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 bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
991 const struct bpf_reg_state *reg = &env->cur_state.regs[regno];
993 return reg->type == PTR_TO_CTX;
996 static int check_pkt_ptr_alignment(const struct bpf_reg_state *reg,
997 int off, int size, bool strict)
1002 /* Byte size accesses are always allowed. */
1003 if (!strict || size == 1)
1006 /* For platforms that do not have a Kconfig enabling
1007 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
1008 * NET_IP_ALIGN is universally set to '2'. And on platforms
1009 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
1010 * to this code only in strict mode where we want to emulate
1011 * the NET_IP_ALIGN==2 checking. Therefore use an
1012 * unconditional IP align value of '2'.
1016 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
1017 if (!tnum_is_aligned(reg_off, size)) {
1020 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1021 verbose("misaligned packet access off %d+%s+%d+%d size %d\n",
1022 ip_align, tn_buf, reg->off, off, size);
1029 static int check_generic_ptr_alignment(const struct bpf_reg_state *reg,
1030 const char *pointer_desc,
1031 int off, int size, bool strict)
1033 struct tnum reg_off;
1035 /* Byte size accesses are always allowed. */
1036 if (!strict || size == 1)
1039 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
1040 if (!tnum_is_aligned(reg_off, size)) {
1043 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1044 verbose("misaligned %saccess off %s+%d+%d size %d\n",
1045 pointer_desc, tn_buf, reg->off, off, size);
1052 static int check_ptr_alignment(struct bpf_verifier_env *env,
1053 const struct bpf_reg_state *reg,
1056 bool strict = env->strict_alignment;
1057 const char *pointer_desc = "";
1059 switch (reg->type) {
1061 /* special case, because of NET_IP_ALIGN */
1062 return check_pkt_ptr_alignment(reg, off, size, strict);
1063 case PTR_TO_MAP_VALUE:
1064 pointer_desc = "value ";
1067 pointer_desc = "context ";
1070 pointer_desc = "stack ";
1071 /* The stack spill tracking logic in check_stack_write()
1072 * and check_stack_read() relies on stack accesses being
1080 return check_generic_ptr_alignment(reg, pointer_desc, off, size, strict);
1083 /* truncate register to smaller size (in bytes)
1084 * must be called with size < BPF_REG_SIZE
1086 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
1090 /* clear high bits in bit representation */
1091 reg->var_off = tnum_cast(reg->var_off, size);
1093 /* fix arithmetic bounds */
1094 mask = ((u64)1 << (size * 8)) - 1;
1095 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
1096 reg->umin_value &= mask;
1097 reg->umax_value &= mask;
1099 reg->umin_value = 0;
1100 reg->umax_value = mask;
1102 reg->smin_value = reg->umin_value;
1103 reg->smax_value = reg->umax_value;
1106 /* check whether memory at (regno + off) is accessible for t = (read | write)
1107 * if t==write, value_regno is a register which value is stored into memory
1108 * if t==read, value_regno is a register which will receive the value from memory
1109 * if t==write && value_regno==-1, some unknown value is stored into memory
1110 * if t==read && value_regno==-1, don't care what we read from memory
1112 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno, int off,
1113 int bpf_size, enum bpf_access_type t,
1116 struct bpf_verifier_state *state = &env->cur_state;
1117 struct bpf_reg_state *reg = &state->regs[regno];
1120 size = bpf_size_to_bytes(bpf_size);
1124 /* alignment checks will add in reg->off themselves */
1125 err = check_ptr_alignment(env, reg, off, size);
1129 /* for access checks, reg->off is just part of off */
1132 if (reg->type == PTR_TO_MAP_VALUE) {
1133 if (t == BPF_WRITE && value_regno >= 0 &&
1134 is_pointer_value(env, value_regno)) {
1135 verbose("R%d leaks addr into map\n", value_regno);
1139 err = check_map_access(env, regno, off, size);
1140 if (!err && t == BPF_READ && value_regno >= 0)
1141 mark_reg_unknown(state->regs, value_regno);
1143 } else if (reg->type == PTR_TO_CTX) {
1144 enum bpf_reg_type reg_type = SCALAR_VALUE;
1146 if (t == BPF_WRITE && value_regno >= 0 &&
1147 is_pointer_value(env, value_regno)) {
1148 verbose("R%d leaks addr into ctx\n", value_regno);
1151 /* ctx accesses must be at a fixed offset, so that we can
1152 * determine what type of data were returned.
1155 verbose("dereference of modified ctx ptr R%d off=%d+%d, ctx+const is allowed, ctx+const+const is not\n",
1156 regno, reg->off, off - reg->off);
1159 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
1162 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1163 verbose("variable ctx access var_off=%s off=%d size=%d",
1167 err = check_ctx_access(env, insn_idx, off, size, t, ®_type);
1168 if (!err && t == BPF_READ && value_regno >= 0) {
1169 /* ctx access returns either a scalar, or a
1170 * PTR_TO_PACKET[_END]. In the latter case, we know
1171 * the offset is zero.
1173 if (reg_type == SCALAR_VALUE)
1174 mark_reg_unknown(state->regs, value_regno);
1176 mark_reg_known_zero(state->regs, value_regno);
1177 state->regs[value_regno].id = 0;
1178 state->regs[value_regno].off = 0;
1179 state->regs[value_regno].range = 0;
1180 state->regs[value_regno].type = reg_type;
1183 } else if (reg->type == PTR_TO_STACK) {
1184 /* stack accesses must be at a fixed offset, so that we can
1185 * determine what type of data were returned.
1186 * See check_stack_read().
1188 if (!tnum_is_const(reg->var_off)) {
1191 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1192 verbose("variable stack access var_off=%s off=%d size=%d",
1196 off += reg->var_off.value;
1197 if (off >= 0 || off < -MAX_BPF_STACK) {
1198 verbose("invalid stack off=%d size=%d\n", off, size);
1202 if (env->prog->aux->stack_depth < -off)
1203 env->prog->aux->stack_depth = -off;
1205 if (t == BPF_WRITE) {
1206 if (!env->allow_ptr_leaks &&
1207 state->stack_slot_type[MAX_BPF_STACK + off] == STACK_SPILL &&
1208 size != BPF_REG_SIZE) {
1209 verbose("attempt to corrupt spilled pointer on stack\n");
1212 err = check_stack_write(state, off, size, value_regno);
1214 err = check_stack_read(state, off, size, value_regno);
1216 } else if (reg->type == PTR_TO_PACKET) {
1217 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
1218 verbose("cannot write into packet\n");
1221 if (t == BPF_WRITE && value_regno >= 0 &&
1222 is_pointer_value(env, value_regno)) {
1223 verbose("R%d leaks addr into packet\n", value_regno);
1226 err = check_packet_access(env, regno, off, size);
1227 if (!err && t == BPF_READ && value_regno >= 0)
1228 mark_reg_unknown(state->regs, value_regno);
1230 verbose("R%d invalid mem access '%s'\n",
1231 regno, reg_type_str[reg->type]);
1235 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
1236 state->regs[value_regno].type == SCALAR_VALUE) {
1237 /* b/h/w load zero-extends, mark upper bits as known 0 */
1238 coerce_reg_to_size(&state->regs[value_regno], size);
1243 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
1247 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
1249 verbose("BPF_XADD uses reserved fields\n");
1253 /* check src1 operand */
1254 err = check_reg_arg(env, insn->src_reg, SRC_OP);
1258 /* check src2 operand */
1259 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
1263 if (is_pointer_value(env, insn->src_reg)) {
1264 verbose("R%d leaks addr into mem\n", insn->src_reg);
1268 if (is_ctx_reg(env, insn->dst_reg)) {
1269 verbose("BPF_XADD stores into R%d context is not allowed\n",
1274 /* check whether atomic_add can read the memory */
1275 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1276 BPF_SIZE(insn->code), BPF_READ, -1);
1280 /* check whether atomic_add can write into the same memory */
1281 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1282 BPF_SIZE(insn->code), BPF_WRITE, -1);
1285 /* Does this register contain a constant zero? */
1286 static bool register_is_null(struct bpf_reg_state reg)
1288 return reg.type == SCALAR_VALUE && tnum_equals_const(reg.var_off, 0);
1291 /* when register 'regno' is passed into function that will read 'access_size'
1292 * bytes from that pointer, make sure that it's within stack boundary
1293 * and all elements of stack are initialized.
1294 * Unlike most pointer bounds-checking functions, this one doesn't take an
1295 * 'off' argument, so it has to add in reg->off itself.
1297 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
1298 int access_size, bool zero_size_allowed,
1299 struct bpf_call_arg_meta *meta)
1301 struct bpf_verifier_state *state = &env->cur_state;
1302 struct bpf_reg_state *regs = state->regs;
1305 if (regs[regno].type != PTR_TO_STACK) {
1306 /* Allow zero-byte read from NULL, regardless of pointer type */
1307 if (zero_size_allowed && access_size == 0 &&
1308 register_is_null(regs[regno]))
1311 verbose("R%d type=%s expected=%s\n", regno,
1312 reg_type_str[regs[regno].type],
1313 reg_type_str[PTR_TO_STACK]);
1317 /* Only allow fixed-offset stack reads */
1318 if (!tnum_is_const(regs[regno].var_off)) {
1321 tnum_strn(tn_buf, sizeof(tn_buf), regs[regno].var_off);
1322 verbose("invalid variable stack read R%d var_off=%s\n",
1326 off = regs[regno].off + regs[regno].var_off.value;
1327 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
1329 verbose("invalid stack type R%d off=%d access_size=%d\n",
1330 regno, off, access_size);
1334 if (env->prog->aux->stack_depth < -off)
1335 env->prog->aux->stack_depth = -off;
1337 if (meta && meta->raw_mode) {
1338 meta->access_size = access_size;
1339 meta->regno = regno;
1343 for (i = 0; i < access_size; i++) {
1344 if (state->stack_slot_type[MAX_BPF_STACK + off + i] != STACK_MISC) {
1345 verbose("invalid indirect read from stack off %d+%d size %d\n",
1346 off, i, access_size);
1353 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
1354 int access_size, bool zero_size_allowed,
1355 struct bpf_call_arg_meta *meta)
1357 struct bpf_reg_state *regs = env->cur_state.regs, *reg = ®s[regno];
1359 switch (reg->type) {
1361 return check_packet_access(env, regno, reg->off, access_size);
1362 case PTR_TO_MAP_VALUE:
1363 return check_map_access(env, regno, reg->off, access_size);
1364 default: /* scalar_value|ptr_to_stack or invalid ptr */
1365 return check_stack_boundary(env, regno, access_size,
1366 zero_size_allowed, meta);
1370 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
1371 enum bpf_arg_type arg_type,
1372 struct bpf_call_arg_meta *meta)
1374 struct bpf_reg_state *regs = env->cur_state.regs, *reg = ®s[regno];
1375 enum bpf_reg_type expected_type, type = reg->type;
1378 if (arg_type == ARG_DONTCARE)
1381 err = check_reg_arg(env, regno, SRC_OP);
1385 if (arg_type == ARG_ANYTHING) {
1386 if (is_pointer_value(env, regno)) {
1387 verbose("R%d leaks addr into helper function\n", regno);
1393 if (type == PTR_TO_PACKET &&
1394 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
1395 verbose("helper access to the packet is not allowed\n");
1399 if (arg_type == ARG_PTR_TO_MAP_KEY ||
1400 arg_type == ARG_PTR_TO_MAP_VALUE) {
1401 expected_type = PTR_TO_STACK;
1402 if (type != PTR_TO_PACKET && type != expected_type)
1404 } else if (arg_type == ARG_CONST_SIZE ||
1405 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1406 expected_type = SCALAR_VALUE;
1407 if (type != expected_type)
1409 } else if (arg_type == ARG_CONST_MAP_PTR) {
1410 expected_type = CONST_PTR_TO_MAP;
1411 if (type != expected_type)
1413 } else if (arg_type == ARG_PTR_TO_CTX) {
1414 expected_type = PTR_TO_CTX;
1415 if (type != expected_type)
1417 } else if (arg_type == ARG_PTR_TO_MEM ||
1418 arg_type == ARG_PTR_TO_UNINIT_MEM) {
1419 expected_type = PTR_TO_STACK;
1420 /* One exception here. In case function allows for NULL to be
1421 * passed in as argument, it's a SCALAR_VALUE type. Final test
1422 * happens during stack boundary checking.
