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;
657 /* if read wasn't screened by an earlier write ... */
658 if (state->regs[regno].live & REG_LIVE_WRITTEN)
660 /* ... then we depend on parent's value */
661 parent->regs[regno].live |= REG_LIVE_READ;
663 parent = state->parent;
667 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
670 struct bpf_reg_state *regs = env->cur_state.regs;
672 if (regno >= MAX_BPF_REG) {
673 verbose("R%d is invalid\n", regno);
678 /* check whether register used as source operand can be read */
679 if (regs[regno].type == NOT_INIT) {
680 verbose("R%d !read_ok\n", regno);
683 mark_reg_read(&env->cur_state, regno);
685 /* check whether register used as dest operand can be written to */
686 if (regno == BPF_REG_FP) {
687 verbose("frame pointer is read only\n");
690 regs[regno].live |= REG_LIVE_WRITTEN;
692 mark_reg_unknown(regs, regno);
697 static bool is_spillable_regtype(enum bpf_reg_type type)
700 case PTR_TO_MAP_VALUE:
701 case PTR_TO_MAP_VALUE_OR_NULL:
705 case PTR_TO_PACKET_END:
706 case CONST_PTR_TO_MAP:
713 /* check_stack_read/write functions track spill/fill of registers,
714 * stack boundary and alignment are checked in check_mem_access()
716 static int check_stack_write(struct bpf_verifier_state *state, int off,
717 int size, int value_regno)
719 int i, spi = (MAX_BPF_STACK + off) / BPF_REG_SIZE;
720 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
721 * so it's aligned access and [off, off + size) are within stack limits
724 if (value_regno >= 0 &&
725 is_spillable_regtype(state->regs[value_regno].type)) {
727 /* register containing pointer is being spilled into stack */
728 if (size != BPF_REG_SIZE) {
729 verbose("invalid size of register spill\n");
733 /* save register state */
734 state->spilled_regs[spi] = state->regs[value_regno];
735 state->spilled_regs[spi].live |= REG_LIVE_WRITTEN;
737 for (i = 0; i < BPF_REG_SIZE; i++)
738 state->stack_slot_type[MAX_BPF_STACK + off + i] = STACK_SPILL;
740 /* regular write of data into stack */
741 state->spilled_regs[spi] = (struct bpf_reg_state) {};
743 for (i = 0; i < size; i++)
744 state->stack_slot_type[MAX_BPF_STACK + off + i] = STACK_MISC;
749 static void mark_stack_slot_read(const struct bpf_verifier_state *state, int slot)
751 struct bpf_verifier_state *parent = state->parent;
754 /* if read wasn't screened by an earlier write ... */
755 if (state->spilled_regs[slot].live & REG_LIVE_WRITTEN)
757 /* ... then we depend on parent's value */
758 parent->spilled_regs[slot].live |= REG_LIVE_READ;
760 parent = state->parent;
764 static int check_stack_read(struct bpf_verifier_state *state, int off, int size,
770 slot_type = &state->stack_slot_type[MAX_BPF_STACK + off];
772 if (slot_type[0] == STACK_SPILL) {
773 if (size != BPF_REG_SIZE) {
774 verbose("invalid size of register spill\n");
777 for (i = 1; i < BPF_REG_SIZE; i++) {
778 if (slot_type[i] != STACK_SPILL) {
779 verbose("corrupted spill memory\n");
784 spi = (MAX_BPF_STACK + off) / BPF_REG_SIZE;
786 if (value_regno >= 0) {
787 /* restore register state from stack */
788 state->regs[value_regno] = state->spilled_regs[spi];
789 mark_stack_slot_read(state, spi);
793 for (i = 0; i < size; i++) {
794 if (slot_type[i] != STACK_MISC) {
795 verbose("invalid read from stack off %d+%d size %d\n",
800 if (value_regno >= 0)
801 /* have read misc data from the stack */
802 mark_reg_unknown(state->regs, value_regno);
807 /* check read/write into map element returned by bpf_map_lookup_elem() */
808 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
811 struct bpf_map *map = env->cur_state.regs[regno].map_ptr;
813 if (off < 0 || size <= 0 || off + size > map->value_size) {
814 verbose("invalid access to map value, value_size=%d off=%d size=%d\n",
815 map->value_size, off, size);
821 /* check read/write into a map element with possible variable offset */
822 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
825 struct bpf_verifier_state *state = &env->cur_state;
826 struct bpf_reg_state *reg = &state->regs[regno];
829 /* We may have adjusted the register to this map value, so we
830 * need to try adding each of min_value and max_value to off
831 * to make sure our theoretical access will be safe.
834 print_verifier_state(state);
835 /* If the offset is variable, we will need to be stricter in state
836 * pruning from now on.
838 if (!tnum_is_const(reg->var_off))
839 env->varlen_map_value_access = true;
840 /* The minimum value is only important with signed
841 * comparisons where we can't assume the floor of a
842 * value is 0. If we are using signed variables for our
843 * index'es we need to make sure that whatever we use
844 * will have a set floor within our range.
846 if (reg->smin_value < 0) {
847 verbose("R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
851 err = __check_map_access(env, regno, reg->smin_value + off, size);
853 verbose("R%d min value is outside of the array range\n", regno);
857 /* If we haven't set a max value then we need to bail since we can't be
858 * sure we won't do bad things.
859 * If reg->umax_value + off could overflow, treat that as unbounded too.
861 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
862 verbose("R%d unbounded memory access, make sure to bounds check any array access into a map\n",
866 err = __check_map_access(env, regno, reg->umax_value + off, size);
868 verbose("R%d max value is outside of the array range\n", regno);
872 #define MAX_PACKET_OFF 0xffff
874 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
875 const struct bpf_call_arg_meta *meta,
876 enum bpf_access_type t)
878 switch (env->prog->type) {
879 case BPF_PROG_TYPE_LWT_IN:
880 case BPF_PROG_TYPE_LWT_OUT:
881 /* dst_input() and dst_output() can't write for now */
885 case BPF_PROG_TYPE_SCHED_CLS:
886 case BPF_PROG_TYPE_SCHED_ACT:
887 case BPF_PROG_TYPE_XDP:
888 case BPF_PROG_TYPE_LWT_XMIT:
889 case BPF_PROG_TYPE_SK_SKB:
891 return meta->pkt_access;
893 env->seen_direct_write = true;
900 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
903 struct bpf_reg_state *regs = env->cur_state.regs;
904 struct bpf_reg_state *reg = ®s[regno];
906 if (off < 0 || size <= 0 || (u64)off + size > reg->range) {
907 verbose("invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
908 off, size, regno, reg->id, reg->off, reg->range);
914 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
917 struct bpf_reg_state *regs = env->cur_state.regs;
918 struct bpf_reg_state *reg = ®s[regno];
921 /* We may have added a variable offset to the packet pointer; but any
922 * reg->range we have comes after that. We are only checking the fixed
926 /* We don't allow negative numbers, because we aren't tracking enough
927 * detail to prove they're safe.
929 if (reg->smin_value < 0) {
930 verbose("R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
934 err = __check_packet_access(env, regno, off, size);
936 verbose("R%d offset is outside of the packet\n", regno);
942 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
943 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
944 enum bpf_access_type t, enum bpf_reg_type *reg_type)
946 struct bpf_insn_access_aux info = {
947 .reg_type = *reg_type,
950 /* for analyzer ctx accesses are already validated and converted */
951 if (env->analyzer_ops)
954 if (env->prog->aux->ops->is_valid_access &&
955 env->prog->aux->ops->is_valid_access(off, size, t, &info)) {
956 /* A non zero info.ctx_field_size indicates that this field is a
957 * candidate for later verifier transformation to load the whole
958 * field and then apply a mask when accessed with a narrower
959 * access than actual ctx access size. A zero info.ctx_field_size
960 * will only allow for whole field access and rejects any other
961 * type of narrower access.
963 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
964 *reg_type = info.reg_type;
966 /* remember the offset of last byte accessed in ctx */
967 if (env->prog->aux->max_ctx_offset < off + size)
968 env->prog->aux->max_ctx_offset = off + size;
972 verbose("invalid bpf_context access off=%d size=%d\n", off, size);
976 static bool __is_pointer_value(bool allow_ptr_leaks,
977 const struct bpf_reg_state *reg)
982 return reg->type != SCALAR_VALUE;
985 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
987 return __is_pointer_value(env->allow_ptr_leaks, &env->cur_state.regs[regno]);
990 static int check_pkt_ptr_alignment(const struct bpf_reg_state *reg,
991 int off, int size, bool strict)
996 /* Byte size accesses are always allowed. */
997 if (!strict || size == 1)
1000 /* For platforms that do not have a Kconfig enabling
1001 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
1002 * NET_IP_ALIGN is universally set to '2'. And on platforms
1003 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
1004 * to this code only in strict mode where we want to emulate
1005 * the NET_IP_ALIGN==2 checking. Therefore use an
1006 * unconditional IP align value of '2'.
1010 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
1011 if (!tnum_is_aligned(reg_off, size)) {
1014 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1015 verbose("misaligned packet access off %d+%s+%d+%d size %d\n",
1016 ip_align, tn_buf, reg->off, off, size);
1023 static int check_generic_ptr_alignment(const struct bpf_reg_state *reg,
1024 const char *pointer_desc,
1025 int off, int size, bool strict)
1027 struct tnum reg_off;
1029 /* Byte size accesses are always allowed. */
1030 if (!strict || size == 1)
1033 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
1034 if (!tnum_is_aligned(reg_off, size)) {
1037 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1038 verbose("misaligned %saccess off %s+%d+%d size %d\n",
1039 pointer_desc, tn_buf, reg->off, off, size);
1046 static int check_ptr_alignment(struct bpf_verifier_env *env,
1047 const struct bpf_reg_state *reg,
1050 bool strict = env->strict_alignment;
1051 const char *pointer_desc = "";
1053 switch (reg->type) {
1055 /* special case, because of NET_IP_ALIGN */
1056 return check_pkt_ptr_alignment(reg, off, size, strict);
1057 case PTR_TO_MAP_VALUE:
1058 pointer_desc = "value ";
1061 pointer_desc = "context ";
1064 pointer_desc = "stack ";
1069 return check_generic_ptr_alignment(reg, pointer_desc, off, size, strict);
1072 /* check whether memory at (regno + off) is accessible for t = (read | write)
1073 * if t==write, value_regno is a register which value is stored into memory
1074 * if t==read, value_regno is a register which will receive the value from memory
1075 * if t==write && value_regno==-1, some unknown value is stored into memory
1076 * if t==read && value_regno==-1, don't care what we read from memory
1078 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno, int off,
1079 int bpf_size, enum bpf_access_type t,
1082 struct bpf_verifier_state *state = &env->cur_state;
1083 struct bpf_reg_state *reg = &state->regs[regno];
1086 size = bpf_size_to_bytes(bpf_size);
1090 /* alignment checks will add in reg->off themselves */
1091 err = check_ptr_alignment(env, reg, off, size);
1095 /* for access checks, reg->off is just part of off */
1098 if (reg->type == PTR_TO_MAP_VALUE) {
1099 if (t == BPF_WRITE && value_regno >= 0 &&
1100 is_pointer_value(env, value_regno)) {
1101 verbose("R%d leaks addr into map\n", value_regno);
1105 err = check_map_access(env, regno, off, size);
1106 if (!err && t == BPF_READ && value_regno >= 0)
1107 mark_reg_unknown(state->regs, value_regno);
1109 } else if (reg->type == PTR_TO_CTX) {
1110 enum bpf_reg_type reg_type = SCALAR_VALUE;
1112 if (t == BPF_WRITE && value_regno >= 0 &&
1113 is_pointer_value(env, value_regno)) {
1114 verbose("R%d leaks addr into ctx\n", value_regno);
1117 /* ctx accesses must be at a fixed offset, so that we can
1118 * determine what type of data were returned.
1120 if (!tnum_is_const(reg->var_off)) {
1123 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1124 verbose("variable ctx access var_off=%s off=%d size=%d",
1128 off += reg->var_off.value;
1129 err = check_ctx_access(env, insn_idx, off, size, t, ®_type);
1130 if (!err && t == BPF_READ && value_regno >= 0) {
1131 /* ctx access returns either a scalar, or a
1132 * PTR_TO_PACKET[_END]. In the latter case, we know
1133 * the offset is zero.
