_PSEUDO_MAP_FD
);
2995 regs
[insn
->dst_reg
].type
= CONST_PTR_TO_MAP
;
2996 regs
[insn
->dst_reg
].map_ptr
= ld_imm64_to_map_ptr(insn
);
3000 static bool may_access_skb(enum bpf_prog_type type
)
3003 case BPF_PROG_TYPE_SOCKET_FILTER
:
3004 case BPF_PROG_TYPE_SCHED_CLS
:
3005 case BPF_PROG_TYPE_SCHED_ACT
:
3012 /* verify safety of LD_ABS|LD_IND instructions:
3013 * - they can only appear in the programs where ctx == skb
3014 * - since they are wrappers of function calls, they scratch R1-R5 registers,
3015 * preserve R6-R9, and store return value into R0
3018 * ctx == skb == R6 == CTX
3021 * SRC == any register
3022 * IMM == 32-bit immediate
3025 * R0 - 8/16/32-bit skb data converted to cpu endianness
3027 static int check_ld_abs(struct bpf_verifier_env
*env
, struct bpf_insn
*insn
)
3029 struct bpf_reg_state
*regs
= env
->cur_state
.regs
;
3030 u8 mode
= BPF_MODE(insn
->code
);
3033 if (!may_access_skb(env
->prog
->type
)) {
3034 verbose("BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
3038 if (insn
->dst_reg
!= BPF_REG_0
|| insn
->off
!= 0 ||
3039 BPF_SIZE(insn
->code
) == BPF_DW
||
3040 (mode
== BPF_ABS
&& insn
->src_reg
!= BPF_REG_0
)) {
3041 verbose("BPF_LD_[ABS|IND] uses reserved fields\n");
3045 /* check whether implicit source operand (register R6) is readable */
3046 err
= check_reg_arg(env
, BPF_REG_6
, SRC_OP
);
3050 if (regs
[BPF_REG_6
].type
!= PTR_TO_CTX
) {
3051 verbose("at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
3055 if (mode
== BPF_IND
) {
3056 /* check explicit source operand */
3057 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
3062 /* reset caller saved regs to unreadable */
3063 for (i
= 0; i
< CALLER_SAVED_REGS
; i
++) {
3064 mark_reg_not_init(regs
, caller_saved
[i
]);
3065 check_reg_arg(env
, caller_saved
[i
], DST_OP_NO_MARK
);
3068 /* mark destination R0 register as readable, since it contains
3069 * the value fetched from the packet.
3070 * Already marked as written above.
3072 mark_reg_unknown(regs
, BPF_REG_0
);
3076 static int check_return_code(struct bpf_verifier_env
*env
)
3078 struct bpf_reg_state
*reg
;
3079 struct tnum range
= tnum_range(0, 1);
3081 switch (env
->prog
->type
) {
3082 case BPF_PROG_TYPE_CGROUP_SKB
:
3083 case BPF_PROG_TYPE_CGROUP_SOCK
:
3084 case BPF_PROG_TYPE_SOCK_OPS
:
3090 reg
= &env
->cur_state
.regs
[BPF_REG_0
];
3091 if (reg
->type
!= SCALAR_VALUE
) {
3092 verbose("At program exit the register R0 is not a known value (%s)\n",
3093 reg_type_str
[reg
->type
]);
3097 if (!tnum_in(range
, reg
->var_off
)) {
3098 verbose("At program exit the register R0 ");
3099 if (!tnum_is_unknown(reg
->var_off
)) {
3102 tnum_strn(tn_buf
, sizeof(tn_buf
), reg
->var_off
);
3103 verbose("has value %s", tn_buf
);
3105 verbose("has unknown scalar value");
3107 verbose(" should have been 0 or 1\n");
3113 /* non-recursive DFS pseudo code
3114 * 1 procedure DFS-iterative(G,v):
3115 * 2 label v as discovered
3116 * 3 let S be a stack
3118 * 5 while S is not empty
3120 * 7 if t is what we're looking for:
3122 * 9 for all edges e in G.adjacentEdges(t) do
3123 * 10 if edge e is already labelled
3124 * 11 continue with the next edge
3125 * 12 w <- G.adjacentVertex(t,e)
3126 * 13 if vertex w is not discovered and not explored
3127 * 14 label e as tree-edge
3128 * 15 label w as discovered
3131 * 18 else if vertex w is discovered
3132 * 19 label e as back-edge
3134 * 21 // vertex w is explored
3135 * 22 label e as forward- or cross-edge
3136 * 23 label t as explored
3141 * 0x11 - discovered and fall-through edge labelled
3142 * 0x12 - discovered and fall-through and branch edges labelled
3153 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
3155 static int *insn_stack
; /* stack of insns to process */
3156 static int cur_stack
; /* current stack index */
3157 static int *insn_state
;
3159 /* t, w, e - match pseudo-code above:
3160 * t - index of current instruction
3161 * w - next instruction
3164 static int push_insn(int t
, int w
, int e
, struct bpf_verifier_env
*env
)
3166 if (e
== FALLTHROUGH
&& insn_state
[t
] >= (DISCOVERED
| FALLTHROUGH
))
3169 if (e
== BRANCH
&& insn_state
[t
] >= (DISCOVERED
| BRANCH
))
3172 if (w
< 0 || w
>= env
->prog
->len
) {
3173 verbose("jump out of range from insn %d to %d\n", t
, w
);
3178 /* mark branch target for state pruning */
3179 env
->explored_states
[w
] = STATE_LIST_MARK
;
3181 if (insn_state
[w
] == 0) {
3183 insn_state
[t
] = DISCOVERED
| e
;
3184 insn_state
[w
] = DISCOVERED
;
3185 if (cur_stack
>= env
->prog
->len
)
3187 insn_stack
[cur_stack
++] = w
;
3189 } else if ((insn_state
[w
] & 0xF0) == DISCOVERED
) {
3190 verbose("back-edge from insn %d to %d\n", t
, w
);
3192 } else if (insn_state
[w
] == EXPLORED
) {
3193 /* forward- or cross-edge */
3194 insn_state
[t
] = DISCOVERED
| e
;
3196 verbose("insn state internal bug\n");
3202 /* non-recursive depth-first-search to detect loops in BPF program
3203 * loop == back-edge in directed graph
3205 static int check_cfg(struct bpf_verifier_env