1424 if (register_is_null(*reg))
1425 /* final test in check_stack_boundary() */;
1426 else if (type != PTR_TO_PACKET && type != PTR_TO_MAP_VALUE &&
1427 type != expected_type)
1429 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
1431 verbose("unsupported arg_type %d\n", arg_type);
1435 if (arg_type == ARG_CONST_MAP_PTR) {
1436 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
1437 meta->map_ptr = reg->map_ptr;
1438 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
1439 /* bpf_map_xxx(..., map_ptr, ..., key) call:
1440 * check that [key, key + map->key_size) are within
1441 * stack limits and initialized
1443 if (!meta->map_ptr) {
1444 /* in function declaration map_ptr must come before
1445 * map_key, so that it's verified and known before
1446 * we have to check map_key here. Otherwise it means
1447 * that kernel subsystem misconfigured verifier
1449 verbose("invalid map_ptr to access map->key\n");
1452 if (type == PTR_TO_PACKET)
1453 err = check_packet_access(env, regno, reg->off,
1454 meta->map_ptr->key_size);
1456 err = check_stack_boundary(env, regno,
1457 meta->map_ptr->key_size,
1459 } else if (arg_type == ARG_PTR_TO_MAP_VALUE) {
1460 /* bpf_map_xxx(..., map_ptr, ..., value) call:
1461 * check [value, value + map->value_size) validity
1463 if (!meta->map_ptr) {
1464 /* kernel subsystem misconfigured verifier */
1465 verbose("invalid map_ptr to access map->value\n");
1468 if (type == PTR_TO_PACKET)
1469 err = check_packet_access(env, regno, reg->off,
1470 meta->map_ptr->value_size);
1472 err = check_stack_boundary(env, regno,
1473 meta->map_ptr->value_size,
1475 } else if (arg_type == ARG_CONST_SIZE ||
1476 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1477 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
1479 /* bpf_xxx(..., buf, len) call will access 'len' bytes
1480 * from stack pointer 'buf'. Check it
1481 * note: regno == len, regno - 1 == buf
1484 /* kernel subsystem misconfigured verifier */
1485 verbose("ARG_CONST_SIZE cannot be first argument\n");
1489 /* The register is SCALAR_VALUE; the access check
1490 * happens using its boundaries.
1493 if (!tnum_is_const(reg->var_off))
1494 /* For unprivileged variable accesses, disable raw
1495 * mode so that the program is required to
1496 * initialize all the memory that the helper could
1497 * just partially fill up.
1501 if (reg->smin_value < 0) {
1502 verbose("R%d min value is negative, either use unsigned or 'var &= const'\n",
1507 if (reg->umin_value == 0) {
1508 err = check_helper_mem_access(env, regno - 1, 0,
1515 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
1516 verbose("R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
1520 err = check_helper_mem_access(env, regno - 1,
1522 zero_size_allowed, meta);
1527 verbose("R%d type=%s expected=%s\n", regno,
1528 reg_type_str[type], reg_type_str[expected_type]);
1532 static int check_map_func_compatibility(struct bpf_map *map, int func_id)
1537 /* We need a two way check, first is from map perspective ... */
1538 switch (map->map_type) {
1539 case BPF_MAP_TYPE_PROG_ARRAY:
1540 if (func_id != BPF_FUNC_tail_call)
1543 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
1544 if (func_id != BPF_FUNC_perf_event_read &&
1545 func_id != BPF_FUNC_perf_event_output)
1548 case BPF_MAP_TYPE_STACK_TRACE:
1549 if (func_id != BPF_FUNC_get_stackid)
1552 case BPF_MAP_TYPE_CGROUP_ARRAY:
1553 if (func_id != BPF_FUNC_skb_under_cgroup &&
1554 func_id != BPF_FUNC_current_task_under_cgroup)
1557 /* devmap returns a pointer to a live net_device ifindex that we cannot
1558 * allow to be modified from bpf side. So do not allow lookup elements
1561 case BPF_MAP_TYPE_DEVMAP:
1562 if (func_id != BPF_FUNC_redirect_map)
1565 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
1566 case BPF_MAP_TYPE_HASH_OF_MAPS:
1567 if (func_id != BPF_FUNC_map_lookup_elem)
1570 case BPF_MAP_TYPE_SOCKMAP:
1571 if (func_id != BPF_FUNC_sk_redirect_map &&
1572 func_id != BPF_FUNC_sock_map_update &&
1573 func_id != BPF_FUNC_map_delete_elem)
1580 /* ... and second from the function itself. */
1582 case BPF_FUNC_tail_call:
1583 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
1586 case BPF_FUNC_perf_event_read:
1587 case BPF_FUNC_perf_event_output:
1588 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
1591 case BPF_FUNC_get_stackid:
1592 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
1595 case BPF_FUNC_current_task_under_cgroup:
1596 case BPF_FUNC_skb_under_cgroup:
1597 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
1600 case BPF_FUNC_redirect_map:
1601 if (map->map_type != BPF_MAP_TYPE_DEVMAP)
1604 case BPF_FUNC_sk_redirect_map:
1605 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
1608 case BPF_FUNC_sock_map_update:
1609 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
1618 verbose("cannot pass map_type %d into func %s#%d\n",
1619 map->map_type, func_id_name(func_id), func_id);
1623 static int check_raw_mode(const struct bpf_func_proto *fn)
1627 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
1629 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
1631 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
1633 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
1635 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
1638 return count > 1 ? -EINVAL : 0;
1641 /* Packet data might have moved, any old PTR_TO_PACKET[_END] are now invalid,
1642 * so turn them into unknown SCALAR_VALUE.
1644 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
1646 struct bpf_verifier_state *state = &env->cur_state;
1647 struct bpf_reg_state *regs = state->regs, *reg;
1650 for (i = 0; i < MAX_BPF_REG; i++)
1651 if (regs[i].type == PTR_TO_PACKET ||
1652 regs[i].type == PTR_TO_PACKET_END)
1653 mark_reg_unknown(regs, i);
1655 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
1656 if (state->stack_slot_type[i] != STACK_SPILL)
1658 reg = &state->spilled_regs[i / BPF_REG_SIZE];
1659 if (reg->type != PTR_TO_PACKET &&
1660 reg->type != PTR_TO_PACKET_END)
1662 __mark_reg_unknown(reg);
1666 static int check_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
1668 struct bpf_verifier_state *state = &env->cur_state;
1669 const struct bpf_func_proto *fn = NULL;
1670 struct bpf_reg_state *regs = state->regs;
1671 struct bpf_call_arg_meta meta;
1675 /* find function prototype */
1676 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
1677 verbose("invalid func %s#%d\n", func_id_name(func_id), func_id);
1681 if (env->prog->aux->ops->get_func_proto)
1682 fn = env->prog->aux->ops->get_func_proto(func_id);
1685 verbose("unknown func %s#%d\n", func_id_name(func_id), func_id);
1689 /* eBPF programs must be GPL compatible to use GPL-ed functions */
1690 if (!env->prog->gpl_compatible && fn->gpl_only) {
1691 verbose("cannot call GPL only function from proprietary program\n");
1695 changes_data = bpf_helper_changes_pkt_data(fn->func);
1697 memset(&meta, 0, sizeof(meta));
1698 meta.pkt_access = fn->pkt_access;
1700 /* We only support one arg being in raw mode at the moment, which
1701 * is sufficient for the helper functions we have right now.
1703 err = check_raw_mode(fn);
1705 verbose("kernel subsystem misconfigured func %s#%d\n",
1706 func_id_name(func_id), func_id);
1711 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
1714 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
1717 if (func_id == BPF_FUNC_tail_call) {
1718 if (meta.map_ptr == NULL) {
1719 verbose("verifier bug\n");
1722 env->insn_aux_data[insn_idx].map_ptr = meta.map_ptr;
1724 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
1727 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
1730 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
1734 /* Mark slots with STACK_MISC in case of raw mode, stack offset
1735 * is inferred from register state.
1737 for (i = 0; i < meta.access_size; i++) {
1738 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B, BPF_WRITE, -1);
1743 /* reset caller saved regs */
1744 for (i = 0; i < CALLER_SAVED_REGS; i++) {
1745 mark_reg_not_init(regs, caller_saved[i]);
1746 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
1749 /* update return register (already marked as written above) */
1750 if (fn->ret_type == RET_INTEGER) {
1751 /* sets type to SCALAR_VALUE */
1752 mark_reg_unknown(regs, BPF_REG_0);
1753 } else if (fn->ret_type == RET_VOID) {
1754 regs[BPF_REG_0].type = NOT_INIT;
1755 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL) {
1756 struct bpf_insn_aux_data *insn_aux;
1758 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
1759 /* There is no offset yet applied, variable or fixed */
1760 mark_reg_known_zero(regs, BPF_REG_0);
1761 regs[BPF_REG_0].off = 0;
1762 /* remember map_ptr, so that check_map_access()
1763 * can check 'value_size' boundary of memory access
1764 * to map element returned from bpf_map_lookup_elem()
1766 if (meta.map_ptr == NULL) {
1767 verbose("kernel subsystem misconfigured verifier\n");
1770 regs[BPF_REG_0].map_ptr = meta.map_ptr;
1771 regs[BPF_REG_0].id = ++env->id_gen;
1772 insn_aux = &env->insn_aux_data[insn_idx];
1773 if (!insn_aux->map_ptr)
1774 insn_aux->map_ptr = meta.map_ptr;
1775 else if (insn_aux->map_ptr != meta.map_ptr)
1776 insn_aux->map_ptr = BPF_MAP_PTR_POISON;
1778 verbose("unknown return type %d of func %s#%d\n",
1779 fn->ret_type, func_id_name(func_id), func_id);
1783 err = check_map_func_compatibility(meta.map_ptr, func_id);
1788 clear_all_pkt_pointers(env);
1792 static bool signed_add_overflows(s64 a, s64 b)
1794 /* Do the add in u64, where overflow is well-defined */
1795 s64 res = (s64)((u64)a + (u64)b);
1802 static bool signed_sub_overflows(s64 a, s64 b)
1804 /* Do the sub in u64, where overflow is well-defined */
1805 s64 res = (s64)((u64)a - (u64)b);
1812 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
1813 const struct bpf_reg_state *reg,
1814 enum bpf_reg_type type)
1816 bool known = tnum_is_const(reg->var_off);
1817 s64 val = reg->var_off.value;
1818 s64 smin = reg->smin_value;
1820 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
1821 verbose("math between %s pointer and %lld is not allowed\n",
1822 reg_type_str[type], val);
1826 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
1827 verbose("%s pointer offset %d is not allowed\n",
1828 reg_type_str[type], reg->off);
1832 if (smin == S64_MIN) {
1833 verbose("math between %s pointer and register with unbounded min value is not allowed\n",
1834 reg_type_str[type]);
1838 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
1839 verbose("value %lld makes %s pointer be out of bounds\n",
1840 smin, reg_type_str[type]);
1847 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
1848 * Caller should also handle BPF_MOV case separately.
1849 * If we return -EACCES, caller may want to try again treating pointer as a
1850 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
1852 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
1853 struct bpf_insn *insn,
1854 const struct bpf_reg_state *ptr_reg,
1855 const struct bpf_reg_state *off_reg)
1857 struct bpf_reg_state *regs = env->cur_state.regs, *dst_reg;
1858 bool known = tnum_is_const(off_reg->var_off);
1859 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
1860 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
1861 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
1862 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
1863 u8 opcode = BPF_OP(insn->code);
1864 u32 dst = insn->dst_reg;
1866 dst_reg = ®s[dst];
1868 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
1869 smin_val > smax_val || umin_val > umax_val) {
1870 /* Taint dst register if offset had invalid bounds derived from
1871 * e.g. dead branches.
1873 __mark_reg_unknown(dst_reg);
1877 if (BPF_CLASS(insn->code) != BPF_ALU64) {
1878 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
1879 if (!env->allow_ptr_leaks)
1880 verbose("R%d 32-bit pointer arithmetic prohibited\n",
1885 if (ptr_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
1886 if (!env->allow_ptr_leaks)
1887 verbose("R%d pointer arithmetic on PTR_TO_MAP_VALUE_OR_NULL prohibited, null-check it first\n",
1891 if (ptr_reg->type == CONST_PTR_TO_MAP) {
1892 if (!env->allow_ptr_leaks)
1893 verbose("R%d pointer arithmetic on CONST_PTR_TO_MAP prohibited\n",
1897 if (ptr_reg->type == PTR_TO_PACKET_END) {
1898 if (!env->allow_ptr_leaks)
1899 verbose("R%d pointer arithmetic on PTR_TO_PACKET_END prohibited\n",
1904 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
1905 * The id may be overwritten later if we create a new variable offset.