1135 if (reg_type == SCALAR_VALUE)
1136 mark_reg_unknown(state->regs, value_regno);
1138 mark_reg_known_zero(state->regs, value_regno);
1139 state->regs[value_regno].id = 0;
1140 state->regs[value_regno].off = 0;
1141 state->regs[value_regno].range = 0;
1142 state->regs[value_regno].type = reg_type;
1145 } else if (reg->type == PTR_TO_STACK) {
1146 /* stack accesses must be at a fixed offset, so that we can
1147 * determine what type of data were returned.
1148 * See check_stack_read().
1150 if (!tnum_is_const(reg->var_off)) {
1153 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1154 verbose("variable stack access var_off=%s off=%d size=%d",
1158 off += reg->var_off.value;
1159 if (off >= 0 || off < -MAX_BPF_STACK) {
1160 verbose("invalid stack off=%d size=%d\n", off, size);
1164 if (env->prog->aux->stack_depth < -off)
1165 env->prog->aux->stack_depth = -off;
1167 if (t == BPF_WRITE) {
1168 if (!env->allow_ptr_leaks &&
1169 state->stack_slot_type[MAX_BPF_STACK + off] == STACK_SPILL &&
1170 size != BPF_REG_SIZE) {
1171 verbose("attempt to corrupt spilled pointer on stack\n");
1174 err = check_stack_write(state, off, size, value_regno);
1176 err = check_stack_read(state, off, size, value_regno);
1178 } else if (reg->type == PTR_TO_PACKET) {
1179 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
1180 verbose("cannot write into packet\n");
1183 if (t == BPF_WRITE && value_regno >= 0 &&
1184 is_pointer_value(env, value_regno)) {
1185 verbose("R%d leaks addr into packet\n", value_regno);
1188 err = check_packet_access(env, regno, off, size);
1189 if (!err && t == BPF_READ && value_regno >= 0)
1190 mark_reg_unknown(state->regs, value_regno);
1192 verbose("R%d invalid mem access '%s'\n",
1193 regno, reg_type_str[reg->type]);
1197 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
1198 state->regs[value_regno].type == SCALAR_VALUE) {
1199 /* b/h/w load zero-extends, mark upper bits as known 0 */
1200 state->regs[value_regno].var_off = tnum_cast(
1201 state->regs[value_regno].var_off, size);
1202 __update_reg_bounds(&state->regs[value_regno]);
1207 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
1211 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
1213 verbose("BPF_XADD uses reserved fields\n");
1217 /* check src1 operand */
1218 err = check_reg_arg(env, insn->src_reg, SRC_OP);
1222 /* check src2 operand */
1223 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
1227 if (is_pointer_value(env, insn->src_reg)) {
1228 verbose("R%d leaks addr into mem\n", insn->src_reg);
1232 /* check whether atomic_add can read the memory */
1233 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1234 BPF_SIZE(insn->code), BPF_READ, -1);
1238 /* check whether atomic_add can write into the same memory */
1239 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1240 BPF_SIZE(insn->code), BPF_WRITE, -1);
1243 /* Does this register contain a constant zero? */
1244 static bool register_is_null(struct bpf_reg_state reg)
1246 return reg.type == SCALAR_VALUE && tnum_equals_const(reg.var_off, 0);
1249 /* when register 'regno' is passed into function that will read 'access_size'
1250 * bytes from that pointer, make sure that it's within stack boundary
1251 * and all elements of stack are initialized.
1252 * Unlike most pointer bounds-checking functions, this one doesn't take an
1253 * 'off' argument, so it has to add in reg->off itself.
1255 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
1256 int access_size, bool zero_size_allowed,
1257 struct bpf_call_arg_meta *meta)
1259 struct bpf_verifier_state *state = &env->cur_state;
1260 struct bpf_reg_state *regs = state->regs;
1263 if (regs[regno].type != PTR_TO_STACK) {
1264 /* Allow zero-byte read from NULL, regardless of pointer type */
1265 if (zero_size_allowed && access_size == 0 &&
1266 register_is_null(regs[regno]))
1269 verbose("R%d type=%s expected=%s\n", regno,
1270 reg_type_str[regs[regno].type],
1271 reg_type_str[PTR_TO_STACK]);
1275 /* Only allow fixed-offset stack reads */
1276 if (!tnum_is_const(regs[regno].var_off)) {
1279 tnum_strn(tn_buf, sizeof(tn_buf), regs[regno].var_off);
1280 verbose("invalid variable stack read R%d var_off=%s\n",
1283 off = regs[regno].off + regs[regno].var_off.value;
1284 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
1286 verbose("invalid stack type R%d off=%d access_size=%d\n",
1287 regno, off, access_size);
1291 if (env->prog->aux->stack_depth < -off)
1292 env->prog->aux->stack_depth = -off;
1294 if (meta && meta->raw_mode) {
1295 meta->access_size = access_size;
1296 meta->regno = regno;
1300 for (i = 0; i < access_size; i++) {
1301 if (state->stack_slot_type[MAX_BPF_STACK + off + i] != STACK_MISC) {
1302 verbose("invalid indirect read from stack off %d+%d size %d\n",
1303 off, i, access_size);
1310 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
1311 int access_size, bool zero_size_allowed,
1312 struct bpf_call_arg_meta *meta)
1314 struct bpf_reg_state *regs = env->cur_state.regs, *reg = ®s[regno];
1316 switch (reg->type) {
1318 return check_packet_access(env, regno, reg->off, access_size);
1319 case PTR_TO_MAP_VALUE:
1320 return check_map_access(env, regno, reg->off, access_size);
1321 default: /* scalar_value|ptr_to_stack or invalid ptr */
1322 return check_stack_boundary(env, regno, access_size,
1323 zero_size_allowed, meta);
1327 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
1328 enum bpf_arg_type arg_type,
1329 struct bpf_call_arg_meta *meta)
1331 struct bpf_reg_state *regs = env->cur_state.regs, *reg = ®s[regno];
1332 enum bpf_reg_type expected_type, type = reg->type;
1335 if (arg_type == ARG_DONTCARE)
1338 err = check_reg_arg(env, regno, SRC_OP);
1342 if (arg_type == ARG_ANYTHING) {
1343 if (is_pointer_value(env, regno)) {
1344 verbose("R%d leaks addr into helper function\n", regno);
1350 if (type == PTR_TO_PACKET &&
1351 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
1352 verbose("helper access to the packet is not allowed\n");
1356 if (arg_type == ARG_PTR_TO_MAP_KEY ||
1357 arg_type == ARG_PTR_TO_MAP_VALUE) {
1358 expected_type = PTR_TO_STACK;
1359 if (type != PTR_TO_PACKET && type != expected_type)
1361 } else if (arg_type == ARG_CONST_SIZE ||
1362 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1363 expected_type = SCALAR_VALUE;
1364 if (type != expected_type)
1366 } else if (arg_type == ARG_CONST_MAP_PTR) {
1367 expected_type = CONST_PTR_TO_MAP;
1368 if (type != expected_type)
1370 } else if (arg_type == ARG_PTR_TO_CTX) {
1371 expected_type = PTR_TO_CTX;
1372 if (type != expected_type)
1374 } else if (arg_type == ARG_PTR_TO_MEM ||
1375 arg_type == ARG_PTR_TO_UNINIT_MEM) {
1376 expected_type = PTR_TO_STACK;
1377 /* One exception here. In case function allows for NULL to be
1378 * passed in as argument, it's a SCALAR_VALUE type. Final test
1379 * happens during stack boundary checking.
1381 if (register_is_null(*reg))
1382 /* final test in check_stack_boundary() */;
1383 else if (type != PTR_TO_PACKET && type != PTR_TO_MAP_VALUE &&
1384 type != expected_type)
1386 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
1388 verbose("unsupported arg_type %d\n", arg_type);
1392 if (arg_type == ARG_CONST_MAP_PTR) {
1393 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
1394 meta->map_ptr = reg->map_ptr;
1395 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
1396 /* bpf_map_xxx(..., map_ptr, ..., key) call:
1397 * check that [key, key + map->key_size) are within
1398 * stack limits and initialized
1400 if (!meta->map_ptr) {
1401 /* in function declaration map_ptr must come before
1402 * map_key, so that it's verified and known before
1403 * we have to check map_key here. Otherwise it means
1404 * that kernel subsystem misconfigured verifier
1406 verbose("invalid map_ptr to access map->key\n");
1409 if (type == PTR_TO_PACKET)
1410 err = check_packet_access(env, regno, reg->off,
1411 meta->map_ptr->key_size);
1413 err = check_stack_boundary(env, regno,
1414 meta->map_ptr->key_size,
1416 } else if (arg_type == ARG_PTR_TO_MAP_VALUE) {
1417 /* bpf_map_xxx(..., map_ptr, ..., value) call:
1418 * check [value, value + map->value_size) validity
1420 if (!meta->map_ptr) {
1421 /* kernel subsystem misconfigured verifier */
1422 verbose("invalid map_ptr to access map->value\n");
1425 if (type == PTR_TO_PACKET)
1426 err = check_packet_access(env, regno, reg->off,
1427 meta->map_ptr->value_size);
1429 err = check_stack_boundary(env, regno,
1430 meta->map_ptr->value_size,
1432 } else if (arg_type == ARG_CONST_SIZE ||
1433 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1434 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
1436 /* bpf_xxx(..., buf, len) call will access 'len' bytes
1437 * from stack pointer 'buf'. Check it
1438 * note: regno == len, regno - 1 == buf
1441 /* kernel subsystem misconfigured verifier */
1442 verbose("ARG_CONST_SIZE cannot be first argument\n");
1446 /* The register is SCALAR_VALUE; the access check
1447 * happens using its boundaries.
1450 if (!tnum_is_const(reg->var_off))
1451 /* For unprivileged variable accesses, disable raw
1452 * mode so that the program is required to
1453 * initialize all the memory that the helper could
1454 * just partially fill up.
1458 if (reg->smin_value < 0) {
1459 verbose("R%d min value is negative, either use unsigned or 'var &= const'\n",
1464 if (reg->umin_value == 0) {
1465 err = check_helper_mem_access(env, regno - 1, 0,
1472 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
1473 verbose("R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
1477 err = check_helper_mem_access(env, regno - 1,
1479 zero_size_allowed, meta);
1484 verbose("R%d type=%s expected=%s\n", regno,
1485 reg_type_str[type], reg_type_str[expected_type]);
1489 static int check_map_func_compatibility(struct bpf_map *map, int func_id)
1494 /* We need a two way check, first is from map perspective ... */
1495 switch (map->map_type) {
1496 case BPF_MAP_TYPE_PROG_ARRAY:
1497 if (func_id != BPF_FUNC_tail_call)
1500 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
1501 if (func_id != BPF_FUNC_perf_event_read &&
1502 func_id != BPF_FUNC_perf_event_output)
1505 case BPF_MAP_TYPE_STACK_TRACE:
1506 if (func_id != BPF_FUNC_get_stackid)
1509 case BPF_MAP_TYPE_CGROUP_ARRAY:
1510 if (func_id != BPF_FUNC_skb_under_cgroup &&
1511 func_id != BPF_FUNC_current_task_under_cgroup)
1514 /* devmap returns a pointer to a live net_device ifindex that we cannot
1515 * allow to be modified from bpf side. So do not allow lookup elements
1518 case BPF_MAP_TYPE_DEVMAP:
1519 if (func_id != BPF_FUNC_redirect_map)
1522 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
1523 case BPF_MAP_TYPE_HASH_OF_MAPS:
1524 if (func_id != BPF_FUNC_map_lookup_elem)
1526 case BPF_MAP_TYPE_SOCKMAP:
1527 if (func_id != BPF_FUNC_sk_redirect_map &&
1528 func_id != BPF_FUNC_sock_map_update &&
1529 func_id != BPF_FUNC_map_delete_elem)
1536 /* ... and second from the function itself. */
1538 case BPF_FUNC_tail_call:
1539 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
1542 case BPF_FUNC_perf_event_read:
1543 case BPF_FUNC_perf_event_output:
1544 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
1547 case BPF_FUNC_get_stackid:
1548 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
1551 case BPF_FUNC_current_task_under_cgroup:
1552 case BPF_FUNC_skb_under_cgroup:
1553 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
1556 case BPF_FUNC_redirect_map:
1557 if (map->map_type != BPF_MAP_TYPE_DEVMAP)
1560 case BPF_FUNC_sk_redirect_map:
1561 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
1564 case BPF_FUNC_sock_map_update:
1565 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
1574 verbose("cannot pass map_type %d into func %s#%d\n",
1575 map->map_type, func_id_name(func_id), func_id);
1579 static int check_raw_mode(const struct bpf_func_proto *fn)
1583 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
1585 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
1587 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
1589 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
1591 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
1594 return count > 1 ? -EINVAL : 0;
1597 /* Packet data might have moved, any old PTR_TO_PACKET[_END] are now invalid,
1598 * so turn them into unknown SCALAR_VALUE.