*env
)
3207 struct bpf_insn
*insns
= env
->prog
->insnsi
;
3208 int insn_cnt
= env
->prog
->len
;
3212 insn_state
= kcalloc(insn_cnt
, sizeof(int), GFP_KERNEL
);
3216 insn_stack
= kcalloc(insn_cnt
, sizeof(int), GFP_KERNEL
);
3222 insn_state
[0] = DISCOVERED
; /* mark 1st insn as discovered */
3223 insn_stack
[0] = 0; /* 0 is the first instruction */
3229 t
= insn_stack
[cur_stack
- 1];
3231 if (BPF_CLASS(insns
[t
].code
) == BPF_JMP
) {
3232 u8 opcode
= BPF_OP(insns
[t
].code
);
3234 if (opcode
== BPF_EXIT
) {
3236 } else if (opcode
== BPF_CALL
) {
3237 ret
= push_insn(t
, t
+ 1, FALLTHROUGH
, env
);
3242 if (t
+ 1 < insn_cnt
)
3243 env
->explored_states
[t
+ 1] = STATE_LIST_MARK
;
3244 } else if (opcode
== BPF_JA
) {
3245 if (BPF_SRC(insns
[t
].code
) != BPF_K
) {
3249 /* unconditional jump with single edge */
3250 ret
= push_insn(t
, t
+ insns
[t
].off
+ 1,
3256 /* tell verifier to check for equivalent states
3257 * after every call and jump
3259 if (t
+ 1 < insn_cnt
)
3260 env
->explored_states
[t
+ 1] = STATE_LIST_MARK
;
3262 /* conditional jump with two edges */
3263 env
->explored_states
[t
] = STATE_LIST_MARK
;
3264 ret
= push_insn(t
, t
+ 1, FALLTHROUGH
, env
);
3270 ret
= push_insn(t
, t
+ insns
[t
].off
+ 1, BRANCH
, env
);
3277 /* all other non-branch instructions with single
3280 ret
= push_insn(t
, t
+ 1, FALLTHROUGH
, env
);
3288 insn_state
[t
] = EXPLORED
;
3289 if (cur_stack
-- <= 0) {
3290 verbose("pop stack internal bug\n");
3297 for (i
= 0; i
< insn_cnt
; i
++) {
3298 if (insn_state
[i
] != EXPLORED
) {
3299 verbose("unreachable insn %d\n", i
);
3304 ret
= 0; /* cfg looks good */
3312 /* check %cur's range satisfies %old's */
3313 static bool range_within(struct bpf_reg_state
*old
,
3314 struct bpf_reg_state
*cur
)
3316 return old
->umin_value
<= cur
->umin_value
&&
3317 old
->umax_value
>= cur
->umax_value
&&
3318 old
->smin_value
<= cur
->smin_value
&&
3319 old
->smax_value
>= cur
->smax_value
;
3322 /* Maximum number of register states that can exist at once */
3323 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
3329 /* If in the old state two registers had the same id, then they need to have
3330 * the same id in the new state as well. But that id could be different from
3331 * the old state, so we need to track the mapping from old to new ids.
3332 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
3333 * regs with old id 5 must also have new id 9 for the new state to be safe. But
3334 * regs with a different old id could still have new id 9, we don't care about
3336 * So we look through our idmap to see if this old id has been seen before. If
3337 * so, we require the new id to match; otherwise, we add the id pair to the map.
3339 static bool check_ids(u32 old_id
, u32 cur_id
, struct idpair
*idmap
)
3343 for (i
= 0; i
< ID_MAP_SIZE
; i
++) {
3344 if (!idmap
[i
].old
) {
3345 /* Reached an empty slot; haven't seen this id before */
3346 idmap
[i
].old
= old_id
;
3347 idmap
[i
].cur
= cur_id
;
3350 if (idmap
[i
].old
== old_id
)
3351 return idmap
[i
].cur
== cur_id
;
3353 /* We ran out of idmap slots, which should be impossible */
3358 /* Returns true if (rold safe implies rcur safe) */
3359 static bool regsafe(struct bpf_reg_state
*rold
, struct bpf_reg_state
*rcur
,
3360 struct idpair
*idmap
)
3362 if (!(rold
->live
& REG_LIVE_READ
))
3363 /* explored state didn't use this */
3366 if (memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, live
)) == 0)
3369 if (rold
->type
== NOT_INIT
)
3370 /* explored state can't have used this */
3372 if (rcur
->type
== NOT_INIT
)
3374 switch (rold
->type
) {
3376 if (rcur
->type
== SCALAR_VALUE
) {
3377 /* new val must satisfy old val knowledge */
3378 return range_within(rold
, rcur
) &&
3379 tnum_in(rold
->var_off
, rcur
->var_off
);
3381 /* if we knew anything about the old value, we're not
3382 * equal, because we can't know anything about the
3383 * scalar value of the pointer in the new value.
3385 return rold
->umin_value
== 0 &&
3386 rold
->umax_value
== U64_MAX
&&
3387 rold
->smin_value
== S64_MIN
&&
3388 rold
->smax_value
== S64_MAX
&&
3389 tnum_is_unknown(rold
->var_off
);
3391 case PTR_TO_MAP_VALUE
:
3392 /* If the new min/max/var_off satisfy the old ones and
3393 * everything else matches, we are OK.
3394 * We don't care about the 'id' value, because nothing
3395 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
3397 return memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, id
)) == 0 &&
3398 range_within(rold
, rcur
) &&
3399 tnum_in(rold
->var_off
, rcur
->var_off
);
3400 case PTR_TO_MAP_VALUE_OR_NULL
:
3401 /* a PTR_TO_MAP_VALUE could be safe to use as a
3402 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
3403 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
3404 * checked, doing so could have affected others with the same
3405 * id, and we can't check for that because we lost the id when
3406 * we converted to a PTR_TO_MAP_VALUE.