1907 dst_reg->type = ptr_reg->type;
1908 dst_reg->id = ptr_reg->id;
1910 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
1911 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
1916 /* We can take a fixed offset as long as it doesn't overflow
1917 * the s32 'off' field
1919 if (known && (ptr_reg->off + smin_val ==
1920 (s64)(s32)(ptr_reg->off + smin_val))) {
1921 /* pointer += K. Accumulate it into fixed offset */
1922 dst_reg->smin_value = smin_ptr;
1923 dst_reg->smax_value = smax_ptr;
1924 dst_reg->umin_value = umin_ptr;
1925 dst_reg->umax_value = umax_ptr;
1926 dst_reg->var_off = ptr_reg->var_off;
1927 dst_reg->off = ptr_reg->off + smin_val;
1928 dst_reg->range = ptr_reg->range;
1931 /* A new variable offset is created. Note that off_reg->off
1932 * == 0, since it's a scalar.
1933 * dst_reg gets the pointer type and since some positive
1934 * integer value was added to the pointer, give it a new 'id'
1935 * if it's a PTR_TO_PACKET.
1936 * this creates a new 'base' pointer, off_reg (variable) gets
1937 * added into the variable offset, and we copy the fixed offset
1940 if (signed_add_overflows(smin_ptr, smin_val) ||
1941 signed_add_overflows(smax_ptr, smax_val)) {
1942 dst_reg->smin_value = S64_MIN;
1943 dst_reg->smax_value = S64_MAX;
1945 dst_reg->smin_value = smin_ptr + smin_val;
1946 dst_reg->smax_value = smax_ptr + smax_val;
1948 if (umin_ptr + umin_val < umin_ptr ||
1949 umax_ptr + umax_val < umax_ptr) {
1950 dst_reg->umin_value = 0;
1951 dst_reg->umax_value = U64_MAX;
1953 dst_reg->umin_value = umin_ptr + umin_val;
1954 dst_reg->umax_value = umax_ptr + umax_val;
1956 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
1957 dst_reg->off = ptr_reg->off;
1958 if (ptr_reg->type == PTR_TO_PACKET) {
1959 dst_reg->id = ++env->id_gen;
1960 /* something was added to pkt_ptr, set range to zero */
1965 if (dst_reg == off_reg) {
1966 /* scalar -= pointer. Creates an unknown scalar */
1967 if (!env->allow_ptr_leaks)
1968 verbose("R%d tried to subtract pointer from scalar\n",
1972 /* We don't allow subtraction from FP, because (according to
1973 * test_verifier.c test "invalid fp arithmetic", JITs might not
1974 * be able to deal with it.
1976 if (ptr_reg->type == PTR_TO_STACK) {
1977 if (!env->allow_ptr_leaks)
1978 verbose("R%d subtraction from stack pointer prohibited\n",
1982 if (known && (ptr_reg->off - smin_val ==
1983 (s64)(s32)(ptr_reg->off - smin_val))) {
1984 /* pointer -= K. Subtract it from fixed offset */
1985 dst_reg->smin_value = smin_ptr;
1986 dst_reg->smax_value = smax_ptr;
1987 dst_reg->umin_value = umin_ptr;
1988 dst_reg->umax_value = umax_ptr;
1989 dst_reg->var_off = ptr_reg->var_off;
1990 dst_reg->id = ptr_reg->id;
1991 dst_reg->off = ptr_reg->off - smin_val;
1992 dst_reg->range = ptr_reg->range;
1995 /* A new variable offset is created. If the subtrahend is known
1996 * nonnegative, then any reg->range we had before is still good.
1998 if (signed_sub_overflows(smin_ptr, smax_val) ||
1999 signed_sub_overflows(smax_ptr, smin_val)) {
2000 /* Overflow possible, we know nothing */
2001 dst_reg->smin_value = S64_MIN;
2002 dst_reg->smax_value = S64_MAX;
2004 dst_reg->smin_value = smin_ptr - smax_val;
2005 dst_reg->smax_value = smax_ptr - smin_val;
2007 if (umin_ptr < umax_val) {
2008 /* Overflow possible, we know nothing */
2009 dst_reg->umin_value = 0;
2010 dst_reg->umax_value = U64_MAX;
2012 /* Cannot overflow (as long as bounds are consistent) */
2013 dst_reg->umin_value = umin_ptr - umax_val;
2014 dst_reg->umax_value = umax_ptr - umin_val;
2016 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
2017 dst_reg->off = ptr_reg->off;
2018 if (ptr_reg->type == PTR_TO_PACKET) {
2019 dst_reg->id = ++env->id_gen;
2020 /* something was added to pkt_ptr, set range to zero */
2028 /* bitwise ops on pointers are troublesome, prohibit for now.
2029 * (However, in principle we could allow some cases, e.g.
2030 * ptr &= ~3 which would reduce min_value by 3.)
2032 if (!env->allow_ptr_leaks)
2033 verbose("R%d bitwise operator %s on pointer prohibited\n",
2034 dst, bpf_alu_string[opcode >> 4]);
2037 /* other operators (e.g. MUL,LSH) produce non-pointer results */
2038 if (!env->allow_ptr_leaks)
2039 verbose("R%d pointer arithmetic with %s operator prohibited\n",
2040 dst, bpf_alu_string[opcode >> 4]);
2044 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
2047 __update_reg_bounds(dst_reg);
2048 __reg_deduce_bounds(dst_reg);
2049 __reg_bound_offset(dst_reg);
2053 /* WARNING: This function does calculations on 64-bit values, but the actual
2054 * execution may occur on 32-bit values. Therefore, things like bitshifts
2055 * need extra checks in the 32-bit case.
2057 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
2058 struct bpf_insn *insn,
2059 struct bpf_reg_state *dst_reg,
2060 struct bpf_reg_state src_reg)
2062 struct bpf_reg_state *regs = env->cur_state.regs;
2063 u8 opcode = BPF_OP(insn->code);
2064 bool src_known, dst_known;
2065 s64 smin_val, smax_val;
2066 u64 umin_val, umax_val;
2067 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
2069 smin_val = src_reg.smin_value;
2070 smax_val = src_reg.smax_value;
2071 umin_val = src_reg.umin_value;
2072 umax_val = src_reg.umax_value;
2073 src_known = tnum_is_const(src_reg.var_off);
2074 dst_known = tnum_is_const(dst_reg->var_off);
2076 if ((src_known && (smin_val != smax_val || umin_val != umax_val)) ||
2077 smin_val > smax_val || umin_val > umax_val) {
2078 /* Taint dst register if offset had invalid bounds derived from
2079 * e.g. dead branches.
2081 __mark_reg_unknown(dst_reg);
2086 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
2087 __mark_reg_unknown(dst_reg);
2093 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
2094 signed_add_overflows(dst_reg->smax_value, smax_val)) {
2095 dst_reg->smin_value = S64_MIN;
2096 dst_reg->smax_value = S64_MAX;
2098 dst_reg->smin_value += smin_val;
2099 dst_reg->smax_value += smax_val;
2101 if (dst_reg->umin_value + umin_val < umin_val ||
2102 dst_reg->umax_value + umax_val < umax_val) {
2103 dst_reg->umin_value = 0;
2104 dst_reg->umax_value = U64_MAX;
2106 dst_reg->umin_value += umin_val;
2107 dst_reg->umax_value += umax_val;
2109 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
2112 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
2113 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
2114 /* Overflow possible, we know nothing */
2115 dst_reg->smin_value = S64_MIN;
2116 dst_reg->smax_value = S64_MAX;
2118 dst_reg->smin_value -= smax_val;
2119 dst_reg->smax_value -= smin_val;
2121 if (dst_reg->umin_value < umax_val) {
2122 /* Overflow possible, we know nothing */
2123 dst_reg->umin_value = 0;
2124 dst_reg->umax_value = U64_MAX;
2126 /* Cannot overflow (as long as bounds are consistent) */
2127 dst_reg->umin_value -= umax_val;
2128 dst_reg->umax_value -= umin_val;
2130 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
2133 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
2134 if (smin_val < 0 || dst_reg->smin_value < 0) {
2135 /* Ain't nobody got time to multiply that sign */
2136 __mark_reg_unbounded(dst_reg);
2137 __update_reg_bounds(dst_reg);
2140 /* Both values are positive, so we can work with unsigned and
2141 * copy the result to signed (unless it exceeds S64_MAX).
2143 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
2144 /* Potential overflow, we know nothing */
2145 __mark_reg_unbounded(dst_reg);
2146 /* (except what we can learn from the var_off) */
2147 __update_reg_bounds(dst_reg);
2150 dst_reg->umin_value *= umin_val;
2151 dst_reg->umax_value *= umax_val;
2152 if (dst_reg->umax_value > S64_MAX) {
2153 /* Overflow possible, we know nothing */
2154 dst_reg->smin_value = S64_MIN;
2155 dst_reg->smax_value = S64_MAX;
2157 dst_reg->smin_value = dst_reg->umin_value;
2158 dst_reg->smax_value = dst_reg->umax_value;
2162 if (src_known && dst_known) {
2163 __mark_reg_known(dst_reg, dst_reg->var_off.value &
2164 src_reg.var_off.value);
2167 /* We get our minimum from the var_off, since that's inherently
2168 * bitwise. Our maximum is the minimum of the operands' maxima.
2170 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
2171 dst_reg->umin_value = dst_reg->var_off.value;
2172 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
2173 if (dst_reg->smin_value < 0 || smin_val < 0) {
2174 /* Lose signed bounds when ANDing negative numbers,
2175 * ain't nobody got time for that.
2177 dst_reg->smin_value = S64_MIN;
2178 dst_reg->smax_value = S64_MAX;
2180 /* ANDing two positives gives a positive, so safe to
2181 * cast result into s64.
2183 dst_reg->smin_value = dst_reg->umin_value;
2184 dst_reg->smax_value = dst_reg->umax_value;
2186 /* We may learn something more from the var_off */
2187 __update_reg_bounds(dst_reg);
2190 if (src_known && dst_known) {
2191 __mark_reg_known(dst_reg, dst_reg->var_off.value |
2192 src_reg.var_off.value);
2195 /* We get our maximum from the var_off, and our minimum is the
2196 * maximum of the operands' minima
2198 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
2199 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
2200 dst_reg->umax_value = dst_reg->var_off.value |
2201 dst_reg->var_off.mask;
2202 if (dst_reg->smin_value < 0 || smin_val < 0) {
2203 /* Lose signed bounds when ORing negative numbers,
2204 * ain't nobody got time for that.
2206 dst_reg->smin_value = S64_MIN;
2207 dst_reg->smax_value = S64_MAX;
2209 /* ORing two positives gives a positive, so safe to
2210 * cast result into s64.
2212 dst_reg->smin_value = dst_reg->umin_value;
2213 dst_reg->smax_value = dst_reg->umax_value;
2215 /* We may learn something more from the var_off */
2216 __update_reg_bounds(dst_reg);
2219 if (umax_val >= insn_bitness) {
2220 /* Shifts greater than 31 or 63 are undefined.
2221 * This includes shifts by a negative number.
2223 mark_reg_unknown(regs, insn->dst_reg);
2226 /* We lose all sign bit information (except what we can pick
2229 dst_reg->smin_value = S64_MIN;
2230 dst_reg->smax_value = S64_MAX;
2231 /* If we might shift our top bit out, then we know nothing */
2232 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
2233 dst_reg->umin_value = 0;
2234 dst_reg->umax_value = U64_MAX;
2236 dst_reg->umin_value <<= umin_val;
2237 dst_reg->umax_value <<= umax_val;
2240 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
2242 dst_reg->var_off = tnum_lshift(tnum_unknown, umin_val);
2243 /* We may learn something more from the var_off */
2244 __update_reg_bounds(dst_reg);
2247 if (umax_val >= insn_bitness) {
2248 /* Shifts greater than 31 or 63 are undefined.
2249 * This includes shifts by a negative number.
2251 mark_reg_unknown(regs, insn->dst_reg);
2254 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
2255 * be negative, then either:
2256 * 1) src_reg might be zero, so the sign bit of the result is
2257 * unknown, so we lose our signed bounds
2258 * 2) it's known negative, thus the unsigned bounds capture the
2260 * 3) the signed bounds cross zero, so they tell us nothing
2262 * If the value in dst_reg is known nonnegative, then again the
2263 * unsigned bounts capture the signed bounds.
