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 /* string representation of 'enum bpf_reg_type' */
181 static const char * const reg_type_str[] = {
183 [SCALAR_VALUE] = "inv",
184 [PTR_TO_CTX] = "ctx",
185 [CONST_PTR_TO_MAP] = "map_ptr",
186 [PTR_TO_MAP_VALUE] = "map_value",
187 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
188 [PTR_TO_STACK] = "fp",
189 [PTR_TO_PACKET] = "pkt",
190 [PTR_TO_PACKET_END] = "pkt_end",
193 #define __BPF_FUNC_STR_FN(x) [BPF_FUNC_ ## x] = __stringify(bpf_ ## x)
194 static const char * const func_id_str[] = {
195 __BPF_FUNC_MAPPER(__BPF_FUNC_STR_FN)
197 #undef __BPF_FUNC_STR_FN
199 static const char *func_id_name(int id)
201 BUILD_BUG_ON(ARRAY_SIZE(func_id_str) != __BPF_FUNC_MAX_ID);
203 if (id >= 0 && id < __BPF_FUNC_MAX_ID && func_id_str[id])
204 return func_id_str[id];
209 static void print_verifier_state(struct bpf_verifier_state *state)
211 struct bpf_reg_state *reg;
215 for (i = 0; i < MAX_BPF_REG; i++) {
216 reg = &state->regs[i];
220 verbose(" R%d=%s", i, reg_type_str[t]);
221 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
222 tnum_is_const(reg->var_off)) {
223 /* reg->off should be 0 for SCALAR_VALUE */
224 verbose("%lld", reg->var_off.value + reg->off);
226 verbose("(id=%d", reg->id);
227 if (t != SCALAR_VALUE)
228 verbose(",off=%d", reg->off);
229 if (t == PTR_TO_PACKET)
230 verbose(",r=%d", reg->range);
231 else if (t == CONST_PTR_TO_MAP ||
232 t == PTR_TO_MAP_VALUE ||
233 t == PTR_TO_MAP_VALUE_OR_NULL)
234 verbose(",ks=%d,vs=%d",
235 reg->map_ptr->key_size,
236 reg->map_ptr->value_size);
237 if (tnum_is_const(reg->var_off)) {
238 /* Typically an immediate SCALAR_VALUE, but
239 * could be a pointer whose offset is too big
242 verbose(",imm=%llx", reg->var_off.value);
244 if (reg->smin_value != reg->umin_value &&
245 reg->smin_value != S64_MIN)
246 verbose(",smin_value=%lld",
247 (long long)reg->smin_value);
248 if (reg->smax_value != reg->umax_value &&
249 reg->smax_value != S64_MAX)
250 verbose(",smax_value=%lld",
251 (long long)reg->smax_value);
252 if (reg->umin_value != 0)
253 verbose(",umin_value=%llu",
254 (unsigned long long)reg->umin_value);
255 if (reg->umax_value != U64_MAX)
256 verbose(",umax_value=%llu",
257 (unsigned long long)reg->umax_value);
258 if (!tnum_is_unknown(reg->var_off)) {
261 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
262 verbose(",var_off=%s", tn_buf);
268 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
269 if (state->stack_slot_type[i] == STACK_SPILL)
270 verbose(" fp%d=%s", -MAX_BPF_STACK + i,
271 reg_type_str[state->spilled_regs[i / BPF_REG_SIZE].type]);
276 static const char *const bpf_class_string[] = {
284 [BPF_ALU64] = "alu64",
287 static const char *const bpf_alu_string[16] = {
288 [BPF_ADD >> 4] = "+=",
289 [BPF_SUB >> 4] = "-=",
290 [BPF_MUL >> 4] = "*=",
291 [BPF_DIV >> 4] = "/=",
292 [BPF_OR >> 4] = "|=",
293 [BPF_AND >> 4] = "&=",
294 [BPF_LSH >> 4] = "<<=",
295 [BPF_RSH >> 4] = ">>=",
296 [BPF_NEG >> 4] = "neg",
297 [BPF_MOD >> 4] = "%=",
298 [BPF_XOR >> 4] = "^=",
299 [BPF_MOV >> 4] = "=",
300 [BPF_ARSH >> 4] = "s>>=",
301 [BPF_END >> 4] = "endian",
304 static const char *const bpf_ldst_string[] = {
305 [BPF_W >> 3] = "u32",
306 [BPF_H >> 3] = "u16",
308 [BPF_DW >> 3] = "u64",
311 static const char *const bpf_jmp_string[16] = {
312 [BPF_JA >> 4] = "jmp",
313 [BPF_JEQ >> 4] = "==",
314 [BPF_JGT >> 4] = ">",
315 [BPF_JLT >> 4] = "<",
316 [BPF_JGE >> 4] = ">=",
317 [BPF_JLE >> 4] = "<=",
318 [BPF_JSET >> 4] = "&",
319 [BPF_JNE >> 4] = "!=",
320 [BPF_JSGT >> 4] = "s>",
321 [BPF_JSLT >> 4] = "s<",
322 [BPF_JSGE >> 4] = "s>=",
323 [BPF_JSLE >> 4] = "s<=",
324 [BPF_CALL >> 4] = "call",
325 [BPF_EXIT >> 4] = "exit",
328 static void print_bpf_insn(const struct bpf_verifier_env *env,
329 const struct bpf_insn *insn)
331 u8 class = BPF_CLASS(insn->code);
333 if (class == BPF_ALU || class == BPF_ALU64) {
334 if (BPF_SRC(insn->code) == BPF_X)
335 verbose("(%02x) %sr%d %s %sr%d\n",
336 insn->code, class == BPF_ALU ? "(u32) " : "",
338 bpf_alu_string[BPF_OP(insn->code) >> 4],
339 class == BPF_ALU ? "(u32) " : "",
342 verbose("(%02x) %sr%d %s %s%d\n",
343 insn->code, class == BPF_ALU ? "(u32) " : "",
345 bpf_alu_string[BPF_OP(insn->code) >> 4],
346 class == BPF_ALU ? "(u32) " : "",
348 } else if (class == BPF_STX) {
349 if (BPF_MODE(insn->code) == BPF_MEM)
350 verbose("(%02x) *(%s *)(r%d %+d) = r%d\n",
352 bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
354 insn->off, insn->src_reg);
355 else if (BPF_MODE(insn->code) == BPF_XADD)
356 verbose("(%02x) lock *(%s *)(r%d %+d) += r%d\n",
358 bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
359 insn->dst_reg, insn->off,
362 verbose("BUG_%02x\n", insn->code);
363 } else if (class == BPF_ST) {
364 if (BPF_MODE(insn->code) != BPF_MEM) {
365 verbose("BUG_st_%02x\n", insn->code);
368 verbose("(%02x) *(%s *)(r%d %+d) = %d\n",
370 bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
372 insn->off, insn->imm);
373 } else if (class == BPF_LDX) {
374 if (BPF_MODE(insn->code) != BPF_MEM) {
375 verbose("BUG_ldx_%02x\n", insn->code);
378 verbose("(%02x) r%d = *(%s *)(r%d %+d)\n",
379 insn->code, insn->dst_reg,
380 bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
381 insn->src_reg, insn->off);
382 } else if (class == BPF_LD) {
383 if (BPF_MODE(insn->code) == BPF_ABS) {
384 verbose("(%02x) r0 = *(%s *)skb[%d]\n",
386 bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
388 } else if (BPF_MODE(insn->code) == BPF_IND) {
389 verbose("(%02x) r0 = *(%s *)skb[r%d + %d]\n",
391 bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
392 insn->src_reg, insn->imm);
393 } else if (BPF_MODE(insn->code) == BPF_IMM &&
394 BPF_SIZE(insn->code) == BPF_DW) {
395 /* At this point, we already made sure that the second
396 * part of the ldimm64 insn is accessible.
398 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
399 bool map_ptr = insn->src_reg == BPF_PSEUDO_MAP_FD;
401 if (map_ptr && !env->allow_ptr_leaks)
404 verbose("(%02x) r%d = 0x%llx\n", insn->code,
405 insn->dst_reg, (unsigned long long)imm);
407 verbose("BUG_ld_%02x\n", insn->code);
410 } else if (class == BPF_JMP) {
411 u8 opcode = BPF_OP(insn->code);
413 if (opcode == BPF_CALL) {
414 verbose("(%02x) call %s#%d\n", insn->code,
415 func_id_name(insn->imm), insn->imm);
416 } else if (insn->code == (BPF_JMP | BPF_JA)) {
417 verbose("(%02x) goto pc%+d\n",
418 insn->code, insn->off);
419 } else if (insn->code == (BPF_JMP | BPF_EXIT)) {
420 verbose("(%02x) exit\n", insn->code);
421 } else if (BPF_SRC(insn->code) == BPF_X) {
422 verbose("(%02x) if r%d %s r%d goto pc%+d\n",
423 insn->code, insn->dst_reg,
424 bpf_jmp_string[BPF_OP(insn->code) >> 4],
425 insn->src_reg, insn->off);
427 verbose("(%02x) if r%d %s 0x%x goto pc%+d\n",
428 insn->code, insn->dst_reg,
429 bpf_jmp_string[BPF_OP(insn->code) >> 4],
430 insn->imm, insn->off);
433 verbose("(%02x) %s\n", insn->code, bpf_class_string[class]);
437 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx)
439 struct bpf_verifier_stack_elem *elem;
442 if (env->head == NULL)
445 memcpy(&env->cur_state, &env->head->st, sizeof(env->cur_state));
446 insn_idx = env->head->insn_idx;
448 *prev_insn_idx = env->head->prev_insn_idx;
449 elem = env->head->next;
456 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
457 int insn_idx, int prev_insn_idx)
459 struct bpf_verifier_stack_elem *elem;
461 elem = kmalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
465 memcpy(&elem->st, &env->cur_state, sizeof(env->cur_state));
466 elem->insn_idx = insn_idx;
467 elem->prev_insn_idx = prev_insn_idx;
468 elem->next = env->head;
471 if (env->stack_size > BPF_COMPLEXITY_LIMIT_STACK) {
472 verbose("BPF program is too complex\n");
477 /* pop all elements and return */
478 while (pop_stack(env, NULL) >= 0);
482 #define CALLER_SAVED_REGS 6
483 static const int caller_saved[CALLER_SAVED_REGS] = {
484 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
487 static void __mark_reg_not_init(struct bpf_reg_state *reg);
489 /* Mark the unknown part of a register (variable offset or scalar value) as
490 * known to have the value @imm.
492 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
495 reg->var_off = tnum_const(imm);
496 reg->smin_value = (s64)imm;
497 reg->smax_value = (s64)imm;
498 reg->umin_value = imm;
499 reg->umax_value = imm;
502 /* Mark the 'variable offset' part of a register as zero. This should be
503 * used only on registers holding a pointer type.
505 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
507 __mark_reg_known(reg, 0);
510 static void mark_reg_known_zero(struct bpf_reg_state *regs, u32 regno)
512 if (WARN_ON(regno >= MAX_BPF_REG)) {
513 verbose("mark_reg_known_zero(regs, %u)\n", regno);
514 /* Something bad happened, let's kill all regs */
515 for (regno = 0; regno < MAX_BPF_REG; regno++)
516 __mark_reg_not_init(regs + regno);
519 __mark_reg_known_zero(regs + regno);
522 /* Attempts to improve min/max values based on var_off information */
523 static void __update_reg_bounds(struct bpf_reg_state *reg)
525 /* min signed is max(sign bit) | min(other bits) */
526 reg->smin_value = max_t(s64, reg->smin_value,
527 reg->var_off.value | (reg->var_off.mask & S64_MIN));
528 /* max signed is min(sign bit) | max(other bits) */
529 reg->smax_value = min_t(s64, reg->smax_value,
530 reg->var_off.value | (reg->var_off.mask & S64_MAX));
531 reg->umin_value = max(reg->umin_value, reg->var_off.value);
532 reg->umax_value = min(reg->umax_value,
533 reg->var_off.value | reg->var_off.mask);
536 /* Uses signed min/max values to inform unsigned, and vice-versa */
537 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
539 /* Learn sign from signed bounds.
540 * If we cannot cross the sign boundary, then signed and unsigned bounds
541 * are the same, so combine. This works even in the negative case, e.g.
542 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
544 if (reg->smin_value >= 0 || reg->smax_value < 0) {
545 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
547 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
551 /* Learn sign from unsigned bounds. Signed bounds cross the sign
552 * boundary, so we must be careful.
554 if ((s64)reg->umax_value >= 0) {
555 /* Positive. We can't learn anything from the smin, but smax
556 * is positive, hence safe.
558 reg->smin_value = reg->umin_value;
559 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
561 } else if ((s64)reg->umin_value < 0) {
562 /* Negative. We can't learn anything from the smax, but smin
563 * is negative, hence safe.
