1 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
2 * Copyright (c) 2016 Facebook
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of version 2 of the GNU General Public
6 * License as published by the Free Software Foundation.
8 * This program is distributed in the hope that it will be useful, but
9 * WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
13 #include <linux/kernel.h>
14 #include <linux/types.h>
15 #include <linux/slab.h>
16 #include <linux/bpf.h>
17 #include <linux/bpf_verifier.h>
18 #include <linux/filter.h>
19 #include <net/netlink.h>
20 #include <linux/file.h>
21 #include <linux/vmalloc.h>
22 #include <linux/stringify.h>
26 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
27 #define BPF_PROG_TYPE(_id, _name) \
28 [_id] = & _name ## _verifier_ops,
29 #define BPF_MAP_TYPE(_id, _ops)
30 #include <linux/bpf_types.h>
35 /* bpf_check() is a static code analyzer that walks eBPF program
36 * instruction by instruction and updates register/stack state.
37 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
39 * The first pass is depth-first-search to check that the program is a DAG.
40 * It rejects the following programs:
41 * - larger than BPF_MAXINSNS insns
42 * - if loop is present (detected via back-edge)
43 * - unreachable insns exist (shouldn't be a forest. program = one function)
44 * - out of bounds or malformed jumps
45 * The second pass is all possible path descent from the 1st insn.
46 * Since it's analyzing all pathes through the program, the length of the
47 * analysis is limited to 64k insn, which may be hit even if total number of
48 * insn is less then 4K, but there are too many branches that change stack/regs.
49 * Number of 'branches to be analyzed' is limited to 1k
51 * On entry to each instruction, each register has a type, and the instruction
52 * changes the types of the registers depending on instruction semantics.
53 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
56 * All registers are 64-bit.
57 * R0 - return register
58 * R1-R5 argument passing registers
59 * R6-R9 callee saved registers
60 * R10 - frame pointer read-only
62 * At the start of BPF program the register R1 contains a pointer to bpf_context
63 * and has type PTR_TO_CTX.
65 * Verifier tracks arithmetic operations on pointers in case:
66 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
67 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
68 * 1st insn copies R10 (which has FRAME_PTR) type into R1
69 * and 2nd arithmetic instruction is pattern matched to recognize
70 * that it wants to construct a pointer to some element within stack.
71 * So after 2nd insn, the register R1 has type PTR_TO_STACK
72 * (and -20 constant is saved for further stack bounds checking).
73 * Meaning that this reg is a pointer to stack plus known immediate constant.
75 * Most of the time the registers have SCALAR_VALUE type, which
76 * means the register has some value, but it's not a valid pointer.
77 * (like pointer plus pointer becomes SCALAR_VALUE type)
79 * When verifier sees load or store instructions the type of base register
80 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK. These are three pointer
81 * types recognized by check_mem_access() function.
83 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
84 * and the range of [ptr, ptr + map's value_size) is accessible.
86 * registers used to pass values to function calls are checked against
87 * function argument constraints.
89 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
90 * It means that the register type passed to this function must be
91 * PTR_TO_STACK and it will be used inside the function as
92 * 'pointer to map element key'
94 * For example the argument constraints for bpf_map_lookup_elem():
95 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
96 * .arg1_type = ARG_CONST_MAP_PTR,
97 * .arg2_type = ARG_PTR_TO_MAP_KEY,
99 * ret_type says that this function returns 'pointer to map elem value or null'
100 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
101 * 2nd argument should be a pointer to stack, which will be used inside
102 * the helper function as a pointer to map element key.
104 * On the kernel side the helper function looks like:
105 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
107 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
108 * void *key = (void *) (unsigned long) r2;
111 * here kernel can access 'key' and 'map' pointers safely, knowing that
112 * [key, key + map->key_size) bytes are valid and were initialized on
113 * the stack of eBPF program.
116 * Corresponding eBPF program may look like:
117 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
118 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
119 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
120 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
121 * here verifier looks at prototype of map_lookup_elem() and sees:
122 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
123 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
125 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
126 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
127 * and were initialized prior to this call.
128 * If it's ok, then verifier allows this BPF_CALL insn and looks at
129 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
130 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
131 * returns ether pointer to map value or NULL.
133 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
134 * insn, the register holding that pointer in the true branch changes state to
135 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
136 * branch. See check_cond_jmp_op().
138 * After the call R0 is set to return type of the function and registers R1-R5
139 * are set to NOT_INIT to indicate that they are no longer readable.
142 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
143 struct bpf_verifier_stack_elem {
144 /* verifer state is 'st'
145 * before processing instruction 'insn_idx'
146 * and after processing instruction 'prev_insn_idx'
148 struct bpf_verifier_state st;
151 struct bpf_verifier_stack_elem *next;
154 #define BPF_COMPLEXITY_LIMIT_INSNS 131072
155 #define BPF_COMPLEXITY_LIMIT_STACK 1024
157 #define BPF_MAP_PTR_POISON ((void *)0xeB9F + POISON_POINTER_DELTA)
159 struct bpf_call_arg_meta {
160 struct bpf_map *map_ptr;
167 static DEFINE_MUTEX(bpf_verifier_lock);
169 /* log_level controls verbosity level of eBPF verifier.
170 * verbose() is used to dump the verification trace to the log, so the user
171 * can figure out what's wrong with the program
173 static __printf(2, 3) void verbose(struct bpf_verifier_env *env,
174 const char *fmt, ...)
176 struct bpf_verifer_log *log = &env->log;
180 if (!log->level || !log->ubuf || bpf_verifier_log_full(log))
184 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
187 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
188 "verifier log line truncated - local buffer too short\n");
190 n = min(log->len_total - log->len_used - 1, n);
193 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
199 static bool type_is_pkt_pointer(enum bpf_reg_type type)
201 return type == PTR_TO_PACKET ||
202 type == PTR_TO_PACKET_META;
205 /* string representation of 'enum bpf_reg_type' */
206 static const char * const reg_type_str[] = {
208 [SCALAR_VALUE] = "inv",
209 [PTR_TO_CTX] = "ctx",
210 [CONST_PTR_TO_MAP] = "map_ptr",
211 [PTR_TO_MAP_VALUE] = "map_value",
212 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
213 [PTR_TO_STACK] = "fp",
214 [PTR_TO_PACKET] = "pkt",
215 [PTR_TO_PACKET_META] = "pkt_meta",
216 [PTR_TO_PACKET_END] = "pkt_end",
219 static void print_verifier_state(struct bpf_verifier_env *env,
220 struct bpf_verifier_state *state)
222 struct bpf_reg_state *reg;
226 for (i = 0; i < MAX_BPF_REG; i++) {
227 reg = &state->regs[i];
231 verbose(env, " R%d=%s", i, reg_type_str[t]);
232 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
233 tnum_is_const(reg->var_off)) {
234 /* reg->off should be 0 for SCALAR_VALUE */
235 verbose(env, "%lld", reg->var_off.value + reg->off);
237 verbose(env, "(id=%d", reg->id);
238 if (t != SCALAR_VALUE)
239 verbose(env, ",off=%d", reg->off);
240 if (type_is_pkt_pointer(t))
241 verbose(env, ",r=%d", reg->range);
242 else if (t == CONST_PTR_TO_MAP ||
243 t == PTR_TO_MAP_VALUE ||
244 t == PTR_TO_MAP_VALUE_OR_NULL)
245 verbose(env, ",ks=%d,vs=%d",
246 reg->map_ptr->key_size,
247 reg->map_ptr->value_size);
248 if (tnum_is_const(reg->var_off)) {
249 /* Typically an immediate SCALAR_VALUE, but
250 * could be a pointer whose offset is too big
253 verbose(env, ",imm=%llx", reg->var_off.value);
255 if (reg->smin_value != reg->umin_value &&
256 reg->smin_value != S64_MIN)
257 verbose(env, ",smin_value=%lld",
258 (long long)reg->smin_value);
259 if (reg->smax_value != reg->umax_value &&
260 reg->smax_value != S64_MAX)
261 verbose(env, ",smax_value=%lld",
262 (long long)reg->smax_value);
263 if (reg->umin_value != 0)
264 verbose(env, ",umin_value=%llu",
265 (unsigned long long)reg->umin_value);
266 if (reg->umax_value != U64_MAX)
267 verbose(env, ",umax_value=%llu",
268 (unsigned long long)reg->umax_value);
269 if (!tnum_is_unknown(reg->var_off)) {
272 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
273 verbose(env, ",var_off=%s", tn_buf);
279 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
280 if (state->stack[i].slot_type[0] == STACK_SPILL)
281 verbose(env, " fp%d=%s",
282 -MAX_BPF_STACK + i * BPF_REG_SIZE,
283 reg_type_str[state->stack[i].spilled_ptr.type]);
288 static int copy_stack_state(struct bpf_verifier_state *dst,
289 const struct bpf_verifier_state *src)
293 if (WARN_ON_ONCE(dst->allocated_stack < src->allocated_stack)) {
294 /* internal bug, make state invalid to reject the program */
295 memset(dst, 0, sizeof(*dst));
298 memcpy(dst->stack, src->stack,
299 sizeof(*src->stack) * (src->allocated_stack / BPF_REG_SIZE));
303 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
304 * make it consume minimal amount of memory. check_stack_write() access from
305 * the program calls into realloc_verifier_state() to grow the stack size.
306 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
307 * which this function copies over. It points to previous bpf_verifier_state
308 * which is never reallocated
310 static int realloc_verifier_state(struct bpf_verifier_state *state, int size,
313 u32 old_size = state->allocated_stack;
314 struct bpf_stack_state *new_stack;
315 int slot = size / BPF_REG_SIZE;
317 if (size <= old_size || !size) {
320 state->allocated_stack = slot * BPF_REG_SIZE;
321 if (!size && old_size) {
327 new_stack = kmalloc_array(slot, sizeof(struct bpf_stack_state),
333 memcpy(new_stack, state->stack,
334 sizeof(*new_stack) * (old_size / BPF_REG_SIZE));
335 memset(new_stack + old_size / BPF_REG_SIZE, 0,
336 sizeof(*new_stack) * (size - old_size) / BPF_REG_SIZE);
338 state->allocated_stack = slot * BPF_REG_SIZE;
340 state->stack = new_stack;
344 static void free_verifier_state(struct bpf_verifier_state *state,
352 /* copy verifier state from src to dst growing dst stack space
353 * when necessary to accommodate larger src stack
355 static int copy_verifier_state(struct bpf_verifier_state *dst,
356 const struct bpf_verifier_state *src)
360 err = realloc_verifier_state(dst, src->allocated_stack, false);
363 memcpy(dst, src, offsetof(struct bpf_verifier_state, allocated_stack));
364 return copy_stack_state(dst, src);
367 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
370 struct bpf_verifier_state *cur = env->cur_state;
371 struct bpf_verifier_stack_elem *elem, *head = env->head;
374 if (env->head == NULL)
378 err = copy_verifier_state(cur, &head->st);
383 *insn_idx = head->insn_idx;
385 *prev_insn_idx = head->prev_insn_idx;
387 free_verifier_state(&head->st, false);
394 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
395 int insn_idx, int prev_insn_idx)
397 struct bpf_verifier_state *cur = env->cur_state;
398 struct bpf_verifier_stack_elem *elem;
401 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
405 elem->insn_idx = insn_idx;
406 elem->prev_insn_idx = prev_insn_idx;
407 elem->next = env->head;
410 err = copy_verifier_state(&elem->st, cur);
413 if (env->stack_size > BPF_COMPLEXITY_LIMIT_STACK) {
414 verbose(env, "BPF program is too complex\n");
419 /* pop all elements and return */
420 while (!pop_stack(env, NULL, NULL));
424 #define CALLER_SAVED_REGS 6
425 static const int caller_saved[CALLER_SAVED_REGS] = {
426 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
429 static void __mark_reg_not_init(struct bpf_reg_state *reg);
431 /* Mark the unknown part of a register (variable offset or scalar value) as
432 * known to have the value @imm.
434 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
437 reg->var_off = tnum_const(imm);
438 reg->smin_value = (s64)imm;
439 reg->smax_value = (s64)imm;
440 reg->umin_value = imm;
441 reg->umax_value = imm;
444 /* Mark the 'variable offset' part of a register as zero. This should be
445 * used only on registers holding a pointer type.
447 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
449 __mark_reg_known(reg, 0);
452 static void mark_reg_known_zero(struct bpf_verifier_env *env,
453 struct bpf_reg_state *regs, u32 regno)
455 if (WARN_ON(regno >= MAX_BPF_REG)) {
456 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
457 /* Something bad happened, let's kill all regs */
458 for (regno = 0; regno < MAX_BPF_REG; regno++)
459 __mark_reg_not_init(regs + regno);
462 __mark_reg_known_zero(regs + regno);
465 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
467 return type_is_pkt_pointer(reg->type);
470 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
472 return reg_is_pkt_pointer(reg) ||
473 reg->type == PTR_TO_PACKET_END;
476 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
477 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
478 enum bpf_reg_type which)
480 /* The register can already have a range from prior markings.
481 * This is fine as long as it hasn't been advanced from its
484 return reg->type == which &&
487 tnum_equals_const(reg->var_off, 0);
490 /* Attempts to improve min/max values based on var_off information */
491 static void __update_reg_bounds(struct bpf_reg_state *reg)
493 /* min signed is max(sign bit) | min(other bits) */
494 reg->smin_value = max_t(s64, reg->smin_value,
495 reg->var_off.value | (reg->var_off.mask & S64_MIN));
496 /* max signed is min(sign bit) | max(other bits) */
497 reg->smax_value = min_t(s64, reg->smax_value,
498 reg->var_off.value | (reg->var_off.mask & S64_MAX));
499 reg->umin_value = max(reg->umin_value, reg->var_off.value);
500 reg->umax_value = min(reg->umax_value,
501 reg->var_off.value | reg->var_off.mask);
504 /* Uses signed min/max values to inform unsigned, and vice-versa */
505 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
507 /* Learn sign from signed bounds.
508 * If we cannot cross the sign boundary, then signed and unsigned bounds
509 * are the same, so combine. This works even in the negative case, e.g.
510 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
512 if (reg->smin_value >= 0 || reg->smax_value < 0) {
513 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
515 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
519 /* Learn sign from unsigned bounds. Signed bounds cross the sign
520 * boundary, so we must be careful.
522 if ((s64)reg->umax_value >= 0) {
523 /* Positive. We can't learn anything from the smin, but smax
524 * is positive, hence safe.
526 reg->smin_value = reg->umin_value;
527 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
529 } else if ((s64)reg->umin_value < 0) {
530 /* Negative. We can't learn anything from the smax, but smin
531 * is negative, hence safe.