2264 * Thus, in all cases it suffices to blow away our signed bounds
2265 * and rely on inferring new ones from the unsigned bounds and
2266 * var_off of the result.
2268 dst_reg->smin_value = S64_MIN;
2269 dst_reg->smax_value = S64_MAX;
2271 dst_reg->var_off = tnum_rshift(dst_reg->var_off,
2274 dst_reg->var_off = tnum_rshift(tnum_unknown, umin_val);
2275 dst_reg->umin_value >>= umax_val;
2276 dst_reg->umax_value >>= umin_val;
2277 /* We may learn something more from the var_off */
2278 __update_reg_bounds(dst_reg);
2281 mark_reg_unknown(regs, insn->dst_reg);
2285 if (BPF_CLASS(insn->code) != BPF_ALU64) {
2286 /* 32-bit ALU ops are (32,32)->32 */
2287 coerce_reg_to_size(dst_reg, 4);
2288 coerce_reg_to_size(&src_reg, 4);
2291 __reg_deduce_bounds(dst_reg);
2292 __reg_bound_offset(dst_reg);
2296 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
2299 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
2300 struct bpf_insn *insn)
2302 struct bpf_reg_state *regs = env->cur_state.regs, *dst_reg, *src_reg;
2303 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
2304 u8 opcode = BPF_OP(insn->code);
2307 dst_reg = ®s[insn->dst_reg];
2309 if (dst_reg->type != SCALAR_VALUE)
2311 if (BPF_SRC(insn->code) == BPF_X) {
2312 src_reg = ®s[insn->src_reg];
2313 if (src_reg->type != SCALAR_VALUE) {
2314 if (dst_reg->type != SCALAR_VALUE) {
2315 /* Combining two pointers by any ALU op yields
2316 * an arbitrary scalar.
2318 if (!env->allow_ptr_leaks) {
2319 verbose("R%d pointer %s pointer prohibited\n",
2321 bpf_alu_string[opcode >> 4]);
2324 mark_reg_unknown(regs, insn->dst_reg);
2327 /* scalar += pointer
2328 * This is legal, but we have to reverse our
2329 * src/dest handling in computing the range
2331 rc = adjust_ptr_min_max_vals(env, insn,
2333 if (rc == -EACCES && env->allow_ptr_leaks) {
2334 /* scalar += unknown scalar */
2335 __mark_reg_unknown(&off_reg);
2336 return adjust_scalar_min_max_vals(
2342 } else if (ptr_reg) {
2343 /* pointer += scalar */
2344 rc = adjust_ptr_min_max_vals(env, insn,
2346 if (rc == -EACCES && env->allow_ptr_leaks) {
2347 /* unknown scalar += scalar */
2348 __mark_reg_unknown(dst_reg);
2349 return adjust_scalar_min_max_vals(
2350 env, insn, dst_reg, *src_reg);
2355 /* Pretend the src is a reg with a known value, since we only
2356 * need to be able to read from this state.
2358 off_reg.type = SCALAR_VALUE;
2359 __mark_reg_known(&off_reg, insn->imm);
2361 if (ptr_reg) { /* pointer += K */
2362 rc = adjust_ptr_min_max_vals(env, insn,
2364 if (rc == -EACCES && env->allow_ptr_leaks) {
2365 /* unknown scalar += K */
2366 __mark_reg_unknown(dst_reg);
2367 return adjust_scalar_min_max_vals(
2368 env, insn, dst_reg, off_reg);
2374 /* Got here implies adding two SCALAR_VALUEs */
2375 if (WARN_ON_ONCE(ptr_reg)) {
2376 print_verifier_state(&env->cur_state);
2377 verbose("verifier internal error: unexpected ptr_reg\n");
2380 if (WARN_ON(!src_reg)) {
2381 print_verifier_state(&env->cur_state);
2382 verbose("verifier internal error: no src_reg\n");
2385 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
2388 /* check validity of 32-bit and 64-bit arithmetic operations */
2389 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
2391 struct bpf_reg_state *regs = env->cur_state.regs;
2392 u8 opcode = BPF_OP(insn->code);
2395 if (opcode == BPF_END || opcode == BPF_NEG) {
2396 if (opcode == BPF_NEG) {
2397 if (BPF_SRC(insn->code) != 0 ||
2398 insn->src_reg != BPF_REG_0 ||
2399 insn->off != 0 || insn->imm != 0) {
2400 verbose("BPF_NEG uses reserved fields\n");
2404 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
2405 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
2406 BPF_CLASS(insn->code) == BPF_ALU64) {
2407 verbose("BPF_END uses reserved fields\n");
2412 /* check src operand */
2413 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2417 if (is_pointer_value(env, insn->dst_reg)) {
2418 verbose("R%d pointer arithmetic prohibited\n",
2423 /* check dest operand */
2424 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2428 } else if (opcode == BPF_MOV) {
2430 if (BPF_SRC(insn->code) == BPF_X) {
2431 if (insn->imm != 0 || insn->off != 0) {
2432 verbose("BPF_MOV uses reserved fields\n");
2436 /* check src operand */
2437 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2441 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
2442 verbose("BPF_MOV uses reserved fields\n");
2447 /* check dest operand */
2448 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2452 if (BPF_SRC(insn->code) == BPF_X) {
2453 if (BPF_CLASS(insn->code) == BPF_ALU64) {
2455 * copy register state to dest reg
2457 regs[insn->dst_reg] = regs[insn->src_reg];
2458 regs[insn->dst_reg].live |= REG_LIVE_WRITTEN;
2461 if (is_pointer_value(env, insn->src_reg)) {
2462 verbose("R%d partial copy of pointer\n",
2466 mark_reg_unknown(regs, insn->dst_reg);
2467 coerce_reg_to_size(®s[insn->dst_reg], 4);
2471 * remember the value we stored into this reg
2473 regs[insn->dst_reg].type = SCALAR_VALUE;
2474 if (BPF_CLASS(insn->code) == BPF_ALU64) {
2475 __mark_reg_known(regs + insn->dst_reg,
2478 __mark_reg_known(regs + insn->dst_reg,
2483 } else if (opcode > BPF_END) {
2484 verbose("invalid BPF_ALU opcode %x\n", opcode);
2487 } else { /* all other ALU ops: and, sub, xor, add, ... */
2489 if (BPF_SRC(insn->code) == BPF_X) {
2490 if (insn->imm != 0 || insn->off != 0) {
2491 verbose("BPF_ALU uses reserved fields\n");
2494 /* check src1 operand */
2495 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2499 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
2500 verbose("BPF_ALU uses reserved fields\n");
2505 /* check src2 operand */
2506 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2510 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
2511 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
2512 verbose("div by zero\n");
2516 if (opcode == BPF_ARSH && BPF_CLASS(insn->code) != BPF_ALU64) {
2517 verbose("BPF_ARSH not supported for 32 bit ALU\n");
2521 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
2522 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
2523 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
2525 if (insn->imm < 0 || insn->imm >= size) {
2526 verbose("invalid shift %d\n", insn->imm);
2531 /* check dest operand */
2532 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
2536 return adjust_reg_min_max_vals(env, insn);
2542 static void find_good_pkt_pointers(struct bpf_verifier_state *state,
2543 struct bpf_reg_state *dst_reg,
2544 bool range_right_open)
2546 struct bpf_reg_state *regs = state->regs, *reg;
2550 if (dst_reg->off < 0 ||
2551 (dst_reg->off == 0 && range_right_open))
2552 /* This doesn't give us any range */
2555 if (dst_reg->umax_value > MAX_PACKET_OFF ||
2556 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
2557 /* Risk of overflow. For instance, ptr + (1<<63) may be less
2558 * than pkt_end, but that's because it's also less than pkt.
2562 new_range = dst_reg->off;
2563 if (range_right_open)
2566 /* Examples for register markings:
2568 * pkt_data in dst register:
2572 * if (r2 > pkt_end) goto <handle exception>
2577 * if (r2 < pkt_end) goto <access okay>
2578 * <handle exception>
2581 * r2 == dst_reg, pkt_end == src_reg
2582 * r2=pkt(id=n,off=8,r=0)
2583 * r3=pkt(id=n,off=0,r=0)
2585 * pkt_data in src register:
2589 * if (pkt_end >= r2) goto <access okay>
2590 * <handle exception>
2594 * if (pkt_end <= r2) goto <handle exception>
2598 * pkt_end == dst_reg, r2 == src_reg
2599 * r2=pkt(id=n,off=8,r=0)
2600 * r3=pkt(id=n,off=0,r=0)
2602 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
2603 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
2604 * and [r3, r3 + 8-1) respectively is safe to access depending on
2608 /* If our ids match, then we must have the same max_value. And we
2609 * don't care about the other reg's fixed offset, since if it's too big
2610 * the range won't allow anything.
2611 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
2613 for (i = 0; i < MAX_BPF_REG; i++)
2614 if (regs[i].type == PTR_TO_PACKET && regs[i].id == dst_reg->id)
2615 /* keep the maximum range already checked */
2616 regs[i].range = max(regs[i].range, new_range);
2618 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
2619 if (state->stack_slot_type[i] != STACK_SPILL)
2621 reg = &state->spilled_regs[i / BPF_REG_SIZE];
2622 if (reg->type == PTR_TO_PACKET && reg->id == dst_reg->id)
2623 reg->range = max(reg->range, new_range);
2627 /* Adjusts the register min/max values in the case that the dst_reg is the
2628 * variable register that we are working on, and src_reg is a constant or we're
2629 * simply doing a BPF_K check.
2630 * In JEQ/JNE cases we also adjust the var_off values.
2632 static void reg_set_min_max(struct bpf_reg_state *true_reg,
2633 struct bpf_reg_state *false_reg, u64 val,
2636 /* If the dst_reg is a pointer, we can't learn anything about its
2637 * variable offset from the compare (unless src_reg were a pointer into
2638 * the same object, but we don't bother with that.
2639 * Since false_reg and true_reg have the same type by construction, we
2640 * only need to check one of them for pointerness.
2642 if (__is_pointer_value(false, false_reg))
2647 /* If this is false then we know nothing Jon Snow, but if it is
2648 * true then we know for sure.
2650 __mark_reg_known(true_reg, val);
2653 /* If this is true we know nothing Jon Snow, but if it is false
2654 * we know the value for sure;
2656 __mark_reg_known(false_reg, val);
2659 false_reg->umax_value = min(false_reg->umax_value, val);
2660 true_reg->umin_value = max(true_reg->umin_value, val + 1);
2663 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
2664 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
2667 false_reg->umin_value = max(false_reg->umin_value, val);
2668 true_reg->umax_value = min(true_reg->umax_value, val - 1);
2671 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
2672 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
2675 false_reg->umax_value = min(false_reg->umax_value, val - 1);
2676 true_reg->umin_value = max(true_reg->umin_value, val);
2679 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
2680 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
2683 false_reg->umin_value = max(false_reg->umin_value, val + 1);
2684 true_reg->umax_value = min(true_reg->umax_value, val);
2687 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
2688 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
2694 __reg_deduce_bounds(false_reg);
2695 __reg_deduce_bounds(true_reg);
2696 /* We might have learned some bits from the bounds. */
2697 __reg_bound_offset(false_reg);
2698 __reg_bound_offset(true_reg);
2699 /* Intersecting with the old var_off might have improved our bounds
2700 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2701 * then new var_off is (0; 0x7f...fc) which improves our umax.
2703 __update_reg_bounds(false_reg);
2704 __update_reg_bounds(true_reg);
2707 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
2710 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
2711 struct bpf_reg_state *false_reg, u64 val,
2714 if (__is_pointer_value(false, false_reg))
2719 /* If this is false then we know nothing Jon Snow, but if it is
2720 * true then we know for sure.