1600 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
1602 struct bpf_verifier_state *state = &env->cur_state;
1603 struct bpf_reg_state *regs = state->regs, *reg;
1606 for (i = 0; i < MAX_BPF_REG; i++)
1607 if (regs[i].type == PTR_TO_PACKET ||
1608 regs[i].type == PTR_TO_PACKET_END)
1609 mark_reg_unknown(regs, i);
1611 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
1612 if (state->stack_slot_type[i] != STACK_SPILL)
1614 reg = &state->spilled_regs[i / BPF_REG_SIZE];
1615 if (reg->type != PTR_TO_PACKET &&
1616 reg->type != PTR_TO_PACKET_END)
1618 __mark_reg_unknown(reg);
1622 static int check_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
1624 struct bpf_verifier_state *state = &env->cur_state;
1625 const struct bpf_func_proto *fn = NULL;
1626 struct bpf_reg_state *regs = state->regs;
1627 struct bpf_call_arg_meta meta;
1631 /* find function prototype */
1632 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
1633 verbose("invalid func %s#%d\n", func_id_name(func_id), func_id);
1637 if (env->prog->aux->ops->get_func_proto)
1638 fn = env->prog->aux->ops->get_func_proto(func_id);
1641 verbose("unknown func %s#%d\n", func_id_name(func_id), func_id);
1645 /* eBPF programs must be GPL compatible to use GPL-ed functions */
1646 if (!env->prog->gpl_compatible && fn->gpl_only) {
1647 verbose("cannot call GPL only function from proprietary program\n");
1651 changes_data = bpf_helper_changes_pkt_data(fn->func);
1653 memset(&meta, 0, sizeof(meta));
1654 meta.pkt_access = fn->pkt_access;
1656 /* We only support one arg being in raw mode at the moment, which
1657 * is sufficient for the helper functions we have right now.
1659 err = check_raw_mode(fn);
1661 verbose("kernel subsystem misconfigured func %s#%d\n",
1662 func_id_name(func_id), func_id);
1667 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
1670 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
1673 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
1676 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
1679 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
1683 /* Mark slots with STACK_MISC in case of raw mode, stack offset
1684 * is inferred from register state.
1686 for (i = 0; i < meta.access_size; i++) {
1687 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B, BPF_WRITE, -1);
1692 /* reset caller saved regs */
1693 for (i = 0; i < CALLER_SAVED_REGS; i++) {
1694 mark_reg_not_init(regs, caller_saved[i]);
1695 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
1698 /* update return register (already marked as written above) */
1699 if (fn->ret_type == RET_INTEGER) {
1700 /* sets type to SCALAR_VALUE */
1701 mark_reg_unknown(regs, BPF_REG_0);
1702 } else if (fn->ret_type == RET_VOID) {
1703 regs[BPF_REG_0].type = NOT_INIT;
1704 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL) {
1705 struct bpf_insn_aux_data *insn_aux;
1707 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
1708 /* There is no offset yet applied, variable or fixed */
1709 mark_reg_known_zero(regs, BPF_REG_0);
1710 regs[BPF_REG_0].off = 0;
1711 /* remember map_ptr, so that check_map_access()
1712 * can check 'value_size' boundary of memory access
1713 * to map element returned from bpf_map_lookup_elem()
1715 if (meta.map_ptr == NULL) {
1716 verbose("kernel subsystem misconfigured verifier\n");
1719 regs[BPF_REG_0].map_ptr = meta.map_ptr;
1720 regs[BPF_REG_0].id = ++env->id_gen;
1721 insn_aux = &env->insn_aux_data[insn_idx];
1722 if (!insn_aux->map_ptr)
1723 insn_aux->map_ptr = meta.map_ptr;
1724 else if (insn_aux->map_ptr != meta.map_ptr)
1725 insn_aux->map_ptr = BPF_MAP_PTR_POISON;
1727 verbose("unknown return type %d of func %s#%d\n",
1728 fn->ret_type, func_id_name(func_id), func_id);
1732 err = check_map_func_compatibility(meta.map_ptr, func_id);
1737 clear_all_pkt_pointers(env);
1741 static void coerce_reg_to_32(struct bpf_reg_state *reg)
1743 /* clear high 32 bits */
1744 reg->var_off = tnum_cast(reg->var_off, 4);
1746 __update_reg_bounds(reg);
1749 static bool signed_add_overflows(s64 a, s64 b)
1751 /* Do the add in u64, where overflow is well-defined */
1752 s64 res = (s64)((u64)a + (u64)b);
1759 static bool signed_sub_overflows(s64 a, s64 b)
1761 /* Do the sub in u64, where overflow is well-defined */
1762 s64 res = (s64)((u64)a - (u64)b);
1769 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
1770 * Caller should also handle BPF_MOV case separately.
1771 * If we return -EACCES, caller may want to try again treating pointer as a
1772 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
1774 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
1775 struct bpf_insn *insn,
1776 const struct bpf_reg_state *ptr_reg,
1777 const struct bpf_reg_state *off_reg)
1779 struct bpf_reg_state *regs = env->cur_state.regs, *dst_reg;
1780 bool known = tnum_is_const(off_reg->var_off);
1781 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
1782 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
1783 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
1784 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
1785 u8 opcode = BPF_OP(insn->code);
1786 u32 dst = insn->dst_reg;
1788 dst_reg = ®s[dst];
1790 if (WARN_ON_ONCE(known && (smin_val != smax_val))) {
1791 print_verifier_state(&env->cur_state);
1792 verbose("verifier internal error: known but bad sbounds\n");
1795 if (WARN_ON_ONCE(known && (umin_val != umax_val))) {
1796 print_verifier_state(&env->cur_state);
1797 verbose("verifier internal error: known but bad ubounds\n");
1801 if (BPF_CLASS(insn->code) != BPF_ALU64) {
1802 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
1803 if (!env->allow_ptr_leaks)
1804 verbose("R%d 32-bit pointer arithmetic prohibited\n",
1809 if (ptr_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
1810 if (!env->allow_ptr_leaks)
1811 verbose("R%d pointer arithmetic on PTR_TO_MAP_VALUE_OR_NULL prohibited, null-check it first\n",
1815 if (ptr_reg->type == CONST_PTR_TO_MAP) {
1816 if (!env->allow_ptr_leaks)
1817 verbose("R%d pointer arithmetic on CONST_PTR_TO_MAP prohibited\n",
1821 if (ptr_reg->type == PTR_TO_PACKET_END) {
1822 if (!env->allow_ptr_leaks)
1823 verbose("R%d pointer arithmetic on PTR_TO_PACKET_END prohibited\n",
1828 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
1829 * The id may be overwritten later if we create a new variable offset.
1831 dst_reg->type = ptr_reg->type;
1832 dst_reg->id = ptr_reg->id;
1836 /* We can take a fixed offset as long as it doesn't overflow
1837 * the s32 'off' field
1839 if (known && (ptr_reg->off + smin_val ==
1840 (s64)(s32)(ptr_reg->off + smin_val))) {
1841 /* pointer += K. Accumulate it into fixed offset */
1842 dst_reg->smin_value = smin_ptr;
1843 dst_reg->smax_value = smax_ptr;
1844 dst_reg->umin_value = umin_ptr;
1845 dst_reg->umax_value = umax_ptr;
1846 dst_reg->var_off = ptr_reg->var_off;
1847 dst_reg->off = ptr_reg->off + smin_val;
1848 dst_reg->range = ptr_reg->range;
1851 /* A new variable offset is created. Note that off_reg->off
1852 * == 0, since it's a scalar.
1853 * dst_reg gets the pointer type and since some positive
1854 * integer value was added to the pointer, give it a new 'id'
1855 * if it's a PTR_TO_PACKET.
1856 * this creates a new 'base' pointer, off_reg (variable) gets
1857 * added into the variable offset, and we copy the fixed offset
1860 if (signed_add_overflows(smin_ptr, smin_val) ||
1861 signed_add_overflows(smax_ptr, smax_val)) {
1862 dst_reg->smin_value = S64_MIN;
1863 dst_reg->smax_value = S64_MAX;
1865 dst_reg->smin_value = smin_ptr + smin_val;
1866 dst_reg->smax_value = smax_ptr + smax_val;
1868 if (umin_ptr + umin_val < umin_ptr ||
1869 umax_ptr + umax_val < umax_ptr) {
1870 dst_reg->umin_value = 0;
1871 dst_reg->umax_value = U64_MAX;
1873 dst_reg->umin_value = umin_ptr + umin_val;
1874 dst_reg->umax_value = umax_ptr + umax_val;
1876 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
1877 dst_reg->off = ptr_reg->off;
1878 if (ptr_reg->type == PTR_TO_PACKET) {
1879 dst_reg->id = ++env->id_gen;
1880 /* something was added to pkt_ptr, set range to zero */
1885 if (dst_reg == off_reg) {
1886 /* scalar -= pointer. Creates an unknown scalar */
1887 if (!env->allow_ptr_leaks)
1888 verbose("R%d tried to subtract pointer from scalar\n",
1892 /* We don't allow subtraction from FP, because (according to
1893 * test_verifier.c test "invalid fp arithmetic", JITs might not
1894 * be able to deal with it.
1896 if (ptr_reg->type == PTR_TO_STACK) {
1897 if (!env->allow_ptr_leaks)
1898 verbose("R%d subtraction from stack pointer prohibited\n",
1902 if (known && (ptr_reg->off - smin_val ==
1903 (s64)(s32)(ptr_reg->off - smin_val))) {
1904 /* pointer -= K. Subtract it from fixed offset */
1905 dst_reg->smin_value = smin_ptr;
1906 dst_reg->smax_value = smax_ptr;
1907 dst_reg->umin_value = umin_ptr;
1908 dst_reg->umax_value = umax_ptr;
1909 dst_reg->var_off = ptr_reg->var_off;
1910 dst_reg->id = ptr_reg->id;
1911 dst_reg->off = ptr_reg->off - smin_val;
1912 dst_reg->range = ptr_reg->range;
1915 /* A new variable offset is created. If the subtrahend is known
1916 * nonnegative, then any reg->range we had before is still good.
1918 if (signed_sub_overflows(smin_ptr, smax_val) ||
1919 signed_sub_overflows(smax_ptr, smin_val)) {
1920 /* Overflow possible, we know nothing */
1921 dst_reg->smin_value = S64_MIN;
1922 dst_reg->smax_value = S64_MAX;
1924 dst_reg->smin_value = smin_ptr - smax_val;
1925 dst_reg->smax_value = smax_ptr - smin_val;
1927 if (umin_ptr < umax_val) {
1928 /* Overflow possible, we know nothing */
1929 dst_reg->umin_value = 0;
1930 dst_reg->umax_value = U64_MAX;
1932 /* Cannot overflow (as long as bounds are consistent) */
1933 dst_reg->umin_value = umin_ptr - umax_val;
1934 dst_reg->umax_value = umax_ptr - umin_val;
1936 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
1937 dst_reg->off = ptr_reg->off;
1938 if (ptr_reg->type == PTR_TO_PACKET) {
1939 dst_reg->id = ++env->id_gen;
1940 /* something was added to pkt_ptr, set range to zero */
1948 /* bitwise ops on pointers are troublesome, prohibit for now.
1949 * (However, in principle we could allow some cases, e.g.
1950 * ptr &= ~3 which would reduce min_value by 3.)