533 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
535 reg->smax_value = reg->umax_value;
539 /* Attempts to improve var_off based on unsigned min/max information */
540 static void __reg_bound_offset(struct bpf_reg_state *reg)
542 reg->var_off = tnum_intersect(reg->var_off,
543 tnum_range(reg->umin_value,
547 /* Reset the min/max bounds of a register */
548 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
550 reg->smin_value = S64_MIN;
551 reg->smax_value = S64_MAX;
553 reg->umax_value = U64_MAX;
556 /* Mark a register as having a completely unknown (scalar) value. */
557 static void __mark_reg_unknown(struct bpf_reg_state *reg)
559 reg->type = SCALAR_VALUE;
562 reg->var_off = tnum_unknown;
563 __mark_reg_unbounded(reg);
566 static void mark_reg_unknown(struct bpf_verifier_env *env,
567 struct bpf_reg_state *regs, u32 regno)
569 if (WARN_ON(regno >= MAX_BPF_REG)) {
570 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
571 /* Something bad happened, let's kill all regs */
572 for (regno = 0; regno < MAX_BPF_REG; regno++)
573 __mark_reg_not_init(regs + regno);
576 __mark_reg_unknown(regs + regno);
579 static void __mark_reg_not_init(struct bpf_reg_state *reg)
581 __mark_reg_unknown(reg);
582 reg->type = NOT_INIT;
585 static void mark_reg_not_init(struct bpf_verifier_env *env,
586 struct bpf_reg_state *regs, u32 regno)
588 if (WARN_ON(regno >= MAX_BPF_REG)) {
589 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
590 /* Something bad happened, let's kill all regs */
591 for (regno = 0; regno < MAX_BPF_REG; regno++)
592 __mark_reg_not_init(regs + regno);
595 __mark_reg_not_init(regs + regno);
598 static void init_reg_state(struct bpf_verifier_env *env,
599 struct bpf_reg_state *regs)
603 for (i = 0; i < MAX_BPF_REG; i++) {
604 mark_reg_not_init(env, regs, i);
605 regs[i].live = REG_LIVE_NONE;
609 regs[BPF_REG_FP].type = PTR_TO_STACK;
610 mark_reg_known_zero(env, regs, BPF_REG_FP);
612 /* 1st arg to a function */
613 regs[BPF_REG_1].type = PTR_TO_CTX;
614 mark_reg_known_zero(env, regs, BPF_REG_1);
618 SRC_OP, /* register is used as source operand */
619 DST_OP, /* register is used as destination operand */
620 DST_OP_NO_MARK /* same as above, check only, don't mark */
623 static void mark_reg_read(const struct bpf_verifier_state *state, u32 regno)
625 struct bpf_verifier_state *parent = state->parent;
627 if (regno == BPF_REG_FP)
628 /* We don't need to worry about FP liveness because it's read-only */
632 /* if read wasn't screened by an earlier write ... */
633 if (state->regs[regno].live & REG_LIVE_WRITTEN)
635 /* ... then we depend on parent's value */
636 parent->regs[regno].live |= REG_LIVE_READ;
638 parent = state->parent;
642 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
645 struct bpf_reg_state *regs = env->cur_state->regs;
647 if (regno >= MAX_BPF_REG) {
648 verbose(env, "R%d is invalid\n", regno);
653 /* check whether register used as source operand can be read */
654 if (regs[regno].type == NOT_INIT) {
655 verbose(env, "R%d !read_ok\n", regno);
658 mark_reg_read(env->cur_state, regno);
660 /* check whether register used as dest operand can be written to */
661 if (regno == BPF_REG_FP) {
662 verbose(env, "frame pointer is read only\n");
665 regs[regno].live |= REG_LIVE_WRITTEN;
667 mark_reg_unknown(env, regs, regno);
672 static bool is_spillable_regtype(enum bpf_reg_type type)
675 case PTR_TO_MAP_VALUE:
676 case PTR_TO_MAP_VALUE_OR_NULL:
680 case PTR_TO_PACKET_META:
681 case PTR_TO_PACKET_END:
682 case CONST_PTR_TO_MAP:
689 /* check_stack_read/write functions track spill/fill of registers,
690 * stack boundary and alignment are checked in check_mem_access()
692 static int check_stack_write(struct bpf_verifier_env *env,
693 struct bpf_verifier_state *state, int off,
694 int size, int value_regno)
696 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
698 err = realloc_verifier_state(state, round_up(slot + 1, BPF_REG_SIZE),
702 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
703 * so it's aligned access and [off, off + size) are within stack limits
705 if (!env->allow_ptr_leaks &&
706 state->stack[spi].slot_type[0] == STACK_SPILL &&
707 size != BPF_REG_SIZE) {
708 verbose(env, "attempt to corrupt spilled pointer on stack\n");
712 if (value_regno >= 0 &&
713 is_spillable_regtype(state->regs[value_regno].type)) {
715 /* register containing pointer is being spilled into stack */
716 if (size != BPF_REG_SIZE) {
717 verbose(env, "invalid size of register spill\n");
721 /* save register state */
722 state->stack[spi].spilled_ptr = state->regs[value_regno];
723 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
725 for (i = 0; i < BPF_REG_SIZE; i++)
726 state->stack[spi].slot_type[i] = STACK_SPILL;
728 /* regular write of data into stack */
729 state->stack[spi].spilled_ptr = (struct bpf_reg_state) {};
731 for (i = 0; i < size; i++)
732 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
738 static void mark_stack_slot_read(const struct bpf_verifier_state *state, int slot)
740 struct bpf_verifier_state *parent = state->parent;
743 /* if read wasn't screened by an earlier write ... */
744 if (state->stack[slot].spilled_ptr.live & REG_LIVE_WRITTEN)
746 /* ... then we depend on parent's value */
747 parent->stack[slot].spilled_ptr.live |= REG_LIVE_READ;
749 parent = state->parent;
753 static int check_stack_read(struct bpf_verifier_env *env,
754 struct bpf_verifier_state *state, int off, int size,
757 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
760 if (state->allocated_stack <= slot) {
761 verbose(env, "invalid read from stack off %d+0 size %d\n",
765 stype = state->stack[spi].slot_type;
767 if (stype[0] == STACK_SPILL) {
768 if (size != BPF_REG_SIZE) {
769 verbose(env, "invalid size of register spill\n");
772 for (i = 1; i < BPF_REG_SIZE; i++) {
773 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
774 verbose(env, "corrupted spill memory\n");
779 if (value_regno >= 0) {
780 /* restore register state from stack */
781 state->regs[value_regno] = state->stack[spi].spilled_ptr;
782 mark_stack_slot_read(state, spi);
786 for (i = 0; i < size; i++) {
787 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_MISC) {
788 verbose(env, "invalid read from stack off %d+%d size %d\n",
793 if (value_regno >= 0)
794 /* have read misc data from the stack */
795 mark_reg_unknown(env, state->regs, value_regno);
800 /* check read/write into map element returned by bpf_map_lookup_elem() */
801 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
802 int size, bool zero_size_allowed)
804 struct bpf_reg_state *regs = cur_regs(env);
805 struct bpf_map *map = regs[regno].map_ptr;
807 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
808 off + size > map->value_size) {
809 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
810 map->value_size, off, size);
816 /* check read/write into a map element with possible variable offset */
817 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
818 int off, int size, bool zero_size_allowed)
820 struct bpf_verifier_state *state = env->cur_state;
821 struct bpf_reg_state *reg = &state->regs[regno];
824 /* We may have adjusted the register to this map value, so we
825 * need to try adding each of min_value and max_value to off
826 * to make sure our theoretical access will be safe.
829 print_verifier_state(env, state);
830 /* The minimum value is only important with signed
831 * comparisons where we can't assume the floor of a
832 * value is 0. If we are using signed variables for our
833 * index'es we need to make sure that whatever we use
834 * will have a set floor within our range.
836 if (reg->smin_value < 0) {
837 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
841 err = __check_map_access(env, regno, reg->smin_value + off, size,
844 verbose(env, "R%d min value is outside of the array range\n",
849 /* If we haven't set a max value then we need to bail since we can't be
850 * sure we won't do bad things.
851 * If reg->umax_value + off could overflow, treat that as unbounded too.
853 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
854 verbose(env, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
858 err = __check_map_access(env, regno, reg->umax_value + off, size,
861 verbose(env, "R%d max value is outside of the array range\n",
866 #define MAX_PACKET_OFF 0xffff
868 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
869 const struct bpf_call_arg_meta *meta,
870 enum bpf_access_type t)
872 switch (env->prog->type) {
873 case BPF_PROG_TYPE_LWT_IN:
874 case BPF_PROG_TYPE_LWT_OUT:
875 /* dst_input() and dst_output() can't write for now */
879 case BPF_PROG_TYPE_SCHED_CLS:
880 case BPF_PROG_TYPE_SCHED_ACT:
881 case BPF_PROG_TYPE_XDP:
882 case BPF_PROG_TYPE_LWT_XMIT:
883 case BPF_PROG_TYPE_SK_SKB:
885 return meta->pkt_access;
887 env->seen_direct_write = true;
894 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
895 int off, int size, bool zero_size_allowed)
897 struct bpf_reg_state *regs = cur_regs(env);
898 struct bpf_reg_state *reg = ®s[regno];
900 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
901 (u64)off + size > reg->range) {
902 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
903 off, size, regno, reg->id, reg->off, reg->range);
909 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
910 int size, bool zero_size_allowed)
912 struct bpf_reg_state *regs = cur_regs(env);
913 struct bpf_reg_state *reg = ®s[regno];
916 /* We may have added a variable offset to the packet pointer; but any
917 * reg->range we have comes after that. We are only checking the fixed
921 /* We don't allow negative numbers, because we aren't tracking enough
922 * detail to prove they're safe.
924 if (reg->smin_value < 0) {
925 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
929 err = __check_packet_access(env, regno, off, size, zero_size_allowed);
931 verbose(env, "R%d offset is outside of the packet\n", regno);
937 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
938 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
939 enum bpf_access_type t, enum bpf_reg_type *reg_type)
941 struct bpf_insn_access_aux info = {
942 .reg_type = *reg_type,
945 if (env->ops->is_valid_access &&
946 env->ops->is_valid_access(off, size, t, &info)) {
947 /* A non zero info.ctx_field_size indicates that this field is a
948 * candidate for later verifier transformation to load the whole
949 * field and then apply a mask when accessed with a narrower
950 * access than actual ctx access size. A zero info.ctx_field_size
951 * will only allow for whole field access and rejects any other
952 * type of narrower access.
954 *reg_type = info.reg_type;
956 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
957 /* remember the offset of last byte accessed in ctx */
958 if (env->prog->aux->max_ctx_offset < off + size)
959 env->prog->aux->max_ctx_offset = off + size;
963 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
967 static bool __is_pointer_value(bool allow_ptr_leaks,
968 const struct bpf_reg_state *reg)
973 return reg->type != SCALAR_VALUE;
976 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
978 return __is_pointer_value(env->allow_ptr_leaks, cur_regs(env) + regno);
981 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
982 const struct bpf_reg_state *reg,
983 int off, int size, bool strict)
988 /* Byte size accesses are always allowed. */
989 if (!strict || size == 1)
992 /* For platforms that do not have a Kconfig enabling
993 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
994 * NET_IP_ALIGN is universally set to '2'. And on platforms
995 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
996 * to this code only in strict mode where we want to emulate
997 * the NET_IP_ALIGN==2 checking. Therefore use an
998 * unconditional IP align value of '2'.
1002 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
1003 if (!tnum_is_aligned(reg_off, size)) {
1006 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1008 "misaligned packet access off %d+%s+%d+%d size %d\n",
1009 ip_align, tn_buf, reg->off, off, size);
1016 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
1017 const struct bpf_reg_state *reg,
1018 const char *pointer_desc,
1019 int off, int size, bool strict)
1021 struct tnum reg_off;
1023 /* Byte size accesses are always allowed. */
1024 if (!strict || size == 1)
1027 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
1028 if (!tnum_is_aligned(reg_off, size)) {
1031 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1032 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
1033 pointer_desc, tn_buf, reg->off, off, size);
1040 static int check_ptr_alignment(struct bpf_verifier_env *env,
1041 const struct bpf_reg_state *reg,
1044 bool strict = env->strict_alignment;
1045 const char *pointer_desc = "";
1047 switch (reg->type) {
1049 case PTR_TO_PACKET_META:
1050 /* Special case, because of NET_IP_ALIGN. Given metadata sits
1051 * right in front, treat it the very same way.
1053 return check_pkt_ptr_alignment(env, reg, off, size, strict);
1054 case PTR_TO_MAP_VALUE:
1055 pointer_desc = "value ";
1058 pointer_desc = "context ";
1061 pointer_desc = "stack ";
1066 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
1070 /* truncate register to smaller size (in bytes)
1071 * must be called with size < BPF_REG_SIZE
1073 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
1077 /* clear high bits in bit representation */
1078 reg->var_off = tnum_cast(reg->var_off, size);
1080 /* fix arithmetic bounds */
1081 mask = ((u64)1 << (size * 8)) - 1;
1082 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
1083 reg->umin_value &= mask;
1084 reg->umax_value &= mask;
1086 reg->umin_value = 0;
1087 reg->umax_value = mask;
1089 reg->smin_value = reg->umin_value;
1090 reg->smax_value = reg->umax_value;
1093 /* check whether memory at (regno + off) is accessible for t = (read | write)
1094 * if t==write, value_regno is a register which value is stored into memory
1095 * if t==read, value_regno is a register which will receive the value from memory
1096 * if t==write && value_regno==-1, some unknown value is stored into memory
1097 * if t==read && value_regno==-1, don't care what we read from memory
1099 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno, int off,
1100 int bpf_size, enum bpf_access_type t,
1103 struct bpf_verifier_state *state = env->cur_state;
1104 struct bpf_reg_state *regs = cur_regs(env);
1105 struct bpf_reg_state *reg = regs + regno;
1108 size = bpf_size_to_bytes(bpf_size);
1112 /* alignment checks will add in reg->off themselves */
1113 err = check_ptr_alignment(env, reg, off, size);
1117 /* for access checks, reg->off is just part of off */
1120 if (reg->type == PTR_TO_MAP_VALUE) {
1121 if (t == BPF_WRITE && value_regno >= 0 &&
1122 is_pointer_value(env, value_regno)) {
1123 verbose(env, "R%d leaks addr into map\n", value_regno);
1127 err = check_map_access(env, regno, off, size, false);
1128 if (!err && t == BPF_READ && value_regno >= 0)
1129 mark_reg_unknown(env, regs, value_regno);
1131 } else if (reg->type == PTR_TO_CTX) {
1132 enum bpf_reg_type reg_type = SCALAR_VALUE;
1134 if (t == BPF_WRITE && value_regno >= 0 &&
1135 is_pointer_value(env, value_regno)) {
1136 verbose(env, "R%d leaks addr into ctx\n", value_regno);
1139 /* ctx accesses must be at a fixed offset, so that we can
1140 * determine what type of data were returned.
1144 "dereference of modified ctx ptr R%d off=%d+%d, ctx+const is allowed, ctx+const+const is not\n",
1145 regno, reg->off, off - reg->off);
1148 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
1151 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1153 "variable ctx access var_off=%s off=%d size=%d",
1157 err = check_ctx_access(env, insn_idx, off, size, t, ®_type);
1158 if (!err && t == BPF_READ && value_regno >= 0) {
1159 /* ctx access returns either a scalar, or a
1160 * PTR_TO_PACKET[_META,_END]. In the latter
1161 * case, we know the offset is zero.
1163 if (reg_type == SCALAR_VALUE)
1164 mark_reg_unknown(env, regs, value_regno);
1166 mark_reg_known_zero(env, regs,
1168 regs[value_regno].id = 0;
1169 regs[value_regno].off = 0;
1170 regs[value_regno].range = 0;
1171 regs[value_regno].type = reg_type;
1174 } else if (reg->type == PTR_TO_STACK) {
1175 /* stack accesses must be at a fixed offset, so that we can
1176 * determine what type of data were returned.