2722 __mark_reg_known(true_reg, val);
2725 /* If this is true we know nothing Jon Snow, but if it is false
2726 * we know the value for sure;
2728 __mark_reg_known(false_reg, val);
2731 true_reg->umax_value = min(true_reg->umax_value, val - 1);
2732 false_reg->umin_value = max(false_reg->umin_value, val);
2735 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
2736 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
2739 true_reg->umin_value = max(true_reg->umin_value, val + 1);
2740 false_reg->umax_value = min(false_reg->umax_value, val);
2743 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
2744 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
2747 true_reg->umax_value = min(true_reg->umax_value, val);
2748 false_reg->umin_value = max(false_reg->umin_value, val + 1);
2751 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
2752 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
2755 true_reg->umin_value = max(true_reg->umin_value, val);
2756 false_reg->umax_value = min(false_reg->umax_value, val - 1);
2759 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
2760 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
2766 __reg_deduce_bounds(false_reg);
2767 __reg_deduce_bounds(true_reg);
2768 /* We might have learned some bits from the bounds. */
2769 __reg_bound_offset(false_reg);
2770 __reg_bound_offset(true_reg);
2771 /* Intersecting with the old var_off might have improved our bounds
2772 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2773 * then new var_off is (0; 0x7f...fc) which improves our umax.
2775 __update_reg_bounds(false_reg);
2776 __update_reg_bounds(true_reg);
2779 /* Regs are known to be equal, so intersect their min/max/var_off */
2780 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
2781 struct bpf_reg_state *dst_reg)
2783 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
2784 dst_reg->umin_value);
2785 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
2786 dst_reg->umax_value);
2787 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
2788 dst_reg->smin_value);
2789 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
2790 dst_reg->smax_value);
2791 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
2793 /* We might have learned new bounds from the var_off. */
2794 __update_reg_bounds(src_reg);
2795 __update_reg_bounds(dst_reg);
2796 /* We might have learned something about the sign bit. */
2797 __reg_deduce_bounds(src_reg);
2798 __reg_deduce_bounds(dst_reg);
2799 /* We might have learned some bits from the bounds. */
2800 __reg_bound_offset(src_reg);
2801 __reg_bound_offset(dst_reg);
2802 /* Intersecting with the old var_off might have improved our bounds
2803 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2804 * then new var_off is (0; 0x7f...fc) which improves our umax.
2806 __update_reg_bounds(src_reg);
2807 __update_reg_bounds(dst_reg);
2810 static void reg_combine_min_max(struct bpf_reg_state *true_src,
2811 struct bpf_reg_state *true_dst,
2812 struct bpf_reg_state *false_src,
2813 struct bpf_reg_state *false_dst,
2818 __reg_combine_min_max(true_src, true_dst);
2821 __reg_combine_min_max(false_src, false_dst);
2826 static void mark_map_reg(struct bpf_reg_state *regs, u32 regno, u32 id,
2829 struct bpf_reg_state *reg = ®s[regno];
2831 if (reg->type == PTR_TO_MAP_VALUE_OR_NULL && reg->id == id) {
2832 /* Old offset (both fixed and variable parts) should
2833 * have been known-zero, because we don't allow pointer
2834 * arithmetic on pointers that might be NULL.
2836 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
2837 !tnum_equals_const(reg->var_off, 0) ||
2839 __mark_reg_known_zero(reg);
2843 reg->type = SCALAR_VALUE;
2844 } else if (reg->map_ptr->inner_map_meta) {
2845 reg->type = CONST_PTR_TO_MAP;
2846 reg->map_ptr = reg->map_ptr->inner_map_meta;
2848 reg->type = PTR_TO_MAP_VALUE;
2850 /* We don't need id from this point onwards anymore, thus we
2851 * should better reset it, so that state pruning has chances
2858 /* The logic is similar to find_good_pkt_pointers(), both could eventually
2859 * be folded together at some point.
2861 static void mark_map_regs(struct bpf_verifier_state *state, u32 regno,
2864 struct bpf_reg_state *regs = state->regs;
2865 u32 id = regs[regno].id;
2868 for (i = 0; i < MAX_BPF_REG; i++)
2869 mark_map_reg(regs, i, id, is_null);
2871 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
2872 if (state->stack_slot_type[i] != STACK_SPILL)
2874 mark_map_reg(state->spilled_regs, i / BPF_REG_SIZE, id, is_null);
2878 static int check_cond_jmp_op(struct bpf_verifier_env *env,
2879 struct bpf_insn *insn, int *insn_idx)
2881 struct bpf_verifier_state *other_branch, *this_branch = &env->cur_state;
2882 struct bpf_reg_state *regs = this_branch->regs, *dst_reg;
2883 u8 opcode = BPF_OP(insn->code);
2886 if (opcode > BPF_JSLE) {
2887 verbose("invalid BPF_JMP opcode %x\n", opcode);
2891 if (BPF_SRC(insn->code) == BPF_X) {
2892 if (insn->imm != 0) {
2893 verbose("BPF_JMP uses reserved fields\n");
2897 /* check src1 operand */
2898 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2902 if (is_pointer_value(env, insn->src_reg)) {
2903 verbose("R%d pointer comparison prohibited\n",
2908 if (insn->src_reg != BPF_REG_0) {
2909 verbose("BPF_JMP uses reserved fields\n");
2914 /* check src2 operand */
2915 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2919 dst_reg = ®s[insn->dst_reg];
2921 /* detect if R == 0 where R was initialized to zero earlier */
2922 if (BPF_SRC(insn->code) == BPF_K &&
2923 (opcode == BPF_JEQ || opcode == BPF_JNE) &&
2924 dst_reg->type == SCALAR_VALUE &&
2925 tnum_equals_const(dst_reg->var_off, insn->imm)) {
2926 if (opcode == BPF_JEQ) {
2927 /* if (imm == imm) goto pc+off;
2928 * only follow the goto, ignore fall-through
2930 *insn_idx += insn->off;
2933 /* if (imm != imm) goto pc+off;
2934 * only follow fall-through branch, since
2935 * that's where the program will go
2941 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx);
2945 /* detect if we are comparing against a constant value so we can adjust
2946 * our min/max values for our dst register.
2947 * this is only legit if both are scalars (or pointers to the same
2948 * object, I suppose, but we don't support that right now), because
2949 * otherwise the different base pointers mean the offsets aren't
2952 if (BPF_SRC(insn->code) == BPF_X) {
2953 if (dst_reg->type == SCALAR_VALUE &&
2954 regs[insn->src_reg].type == SCALAR_VALUE) {
2955 if (tnum_is_const(regs[insn->src_reg].var_off))
2956 reg_set_min_max(&other_branch->regs[insn->dst_reg],
2957 dst_reg, regs[insn->src_reg].var_off.value,
2959 else if (tnum_is_const(dst_reg->var_off))
2960 reg_set_min_max_inv(&other_branch->regs[insn->src_reg],
2961 ®s[insn->src_reg],
2962 dst_reg->var_off.value, opcode);
2963 else if (opcode == BPF_JEQ || opcode == BPF_JNE)
2964 /* Comparing for equality, we can combine knowledge */
2965 reg_combine_min_max(&other_branch->regs[insn->src_reg],
2966 &other_branch->regs[insn->dst_reg],
2967 ®s[insn->src_reg],
2968 ®s[insn->dst_reg], opcode);
2970 } else if (dst_reg->type == SCALAR_VALUE) {
2971 reg_set_min_max(&other_branch->regs[insn->dst_reg],
2972 dst_reg, insn->imm, opcode);
2975 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
2976 if (BPF_SRC(insn->code) == BPF_K &&
2977 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
2978 dst_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
2979 /* Mark all identical map registers in each branch as either
2980 * safe or unknown depending R == 0 or R != 0 conditional.
2982 mark_map_regs(this_branch, insn->dst_reg, opcode == BPF_JNE);
2983 mark_map_regs(other_branch, insn->dst_reg, opcode == BPF_JEQ);
2984 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGT &&
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, false);
2989 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGT &&
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], true);
2994 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JLT &&
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, true);
2999 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JLT &&
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], false);
3004 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGE &&
3005 dst_reg->type == PTR_TO_PACKET &&
3006 regs[insn->src_reg].type == PTR_TO_PACKET_END) {
3007 /* pkt_data' >= pkt_end */
3008 find_good_pkt_pointers(this_branch, dst_reg, true);
3009 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGE &&
3010 dst_reg->type == PTR_TO_PACKET_END &&
3011 regs[insn->src_reg].type == PTR_TO_PACKET) {
3012 /* pkt_end >= pkt_data' */
3013 find_good_pkt_pointers(other_branch, ®s[insn->src_reg], false);
3014 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JLE &&
3015 dst_reg->type == PTR_TO_PACKET &&
3016 regs[insn->src_reg].type == PTR_TO_PACKET_END) {
3017 /* pkt_data' <= pkt_end */
3018 find_good_pkt_pointers(other_branch, dst_reg, false);
3019 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JLE &&
3020 dst_reg->type == PTR_TO_PACKET_END &&
3021 regs[insn->src_reg].type == PTR_TO_PACKET) {
3022 /* pkt_end <= pkt_data' */
3023 find_good_pkt_pointers(this_branch, ®s[insn->src_reg], true);
3024 } else if (is_pointer_value(env, insn->dst_reg)) {
3025 verbose("R%d pointer comparison prohibited\n", insn->dst_reg);
3029 print_verifier_state(this_branch);
3033 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
3034 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
3036 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
3038 return (struct bpf_map *) (unsigned long) imm64;
3041 /* verify BPF_LD_IMM64 instruction */
3042 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
3044 struct bpf_reg_state *regs = env->cur_state.regs;
3047 if (BPF_SIZE(insn->code) != BPF_DW) {
3048 verbose("invalid BPF_LD_IMM insn\n");
3051 if (insn->off != 0) {
3052 verbose("BPF_LD_IMM64 uses reserved fields\n");
3056 err = check_reg_arg(env, insn->dst_reg, DST_OP);
3060 if (insn->src_reg == 0) {
3061 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
3063 regs[insn->dst_reg].type = SCALAR_VALUE;
3064 __mark_reg_known(®s[insn->dst_reg], imm);
3068 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
3069 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
3071 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
3072 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
3076 static bool may_access_skb(enum bpf_prog_type type)
3079 case BPF_PROG_TYPE_SOCKET_FILTER:
3080 case BPF_PROG_TYPE_SCHED_CLS:
3081 case BPF_PROG_TYPE_SCHED_ACT:
3088 /* verify safety of LD_ABS|LD_IND instructions:
3089 * - they can only appear in the programs where ctx == skb
3090 * - since they are wrappers of function calls, they scratch R1-R5 registers,
3091 * preserve R6-R9, and store return value into R0
3094 * ctx == skb == R6 == CTX
3097 * SRC == any register
3098 * IMM == 32-bit immediate
3101 * R0 - 8/16/32-bit skb data converted to cpu endianness
3103 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
3105 struct bpf_reg_state *regs = env->cur_state.regs;
3106 u8 mode = BPF_MODE(insn->code);
3109 if (!may_access_skb(env->prog->type)) {
3110 verbose("BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
3114 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
3115 BPF_SIZE(insn->code) == BPF_DW ||
3116 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
3117 verbose("BPF_LD_[ABS|IND] uses reserved fields\n");
3121 /* check whether implicit source operand (register R6) is readable */
3122 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
3126 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
3127 verbose("at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
3131 if (mode == BPF_IND) {
3132 /* check explicit source operand */
3133 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3138 /* reset caller saved regs to unreadable */
3139 for (i = 0; i < CALLER_SAVED_REGS; i++) {
3140 mark_reg_not_init(regs, caller_saved[i]);
3141 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
3144 /* mark destination R0 register as readable, since it contains
3145 * the value fetched from the packet.
3146 * Already marked as written above.