565 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
567 reg->smax_value = reg->umax_value;
571 /* Attempts to improve var_off based on unsigned min/max information */
572 static void __reg_bound_offset(struct bpf_reg_state *reg)
574 reg->var_off = tnum_intersect(reg->var_off,
575 tnum_range(reg->umin_value,
579 /* Reset the min/max bounds of a register */
580 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
582 reg->smin_value = S64_MIN;
583 reg->smax_value = S64_MAX;
585 reg->umax_value = U64_MAX;
588 /* Mark a register as having a completely unknown (scalar) value. */
589 static void __mark_reg_unknown(struct bpf_reg_state *reg)
591 reg->type = SCALAR_VALUE;
594 reg->var_off = tnum_unknown;
595 __mark_reg_unbounded(reg);
598 static void mark_reg_unknown(struct bpf_reg_state *regs, u32 regno)
600 if (WARN_ON(regno >= MAX_BPF_REG)) {
601 verbose("mark_reg_unknown(regs, %u)\n", regno);
602 /* Something bad happened, let's kill all regs */
603 for (regno = 0; regno < MAX_BPF_REG; regno++)
604 __mark_reg_not_init(regs + regno);
607 __mark_reg_unknown(regs + regno);
610 static void __mark_reg_not_init(struct bpf_reg_state *reg)
612 __mark_reg_unknown(reg);
613 reg->type = NOT_INIT;
616 static void mark_reg_not_init(struct bpf_reg_state *regs, u32 regno)
618 if (WARN_ON(regno >= MAX_BPF_REG)) {
619 verbose("mark_reg_not_init(regs, %u)\n", regno);
620 /* Something bad happened, let's kill all regs */
621 for (regno = 0; regno < MAX_BPF_REG; regno++)
622 __mark_reg_not_init(regs + regno);
625 __mark_reg_not_init(regs + regno);
628 static void init_reg_state(struct bpf_reg_state *regs)
632 for (i = 0; i < MAX_BPF_REG; i++) {
633 mark_reg_not_init(regs, i);
634 regs[i].live = REG_LIVE_NONE;
638 regs[BPF_REG_FP].type = PTR_TO_STACK;
639 mark_reg_known_zero(regs, BPF_REG_FP);
641 /* 1st arg to a function */
642 regs[BPF_REG_1].type = PTR_TO_CTX;
643 mark_reg_known_zero(regs, BPF_REG_1);
647 SRC_OP, /* register is used as source operand */
648 DST_OP, /* register is used as destination operand */
649 DST_OP_NO_MARK /* same as above, check only, don't mark */
652 static void mark_reg_read(const struct bpf_verifier_state *state, u32 regno)
654 struct bpf_verifier_state *parent = state->parent;
656 if (regno == BPF_REG_FP)
657 /* We don't need to worry about FP liveness because it's read-only */
661 /* if read wasn't screened by an earlier write ... */
662 if (state->regs[regno].live & REG_LIVE_WRITTEN)
664 /* ... then we depend on parent's value */
665 parent->regs[regno].live |= REG_LIVE_READ;
667 parent = state->parent;
671 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
674 struct bpf_reg_state *regs = env->cur_state.regs;
676 if (regno >= MAX_BPF_REG) {
677 verbose("R%d is invalid\n", regno);
682 /* check whether register used as source operand can be read */
683 if (regs[regno].type == NOT_INIT) {
684 verbose("R%d !read_ok\n", regno);
687 mark_reg_read(&env->cur_state, regno);
689 /* check whether register used as dest operand can be written to */
690 if (regno == BPF_REG_FP) {
691 verbose("frame pointer is read only\n");
694 regs[regno].live |= REG_LIVE_WRITTEN;
696 mark_reg_unknown(regs, regno);
701 static bool is_spillable_regtype(enum bpf_reg_type type)
704 case PTR_TO_MAP_VALUE:
705 case PTR_TO_MAP_VALUE_OR_NULL:
709 case PTR_TO_PACKET_END:
710 case CONST_PTR_TO_MAP:
717 /* check_stack_read/write functions track spill/fill of registers,
718 * stack boundary and alignment are checked in check_mem_access()
720 static int check_stack_write(struct bpf_verifier_state *state, int off,
721 int size, int value_regno)
723 int i, spi = (MAX_BPF_STACK + off) / BPF_REG_SIZE;
724 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
725 * so it's aligned access and [off, off + size) are within stack limits
728 if (value_regno >= 0 &&
729 is_spillable_regtype(state->regs[value_regno].type)) {
731 /* register containing pointer is being spilled into stack */
732 if (size != BPF_REG_SIZE) {
733 verbose("invalid size of register spill\n");
737 /* save register state */
738 state->spilled_regs[spi] = state->regs[value_regno];
739 state->spilled_regs[spi].live |= REG_LIVE_WRITTEN;
741 for (i = 0; i < BPF_REG_SIZE; i++)
742 state->stack_slot_type[MAX_BPF_STACK + off + i] = STACK_SPILL;
744 /* regular write of data into stack */
745 state->spilled_regs[spi] = (struct bpf_reg_state) {};
747 for (i = 0; i < size; i++)
748 state->stack_slot_type[MAX_BPF_STACK + off + i] = STACK_MISC;
753 static void mark_stack_slot_read(const struct bpf_verifier_state *state, int slot)
755 struct bpf_verifier_state *parent = state->parent;
758 /* if read wasn't screened by an earlier write ... */
759 if (state->spilled_regs[slot].live & REG_LIVE_WRITTEN)
761 /* ... then we depend on parent's value */
762 parent->spilled_regs[slot].live |= REG_LIVE_READ;
764 parent = state->parent;
768 static int check_stack_read(struct bpf_verifier_state *state, int off, int size,
774 slot_type = &state->stack_slot_type[MAX_BPF_STACK + off];
776 if (slot_type[0] == STACK_SPILL) {
777 if (size != BPF_REG_SIZE) {
778 verbose("invalid size of register spill\n");
781 for (i = 1; i < BPF_REG_SIZE; i++) {
782 if (slot_type[i] != STACK_SPILL) {
783 verbose("corrupted spill memory\n");
788 spi = (MAX_BPF_STACK + off) / BPF_REG_SIZE;
790 if (value_regno >= 0) {
791 /* restore register state from stack */
792 state->regs[value_regno] = state->spilled_regs[spi];
793 mark_stack_slot_read(state, spi);
797 for (i = 0; i < size; i++) {
798 if (slot_type[i] != STACK_MISC) {
799 verbose("invalid read from stack off %d+%d size %d\n",
804 if (value_regno >= 0)
805 /* have read misc data from the stack */
806 mark_reg_unknown(state->regs, value_regno);
811 /* check read/write into map element returned by bpf_map_lookup_elem() */
812 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
815 struct bpf_map *map = env->cur_state.regs[regno].map_ptr;
817 if (off < 0 || size <= 0 || off + size > map->value_size) {
818 verbose("invalid access to map value, value_size=%d off=%d size=%d\n",
819 map->value_size, off, size);
825 /* check read/write into a map element with possible variable offset */
826 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
829 struct bpf_verifier_state *state = &env->cur_state;
830 struct bpf_reg_state *reg = &state->regs[regno];
833 /* We may have adjusted the register to this map value, so we
834 * need to try adding each of min_value and max_value to off
835 * to make sure our theoretical access will be safe.
838 print_verifier_state(state);
839 /* The minimum value is only important with signed
840 * comparisons where we can't assume the floor of a
841 * value is 0. If we are using signed variables for our
842 * index'es we need to make sure that whatever we use
843 * will have a set floor within our range.
845 if (reg->smin_value < 0) {
846 verbose("R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
850 err = __check_map_access(env, regno, reg->smin_value + off, size);
852 verbose("R%d min value is outside of the array range\n", regno);
856 /* If we haven't set a max value then we need to bail since we can't be
857 * sure we won't do bad things.
858 * If reg->umax_value + off could overflow, treat that as unbounded too.
860 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
861 verbose("R%d unbounded memory access, make sure to bounds check any array access into a map\n",
865 err = __check_map_access(env, regno, reg->umax_value + off, size);
867 verbose("R%d max value is outside of the array range\n", regno);
871 #define MAX_PACKET_OFF 0xffff
873 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
874 const struct bpf_call_arg_meta *meta,
875 enum bpf_access_type t)
877 switch (env->prog->type) {
878 case BPF_PROG_TYPE_LWT_IN:
879 case BPF_PROG_TYPE_LWT_OUT:
880 /* dst_input() and dst_output() can't write for now */
884 case BPF_PROG_TYPE_SCHED_CLS:
885 case BPF_PROG_TYPE_SCHED_ACT:
886 case BPF_PROG_TYPE_XDP:
887 case BPF_PROG_TYPE_LWT_XMIT:
888 case BPF_PROG_TYPE_SK_SKB:
890 return meta->pkt_access;
892 env->seen_direct_write = true;
899 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
902 struct bpf_reg_state *regs = env->cur_state.regs;
903 struct bpf_reg_state *reg = ®s[regno];
905 if (off < 0 || size <= 0 || (u64)off + size > reg->range) {
906 verbose("invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
907 off, size, regno, reg->id, reg->off, reg->range);
913 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
916 struct bpf_reg_state *regs = env->cur_state.regs;
917 struct bpf_reg_state *reg = ®s[regno];
920 /* We may have added a variable offset to the packet pointer; but any
921 * reg->range we have comes after that. We are only checking the fixed
925 /* We don't allow negative numbers, because we aren't tracking enough
926 * detail to prove they're safe.
928 if (reg->smin_value < 0) {
929 verbose("R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
933 err = __check_packet_access(env, regno, off, size);
935 verbose("R%d offset is outside of the packet\n", regno);
941 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
942 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
943 enum bpf_access_type t, enum bpf_reg_type *reg_type)
945 struct bpf_insn_access_aux info = {
946 .reg_type = *reg_type,
949 /* for analyzer ctx accesses are already validated and converted */
950 if (env->analyzer_ops)
953 if (env->prog->aux->ops->is_valid_access &&
954 env->prog->aux->ops->is_valid_access(off, size, t, &info)) {
955 /* A non zero info.ctx_field_size indicates that this field is a
956 * candidate for later verifier transformation to load the whole
957 * field and then apply a mask when accessed with a narrower
958 * access than actual ctx access size. A zero info.ctx_field_size
959 * will only allow for whole field access and rejects any other
960 * type of narrower access.
962 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
963 *reg_type = info.reg_type;
965 /* remember the offset of last byte accessed in ctx */
966 if (env->prog->aux->max_ctx_offset < off + size)
967 env->prog->aux->max_ctx_offset = off + size;
971 verbose("invalid bpf_context access off=%d size=%d\n", off, size);
975 static bool __is_pointer_value(bool allow_ptr_leaks,
976 const struct bpf_reg_state *reg)
981 return reg->type != SCALAR_VALUE;
984 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
986 return __is_pointer_value(env->allow_ptr_leaks, &env->cur_state.regs[regno]);
989 static int check_pkt_ptr_alignment(const struct bpf_reg_state *reg,
990 int off, int size, bool strict)
995 /* Byte size accesses are always allowed. */
996 if (!strict || size == 1)
999 /* For platforms that do not have a Kconfig enabling
1000 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
1001 * NET_IP_ALIGN is universally set to '2'. And on platforms
1002 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
1003 * to this code only in strict mode where we want to emulate
1004 * the NET_IP_ALIGN==2 checking. Therefore use an
1005 * unconditional IP align value of '2'.
1009 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
1010 if (!tnum_is_aligned(reg_off, size)) {
1013 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1014 verbose("misaligned packet access off %d+%s+%d+%d size %d\n",
1015 ip_align, tn_buf, reg->off, off, size);
1022 static int check_generic_ptr_alignment(const struct bpf_reg_state *reg,
1023 const char *pointer_desc,
1024 int off, int size, bool strict)
1026 struct tnum reg_off;
1028 /* Byte size accesses are always allowed. */
1029 if (!strict || size == 1)
1032 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
1033 if (!tnum_is_aligned(reg_off, size)) {
1036 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1037 verbose("misaligned %saccess off %s+%d+%d size %d\n",
1038 pointer_desc, tn_buf, reg->off, off, size);
1045 static int check_ptr_alignment(struct bpf_verifier_env *env,
1046 const struct bpf_reg_state *reg,
1049 bool strict = env->strict_alignment;
1050 const char *pointer_desc = "";
1052 switch (reg->type) {
1054 /* special case, because of NET_IP_ALIGN */
1055 return check_pkt_ptr_alignment(reg, off, size, strict);
1056 case PTR_TO_MAP_VALUE:
1057 pointer_desc = "value ";
1060 pointer_desc = "context ";
1063 pointer_desc = "stack ";
1064 /* The stack spill tracking logic in check_stack_write()
1065 * and check_stack_read() relies on stack accesses being
1073 return check_generic_ptr_alignment(reg, pointer_desc, off, size, strict);
1076 /* truncate register to smaller size (in bytes)
1077 * must be called with size < BPF_REG_SIZE
1079 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
1083 /* clear high bits in bit representation */
1084 reg->var_off = tnum_cast(reg->var_off, size);
1086 /* fix arithmetic bounds */
1087 mask = ((u64)1 << (size * 8)) - 1;
1088 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
1089 reg->umin_value &= mask;
1090 reg->umax_value &= mask;
1092 reg->umin_value = 0;
1093 reg->umax_value = mask;
1095 reg->smin_value = reg->umin_value;
1096 reg->smax_value = reg->umax_value;
1099 /* check whether memory at (regno + off) is accessible for t = (read | write)
1100 * if t==write, value_regno is a register which value is stored into memory
1101 * if t==read, value_regno is a register which will receive the value from memory
1102 * if t==write && value_regno==-1, some unknown value is stored into memory
1103 * if t==read && value_regno==-1, don't care what we read from memory
1105 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno, int off,
1106 int bpf_size, enum bpf_access_type t,
1109 struct bpf_verifier_state *state = &env->cur_state;
1110 struct bpf_reg_state *reg = &state->regs[regno];
1113 size = bpf_size_to_bytes(bpf_size);
1117 /* alignment checks will add in reg->off themselves */
1118 err = check_ptr_alignment(env, reg, off, size);
1122 /* for access checks, reg->off is just part of off */
1125 if (reg->type == PTR_TO_MAP_VALUE) {
1126 if (t == BPF_WRITE && value_regno >= 0 &&
1127 is_pointer_value(env, value_regno)) {
1128 verbose("R%d leaks addr into map\n", value_regno);
1132 err = check_map_access(env, regno, off, size);
1133 if (!err && t == BPF_READ && value_regno >= 0)
1134 mark_reg_unknown(state->regs, value_regno);
1136 } else if (reg->type == PTR_TO_CTX) {
1137 enum bpf_reg_type reg_type = SCALAR_VALUE;
1139 if (t == BPF_WRITE && value_regno >= 0 &&
1140 is_pointer_value(env, value_regno)) {
1141 verbose("R%d leaks addr into ctx\n", value_regno);
1144 /* ctx accesses must be at a fixed offset, so that we can
1145 * determine what type of data were returned.
1148 verbose("dereference of modified ctx ptr R%d off=%d+%d, ctx+const is allowed, ctx+const+const is not\n",
1149 regno, reg->off, off - reg->off);
1152 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
1155 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1156 verbose("variable ctx access var_off=%s off=%d size=%d",
1160 err = check_ctx_access(env, insn_idx, off, size, t, ®_type);
1161 if (!err && t == BPF_READ && value_regno >= 0) {
1162 /* ctx access returns either a scalar, or a
1163 * PTR_TO_PACKET[_END]. In the latter case, we know
1164 * the offset is zero.
1166 if (reg_type == SCALAR_VALUE)
1167 mark_reg_unknown(state->regs, value_regno);
1169 mark_reg_known_zero(state->regs, value_regno);
1170 state->regs[value_regno].id = 0;
1171 state->regs[value_regno].off = 0;
1172 state->regs[value_regno].range = 0;
1173 state->regs[value_regno].type = reg_type;
1176 } else if (reg->type == PTR_TO_STACK) {
1177 /* stack accesses must be at a fixed offset, so that we can
1178 * determine what type of data were returned.
1179 * See check_stack_read().
1181 if (!tnum_is_const(reg->var_off)) {
1184 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1185 verbose("variable stack access var_off=%s off=%d size=%d",
1189 off += reg->var_off.value;
1190 if (off >= 0 || off < -MAX_BPF_STACK) {
1191 verbose("invalid stack off=%d size=%d\n", off, size);
1195 if (env->prog->aux->stack_depth < -off)
1196 env->prog->aux->stack_depth = -off;
1198 if (t == BPF_WRITE) {
1199 if (!env->allow_ptr_leaks &&
1200 state->stack_slot_type[MAX_BPF_STACK + off] == STACK_SPILL &&
1201 size != BPF_REG_SIZE) {
1202 verbose("attempt to corrupt spilled pointer on stack\n");
1205 err = check_stack_write(state, off, size, value_regno);
1207 err = check_stack_read(state, off, size, value_regno);
1209 } else if (reg->type == PTR_TO_PACKET) {
1210 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
1211 verbose("cannot write into packet\n");
1214 if (t == BPF_WRITE && value_regno >= 0 &&
1215 is_pointer_value(env, value_regno)) {
1216 verbose("R%d leaks addr into packet\n", value_regno);
1219 err = check_packet_access(env, regno, off, size);
1220 if (!err && t == BPF_READ && value_regno >= 0)
1221 mark_reg_unknown(state->regs, value_regno);
1223 verbose("R%d invalid mem access '%s'\n",
1224 regno, reg_type_str[reg->type]);
1228 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
1229 state->regs[value_regno].type == SCALAR_VALUE) {
1230 /* b/h/w load zero-extends, mark upper bits as known 0 */
1231 coerce_reg_to_size(&state->regs[value_regno], size);
1236 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
1240 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
1242 verbose("BPF_XADD uses reserved fields\n");
1246 /* check src1 operand */
1247 err = check_reg_arg(env, insn->src_reg, SRC_OP);
1251 /* check src2 operand */
1252 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
1256 if (is_pointer_value(env, insn->src_reg)) {
1257 verbose("R%d leaks addr into mem\n", insn->src_reg);
1261 /* check whether atomic_add can read the memory */
1262 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1263 BPF_SIZE(insn->code), BPF_READ, -1);
1267 /* check whether atomic_add can write into the same memory */
1268 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1269 BPF_SIZE(insn->code), BPF_WRITE, -1);
1272 /* Does this register contain a constant zero? */
1273 static bool register_is_null(struct bpf_reg_state reg)
1275 return reg.type == SCALAR_VALUE && tnum_equals_const(reg.var_off, 0);
1278 /* when register 'regno' is passed into function that will read 'access_size'
1279 * bytes from that pointer, make sure that it's within stack boundary
1280 * and all elements of stack are initialized.