1177 * See check_stack_read().
1179 if (!tnum_is_const(reg->var_off)) {
1182 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1183 verbose(env, "variable stack access var_off=%s off=%d size=%d",
1187 off += reg->var_off.value;
1188 if (off >= 0 || off < -MAX_BPF_STACK) {
1189 verbose(env, "invalid stack off=%d size=%d\n", off,
1194 if (env->prog->aux->stack_depth < -off)
1195 env->prog->aux->stack_depth = -off;
1198 err = check_stack_write(env, state, off, size,
1201 err = check_stack_read(env, state, off, size,
1203 } else if (reg_is_pkt_pointer(reg)) {
1204 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
1205 verbose(env, "cannot write into packet\n");
1208 if (t == BPF_WRITE && value_regno >= 0 &&
1209 is_pointer_value(env, value_regno)) {
1210 verbose(env, "R%d leaks addr into packet\n",
1214 err = check_packet_access(env, regno, off, size, false);
1215 if (!err && t == BPF_READ && value_regno >= 0)
1216 mark_reg_unknown(env, regs, value_regno);
1218 verbose(env, "R%d invalid mem access '%s'\n", regno,
1219 reg_type_str[reg->type]);
1223 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
1224 regs[value_regno].type == SCALAR_VALUE) {
1225 /* b/h/w load zero-extends, mark upper bits as known 0 */
1226 coerce_reg_to_size(®s[value_regno], size);
1231 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
1235 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
1237 verbose(env, "BPF_XADD uses reserved fields\n");
1241 /* check src1 operand */
1242 err = check_reg_arg(env, insn->src_reg, SRC_OP);
1246 /* check src2 operand */
1247 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
1251 if (is_pointer_value(env, insn->src_reg)) {
1252 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
1256 /* check whether atomic_add can read the memory */
1257 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1258 BPF_SIZE(insn->code), BPF_READ, -1);
1262 /* check whether atomic_add can write into the same memory */
1263 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1264 BPF_SIZE(insn->code), BPF_WRITE, -1);
1267 /* Does this register contain a constant zero? */
1268 static bool register_is_null(struct bpf_reg_state reg)
1270 return reg.type == SCALAR_VALUE && tnum_equals_const(reg.var_off, 0);
1273 /* when register 'regno' is passed into function that will read 'access_size'
1274 * bytes from that pointer, make sure that it's within stack boundary
1275 * and all elements of stack are initialized.
1276 * Unlike most pointer bounds-checking functions, this one doesn't take an
1277 * 'off' argument, so it has to add in reg->off itself.
1279 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
1280 int access_size, bool zero_size_allowed,
1281 struct bpf_call_arg_meta *meta)
1283 struct bpf_verifier_state *state = env->cur_state;
1284 struct bpf_reg_state *regs = state->regs;
1285 int off, i, slot, spi;
1287 if (regs[regno].type != PTR_TO_STACK) {
1288 /* Allow zero-byte read from NULL, regardless of pointer type */
1289 if (zero_size_allowed && access_size == 0 &&
1290 register_is_null(regs[regno]))
1293 verbose(env, "R%d type=%s expected=%s\n", regno,
1294 reg_type_str[regs[regno].type],
1295 reg_type_str[PTR_TO_STACK]);
1299 /* Only allow fixed-offset stack reads */
1300 if (!tnum_is_const(regs[regno].var_off)) {
1303 tnum_strn(tn_buf, sizeof(tn_buf), regs[regno].var_off);
1304 verbose(env, "invalid variable stack read R%d var_off=%s\n",
1308 off = regs[regno].off + regs[regno].var_off.value;
1309 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
1310 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
1311 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
1312 regno, off, access_size);
1316 if (env->prog->aux->stack_depth < -off)
1317 env->prog->aux->stack_depth = -off;
1319 if (meta && meta->raw_mode) {
1320 meta->access_size = access_size;
1321 meta->regno = regno;
1325 for (i = 0; i < access_size; i++) {
1326 slot = -(off + i) - 1;
1327 spi = slot / BPF_REG_SIZE;
1328 if (state->allocated_stack <= slot ||
1329 state->stack[spi].slot_type[slot % BPF_REG_SIZE] !=
1331 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
1332 off, i, access_size);
1339 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
1340 int access_size, bool zero_size_allowed,
1341 struct bpf_call_arg_meta *meta)
1343 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
1345 switch (reg->type) {
1347 case PTR_TO_PACKET_META:
1348 return check_packet_access(env, regno, reg->off, access_size,
1350 case PTR_TO_MAP_VALUE:
1351 return check_map_access(env, regno, reg->off, access_size,
1353 default: /* scalar_value|ptr_to_stack or invalid ptr */
1354 return check_stack_boundary(env, regno, access_size,
1355 zero_size_allowed, meta);
1359 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
1360 enum bpf_arg_type arg_type,
1361 struct bpf_call_arg_meta *meta)
1363 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
1364 enum bpf_reg_type expected_type, type = reg->type;
1367 if (arg_type == ARG_DONTCARE)
1370 err = check_reg_arg(env, regno, SRC_OP);
1374 if (arg_type == ARG_ANYTHING) {
1375 if (is_pointer_value(env, regno)) {
1376 verbose(env, "R%d leaks addr into helper function\n",
1383 if (type_is_pkt_pointer(type) &&
1384 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
1385 verbose(env, "helper access to the packet is not allowed\n");
1389 if (arg_type == ARG_PTR_TO_MAP_KEY ||
1390 arg_type == ARG_PTR_TO_MAP_VALUE) {
1391 expected_type = PTR_TO_STACK;
1392 if (!type_is_pkt_pointer(type) &&
1393 type != expected_type)
1395 } else if (arg_type == ARG_CONST_SIZE ||
1396 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1397 expected_type = SCALAR_VALUE;
1398 if (type != expected_type)
1400 } else if (arg_type == ARG_CONST_MAP_PTR) {
1401 expected_type = CONST_PTR_TO_MAP;
1402 if (type != expected_type)
1404 } else if (arg_type == ARG_PTR_TO_CTX) {
1405 expected_type = PTR_TO_CTX;
1406 if (type != expected_type)
1408 } else if (arg_type == ARG_PTR_TO_MEM ||
1409 arg_type == ARG_PTR_TO_MEM_OR_NULL ||
1410 arg_type == ARG_PTR_TO_UNINIT_MEM) {
1411 expected_type = PTR_TO_STACK;
1412 /* One exception here. In case function allows for NULL to be
1413 * passed in as argument, it's a SCALAR_VALUE type. Final test
1414 * happens during stack boundary checking.
1416 if (register_is_null(*reg) &&
1417 arg_type == ARG_PTR_TO_MEM_OR_NULL)
1418 /* final test in check_stack_boundary() */;
1419 else if (!type_is_pkt_pointer(type) &&
1420 type != PTR_TO_MAP_VALUE &&
1421 type != expected_type)
1423 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
1425 verbose(env, "unsupported arg_type %d\n", arg_type);
1429 if (arg_type == ARG_CONST_MAP_PTR) {
1430 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
1431 meta->map_ptr = reg->map_ptr;
1432 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
1433 /* bpf_map_xxx(..., map_ptr, ..., key) call:
1434 * check that [key, key + map->key_size) are within
1435 * stack limits and initialized
1437 if (!meta->map_ptr) {
1438 /* in function declaration map_ptr must come before
1439 * map_key, so that it's verified and known before
1440 * we have to check map_key here. Otherwise it means
1441 * that kernel subsystem misconfigured verifier
1443 verbose(env, "invalid map_ptr to access map->key\n");
1446 if (type_is_pkt_pointer(type))
1447 err = check_packet_access(env, regno, reg->off,
1448 meta->map_ptr->key_size,
1451 err = check_stack_boundary(env, regno,
1452 meta->map_ptr->key_size,
1454 } else if (arg_type == ARG_PTR_TO_MAP_VALUE) {
1455 /* bpf_map_xxx(..., map_ptr, ..., value) call:
1456 * check [value, value + map->value_size) validity
1458 if (!meta->map_ptr) {
1459 /* kernel subsystem misconfigured verifier */
1460 verbose(env, "invalid map_ptr to access map->value\n");
1463 if (type_is_pkt_pointer(type))
1464 err = check_packet_access(env, regno, reg->off,
1465 meta->map_ptr->value_size,
1468 err = check_stack_boundary(env, regno,
1469 meta->map_ptr->value_size,
1471 } else if (arg_type == ARG_CONST_SIZE ||
1472 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1473 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
1475 /* bpf_xxx(..., buf, len) call will access 'len' bytes
1476 * from stack pointer 'buf'. Check it
1477 * note: regno == len, regno - 1 == buf
1480 /* kernel subsystem misconfigured verifier */
1482 "ARG_CONST_SIZE cannot be first argument\n");
1486 /* The register is SCALAR_VALUE; the access check
1487 * happens using its boundaries.
1490 if (!tnum_is_const(reg->var_off))
1491 /* For unprivileged variable accesses, disable raw
1492 * mode so that the program is required to
1493 * initialize all the memory that the helper could
1494 * just partially fill up.
1498 if (reg->smin_value < 0) {
1499 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
1504 if (reg->umin_value == 0) {
1505 err = check_helper_mem_access(env, regno - 1, 0,
1512 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
1513 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
1517 err = check_helper_mem_access(env, regno - 1,
1519 zero_size_allowed, meta);
1524 verbose(env, "R%d type=%s expected=%s\n", regno,
1525 reg_type_str[type], reg_type_str[expected_type]);
1529 static int check_map_func_compatibility(struct bpf_verifier_env *env,
1530 struct bpf_map *map, int func_id)
1535 /* We need a two way check, first is from map perspective ... */
1536 switch (map->map_type) {
1537 case BPF_MAP_TYPE_PROG_ARRAY:
1538 if (func_id != BPF_FUNC_tail_call)
1541 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
1542 if (func_id != BPF_FUNC_perf_event_read &&
1543 func_id != BPF_FUNC_perf_event_output &&
1544 func_id != BPF_FUNC_perf_event_read_value)
1547 case BPF_MAP_TYPE_STACK_TRACE:
1548 if (func_id != BPF_FUNC_get_stackid)
1551 case BPF_MAP_TYPE_CGROUP_ARRAY:
1552 if (func_id != BPF_FUNC_skb_under_cgroup &&
1553 func_id != BPF_FUNC_current_task_under_cgroup)
1556 /* devmap returns a pointer to a live net_device ifindex that we cannot
1557 * allow to be modified from bpf side. So do not allow lookup elements
1560 case BPF_MAP_TYPE_DEVMAP:
1561 if (func_id != BPF_FUNC_redirect_map)
1564 /* Restrict bpf side of cpumap, open when use-cases appear */
1565 case BPF_MAP_TYPE_CPUMAP:
1566 if (func_id != BPF_FUNC_redirect_map)
1569 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
1570 case BPF_MAP_TYPE_HASH_OF_MAPS:
1571 if (func_id != BPF_FUNC_map_lookup_elem)
1574 case BPF_MAP_TYPE_SOCKMAP:
1575 if (func_id != BPF_FUNC_sk_redirect_map &&
1576 func_id != BPF_FUNC_sock_map_update &&
1577 func_id != BPF_FUNC_map_delete_elem)
1584 /* ... and second from the function itself. */
1586 case BPF_FUNC_tail_call:
1587 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
1590 case BPF_FUNC_perf_event_read:
1591 case BPF_FUNC_perf_event_output:
1592 case BPF_FUNC_perf_event_read_value:
1593 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
1596 case BPF_FUNC_get_stackid:
1597 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
1600 case BPF_FUNC_current_task_under_cgroup:
1601 case BPF_FUNC_skb_under_cgroup:
1602 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
1605 case BPF_FUNC_redirect_map:
1606 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
1607 map->map_type != BPF_MAP_TYPE_CPUMAP)
1610 case BPF_FUNC_sk_redirect_map:
1611 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
1614 case BPF_FUNC_sock_map_update:
1615 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
1624 verbose(env, "cannot pass map_type %d into func %s#%d\n",
1625 map->map_type, func_id_name(func_id), func_id);
1629 static int check_raw_mode(const struct bpf_func_proto *fn)
1633 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
1635 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
1637 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
1639 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
1641 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
1644 return count > 1 ? -EINVAL : 0;
1647 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
1648 * are now invalid, so turn them into unknown SCALAR_VALUE.
1650 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
1652 struct bpf_verifier_state *state = env->cur_state;
1653 struct bpf_reg_state *regs = state->regs, *reg;
1656 for (i = 0; i < MAX_BPF_REG; i++)
1657 if (reg_is_pkt_pointer_any(®s[i]))
1658 mark_reg_unknown(env, regs, i);
1660 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
1661 if (state->stack[i].slot_type[0] != STACK_SPILL)
1663 reg = &state->stack[i].spilled_ptr;
1664 if (reg_is_pkt_pointer_any(reg))
1665 __mark_reg_unknown(reg);
1669 static int check_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
1671 const struct bpf_func_proto *fn = NULL;
1672 struct bpf_reg_state *regs;
1673 struct bpf_call_arg_meta meta;
1677 /* find function prototype */
1678 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
1679 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
1684 if (env->ops->get_func_proto)
1685 fn = env->ops->get_func_proto(func_id);
1688 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
1693 /* eBPF programs must be GPL compatible to use GPL-ed functions */
1694 if (!env->prog->gpl_compatible && fn->gpl_only) {
1695 verbose(env, "cannot call GPL only function from proprietary program\n");
1699 /* With LD_ABS/IND some JITs save/restore skb from r1. */
1700 changes_data = bpf_helper_changes_pkt_data(fn->func);
1701 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
1702 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
1703 func_id_name(func_id), func_id);
1707 memset(&meta, 0, sizeof(meta));
1708 meta.pkt_access = fn->pkt_access;
1710 /* We only support one arg being in raw mode at the moment, which
1711 * is sufficient for the helper functions we have right now.
1713 err = check_raw_mode(fn);
1715 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
1716 func_id_name(func_id), func_id);
1721 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
1724 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
1727 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
1730 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
1733 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
1737 /* Mark slots with STACK_MISC in case of raw mode, stack offset
1738 * is inferred from register state.
1740 for (i = 0; i < meta.access_size; i++) {
1741 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B, BPF_WRITE, -1);
1746 regs = cur_regs(env);
1747 /* reset caller saved regs */
1748 for (i = 0; i < CALLER_SAVED_REGS; i++) {
1749 mark_reg_not_init(env, regs, caller_saved[i]);
1750 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
1753 /* update return register (already marked as written above) */
1754 if (fn->ret_type == RET_INTEGER) {
1755 /* sets type to SCALAR_VALUE */
1756 mark_reg_unknown(env, regs, BPF_REG_0);
1757 } else if (fn->ret_type == RET_VOID) {
1758 regs[BPF_REG_0].type = NOT_INIT;
1759 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL) {
1760 struct bpf_insn_aux_data *insn_aux;
1762 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
1763 /* There is no offset yet applied, variable or fixed */
1764 mark_reg_known_zero(env, regs, BPF_REG_0);
1765 regs[BPF_REG_0].off = 0;
1766 /* remember map_ptr, so that check_map_access()
1767 * can check 'value_size' boundary of memory access
1768 * to map element returned from bpf_map_lookup_elem()
1770 if (meta.map_ptr == NULL) {
1772 "kernel subsystem misconfigured verifier\n");
1775 regs[BPF_REG_0].map_ptr = meta.map_ptr;
1776 regs[BPF_REG_0].id = ++env->id_gen;
1777 insn_aux = &env->insn_aux_data[insn_idx];
1778 if (!insn_aux->map_ptr)
1779 insn_aux->map_ptr = meta.map_ptr;
1780 else if (insn_aux->map_ptr != meta.map_ptr)
1781 insn_aux->map_ptr = BPF_MAP_PTR_POISON;
1783 verbose(env, "unknown return type %d of func %s#%d\n",
1784 fn->ret_type, func_id_name(func_id), func_id);
1788 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
1793 clear_all_pkt_pointers(env);
1797 static bool signed_add_overflows(s64 a, s64 b)
1799 /* Do the add in u64, where overflow is well-defined */
1800 s64 res = (s64)((u64)a + (u64)b);
1807 static bool signed_sub_overflows(s64 a, s64 b)
1809 /* Do the sub in u64, where overflow is well-defined */
1810 s64 res = (s64)((u64)a - (u64)b);
1817 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
1818 * Caller should also handle BPF_MOV case separately.