3408 if (rcur
->type
!= PTR_TO_MAP_VALUE_OR_NULL
)
3410 if (memcmp(rold
, rcur
, offsetof(struct bpf_reg_state
, id
)))
3412 /* Check our ids match any regs they're supposed to */
3413 return check_ids(rold
->id
, rcur
->id
, idmap
);
3414 case PTR_TO_PACKET_META
:
3416 if (rcur
->type
!= rold
->type
)
3418 /* We must have at least as much range as the old ptr
3419 * did, so that any accesses which were safe before are
3420 * still safe. This is true even if old range < old off,
3421 * since someone could have accessed through (ptr - k), or
3422 * even done ptr -= k in a register, to get a safe access.
3424 if (rold
->range
> rcur
->range
)
3426 /* If the offsets don't match, we can't trust our alignment;
3427 * nor can we be sure that we won't fall out of range.
3429 if (rold
->off
!= rcur
->off
)
3431 /* id relations must be preserved */
3432 if (rold
->id
&& !check_ids(rold
->id
, rcur
->id
, idmap
))
3434 /* new val must satisfy old val knowledge */
3435 return range_within(rold
, rcur
) &&
3436 tnum_in(rold
->var_off
, rcur
->var_off
);
3438 case CONST_PTR_TO_MAP
:
3440 case PTR_TO_PACKET_END
:
3441 /* Only valid matches are exact, which memcmp() above
3442 * would have accepted
3445 /* Don't know what's going on, just say it's not safe */
3449 /* Shouldn't get here; if we do, say it's not safe */
3454 /* compare two verifier states
3456 * all states stored in state_list are known to be valid, since
3457 * verifier reached 'bpf_exit' instruction through them
3459 * this function is called when verifier exploring different branches of
3460 * execution popped from the state stack. If it sees an old state that has
3461 * more strict register state and more strict stack state then this execution
3462 * branch doesn't need to be explored further, since verifier already
3463 * concluded that more strict state leads to valid finish.
3465 * Therefore two states are equivalent if register state is more conservative
3466 * and explored stack state is more conservative than the current one.
3469 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
3470 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
3472 * In other words if current stack state (one being explored) has more
3473 * valid slots than old one that already passed validation, it means
3474 * the verifier can stop exploring and conclude that current state is valid too
3476 * Similarly with registers. If explored state has register type as invalid
3477 * whereas register type in current state is meaningful, it means that
3478 * the current state will reach 'bpf_exit' instruction safely
3480 static bool states_equal(struct bpf_verifier_env
*env
,
3481 struct bpf_verifier_state
*old
,
3482 struct bpf_verifier_state
*cur
)
3484 struct idpair
*idmap
;
3488 idmap
= kcalloc(ID_MAP_SIZE
, sizeof(struct idpair
), GFP_KERNEL
);
3489 /* If we failed to allocate the idmap, just say it's not safe */
3493 for (i
= 0; i
< MAX_BPF_REG
; i
++) {
3494 if (!regsafe(&old
->regs
[i
], &cur
->regs
[i
], idmap
))
3498 for (i
= 0; i
< MAX_BPF_STACK
; i
++) {
3499 if (old
->stack_slot_type
[i
] == STACK_INVALID
)
3501 if (old
->stack_slot_type
[i
] != cur
->stack_slot_type
[i
])
3502 /* Ex: old explored (safe) state has STACK_SPILL in
3503 * this stack slot, but current has has STACK_MISC ->
3504 * this verifier states are not equivalent,
3505 * return false to continue verification of this path
3508 if (i
% BPF_REG_SIZE
)
3510 if (old
->stack_slot_type
[i
] != STACK_SPILL
)
3512 if (!regsafe(&old
->spilled_regs
[i
/ BPF_REG_SIZE
],
3513 &cur
->spilled_regs
[i
/ BPF_REG_SIZE
],
3515 /* when explored and current stack slot are both storing
3516 * spilled registers, check that stored pointers types
3517 * are the same as well.
3518 * Ex: explored safe path could have stored
3519 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
3520 * but current path has stored:
3521 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
3522 * such verifier states are not equivalent.
3523 * return false to continue verification of this path
3535 /* A write screens off any subsequent reads; but write marks come from the
3536 * straight-line code between a state and its parent. When we arrive at a
3537 * jump target (in the first iteration of the propagate_liveness() loop),
3538 * we didn't arrive by the straight-line code, so read marks in state must
3539 * propagate to parent regardless of state's write marks.
3541 static bool do_propagate_liveness(const struct bpf_verifier_state
*state
,
3542 struct bpf_verifier_state
*parent
)
3544 bool writes
= parent
== state
->parent
; /* Observe write marks */
3545 bool touched
= false; /* any changes made? */
3550 /* Propagate read liveness of registers... */
3551 BUILD_BUG_ON(BPF_REG_FP
+ 1 != MAX_BPF_REG
);
3552 /* We don't need to worry about FP liveness because it's read-only */
3553 for (i
= 0; i
< BPF_REG_FP
; i
++) {
3554 if (parent
->regs
[i
].live
& REG_LIVE_READ
)
3556 if (writes
&& (state
->regs
[i
].live
& REG_LIVE_WRITTEN
))
3558 if (state
->regs
[i
].live
& REG_LIVE_READ
) {
3559 parent
->regs
[i
].live
|= REG_LIVE_READ
;
3563 /* ... and stack slots */
3564 for (i
= 0; i
< MAX_BPF_STACK
/ BPF_REG_SIZE
; i
++) {
3565 if (parent
->stack_slot_type
[i
* BPF_REG_SIZE
] != STACK_SPILL
)
3567 if (state
->stack_slot_type
[i
* BPF_REG_SIZE
] != STACK_SPILL
)
3569 if (parent
->spilled_regs
[i
].live
& REG_LIVE_READ
)
3571 if (writes
&& (state
->spilled_regs
[i
].live
& REG_LIVE_WRITTEN
))
3573 if (state
->spilled_regs
[i
].live
& REG_LIVE_READ
) {
3574 parent
->spilled_regs
[i
].live
|= REG_LIVE_READ
;
3581 /* "parent" is "a state from which we reach the current state", but initially
3582 * it is not the state->parent (i.e. "the state whose straight-line code leads
3583 * to the current state"), instead it is the state that happened to arrive at
3584 * a (prunable) equivalent of the current state. See comment above
3585 * do_propagate_liveness() for consequences of this.