1952 if (!env->allow_ptr_leaks)
1953 verbose("R%d bitwise operator %s on pointer prohibited\n",
1954 dst, bpf_alu_string[opcode >> 4]);
1957 /* other operators (e.g. MUL,LSH) produce non-pointer results */
1958 if (!env->allow_ptr_leaks)
1959 verbose("R%d pointer arithmetic with %s operator prohibited\n",
1960 dst, bpf_alu_string[opcode >> 4]);
1964 __update_reg_bounds(dst_reg);
1965 __reg_deduce_bounds(dst_reg);
1966 __reg_bound_offset(dst_reg);
1970 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
1971 struct bpf_insn *insn,
1972 struct bpf_reg_state *dst_reg,
1973 struct bpf_reg_state src_reg)
1975 struct bpf_reg_state *regs = env->cur_state.regs;
1976 u8 opcode = BPF_OP(insn->code);
1977 bool src_known, dst_known;
1978 s64 smin_val, smax_val;
1979 u64 umin_val, umax_val;
1981 if (BPF_CLASS(insn->code) != BPF_ALU64) {
1982 /* 32-bit ALU ops are (32,32)->64 */
1983 coerce_reg_to_32(dst_reg);
1984 coerce_reg_to_32(&src_reg);
1986 smin_val = src_reg.smin_value;
1987 smax_val = src_reg.smax_value;
1988 umin_val = src_reg.umin_value;
1989 umax_val = src_reg.umax_value;
1990 src_known = tnum_is_const(src_reg.var_off);
1991 dst_known = tnum_is_const(dst_reg->var_off);
1995 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
1996 signed_add_overflows(dst_reg->smax_value, smax_val)) {
1997 dst_reg->smin_value = S64_MIN;
1998 dst_reg->smax_value = S64_MAX;
2000 dst_reg->smin_value += smin_val;
2001 dst_reg->smax_value += smax_val;
2003 if (dst_reg->umin_value + umin_val < umin_val ||
2004 dst_reg->umax_value + umax_val < umax_val) {
2005 dst_reg->umin_value = 0;
2006 dst_reg->umax_value = U64_MAX;
2008 dst_reg->umin_value += umin_val;
2009 dst_reg->umax_value += umax_val;
2011 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
2014 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
2015 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
2016 /* Overflow possible, we know nothing */
2017 dst_reg->smin_value = S64_MIN;
2018 dst_reg->smax_value = S64_MAX;
2020 dst_reg->smin_value -= smax_val;
2021 dst_reg->smax_value -= smin_val;
2023 if (dst_reg->umin_value < umax_val) {
2024 /* Overflow possible, we know nothing */
2025 dst_reg->umin_value = 0;
2026 dst_reg->umax_value = U64_MAX;
2028 /* Cannot overflow (as long as bounds are consistent) */
2029 dst_reg->umin_value -= umax_val;
2030 dst_reg->umax_value -= umin_val;
2032 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
2035 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
2036 if (smin_val < 0 || dst_reg->smin_value < 0) {
2037 /* Ain't nobody got time to multiply that sign */
2038 __mark_reg_unbounded(dst_reg);
2039 __update_reg_bounds(dst_reg);
2042 /* Both values are positive, so we can work with unsigned and
2043 * copy the result to signed (unless it exceeds S64_MAX).
2045 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
2046 /* Potential overflow, we know nothing */
2047 __mark_reg_unbounded(dst_reg);
2048 /* (except what we can learn from the var_off) */
2049 __update_reg_bounds(dst_reg);
2052 dst_reg->umin_value *= umin_val;
2053 dst_reg->umax_value *= umax_val;
2054 if (dst_reg->umax_value > S64_MAX) {
2055 /* Overflow possible, we know nothing */
2056 dst_reg->smin_value = S64_MIN;
2057 dst_reg->smax_value = S64_MAX;
2059 dst_reg->smin_value = dst_reg->umin_value;
2060 dst_reg->smax_value = dst_reg->umax_value;
2064 if (src_known && dst_known) {
2065 __mark_reg_known(dst_reg, dst_reg->var_off.value &
2066 src_reg.var_off.value);
2069 /* We get our minimum from the var_off, since that's inherently
2070 * bitwise. Our maximum is the minimum of the operands' maxima.
2072 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
2073 dst_reg->umin_value = dst_reg->var_off.value;
2074 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
2075 if (dst_reg->smin_value < 0 || smin_val < 0) {
2076 /* Lose signed bounds when ANDing negative numbers,
2077 * ain't nobody got time for that.
2079 dst_reg->smin_value = S64_MIN;
2080 dst_reg->smax_value = S64_MAX;
2082 /* ANDing two positives gives a positive, so safe to
2083 * cast result into s64.
2085 dst_reg->smin_value = dst_reg->umin_value;
2086 dst_reg->smax_value = dst_reg->umax_value;
2088 /* We may learn something more from the var_off */
2089 __update_reg_bounds(dst_reg);
2092 if (src_known && dst_known) {
2093 __mark_reg_known(dst_reg, dst_reg->var_off.value |
2094 src_reg.var_off.value);
2097 /* We get our maximum from the var_off, and our minimum is the
2098 * maximum of the operands' minima
2100 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
2101 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
2102 dst_reg->umax_value = dst_reg->var_off.value |
2103 dst_reg->var_off.mask;
2104 if (dst_reg->smin_value < 0 || smin_val < 0) {
2105 /* Lose signed bounds when ORing negative numbers,
2106 * ain't nobody got time for that.
2108 dst_reg->smin_value = S64_MIN;
2109 dst_reg->smax_value = S64_MAX;
2111 /* ORing two positives gives a positive, so safe to
2112 * cast result into s64.
2114 dst_reg->smin_value = dst_reg->umin_value;
2115 dst_reg->smax_value = dst_reg->umax_value;
2117 /* We may learn something more from the var_off */
2118 __update_reg_bounds(dst_reg);
2121 if (umax_val > 63) {
2122 /* Shifts greater than 63 are undefined. This includes
2123 * shifts by a negative number.
2125 mark_reg_unknown(regs, insn->dst_reg);
2128 /* We lose all sign bit information (except what we can pick
2131 dst_reg->smin_value = S64_MIN;
2132 dst_reg->smax_value = S64_MAX;
2133 /* If we might shift our top bit out, then we know nothing */
2134 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
2135 dst_reg->umin_value = 0;
2136 dst_reg->umax_value = U64_MAX;
2138 dst_reg->umin_value <<= umin_val;
2139 dst_reg->umax_value <<= umax_val;
2142 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
2144 dst_reg->var_off = tnum_lshift(tnum_unknown, umin_val);
2145 /* We may learn something more from the var_off */
2146 __update_reg_bounds(dst_reg);
2149 if (umax_val > 63) {
2150 /* Shifts greater than 63 are undefined. This includes
2151 * shifts by a negative number.
2153 mark_reg_unknown(regs, insn->dst_reg);
2156 /* BPF_RSH is an unsigned shift, so make the appropriate casts */
2157 if (dst_reg->smin_value < 0) {
2159 /* Sign bit will be cleared */
2160 dst_reg->smin_value = 0;
2162 /* Lost sign bit information */
2163 dst_reg->smin_value = S64_MIN;
2164 dst_reg->smax_value = S64_MAX;
2167 dst_reg->smin_value =
2168 (u64)(dst_reg->smin_value) >> umax_val;
2171 dst_reg->var_off = tnum_rshift(dst_reg->var_off,
2174 dst_reg->var_off = tnum_rshift(tnum_unknown, umin_val);
2175 dst_reg->umin_value >>= umax_val;
2176 dst_reg->umax_value >>= umin_val;
2177 /* We may learn something more from the var_off */
2178 __update_reg_bounds(dst_reg);
2181 mark_reg_unknown(regs, insn->dst_reg);
2185 __reg_deduce_bounds(dst_reg);
2186 __reg_bound_offset(dst_reg);
2190 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
2193 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
2194 struct bpf_insn *insn)
2196 struct bpf_reg_state *regs = env->cur_state.regs, *dst_reg, *src_reg;
2197 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
2198 u8 opcode = BPF_OP(insn->code);
2201 dst_reg = ®s[insn->dst_reg];
2203 if (dst_reg->type != SCALAR_VALUE)
2205 if (BPF_SRC(insn->code) == BPF_X) {
2206 src_reg = ®s[insn->src_reg];
2207 if (src_reg->type != SCALAR_VALUE) {
2208 if (dst_reg->type != SCALAR_VALUE) {
2209 /* Combining two pointers by any ALU op yields
2210 * an arbitrary scalar.
2212 if (!env->allow_ptr_leaks) {
2213 verbose("R%d pointer %s pointer prohibited\n",
2215 bpf_alu_string[opcode >> 4]);
2218 mark_reg_unknown(regs, insn->dst_reg);
2221 /* scalar += pointer
2222 * This is legal, but we have to reverse our
2223 * src/dest handling in computing the range
2225 rc = adjust_ptr_min_max_vals(env, insn,
2227 if (rc == -EACCES && env->allow_ptr_leaks) {
2228 /* scalar += unknown scalar */
2229 __mark_reg_unknown(&off_reg);
2230 return adjust_scalar_min_max_vals(
2236 } else if (ptr_reg) {
2237 /* pointer += scalar */
2238 rc = adjust_ptr_min_max_vals(env, insn,
2240 if (rc == -EACCES && env->allow_ptr_leaks) {
2241 /* unknown scalar += scalar */
2242 __mark_reg_unknown(dst_reg);
2243 return adjust_scalar_min_max_vals(
2244 env, insn, dst_reg, *src_reg);
2249 /* Pretend the src is a reg with a known value, since we only
2250 * need to be able to read from this state.
2252 off_reg.type = SCALAR_VALUE;
2253 __mark_reg_known(&off_reg, insn->imm);
2255 if (ptr_reg) { /* pointer += K */
2256 rc = adjust_ptr_min_max_vals(env, insn,
2258 if (rc == -EACCES && env->allow_ptr_leaks) {
2259 /* unknown scalar += K */
2260 __mark_reg_unknown(dst_reg);
2261 return adjust_scalar_min_max_vals(
2262 env, insn, dst_reg, off_reg);
2268 /* Got here implies adding two SCALAR_VALUEs */
2269 if (WARN_ON_ONCE(ptr_reg)) {
2270 print_verifier_state(&env->cur_state);
2271 verbose("verifier internal error: unexpected ptr_reg\n");
2274 if (WARN_ON(!src_reg)) {
2275 print_verifier_state(&env->cur_state);
2276 verbose("verifier internal error: no src_reg\n");
2279 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
2282 /* check validity of 32-bit and 64-bit arithmetic operations */
2283 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
2285 struct bpf_reg_state *regs = env->cur_state.regs;
2286 u8 opcode = BPF_OP(insn->code);
2289 if (opcode == BPF_END || opcode == BPF_NEG) {
2290 if (opcode == BPF_NEG) {
2291 if (BPF_SRC(insn->code) != 0 ||
2292 insn->src_reg != BPF_REG_0 ||
2293 insn->off != 0 || insn->imm != 0) {
2294 verbose("BPF_NEG uses reserved fields\n");
2298 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
2299 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64)) {
2300 verbose("BPF_END uses reserved fields\n");
2305 /* check src operand */
2306 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2310 if (is_pointer_value(env, insn->dst_reg)) {
2311 verbose("R%d pointer arithmetic prohibited\n",
2316 /* check dest operand */
2317 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2321 } else if (opcode == BPF_MOV) {
2323 if (BPF_SRC(insn->code) == BPF_X) {
2324 if (insn->imm != 0 || insn->off != 0) {
2325 verbose("BPF_MOV uses reserved fields\n");
2329 /* check src operand */
2330 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2334 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
2335 verbose("BPF_MOV uses reserved fields\n");
2340 /* check dest operand */
2341 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2345 if (BPF_SRC(insn->code) == BPF_X) {
2346 if (BPF_CLASS(insn->code) == BPF_ALU64) {
2348 * copy register state to dest reg
2350 regs[insn->dst_reg] = regs[insn->src_reg];
2353 if (is_pointer_value(env, insn->src_reg)) {
2354 verbose("R%d partial copy of pointer\n",
2358 mark_reg_unknown(regs, insn->dst_reg);
2359 /* high 32 bits are known zero. */
2360 regs[insn->dst_reg].var_off = tnum_cast(
2361 regs[insn->dst_reg].var_off, 4);
2362 __update_reg_bounds(®s[insn->dst_reg]);
2366 * remember the value we stored into this reg
2368 regs[insn->dst_reg].type = SCALAR_VALUE;
2369 __mark_reg_known(regs + insn->dst_reg, insn->imm);
2372 } else if (opcode > BPF_END) {
2373 verbose("invalid BPF_ALU opcode %x\n", opcode);
2376 } else { /* all other ALU ops: and, sub, xor, add, ... */
2378 if (BPF_SRC(insn->code) == BPF_X) {
2379 if (insn->imm != 0 || insn->off != 0) {
2380 verbose("BPF_ALU uses reserved fields\n");
2383 /* check src1 operand */
2384 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2388 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
2389 verbose("BPF_ALU uses reserved fields\n");
2394 /* check src2 operand */
2395 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2399 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
2400 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
2401 verbose("div by zero\n");
2405 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
2406 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
2407 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
2409 if (insn->imm < 0 || insn->imm >= size) {
2410 verbose("invalid shift %d\n", insn->imm);
2415 /* check dest operand */
2416 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
2420 return adjust_reg_min_max_vals(env, insn);
2426 static void find_good_pkt_pointers(struct bpf_verifier_state *state,
2427 struct bpf_reg_state *dst_reg)
2429 struct bpf_reg_state *regs = state->regs, *reg;
2432 if (dst_reg->off < 0)
2433 /* This doesn't give us any range */
2436 if (dst_reg->umax_value > MAX_PACKET_OFF ||
2437 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
2438 /* Risk of overflow. For instance, ptr + (1<<63) may be less
2439 * than pkt_end, but that's because it's also less than pkt.