1819 * If we return -EACCES, caller may want to try again treating pointer as a
1820 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
1822 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
1823 struct bpf_insn *insn,
1824 const struct bpf_reg_state *ptr_reg,
1825 const struct bpf_reg_state *off_reg)
1827 struct bpf_reg_state *regs = cur_regs(env), *dst_reg;
1828 bool known = tnum_is_const(off_reg->var_off);
1829 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
1830 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
1831 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
1832 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
1833 u8 opcode = BPF_OP(insn->code);
1834 u32 dst = insn->dst_reg;
1836 dst_reg = ®s[dst];
1838 if (WARN_ON_ONCE(known && (smin_val != smax_val))) {
1839 print_verifier_state(env, env->cur_state);
1841 "verifier internal error: known but bad sbounds\n");
1844 if (WARN_ON_ONCE(known && (umin_val != umax_val))) {
1845 print_verifier_state(env, env->cur_state);
1847 "verifier internal error: known but bad ubounds\n");
1851 if (BPF_CLASS(insn->code) != BPF_ALU64) {
1852 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
1853 if (!env->allow_ptr_leaks)
1855 "R%d 32-bit pointer arithmetic prohibited\n",
1860 if (ptr_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
1861 if (!env->allow_ptr_leaks)
1862 verbose(env, "R%d pointer arithmetic on PTR_TO_MAP_VALUE_OR_NULL prohibited, null-check it first\n",
1866 if (ptr_reg->type == CONST_PTR_TO_MAP) {
1867 if (!env->allow_ptr_leaks)
1868 verbose(env, "R%d pointer arithmetic on CONST_PTR_TO_MAP prohibited\n",
1872 if (ptr_reg->type == PTR_TO_PACKET_END) {
1873 if (!env->allow_ptr_leaks)
1874 verbose(env, "R%d pointer arithmetic on PTR_TO_PACKET_END prohibited\n",
1879 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
1880 * The id may be overwritten later if we create a new variable offset.
1882 dst_reg->type = ptr_reg->type;
1883 dst_reg->id = ptr_reg->id;
1887 /* We can take a fixed offset as long as it doesn't overflow
1888 * the s32 'off' field
1890 if (known && (ptr_reg->off + smin_val ==
1891 (s64)(s32)(ptr_reg->off + smin_val))) {
1892 /* pointer += K. Accumulate it into fixed offset */
1893 dst_reg->smin_value = smin_ptr;
1894 dst_reg->smax_value = smax_ptr;
1895 dst_reg->umin_value = umin_ptr;
1896 dst_reg->umax_value = umax_ptr;
1897 dst_reg->var_off = ptr_reg->var_off;
1898 dst_reg->off = ptr_reg->off + smin_val;
1899 dst_reg->range = ptr_reg->range;
1902 /* A new variable offset is created. Note that off_reg->off
1903 * == 0, since it's a scalar.
1904 * dst_reg gets the pointer type and since some positive
1905 * integer value was added to the pointer, give it a new 'id'
1906 * if it's a PTR_TO_PACKET.
1907 * this creates a new 'base' pointer, off_reg (variable) gets
1908 * added into the variable offset, and we copy the fixed offset
1911 if (signed_add_overflows(smin_ptr, smin_val) ||
1912 signed_add_overflows(smax_ptr, smax_val)) {
1913 dst_reg->smin_value = S64_MIN;
1914 dst_reg->smax_value = S64_MAX;
1916 dst_reg->smin_value = smin_ptr + smin_val;
1917 dst_reg->smax_value = smax_ptr + smax_val;
1919 if (umin_ptr + umin_val < umin_ptr ||
1920 umax_ptr + umax_val < umax_ptr) {
1921 dst_reg->umin_value = 0;
1922 dst_reg->umax_value = U64_MAX;
1924 dst_reg->umin_value = umin_ptr + umin_val;
1925 dst_reg->umax_value = umax_ptr + umax_val;
1927 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
1928 dst_reg->off = ptr_reg->off;
1929 if (reg_is_pkt_pointer(ptr_reg)) {
1930 dst_reg->id = ++env->id_gen;
1931 /* something was added to pkt_ptr, set range to zero */
1936 if (dst_reg == off_reg) {
1937 /* scalar -= pointer. Creates an unknown scalar */
1938 if (!env->allow_ptr_leaks)
1939 verbose(env, "R%d tried to subtract pointer from scalar\n",
1943 /* We don't allow subtraction from FP, because (according to
1944 * test_verifier.c test "invalid fp arithmetic", JITs might not
1945 * be able to deal with it.
1947 if (ptr_reg->type == PTR_TO_STACK) {
1948 if (!env->allow_ptr_leaks)
1949 verbose(env, "R%d subtraction from stack pointer prohibited\n",
1953 if (known && (ptr_reg->off - smin_val ==
1954 (s64)(s32)(ptr_reg->off - smin_val))) {
1955 /* pointer -= K. Subtract it from fixed offset */
1956 dst_reg->smin_value = smin_ptr;
1957 dst_reg->smax_value = smax_ptr;
1958 dst_reg->umin_value = umin_ptr;
1959 dst_reg->umax_value = umax_ptr;
1960 dst_reg->var_off = ptr_reg->var_off;
1961 dst_reg->id = ptr_reg->id;
1962 dst_reg->off = ptr_reg->off - smin_val;
1963 dst_reg->range = ptr_reg->range;
1966 /* A new variable offset is created. If the subtrahend is known
1967 * nonnegative, then any reg->range we had before is still good.
1969 if (signed_sub_overflows(smin_ptr, smax_val) ||
1970 signed_sub_overflows(smax_ptr, smin_val)) {
1971 /* Overflow possible, we know nothing */
1972 dst_reg->smin_value = S64_MIN;
1973 dst_reg->smax_value = S64_MAX;
1975 dst_reg->smin_value = smin_ptr - smax_val;
1976 dst_reg->smax_value = smax_ptr - smin_val;
1978 if (umin_ptr < umax_val) {
1979 /* Overflow possible, we know nothing */
1980 dst_reg->umin_value = 0;
1981 dst_reg->umax_value = U64_MAX;
1983 /* Cannot overflow (as long as bounds are consistent) */
1984 dst_reg->umin_value = umin_ptr - umax_val;
1985 dst_reg->umax_value = umax_ptr - umin_val;
1987 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
1988 dst_reg->off = ptr_reg->off;
1989 if (reg_is_pkt_pointer(ptr_reg)) {
1990 dst_reg->id = ++env->id_gen;
1991 /* something was added to pkt_ptr, set range to zero */
1999 /* bitwise ops on pointers are troublesome, prohibit for now.
2000 * (However, in principle we could allow some cases, e.g.
2001 * ptr &= ~3 which would reduce min_value by 3.)
2003 if (!env->allow_ptr_leaks)
2004 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
2005 dst, bpf_alu_string[opcode >> 4]);
2008 /* other operators (e.g. MUL,LSH) produce non-pointer results */
2009 if (!env->allow_ptr_leaks)
2010 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
2011 dst, bpf_alu_string[opcode >> 4]);
2015 __update_reg_bounds(dst_reg);
2016 __reg_deduce_bounds(dst_reg);
2017 __reg_bound_offset(dst_reg);
2021 /* WARNING: This function does calculations on 64-bit values, but the actual
2022 * execution may occur on 32-bit values. Therefore, things like bitshifts
2023 * need extra checks in the 32-bit case.
2025 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
2026 struct bpf_insn *insn,
2027 struct bpf_reg_state *dst_reg,
2028 struct bpf_reg_state src_reg)
2030 struct bpf_reg_state *regs = cur_regs(env);
2031 u8 opcode = BPF_OP(insn->code);
2032 bool src_known, dst_known;
2033 s64 smin_val, smax_val;
2034 u64 umin_val, umax_val;
2035 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
2037 smin_val = src_reg.smin_value;
2038 smax_val = src_reg.smax_value;
2039 umin_val = src_reg.umin_value;
2040 umax_val = src_reg.umax_value;
2041 src_known = tnum_is_const(src_reg.var_off);
2042 dst_known = tnum_is_const(dst_reg->var_off);
2046 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
2047 signed_add_overflows(dst_reg->smax_value, smax_val)) {
2048 dst_reg->smin_value = S64_MIN;
2049 dst_reg->smax_value = S64_MAX;
2051 dst_reg->smin_value += smin_val;
2052 dst_reg->smax_value += smax_val;
2054 if (dst_reg->umin_value + umin_val < umin_val ||
2055 dst_reg->umax_value + umax_val < umax_val) {
2056 dst_reg->umin_value = 0;
2057 dst_reg->umax_value = U64_MAX;
2059 dst_reg->umin_value += umin_val;
2060 dst_reg->umax_value += umax_val;
2062 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
2065 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
2066 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
2067 /* Overflow possible, we know nothing */
2068 dst_reg->smin_value = S64_MIN;
2069 dst_reg->smax_value = S64_MAX;
2071 dst_reg->smin_value -= smax_val;
2072 dst_reg->smax_value -= smin_val;
2074 if (dst_reg->umin_value < umax_val) {
2075 /* Overflow possible, we know nothing */
2076 dst_reg->umin_value = 0;
2077 dst_reg->umax_value = U64_MAX;
2079 /* Cannot overflow (as long as bounds are consistent) */
2080 dst_reg->umin_value -= umax_val;
2081 dst_reg->umax_value -= umin_val;
2083 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
2086 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
2087 if (smin_val < 0 || dst_reg->smin_value < 0) {
2088 /* Ain't nobody got time to multiply that sign */
2089 __mark_reg_unbounded(dst_reg);
2090 __update_reg_bounds(dst_reg);
2093 /* Both values are positive, so we can work with unsigned and
2094 * copy the result to signed (unless it exceeds S64_MAX).
2096 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
2097 /* Potential overflow, we know nothing */
2098 __mark_reg_unbounded(dst_reg);
2099 /* (except what we can learn from the var_off) */
2100 __update_reg_bounds(dst_reg);
2103 dst_reg->umin_value *= umin_val;
2104 dst_reg->umax_value *= umax_val;
2105 if (dst_reg->umax_value > S64_MAX) {
2106 /* Overflow possible, we know nothing */
2107 dst_reg->smin_value = S64_MIN;
2108 dst_reg->smax_value = S64_MAX;
2110 dst_reg->smin_value = dst_reg->umin_value;
2111 dst_reg->smax_value = dst_reg->umax_value;
2115 if (src_known && dst_known) {
2116 __mark_reg_known(dst_reg, dst_reg->var_off.value &
2117 src_reg.var_off.value);
2120 /* We get our minimum from the var_off, since that's inherently
2121 * bitwise. Our maximum is the minimum of the operands' maxima.
2123 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
2124 dst_reg->umin_value = dst_reg->var_off.value;
2125 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
2126 if (dst_reg->smin_value < 0 || smin_val < 0) {
2127 /* Lose signed bounds when ANDing negative numbers,
2128 * ain't nobody got time for that.
2130 dst_reg->smin_value = S64_MIN;
2131 dst_reg->smax_value = S64_MAX;
2133 /* ANDing two positives gives a positive, so safe to
2134 * cast result into s64.
2136 dst_reg->smin_value = dst_reg->umin_value;
2137 dst_reg->smax_value = dst_reg->umax_value;
2139 /* We may learn something more from the var_off */
2140 __update_reg_bounds(dst_reg);
2143 if (src_known && dst_known) {
2144 __mark_reg_known(dst_reg, dst_reg->var_off.value |
2145 src_reg.var_off.value);
2148 /* We get our maximum from the var_off, and our minimum is the
2149 * maximum of the operands' minima
2151 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
2152 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
2153 dst_reg->umax_value = dst_reg->var_off.value |
2154 dst_reg->var_off.mask;
2155 if (dst_reg->smin_value < 0 || smin_val < 0) {
2156 /* Lose signed bounds when ORing negative numbers,
2157 * ain't nobody got time for that.
2159 dst_reg->smin_value = S64_MIN;
2160 dst_reg->smax_value = S64_MAX;
2162 /* ORing two positives gives a positive, so safe to
2163 * cast result into s64.
2165 dst_reg->smin_value = dst_reg->umin_value;
2166 dst_reg->smax_value = dst_reg->umax_value;
2168 /* We may learn something more from the var_off */
2169 __update_reg_bounds(dst_reg);
2172 if (umax_val >= insn_bitness) {
2173 /* Shifts greater than 31 or 63 are undefined.
2174 * This includes shifts by a negative number.
2176 mark_reg_unknown(env, regs, insn->dst_reg);
2179 /* We lose all sign bit information (except what we can pick
2182 dst_reg->smin_value = S64_MIN;
2183 dst_reg->smax_value = S64_MAX;
2184 /* If we might shift our top bit out, then we know nothing */
2185 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
2186 dst_reg->umin_value = 0;
2187 dst_reg->umax_value = U64_MAX;
2189 dst_reg->umin_value <<= umin_val;
2190 dst_reg->umax_value <<= umax_val;
2193 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
2195 dst_reg->var_off = tnum_lshift(tnum_unknown, umin_val);
2196 /* We may learn something more from the var_off */
2197 __update_reg_bounds(dst_reg);
2200 if (umax_val >= insn_bitness) {
2201 /* Shifts greater than 31 or 63 are undefined.
2202 * This includes shifts by a negative number.
2204 mark_reg_unknown(env, regs, insn->dst_reg);
2207 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
2208 * be negative, then either:
2209 * 1) src_reg might be zero, so the sign bit of the result is
2210 * unknown, so we lose our signed bounds
2211 * 2) it's known negative, thus the unsigned bounds capture the
2213 * 3) the signed bounds cross zero, so they tell us nothing
2215 * If the value in dst_reg is known nonnegative, then again the
2216 * unsigned bounts capture the signed bounds.
2217 * Thus, in all cases it suffices to blow away our signed bounds
2218 * and rely on inferring new ones from the unsigned bounds and
2219 * var_off of the result.