1281 * Unlike most pointer bounds-checking functions, this one doesn't take an
1282 * 'off' argument, so it has to add in reg->off itself.
1284 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
1285 int access_size, bool zero_size_allowed,
1286 struct bpf_call_arg_meta *meta)
1288 struct bpf_verifier_state *state = &env->cur_state;
1289 struct bpf_reg_state *regs = state->regs;
1292 if (regs[regno].type != PTR_TO_STACK) {
1293 /* Allow zero-byte read from NULL, regardless of pointer type */
1294 if (zero_size_allowed && access_size == 0 &&
1295 register_is_null(regs[regno]))
1298 verbose("R%d type=%s expected=%s\n", regno,
1299 reg_type_str[regs[regno].type],
1300 reg_type_str[PTR_TO_STACK]);
1304 /* Only allow fixed-offset stack reads */
1305 if (!tnum_is_const(regs[regno].var_off)) {
1308 tnum_strn(tn_buf, sizeof(tn_buf), regs[regno].var_off);
1309 verbose("invalid variable stack read R%d var_off=%s\n",
1313 off = regs[regno].off + regs[regno].var_off.value;
1314 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
1316 verbose("invalid stack type R%d off=%d access_size=%d\n",
1317 regno, off, access_size);
1321 if (env->prog->aux->stack_depth < -off)
1322 env->prog->aux->stack_depth = -off;
1324 if (meta && meta->raw_mode) {
1325 meta->access_size = access_size;
1326 meta->regno = regno;
1330 for (i = 0; i < access_size; i++) {
1331 if (state->stack_slot_type[MAX_BPF_STACK + off + i] != STACK_MISC) {
1332 verbose("invalid indirect read from stack off %d+%d size %d\n",
1333 off, i, access_size);
1340 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
1341 int access_size, bool zero_size_allowed,
1342 struct bpf_call_arg_meta *meta)
1344 struct bpf_reg_state *regs = env->cur_state.regs, *reg = ®s[regno];
1346 switch (reg->type) {
1348 return check_packet_access(env, regno, reg->off, access_size);
1349 case PTR_TO_MAP_VALUE:
1350 return check_map_access(env, regno, reg->off, access_size);
1351 default: /* scalar_value|ptr_to_stack or invalid ptr */
1352 return check_stack_boundary(env, regno, access_size,
1353 zero_size_allowed, meta);
1357 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
1358 enum bpf_arg_type arg_type,
1359 struct bpf_call_arg_meta *meta)
1361 struct bpf_reg_state *regs = env->cur_state.regs, *reg = ®s[regno];
1362 enum bpf_reg_type expected_type, type = reg->type;
1365 if (arg_type == ARG_DONTCARE)
1368 err = check_reg_arg(env, regno, SRC_OP);
1372 if (arg_type == ARG_ANYTHING) {
1373 if (is_pointer_value(env, regno)) {
1374 verbose("R%d leaks addr into helper function\n", regno);
1380 if (type == PTR_TO_PACKET &&
1381 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
1382 verbose("helper access to the packet is not allowed\n");
1386 if (arg_type == ARG_PTR_TO_MAP_KEY ||
1387 arg_type == ARG_PTR_TO_MAP_VALUE) {
1388 expected_type = PTR_TO_STACK;
1389 if (type != PTR_TO_PACKET && type != expected_type)
1391 } else if (arg_type == ARG_CONST_SIZE ||
1392 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1393 expected_type = SCALAR_VALUE;
1394 if (type != expected_type)
1396 } else if (arg_type == ARG_CONST_MAP_PTR) {
1397 expected_type = CONST_PTR_TO_MAP;
1398 if (type != expected_type)
1400 } else if (arg_type == ARG_PTR_TO_CTX) {
1401 expected_type = PTR_TO_CTX;
1402 if (type != expected_type)
1404 } else if (arg_type == ARG_PTR_TO_MEM ||
1405 arg_type == ARG_PTR_TO_UNINIT_MEM) {
1406 expected_type = PTR_TO_STACK;
1407 /* One exception here. In case function allows for NULL to be
1408 * passed in as argument, it's a SCALAR_VALUE type. Final test
1409 * happens during stack boundary checking.
1411 if (register_is_null(*reg))
1412 /* final test in check_stack_boundary() */;
1413 else if (type != PTR_TO_PACKET && type != PTR_TO_MAP_VALUE &&
1414 type != expected_type)
1416 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
1418 verbose("unsupported arg_type %d\n", arg_type);
1422 if (arg_type == ARG_CONST_MAP_PTR) {
1423 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
1424 meta->map_ptr = reg->map_ptr;
1425 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
1426 /* bpf_map_xxx(..., map_ptr, ..., key) call:
1427 * check that [key, key + map->key_size) are within
1428 * stack limits and initialized
1430 if (!meta->map_ptr) {
1431 /* in function declaration map_ptr must come before
1432 * map_key, so that it's verified and known before
1433 * we have to check map_key here. Otherwise it means
1434 * that kernel subsystem misconfigured verifier
1436 verbose("invalid map_ptr to access map->key\n");
1439 if (type == PTR_TO_PACKET)
1440 err = check_packet_access(env, regno, reg->off,
1441 meta->map_ptr->key_size);
1443 err = check_stack_boundary(env, regno,
1444 meta->map_ptr->key_size,
1446 } else if (arg_type == ARG_PTR_TO_MAP_VALUE) {
1447 /* bpf_map_xxx(..., map_ptr, ..., value) call:
1448 * check [value, value + map->value_size) validity
1450 if (!meta->map_ptr) {
1451 /* kernel subsystem misconfigured verifier */
1452 verbose("invalid map_ptr to access map->value\n");
1455 if (type == PTR_TO_PACKET)
1456 err = check_packet_access(env, regno, reg->off,
1457 meta->map_ptr->value_size);
1459 err = check_stack_boundary(env, regno,
1460 meta->map_ptr->value_size,
1462 } else if (arg_type == ARG_CONST_SIZE ||
1463 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1464 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
1466 /* bpf_xxx(..., buf, len) call will access 'len' bytes
1467 * from stack pointer 'buf'. Check it
1468 * note: regno == len, regno - 1 == buf
1471 /* kernel subsystem misconfigured verifier */
1472 verbose("ARG_CONST_SIZE cannot be first argument\n");
1476 /* The register is SCALAR_VALUE; the access check
1477 * happens using its boundaries.
1480 if (!tnum_is_const(reg->var_off))
1481 /* For unprivileged variable accesses, disable raw
1482 * mode so that the program is required to
1483 * initialize all the memory that the helper could
1484 * just partially fill up.
1488 if (reg->smin_value < 0) {
1489 verbose("R%d min value is negative, either use unsigned or 'var &= const'\n",
1494 if (reg->umin_value == 0) {
1495 err = check_helper_mem_access(env, regno - 1, 0,
1502 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
1503 verbose("R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
1507 err = check_helper_mem_access(env, regno - 1,
1509 zero_size_allowed, meta);
1514 verbose("R%d type=%s expected=%s\n", regno,
1515 reg_type_str[type], reg_type_str[expected_type]);
1519 static int check_map_func_compatibility(struct bpf_map *map, int func_id)
1524 /* We need a two way check, first is from map perspective ... */
1525 switch (map->map_type) {
1526 case BPF_MAP_TYPE_PROG_ARRAY:
1527 if (func_id != BPF_FUNC_tail_call)
1530 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
1531 if (func_id != BPF_FUNC_perf_event_read &&
1532 func_id != BPF_FUNC_perf_event_output)
1535 case BPF_MAP_TYPE_STACK_TRACE:
1536 if (func_id != BPF_FUNC_get_stackid)
1539 case BPF_MAP_TYPE_CGROUP_ARRAY:
1540 if (func_id != BPF_FUNC_skb_under_cgroup &&
1541 func_id != BPF_FUNC_current_task_under_cgroup)
1544 /* devmap returns a pointer to a live net_device ifindex that we cannot
1545 * allow to be modified from bpf side. So do not allow lookup elements
1548 case BPF_MAP_TYPE_DEVMAP:
1549 if (func_id != BPF_FUNC_redirect_map)
1552 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
1553 case BPF_MAP_TYPE_HASH_OF_MAPS:
1554 if (func_id != BPF_FUNC_map_lookup_elem)
1557 case BPF_MAP_TYPE_SOCKMAP:
1558 if (func_id != BPF_FUNC_sk_redirect_map &&
1559 func_id != BPF_FUNC_sock_map_update &&
1560 func_id != BPF_FUNC_map_delete_elem)
1567 /* ... and second from the function itself. */
1569 case BPF_FUNC_tail_call:
1570 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
1573 case BPF_FUNC_perf_event_read:
1574 case BPF_FUNC_perf_event_output:
1575 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
1578 case BPF_FUNC_get_stackid:
1579 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
1582 case BPF_FUNC_current_task_under_cgroup:
1583 case BPF_FUNC_skb_under_cgroup:
1584 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
1587 case BPF_FUNC_redirect_map:
1588 if (map->map_type != BPF_MAP_TYPE_DEVMAP)
1591 case BPF_FUNC_sk_redirect_map:
1592 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
1595 case BPF_FUNC_sock_map_update:
1596 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
1605 verbose("cannot pass map_type %d into func %s#%d\n",
1606 map->map_type, func_id_name(func_id), func_id);
1610 static int check_raw_mode(const struct bpf_func_proto *fn)
1614 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
1616 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
1618 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
1620 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
1622 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
1625 return count > 1 ? -EINVAL : 0;
1628 /* Packet data might have moved, any old PTR_TO_PACKET[_END] are now invalid,
1629 * so turn them into unknown SCALAR_VALUE.
1631 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
1633 struct bpf_verifier_state *state = &env->cur_state;
1634 struct bpf_reg_state *regs = state->regs, *reg;
1637 for (i = 0; i < MAX_BPF_REG; i++)
1638 if (regs[i].type == PTR_TO_PACKET ||
1639 regs[i].type == PTR_TO_PACKET_END)
1640 mark_reg_unknown(regs, i);
1642 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
1643 if (state->stack_slot_type[i] != STACK_SPILL)
1645 reg = &state->spilled_regs[i / BPF_REG_SIZE];
1646 if (reg->type != PTR_TO_PACKET &&
1647 reg->type != PTR_TO_PACKET_END)
1649 __mark_reg_unknown(reg);
1653 static int check_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
1655 struct bpf_verifier_state *state = &env->cur_state;
1656 const struct bpf_func_proto *fn = NULL;
1657 struct bpf_reg_state *regs = state->regs;
1658 struct bpf_call_arg_meta meta;
1662 /* find function prototype */
1663 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
1664 verbose("invalid func %s#%d\n", func_id_name(func_id), func_id);
1668 if (env->prog->aux->ops->get_func_proto)
1669 fn = env->prog->aux->ops->get_func_proto(func_id);
1672 verbose("unknown func %s#%d\n", func_id_name(func_id), func_id);
1676 /* eBPF programs must be GPL compatible to use GPL-ed functions */
1677 if (!env->prog->gpl_compatible && fn->gpl_only) {
1678 verbose("cannot call GPL only function from proprietary program\n");
1682 changes_data = bpf_helper_changes_pkt_data(fn->func);
1684 memset(&meta, 0, sizeof(meta));
1685 meta.pkt_access = fn->pkt_access;
1687 /* We only support one arg being in raw mode at the moment, which
1688 * is sufficient for the helper functions we have right now.
1690 err = check_raw_mode(fn);
1692 verbose("kernel subsystem misconfigured func %s#%d\n",
1693 func_id_name(func_id), func_id);
1698 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
1701 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
1704 if (func_id == BPF_FUNC_tail_call) {
1705 if (meta.map_ptr == NULL) {
1706 verbose("verifier bug\n");
1709 env->insn_aux_data[insn_idx].map_ptr = meta.map_ptr;
1711 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
1714 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
1717 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
1721 /* Mark slots with STACK_MISC in case of raw mode, stack offset
1722 * is inferred from register state.
1724 for (i = 0; i < meta.access_size; i++) {
1725 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B, BPF_WRITE, -1);
1730 /* reset caller saved regs */
1731 for (i = 0; i < CALLER_SAVED_REGS; i++) {
1732 mark_reg_not_init(regs, caller_saved[i]);
1733 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
1736 /* update return register (already marked as written above) */
1737 if (fn->ret_type == RET_INTEGER) {
1738 /* sets type to SCALAR_VALUE */
1739 mark_reg_unknown(regs, BPF_REG_0);
1740 } else if (fn->ret_type == RET_VOID) {
1741 regs[BPF_REG_0].type = NOT_INIT;
1742 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL) {
1743 struct bpf_insn_aux_data *insn_aux;
1745 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
1746 /* There is no offset yet applied, variable or fixed */
1747 mark_reg_known_zero(regs, BPF_REG_0);
1748 regs[BPF_REG_0].off = 0;
1749 /* remember map_ptr, so that check_map_access()
1750 * can check 'value_size' boundary of memory access
1751 * to map element returned from bpf_map_lookup_elem()
1753 if (meta.map_ptr == NULL) {
1754 verbose("kernel subsystem misconfigured verifier\n");
1757 regs[BPF_REG_0].map_ptr = meta.map_ptr;
1758 regs[BPF_REG_0].id = ++env->id_gen;
1759 insn_aux = &env->insn_aux_data[insn_idx];
1760 if (!insn_aux->map_ptr)
1761 insn_aux->map_ptr = meta.map_ptr;
1762 else if (insn_aux->map_ptr != meta.map_ptr)
1763 insn_aux->map_ptr = BPF_MAP_PTR_POISON;
1765 verbose("unknown return type %d of func %s#%d\n",
1766 fn->ret_type, func_id_name(func_id), func_id);
1770 err = check_map_func_compatibility(meta.map_ptr, func_id);
1775 clear_all_pkt_pointers(env);
1779 static bool signed_add_overflows(s64 a, s64 b)
1781 /* Do the add in u64, where overflow is well-defined */
1782 s64 res = (s64)((u64)a + (u64)b);
1789 static bool signed_sub_overflows(s64 a, s64 b)
1791 /* Do the sub in u64, where overflow is well-defined */
1792 s64 res = (s64)((u64)a - (u64)b);
1799 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
1800 const struct bpf_reg_state *reg,
1801 enum bpf_reg_type type)
1803 bool known = tnum_is_const(reg->var_off);
1804 s64 val = reg->var_off.value;
1805 s64 smin = reg->smin_value;
1807 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
1808 verbose("math between %s pointer and %lld is not allowed\n",
1809 reg_type_str[type], val);
1813 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
1814 verbose("%s pointer offset %d is not allowed\n",
1815 reg_type_str[type], reg->off);
1819 if (smin == S64_MIN) {
1820 verbose("math between %s pointer and register with unbounded min value is not allowed\n",
1821 reg_type_str[type]);
1825 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
1826 verbose("value %lld makes %s pointer be out of bounds\n",
1827 smin, reg_type_str[type]);
1834 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
1835 * Caller should also handle BPF_MOV case separately.