2443 /* LLVM can generate four kind of checks:
2449 * if (r2 > pkt_end) goto <handle exception>
2454 * if (r2 < pkt_end) goto <access okay>
2455 * <handle exception>
2458 * r2 == dst_reg, pkt_end == src_reg
2459 * r2=pkt(id=n,off=8,r=0)
2460 * r3=pkt(id=n,off=0,r=0)
2466 * if (pkt_end >= r2) goto <access okay>
2467 * <handle exception>
2471 * if (pkt_end <= r2) goto <handle exception>
2475 * pkt_end == dst_reg, r2 == src_reg
2476 * r2=pkt(id=n,off=8,r=0)
2477 * r3=pkt(id=n,off=0,r=0)
2479 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
2480 * so that range of bytes [r3, r3 + 8) is safe to access.
2483 /* If our ids match, then we must have the same max_value. And we
2484 * don't care about the other reg's fixed offset, since if it's too big
2485 * the range won't allow anything.
2486 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
2488 for (i = 0; i < MAX_BPF_REG; i++)
2489 if (regs[i].type == PTR_TO_PACKET && regs[i].id == dst_reg->id)
2490 /* keep the maximum range already checked */
2491 regs[i].range = max_t(u16, regs[i].range, dst_reg->off);
2493 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
2494 if (state->stack_slot_type[i] != STACK_SPILL)
2496 reg = &state->spilled_regs[i / BPF_REG_SIZE];
2497 if (reg->type == PTR_TO_PACKET && reg->id == dst_reg->id)
2498 reg->range = max_t(u16, reg->range, dst_reg->off);
2502 /* Adjusts the register min/max values in the case that the dst_reg is the
2503 * variable register that we are working on, and src_reg is a constant or we're
2504 * simply doing a BPF_K check.
2505 * In JEQ/JNE cases we also adjust the var_off values.
2507 static void reg_set_min_max(struct bpf_reg_state *true_reg,
2508 struct bpf_reg_state *false_reg, u64 val,
2511 /* If the dst_reg is a pointer, we can't learn anything about its
2512 * variable offset from the compare (unless src_reg were a pointer into
2513 * the same object, but we don't bother with that.
2514 * Since false_reg and true_reg have the same type by construction, we
2515 * only need to check one of them for pointerness.
2517 if (__is_pointer_value(false, false_reg))
2522 /* If this is false then we know nothing Jon Snow, but if it is
2523 * true then we know for sure.
2525 __mark_reg_known(true_reg, val);
2528 /* If this is true we know nothing Jon Snow, but if it is false
2529 * we know the value for sure;
2531 __mark_reg_known(false_reg, val);
2534 false_reg->umax_value = min(false_reg->umax_value, val);
2535 true_reg->umin_value = max(true_reg->umin_value, val + 1);
2538 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
2539 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
2542 false_reg->umin_value = max(false_reg->umin_value, val);
2543 true_reg->umax_value = min(true_reg->umax_value, val - 1);
2546 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
2547 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
2550 false_reg->umax_value = min(false_reg->umax_value, val - 1);
2551 true_reg->umin_value = max(true_reg->umin_value, val);
2554 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
2555 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
2558 false_reg->umin_value = max(false_reg->umin_value, val + 1);
2559 true_reg->umax_value = min(true_reg->umax_value, val);
2562 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
2563 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
2569 __reg_deduce_bounds(false_reg);
2570 __reg_deduce_bounds(true_reg);
2571 /* We might have learned some bits from the bounds. */
2572 __reg_bound_offset(false_reg);
2573 __reg_bound_offset(true_reg);
2574 /* Intersecting with the old var_off might have improved our bounds
2575 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2576 * then new var_off is (0; 0x7f...fc) which improves our umax.
2578 __update_reg_bounds(false_reg);
2579 __update_reg_bounds(true_reg);
2582 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
2585 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
2586 struct bpf_reg_state *false_reg, u64 val,
2589 if (__is_pointer_value(false, false_reg))
2594 /* If this is false then we know nothing Jon Snow, but if it is
2595 * true then we know for sure.
2597 __mark_reg_known(true_reg, val);
2600 /* If this is true we know nothing Jon Snow, but if it is false
2601 * we know the value for sure;
2603 __mark_reg_known(false_reg, val);
2606 true_reg->umax_value = min(true_reg->umax_value, val - 1);
2607 false_reg->umin_value = max(false_reg->umin_value, val);
2610 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
2611 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
2614 true_reg->umin_value = max(true_reg->umin_value, val + 1);
2615 false_reg->umax_value = min(false_reg->umax_value, val);
2618 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
2619 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
2622 true_reg->umax_value = min(true_reg->umax_value, val);
2623 false_reg->umin_value = max(false_reg->umin_value, val + 1);
2626 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
2627 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
2630 true_reg->umin_value = max(true_reg->umin_value, val);
2631 false_reg->umax_value = min(false_reg->umax_value, val - 1);
2634 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
2635 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
2641 __reg_deduce_bounds(false_reg);
2642 __reg_deduce_bounds(true_reg);
2643 /* We might have learned some bits from the bounds. */
2644 __reg_bound_offset(false_reg);
2645 __reg_bound_offset(true_reg);
2646 /* Intersecting with the old var_off might have improved our bounds
2647 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2648 * then new var_off is (0; 0x7f...fc) which improves our umax.
2650 __update_reg_bounds(false_reg);
2651 __update_reg_bounds(true_reg);
2654 /* Regs are known to be equal, so intersect their min/max/var_off */
2655 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
2656 struct bpf_reg_state *dst_reg)
2658 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
2659 dst_reg->umin_value);
2660 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
2661 dst_reg->umax_value);
2662 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
2663 dst_reg->smin_value);
2664 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
2665 dst_reg->smax_value);
2666 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
2668 /* We might have learned new bounds from the var_off. */
2669 __update_reg_bounds(src_reg);
2670 __update_reg_bounds(dst_reg);
2671 /* We might have learned something about the sign bit. */
2672 __reg_deduce_bounds(src_reg);
2673 __reg_deduce_bounds(dst_reg);
2674 /* We might have learned some bits from the bounds. */
2675 __reg_bound_offset(src_reg);
2676 __reg_bound_offset(dst_reg);
2677 /* Intersecting with the old var_off might have improved our bounds
2678 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2679 * then new var_off is (0; 0x7f...fc) which improves our umax.
2681 __update_reg_bounds(src_reg);
2682 __update_reg_bounds(dst_reg);
2685 static void reg_combine_min_max(struct bpf_reg_state *true_src,
2686 struct bpf_reg_state *true_dst,
2687 struct bpf_reg_state *false_src,
2688 struct bpf_reg_state *false_dst,
2693 __reg_combine_min_max(true_src, true_dst);
2696 __reg_combine_min_max(false_src, false_dst);
2701 static void mark_map_reg(struct bpf_reg_state *regs, u32 regno, u32 id,
2704 struct bpf_reg_state *reg = ®s[regno];
2706 if (reg->type == PTR_TO_MAP_VALUE_OR_NULL && reg->id == id) {
2707 /* Old offset (both fixed and variable parts) should
2708 * have been known-zero, because we don't allow pointer
2709 * arithmetic on pointers that might be NULL.
2711 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
2712 !tnum_equals_const(reg->var_off, 0) ||
2714 __mark_reg_known_zero(reg);
2718 reg->type = SCALAR_VALUE;
2719 } else if (reg->map_ptr->inner_map_meta) {
2720 reg->type = CONST_PTR_TO_MAP;
2721 reg->map_ptr = reg->map_ptr->inner_map_meta;
2723 reg->type = PTR_TO_MAP_VALUE;
2725 /* We don't need id from this point onwards anymore, thus we
2726 * should better reset it, so that state pruning has chances
2733 /* The logic is similar to find_good_pkt_pointers(), both could eventually
2734 * be folded together at some point.
2736 static void mark_map_regs(struct bpf_verifier_state *state, u32 regno,
2739 struct bpf_reg_state *regs = state->regs;
2740 u32 id = regs[regno].id;
2743 for (i = 0; i < MAX_BPF_REG; i++)
2744 mark_map_reg(regs, i, id, is_null);
2746 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
2747 if (state->stack_slot_type[i] != STACK_SPILL)
2749 mark_map_reg(state->spilled_regs, i / BPF_REG_SIZE, id, is_null);
2753 static int check_cond_jmp_op(struct bpf_verifier_env *env,
2754 struct bpf_insn *insn, int *insn_idx)
2756 struct bpf_verifier_state *other_branch, *this_branch = &env->cur_state;
2757 struct bpf_reg_state *regs = this_branch->regs, *dst_reg;
2758 u8 opcode = BPF_OP(insn->code);
2761 if (opcode > BPF_JSLE) {
2762 verbose("invalid BPF_JMP opcode %x\n", opcode);
2766 if (BPF_SRC(insn->code) == BPF_X) {
2767 if (insn->imm != 0) {
2768 verbose("BPF_JMP uses reserved fields\n");
2772 /* check src1 operand */
2773 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2777 if (is_pointer_value(env, insn->src_reg)) {
2778 verbose("R%d pointer comparison prohibited\n",
2783 if (insn->src_reg != BPF_REG_0) {
2784 verbose("BPF_JMP uses reserved fields\n");
2789 /* check src2 operand */
2790 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2794 dst_reg = ®s[insn->dst_reg];
2796 /* detect if R == 0 where R was initialized to zero earlier */
2797 if (BPF_SRC(insn->code) == BPF_K &&
2798 (opcode == BPF_JEQ || opcode == BPF_JNE) &&
2799 dst_reg->type == SCALAR_VALUE &&
2800 tnum_equals_const(dst_reg->var_off, insn->imm)) {
2801 if (opcode == BPF_JEQ) {
2802 /* if (imm == imm) goto pc+off;
2803 * only follow the goto, ignore fall-through
2805 *insn_idx += insn->off;
2808 /* if (imm != imm) goto pc+off;
2809 * only follow fall-through branch, since
2810 * that's where the program will go
2816 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx);
2820 /* detect if we are comparing against a constant value so we can adjust
2821 * our min/max values for our dst register.
2822 * this is only legit if both are scalars (or pointers to the same
2823 * object, I suppose, but we don't support that right now), because
2824 * otherwise the different base pointers mean the offsets aren't
2827 if (BPF_SRC(insn->code) == BPF_X) {
2828 if (dst_reg->type == SCALAR_VALUE &&
2829 regs[insn->src_reg].type == SCALAR_VALUE) {
2830 if (tnum_is_const(regs[insn->src_reg].var_off))
2831 reg_set_min_max(&other_branch->regs[insn->dst_reg],
2832 dst_reg, regs[insn->src_reg].var_off.value,
2834 else if (tnum_is_const(dst_reg->var_off))
2835 reg_set_min_max_inv(&other_branch->regs[insn->src_reg],
2836 ®s[insn->src_reg],
2837 dst_reg->var_off.value, opcode);
2838 else if (opcode == BPF_JEQ || opcode == BPF_JNE)
2839 /* Comparing for equality, we can combine knowledge */
2840 reg_combine_min_max(&other_branch->regs[insn->src_reg],
2841 &other_branch->regs[insn->dst_reg],
2842 ®s[insn->src_reg],
2843 ®s[insn->dst_reg], opcode);
2845 } else if (dst_reg->type == SCALAR_VALUE) {
2846 reg_set_min_max(&other_branch->regs[insn->dst_reg],
2847 dst_reg, insn->imm, opcode);
2850 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
2851 if (BPF_SRC(insn->code) == BPF_K &&
2852 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
2853 dst_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
2854 /* Mark all identical map registers in each branch as either
2855 * safe or unknown depending R == 0 or R != 0 conditional.