3148 mark_reg_unknown(regs, BPF_REG_0);
3152 /* non-recursive DFS pseudo code
3153 * 1 procedure DFS-iterative(G,v):
3154 * 2 label v as discovered
3155 * 3 let S be a stack
3157 * 5 while S is not empty
3159 * 7 if t is what we're looking for:
3161 * 9 for all edges e in G.adjacentEdges(t) do
3162 * 10 if edge e is already labelled
3163 * 11 continue with the next edge
3164 * 12 w <- G.adjacentVertex(t,e)
3165 * 13 if vertex w is not discovered and not explored
3166 * 14 label e as tree-edge
3167 * 15 label w as discovered
3170 * 18 else if vertex w is discovered
3171 * 19 label e as back-edge
3173 * 21 // vertex w is explored
3174 * 22 label e as forward- or cross-edge
3175 * 23 label t as explored
3180 * 0x11 - discovered and fall-through edge labelled
3181 * 0x12 - discovered and fall-through and branch edges labelled
3192 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
3194 static int *insn_stack; /* stack of insns to process */
3195 static int cur_stack; /* current stack index */
3196 static int *insn_state;
3198 /* t, w, e - match pseudo-code above:
3199 * t - index of current instruction
3200 * w - next instruction
3203 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
3205 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
3208 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
3211 if (w < 0 || w >= env->prog->len) {
3212 verbose("jump out of range from insn %d to %d\n", t, w);
3217 /* mark branch target for state pruning */
3218 env->explored_states[w] = STATE_LIST_MARK;
3220 if (insn_state[w] == 0) {
3222 insn_state[t] = DISCOVERED | e;
3223 insn_state[w] = DISCOVERED;
3224 if (cur_stack >= env->prog->len)
3226 insn_stack[cur_stack++] = w;
3228 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
3229 verbose("back-edge from insn %d to %d\n", t, w);
3231 } else if (insn_state[w] == EXPLORED) {
3232 /* forward- or cross-edge */
3233 insn_state[t] = DISCOVERED | e;
3235 verbose("insn state internal bug\n");
3241 /* non-recursive depth-first-search to detect loops in BPF program
3242 * loop == back-edge in directed graph
3244 static int check_cfg(struct bpf_verifier_env *env)
3246 struct bpf_insn *insns = env->prog->insnsi;
3247 int insn_cnt = env->prog->len;
3251 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
3255 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
3261 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
3262 insn_stack[0] = 0; /* 0 is the first instruction */
3268 t = insn_stack[cur_stack - 1];
3270 if (BPF_CLASS(insns[t].code) == BPF_JMP) {
3271 u8 opcode = BPF_OP(insns[t].code);
3273 if (opcode == BPF_EXIT) {
3275 } else if (opcode == BPF_CALL) {
3276 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3281 if (t + 1 < insn_cnt)
3282 env->explored_states[t + 1] = STATE_LIST_MARK;
3283 } else if (opcode == BPF_JA) {
3284 if (BPF_SRC(insns[t].code) != BPF_K) {
3288 /* unconditional jump with single edge */
3289 ret = push_insn(t, t + insns[t].off + 1,
3295 /* tell verifier to check for equivalent states
3296 * after every call and jump
3298 if (t + 1 < insn_cnt)
3299 env->explored_states[t + 1] = STATE_LIST_MARK;
3301 /* conditional jump with two edges */
3302 env->explored_states[t] = STATE_LIST_MARK;
3303 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3309 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
3316 /* all other non-branch instructions with single
3319 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3327 insn_state[t] = EXPLORED;
3328 if (cur_stack-- <= 0) {
3329 verbose("pop stack internal bug\n");
3336 for (i = 0; i < insn_cnt; i++) {
3337 if (insn_state[i] != EXPLORED) {
3338 verbose("unreachable insn %d\n", i);
3343 ret = 0; /* cfg looks good */
3351 /* check %cur's range satisfies %old's */
3352 static bool range_within(struct bpf_reg_state *old,
3353 struct bpf_reg_state *cur)
3355 return old->umin_value <= cur->umin_value &&
3356 old->umax_value >= cur->umax_value &&
3357 old->smin_value <= cur->smin_value &&
3358 old->smax_value >= cur->smax_value;
3361 /* Maximum number of register states that can exist at once */
3362 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
3368 /* If in the old state two registers had the same id, then they need to have
3369 * the same id in the new state as well. But that id could be different from
3370 * the old state, so we need to track the mapping from old to new ids.
3371 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
3372 * regs with old id 5 must also have new id 9 for the new state to be safe. But
3373 * regs with a different old id could still have new id 9, we don't care about
3375 * So we look through our idmap to see if this old id has been seen before. If
3376 * so, we require the new id to match; otherwise, we add the id pair to the map.
3378 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
3382 for (i = 0; i < ID_MAP_SIZE; i++) {
3383 if (!idmap[i].old) {
3384 /* Reached an empty slot; haven't seen this id before */
3385 idmap[i].old = old_id;
3386 idmap[i].cur = cur_id;
3389 if (idmap[i].old == old_id)
3390 return idmap[i].cur == cur_id;
3392 /* We ran out of idmap slots, which should be impossible */
3397 /* Returns true if (rold safe implies rcur safe) */
3398 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
3399 struct idpair *idmap)
3401 if (!(rold->live & REG_LIVE_READ))
3402 /* explored state didn't use this */
3405 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, live)) == 0)
3408 if (rold->type == NOT_INIT)
3409 /* explored state can't have used this */
3411 if (rcur->type == NOT_INIT)
3413 switch (rold->type) {
3415 if (rcur->type == SCALAR_VALUE) {
3416 /* new val must satisfy old val knowledge */
3417 return range_within(rold, rcur) &&
3418 tnum_in(rold->var_off, rcur->var_off);
3420 /* We're trying to use a pointer in place of a scalar.
3421 * Even if the scalar was unbounded, this could lead to
3422 * pointer leaks because scalars are allowed to leak
3423 * while pointers are not. We could make this safe in
3424 * special cases if root is calling us, but it's
3425 * probably not worth the hassle.
3429 case PTR_TO_MAP_VALUE:
3430 /* If the new min/max/var_off satisfy the old ones and
3431 * everything else matches, we are OK.
3432 * We don't care about the 'id' value, because nothing
3433 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
3435 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
3436 range_within(rold, rcur) &&
3437 tnum_in(rold->var_off, rcur->var_off);
3438 case PTR_TO_MAP_VALUE_OR_NULL:
3439 /* a PTR_TO_MAP_VALUE could be safe to use as a
3440 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
3441 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
3442 * checked, doing so could have affected others with the same
3443 * id, and we can't check for that because we lost the id when
3444 * we converted to a PTR_TO_MAP_VALUE.
3446 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
3448 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
3450 /* Check our ids match any regs they're supposed to */
3451 return check_ids(rold->id, rcur->id, idmap);
3453 if (rcur->type != PTR_TO_PACKET)
3455 /* We must have at least as much range as the old ptr
3456 * did, so that any accesses which were safe before are
3457 * still safe. This is true even if old range < old off,
3458 * since someone could have accessed through (ptr - k), or
3459 * even done ptr -= k in a register, to get a safe access.
3461 if (rold->range > rcur->range)
3463 /* If the offsets don't match, we can't trust our alignment;
3464 * nor can we be sure that we won't fall out of range.
3466 if (rold->off != rcur->off)
3468 /* id relations must be preserved */
3469 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
3471 /* new val must satisfy old val knowledge */
3472 return range_within(rold, rcur) &&
3473 tnum_in(rold->var_off, rcur->var_off);
3475 case CONST_PTR_TO_MAP:
3477 case PTR_TO_PACKET_END:
3478 /* Only valid matches are exact, which memcmp() above
3479 * would have accepted
3482 /* Don't know what's going on, just say it's not safe */
3486 /* Shouldn't get here; if we do, say it's not safe */
3491 /* compare two verifier states
3493 * all states stored in state_list are known to be valid, since
3494 * verifier reached 'bpf_exit' instruction through them
3496 * this function is called when verifier exploring different branches of
3497 * execution popped from the state stack. If it sees an old state that has
3498 * more strict register state and more strict stack state then this execution
3499 * branch doesn't need to be explored further, since verifier already
3500 * concluded that more strict state leads to valid finish.
3502 * Therefore two states are equivalent if register state is more conservative
3503 * and explored stack state is more conservative than the current one.
3506 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
3507 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
3509 * In other words if current stack state (one being explored) has more
3510 * valid slots than old one that already passed validation, it means
3511 * the verifier can stop exploring and conclude that current state is valid too
3513 * Similarly with registers. If explored state has register type as invalid
3514 * whereas register type in current state is meaningful, it means that
3515 * the current state will reach 'bpf_exit' instruction safely
3517 static bool states_equal(struct bpf_verifier_env *env,
3518 struct bpf_verifier_state *old,
3519 struct bpf_verifier_state *cur)
3521 struct idpair *idmap;
3525 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
3526 /* If we failed to allocate the idmap, just say it's not safe */
3530 for (i = 0; i < MAX_BPF_REG; i++) {
3531 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
3535 for (i = 0; i < MAX_BPF_STACK; i++) {
3536 if (old->stack_slot_type[i] == STACK_INVALID)
3538 if (old->stack_slot_type[i] != cur->stack_slot_type[i])
3539 /* Ex: old explored (safe) state has STACK_SPILL in
3540 * this stack slot, but current has has STACK_MISC ->
3541 * this verifier states are not equivalent,
3542 * return false to continue verification of this path
3545 if (i % BPF_REG_SIZE)
3547 if (old->stack_slot_type[i] != STACK_SPILL)
3549 if (!regsafe(&old->spilled_regs[i / BPF_REG_SIZE],
3550 &cur->spilled_regs[i / BPF_REG_SIZE],
3552 /* when explored and current stack slot are both storing
3553 * spilled registers, check that stored pointers types
3554 * are the same as well.
3555 * Ex: explored safe path could have stored
3556 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
3557 * but current path has stored:
3558 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
3559 * such verifier states are not equivalent.
3560 * return false to continue verification of this path
3572 /* A write screens off any subsequent reads; but write marks come from the
3573 * straight-line code between a state and its parent. When we arrive at a
3574 * jump target (in the first iteration of the propagate_liveness() loop),
3575 * we didn't arrive by the straight-line code, so read marks in state must
3576 * propagate to parent regardless of state's write marks.
3578 static bool do_propagate_liveness(const struct bpf_verifier_state *state,
3579 struct bpf_verifier_state *parent)
3581 bool writes = parent == state->parent; /* Observe write marks */
3582 bool touched = false; /* any changes made? */
3587 /* Propagate read liveness of registers... */
3588 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
3589 /* We don't need to worry about FP liveness because it's read-only */
3590 for (i = 0; i < BPF_REG_FP; i++) {
3591 if (parent->regs[i].live & REG_LIVE_READ)
3593 if (writes && (state->regs[i].live & REG_LIVE_WRITTEN))
3595 if (state->regs[i].live & REG_LIVE_READ) {
3596 parent->regs[i].live |= REG_LIVE_READ;
3600 /* ... and stack slots */
3601 for (i = 0; i < MAX_BPF_STACK / BPF_REG_SIZE; i++) {
3602 if (parent->stack_slot_type[i * BPF_REG_SIZE] != STACK_SPILL)
3604 if (state->stack_slot_type[i * BPF_REG_SIZE] != STACK_SPILL)
3606 if (parent->spilled_regs[i].live & REG_LIVE_READ)
3608 if (writes && (state->spilled_regs[i].live & REG_LIVE_WRITTEN))
3610 if (state->spilled_regs[i].live & REG_LIVE_READ) {
3611 parent->spilled_regs[i].live |= REG_LIVE_READ;
3618 /* "parent" is "a state from which we reach the current state", but initially
3619 * it is not the state->parent (i.e. "the state whose straight-line code leads
3620 * to the current state"), instead it is the state that happened to arrive at
3621 * a (prunable) equivalent of the current state. See comment above
3622 * do_propagate_liveness() for consequences of this.
3623 * This function is just a more efficient way of calling mark_reg_read() or
3624 * mark_stack_slot_read() on each reg in "parent" that is read in "state",
3625 * though it requires that parent != state->parent in the call arguments.
3627 static void propagate_liveness(const struct bpf_verifier_state *state,
3628 struct bpf_verifier_state *parent)
3630 while (do_propagate_liveness(state, parent)) {
3631 /* Something changed, so we need to feed those changes onward */
3633 parent = state->parent;
3637 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
3639 struct bpf_verifier_state_list *new_sl;
3640 struct bpf_verifier_state_list *sl;
3643 sl = env->explored_states[insn_idx];
3645 /* this 'insn_idx' instruction wasn't marked, so we will not
3646 * be doing state search here
3650 while (sl != STATE_LIST_MARK) {
3651 if (states_equal(env, &sl->state, &env->cur_state)) {
3652 /* reached equivalent register/stack state,
3654 * Registers read by the continuation are read by us.
3655 * If we have any write marks in env->cur_state, they
3656 * will prevent corresponding reads in the continuation
3657 * from reaching our parent (an explored_state). Our
3658 * own state will get the read marks recorded, but
3659 * they'll be immediately forgotten as we're pruning
3660 * this state and will pop a new one.
3662 propagate_liveness(&sl->state, &env->cur_state);
3668 /* there were no equivalent states, remember current one.