1836 * If we return -EACCES, caller may want to try again treating pointer as a
1837 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
1839 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
1840 struct bpf_insn *insn,
1841 const struct bpf_reg_state *ptr_reg,
1842 const struct bpf_reg_state *off_reg)
1844 struct bpf_reg_state *regs = env->cur_state.regs, *dst_reg;
1845 bool known = tnum_is_const(off_reg->var_off);
1846 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
1847 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
1848 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
1849 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
1850 u8 opcode = BPF_OP(insn->code);
1851 u32 dst = insn->dst_reg;
1853 dst_reg = ®s[dst];
1855 if (WARN_ON_ONCE(known && (smin_val != smax_val))) {
1856 print_verifier_state(&env->cur_state);
1857 verbose("verifier internal error: known but bad sbounds\n");
1860 if (WARN_ON_ONCE(known && (umin_val != umax_val))) {
1861 print_verifier_state(&env->cur_state);
1862 verbose("verifier internal error: known but bad ubounds\n");
1866 if (BPF_CLASS(insn->code) != BPF_ALU64) {
1867 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
1868 if (!env->allow_ptr_leaks)
1869 verbose("R%d 32-bit pointer arithmetic prohibited\n",
1874 if (ptr_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
1875 if (!env->allow_ptr_leaks)
1876 verbose("R%d pointer arithmetic on PTR_TO_MAP_VALUE_OR_NULL prohibited, null-check it first\n",
1880 if (ptr_reg->type == CONST_PTR_TO_MAP) {
1881 if (!env->allow_ptr_leaks)
1882 verbose("R%d pointer arithmetic on CONST_PTR_TO_MAP prohibited\n",
1886 if (ptr_reg->type == PTR_TO_PACKET_END) {
1887 if (!env->allow_ptr_leaks)
1888 verbose("R%d pointer arithmetic on PTR_TO_PACKET_END prohibited\n",
1893 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
1894 * The id may be overwritten later if we create a new variable offset.
1896 dst_reg->type = ptr_reg->type;
1897 dst_reg->id = ptr_reg->id;
1899 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
1900 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
1905 /* We can take a fixed offset as long as it doesn't overflow
1906 * the s32 'off' field
1908 if (known && (ptr_reg->off + smin_val ==
1909 (s64)(s32)(ptr_reg->off + smin_val))) {
1910 /* pointer += K. Accumulate it into fixed offset */
1911 dst_reg->smin_value = smin_ptr;
1912 dst_reg->smax_value = smax_ptr;
1913 dst_reg->umin_value = umin_ptr;
1914 dst_reg->umax_value = umax_ptr;
1915 dst_reg->var_off = ptr_reg->var_off;
1916 dst_reg->off = ptr_reg->off + smin_val;
1917 dst_reg->range = ptr_reg->range;
1920 /* A new variable offset is created. Note that off_reg->off
1921 * == 0, since it's a scalar.
1922 * dst_reg gets the pointer type and since some positive
1923 * integer value was added to the pointer, give it a new 'id'
1924 * if it's a PTR_TO_PACKET.
1925 * this creates a new 'base' pointer, off_reg (variable) gets
1926 * added into the variable offset, and we copy the fixed offset
1929 if (signed_add_overflows(smin_ptr, smin_val) ||
1930 signed_add_overflows(smax_ptr, smax_val)) {
1931 dst_reg->smin_value = S64_MIN;
1932 dst_reg->smax_value = S64_MAX;
1934 dst_reg->smin_value = smin_ptr + smin_val;
1935 dst_reg->smax_value = smax_ptr + smax_val;
1937 if (umin_ptr + umin_val < umin_ptr ||
1938 umax_ptr + umax_val < umax_ptr) {
1939 dst_reg->umin_value = 0;
1940 dst_reg->umax_value = U64_MAX;
1942 dst_reg->umin_value = umin_ptr + umin_val;
1943 dst_reg->umax_value = umax_ptr + umax_val;
1945 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
1946 dst_reg->off = ptr_reg->off;
1947 if (ptr_reg->type == PTR_TO_PACKET) {
1948 dst_reg->id = ++env->id_gen;
1949 /* something was added to pkt_ptr, set range to zero */
1954 if (dst_reg == off_reg) {
1955 /* scalar -= pointer. Creates an unknown scalar */
1956 if (!env->allow_ptr_leaks)
1957 verbose("R%d tried to subtract pointer from scalar\n",
1961 /* We don't allow subtraction from FP, because (according to
1962 * test_verifier.c test "invalid fp arithmetic", JITs might not
1963 * be able to deal with it.
1965 if (ptr_reg->type == PTR_TO_STACK) {
1966 if (!env->allow_ptr_leaks)
1967 verbose("R%d subtraction from stack pointer prohibited\n",
1971 if (known && (ptr_reg->off - smin_val ==
1972 (s64)(s32)(ptr_reg->off - smin_val))) {
1973 /* pointer -= K. Subtract it from fixed offset */
1974 dst_reg->smin_value = smin_ptr;
1975 dst_reg->smax_value = smax_ptr;
1976 dst_reg->umin_value = umin_ptr;
1977 dst_reg->umax_value = umax_ptr;
1978 dst_reg->var_off = ptr_reg->var_off;
1979 dst_reg->id = ptr_reg->id;
1980 dst_reg->off = ptr_reg->off - smin_val;
1981 dst_reg->range = ptr_reg->range;
1984 /* A new variable offset is created. If the subtrahend is known
1985 * nonnegative, then any reg->range we had before is still good.
1987 if (signed_sub_overflows(smin_ptr, smax_val) ||
1988 signed_sub_overflows(smax_ptr, smin_val)) {
1989 /* Overflow possible, we know nothing */
1990 dst_reg->smin_value = S64_MIN;
1991 dst_reg->smax_value = S64_MAX;
1993 dst_reg->smin_value = smin_ptr - smax_val;
1994 dst_reg->smax_value = smax_ptr - smin_val;
1996 if (umin_ptr < umax_val) {
1997 /* Overflow possible, we know nothing */
1998 dst_reg->umin_value = 0;
1999 dst_reg->umax_value = U64_MAX;
2001 /* Cannot overflow (as long as bounds are consistent) */
2002 dst_reg->umin_value = umin_ptr - umax_val;
2003 dst_reg->umax_value = umax_ptr - umin_val;
2005 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
2006 dst_reg->off = ptr_reg->off;
2007 if (ptr_reg->type == PTR_TO_PACKET) {
2008 dst_reg->id = ++env->id_gen;
2009 /* something was added to pkt_ptr, set range to zero */
2017 /* bitwise ops on pointers are troublesome, prohibit for now.
2018 * (However, in principle we could allow some cases, e.g.
2019 * ptr &= ~3 which would reduce min_value by 3.)
2021 if (!env->allow_ptr_leaks)
2022 verbose("R%d bitwise operator %s on pointer prohibited\n",
2023 dst, bpf_alu_string[opcode >> 4]);
2026 /* other operators (e.g. MUL,LSH) produce non-pointer results */
2027 if (!env->allow_ptr_leaks)
2028 verbose("R%d pointer arithmetic with %s operator prohibited\n",
2029 dst, bpf_alu_string[opcode >> 4]);
2033 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
2036 __update_reg_bounds(dst_reg);
2037 __reg_deduce_bounds(dst_reg);
2038 __reg_bound_offset(dst_reg);
2042 /* WARNING: This function does calculations on 64-bit values, but the actual
2043 * execution may occur on 32-bit values. Therefore, things like bitshifts
2044 * need extra checks in the 32-bit case.
2046 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
2047 struct bpf_insn *insn,
2048 struct bpf_reg_state *dst_reg,
2049 struct bpf_reg_state src_reg)
2051 struct bpf_reg_state *regs = env->cur_state.regs;
2052 u8 opcode = BPF_OP(insn->code);
2053 bool src_known, dst_known;
2054 s64 smin_val, smax_val;
2055 u64 umin_val, umax_val;
2056 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
2058 smin_val = src_reg.smin_value;
2059 smax_val = src_reg.smax_value;
2060 umin_val = src_reg.umin_value;
2061 umax_val = src_reg.umax_value;
2062 src_known = tnum_is_const(src_reg.var_off);
2063 dst_known = tnum_is_const(dst_reg->var_off);
2066 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
2067 __mark_reg_unknown(dst_reg);
2073 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
2074 signed_add_overflows(dst_reg->smax_value, smax_val)) {
2075 dst_reg->smin_value = S64_MIN;
2076 dst_reg->smax_value = S64_MAX;
2078 dst_reg->smin_value += smin_val;
2079 dst_reg->smax_value += smax_val;
2081 if (dst_reg->umin_value + umin_val < umin_val ||
2082 dst_reg->umax_value + umax_val < umax_val) {
2083 dst_reg->umin_value = 0;
2084 dst_reg->umax_value = U64_MAX;
2086 dst_reg->umin_value += umin_val;
2087 dst_reg->umax_value += umax_val;
2089 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
2092 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
2093 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
2094 /* Overflow possible, we know nothing */
2095 dst_reg->smin_value = S64_MIN;
2096 dst_reg->smax_value = S64_MAX;
2098 dst_reg->smin_value -= smax_val;
2099 dst_reg->smax_value -= smin_val;
2101 if (dst_reg->umin_value < umax_val) {
2102 /* Overflow possible, we know nothing */
2103 dst_reg->umin_value = 0;
2104 dst_reg->umax_value = U64_MAX;
2106 /* Cannot overflow (as long as bounds are consistent) */
2107 dst_reg->umin_value -= umax_val;
2108 dst_reg->umax_value -= umin_val;
2110 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
2113 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
2114 if (smin_val < 0 || dst_reg->smin_value < 0) {
2115 /* Ain't nobody got time to multiply that sign */
2116 __mark_reg_unbounded(dst_reg);
2117 __update_reg_bounds(dst_reg);
2120 /* Both values are positive, so we can work with unsigned and
2121 * copy the result to signed (unless it exceeds S64_MAX).
2123 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
2124 /* Potential overflow, we know nothing */
2125 __mark_reg_unbounded(dst_reg);
2126 /* (except what we can learn from the var_off) */
2127 __update_reg_bounds(dst_reg);
2130 dst_reg->umin_value *= umin_val;
2131 dst_reg->umax_value *= umax_val;
2132 if (dst_reg->umax_value > S64_MAX) {
2133 /* Overflow possible, we know nothing */
2134 dst_reg->smin_value = S64_MIN;
2135 dst_reg->smax_value = S64_MAX;
2137 dst_reg->smin_value = dst_reg->umin_value;
2138 dst_reg->smax_value = dst_reg->umax_value;
2142 if (src_known && dst_known) {
2143 __mark_reg_known(dst_reg, dst_reg->var_off.value &
2144 src_reg.var_off.value);
2147 /* We get our minimum from the var_off, since that's inherently
2148 * bitwise. Our maximum is the minimum of the operands' maxima.
2150 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
2151 dst_reg->umin_value = dst_reg->var_off.value;
2152 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
2153 if (dst_reg->smin_value < 0 || smin_val < 0) {
2154 /* Lose signed bounds when ANDing negative numbers,
2155 * ain't nobody got time for that.
2157 dst_reg->smin_value = S64_MIN;
2158 dst_reg->smax_value = S64_MAX;
2160 /* ANDing two positives gives a positive, so safe to
2161 * cast result into s64.
2163 dst_reg->smin_value = dst_reg->umin_value;
2164 dst_reg->smax_value = dst_reg->umax_value;
2166 /* We may learn something more from the var_off */
2167 __update_reg_bounds(dst_reg);
2170 if (src_known && dst_known) {
2171 __mark_reg_known(dst_reg, dst_reg->var_off.value |
2172 src_reg.var_off.value);
2175 /* We get our maximum from the var_off, and our minimum is the
2176 * maximum of the operands' minima
2178 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
2179 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
2180 dst_reg->umax_value = dst_reg->var_off.value |
2181 dst_reg->var_off.mask;
2182 if (dst_reg->smin_value < 0 || smin_val < 0) {
2183 /* Lose signed bounds when ORing negative numbers,
2184 * ain't nobody got time for that.
2186 dst_reg->smin_value = S64_MIN;
2187 dst_reg->smax_value = S64_MAX;
2189 /* ORing two positives gives a positive, so safe to
2190 * cast result into s64.
2192 dst_reg->smin_value = dst_reg->umin_value;
2193 dst_reg->smax_value = dst_reg->umax_value;
2195 /* We may learn something more from the var_off */
2196 __update_reg_bounds(dst_reg);
2199 if (umax_val >= insn_bitness) {
2200 /* Shifts greater than 31 or 63 are undefined.
2201 * This includes shifts by a negative number.
2203 mark_reg_unknown(regs, insn->dst_reg);
2206 /* We lose all sign bit information (except what we can pick
2209 dst_reg->smin_value = S64_MIN;
2210 dst_reg->smax_value = S64_MAX;
2211 /* If we might shift our top bit out, then we know nothing */
2212 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
2213 dst_reg->umin_value = 0;
2214 dst_reg->umax_value = U64_MAX;
2216 dst_reg->umin_value <<= umin_val;
2217 dst_reg->umax_value <<= umax_val;
2220 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
2222 dst_reg->var_off = tnum_lshift(tnum_unknown, umin_val);
2223 /* We may learn something more from the var_off */
2224 __update_reg_bounds(dst_reg);
2227 if (umax_val >= insn_bitness) {
2228 /* Shifts greater than 31 or 63 are undefined.
2229 * This includes shifts by a negative number.
2231 mark_reg_unknown(regs, insn->dst_reg);
2234 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
2235 * be negative, then either:
2236 * 1) src_reg might be zero, so the sign bit of the result is
2237 * unknown, so we lose our signed bounds
2238 * 2) it's known negative, thus the unsigned bounds capture the
2240 * 3) the signed bounds cross zero, so they tell us nothing
2242 * If the value in dst_reg is known nonnegative, then again the
2243 * unsigned bounts capture the signed bounds.
2244 * Thus, in all cases it suffices to blow away our signed bounds
2245 * and rely on inferring new ones from the unsigned bounds and
2246 * var_off of the result.