3586 * This function is just a more efficient way of calling mark_reg_read() or
3587 * mark_stack_slot_read() on each reg in "parent" that is read in "state",
3588 * though it requires that parent != state->parent in the call arguments.
3590 static void propagate_liveness(const struct bpf_verifier_state
*state
,
3591 struct bpf_verifier_state
*parent
)
3593 while (do_propagate_liveness(state
, parent
)) {
3594 /* Something changed, so we need to feed those changes onward */
3596 parent
= state
->parent
;
3600 static int is_state_visited(struct bpf_verifier_env
*env
, int insn_idx
)
3602 struct bpf_verifier_state_list
*new_sl
;
3603 struct bpf_verifier_state_list
*sl
;
3606 sl
= env
->explored_states
[insn_idx
];
3608 /* this 'insn_idx' instruction wasn't marked, so we will not
3609 * be doing state search here
3613 while (sl
!= STATE_LIST_MARK
) {
3614 if (states_equal(env
, &sl
->state
, &env
->cur_state
)) {
3615 /* reached equivalent register/stack state,
3617 * Registers read by the continuation are read by us.
3618 * If we have any write marks in env->cur_state, they
3619 * will prevent corresponding reads in the continuation
3620 * from reaching our parent (an explored_state). Our
3621 * own state will get the read marks recorded, but
3622 * they'll be immediately forgotten as we're pruning
3623 * this state and will pop a new one.
3625 propagate_liveness(&sl
->state
, &env
->cur_state
);
3631 /* there were no equivalent states, remember current one.
3632 * technically the current state is not proven to be safe yet,
3633 * but it will either reach bpf_exit (which means it's safe) or
3634 * it will be rejected. Since there are no loops, we won't be
3635 * seeing this 'insn_idx' instruction again on the way to bpf_exit
3637 new_sl
= kmalloc(sizeof(struct bpf_verifier_state_list
), GFP_USER
);
3641 /* add new state to the head of linked list */
3642 memcpy(&new_sl
->state
, &env
->cur_state
, sizeof(env
->cur_state
));
3643 new_sl
->next
= env
->explored_states
[insn_idx
];
3644 env
->explored_states
[insn_idx
] = new_sl
;
3645 /* connect new state to parentage chain */
3646 env
->cur_state
.parent
= &new_sl
->state
;
3647 /* clear write marks in current state: the writes we did are not writes
3648 * our child did, so they don't screen off its reads from us.
3649 * (There are no read marks in current state, because reads always mark
3650 * their parent and current state never has children yet. Only
3651 * explored_states can get read marks.)
3653 for (i
= 0; i
< BPF_REG_FP
; i
++)
3654 env
->cur_state
.regs
[i
].live
= REG_LIVE_NONE
;
3655 for (i
= 0; i
< MAX_BPF_STACK
/ BPF_REG_SIZE
; i
++)
3656 if (env
->cur_state
.stack_slot_type
[i
* BPF_REG_SIZE
] == STACK_SPILL
)
3657 env
->cur_state
.spilled_regs
[i
].live
= REG_LIVE_NONE
;
3661 static int ext_analyzer_insn_hook(struct bpf_verifier_env
*env
,
3662 int insn_idx
, int prev_insn_idx
)
3664 if (!env
->analyzer_ops
|| !env
->analyzer_ops
->insn_hook
)
3667 return env
->analyzer_ops
->insn_hook(env
, insn_idx
, prev_insn_idx
);
3670 static int do_check(struct bpf_verifier_env
*env
)
3672 struct bpf_verifier_state
*state
= &env
->cur_state
;
3673 struct bpf_insn
*insns
= env
->prog
->insnsi
;
3674 struct bpf_reg_state
*regs
= state
->regs
;
3675 int insn_cnt
= env
->prog
->len
;
3676 int insn_idx
, prev_insn_idx
= 0;
3677 int insn_processed
= 0;
3678 bool do_print_state
= false;
3680 init_reg_state(regs
);
3681 state
->parent
= NULL
;
3684 struct bpf_insn
*insn
;
3688 if (insn_idx
>= insn_cnt
) {
3689 verbose("invalid insn idx %d insn_cnt %d\n",
3690 insn_idx
, insn_cnt
);
3694 insn
= &insns
[insn_idx
];
3695 class = BPF_CLASS(insn
->code
);
3697 if (++insn_processed
> BPF_COMPLEXITY_LIMIT_INSNS
) {
3698 verbose("BPF program is too large. Processed %d insn\n",
3703 err
= is_state_visited(env
, insn_idx
);
3707 /* found equivalent state, can prune the search */
3710 verbose("\nfrom %d to %d: safe\n",
3711 prev_insn_idx
, insn_idx
);
3713 verbose("%d: safe\n", insn_idx
);
3715 goto process_bpf_exit
;
3721 if (log_level
> 1 || (log_level
&& do_print_state
)) {
3723 verbose("%d:", insn_idx
);
3725 verbose("\nfrom %d to %d:",
3726 prev_insn_idx
, insn_idx
);
3727 print_verifier_state(&env
->cur_state
);
3728 do_print_state
= false;
3732 verbose("%d: ", insn_idx
);
3733 print_bpf_insn(env
, insn
);
3736 err
= ext_analyzer_insn_hook(env
, insn_idx
, prev_insn_idx
);
3740 if (class == BPF_ALU
|| class == BPF_ALU64
) {
3741 err
= check_alu_op(env
, insn
);
3745 } else if (class == BPF_LDX
) {
3746 enum bpf_reg_type
*prev_src_type
, src_reg_type
;
3748 /* check for reserved fields is already done */
3750 /* check src operand */
3751 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
3755 err
= check_reg_arg(env
, insn
->dst_reg
, DST_OP_NO_MARK
);
3759 src_reg_type
= regs
[insn
->src_reg
].type
;
3761 /* check that memory (src_reg + off) is readable,
3762 * the state of dst_reg will be updated by this func
3764 err
= check_mem_access(env
, insn_idx
, insn
->src_reg
, insn
->off
,
3765 BPF_SIZE(insn
->code
), BPF_READ
,
3770 prev_src_type
= &env
->insn_aux_data
[insn_idx
].ptr_type
;
3772 if (*prev_src_type
== NOT_INIT
) {
3774 * dst_reg = *(u32 *)(src_reg + off)
3775 * save type to validate intersecting paths
3777 *prev_src_type
= src_reg_type
;
3779 } else if (src_reg_type
!= *prev_src_type
&&
3780 (src_reg_type
== PTR_TO_CTX
||
3781 *prev_src_type
== PTR_TO_CTX
)) {
3782 /* ABuser program is trying to use the same insn
3783 * dst_reg = *(u32*) (src_reg + off)
3784 * with different pointer types:
3785 * src_reg == ctx in one branch and
3786 * src_reg == stack|map in some other branch.