2857 mark_map_regs(this_branch, insn->dst_reg, opcode == BPF_JNE);
2858 mark_map_regs(other_branch, insn->dst_reg, opcode == BPF_JEQ);
2859 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGT &&
2860 dst_reg->type == PTR_TO_PACKET &&
2861 regs[insn->src_reg].type == PTR_TO_PACKET_END) {
2862 find_good_pkt_pointers(this_branch, dst_reg);
2863 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JLT &&
2864 dst_reg->type == PTR_TO_PACKET &&
2865 regs[insn->src_reg].type == PTR_TO_PACKET_END) {
2866 find_good_pkt_pointers(other_branch, dst_reg);
2867 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGE &&
2868 dst_reg->type == PTR_TO_PACKET_END &&
2869 regs[insn->src_reg].type == PTR_TO_PACKET) {
2870 find_good_pkt_pointers(other_branch, ®s[insn->src_reg]);
2871 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JLE &&
2872 dst_reg->type == PTR_TO_PACKET_END &&
2873 regs[insn->src_reg].type == PTR_TO_PACKET) {
2874 find_good_pkt_pointers(this_branch, ®s[insn->src_reg]);
2875 } else if (is_pointer_value(env, insn->dst_reg)) {
2876 verbose("R%d pointer comparison prohibited\n", insn->dst_reg);
2880 print_verifier_state(this_branch);
2884 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
2885 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
2887 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
2889 return (struct bpf_map *) (unsigned long) imm64;
2892 /* verify BPF_LD_IMM64 instruction */
2893 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
2895 struct bpf_reg_state *regs = env->cur_state.regs;
2898 if (BPF_SIZE(insn->code) != BPF_DW) {
2899 verbose("invalid BPF_LD_IMM insn\n");
2902 if (insn->off != 0) {
2903 verbose("BPF_LD_IMM64 uses reserved fields\n");
2907 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2911 if (insn->src_reg == 0) {
2912 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
2914 regs[insn->dst_reg].type = SCALAR_VALUE;
2915 __mark_reg_known(®s[insn->dst_reg], imm);
2919 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
2920 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
2922 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
2923 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
2927 static bool may_access_skb(enum bpf_prog_type type)
2930 case BPF_PROG_TYPE_SOCKET_FILTER:
2931 case BPF_PROG_TYPE_SCHED_CLS:
2932 case BPF_PROG_TYPE_SCHED_ACT:
2939 /* verify safety of LD_ABS|LD_IND instructions:
2940 * - they can only appear in the programs where ctx == skb
2941 * - since they are wrappers of function calls, they scratch R1-R5 registers,
2942 * preserve R6-R9, and store return value into R0
2945 * ctx == skb == R6 == CTX
2948 * SRC == any register
2949 * IMM == 32-bit immediate
2952 * R0 - 8/16/32-bit skb data converted to cpu endianness
2954 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
2956 struct bpf_reg_state *regs = env->cur_state.regs;
2957 u8 mode = BPF_MODE(insn->code);
2960 if (!may_access_skb(env->prog->type)) {
2961 verbose("BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
2965 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
2966 BPF_SIZE(insn->code) == BPF_DW ||
2967 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
2968 verbose("BPF_LD_[ABS|IND] uses reserved fields\n");
2972 /* check whether implicit source operand (register R6) is readable */
2973 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
2977 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
2978 verbose("at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
2982 if (mode == BPF_IND) {
2983 /* check explicit source operand */
2984 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2989 /* reset caller saved regs to unreadable */
2990 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2991 mark_reg_not_init(regs, caller_saved[i]);
2992 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2995 /* mark destination R0 register as readable, since it contains
2996 * the value fetched from the packet.
2997 * Already marked as written above.
2999 mark_reg_unknown(regs, BPF_REG_0);
3003 /* non-recursive DFS pseudo code
3004 * 1 procedure DFS-iterative(G,v):
3005 * 2 label v as discovered
3006 * 3 let S be a stack
3008 * 5 while S is not empty
3010 * 7 if t is what we're looking for:
3012 * 9 for all edges e in G.adjacentEdges(t) do
3013 * 10 if edge e is already labelled
3014 * 11 continue with the next edge
3015 * 12 w <- G.adjacentVertex(t,e)
3016 * 13 if vertex w is not discovered and not explored
3017 * 14 label e as tree-edge
3018 * 15 label w as discovered
3021 * 18 else if vertex w is discovered
3022 * 19 label e as back-edge
3024 * 21 // vertex w is explored
3025 * 22 label e as forward- or cross-edge
3026 * 23 label t as explored
3031 * 0x11 - discovered and fall-through edge labelled
3032 * 0x12 - discovered and fall-through and branch edges labelled
3043 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
3045 static int *insn_stack; /* stack of insns to process */
3046 static int cur_stack; /* current stack index */
3047 static int *insn_state;
3049 /* t, w, e - match pseudo-code above:
3050 * t - index of current instruction
3051 * w - next instruction
3054 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
3056 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
3059 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
3062 if (w < 0 || w >= env->prog->len) {
3063 verbose("jump out of range from insn %d to %d\n", t, w);
3068 /* mark branch target for state pruning */
3069 env->explored_states[w] = STATE_LIST_MARK;
3071 if (insn_state[w] == 0) {
3073 insn_state[t] = DISCOVERED | e;
3074 insn_state[w] = DISCOVERED;
3075 if (cur_stack >= env->prog->len)
3077 insn_stack[cur_stack++] = w;
3079 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
3080 verbose("back-edge from insn %d to %d\n", t, w);
3082 } else if (insn_state[w] == EXPLORED) {
3083 /* forward- or cross-edge */
3084 insn_state[t] = DISCOVERED | e;
3086 verbose("insn state internal bug\n");
3092 /* non-recursive depth-first-search to detect loops in BPF program
3093 * loop == back-edge in directed graph
3095 static int check_cfg(struct bpf_verifier_env *env)
3097 struct bpf_insn *insns = env->prog->insnsi;
3098 int insn_cnt = env->prog->len;
3102 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
3106 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
3112 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
3113 insn_stack[0] = 0; /* 0 is the first instruction */
3119 t = insn_stack[cur_stack - 1];
3121 if (BPF_CLASS(insns[t].code) == BPF_JMP) {
3122 u8 opcode = BPF_OP(insns[t].code);
3124 if (opcode == BPF_EXIT) {
3126 } else if (opcode == BPF_CALL) {
3127 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3132 if (t + 1 < insn_cnt)
3133 env->explored_states[t + 1] = STATE_LIST_MARK;
3134 } else if (opcode == BPF_JA) {
3135 if (BPF_SRC(insns[t].code) != BPF_K) {
3139 /* unconditional jump with single edge */
3140 ret = push_insn(t, t + insns[t].off + 1,
3146 /* tell verifier to check for equivalent states
3147 * after every call and jump
3149 if (t + 1 < insn_cnt)
3150 env->explored_states[t + 1] = STATE_LIST_MARK;
3152 /* conditional jump with two edges */
3153 env->explored_states[t] = STATE_LIST_MARK;
3154 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3160 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
3167 /* all other non-branch instructions with single
3170 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3178 insn_state[t] = EXPLORED;
3179 if (cur_stack-- <= 0) {
3180 verbose("pop stack internal bug\n");
3187 for (i = 0; i < insn_cnt; i++) {
3188 if (insn_state[i] != EXPLORED) {
3189 verbose("unreachable insn %d\n", i);
3194 ret = 0; /* cfg looks good */
3202 /* check %cur's range satisfies %old's */
3203 static bool range_within(struct bpf_reg_state *old,
3204 struct bpf_reg_state *cur)
3206 return old->umin_value <= cur->umin_value &&
3207 old->umax_value >= cur->umax_value &&
3208 old->smin_value <= cur->smin_value &&
3209 old->smax_value >= cur->smax_value;
3212 /* Maximum number of register states that can exist at once */
3213 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
3219 /* If in the old state two registers had the same id, then they need to have
3220 * the same id in the new state as well. But that id could be different from
3221 * the old state, so we need to track the mapping from old to new ids.
3222 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
3223 * regs with old id 5 must also have new id 9 for the new state to be safe. But
3224 * regs with a different old id could still have new id 9, we don't care about
3226 * So we look through our idmap to see if this old id has been seen before. If
3227 * so, we require the new id to match; otherwise, we add the id pair to the map.
3229 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
3233 for (i = 0; i < ID_MAP_SIZE; i++) {
3234 if (!idmap[i].old) {
3235 /* Reached an empty slot; haven't seen this id before */
3236 idmap[i].old = old_id;
3237 idmap[i].cur = cur_id;
3240 if (idmap[i].old == old_id)
3241 return idmap[i].cur == cur_id;
3243 /* We ran out of idmap slots, which should be impossible */
3248 /* Returns true if (rold safe implies rcur safe) */
3249 static bool regsafe(struct bpf_reg_state *rold,
3250 struct bpf_reg_state *rcur,
3251 bool varlen_map_access, struct idpair *idmap)
3253 if (!(rold->live & REG_LIVE_READ))
3254 /* explored state didn't use this */
3257 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, live)) == 0)
3260 if (rold->type == NOT_INIT)
3261 /* explored state can't have used this */
3263 if (rcur->type == NOT_INIT)
3265 switch (rold->type) {
3267 if (rcur->type == SCALAR_VALUE) {
3268 /* new val must satisfy old val knowledge */
3269 return range_within(rold, rcur) &&
3270 tnum_in(rold->var_off, rcur->var_off);
3272 /* if we knew anything about the old value, we're not
3273 * equal, because we can't know anything about the
3274 * scalar value of the pointer in the new value.
3276 return rold->umin_value == 0 &&
3277 rold->umax_value == U64_MAX &&
3278 rold->smin_value == S64_MIN &&
3279 rold->smax_value == S64_MAX &&
3280 tnum_is_unknown(rold->var_off);
3282 case PTR_TO_MAP_VALUE:
3283 if (varlen_map_access) {
3284 /* If the new min/max/var_off satisfy the old ones and
3285 * everything else matches, we are OK.
3286 * We don't care about the 'id' value, because nothing
3287 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
3289 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
3290 range_within(rold, rcur) &&
3291 tnum_in(rold->var_off, rcur->var_off);
3293 /* If the ranges/var_off were not the same, but
3294 * everything else was and we didn't do a variable
3295 * access into a map then we are a-ok.
3297 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0;
3299 case PTR_TO_MAP_VALUE_OR_NULL:
3300 /* a PTR_TO_MAP_VALUE could be safe to use as a
3301 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
3302 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
3303 * checked, doing so could have affected others with the same
3304 * id, and we can't check for that because we lost the id when
3305 * we converted to a PTR_TO_MAP_VALUE.
3307 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
3309 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
3311 /* Check our ids match any regs they're supposed to */
3312 return check_ids(rold->id, rcur->id, idmap);
3314 if (rcur->type != PTR_TO_PACKET)
3316 /* We must have at least as much range as the old ptr
3317 * did, so that any accesses which were safe before are
3318 * still safe. This is true even if old range < old off,
3319 * since someone could have accessed through (ptr - k), or
3320 * even done ptr -= k in a register, to get a safe access.
3322 if (rold->range > rcur->range)
3324 /* If the offsets don't match, we can't trust our alignment;
3325 * nor can we be sure that we won't fall out of range.
3327 if (rold->off != rcur->off)
3329 /* id relations must be preserved */
3330 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
3332 /* new val must satisfy old val knowledge */
3333 return range_within(rold, rcur) &&
3334 tnum_in(rold->var_off, rcur->var_off);
3336 case CONST_PTR_TO_MAP:
3338 case PTR_TO_PACKET_END:
3339 /* Only valid matches are exact, which memcmp() above
3340 * would have accepted
3343 /* Don't know what's going on, just say it's not safe */
3347 /* Shouldn't get here; if we do, say it's not safe */
3352 /* compare two verifier states
3354 * all states stored in state_list are known to be valid, since
3355 * verifier reached 'bpf_exit' instruction through them
3357 * this function is called when verifier exploring different branches of
3358 * execution popped from the state stack. If it sees an old state that has
3359 * more strict register state and more strict stack state then this execution
3360 * branch doesn't need to be explored further, since verifier already
3361 * concluded that more strict state leads to valid finish.
3363 * Therefore two states are equivalent if register state is more conservative
3364 * and explored stack state is more conservative than the current one.