2248 dst_reg->smin_value = S64_MIN;
2249 dst_reg->smax_value = S64_MAX;
2251 dst_reg->var_off = tnum_rshift(dst_reg->var_off,
2254 dst_reg->var_off = tnum_rshift(tnum_unknown, umin_val);
2255 dst_reg->umin_value >>= umax_val;
2256 dst_reg->umax_value >>= umin_val;
2257 /* We may learn something more from the var_off */
2258 __update_reg_bounds(dst_reg);
2261 mark_reg_unknown(regs, insn->dst_reg);
2265 if (BPF_CLASS(insn->code) != BPF_ALU64) {
2266 /* 32-bit ALU ops are (32,32)->32 */
2267 coerce_reg_to_size(dst_reg, 4);
2268 coerce_reg_to_size(&src_reg, 4);
2271 __reg_deduce_bounds(dst_reg);
2272 __reg_bound_offset(dst_reg);
2276 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
2279 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
2280 struct bpf_insn *insn)
2282 struct bpf_reg_state *regs = env->cur_state.regs, *dst_reg, *src_reg;
2283 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
2284 u8 opcode = BPF_OP(insn->code);
2287 dst_reg = ®s[insn->dst_reg];
2289 if (dst_reg->type != SCALAR_VALUE)
2291 if (BPF_SRC(insn->code) == BPF_X) {
2292 src_reg = ®s[insn->src_reg];
2293 if (src_reg->type != SCALAR_VALUE) {
2294 if (dst_reg->type != SCALAR_VALUE) {
2295 /* Combining two pointers by any ALU op yields
2296 * an arbitrary scalar.
2298 if (!env->allow_ptr_leaks) {
2299 verbose("R%d pointer %s pointer prohibited\n",
2301 bpf_alu_string[opcode >> 4]);
2304 mark_reg_unknown(regs, insn->dst_reg);
2307 /* scalar += pointer
2308 * This is legal, but we have to reverse our
2309 * src/dest handling in computing the range
2311 rc = adjust_ptr_min_max_vals(env, insn,
2313 if (rc == -EACCES && env->allow_ptr_leaks) {
2314 /* scalar += unknown scalar */
2315 __mark_reg_unknown(&off_reg);
2316 return adjust_scalar_min_max_vals(
2322 } else if (ptr_reg) {
2323 /* pointer += scalar */
2324 rc = adjust_ptr_min_max_vals(env, insn,
2326 if (rc == -EACCES && env->allow_ptr_leaks) {
2327 /* unknown scalar += scalar */
2328 __mark_reg_unknown(dst_reg);
2329 return adjust_scalar_min_max_vals(
2330 env, insn, dst_reg, *src_reg);
2335 /* Pretend the src is a reg with a known value, since we only
2336 * need to be able to read from this state.
2338 off_reg.type = SCALAR_VALUE;
2339 __mark_reg_known(&off_reg, insn->imm);
2341 if (ptr_reg) { /* pointer += K */
2342 rc = adjust_ptr_min_max_vals(env, insn,
2344 if (rc == -EACCES && env->allow_ptr_leaks) {
2345 /* unknown scalar += K */
2346 __mark_reg_unknown(dst_reg);
2347 return adjust_scalar_min_max_vals(
2348 env, insn, dst_reg, off_reg);
2354 /* Got here implies adding two SCALAR_VALUEs */
2355 if (WARN_ON_ONCE(ptr_reg)) {
2356 print_verifier_state(&env->cur_state);
2357 verbose("verifier internal error: unexpected ptr_reg\n");
2360 if (WARN_ON(!src_reg)) {
2361 print_verifier_state(&env->cur_state);
2362 verbose("verifier internal error: no src_reg\n");
2365 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
2368 /* check validity of 32-bit and 64-bit arithmetic operations */
2369 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
2371 struct bpf_reg_state *regs = env->cur_state.regs;
2372 u8 opcode = BPF_OP(insn->code);
2375 if (opcode == BPF_END || opcode == BPF_NEG) {
2376 if (opcode == BPF_NEG) {
2377 if (BPF_SRC(insn->code) != 0 ||
2378 insn->src_reg != BPF_REG_0 ||
2379 insn->off != 0 || insn->imm != 0) {
2380 verbose("BPF_NEG uses reserved fields\n");
2384 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
2385 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
2386 BPF_CLASS(insn->code) == BPF_ALU64) {
2387 verbose("BPF_END uses reserved fields\n");
2392 /* check src operand */
2393 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2397 if (is_pointer_value(env, insn->dst_reg)) {
2398 verbose("R%d pointer arithmetic prohibited\n",
2403 /* check dest operand */
2404 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2408 } else if (opcode == BPF_MOV) {
2410 if (BPF_SRC(insn->code) == BPF_X) {
2411 if (insn->imm != 0 || insn->off != 0) {
2412 verbose("BPF_MOV uses reserved fields\n");
2416 /* check src operand */
2417 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2421 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
2422 verbose("BPF_MOV uses reserved fields\n");
2427 /* check dest operand */
2428 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2432 if (BPF_SRC(insn->code) == BPF_X) {
2433 if (BPF_CLASS(insn->code) == BPF_ALU64) {
2435 * copy register state to dest reg
2437 regs[insn->dst_reg] = regs[insn->src_reg];
2438 regs[insn->dst_reg].live |= REG_LIVE_WRITTEN;
2441 if (is_pointer_value(env, insn->src_reg)) {
2442 verbose("R%d partial copy of pointer\n",
2446 mark_reg_unknown(regs, insn->dst_reg);
2447 coerce_reg_to_size(®s[insn->dst_reg], 4);
2451 * remember the value we stored into this reg
2453 regs[insn->dst_reg].type = SCALAR_VALUE;
2454 if (BPF_CLASS(insn->code) == BPF_ALU64) {
2455 __mark_reg_known(regs + insn->dst_reg,
2458 __mark_reg_known(regs + insn->dst_reg,
2463 } else if (opcode > BPF_END) {
2464 verbose("invalid BPF_ALU opcode %x\n", opcode);
2467 } else { /* all other ALU ops: and, sub, xor, add, ... */
2469 if (BPF_SRC(insn->code) == BPF_X) {
2470 if (insn->imm != 0 || insn->off != 0) {
2471 verbose("BPF_ALU uses reserved fields\n");
2474 /* check src1 operand */
2475 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2479 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
2480 verbose("BPF_ALU uses reserved fields\n");
2485 /* check src2 operand */
2486 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2490 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
2491 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
2492 verbose("div by zero\n");
2496 if (opcode == BPF_ARSH && BPF_CLASS(insn->code) != BPF_ALU64) {
2497 verbose("BPF_ARSH not supported for 32 bit ALU\n");
2501 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
2502 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
2503 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
2505 if (insn->imm < 0 || insn->imm >= size) {
2506 verbose("invalid shift %d\n", insn->imm);
2511 /* check dest operand */
2512 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
2516 return adjust_reg_min_max_vals(env, insn);
2522 static void find_good_pkt_pointers(struct bpf_verifier_state *state,
2523 struct bpf_reg_state *dst_reg,
2524 bool range_right_open)
2526 struct bpf_reg_state *regs = state->regs, *reg;
2530 if (dst_reg->off < 0 ||
2531 (dst_reg->off == 0 && range_right_open))
2532 /* This doesn't give us any range */
2535 if (dst_reg->umax_value > MAX_PACKET_OFF ||
2536 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
2537 /* Risk of overflow. For instance, ptr + (1<<63) may be less
2538 * than pkt_end, but that's because it's also less than pkt.
2542 new_range = dst_reg->off;
2543 if (range_right_open)
2546 /* Examples for register markings:
2548 * pkt_data in dst register:
2552 * if (r2 > pkt_end) goto <handle exception>
2557 * if (r2 < pkt_end) goto <access okay>
2558 * <handle exception>
2561 * r2 == dst_reg, pkt_end == src_reg
2562 * r2=pkt(id=n,off=8,r=0)
2563 * r3=pkt(id=n,off=0,r=0)
2565 * pkt_data in src register:
2569 * if (pkt_end >= r2) goto <access okay>
2570 * <handle exception>
2574 * if (pkt_end <= r2) goto <handle exception>
2578 * pkt_end == dst_reg, r2 == src_reg
2579 * r2=pkt(id=n,off=8,r=0)
2580 * r3=pkt(id=n,off=0,r=0)
2582 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
2583 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
2584 * and [r3, r3 + 8-1) respectively is safe to access depending on
2588 /* If our ids match, then we must have the same max_value. And we
2589 * don't care about the other reg's fixed offset, since if it's too big
2590 * the range won't allow anything.
2591 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
2593 for (i = 0; i < MAX_BPF_REG; i++)
2594 if (regs[i].type == PTR_TO_PACKET && regs[i].id == dst_reg->id)
2595 /* keep the maximum range already checked */
2596 regs[i].range = max(regs[i].range, new_range);
2598 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
2599 if (state->stack_slot_type[i] != STACK_SPILL)
2601 reg = &state->spilled_regs[i / BPF_REG_SIZE];
2602 if (reg->type == PTR_TO_PACKET && reg->id == dst_reg->id)
2603 reg->range = max(reg->range, new_range);
2607 /* Adjusts the register min/max values in the case that the dst_reg is the
2608 * variable register that we are working on, and src_reg is a constant or we're
2609 * simply doing a BPF_K check.
2610 * In JEQ/JNE cases we also adjust the var_off values.
2612 static void reg_set_min_max(struct bpf_reg_state *true_reg,
2613 struct bpf_reg_state *false_reg, u64 val,
2616 /* If the dst_reg is a pointer, we can't learn anything about its
2617 * variable offset from the compare (unless src_reg were a pointer into
2618 * the same object, but we don't bother with that.
2619 * Since false_reg and true_reg have the same type by construction, we
2620 * only need to check one of them for pointerness.
2622 if (__is_pointer_value(false, false_reg))
2627 /* If this is false then we know nothing Jon Snow, but if it is
2628 * true then we know for sure.
2630 __mark_reg_known(true_reg, val);
2633 /* If this is true we know nothing Jon Snow, but if it is false
2634 * we know the value for sure;
2636 __mark_reg_known(false_reg, val);
2639 false_reg->umax_value = min(false_reg->umax_value, val);
2640 true_reg->umin_value = max(true_reg->umin_value, val + 1);
2643 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
2644 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
2647 false_reg->umin_value = max(false_reg->umin_value, val);
2648 true_reg->umax_value = min(true_reg->umax_value, val - 1);
2651 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
2652 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
2655 false_reg->umax_value = min(false_reg->umax_value, val - 1);
2656 true_reg->umin_value = max(true_reg->umin_value, val);
2659 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
2660 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
2663 false_reg->umin_value = max(false_reg->umin_value, val + 1);
2664 true_reg->umax_value = min(true_reg->umax_value, val);
2667 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
2668 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
2674 __reg_deduce_bounds(false_reg);
2675 __reg_deduce_bounds(true_reg);
2676 /* We might have learned some bits from the bounds. */
2677 __reg_bound_offset(false_reg);
2678 __reg_bound_offset(true_reg);
2679 /* Intersecting with the old var_off might have improved our bounds
2680 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2681 * then new var_off is (0; 0x7f...fc) which improves our umax.
2683 __update_reg_bounds(false_reg);
2684 __update_reg_bounds(true_reg);
2687 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
2690 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
2691 struct bpf_reg_state *false_reg, u64 val,
2694 if (__is_pointer_value(false, false_reg))
2699 /* If this is false then we know nothing Jon Snow, but if it is
2700 * true then we know for sure.
2702 __mark_reg_known(true_reg, val);
2705 /* If this is true we know nothing Jon Snow, but if it is false
2706 * we know the value for sure;
2708 __mark_reg_known(false_reg, val);
2711 true_reg->umax_value = min(true_reg->umax_value, val - 1);
2712 false_reg->umin_value = max(false_reg->umin_value, val);
2715 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
2716 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
2719 true_reg->umin_value = max(true_reg->umin_value, val + 1);
2720 false_reg->umax_value = min(false_reg->umax_value, val);
2723 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
2724 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
2727 true_reg->umax_value = min(true_reg->umax_value, val);
2728 false_reg->umin_value = max(false_reg->umin_value, val + 1);
2731 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
2732 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
2735 true_reg->umin_value = max(true_reg->umin_value, val);
2736 false_reg->umax_value = min(false_reg->umax_value, val - 1);
2739 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
2740 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
2746 __reg_deduce_bounds(false_reg);
2747 __reg_deduce_bounds(true_reg);
2748 /* We might have learned some bits from the bounds. */
2749 __reg_bound_offset(false_reg);
2750 __reg_bound_offset(true_reg);
2751 /* Intersecting with the old var_off might have improved our bounds
2752 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2753 * then new var_off is (0; 0x7f...fc) which improves our umax.
2755 __update_reg_bounds(false_reg);
2756 __update_reg_bounds(true_reg);
2759 /* Regs are known to be equal, so intersect their min/max/var_off */
2760 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
2761 struct bpf_reg_state *dst_reg)
2763 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
2764 dst_reg->umin_value);
2765 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
2766 dst_reg->umax_value);
2767 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
2768 dst_reg->smin_value);
2769 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
2770 dst_reg->smax_value);
2771 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
2773 /* We might have learned new bounds from the var_off. */
2774 __update_reg_bounds(src_reg);
2775 __update_reg_bounds(dst_reg);
2776 /* We might have learned something about the sign bit. */
2777 __reg_deduce_bounds(src_reg);
2778 __reg_deduce_bounds(dst_reg);
2779 /* We might have learned some bits from the bounds. */
2780 __reg_bound_offset(src_reg);
2781 __reg_bound_offset(dst_reg);
2782 /* Intersecting with the old var_off might have improved our bounds
2783 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2784 * then new var_off is (0; 0x7f...fc) which improves our umax.
2786 __update_reg_bounds(src_reg);
2787 __update_reg_bounds(dst_reg);
2790 static void reg_combine_min_max(struct bpf_reg_state *true_src,
2791 struct bpf_reg_state *true_dst,
2792 struct bpf_reg_state *false_src,
2793 struct bpf_reg_state *false_dst,
2798 __reg_combine_min_max(true_src, true_dst);
2801 __reg_combine_min_max(false_src, false_dst);
2806 static void mark_map_reg(struct bpf_reg_state *regs, u32 regno, u32 id,
2809 struct bpf_reg_state *reg = ®s[regno];
2811 if (reg->type == PTR_TO_MAP_VALUE_OR_NULL && reg->id == id) {
2812 /* Old offset (both fixed and variable parts) should
2813 * have been known-zero, because we don't allow pointer
2814 * arithmetic on pointers that might be NULL.
2816 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
2817 !tnum_equals_const(reg->var_off, 0) ||
2819 __mark_reg_known_zero(reg);
2823 reg->type = SCALAR_VALUE;
2824 } else if (reg->map_ptr->inner_map_meta) {
2825 reg->type = CONST_PTR_TO_MAP;
2826 reg->map_ptr = reg->map_ptr->inner_map_meta;
2828 reg->type = PTR_TO_MAP_VALUE;
2830 /* We don't need id from this point onwards anymore, thus we
2831 * should better reset it, so that state pruning has chances
2838 /* The logic is similar to find_good_pkt_pointers(), both could eventually
2839 * be folded together at some point.