2221 dst_reg->smin_value = S64_MIN;
2222 dst_reg->smax_value = S64_MAX;
2224 dst_reg->var_off = tnum_rshift(dst_reg->var_off,
2227 dst_reg->var_off = tnum_rshift(tnum_unknown, umin_val);
2228 dst_reg->umin_value >>= umax_val;
2229 dst_reg->umax_value >>= umin_val;
2230 /* We may learn something more from the var_off */
2231 __update_reg_bounds(dst_reg);
2234 mark_reg_unknown(env, regs, insn->dst_reg);
2238 if (BPF_CLASS(insn->code) != BPF_ALU64) {
2239 /* 32-bit ALU ops are (32,32)->32 */
2240 coerce_reg_to_size(dst_reg, 4);
2241 coerce_reg_to_size(&src_reg, 4);
2244 __reg_deduce_bounds(dst_reg);
2245 __reg_bound_offset(dst_reg);
2249 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
2252 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
2253 struct bpf_insn *insn)
2255 struct bpf_reg_state *regs = cur_regs(env), *dst_reg, *src_reg;
2256 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
2257 u8 opcode = BPF_OP(insn->code);
2260 dst_reg = ®s[insn->dst_reg];
2262 if (dst_reg->type != SCALAR_VALUE)
2264 if (BPF_SRC(insn->code) == BPF_X) {
2265 src_reg = ®s[insn->src_reg];
2266 if (src_reg->type != SCALAR_VALUE) {
2267 if (dst_reg->type != SCALAR_VALUE) {
2268 /* Combining two pointers by any ALU op yields
2269 * an arbitrary scalar.
2271 if (!env->allow_ptr_leaks) {
2272 verbose(env, "R%d pointer %s pointer prohibited\n",
2274 bpf_alu_string[opcode >> 4]);
2277 mark_reg_unknown(env, regs, insn->dst_reg);
2280 /* scalar += pointer
2281 * This is legal, but we have to reverse our
2282 * src/dest handling in computing the range
2284 rc = adjust_ptr_min_max_vals(env, insn,
2286 if (rc == -EACCES && env->allow_ptr_leaks) {
2287 /* scalar += unknown scalar */
2288 __mark_reg_unknown(&off_reg);
2289 return adjust_scalar_min_max_vals(
2295 } else if (ptr_reg) {
2296 /* pointer += scalar */
2297 rc = adjust_ptr_min_max_vals(env, insn,
2299 if (rc == -EACCES && env->allow_ptr_leaks) {
2300 /* unknown scalar += scalar */
2301 __mark_reg_unknown(dst_reg);
2302 return adjust_scalar_min_max_vals(
2303 env, insn, dst_reg, *src_reg);
2308 /* Pretend the src is a reg with a known value, since we only
2309 * need to be able to read from this state.
2311 off_reg.type = SCALAR_VALUE;
2312 __mark_reg_known(&off_reg, insn->imm);
2314 if (ptr_reg) { /* pointer += K */
2315 rc = adjust_ptr_min_max_vals(env, insn,
2317 if (rc == -EACCES && env->allow_ptr_leaks) {
2318 /* unknown scalar += K */
2319 __mark_reg_unknown(dst_reg);
2320 return adjust_scalar_min_max_vals(
2321 env, insn, dst_reg, off_reg);
2327 /* Got here implies adding two SCALAR_VALUEs */
2328 if (WARN_ON_ONCE(ptr_reg)) {
2329 print_verifier_state(env, env->cur_state);
2330 verbose(env, "verifier internal error: unexpected ptr_reg\n");
2333 if (WARN_ON(!src_reg)) {
2334 print_verifier_state(env, env->cur_state);
2335 verbose(env, "verifier internal error: no src_reg\n");
2338 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
2341 /* check validity of 32-bit and 64-bit arithmetic operations */
2342 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
2344 struct bpf_reg_state *regs = cur_regs(env);
2345 u8 opcode = BPF_OP(insn->code);
2348 if (opcode == BPF_END || opcode == BPF_NEG) {
2349 if (opcode == BPF_NEG) {
2350 if (BPF_SRC(insn->code) != 0 ||
2351 insn->src_reg != BPF_REG_0 ||
2352 insn->off != 0 || insn->imm != 0) {
2353 verbose(env, "BPF_NEG uses reserved fields\n");
2357 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
2358 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
2359 BPF_CLASS(insn->code) == BPF_ALU64) {
2360 verbose(env, "BPF_END uses reserved fields\n");
2365 /* check src operand */
2366 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2370 if (is_pointer_value(env, insn->dst_reg)) {
2371 verbose(env, "R%d pointer arithmetic prohibited\n",
2376 /* check dest operand */
2377 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2381 } else if (opcode == BPF_MOV) {
2383 if (BPF_SRC(insn->code) == BPF_X) {
2384 if (insn->imm != 0 || insn->off != 0) {
2385 verbose(env, "BPF_MOV uses reserved fields\n");
2389 /* check src operand */
2390 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2394 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
2395 verbose(env, "BPF_MOV uses reserved fields\n");
2400 /* check dest operand */
2401 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2405 if (BPF_SRC(insn->code) == BPF_X) {
2406 if (BPF_CLASS(insn->code) == BPF_ALU64) {
2408 * copy register state to dest reg
2410 regs[insn->dst_reg] = regs[insn->src_reg];
2411 regs[insn->dst_reg].live |= REG_LIVE_WRITTEN;
2414 if (is_pointer_value(env, insn->src_reg)) {
2416 "R%d partial copy of pointer\n",
2420 mark_reg_unknown(env, regs, insn->dst_reg);
2421 coerce_reg_to_size(®s[insn->dst_reg], 4);
2425 * remember the value we stored into this reg
2427 regs[insn->dst_reg].type = SCALAR_VALUE;
2428 if (BPF_CLASS(insn->code) == BPF_ALU64) {
2429 __mark_reg_known(regs + insn->dst_reg,
2432 __mark_reg_known(regs + insn->dst_reg,
2437 } else if (opcode > BPF_END) {
2438 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
2441 } else { /* all other ALU ops: and, sub, xor, add, ... */
2443 if (BPF_SRC(insn->code) == BPF_X) {
2444 if (insn->imm != 0 || insn->off != 0) {
2445 verbose(env, "BPF_ALU uses reserved fields\n");
2448 /* check src1 operand */
2449 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2453 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
2454 verbose(env, "BPF_ALU uses reserved fields\n");
2459 /* check src2 operand */
2460 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2464 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
2465 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
2466 verbose(env, "div by zero\n");
2470 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
2471 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
2472 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
2474 if (insn->imm < 0 || insn->imm >= size) {
2475 verbose(env, "invalid shift %d\n", insn->imm);
2480 /* check dest operand */
2481 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
2485 return adjust_reg_min_max_vals(env, insn);
2491 static void find_good_pkt_pointers(struct bpf_verifier_state *state,
2492 struct bpf_reg_state *dst_reg,
2493 enum bpf_reg_type type,
2494 bool range_right_open)
2496 struct bpf_reg_state *regs = state->regs, *reg;
2500 if (dst_reg->off < 0 ||
2501 (dst_reg->off == 0 && range_right_open))
2502 /* This doesn't give us any range */
2505 if (dst_reg->umax_value > MAX_PACKET_OFF ||
2506 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
2507 /* Risk of overflow. For instance, ptr + (1<<63) may be less
2508 * than pkt_end, but that's because it's also less than pkt.
2512 new_range = dst_reg->off;
2513 if (range_right_open)
2516 /* Examples for register markings:
2518 * pkt_data in dst register:
2522 * if (r2 > pkt_end) goto <handle exception>
2527 * if (r2 < pkt_end) goto <access okay>
2528 * <handle exception>
2531 * r2 == dst_reg, pkt_end == src_reg
2532 * r2=pkt(id=n,off=8,r=0)
2533 * r3=pkt(id=n,off=0,r=0)
2535 * pkt_data in src register:
2539 * if (pkt_end >= r2) goto <access okay>
2540 * <handle exception>
2544 * if (pkt_end <= r2) goto <handle exception>
2548 * pkt_end == dst_reg, r2 == src_reg
2549 * r2=pkt(id=n,off=8,r=0)
2550 * r3=pkt(id=n,off=0,r=0)
2552 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
2553 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
2554 * and [r3, r3 + 8-1) respectively is safe to access depending on
2558 /* If our ids match, then we must have the same max_value. And we
2559 * don't care about the other reg's fixed offset, since if it's too big
2560 * the range won't allow anything.
2561 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
2563 for (i = 0; i < MAX_BPF_REG; i++)
2564 if (regs[i].type == type && regs[i].id == dst_reg->id)
2565 /* keep the maximum range already checked */
2566 regs[i].range = max(regs[i].range, new_range);
2568 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
2569 if (state->stack[i].slot_type[0] != STACK_SPILL)
2571 reg = &state->stack[i].spilled_ptr;
2572 if (reg->type == type && reg->id == dst_reg->id)
2573 reg->range = max(reg->range, new_range);
2577 /* Adjusts the register min/max values in the case that the dst_reg is the
2578 * variable register that we are working on, and src_reg is a constant or we're
2579 * simply doing a BPF_K check.
2580 * In JEQ/JNE cases we also adjust the var_off values.
2582 static void reg_set_min_max(struct bpf_reg_state *true_reg,
2583 struct bpf_reg_state *false_reg, u64 val,
2586 /* If the dst_reg is a pointer, we can't learn anything about its
2587 * variable offset from the compare (unless src_reg were a pointer into
2588 * the same object, but we don't bother with that.
2589 * Since false_reg and true_reg have the same type by construction, we
2590 * only need to check one of them for pointerness.
2592 if (__is_pointer_value(false, false_reg))
2597 /* If this is false then we know nothing Jon Snow, but if it is
2598 * true then we know for sure.
2600 __mark_reg_known(true_reg, val);
2603 /* If this is true we know nothing Jon Snow, but if it is false
2604 * we know the value for sure;
2606 __mark_reg_known(false_reg, val);
2609 false_reg->umax_value = min(false_reg->umax_value, val);
2610 true_reg->umin_value = max(true_reg->umin_value, val + 1);
2613 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
2614 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
2617 false_reg->umin_value = max(false_reg->umin_value, val);
2618 true_reg->umax_value = min(true_reg->umax_value, val - 1);
2621 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
2622 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
2625 false_reg->umax_value = min(false_reg->umax_value, val - 1);
2626 true_reg->umin_value = max(true_reg->umin_value, val);
2629 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
2630 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
2633 false_reg->umin_value = max(false_reg->umin_value, val + 1);
2634 true_reg->umax_value = min(true_reg->umax_value, val);
2637 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
2638 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
2644 __reg_deduce_bounds(false_reg);
2645 __reg_deduce_bounds(true_reg);
2646 /* We might have learned some bits from the bounds. */
2647 __reg_bound_offset(false_reg);
2648 __reg_bound_offset(true_reg);
2649 /* Intersecting with the old var_off might have improved our bounds
2650 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2651 * then new var_off is (0; 0x7f...fc) which improves our umax.
2653 __update_reg_bounds(false_reg);
2654 __update_reg_bounds(true_reg);
2657 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
2660 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
2661 struct bpf_reg_state *false_reg, u64 val,
2664 if (__is_pointer_value(false, false_reg))
2669 /* If this is false then we know nothing Jon Snow, but if it is
2670 * true then we know for sure.
2672 __mark_reg_known(true_reg, val);
2675 /* If this is true we know nothing Jon Snow, but if it is false
2676 * we know the value for sure;
2678 __mark_reg_known(false_reg, val);
2681 true_reg->umax_value = min(true_reg->umax_value, val - 1);
2682 false_reg->umin_value = max(false_reg->umin_value, val);
2685 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
2686 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
2689 true_reg->umin_value = max(true_reg->umin_value, val + 1);
2690 false_reg->umax_value = min(false_reg->umax_value, val);
2693 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
2694 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
2697 true_reg->umax_value = min(true_reg->umax_value, val);
2698 false_reg->umin_value = max(false_reg->umin_value, val + 1);
2701 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
2702 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
2705 true_reg->umin_value = max(true_reg->umin_value, val);
2706 false_reg->umax_value = min(false_reg->umax_value, val - 1);
2709 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
2710 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
2716 __reg_deduce_bounds(false_reg);
2717 __reg_deduce_bounds(true_reg);
2718 /* We might have learned some bits from the bounds. */
2719 __reg_bound_offset(false_reg);
2720 __reg_bound_offset(true_reg);
2721 /* Intersecting with the old var_off might have improved our bounds
2722 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2723 * then new var_off is (0; 0x7f...fc) which improves our umax.
2725 __update_reg_bounds(false_reg);
2726 __update_reg_bounds(true_reg);
2729 /* Regs are known to be equal, so intersect their min/max/var_off */
2730 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
2731 struct bpf_reg_state *dst_reg)
2733 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
2734 dst_reg->umin_value);
2735 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
2736 dst_reg->umax_value);
2737 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
2738 dst_reg->smin_value);
2739 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
2740 dst_reg->smax_value);
2741 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
2743 /* We might have learned new bounds from the var_off. */
2744 __update_reg_bounds(src_reg);
2745 __update_reg_bounds(dst_reg);
2746 /* We might have learned something about the sign bit. */
2747 __reg_deduce_bounds(src_reg);
2748 __reg_deduce_bounds(dst_reg);
2749 /* We might have learned some bits from the bounds. */
2750 __reg_bound_offset(src_reg);
2751 __reg_bound_offset(dst_reg);
2752 /* Intersecting with the old var_off might have improved our bounds
2753 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2754 * then new var_off is (0; 0x7f...fc) which improves our umax.
2756 __update_reg_bounds(src_reg);
2757 __update_reg_bounds(dst_reg);
2760 static void reg_combine_min_max(struct bpf_reg_state *true_src,
2761 struct bpf_reg_state *true_dst,
2762 struct bpf_reg_state *false_src,
2763 struct bpf_reg_state *false_dst,
2768 __reg_combine_min_max(true_src, true_dst);
2771 __reg_combine_min_max(false_src, false_dst);
2776 static void mark_map_reg(struct bpf_reg_state *regs, u32 regno, u32 id,
2779 struct bpf_reg_state *reg = ®s[regno];
2781 if (reg->type == PTR_TO_MAP_VALUE_OR_NULL && reg->id == id) {
2782 /* Old offset (both fixed and variable parts) should
2783 * have been known-zero, because we don't allow pointer
2784 * arithmetic on pointers that might be NULL.
2786 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
2787 !tnum_equals_const(reg->var_off, 0) ||
2789 __mark_reg_known_zero(reg);
2793 reg->type = SCALAR_VALUE;
2794 } else if (reg->map_ptr->inner_map_meta) {
2795 reg->type = CONST_PTR_TO_MAP;
2796 reg->map_ptr = reg->map_ptr->inner_map_meta;
2798 reg->type = PTR_TO_MAP_VALUE;
2800 /* We don't need id from this point onwards anymore, thus we
2801 * should better reset it, so that state pruning has chances
2808 /* The logic is similar to find_good_pkt_pointers(), both could eventually
2809 * be folded together at some point.