3669 * technically the current state is not proven to be safe yet,
3670 * but it will either reach bpf_exit (which means it's safe) or
3671 * it will be rejected. Since there are no loops, we won't be
3672 * seeing this 'insn_idx' instruction again on the way to bpf_exit
3674 new_sl = kmalloc(sizeof(struct bpf_verifier_state_list), GFP_USER);
3678 /* add new state to the head of linked list */
3679 memcpy(&new_sl->state, &env->cur_state, sizeof(env->cur_state));
3680 new_sl->next = env->explored_states[insn_idx];
3681 env->explored_states[insn_idx] = new_sl;
3682 /* connect new state to parentage chain */
3683 env->cur_state.parent = &new_sl->state;
3684 /* clear write marks in current state: the writes we did are not writes
3685 * our child did, so they don't screen off its reads from us.
3686 * (There are no read marks in current state, because reads always mark
3687 * their parent and current state never has children yet. Only
3688 * explored_states can get read marks.)
3690 for (i = 0; i < BPF_REG_FP; i++)
3691 env->cur_state.regs[i].live = REG_LIVE_NONE;
3692 for (i = 0; i < MAX_BPF_STACK / BPF_REG_SIZE; i++)
3693 if (env->cur_state.stack_slot_type[i * BPF_REG_SIZE] == STACK_SPILL)
3694 env->cur_state.spilled_regs[i].live = REG_LIVE_NONE;
3698 static int ext_analyzer_insn_hook(struct bpf_verifier_env *env,
3699 int insn_idx, int prev_insn_idx)
3701 if (!env->analyzer_ops || !env->analyzer_ops->insn_hook)
3704 return env->analyzer_ops->insn_hook(env, insn_idx, prev_insn_idx);
3707 static int do_check(struct bpf_verifier_env *env)
3709 struct bpf_verifier_state *state = &env->cur_state;
3710 struct bpf_insn *insns = env->prog->insnsi;
3711 struct bpf_reg_state *regs = state->regs;
3712 int insn_cnt = env->prog->len;
3713 int insn_idx, prev_insn_idx = 0;
3714 int insn_processed = 0;
3715 bool do_print_state = false;
3717 init_reg_state(regs);
3718 state->parent = NULL;
3721 struct bpf_insn *insn;
3725 if (insn_idx >= insn_cnt) {
3726 verbose("invalid insn idx %d insn_cnt %d\n",
3727 insn_idx, insn_cnt);
3731 insn = &insns[insn_idx];
3732 class = BPF_CLASS(insn->code);
3734 if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
3735 verbose("BPF program is too large. Processed %d insn\n",
3740 err = is_state_visited(env, insn_idx);
3744 /* found equivalent state, can prune the search */
3747 verbose("\nfrom %d to %d: safe\n",
3748 prev_insn_idx, insn_idx);
3750 verbose("%d: safe\n", insn_idx);
3752 goto process_bpf_exit;
3758 if (log_level > 1 || (log_level && do_print_state)) {
3760 verbose("%d:", insn_idx);
3762 verbose("\nfrom %d to %d:",
3763 prev_insn_idx, insn_idx);
3764 print_verifier_state(&env->cur_state);
3765 do_print_state = false;
3769 verbose("%d: ", insn_idx);
3770 print_bpf_insn(env, insn);
3773 err = ext_analyzer_insn_hook(env, insn_idx, prev_insn_idx);
3777 env->insn_aux_data[insn_idx].seen = true;
3778 if (class == BPF_ALU || class == BPF_ALU64) {
3779 err = check_alu_op(env, insn);
3783 } else if (class == BPF_LDX) {
3784 enum bpf_reg_type *prev_src_type, src_reg_type;
3786 /* check for reserved fields is already done */
3788 /* check src operand */
3789 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3793 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3797 src_reg_type = regs[insn->src_reg].type;
3799 /* check that memory (src_reg + off) is readable,
3800 * the state of dst_reg will be updated by this func
3802 err = check_mem_access(env, insn_idx, insn->src_reg, insn->off,
3803 BPF_SIZE(insn->code), BPF_READ,
3808 prev_src_type = &env->insn_aux_data[insn_idx].ptr_type;
3810 if (*prev_src_type == NOT_INIT) {
3812 * dst_reg = *(u32 *)(src_reg + off)
3813 * save type to validate intersecting paths
3815 *prev_src_type = src_reg_type;
3817 } else if (src_reg_type != *prev_src_type &&
3818 (src_reg_type == PTR_TO_CTX ||
3819 *prev_src_type == PTR_TO_CTX)) {
3820 /* ABuser program is trying to use the same insn
3821 * dst_reg = *(u32*) (src_reg + off)
3822 * with different pointer types:
3823 * src_reg == ctx in one branch and
3824 * src_reg == stack|map in some other branch.
3827 verbose("same insn cannot be used with different pointers\n");
3831 } else if (class == BPF_STX) {
3832 enum bpf_reg_type *prev_dst_type, dst_reg_type;
3834 if (BPF_MODE(insn->code) == BPF_XADD) {
3835 err = check_xadd(env, insn_idx, insn);
3842 /* check src1 operand */
3843 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3846 /* check src2 operand */
3847 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3851 dst_reg_type = regs[insn->dst_reg].type;
3853 /* check that memory (dst_reg + off) is writeable */
3854 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3855 BPF_SIZE(insn->code), BPF_WRITE,
3860 prev_dst_type = &env->insn_aux_data[insn_idx].ptr_type;
3862 if (*prev_dst_type == NOT_INIT) {
3863 *prev_dst_type = dst_reg_type;
3864 } else if (dst_reg_type != *prev_dst_type &&
3865 (dst_reg_type == PTR_TO_CTX ||
3866 *prev_dst_type == PTR_TO_CTX)) {
3867 verbose("same insn cannot be used with different pointers\n");
3871 } else if (class == BPF_ST) {
3872 if (BPF_MODE(insn->code) != BPF_MEM ||
3873 insn->src_reg != BPF_REG_0) {
3874 verbose("BPF_ST uses reserved fields\n");
3877 /* check src operand */
3878 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3882 if (is_ctx_reg(env, insn->dst_reg)) {
3883 verbose("BPF_ST stores into R%d context is not allowed\n",
3888 /* check that memory (dst_reg + off) is writeable */
3889 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3890 BPF_SIZE(insn->code), BPF_WRITE,
3895 } else if (class == BPF_JMP) {
3896 u8 opcode = BPF_OP(insn->code);
3898 if (opcode == BPF_CALL) {
3899 if (BPF_SRC(insn->code) != BPF_K ||
3901 insn->src_reg != BPF_REG_0 ||
3902 insn->dst_reg != BPF_REG_0) {
3903 verbose("BPF_CALL uses reserved fields\n");
3907 err = check_call(env, insn->imm, insn_idx);
3911 } else if (opcode == BPF_JA) {
3912 if (BPF_SRC(insn->code) != BPF_K ||
3914 insn->src_reg != BPF_REG_0 ||
3915 insn->dst_reg != BPF_REG_0) {
3916 verbose("BPF_JA uses reserved fields\n");
3920 insn_idx += insn->off + 1;
3923 } else if (opcode == BPF_EXIT) {
3924 if (BPF_SRC(insn->code) != BPF_K ||
3926 insn->src_reg != BPF_REG_0 ||
3927 insn->dst_reg != BPF_REG_0) {
3928 verbose("BPF_EXIT uses reserved fields\n");
3932 /* eBPF calling convetion is such that R0 is used
3933 * to return the value from eBPF program.
3934 * Make sure that it's readable at this time
3935 * of bpf_exit, which means that program wrote
3936 * something into it earlier
3938 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
3942 if (is_pointer_value(env, BPF_REG_0)) {
3943 verbose("R0 leaks addr as return value\n");
3948 insn_idx = pop_stack(env, &prev_insn_idx);
3952 do_print_state = true;
3956 err = check_cond_jmp_op(env, insn, &insn_idx);
3960 } else if (class == BPF_LD) {
3961 u8 mode = BPF_MODE(insn->code);
3963 if (mode == BPF_ABS || mode == BPF_IND) {
3964 err = check_ld_abs(env, insn);
3968 } else if (mode == BPF_IMM) {
3969 err = check_ld_imm(env, insn);
3974 env->insn_aux_data[insn_idx].seen = true;
3976 verbose("invalid BPF_LD mode\n");
3980 verbose("unknown insn class %d\n", class);
3987 verbose("processed %d insns, stack depth %d\n",
3988 insn_processed, env->prog->aux->stack_depth);
3992 static int check_map_prealloc(struct bpf_map *map)
3994 return (map->map_type != BPF_MAP_TYPE_HASH &&
3995 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
3996 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
3997 !(map->map_flags & BPF_F_NO_PREALLOC);
4000 static int check_map_prog_compatibility(struct bpf_map *map,
4001 struct bpf_prog *prog)
4004 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
4005 * preallocated hash maps, since doing memory allocation
4006 * in overflow_handler can crash depending on where nmi got
4009 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
4010 if (!check_map_prealloc(map)) {
4011 verbose("perf_event programs can only use preallocated hash map\n");
4014 if (map->inner_map_meta &&
4015 !check_map_prealloc(map->inner_map_meta)) {
4016 verbose("perf_event programs can only use preallocated inner hash map\n");
4023 /* look for pseudo eBPF instructions that access map FDs and
4024 * replace them with actual map pointers
4026 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
4028 struct bpf_insn *insn = env->prog->insnsi;
4029 int insn_cnt = env->prog->len;
4032 err = bpf_prog_calc_tag(env->prog);
4036 for (i = 0; i < insn_cnt; i++, insn++) {
4037 if (BPF_CLASS(insn->code) == BPF_LDX &&
4038 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
4039 verbose("BPF_LDX uses reserved fields\n");
4043 if (BPF_CLASS(insn->code) == BPF_STX &&
4044 ((BPF_MODE(insn->code) != BPF_MEM &&
4045 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
4046 verbose("BPF_STX uses reserved fields\n");
4050 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
4051 struct bpf_map *map;
4054 if (i == insn_cnt - 1 || insn[1].code != 0 ||
4055 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
4057 verbose("invalid bpf_ld_imm64 insn\n");
4061 if (insn->src_reg == 0)
4062 /* valid generic load 64-bit imm */
4065 if (insn->src_reg != BPF_PSEUDO_MAP_FD) {
4066 verbose("unrecognized bpf_ld_imm64 insn\n");
4070 f = fdget(insn->imm);
4071 map = __bpf_map_get(f);
4073 verbose("fd %d is not pointing to valid bpf_map\n",
4075 return PTR_ERR(map);
4078 err = check_map_prog_compatibility(map, env->prog);
4084 /* store map pointer inside BPF_LD_IMM64 instruction */
4085 insn[0].imm = (u32) (unsigned long) map;
4086 insn[1].imm = ((u64) (unsigned long) map) >> 32;
4088 /* check whether we recorded this map already */
4089 for (j = 0; j < env->used_map_cnt; j++)
4090 if (env->used_maps[j] == map) {
4095 if (env->used_map_cnt >= MAX_USED_MAPS) {
4100 /* hold the map. If the program is rejected by verifier,
4101 * the map will be released by release_maps() or it
4102 * will be used by the valid program until it's unloaded
4103 * and all maps are released in free_bpf_prog_info()
4105 map = bpf_map_inc(map, false);
4108 return PTR_ERR(map);
4110 env->used_maps[env->used_map_cnt++] = map;
4119 /* now all pseudo BPF_LD_IMM64 instructions load valid
4120 * 'struct bpf_map *' into a register instead of user map_fd.
4121 * These pointers will be used later by verifier to validate map access.
4126 /* drop refcnt of maps used by the rejected program */
4127 static void release_maps(struct bpf_verifier_env *env)
4131 for (i = 0; i < env->used_map_cnt; i++)
4132 bpf_map_put(env->used_maps[i]);
4135 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
4136 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
4138 struct bpf_insn *insn = env->prog->insnsi;
4139 int insn_cnt = env->prog->len;
4142 for (i = 0; i < insn_cnt; i++, insn++)
4143 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
4147 /* single env->prog->insni[off] instruction was replaced with the range
4148 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
4149 * [0, off) and [off, end) to new locations, so the patched range stays zero
4151 static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len,
4154 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
4159 new_data = vzalloc(sizeof(struct bpf_insn_aux_data) * prog_len);
4162 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
4163 memcpy(new_data + off + cnt - 1, old_data + off,
4164 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
4165 for (i = off; i < off + cnt - 1; i++)
4166 new_data[i].seen = true;
4167 env->insn_aux_data = new_data;
4172 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
4173 const struct bpf_insn *patch, u32 len)
4175 struct bpf_prog *new_prog;
4177 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
4180 if (adjust_insn_aux_data(env, new_prog->len, off, len))
4185 /* The verifier does more data flow analysis than llvm and will not explore
4186 * branches that are dead at run time. Malicious programs can have dead code
4187 * too. Therefore replace all dead at-run-time code with nops.