2841 static void mark_map_regs(struct bpf_verifier_state *state, u32 regno,
2844 struct bpf_reg_state *regs = state->regs;
2845 u32 id = regs[regno].id;
2848 for (i = 0; i < MAX_BPF_REG; i++)
2849 mark_map_reg(regs, i, id, is_null);
2851 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
2852 if (state->stack_slot_type[i] != STACK_SPILL)
2854 mark_map_reg(state->spilled_regs, i / BPF_REG_SIZE, id, is_null);
2858 static int check_cond_jmp_op(struct bpf_verifier_env *env,
2859 struct bpf_insn *insn, int *insn_idx)
2861 struct bpf_verifier_state *other_branch, *this_branch = &env->cur_state;
2862 struct bpf_reg_state *regs = this_branch->regs, *dst_reg;
2863 u8 opcode = BPF_OP(insn->code);
2866 if (opcode > BPF_JSLE) {
2867 verbose("invalid BPF_JMP opcode %x\n", opcode);
2871 if (BPF_SRC(insn->code) == BPF_X) {
2872 if (insn->imm != 0) {
2873 verbose("BPF_JMP uses reserved fields\n");
2877 /* check src1 operand */
2878 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2882 if (is_pointer_value(env, insn->src_reg)) {
2883 verbose("R%d pointer comparison prohibited\n",
2888 if (insn->src_reg != BPF_REG_0) {
2889 verbose("BPF_JMP uses reserved fields\n");
2894 /* check src2 operand */
2895 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2899 dst_reg = ®s[insn->dst_reg];
2901 /* detect if R == 0 where R was initialized to zero earlier */
2902 if (BPF_SRC(insn->code) == BPF_K &&
2903 (opcode == BPF_JEQ || opcode == BPF_JNE) &&
2904 dst_reg->type == SCALAR_VALUE &&
2905 tnum_equals_const(dst_reg->var_off, insn->imm)) {
2906 if (opcode == BPF_JEQ) {
2907 /* if (imm == imm) goto pc+off;
2908 * only follow the goto, ignore fall-through
2910 *insn_idx += insn->off;
2913 /* if (imm != imm) goto pc+off;
2914 * only follow fall-through branch, since
2915 * that's where the program will go
2921 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx);
2925 /* detect if we are comparing against a constant value so we can adjust
2926 * our min/max values for our dst register.
2927 * this is only legit if both are scalars (or pointers to the same
2928 * object, I suppose, but we don't support that right now), because
2929 * otherwise the different base pointers mean the offsets aren't
2932 if (BPF_SRC(insn->code) == BPF_X) {
2933 if (dst_reg->type == SCALAR_VALUE &&
2934 regs[insn->src_reg].type == SCALAR_VALUE) {
2935 if (tnum_is_const(regs[insn->src_reg].var_off))
2936 reg_set_min_max(&other_branch->regs[insn->dst_reg],
2937 dst_reg, regs[insn->src_reg].var_off.value,
2939 else if (tnum_is_const(dst_reg->var_off))
2940 reg_set_min_max_inv(&other_branch->regs[insn->src_reg],
2941 ®s[insn->src_reg],
2942 dst_reg->var_off.value, opcode);
2943 else if (opcode == BPF_JEQ || opcode == BPF_JNE)
2944 /* Comparing for equality, we can combine knowledge */
2945 reg_combine_min_max(&other_branch->regs[insn->src_reg],
2946 &other_branch->regs[insn->dst_reg],
2947 ®s[insn->src_reg],
2948 ®s[insn->dst_reg], opcode);
2950 } else if (dst_reg->type == SCALAR_VALUE) {
2951 reg_set_min_max(&other_branch->regs[insn->dst_reg],
2952 dst_reg, insn->imm, opcode);
2955 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
2956 if (BPF_SRC(insn->code) == BPF_K &&
2957 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
2958 dst_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
2959 /* Mark all identical map registers in each branch as either
2960 * safe or unknown depending R == 0 or R != 0 conditional.
2962 mark_map_regs(this_branch, insn->dst_reg, opcode == BPF_JNE);
2963 mark_map_regs(other_branch, insn->dst_reg, opcode == BPF_JEQ);
2964 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGT &&
2965 dst_reg->type == PTR_TO_PACKET &&
2966 regs[insn->src_reg].type == PTR_TO_PACKET_END) {
2967 /* pkt_data' > pkt_end */
2968 find_good_pkt_pointers(this_branch, dst_reg, false);
2969 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGT &&
2970 dst_reg->type == PTR_TO_PACKET_END &&
2971 regs[insn->src_reg].type == PTR_TO_PACKET) {
2972 /* pkt_end > pkt_data' */
2973 find_good_pkt_pointers(other_branch, ®s[insn->src_reg], true);
2974 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JLT &&
2975 dst_reg->type == PTR_TO_PACKET &&
2976 regs[insn->src_reg].type == PTR_TO_PACKET_END) {
2977 /* pkt_data' < pkt_end */
2978 find_good_pkt_pointers(other_branch, dst_reg, true);
2979 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JLT &&
2980 dst_reg->type == PTR_TO_PACKET_END &&
2981 regs[insn->src_reg].type == PTR_TO_PACKET) {
2982 /* pkt_end < pkt_data' */
2983 find_good_pkt_pointers(this_branch, ®s[insn->src_reg], false);
2984 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGE &&
2985 dst_reg->type == PTR_TO_PACKET &&
2986 regs[insn->src_reg].type == PTR_TO_PACKET_END) {
2987 /* pkt_data' >= pkt_end */
2988 find_good_pkt_pointers(this_branch, dst_reg, true);
2989 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGE &&
2990 dst_reg->type == PTR_TO_PACKET_END &&
2991 regs[insn->src_reg].type == PTR_TO_PACKET) {
2992 /* pkt_end >= pkt_data' */
2993 find_good_pkt_pointers(other_branch, ®s[insn->src_reg], false);
2994 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JLE &&
2995 dst_reg->type == PTR_TO_PACKET &&
2996 regs[insn->src_reg].type == PTR_TO_PACKET_END) {
2997 /* pkt_data' <= pkt_end */
2998 find_good_pkt_pointers(other_branch, dst_reg, false);
2999 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JLE &&
3000 dst_reg->type == PTR_TO_PACKET_END &&
3001 regs[insn->src_reg].type == PTR_TO_PACKET) {
3002 /* pkt_end <= pkt_data' */
3003 find_good_pkt_pointers(this_branch, ®s[insn->src_reg], true);
3004 } else if (is_pointer_value(env, insn->dst_reg)) {
3005 verbose("R%d pointer comparison prohibited\n", insn->dst_reg);
3009 print_verifier_state(this_branch);
3013 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
3014 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
3016 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
3018 return (struct bpf_map *) (unsigned long) imm64;
3021 /* verify BPF_LD_IMM64 instruction */
3022 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
3024 struct bpf_reg_state *regs = env->cur_state.regs;
3027 if (BPF_SIZE(insn->code) != BPF_DW) {
3028 verbose("invalid BPF_LD_IMM insn\n");
3031 if (insn->off != 0) {
3032 verbose("BPF_LD_IMM64 uses reserved fields\n");
3036 err = check_reg_arg(env, insn->dst_reg, DST_OP);
3040 if (insn->src_reg == 0) {
3041 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
3043 regs[insn->dst_reg].type = SCALAR_VALUE;
3044 __mark_reg_known(®s[insn->dst_reg], imm);
3048 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
3049 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
3051 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
3052 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
3056 static bool may_access_skb(enum bpf_prog_type type)
3059 case BPF_PROG_TYPE_SOCKET_FILTER:
3060 case BPF_PROG_TYPE_SCHED_CLS:
3061 case BPF_PROG_TYPE_SCHED_ACT:
3068 /* verify safety of LD_ABS|LD_IND instructions:
3069 * - they can only appear in the programs where ctx == skb
3070 * - since they are wrappers of function calls, they scratch R1-R5 registers,
3071 * preserve R6-R9, and store return value into R0
3074 * ctx == skb == R6 == CTX
3077 * SRC == any register
3078 * IMM == 32-bit immediate
3081 * R0 - 8/16/32-bit skb data converted to cpu endianness
3083 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
3085 struct bpf_reg_state *regs = env->cur_state.regs;
3086 u8 mode = BPF_MODE(insn->code);
3089 if (!may_access_skb(env->prog->type)) {
3090 verbose("BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
3094 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
3095 BPF_SIZE(insn->code) == BPF_DW ||
3096 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
3097 verbose("BPF_LD_[ABS|IND] uses reserved fields\n");
3101 /* check whether implicit source operand (register R6) is readable */
3102 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
3106 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
3107 verbose("at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
3111 if (mode == BPF_IND) {
3112 /* check explicit source operand */
3113 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3118 /* reset caller saved regs to unreadable */
3119 for (i = 0; i < CALLER_SAVED_REGS; i++) {
3120 mark_reg_not_init(regs, caller_saved[i]);
3121 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
3124 /* mark destination R0 register as readable, since it contains
3125 * the value fetched from the packet.
3126 * Already marked as written above.
3128 mark_reg_unknown(regs, BPF_REG_0);
3132 /* non-recursive DFS pseudo code
3133 * 1 procedure DFS-iterative(G,v):
3134 * 2 label v as discovered
3135 * 3 let S be a stack
3137 * 5 while S is not empty
3139 * 7 if t is what we're looking for:
3141 * 9 for all edges e in G.adjacentEdges(t) do
3142 * 10 if edge e is already labelled
3143 * 11 continue with the next edge
3144 * 12 w <- G.adjacentVertex(t,e)
3145 * 13 if vertex w is not discovered and not explored
3146 * 14 label e as tree-edge
3147 * 15 label w as discovered
3150 * 18 else if vertex w is discovered
3151 * 19 label e as back-edge
3153 * 21 // vertex w is explored
3154 * 22 label e as forward- or cross-edge
3155 * 23 label t as explored
3160 * 0x11 - discovered and fall-through edge labelled
3161 * 0x12 - discovered and fall-through and branch edges labelled
3172 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
3174 static int *insn_stack; /* stack of insns to process */
3175 static int cur_stack; /* current stack index */
3176 static int *insn_state;
3178 /* t, w, e - match pseudo-code above:
3179 * t - index of current instruction
3180 * w - next instruction
3183 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
3185 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
3188 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
3191 if (w < 0 || w >= env->prog->len) {
3192 verbose("jump out of range from insn %d to %d\n", t, w);
3197 /* mark branch target for state pruning */
3198 env->explored_states[w] = STATE_LIST_MARK;
3200 if (insn_state[w] == 0) {
3202 insn_state[t] = DISCOVERED | e;
3203 insn_state[w] = DISCOVERED;
3204 if (cur_stack >= env->prog->len)
3206 insn_stack[cur_stack++] = w;
3208 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
3209 verbose("back-edge from insn %d to %d\n", t, w);
3211 } else if (insn_state[w] == EXPLORED) {
3212 /* forward- or cross-edge */
3213 insn_state[t] = DISCOVERED | e;
3215 verbose("insn state internal bug\n");
3221 /* non-recursive depth-first-search to detect loops in BPF program
3222 * loop == back-edge in directed graph
3224 static int check_cfg(struct bpf_verifier_env *env)
3226 struct bpf_insn *insns = env->prog->insnsi;
3227 int insn_cnt = env->prog->len;
3231 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
3235 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
3241 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
3242 insn_stack[0] = 0; /* 0 is the first instruction */
3248 t = insn_stack[cur_stack - 1];
3250 if (BPF_CLASS(insns[t].code) == BPF_JMP) {
3251 u8 opcode = BPF_OP(insns[t].code);
3253 if (opcode == BPF_EXIT) {
3255 } else if (opcode == BPF_CALL) {
3256 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3261 if (t + 1 < insn_cnt)
3262 env->explored_states[t + 1] = STATE_LIST_MARK;
3263 } else if (opcode == BPF_JA) {
3264 if (BPF_SRC(insns[t].code) != BPF_K) {
3268 /* unconditional jump with single edge */
3269 ret = push_insn(t, t + insns[t].off + 1,
3275 /* tell verifier to check for equivalent states
3276 * after every call and jump
3278 if (t + 1 < insn_cnt)
3279 env->explored_states[t + 1] = STATE_LIST_MARK;
3281 /* conditional jump with two edges */
3282 env->explored_states[t] = STATE_LIST_MARK;
3283 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3289 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
3296 /* all other non-branch instructions with single
3299 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3307 insn_state[t] = EXPLORED;
3308 if (cur_stack-- <= 0) {
3309 verbose("pop stack internal bug\n");
3316 for (i = 0; i < insn_cnt; i++) {
3317 if (insn_state[i] != EXPLORED) {
3318 verbose("unreachable insn %d\n", i);
3323 ret = 0; /* cfg looks good */
3331 /* check %cur's range satisfies %old's */
3332 static bool range_within(struct bpf_reg_state *old,
3333 struct bpf_reg_state *cur)
3335 return old->umin_value <= cur->umin_value &&
3336 old->umax_value >= cur->umax_value &&
3337 old->smin_value <= cur->smin_value &&
3338 old->smax_value >= cur->smax_value;
3341 /* Maximum number of register states that can exist at once */
3342 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
3348 /* If in the old state two registers had the same id, then they need to have
3349 * the same id in the new state as well. But that id could be different from
3350 * the old state, so we need to track the mapping from old to new ids.
3351 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
3352 * regs with old id 5 must also have new id 9 for the new state to be safe. But
3353 * regs with a different old id could still have new id 9, we don't care about
3355 * So we look through our idmap to see if this old id has been seen before. If
3356 * so, we require the new id to match; otherwise, we add the id pair to the map.
3358 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
3362 for (i = 0; i < ID_MAP_SIZE; i++) {
3363 if (!idmap[i].old) {
3364 /* Reached an empty slot; haven't seen this id before */
3365 idmap[i].old = old_id;
3366 idmap[i].cur = cur_id;
3369 if (idmap[i].old == old_id)
3370 return idmap[i].cur == cur_id;
3372 /* We ran out of idmap slots, which should be impossible */
3377 /* Returns true if (rold safe implies rcur safe) */
3378 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
3379 struct idpair *idmap)
3381 if (!(rold->live & REG_LIVE_READ))
3382 /* explored state didn't use this */
3385 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, live)) == 0)
3388 if (rold->type == NOT_INIT)
3389 /* explored state can't have used this */
3391 if (rcur->type == NOT_INIT)
3393 switch (rold->type) {
3395 if (rcur->type == SCALAR_VALUE) {
3396 /* new val must satisfy old val knowledge */
3397 return range_within(rold, rcur) &&
3398 tnum_in(rold->var_off, rcur->var_off);
3400 /* We're trying to use a pointer in place of a scalar.
3401 * Even if the scalar was unbounded, this could lead to
3402 * pointer leaks because scalars are allowed to leak
3403 * while pointers are not. We could make this safe in
3404 * special cases if root is calling us, but it's
3405 * probably not worth the hassle.
3409 case PTR_TO_MAP_VALUE:
3410 /* If the new min/max/var_off satisfy the old ones and
3411 * everything else matches, we are OK.
3412 * We don't care about the 'id' value, because nothing
3413 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
3415 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
3416 range_within(rold, rcur) &&
3417 tnum_in(rold->var_off, rcur->var_off);
3418 case PTR_TO_MAP_VALUE_OR_NULL:
3419 /* a PTR_TO_MAP_VALUE could be safe to use as a
3420 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
3421 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
3422 * checked, doing so could have affected others with the same
3423 * id, and we can't check for that because we lost the id when
3424 * we converted to a PTR_TO_MAP_VALUE.
3426 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
3428 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
3430 /* Check our ids match any regs they're supposed to */
3431 return check_ids(rold->id, rcur->id, idmap);
3433 if (rcur->type != PTR_TO_PACKET)
3435 /* We must have at least as much range as the old ptr
3436 * did, so that any accesses which were safe before are
3437 * still safe. This is true even if old range < old off,
3438 * since someone could have accessed through (ptr - k), or
3439 * even done ptr -= k in a register, to get a safe access.
3441 if (rold->range > rcur->range)
3443 /* If the offsets don't match, we can't trust our alignment;
3444 * nor can we be sure that we won't fall out of range.