2811 static void mark_map_regs(struct bpf_verifier_state *state, u32 regno,
2814 struct bpf_reg_state *regs = state->regs;
2815 u32 id = regs[regno].id;
2818 for (i = 0; i < MAX_BPF_REG; i++)
2819 mark_map_reg(regs, i, id, is_null);
2821 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
2822 if (state->stack[i].slot_type[0] != STACK_SPILL)
2824 mark_map_reg(&state->stack[i].spilled_ptr, 0, id, is_null);
2828 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
2829 struct bpf_reg_state *dst_reg,
2830 struct bpf_reg_state *src_reg,
2831 struct bpf_verifier_state *this_branch,
2832 struct bpf_verifier_state *other_branch)
2834 if (BPF_SRC(insn->code) != BPF_X)
2837 switch (BPF_OP(insn->code)) {
2839 if ((dst_reg->type == PTR_TO_PACKET &&
2840 src_reg->type == PTR_TO_PACKET_END) ||
2841 (dst_reg->type == PTR_TO_PACKET_META &&
2842 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
2843 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
2844 find_good_pkt_pointers(this_branch, dst_reg,
2845 dst_reg->type, false);
2846 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
2847 src_reg->type == PTR_TO_PACKET) ||
2848 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
2849 src_reg->type == PTR_TO_PACKET_META)) {
2850 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
2851 find_good_pkt_pointers(other_branch, src_reg,
2852 src_reg->type, true);
2858 if ((dst_reg->type == PTR_TO_PACKET &&
2859 src_reg->type == PTR_TO_PACKET_END) ||
2860 (dst_reg->type == PTR_TO_PACKET_META &&
2861 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
2862 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
2863 find_good_pkt_pointers(other_branch, dst_reg,
2864 dst_reg->type, true);
2865 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
2866 src_reg->type == PTR_TO_PACKET) ||
2867 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
2868 src_reg->type == PTR_TO_PACKET_META)) {
2869 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
2870 find_good_pkt_pointers(this_branch, src_reg,
2871 src_reg->type, false);
2877 if ((dst_reg->type == PTR_TO_PACKET &&
2878 src_reg->type == PTR_TO_PACKET_END) ||
2879 (dst_reg->type == PTR_TO_PACKET_META &&
2880 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
2881 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
2882 find_good_pkt_pointers(this_branch, dst_reg,
2883 dst_reg->type, true);
2884 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
2885 src_reg->type == PTR_TO_PACKET) ||
2886 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
2887 src_reg->type == PTR_TO_PACKET_META)) {
2888 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
2889 find_good_pkt_pointers(other_branch, src_reg,
2890 src_reg->type, false);
2896 if ((dst_reg->type == PTR_TO_PACKET &&
2897 src_reg->type == PTR_TO_PACKET_END) ||
2898 (dst_reg->type == PTR_TO_PACKET_META &&
2899 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
2900 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
2901 find_good_pkt_pointers(other_branch, dst_reg,
2902 dst_reg->type, false);
2903 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
2904 src_reg->type == PTR_TO_PACKET) ||
2905 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
2906 src_reg->type == PTR_TO_PACKET_META)) {
2907 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
2908 find_good_pkt_pointers(this_branch, src_reg,
2909 src_reg->type, true);
2921 static int check_cond_jmp_op(struct bpf_verifier_env *env,
2922 struct bpf_insn *insn, int *insn_idx)
2924 struct bpf_verifier_state *other_branch, *this_branch = env->cur_state;
2925 struct bpf_reg_state *regs = this_branch->regs, *dst_reg;
2926 u8 opcode = BPF_OP(insn->code);
2929 if (opcode > BPF_JSLE) {
2930 verbose(env, "invalid BPF_JMP opcode %x\n", opcode);
2934 if (BPF_SRC(insn->code) == BPF_X) {
2935 if (insn->imm != 0) {
2936 verbose(env, "BPF_JMP uses reserved fields\n");
2940 /* check src1 operand */
2941 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2945 if (is_pointer_value(env, insn->src_reg)) {
2946 verbose(env, "R%d pointer comparison prohibited\n",
2951 if (insn->src_reg != BPF_REG_0) {
2952 verbose(env, "BPF_JMP uses reserved fields\n");
2957 /* check src2 operand */
2958 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2962 dst_reg = ®s[insn->dst_reg];
2964 /* detect if R == 0 where R was initialized to zero earlier */
2965 if (BPF_SRC(insn->code) == BPF_K &&
2966 (opcode == BPF_JEQ || opcode == BPF_JNE) &&
2967 dst_reg->type == SCALAR_VALUE &&
2968 tnum_equals_const(dst_reg->var_off, insn->imm)) {
2969 if (opcode == BPF_JEQ) {
2970 /* if (imm == imm) goto pc+off;
2971 * only follow the goto, ignore fall-through
2973 *insn_idx += insn->off;
2976 /* if (imm != imm) goto pc+off;
2977 * only follow fall-through branch, since
2978 * that's where the program will go
2984 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx);
2988 /* detect if we are comparing against a constant value so we can adjust
2989 * our min/max values for our dst register.
2990 * this is only legit if both are scalars (or pointers to the same
2991 * object, I suppose, but we don't support that right now), because
2992 * otherwise the different base pointers mean the offsets aren't
2995 if (BPF_SRC(insn->code) == BPF_X) {
2996 if (dst_reg->type == SCALAR_VALUE &&
2997 regs[insn->src_reg].type == SCALAR_VALUE) {
2998 if (tnum_is_const(regs[insn->src_reg].var_off))
2999 reg_set_min_max(&other_branch->regs[insn->dst_reg],
3000 dst_reg, regs[insn->src_reg].var_off.value,
3002 else if (tnum_is_const(dst_reg->var_off))
3003 reg_set_min_max_inv(&other_branch->regs[insn->src_reg],
3004 ®s[insn->src_reg],
3005 dst_reg->var_off.value, opcode);
3006 else if (opcode == BPF_JEQ || opcode == BPF_JNE)
3007 /* Comparing for equality, we can combine knowledge */
3008 reg_combine_min_max(&other_branch->regs[insn->src_reg],
3009 &other_branch->regs[insn->dst_reg],
3010 ®s[insn->src_reg],
3011 ®s[insn->dst_reg], opcode);
3013 } else if (dst_reg->type == SCALAR_VALUE) {
3014 reg_set_min_max(&other_branch->regs[insn->dst_reg],
3015 dst_reg, insn->imm, opcode);
3018 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
3019 if (BPF_SRC(insn->code) == BPF_K &&
3020 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
3021 dst_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
3022 /* Mark all identical map registers in each branch as either
3023 * safe or unknown depending R == 0 or R != 0 conditional.
3025 mark_map_regs(this_branch, insn->dst_reg, opcode == BPF_JNE);
3026 mark_map_regs(other_branch, insn->dst_reg, opcode == BPF_JEQ);
3027 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
3028 this_branch, other_branch) &&
3029 is_pointer_value(env, insn->dst_reg)) {
3030 verbose(env, "R%d pointer comparison prohibited\n",
3035 print_verifier_state(env, this_branch);
3039 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
3040 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
3042 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
3044 return (struct bpf_map *) (unsigned long) imm64;
3047 /* verify BPF_LD_IMM64 instruction */
3048 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
3050 struct bpf_reg_state *regs = cur_regs(env);
3053 if (BPF_SIZE(insn->code) != BPF_DW) {
3054 verbose(env, "invalid BPF_LD_IMM insn\n");
3057 if (insn->off != 0) {
3058 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
3062 err = check_reg_arg(env, insn->dst_reg, DST_OP);
3066 if (insn->src_reg == 0) {
3067 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
3069 regs[insn->dst_reg].type = SCALAR_VALUE;
3070 __mark_reg_known(®s[insn->dst_reg], imm);
3074 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
3075 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
3077 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
3078 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
3082 static bool may_access_skb(enum bpf_prog_type type)
3085 case BPF_PROG_TYPE_SOCKET_FILTER:
3086 case BPF_PROG_TYPE_SCHED_CLS:
3087 case BPF_PROG_TYPE_SCHED_ACT:
3094 /* verify safety of LD_ABS|LD_IND instructions:
3095 * - they can only appear in the programs where ctx == skb
3096 * - since they are wrappers of function calls, they scratch R1-R5 registers,
3097 * preserve R6-R9, and store return value into R0
3100 * ctx == skb == R6 == CTX
3103 * SRC == any register
3104 * IMM == 32-bit immediate
3107 * R0 - 8/16/32-bit skb data converted to cpu endianness
3109 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
3111 struct bpf_reg_state *regs = cur_regs(env);
3112 u8 mode = BPF_MODE(insn->code);
3115 if (!may_access_skb(env->prog->type)) {
3116 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
3120 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
3121 BPF_SIZE(insn->code) == BPF_DW ||
3122 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
3123 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
3127 /* check whether implicit source operand (register R6) is readable */
3128 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
3132 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
3134 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
3138 if (mode == BPF_IND) {
3139 /* check explicit source operand */
3140 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3145 /* reset caller saved regs to unreadable */
3146 for (i = 0; i < CALLER_SAVED_REGS; i++) {
3147 mark_reg_not_init(env, regs, caller_saved[i]);
3148 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
3151 /* mark destination R0 register as readable, since it contains
3152 * the value fetched from the packet.
3153 * Already marked as written above.
3155 mark_reg_unknown(env, regs, BPF_REG_0);
3159 static int check_return_code(struct bpf_verifier_env *env)
3161 struct bpf_reg_state *reg;
3162 struct tnum range = tnum_range(0, 1);
3164 switch (env->prog->type) {
3165 case BPF_PROG_TYPE_CGROUP_SKB:
3166 case BPF_PROG_TYPE_CGROUP_SOCK:
3167 case BPF_PROG_TYPE_SOCK_OPS:
3168 case BPF_PROG_TYPE_CGROUP_DEVICE:
3174 reg = cur_regs(env) + BPF_REG_0;
3175 if (reg->type != SCALAR_VALUE) {
3176 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
3177 reg_type_str[reg->type]);
3181 if (!tnum_in(range, reg->var_off)) {
3182 verbose(env, "At program exit the register R0 ");
3183 if (!tnum_is_unknown(reg->var_off)) {
3186 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3187 verbose(env, "has value %s", tn_buf);
3189 verbose(env, "has unknown scalar value");
3191 verbose(env, " should have been 0 or 1\n");
3197 /* non-recursive DFS pseudo code
3198 * 1 procedure DFS-iterative(G,v):
3199 * 2 label v as discovered
3200 * 3 let S be a stack
3202 * 5 while S is not empty
3204 * 7 if t is what we're looking for:
3206 * 9 for all edges e in G.adjacentEdges(t) do
3207 * 10 if edge e is already labelled
3208 * 11 continue with the next edge
3209 * 12 w <- G.adjacentVertex(t,e)
3210 * 13 if vertex w is not discovered and not explored
3211 * 14 label e as tree-edge
3212 * 15 label w as discovered
3215 * 18 else if vertex w is discovered
3216 * 19 label e as back-edge
3218 * 21 // vertex w is explored
3219 * 22 label e as forward- or cross-edge
3220 * 23 label t as explored
3225 * 0x11 - discovered and fall-through edge labelled
3226 * 0x12 - discovered and fall-through and branch edges labelled
3237 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
3239 static int *insn_stack; /* stack of insns to process */
3240 static int cur_stack; /* current stack index */
3241 static int *insn_state;
3243 /* t, w, e - match pseudo-code above:
3244 * t - index of current instruction
3245 * w - next instruction
3248 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
3250 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
3253 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
3256 if (w < 0 || w >= env->prog->len) {
3257 verbose(env, "jump out of range from insn %d to %d\n", t, w);
3262 /* mark branch target for state pruning */
3263 env->explored_states[w] = STATE_LIST_MARK;
3265 if (insn_state[w] == 0) {
3267 insn_state[t] = DISCOVERED | e;
3268 insn_state[w] = DISCOVERED;
3269 if (cur_stack >= env->prog->len)
3271 insn_stack[cur_stack++] = w;
3273 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
3274 verbose(env, "back-edge from insn %d to %d\n", t, w);
3276 } else if (insn_state[w] == EXPLORED) {
3277 /* forward- or cross-edge */
3278 insn_state[t] = DISCOVERED | e;
3280 verbose(env, "insn state internal bug\n");
3286 /* non-recursive depth-first-search to detect loops in BPF program
3287 * loop == back-edge in directed graph
3289 static int check_cfg(struct bpf_verifier_env *env)
3291 struct bpf_insn *insns = env->prog->insnsi;
3292 int insn_cnt = env->prog->len;
3296 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
3300 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
3306 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
3307 insn_stack[0] = 0; /* 0 is the first instruction */
3313 t = insn_stack[cur_stack - 1];
3315 if (BPF_CLASS(insns[t].code) == BPF_JMP) {
3316 u8 opcode = BPF_OP(insns[t].code);
3318 if (opcode == BPF_EXIT) {
3320 } else if (opcode == BPF_CALL) {
3321 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3326 if (t + 1 < insn_cnt)
3327 env->explored_states[t + 1] = STATE_LIST_MARK;
3328 } else if (opcode == BPF_JA) {
3329 if (BPF_SRC(insns[t].code) != BPF_K) {
3333 /* unconditional jump with single edge */
3334 ret = push_insn(t, t + insns[t].off + 1,
3340 /* tell verifier to check for equivalent states
3341 * after every call and jump
3343 if (t + 1 < insn_cnt)
3344 env->explored_states[t + 1] = STATE_LIST_MARK;
3346 /* conditional jump with two edges */
3347 env->explored_states[t] = STATE_LIST_MARK;
3348 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3354 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
3361 /* all other non-branch instructions with single
3364 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3372 insn_state[t] = EXPLORED;
3373 if (cur_stack-- <= 0) {
3374 verbose(env, "pop stack internal bug\n");
3381 for (i = 0; i < insn_cnt; i++) {
3382 if (insn_state[i] != EXPLORED) {
3383 verbose(env, "unreachable insn %d\n", i);
3388 ret = 0; /* cfg looks good */
3396 /* check %cur's range satisfies %old's */
3397 static bool range_within(struct bpf_reg_state *old,
3398 struct bpf_reg_state *cur)
3400 return old->umin_value <= cur->umin_value &&
3401 old->umax_value >= cur->umax_value &&
3402 old->smin_value <= cur->smin_value &&
3403 old->smax_value >= cur->smax_value;
3406 /* Maximum number of register states that can exist at once */
3407 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
3413 /* If in the old state two registers had the same id, then they need to have
3414 * the same id in the new state as well. But that id could be different from
3415 * the old state, so we need to track the mapping from old to new ids.
3416 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
3417 * regs with old id 5 must also have new id 9 for the new state to be safe. But
3418 * regs with a different old id could still have new id 9, we don't care about
3420 * So we look through our idmap to see if this old id has been seen before. If
3421 * so, we require the new id to match; otherwise, we add the id pair to the map.
3423 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
3427 for (i = 0; i < ID_MAP_SIZE; i++) {
3428 if (!idmap[i].old) {
3429 /* Reached an empty slot; haven't seen this id before */
3430 idmap[i].old = old_id;
3431 idmap[i].cur = cur_id;
3434 if (idmap[i].old == old_id)
3435 return idmap[i].cur == cur_id;
3437 /* We ran out of idmap slots, which should be impossible */
3442 /* Returns true if (rold safe implies rcur safe) */
3443 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
3444 struct idpair *idmap)
3446 if (!(rold->live & REG_LIVE_READ))
3447 /* explored state didn't use this */
3450 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, live)) == 0)
3453 if (rold->type == NOT_INIT)
3454 /* explored state can't have used this */
3456 if (rcur->type == NOT_INIT)
3458 switch (rold->type) {
3460 if (rcur->type == SCALAR_VALUE) {
3461 /* new val must satisfy old val knowledge */
3462 return range_within(rold, rcur) &&
3463 tnum_in(rold->var_off, rcur->var_off);
3465 /* if we knew anything about the old value, we're not
3466 * equal, because we can't know anything about the
3467 * scalar value of the pointer in the new value.