3789 verbose("same insn cannot be used with different pointers\n");
3793 } else if (class == BPF_STX
) {
3794 enum bpf_reg_type
*prev_dst_type
, dst_reg_type
;
3796 if (BPF_MODE(insn
->code
) == BPF_XADD
) {
3797 err
= check_xadd(env
, insn_idx
, insn
);
3804 /* check src1 operand */
3805 err
= check_reg_arg(env
, insn
->src_reg
, SRC_OP
);
3808 /* check src2 operand */
3809 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
3813 dst_reg_type
= regs
[insn
->dst_reg
].type
;
3815 /* check that memory (dst_reg + off) is writeable */
3816 err
= check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
3817 BPF_SIZE(insn
->code
), BPF_WRITE
,
3822 prev_dst_type
= &env
->insn_aux_data
[insn_idx
].ptr_type
;
3824 if (*prev_dst_type
== NOT_INIT
) {
3825 *prev_dst_type
= dst_reg_type
;
3826 } else if (dst_reg_type
!= *prev_dst_type
&&
3827 (dst_reg_type
== PTR_TO_CTX
||
3828 *prev_dst_type
== PTR_TO_CTX
)) {
3829 verbose("same insn cannot be used with different pointers\n");
3833 } else if (class == BPF_ST
) {
3834 if (BPF_MODE(insn
->code
) != BPF_MEM
||
3835 insn
->src_reg
!= BPF_REG_0
) {
3836 verbose("BPF_ST uses reserved fields\n");
3839 /* check src operand */
3840 err
= check_reg_arg(env
, insn
->dst_reg
, SRC_OP
);
3844 /* check that memory (dst_reg + off) is writeable */
3845 err
= check_mem_access(env
, insn_idx
, insn
->dst_reg
, insn
->off
,
3846 BPF_SIZE(insn
->code
), BPF_WRITE
,
3851 } else if (class == BPF_JMP
) {
3852 u8 opcode
= BPF_OP(insn
->code
);
3854 if (opcode
== BPF_CALL
) {
3855 if (BPF_SRC(insn
->code
) != BPF_K
||
3857 insn
->src_reg
!= BPF_REG_0
||
3858 insn
->dst_reg
!= BPF_REG_0
) {
3859 verbose("BPF_CALL uses reserved fields\n");
3863 err
= check_call(env
, insn
->imm
, insn_idx
);
3867 } else if (opcode
== BPF_JA
) {
3868 if (BPF_SRC(insn
->code
) != BPF_K
||
3870 insn
->src_reg
!= BPF_REG_0
||
3871 insn
->dst_reg
!= BPF_REG_0
) {
3872 verbose("BPF_JA uses reserved fields\n");
3876 insn_idx
+= insn
->off
+ 1;
3879 } else if (opcode
== BPF_EXIT
) {
3880 if (BPF_SRC(insn
->code
) != BPF_K
||
3882 insn
->src_reg
!= BPF_REG_0
||
3883 insn
->dst_reg
!= BPF_REG_0
) {
3884 verbose("BPF_EXIT uses reserved fields\n");
3888 /* eBPF calling convetion is such that R0 is used
3889 * to return the value from eBPF program.
3890 * Make sure that it's readable at this time
3891 * of bpf_exit, which means that program wrote
3892 * something into it earlier
3894 err
= check_reg_arg(env
, BPF_REG_0
, SRC_OP
);
3898 if (is_pointer_value(env
, BPF_REG_0
)) {
3899 verbose("R0 leaks addr as return value\n");
3903 err
= check_return_code(env
);
3907 insn_idx
= pop_stack(env
, &prev_insn_idx
);
3911 do_print_state
= true;
3915 err
= check_cond_jmp_op(env
, insn
, &insn_idx
);
3919 } else if (class == BPF_LD
) {
3920 u8 mode
= BPF_MODE(insn
->code
);
3922 if (mode
== BPF_ABS
|| mode
== BPF_IND
) {
3923 err
= check_ld_abs(env
, insn
);
3927 } else if (mode
== BPF_IMM
) {
3928 err
= check_ld_imm(env
, insn
);
3934 verbose("invalid BPF_LD mode\n");
3938 verbose("unknown insn class %d\n", class);
3945 verbose("processed %d insns, stack depth %d\n",
3946 insn_processed
, env
->prog
->aux
->stack_depth
);
3950 static int check_map_prealloc(struct bpf_map
*map
)
3952 return (map
->map_type
!= BPF_MAP_TYPE_HASH
&&
3953 map
->map_type
!= BPF_MAP_TYPE_PERCPU_HASH
&&
3954 map
->map_type
!= BPF_MAP_TYPE_HASH_OF_MAPS
) ||
3955 !(map
->map_flags
& BPF_F_NO_PREALLOC
);
3958 static int check_map_prog_compatibility(struct bpf_map
*map
,
3959 struct bpf_prog
*prog
)
3962 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
3963 * preallocated hash maps, since doing memory allocation
3964 * in overflow_handler can crash depending on where nmi got
3967 if (prog
->type
== BPF_PROG_TYPE_PERF_EVENT
) {
3968 if (!check_map_prealloc(map
)) {
3969 verbose("perf_event programs can only use preallocated hash map\n");
3972 if (map
->inner_map_meta
&&
3973 !check_map_prealloc(map
->inner_map_meta
)) {
3974 verbose("perf_event programs can only use preallocated inner hash map\n");
3981 /* look for pseudo eBPF instructions that access map FDs and
3982 * replace them with actual map pointers
3984 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env
*env
)
3986 struct bpf_insn
*insn
= env
->prog
->insnsi
;
3987 int insn_cnt
= env
->prog
->len
;
3990 err
= bpf_prog_calc_tag(env
->prog
);
3994 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
3995 if (BPF_CLASS(insn
->code
) == BPF_LDX
&&
3996 (BPF_MODE(insn
->code
) != BPF_MEM
|| insn
->imm
!= 0)) {
3997 verbose("BPF_LDX uses reserved fields\n");
4001 if (BPF_CLASS(insn
->code
) == BPF_STX
&&
4002 ((BPF_MODE(insn
->code
) != BPF_MEM
&&
4003 BPF_MODE(insn
->code
) != BPF_XADD
) || insn
->imm
!= 0)) {
4004 verbose("BPF_STX uses reserved fields\n");
4008 if (insn
[0].code
== (BPF_LD
| BPF_IMM
| BPF_DW
)) {
4009 struct bpf_map