3367 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
3368 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
3370 * In other words if current stack state (one being explored) has more
3371 * valid slots than old one that already passed validation, it means
3372 * the verifier can stop exploring and conclude that current state is valid too
3374 * Similarly with registers. If explored state has register type as invalid
3375 * whereas register type in current state is meaningful, it means that
3376 * the current state will reach 'bpf_exit' instruction safely
3378 static bool states_equal(struct bpf_verifier_env *env,
3379 struct bpf_verifier_state *old,
3380 struct bpf_verifier_state *cur)
3382 bool varlen_map_access = env->varlen_map_value_access;
3383 struct idpair *idmap;
3387 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
3388 /* If we failed to allocate the idmap, just say it's not safe */
3392 for (i = 0; i < MAX_BPF_REG; i++) {
3393 if (!regsafe(&old->regs[i], &cur->regs[i], varlen_map_access,
3398 for (i = 0; i < MAX_BPF_STACK; i++) {
3399 if (old->stack_slot_type[i] == STACK_INVALID)
3401 if (old->stack_slot_type[i] != cur->stack_slot_type[i])
3402 /* Ex: old explored (safe) state has STACK_SPILL in
3403 * this stack slot, but current has has STACK_MISC ->
3404 * this verifier states are not equivalent,
3405 * return false to continue verification of this path
3408 if (i % BPF_REG_SIZE)
3410 if (old->stack_slot_type[i] != STACK_SPILL)
3412 if (!regsafe(&old->spilled_regs[i / BPF_REG_SIZE],
3413 &cur->spilled_regs[i / BPF_REG_SIZE],
3414 varlen_map_access, idmap))
3415 /* when explored and current stack slot are both storing
3416 * spilled registers, check that stored pointers types
3417 * are the same as well.
3418 * Ex: explored safe path could have stored
3419 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
3420 * but current path has stored:
3421 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
3422 * such verifier states are not equivalent.
3423 * return false to continue verification of this path
3435 static bool do_propagate_liveness(const struct bpf_verifier_state *state,
3436 struct bpf_verifier_state *parent)
3438 bool touched = false; /* any changes made? */
3443 /* Propagate read liveness of registers... */
3444 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
3445 /* We don't need to worry about FP liveness because it's read-only */
3446 for (i = 0; i < BPF_REG_FP; i++) {
3447 if (parent->regs[i].live & REG_LIVE_READ)
3449 if (state->regs[i].live == REG_LIVE_READ) {
3450 parent->regs[i].live |= REG_LIVE_READ;
3454 /* ... and stack slots */
3455 for (i = 0; i < MAX_BPF_STACK / BPF_REG_SIZE; i++) {
3456 if (parent->stack_slot_type[i * BPF_REG_SIZE] != STACK_SPILL)
3458 if (state->stack_slot_type[i * BPF_REG_SIZE] != STACK_SPILL)
3460 if (parent->spilled_regs[i].live & REG_LIVE_READ)
3462 if (state->spilled_regs[i].live == REG_LIVE_READ) {
3463 parent->spilled_regs[i].live |= REG_LIVE_READ;
3470 static void propagate_liveness(const struct bpf_verifier_state *state,
3471 struct bpf_verifier_state *parent)
3473 while (do_propagate_liveness(state, parent)) {
3474 /* Something changed, so we need to feed those changes onward */
3476 parent = state->parent;
3480 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
3482 struct bpf_verifier_state_list *new_sl;
3483 struct bpf_verifier_state_list *sl;
3486 sl = env->explored_states[insn_idx];
3488 /* this 'insn_idx' instruction wasn't marked, so we will not
3489 * be doing state search here
3493 while (sl != STATE_LIST_MARK) {
3494 if (states_equal(env, &sl->state, &env->cur_state)) {
3495 /* reached equivalent register/stack state,
3497 * Registers read by the continuation are read by us.
3499 propagate_liveness(&sl->state, &env->cur_state);
3505 /* there were no equivalent states, remember current one.
3506 * technically the current state is not proven to be safe yet,
3507 * but it will either reach bpf_exit (which means it's safe) or
3508 * it will be rejected. Since there are no loops, we won't be
3509 * seeing this 'insn_idx' instruction again on the way to bpf_exit
3511 new_sl = kmalloc(sizeof(struct bpf_verifier_state_list), GFP_USER);
3515 /* add new state to the head of linked list */
3516 memcpy(&new_sl->state, &env->cur_state, sizeof(env->cur_state));
3517 new_sl->next = env->explored_states[insn_idx];
3518 env->explored_states[insn_idx] = new_sl;
3519 /* connect new state to parentage chain */
3520 env->cur_state.parent = &new_sl->state;
3521 /* clear liveness marks in current state */
3522 for (i = 0; i < BPF_REG_FP; i++)
3523 env->cur_state.regs[i].live = REG_LIVE_NONE;
3524 for (i = 0; i < MAX_BPF_STACK / BPF_REG_SIZE; i++)
3525 if (env->cur_state.stack_slot_type[i * BPF_REG_SIZE] == STACK_SPILL)
3526 env->cur_state.spilled_regs[i].live = REG_LIVE_NONE;
3530 static int ext_analyzer_insn_hook(struct bpf_verifier_env *env,
3531 int insn_idx, int prev_insn_idx)
3533 if (!env->analyzer_ops || !env->analyzer_ops->insn_hook)
3536 return env->analyzer_ops->insn_hook(env, insn_idx, prev_insn_idx);
3539 static int do_check(struct bpf_verifier_env *env)
3541 struct bpf_verifier_state *state = &env->cur_state;
3542 struct bpf_insn *insns = env->prog->insnsi;
3543 struct bpf_reg_state *regs = state->regs;
3544 int insn_cnt = env->prog->len;
3545 int insn_idx, prev_insn_idx = 0;
3546 int insn_processed = 0;
3547 bool do_print_state = false;
3549 init_reg_state(regs);
3550 state->parent = NULL;
3552 env->varlen_map_value_access = false;
3554 struct bpf_insn *insn;
3558 if (insn_idx >= insn_cnt) {
3559 verbose("invalid insn idx %d insn_cnt %d\n",
3560 insn_idx, insn_cnt);
3564 insn = &insns[insn_idx];
3565 class = BPF_CLASS(insn->code);
3567 if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
3568 verbose("BPF program is too large. Processed %d insn\n",
3573 err = is_state_visited(env, insn_idx);
3577 /* found equivalent state, can prune the search */
3580 verbose("\nfrom %d to %d: safe\n",
3581 prev_insn_idx, insn_idx);
3583 verbose("%d: safe\n", insn_idx);
3585 goto process_bpf_exit;
3591 if (log_level > 1 || (log_level && do_print_state)) {
3593 verbose("%d:", insn_idx);
3595 verbose("\nfrom %d to %d:",
3596 prev_insn_idx, insn_idx);
3597 print_verifier_state(&env->cur_state);
3598 do_print_state = false;
3602 verbose("%d: ", insn_idx);
3603 print_bpf_insn(env, insn);
3606 err = ext_analyzer_insn_hook(env, insn_idx, prev_insn_idx);
3610 if (class == BPF_ALU || class == BPF_ALU64) {
3611 err = check_alu_op(env, insn);
3615 } else if (class == BPF_LDX) {
3616 enum bpf_reg_type *prev_src_type, src_reg_type;
3618 /* check for reserved fields is already done */
3620 /* check src operand */
3621 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3625 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3629 src_reg_type = regs[insn->src_reg].type;
3631 /* check that memory (src_reg + off) is readable,
3632 * the state of dst_reg will be updated by this func
3634 err = check_mem_access(env, insn_idx, insn->src_reg, insn->off,
3635 BPF_SIZE(insn->code), BPF_READ,
3640 prev_src_type = &env->insn_aux_data[insn_idx].ptr_type;
3642 if (*prev_src_type == NOT_INIT) {
3644 * dst_reg = *(u32 *)(src_reg + off)
3645 * save type to validate intersecting paths
3647 *prev_src_type = src_reg_type;
3649 } else if (src_reg_type != *prev_src_type &&
3650 (src_reg_type == PTR_TO_CTX ||
3651 *prev_src_type == PTR_TO_CTX)) {
3652 /* ABuser program is trying to use the same insn
3653 * dst_reg = *(u32*) (src_reg + off)
3654 * with different pointer types:
3655 * src_reg == ctx in one branch and
3656 * src_reg == stack|map in some other branch.
3659 verbose("same insn cannot be used with different pointers\n");
3663 } else if (class == BPF_STX) {
3664 enum bpf_reg_type *prev_dst_type, dst_reg_type;
3666 if (BPF_MODE(insn->code) == BPF_XADD) {
3667 err = check_xadd(env, insn_idx, insn);
3674 /* check src1 operand */
3675 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3678 /* check src2 operand */
3679 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3683 dst_reg_type = regs[insn->dst_reg].type;
3685 /* check that memory (dst_reg + off) is writeable */
3686 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3687 BPF_SIZE(insn->code), BPF_WRITE,
3692 prev_dst_type = &env->insn_aux_data[insn_idx].ptr_type;
3694 if (*prev_dst_type == NOT_INIT) {
3695 *prev_dst_type = dst_reg_type;
3696 } else if (dst_reg_type != *prev_dst_type &&
3697 (dst_reg_type == PTR_TO_CTX ||
3698 *prev_dst_type == PTR_TO_CTX)) {
3699 verbose("same insn cannot be used with different pointers\n");
3703 } else if (class == BPF_ST) {
3704 if (BPF_MODE(insn->code) != BPF_MEM ||
3705 insn->src_reg != BPF_REG_0) {
3706 verbose("BPF_ST uses reserved fields\n");
3709 /* check src operand */
3710 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3714 /* check that memory (dst_reg + off) is writeable */
3715 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3716 BPF_SIZE(insn->code), BPF_WRITE,
3721 } else if (class == BPF_JMP) {
3722 u8 opcode = BPF_OP(insn->code);
3724 if (opcode == BPF_CALL) {
3725 if (BPF_SRC(insn->code) != BPF_K ||
3727 insn->src_reg != BPF_REG_0 ||
3728 insn->dst_reg != BPF_REG_0) {
3729 verbose("BPF_CALL uses reserved fields\n");
3733 err = check_call(env, insn->imm, insn_idx);
3737 } else if (opcode == BPF_JA) {
3738 if (BPF_SRC(insn->code) != BPF_K ||
3740 insn->src_reg != BPF_REG_0 ||
3741 insn->dst_reg != BPF_REG_0) {
3742 verbose("BPF_JA uses reserved fields\n");
3746 insn_idx += insn->off + 1;
3749 } else if (opcode == BPF_EXIT) {
3750 if (BPF_SRC(insn->code) != BPF_K ||
3752 insn->src_reg != BPF_REG_0 ||
3753 insn->dst_reg != BPF_REG_0) {
3754 verbose("BPF_EXIT uses reserved fields\n");
3758 /* eBPF calling convetion is such that R0 is used
3759 * to return the value from eBPF program.
3760 * Make sure that it's readable at this time
3761 * of bpf_exit, which means that program wrote
3762 * something into it earlier
3764 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
3768 if (is_pointer_value(env, BPF_REG_0)) {
3769 verbose("R0 leaks addr as return value\n");
3774 insn_idx = pop_stack(env, &prev_insn_idx);
3778 do_print_state = true;
3782 err = check_cond_jmp_op(env, insn, &insn_idx);
3786 } else if (class == BPF_LD) {
3787 u8 mode = BPF_MODE(insn->code);
3789 if (mode == BPF_ABS || mode == BPF_IND) {
3790 err = check_ld_abs(env, insn);
3794 } else if (mode == BPF_IMM) {
3795 err = check_ld_imm(env, insn);
3801 verbose("invalid BPF_LD mode\n");
3805 verbose("unknown insn class %d\n", class);
3812 verbose("processed %d insns, stack depth %d\n",
3813 insn_processed, env->prog->aux->stack_depth);
3817 static int check_map_prealloc(struct bpf_map *map)
3819 return (map->map_type != BPF_MAP_TYPE_HASH &&
3820 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
3821 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
3822 !(map->map_flags & BPF_F_NO_PREALLOC);
3825 static int check_map_prog_compatibility(struct bpf_map *map,
3826 struct bpf_prog *prog)
3829 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
3830 * preallocated hash maps, since doing memory allocation
3831 * in overflow_handler can crash depending on where nmi got
3834 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
3835 if (!check_map_prealloc(map)) {
3836 verbose("perf_event programs can only use preallocated hash map\n");
3839 if (map->inner_map_meta &&
3840 !check_map_prealloc(map->inner_map_meta)) {
3841 verbose("perf_event programs can only use preallocated inner hash map\n");
3848 /* look for pseudo eBPF instructions that access map FDs and
3849 * replace them with actual map pointers
3851 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
3853 struct bpf_insn *insn = env->prog->insnsi;
3854 int insn_cnt = env->prog->len;
3857 err = bpf_prog_calc_tag(env->prog);
3861 for (i = 0; i < insn_cnt; i++, insn++) {
3862 if (BPF_CLASS(insn->code) == BPF_LDX &&
3863 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
3864 verbose("BPF_LDX uses reserved fields\n");
3868 if (BPF_CLASS(insn->code) == BPF_STX &&
3869 ((BPF_MODE(insn->code) != BPF_MEM &&
3870 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
3871 verbose("BPF_STX uses reserved fields\n");
3875 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
3876 struct bpf_map *map;
3879 if (i == insn_cnt - 1 || insn[1].code != 0 ||
3880 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
3882 verbose("invalid bpf_ld_imm64 insn\n");
3886 if (insn->src_reg == 0)
3887 /* valid generic load 64-bit imm */
3890 if (insn->src_reg != BPF_PSEUDO_MAP_FD) {
3891 verbose("unrecognized bpf_ld_imm64 insn\n");
3895 f = fdget(insn->imm);
3896 map = __bpf_map_get(f);
3898 verbose("fd %d is not pointing to valid bpf_map\n",
3900 return PTR_ERR(map);
3903 err = check_map_prog_compatibility(map, env->prog);
3909 /* store map pointer inside BPF_LD_IMM64 instruction */
3910 insn[0].imm = (u32) (unsigned long) map;
3911 insn[1].imm = ((u64) (unsigned long) map) >> 32;
3913 /* check whether we recorded this map already */
3914 for (j = 0; j < env->used_map_cnt; j++)
3915 if (env->used_maps[j] == map) {
3920 if (env->used_map_cnt >= MAX_USED_MAPS) {
3925 /* hold the map. If the program is rejected by verifier,
3926 * the map will be released by release_maps() or it
3927 * will be used by the valid program until it's unloaded
3928 * and all maps are released in free_bpf_prog_info()
3930 map = bpf_map_inc(map, false);
3933 return PTR_ERR(map);
3935 env->used_maps[env->used_map_cnt++] = map;
3944 /* now all pseudo BPF_LD_IMM64 instructions load valid
3945 * 'struct bpf_map *' into a register instead of user map_fd.