3469 return rold->umin_value == 0 &&
3470 rold->umax_value == U64_MAX &&
3471 rold->smin_value == S64_MIN &&
3472 rold->smax_value == S64_MAX &&
3473 tnum_is_unknown(rold->var_off);
3475 case PTR_TO_MAP_VALUE:
3476 /* If the new min/max/var_off satisfy the old ones and
3477 * everything else matches, we are OK.
3478 * We don't care about the 'id' value, because nothing
3479 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
3481 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
3482 range_within(rold, rcur) &&
3483 tnum_in(rold->var_off, rcur->var_off);
3484 case PTR_TO_MAP_VALUE_OR_NULL:
3485 /* a PTR_TO_MAP_VALUE could be safe to use as a
3486 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
3487 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
3488 * checked, doing so could have affected others with the same
3489 * id, and we can't check for that because we lost the id when
3490 * we converted to a PTR_TO_MAP_VALUE.
3492 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
3494 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
3496 /* Check our ids match any regs they're supposed to */
3497 return check_ids(rold->id, rcur->id, idmap);
3498 case PTR_TO_PACKET_META:
3500 if (rcur->type != rold->type)
3502 /* We must have at least as much range as the old ptr
3503 * did, so that any accesses which were safe before are
3504 * still safe. This is true even if old range < old off,
3505 * since someone could have accessed through (ptr - k), or
3506 * even done ptr -= k in a register, to get a safe access.
3508 if (rold->range > rcur->range)
3510 /* If the offsets don't match, we can't trust our alignment;
3511 * nor can we be sure that we won't fall out of range.
3513 if (rold->off != rcur->off)
3515 /* id relations must be preserved */
3516 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
3518 /* new val must satisfy old val knowledge */
3519 return range_within(rold, rcur) &&
3520 tnum_in(rold->var_off, rcur->var_off);
3522 case CONST_PTR_TO_MAP:
3524 case PTR_TO_PACKET_END:
3525 /* Only valid matches are exact, which memcmp() above
3526 * would have accepted
3529 /* Don't know what's going on, just say it's not safe */
3533 /* Shouldn't get here; if we do, say it's not safe */
3538 static bool stacksafe(struct bpf_verifier_state *old,
3539 struct bpf_verifier_state *cur,
3540 struct idpair *idmap)
3544 /* if explored stack has more populated slots than current stack
3545 * such stacks are not equivalent
3547 if (old->allocated_stack > cur->allocated_stack)
3550 /* walk slots of the explored stack and ignore any additional
3551 * slots in the current stack, since explored(safe) state
3554 for (i = 0; i < old->allocated_stack; i++) {
3555 spi = i / BPF_REG_SIZE;
3557 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
3559 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
3560 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
3561 /* Ex: old explored (safe) state has STACK_SPILL in
3562 * this stack slot, but current has has STACK_MISC ->
3563 * this verifier states are not equivalent,
3564 * return false to continue verification of this path
3567 if (i % BPF_REG_SIZE)
3569 if (old->stack[spi].slot_type[0] != STACK_SPILL)
3571 if (!regsafe(&old->stack[spi].spilled_ptr,
3572 &cur->stack[spi].spilled_ptr,
3574 /* when explored and current stack slot are both storing
3575 * spilled registers, check that stored pointers types
3576 * are the same as well.
3577 * Ex: explored safe path could have stored
3578 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
3579 * but current path has stored:
3580 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
3581 * such verifier states are not equivalent.
3582 * return false to continue verification of this path
3589 /* compare two verifier states
3591 * all states stored in state_list are known to be valid, since
3592 * verifier reached 'bpf_exit' instruction through them
3594 * this function is called when verifier exploring different branches of
3595 * execution popped from the state stack. If it sees an old state that has
3596 * more strict register state and more strict stack state then this execution
3597 * branch doesn't need to be explored further, since verifier already
3598 * concluded that more strict state leads to valid finish.
3600 * Therefore two states are equivalent if register state is more conservative
3601 * and explored stack state is more conservative than the current one.
3604 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
3605 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
3607 * In other words if current stack state (one being explored) has more
3608 * valid slots than old one that already passed validation, it means
3609 * the verifier can stop exploring and conclude that current state is valid too
3611 * Similarly with registers. If explored state has register type as invalid
3612 * whereas register type in current state is meaningful, it means that
3613 * the current state will reach 'bpf_exit' instruction safely
3615 static bool states_equal(struct bpf_verifier_env *env,
3616 struct bpf_verifier_state *old,
3617 struct bpf_verifier_state *cur)
3619 struct idpair *idmap;
3623 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
3624 /* If we failed to allocate the idmap, just say it's not safe */
3628 for (i = 0; i < MAX_BPF_REG; i++) {
3629 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
3633 if (!stacksafe(old, cur, idmap))
3641 /* A write screens off any subsequent reads; but write marks come from the
3642 * straight-line code between a state and its parent. When we arrive at a
3643 * jump target (in the first iteration of the propagate_liveness() loop),
3644 * we didn't arrive by the straight-line code, so read marks in state must
3645 * propagate to parent regardless of state's write marks.
3647 static bool do_propagate_liveness(const struct bpf_verifier_state *state,
3648 struct bpf_verifier_state *parent)
3650 bool writes = parent == state->parent; /* Observe write marks */
3651 bool touched = false; /* any changes made? */
3656 /* Propagate read liveness of registers... */
3657 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
3658 /* We don't need to worry about FP liveness because it's read-only */
3659 for (i = 0; i < BPF_REG_FP; i++) {
3660 if (parent->regs[i].live & REG_LIVE_READ)
3662 if (writes && (state->regs[i].live & REG_LIVE_WRITTEN))
3664 if (state->regs[i].live & REG_LIVE_READ) {
3665 parent->regs[i].live |= REG_LIVE_READ;
3669 /* ... and stack slots */
3670 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
3671 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
3672 if (parent->stack[i].slot_type[0] != STACK_SPILL)
3674 if (state->stack[i].slot_type[0] != STACK_SPILL)
3676 if (parent->stack[i].spilled_ptr.live & REG_LIVE_READ)
3679 (state->stack[i].spilled_ptr.live & REG_LIVE_WRITTEN))
3681 if (state->stack[i].spilled_ptr.live & REG_LIVE_READ) {
3682 parent->stack[i].spilled_ptr.live |= REG_LIVE_READ;
3689 /* "parent" is "a state from which we reach the current state", but initially
3690 * it is not the state->parent (i.e. "the state whose straight-line code leads
3691 * to the current state"), instead it is the state that happened to arrive at
3692 * a (prunable) equivalent of the current state. See comment above
3693 * do_propagate_liveness() for consequences of this.
3694 * This function is just a more efficient way of calling mark_reg_read() or
3695 * mark_stack_slot_read() on each reg in "parent" that is read in "state",
3696 * though it requires that parent != state->parent in the call arguments.
3698 static void propagate_liveness(const struct bpf_verifier_state *state,
3699 struct bpf_verifier_state *parent)
3701 while (do_propagate_liveness(state, parent)) {
3702 /* Something changed, so we need to feed those changes onward */
3704 parent = state->parent;
3708 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
3710 struct bpf_verifier_state_list *new_sl;
3711 struct bpf_verifier_state_list *sl;
3712 struct bpf_verifier_state *cur = env->cur_state;
3715 sl = env->explored_states[insn_idx];
3717 /* this 'insn_idx' instruction wasn't marked, so we will not
3718 * be doing state search here
3722 while (sl != STATE_LIST_MARK) {
3723 if (states_equal(env, &sl->state, cur)) {
3724 /* reached equivalent register/stack state,
3726 * Registers read by the continuation are read by us.
3727 * If we have any write marks in env->cur_state, they
3728 * will prevent corresponding reads in the continuation
3729 * from reaching our parent (an explored_state). Our
3730 * own state will get the read marks recorded, but
3731 * they'll be immediately forgotten as we're pruning
3732 * this state and will pop a new one.
3734 propagate_liveness(&sl->state, cur);
3740 /* there were no equivalent states, remember current one.
3741 * technically the current state is not proven to be safe yet,
3742 * but it will either reach bpf_exit (which means it's safe) or
3743 * it will be rejected. Since there are no loops, we won't be
3744 * seeing this 'insn_idx' instruction again on the way to bpf_exit
3746 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
3750 /* add new state to the head of linked list */
3751 err = copy_verifier_state(&new_sl->state, cur);
3753 free_verifier_state(&new_sl->state, false);
3757 new_sl->next = env->explored_states[insn_idx];
3758 env->explored_states[insn_idx] = new_sl;
3759 /* connect new state to parentage chain */
3760 cur->parent = &new_sl->state;
3761 /* clear write marks in current state: the writes we did are not writes
3762 * our child did, so they don't screen off its reads from us.
3763 * (There are no read marks in current state, because reads always mark
3764 * their parent and current state never has children yet. Only
3765 * explored_states can get read marks.)
3767 for (i = 0; i < BPF_REG_FP; i++)
3768 cur->regs[i].live = REG_LIVE_NONE;
3769 for (i = 0; i < cur->allocated_stack / BPF_REG_SIZE; i++)
3770 if (cur->stack[i].slot_type[0] == STACK_SPILL)
3771 cur->stack[i].spilled_ptr.live = REG_LIVE_NONE;
3775 static int ext_analyzer_insn_hook(struct bpf_verifier_env *env,
3776 int insn_idx, int prev_insn_idx)
3778 if (env->dev_ops && env->dev_ops->insn_hook)
3779 return env->dev_ops->insn_hook(env, insn_idx, prev_insn_idx);
3784 static int do_check(struct bpf_verifier_env *env)
3786 struct bpf_verifier_state *state;
3787 struct bpf_insn *insns = env->prog->insnsi;
3788 struct bpf_reg_state *regs;
3789 int insn_cnt = env->prog->len;
3790 int insn_idx, prev_insn_idx = 0;
3791 int insn_processed = 0;
3792 bool do_print_state = false;
3794 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
3797 env->cur_state = state;
3798 init_reg_state(env, state->regs);
3799 state->parent = NULL;
3802 struct bpf_insn *insn;
3806 if (insn_idx >= insn_cnt) {
3807 verbose(env, "invalid insn idx %d insn_cnt %d\n",
3808 insn_idx, insn_cnt);
3812 insn = &insns[insn_idx];
3813 class = BPF_CLASS(insn->code);
3815 if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
3817 "BPF program is too large. Processed %d insn\n",
3822 err = is_state_visited(env, insn_idx);
3826 /* found equivalent state, can prune the search */
3827 if (env->log.level) {
3829 verbose(env, "\nfrom %d to %d: safe\n",
3830 prev_insn_idx, insn_idx);
3832 verbose(env, "%d: safe\n", insn_idx);
3834 goto process_bpf_exit;
3840 if (env->log.level > 1 || (env->log.level && do_print_state)) {
3841 if (env->log.level > 1)
3842 verbose(env, "%d:", insn_idx);
3844 verbose(env, "\nfrom %d to %d:",
3845 prev_insn_idx, insn_idx);
3846 print_verifier_state(env, state);
3847 do_print_state = false;
3850 if (env->log.level) {
3851 verbose(env, "%d: ", insn_idx);
3852 print_bpf_insn(verbose, env, insn,
3853 env->allow_ptr_leaks);
3856 err = ext_analyzer_insn_hook(env, insn_idx, prev_insn_idx);
3860 regs = cur_regs(env);
3861 env->insn_aux_data[insn_idx].seen = true;
3862 if (class == BPF_ALU || class == BPF_ALU64) {
3863 err = check_alu_op(env, insn);
3867 } else if (class == BPF_LDX) {
3868 enum bpf_reg_type *prev_src_type, src_reg_type;
3870 /* check for reserved fields is already done */
3872 /* check src operand */
3873 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3877 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3881 src_reg_type = regs[insn->src_reg].type;
3883 /* check that memory (src_reg + off) is readable,
3884 * the state of dst_reg will be updated by this func
3886 err = check_mem_access(env, insn_idx, insn->src_reg, insn->off,
3887 BPF_SIZE(insn->code), BPF_READ,
3892 prev_src_type = &env->insn_aux_data[insn_idx].ptr_type;
3894 if (*prev_src_type == NOT_INIT) {
3896 * dst_reg = *(u32 *)(src_reg + off)
3897 * save type to validate intersecting paths
3899 *prev_src_type = src_reg_type;
3901 } else if (src_reg_type != *prev_src_type &&
3902 (src_reg_type == PTR_TO_CTX ||
3903 *prev_src_type == PTR_TO_CTX)) {
3904 /* ABuser program is trying to use the same insn
3905 * dst_reg = *(u32*) (src_reg + off)
3906 * with different pointer types:
3907 * src_reg == ctx in one branch and
3908 * src_reg == stack|map in some other branch.
3911 verbose(env, "same insn cannot be used with different pointers\n");
3915 } else if (class == BPF_STX) {
3916 enum bpf_reg_type *prev_dst_type, dst_reg_type;
3918 if (BPF_MODE(insn->code) == BPF_XADD) {
3919 err = check_xadd(env, insn_idx, insn);
3926 /* check src1 operand */
3927 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3930 /* check src2 operand */
3931 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3935 dst_reg_type = regs[insn->dst_reg].type;
3937 /* check that memory (dst_reg + off) is writeable */
3938 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3939 BPF_SIZE(insn->code), BPF_WRITE,
3944 prev_dst_type = &env->insn_aux_data[insn_idx].ptr_type;
3946 if (*prev_dst_type == NOT_INIT) {
3947 *prev_dst_type = dst_reg_type;
3948 } else if (dst_reg_type != *prev_dst_type &&
3949 (dst_reg_type == PTR_TO_CTX ||
3950 *prev_dst_type == PTR_TO_CTX)) {
3951 verbose(env, "same insn cannot be used with different pointers\n");
3955 } else if (class == BPF_ST) {
3956 if (BPF_MODE(insn->code) != BPF_MEM ||
3957 insn->src_reg != BPF_REG_0) {
3958 verbose(env, "BPF_ST uses reserved fields\n");
3961 /* check src operand */
3962 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3966 /* check that memory (dst_reg + off) is writeable */
3967 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3968 BPF_SIZE(insn->code), BPF_WRITE,
3973 } else if (class == BPF_JMP) {
3974 u8 opcode = BPF_OP(insn->code);
3976 if (opcode == BPF_CALL) {
3977 if (BPF_SRC(insn->code) != BPF_K ||
3979 insn->src_reg != BPF_REG_0 ||
3980 insn->dst_reg != BPF_REG_0) {
3981 verbose(env, "BPF_CALL uses reserved fields\n");
3985 err = check_call(env, insn->imm, insn_idx);
3989 } else if (opcode == BPF_JA) {
3990 if (BPF_SRC(insn->code) != BPF_K ||
3992 insn->src_reg != BPF_REG_0 ||
3993 insn->dst_reg != BPF_REG_0) {
3994 verbose(env, "BPF_JA uses reserved fields\n");
3998 insn_idx += insn->off + 1;
4001 } else if (opcode == BPF_EXIT) {
4002 if (BPF_SRC(insn->code) != BPF_K ||
4004 insn->src_reg != BPF_REG_0 ||
4005 insn->dst_reg != BPF_REG_0) {
4006 verbose(env, "BPF_EXIT uses reserved fields\n");
4010 /* eBPF calling convetion is such that R0 is used
4011 * to return the value from eBPF program.