*map
;
4012 if (i
== insn_cnt
- 1 || insn
[1].code
!= 0 ||
4013 insn
[1].dst_reg
!= 0 || insn
[1].src_reg
!= 0 ||
4015 verbose("invalid bpf_ld_imm64 insn\n");
4019 if (insn
->src_reg
== 0)
4020 /* valid generic load 64-bit imm */
4023 if (insn
->src_reg
!= BPF_PSEUDO_MAP_FD
) {
4024 verbose("unrecognized bpf_ld_imm64 insn\n");
4028 f
= fdget(insn
->imm
);
4029 map
= __bpf_map_get(f
);
4031 verbose("fd %d is not pointing to valid bpf_map\n",
4033 return PTR_ERR(map
);
4036 err
= check_map_prog_compatibility(map
, env
->prog
);
4042 /* store map pointer inside BPF_LD_IMM64 instruction */
4043 insn
[0].imm
= (u32
) (unsigned long) map
;
4044 insn
[1].imm
= ((u64
) (unsigned long) map
) >> 32;
4046 /* check whether we recorded this map already */
4047 for (j
= 0; j
< env
->used_map_cnt
; j
++)
4048 if (env
->used_maps
[j
] == map
) {
4053 if (env
->used_map_cnt
>= MAX_USED_MAPS
) {
4058 /* hold the map. If the program is rejected by verifier,
4059 * the map will be released by release_maps() or it
4060 * will be used by the valid program until it's unloaded
4061 * and all maps are released in free_bpf_prog_info()
4063 map
= bpf_map_inc(map
, false);
4066 return PTR_ERR(map
);
4068 env
->used_maps
[env
->used_map_cnt
++] = map
;
4077 /* now all pseudo BPF_LD_IMM64 instructions load valid
4078 * 'struct bpf_map *' into a register instead of user map_fd.
4079 * These pointers will be used later by verifier to validate map access.
4084 /* drop refcnt of maps used by the rejected program */
4085 static void release_maps(struct bpf_verifier_env
*env
)
4089 for (i
= 0; i
< env
->used_map_cnt
; i
++)
4090 bpf_map_put(env
->used_maps
[i
]);
4093 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
4094 static void convert_pseudo_ld_imm64(struct bpf_verifier_env
*env
)
4096 struct bpf_insn
*insn
= env
->prog
->insnsi
;
4097 int insn_cnt
= env
->prog
->len
;
4100 for (i
= 0; i
< insn_cnt
; i
++, insn
++)
4101 if (insn
->code
== (BPF_LD
| BPF_IMM
| BPF_DW
))
4105 /* single env->prog->insni[off] instruction was replaced with the range
4106 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
4107 * [0, off) and [off, end) to new locations, so the patched range stays zero
4109 static int adjust_insn_aux_data(struct bpf_verifier_env
*env
, u32 prog_len
,
4112 struct bpf_insn_aux_data
*new_data
, *old_data
= env
->insn_aux_data
;
4116 new_data
= vzalloc(sizeof(struct bpf_insn_aux_data
) * prog_len
);
4119 memcpy(new_data
, old_data
, sizeof(struct bpf_insn_aux_data
) * off
);
4120 memcpy(new_data
+ off
+ cnt
- 1, old_data
+ off
,
4121 sizeof(struct bpf_insn_aux_data
) * (prog_len
- off
- cnt
+ 1));
4122 env
->insn_aux_data
= new_data
;
4127 static struct bpf_prog
*bpf_patch_insn_data(struct bpf_verifier_env
*env
, u32 off
,
4128 const struct bpf_insn
*patch
, u32 len
)
4130 struct bpf_prog
*new_prog
;
4132 new_prog
= bpf_patch_insn_single(env
->prog
, off
, patch
, len
);
4135 if (adjust_insn_aux_data(env
, new_prog
->len
, off
, len
))
4140 /* convert load instructions that access fields of 'struct __sk_buff'
4141 * into sequence of instructions that access fields of 'struct sk_buff'
4143 static int convert_ctx_accesses(struct bpf_verifier_env
*env
)
4145 const struct bpf_verifier_ops
*ops
= env
->prog
->aux
->ops
;
4146 int i
, cnt
, size
, ctx_field_size
, delta
= 0;
4147 const int insn_cnt
= env
->prog
->len
;
4148 struct bpf_insn insn_buf
[16], *insn
;
4149 struct bpf_prog
*new_prog
;
4150 enum bpf_access_type type
;
4151 bool is_narrower_load
;
4154 if (ops
->gen_prologue
) {
4155 cnt
= ops
->gen_prologue(insn_buf
, env
->seen_direct_write
,
4157 if (cnt
>= ARRAY_SIZE(insn_buf
)) {
4158 verbose("bpf verifier is misconfigured\n");
4161 new_prog
= bpf_patch_insn_data(env
, 0, insn_buf
, cnt
);
4165 env
->prog
= new_prog
;
4170 if (!ops
->convert_ctx_access
)
4173 insn
= env
->prog
->insnsi
+ delta
;
4175 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
4176 if (insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_B
) ||
4177 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_H
) ||
4178 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_W
) ||
4179 insn
->code
== (BPF_LDX
| BPF_MEM
| BPF_DW
))
4181 else if (insn
->code
== (BPF_STX
| BPF_MEM
| BPF_B
) ||
4182 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_H
) ||
4183 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_W
) ||
4184 insn
->code
== (BPF_STX
| BPF_MEM
| BPF_DW
))
4189 if (env
->insn_aux_data
[i
+ delta
].ptr_type
!= PTR_TO_CTX
)
4192 ctx_field_size
= env
->insn_aux_data
[i
+ delta
].ctx_field_size
;
4193 size
= BPF_LDST_BYTES(insn
);
4195 /* If the read access is a narrower load of the field,
4196 * convert to a 4/8-byte load, to minimum program type specific
4197 * convert_ctx_access changes. If conversion is successful,
4198 * we will apply proper mask to the result.