3946 * These pointers will be used later by verifier to validate map access.
3951 /* drop refcnt of maps used by the rejected program */
3952 static void release_maps(struct bpf_verifier_env *env)
3956 for (i = 0; i < env->used_map_cnt; i++)
3957 bpf_map_put(env->used_maps[i]);
3960 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
3961 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
3963 struct bpf_insn *insn = env->prog->insnsi;
3964 int insn_cnt = env->prog->len;
3967 for (i = 0; i < insn_cnt; i++, insn++)
3968 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
3972 /* single env->prog->insni[off] instruction was replaced with the range
3973 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
3974 * [0, off) and [off, end) to new locations, so the patched range stays zero
3976 static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len,
3979 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
3983 new_data = vzalloc(sizeof(struct bpf_insn_aux_data) * prog_len);
3986 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
3987 memcpy(new_data + off + cnt - 1, old_data + off,
3988 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
3989 env->insn_aux_data = new_data;
3994 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
3995 const struct bpf_insn *patch, u32 len)
3997 struct bpf_prog *new_prog;
3999 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
4002 if (adjust_insn_aux_data(env, new_prog->len, off, len))
4007 /* convert load instructions that access fields of 'struct __sk_buff'
4008 * into sequence of instructions that access fields of 'struct sk_buff'
4010 static int convert_ctx_accesses(struct bpf_verifier_env *env)
4012 const struct bpf_verifier_ops *ops = env->prog->aux->ops;
4013 int i, cnt, size, ctx_field_size, delta = 0;
4014 const int insn_cnt = env->prog->len;
4015 struct bpf_insn insn_buf[16], *insn;
4016 struct bpf_prog *new_prog;
4017 enum bpf_access_type type;
4018 bool is_narrower_load;
4021 if (ops->gen_prologue) {
4022 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
4024 if (cnt >= ARRAY_SIZE(insn_buf)) {
4025 verbose("bpf verifier is misconfigured\n");
4028 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
4032 env->prog = new_prog;
4037 if (!ops->convert_ctx_access)
4040 insn = env->prog->insnsi + delta;
4042 for (i = 0; i < insn_cnt; i++, insn++) {
4043 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
4044 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
4045 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
4046 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
4048 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
4049 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
4050 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
4051 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
4056 if (env->insn_aux_data[i + delta].ptr_type != PTR_TO_CTX)
4059 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
4060 size = BPF_LDST_BYTES(insn);
4062 /* If the read access is a narrower load of the field,
4063 * convert to a 4/8-byte load, to minimum program type specific
4064 * convert_ctx_access changes. If conversion is successful,
4065 * we will apply proper mask to the result.
4067 is_narrower_load = size < ctx_field_size;
4068 if (is_narrower_load) {
4069 u32 off = insn->off;
4072 if (type == BPF_WRITE) {
4073 verbose("bpf verifier narrow ctx access misconfigured\n");
4078 if (ctx_field_size == 4)
4080 else if (ctx_field_size == 8)
4083 insn->off = off & ~(ctx_field_size - 1);
4084 insn->code = BPF_LDX | BPF_MEM | size_code;
4088 cnt = ops->convert_ctx_access(type, insn, insn_buf, env->prog,
4090 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
4091 (ctx_field_size && !target_size)) {
4092 verbose("bpf verifier is misconfigured\n");
4096 if (is_narrower_load && size < target_size) {
4097 if (ctx_field_size <= 4)
4098 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
4099 (1 << size * 8) - 1);
4101 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
4102 (1 << size * 8) - 1);
4105 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
4111 /* keep walking new program and skip insns we just inserted */
4112 env->prog = new_prog;
4113 insn = new_prog->insnsi + i + delta;
4119 /* fixup insn->imm field of bpf_call instructions
4120 * and inline eligible helpers as explicit sequence of BPF instructions
4122 * this function is called after eBPF program passed verification
4124 static int fixup_bpf_calls(struct bpf_verifier_env *env)
4126 struct bpf_prog *prog = env->prog;
4127 struct bpf_insn *insn = prog->insnsi;
4128 const struct bpf_func_proto *fn;
4129 const int insn_cnt = prog->len;
4130 struct bpf_insn insn_buf[16];
4131 struct bpf_prog *new_prog;
4132 struct bpf_map *map_ptr;
4133 int i, cnt, delta = 0;
4135 for (i = 0; i < insn_cnt; i++, insn++) {
4136 if (insn->code != (BPF_JMP | BPF_CALL))
4139 if (insn->imm == BPF_FUNC_get_route_realm)
4140 prog->dst_needed = 1;
4141 if (insn->imm == BPF_FUNC_get_prandom_u32)
4142 bpf_user_rnd_init_once();
4143 if (insn->imm == BPF_FUNC_tail_call) {
4144 /* If we tail call into other programs, we
4145 * cannot make any assumptions since they can
4146 * be replaced dynamically during runtime in
4147 * the program array.
4149 prog->cb_access = 1;
4150 env->prog->aux->stack_depth = MAX_BPF_STACK;
4152 /* mark bpf_tail_call as different opcode to avoid
4153 * conditional branch in the interpeter for every normal
4154 * call and to prevent accidental JITing by JIT compiler
4155 * that doesn't support bpf_tail_call yet
4158 insn->code = BPF_JMP | BPF_TAIL_CALL;
4162 if (ebpf_jit_enabled() && insn->imm == BPF_FUNC_map_lookup_elem) {
4163 map_ptr = env->insn_aux_data[i + delta].map_ptr;
4164 if (map_ptr == BPF_MAP_PTR_POISON ||
4165 !map_ptr->ops->map_gen_lookup)
4166 goto patch_call_imm;
4168 cnt = map_ptr->ops->map_gen_lookup(map_ptr, insn_buf);
4169 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
4170 verbose("bpf verifier is misconfigured\n");
4174 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
4181 /* keep walking new program and skip insns we just inserted */
4182 env->prog = prog = new_prog;
4183 insn = new_prog->insnsi + i + delta;
4188 fn = prog->aux->ops->get_func_proto(insn->imm);
4189 /* all functions that have prototype and verifier allowed
4190 * programs to call them, must be real in-kernel functions
4193 verbose("kernel subsystem misconfigured func %s#%d\n",
4194 func_id_name(insn->imm), insn->imm);
4197 insn->imm = fn->func - __bpf_call_base;
4203 static void free_states(struct bpf_verifier_env *env)
4205 struct bpf_verifier_state_list *sl, *sln;
4208 if (!env->explored_states)
4211 for (i = 0; i < env->prog->len; i++) {
4212 sl = env->explored_states[i];
4215 while (sl != STATE_LIST_MARK) {
4222 kfree(env->explored_states);
4225 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr)
4227 char __user *log_ubuf = NULL;
4228 struct bpf_verifier_env *env;
4231 /* 'struct bpf_verifier_env' can be global, but since it's not small,
4232 * allocate/free it every time bpf_check() is called
4234 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
4238 env->insn_aux_data = vzalloc(sizeof(struct bpf_insn_aux_data) *
4241 if (!env->insn_aux_data)
4245 /* grab the mutex to protect few globals used by verifier */
4246 mutex_lock(&bpf_verifier_lock);
4248 if (attr->log_level || attr->log_buf || attr->log_size) {
4249 /* user requested verbose verifier output
4250 * and supplied buffer to store the verification trace
4252 log_level = attr->log_level;
4253 log_ubuf = (char __user *) (unsigned long) attr->log_buf;
4254 log_size = attr->log_size;
4258 /* log_* values have to be sane */
4259 if (log_size < 128 || log_size > UINT_MAX >> 8 ||
4260 log_level == 0 || log_ubuf == NULL)
4264 log_buf = vmalloc(log_size);
4271 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
4272 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
4273 env->strict_alignment = true;
4275 ret = replace_map_fd_with_map_ptr(env);
4277 goto skip_full_check;
4279 env->explored_states = kcalloc(env->prog->len,
4280 sizeof(struct bpf_verifier_state_list *),
4283 if (!env->explored_states)
4284 goto skip_full_check;
4286 ret = check_cfg(env);
4288 goto skip_full_check;
4290 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
4292 ret = do_check(env);
4295 while (pop_stack(env, NULL) >= 0);
4299 /* program is valid, convert *(u32*)(ctx + off) accesses */
4300 ret = convert_ctx_accesses(env);
4303 ret = fixup_bpf_calls(env);
4305 if (log_level && log_len >= log_size - 1) {
4306 BUG_ON(log_len >= log_size);
4307 /* verifier log exceeded user supplied buffer */
4309 /* fall through to return what was recorded */
4312 /* copy verifier log back to user space including trailing zero */
4313 if (log_level && copy_to_user(log_ubuf, log_buf, log_len + 1) != 0) {
4318 if (ret == 0 && env->used_map_cnt) {
4319 /* if program passed verifier, update used_maps in bpf_prog_info */
4320 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
4321 sizeof(env->used_maps[0]),
4324 if (!env->prog->aux->used_maps) {
4329 memcpy(env->prog->aux->used_maps, env->used_maps,
4330 sizeof(env->used_maps[0]) * env->used_map_cnt);
4331 env->prog->aux->used_map_cnt = env->used_map_cnt;
4333 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
4334 * bpf_ld_imm64 instructions
4336 convert_pseudo_ld_imm64(env);
4342 if (!env->prog->aux->used_maps)
4343 /* if we didn't copy map pointers into bpf_prog_info, release
4344 * them now. Otherwise free_bpf_prog_info() will release them.
4349 mutex_unlock(&bpf_verifier_lock);
4350 vfree(env->insn_aux_data);
4356 int bpf_analyzer(struct bpf_prog *prog, const struct bpf_ext_analyzer_ops *ops,
4359 struct bpf_verifier_env *env;
4362 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
4366 env->insn_aux_data = vzalloc(sizeof(struct bpf_insn_aux_data) *
4369 if (!env->insn_aux_data)
4372 env->analyzer_ops = ops;
4373 env->analyzer_priv = priv;
4375 /* grab the mutex to protect few globals used by verifier */
4376 mutex_lock(&bpf_verifier_lock);
4380 env->strict_alignment = false;
4381 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
4382 env->strict_alignment = true;
4384 env->explored_states = kcalloc(env->prog->len,
4385 sizeof(struct bpf_verifier_state_list *),
4388 if (!env->explored_states)
4389 goto skip_full_check;
4391 ret = check_cfg(env);
4393 goto skip_full_check;
4395 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
4397 ret = do_check(env);
4400 while (pop_stack(env, NULL) >= 0);
4403 mutex_unlock(&bpf_verifier_lock);
4404 vfree(env->insn_aux_data);
4409 EXPORT_SYMBOL_GPL(bpf_analyzer);