4012 * Make sure that it's readable at this time
4013 * of bpf_exit, which means that program wrote
4014 * something into it earlier
4016 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
4020 if (is_pointer_value(env, BPF_REG_0)) {
4021 verbose(env, "R0 leaks addr as return value\n");
4025 err = check_return_code(env);
4029 err = pop_stack(env, &prev_insn_idx, &insn_idx);
4035 do_print_state = true;
4039 err = check_cond_jmp_op(env, insn, &insn_idx);
4043 } else if (class == BPF_LD) {
4044 u8 mode = BPF_MODE(insn->code);
4046 if (mode == BPF_ABS || mode == BPF_IND) {
4047 err = check_ld_abs(env, insn);
4051 } else if (mode == BPF_IMM) {
4052 err = check_ld_imm(env, insn);
4057 env->insn_aux_data[insn_idx].seen = true;
4059 verbose(env, "invalid BPF_LD mode\n");
4063 verbose(env, "unknown insn class %d\n", class);
4070 verbose(env, "processed %d insns, stack depth %d\n", insn_processed,
4071 env->prog->aux->stack_depth);
4075 static int check_map_prealloc(struct bpf_map *map)
4077 return (map->map_type != BPF_MAP_TYPE_HASH &&
4078 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
4079 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
4080 !(map->map_flags & BPF_F_NO_PREALLOC);
4083 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
4084 struct bpf_map *map,
4085 struct bpf_prog *prog)
4088 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
4089 * preallocated hash maps, since doing memory allocation
4090 * in overflow_handler can crash depending on where nmi got
4093 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
4094 if (!check_map_prealloc(map)) {
4095 verbose(env, "perf_event programs can only use preallocated hash map\n");
4098 if (map->inner_map_meta &&
4099 !check_map_prealloc(map->inner_map_meta)) {
4100 verbose(env, "perf_event programs can only use preallocated inner hash map\n");
4107 /* look for pseudo eBPF instructions that access map FDs and
4108 * replace them with actual map pointers
4110 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
4112 struct bpf_insn *insn = env->prog->insnsi;
4113 int insn_cnt = env->prog->len;
4116 err = bpf_prog_calc_tag(env->prog);
4120 for (i = 0; i < insn_cnt; i++, insn++) {
4121 if (BPF_CLASS(insn->code) == BPF_LDX &&
4122 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
4123 verbose(env, "BPF_LDX uses reserved fields\n");
4127 if (BPF_CLASS(insn->code) == BPF_STX &&
4128 ((BPF_MODE(insn->code) != BPF_MEM &&
4129 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
4130 verbose(env, "BPF_STX uses reserved fields\n");
4134 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
4135 struct bpf_map *map;
4138 if (i == insn_cnt - 1 || insn[1].code != 0 ||
4139 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
4141 verbose(env, "invalid bpf_ld_imm64 insn\n");
4145 if (insn->src_reg == 0)
4146 /* valid generic load 64-bit imm */
4149 if (insn->src_reg != BPF_PSEUDO_MAP_FD) {
4151 "unrecognized bpf_ld_imm64 insn\n");
4155 f = fdget(insn->imm);
4156 map = __bpf_map_get(f);
4158 verbose(env, "fd %d is not pointing to valid bpf_map\n",
4160 return PTR_ERR(map);
4163 err = check_map_prog_compatibility(env, map, env->prog);
4169 /* store map pointer inside BPF_LD_IMM64 instruction */
4170 insn[0].imm = (u32) (unsigned long) map;
4171 insn[1].imm = ((u64) (unsigned long) map) >> 32;
4173 /* check whether we recorded this map already */
4174 for (j = 0; j < env->used_map_cnt; j++)
4175 if (env->used_maps[j] == map) {
4180 if (env->used_map_cnt >= MAX_USED_MAPS) {
4185 /* hold the map. If the program is rejected by verifier,
4186 * the map will be released by release_maps() or it
4187 * will be used by the valid program until it's unloaded
4188 * and all maps are released in free_bpf_prog_info()
4190 map = bpf_map_inc(map, false);
4193 return PTR_ERR(map);
4195 env->used_maps[env->used_map_cnt++] = map;
4204 /* now all pseudo BPF_LD_IMM64 instructions load valid
4205 * 'struct bpf_map *' into a register instead of user map_fd.
4206 * These pointers will be used later by verifier to validate map access.
4211 /* drop refcnt of maps used by the rejected program */
4212 static void release_maps(struct bpf_verifier_env *env)
4216 for (i = 0; i < env->used_map_cnt; i++)
4217 bpf_map_put(env->used_maps[i]);
4220 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
4221 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
4223 struct bpf_insn *insn = env->prog->insnsi;
4224 int insn_cnt = env->prog->len;
4227 for (i = 0; i < insn_cnt; i++, insn++)
4228 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
4232 /* single env->prog->insni[off] instruction was replaced with the range
4233 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
4234 * [0, off) and [off, end) to new locations, so the patched range stays zero
4236 static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len,
4239 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
4244 new_data = vzalloc(sizeof(struct bpf_insn_aux_data) * prog_len);
4247 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
4248 memcpy(new_data + off + cnt - 1, old_data + off,
4249 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
4250 for (i = off; i < off + cnt - 1; i++)
4251 new_data[i].seen = true;
4252 env->insn_aux_data = new_data;
4257 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
4258 const struct bpf_insn *patch, u32 len)
4260 struct bpf_prog *new_prog;
4262 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
4265 if (adjust_insn_aux_data(env, new_prog->len, off, len))
4270 /* The verifier does more data flow analysis than llvm and will not explore
4271 * branches that are dead at run time. Malicious programs can have dead code
4272 * too. Therefore replace all dead at-run-time code with nops.
4274 static void sanitize_dead_code(struct bpf_verifier_env *env)
4276 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
4277 struct bpf_insn nop = BPF_MOV64_REG(BPF_REG_0, BPF_REG_0);
4278 struct bpf_insn *insn = env->prog->insnsi;
4279 const int insn_cnt = env->prog->len;
4282 for (i = 0; i < insn_cnt; i++) {
4283 if (aux_data[i].seen)
4285 memcpy(insn + i, &nop, sizeof(nop));
4289 /* convert load instructions that access fields of 'struct __sk_buff'
4290 * into sequence of instructions that access fields of 'struct sk_buff'
4292 static int convert_ctx_accesses(struct bpf_verifier_env *env)
4294 const struct bpf_verifier_ops *ops = env->ops;
4295 int i, cnt, size, ctx_field_size, delta = 0;
4296 const int insn_cnt = env->prog->len;
4297 struct bpf_insn insn_buf[16], *insn;
4298 struct bpf_prog *new_prog;
4299 enum bpf_access_type type;
4300 bool is_narrower_load;
4303 if (ops->gen_prologue) {
4304 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
4306 if (cnt >= ARRAY_SIZE(insn_buf)) {
4307 verbose(env, "bpf verifier is misconfigured\n");
4310 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
4314 env->prog = new_prog;
4319 if (!ops->convert_ctx_access)
4322 insn = env->prog->insnsi + delta;
4324 for (i = 0; i < insn_cnt; i++, insn++) {
4325 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
4326 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
4327 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
4328 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
4330 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
4331 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
4332 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
4333 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
4338 if (env->insn_aux_data[i + delta].ptr_type != PTR_TO_CTX)
4341 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
4342 size = BPF_LDST_BYTES(insn);
4344 /* If the read access is a narrower load of the field,
4345 * convert to a 4/8-byte load, to minimum program type specific
4346 * convert_ctx_access changes. If conversion is successful,
4347 * we will apply proper mask to the result.
4349 is_narrower_load = size < ctx_field_size;
4350 if (is_narrower_load) {
4351 u32 off = insn->off;
4354 if (type == BPF_WRITE) {
4355 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
4360 if (ctx_field_size == 4)
4362 else if (ctx_field_size == 8)
4365 insn->off = off & ~(ctx_field_size - 1);
4366 insn->code = BPF_LDX | BPF_MEM | size_code;
4370 cnt = ops->convert_ctx_access(type, insn, insn_buf, env->prog,
4372 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
4373 (ctx_field_size && !target_size)) {
4374 verbose(env, "bpf verifier is misconfigured\n");
4378 if (is_narrower_load && size < target_size) {
4379 if (ctx_field_size <= 4)
4380 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
4381 (1 << size * 8) - 1);
4383 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
4384 (1 << size * 8) - 1);
4387 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
4393 /* keep walking new program and skip insns we just inserted */
4394 env->prog = new_prog;
4395 insn = new_prog->insnsi + i + delta;
4401 /* fixup insn->imm field of bpf_call instructions
4402 * and inline eligible helpers as explicit sequence of BPF instructions
4404 * this function is called after eBPF program passed verification
4406 static int fixup_bpf_calls(struct bpf_verifier_env *env)
4408 struct bpf_prog *prog = env->prog;
4409 struct bpf_insn *insn = prog->insnsi;
4410 const struct bpf_func_proto *fn;
4411 const int insn_cnt = prog->len;
4412 struct bpf_insn insn_buf[16];
4413 struct bpf_prog *new_prog;
4414 struct bpf_map *map_ptr;
4415 int i, cnt, delta = 0;
4417 for (i = 0; i < insn_cnt; i++, insn++) {
4418 if (insn->code != (BPF_JMP | BPF_CALL))
4421 if (insn->imm == BPF_FUNC_get_route_realm)
4422 prog->dst_needed = 1;
4423 if (insn->imm == BPF_FUNC_get_prandom_u32)
4424 bpf_user_rnd_init_once();
4425 if (insn->imm == BPF_FUNC_tail_call) {
4426 /* If we tail call into other programs, we
4427 * cannot make any assumptions since they can
4428 * be replaced dynamically during runtime in
4429 * the program array.
4431 prog->cb_access = 1;
4432 env->prog->aux->stack_depth = MAX_BPF_STACK;
4434 /* mark bpf_tail_call as different opcode to avoid
4435 * conditional branch in the interpeter for every normal
4436 * call and to prevent accidental JITing by JIT compiler
4437 * that doesn't support bpf_tail_call yet
4440 insn->code = BPF_JMP | BPF_TAIL_CALL;
4444 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
4445 * handlers are currently limited to 64 bit only.
4447 if (ebpf_jit_enabled() && BITS_PER_LONG == 64 &&
4448 insn->imm == BPF_FUNC_map_lookup_elem) {
4449 map_ptr = env->insn_aux_data[i + delta].map_ptr;
4450 if (map_ptr == BPF_MAP_PTR_POISON ||
4451 !map_ptr->ops->map_gen_lookup)
4452 goto patch_call_imm;
4454 cnt = map_ptr->ops->map_gen_lookup(map_ptr, insn_buf);
4455 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
4456 verbose(env, "bpf verifier is misconfigured\n");
4460 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
4467 /* keep walking new program and skip insns we just inserted */
4468 env->prog = prog = new_prog;
4469 insn = new_prog->insnsi + i + delta;
4473 if (insn->imm == BPF_FUNC_redirect_map) {
4474 /* Note, we cannot use prog directly as imm as subsequent
4475 * rewrites would still change the prog pointer. The only
4476 * stable address we can use is aux, which also works with
4477 * prog clones during blinding.
4479 u64 addr = (unsigned long)prog->aux;
4480 struct bpf_insn r4_ld[] = {
4481 BPF_LD_IMM64(BPF_REG_4, addr),
4484 cnt = ARRAY_SIZE(r4_ld);
4486 new_prog = bpf_patch_insn_data(env, i + delta, r4_ld, cnt);
4491 env->prog = prog = new_prog;
4492 insn = new_prog->insnsi + i + delta;
4495 fn = env->ops->get_func_proto(insn->imm);
4496 /* all functions that have prototype and verifier allowed
4497 * programs to call them, must be real in-kernel functions
4501 "kernel subsystem misconfigured func %s#%d\n",
4502 func_id_name(insn->imm), insn->imm);
4505 insn->imm = fn->func - __bpf_call_base;
4511 static void free_states(struct bpf_verifier_env *env)
4513 struct bpf_verifier_state_list *sl, *sln;
4516 if (!env->explored_states)
4519 for (i = 0; i < env->prog->len; i++) {
4520 sl = env->explored_states[i];
4523 while (sl != STATE_LIST_MARK) {
4525 free_verifier_state(&sl->state, false);
4531 kfree(env->explored_states);
4534 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr)
4536 struct bpf_verifier_env *env;
4537 struct bpf_verifer_log *log;
4540 /* no program is valid */
4541 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
4544 /* 'struct bpf_verifier_env' can be global, but since it's not small,
4545 * allocate/free it every time bpf_check() is called
4547 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
4552 env->insn_aux_data = vzalloc(sizeof(struct bpf_insn_aux_data) *
4555 if (!env->insn_aux_data)
4558 env->ops = bpf_verifier_ops[env->prog->type];
4560 /* grab the mutex to protect few globals used by verifier */
4561 mutex_lock(&bpf_verifier_lock);
4563 if (attr->log_level || attr->log_buf || attr->log_size) {
4564 /* user requested verbose verifier output
4565 * and supplied buffer to store the verification trace
4567 log->level = attr->log_level;
4568 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
4569 log->len_total = attr->log_size;
4572 /* log attributes have to be sane */
4573 if (log->len_total < 128 || log->len_total > UINT_MAX >> 8 ||
4574 !log->level || !log->ubuf)
4578 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
4579 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
4580 env->strict_alignment = true;
4582 if (env->prog->aux->offload) {
4583 ret = bpf_prog_offload_verifier_prep(env);
4588 ret = replace_map_fd_with_map_ptr(env);
4590 goto skip_full_check;
4592 env->explored_states = kcalloc(env->prog->len,
4593 sizeof(struct bpf_verifier_state_list *),
4596 if (!env->explored_states)
4597 goto skip_full_check;
4599 ret = check_cfg(env);
4601 goto skip_full_check;
4603 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
4605 ret = do_check(env);
4606 if (env->cur_state) {
4607 free_verifier_state(env->cur_state, true);
4608 env->cur_state = NULL;
4612 while (!pop_stack(env, NULL, NULL));
4616 sanitize_dead_code(env);
4619 /* program is valid, convert *(u32*)(ctx + off) accesses */
4620 ret = convert_ctx_accesses(env);
4623 ret = fixup_bpf_calls(env);
4625 if (log->level && bpf_verifier_log_full(log))
4627 if (log->level && !log->ubuf) {
4629 goto err_release_maps;
4632 if (ret == 0 && env->used_map_cnt) {
4633 /* if program passed verifier, update used_maps in bpf_prog_info */
4634 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
4635 sizeof(env->used_maps[0]),
4638 if (!env->prog->aux->used_maps) {
4640 goto err_release_maps;
4643 memcpy(env->prog->aux->used_maps, env->used_maps,
4644 sizeof(env->used_maps[0]) * env->used_map_cnt);
4645 env->prog->aux->used_map_cnt = env->used_map_cnt;
4647 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
4648 * bpf_ld_imm64 instructions
4650 convert_pseudo_ld_imm64(env);
4654 if (!env->prog->aux->used_maps)
4655 /* if we didn't copy map pointers into bpf_prog_info, release
4656 * them now. Otherwise free_bpf_prog_info() will release them.
4661 mutex_unlock(&bpf_verifier_lock);
4662 vfree(env->insn_aux_data);