4200 is_narrower_load
= size
< ctx_field_size
;
4201 if (is_narrower_load
) {
4202 u32 off
= insn
->off
;
4205 if (type
== BPF_WRITE
) {
4206 verbose("bpf verifier narrow ctx access misconfigured\n");
4211 if (ctx_field_size
== 4)
4213 else if (ctx_field_size
== 8)
4216 insn
->off
= off
& ~(ctx_field_size
- 1);
4217 insn
->code
= BPF_LDX
| BPF_MEM
| size_code
;
4221 cnt
= ops
->convert_ctx_access(type
, insn
, insn_buf
, env
->prog
,
4223 if (cnt
== 0 || cnt
>= ARRAY_SIZE(insn_buf
) ||
4224 (ctx_field_size
&& !target_size
)) {
4225 verbose("bpf verifier is misconfigured\n");
4229 if (is_narrower_load
&& size
< target_size
) {
4230 if (ctx_field_size
<= 4)
4231 insn_buf
[cnt
++] = BPF_ALU32_IMM(BPF_AND
, insn
->dst_reg
,
4232 (1 << size
* 8) - 1);
4234 insn_buf
[cnt
++] = BPF_ALU64_IMM(BPF_AND
, insn
->dst_reg
,
4235 (1 << size
* 8) - 1);
4238 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
, cnt
);
4244 /* keep walking new program and skip insns we just inserted */
4245 env
->prog
= new_prog
;
4246 insn
= new_prog
->insnsi
+ i
+ delta
;
4252 /* fixup insn->imm field of bpf_call instructions
4253 * and inline eligible helpers as explicit sequence of BPF instructions
4255 * this function is called after eBPF program passed verification
4257 static int fixup_bpf_calls(struct bpf_verifier_env
*env
)
4259 struct bpf_prog
*prog
= env
->prog
;
4260 struct bpf_insn
*insn
= prog
->insnsi
;
4261 const struct bpf_func_proto
*fn
;
4262 const int insn_cnt
= prog
->len
;
4263 struct bpf_insn insn_buf
[16];
4264 struct bpf_prog
*new_prog
;
4265 struct bpf_map
*map_ptr
;
4266 int i
, cnt
, delta
= 0;
4268 for (i
= 0; i
< insn_cnt
; i
++, insn
++) {
4269 if (insn
->code
!= (BPF_JMP
| BPF_CALL
))
4272 if (insn
->imm
== BPF_FUNC_get_route_realm
)
4273 prog
->dst_needed
= 1;
4274 if (insn
->imm
== BPF_FUNC_get_prandom_u32
)
4275 bpf_user_rnd_init_once();
4276 if (insn
->imm
== BPF_FUNC_tail_call
) {
4277 /* If we tail call into other programs, we
4278 * cannot make any assumptions since they can
4279 * be replaced dynamically during runtime in
4280 * the program array.
4282 prog
->cb_access
= 1;
4283 env
->prog
->aux
->stack_depth
= MAX_BPF_STACK
;
4285 /* mark bpf_tail_call as different opcode to avoid
4286 * conditional branch in the interpeter for every normal
4287 * call and to prevent accidental JITing by JIT compiler
4288 * that doesn't support bpf_tail_call yet
4291 insn
->code
= BPF_JMP
| BPF_TAIL_CALL
;
4295 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
4296 * handlers are currently limited to 64 bit only.
4298 if (ebpf_jit_enabled() && BITS_PER_LONG
== 64 &&
4299 insn
->imm
== BPF_FUNC_map_lookup_elem
) {
4300 map_ptr
= env
->insn_aux_data
[i
+ delta
].map_ptr
;
4301 if (map_ptr
== BPF_MAP_PTR_POISON
||
4302 !map_ptr
->ops
->map_gen_lookup
)
4303 goto patch_call_imm
;
4305 cnt
= map_ptr
->ops
->map_gen_lookup(map_ptr
, insn_buf
);
4306 if (cnt
== 0 || cnt
>= ARRAY_SIZE(insn_buf
)) {
4307 verbose("bpf verifier is misconfigured\n");
4311 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, insn_buf
,
4318 /* keep walking new program and skip insns we just inserted */
4319 env
->prog
= prog
= new_prog
;
4320 insn
= new_prog
->insnsi
+ i
+ delta
;
4324 if (insn
->imm
== BPF_FUNC_redirect_map
) {
4325 /* Note, we cannot use prog directly as imm as subsequent
4326 * rewrites would still change the prog pointer. The only
4327 * stable address we can use is aux, which also works with
4328 * prog clones during blinding.