3446 if (rold->off != rcur->off)
3448 /* id relations must be preserved */
3449 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
3451 /* new val must satisfy old val knowledge */
3452 return range_within(rold, rcur) &&
3453 tnum_in(rold->var_off, rcur->var_off);
3455 case CONST_PTR_TO_MAP:
3457 case PTR_TO_PACKET_END:
3458 /* Only valid matches are exact, which memcmp() above
3459 * would have accepted
3462 /* Don't know what's going on, just say it's not safe */
3466 /* Shouldn't get here; if we do, say it's not safe */
3471 /* compare two verifier states
3473 * all states stored in state_list are known to be valid, since
3474 * verifier reached 'bpf_exit' instruction through them
3476 * this function is called when verifier exploring different branches of
3477 * execution popped from the state stack. If it sees an old state that has
3478 * more strict register state and more strict stack state then this execution
3479 * branch doesn't need to be explored further, since verifier already
3480 * concluded that more strict state leads to valid finish.
3482 * Therefore two states are equivalent if register state is more conservative
3483 * and explored stack state is more conservative than the current one.
3486 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
3487 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
3489 * In other words if current stack state (one being explored) has more
3490 * valid slots than old one that already passed validation, it means
3491 * the verifier can stop exploring and conclude that current state is valid too
3493 * Similarly with registers. If explored state has register type as invalid
3494 * whereas register type in current state is meaningful, it means that
3495 * the current state will reach 'bpf_exit' instruction safely
3497 static bool states_equal(struct bpf_verifier_env *env,
3498 struct bpf_verifier_state *old,
3499 struct bpf_verifier_state *cur)
3501 struct idpair *idmap;
3505 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
3506 /* If we failed to allocate the idmap, just say it's not safe */
3510 for (i = 0; i < MAX_BPF_REG; i++) {
3511 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
3515 for (i = 0; i < MAX_BPF_STACK; i++) {
3516 if (old->stack_slot_type[i] == STACK_INVALID)
3518 if (old->stack_slot_type[i] != cur->stack_slot_type[i])
3519 /* Ex: old explored (safe) state has STACK_SPILL in
3520 * this stack slot, but current has has STACK_MISC ->
3521 * this verifier states are not equivalent,
3522 * return false to continue verification of this path
3525 if (i % BPF_REG_SIZE)
3527 if (old->stack_slot_type[i] != STACK_SPILL)
3529 if (!regsafe(&old->spilled_regs[i / BPF_REG_SIZE],
3530 &cur->spilled_regs[i / BPF_REG_SIZE],
3532 /* when explored and current stack slot are both storing
3533 * spilled registers, check that stored pointers types
3534 * are the same as well.
3535 * Ex: explored safe path could have stored
3536 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
3537 * but current path has stored:
3538 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
3539 * such verifier states are not equivalent.
3540 * return false to continue verification of this path
3552 /* A write screens off any subsequent reads; but write marks come from the
3553 * straight-line code between a state and its parent. When we arrive at a
3554 * jump target (in the first iteration of the propagate_liveness() loop),
3555 * we didn't arrive by the straight-line code, so read marks in state must
3556 * propagate to parent regardless of state's write marks.
3558 static bool do_propagate_liveness(const struct bpf_verifier_state *state,
3559 struct bpf_verifier_state *parent)
3561 bool writes = parent == state->parent; /* Observe write marks */
3562 bool touched = false; /* any changes made? */
3567 /* Propagate read liveness of registers... */
3568 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
3569 /* We don't need to worry about FP liveness because it's read-only */
3570 for (i = 0; i < BPF_REG_FP; i++) {
3571 if (parent->regs[i].live & REG_LIVE_READ)
3573 if (writes && (state->regs[i].live & REG_LIVE_WRITTEN))
3575 if (state->regs[i].live & REG_LIVE_READ) {
3576 parent->regs[i].live |= REG_LIVE_READ;
3580 /* ... and stack slots */
3581 for (i = 0; i < MAX_BPF_STACK / BPF_REG_SIZE; i++) {
3582 if (parent->stack_slot_type[i * BPF_REG_SIZE] != STACK_SPILL)
3584 if (state->stack_slot_type[i * BPF_REG_SIZE] != STACK_SPILL)
3586 if (parent->spilled_regs[i].live & REG_LIVE_READ)
3588 if (writes && (state->spilled_regs[i].live & REG_LIVE_WRITTEN))
3590 if (state->spilled_regs[i].live & REG_LIVE_READ) {
3591 parent->spilled_regs[i].live |= REG_LIVE_READ;
3598 /* "parent" is "a state from which we reach the current state", but initially
3599 * it is not the state->parent (i.e. "the state whose straight-line code leads
3600 * to the current state"), instead it is the state that happened to arrive at
3601 * a (prunable) equivalent of the current state. See comment above
3602 * do_propagate_liveness() for consequences of this.
3603 * This function is just a more efficient way of calling mark_reg_read() or
3604 * mark_stack_slot_read() on each reg in "parent" that is read in "state",
3605 * though it requires that parent != state->parent in the call arguments.
3607 static void propagate_liveness(const struct bpf_verifier_state *state,
3608 struct bpf_verifier_state *parent)
3610 while (do_propagate_liveness(state, parent)) {
3611 /* Something changed, so we need to feed those changes onward */
3613 parent = state->parent;
3617 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
3619 struct bpf_verifier_state_list *new_sl;
3620 struct bpf_verifier_state_list *sl;
3623 sl = env->explored_states[insn_idx];
3625 /* this 'insn_idx' instruction wasn't marked, so we will not
3626 * be doing state search here
3630 while (sl != STATE_LIST_MARK) {
3631 if (states_equal(env, &sl->state, &env->cur_state)) {
3632 /* reached equivalent register/stack state,
3634 * Registers read by the continuation are read by us.
3635 * If we have any write marks in env->cur_state, they
3636 * will prevent corresponding reads in the continuation
3637 * from reaching our parent (an explored_state). Our
3638 * own state will get the read marks recorded, but
3639 * they'll be immediately forgotten as we're pruning
3640 * this state and will pop a new one.
3642 propagate_liveness(&sl->state, &env->cur_state);
3648 /* there were no equivalent states, remember current one.
3649 * technically the current state is not proven to be safe yet,
3650 * but it will either reach bpf_exit (which means it's safe) or
3651 * it will be rejected. Since there are no loops, we won't be
3652 * seeing this 'insn_idx' instruction again on the way to bpf_exit
3654 new_sl = kmalloc(sizeof(struct bpf_verifier_state_list), GFP_USER);
3658 /* add new state to the head of linked list */
3659 memcpy(&new_sl->state, &env->cur_state, sizeof(env->cur_state));
3660 new_sl->next = env->explored_states[insn_idx];
3661 env->explored_states[insn_idx] = new_sl;
3662 /* connect new state to parentage chain */
3663 env->cur_state.parent = &new_sl->state;
3664 /* clear write marks in current state: the writes we did are not writes
3665 * our child did, so they don't screen off its reads from us.
3666 * (There are no read marks in current state, because reads always mark
3667 * their parent and current state never has children yet. Only
3668 * explored_states can get read marks.)
3670 for (i = 0; i < BPF_REG_FP; i++)
3671 env->cur_state.regs[i].live = REG_LIVE_NONE;
3672 for (i = 0; i < MAX_BPF_STACK / BPF_REG_SIZE; i++)
3673 if (env->cur_state.stack_slot_type[i * BPF_REG_SIZE] == STACK_SPILL)
3674 env->cur_state.spilled_regs[i].live = REG_LIVE_NONE;
3678 static int ext_analyzer_insn_hook(struct bpf_verifier_env *env,
3679 int insn_idx, int prev_insn_idx)
3681 if (!env->analyzer_ops || !env->analyzer_ops->insn_hook)
3684 return env->analyzer_ops->insn_hook(env, insn_idx, prev_insn_idx);
3687 static int do_check(struct bpf_verifier_env *env)
3689 struct bpf_verifier_state *state = &env->cur_state;
3690 struct bpf_insn *insns = env->prog->insnsi;
3691 struct bpf_reg_state *regs = state->regs;
3692 int insn_cnt = env->prog->len;
3693 int insn_idx, prev_insn_idx = 0;
3694 int insn_processed = 0;
3695 bool do_print_state = false;
3697 init_reg_state(regs);
3698 state->parent = NULL;
3701 struct bpf_insn *insn;
3705 if (insn_idx >= insn_cnt) {
3706 verbose("invalid insn idx %d insn_cnt %d\n",
3707 insn_idx, insn_cnt);
3711 insn = &insns[insn_idx];
3712 class = BPF_CLASS(insn->code);
3714 if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
3715 verbose("BPF program is too large. Processed %d insn\n",
3720 err = is_state_visited(env, insn_idx);
3724 /* found equivalent state, can prune the search */
3727 verbose("\nfrom %d to %d: safe\n",
3728 prev_insn_idx, insn_idx);
3730 verbose("%d: safe\n", insn_idx);
3732 goto process_bpf_exit;
3738 if (log_level > 1 || (log_level && do_print_state)) {
3740 verbose("%d:", insn_idx);
3742 verbose("\nfrom %d to %d:",
3743 prev_insn_idx, insn_idx);
3744 print_verifier_state(&env->cur_state);
3745 do_print_state = false;
3749 verbose("%d: ", insn_idx);
3750 print_bpf_insn(env, insn);
3753 err = ext_analyzer_insn_hook(env, insn_idx, prev_insn_idx);
3757 env->insn_aux_data[insn_idx].seen = true;
3758 if (class == BPF_ALU || class == BPF_ALU64) {
3759 err = check_alu_op(env, insn);
3763 } else if (class == BPF_LDX) {
3764 enum bpf_reg_type *prev_src_type, src_reg_type;
3766 /* check for reserved fields is already done */
3768 /* check src operand */
3769 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3773 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3777 src_reg_type = regs[insn->src_reg].type;
3779 /* check that memory (src_reg + off) is readable,
3780 * the state of dst_reg will be updated by this func
3782 err = check_mem_access(env, insn_idx, insn->src_reg, insn->off,
3783 BPF_SIZE(insn->code), BPF_READ,
3788 prev_src_type = &env->insn_aux_data[insn_idx].ptr_type;
3790 if (*prev_src_type == NOT_INIT) {
3792 * dst_reg = *(u32 *)(src_reg + off)
3793 * save type to validate intersecting paths
3795 *prev_src_type = src_reg_type;
3797 } else if (src_reg_type != *prev_src_type &&
3798 (src_reg_type == PTR_TO_CTX ||
3799 *prev_src_type == PTR_TO_CTX)) {
3800 /* ABuser program is trying to use the same insn
3801 * dst_reg = *(u32*) (src_reg + off)
3802 * with different pointer types:
3803 * src_reg == ctx in one branch and
3804 * src_reg == stack|map in some other branch.
3807 verbose("same insn cannot be used with different pointers\n");
3811 } else if (class == BPF_STX) {
3812 enum bpf_reg_type *prev_dst_type, dst_reg_type;
3814 if (BPF_MODE(insn->code) == BPF_XADD) {
3815 err = check_xadd(env, insn_idx, insn);
3822 /* check src1 operand */
3823 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3826 /* check src2 operand */
3827 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3831 dst_reg_type = regs[insn->dst_reg].type;
3833 /* check that memory (dst_reg + off) is writeable */
3834 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3835 BPF_SIZE(insn->code), BPF_WRITE,
3840 prev_dst_type = &env->insn_aux_data[insn_idx].ptr_type;
3842 if (*prev_dst_type == NOT_INIT) {
3843 *prev_dst_type = dst_reg_type;
3844 } else if (dst_reg_type != *prev_dst_type &&
3845 (dst_reg_type == PTR_TO_CTX ||
3846 *prev_dst_type == PTR_TO_CTX)) {
3847 verbose("same insn cannot be used with different pointers\n");
3851 } else if (class == BPF_ST) {
3852 if (BPF_MODE(insn->code) != BPF_MEM ||
3853 insn->src_reg != BPF_REG_0) {
3854 verbose("BPF_ST uses reserved fields\n");
3857 /* check src operand */
3858 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3862 /* check that memory (dst_reg + off) is writeable */
3863 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3864 BPF_SIZE(insn->code), BPF_WRITE,
3869 } else if (class == BPF_JMP) {
3870 u8 opcode = BPF_OP(insn->code);
3872 if (opcode == BPF_CALL) {
3873 if (BPF_SRC(insn->code) != BPF_K ||
3875 insn->src_reg != BPF_REG_0 ||
3876 insn->dst_reg != BPF_REG_0) {
3877 verbose("BPF_CALL uses reserved fields\n");
3881 err = check_call(env, insn->imm, insn_idx);
3885 } else if (opcode == BPF_JA) {
3886 if (BPF_SRC(insn->code) != BPF_K ||
3888 insn->src_reg != BPF_REG_0 ||
3889 insn->dst_reg != BPF_REG_0) {
3890 verbose("BPF_JA uses reserved fields\n");
3894 insn_idx += insn->off + 1;
3897 } else if (opcode == BPF_EXIT) {
3898 if (BPF_SRC(insn->code) != BPF_K ||
3900 insn->src_reg != BPF_REG_0 ||
3901 insn->dst_reg != BPF_REG_0) {
3902 verbose("BPF_EXIT uses reserved fields\n");
3906 /* eBPF calling convetion is such that R0 is used
3907 * to return the value from eBPF program.