4189 static void sanitize_dead_code(struct bpf_verifier_env *env)
4191 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
4192 struct bpf_insn nop = BPF_MOV64_REG(BPF_REG_0, BPF_REG_0);
4193 struct bpf_insn *insn = env->prog->insnsi;
4194 const int insn_cnt = env->prog->len;
4197 for (i = 0; i < insn_cnt; i++) {
4198 if (aux_data[i].seen)
4200 memcpy(insn + i, &nop, sizeof(nop));
4204 /* convert load instructions that access fields of 'struct __sk_buff'
4205 * into sequence of instructions that access fields of 'struct sk_buff'
4207 static int convert_ctx_accesses(struct bpf_verifier_env *env)
4209 const struct bpf_verifier_ops *ops = env->prog->aux->ops;
4210 int i, cnt, size, ctx_field_size, delta = 0;
4211 const int insn_cnt = env->prog->len;
4212 struct bpf_insn insn_buf[16], *insn;
4213 struct bpf_prog *new_prog;
4214 enum bpf_access_type type;
4215 bool is_narrower_load;
4218 if (ops->gen_prologue) {
4219 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
4221 if (cnt >= ARRAY_SIZE(insn_buf)) {
4222 verbose("bpf verifier is misconfigured\n");
4225 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
4229 env->prog = new_prog;
4234 if (!ops->convert_ctx_access)
4237 insn = env->prog->insnsi + delta;
4239 for (i = 0; i < insn_cnt; i++, insn++) {
4240 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
4241 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
4242 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
4243 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
4245 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
4246 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
4247 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
4248 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
4253 if (env->insn_aux_data[i + delta].ptr_type != PTR_TO_CTX)
4256 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
4257 size = BPF_LDST_BYTES(insn);
4259 /* If the read access is a narrower load of the field,
4260 * convert to a 4/8-byte load, to minimum program type specific
4261 * convert_ctx_access changes. If conversion is successful,
4262 * we will apply proper mask to the result.
4264 is_narrower_load = size < ctx_field_size;
4265 if (is_narrower_load) {
4266 u32 off = insn->off;
4269 if (type == BPF_WRITE) {
4270 verbose("bpf verifier narrow ctx access misconfigured\n");
4275 if (ctx_field_size == 4)
4277 else if (ctx_field_size == 8)
4280 insn->off = off & ~(ctx_field_size - 1);
4281 insn->code = BPF_LDX | BPF_MEM | size_code;
4285 cnt = ops->convert_ctx_access(type, insn, insn_buf, env->prog,
4287 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
4288 (ctx_field_size && !target_size)) {
4289 verbose("bpf verifier is misconfigured\n");
4293 if (is_narrower_load && size < target_size) {
4294 if (ctx_field_size <= 4)
4295 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
4296 (1 << size * 8) - 1);
4298 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
4299 (1 << size * 8) - 1);
4302 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
4308 /* keep walking new program and skip insns we just inserted */
4309 env->prog = new_prog;
4310 insn = new_prog->insnsi + i + delta;
4316 /* fixup insn->imm field of bpf_call instructions
4317 * and inline eligible helpers as explicit sequence of BPF instructions
4319 * this function is called after eBPF program passed verification
4321 static int fixup_bpf_calls(struct bpf_verifier_env *env)
4323 struct bpf_prog *prog = env->prog;
4324 struct bpf_insn *insn = prog->insnsi;
4325 const struct bpf_func_proto *fn;
4326 const int insn_cnt = prog->len;
4327 struct bpf_insn insn_buf[16];
4328 struct bpf_prog *new_prog;
4329 struct bpf_map *map_ptr;
4330 int i, cnt, delta = 0;
4332 for (i = 0; i < insn_cnt; i++, insn++) {
4333 if (insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
4334 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
4335 /* due to JIT bugs clear upper 32-bits of src register
4336 * before div/mod operation
4338 insn_buf[0] = BPF_MOV32_REG(insn->src_reg, insn->src_reg);
4339 insn_buf[1] = *insn;
4341 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
4346 env->prog = prog = new_prog;
4347 insn = new_prog->insnsi + i + delta;
4351 if (insn->code != (BPF_JMP | BPF_CALL))
4354 if (insn->imm == BPF_FUNC_get_route_realm)
4355 prog->dst_needed = 1;
4356 if (insn->imm == BPF_FUNC_get_prandom_u32)
4357 bpf_user_rnd_init_once();
4358 if (insn->imm == BPF_FUNC_tail_call) {
4359 /* If we tail call into other programs, we
4360 * cannot make any assumptions since they can
4361 * be replaced dynamically during runtime in
4362 * the program array.
4364 prog->cb_access = 1;
4365 env->prog->aux->stack_depth = MAX_BPF_STACK;
4367 /* mark bpf_tail_call as different opcode to avoid
4368 * conditional branch in the interpeter for every normal
4369 * call and to prevent accidental JITing by JIT compiler
4370 * that doesn't support bpf_tail_call yet
4373 insn->code = BPF_JMP | BPF_TAIL_CALL;
4375 /* instead of changing every JIT dealing with tail_call
4376 * emit two extra insns:
4377 * if (index >= max_entries) goto out;
4378 * index &= array->index_mask;
4379 * to avoid out-of-bounds cpu speculation
4381 map_ptr = env->insn_aux_data[i + delta].map_ptr;
4382 if (map_ptr == BPF_MAP_PTR_POISON) {
4383 verbose("tail_call obusing map_ptr\n");
4386 if (!map_ptr->unpriv_array)
4388 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
4389 map_ptr->max_entries, 2);
4390 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
4391 container_of(map_ptr,
4394 insn_buf[2] = *insn;
4396 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
4401 env->prog = prog = new_prog;
4402 insn = new_prog->insnsi + i + delta;
4406 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
4407 * handlers are currently limited to 64 bit only.
4409 if (ebpf_jit_enabled() && BITS_PER_LONG == 64 &&
4410 insn->imm == BPF_FUNC_map_lookup_elem) {
4411 map_ptr = env->insn_aux_data[i + delta].map_ptr;
4412 if (map_ptr == BPF_MAP_PTR_POISON ||
4413 !map_ptr->ops->map_gen_lookup)
4414 goto patch_call_imm;
4416 cnt = map_ptr->ops->map_gen_lookup(map_ptr, insn_buf);
4417 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
4418 verbose("bpf verifier is misconfigured\n");
4422 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
4429 /* keep walking new program and skip insns we just inserted */
4430 env->prog = prog = new_prog;
4431 insn = new_prog->insnsi + i + delta;
4435 if (insn->imm == BPF_FUNC_redirect_map) {
4436 /* Note, we cannot use prog directly as imm as subsequent
4437 * rewrites would still change the prog pointer. The only
4438 * stable address we can use is aux, which also works with
4439 * prog clones during blinding.
4441 u64 addr = (unsigned long)prog->aux;
4442 struct bpf_insn r4_ld[] = {
4443 BPF_LD_IMM64(BPF_REG_4, addr),
4446 cnt = ARRAY_SIZE(r4_ld);
4448 new_prog = bpf_patch_insn_data(env, i + delta, r4_ld, cnt);
4453 env->prog = prog = new_prog;
4454 insn = new_prog->insnsi + i + delta;
4457 fn = prog->aux->ops->get_func_proto(insn->imm);
4458 /* all functions that have prototype and verifier allowed
4459 * programs to call them, must be real in-kernel functions
4462 verbose("kernel subsystem misconfigured func %s#%d\n",
4463 func_id_name(insn->imm), insn->imm);
4466 insn->imm = fn->func - __bpf_call_base;
4472 static void free_states(struct bpf_verifier_env *env)
4474 struct bpf_verifier_state_list *sl, *sln;
4477 if (!env->explored_states)
4480 for (i = 0; i < env->prog->len; i++) {
4481 sl = env->explored_states[i];
4484 while (sl != STATE_LIST_MARK) {
4491 kfree(env->explored_states);
4494 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr)
4496 char __user *log_ubuf = NULL;
4497 struct bpf_verifier_env *env;
4500 /* 'struct bpf_verifier_env' can be global, but since it's not small,
4501 * allocate/free it every time bpf_check() is called
4503 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
4507 env->insn_aux_data = vzalloc(sizeof(struct bpf_insn_aux_data) *
4510 if (!env->insn_aux_data)
4514 /* grab the mutex to protect few globals used by verifier */
4515 mutex_lock(&bpf_verifier_lock);
4517 if (attr->log_level || attr->log_buf || attr->log_size) {
4518 /* user requested verbose verifier output
4519 * and supplied buffer to store the verification trace
4521 log_level = attr->log_level;
4522 log_ubuf = (char __user *) (unsigned long) attr->log_buf;
4523 log_size = attr->log_size;
4527 /* log_* values have to be sane */
4528 if (log_size < 128 || log_size > UINT_MAX >> 8 ||
4529 log_level == 0 || log_ubuf == NULL)
4533 log_buf = vmalloc(log_size);
4540 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
4541 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
4542 env->strict_alignment = true;
4544 ret = replace_map_fd_with_map_ptr(env);
4546 goto skip_full_check;
4548 env->explored_states = kcalloc(env->prog->len,
4549 sizeof(struct bpf_verifier_state_list *),
4552 if (!env->explored_states)
4553 goto skip_full_check;
4555 ret = check_cfg(env);
4557 goto skip_full_check;
4559 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
4561 ret = do_check(env);
4564 while (pop_stack(env, NULL) >= 0);
4568 sanitize_dead_code(env);
4571 /* program is valid, convert *(u32*)(ctx + off) accesses */
4572 ret = convert_ctx_accesses(env);
4575 ret = fixup_bpf_calls(env);
4577 if (log_level && log_len >= log_size - 1) {
4578 BUG_ON(log_len >= log_size);
4579 /* verifier log exceeded user supplied buffer */
4581 /* fall through to return what was recorded */
4584 /* copy verifier log back to user space including trailing zero */
4585 if (log_level && copy_to_user(log_ubuf, log_buf, log_len + 1) != 0) {
4590 if (ret == 0 && env->used_map_cnt) {
4591 /* if program passed verifier, update used_maps in bpf_prog_info */
4592 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
4593 sizeof(env->used_maps[0]),
4596 if (!env->prog->aux->used_maps) {
4601 memcpy(env->prog->aux->used_maps, env->used_maps,
4602 sizeof(env->used_maps[0]) * env->used_map_cnt);
4603 env->prog->aux->used_map_cnt = env->used_map_cnt;
4605 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
4606 * bpf_ld_imm64 instructions
4608 convert_pseudo_ld_imm64(env);
4614 if (!env->prog->aux->used_maps)
4615 /* if we didn't copy map pointers into bpf_prog_info, release
4616 * them now. Otherwise free_bpf_prog_info() will release them.
4621 mutex_unlock(&bpf_verifier_lock);
4622 vfree(env->insn_aux_data);
4628 int bpf_analyzer(struct bpf_prog *prog, const struct bpf_ext_analyzer_ops *ops,
4631 struct bpf_verifier_env *env;
4634 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
4638 env->insn_aux_data = vzalloc(sizeof(struct bpf_insn_aux_data) *
4641 if (!env->insn_aux_data)
4644 env->analyzer_ops = ops;
4645 env->analyzer_priv = priv;
4647 /* grab the mutex to protect few globals used by verifier */
4648 mutex_lock(&bpf_verifier_lock);
4652 env->strict_alignment = false;
4653 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
4654 env->strict_alignment = true;
4656 env->explored_states = kcalloc(env->prog->len,
4657 sizeof(struct bpf_verifier_state_list *),
4660 if (!env->explored_states)
4661 goto skip_full_check;
4663 ret = check_cfg(env);
4665 goto skip_full_check;
4667 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
4669 ret = do_check(env);
4672 while (pop_stack(env, NULL) >= 0);
4675 mutex_unlock(&bpf_verifier_lock);
4676 vfree(env->insn_aux_data);
4681 EXPORT_SYMBOL_GPL(bpf_analyzer);