4330 u64 addr
= (unsigned long)prog
->aux
;
4331 struct bpf_insn r4_ld
[] = {
4332 BPF_LD_IMM64(BPF_REG_4
, addr
),
4335 cnt
= ARRAY_SIZE(r4_ld
);
4337 new_prog
= bpf_patch_insn_data(env
, i
+ delta
, r4_ld
, cnt
);
4342 env
->prog
= prog
= new_prog
;
4343 insn
= new_prog
->insnsi
+ i
+ delta
;
4346 fn
= prog
->aux
->ops
->get_func_proto(insn
->imm
);
4347 /* all functions that have prototype and verifier allowed
4348 * programs to call them, must be real in-kernel functions
4351 verbose("kernel subsystem misconfigured func %s#%d\n",
4352 func_id_name(insn
->imm
), insn
->imm
);
4355 insn
->imm
= fn
->func
- __bpf_call_base
;
4361 static void free_states(struct bpf_verifier_env
*env
)
4363 struct bpf_verifier_state_list
*sl
, *sln
;
4366 if (!env
->explored_states
)
4369 for (i
= 0; i
< env
->prog
->len
; i
++) {
4370 sl
= env
->explored_states
[i
];
4373 while (sl
!= STATE_LIST_MARK
) {
4380 kfree(env
->explored_states
);
4383 int bpf_check(struct bpf_prog
**prog
, union bpf_attr
*attr
)
4385 char __user
*log_ubuf
= NULL
;
4386 struct bpf_verifier_env
*env
;
4389 /* 'struct bpf_verifier_env' can be global, but since it's not small,
4390 * allocate/free it every time bpf_check() is called
4392 env
= kzalloc(sizeof(struct bpf_verifier_env
), GFP_KERNEL
);
4396 env
->insn_aux_data
= vzalloc(sizeof(struct bpf_insn_aux_data
) *
4399 if (!env
->insn_aux_data
)
4403 /* grab the mutex to protect few globals used by verifier */
4404 mutex_lock(&bpf_verifier_lock
);
4406 if (attr
->log_level
|| attr
->log_buf
|| attr
->log_size
) {
4407 /* user requested verbose verifier output
4408 * and supplied buffer to store the verification trace
4410 log_level
= attr
->log_level
;
4411 log_ubuf
= (char __user
*) (unsigned long) attr
->log_buf
;
4412 log_size
= attr
->log_size
;
4416 /* log_* values have to be sane */
4417 if (log_size
< 128 || log_size
> UINT_MAX
>> 8 ||
4418 log_level
== 0 || log_ubuf
== NULL
)
4422 log_buf
= vmalloc(log_size
);
4429 env
->strict_alignment
= !!(attr
->prog_flags
& BPF_F_STRICT_ALIGNMENT
);
4430 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
))
4431 env
->strict_alignment
= true;
4433 ret
= replace_map_fd_with_map_ptr(env
);
4435 goto skip_full_check
;
4437 env
->explored_states
= kcalloc(env
->prog
->len
,
4438 sizeof(struct bpf_verifier_state_list
*),
4441 if (!env
->explored_states
)
4442 goto skip_full_check
;
4444 ret
= check_cfg(env
);
4446 goto skip_full_check
;
4448 env
->allow_ptr_leaks
= capable(CAP_SYS_ADMIN
);
4450 ret
= do_check(env
);
4453 while (pop_stack(env
, NULL
) >= 0);
4457 /* program is valid, convert *(u32*)(ctx + off) accesses */
4458 ret
= convert_ctx_accesses(env
);
4461 ret
= fixup_bpf_calls(env
);
4463 if (log_level
&& log_len
>= log_size
- 1) {
4464 BUG_ON(log_len
>= log_size
);
4465 /* verifier log exceeded user supplied buffer */
4467 /* fall through to return what was recorded */
4470 /* copy verifier log back to user space including trailing zero */
4471 if (log_level
&& copy_to_user(log_ubuf
, log_buf
, log_len
+ 1) != 0) {
4476 if (ret
== 0 && env
->used_map_cnt
) {
4477 /* if program passed verifier, update used_maps in bpf_prog_info */
4478 env
->prog
->aux
->used_maps
= kmalloc_array(env
->used_map_cnt
,
4479 sizeof(env
->used_maps
[0]),
4482 if (!env
->prog
->aux
->used_maps
) {
4487 memcpy(env
->prog
->aux
->used_maps
, env
->used_maps
,
4488 sizeof(env
->used_maps
[0]) * env
->used_map_cnt
);
4489 env
->prog
->aux
->used_map_cnt
= env
->used_map_cnt
;
4491 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
4492 * bpf_ld_imm64 instructions
4494 convert_pseudo_ld_imm64(env
);
4500 if (!env
->prog
->aux
->used_maps
)
4501 /* if we didn't copy map pointers into bpf_prog_info, release
4502 * them now. Otherwise free_bpf_prog_info() will release them.
4507 mutex_unlock(&bpf_verifier_lock
);
4508 vfree(env
->insn_aux_data
);
4514 int bpf_analyzer(struct bpf_prog
*prog
, const struct bpf_ext_analyzer_ops
*ops
,
4517 struct bpf_verifier_env
*env
;
4520 env
= kzalloc(sizeof(struct bpf_verifier_env
), GFP_KERNEL
);
4524 env
->insn_aux_data
= vzalloc(sizeof(struct bpf_insn_aux_data
) *
4527 if (!env
->insn_aux_data
)
4530 env
->analyzer_ops
= ops
;
4531 env
->analyzer_priv
= priv
;
4533 /* grab the mutex to protect few globals used by verifier */
4534 mutex_lock(&bpf_verifier_lock
);
4538 env
->strict_alignment
= false;
4539 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
))
4540 env
->strict_alignment
= true;
4542 env
->explored_states
= kcalloc(env
->prog
->len
,
4543 sizeof(struct bpf_verifier_state_list
*),
4546 if (!env
->explored_states
)
4547 goto skip_full_check
;
4549 ret
= check_cfg(env
);
4551 goto skip_full_check
;
4553 env
->allow_ptr_leaks
= capable(CAP_SYS_ADMIN
);
4555 ret
= do_check(env
);
4558 while (pop_stack(env
, NULL
) >= 0);
4561 mutex_unlock(&bpf_verifier_lock
);
4562 vfree(env
->insn_aux_data
);
4567 EXPORT_SYMBOL_GPL(bpf_analyzer
);