3908 * Make sure that it's readable at this time
3909 * of bpf_exit, which means that program wrote
3910 * something into it earlier
3912 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
3916 if (is_pointer_value(env, BPF_REG_0)) {
3917 verbose("R0 leaks addr as return value\n");
3922 insn_idx = pop_stack(env, &prev_insn_idx);
3926 do_print_state = true;
3930 err = check_cond_jmp_op(env, insn, &insn_idx);
3934 } else if (class == BPF_LD) {
3935 u8 mode = BPF_MODE(insn->code);
3937 if (mode == BPF_ABS || mode == BPF_IND) {
3938 err = check_ld_abs(env, insn);
3942 } else if (mode == BPF_IMM) {
3943 err = check_ld_imm(env, insn);
3948 env->insn_aux_data[insn_idx].seen = true;
3950 verbose("invalid BPF_LD mode\n");
3954 verbose("unknown insn class %d\n", class);
3961 verbose("processed %d insns, stack depth %d\n",
3962 insn_processed, env->prog->aux->stack_depth);
3966 static int check_map_prealloc(struct bpf_map *map)
3968 return (map->map_type != BPF_MAP_TYPE_HASH &&
3969 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
3970 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
3971 !(map->map_flags & BPF_F_NO_PREALLOC);
3974 static int check_map_prog_compatibility(struct bpf_map *map,
3975 struct bpf_prog *prog)
3978 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
3979 * preallocated hash maps, since doing memory allocation
3980 * in overflow_handler can crash depending on where nmi got
3983 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
3984 if (!check_map_prealloc(map)) {
3985 verbose("perf_event programs can only use preallocated hash map\n");
3988 if (map->inner_map_meta &&
3989 !check_map_prealloc(map->inner_map_meta)) {
3990 verbose("perf_event programs can only use preallocated inner hash map\n");
3997 /* look for pseudo eBPF instructions that access map FDs and
3998 * replace them with actual map pointers
4000 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
4002 struct bpf_insn *insn = env->prog->insnsi;
4003 int insn_cnt = env->prog->len;
4006 err = bpf_prog_calc_tag(env->prog);
4010 for (i = 0; i < insn_cnt; i++, insn++) {
4011 if (BPF_CLASS(insn->code) == BPF_LDX &&
4012 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
4013 verbose("BPF_LDX uses reserved fields\n");
4017 if (BPF_CLASS(insn->code) == BPF_STX &&
4018 ((BPF_MODE(insn->code) != BPF_MEM &&
4019 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
4020 verbose("BPF_STX uses reserved fields\n");
4024 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
4025 struct bpf_map *map;
4028 if (i == insn_cnt - 1 || insn[1].code != 0 ||
4029 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
4031 verbose("invalid bpf_ld_imm64 insn\n");
4035 if (insn->src_reg == 0)
4036 /* valid generic load 64-bit imm */
4039 if (insn->src_reg != BPF_PSEUDO_MAP_FD) {
4040 verbose("unrecognized bpf_ld_imm64 insn\n");
4044 f = fdget(insn->imm);
4045 map = __bpf_map_get(f);
4047 verbose("fd %d is not pointing to valid bpf_map\n",
4049 return PTR_ERR(map);
4052 err = check_map_prog_compatibility(map, env->prog);
4058 /* store map pointer inside BPF_LD_IMM64 instruction */
4059 insn[0].imm = (u32) (unsigned long) map;
4060 insn[1].imm = ((u64) (unsigned long) map) >> 32;
4062 /* check whether we recorded this map already */
4063 for (j = 0; j < env->used_map_cnt; j++)
4064 if (env->used_maps[j] == map) {
4069 if (env->used_map_cnt >= MAX_USED_MAPS) {
4074 /* hold the map. If the program is rejected by verifier,
4075 * the map will be released by release_maps() or it
4076 * will be used by the valid program until it's unloaded
4077 * and all maps are released in free_bpf_prog_info()
4079 map = bpf_map_inc(map, false);
4082 return PTR_ERR(map);
4084 env->used_maps[env->used_map_cnt++] = map;
4093 /* now all pseudo BPF_LD_IMM64 instructions load valid
4094 * 'struct bpf_map *' into a register instead of user map_fd.
4095 * These pointers will be used later by verifier to validate map access.
4100 /* drop refcnt of maps used by the rejected program */
4101 static void release_maps(struct bpf_verifier_env *env)
4105 for (i = 0; i < env->used_map_cnt; i++)
4106 bpf_map_put(env->used_maps[i]);
4109 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
4110 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
4112 struct bpf_insn *insn = env->prog->insnsi;
4113 int insn_cnt = env->prog->len;
4116 for (i = 0; i < insn_cnt; i++, insn++)
4117 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
4121 /* single env->prog->insni[off] instruction was replaced with the range
4122 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
4123 * [0, off) and [off, end) to new locations, so the patched range stays zero
4125 static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len,
4128 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
4133 new_data = vzalloc(sizeof(struct bpf_insn_aux_data) * prog_len);
4136 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
4137 memcpy(new_data + off + cnt - 1, old_data + off,
4138 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
4139 for (i = off; i < off + cnt - 1; i++)
4140 new_data[i].seen = true;
4141 env->insn_aux_data = new_data;
4146 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
4147 const struct bpf_insn *patch, u32 len)
4149 struct bpf_prog *new_prog;
4151 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
4154 if (adjust_insn_aux_data(env, new_prog->len, off, len))
4159 /* The verifier does more data flow analysis than llvm and will not explore
4160 * branches that are dead at run time. Malicious programs can have dead code
4161 * too. Therefore replace all dead at-run-time code with nops.
4163 static void sanitize_dead_code(struct bpf_verifier_env *env)
4165 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
4166 struct bpf_insn nop = BPF_MOV64_REG(BPF_REG_0, BPF_REG_0);
4167 struct bpf_insn *insn = env->prog->insnsi;
4168 const int insn_cnt = env->prog->len;
4171 for (i = 0; i < insn_cnt; i++) {
4172 if (aux_data[i].seen)
4174 memcpy(insn + i, &nop, sizeof(nop));
4178 /* convert load instructions that access fields of 'struct __sk_buff'
4179 * into sequence of instructions that access fields of 'struct sk_buff'
4181 static int convert_ctx_accesses(struct bpf_verifier_env *env)
4183 const struct bpf_verifier_ops *ops = env->prog->aux->ops;
4184 int i, cnt, size, ctx_field_size, delta = 0;
4185 const int insn_cnt = env->prog->len;
4186 struct bpf_insn insn_buf[16], *insn;
4187 struct bpf_prog *new_prog;
4188 enum bpf_access_type type;
4189 bool is_narrower_load;
4192 if (ops->gen_prologue) {
4193 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
4195 if (cnt >= ARRAY_SIZE(insn_buf)) {
4196 verbose("bpf verifier is misconfigured\n");
4199 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
4203 env->prog = new_prog;
4208 if (!ops->convert_ctx_access)
4211 insn = env->prog->insnsi + delta;
4213 for (i = 0; i < insn_cnt; i++, insn++) {
4214 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
4215 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
4216 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
4217 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
4219 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
4220 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
4221 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
4222 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
4227 if (env->insn_aux_data[i + delta].ptr_type != PTR_TO_CTX)
4230 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
4231 size = BPF_LDST_BYTES(insn);
4233 /* If the read access is a narrower load of the field,
4234 * convert to a 4/8-byte load, to minimum program type specific
4235 * convert_ctx_access changes. If conversion is successful,
4236 * we will apply proper mask to the result.
4238 is_narrower_load = size < ctx_field_size;
4239 if (is_narrower_load) {
4240 u32 off = insn->off;
4243 if (type == BPF_WRITE) {
4244 verbose("bpf verifier narrow ctx access misconfigured\n");
4249 if (ctx_field_size == 4)
4251 else if (ctx_field_size == 8)
4254 insn->off = off & ~(ctx_field_size - 1);
4255 insn->code = BPF_LDX | BPF_MEM | size_code;
4259 cnt = ops->convert_ctx_access(type, insn, insn_buf, env->prog,
4261 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
4262 (ctx_field_size && !target_size)) {
4263 verbose("bpf verifier is misconfigured\n");
4267 if (is_narrower_load && size < target_size) {
4268 if (ctx_field_size <= 4)
4269 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
4270 (1 << size * 8) - 1);
4272 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
4273 (1 << size * 8) - 1);
4276 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
4282 /* keep walking new program and skip insns we just inserted */
4283 env->prog = new_prog;
4284 insn = new_prog->insnsi + i + delta;
4290 /* fixup insn->imm field of bpf_call instructions
4291 * and inline eligible helpers as explicit sequence of BPF instructions
4293 * this function is called after eBPF program passed verification
4295 static int fixup_bpf_calls(struct bpf_verifier_env *env)
4297 struct bpf_prog *prog = env->prog;
4298 struct bpf_insn *insn = prog->insnsi;
4299 const struct bpf_func_proto *fn;
4300 const int insn_cnt = prog->len;
4301 struct bpf_insn insn_buf[16];
4302 struct bpf_prog *new_prog;
4303 struct bpf_map *map_ptr;
4304 int i, cnt, delta = 0;
4306 for (i = 0; i < insn_cnt; i++, insn++) {
4307 if (insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
4308 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
4309 /* due to JIT bugs clear upper 32-bits of src register
4310 * before div/mod operation
4312 insn_buf[0] = BPF_MOV32_REG(insn->src_reg, insn->src_reg);
4313 insn_buf[1] = *insn;
4315 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
4320 env->prog = prog = new_prog;
4321 insn = new_prog->insnsi + i + delta;
4325 if (insn->code != (BPF_JMP | BPF_CALL))
4328 if (insn->imm == BPF_FUNC_get_route_realm)
4329 prog->dst_needed = 1;
4330 if (insn->imm == BPF_FUNC_get_prandom_u32)
4331 bpf_user_rnd_init_once();
4332 if (insn->imm == BPF_FUNC_tail_call) {
4333 /* If we tail call into other programs, we
4334 * cannot make any assumptions since they can
4335 * be replaced dynamically during runtime in
4336 * the program array.
4338 prog->cb_access = 1;
4339 env->prog->aux->stack_depth = MAX_BPF_STACK;
4341 /* mark bpf_tail_call as different opcode to avoid
4342 * conditional branch in the interpeter for every normal
4343 * call and to prevent accidental JITing by JIT compiler
4344 * that doesn't support bpf_tail_call yet
4347 insn->code = BPF_JMP | BPF_TAIL_CALL;
4349 /* instead of changing every JIT dealing with tail_call
4350 * emit two extra insns:
4351 * if (index >= max_entries) goto out;
4352 * index &= array->index_mask;
4353 * to avoid out-of-bounds cpu speculation
4355 map_ptr = env->insn_aux_data[i + delta].map_ptr;
4356 if (map_ptr == BPF_MAP_PTR_POISON) {
4357 verbose("tail_call obusing map_ptr\n");
4360 if (!map_ptr->unpriv_array)
4362 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
4363 map_ptr->max_entries, 2);
4364 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
4365 container_of(map_ptr,
4368 insn_buf[2] = *insn;
4370 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
4375 env->prog = prog = new_prog;
4376 insn = new_prog->insnsi + i + delta;
4380 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
4381 * handlers are currently limited to 64 bit only.
4383 if (ebpf_jit_enabled() && BITS_PER_LONG == 64 &&
4384 insn->imm == BPF_FUNC_map_lookup_elem) {
4385 map_ptr = env->insn_aux_data[i + delta].map_ptr;
4386 if (map_ptr == BPF_MAP_PTR_POISON ||
4387 !map_ptr->ops->map_gen_lookup)
4388 goto patch_call_imm;
4390 cnt = map_ptr->ops->map_gen_lookup(map_ptr, insn_buf);
4391 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
4392 verbose("bpf verifier is misconfigured\n");
4396 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
4403 /* keep walking new program and skip insns we just inserted */
4404 env->prog = prog = new_prog;
4405 insn = new_prog->insnsi + i + delta;
4409 if (insn->imm == BPF_FUNC_redirect_map) {
4410 /* Note, we cannot use prog directly as imm as subsequent
4411 * rewrites would still change the prog pointer. The only
4412 * stable address we can use is aux, which also works with
4413 * prog clones during blinding.
4415 u64 addr = (unsigned long)prog->aux;
4416 struct bpf_insn r4_ld[] = {
4417 BPF_LD_IMM64(BPF_REG_4, addr),
4420 cnt = ARRAY_SIZE(r4_ld);
4422 new_prog = bpf_patch_insn_data(env, i + delta, r4_ld, cnt);
4427 env->prog = prog = new_prog;
4428 insn = new_prog->insnsi + i + delta;
4431 fn = prog->aux->ops->get_func_proto(insn->imm);
4432 /* all functions that have prototype and verifier allowed
4433 * programs to call them, must be real in-kernel functions
4436 verbose("kernel subsystem misconfigured func %s#%d\n",
4437 func_id_name(insn->imm), insn->imm);
4440 insn->imm = fn->func - __bpf_call_base;
4446 static void free_states(struct bpf_verifier_env *env)
4448 struct bpf_verifier_state_list *sl, *sln;
4451 if (!env->explored_states)
4454 for (i = 0; i < env->prog->len; i++) {
4455 sl = env->explored_states[i];
4458 while (sl != STATE_LIST_MARK) {
4465 kfree(env->explored_states);
4468 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr)
4470 char __user *log_ubuf = NULL;
4471 struct bpf_verifier_env *env;
4474 /* 'struct bpf_verifier_env' can be global, but since it's not small,
4475 * allocate/free it every time bpf_check() is called
4477 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
4481 env->insn_aux_data = vzalloc(sizeof(struct bpf_insn_aux_data) *
4484 if (!env->insn_aux_data)
4488 /* grab the mutex to protect few globals used by verifier */
4489 mutex_lock(&bpf_verifier_lock);
4491 if (attr->log_level || attr->log_buf || attr->log_size) {
4492 /* user requested verbose verifier output
4493 * and supplied buffer to store the verification trace
4495 log_level = attr->log_level;
4496 log_ubuf = (char __user *) (unsigned long) attr->log_buf;
4497 log_size = attr->log_size;
4501 /* log_* values have to be sane */
4502 if (log_size < 128 || log_size > UINT_MAX >> 8 ||
4503 log_level == 0 || log_ubuf == NULL)
4507 log_buf = vmalloc(log_size);
4514 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
4515 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
4516 env->strict_alignment = true;
4518 ret = replace_map_fd_with_map_ptr(env);
4520 goto skip_full_check;
4522 env->explored_states = kcalloc(env->prog->len,
4523 sizeof(struct bpf_verifier_state_list *),
4526 if (!env->explored_states)
4527 goto skip_full_check;
4529 ret = check_cfg(env);
4531 goto skip_full_check;
4533 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
4535 ret = do_check(env);
4538 while (pop_stack(env, NULL) >= 0);
4542 sanitize_dead_code(env);
4545 /* program is valid, convert *(u32*)(ctx + off) accesses */
4546 ret = convert_ctx_accesses(env);
4549 ret = fixup_bpf_calls(env);
4551 if (log_level && log_len >= log_size - 1) {
4552 BUG_ON(log_len >= log_size);
4553 /* verifier log exceeded user supplied buffer */
4555 /* fall through to return what was recorded */
4558 /* copy verifier log back to user space including trailing zero */
4559 if (log_level && copy_to_user(log_ubuf, log_buf, log_len + 1) != 0) {
4564 if (ret == 0 && env->used_map_cnt) {
4565 /* if program passed verifier, update used_maps in bpf_prog_info */
4566 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
4567 sizeof(env->used_maps[0]),
4570 if (!env->prog->aux->used_maps) {
4575 memcpy(env->prog->aux->used_maps, env->used_maps,
4576 sizeof(env->used_maps[0]) * env->used_map_cnt);
4577 env->prog->aux->used_map_cnt = env->used_map_cnt;
4579 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
4580 * bpf_ld_imm64 instructions
4582 convert_pseudo_ld_imm64(env);
4588 if (!env->prog->aux->used_maps)
4589 /* if we didn't copy map pointers into bpf_prog_info, release
4590 * them now. Otherwise free_bpf_prog_info() will release them.
4595 mutex_unlock(&bpf_verifier_lock);
4596 vfree(env->insn_aux_data);
4602 int bpf_analyzer(struct bpf_prog *prog, const struct bpf_ext_analyzer_ops *ops,
4605 struct bpf_verifier_env *env;
4608 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
4612 env->insn_aux_data = vzalloc(sizeof(struct bpf_insn_aux_data) *
4615 if (!env->insn_aux_data)
4618 env->analyzer_ops = ops;
4619 env->analyzer_priv = priv;
4621 /* grab the mutex to protect few globals used by verifier */
4622 mutex_lock(&bpf_verifier_lock);
4626 env->strict_alignment = false;
4627 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
4628 env->strict_alignment = true;
4630 env->explored_states = kcalloc(env->prog->len,
4631 sizeof(struct bpf_verifier_state_list *),
4634 if (!env->explored_states)
4635 goto skip_full_check;
4637 ret = check_cfg(env);
4639 goto skip_full_check;
4641 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
4643 ret = do_check(env);
4646 while (pop_stack(env, NULL) >= 0);
4649 mutex_unlock(&bpf_verifier_lock);
4650 vfree(env->insn_aux_data);
4655 EXPORT_SYMBOL_GPL(bpf_analyzer);